Power Systems

Morteza Nazari-Heris · Somayeh Asadi

Behnam Mohammadi-Ivatloo
Moloud Abdar · Houtan Jebelli
Milad Sadat-Mohammadi Editors

Application
of Machine Learning
and Deep Learning
Methods to Power
System Problems
Power Systems
Electrical power has been the technological foundation of industrial societies for
many years. Although the systems designed to provide and apply electrical energy
have reached a high degree of maturity, unforeseen problems are constantly encoun-
tered, necessitating the design of more efficient and reliable systems based on novel
technologies. The book series Power Systems is aimed at providing detailed, accu-
rate and sound technical information about these new developments in electrical
power engineering. It includes topics on power generation, storage and transmission
as well as electrical machines. The monographs and advanced textbooks in this
series address researchers, lecturers, industrial engineers and senior students in
electrical engineering.
**Power Systems is indexed in Scopus**

More information about this series at http://www.springer.com/series/4622

Morteza Nazari-Heris • Somayeh Asadi
Behnam Mohammadi-Ivatloo • Moloud Abdar
Houtan Jebelli • Milad Sadat-Mohammadi
Editors

Application of Machine
Learning and Deep Learning
Methods to Power System
Problems
Editors
Morteza Nazari-Heris Somayeh Asadi
Department of Architectural Engineering Department of Architectural Engineering
Pennsylvania State University Pennsylvania State University
University Park, PA, USA University Park, PA, USA

Behnam Mohammadi-Ivatloo Moloud Abdar

Department of Energy Technology Deakin University
Aalborg University Geelong, VIC, Australia
Aalborg, Denmark
Faculty of Electrical
and Computer Engineering
University of Tabriz
Tabriz, Iran

Houtan Jebelli Milad Sadat-Mohammadi

Department of Architectural Engineering Department of Architectural Engineering
Pennsylvania State University Pennsylvania State University
University Park, PA, USA University Park, PA, USA

ISSN 1612-1287 ISSN 1860-4676 (electronic)

Power Systems
ISBN 978-3-030-77695-4 ISBN 978-3-030-77696-1 (eBook)
https://doi.org/10.1007/978-3-030-77696-1

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by
similar or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

Considering the modern power systems and their developments in terms of smart
operation, high rate of integration of renewable energy sources and emerging
technologies in power systems, the importance of the efficiency and quality indexes
of power systems, data prediction and energy liberalization, the operation, plan, and
control of power networks should be investigated. Additionally, new issues have
been observed in the power industry considering the growing rate of load demands
and increasing competition in this industry. Accordingly, the application of various
machine learning and deep learning methods such as artificial neural networks
(ANNs), expert systems, fuzzy systems, evolutionary-based methods, deep neural
network (DNN), convolutional neural network (CNN), and long short-term memory
(LSTM) has been introduced as effective methods to handle the decision-making
process of modeling power systems. The use of machine learning and deep learning,
which are data analysis techniques for building analytical models for a variety of
subjects (e.g., energy, healthcare, bioinformatics, transportation), is a promising
solution to overcome the current challenges of power systems. Machine learning
and deep learning, as a part of artificial intelligence family, are very effective
methods for facilitating the decision-making process of power systems operation,
planning, and control by learning from the raw data, identifying patterns, and
making decisions with minimum human intervention.
The advancement and development of power systems, as well as considerable
challenges such as the uncertain nature of renewable energy sources such as power
output of wind turbines and photovoltaic cells, load demand, and electrical energy
market, requires a high-performance approach for appropriate decision-making on
the operation, planning, and control of such systems. At the same time, the impor-
tance of data clustering and security of power systems highlights the need for a high-
performance method to handle the operation, planning, and control of such systems.
Accordingly, the use of different machine learning and deep learning methods as
effective techniques with acceptable performance (high accuracy) and reliability in
dealing with current issues in power systems in terms of management and operation
approaches of the system and forecasting the system parameters is discussed in this

v
vi Preface

book. Application of Machine Learning and Deep Learning Methods to Power

System Problems aims to evaluate the application of machine learning/deep learning
to issues and challenges of power systems considering recent developments and
advances in planning, operation, and control of such systems in both collecting a
comprehensive review based on the literature and applying machine learning to
power system planning, operation, and control. The topics covered in this book are
as follows:
• Power System Challenges and Issues
• Machine Learning Methods in Energy Engineering
• Overview of the Application of Machine Learning/Deep Learning for Controlling
and Planning Power Systems
• Application of Machine Learning/Deep Learning Methods for Clustering in
Power Systems
• Application of Machine Learning/Deep Learning Methods for Forecasting Power
System Parameters

University Park, PA, USA Morteza Nazari-Heris

University Park, PA, USA Somayeh Asadi
Aalborg, Denmark Behnam Mohammadi-Ivatloo
Geelong, VIC, Austrlia Moloud Abdar
University Park, PA, USA Houtan Jebelli
University Park, PA, USA Milad Sadat-Mohammadi
Contents

1 Power System Challenges and Issues . . . . . . . . . . . . . . . . . . . . . . . . 1

Ali Sharifzadeh, Mohammad Taghi Ameli, and Sasan Azad
2 Introduction and Literature Review of Power System
Challenges and Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Ali Ardeshiri, Amir Lotfi, Reza Behkam, Arash Moradzadeh,
and Ashkan Barzkar
3 Machine Learning and Power System Planning:
Opportunities and Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Mohammad Hosein Asgharinejad Keisami, Sasan Azad,
Reza Mohammadi Chabanloo, Morteza Nazari-Heris,
and Somayeh Asadi
4 Introduction to Machine Learning Methods in Energy
Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Arash Moradzadeh, Behnam Mohammadi-Ivatloo, Kazem
Pourhossein, Morteza Nazari-Heris, and Somayeh Asadi
5 Introduction and Literature Review of the Application
of Machine Learning/Deep Learning to Control Problems
of Power Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Samira Sadeghi, Ali Hesami Naghshbandy, Parham Moradi,
and Navid Rezaei
6 Introduction and Literature Review of the Application
of Machine Learning/Deep Learning to Load Forecasting
in Power System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Arash Moradzadeh, Amin Mansour-Saatloo, Morteza Nazari-Heris,
Behnam Mohammadi-Ivatloo, and Somayeh Asadi

vii
viii Contents

7 A Survey of Recent Particle Swarm Optimization

(PSO)-Based Clustering Approaches to Energy Efficiency
in Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Emrah Hancer
8 Clustering in Power Systems Using Innovative Machine
Learning/Deep Learning Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Mohammad Hossein Rezaeian Koochi,
Mohammad Hasan Hemmatpour, and Payman Dehghanian
9 Voltage Stability Assessment in Power Grids Using
Novel Machine Learning-Based Methods . . . . . . . . . . . . . . . . . . . . 177
Ali Mollaiee, Sasan Azad, Mohammad Taghi Ameli,
and Morteza Nazari-Heris
10 Evaluation and Classification of Cascading Failure Occurrence
Potential Due to Line Outage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Morteza Abedi, Mohammad Reza Aghamohammadi,
and Mohammad Taghi Ameli
11 LSTM-Assisted Heating Energy Demand Management
in Residential Buildings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Amin Mansour-Saatloo, Arash Moradzadeh, Sahar Zakeri,
and Behnam Mohammadi-Ivatloo
12 Wind Speed Forecasting Using Innovative Regression
Applications of Machine Learning Techniques . . . . . . . . . . . . . . . . 249
Arash Moradzadeh, Amin Mansour-Saatloo, Morteza Nazari-Heris,
Behnam Mohammadi-Ivatloo, and Somayeh Asadi
13 Effective Load Pattern Classification by Processing
the Smart Meter Data Based on Event-Driven Processing
and Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Saeed Mian Qaisar and Futoon Alsharif
14 Prediction of Out-of-Step Condition for Synchronous
Generators Using Decision Tree Based on the Dynamic
Data by WAMS/PMU . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
Morteza Abedi, Mohammad Reza Aghamohammadi, Sasan Azad,
Morteza Nazari-Heris, and Somayeh Asadi
15 The Adaptive Neuro-Fuzzy Inference System Model
for Short-Term Load, Price, and Topology Forecasting
of Distribution System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Mehrdad Setayesh Nazar and Ashkan Eslami Fard
Contents ix

16 Application of Machine Learning for Predicting User

Preferences in Optimal Scheduling of Smart Appliances . . . . . . . . . 345
Milad Sadat-Mohammadi, Morteza Nazari-Heris, Alireza Ameli,
Somayeh Asadi, Behnam Mohammadi-Ivatloo, and Houtan Jebelli
17 Machine Learning Approaches in a Real Power System
and Power Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
Peyman Razmi and Mahdi Ghaemi Asl

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
Chapter 1
Power System Challenges and Issues

Ali Sharifzadeh, Mohammad Taghi Ameli, and Sasan Azad

1.1 Introduction

In conventional power systems, large power plants have provided balancing in

the network parameters and its exchanges. Among different system requirements,
a priority after a basic balancing of power and energy is to ensure that power
flows and dynamics are within bounds and stable (for the angle, voltage, and
frequency) in normal and after events (faults, failures). However, conventional
planning and control models, which are well known and standardized, will be
challenged by all new concepts and future structures. Thus, the role of modeling
and calculations related to power system analysis with new concepts remains of
central importance[1].
On the other hand, large blackouts, although infrequent, are costly to society with
estimates of direct costs up to billions of dollars [2, 3]. Some other indirect costs such
as the failures of communications [4], natural gas network [5, 6], water supply,
transportation, and etc. are also causing the challenge. For existing and future
networks with extensive energy and money exchanges, blackout is no longer
acceptable [2]. Therefore, it is necessary to survey power system security and all
network parameters related to stability fast and accurate to be online and real time. In
addition, other issues such as greenhouse gas emissions, smart contracts, energy
management, and energy storage, each impose new requirements on the power grid.
Lack of proper consideration of new concepts and issues and lack of correct
prediction in this case make it difficult to manage the network.
All previous tips and the details will be mentioned later, such as renewables,
extensive load changes, load forecasting, resiliency, energy transactions, and

A. Sharifzadeh · M. T. Ameli (*) · S. Azad

Department of Electrical Engineering, Shahid Beheshti University, Tehran, Iran
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1

M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_1
2 A. Sharifzadeh et al.

uncertainties, may challenge a power system in the present and future [7]. Also, huge
data and increasing interdependence of information technology and energy manage-
ment infuse much of the current solutions about power systems in the future.
Therefore, new challenges required new and intelligent solutions. Older methods
will not work properly with new concepts, and new methods will need it.
The most attractive art for power system analysis is machine learning and its
various advanced methods, eliminating the need for complex calculations and
models. These stately methods, which are fast, robust, and adaptive, can overcome
the drawbacks of traditional solutions for several power systems problems [8].
It eliminates the need for system time allocation to perform continuous calculations.
The remainder of this chapter is organized as follows: Sect. 1.1 Introduction,
presents a brief of the work, and Sect. 1.2 presents existing and future system
challenges and classified according to the types. An overview of machine learning
challenges and issues is provided in Sect. 1.3. Finally, Sect. 1.4 presents a conclusion
for the chapter.

1.2 Present and Future Challenges

Construction and operation management of a power network always has challenges

within itself or various general policies. Addressing issues lead to planning to
mitigate the consequences. These challenges may be endogenous or exogenous.
The main issues were how to establish connections between different local networks
and increase equipment efficiency. In recent years, many issues that can reduce
network security or its reliability can make challenges. For example, increasing the
penetration rate of distributed generations and renewables and increasing uncer-
tainties on the demand side, such as the number of electric vehicles, have led to new
problems. Also, environmental constraints or restructuring to increase all partici-
pants’ profitability can lead to some new problems. In the future, issues such as
online monitoring of the network reduce transmission and distribution of technical or
financial costs, as well as online. However, increasing the use of distributed gener-
ators (DG) with uncertainty output will remain the major issues.
In the coming years, local energy transactions will be highly regarded. For
example, an electric vehicle can be used to supply some of a building’s energy
consumption or charged by local generator. This issue will be given a lot of attention
due to the importance of local exchanges. Due to the decentralization of these
exchanges, new platforms will be developed [9]. This subject may not be a chal-
lenge, but the multiplicity of these transactions may lead to issues such as legal
disputes, local technical or simulation problems, management, and decision-making
interferences in distribution networks or microgrids [10]. Figure 1.1 shows the
general challenges of future networks at a glance. In the following, the main
challenges and issues will be discussed.
1 Power System Challenges and Issues 3

DG
Penetration
Smart Big Data and
Contracts Data Sources

Power
Uncertainities
Electronics

Power System
Future Challenges
Local Energy Protection
Transaction Paradigm

Electrical
Resiliency
Vehicles
Storages

Fig. 1.1 Future challenges to power systems

1.2.1 Greenhouse Gases Emission

One of the most important issues of energy management is the reduction of green-
house gases and their consequences. One of the sources of greenhouse gas emissions
is fossil power plants and industrial consumers that emit large amounts of carbon
dioxide and gases. From 1990 to 2013, global energy demand increased by 55%.
The European Union, after China (3036 Mtoe), the United States (2188 Mtoe), and
Asia (excluding China) (1655 Mtoe), with an estimated gross domestic product
(GDP) of 1626 Mtoe, is the fourth most energy-consuming region in the world
[11]. In the field of energy production, despite efforts to decarbonize, the share of
fossil fuels in world energy has changed slightly over the years [12].
The Lisbon Treaty [13], adopted in 2009, brought about fundamental changes in
the main EU laws in the field of energy and European networks of the EU member
states. In other words, the treaty sets out four main objectives of EU energy policy.
• Achieve the proper performance of the energy market.
• Ensure the security of energy production in the union.
• Provide solutions to improve energy efficiency and energy savings and create
new and renewable forms of energy.
• Enable better connections between different networks.
4 A. Sharifzadeh et al.

In general, the European Union’s solutions do not affect the right of a member
state to determine the conditions of use of its energy resources, choose between
different energy sources, and the overall structure of the energy source. However, the
Lisbon Treaty refers to measures tailored to the economic situation and technical
issues. Especially if there are serious problems in different sectors, especially energy
[13], in this regard, the European Commission has begun to develop a strategy for a
flexible energy structure related to climate policy measures. The stated goal for
families and businesses and consumers is to provide safe, sustainable, competitive,
and cost-effective energy. Achieving this goal requires a major change of the
planned European energy system. These issues can cause problems in both produc-
tion and consumption. For this reason, related solutions were provided to overcome
the challenges [14].
In particular, the political challenges of recent years have led to a great deal of
attention being paid to the diversity of energy sources, suppliers, and safe and
sustainable energy supply routes for citizens and companies and to expectations
for affordable energy access.
EU policies of energy are guided by three mains and reflect the objectives set out
in the Lisbon Treaty.
Cost-effectiveness ensures the activity of energy suppliers in a competitive
environment that ensures reasonable and competitive prices for homes, businesses,
and industries.
• Security: improving reliability and continuous energy supply
• Sustainability: sustainable energy consumption, with reduced greenhouse gas
emissions, pollution, and dependence on fossil fuels
To pursue energy and climate change goals in a long-term strategy, the European
Union has agreed on specific targets for 2020–2030 and has drawn up a road map for
2050. The goal of greenhouse gas emissions seems to emerge as a major factor in
shaping the EU’s domestic/foreign energy policies and climate change.
Greenhouse gas emissions in Europe are expected to decline as follows [15, 16]:
• By 2020: 20% below the 1990 level
• By 2030: 40% reduction
• By 2050: Decrease of at least 80%
Pursuing greenhouse gas reduction programs requires the use of new technolo-
gies and concepts in the management and structures of power grids. Given the above
and looking at the future of Europe plans, energy by 2040 will rely significantly on
renewable resources and new technologies. Some of these new concepts can create
uncertainties in the network.
1 Power System Challenges and Issues 5

1.2.2 Distributed Generation and Renewables

With the development of industry and the increase of domestic and industrial
consumption, the penetration rate of distributed generators and renewables will
increase. Therefore, the use of large power plants and fossil fuels will be less.
This can reduce the rotational inertia in the network and can lead to instabilities.
To this end, changes will occur in the management and operation of the network.
This rotational inertia, network requirement, will be partially compensated by
electronic power capabilities, and network stability will be improved [17, 18]
(Fig. 1.2).
Due to the fact that renewables create an uncertainty in power generation, the use
of equipment that can store energy and inject into the network at the appropriate time
has been considered. Future grids with high penetration of renewables use
microgrids to enable widespread utilization of renewable energy sources for increas-
ing efficiency and reliability [19].

Fig. 1.2 Challenges of renewable energies

6 A. Sharifzadeh et al.

1.2.3 Energy Storages Integrity

One of the equipment groups that is widely used is energy storage. With the integrity
of storage equipment, operators can perform peak shaving and transient damping
better with more flexibility [20, 21]. This equipment usage in the network can cause
problems in power flow calculations, and it is necessary for considerations in this
regard. Developments in batteries, superconductors, compressed-air energy storage
(CAES), power electronic, and etc. have led to the development of storage equip-
ment and are expected to be widely used in transmission and distribution networks in
the future (Fig. 1.3).
Storage solutions and technologies available in medium and large capacities
create new concepts in general utilization of power networks. In the same way,
with the expansion of storage equipment in the network, it is necessary to examine
the interactions and effects [22].
The allocation of energy storage in distribution systems is widely considered
worldwide. This problem poses two major challenges: storage location and its
capacity. If wrong decisions are made in both cases or one of them, the distribution
system’s performance may be compromised. The placement and selection of storage
capacities are done according to the network requirements or depth and optimized.
Traditional and intelligent methods are also used in this field [23, 24]. This will also
be used extensively in smart home energy management systems [25].
Storages may increase the level of energy dissipation, decrease the voltage pro-
file’s status, and negatively affect the network’s technical operating conditions.
Determining the capacity of this equipment should also be done in a way that
optimizes economic issues. In other words, it can store a suitable amount when the

Fig. 1.3 Energy storage technologies

1 Power System Challenges and Issues 7

Fig. 1.4 Benefits of power storages

energy price is low, and when it is expensive or in case of emergency, it can inject
the stored energy into the grid.
A multi-energy system provides more system operation flexibility. This concept
will improve reliability and efficiency. Energy storages can integrate various energy
sources such as solar, wind, gas, heat, and combined heat and power units at different
levels [26]. Additionally, coordinated scheduling of flexible loads and energy
storage systems can play an important role in the optimal scheduling of microgrids
and lowering the costs [19, 27].
The development of storage equipment and its integration into the power grid will
increase maintenance costs because this equipment mainly has expensive parts and
impose additional costs. For this purpose, it will be necessary to develop mainte-
nance programs and network operation plans based on these activities (Fig. 1.4).

1.2.4 Electric Vehicles

The spread of electric vehicles, like a double-edged sword, can be used to improve
the condition of the network or cause problems if mismanaged. Frequent use of
8 A. Sharifzadeh et al.

electric vehicles and plugging to the network creates an unpredictable situation.

However, managing the charge or discharge of these vehicles and applying new
control methods in smart grids can improve the situation.
In future networks, there may be developments that will lead to major changes in
this field. New researches cover topics such as wireless charging [28, 29], vehicle to
grid (V2G) block-chain management [30], privacy, and security, and smart charge
optimization is considered. In the future, with the development of electric vehicles
and charging stations, a change in the management of distribution and power
networks will be required.
As mentioned above, the advantage of electric vehicle application is that if they
are properly managed to charge and discharge. Therefore, network parameters such
as voltage and frequency can be controlled. For this reason, the location and capacity
of charging stations are of particular importance. Various optimization or intelligent
methods have been provided in the placement of this equipment and interact with the
network. One of the most important issues addressed in new studies is car charging
stations allocation [31]. Point placements are based on optimization methods, but
other factors such as feasibility of installation, economic issues, etc. are also
involved in this issue. However, all components within the network are currently
being used to improve its situation.

1.2.5 Decentralization and Smart Contract

In two-way interactions of consumers and networks owners, new financial structures

and different settlement methods are formed. Therefore, by increasing the profitabil-
ity of all components, the various capabilities that consumers can create will be used
properly. The volume of financial transactions and its optimization with different
conditions will develop the use of block-chain technology, digital currencies, and
smart contract structures [32, 33]. This requires the creation of various legal laws as
well as new financial structures. Creating communication platforms and a secure
database will also be required although focusing on financial structures may led to
some challenges for power network security. But creating flexible rules can drive
interactions to improve network parameters and conditions. Smart contracts will
create a new perspective on the electricity market. By observing the options of these
contracts by the parties, the agreed outputs will be applied, and by deviating from it,
some presuppositions or penalties will be activated. By properly designing the
various dimensions of this type of contract, the network situation in relation to the
market can be controlled appropriately and optimized. Developments and increasing
interactions between network components may affect network security and stability.
Rapid changes in load values, as well as planning to increase the profitability of
components, can reduce operating safety margins. Therefore, software related to
controlling, monitoring, and predicting future status should be developed to meet
operators’ new requirements. To this end, new security features and new methods
without modeling will be developed and used. System reliability will also be
1 Power System Challenges and Issues 9

SMART CONTRACT

INPUT DATA AUTOMATIC

EXECUATION
PRESET RESPONSE
DATA STATE MACHINE
RATE

PRESET RESPONSE CONTRACT STATUS CONTRACT VALUE

CONDITION

BLOCK CHAIN

B1 B2 B3

Fig. 1.5 Smart contract configuration

affected, and correction plans will be required. Currently, network status can be
extracted using calculations and simulations. Different network modes and emer-
gencies must also be modeled to predict or evaluate the situation (Fig. 1.5).

1.2.6 Reliability and Security Problems

The availability of all components of a power grid and increasing its reliability is one
of the main concepts in multi-energy systems. Contribution in different exchange
forms of energy such as electricity, heat, and natural gas plays main rules for future
consumers and industries. Multi-energy networks can provide more opportunities
for players to increase their profit [34]. As the number of devices, market players,
and energy forms in the network increases, the network’s reliability and security will
be affected. These effects may be due to changes in voltage or frequency or other
network parameters. Players outside the power grid will also have a big impact on its
performance. For this reason, network operators consider it desirable to increase
stability in network parameters (Fig. 1.6).
Power system transient stability assessment is important to determine its reliabil-
ity and continuity of performance. For this purpose, Phasor Measurement Units
(PMU) and Remote Terminal Units (RTU) may be widely use in the network. The
data collected from these units, after analysis, is used for specific purposes, such as
determining the stability of the network or its orientation. This is Supervisory
10 A. Sharifzadeh et al.

Fig. 1.6 Classification of

power system stability
Rotor
angle
stability

Power
system
stability
Voltage Frequency
stability stability

Control and Data Acquisition (SCADA) system responsibility. The amount of

processing time of the network status assessment and the accuracy of these calcula-
tions can cause some problems. Increasing processing times or decreasing results
accuracy can lead to incorrect control decisions or grid blackouts. Also, the network
data gathering through PMUs and high-speed communication media is important in
the network’s proper operation. Using reliable telecommunication infrastructures
and capturing network parameters at all levels of voltage, huge volumes of data
transmitting pose a great challenge to network designers and operators.
For this reason, operation and optimization of various power network issues and
problems led to the new fast and smart solutions. New methods are used to assess the
stability and security of the network in studies or in practice. These methods, which
have recently used image processing and deep learning processes [35], can provide
the desired output with appropriate accuracy. A pattern of network status is provided
that identifies and classifies insecure or safe points. These patterns are updated at any
time using the information received from PMUs [36]. The problem is the accuracy of
the outputs, which, if the network is misdiagnosed as insecure or vice versa, will lead
to erroneous decisions and lead to irreparable accidents. For this reason, it is
necessary to check the results accuracy so as not to challenge the network. There
are several methods for evaluating the accuracy of intelligent methods that have been
addressed in recent researches. However, increasing the accuracy and reducing the
computation time is very important in evaluating network security. On the other
hand, generating train and test data for neural networks is one of the new methods
challenges. In recent years, various methods of data generation have been proposed
in this field, which are considered in research [37].
Voltage regulation is an important aspect of reliability. Power grid operating
regulations require all installations to maintain voltage within a suitable and stable
range. This is traditionally achieved using capacitors, tap changers, and distributed
generators. All of these methods may be relatively slow. But modern inverters with
new technology can enable fast voltage regulation in the network.
1 Power System Challenges and Issues 11

1.2.7 Load Forecasting Problems

Extensive use of renewable energy and various loads with high uncertainty in power
grids, and the importance of profitability, disrupt short-term and long-term load
forecasting. In other words, one of the main challenges in future networks is load
management and generation planning based on high uncertainty load forecasts and
stochastic energy generation. Therefore, the use of neural networks and deep
learning methods in this field is considered important in recent research [38–40]
(Fig. 1.7).

1.2.8 Big Data

As a significant application of energy, future networks are complicated

interconnected power grid that involves high penetration distributed generations,
highly variable loads, smart meters, real-time controllers, and data processors. It
continuously generates data with large volume, high velocity, and diverse variety.
We need the high volume of data and information exchanged for monitoring and
controlling power networks [41].
To this end, the applications of big data and neural networks are becoming more
widespread. Free model methods were introduced and used in research. These

Fig. 1.7 Load forecasting

techniques Load Forcasting
Techniques

Base on the
Base on the
duration of
models used
forcasting

Linear methods Short term

Non-linear
Medum term
methods

Long term
12 A. Sharifzadeh et al.

Fig. 1.8 Big data in future

networks

PMU/RTU

Smart
Power Market
Meteres

Big Data in Future

Networks

Weather SCADA

Simulations

methods led to the development of control center applications and improved the
system. Relying on telecommunications and large amounts of data can compromise
the communication cybersecurity of the system. Creating and increasing telecom-
munication cybersecurity is one of the most important issues that is addressed. This
subject can cause local or major problems (Fig. 1.8).

1.3 Machine Learning (ML) Application and Challenges

Power network management and solving various problems require the use of
different methods of complex modeling and calculations. Sometimes, the volume
of these calculations may be large enough to interrupt the decision. To solve
problems, model-free methods were proposed [42]. Also, artificial neural networks
and machine learning methods were used to reduce the volume of calculations.
However, these methods did not eliminate much of the need for simulation because
true data is required for different learning methods or results testing, which must be
either available in the control center’s database or generated by simulations. There-
fore, big data applications and data mining are also considered. In recent years,
focusing on increasing accuracy and reducing processing time, reliable methods
based on machine learning models have been proposed. With the development of the
proposed models, new processing capabilities and tools have been made available to
researchers and engineers. These capabilities are used in data generation, status
1 Power System Challenges and Issues 13

Fig. 1.9 Machine learning

applications

Load
management

Fault
Pricing classification
and location

Machine learning
application

Reconstruction Restoration

Security
assessment

assessment, or status forecasting. Figure 1.9 shows a number of applications of

machine learning in the field of power engineering.
As shown in Fig. 1.9, there are several applications for machine learning in power
grids. In other words, all the challenges in power networks can be solved using
different machine learning methods. In recent years, much attention has been paid to
develop and improve problem-solving with these methods. For example, in short-
term forecasts of load consumption and the presence of large uncertainties, such as
renewable generators and electric vehicles, intelligent methods can provide good
results in this field. Also, in finding suitable routes to restoring the network after
blackouts, intelligent methods can provide suitable patterns based on different grid
situations.
In power transactions and electricity markets, machine learning methods can play
an effective role in both finding the optimal price and increasing network capabil-
ities. This role will be highlighted when the rate of these transactions is greatly
increased and the profitability of network components in interaction with operating
conditions is considered. However, given the breadth of new power grid issues and
the new complexities that have arisen, the use of these intelligent methods has been
considered. However, it needs to be addressed more in the industry. The problem is
that many operators are still interested in using traditional methods, and they may
prevent the use of smart methods in the operation of power networks.
On the other hand, machine learning and data mining methods led to new
challenges in the power system processing. Some of these challenges may be in
terms of operating costs, and some may be related to the data usages and applications
14 A. Sharifzadeh et al.

Fig. 1.10 Machine learning

challenges
Talent
Deficit
Cyber
High Cost
attacks

Data and Machine Learning Historical

Computer
Challenges Data
Science

Results New
Evaluation Technology
Operators

[43]. Figure 1.10 shows some of these challenges. These challenges have not
prevented the use of intelligent methods in studies. These challenges may arise
due to the requirement for new concepts. These new concepts, including data
science, big data, artificial intelligence, and image processing, are used in simulation
processes.
For this reason, it may not be easy for older operators to understand. Sometimes,
the use of these methods may be opposed by operators. It also requires some
structures such as historical databases or result evaluation. Evaluating the results is
important because the accuracy of the outputs must be within the appropriate range.
Otherwise, decisions may be made with a lot of deviation. Also, if there are
deviations in the results, the actual state of the system may be presented incorrectly.
Therefore, evaluating the results and its different methods is of great importance.
Engineers and managers who have to work with these structures required data
science and computer science training. They should have a wealth of professional
talent in this field. On the other hand, creating and using any new structure requires
time and money. Solving possible structural problems and improving methods
depend on the application of researches in the industries.
The development of machine learning methods and models and dependence on
databases and software methods increase hackers and cyberattacks. The relevant
protection structures and firewalls can control these attacks.
Different methods of machine learning are rapidly penetrating the power grid
operation. This rapid development will lead to a new revolution in various industries
and problem-solving strategies.
1 Power System Challenges and Issues 15

1.4 Conclusion

As discussed in this chapter, the future power grids will face various challenges and
issues. Some of these challenges are related to new equipment, some are related to
new concepts, and some are related to new simulation and calculation methods.
Also, the large volume of big data and the breadth of calculations will greatly impact
operations and decisions. Examining the impact of using new equipment such as
storage, renewables, and predicting future network states with many power elec-
tronics helps a lot to understand future issues of the power grid.
These challenges will lead to the following main items:
• Renewable and storage penetration
• The impact of consumers on network management
• Big data and large volumes of data exchanges
• New concepts, technologies, and methods
These changes in structure and concepts may also create resistance to the use of
new methods. Decisions at the highest levels of management can be helpful in this
regard. Using some smart methods can also help provide new solutions. At this
stage, new methods have some problems and need more development. However,
focusing on new concepts and developing intelligent algorithms such as machine
learning effectively reduce the challenges of power grids. In this chapter, by
examining the future network, some concepts and methods were evaluated, and
general pre-enlightenment challenges were obtained.

References

1. J.P. Lopes, N. Hatziargyriou, J. Mutale, P. Djapic, N. Jenkins, Integrating distributed generation

into electric power systems: a review of drivers, challenges and opportunities. Electr. Pow. Syst.
Res. 77(9), 1189–1203 (2007)
2. H. Haes Alhelou, M.E. Hamedani-Golshan, T.C. Njenda, P. Siano, A survey on power system
blackout and cascading events: research motivations and challenges. Energies 12(4), 682 (2019)
3. B.A. Carreras, D.E. Newman, I. Dobson, North American blackout time series statistics and
implications for blackout risk. IEEE Trans. Power Syst. 31(6), 4406–4414 (2016)
4. P.-Y. Kong, Optimal configuration of interdependence between communication network and
power grid. IEEE Trans. Ind. Inform. 15(7), 4054–4065 (2019)
5. H. Ameli, M. Qadrdan, G. Strbac, Coordinated operation strategies for natural gas and power
systems in presence of gas-related flexibilities. Energy Syst. Integration 1(1), 3–13 (2019)
6. V. Shabazbegian, H. Ameli, M.T. Ameli, G. Strbac, Stochastic optimization model for coordi-
nated operation of natural gas and electricity networks. Comput. Chem. Eng. 142, 107060
(2020)
7. C. Syranidou, J. Linssen, D. Stolten, M. Robinius, Integration of large-scale variable renewable
energy sources into the future european power system: on the curtailment challenge. Energies
13(20), 5490 (2020)
8. S.M. Miraftabzadeh, F. Foiadelli, M. Longo, M. Pasetti, A survey of machine learning appli-
cations for power system analytics, in 2019 IEEE International Conference on Environment
16 A. Sharifzadeh et al.

and Electrical Engineering and 2019 IEEE Industrial and Commercial Power Systems Europe
(EEEIC/I&CPS Europe), (IEEE, 2019), pp. 1–5
9. D. Vangulick, B. Cornélusse, D. Ernst, Blockchain for peer-to-peer energy exchanges: design
and recommendations, in 2018 Power Systems Computation Conference (PSCC), (IEEE, 2018),
pp. 1–7
10. W. Tushar, C. Yuen, H. Mohsenian-Rad, T. Saha, H.V. Poor, K.L. Wood, Transforming energy
networks via peer-to-peer energy trading: The potential of game-theoretic approaches. IEEE
Signal Process. Mag. 35(4), 90–111 (2018)
11. F. Profumo, E. Bompard, Fulli, G, Electricity security: models and methods for supporting the
policy decision making in the european union (2016)
12. R.G. Newell, Y. Qian, D. Raimi, Global Energy Outlook 2015, National Bureau of Economic
Research, 0898-2937 (2016)
13. E. Union, Treaty of Lisbon: Amending the Treaty on European Union and the Treaty
Establishing the European Community (Office for Official Publications of the European
Community, 2007)
14. E.U. Package, A framework strategy for a resilient energy union with a forward-looking climate
change policy, in Communication from the Commission to the European Parliament, the
Council, the European Economic and Social Committee, the Committee of the Regions and
the European Investment Bank, COM, vol. 80 (2015)
15. A. Froggatt, A. Hadfield, Deconstructing the European Energy Union: Governance and 2030
Goals, EPG Working Paper: EPG 1507 (2015)
16. P. Capros et al., EU Energy, Transport and GHG Emissions Trends to 2050-Reference Scenario
2013 (European Commission, 2013)
17. K.S. Ratnam, K. Palanisamy, G. Yang, Future low-inertia power systems: requirements, issues,
and solutions-a review. Renew. Sustain. Energy Rev. 124, 109773 (2020)
18. L. Mehigan, D. Al Kez, S. Collins, A. Foley, B. Ó’Gallachóir, P. Deane, Renewables in the
European power system and the impact on system rotational inertia. Energy 203, 117776 (2020)
19. M.A. Mirzaei et al., A novel hybrid two-stage framework for flexible bidding strategy of
reconfigurable micro-grid in day-ahead and real-time markets. Int. J. Electr. Power Energy
Syst. 123, 106293 (2020)
20. H.E. Murdock et al., Renewables 2019 Global Status Report, 2019.
21. K.M. Muttaqi, M.R. Islam, D. Sutanto, Future power distribution grids: integration of renew-
able energy, energy storage, electric vehicles, superconductor, and magnetic bus. IEEE Trans.
Appl. Superconductivity 29(2), 1–5 (2019)
22. J.E. Bistline, D.T. Young, Emissions impacts of future battery storage deployment on regional
power systems. Appl. Energy 264, 114678 (2020)
23. M. Baza, M. Nabil, M. Ismail, M. Mahmoud, E. Serpedin, M.A. Rahman, Blockchain-based
charging coordination mechanism for smart grid energy storage units, in 2019 IEEE Interna-
tional Conference on Blockchain (Blockchain), (IEEE, 2019), pp. 504–509
24. T. Fu, C. Wang, N. Cheng, Deep learning based joint optimization of renewable energy storage
and routing in vehicular energy network. IEEE Internet Things J. (2020)
25. H. Jang, T. Lee, S.M. Kim, J. Lee, S. Park, Energy storage system management method based
on deep learning for energy-efficient smart home, in 2020 IEEE International Conference on
Consumer Electronics (ICCE), (IEEE, 2020), pp. 1–2
26. N. Nasiri et al., A bi-level market-clearing for coordinated regional-local multi-carrier systems
in presence of energy storage technologies. Sustain. Cities Soc. 63, 102439 (2020)
27. M.-N. Heris et al., Evaluation of hydrogen storage technology in risk-constrained stochastic
scheduling of multi-carrier energy systems considering power, gas and heating network con-
straints. Int. J. Hydrogen Energy 45(55), 30129–30141 (2020)
28. J. Chen, C.W. Yu, W. Ouyang, Efficient wireless charging pad deployment in wireless
rechargeable sensor networks. IEEE Access 8, 39056–39077 (2020)
29. O. Nezamuddin, E.C. dos Santos, Vehicle-to-vehicle in-route wireless charging system, in 2020
IEEE Transportation Electrification Conference & Expo (ITEC), (IEEE, 2020), pp. 371–376
1 Power System Challenges and Issues 17

30. V. Hassija, V. Chamola, S. Garg, N.G.K. Dara, G. Kaddoum, D.N.K. Jayakody, A blockchain-
based framework for lightweight data sharing and energy trading in V2G network. IEEE Trans.
Vehicular Technol. (2020)
31. Y. Motoaki, Location-allocation of electric vehicle fast chargers—research and practice. World
Electr. Vehicle J. 10(1), 12 (2019)
32. D. Han, C. Zhang, J. Ping, Z. Yan, Smart contract architecture for decentralized energy trading
and management based on blockchains. Energy, 117417 (2020)
33. S. Yu, S. Yang, Y. Li, J. Geng, Distributed energy transaction mechanism design based on
smart contract, in 2018 China International Conference on Electricity Distribution (CICED),
(IEEE, 2018), pp. 2790–2793
34. M.Z. Oskouei, M.A. Mirzaei, B. Mohammadi-Ivatloo, M. Shafiee, M. Marzband, A. Anvari-
Moghaddam, A hybrid robust-stochastic approach to evaluate the profit of a multi-energy
retailer in tri-layer energy markets. Energy 214, 118948 (2020)
35. J.-M.H. Arteaga, F. Hancharou, F. Thams, S. Chatzivasileiadis, Deep learning for power system
security assessment, in 2019 IEEE Milan PowerTech, (IEEE, 2019), pp. 1–6
36. P.K. Jaiswal, S. Das, B.K. Panigrahi, PMU based data driven approach for online dynamic
security assessment in power systems, in 2019 20th International Conference on Intelligent
System Application to Power Systems (ISAP), (IEEE, 2019), pp. 1–7
37. F. Thams, A. Venzke, R. Eriksson, S. Chatzivasileiadis, Efficient database generation for data-
driven security assessment of power systems. IEEE Trans. Power Syst. 35(1), 30–41 (2019)
38. A. Al Mamun, M. Sohel, N. Mohammad, M.S.H. Sunny, D.R. Dipta, E. Hossain, A compre-
hensive review of the load forecasting techniques using single and hybrid predictive models.
IEEE Access 8, 134911–134939 (2020)
39. M.A. Hammad, B. Jereb, B. Rosi, D. Dragan, Methods and models for electric load forecasting:
a comprehensive review. Logist. Sustain. Transp. 11(1), 51–76 (2020)
40. N.M.M. Bendaoud, N. Farah, Using deep learning for short-term load forecasting. Neural
Comput. Appl. 32(18), 15029–15041 (2020)
41. M. Ghorbanian, S.H. Dolatabadi, P. Siano, Big data issues in smart grids: a survey. IEEE
Syst. J. 13(4), 4158–4168 (2019)
42. X. Li, J. Wen, Review of building energy modeling for control and operation. Renew. Sustain.
Energy Rev. 37, 517–537 (2014)
43. M. Shafique et al., Adaptive and energy-efficient architectures for machine learning: challenges,
opportunities, and research roadmap, in 2017 IEEE Computer Society Annual Symposium on
VLSI (ISVLSI), (IEEE, 2017), pp. 627–632
Chapter 2
Introduction and Literature Review
of Power System Challenges and Issues

Ali Ardeshiri, Amir Lotfi, Reza Behkam, Arash Moradzadeh,

and Ashkan Barzkar

2.1 Introduction

Since the beginning of electrical power system in 1880s, when lamps were used for
lighthouse and street lighting purposes and the commercial use of electricity started
[1], it has been developed into a great industry and economy. Having a fundamental
role in modern era lifestyle, the consumption of electrical power has risen sharply in
the twenty-first century, and as a response do demand growth, electricity generation
has increased accordingly [2]. Worldwide electrical power demand increases 2.1%
per year until 2040, while total energy demand growth is half of electricity. This
matter raises the electricity share in total energy demand from 19% in 2018 to 24% in
2040. The electricity consumption is set to grow due to the electrification of transport
and heat, increasing incomes of households, and rising demand of air-conditioning,
digital, and electronic devices [3, 4].
Nowadays, the increasing energy demand, development of smart grids, and the
combination of different types of energy systems have led to complexity of power
systems. On the other hand, ever-expanding energy consumption, development of
industry and technology systems, high penetration of solar and wind energies have
made electricity networks operate in more complex and uncertain conditions.

A. Ardeshiri (*) · A. Lotfi · A. Barzkar

Electrical Engineering Department, Sharif University of Technology, Tehran, Iran
e-mail: [email protected]; [email protected]
R. Behkam
Electrical Engineering Department, Amirkabir University of Technology, Tehran, Iran
e-mail: [email protected]
A. Moradzadeh
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 19

M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_2
20 A. Ardeshiri et al.

Therefore, analysis of traditional power systems requires physical modeling and

extensive numerical computation. To analyze behavior of these systems, advanced
metering and monitoring systems are utilized which generate huge amount of data.
Machine learning, deep learning, and variety of regression, classification, and
clustering algorithms are powerful tools to use in these issues. These procedures
can be utilized to solve the power system problems and challenges such as planning,
operation, fault detection and protection, power system analysis and control, and
cybersecurity.
Globally, millions of people around the world have no access or limited access to
electricity. This problem is not limited to developing countries. Even developed
countries need to deliver electricity to remote areas which are far from the grid, such
as islands, deserts, and places surrounded by mountains. Available energy is vital to
reach development targets [5, 6]. Moreover, electricity plays an important role in the
interconnection of food, energy, and water, which is known as FEW Nexus. Thus,
the availability and resilience of electricity is of paramount importance [7, 8].
The main goal of power system is to supply the demand as much as possible.
However, this matter is not as simple as it appears. To address the challenges of
power system, three categories of challenges are considered in this chapter which are
issues regarding power system planning, operation, and control. Planning is
the activity and decision-making associated with the development of plans regarding
the construction, design, and expansion of power system elements which meets the
electricity demand in the future, considering the present situation [9], by predicting
the condition and making some assumption of the future [10]. Power system
planning can be mainly about reducing the investment costs, pollution, and power
outage or increasing reliability, resiliency, and security [11]. Planning phase vary
from 1 to 20 years. Once the power system elements are built, these elements need to
be operated to supply the demand while following environmental and economic
objectives and satisfying technical constraints. At the end, in power system control, a
short time before delivering the power to the loads, considering all events, for
example, contingencies and outages, happening in generation, transmission, or
distribution, the main focus is to maintain the robustness and security of the power
grid, while the economy and environment may not be as important as it was in
planning and operation [12].
In the following subsections, the challenges and issues regarding planning,
operation, and control are explained in details separately.

2.2 Power System Planning

In the other part, which is known as planning studies, the energy supply required by
consumers is not provided properly because parts of the power system have lost their
adequacy and are not able to provide the desired loads. The purpose of these studies
is to determine the type, installation location, number, and time of installation of new
2 Introduction and Literature Review of Power System Challenges and Issues 21

Fig. 2.1 The basic structure of the power system

equipment so that the required network adequacy is retrieved with the least cost
[13, 14].
As stated, the purpose of planning studies in power systems is the recovery of the
adequacy of equipment to meet the load required by consumers in an optimal way.
The inability of power plants to produce the required power, the lack of suitable
capacity of lines to transfer the required power, and the lack of capacity or the
number of network substations are the inadequacies of the power system. Therefore,
it can be concluded that determining the time and capacity required to expand
generation, transmission, and distribution to properly supply the load required by
consumers to electricity at the lowest possible cost is the goal of the planning
issue [15].
Due to the complexity and dimensions of the problem, expansion planning
studies are usually considered in three separate sections [16]:
1. Generation Expansion Planning (GEP).
2. Network Expansion Planning (NEP).
3. Substation Expansion Planning (SEP).
In Generation Expansion Planning, the goal is to determine the capacity of
generation resources, their installation location, and installation time at the lowest
cost. In this purpose, constraints are also considered in such a way that by consid-
ering the failures and units leaving the network, the total production capacity in the
system is able to supply the maximum amount of consumer load, and their
non-supply does not exceed a certain amount [13, 17].
Line expansion planning can also be considered in three sections: transmission
lines, sub-transmission, and distribution. Assuming that load of substations and
generation centers in the line expansion planning is clear and the generation sector
and substations have the necessary adequacy, the type, capacity, location, and time
of need for lines are considered [16, 18].
In a power system, as shown in Fig. 2.1, power is generated in energy-generating
sources and transmitted through transmission lines to transmission substations. Each
transmission substation feeds a number of sub-transmission substations through the
sub-transmission lines. Finally, 20 kV feeders deliver power from the
sub-transmission substations to the distribution substations, and through this, the
22 A. Ardeshiri et al.

consumers are fed. In order to feed consumers adequately in normal conditions or in

the event of an accident, the total load of the substations must be calculated taking
into account the allowable capacity of the equipment installed in the substation. As
the load grows, the capacity of the existing substation equipment may not meet the
load consumption. To solve this problem, existing substations must be expanded or
solved by installing new substations on the network.
In general, the amount of network load downstream and the position of the
network affect the capacity of each substation. With the implementation of SEP in
a network, it is determined what equipment of the mentioned substations should be
increased in capacity when and to what extent or in what time and place how many
new posts should be built with what capacity that the network will regain its lost
adequacy [16, 18].
As mentioned, the purpose of implementing SEP is to determine a combination of
expanding existing substations or constructing new substations to provide safe and
desirable supply to loads at the lowest cost.
In a given geographical area, consumers are scattered throughout the study area.
Using transmission lines, sub-transmission, and distribution, electrical energy must
be delivered to all consumers in the desired geographical area. After generating
electricity in power plants, it is transferred to the sub-transmission substations
through transmission lines, transmission substations, and sub-transmission lines.
The sub-transmission substations are closer to the consumers and transfer the
power to several distribution substations through 20 kV feeders after reducing the
voltage level. Transmission substations with a capacity of several hundred MVA,
sub-transmission substations with a capacity of several tens of MVA, and distribu-
tion substations with a capacity of about a fraction of MVA to a few MVA are
constructed. The geographical distribution of consumers and the amount of their
load determine the number and density of required substations. The manner of
establishing the sub-transmission substations is effective on the construction site
of the transmission substations, and the manner of establishing the distribution
substations is also effective on the construction site of the sub-transmission sub-
stations. Given this dependency, distribution planning for distribution,
sub-transmission, and transmission substations should be done simultaneously.
But because the problem of simultaneous development planning of these posts is
very complex due to the existence of many variables and constraints, in order to
achieve a practical answer in a certain time, the designers of the power system divide
the problem of planning the development of posts into three separate parts. These are
the expansion of distribution substations, expansion of sub-transmission substations,
and expansion of transmission substations, each of which is solved separately [16].
Thus, first, according to the location of low-pressure loads, the arrangement of
low-pressure lines and the capacity and location of distribution substations are
determined. Then these substations are considered as load points, and their capacity
is also considered as the value of these load points, and according to this, the
arrangement of medium pressure lines and the location and capacity of the
sub-transmission substations are determined optimally. Downstream (distribution)
and upstream (transmission) networks are not considered to simplify the task at this
2 Introduction and Literature Review of Power System Challenges and Issues 23

stage. After obtaining the location of the two capacities of the sub-transmission
substations, these substations are considered as load points with the amount of load,
and according to that, the arrangement of the above distribution lines, the location,
and capacity of the transfer substations are determined.
As mentioned, in order to reduce the complexity and dimensions of the problem
in each of the stages of separate expansion, the arrangement of downstream or
upstream networks is not considered. However, this can take the answers to the
problem of developing the whole system away from the optimal value. For example,
suppose that in planning the expansion of a sub-transmission network including
transmission substations, sub-transmission lines, and sub-transmission substations,
the problem of simultaneous expansion is divided into two subthemes of the
expansion of sub-transmission substations and the development of transmission
substations. In this case, in planning the sub-transmission substations, the designer
tries to reduce the length of the feeders and reduce costs, the substations should be
built as close as possible to the load centers, and this will cause these substations
have more distance from the transmission substations, which in addition to increas-
ing the cost of connection to the transmission system will lead to other technical
problems. This is shown schematically in Fig. 2.2.
Network expansion planning is based on projected loads, while projected loads
are not based solely on network performance in the future and will be affected by
factors based on network performance in the past and influential factors in the future.
In this regard, examining the effect of uncertainty on effective parameters will have a
significant impact on the planning response [19].
Load uncertainty studies are one of the important inputs in estimating the time of
equipment addition to maintain network adequacy. As mentioned, the main causes
of uncertainties are the emergence of new loads, rising prices, changing laws, and
weather conditions (temperature and wind, etc.) [20]. In addition, in recent years,
there has been a wave of changes in power networks that have led to uncertainties.
These include restructuring of distribution networks, privatization, free market or
competitive market, alternative energy sources, distributed generation, and new
protection and communication technologies in the field of energy [21].
In most of the proposed models for network development planning for simplifi-
cation, the effect of uncertainty on the problem parameters has been neglected [22].
Expansion planning for large networks will not have a real answer without
considering these factors because as the network grows, both the probability of
uncertainty and the amount of uncertainty in the parameters of the problem increase.
Due to recent developments in the field of the logic of power systems operation,
in the face of increasing system capacity, innovative options have been proposed.
One of the options that have been considered by many researchers recently is
distributed generation (DG) [23].
DGs represent production units that are connected to the grid near low-load
capacities. The use of DGs improves the voltage profile and reduces losses, which
in turn reduces the cost of operating the network. Also, with the installation of DGs
in the network, investment costs are reduced because the installation of DGs can
delay the construction of new equipment or strengthen existing equipment [24].
24 A. Ardeshiri et al.

possible candidates for network expansion (a)

Available Transmission Substations

Candidate Sub-transmission Substations

Candidate Sub-transmission Power Lines

Candidate Feeders

Load Points

Short feeders but long sub-transmission lines separate expansion (b)

relatively short feeders and short sub-transmission lines simultaneous expansion (c)

Fig. 2.2 Comparison of separate and simultaneous expansion schematically

2 Introduction and Literature Review of Power System Challenges and Issues 25

In addition, DGs will play an important role in increasing the quality and
reliability of customer service. Given the potential benefits of DGs and the declining
price of DG resources, these resources are expected to play an important role in the
future of power systems [25].

2.3 Power System Operation Challenges

The power system operation deals with several challenges all the time. Some of these
challenges, i.e., disturbance/unexpected events and cyberattack, could happen
every current time, but there is challenge, i.e., climate change, which will effect
gradually on power system operation, and its result will be noticeable after several
decades. These challenges and their impacts on the power system operation are
summarized in the following.

2.3.1 Climate Change Impact on Power System Operation

Many various aspects (especially generation aspects) of energy sector is susceptible

to the influences of climate changes and environmental conditions. Some of these
interactions are summarized in below.
• Solar. Cloud cover will decline in low- to midlatitude regions [26]. While due to
rising temperature, decreasing efficiency will counterbalance the increase in solar
resources. Based on prediction of regional studies, solar generation changes will
be less than 10% by the end of the century [27–29].
• Wind. According to the several different studies, it is predicted that around
Europe and North America, average wind speeds would remain within three
different percent, 15% [30], 20% [31, 32], and 30% [33] of current values.
Moreover, it is evaluated that there will be no considerable change in wind
resources over China [34] or Southern Africa [35].
• Wave. Due to changes in wind pattern and sea-level rise, wave resources could be
affected. Studies indicated that wave generations have no changes in the Persian
Gulf [36] and Menorca [37]. In addition, wave generation changes in the UK will
be less than 3% [38].
• Hydropower. Global studies were done all over the world, and according to
results, their conclusions differ about the impact of rising temperatures and
changing precipitation patterns. In some studies, it is concluded that climate
changes will have little influences on total global resources potential [39, 40],
while the conclusion of other study indicated that due to climate changes, global
hydropower volume will decline up to 6.1% in the 2080s [41].
• Bioenergy. Rising temperature and changing precipitation patterns affected the
areas that appreciate to grow bioenergy and average yields. Yields will increase as
26 A. Ardeshiri et al.

much as +28% at high latitude and will decline as much as 16% at low latitude
in the 2050s [42]. Moreover, it is predicted to shift the suitable land northward for
several crops [43], switchgrass in the USA [44], and miscanthus globally [45].
• Thermal power stations. As a result of reduction in the thermal efficiencies due
to rising temperatures, power plant output will reduce about 0.4–0.7% per degree
[46–48]. The consequence of the decline in water resources for cooling of power
stations is reduction in the load, and it could be more serious which causes
shutdown of power stations.
• Transmission lines. Rising temperature will decrease the overhead lines’ trans-
mission capacity. For an instance in the USA, it is expected reduction in trans-
mission capacity of overhead lines will be approximately 5.8% at the peak
summer demand times [49].

2.3.2 Disturbance/Unexpected Events Impact on Power

System Operation

Unexpected events have considerable influences on power operation. Large or small

disturbances happen due to several reasons such as human fault, impact of extreme
weather events [50], and failure of power equipment due to production/installation
error or equipment aging [51]. As a result of improvement in the life span of power
equipment due to utilizing superconducting technologies, it is possible to reach more
reliability of power system operation and reduce the failure possibility of power
applications because of the aging factor [52, 53]. Moreover, some of other type of
disturbances, their possible causes, and their effect are listed in Table 2.1 [54]:
Numerous studies have so far addressed the issues of protection and fault
detection in transmission lines and distribution networks in the power systems
using deep learning and machine learning techniques.
Power systems mainly due to various environmental, mechanical, and electrical
factors suffer from two categories (11 types) of short circuit faults called asymmetric
faults (AG, BG, CG, ABG, BCG, ACG, AB, BC, and AC) and symmetric faults
(ABC and ABCG). In [55], a hybrid machine learning application based on mor-
phological median filter (MMF) and decision tree (DT) is utilized to identify and
classify the types of symmetric and asymmetric faults in the transmission line. In
other similar studies [56, 57], the DT method has been employed to identify the type
and location of symmetric and asymmetric faults related to the transmission line. In a
valuable study [58], deep learning applications called convolutional neural network
(CNN) and deep reinforcement learning have been used to identify the location of
single-phase to ground short-circuit faults in the transmission line. In this paper,
comparing and evaluating the results introduces the deep reinforcement learning
procedure as the superior method.
Identification and classification of tested faults in the IEEE 123 bus benchmark
distribution system using deep learning applications called the graph-CNN is done in
2 Introduction and Literature Review of Power System Challenges and Issues 27

Table 2.1 Disturbance and unexpected events and their features

Disturbance Symptoms/ Potential
type Description Possible causes effects solutions
Power outage Total interruption – Accidents, acts of – System shut- Uninterruptible
of electrical supply: nature, etc. which down power supply
– Momentary require the proper – Loss of com- (UPS)
outage lasts operation of utility puter/controller
03.5 cycles, 3 s equipment (fuses, memory
– Temporary reclosers, etc.) – Hardware
outage last from 3 s – Internal short cir- damage
to 1 min cuits requiring the – Product loss
– long-term out- proper operation of or damage
age last longer than a customer’s brea-
1 min kers and fuses.
Transient A sub-cycle distur- Surges are caused – Computer – Surge protec-
(surge) bance in the AC by storms (light- lockup, tors
waveform, ning); operation of processing –
evidenced by a utility fuses, errors and data Uninterruptible
sharp, brief discon- reclosers, and brea- loss – Power sup-
tinuity of the kers; and turning on – Burned cir- plies with built-
waveform or off large equip- cuit boards, in surge sup-
ment and capacitor electrical insu- pression
switching (cus- lation damage, – Isolation
tomer and utility). and equipment transformers
damage – Constant
voltage trans-
former
– Line reactors
Sag/swell Any short-term – Major equipment – Memory loss –
(0.5 cycles, 1 min) shutdown and/or and data errors Uninterruptible
decrease (sag) or restart – Equipment power supply
increase (swell) in – Short circuits shutdown – Constant
voltage. Sags – Utility equipment – Flickering voltage trans-
account for up to failure or utility lights formers
87% of all power switching – Motors – Voltage regu-
disturbances stalling or stop- lators
(according to a Bell ping and – Power elec-
Labs study). decreased tronic sag
motor life correctors
Noise An unwanted high- – Interference from – Lockup of –
frequency electrical radio or TV trans- sensitive equip- Uninterruptible
signal that alters the mission ment power supply
normal voltage pat- – Operation of – Data loss and – Isolation
tern (sine wave) electronic processing transformers
equipment errors – Power line
– Distorted filters
audio and
video reception
Harmonic The alteration of the Electronic ballasts – Overheating – Harmonic fil-
distortion normal voltage pat- and other nonlinear of electrical ters
tern (sine wave) due loads like switch- equipment and – Isolation
(continued)
28 A. Ardeshiri et al.

Table 2.1 (continued)

Disturbance Symptoms/ Potential
type Description Possible causes effects solutions
to equipment gen- mode power sup- wiring transformers
erating frequencies plies and variable – Decreased – Improved
other than the stan- frequency drives motor perfor- wiring and
dard 60 cycles per mance grounding
second – Improper – Isolated loads
operation of – Line reactors
breakers,
relays, or fuses
Undervoltage/ Any long-term Overloaded wiring – Dim or bright –
overvoltage change lasting more or equipment. lights Uninterruptible
than a minute Large load swings – Equipment power supply
below or above or improper trans- shutdown – Constant
normal voltage former settings. – Overheating voltage trans-
Undersized wiring of motors or formers
and faulty or poor lights – Verify electri-
electrical – Reduced effi- cal connections
connections ciency or life of and wiring
electrical – Relocate
equipment equipment
– Reduced volt-
age motor
starters and
voltage ride-
through
equipment

[59]. The method proposed in this paper converts the signals received from the
system into two-dimensional images and then considers them as the CNN input. In
[60], trip fault is identified in the China Southern power grid using the Support
Vector Machine (SVM) and Long Short-Term Memory (LSTM) techniques. The
performance evaluation of the proposed models in this paper introduces the LSTM
method as the superior solution. The localization of single-phase to ground short-
circuit faults in the distribution network is done by one of the deep learning
applications called Stacked Auto-Encoder in [61]. In another valuable study [62],
the identification of symmetric and asymmetric faults in the IEEE 39-bus distribu-
tion system is performed using the LSTM method. In this paper, the performance of
the proposed method is compared with other machine learning techniques called
SVM, DT, random forest (RF), and k-nearest neighbor (k-NN).
In addition to the abovementioned damages related to transmission lines and
distribution networks, equipment installed in the power systems such as power
transformers, electrical machines, and power electronics equipment are also dam-
aged by various factors. Timely detection and elimination of abnormalities in this
equipment can prevent serious damage to the power system. In this regard, the
applications of machine learning and deep learning are mainly utilized.
2 Introduction and Literature Review of Power System Challenges and Issues 29

The detection of mechanical faults related to the power transformer windings in

[63] has been done using the SVM technique. In [64], the CNN procedure is selected
as one of the deep learning applications for locating high impedance and low
impedance short-circuit faults in transformer windings. Minor short circuit fault
localization in transformer windings has been performed in [65] by using regression
version of SVM. In this study, early detection of fault location prevents serious
damage to the transformer windings and power system security.
In [66], the high-resistance grounding faults detection in a high-voltage direct
current (HVDC) transmission lines with considering double-end unsynchronized has
been performed by employing Hilbert-Huang (HH) transform and one-dimensional
CNN techniques. An optimized adaptive neuro-fuzzy inference system (ANFIS)
neural network and HH transform have been introduced in [67] for localization of
internal and external faults in a voltage-sourced converter-HVDC (VSC-HVDC).
Diagnosing three common transients including GF, lightning fault, and lightning
disturbance on a VSC-HVDC transmission lines has been done by employing one of
the deep learning applications called deep belief network in [68].

2.3.3 Cyberattack Impact on Power System Operation

Nowadays, in our modern digital society with complex and modern infrastructure,
the power systems are one of the most crucial components which play a role as a
backbone for its economic activities and securities. Modern power systems are
considerably dependent on communication systems for their operation, as a result,
susceptive to cyberattacks. Consequently, every country is interested to boost the
security of their power system operation against cyber risks and threats which could
cause cascading effects, power outage, and personal data breaches [69–71].
These cyberattack challenges affect sufficiency of TSOs and DSOs (Transmission
and Distribution System Operators) to insure flexibility and resilience, dependabil-
ity, stableness, security of the supply, and power quality for electricity’s consumers.
Power systems functions against risks of cyberattacks include several aspects, i.e.,
generation, transmission, and distribution stability and reliability, communication
between systems and equipment, information on the operating circumstance of
generation, transmission, distribution equipment, black start capability, and appli-
cation performance and ability to recover [72].
So far, many studies have investigated and detected the types of cyberattacks in
the power systems. Introducing and reviewing the challenges in cyber-physical
systems, especially the security issues of these systems, have been discussed and
evaluated in some valuable review papers [73–78]. The cyber-physical system
generalities have been described in detail in [73]. In [74], the reliability effects of
cyber-physical layers in the power system has been evaluated. In [75], after evalu-
ating the security of the cyber-physical system, the existing challenges in this regard
has been reviewed, and appropriate solutions are introduced. The effects of cyber-
physical attacks in the smart grid has been investigated in [76]. In a valuable review
30 A. Ardeshiri et al.

paper [77], the limitations, issues, and future trends of cyber-physical systems
security have been surveyed. Projection, prediction, and forecasting of various
types of cyberattacks in the power system are introduced and evaluated in [78]. A
review of the literature and evaluation of studies conducted in recent years shows
that many solutions have been introduced to cyberattacks detection in the power
systems so that machine learning and deep learning methods have found a special
place in this field [79–83].
In the following, susceptibility associated with three important power system
applications, i.e., state estimation, automatic generation control, and voltage control,
against cyberattack are summarized.
1. Attacks on State Estimation.
State estimation plays an important role in energy management application
(control centers) and many operational decisions [84, 85]. Control centers run an
elaborate algorithm to process measurement data from different locations such as
power injections, line flows, voltage measurements, and the status information of
breakers and switches. Based on the process result, the system states, i.e., voltage
magnitude and angle, were estimated. According to the mentioned estimation,
decisions such as ramping of generators, opening or closing of lines, and chang-
ing of transformer tapes are made. The attacks to state estimation can be done in
several ways like modifying the sensor data at the meter level, a cyber-
interference at the communication layer, and deceiving the operator with wrong
configuration of the topology. The operator can figure out the presence of
unreasonable data and omit them for estimation target if the attack is not arranged.
Although the estimation cannot be conceivable, in the case of system, it becomes
unobservable as a result of discarding too many measurements [86]. Totally, the
purpose behind leading attacks on state estimation of power system can cause
large-scale blackouts, for an instance, the case of Ukrainian attack of 2015.
Table 2.1 shows the disturbance type, the description, possible causes, symptoms,
effects, and potential solutions of these disturbances.
2. Attacks on Automatic Generation Control.
In an interconnected power system, due to automatic generation control
(AGC), the power output is regulated in which power system frequency changes
in defined limits, and the power exchange between neighboring zones remain
within acceptable scheduled values [85, 87]. The AGC operates such as a closed-
loop feedback control system and results in minimization of human intervening.
Control signals duty from the AGC should be in the range of several seconds.
Consequently, it is not possible that data validation algorithms are utilized, so
attackers are able to manipulate the measurements without detailed mathematics
[88]. As a result of the attack on the AGC, a loud generation mismatch will be
occurring which causes center of inertia motion in the generation frequencies. If
the frequency falls sufficient, underfrequency relays will react and isolate large
regions from the main grid. If maintenance of load generation balance in the
isolated regions will not happen, the insolation regions will go on the cumulative
manner and cause a large-scale load shedding or blackout [86].
2 Introduction and Literature Review of Power System Challenges and Issues 31

3. Attack on Voltage Control.

In the long distribution network, stability of the voltage is vital component.
The increase in loading and voltage drop across the lines can cause a cascading
effect and finally lead to collapse. Load Tap Changing (LTC) transformer are
utilized to enhance voltage stability of a distribution system. According to the
value of the voltage sensed by LTC transformer, the position of the tapes changes.
The voltage value at the end of line is sent through communication channel to the
control center where the LTC transformer is placed. As a result, the system is
vulnerable to attacks. In the two following ways, the sensor data can be played
[86]:
(a) The sensors are comprised to report values lower than the actual ones under
normal operation situations. Consequently, the actuators are excited to increase
the tap setting, and system operates inefficiency at a higher voltage.
(b) The voltage drops actually happen but the value data tampered in way that the
base case values are read. According to the read data, it is concluded that
everything is fine, but the voltage drops could increase as much as result in
voltage collapse.

2.4 Power System Control

One of the main functions of electricity is to convert energy from other sources to
electrical power and then deliver the energy to the consumers in near or far distances.
Usually, electricity is used in other converted types of energy, for example, mechan-
ical energy, light, and heat. One of the main advantages of electricity is that it can be
controlled relatively easier than other forms of energy with high reliability and
efficiency. A professionally controlled power system must have the following
conditions [1, 89]:
1. The power system must meet the fluctuating demand considering both active and
reactive power. Dissimilar to other forms of energy, electrical power cannot be
easily and efficiently stored in adequate amounts yet. Thus, sufficient reserve
including both spinning and non-spinning reserve must be considered and
controlled.
2. The predefined power system standards must be maintained at all times, known as
“Power System Quality,” which are frequency consistency, voltage consistence,
and high reliability [1].
In order to meet the aforementioned necessities, several complex devices are
needed to be integrated which is depicted in Fig. 2.3.
In a generator, the controlling system includes prime mover (governor) and
excitation controllers. Prime mover controller deals with speed standards and con-
trols input energy variables such as boiler temperature, flows, and pressure.
32 A. Ardeshiri et al.

Fig. 2.3 Power system control

Excitation control regulates and maintains the voltage of the generator and injects
reactive power [1].
System generation control sets the output power of the generators by calculating
the network losses and loads and then balancing the generator output power against
them. This helps the system to achieve the desired power exchange with neighboring
networks (tie flows) [1].
The control of transmission consists of voltage and power control elements,
which are switched reactors and capacitors, synchronous condensers, phase-shifter
transformers, Static Var Compensator (SVC), tap-changing transformers, and
HVDC transmission controllers [1, 89].
The aforementioned system concerns about the perfect performance of power
system by controlling the frequency and voltages of the grid while keeping the other
variables within a certain limit. In addition, these controllers have a huge effect on
the dynamic performance of the grid, which demonstrates the tolerance of the grid
against disturbances and faults [1, 90].
The objective of power systems control may be different, depending on the
system operating condition. For normal conditions, the aim of the controlling system
is to keep the voltage and frequency in the predetermined range and run the grid with
efficiency. However, in abnormal conditions like contingencies or blackouts, the
primary goal of the controlling system is to restore the grid and try to go back to the
normal condition. In this case, efficiency and cost are the secondary concerns. These
abnormal conditions are caused by single disturbances, leading to the collapse of the
grid, which was known to be a secure system earlier. Human error, equipment and
element malfunction, extreme climate events (such as heavy snow or rain, thunder,
2 Introduction and Literature Review of Power System Challenges and Issues 33

Fig. 2.4 Power system

conditions Normal Emergency

Alert

Restorave Extreme

or tornado), and poor and insecure power network design may result in cascading
breakdown and outages [1, 90].
In this chapter, five system conditions are introduced which are Normal, Alert,
Emergency, Extreme, and Restorative. These condition are depicted in Fig. 2.4
[1, 90].
The system parameters are within the normal limit, and no element is overloaded.
The grid performs efficiently with security and is able to handle a single contingency
with no constraint violation [1, 90].
In alert condition, the security of the grid reduces although the system parameters
are still within the standard range. However, the parameters are prone to go below or
above the normal limits due to possible upcoming conditions such as thunder or
storms. An equipment overload leads the grid to emergency condition; however, if
the disturbance is harsh, it will place the grid directly in extreme condition. In order
to bring the grid back to the normal condition, preventive actions such as increased
generation or generation shifting can be considered [1].
The system will move from alert to emergency condition if enough harsh
disturbances happen in alert mode. In emergency condition, the equipment are
overloaded, and voltage of buses are out of the specified range; however, the system
can be returned to alert condition by commencing emergency control, which are load
curtailment, excitation control, fault clearing, generation runback and tripping,
HVDC modulation, and fast valving [1].
If the aforementioned actions are not carried out on time or ineffectively, the grid
will enter the extreme condition, in which cascading outages and partial grid
shutdown will happen. In this condition, system separation and load shedding are
the solution to prevent the blackout of the entire grid [1].
After extreme condition, a set of measures are taken to reconnect all the equip-
ment and loads to the grid and restore the system. The system might enter normal or
alert condition based on the taken actions and grid condition [1].
Obtaining a framework for control strategies and operator manual for each
condition for power system is necessary to handle each condition fast and efficiently.
Power system control will help the operator to restore the grid back to its normal
condition. If the disturbance is not large, power system controls can deal with the
34 A. Ardeshiri et al.

matter; nonetheless, if the disturbance is huge, the operator needs to take measures,
such as equipment switching or generator dispatch to bring the grid back to normal
condition [1].

2.4.1 Power Grid Criteria for Control and Stability

A reliable power grid must be intact and solid against a different number of
disturbances. So it is vital that the decision makers design the grid and operate it
so that system can tolerate the disturbances and still be able to operate within the
predetermined boundaries and supple the demand with minimum load loss, espe-
cially without cascading outages [1, 91].
Determining design, operating, and controlling criteria is of paramount impor-
tance in avoiding system disturbances after harsh contingencies. The implementation
of criteria will guarantee that in the worst-case scenario, for all regular happening
contingencies, the system will enter from normal condition to alert, instead of
entering the other critical conditions, which are emergency and extreme conditions.
After the grid enters the alert condition, the operators will commence taking steps to
return the grid back to normal condition [1].

2.4.2 Regular Contingencies

These types of contingencies have common and have high probability of happening.
These contingencies include:
1. Three-phase fault on any transmission line, generator, transformer, or bus bar,
including fault clearing and reclosing.
2. Phase to ground fault on any transmission line, generator, transformer, circuit
breaker, or bus bar due to circuit breaker’s malfunction, signal channel, or relay.
3. Phase to ground fault on any phase, on any tower, which is cleared quickly and in
normal time.
4. Any equipment loss without fault.
5. Simultaneous loss of DC bipolar generator poles.
Following the aforementioned faults, the criteria necessitates that the grid be
sustained and loading and voltages of equipment be kept within the standard range.
The standards are applied to the following situations [1].
1. All live and active equipment.
2. An equipment out of service, including a transmission circuit, generator, or a
transformer.
2 Introduction and Literature Review of Power System Challenges and Issues 35

2.4.3 Extreme Contingencies

These contingencies increase the risk of exceeding the voltage or loading of equip-
ment. However, the probability of these contingencies happening is low. The
extreme contingencies are as follows:
1. Loss of a whole generation station capacity.
2. Loss of all transmission lines in a right of way, connecting the generation,
substation, and switching station to each other.
3. A three-phase fault on any transmission line, generator, transformer, or bus bar,
without quick fault clearing due to malfunction in reclosing equipment, causing
angle swing in generators.
4. An unexpected adding or removal of a large load, or malfunction of a special
protection system, which is generation or load rejection, to name a few [1].

2.4.4 Frequency Control

Frequency in power grid is a fluctuating variable, which demonstrates the balance of

load and generation [92]. In fact, frequency deviation is caused by the imbalance of
load and generation, which will cause equipment damaging, protection devices
triggering, transmission lines overloading, and deficiency in power system opera-
tion, reliability, and security [90]. Frequency is controlled and kept within the
predetermined and standard range by the system operator. In the UK, two frequency
ranges are introduced, which are operational range (49.8–50.2 Hz) and statutory
range (49.5–50.5 Hz). Table 2.2 shows the case description of frequency deviations
in UK [92, 93]. As a result of integration of renewable energy generators, frequency
control by controlling conventional power plants becomes more difficult.
The frequency of power grid is proportional to the generator’s rotor speed, so
frequency control can be done by controlling the speed of generator turbine. This can
be done by implementing governing control systems that monitor the rotor speed and
follow the load change and redirect the frequency back to the standard value [90, 94].
Depending on the range of frequency deviation, different control loops are
needed to restore the grid frequency stability, which are presented in Fig. 2.5.

Table 2.2 Frequency range and the description of required action to restore the frequency
Frequency range
(Hz) Description
49.8–50.2 Frequency within the normal range and maximum frequency deviation hap-
pens with a loss of generation or load up to 300 MW
49.5–50.5 Maximum frequency deviation for loss of generation or load more than
300 MW and less than 1320 MW
Less than 49.2 Maximum frequency deviation for loss of generation more than 1320 MW
and less than 1800 MW. Frequency must be returned to 49.5 Hz in 1 min
36 A. Ardeshiri et al.

Fig. 2.5 Frequency control system [90]

When the load and generation become unbalanced, at the first step for very small
frequency deviation, inertial response is activated, where the kinetic energy is stored
in the rotor of the generators and lasts for a few seconds [95]. If the frequency
deviation goes beyond a certain limit, primary control responds for small frequency
deviation in normal condition. Secondary control, known as Load Frequency Con-
trol (LFC), comes into action considering the available reserve power when larger
frequency deviations happen in non-normal conditions [90, 94]. If the secondary
control loop is not enough, a tertiary control is implemented manually, and the
frequency is brought back to the standard value by unit commitment and generation
dispatch [95].
Two of the main objectives of LFC is to maintain grid frequency and control
power interchanges with neighboring areas [90].
Nowadays, frequency control has become more important due to the changing of
structure, complexity, and size of power grid. In addition, LFC has led to reducing
power system economy and reliability pressure by controlling grid frequency and
tie-line flows closer to predetermined values. This helps power grid with energy
exchanges and electricity trading.
2 Introduction and Literature Review of Power System Challenges and Issues 37

2.4.5 Voltage and Angle Control

The three integral parameters of power grid are nodal voltage magnitudes, nodal
voltage angles, and grid frequency. Also, these parameters define the condition in
which the grid situates in a particular time. Voltage and angle stability can be
grouped into small and large disturbance stability. This stability points out to the
damping of power swings in any part of the grid and voltage deviation beyond a
certain predetermined value. Voltage and angle stability can be maintained sharply
by using specific control devices which are implemented in power grid and acquire
the dynamic response of the grid. These devices are FACTS, AVR, and PSS [90].
The generators operate at a fixed voltage by implementing AVR. AVR controls
the excitation current of the generator. The direct current of the generator generates
flux for the rotor. PSS is a controller that is located within the turbine-governing
system and creates a complementary control loop to AVR system in a generator. A
general implementation of PSS and AVR is shown in Fig. 2.6 [90].
ΔVPSS is zero in the steady state so that it does not cause any distortion in the
voltage regulation process. However, the generator speed is not constant in the
transient state as the rotor swing cause ΔV to change. The voltage change is reduced
when PSS generates a damping signal ΔVPSS that is in phase with generator speed
change Δw [90].
Voltage control consists of multiple control loops on several system levels. AVR
loop maintains the generator terminal voltages besides on lower system levels and
reacts in less than a second. Meanwhile, the secondary voltage control that set the
voltage reference value is activated in tens of seconds or minutes. Secondary voltage
control is used to coordinate setting of set points of AVR and several reactive power
sources in a power grid to maintain the stability. Moreover, voltage stability can be
increased by implementing higher control levels which respond in several minutes
and are called as tertiary voltage control. Tertiary voltage control mainly considers
the grid economic optimization and tries to minimize the cost of generation [90, 95].
Many challenges exist in voltage and frequency control in power grid. A number
of these challenges are explained below [95, 96]:

Input

PSS

vref Emax
ΔvPSS
–
v + + AVR Exciter ΔEf
Δv ΔΣ
Emin

Fig. 2.6 Voltage control system [90]

38 A. Ardeshiri et al.

1. Calculating and specifying AVR set points and input signals:

For voltage control, AVR set points are not determined globally for the best
performance of the power grid. Also, supplementary signals are on local contem-
plation in order to reach better stability. The need exists to create a thorough and
global strategy for voltage stability and control for dynamic and static conditions.
2. Placement and control of FACTS:
One of the biggest concerns of power system voltage control is to figure out
the best locations and sites for reactive power control systems to support the grid
during normal conditions and contingencies.
3. Setting the standard limits for grid parameters:
Different components and equipment of power grid must be held within a
certain standard limit. Having in mind that different standard limits have been
introduced, these standard limits must be determined through analysis tools
which lead to the best operation condition.
4. Novel security idea:
The minimum number of equipment and system components that can be lost
without cascading failure must be determined. This is separate from N-1 security
method. This novel security idea introduces the number of outages and contin-
gencies need to cause a blackout.
5. Fast reactive power calculation in contingencies:
Traditional linear contingency analysis overlooks reactive power and voltage
challenges. The issue of fast computation of contingencies’ impact on voltage
exists.
6. New control necessities:
Due to the high penetration of renewable energy sources and nonsynchronous
generators which are based on converters in order to be able to connect to the grid,
these generators do not provide high inertia during frequency deviations. Thus,
new control systems are needed to respond faster with higher precision to the
disturbances.
The power system is a nonlinear system that its performance is affected by a wide
variety of equipment and devices, which have various response time and character-
istics. The features and characteristics of every major equipment and device of the
power grid have an impact on the system stability. Thus, in order to examine and
learn power system stability and control, a good grasp of these characteristics is vital.
Another challenge of power system control is the complexity of physical side of
different categories of power system, their assessment methods, and action to
increase the performance of power system stability [1].
2 Introduction and Literature Review of Power System Challenges and Issues 39

References

1. P. Kundur, Power System Stability and Control (McGraw-Hill, New York, 1993)
2. M. Fotuhi-Firuzabad, A. Safdarian, M. Moeini-Aghtaie, R. Ghorani, M. Rastegar, H. Farzin,
Upcoming challenges of future electric power systems: sustainability and resiliency. Sci. Iranica
23, 1565 (2016)
3. IEA, World Energy Outlook (2019)
4. A. Moradzadeh, O. Sadeghian, K. Pourhossein, B. Mohammadi-Ivatloo, A. Anvari-
Moghaddam, Improving residential load disaggregation for sustainable development of energy
via principal component analysis. Sustainability (Switzerland) 12(8), 3158 (2020). https://doi.
org/10.3390/SU12083158
5. C. Canizares, J. Nathwani, D. Kammen, Electricity for all: issues, challenges, and solutions for
energy-disadvantaged communities, in Proceedings of the IEEE, vol. 107 (2019)
6. A. Moradzadeh, A. Mansour-Saatloo, B. Mohammadi-Ivatloo, A. Anvari-Moghaddam, Perfor-
mance evaluation of two machine learning techniques in heating and cooling loads forecasting
of residential buildings. Appl. Sci. (Switzerland) 10(11), 3829 (2020). https://doi.org/10.3390/
app10113829
7. A. Lotfi, B. Mohammadi-Ivatloo, S. Asadi, Introduction to FEW Nexus, in Food-Energy-Water
Nexus Resilience and Sustainable Development, ed. by S. Asadi, B. Mohammadi-Ivatloo,
(Springer, Switzerland, 2020)
8. O. Sadeghian, A. Moradzadeh, B. Mohammadi-Ivatloo, M. Abapour, F.P.G. Marquez, Gener-
ation units maintenance in combined heat and power integrated systems using the mixed integer
quadratic programming approach. Energies 13(11), 2840 (2020). https://doi.org/10.3390/
en13112840
9. A. Demir, N. Hadžijahić, Power system planning: part I—basic principles, in Advanced
Technologies, Systems, and Applications II, ed. by M. Hadžikadić, S. Avdaković, vol.
28, (Springer, New York, 2018)
10. A.M. Al-Shaalan, Essential aspects of power system planning in developing countries. J. King
Saud Univ. Eng. Sci. 23, 27–32 (2011)
11. A. Lotfi, S.H. Hosseini, Composite distributed generation and transmission expansion planning
considering security. World Acad. Sci. Eng. Technol. Int. J. Energy Power Eng. 11 (2017)
12. A.J. Conejo, L. Baringo, Power Systems. In Power System Operations, pp. 1–15, (Springer,
Cham, 2018)
13. X. Wang, J.R. McDonald, Modern Power System Planning (McGraw-Hill, New York, 1994)
14. A. Moradzadeh, K. Khaffafi, Comparison and evaluation of the performance of various types of
neural networks for planning issues related to optimal management of charging and discharging
electric cars in intelligent power grids. Emerg. Sci. J. 1(4), 201–207 (2017). https://doi.org/10.
28991/ijse-01123
15. A. Moradzadeh, B. Mohammadi-Ivatloo, M. Abapour, A. Anvari-Moghaddam, S. Gholami
Farkoush, S.B. Rhee, A practical solution based on convolutional neural network for
non-intrusive load monitoring. J. Ambient Intell. Humaniz. Comput. (2021). https://doi.org/
10.1007/s12652-020-02720-6
16. H. Seifi, M.S. Sepasian, Electric Power System Planning: Issues, Algorithms and Solutions
(Springer, New York, 2011)
17. R. Hemmati, R.A. Hooshmand, A. Khodabakhshian, Comprehensive review of generation and
transmission expansion planning. IET Gener. Transm. Distrib. 7(9), 955–964 (2013)
18. G. Latorre, R.D. Crus, J.M. Areiza, A. Villegas, Classification of publications and models on
transmission expansion planning. IEEE Trans. Power Syst. 18(2), 938–946 (2003)
19. R.S. Najafi, H. Khatami, Optimal and robust distribution system planning to forecasting
uncertainty. Tabriz J. Electr. Eng. 46(2), 323–332 (2016)
20. A. Moradzadeh, K. Pourhossein, Early detection of turn-to-turn faults in power transformer
winding: an experimental study, in Proceedings 2019 International Aegean Conference on
Electrical Machines and Power Electronics, ACEMP 2019 and 2019 International Conference
40 A. Ardeshiri et al.

on Optimization of Electrical and Electronic Equipment, OPTIM 2019 (2019), pp. 199–204,
https://doi.org/10.1109/ACEMP-OPTIM44294.2019.9007169
21. A. Mansour-Saatloo, A. Moradzadeh, B. Mohammadi-Ivatloo, A. Ahmadian, A. Elkamel,
Machine learning based PEVs load extraction and analysis. Electronics (Switzerland) 9(7),
1–15 (2020). https://doi.org/10.3390/electronics9071150
22. S.N. Ravandanegh, N. Jahanyari, A. Amini, N. Taghizadeghan, Smart distribution grid multi-
stage expansion planning under load forecasting uncertainty. IET Gener. Transm. Distrib. 10
(5), 1136–1144 (2016)
23. P. Prakash, D.K. Khatod, Optimal sizing and siting techniques for distributed generation in
distribution systems: A review. Renew. Sust. Energ. Rev. 57, 111–130 (2016)
24. A.R. Jordehi, Allocation of distributed generation units in electric power systems: A review.
Renew. Sust. Energ. Rev. 56, 893–905 (2016)
25. J.P. Lopes, N. Hatziargyriou, J. Mutale, P. Djapic, N. Jenkins, Integrating distributed generation
into electric power systems: A review of drivers, challenges and opportunities. Electr. Power
Syst. Res. 77(9), 1189–1203 (2007)
26. A. Patt, S. Pfenninger, J. Lilliestam, Vulnerability of solar energy infrastructure and output to
climate change. Clim. Change 121, 93–102 (2013). https://doi.org/10.1007/s10584-013-0887-0
27. J.A. Crook, L.A. Jones, M. Forster, R. Crook, Climate change impacts on future photovoltaic
and concentrated solar power energy output. Energy Environ. Sci. 4, 3101–3109 (2011). https://
doi.org/10.1039/c1ee01495a
28. M. Gaetani, T. Huld, E. Vignati, F. Monforti-ferrario, A. Dosio, F. Raes, The near future
availability of photovoltaic energy in Europe and Africa in climate-aerosol modeling experi-
ments. Renew. Sust. Energ. Rev. 38, 706–716 (2014). https://doi.org/10.1016/j.rser.2014.07.
041
29. I.S. Panagea, I.K. Tsanis, A.G. Koutroulis, M.G. Grillakis, Climate change impact on photo-
voltaic energy output : the case of Greece. Adv Meteorol 2014, 63–86 (2014)
30. S.C. Pryor, R.J. Barthelmie, Climate change impacts on wind energy: a review. Renewable and
sustainable energy reviews 14, 430–437 (2010). https://doi.org/10.1016/j.rser.2009.07.028
31. I. Tobin et al., Assessing climate change impacts on European wind energy from ENSEMBLES
high-resolution climate projections. Clim. Change 128, 99–112 (2015). https://doi.org/10.1007/
s10584-014-1291-0
32. R. Davy, N. Gnatiuk, L. Pettersson, L. Bobylev, Climate change impacts on wind energy
potential in the European domain with a focus on the Black Sea. Renew. Sustain. Energy Rev.
2016, 1–8 (2017). https://doi.org/10.1016/j.rser.2017.05.253
33. C.S. Santos, D. Carvalho, A. Rocha, M. Gómez-Gesteira, Potential impacts of climate change
on European wind energy resource under the CMIP5 future climate projections. Renew. Energy
101(2017), 29–40 (2020). https://doi.org/10.1016/j.renene.2016.08.036
34. L. Chen, S.C. Pryor, D. Li, Assessing the performance of Intergovernmental Panel on Climate
Change AR5 climate models in simulating and projecting wind speeds over China. Journal of
Geophysical Research: Atmospheres 117, 1–15 (2012). https://doi.org/10.1029/2012JD017533
35. C. Fant, C.A. Schlosser, K. Strzepek, The impact of climate change on wind and solar resources
in southern Africa. Appl. Energy (2015). https://doi.org/10.1016/j.apenergy.2015.03.042
36. B. Kamranzad, A. Etemad-shahidi, V. Chegini, Climate change impact on wave energy in the
Persian Gulf. (2015). https://doi.org/10.1007/s10236-015-0833-y
37. J.P. Sierra, M. Casas-prat, E. Campins, Impact of climate change on wave energy resource : the
case of Menorca (Spain). Renew. Energy 101, 275–285 (2017). https://doi.org/10.1016/j.
renene.2016.08.060
38. D.E. Reeve, Y. Chen, S. Pan, V. Magar, D.J. Simmonds, A. Zacharioudaki, An investigation of
the impacts of climate change on wave energy generation : The Wave Hub, Cornwall,
UK. Renew. Energy 36(9), 2404–2413 (2011). https://doi.org/10.1016/j.renene.2011.02.020
39. B. Hamududu, A. Killingtveit, E. Engineering, Assessing climate change impacts on global
hydropower. Energies 5(2), 305–322 (2012). https://doi.org/10.3390/en5020305
2 Introduction and Literature Review of Power System Challenges and Issues 41

40. S.W.D. Turner, J. Yi, S. Galelli, Science of the total environment examining global electricity
supply vulnerability to climate change using a high-fidelity hydropower dam model. Sci. Total
Environ. (2017). https://doi.org/10.1016/j.scitotenv.2017.03.022
41. M.T.H. Van Vliet, D. Wiberg, S. Leduc, K. Riahi, Power-generation system vulnerability and
adaptation to changes in climate and water resources. Nature Climate Change 6(4), 375–380
(2016). https://doi.org/10.1038/NCLIMATE2903
42. H. Haberl et al., Global bioenergy potentials from agricultural land in 2050: sensitivity to
climate change, diets and yields. Biomass and bioenergy 35(12), 4753–4769 (2011). https://doi.
org/10.1016/j.biombioe.2011.04.035
43. G. Tuck, M.J. Glendining, P. Smith, J.I. House, M. Wattenbach, The potential distribution of
bioenergy crops in Europe under present and future climate. Biomass Bioenergy 30, 183–197
(2006). https://doi.org/10.1016/j.biombioe.2005.11.019
44. J.N. Barney, J.M. Ditomaso, Bioclimatic predictions of habitat suitability for the biofuel
switchgrass in North America under current and future climate scenarios. Biomass Bioenergy
34(1), 124–133 (2010). https://doi.org/10.1016/j.biombioe.2009.10.009
45. H.A. Hager, S.E. Sinasac, Z. Gedalof, J.A. Newman, Predicting potential global distributions of
two miscanthus grasses : implications for horticulture, biofuel production, and biological
invasions. PLoS One 9(6), e100032 (2014). https://doi.org/10.1371/journal.pone.0100032
46. C. Chuang, D. Sue, Performance effects of combined cycle power plant with variable condenser
pressure and loading. Energy 30, 1793–1801 (2005). https://doi.org/10.1016/j.energy.2004.10.
003
47. A. Durmayaz, O.S. Sogut, Influence of cooling water temperature on the efficiency of a
pressurized-water reactor nuclear-power plant. International Journal of Energy Research,
2005, 799–810 (2006). https://doi.org/10.1002/er.1186
48. K. Linnerud, T.K. Mideksa, G.S. Eskeland, The impact of climate change on nuclear power
supply. Energy J. 32, 149–168 (2011)
49. M. Bartos et al., Environ. Res. Lett. 11 (2016)
50. R. Contreras-Lisperguer, K. De-Cuba, The potential impact of climate change on the energy
sector in the Caribbean region. Organization of American States, Washington DC (2008)
51. W. Li, E. Vaahedi, P. Choudhury, Power system equipment aging. IEEE Power Energy Mag 4
(3), 52–58 (2006). https://doi.org/10.1109/MPAE.2006.1632454
52. A. Moradnouri, A. Ardeshiri, M. Vakilian, A. Hekmati, M. Fardmanesh, Survey on high-
temperature superconducting transformer windings design. J. Superconductivity Novel Magnet.
33, 2581–2599 (2020). https://doi.org/10.1007/s10948-020-05539-6
53. S.S. Kalsi, Application of High-Temperature Superconductors to Electric Power Equipment
(IEEE Press, Wiley, 2011)
54. We Energies, Disturbance types and solutions [Online], https://www.we-energies.com/safety/
power-quality/disturbance-types
55. R. Godse, S. Bhat, Mathematical morphology-based feature-extraction technique for detection
and classification of faults on power transmission line. IEEE Access 8, 38459–38471 (2020).
https://doi.org/10.1109/ACCESS.2020.2975431
56. M.M. Taheri, H. Seyedi, B. Mohammadi-ivatloo, DT-based relaying scheme for fault classifi-
cation in transmission lines using MODP. IET Generation Transm. Distrib. 11(11), 2796–2804
(2017). https://doi.org/10.1049/iet-gtd.2016.1821
57. M. Mohammad Taheri, H. Seyedi, M. Nojavan, M. Khoshbouy, B. Mohammadi Ivatloo, High-
speed decision tree based series-compensated transmission lines protection using differential
phase angle of superimposed current. IEEE Trans. Power Deliv. 33(6), 3130–3138 (2018).
https://doi.org/10.1109/TPWRD.2018.2861841
58. H. Teimourzadeh, A. Moradzadeh, M. Shoaran, B. Mohammadi-Ivatloo, R. Razzaghi, High
impedance single-phase faults diagnosis in transmission lines via deep reinforcement learning
of transfer functions. IEEE Access (2021). https://doi.org/10.1109/ACCESS.2021.3051411
42 A. Ardeshiri et al.

59. K. Chen, J. Hu, Y. Zhang, Z. Yu, J. He, Fault location in power distribution systems via deep
graph convolutional networks. IEEE J. Sel. Areas Commun. 38(1), 119–131 (2020). https://doi.
org/10.1109/JSAC.2019.2951964
60. S. Zhang, Y. Wang, M. Liu, Z. Bao, Data-based line trip fault prediction in power systems using
LSTM networks and SVM. IEEE Access 6, 7675–7686 (2018). https://doi.org/10.1109/
ACCESS.2017.2785763
61. G. Luo, Y. Tan, M. Li, M. Cheng, Y. Liu, J. He, Stacked auto-encoder-based fault location in
distribution network. IEEE Access 8, 28043–28053 (2020). https://doi.org/10.1109/ACCESS.
2020.2971582
62. B. Li, J. Wu, L. Hao, M. Shao, R. Zhang, W. Zhao, Anti-jitter and refined power system
transient stability assessment based on long-short term memory network. IEEE Access 8,
35231–35244 (2020). https://doi.org/10.1109/ACCESS.2020.2974915
63. J. Liu, Z. Zhao, C. Tang, C. Yao, C. Li, S. Islam, Classifying transformer winding deformation
fault types and degrees using FRA based on support vector machine. IEEE Access 7,
112494–112504 (2019). https://doi.org/10.1109/access.2019.2932497
64. A. Moradzadeh, K. Pourhossein, Short circuit location in transformer winding using deep
learning of its frequency responses, in Proceedings 2019 International Aegean Conference
on Electrical Machines and Power Electronics, ACEMP 2019 and 2019 International Confer-
ence on Optimization of Electrical and Electronic Equipment, OPTIM 2019 (2019),
pp. 268–273, https://doi.org/10.1109/ACEMP-OPTIM44294.2019.9007176
65. A. Moradzadeh, K. Pourhossein, Application of support vector machines to locate minor short
circuits in transformer windings, in 2019 54th International Universities Power Engineering
Conference (UPEC), (2019), pp. 1–6
66. S. Lan, M.-J. Chen, D.-Y. Chen, A novel HVDC double-terminal non-synchronous fault
location method based on convolutional neural network. IEEE Trans. Power Deliv. 34(3),
848–857 (2019). https://doi.org/10.1109/TPWRD.2019.2901594
67. R. Rohani, A. Koochaki, A hybrid method based on optimized neuro-fuzzy system and
effective features for fault location in VSC-HVDC systems. IEEE Access 8, 70861–70869
(2020). https://doi.org/10.1109/ACCESS.2020.2986919
68. G. Luo, J. Hei, C. Yao, J. He, M. Li, An end-to-end transient recognition method for
VSC-HVDC based on deep belief network. J. Mod. Power Syst. Clean Energy 8(6),
1070–1079 (2020). https://doi.org/10.35833/MPCE.2020.000190
69. SGTF_EG2, 2nd Interim Report Recommendations for the European Commission on Imple-
mentation of a Network Code on Cybersecurity (2018)
70. The European Economic and Social Committee and the Committee of the Regions Cybersecu-
rity strategy of the E. U. European Commission. Joint communication to the European parlia-
ment, the council, An open, safe and secure cyberspace (2013)
71. ANL_GSS_15/4, Analysis of critical infrastructure dependencies and interdependencies,
Argonne-risk and infrastructure science center, (2015)
72. A. Dagoumas, Assessing the impact of cybersecurity attacks on power systems. Energies
(2019). https://doi.org/10.3390/en12040725
73. A. Humayed, J. Lin, F. Li, B. Luo, Cyber-physical systems security - a survey. IEEE Internet
Things J. 4(6), 1802–1831 (2017). https://doi.org/10.1109/JIOT.2017.2703172
74. B. Jimada-Ojuolape, J. Teh, Surveys on the reliability impacts of power system cyber–physical
layers. Sustain. Cities Soc. 62, 102384 (2020). https://doi.org/10.1016/j.scs.2020.102384
75. Y. Ashibani, Q.H. Mahmoud, Cyber physical systems security: Analysis, challenges and
solutions. Comput. Secur. 68, 81–97 (2017). https://doi.org/10.1016/j.cose.2017.04.005
76. H. He, J. Yan, Cyber-physical attacks and defences in the smart grid: A survey. IET Cyber-
Phys. Syst. Theory Appl. 1(1), 13–27 (2016). https://doi.org/10.1049/iet-cps.2016.0019
77. J.P.A. Yaacoub, O. Salman, H.N. Noura, N. Kaaniche, A. Chehab, M. Malli, Cyber-physical
systems security: Limitations, issues and future trends. Microprocess. Microsyst. 77, 103201
(2020). https://doi.org/10.1016/j.micpro.2020.103201
2 Introduction and Literature Review of Power System Challenges and Issues 43

78. M. Husak, J. Komarkova, E. Bou-Harb, P. Celeda, Survey of attack projection, prediction, and
forecasting in cyber security. IEEE Commun. Surv. Tutorials 21(1), 640–660 (2019). https://
doi.org/10.1109/COMST.2018.2871866
79. Y. Wang, M.M. Amin, J. Fu, H.B. Moussa, A novel data analytical approach for false data
injection cyber-physical attack mitigation in smart grids. IEEE Access 5, 26022–26033 (2017).
https://doi.org/10.1109/ACCESS.2017.2769099
80. H. Karimipour, A. Dehghantanha, R.M. Parizi, K.-K.R. Choo, H. Leung, A deep and scalable
unsupervised machine learning system for cyber-attack detection in large-scale smart grids.
IEEE Access 7, 80778–80788 (2019). https://doi.org/10.1109/ACCESS.2019.2920326
81. J.J.Q. Yu, Y. Hou, V.O.K. Li, Online false data injection attack detection with wavelet
transform and deep neural networks. IEEE Trans. Ind. Informat. 14(7), 3271–3280 (2018).
https://doi.org/10.1109/TII.2018.2825243
82. A. Al-Abassi, H. Karimipour, A. Dehghantanha, R.M. Parizi, An ensemble deep learning-based
cyber-attack detection in industrial control system. IEEE Access 8, 83965–83973 (2020).
https://doi.org/10.1109/ACCESS.2020.2992249
83. S. Soltan, P. Mittal, H.V. Poor, Line failure detection after a cyber-physical attack on the grid
using Bayesian regression. IEEE Trans. Power Syst. 34(5), 3758–3768 (2019). https://doi.org/
10.1109/TPWRS.2019.2910396
84. F.C. Schweppe, J. Wildes, Power system static-state estimation, part i: Exact model. IEEE
Trans. Power Apparatus Syst. 59(1), 120–125 (1970)
85. A.J. Wood, B.F. Wollenberg, Power Generation Operation and Control (Wiley, New York,
2003)
86. K. Chatterjee, V. Padmini, S.A. Khaparde, Review of cyber attacks on power system opera-
tions, in IEEE Region 10 Symposium, Conference Paper, (2017)
87. D. P. Kothari and I. J. Padmini, Power System Engineering, New Delhi: Tata McGraw Hill
Education, 2008
88. P.M. Esfahani, M. Vrakopoulou, K. Margellos, J. Lygeros, G. Andersson, Cyber Attack in a
Two-Area Power System : Impact Identification using Reachability, In Proceedings of the 2010
American control conference, pp. 962–967. IEEE (2010)
89. B.F. Wollenberg, Power system operation and control, in Power System Stability and Control,
3rd edn., (CRC Press, 2017). https://doi.org/10.4324/b12113
90. H. Bevrani, Robust Power System Frequency Control (Power Electronics and Power Systems)
(Springer, New York, 2009)
91. A. Moradzadeh, K. Pourhossein, B. Mohammadi-Ivatloo, F. Mohammadi, Locating inter-turn
faults in transformer windings using isometric feature mapping of frequency response traces.
IEEE Trans. Ind. Informat., 17, 1–1 (2020). https://doi.org/10.1109/tii.2020.3016966
92. Z.A. Obaid, L.M. Cipcigan, L. Abrahim, M.T. Muhssin, Frequency control of future power
systems: Reviewing and evaluating challenges and new control methods. J. Mod. Power Syst.
Clean Energy 7(1), 9–25 (2019). https://doi.org/10.1007/s40565-018-0441-1
93. F. Teng, Y. Mu, H. Jia, J. Wu, P. Zeng, G. Strbac, Challenges of primary frequency control and
benefits of primary frequency response support from electric vehicles. Energy Procedia 88,
985–990 (2016). https://doi.org/10.1016/j.egypro.2016.06.123
94. M.J. Bryant, R. Ghanbari, M. Jalili, P. Sokolowski, L. Meegahapola, Frequency Control
Challenges in Power Systems with High Renewable Power Generation: An Australian Perspec-
tive, RMIT University (2019)
95. H.T. Nguyen, G. Yang, A.H. Nielsen, P.H. Jensen, Challenges and research opportunities of
frequency control in low inertia systems, in E3S Web of Conferences, vol. 115, (2019). https://
doi.org/10.1051/e3sconf/201911502001
96. P.W. Sauer, Reactive power and voltage control issues in electric power systems, in Applied
Mathematics for Restructured Electric Power Systems. Power Electronics and Power Systems,
ed. by J. H. Chow, F. F. Wu, J. Momoh, (Springer, Boston, 2005)
Chapter 3
Machine Learning and Power System
Planning: Opportunities and Challenges

Mohammad Hosein Asgharinejad Keisami, Sasan Azad,

Reza Mohammadi Chabanloo, Morteza Nazari-Heris, and Somayeh Asadi

Nomenclature

b, bi, bC, bf, bo Corresponding biases

Ct Cell state at time t
et
C Vector of new candidate values
f(x) The plane that fits closest to the data
Ht Output at timestamp t
n Total number of features
st State at timestamp t
Wi, W f Weight matrixes corresponding to the input and forget gates
WT Weights
xi Input label at time stamp i
xt Input at timestamp t
yi Output label at time stamp i
y
j,t The predicted value of feature j at time t
yj, t Actual value of feature j at time t
ρ Correlation coefficient
σ Standard deviation

M. H. Asgharinejad Keisami · S. Azad (*) · R. Mohammadi Chabanloo

Department of Electrical Engineering, Shahid Beheshti University, Tehran, Iran
e-mail: [email protected]; [email protected]; [email protected]
M. Nazari-Heris · S. Asadi
Department of Architectural Engineering, Pennsylvania State University,
University Park, PA, USA
e-mail: [email protected]; [email protected]; [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 45

M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_3
46 M. H. Asgharinejad Keisami et al.

3.1 Introduction

The application of machine learning (ML) as a modern method to solve engineering

problems has been the center of attention in most research topics, especially power
systems. The underlying need for using such methods can be traced to the rise of
intermittency and uncertainty in the power systems. This increased unpredictability
of power consumption, of course, can be attributed to a variety of different factors.
These factors include but are not limited to the unpredictable nature of consumer
behavior, the economic development of countries and increased gross domestic
product (GDP), the increasing penetration of modern technologies in developing
societies such as the Internet, and the subsequent increase in demand. The exponen-
tial growth of using renewables as a clean, free alternative to fossil fuel resources,
climate change, and shifting weather patterns is also a major contributing factor.
Such significant changes pose a formidable challenge to the power engineering
community regarding how to plan and operate such power systems with acceptable
reliability and how to predict and be prepared for the future. Here is where, in
particular, the ML methods shine bright. ML allows us to infer knowledge from data,
and by implementing different algorithms, such as artificial neural networks, the
unpredictable events can be predicted by an acceptable degree [1]. This is an
essential feature of ML in various fields, from science to engineering and humanities.
ML methods are widely used in power engineering in recent years because of their
effectiveness and the immense variety of ML approaches, each with its benefits and
costs.
It is noteworthy that terms like machine learning, artificial intelligence, and deep
learning are sometimes incorrectly used interchangeably in the literatus. Figure 3.1

Fig. 3.1 The Van diagram

of artificial intelligence,
machine learning, and deep
learning
3 Machine Learning and Power System Planning: Opportunities and Challenges 47

demonstrates the relation between artificial intelligence, machine learning, and deep
learning to clear such misconceptions.
In this chapter, in particular, the authors will inspect the use of ML in power
systems planning. First, the most used methods of ML and deep learning will be
briefly introduced. The authors will then show and examine the papers and studies
done on applying such networks and techniques in solving the power planning
problems. The examples of applications of ML in load forecasting and optimization
problems will be studied, and in the final section, the concluding remarks are
provided.

3.1.1 Machine Learning Methods

ML approaches vary in their architecture and methods, but in general, they can be
categorized as three different main categories as follows [2]:
1. Supervised machine learning.
2. Unsupervised ML.
3. Reinforcement machine learning.
In short, the supervised learning methods require a labeled data set for training in
order to create a link between the input and the expected output. Examples of these
approaches include neural networks, decision tree networks, support vector machine
learning, linear regression, logistic regression, and nearest neighbor.
On the other side, unsupervised learning approaches do not require a labeled data
set to train and attempt to identify existing patterns. Examples for this approach
include a self-organizing map (SOM), adaptive resonance theory (ART), K-means,
principal component analysis (PCA), and support vector machine (SVM) [3].
In the reinforcement learning approaches, artificial intelligent agents interact with
the system and observe the results. The feedback provided by the interaction
between the dynamic system and the agents guides them into reaching the desired
output. This mechanism is often simulated as a concept of a cumulative reward.
Examples of these approaches are Q-learning, Monte Carlo, state-action-reward-
state-action (SARSA), deep Q-learning, and deep belief networks.
Of course, other approaches implement two or more of the mentioned methods to
increase the performance or the accuracy of the result. These methods, known as
ensemble methods, can also be found in the literature and are worth mentioning.
Figure 3.2 shows the basic overview of ML methods and their classifications [4].
Different algorithms are used in the field of machine learning, such as artificial
neural networks [4], generalized neural networks [5], and fuzzy logic models
[6]. These algorithms are not necessarily a subset of machine learning but computer
systems that are popular and used in machine learning methods.
One of the most popular methods of ML in the power system is supervised
methods. These methods usually implement the different architectures of ANN.
ANNs are computer models inspired by natural neural networks such as the
human brain and can be used for recognition purposes based on prior knowledge
48 M. H. Asgharinejad Keisami et al.

Fig. 3.2 Classification of different ML methods

and training networks. These models are useful and adoptive mathematical solutions
to various problems that can estimate the relationships between data. The underlying
architecture of ANN is made of layers of mathematical neurons connected through
links. These links have an attributed weight, which determines the strength of the
connection between neurons. As the network undergoes training, the weight attrib-
uted to these links is adjusted and weakened or strengthened. Models that use ANN
with multiple hidden layers of neurons are often called deep networks. These models
usually have connections between consecutive layers but do not have connections
within the same layer [7].
There are a variety of models used in power systems planning. These ML models
include support vector machine (SVM), long short-term memory (LSTM), and other
regression models. These models have found a favorite place among researchers
who use these tools in power system planning problems.

3.2 Literature Review

Machine learning methods are used in various applications and are related directly or
indirectly to the problem of power system planning. For example, in renewable
energy planning, machine learning methods are heavily used in Wind Power Fore-
casting applications and wind power estimations [8–10].
3 Machine Learning and Power System Planning: Opportunities and Challenges 49

They are also used in Photo Voltaic (PV) power forecasting, where ML methods
can be hugely beneficial for predicting and countering intermittency. Examples of
these applications are introduced in [11–14].
The authors in [4] have surveyed recent publications on using ML models in
smart grid applications. The reviewed papers cover a variety of methods and models
used. They are divided into four main categories: electric load and price forecasting,
ML application in fault and failure analysis, ML application in demand-side man-
agement, ML application in cyberspace security, and other miscellaneous
applications.
For example, authors in [15] have proposed a framework for planning and
operating distributed energy resources by decentralizing optimal power flow and
learning control policies for each DER using machine learning.
In [16], based on Deep Neural Networks, a scenario generation model for a
system with high penetration of renewable resources has been proposed. Authors
have used data-driven approaches to capture renewable energy production patterns.
In the subject of clustering consumer load data, authors of [17] have used a Deep
Learning-based Yearly Load Profile (YLP) feature extraction method from con-
sumer load data. The feature extraction methods can be used to cluster consumers
and compress metering data and is a useful tool for power system planning.
Authors in [18] have proposed a deep learning approach that combines the EMD
method with the long short-term memory network. They have used it for electricity
demand estimation in a day-ahead market.
To schedule solutions of multiple electric vehicle charging stations in a system in
the presence of PV and Energy Storage System (ESS), authors have implemented a
deep reinforcement method in [19].
Different types of loads must be considered when approaching a load planning
problem. Moreover, different consumers have different consumption patterns. For
this reason, a deep regression and stump tree-based ensembles models (DRTSEM)
have been introduced by authors in [20] that are used for load planning. The model
considers various parameters such as weight decay and leaning rate and the total
number of hidden layers to achieve accurate prediction.
Compared to other applications of ML in engineering problems, load forecasting
is a mature area. ML tools have played a critical role in solving the load forecasting
problems and moving toward more accurate predictions using weather data and
hierarchical forecasts of various zones and regions. The ML-based short-term
forecast models have done a great job predicting the demand and countering the
intermittent and uncertain nature of renewables in networks with high penetration of
renewable power generations. Popular ML models used in the area of load forecast-
ing include long short-term memory (LSTM), recurrent neural network (RNN), and
random forest models [4]. However, one of the areas with excellent ML potential is
the planning frameworks of the energy systems. In the authors’ opinion in [4], the
smart grid planning and operation problem can greatly benefit from using ML
methods.
These studies further emphasize that the ML methods offer an excellent solution
to various planning problems, and such methods encourage the researchers to
implement ML tools on more problems. ML methods have been the center of
50 M. H. Asgharinejad Keisami et al.

attention for their application in energy management and smart cities [21] and smart
dispatch systems that manage the generation by implementing artificial intelligence
[22]. Some ML methods, such as reinforcement learning, are also favored by
researchers in building energy control and power system planning. Using a game
theory-based multi-agent method, the authors propose a demand-side management
scheme that aims to control, operate, and plan the heating, ventilation, and
air-conditioning systems (HVAC) and minimize the social cost in [23].
Using ML methods in an online regime is a subject of various studies, such as
voltage stability observation methods [24]; however, most ML applications are made
in an offline state in the planning framework. These applications include short-term
[25], medium-term, and long-term load forecasting. The main challenge in load
forecasting is to improve prediction accuracy when faced with big data gathered
from multiple sources or various uncertainties involving weather forecast and the
intermittent nature of renewable energies [26].
In [27], the authors have studied a generation expansion plan in a multi-period
timescale with the objective function of minimizing the expansion costs and
pollution.
The competitive aspect of such a plan in the presence of environmental policies
and penalties is considered in [28]. In alignment with the subject of ML methods, the
authors have considered the uncertainties of electricity price and wind speed, where
ML methods are used to forecast the load using rough neural networks. The presence
of environmental restrictions has made finding optimal generation expansion plan-
ning in a competitive environment challenging.
The authors in [29] believe that the total emissions can be reduced to 45% of the
original amount and maintain an adequate income for generation companies partic-
ipating in nearby electricity markets by conforming to pollution control policies. The
study uses an ANN architecture named a multilayer perceptron. The perceptron
consists of two layers, a hidden layer with ten rough neurons and a sigmoid
activation function and one output layer with a linear activation function and one
rough neuron. This paper’s assumed neurons are based on a rough structure or RNN
[27]. The basic principle of RNN is to apply the rough set theory to the neural
networks in one or more links such as designing, training, or learning links. The
overall purpose of these networks is to improve the performance of traditional neural
networks. Neuron models based on the rough set theory are named rough neurons,
and neural networks made from rough neurons are called rough neural networks.
The rough neural networks accelerate the training speed of models and increase
neural networks’ performance but do not pose too much change in the structure of
neural networks in the process [28].

3.3 Load Forecasting

Load forecasting is one of the most important studies done for power system
planning and often the first stage of future expansion plans [29]. It is evident that
to propose a plan for further developing a given network, the planners must
3 Machine Learning and Power System Planning: Opportunities and Challenges 51

understand the behavior and future electricity consumption trends. Without this
understanding, proposed expansion plans will either underestimate the required
energy and not provide the demand or overestimate, resulting in unnecessary
investment costs.
Load forecasting is often categorized into four different categories: long-term
forecasting (years), medium-term forecasting (month to a year), short-term forecast-
ing (a day to a week), and very short-term forecasting (minutes to hours) [30].
Load forecasting can be used for both system operation and power system
planning. In the case of power system operation problems, the system operators
usually require short-term to very-short term studies. In power system planning, such
as generation expansion planning, it is the long-term load forecasting study. Various
factors affect the future of load demands. They range from the upcoming trends in
technology to social welfare and economic issues and policies. This uncertainty has
made load forecasting the right subject for ML methods that specialize in predicting
the future and dealing with uncertainties.
In [31], authors have proposed a hybrid AI and DL method for load forecasting.
The results of the load forecasting problem are implemented to inform a
maintenance plan.
Authors use deep learning framework and convolutional neural network (CNN)
and long short-term memory (LSTM) in [32] for short-term electric load forecasting
and individual residential customers and compare the effectiveness of the proposed
method with other conventional methods used in the literature.
Two methods named input attention mechanism (IAM) and hidden connection
mechanism (HCM) are applied to the problem of short-term load forecasting in [33]
to enhance the accuracy and efficiency of RNN-based load forecasting models.
For customer short-term load forecasting, an LSTM framework has been
implemented in [34], in which, according to the authors, this method outperformed
similar state-of-the-art methods in speed and accuracy.
The daily load is nonlinear and dynamic and therefore hard to predict. For this
reason, authors in [35] have used a deep learning method named bespoke-gated
recurrent neural network for daily peak load forecasting.
The following section examines a method introduced in [36] for the long-term
load forecasting done with supervised ML. There are multiple regression models
used in load forecasting, and each one has its advantage and disadvantages. The
authors in [36] have used four different ML models for load forecasting, and they
have showcased the results. We will introduce the fundamentals of two different
models in the following section: subsets of supervised machine learning.

3.3.1 SVM (Support Vector Machine)

One of the popular algorithms used for classification and regression problems is the
SVM. Vapnik and AT&T Bell Laboratories developed SVM as one of the most
robust and widely used algorithms in the field. The SVM, in general, has two unique
52 M. H. Asgharinejad Keisami et al.

properties. It can maximize the margin of separation between two classes, and it
supports nonlinear functions by utilizing different kernels [37]. There are similarities
between SVM and ANN. Simultaneously, the ANN aims at estimation error mini-
mization of training data; the SVM, on the other hand, is trying to follow the
structural risk minimization principle to improve generalization error [37]. The
SVM creates an optimal hyperplane classifier. By maximizing the margin of sepa-
ration, the classifiers aim to classify the data without the error for minimizing the
empirical risk. The empirical risk is defined as the average loss of an estimator for a
finite set of data. It also minimizes the expected risk, described as the hypothesis
value of loss function. The empirical risk minimizer is then used to define the SVM
performance’s theoretical boundaries by minimizing the training data’s error. Using
risk minimization is to measure the estimator’s performance and find the estimator
that minimizes risk over distribution [37].
First, assume that we have a training set {(x1, y1), (x1, y2). . .(xi, yi)} xi 2 Rn. The
dependent variables are denoted yi 2 R that are associated with each input, and n is
the number of features. The main objective of SVM is to map a plane f(x) : Rn ! R f
(x) such that it fits closest to the data and is formulated as follows:

f ð xÞ ¼ W T  X þ b ð3:1Þ

Here, b is defined as the bias.

f(x) is the function that has the most deviations σ from targets yi. It should be
noted that deviations of the resulting hyperplane that are smaller than σ are not
considered. It is required that the transformation function ϕ : Rn ! Rs(s > n) should
be applied to all of the points in the input space. This is done so that we capture the
nonlinear nature of the data.

f ðsÞ ¼ W T  ϕðX Þ þ b ð3:2Þ

The input space is therefore mapped to the linear feature space via the transfor-
mation and with higher dimensionality. We can write the SVR solution problem as a
convex optimization problem. After the calculation, a pair of Lagrange multipliers
are associated with each training point. After training the SVM and finding the
optimal hyperplanes, support vectors are defined as points outside σ band and
distinctly have at least one zero Lagrange multipliers [38].

3.3.2 Long Short-Term Memory Network (LSTM)

Recurrent neural networks’ main idea arises from the fact that traditional ANN is not
able to make a direct connection between the previous data and the following
information and correcting the error. The idea of backpropagation can solve this
shortcoming of the traditional ANN. These networks, commonly referred to as
3 Machine Learning and Power System Planning: Opportunities and Challenges 53

Fig. 3.3 The architecture of the RNN unit [36]

recurrent neural networks (RNN), use backpropagation to compare the error from the
network’s input and output until the error is below a certain threshold [38]. One
instance of using such networks is in text prediction software, where the network is
trained to predict the following words based on the typed words so far.
The backpropagation is done through network loops. The output of previous
timestamps is used as input for the current timestamps. A simple structure of RNN
can be seen in Fig. 3.3. In Fig. 3.3, xt is input at timestamp t, st represents the state at
timestamp t, and ht is output at timestamp t. st is the current state as calculated based
on the xt and the previous hidden state st
1 [36, 38].
RNN performs very well in situations where dependencies in the data are short
term. However, as the gap between dependencies increases, the performance of the
RNN networks decreases drastically. This problem was addressed by introducing
LSTM networks by Hochreiter and Schmidhuber (1997); these networks include
memory cells and gates responsible for regulating information flow across the
network.
Each unit of LSTM consists of a memory part and three different types of gates
responsible for handling information flow. The mechanism of an LSTM unit can be
described in the following steps. First, the LSTM decides what information should
be forgotten and thus removed from the cell state. This operation is performed
through a sigmoid layer named the “forget gate.” It produces an output number
between 0 and 1 based on the ht
1 and xt in the cell state Ct
1. Number 1 represents
keeping the information, and number 0 represents forgetting this information, while
any number between represents the amount of information percentage-wise that
should be overlooked [39]. Please note that Ct is the cell state at time t, and Wi
and Wf are weight matrixes corresponding to the input and forget gates. bi, bC, bf, and
bo are all corresponding biases to the equations.
The following step is deciding which new information is going to be stored in the
cell state. First, a sigmoid layer named “input gate” generates a number determining
what values should be updated. Then a tanh layer creates new candidate values C e t in
the form of a vector that is suitable to be added to the state.

it ¼ σ ðW i  ½ht
1 , xt  þ bi Þ ð3:3Þ
54 M. H. Asgharinejad Keisami et al.

e t ¼ tanh ðW C  ½ht
1 , xt  þ bC Þ
C ð3:4Þ

The old cell Ct
1 is then updated to the new cell state Ct [38].
The next step is to forget the information that was determined to be unnecessary.
This is done by multiplying the old state by ft. Next, the Ct is added, and new
candidate values are generated and are scaled for how much each state value is
updated.

f t ¼ σ ðW f  ½xt :ht
1  þ bf Þ ð3:5Þ

K M K
Ct ¼ f t C t
1 it et
C ð3:6Þ

The last step is the output step that is based on the cell state. The output gate, a
sigmoid layer, decides the cells’ parts that will be included in the output. The tanh
operator is then acting on the cell state, which generates values between
1 and
1. These results are then multiplied by the sigmoid gate’s output, and thus, the
desired outputs are generated [38].

ot ¼ σ ðW o ½ht
1 , xt  þ bo Þ ð3:7Þ
O
ht ¼ ot tanh ðC t Þ ð3:8Þ

Figure 3.4. describes the general structure of a typical LSTM unit with different
gates and inputs and outputs.
The authors in [36] then collect data on electricity consumption, and they use
clustering techniques to categorize them. Specifically, the method of clustering used
in the reference is K-means clustering. The main goal of clustering is categorizing
and classifying data based on similarities in one or more properties of objects.
Using data clustering methods, system planners can find trends and patterns
between consumers and data groups. This trend characterization can be implemented

Fig. 3.4 Structure of an LSTM unit [36]

3 Machine Learning and Power System Planning: Opportunities and Challenges 55

to understand the behavior of electricity consumption and load demand. By identi-

fying similar groups, ML solutions can provide a better understanding of energy
planning problems. As previously mentioned, the authors use ML algorithms to
analyze the load characteristics and predict the future. The ML algorithms’ primary
basis, mainly the SVM and LTSM algorithm, was discussed in the previous section.
In [36], the authors have implemented four regression models, LSTM, SVM,
ANN, and RNN, to perform a load demand estimation for the union territory
Chandigarh in India.
To compare different ML methods, the authors in [36] have used three different
evaluation measures. The first root means squared error (RMSE), one of the most
used and common measures, is defined as the standard deviation of the differences
between actual and predicted values.
This measure can be formulated as follows:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u n  2
u1 X 2 X
RMSE ¼ t y j,t
y
j,t ð3:9Þ
2n t¼1 j¼1

y
j,t is defined as the predicted value, and the yj, t is the actual value at time t.
The second correlation coefficient approximates the strength of the relationship
between real-time observations and the predicted values and is defined as follows:

Covarianceðy, y
Þ
ρ ¼ Correlðy, y
Þ ¼ ð3:10Þ
σ y  σ y

Finally, the amount of deviation of the forecasted values from the actual values
can be expressed as the mean absolute percentage error (% Error), which can be
mathematically described as:
 
X2 X n  y

y 

1  j,t j,t

%Error ¼   ð3:11Þ
2n t¼1 j¼1  y j,t 

3.3.3 Results of Load Forecasting Using Machine Learning

According to [36], after comparing the prediction of ML methods and actual data for
a given cluster of data, the LSTM network model seems to have made a more
accurate prediction than SVM, ANN, and RNN models. For example, using the
LSTM model, the average error is between 7 and 10% and considered acceptable.
LSTM networks’ advantages over SVM can be summarized as follows: LSTM can
access current input timesteps. Simultaneously, for the SVM to be trained, several
56 M. H. Asgharinejad Keisami et al.

successive inputs are required from a properly selected time window. By including
real-time demand observation, the LSTM method can support active learning;
however, the SVM regression model only supports static learning.

3.4 ML in Power Systems Optimization Problems

The application of ML methods in power system planning is not limited to load

forecasting. The long-term planning problems often require an optimization method
to minimize the investment and operation cost or find the most suitable location
capacity or other variables in the scheme of a planning problem for the planning
horizon. By elongating the planning problems in years, technology advancement’s
inherent nature introduces uncertainties to the problem. The two most common
optimization methods used for planning problems under uncertainties include sto-
chastic programming (SP) and robust optimization (RO). The SP aims to optimize
the expected value’s objective functions, where the expectation is calculated regard-
ing the probability distribution of the random variables. These variables represent the
uncertainty in the problem. On the other hand, RO methods have a min-max
approach to the uncertainty, and in contrast to the SP, they do not require defining
the specific probability distribution functions.
In [40], the authors aim to solve long-term electric peak load and demand
forecasting using optimization methods and ML tools. These studies are a critical
part of generation expansion planning studies, where having a better understanding
of the yearly peak load is directly responsible for better planning solutions. There-
fore, the authors first implement an SVM method for long-term load forecasting
problems and particle swarm optimization (PSO) method as a heuristic optimization
algorithm.
Moreover, the authors in [40] try to minimize the forecasting error using a hybrid
forecasting method, which uses three different ML models. These models using as a
hybrid model are auto-regressive integrated moving average (ARIMA), artificial
neural network (ANN), and the proposed support vector regression technique. The
introduced method prioritizes each forecasting method based on the resulted errors
over the existing data. The yearly peak load and total energy demand of Iran
National Electric Energy system are the aimed results.
Therefore, RO establishes a set of possible realizations such as ranges of variation
based on the known information about the uncertain parameters and then optimizes
toward feasible solutions for all of the uncertain parameters within the uncertainty
set [40]. These uncertainties open an opportunity for the ML models to be introduced
to the planning problem. The authors in [40] have proposed a method for the
problem of long-term energy planning. To show that there is a discrepancy among
the associated stochastic solutions and other robust solutions published in the
literature, they first devise an SP method considering different sources of uncertainty
and multiple different PDFs to emphasize and manifest their potentials and their
weaknesses in the long-term investment models in the energy planning problem. The
3 Machine Learning and Power System Planning: Opportunities and Challenges 57

authors then proposed a distributionally robust optimization method combined with

ML that allows them to “numerically tractable, recourse-based robust formulation of
the EnergyScope model,” with less sensitivity to the chosen type of PDF. The
authors of [40] proposed that DRO solutions vary less regarding the underlying
distribution, and they claimed that it produces a more robust decision. Compared to
the other DRO uses, a large dimension of underlying uncertainty has been
implemented in their work that poses a significant computational challenge. There-
fore, ML tools have been used to rank and decide which uncertainty parameter is the
most significant and should be included in the ambiguity set definition.
Combining the ML algorithms and the DRO approach enables the authors to
select the critical variables more systematically. The ML algorithm used is the
Extreme Gradient Boosting (XGBoost) method, one of the models based on the
regression tree model and therefore considered a supervised machine learning. In
short, the XGboost algorithms sequentially makes decision trees according to the
margin of error left by the predictive variables of the previous tree. This is repeated
until a maximum threshold of adjustment is reached for all trees’ combined perfor-
mance, and the algorithm is stabilized [39]. As a result, they have been able to use
their DRO method and, at the same time, cut the computational burden of consid-
ering a large number of uncertainties with the help of machine learning. Using ML
algorithms to rank the variables in the problem and reduce the computational power
required and time needed in an optimization problem in power system planning with
multiple sources of uncertainty is one of the applications of ML in the power system
planning problem.

3.5 Conclusion

In this chapter, first, the authors discussed the necessity of ML in modern-day

engineering problems by focusing on different types of ML and four different
categories of ML methods. In the next section, overall applications of ML in
power systems engineering were investigated with a focus on power system plan-
ning. It seemed that the problem of load forecasting had been an excellent topic to
introduce the application of ML in power system planning. After manifesting the
structure of two supervised ML methods (i.e., SVM and LSTM), the authors
reviewed the load forecasting problem’s results with different ML methods
according to referenced sources. In the end, the authors briefly mentioned other
applications of ML tools in the power planning problem, such as the long-term
optimization problem.
In conclusion, the ML methods are excellent tools in power system engineers’
hands to tackle various problems. In that regard, by correctly using and
implementing ML, researchers can solve problems infused with uncertainty using
artificial intelligence’s immense potential. As some researchers in the field have
pointed out, artificial intelligence systems in power engineering and distribution and
planning have great potential and require more research and resources [20]. The field
58 M. H. Asgharinejad Keisami et al.

of artificial intelligence and, therefore, ML has been exponentially progressing for

decades. The achievements often result in unexpected researchers; hence, this field’s
only expected trajectory seems to be upward.

References

1. Y. Chen. Bridging Machine Learning to Power System Operation and Control (2020). https://
doi.org/10.13140/RG.2.2.25463.98720. https://www.researchgate.net/publication/339988648_
Bridging_Machine_Learning_to_Power_System_Operation_and_Control
2. C.M. Bishop, Pattern Recognition and Machine Learning (Springer, New York, 2006)
3. X. Kong, X. Liu, R. Shi, K.Y. Lee, Wind speed prediction using reduced support vector
machines with feature selection. Neurocomputing 169, 449–456 (2015)
4. M.S. Ibrahim, W. Dong, Q. Yang, Machine learning driven smart electric power systems:
current trends and new perspectives. Appl. Energy 272, 115237 (2020)
5. A. Mansour-Saatloo, A. Moradzadeh, B. Mohammadi-Ivatloo, A. Ahmadian, A. Elkamel,
Machine learning based PEVs load extraction and analysis. Electronics 9(7), 1150 (2020)
6. H. Yang, Z. Jiang, L. Haiyan, A hybrid wind speed forecasting system based on a ‘decompo-
sition and ensemble’ strategy and fuzzy time series. Energies 10(9), 1422 (2017)
7. E. Mocanu, P.H. Nguyen, M. Gibescu, Deep learning for power system data analysis, in Big
Data Application in Power Systems, (Elsevier, Amsterdam, 2018), pp. 125–158
8. X. He et al., Research on a novel combination system on the basis of deep learning and swarm
intelligence optimization algorithm for wind speed forecasting. IEEE Access 8, 51482–51499
(2020)
9. J. Yan et al., Forecasting the high penetration of wind power on multiple scales using multi-to-
multi mapping. IEEE Trans. Power Syst. 33(3), 3276–3284 (2017)
10. T. Hu et al., Distribution-free probability density forecast through deep neural networks. IEEE
Trans. Neural Netw. Learn. Syst. 31(2), 612–625 (2019)
11. H. Zhou et al., Short-term photovoltaic power forecasting based on long short term memory
neural network and attention mechanism. IEEE Access 7, 78063–78074 (2019)
12. Ray, Biplob, et al. A New Data Driven Long-Term Solar Yield Analysis Model of Photovoltaic
Power Plants. IEEE Access 8 (2020): 136223–136233. https://ieeexplore.ieee.org/document/
9149581
13. Z. Zhen et al., Deep learning based surface irradiance mapping model for solar PV power
forecasting using sky image. IEEE Trans. Ind. Appl. (2020)
14. G.W. Chang, L. Heng-Jiu, Integrating gray data preprocessor and deep belief network for
day-ahead PV power output forecast. IEEE Trans. Sustain. Energy 11(1), 185–194 (2018)
15. R. Dobbe et al., Toward distributed energy services: decentralizing optimal power flow with
machine learning. IEEE Trans. on Smart Grid 11(2), 1296–1306 (2019)
16. Y. Chen et al., Model-free renewable scenario generation using generative adversarial net-
works. IEEE Trans. Power Syst. 33(3), 3265–3275 (2018)
17. S. Ryu et al., Convolutional autoencoder based feature extraction and clustering for customer
load analysis. IEEE Trans. Power Syst. 35(2), 1048–1060 (2019)
18. J. Bedi, D. Toshniwal, Empirical mode decomposition based deep learning for electricity
demand forecasting. IEEE Access 6, 49144–49156 (2018)
19. M.J. Shin, D.-H. Choi, J. Kim, Cooperative management for PV/ESS-enabled electric vehicle
charging stations: a multiagent deep reinforcement learning approach. IEEE Trans. Ind. Inform.
16(5), 3493–3503 (2019)
20. T. Ahmad, D. Zhang, Novel deep regression and stump tree-based ensemble models for real-
time load demand planning and management. IEEE Access 8, 48030–48048 (2020)
3 Machine Learning and Power System Planning: Opportunities and Challenges 59

21. M. Zekić-Sušac, S. Mitrović, A. Has, Machine learning based system for managing energy
efficiency of public sector as an approach towards smart cities. Int. J. Inf. Manag. 50, 102074
(2020)
22. L. Yin et al., A review of machine learning for new generation smart dispatch in power systems.
Eng. Appl. Artif. Intell. 88, 103372 (2020)
23. J. Hao, Multi-agent reinforcement learning embedded game for the optimization of building
energy control and power system planning. arXiv preprint arXiv:1901.07333 (2019). https://
arxiv.org/abs/1901.07333
24. V. Malbasa et al., Voltage stability prediction using active machine learning. IEEE Trans. Smart
Grid 8(6), 3117–3124 (2017)
25. U. Munawar, Z. Wang, A framework of using machine learning approaches for short-term solar
power forecasting. J. Electr. Eng. Technol. 15(2), 561–569 (2020)
26. D. Zhang, X. Han, C. Deng, Review on the research and practice of deep learning and
reinforcement learning in smart grids. CSEE J. Power Energy Syst. 4(3), 362–370 (2018)
27. R.A. Mehrabadi, M.P. Moghaddam, M.K. Sheikh-El-Eslami, Generation expansion planning in
multi electricity markets considering environmental impacts. J. Clean. Prod. 243, 118611
(2020)
28. H. Liao et al., An overview on rough neural networks. Neural Comput. & Applic. 27(7),
1805–1816 (2016)
29. S. Madadi, M. Nazari-Heris, B. Mohammadi-Ivatloo, S. Tohidi, Implementation of genetic-
algorithm-based forecasting model to power system problems, in Handbook of Research on
Predictive Modeling and Optimization Methods in Science and Engineering, (IGI Global,
Philadelphia, 2018), pp. 140–155
30. I. Koprinska, M. Rana, V.G. Agelidis, Correlation and instance based feature selection for
electricity load forecasting. Knowl.-Based Syst. 82, 29–40 (2015)
31. S. Motepe, A.N. Hasan, R. Stopforth, Improving load forecasting process for a power distri-
bution network using hybrid AI and deep learning algorithms. IEEE Access 7, 82584–82598
(2019)
32. M. Alhussein, K. Aurangzeb, S.I. Haider, Hybrid CNN-LSTM model for short-term individual
household load forecasting. IEEE Access 8, 180544–180557 (2020)
33. M. Zhang, Z. Yu, X. Zhenghua, Short-term load forecasting using recurrent neural networks
with input attention mechanism and hidden connection mechanism. IEEE Access 8,
186514–186529 (2020)
34. W. Kong et al., Short-term residential load forecasting based on LSTM recurrent neural
network. IEEE Trans. Smart Grid 10(1), 841–851 (2017)
35. Z. Yu et al., Deep learning for daily peak load forecasting–a novel gated recurrent neural
network combining dynamic time warping. IEEE Access 7, 17184–17194 (2019)
36. J. Bedi, D. Toshniwal, Deep learning framework to forecast electricity demand. Appl. Energy
238, 1312–1326 (2019)
37. Rastgoufard, Samin. Applications of Artificial Intelligence in Power Systems. (2018). https://
scholarworks.uno.edu/td/2487/
38. C. Olah, Understanding lstm networks (2015), https://colah.github.io/posts/2015-08-Under
standing-LSTMs/
39. E. Guevara et al., A machine learning and distributionally robust optimization framework for
strategic energy planning under uncertainty. Appl. Energy 271, 115005 (2020)
40. M.-R. Kazemzadeh, A. Amjadian, T. Amraee, A hybrid data mining driven algorithm for long
term electric peak load and energy demand forecasting. Energy 204, 117948 (2020)
Chapter 4
Introduction to Machine Learning Methods
in Energy Engineering

Arash Moradzadeh, Behnam Mohammadi-Ivatloo, Kazem Pourhossein,

Morteza Nazari-Heris, and Somayeh Asadi

4.1 Introduction

Increasing energy consumption and the penetration of renewable energy sources in

traditional power systems have complicated power systems. Hence, the traditional
power system has moved toward smartening in order to increase sustainability as
well as supply energy demand [1, 2]. These smartening are performed through
information infrastructures such as the Internet, which is also sometimes exposed
to cyberattacks. On the other hand, the intelligent power system and the use of

A. Moradzadeh
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
e-mail: [email protected]
B. Mohammadi-Ivatloo (*)
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
Department of Energy Technology, Aalborg University, Aalborg, Denmark
e-mail: [email protected]; [email protected]
K. Pourhossein
Department of Electrical Engineering, Tabriz Branch, Islamic Azad University, Tabriz, Iran
e-mail: [email protected]
M. Nazari-Heris
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
Department of Architectural Engineering, Pennsylvania State University,
University Park, PA, USA
e-mail: [email protected]; [email protected]
S. Asadi
Department of Architectural Engineering, Pennsylvania State University,
University Park, PA, USA
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 61

M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_4
62 A. Moradzadeh et al.

advanced monitoring systems have caused the users of power systems to face big
data [3, 4]. One beneficial way to deal with big data of the power systems for proper
operation and planning methodologies is the application of machine learning and
deep learning methods. Data mining has been introduced as a powerful tool for
analyzing huge volumes of data. This branch of data science is named based on
searching for valuable information and features in a large database and identifying its
patterns. Data mining has easily answered questions in science and technology that
have been impossible or time consuming to resolve manually. Using statistical
techniques for data assessment, feature extraction, pattern recognition, and finding
relationships between data is the main idea of data mining [5, 6]. Data mining can be
used in all data-based issues in science and technology. Today, this science is used
dramatically in many research and industrial fields such as medicine, engineering,
and even security issues related to face recognition and fingerprinting. Accordingly,
in this chapter, the applications of data mining in issues related to energy engineering
are discussed [7]. The fundamental of machine learning methods is automated
learning from past experiences and examples without explicit programming. Various
machine learning approaches have been introduced to apply to engineering problems
such as supervised and unsupervised learning methods, reinforcement learning, etc.
The high performance of machine learning methods in data process, classification,
and prediction have made these methods popular in several engineering areas. Such
methodologies have been widely applied in computer vision, speech recognition,
object recognition, and content-based retrieval from multimedia [6, 8]. In the energy
industry, machine learning methods have been widely used for dealing with explo-
ration and processes of oil and gas, prediction of solar radiation, obtaining optimal
operation of reactors, prediction of wind power output, forecasting of fault in power
systems, and prediction of energy demands [7].
The remainder of this chapter is as follows: Section 4.2 has focused on data
mining and its applications in energy engineering. Then the machine learning
approaches in energy engineering is studied in Sect. 4.3, and deep learning
approaches in energy engineering is investigated in Sect. 4.4. The evaluation metrics
of regression and classification applications is analyzed in Sect. 4.5, and the chapter
is concluded in Sect. 4.6.

4.2 Data Mining and Its Applications in Energy

Engineering

The wide range of issues related to energy and power systems and especially the big
data related to each of the functions of these systems has led to data mining that has
high applications in power and energy systems. Data mining has a variety of
techniques and algorithms, each of which is used in some way to solve problems
related to energy engineering [3, 9].
4 Introduction to Machine Learning Methods in Energy Engineering 63

Today, generating energy from renewable energy sources, transmission, and

injecting it into the power grid and its distribution is fraught with many problems
that have posed many challenges. On the other hand, providing long-term and short-
term planning for traditional and renewable power plants to meet the demand for
load and supply the required and sustainable energy is another challenging issue that
is widely discussed today [10, 11]. The literature mainly deals with issues related to
energy production in the power system. Today, the most widely covered issues in
energy engineering are energy management and conservation in consumers’ side. So
that extensive studies and projects are done annually to manage and increase energy
efficiency in residential and industrial applications [12–14].
In recent decades, solving such problems related to power systems has been done
in various ways. A review of the literature in recent years shows the significant use of
data mining applications such as artificial neural network (ANN) algorithms,
machine learning, and deep learning. Due to their high capabilities in data
processing, each of these algorithms has been able to solve problems related to
power and energy systems. Data mining applications in energy and power systems
can be classified into a variety of forecasts, detections, and analyses. Forecasting
programs include load forecasting [15–17], load disaggregation [18–20], forecasting
of renewable energy sources such as wind and solar energies [21–24], plug-in
electric vehicles load forecasting [25–28], price forecasting in various electricity
markets [29–31], and measuring the potential of different regions for the construc-
tion of the various wind and solar power plants. Diagnostic operations by data
mining applications can be considered as detecting various types of faults and
damages related to equipment in energy and power systems [32–36], as well as its
various parts such as transmission lines and distribution networks [37, 38]. In
addition, today, the cyber-physical system security and the cyberattacks detections
on power and energy systems are dramatically done by data mining applications
[39, 40].
Each of these programs is performed by specific algorithms. All data mining
techniques are able to solve the abovementioned problems, but choosing the correct
algorithm to do anything depends on the data and the type of problem. For example,
load forecasting problems are mainly solved with regression algorithms due to the
fact that they are based on continuous-type data. Classification algorithms mainly do
detection and localization of faults and damages. Each of the regression and classi-
fication applications has suitable algorithms that can be used depending on the
problem data type.
As mentioned in the above literature, machine learning and deep learning algo-
rithms are used extensively in power and energy systems issues. The most important
problem that power and energy system users mainly face is the lack of complete
familiarity with the types and applications of machine learning and deep learning
algorithms. Accordingly, in the following sections of this chapter, various tech-
niques of the machine learning and deep learning will be introduced so that the
structure, formulation, and application of each of them are described in detail to be
used as a guide for users of power and energy systems in the employ of data mining
applications.
64 A. Moradzadeh et al.

4.3 Machine Learning Schemes in Energy Engineering

Machine learning is a tool for presenting a model or an estimate of the future based
on the learnings of past events. In machine learning applications, computer algo-
rithms predict future behavior based on learning from past experiences [41]. Machine
learning algorithms based on the performed training create a mathematical model for
the training data to make this prediction. The use of machine learning tools for
emerging fields such as smart buildings is one of the research trends that has recently
attracted the attention of research communities in several disciplines, including
computer science and electrical engineering, civil engineering, and architecture.
Machine learning has many training algorithms, each of which is used for specific
applications of data mining science [42].
This section introduces the various machine learning algorithms used in energy
engineering applications.

4.3.1 Machine Learning Methods: Application, Formulation,

and Structure

Machine learning techniques are divided into two categories of regression and
classification applications. In the continuation of this section, we will get acquainted
with the types and applications of each of the machine learning algorithms in energy
engineering.

4.3.1.1 Support Vector Machine (SMV)

SVM was introduced in 1995 by Cortes and Vapnik as a nonparametric statistical

learning method [43]. The SVM is one of the most widely used supervised machine
learning algorithms to solve classification, pattern recognition, and prediction prob-
lems. The main purpose of the SVM is to create a linear mapping and transform the
nonlinear input area into a high-dimensional space to find a hyperplane. This
hyperplane is achieved by maximizing the margin between the separating data
[44]. Thus, by achieving this goal, the dataset is broken down into a number of
predefined classes and separated in a manner consistent with the training samples.
Figure 4.1 shows an example of an SVM structure for classification in a
two-dimensional input space [45].
Linear mapping in SVM for the training dataset Z ¼ {Xi, Yi| i ¼ 1, 2, 3, . . ., n} can
be obtained as [44]:

γ ¼ ωT θ ð x Þ þ b ð4:1Þ
4 Introduction to Machine Learning Methods in Energy Engineering 65

Fig. 4.1 Main structure of

the SVM [45]

where ω shows the weight vector, b represents the bias, and θ(X) demonstrates the
agent of a nonlinear mapping function. Calculation of support vectors related to each
class are described as:


b þ W T  X i ¼ þ1, for di ¼ þ1
ð4:2Þ
b þ W T  X i ¼ 1, for di ¼ 1

where di is the related class, i.e., di ¼ + 1 corresponding to class A and di ¼  1

related to class B.
Solving the following inequality determines the training data for classes A and B,
and the last objective function is achieved by Eq. (4.4) [45]:

fdi ðb þ wt xi Þ  1 ð4:3Þ
!
X
N  
f ðxÞ ¼ sign a0,i xT xi þ b ð4:4Þ
i¼1

where X and N show the input vectors and the support vector numbers attained in
training stage, respectively. The positive parameters a0, i are employed to represent
the support vectors between the input vectors.
A vector-mapping function as φ(x) is employed to transfer inseparable data to a
high-dimensional linear space and separate them. This transform allows data to be
categorized using the drawn linear hyperplane. Finally, the decision function is used
as follows [45]:
66 A. Moradzadeh et al.

!
X
N
f ðxÞ ¼ sign a0,i ðφðxÞφðxi ÞÞ þ b ð4:5Þ
i¼1

4.3.1.2 Group Method Data Handling (GMDH)

GMDH is a machine learning application and is one of the nonlinear regression

modeling algorithms that was introduced in 1968 by Ivakhnenko [46]. Modeling
complex systems, performance approximation, feature extraction, and discovering
the relationship between input and output data in them are the most important
applications of the GMDH [47]. Building a forward-forward neural network based
on a quadratic transfer function through supervised learning is the main objective of
the GMDH. Figure 4.2 shows the structure of the GMDH network. The mapping
between the input and output variables in this structure is performed by the nonlinear
function of the Volterra series according to Eq. (4.6). Volterra series analysis is
performed using Eq. (4.7) [48].

X
m m X
X m m X
X m X
m
b ¼ a0 þ
Y ai X i þ aij X i X j þ X j aijk X i X j X k þ ⋯ ð4:6Þ
i¼1 i¼1 j¼1 i¼1 j¼1 k¼1
 
G X i , X j ¼ a0 þ a1 X i þ a2 X j þ a3 X 2i þ a1 X 2j þ a5 X i X j ð4:7Þ

Fig. 4.2 Principle structure of the GMDH

4 Introduction to Machine Learning Methods in Energy Engineering 67

where Yb shows the model prediction of output, X represents the inputs, m is the
number of inputs, and a demonstrates the weights. Finally, by considering the
principle of least squares error [49], the G function is defined to estimate the output
parameter y as follows:
PM
i¼1 ðY i  Gi OÞ2
E¼ ð4:8Þ
M
yi ¼ f ðX i1 , X i2 , X i3 , . . . , X im Þ, i ¼ 1, 2, 3, . . . , m ð4:9Þ

4.3.1.3 Support Vector Regression (SVR)

SVR was introduced as a regression version of the SVM in 1995 by Vapnik

[50]. The SVR as a supportive regression suggests the e-intensive loss function
(e-SVR) performance. Exact linear mapping between input and output variables,
nonlinear data modeling without the overfitting problems in the training phase, and
high processing speed are the most important advantages of the SVR [15, 51]. Based
on mathematical relations, the general model of linear regression is as follows:

f ðxÞ ¼ hw, xi þ b ð4:10Þ

Calculating the weight vector based on the reduction of quadratic errors is the
basis principle of the traditional regression models, while the e-SVR is based on
absolute error optimization. The main objective of the e-SVR is to develop a
function in which all errors lie under a certain value e but with the best generalization
capacity possible and is applied based on the following relationships [15]:

1
Minimize kwk2
2(
yi  ðhW, X i i þ bÞ  ε ð4:11Þ
Subject to
ðhW, X i i þ bÞ  yi  ε


0 if jyi  byi j < ε
ξjεj ¼ ð4:12Þ
jyi  byi j  ε otherwise

where ξi and ξi represent the loose variables, yi represents the measured output, byi is
the predicted output, and e is a user-defined parameter. The points inside and outside
the e-intensive contain the values of Eqs. (4.13) and (4.14), respectively [15, 52]:

ξi ¼ 0 and ξi ¼ 0 ð4:13Þ

68 A. Moradzadeh et al.

   
ξi > 0 and ξi ¼ 0 or ξi ¼ 0 and ξi > 0 ð4:14Þ

Equation (4.11) is optimized by the Lagrangian coefficient during a dual standard

optimization. Finally, after calculating the Lagrangian coefficient, the following
optimized expression is obtained:

X
n  
f ðX Þ ¼ αi  αi hX i , X i þ b ð4:15Þ
i¼1

where αi and αi represent the Lagrange coefficients.

4.3.1.4 General Regression Neural Network (GRNN)

The GRNN was introduced in 1999 by Specht as one of the radial basic function
(RBF) networks to establish nonlinear relationships between input and output vari-
ables [53, 54]. This network was a suitable alternative to the feedback error propa-
gation training algorithm in the feed-forward neural network with very strong
efficiency in regression applications. One of the most important advantages of the
GRNN is its ability to present good performance for very small datasets. Despite an
input parameter, it still can predict well, whereas the available data from the
measurements of an operating system is usually never enough for a backpropagation
neural network (BPNN) [25, 55]. In addition, the GRNN is a time series-based
regression algorithm that can extract and model a continuous time series relationship
between input and output variables. Figure 4.3 shows the interconnected structure of
the GRNN.

Fig. 4.3 Basic principle of the GRNN

4 Introduction to Machine Learning Methods in Energy Engineering 69

In general, estimating the surface of regression types on the independent variable

via GRNN is described in terms of mathematical formulation as follows [55]:

  R 1 yf ðX, yÞdy
b y
Y¼E ¼ R1
1 ð4:16Þ
1 f ðX, yÞdy
X

where X shows an input vector, and Y is the forecasted output value by the GRNN.
The density function should be estimated by the Parzen nonparametric prediction
using the dataset of samples. This prediction is based on Gaussian kernel estimator
and can be written as [56]:

bf ðX, yÞ ¼ 1
pþ1
nð2π Þ 2 σ pþ1
X n    
ðX  X i Þ2 ðX  X i Þ ðX  X i Þ2
 exp  exp  ð4:17Þ
i¼1
2σ 2 2σ 2

where σ represents the smoothing parameter, and Xi and Yi show the data point of
random variable x and y, respectively. The n is the sample sizes and p is the
dimension number of random variable x.
Finally, the Euclidean distance between the selected input sample and the training
dataset, the pattern for estimating the final output, is determined as follows [25]:

D i ¼ ðX  X i ÞT ðX  X i Þ ð4:18Þ

4.3.1.5 Decision Tree

Decision tree is one of the most common data mining techniques that is produced
through a two-step process of learning and classification. The decision tree uses a
tree structure to describe, classify, and generalize the dataset [57]. This method’s
learning process is done by dividing the input dataset into two sets of training and
test. After extracting the features and completing the training process, a decision tree
production algorithm generates the decision tree by considering the training data as
input [58]. So far, various algorithms have been introduced and used to generate
decision trees. C4.5 [59], classification and regression trees [60], and ID3 [61] are
widely used examples of these algorithms.

4.3.1.6 k-Means

k-means is one of the fastest unsupervised learning algorithms in machine learning

applications used for clustering issues. This algorithm divides the input data into
70 A. Moradzadeh et al.

separate clusters through its feature vectors. The basic principle of data clustering is
eigenvalues of data. So that data with the same eigenvalues are placed in a cluster
[62]. The integer k must determine the number of clusters before starting the
clustering operation. After determining the number of clusters, the main goal of k-
means is to minimize the sum of squares within the cluster (the sum of the functions
of the distance from each point of the cluster to the center k) [63, 64]. To do this, k-
mean clustering for the input dataset X1, X2, . . ., Xn, where each X corresponds to an
attribute vector, divides n objects into k(n) and sets S ¼ {S1, S2, . . ., Sk}. Equation
(4.1) describes this behavior in the mathematical model [63]:

arg min Xk X
k X  U i k2 ð4:19Þ
s i¼1 x2Si

where Ui shows a typical feature vector of Si and is the centroid of Si.

In the next step, the distance between two data or two objects is calculated by a
distance function. So far, various distance functions such as Euclidean and cosine
distances have been employed to do this. Each of these distance functions has unique
functions, the choice of which depends on the type of problem [62, 63].

4.4 Deep Learning Schemes in Energy Engineering

In the last decade, deep learning has been considered as the most important advance-
ment of science in the field of computer science so that today its impact and the
significant role are felt in all scientific fields and industrial applications. Deep
learning can be considered an evolution of artificial intelligence based on deep and
continuous learnings [65, 66]. Today, the special position of deep learning applica-
tions in issues related to energy and power systems is clearly visible. In addition,
deep learning has been able to practically compensate for some of the machine
learning disabilities. Various deep learning algorithms are implemented in four
popular ways such as a convolutional neural network (CNN), sparse autoencoders,
restricted Boltzmann machine (RBM), and long short-term memory (LSTM)
[66, 67]. This section introduces the structure, formulation, and applications of
these algorithms in energy engineering.

4.4.1 Deep Learning Methods: Application, Formulation,

and Structure

Deep learning algorithms are divided into two categories of regression and classifi-
cation applications. In the continuation of this section, we will get acquainted with
4 Introduction to Machine Learning Methods in Energy Engineering 71

each of the four types and applications of each of the deep learning algorithms in
energy engineering.

4.4.1.1 Convolutional Neural Network (CNN)

CNN was first introduced and used by LeCun in 1990 for handwritten digit identi-
fication. Since then, with the efforts of researchers, this algorithm has improved day
by day and is used in other applications such as computer vision, image processing,
speech recognition, and some predictive applications [66, 68]. The CNN has a
default structure like Fig. 4.4. As can be seen, this structure consists of convolution
layers, pooling layers, fully connected layers, and, finally, the classification layer.
The basic principle in the structure of the CNN is the automatic extraction of hidden
features in the input data structure through several layers in a row by the filters used
in the convolution layers [69].
Each of these layers has unique applications in the structure of the CNN.
Convolution layers, which use multiple kernels as filters in their structure, are
responsible for extracting the features of the input data. The pooling layers are
employed to aggregate the extracted features from convolution layers in order to
prevent overfitting. In the CNN structure, the features pooled by each pooling layer
are used as input for the next convolution layer. The last pooling layer collects the
final extracted features and converts them into a feature space as input for the fully
connected layers. The structure of fully connected layers acts like a feed-forward
neural network and is responsible for determining the weight and bias associated
with data connections. After weighing and passing the training step in this layer,
finally, in the last layer of the CNN structure, a Softmax function is utilized to
classify the data and present the final output [70, 71].

Fig. 4.4 Layer-to-layer structure of the CNN [69]

72 A. Moradzadeh et al.

4.4.1.2 Autoencoders

Autoencoders are one of the deep learning applications used for unsupervised
training of data. These networks have a structure similar to feed-forward neural
networks and consist of two components, including an encoder and a decoder. As
Fig. 4.5 shows, an input layer, several hidden layers, and an output layer of the same
size as the input layer form the structure of these networks [72]. Minimizing the
mean reconstruction loss is one of the most important issues to consider in the
training of autoencoders. In the training phase of these networks, with the proper
reconstruction of the input data, the maximum information about the original input
data is stored by the neurons in the hidden layers. To activate in the layers of this
network, nonlinear activation functions such as ReLU, tanh, and sigmoid are
employed. These functions make the network architecture more abstract for learning
ideal and hierarchical features and enable the network to identify complex and useful
patterns [66, 68]. Autoencoders are divided into three categories including sparse
autoencoder (SAE) [73], denoising autoencoder (DAE) [74], and contractive
autoencoder (CAE) [75].

4.4.1.3 Recurrent Neural Network (RNN)

RNNs are one of the deep learning algorithms and one of the most interesting
applicationsof backpropagation networks. Based on their structure, these networks
are used for problems that have consecutive and interconnected inputs. The RNNs
are very powerful dynamic systems and perform better than other deep learning
algorithms for the series and sequential data [66]. Each input sequence is processed
separately in the training phase and held as a state vector in its hidden units. At the
end of the training phase, a strong database of the history of all past elements is
available. When the outputs of the hidden layer are considered as discrete over time,
it is as if they are the output of different neurons from a deep multilayer network

Fig. 4.5 Principle structure

of the autoencoders
4 Introduction to Machine Learning Methods in Energy Engineering 73

Fig. 4.6 Structure of an LSTM unit

[65]. The main purpose of the RNNs is to learn the long-term dependencies of data
and to model them for future predictions. But theoretical and empirical evidence has
shown that learning and modeling information for a long time has some problems.
Despite all the advantages that these networks have, it has been proven that their
training is problematic due to the growth or shrinkage of the backpropagated
gradients at each time step. So during different stages, they usually have problems
such as explosion or overfitting [76, 77]. To solve this problem and process long-
term data, the idea of network reinforcement was used. The first proposal to do so
was the long short-term memory (LSTM) networks, introduced in 1997 by
Hockeriter et al. [78]. The LSTM has the ability to train for tasks that require
knowledge of multiple previous modes by retaining information about the previous
status. The network has dramatically reduced the RNN restrictions and eliminated
problems with gradients loss and explosions. Figure 4.6 shows the structure of an
LSTM unit. It is observed that the signal flow in this structure is done by blocks of
the memory cell state, and the adjustment of this structure is done using input, forget,
and output gates. Everything stored, read, and written in this cell is controlled by
these gates [15, 79]. The mathematical formulation of the LSTM is described in
detail in [15].

4.4.1.4 Restricted Boltzmann Machine (RBM)

An RBM is an advanced neural network and one of the deep learning applications.
RBM can utilize unsupervised learning algorithms to solve regression problems,
74 A. Moradzadeh et al.

Fig. 4.7 Interconnected

structure of the RBM

categorize, predict, and construct nonlinear models of unlabeled data [65]. As shown
in Fig. 4.7, the RBM consists of a visible input layer and a single hidden layer that
are completely interconnected, but there is no interlayer communication in this
diagram. The visible layers receive the input data, and the input-related features
are extracted in the hidden layers [65, 72]. The network is trained to perform possible
reconstruction of the inputs by increasing the performance of the vector probability
in the visible units. In the training phase, by repeatedly updating the weight and
network bias connections using an algorithm called contrastive divergence, the
probability of entering a specific dataset according to network parameters can be
maximized. This helps to extract useful and new features from the inputs in the
hidden layer. The RBMs are divided into two types of Deep Belief Network (DBN)
and Deep Boltzmann Machine (DBM) [68]. The structure and function of each of
these types are different, and each of which are described in detail in [80, 81].

4.5 Evaluating the Results of Regression and Classification

Applications

Machine learning and deep learning techniques are used based on the various
algorithms introduced in this chapter. Each of these algorithms specifically deals
with classification and regression problems. The results of machine learning and
deep learning algorithms will not be equal for the same data. But their effectiveness
and performance can be achieved by evaluating the results [15, 82]. In comparing
and evaluating the results, care should be taken to make comparisons for the same
data. So far, many statistical evaluation indicators have been introduced and utilized
for this evaluation [83]. It should be noted that each of the classification and
regression functions has its own unique performance evaluation indicators. Indica-
tors that evaluate the results of regression applications include the Coefficient of
4 Introduction to Machine Learning Methods in Energy Engineering 75

Variation (CV), Mean Absolute Error (MAE), Mean Absolute Percentage Error
(MAPE), Mean Bias Error (MBE), Mean Squared Error (MSE), Root Mean Square
Error (RMSE), R  Squared (R2), Error Rate (δi), Relative Absolute Error (RAE),
Coefficient of Variation of Root Mean Square Error (CVRMSE), Weighted Mean
Absolute percent Error (WMAPE), and Mean Bias Error (MBR) [83–86]. Each of
these indicators has unique definitions and functions. CV is defined as the relative
standard deviation and a standardized measure of dispersion of a probability distri-
bution. MAE is the measure of errors between paired observations. MBE is used as
an indicator to estimate the average bias in the model and to decide on measures to
correct the model bias. MSE is used to measure the difference between the mean
squares of the predicted and actual values. R2 is one of the most important statistical
indicators in regression applications. This index is used to measure the variance ratio
of a dependent variable as described by an independent variable in the model. RAE
is used to evaluate the performance of regression models and is expressed in a ratio
by comparing an average error with the errors for each predicted value. Each of these
indicators is calculated based on the following equations:
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P ffi
n
i¼1
ðyi xi Þ2
n
CV ¼ ð4:20Þ
x
1X
n
MAE ¼ jy  xi j ð4:21Þ
n i¼1 i

1 X yi  xi
n
MAPE ¼  100 ð4:22Þ
n i¼1 xi
Pn
i¼1
ðyi xi Þ
MBE ¼ n
ð4:23Þ
x
1X
n
MSE ¼ ðy  xi Þ2 ð4:24Þ
n i¼1 i
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn 2
i¼1 ðyi  xi Þ
RMSE ¼ ð4:25Þ
n
Pn
ðyi  xi Þ2
R2 ¼ 1  Pi¼1
n 2
ð4:26Þ
i¼1 ðxi  xÞ

δ i ¼ xi ð 1  xi Þ ð yi  xi Þ ð4:27Þ
Pn
j yi  xi j
RAE ¼ Pi¼1 n ð4:28Þ
i¼1 jxi  xj
76 A. Moradzadeh et al.

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn ffi
2
i¼1 ið x  y i Þ
CVRMSE ¼ Pn  ð4:29Þ
i¼1 ix =n
Pn xi yi
i¼1 xi  xi
WMAPE ¼ Pn ð4:30Þ
i¼1 xi

1X
n
MBR ¼ ð y  xi Þ ð4:31Þ
n i¼1 i

where yi and xi represent the predicted value and target value at ith time point,
respectively, x shows the average of target values, and n illustrates the number of
total data samples.
Statistical evaluation indicators such as accuracy (Acc), error rate (err), precision
( p), recall (r), and F1-Score are used to evaluate the performance of the results
obtained from classification algorithms [87, 88]. The computational formulation of
each of these performance evaluation indicators is as follows:

tp þ tn
Acc ¼ ð4:32Þ
tp þ fp þ tn þ fn
fp þ fn
err ¼ ð4:33Þ
tp þ fp þ tn þ fn
tp
p¼ ð4:34Þ
tp þ fp
tp
r¼ ð4:35Þ
tp þ fn
2pr
F1‐Score ¼ ð4:36Þ
pþr

where tp and tn are the true positive and true negative, respectively. fp and fn shows
the false positive and false negative, respectively.
As mentioned in the literature, data mining applications have found a special
position in the various sciences and industries. Accordingly, using machine learning
applications to predict the energy price in order to participate in electricity markets
can be of great benefit to both consumers and energy producers. In addition, the use
of these intelligent algorithms as a tool to estimate the impact of electric vehicles on
reducing environmental pollution in metropolitan areas and improving the monitor-
ing of various components of power systems can be considered as future research on
this issue.
4 Introduction to Machine Learning Methods in Energy Engineering 77

4.6 Conclusions

This chapter studied the application of machine learning and deep learning methods
in energy system problems by concentrating on the literature review and studies of
application, formulation, and structure of such methods. Accordingly, the introduc-
tion and main areas of studies around machine learning and deep learning methods
were provided. Then the basics of various machine learning and deep learning
methods were discussed. Literature on machine learning and deep learning methods,
training process, and applications of each method were introduced and analyzed. To
introduce and highlight the role of such methods in energy systems problems, some
methods were selected, and their learning, structure, formulation, mode of operation,
and application were discussed. The modeling and formulation of machine learning
and deep learning were introduced in detail by focusing on their basics. This chapter
can be beneficial for researchers working on the application of machine learning and
deep learning approaches to deal with problems of energy systems.

References

1. O. Sadeghian, A. Moradzadeh, B. Mohammadi-Ivatloo, B. Mohammadi-Ivatloo, M. Abapour,

F.P.G. Marquez, Generation units maintenance in combined heat and power integrated systems
using the mixed integer quadratic programming approach. Energies 13(11), 2840 (2020).
https://doi.org/10.3390/en13112840
2. S. Abapour, M. Nazari-Heris, B. Mohammadi-Ivatloo, M. Tarafdar Hagh, Game theory
approaches for the solution of power system problems: a comprehensive review. Arch. Comput.
Methods Eng. 27(1), 81–103 (2020). https://doi.org/10.1007/s11831-018-9299-7
3. S. Pan, T. Morris, U. Adhikari, Developing a hybrid intrusion detection system using data
mining for power systems. IEEE Trans. Smart Grid 6(6), 3104–3113 (2015). https://doi.org/10.
1109/TSG.2015.2409775
4. E. Hossain, I. Khan, F. Un-Noor, S.S. Sikander, M.S.H. Sunny, Application of big data and
machine learning in smart grid, and associated security concerns: a review. IEEE Access 7,
13960–13988 (2019). https://doi.org/10.1109/ACCESS.2019.2894819
5. I.H. Witten, E. Frank, M.A. Hall, C.J. Pal, Data Mining: Practical Machine Learning Tools and
Techniques (Elsevier, New York, 2016)
6. Z. Feng, Y. Zhu, A survey on trajectory data mining: techniques and applications. IEEE Access
4, 2056–2067 (2016). https://doi.org/10.1109/ACCESS.2016.2553681
7. Y. Zheng, Trajectory data mining: an overview. ACM Trans. Intell. Syst. Technol. 6(3), 1–41
(2015). https://doi.org/10.1145/2743025
8. A. Moradzadeh, K. Pourhossein, PCA-assisted location of small short circuit in transformer
winding, in 2020 28th Iranian Conference on Electrical Engineering (ICEE), (2020), pp. 1–6.
https://doi.org/10.1109/icee50131.2020.9260815
9. M. Bagheri, R. Esfilar, M.S. Golchi, C.A. Kennedy, A comparative data mining approach for
the prediction of energy recovery potential from various municipal solid waste. Renew. Sust.
Energ. Rev. 116, 109423 (2019). https://doi.org/10.1016/j.rser.2019.109423
10. Y. Noorollahi, A. Golshanfard, A. Aligholian, B. Mohammadi-ivatloo, S. Nielsen,
A. Hajinezhad, Sustainable energy system planning for an industrial zone by integrating electric
vehicles as energy storage. J. Energy Storage 30, 101553 (2020). https://doi.org/10.1016/j.est.
2020.101553
78 A. Moradzadeh et al.

11. M. Ghahramani, M. Nazari-Heris, K. Zare, B. Mohammadi-Ivatloo, Robust Optimal Planning

and Operation of Electrical Energy Systems (Springer, Cham, 2019)
12. D.A.C. Narciso, F.G. Martins, Application of machine learning tools for energy efficiency in
industry: a review. Energy Rep. 6, 1181–1199 (2020). https://doi.org/10.1016/j.egyr.2020.04.
035
13. A. Moradzadeh, A. Mansour-Saatloo, B. Mohammadi-Ivatloo, A. Anvari-Moghaddam, Perfor-
mance evaluation of two machine learning techniques in heating and cooling loads forecasting
of residential buildings. Appl Sci (Switzerland) 10(11), 3829 (2020). https://doi.org/10.3390/
app10113829
14. S. Fathi, R. Srinivasan, A. Fenner, S. Fathi, Machine learning applications in urban building
energy performance forecasting: a systematic review. Renew. Sust. Energ. Rev. 133, 110287
(2020). https://doi.org/10.1016/j.rser.2020.110287
15. A. Moradzadeh, S. Zakeri, M. Shoaran, B. Mohammadi-Ivatloo, F. Mohamamdi, Short-term
load forecasting of microgrid via hybrid support vector regression and long short-term memory
algorithms. Sustainability (Switzerland) 12(17), 7076 (2020). https://doi.org/10.3390/
su12177076
16. S.S. Roy, P. Samui, I. Nagtode, H. Jain, V. Shivaramakrishnan, B. Mohammadi-ivatloo,
Forecasting heating and cooling loads of buildings: A comparative performance
analysis. J. Ambient. Intell. Humaniz. Comput. 11(3), 1253–1264 (2020). https://doi.org/10.
1007/s12652-019-01317-y
17. G. Chitalia, M. Pipattanasomporn, V. Garg, S. Rahman, Robust short-term electrical load
forecasting framework for commercial buildings using deep recurrent neural networks. Appl.
Energy 278, 115410 (2020). https://doi.org/10.1016/j.apenergy.2020.115410
18. A. Moradzadeh, O. Sadeghian, K. Pourhossein, B. Mohammadi-Ivatloo, A. Anvari-
Moghaddam, Improving residential load disaggregation for sustainable development of energy
via principal component analysis. Sustainability (Switzerland) 12(8), 3158 (2020). https://doi.
org/10.3390/SU12083158
19. Y.T. Quek, W.L. Woo, T. Logenthiran, Load disaggregation using one-directional
convolutional stacked long short-term memory recurrent neural network. IEEE Syst. J. 14(1),
1395–1404 (2020). https://doi.org/10.1109/JSYST.2019.2919668
20. M. Kaselimi, E. Protopapadakis, A. Voulodimos, N. Doulamis, A. Doulamis, Multi-channel
recurrent convolutional neural networks for energy disaggregation. IEEE Access 7,
81047–81056 (2019). https://doi.org/10.1109/ACCESS.2019.2923742
21. A. Ahmadi, M. Nabipour, B. Mohammadi-Ivatloo, A.M. Amani, S. Rho, M.J. Piran, Long-term
wind power forecasting using tree-based learning algorithms. IEEE Access 8, 151511–151522
(2020). https://doi.org/10.1109/ACCESS.2020.3017442
22. M. AlKandari, I. Ahmad, Solar power generation forecasting using ensemble approach based on
deep learning and statistical methods. Appl. Comput. Inform. (2019). https://doi.org/10.1016/j.
aci.2019.11.002
23. H. Demolli, A.S. Dokuz, A. Ecemis, M. Gokcek, Wind power forecasting based on daily wind
speed data using machine learning algorithms. Energy Convers. Manag. 198, 111823 (2019).
https://doi.org/10.1016/j.enconman.2019.111823
24. C. Voyant et al., Machine learning methods for solar radiation forecasting: A review. Renew.
Energy 105, 569–582 (2017). https://doi.org/10.1016/j.renene.2016.12.095
25. A. Mansour-Saatloo, A. Moradzadeh, B. Mohammadi-Ivatloo, A. Ahmadian, A. Elkamel,
Machine learning based PEVs load extraction and analysis. Electronics (Switzerland) 9(7),
1–15 (2020). https://doi.org/10.3390/electronics9071150
26. A. Ahmadian, M. Sedghi, H. Fgaier, B. Mohammadi-ivatloo, M.A. Golkar, A. Elkamel, PEVs
data mining based on factor analysis method for energy storage and DG planning in active
distribution network: introducing S2S effect. Energy 175, 265–277 (2019). https://doi.org/10.
1016/j.energy.2019.03.097
4 Introduction to Machine Learning Methods in Energy Engineering 79

27. H. Jahangir et al., Charging demand of plug-in electric vehicles: forecasting travel behavior
based on a novel rough artificial neural network approach. J. Clean. Prod. 229, 1029–1044
(2019). https://doi.org/10.1016/j.jclepro.2019.04.345
28. A. Moradzadeh, K. Khaffafi, Comparison and evaluation of the performance of various types of
neural networks for planning issues related to optimal management of charging and discharging
electric cars in intelligent power grids. Emerg. Sci. J. 1(4), 201–207 (2017). https://doi.org/10.
28991/ijse-01123
29. J. Nowotarski, R. Weron, Recent advances in electricity price forecasting: a review of proba-
bilistic forecasting. Renew. Sust. Energ. Rev. 81, 1548–1568 (2018). https://doi.org/10.1016/j.
rser.2017.05.234
30. R. Weron, Electricity price forecasting: a review of the state-of-the-art with a look into the
future. Int. J. Forecast. 30(4), 1030–1081 (2014). https://doi.org/10.1016/j.ijforecast.2014.08.
008
31. K. Wang, C. Xu, Y. Zhang, S. Guo, A.Y. Zomaya, Robust big data analytics for electricity price
forecasting in the smart grid. IEEE Trans. Big Data 5(1), 34–45 (2017). https://doi.org/10.1109/
tbdata.2017.2723563
32. A. Moradzadeh, K. Pourhossein, B. Mohammadi-Ivatloo, F. Mohammadi, Locating inter-turn
faults in transformer windings using isometric feature mapping of frequency response traces.
IEEE Trans. Ind. Inform. (2020). https://doi.org/10.1109/tii.2020.3016966
33. A. Moradzadeh, K. Pourhossein, Application of support vector machines to locate minor short
circuits in transformer windings, in 2019 54th International Universities Power Engineering
Conference (UPEC), (2019), pp. 1–6
34. H. Momeni, N. Sadoogi, M. Farrokhifar, H.F. Gharibeh, Fault diagnosis in photovoltaic arrays
using GBSSL method and proposing a fault correction system. IEEE Trans. Ind. Inform. 16(8),
5300–5308 (2020). https://doi.org/10.1109/TII.2019.2908992
35. D.N. Coelho, G.A. Barreto, C.M.S. Medeiros, J.D.A. Santos, Performance comparison of
classifiers in the detection of short circuit incipient fault in a three-phase induction motor, in
2014 IEEE Symposium on Computational Intelligence for Engineering Solutions (CIES),
(2014), pp. 42–48. https://doi.org/10.1109/CIES.2014.7011829
36. A. Moradzadeh, K. Pourhossein, Early detection of turn-to-turn faults in power transformer
winding: an experimental study, in Proceedings 2019 International Aegean Conference on
Electrical Machines and Power Electronics, ACEMP 2019 and 2019 International Conference
on Optimization of Electrical and Electronic Equipment, OPTIM 2019, (2019), pp. 199–204.
https://doi.org/10.1109/ACEMP-OPTIM44294.2019.9007169
37. S. Zhang, Y. Wang, M. Liu, Z. Bao, Data-based line trip fault prediction in power systems using
LSTM networks and SVM. IEEE Access 6, 7675–7686 (2018). https://doi.org/10.1109/
ACCESS.2017.2785763
38. M. Mohammad Taheri, H. Seyedi, M. Nojavan, M. Khoshbouy, B. Mohammadi Ivatloo, High-
speed decision tree based series-compensated transmission lines protection using differential
phase angle of superimposed current. IEEE Trans. Power Deliv. 33(6), 3130–3138 (2018).
https://doi.org/10.1109/TPWRD.2018.2861841
39. J.J.Q. Yu, Y. Hou, V.O.K. Li, Online false data injection attack detection with wavelet
transform and deep neural networks. IEEE Trans. Ind. Inform. 14(7), 3271–3280 (2018).
https://doi.org/10.1109/TII.2018.2825243
40. A. Al-Abassi, H. Karimipour, A. Dehghantanha, R.M. Parizi, An ensemble deep learning-based
cyber-attack detection in industrial control system. IEEE Access 8, 83965–83973 (2020).
https://doi.org/10.1109/ACCESS.2020.2992249
41. D. Djenouri, R. Laidi, Y. Djenouri, I. Balasingham, Machine learning for smart building
applications. ACM Comput. Surv. 52(2), 1–36 (2019). https://doi.org/10.1145/3311950
42. I.H. Witten, E. Frank, M.A. Hall, Data Mining: Practical Machine learning (Elsevier,
New York, 2011)
43. C. Cortes, V. Vapnik, Support-vector networks. Mach. Learn. 20(3), 273–297 (1995). https://
doi.org/10.1023/A:1022627411411
80 A. Moradzadeh et al.

44. A. Zendehboudi, M.A. Baseer, R. Saidur, Application of support vector machine models for
forecasting solar and wind energy resources: a review. J. Clean. Prod. 199, 272–285 (2018).
https://doi.org/10.1016/j.jclepro.2018.07.164
45. A. Moradzadeh, S. Zeinal-Kheiri, B. Mohammadi-Ivatloo, M. Abapour, A. Anvari-
Moghaddam, Support vector machine-assisted improvement residential load disaggregation,
in 2020 28th Iranian Conference on Electrical Engineering (ICEE), (2020), pp. 1–6. https://doi.
org/10.1109/icee50131.2020.9260869
46. A.G. Ivakhnenko, Polynomial theory of complex systems. IEEE Trans. Syst. Man Cybern. 1(4),
364–378 (1971). https://doi.org/10.1109/TSMC.1971.4308320
47. I. Ebtehaj, H. Bonakdari, A.H. Zaji, H. Azimi, F. Khoshbin, GMDH-type neural network
approach for modeling the discharge coefficient of rectangular sharp-crested side weirs. Eng.
Sci. Technol. 18(4), 746–757 (2015). https://doi.org/10.1016/j.jestch.2015.04.012
48. H. Jafarian, H. Sayyaadi, F. Torabi, Modeling and optimization of dew-point evaporative
coolers based on a developed GMDH-type neural network. Energy Convers. Manag. 143,
49–65 (2017). https://doi.org/10.1016/j.enconman.2017.03.015
49. N. Nariman-Zadeh, A. Darvizeh, A. Jamali, A. Moeini, Evolutionary design of generalized
polynomial neural networks for modelling and prediction of explosive forming
process. J. Mater. Process. Technol. 164–165, 1561–1571 (2005). https://doi.org/10.1016/j.
jmatprotec.2005.02.020
50. V.N. Vapnik, The Nature of Statistical Learning Theory (Springer, New York, 1995)
51. J. Antonanzas, R. Urraca, F.J. Martinez-De-Pison, F. Antonanzas-Torres, Solar irradiation
mapping with exogenous data from support vector regression machines estimations. Energy
Convers. Manag. 100, 380–390 (2015). https://doi.org/10.1016/j.enconman.2015.05.028
52. F. Antonanzas-Torres, R. Urraca, J. Antonanzas, J. Fernandez-Ceniceros, F.J. Martinez-de-
Pison, Generation of daily global solar irradiation with support vector machines for regression.
Energy Convers. Manag. 96, 277–286 (2015). https://doi.org/10.1016/j.enconman.2015.02.086
53. Specht, Probabilistic neural networks for classification, mapping, or associative memory, in
IEEE International Conference on Neural Networks, (1988), pp. 525–532. https://doi.org/10.
1109/ICNN.1988.23887
54. C.M. Hong, F.S. Cheng, C.H. Chen, Optimal control for variable-speed wind generation
systems using general regression neural network. Int. J. Electr. Power Energy Syst. 60, 14–23
(2014). https://doi.org/10.1016/j.ijepes.2014.02.015
55. Y.W. Huang, M.Q. Chen, Y. Li, J. Guo, Modeling of chemical exergy of agricultural biomass
using improved general regression neural network. Energy 114, 1164–1175 (2016). https://doi.
org/10.1016/j.energy.2016.08.090
56. J. Nirmal, M. Zaveri, S. Patnaik, P. Kachare, Voice conversion using general regression neural
network. Appl. Soft Comput. 24, 1–12 (2014). https://doi.org/10.1016/j.asoc.2014.06.040
57. Z. Yu, F. Haghighat, B.C.M. Fung, H. Yoshino, A decision tree method for building energy
demand modeling. Energ. Buildings 42(10), 1637–1646 (2010). https://doi.org/10.1016/j.
enbuild.2010.04.006
58. P. Moutis, S. Skarvelis-Kazakos, M. Brucoli, Decision tree aided planning and energy
balancing of planned community microgrids. Appl. Energy 161, 197–205 (2016). https://doi.
org/10.1016/j.apenergy.2015.10.002
59. S. Salzberg, Book Review-C4. 5: Programs for Machine Learning (Morgan Kaufmann, Bur-
lington, 1993)
60. R. Yan, Z. Ma, Y. Zhao, G. Kokogiannakis, A decision tree based data-driven diagnostic
strategy for air handling units. Energ. Buildings 133, 37–45 (2016). https://doi.org/10.1016/j.
enbuild.2016.09.039
61. J.R. Quinlan, Induction of decision trees. Mach. Learn. 1(1), 81–106 (1986). https://doi.org/10.
1023/A:1022643204877
62. K. Benmouiza, A. Cheknane, Forecasting hourly global solar radiation using hybrid k-means
and nonlinear autoregressive neural network models. Energy Convers. Manag. 75, 561–569
(2013). https://doi.org/10.1016/j.enconman.2013.07.003
4 Introduction to Machine Learning Methods in Energy Engineering 81

63. S. Li, H. Ma, W. Li, Typical solar radiation year construction using k-means clustering and
discrete-time Markov chain. Appl. Energy 205, 720–731 (2017). https://doi.org/10.1016/j.
apenergy.2017.08.067
64. K. Wang, X. Qi, H. Liu, J. Song, Deep belief network based k-means cluster approach for short-
term wind power forecasting. Energy 165, 840–852 (2018). https://doi.org/10.1016/j.energy.
2018.09.118
65. A. Shrestha, A. Mahmood, Review of deep learning algorithms and architectures. IEEE Access
7, 53040–53065 (2019). https://doi.org/10.1109/ACCESS.2019.2912200
66. Y. Lecun, Y. Bengio, G. Hinton, Deep learning. Nature 521(7553), 436–444 (2015). https://doi.
org/10.1038/nature14539
67. D. Zhang, X. Han, C. Deng, Review on the research and practice of deep learning and
reinforcement learning in smart grids. CSEE J. Power Energy Syst. 4(3), 362–370 (2018).
https://doi.org/10.17775/CSEEJPES.2018.00520
68. L. Zhang, J. Lin, B. Liu, Z. Zhang, X. Yan, M. Wei, A review on deep learning applications in
prognostics and health management. IEEE Access 7, 162415–162438 (2019). https://doi.org/
10.1109/ACCESS.2019.2950985
69. A. Moradzadeh, K. Pourhossein, Location of disk space variations in transformer winding using
convolutional neural networks, in 2019 54th International Universities Power Engineering
Conference, UPEC 2019 - Proceedings, (2019), pp. 1–5. https://doi.org/10.1109/UPEC.2019.
8893596
70. A. Moradzadeh, K. Pourhossein, Short circuit location in transformer winding using deep
learning of its frequency responses, in Proceedings 2019 International Aegean Conference
on Electrical Machines and Power Electronics, ACEMP 2019 and 2019 International Confer-
ence on Optimization of Electrical and Electronic Equipment, OPTIM 2019, (2019),
pp. 268–273. https://doi.org/10.1109/ACEMP-OPTIM44294.2019.9007176
71. P. Li, Z. Chen, L.T. Yang, Q. Zhang, M.J. Deen, Deep convolutional computation model for
feature learning on big data in internet of things. IEEE Trans. Ind. Inform. 14(2), 790–798
(2018). https://doi.org/10.1109/TII.2017.2739340
72. N. Koroniotis, N. Moustafa, E. Sitnikova, Forensics and deep learning mechanisms for botnets
in internet of things: A survey of challenges and solutions. IEEE Access 7, 61764–61785
(2019). https://doi.org/10.1109/ACCESS.2019.2916717
73. J. Han, S. Miao, Y. Li, W. Yang, H. Yin, A wind farm equivalent method based on multi-view
transfer clustering and stack sparse auto encoder. IEEE Access 8, 92827–92841 (2020). https://
doi.org/10.1109/ACCESS.2020.2993808
74. Z.A. Khan, S. Zubair, K. Imran, R. Ahmad, S.A. Butt, N.I. Chaudhary, A new users rating-trend
based collaborative denoising auto-encoder for top-N recommender systems. IEEE Access 7,
141287–141310 (2019). https://doi.org/10.1109/ACCESS.2019.2940603
75. W. Wang, X. Du, D. Shan, R. Qin, N. Wang, Cloud intrusion detection method based on
stacked contractive auto-encoder and support vector machine. IEEE Trans. Cloud Comput.
(2020). https://doi.org/10.1109/TCC.2020.3001017
76. D.A. Clevert, T. Unterthiner, S. Hochreiter, Fast and accurate deep network learning by
exponential linear units (ELUs), in 4th International Conference on Learning Representations,
ICLR 2016 - Conference Track Proceedings, (2016)
77. Y. Bengio, P. Simard, P. Frasconi, Learning long-term dependencies with gradient descent is
difficult. IEEE Trans. Neural Netw. 5(2), 157–166 (1994). https://doi.org/10.1109/72.279181
78. S. Hochreiter, J. Schmidhuber, Long short-term memory. Neural Comput. 9(8), 1735–1780
(1997). https://doi.org/10.1162/neco.1997.9.8.1735
79. K. Wang, X. Qi, H. Liu, Photovoltaic power forecasting based LSTM-convolutional network.
Energy 189, 116225 (2019). https://doi.org/10.1016/j.energy.2019.116225
80. A. Mohamed, G.E. Dahl, G. Hinton, Acoustic modeling using deep belief networks. IEEE
Trans. Audio Speech Lang. Process. 20(1), 14–22 (2012). https://doi.org/10.1109/TASL.2011.
2109382
82 A. Moradzadeh et al.

81. C.-Y. Zhang, C.L.P. Chen, M. Gan, L. Chen, Predictive deep Boltzmann machine for
multiperiod wind speed forecasting. IEEE Trans. Sustain. Energy 6(4), 1416–1425 (2015).
https://doi.org/10.1109/TSTE.2015.2434387
82. B. Choubin, S. Khalighi-Sigaroodi, A. Malekian, Ö. Kişi, Multiple linear regression, multi-
layer perceptron network and adaptive neuro-fuzzy inference system for forecasting precipita-
tion based on large-scale climate signals. Hydrol. Sci. J. 61(6), 1001–1009 (2016). https://doi.
org/10.1080/02626667.2014.966721
83. R. Wang, S. Lu, W. Feng, A novel improved model for building energy consumption prediction
based on model integration. Appl. Energy 262, 114561 (2020). https://doi.org/10.1016/j.
apenergy.2020.114561
84. K. Amasyali, N.M. El-Gohary, A review of data-driven building energy consumption prediction
studies. Renew. Sustain. Energy Rev. 81, 1192–1205 (2018). https://doi.org/10.1016/j.rser.
2017.04.095
85. Z. Xuan, Z. Xuehui, L. Liequan, F. Zubing, Y. Junwei, P. Dongmei, Forecasting performance
comparison of two hybrid machine learning models for cooling load of a large-scale commercial
building. J. Build. Eng. 21, 64–73 (2019). https://doi.org/10.1016/j.jobe.2018.10.006
86. S. Sekhar Roy, R. Roy, V.E. Balas, Estimating heating load in buildings using multivariate
adaptive regression splines, extreme learning machine, a hybrid model of MARS and ELM.
Renew. Sustain. Energy Rev. 82, 4256–4268 (2018). https://doi.org/10.1016/j.rser.2017.05.249
87. M. Hossain, M.N. Sulaiman, A review on evaluation metrics for data classification evaluations.
Int. J. Data Min. Knowl. Manage. Process 5(2), 01–11 (Mar. 2015). https://doi.org/10.5121/
ijdkp.2015.5201
88. W. Kong, Z.Y. Dong, B. Wang, J. Zhao, J. Huang, A practical solution for non-intrusive type II
load monitoring based on deep learning and post-processing. IEEE Trans. Smart Grid 11(1),
148–160 (2020). https://doi.org/10.1109/TSG.2019.2918330
Chapter 5
Introduction and Literature Review
of the Application of Machine
Learning/Deep Learning to Control
Problems of Power Systems

Samira Sadeghi, Ali Hesami Naghshbandy, Parham Moradi,

and Navid Rezaei

5.1 Introduction

Due to the significant change in the structure of the power system and its transfor-
mation into a modern power system, which is expected to be more pronounced in the
future, the need to find advanced techniques to solve complex problems of the power
system has become very prominent [1]. The evolution of the power system can be
summarized as follows:
Changes in power generation sources have shifted power generation from large
thermal power plants to smaller and distributed generations, such as wind [2] and
solar [3], at all levels of transmission and distribution. The presence of small
distributed generations and renewable energy sources along with electronic power
converters at all levels of the power system causes uncertainty in power generation
and reduces system inertia resulting in faster dynamics because the presence of
renewable energy sources challenges frequency regulation and control [4].
The existence of new energy storage technologies, including types of large
storage devices connected to the transmission network and small storage devices
connected to the distribution network and microgrids, are the main factors in
identifying a modern power system [5]. These new technologies make it possible

S. Sadeghi · A. Hesami Naghshbandy · N. Rezaei (*)

Faculty of Engineering, Department of Electrical Engineering, University of Kurdistan,
Sanandaj, Iran
e-mail: [email protected]; [email protected]; [email protected]
P. Moradi
Faculty of Engineering, Department of Computer Engineering, University of Kurdistan,
Sanandaj, Iran
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 83

M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_5
84 S. Sadeghi et al.

to balance generation and consumption in sensitive and uncertain conditions of

power system performance and can act as important control devices in different
operating modes [1].
The loads of the power system, which are also considered as a kind of producer,
have caused a change in the load profile. These loads include electronic devices that
are available in all homes and industrial buildings and are used to meet load demand
response and participate in electricity markets and control the power system. New
types of electric loads include electric vehicles with charging stations installed
anywhere in the power system, which have changed the load profile with the
possibility of charging on the load side [6].
Increasing the use of Flexible Alternating Current Transmission Systems
(FACTS) and various types of static and dynamic compensators for more efficient
control of the power system [7] and also increasing the establishment of High
Voltage Direct Current (HVDC) lines for easy voltage transfer are important and
essential changes in the power system [8], which have made it necessary to evaluate
the power system and check its control.
The creation of micro-energy networks as part of the power system is one of the
undeniable factors in evaluating the control of the power system. These controllable
microgrids are considered as an independent and single geographical area with the
ability to produce and consume together, which includes a variety of small, renew-
able and distributed sources [9]. The existence of microgrids with the ability to
operate as connected to the grid and separate from that has a significant role in
creating flexibility in controlling the system and after the occurrence of faults and
during network restoration [4].
Finally, all the changes in the power network have led to the modernization of the
power system and the creation of the smart grid. Outstanding technologies for smart
grids include advanced communication and measurement infrastructure, Phasor
Measurement Units (PMU) in transmission and distribution systems [10], and
smart meters in distribution networks. Thus, the emergence of loads with the ability
to produce and consume, a variety of Internet of Things (IOT)-based equipment,
advanced telecommunications equipment, etc. have caused the smart grid to face a
variety of data and many complexities [11]. Therefore, the need for complex
calculations and the introduction of new methods in controlling the modern power
grid at any time and in any mode of operation have become important [12].
In power system control problems, due to the large amount of data (big data), the
need of algorithms based on data analysis is essential. These types of algorithms that
are based on data measurement are called machine learning algorithms. In other
words, machine learning methods are applied to analyzing, processing, predicting,
and categorizing big data in complex problems [13]. The purpose of using these
types of algorithms is to extract patterns and order in the data. In fact, learning means
improving behavior based on past experience. Therefore, the reason for turning to
machine learning in the field of power systems includes the following [14]:
• Strengthening computing environments and producing sufficient
information data.
5 Introduction and Literature Review of the Application of Machine. . . 85

• Complexity and nonlinearity of power system security issues.

• Generation of sufficiently rich security information databases with acceptable
responses.
• Contestations of power companies in order to apply the approach of their
proposed power system and meeting and responding the practical needs.
• Rapid and dynamic changes in the organization of the power system that will
require the use of regular approaches in assessing dynamic security to maintain
reliability.
• Recent events and blackouts in various parts of the world have highlighted the use
of machine learning techniques to predict various disturbances.
• Generating large volumes of diverse data in various fields of the power system by
moving toward smart power networks and its rapid development.
Recently, machine learning methods have been shown in numerous applications,
such as power system control and stability studies and Dynamic Security
Assessment.
Over the past few decades, with the development of advanced infrastructure and
extensive interconnection of modern applications in the power system including
Supervisory Control and Data Acquisition (SCADA), Phasor Measurement Unit
(PMU), and modern telecommunication and communication technologies, the
power grid is increasingly exposed to external damage and threats [15]. Monitoring
on the condition of the power system, especially after disruptions and when the loads
are changing, has always been one of the most important issues and concerns. Also,
connecting renewable energies and developing the network with the ability to
produce and consume in one place, increasing the number of consumers and,
consequently, transmission lines, have faced the security assessment and stability
with many challenges [13].

5.1.1 Reliability and Security

The probability of favorable power system performance and with minimum inter-
ruptions over a long period is called reliability. The ability of the power system to
deal with impendent disturbances without interrupting services to electricity con-
sumers is called security, which depends on the operating conditions of the system as
well as the type of disturbance [16]. Mathematical analysis of the system response to
changes made after a fault in the system and the new equilibrium conditions is called
system security analysis. This analysis is called Static Security Assessment (SSA) if
the analysis only assesses equilibrium conditions after turbulence (steady-state
operating point). If the system evaluates the transient performance of the system
after it malfunctions, it is called a Dynamic Security Assessment (DSA) [17].
In DSA offline analysis, accurate time domain stability analysis is performed for
all valid disturbances and types of operating conditions, there is no severe limit in
calculations time, and detailed analysis is performed for a wide range of conditions
86 S. Sadeghi et al.

and probabilities [14]. Online DSA analysis is used to complete and update offline
DSA to consider current operating conditions. In DSA online, a rapid screening
process is performed to limit the number of occurrences, rapidly assess stability by
predicting stability or calculating instability margins, and assess the severity of
turbulence. Traditional time domain simulations also involve extensive numerical
integrals to detect oscillation paths and voltage changes. In DSA online, calculation
time is of particular importance [18].

5.1.2 Stability

The continued smooth operation of the power system performed by the occurrence
of a perturbation is called stability, which depends on the operating conditions and
the nature of the physical disturbance. Power system stability is divided into three
categories: steady-state stability, dynamic stability, and transient stability
[19]. Steady-state stability, the ability of different machines in the power system to
restore and synchronize after a small and quiet disturbance, is like a gradual change
of load. Transient stability refers to the stability after a sudden large disturbance such
as a fault, generator failure, switching operation, and large load change. Dynamic
stability is a state between steady-state and transient stability and is overcome by
voltage regulators, controllers, and governors. In this type of stability, small pertur-
bations of 10–30 s are desired [16]. Types of stability include transient and small
signal stability for rotor angle and voltage and frequency stability [19]. Stability
assessment methods include power flow analysis, eigenvalue analysis, time domain
simulation, numerical integration method, direct or Lyapunov method, probabilistic
methods, expert system and metaheuristic methods, and database analysis and
machine learning approaches [18].
Therefore, models and various methods in security and stability studies of power
system have been presented. Despite the almost satisfactory performance of con-
ventional methods, these techniques are computationally complex, fiscally expen-
sive, inaccurate, and time-consuming. In this regard, recently, machine learning and
deep learning techniques have been highly regarded in monitoring complex appli-
cations and their modeling [20]. Unlike traditional methods, machine learning
methods are computationally robust, principled, and explicitly reliable when used
in classification studies. Machine learning methods have the ability of learning and
understanding the features of changing different loads, grid data, etc. in normal,
emergency, and post-fault conditions. Therefore, these methods are specific to the
dynamic smart grid system [21].
5 Introduction and Literature Review of the Application of Machine. . . 87

5.2 Overview of Machine Learning and Deep Learning

Machine learning (ML) and deep learning (DL) are two topics that are widely used
these days in computer science and other engineering disciplines.
Machine learning is used in computational problems in which designing and
programming explicit algorithms with appropriate performance is difficult or impos-
sible. The fundamental goal of ML algorithms is to generalize learning beyond the
trained examples that is the successful interpretation of data. In ML, systems learn by
receiving various inputs and performing statistical analysis to generate outputs in a
specific statistical range. Therefore, ML allows computers to automatically perform
the decision-making process on new input data after receiving the sample data and
modeling it [22]. Deep learning is a subset of machine learning which is able to
predict outputs and make decision-making patterns by artificial intelligence and
mimicking the function of the human brain. DL refers to the Artificial Neural
Network that consists of several layers of learning. This method allows machines
to solve complex problems even when using very diverse, unstructured, and
interconnected data. The DL algorithm will perform better when it learns more
deeply [23]. Each machine learning algorithm is a subset of the field of artificial
intelligence. Deep learning is also a subset of machine learning and tries to extract
information based on multilevel learning. In fact, the place of artificial intelligence,
machine learning, and deep learning are shown in Fig. 5.1. It should be noted that
artificial intelligence is any code, technique, or algorithm that enables machines to
behave like humans. Now this code may consist of some condition written by the
programmer, or it may contain a mathematical relation that produces a fixed value
for a particular input [24].

Fig. 5.1 Position of

artificial intelligence,
machine learning, and deep
learning [24] Field of Artificial Intelligence

Field of Machine Learning

Field of
Deep Learning
88 S. Sadeghi et al.

5.2.1 Machine Learning Technique

Over the past four decades, machine learning has made great strides in terms of
theoretical understanding and actual applications in many different fields, the main
reason for achieving these important successes being its dramatic increase in the
computing potency.
ML is a category of techniques applied in big data processing with the develop-
ment of algorithms and a category of principles to provide the outcomes needed by
users. This method is mostly used for the development of automatic machines by
implementing algorithms and a category of predefined principles [22]. In ML, data is
used, and a set of rules will be executed by the mentioned algorithm. Therefore,
machine learning techniques are classified as automatic instructions to produce the
desired results. However, this operation can be done without human role and
automatically converts the data into a template and will automatically detect the
production problem. Machine learning is a more dynamic and robust method. Also,
machine learning algorithms try to minimize the error by defining the error function.
Before learning begins, the parameters start randomly and then gradually adapt
[23]. Machine learning consists of several parts; first, the data is entered into the
system, then the data is preprocessed, and if it has noise or a perturbation, it is first
deleted, and then a series of properties are extracted from the raw data. These
properties will represent the raw data. This data is then passed to a classification or
clustering algorithm that plays a decision-maker role. The characteristics of a
machine learning method depend on the following [14]:
Feature Selection: The purpose of this item is to decrease the data and dimen-
sions of the input space, which, by reducing the attributes that have not any
beneficial information to forecast the desired output information, the dimensions
are reduced.
Model Selection: Usually, models are selected that have the best fit of learning
situations in the predefined class. This requires the organized selection of the
parameters and building of model using the temporary search technique or optimi-
zation approaches appropriate for the type of the desired model.
Exegesis and Validation: This item is synthesized to understand the systemic
meaning of the model, and it is very important to determine its validity. Comparison
of its information can be obtained from the model with previous allocation and tested
on a set of invisible experimental samples.
Model Using: The application of the model is to predict the output using the
assumed values of the input parameters. Sometimes, the inversion of the model is
performed to prepare the information to change the input parameters in order to
attain a certain output.
Machine learning methods are used in practical problems for the purposes of
diagnosis or prediction. In diagnosis, the probabilistic values assumed by the
attributes are the result of the causality of membership in the classification, but in
predicting, the probabilistic values are consistent with the future state of the system.
5 Introduction and Literature Review of the Application of Machine. . . 89

Data selection is another important issue in machine learning methods for which the
following should be considered:
• The data must be a good representation of the problem because it is to be modeled
by an algorithm.
• Poor data quality (for example, data that is noisy, out of data, or that not much
data is available) cannot lead to a suitable model. The solution here is to remove
data noise, outdated data, and unspecified data before learning.
• There are a lot of unrelated and redundant (features that are similar and some of
which are not necessary) features for the data.
Practical limitations in implementing machine learning methods include the high
cost of collecting databases, complexity of learning according to computational
needs in terms of time and CPU memory, and execution cost due to the complexity
of the used algorithms. However, challenges of machine learning algorithms in
different issues include wrong choice of learning algorithm, inadequate and incorrect
selection of data for the training process, and overlearning of algorithm.

5.2.2 Deep Learning Technique

Deep learning is a type of neural network which attracts, processes, and calculates
metadata as input through some nonlinear conversion layers and returns it as output
data. This algorithm has a unique feature which is the feature of automatic extrac-
tion. This means that the algorithm understands the required and related attributes to
solve the problem. Also, deep learning algorithm reduces the tasks of programmers
to explicitly select features. This algorithm is even used to solve supervised or
unsupervised challenges [23]. In DL, each hidden layer teaches a set of special
and unique attributes that the performance of each layer is based on the output of the
former layer. This type of hierarchical learning also converts low-level features into
high-level features. In doing so, the DL is applied to solve the intricate subjects
involving multiple nonlinear layers [22]. Large neural networks are called deep
learning. The core of deep learning is based on having fast computers and enough
data to train large neural networks. In deep learning, the model may also achieve
greater accuracy than human accuracy [23]. Neural networks could have two or three
layers, but today, Deep Neural Networks can have as many as 150 layers. Each
neuron tries to store the lowest layer of information and transfer it to the next layers
of neurons. This is done in such a way that the information inside the neurons is
extracted along with a hidden layer of neuronal information. With this account, it can
be concluded that the data goes from the lowest layer to the highest layer and collects
information [25].
Deep learning models require large volumes of labeled data and neural network
architecture. These models extract features automatically and do not require manual
feature extraction. One of the most common models of Deep Neural Networks is the
Convolutional Neural Networks, abbreviated CNN or Conv Net. In deep learning,
90 S. Sadeghi et al.

the feature extraction and classification sections do not exist separately, and the data
is applied directly to the algorithm and is obtained at the output of the data label. In
fact, the classification section and the feature extraction section are merged [22].

5.2.3 Categorization of ML and DL

Deep learning is a special mode of machine learning. In machine learning, work

begins with the manual extraction of features, and features are used to build a model
that does the job of categorization, but in deep learning, the feature extraction
process is not manual and is done automatically. Another difference is that in deep
learning, the model expands as the amount of data increases, but in machine
learning, the model expands to a certain extent and then does not change as the
data increases. One of the key features of deep learning is that it improves with
increasing data [22]. The most general categorizations for machine learning and deep
learning algorithms include supervised learning, unsupervised learning, and rein-
forcement learning [23]. Another categorization of machine learning algorithms
include semilogarithmic learning, offline learning, online learning, sample-based
learning, and model-based learning [25].
Supervised Learning: The model is trained using specific inputs and outputs and
consists of two general categories [23]:
Countable (discrete) outputs: that use classification algorithms.
Uncountable (continuous) outputs: that use regression algorithms.
Some of the most important supervised classification methods are as follows:
Linear Regression, Logistic Regression, Support Vector Machine (SVM), k-Nearest
Neighbor (k-NN), Decision Tree (DT), Random Forest (RF), Naive Bayes, Concept
Learning, Neural Network, and Gradient Boosting.
Unsupervised Learning: The model is trained using a series of data whose
output is not known and includes two general categories as follows [25]:
Clustering: Only similarities are used for clustering.
Association rules (dimensional reduction algorithms): This category includes the
analysis of input data and the extraction of relationships between features and
attributes.
Some of the most important unsupervised clustering methods are as follows:
Hierarchical Agglomerative Clustering (HAC), k-Means, C-Means, Density-Based
Gaussian (DBSCAN), Community Detection, Mean-Shelf Clustering, and Gaussian
Mixture Models (GMM).
Reinforcement Learning: This is one of the machine learning trends that is
inspired by behavioral psychology. This method focuses on the behaviors that
the machine must perform to maximize its reward. Here, the algorithm learns the
optimal behavior based on the feedback (reward or penalty) it receives from the
environment. Reinforcement learning has two major differences from supervised
learning: The first is that it does not have the right input and output pairs, and the
dysfunctional behaviors are not corrected from the outside, and the second is that
5 Introduction and Literature Review of the Application of Machine. . . 91

there is a strong focus on live performance that requires finding the proper balance
between discovering new things and utilizing stored knowledge [26].
Also, it should be noted that the choice of machine learning method to solve
the problem depends on factors such as the nature and complexity of the problem,
the type and nature of the required data, the limitations and scope of learning, and the
expected outcome and results of the problem. In this regard, the tasks of machine
learning techniques include forecasting, modeling, categorization, and classification.
In addition, most research contains machine learning algorithms based on methods
for selecting, extracting, reducing, and optimizing the features of problem data.
Some of important selection, extraction, and reduction methods include Principal
Component Analysis (PCA), Forward Feature Selection (FFS), Backward Feature
Selection (BFS), Wavelet Transform (WT), Partial Least Squares Regression
(PLSR), and Factor Analysis (FA). Furthermore, data preprocessing and optimiza-
tion methods in most cases include metaheuristic algorithms such as Genetic Algo-
rithm (GA), Artificial Bee Colony (ABC), Particle Swarm Optimization (PSO), and
Ant Colony Optimization (ACO).
According to the above categories, many machine learning methods are classified
in Fig. 5.2. Also, most of the scholars’ research is devoted to hybrid machine
learning methods or multiple strategies.

5.3 Overview of Power System Control Problems

The power system is the most important vital infrastructure in the modern world.
Thus, this system with proper design and operation must meet the following basic
requirements:
• It must be able to meet the growing demand for active and reactive powers and
should maintain and control the spinning reserve requirements.
• The quality of the supplied power must satisfy the minimum required standards
according to the frequency stabilization, voltage stabilization, and reliability
level.
• The system must provide energy at the lowest cost [16].
Also, preventing partial and total blackouts caused by various disturbances in the
power system are the important goals in controlling the system against all kinds of
faults. In assessing the security of the power system, different operational modes are
investigated [13]. To meet the above needs, different levels of control are used that
includes a complex set of equipment. Figure 5.3 shows different parts of a power
system with their corresponding controls. The controllers create the proper operation
of the system by maintaining the voltage, frequency, and other variables of the
system within the allowable range. These controllers have critical impacts on the
dynamical performance of the power system. The goals of controlling a power
system depend on its operating conditions. Typically, the aim is to operate the
 92

Naive Bayes (NB)

k-Nearest Neighbor (k-NN) Averaged One -Dependence Estimators (AODE)
Learning Vector Quantization (LVQ) Bayesian Belief Network (BBN)
Radial Basis Function Network (RBFN)
Self-Organizing Map (SOM) Gaussian Naive Bayes (GNB)
Bayesian

Perceptron
Locally Weighted Learning (LWL) Multinomial Naive Bayes (MNB)

Instance Based
Back-Propagation
Bayesian Network (BN)
Hopfield Network (HN)

Neural Networks
k-Means
Classification and Regression Tree (CART)
C-Means
C4.5
Ridge Regression X-Means
C5.0
Least Absolute Shrinkage and Selection Operator (LASSO) PG-Means
Chi-squared Automatic Interaction Detection (CHAID)
Elastic Net k-Medians
Decision Stump (DS)

Clustering
Decision Tree

Expectation Maximization

Regularization
Least Angle Regression (LARS)
Conditional Decision Trees (CDT)
Hierarchical Clustering (HC)
MS
Mean-Shelf Clustering
Cubist Graph Clustering Principal Component Analysis (PCA)
One Rule (One R) Forward Feature Selection (FFS)
Deep Boltzmann Machin (DBM)
Zero Rule (Zero R) Backward Feature Selection (BFS)
Deep Belief Networks (DBN)

Rule System
Repeated Incremental Pruning to Produce Error Reduction (RIPPER) Partial Least Squares Regression (PLSR)
Convolutional Neural Network (CNN)
Factor Analysis (FA)
Deep Learning

Fig. 5.2 Machine learning methods categorization [14, 22, 23, 25]
Stacked Auto-Encoders
Multidimensional Scaling (MDS)
Linear Regression Projection Pursuit (PP)
Random Forest (RF)
Ordinary Least Squares Regression (OLSR) Gradient Boosting Machines (GBM) Principal Component Regression (PCR)
Stepwise Regression (SR) Boosting Partial Least Squares Discriminant (PLSD)
Multivariate Adaptive Regression Splines (MARS) Bootstrapped Aggregation (Bagging) Mixture Discriminant Analysis (MDA)

Regression
Dimensionality Reduction

Locally Estimated Scatterplot Smoothing (LOESS) Quadratic Discriminant Analysis (QDA)

Ensemble

Ada Boost
Logistic Regression Regularized Discriminant Analysis (RDA)
Stacked Generalization (Blending)
Gradient Boosted Regression Trees (GBRT) Flexible Discriminant Analysis (FDA)
Linear Discriminant Analysis (LDA)
S. Sadeghi et al.
5 Introduction and Literature Review of the Application of Machine. . . 93

Generator
Frequency Tie Flows Power

System Generation Control

Schedule
Economical Load-Frequency Control

Supplementary
Control

Prime Mover and

Generating Unit Controls Other
Control
Generating
Unit and
Shaft Associated
Field Power Controls
Excitation Current
System and Generator
Control
Voltage Speed

Speed/ Power

Electrical Power

Transmission Control
Reactive Power and Voltage Control, HVDC
Transmission and Control

Generator
Frequency Tie Flows Power

Fig. 5.3 Subsystems of a power system and their control [16]

94 S. Sadeghi et al.

power system with the best possible efficiency, while the voltage and frequency are
close to nominal values [18].
When an unusual situation occurs, new goals must be defined to get the system
back to the normal condition. It is rare for a single disturbance in a power system to
lead to a major breakdown and the collapse of a seemingly secure system. Such
interruptions usually occur by combining a set of events that put pressure on the
system beyond its capabilities. Under normal circumstances, all power system vari-
ables are within the allowable range. The goal of controllers of a power system is to
restore a disrupted system to the normal situation. Hence, it should be noted that
dynamical performance of a large interconnected power system is affected by a wide
range of equipment failures and outages [16]. In relation to controlling various
issues, it should be noted that most control functions such as automatic generation
control are performed in control centers located in the power plant. Substations also
allow the introduction of some local control functions such as control of distributed
generation sources and load management. The control process is often hierarchically
and through the exchange of information and decision-making and control signals.
The control process must be set instantaneously, accurately, and in coordination
between the control commands [14].

5.4 Evaluation of Operating Modes of Power System

Control Using Machine Learning Methods

It is an undeniable fact that a power system cannot always operates in its sustained
state because in a power system, there is always the possibility of various types of
faults such as three-phase to ground, single-phase to ground, etc. Therefore, after
identifying the behavior of the system when these faults occur, appropriate control
strategies should be established [18].
Control and stability assessment of the power system is in the face of various
disturbances and suggests appropriate treatment measures to resolve its main weak-
nesses at any time. Perturbations and faults may be due to outer or inner occurrences.
For example, these faults are caused by lightning (external) and switching
(internal) [14].
Before a disturbance occurs, the necessary preventives must be taken to deal with
any type of disturbance so that the system can continue to operate. Also, generation
and demand of energy should be balanced, and all variables of the power system
should be in their operational range. When a disturbance occurs and in an emergency
state, the necessary arrangements must be adopted to maintain synchronicity and
protect the power system. After a disturbance occurs, when a part of the network or
the entire network is removed, the necessary operations must be performed to
quickly restore the power system [20]. Therefore, the control system is examined
in preventive, normal, emergency, and restorative operating modes, which most
studies in controlling the power system are done in two control modes including
5 Introduction and Literature Review of the Application of Machine. . . 95

Preventive State

Secure
Normal State
Maximizing the economy and minimizing the contingency uncertain effect

Preventive
control

Insecure
Restorative Alert
Resynchronization Tradeoff of
Load Pickup Preventive and
Corrective Control

Emergency
Control
(Corrective)

Protections Emergency
In Extremis Overloads,

Asecure
Partial or Total Under voltages,
Split
Service Interruption Under frequency
Load
Instabilities
Shedding

System Not Intact

Control and/ or protective actions

Foreseen or unforeseen disturbances

Fig. 5.4 Operating states and transitions [14]

prevention and emergency. Figure 5.4 displays a detailed explanation of the scheme
provided by Dy Liacco, which defines the operating modes of a power system [14].
The use of PMUs in modern power systems has led to significant advances in
researches based on measurable data. Nevertheless, main subjects such as big data
related to PMUs, uncertainties related to measurement faults, and nonlinearity of
loads in future and current power system, and computationally complexity of them,
expose the constraints of traditional control methods. Therefore, the need for fast,
reliable, advanced, scanning, calibrated, and dynamic methods in controlling the
modern power system and in the face of continuous blackouts and other threats to the
power system has become essential [10, 27].
Some fast system controls include protective relays, turbine governors, and
automatic voltage regulators on generators that operate in fractions of seconds,
while subsystem level controls such as secondary voltage control and automatic
generation control during act for a few seconds. Other types of controls (preventive
and emergency) are activated by detecting system limitations and disturbances [26].
96 S. Sadeghi et al.

5.4.1 Preventive Mode

In this case, the question is whether the system is able to withstand any possible
disturbance in its normal operation. If not controlled, prevention transfers this state
of the system to the safe operating area. Because instantaneous disturbances are
difficult to predict, preventive security assessments are primarily aimed at balancing
and reducing the likelihood of blackout and inconsistencies with lower operating
costs [20].
Some of the researches related to power system control that have been done using
machine learning methods in preventive operational mode include line trip fault
prediction [28] at device level, maintaining frequency stability [29], voltage insta-
bility prediction [30], predicting post-fault transient instability [31], and proactive
frequency control based on anticipation of sudden power fluctuation [32] at
subsystem level and predicting transient stability [33, 34] in smart grid and wide
area level, which these preventions avoid potential future hazards. For this purpose,
in Table 5.1, for preventive control state in power systems, several levels (device,
subsystem, microgrid, smart grid, wide area) are considered, and at each level, some
papers have examined various control issues, which researchers have used machine
learning methods to solve these problems.

Table 5.1 Summary of applications of machine learning methods in preventive control mode
Control
level Control problem Learning algorithm Refs
Device Define fault current direction Decision Trees (DT) and Support [35]
Vector Machine (SVM)
Line trip fault prediction Long Short-Term Memory [28]
(LSTM) Networks and Support
Vector Machine (SVM)
Subsystem Maintaining frequency stability Extreme Learning Machine (ELM) [29]
Voltage instability prediction Deep Recurrent Neural Network [30]
(DRNN)
Transient stability Classification and Regression Tree [36]
(CART) and Multilayer Perceptron
(MLP)
Predicting post-fault transient insta- Ensemble Online Sequential [31]
bility and develop emergency Learning Machine (E-OS-ELM)
generator-shedding control
Proactive frequency control based on Extreme Learning Machine (ELM) [32]
anticipation of sudden power
fluctuation
Smart grid Predicting transient stability Extreme Learning Machine (ELM) [33]
Wide area Determine the transient stability Radial Basis Function Neural Net- [34]
aspects work (RBFNN)
Real-time transient stability Least Square Support Vector [37]
assessment Machine (LS-SVM)
5 Introduction and Literature Review of the Application of Machine. . . 97

5.4.2 Normal Mode

Due to small and continuous changes in the power generation of generators and
loads in the power system, it is necessary to control the system in normal operating
state. In normal operating mode, the behavior of all system loads and the amount of
generated energy by the generators at all control levels produce a certain frequency
and voltage without interruption, and the system does not see any physical or
security faults. Also, all nominal limitations in each power device are met [20].
Some of the important control problems of different levels that are investigated
under normal conditions, in this part, include issues such as improvement in the
performance indices of relays [38] and controlling maximum power point tracking
(MPPT) for photovoltaic systems [3] at the device and local level; voltage stability
margin monitoring [39], automatic generation control (AGC) [40], damping
low-frequency oscillations by power system stabilizer (PSS) [41], power quality
disturbances control [42], and online rotor angle stability prediction [43] at the
subsystem level; load response [44], dynamic energy management [9], and droop
control, automatic generation control, and economic dispatch [45] at the microgrid
and smart grid level; and transient stability assessment [46] at the wide area level.
Table 5.2 summarizes some of the researches that examine various problems of
normal power system control. In this table, different machine learning methods have

Table 5.2 Summary of applications of machine learning methods in normal control mode
Control
level Control problem Learning algorithm Refs
Device Improving performance indices of protec- Extreme Learning Machine [38]
tion relays (ELM)
Maximum power point tracking (MPPT) Markov Decision Process [3]
(MDP)
Subsystem Voltage stability margin (VSM) Local Regression [39]
monitoring
Automatic generation control (AGC) Long Short-Term Memory [40]
Recurrent Neural Network
Damping low-frequency oscillations Artificial Neural Networks [41]
(ANN)
Power quality disturbances control Deep Neural Network [42]
(DNN)
Online rotor angle stability prediction Ensemble Decision Tree [43]
Microgrid/ Demand response Deep Neural Network [44]
smart grid (DNN)
Dynamic energy management Decision Tree (DT) [9]
Droop control, automatic generation con- Extreme Learning Machine [45]
trol, and economic dispatch (ELM)
Wide area Transient stability assessment Core Vector Machine [46]
(CVM)
98 S. Sadeghi et al.

been used to investigate various issues at the control levels of the device, subsystem,
microgrid, smart grid, and wide area.

5.4.3 Emergency Mode

The purpose of identifying an emergency state is to whether the system seeks to lose
integrity after the actual malfunction begins. There is an important interaction here
that response time is much more important, while economic considerations are
temporarily given secondary priority [20]. Emergency control is performed with
the aim of taking urgent measures to prevent partial or complete interruption of
services. In emergency control mode, some of the normal operating limits of the
system are violated. For example, lines are overloaded, and the frequency is lower
than the allowable value. Also, the frequency may exceed the allowable value due to
the departure of a large part of the load, or due to various faults or lack of precise
control of the reactive power, the voltage may be out of the allowable range [14].
Some important problems of emergency operating control mode examined in
some studies are summarized in Table 5.3. These issues have also been investigated
using machine learning methods at various levels. These problems include switch
PWM fault distinction in rectifier [47] and detecting fault by relays [48] in device
level; transient stability [49], online voltage stability monitoring [50], fault detection
[51], enhancing electromechanical oscillations damping [52], emergency load-
shedding control [53], and frequency control to prevent system collapse [54] in
subsystem level; short-term voltage instability [55], fault detection [56], islanding
detection approach [57], and mitigating cascading failures, preventing blackout, and
adaptive output power regulation of generators via frequency control [58] in smart
grid and microgrid level; and enhancing transient stability and damping the inter-
area oscillations [2], oscillatory angle stability [59], damping low-frequency oscil-
lation [60], and short-term voltage stability [61] in wide area level.

5.4.4 Restoration Mode

When both preventive and emergency control have failed to return system param-
eters to their inequality limits, local protection devices will operate to protect
components of the electrical system from irreparable damage. This may cause
further disturbances, system failures, and even partial or complete blackouts.
In restorative mode, some system loads may not be supplied so that partial or total
blackouts occurs in this mode, but the operational parts of the system are returned to
normal. Restoration at each control level and for each type of load will be different in
terms of speed and accuracy. Also, the restorative method in smart grids is different
from the current power grid. In this mode, the task of operator is minimizing the
5 Introduction and Literature Review of the Application of Machine. . . 99

Table 5.3 Summary of applications of machine learning methods in emergency control mode
Control
level Control problem Learning algorithm Refs
Device Open switch fault diagnosis PWM in Least Square Support Vector [47]
voltage source rectifier Machine (LSSVM)
Detecting fault by relays Support Vector Machines (SVM) [48]
Subsystem Transient stability High-Performance Back Propaga- [49]
tion Neural Network (HBPNN)
Online voltage stability monitoring Ensemble Ada Boost Classifier [50]
Fault detection Nearest Neighbor (KNN), Decision [51]
Trees, and Support Vector
Machines (SVM)
Enhancing electromechanical oscilla- Tree-Based Batch Mode Rein- [52]
tions damping forcement Learning (RL)
Emergency load-shedding control Artificial Neural Network and [53]
Analytic Hierarchy Process
Algorithm
Frequency control to prevent system Multi Q-Learning—Reinforcement [54]
collapse Learning
Microgrid/ Short-Term Voltage instability Random-Weights Neural Networks [55]
smart grid Mode detection and fault detection Artificial Neural Network (ANN), [56]
Support Vector Machine (SVM),
and Decision Tree
Islanding detection approach Support Vector Machine (SVM) [57]
Mitigating cascading failures, Artificial Neural Networks (ANN) [58]
preventing blackout, adaptive adjust-
ment of generators’ output power
through frequency control
Wide area Enhancing transient stability and Reinforcement Learning (RL), [2]
damping the inter-area oscillations Neural Network (NN)
Oscillatory angle stability Actor-Critic Neural Network [59]
Damping low-frequency oscillation Artificial Neural Network (ANN) [60]
Short-term voltage stability Random Forest [61]

value of unreleased power via recoordinating the lost generation in the shortest
feasible time and adding the interrupted load in order of primacy.
In Table 5.4, for restorative control mode in power systems, few numbers of
researches are done with machine learning methods. In this state, several levels of
subsystem, microgrid, and wide area are considered, and at each level, some papers
have examined various control issues. The control problems investigated in this
mode include generation rescheduling and load shedding [62], fault-induced delayed
voltage recovery [63], and restoration of power grid systems [64] in subsystem level,
post-fault restoration [65] in microgrid level, and real-time short-term voltage
stability assessment [66] in wide area level.
In some studies, several control modes are examined simultaneously. For exam-
ple, in [67], two modes of emergency and restoration control are performed using the
reinforcement learning method (multi-agent Q-learning) in power distribution
100 S. Sadeghi et al.

Table 5.4 Summary of applications of machine learning methods in restoration control mode
Control
level Control problem Learning algorithm Refs
Subsystem Generation rescheduling and load Decision Tree (DT) [62]
shedding
Fault-induced delayed voltage Weighted Kernel Extreme Learning [63]
recovery (FIDVR) Machine (WKELM)
Restoration of power grid systems Q-Learning [64]
Microgrid Post-fault restoration Feature Selection [65]
Wide area Real-time short-term voltage stabil- Extreme Learning Machine (ELM) [66]
ity (STVS) assessment

systems. So the fault location detection is done in an emergency state, and then the
isolation and restoration are done.
According to the tables, it can be seen that in most research, machine learning
methods may be helpful in restorative state but have focused on preventative,
normal, and emergency situations. In fact, machine learning and deep learning
methods are focused on predictive modes, and these methods are more effective
for preventive and emergency modes than other modes.

5.5 Application of Machine Learning Methods

in Evaluating the Security and Stability of the Power
System

Since various social, economic, and political activities are linked to the national
power system, the safe and sustainable operation of the power system are also
determined by governments and public industry stakeholders and have the highest
priority. Sometimes, enemies or rivals can access network information and change
control commands, thus destabilizing network performance and causing blackouts
and financial losses, and national security can be compromised [15]. In addition, the
increasing growth of energy demand, the presence of annoying loads,
rearrangements and changes in topology and architecture of network, overloading
of transmission lines, etc. have led to the power system deviating from its safe range
so they cause instability and disruption in power quality [68].
Since the purpose of controlling the power system is to create a stable and secure
system in all control modes, this section examines in detail the stability and security
of the power system. The main areas of power system security and stability include
transient stability, voltage stability, frequency stability, and power quality distur-
bances, in which machine learning are widely developed [16]. Evaluation of
machine learning methods in various areas of power system sustainability concen-
trates on highlighting methods, attainments, and restrictions in the scheme of
classifiers, the production of data sets, and the test systems that are used. Also,
5 Introduction and Literature Review of the Application of Machine. . . 101

machine learning methods are widely used to monitor, detect intrusion, predict, and
classify various threats of the power system.
Ensuring the security and stability of the power grid is a major challenge today,
especially after being exposed to various pressures and disturbances. Therefore,
important issues that play an important role in assessing the stable and safe operation
of the power system are as follows:
• Transient stability assessment.
• Voltage stability assessment.
• Power quality disturbances.
• Frequency stability assessment.
The three main categories of transient stability, frequency stability, and voltage
stability are recognized as important issues in dynamical security assessment [69].

5.5.1 Transient Stability Assessment

The capability of power system generators to keep coordination after a large

perturbation such as a severe external fault, sudden and immediate loss of loads,
or generators of the power system is called transient stability. Transient instability is
one of the main reasons of power instability, which will lead to island construction
and widespread blackouts [70]. Conventional methods for evaluating transient
stability in a power system, including time domain simulation, equal areas criterion,
transient energy function, and Lyapunov, do not meet its new and modern require-
ments because evaluating the swing curves of generators, different loads, faults, and
fault clearance times requires extensive computational effort [71].
Monitoring and evaluating the stability of the power system is very important for
its efficient operation in the margins of its stability. If there is a huge amount of high-
speed data collected by PMU, using conventional methods of transient stability
assessment alone may not meet the needs of transient stability in real time and will
be computationally time-consuming. Thus, machine learning techniques have
recently been widely used in power system control issues, including transient
stability studies. These methods process the great values of PMU data, then analyze
data, and finally classify and detect the stability state of the power system. In
machine learning methods, due to the generalizability, explicitly trained data-driven
models can accurately predict the stability [46]. Transient stability is assessed using
machine learning methods in three stages: feature generation, feature preprocessing
and optimization, and classification/prediction.
Feature Generation: The first and most important step in organizing the analysis
of reliable transient stability assessment models is the generation of input data set
vectors. Synchronous variables of the power system sampled by PMUs based on the
wide area measurement system (WAMS) that were collected before or immediately
after clearing a fault provide the possibility of performing advanced protect opera-
tions a wide area for control and decision [10].
102 S. Sadeghi et al.

Choosing the proper features is an important criterion for the transient stability
assessment. Therefore, generating/extracting feature data through time domain sim-
ulation processes is an important issue [72]. The major concerns in transient stability
assessment is the selection of appropriate trajectories characteristics [73]. To predict
the stability or instability of the system, various procedures such as rotor angles,
rotor speed, and voltage amplitude are applied as forecasters [74].
Feature preprocessing and Optimization: Various techniques for reducing,
selecting, and optimizing attributes appropriate to the type of each data set and
each classification algorithms are proposed to eliminate redundancies, improve
classification, and investigate transient instability [75].
Classification/Prediction: Typically, “offline training, online application”
methods are used in many transient stability assessment studies based on machine
learning methods. This method is that first the model is taught offline and then the
transient stability test is done online [72]. In the study of stability with offline
training, for example, generators and different load changes are usually modeled
with different possible disturbances, including three-phase faults to ground, etc. with
different fault clearance times. But despite the innumerable successes of using the
“offline training, online application” model, this model is not applicable in the real
world because the set of training data generated by the offline method cannot exhibit
all the features and characteristics of a modern power system [10].
In the study of transient stability, the results of neural network classifiers can be
continuous so that the transient stability margins and boundaries appear softer [36];
decision trees are well interpretable [76]; support vector machine can calculate the
distance between a moment and a stable boundary, in which case it is mostly used to
define a certainty index [68]; Random Forest algorithms can regularly evaluate
feature weights and arrange features according to sorting rate [77]; and Extreme
Learning Machine has a fast learning speed [78].
Table 5.5 summarizes the machine learning algorithms and preprocessing/opti-
mization methods adopted in each learning method for some recent research works
in the field of transient stability. Different types of power systems have been used in
various studies to investigate transient stability. According to Table 5.5, the 39-bus
system is used in most articles.
Comparison of machine learning methods in the table is done based on the
obtained accuracy criterion, which according to the references, this value is usually
between 85% and 100%, and the accuracy criterion with higher percentage shows
the superiority of the method.

5.5.2 Voltage Stability Assessment

Transient stability is related to generator synchronization, but voltage stability is

related to load dynamics and reactive power management. Voltage stability indicates
the capability of the power system to retain the buses voltage at acceptable values
after perturbation in a particular operating condition [79]. Reactive power
5 Introduction and Literature Review of the Application of Machine. . . 103

Table 5.5 A summary of some of the offered machine learning methods to classify transient
stability assessment
Refs Machine learning algorithm Preprocessing/optimization technique
[10] Ensemble of Online Sequential-Extreme Learn- Binary Jaya (Bin Jaya)
ing Machine (EOS-ELM)
[36] Classification and Regression Tree (CART) and –
Artificial Neural Network (ANN)
[46] Core Vector Machine (CVM) –
[68] Ensemble of SVM Min-max normalization
[72] Bayesian Multiple Kernels Learning (BMKL) –
[73] Twin Convolutional SVM (TWCSVM) –
[74] Aggressive SVM (ASVM) and Conservative –
SVM (CSVM)
[75] Extreme Learning Machine (ELM) PSO
[76] Decision Trees (DT) –
[77] Random Forest Recursive Feature Elimination
Strategy
[78] ELM Kernelized Fuzzy Rough Sets
(KFRS) and Memetic Algorithm

management and consequently monitoring and evaluating the voltage stability

situation are important issues to prevent various blackouts and maintain the overall
stability of the modern power system. When a significant fault occurs in the power
system, the voltage profile loses stability, the speed of the induction motors
decreases rapidly, and they receive a high reactive current, which ultimately leads
to an uncontrollable voltage drop and blackouts. Voltage instability is divided into
short term and long term [80].
Most conventional methods do not prepare accurate useful information about
stability issues although they require relatively complex calculations, but they ignore
the dynamics of the modern power system. Also, calculating a large number of loads
bus currents has many limitations in conventional models of assessing P-V and Q-V
curves of voltage stability analysis. Therefore, machine learning methods have
recently been used to overcome shortcomings and violations of conventional voltage
stability assessment methods [81]. The study of machine learning methods in
assessing voltage stability is done in two stages, including the stages of generation
and selection of features and the classification of voltage stability margin index.
Generation and Selection of Features: If the power system is able to keep
admissible stable voltages at all system bases under operating conditions of normal
and after a fault has occurred, it will be classified in the voltage stability zone. In this
regard, to evaluate the voltage stability using machine learning methods, the data of
the Phasor Measurement Units is used that ordinary input vectors are considered as
the voltage phasor. Here, as in the case of transient stability, the input vectors for
classifier algorithms are usually educated via “offline training,” and the generated
output vector is considered as the index vector for voltage stability margin [82].
104 S. Sadeghi et al.

Table 5.6 A summary of some of the offered machine learning methods to classify voltage
stability assessment
Refs Machine learning algorithm Preprocessing/optimization technique
[27] Decision Tree (DT) Principal Component Analysis (PCA) and Cor-
relation Techniques
[55] Neural Network with Random Weight Relief Algorithm
(NNRW)
[80] Feed Forward Back Propagation –
(FFBP)
[81] SVM k-Means
[82] Artificial Neural Network (ANN) Imperialist Competitive Algorithm (ICA)
[83] SVM Genetic Algorithm (GA)
[84] ANN Gram-Schmidt Orthogonalization (GSO)
[85] Probabilistic Fuzzy Decision Tree Case-Based Reasoning (CBR)
(PFDT)
[86] SVM Multi-Objective Biogeography-Based Optimiza-
tion (MOBBO)
[87] Feed Forward Back Propagation Net- Linear Optimization
work (FFBPN)

Voltage stability evaluation models are very complex and nonlinear and involve a
large amount of data sets, so the selection and reduction of features in these models is
a very important issue. In this case, many features of the power system are not
suitable, and it is better not to use them directly as classifier inputs. Therefore, in
most models, feature extraction and reduction methods are used [27]. In addition,
different optimization methods have been used to increase exactitude, improve the
parameters adjustment of machine learning algorithms, and decrease the data train-
ing time [83].
Classification of Voltage Stability Margin Index: Using ML algorithms, the
nonlinear relationship between input vector and output vector can be extracted and
adjusted, in which the input vector is the operating parameters related to the power
system, and the output vector is the voltage stability margin. In the problem of
voltage stability assessment, some of the features of machine learning algorithms are
as follows:
The calculation time of the Artificial Neural Network algorithm is very short, and
it predicts the stability margin index very accurately, but this algorithm suffers from
overtraining and in which setting parameters can be a concern [84]. DT has simple
division rules due to fewer data samples and has excellent performance in online
voltage rating classification [27].
Table 5.6 summarizes the machine learning methods for predicting, monitoring,
and analyzing voltage stability assessments, in which some machine learning algo-
rithms and data preprocessing/optimization techniques are adopted. Also, in this
case, comparison of machine learning methods is done based on the obtained
accuracy criterion, and the accuracy criterion with higher percentage shows the
superiority of the method.
5 Introduction and Literature Review of the Application of Machine. . . 105

5.5.3 Power Quality Disturbances Assessment

Sudden deviation of the voltage amplitude, frequency, and phase angle from the
defined standard rate is called power quality disturbances. These disturbances are
often caused by the increased use of switching devices, nonlinear loads, inverters,
and rectifiers in the power system. Types of power quality disturbances include
harmonic distortion, voltage sag, flicker, interruption, swell, transient, etc. [88]. In
each power system, there are different types of disturbances in power quality, so
accurate diagnosis and classification of events in the study of disturbances related to
power quality are done in three stages of feature extraction/selection, feature opti-
mization, and classification of events [89].
Feature Selection: In the first step of feature selection, the volume of data must
be considered because generated feature data with large dimension is not ideal for
use as a classifier input. Duplicate data with unrelated features significantly increases
the time of calculations and decreases the accuracy of classification. Therefore,
extraction of dominant features in the subject of machine learning is usually neces-
sary for classification [90]. In the study of power quality disturbances using machine
learning methods, frequency domain techniques such as Fourier Transform (FT) [91]
and various time domain signal processing methods such as Empirical Mode
Decomposition (EMD) [92] have been used successfully to select features of
waveforms. In many studies, the Wavelet Transform (WT) method [88] is used
because they provide good time-frequency properties and have a very good ability to
analyze local discontinuities of signals.
Feature Optimization: Various feature optimization techniques are used to
improve the performance of power quality disturbances events classification. The
main purpose of using these methods is to remove additional attributes such as noise
in the adopted features and optimize the accuracy of the classifiers. Most attribute
selection methods for events can be based on wrapper or based on filter [89]. Filter-
based feature selection is fast because it ranks attributes according to their inherent
properties, while feature selection based on the wrapper needs more time, but it is a
more efficient option. If no feature optimization method is used, classification may
need complex and time-consuming simulations, voluminous calculations sources,
and more time [89].
Classification of Events: Countless machine learning and deep learning algo-
rithms have been used in the power quality disturbances classification. Among the
various types of classifiers, SVM [89] and PNN [88] are the most widely used due to
some of the salient features. The PNN technique has a good performance for
classification and has high accuracy in signal distance studies. Also, PNN technique
does not require initial weight adjustment. The SVM technique is also ranked as a
powerful classifier.
Table 5.7 shows some of the machine learning techniques along with the method
of extracting and optimizing features in some articles in the field of power quality
disturbances. As shown in the table, the most used machine learning tool in power
quality disturbances assessment and classification is the SVM technique because it
106 S. Sadeghi et al.

Table 5.7 A summary of some of the offered machine learning methods to classify power quality
disturbances events
Machine learning Feature extraction Feature optimization/reduction
Refs algorithm technique technique
[42] Deep Neural Net- Singular Spectrum Analy- Compressive Sensing (CS)
work (DNN) sis (SSA) and Wavelet
Transform (WT)
[88] Probabilistic Neu- Wavelet Transform (WT) Artificial Bee Colony (ABC)
ral Network (PNN)
[89] Support Vector Variational Mode Decom- Sequential Forward Selection (SFS),
Machine (SVM) position (VMD) and Sequential Backward Selection
S-Transform (ST) (SBS), and Gram-Schmidt Orthogo-
nalization (GSO)
[90] SVM Wavelet Packet Transform Genetic Algorithm (GA)
(WPT)
[93] SVM WT Particle Swarm Optimization (PSO)
[94] k-Means, Decision ST Ant Colony Optimization (ACO)
Tree (DT), and
SVM
[95] SVM WT –
[96] Radial Basis Func- Empirical Mode Decom- PSO
tion Neural Net- position (EMD) and Hil-
work (RBFNN) bert Transform (HT)
[97] SVM Wavelet Multiresolution –
Analysis (WMRA)
[98] Convolutional Curvelet Transform (CT) Static Single Assignment
Neural Network
(CNN)

ensures high accuracy and efficiency. Here, too, the percentage of accuracy obtained
is considered as a criterion for evaluating the superiority of one method over other
methods. So the closer the accuracy to 100%, the superiority of the method.

5.5.4 Frequency Stability Assessment

The most important task of a power grid is to generate and transmit electricity to
consumers while maintaining the quality standards of electrical power. One of the
most important of these criteria is to maintain the network frequency close to the
nominal value, in which, in order to maintain the frequency, a balance of production
and consumption must be established at all times. Otherwise, the network frequency
will change and may become unstable and cause severe damage to network equip-
ment such as turbines, motors, transformers, etc. Therefore, special control and
protection operations are used to maintain the balance between production and
consumption and to maintain the network frequency close to the nominal value.
The primary frequency control is the first control operation used for this purpose, and
5 Introduction and Literature Review of the Application of Machine. . . 107

the secondary and tertiary controls are the backup controls. Primary and secondary
frequency control characteristics, including dynamic behavior and their require-
ments, depend on various parameters (inertia constant, the degree of load depen-
dence on frequency, and the speed and number of units participating in frequency
control and their available capacity), which are basically the inherent parameters of
the system and are different for various networks. These parameters are different for
various networks and play a very important role in the operation and control of
network frequency and the allocation of primary and secondary storage. Also, by
using spinning reserve with the help of primary frequency control, frequency drop to
frequency loading thresholds can be prevented to protect against frequency instabil-
ity [99]. If, due to major disturbances, such as power plant outages, power imbal-
ances persist and the system frequency deviates significantly from the nominal
value, and underfrequency load shedding steps begin, in this regard, in order to
increase the accuracy of frequency security assessment of systems that are under
disturbances, new methods are needed [53].
Frequency stability assessment is based on the degree of frequency deviation and
the rate of change of frequency from the nominal value. Then according to the
amount of changes, the necessary measures are taken for stability, and even correc-
tive measures may be taken after instability. In frequency security assessment using
machine learning methods, the model is trained offline and used for online
applications [100].
Frequency assessment in three issues of automatic generation control (AGC)
[32, 40, 101], frequency load control (LFC) [99, 102, 103], and underfrequency
load shedding (UFLS) [29, 53, 54, 100] is done by using machine learning methods
in two stages of generation and optimization of features and classification based on
frequency deviation.
Generation and Optimization of Features: The initial data set for determining
the frequency stability margin includes the frequency value and the frequency
change rate. A variety of optimization methods are used to reduce redundant features
and initial data. The initial training of the data is in the form of measuring the
distance of the samples in order to identify the features and group the samples in
different categories [100].
Classification Based on Frequency Stability Margin: Compared to traditional
and common techniques, machine learning methods have the ability to make quick
decisions in choosing the type of strategy, reducing the time of decision and
restoration, and improving frequency stability. The relationship between the fre-
quency deviation from the nominal value and other attributes is used to determine
stability [100]. Some methods, such as Artificial Neural Network and extreme
learning machine, have high speed in teaching and learning process, but extreme
learning machine algorithm is faster [53]. ELM also has fewer learning limitations
and does not fall into local optimizations [29]. Frequency assessment in multi-area
and wide area systems is an important criterion in determining system stability [102].
Table 5.8 illustrates some of the machine learning methods along with the
techniques of preprocessing and optimizing features in some articles in the field of
frequency stability. As shown, the most used machine learning tool in frequency
108 S. Sadeghi et al.

Table 5.8 A summary of some of the offered machine learning methods to frequency assessment
Preprocessing/optimization
Refs Machine learning algorithm technique
[29] Extreme Learning Machine (ELM) –
[32] ELM –
[40] Long Short-Term emory Recurrent Neural Network Backward Propagation with the
Gradient Optimizer
[53] Artificial Neural Network (ANN) and Analytic Hier- k-Means
archy Process (AHP)
[54] Deep Q Network (Multi-Q-Learning) and Deep Deep Deterministic Policy Gra-
Reinforcement Learning (DRL) dient (DDPG)
[99] Multi-Agent Reinforcement Learning (MARL) Genetic Algorithm (GA)
[100] Support Vector Regression (SVR) and Core Vector Mini-Batch Gradient Descent
Regression (CVR) (MBGD) and Fuzzy k-Means
[101] Artificial Emotional Reinforcement Learning (ERL) –
[102] Integral Reinforcement Learning –
[103] Multi-Agent Deep Reinforcement Learning Deep Deterministic Policy Gra-
(MA-DRL) dient (DDPG)
[104] ELM –
[105] Principal Component Analysis (PCA), Regression –
Trees, and Support Vector Machine (SVM)

stability assessment and classification is the reinforcement learning technique. Since

reinforcement methods provide the learning with feedback from the environment,
they have a special application for online training.

5.6 Challenges, Comparative Discussion, and Future

Perspectives

Today, practical complexities in the power system have increased, so control and
evaluation of stability have become the most important concern in the power system.
In this regard, data-based methods should replace model-based methods because
data-based methods have online decision-making capabilities, but model-based
methods have the contradiction between efficiency and accuracy.
Much progress has been made in using machine learning and deep learning
methods related to power system control and stability studies, but some challenges
remain unresolved. The quantity and quality of input data sets and experimental
systems used in each research will affect the prediction and accuracy of machine
learning methods.
Researchers obtain the data, which they need for their research, from open source
data and simulated data sets because real power system data may be unavailable for
reasons such as security issues, attacks, overcrowding, and breaches, or the number
of available data is very small. Therefore, this inconsistency between the available
5 Introduction and Literature Review of the Application of Machine. . . 109

data and the actual data of the power system will cause inconsistencies in forecasts
and classifications. In addition, fine-tuning the parameters and measurements of each
control problem is an important issue in the application of the machine learning
algorithm, and it can be a time-consuming problem. Also, the use of “offline
training, online application” method has become a bit challenging due to inconsis-
tency and possible imbalance crisis of data, measurements, and status of power
system variables. However, the robust status of the power system is a point of hope
for the use of machine learning methods in preventive control measures, protection
of the power system in an emergency situation, and its restoration after a fault and
blackout. It should be noted that the “offline training, online application” method is
more challenging in examining the state of transient stability than other areas
because its evaluation is related only to static data after the fault, and also, simulated
training samples produced offline mode may not display the current or future state of
the power system correctly.
To control the power system and related issues, the following issues are observed:
Most control considerations relate to control issues in normal and emergency
situations, while relatively few considerations related to preventive and restorative
controls. However, studies on control issues in the operating mode of preventive
using machine learning techniques have been surprisingly abundantly presented.
The reason for this is to anticipate possible security and stability disruptions and to
prevent and plan as much as possible to take the best action in the event of a
disruption.
Different control issues have been discussed at different levels of the wide area,
smart grids, microgrids, subsystems, and devices. It should be noted that the move
toward smart power grids has highlighted the efficiency of machine learning
methods.
The most important security and stability problems that can be examined with the
largest volume of studies and researches include transient stability assessment after
perturbations, voltage stability assessment, and power quality disturbances assess-
ment. It is worth mentioning that power quality assessment includes frequency,
voltage, and angle stability.
The predominant machine learning methods used to solve the control problems in
this chapter include Support Vector Machine (SVM), k-Nearest Neighbor (k-NN),
Decision Tree (DT), Random Forest (RF), Neural Network, and k-Means, which
sometimes adopt the same important feature selection techniques such as Factor
Analysis, Forward Feature Selection (FFS), Backward Feature Selection (BFS), and
Principal Component Analysis (PCA).
A very important area that has received a lot of attention in the last decade, and in
which issues related to security and stability of control issues are discussed, is the
study and prediction of various types of attacks and injuries in control of SCADA
power network infrastructure. The SCADA system enables the automatic coordina-
tion of control functions such as monitoring the security of the power system,
economic operations of load dispatch among power plants, short-term forecasts,
automatic production control, emergency control, and etc. across the power system
in a coordinated manner. Control systems include a variety of electronic and
110 S. Sadeghi et al.

processing equipment, the most important of which are actuators, sensors, Program-
mable Logic Controllers (PLCs), and Remote Terminal Units (RTUs). By accessing
the software or hardware code of this equipment, attackers can easily infiltrate the
control system and take control of the process. Also, the widespread presence of
cyber in the form of advanced communication tools and the Internet of Things in the
SCADA network has made the power system vulnerable to attacks and security
threats. So safety solutions and measures such as antivirus programs, firewalls,
encryption algorithms, authentication, traditional intrusion detection systems, etc.
are unable to deal with cyberattacks [11].
Recently, cyberattacks on the SCADA network have increased. Therefore, intru-
sion detection systems are installed for early detection of attacks. In this regard, the
ability of machine learning algorithms in autonomous learning, adaptation to change
and performance without any prior planning, has led to their use as valid methods for
intelligent and efficient intrusion detection systems [15]. Techniques of machine
learning algorithms for investigating power grid attacks in SCADA include three
main steps including data set generation, data processing, and classification/detec-
tion. SCADA network data is recorded and analyzed using machine learning
methods to create training data sets to reduce attacks. SCADA data is not available
for real-time security reasons. Therefore, predicting the occurrence of instabilities is
one of the important issues in controlling the power system. In order to prevent
cyberattacks and injuries on SCADA, machine learning methods are used to identify
the types of attacks.

5.7 Conclusions

Control of the system in different operating modes in order to provide electrical

energy with high reliability for consumers and create a secure and stable situation in
the event of any fault in the power system has been considered by all energy
stakeholders, especially operators. However, the move toward smart power system
and increasing use of distributed generation (DG) and renewable energy sources and
converter-based resources, despite their many benefits, have increased instability at
the power system level, in such a way that it has the greatest effect on its frequency
and changes.
Increasing complexity in large interconnected power systems requires advanced
control techniques that effectively control power systems. Also, the hardware cannot
evaluate in real time on large networks, and the use of numerical methods and
modeling of the power system is complex and time-consuming. Thus, researchers
have turned to use machine learning methods. In machine learning, a model is first
trained, and then it is used for predicting the class of newly arrived data. Till now,
several machine learning techniques have been used to evaluate the power system
control. Embedding the machine learning methods in control schemes of power
systems is an important and effective method to create controllers with the capability
to learn and update decision-making features.
5 Introduction and Literature Review of the Application of Machine. . . 111

Machine learning and deep learning methods provide suitable solutions to solve
each of the control problems at each control level of the operating modes of the
electrical power system. Considerations of control levels vary from local to wide
area, and to solve each problem using the machine learning methods in this chapter,
these results are obtained, that there is no method that is guaranteed for all or even
most control problems. However, there are enough ways to evaluate a challenging
issue that are reasonably successful. A comparison method is also used to select the
best learning method so that several machine learning methods are used to solve a
problem, and finally, the method that can offer the most accurate, the best classifi-
cation with a sufficient number of features, the least complexity, and the shortest
simulation time is selected as the method with the best performance in solving the
desired problem. In addition, selecting the best method for extracting, reducing, and
optimizing features is also a consideration in machine learning and deep learning
algorithms.

References

1. DOE-USA, Chapter 5: Increasing efficiency of buildings systems and technologies, in The

Quadrennial Technology Review: An Assessment of Energy Technologies and Research
Opportunities (2015)
2. R. Yousefian, R. Bhattarai, S. Kamalasadan, Transient stability enhancement of power grid
with integrated wide area control of wind farms and synchronous generators. IEEE Trans.
Power Syst. (2017). https://doi.org/10.1109/TPWRS.2017.2676138
3. P. Kofinas, S. Doltsinis, A.I. Dounis, G.A. Vouros, A reinforcement learning approach for
MPPT control method of photovoltaic sources. Renew. Energy 108, 461–473 (2017). https://
doi.org/10.1016/j.renene.2017.03.008
4. A. Rosato, M. Panella, R. Araneo, A. Andreotti, A Neural Network Based Prediction System
of Distributed Generation for the Management of Microgrids. IEEE Trans. Ind. Appl. 55,
922 (2019). https://doi.org/10.1109/TIA.2019.2916758
5. X. Xu, Y. Xu, M.-H. Wang, J. Li, Z. Xu, S. Chai, Y. He, Data-driven game-based pricing for
sharing rooftop photovoltaic generation and energy storage in the residential building cluster
under uncertainties. IEEE Trans. Ind. Informat. (2020). https://doi.org/10.1109/tii.2020.
3016336
6. H. Jahangir, S.S. Gougheri, B. Vatandoust, M.A. Golkar, A. Ahmadian, A. Hajizadeh, Plug-in
electric vehicle behavior modeling in energy market: a novel deep learning-based approach
with clustering technique. IEEE Trans. Smart Grid (2020). https://doi.org/10.1109/tsg.2020.
2998072
7. R. Badar, M.Z. Khan, M.A. Javed, MIMO adaptive bspline-based wavelet neurofuzzy control
for multi-type facts. IEEE Access 8 (2020). https://doi.org/10.1109/ACCESS.2020.2969387
8. S. Lan, M.J. Chen, D.Y. Chen, A novel HVDC double-terminal non-synchronous fault
location method based on convolutional neural network. IEEE Trans. Power Deliv. 34
(2019). https://doi.org/10.1109/TPWRD.2019.2901594
9. G.K. Venayagamoorthy, R.K. Sharma, P.K. Gautam, A. Ahmadi, Dynamic energy manage-
ment system for a smart microgrid. IEEE Trans. Neural Netw. Learn. Syst. (2016). https://doi.
org/10.1109/TNNLS.2016.2514358
10. Y. Li, Z. Yang, Application of EOS-ELM with binary Jaya-based feature selection to real-time
transient stability assessment using PMU data. IEEE Access 5 (2017). https://doi.org/10.1109/
ACCESS.2017.2765626
112 S. Sadeghi et al.

11. J. Gao, L. Gan, F. Buschendorf, L. Zhang, H. Liu, P. Li, X. Dong, T. Lu, Omni SCADA
intrusion detection using deep learning algorithms. IEEE Internet Things J. (2020). https://doi.
org/10.1109/jiot.2020.3009180
12. E. Hossain, I. Khan, F. Un-Noor, S.S. Sikander, M.S.H. Sunny, Application of big data and
machine learning in smart grid, and associated security concerns: a review. IEEE Access
7 (2019)
13. O.A. Alimi, K. Ouahada, A.M. Abu-Mahfouz, A review of machine learning approaches to
power system security and stability. IEEE Access 8 (2020)
14. L.A. Wehenkel, Automatic learning techniques in power systems (1998)
15. M. Kalech, Cyber-attack detection in SCADA systems using temporal pattern recognition
techniques. Comput. Secur. 84 (2019). https://doi.org/10.1016/j.cose.2019.03.007
16. P. Kundur, Power System Stability and Control (McGraw-Hill, New York, 1993)
17. A. Dissanayaka, U.D. Annakkage, B. Jayasekara, B. Bagen, Risk-based dynamic security
assessment. IEEE Trans. Power Syst. 26 (2011). https://doi.org/10.1109/TPWRS.2010.
2089809
18. L. Wang, P. Pourbeik, Assessment of power system stability and dynamic security perfor-
mance, in Power System Stability and Control, 3rd edn. (CRC Press, 2017)
19. P. Kundur, J. Paserba, V. Ajjarapu, G. Andersson, A. Bose, C. Canizares, N. Hatziargyriou,
D. Hill, A. Stankovic, C. Taylor, T. Van Cursem, V. Vittal, Definition and classification of
power system stability. IEEE Trans. Power Syst. 19 (2004). https://doi.org/10.1109/TPWRS.
2004.825981
20. M. Glavic, (Deep) reinforcement learning for electric power system control and related
problems: a short review and perspectives. Annu. Rev. Control 48, 22–35 (2019)
21. X. Wang, X. Luo, M. Zhang, X. Guan, Distributed detection and isolation of false data
injection attacks in smart grids via nonlinear unknown input observers. Int. J. Electr. Power
Energy Syst. 110 (2019). https://doi.org/10.1016/j.ijepes.2019.03.008
22. C.C. Aggarwal, Neural networks and deep learning (2018)
23. E. Alpaydin, Introduction to Machine Learning, 3rd edn. (2014)
24. Miraftabzadeh SM, Foiadelli F, Longo M, Pasetti M (2019) A survey of machine learning
applications for power system analytics. In: Proceedings - 2019 IEEE International Conference
on Environment and Electrical Engineering and 2019 IEEE Industrial and Commercial Power
Systems Europe, EEEIC/I and CPS Europe 2019
25. P. Mehta, M. Bukov, C.-H. Wang, A.G.R. Day, C. Richardson, C.K. Fisher, D.J. Schwab
Review machine learning. arXiv:180308823 [cond-mat, physics:Physics, stat]. (2018). https://
doi.org/arXiv:1803.08823v1
26. M. Glavic, R. Fonteneau, D. Ernst Reinforcement learning for electric power system decision
and control: past considerations and perspectives. IFAC-Papers OnLine. (2017) https://doi.
org/10.1016/j.ifacol.2017.08.1217
27. H. Mohammadi, M. Dehghani, PMU based voltage security assessment of power systems
exploiting principal component analysis and decision trees. Int. J. Electr. Power Energy Syst.
64 (2015). https://doi.org/10.1016/j.ijepes.2014.07.077
28. S. Zhang, Y. Wang, M. Liu, Z. Bao, Data-based line trip fault prediction in power systems
using LSTM networks and SVM. IEEE Access 6 (2017). https://doi.org/10.1109/ACCESS.
2017.2785763
29. Y. Dai, Y. Xu, Z.Y. Dong, K.P. Wong, L. Zhuang, Real-time prediction of event-driven load
shedding for frequency stability enhancement of power systems. IET Gener. Transm. Distrib.
6 (2012). https://doi.org/10.1049/iet-gtd.2011.0810
30. H. Hagmar, L. Tong, R. Eriksson, L.A. Tuan, Voltage instability prediction using a deep
recurrent neural network. (2019). https://doi.org/10.1109/TPWRS.2020.3008801
31. H. Yang, W. Zhang, F. Shi, J. Xie, W. Ju, PMU-based model-free method for transient
instability prediction and emergency generator-shedding control. Int. J. Electr. Power Energy
Syst. 105 (2019). https://doi.org/10.1016/j.ijepes.2018.08.031
5 Introduction and Literature Review of the Application of Machine. . . 113

32. S. Wen, Y. Wang, Y. Tang, Y. Xu, P. Li, Proactive frequency control based on ultra-short-term
power fluctuation forecasting for high renewables penetrated power systems. IET Renew.
Power Gener. 13 (2019). https://doi.org/10.1049/iet-rpg.2019.0234
33. Y. Xu, Z.Y. Dong, K. Meng, R. Zhang, K.P. Wong, Real-time transient stability assessment
model using extreme learning machine. IET Gener. Transm. Distrib. 5 (2011). https://doi.org/
10.1049/iet-gtd.2010.0355
34. B.P. Soni, A. Saxena, V. Gupta, S.L. Surana, Identification of generator criticality and
transient instability by supervising real-time rotor angle trajectories employing RBFNN. ISA
Trans. (2018). https://doi.org/10.1016/j.isatra.2018.08.008
35. J. Morales, E. Orduña, H. Villarroel, J.C. Quispe, High-speed directional protection without
voltage sensors for distribution feeders with distributed generation integration based on the
correlation of signals and machine learning. Electr. Power Syst. Res. 184 (2020). https://doi.
org/10.1016/j.epsr.2020.106295
36. Y.J. Lin, Comparison of CART- and MLP-based power system transient stability preventive
control. Int. J. Electr. Power Energy Syst. 45 (2013). https://doi.org/10.1016/j.ijepes.2012.08.
066
37. B.P. Soni, A. Saxena, V. Gupta, S.L. Surana, Transient stability-oriented assessment and
application of preventive control action for power system. J. Eng. 2019 (2019). https://doi.org/
10.1049/joe.2018.9353
38. R. Dubey, S.R. Samantaray, B.K. Panigrahi, An extreme learning machine based fast and
accurate adaptive distance relaying scheme. Int. J. Electr. Power Energy Syst. (2015). https://
doi.org/10.1016/j.ijepes.2015.06.024
39. S. Li, V. Ajjarapu, M. Djukanovic, Adaptive online monitoring of voltage stability margin via
local regression. IEEE Trans. Power Syst. 33 (2017). https://doi.org/10.1109/tpwrs.2017.
2698065
40. S. Wen, T. Zhao, Y. Wang, Y. Tang, Y. Xu, P. Li, A deep learning method for power
fluctuation identification from frequency fluctuations, in IEEE Power and Energy Society
General Meeting, (2019)
41. D.K. Chaturvedi, O.P. Malik, Generalized neuron-based adaptive PSS for multimachine
environment. IEEE Trans. Power Syst. 20 (2005). https://doi.org/10.1109/TPWRS.2004.
840410
42. H. Liu, F. Hussain, Y. Shen, R. Morales-Menendez, M. Abubakar, S. Junaid Yawar,
H.J. Arain, Signal processing and deep learning techniques for power quality events monitor-
ing and classification. Electr. Power Components Syst. (2019). https://doi.org/10.1080/
15325008.2019.1666178
43. Y. Chen, M. Mazhari, C.Y. Chung, S.O. Faried, B.C. Pal, Rotor angle stability prediction of
power systems with high wind power penetration using a stability index vector. IEEE Trans.
Power Syst. (2020). https://doi.org/10.1109/tpwrs.2020.2989725
44. R. Lu, S.H. Hong, Incentive-based demand response for smart grid with reinforcement
learning and deep neural network. Appl. Energy (2019). https://doi.org/10.1016/j.apenergy.
2018.12.061
45. A unified time scale intelligent control algorithm for micro grid based on extreme dynamic
programming. CSEE J. Power Energy Syst. (2019). https://doi.org/10.17775/cseejpes.2019.
00100
46. B. Wang, B. Fang, Y. Wang, H. Liu, Y. Liu, Power system transient stability assessment based
on big data and the core vector machine. IEEE Trans. Smart Grid 7 (2016). https://doi.org/10.
1109/TSG.2016.2549063
47. T. Shi, Y. He, T. Wang, B. Li, Open switch fault diagnosis method for PWM voltage source
rectifier based on deep learning approach. IEEE Access (2019). https://doi.org/10.1109/
ACCESS.2019.2917311
48. M. Tasdighi, M. Kezunovic, Preventing transmission distance relays maloperation under
unintended bulk DG tripping using SVM-based approach. Electr. Power Syst. Res. 142
(2017). https://doi.org/10.1016/j.epsr.2016.09.024
114 S. Sadeghi et al.

49. Y. Liu, Y. Liu, J. Liu, M. Li, T. Liu, G. Taylor, K. Zuo, A MapReduce based high performance
neural network in enabling fast stability assessment of power systems. Math. Probl. Eng.
(2017). https://doi.org/10.1155/2017/4030146
50. S.S. Maaji, G. Cosma, A. Taherkhani, A.A. Alani, T.M. McGinnity, On-line voltage stability
monitoring using an ensemble AdaBoost classifier, in 2018 4th International Conference on
Information Management, (ICIM, 2018, 2018)
51. R.A. Sowah, N.A. Dzabeng, A.R. Ofoli, A. Acakpovi, K.M. Koumadi, J. Ocrah, D. Martin,
Design of power distribution network fault data collector for fault detection, location and
classification using machine learning, in IEEE International Conference on Adaptive Science
and Technology, ICAST, (2018)
52. D. Wang, M. Glavic, L. Wehenkel, Trajectory-based supplementary damping control for
power system electromechanical oscillations. IEEE Trans. Power Syst. 29 (2014). https://
doi.org/10.1109/TPWRS.2014.2314359
53. T.N. Le, N.A. Nguyen, H.A. Quyen, Emergency control of load shedding based on coordina-
tion of artificial neural network and analytic hierarchy process algorithm, in Proceedings -
2017 International Conference on System Science and Engineering, ICSSE 2017, (2017)
54. C. Chen, M. Cui, F.F. Li, S. Yin, X. Wang, Model-free emergency frequency control based on
reinforcement learning. IEEE Trans. Ind. Informat. (2020). https://doi.org/10.1109/tii.2020.
3001095
55. Y. Xu, R. Zhang, J. Zhao, Z.Y. Dong, D. Wang, H. Yang, K.P. Wong, Assessing short-term
voltage stability of electric power systems by a hierarchical intelligent system. IEEE Trans.
Neural Netw. Learn. Syst. (2016). https://doi.org/10.1109/TNNLS.2015.2441706
56. M. Manohar, E. Koley, S. Ghosh, Enhancing the reliability of protection scheme for PV
integrated microgrid by discriminating between array faults and symmetrical line faults using
sparse auto encoder. IET Renew. Power Gener. 13 (2019). https://doi.org/10.1049/iet-rpg.
2018.5627
57. M.R. Alam, K.M. Muttaqi, A. Bouzerdoum, Evaluating the effectiveness of a machine
learning approach based on response time and reliability for islanding detection of distributed
generation. IET Renew. Power Gener. 11 (2017). https://doi.org/10.1049/iet-rpg.2016.0987
58. S. Zarrabian, R. Belkacemi, A.A. Babalola, Real-time smart grids control for preventing
cascading failures and blackout using neural networks: experimental approach for N-1-1
contingency. Int. J. Emerg. Electr. Power Syst. (2016). https://doi.org/10.1515/ijeeps-
2016-0039
59. R. Yousefian, S. Kamalasadan, Energy function inspired value priority based global wide-area
control of power grid. IEEE Trans. Smart Grid (2018). https://doi.org/10.1109/TSG.2016.
2555909
60. S.S. Jhang, H.J. Lee, C.N. Kim, C.H. Song, W.K. Yu, ANN Control for damping
low-frequency oscillation using deep learning, in Australasian Universities Power Engineer-
ing Conference, AUPEC 2018, (2018)
61. J.D. Pinzón, D.G. Colomé, Real-time multi-state classification of short-term voltage stability
based on multivariate time series machine learning. Int. J. Electr. Power Energy Syst. (2019).
https://doi.org/10.1016/j.ijepes.2019.01.022
62. I. Genc, R. Diao, V. Vittal, S. Kolluri, S. Mandal, Decision tree-based preventive and
corrective control applications for dynamic security enhancement in power systems. IEEE
Trans. Power Syst. (2010). https://doi.org/10.1109/TPWRS.2009.2037006
63. Q. Li, Y. Xu, C. Ren, A hierarchical data-driven method for event-based load shedding against
fault-induced delayed voltage recovery in power systems. IEEE Trans. Ind. Informat. (2020).
https://doi.org/10.1109/tii.2020.2993807
64. D. Ye, M. Zhang, D. Sutanto, A hybrid multiagent framework with Q-learning for power grid
systems restoration. IEEE Trans. Power Syst. (2011). https://doi.org/10.1109/TPWRS.2011.
2157180
5 Introduction and Literature Review of the Application of Machine. . . 115

65. M. Al Karim, J. Currie, T.T. Lie, A feature based distributed machine learning for post fault
restoration of a microgrid under different stochastic scenarios, in 2017 IEEE Innovative Smart
Grid Technologies - Asia: Smart Grid for Smart Community, ISGT-Asia 2017, (2018)
66. Y. Zhang, Y. Xu, Z.Y. Dong, R. Zhang, A hierarchical self-adaptive data-analytics method for
real-time power system short-term voltage stability assessment. IEEE Trans. Ind. Informat. 15
(2019). https://doi.org/10.1109/TII.2018.2829818
67. M.J. Ghorbani, M.A. Choudhry, A. Feliachi, A multiagent design for power distribution
systems automation. IEEE Trans. Smart Grid 7 (2016). https://doi.org/10.1109/TSG.2015.
2453884
68. Y. Zhou, J. Wu, Z. Yu, L. Ji, L. Hao, A hierarchical method for transient stability prediction of
power systems using the confidence of a SVM-based ensemble classifier. Energies (2016).
https://doi.org/10.3390/en9100778
69. A. Sharifian, S. Sharifian, A new power system transient stability assessment method based on
Type-2 fuzzy neural network estimation. Int. J. Electr. Power Energy Syst. 64 (2015). https://
doi.org/10.1016/j.ijepes.2014.07.007
70. J.J.Q. Yu, D.J. Hill, A.Y.S. Lam, J. Gu, V.O.K. Li, Intelligent time-adaptive transient stability
assessment system. IEEE Trans. Power Syst. 33 (2018). https://doi.org/10.1109/TPWRS.
2017.2707501
71. J.J.Q. Yu, A.Y.S. Lam, D.J. Hill, V.O.K. Li, Delay aware intelligent transient stability
assessment system. IEEE Access 5 (2017). https://doi.org/10.1109/ACCESS.2017.2746093
72. X. Gu, Y. Li, Bayesian multiple kernels learning-based transient stability assessment of power
systems using synchronized measurements, in IEEE Power and Energy Society General
Meeting, (2013)
73. A.B. Mosavi, A. Amiri, H. Hosseini, A learning framework for size and type independent
transient stability prediction of power system using twin convolutional support vector
machine. IEEE Access 6 (2018). https://doi.org/10.1109/ACCESS.2018.2880273
74. W. Hu, Z. Lu, S. Wu, W. Zhang, Y. Dong, R. Yu, B. Liu, Real-time transient stability
assessment in power system based on improved SVM. J. Mod. Power Syst. Clean. Energy
7 (2019). https://doi.org/10.1007/s40565-018-0453-x
75. Y. Zhang, T. Li, G. Na, G. Li, Y. Li, Optimized extreme learning machine for power system
transient stability prediction using synchrophasors. Math. Probl. Eng. (2015). https://doi.org/
10.1155/2015/529724
76. M. He, J. Zhang, V. Vittal, Robust online dynamic security assessment using adaptive
ensemble decision-tree learning. IEEE Trans. Power Syst. 28 (2013). https://doi.org/10.
1109/TPWRS.2013.2266617
77. C. Zhang, Y. Li, Z. Yu, F. Tian, Feature selection of power system transient stability
assessment based on random forest and recursive feature elimination, in Asia-Pacific Power
and Energy Engineering Conference, APPEEC, (2016)
78. Y. Li, G. Li, Z. Wang, Rule extraction based on extreme learning machine and an improved
ant-miner algorithm for transient stability assessment. PLoS One 10 (2015). https://doi.org/10.
1371/journal.pone.0130814
79. R. Zhang, Y. Xu, Z.Y. Dong, P. Zhang, K.P. Wong, Voltage stability margin prediction by
ensemble based extreme learning machine, in IEEE Power and Energy Society General
Meeting, (2013)
80. H.H. Goh, Q.S. Chua, S.W. Lee, B.C. Kok, K.C. Goh, K.T.K. Teo, Evaluation for voltage
stability indices in power system using artificial neural network, in Procedia Engineering,
(2015)
81. S.M. Pérez-Londoño, G. Olivar-Tost, J.J. Mora-Florez, Online determination of voltage
stability weak areas for situational awareness improvement. Electr. Power Syst. Res. 145
(2017). https://doi.org/10.1016/j.epsr.2016.12.026
82. K.S. Sajan, V. Kumar, B. Tyagi, ICA based artificial neural network model for voltage
stability monitoring, in IEEE Region 10 Annual International Conference, Proceedings/
TENCON, (2016)
116 S. Sadeghi et al.

83. K.S. Sajan, V. Kumar, B. Tyagi, Genetic algorithm based support vector machine for on-line
voltage stability monitoring. Int. J. Electr. Power Energy Syst. 73 (2015). https://doi.org/10.
1016/j.ijepes.2015.05.002
84. A.R. Bahmanyar, A. Karami, Power system voltage stability monitoring using artificial neural
networks with a reduced set of inputs. Int. J. Electr. Power Energy Syst. 58 (2014). https://doi.
org/10.1016/j.ijepes.2014.01.019
85. S.R. Nandanwar, M.L. Kolhe, S.B. Warkad, N.P. Patidar, V.K. Singh, Voltage security
assessment by using PFDT and CBR methods in emerging power system, in Energy Procedia,
(2018)
86. H. Mohammadi, G. Khademi, M. Dehghani, D. Simon, Voltage stability assessment using
multi-objective biogeography-based subset selection. Int. J. Electr. Power Energy Syst. 103
(2018). https://doi.org/10.1016/j.ijepes.2018.06.017
87. V. Jayasankar, N. Kamaraj, N. Vanaja, Estimation of voltage stability index for power system
employing artificial neural network technique and TCSC placement. Neurocomputing 73
(2010). https://doi.org/10.1016/j.neucom.2010.07.006
88. S. Khokhar, A.A. Mohd Zin, A.P. Memon, A.S. Mokhtar, A new optimal feature selection
algorithm for classification of power quality disturbances using discrete wavelet transform and
probabilistic neural network. J. Int. Meas. Confed. 95 (2017). https://doi.org/10.1016/j.
measurement.2016.10.013
89. A.A. Abdoos, P. Khorshidian Mianaei, M. Rayatpanah Ghadikolaei, Combined VMD-SVM
based feature selection method for classification of power quality events. Appl. Soft
Comput. J. 38 (2016). https://doi.org/10.1016/j.asoc.2015.10.038
90. K. Manimala, K. Selvi, R. Ahila, Optimization techniques for improving power quality data
mining using wavelet packet based support vector machine. Neurocomputing 77 (2012).
https://doi.org/10.1016/j.neucom.2011.08.010
91. U. Singh, S.N. Singh, Application of fractional Fourier transform for classification of power
quality disturbances. IET Sci. Meas. Technol. 11 (2017). https://doi.org/10.1049/iet-smt.2016.
0194
92. Z. Liu, Y. Cui, W. Li, A classification method for complex power quality disturbances using
EEMD and rank wavelet SVM. IEEE Trans. Smart Grid 6 (2015). https://doi.org/10.1109/
TSG.2015.2397431
93. Z. Liquan, G. Meijiao, W. Lin, Classification of multiple power quality disturbances based on
the improved SVM, in Proceedings of the 2017 International Conference on Wireless
Communications, Signal Processing and Networking, WiSPNET 2017, (2018)
94. U. Singh, S.N. Singh, A new optimal feature selection scheme for classification of power
quality disturbances based on ant colony framework. Appl. Soft Comput. J. 74 (2019). https://
doi.org/10.1016/j.asoc.2018.10.017
95. D. De Yong, S. Bhowmik, F. Magnago, An effective power quality classifier using wavelet
transform and support vector machines. Expert Syst. Appl. 42 (2015). https://doi.org/10.1016/
j.eswa.2015.04.002
96. S. Shukla, S. Mishra, B. Singh, Empirical-mode decomposition with hilbert transform for
power-quality assessment. IEEE Trans. Power Deliv. 24 (2009). https://doi.org/10.1109/
TPWRD.2009.2028792
97. H. Erişti, Y. Demir, A new algorithm for automatic classification of power quality events
based on wavelet transform and SVM. Expert Syst. Appl. 37 (2010). https://doi.org/10.1016/j.
eswa.2009.11.015
98. H. Liu, F. Hussain, Y. Shen, S. Arif, A. Nazir, M. Abubakar, Complex power quality
disturbances classification via curvelet transform and deep learning. Electr. Power Syst. Res.
163 (2018). https://doi.org/10.1016/j.epsr.2018.05.018
99. F. Daneshfar, H. Bevrani, Load-frequency control: a GA-based multi-agent reinforcement
learning. IET Gener. Transm. Distrib. (2010). https://doi.org/10.1049/iet-gtd.2009.0168
5 Introduction and Literature Review of the Application of Machine. . . 117

100. H. Li, C. Li, Y. Liu, Maximum frequency deviation assessment with clustering based on
metric learning. Int. J. Electr. Power Energy Syst. 120 (2020). https://doi.org/10.1016/j.ijepes.
2020.105980
101. L. Yin, T. Yu, L. Zhou, L. Huang, X. Zhang, B. Zheng, Artificial emotional reinforcement
learning for automatic generation control of large-scale interconnected power grids. IET
Gener. Transm. Distrib. 11, 2305–2313 (2017). https://doi.org/10.1049/iet-gtd.2016.1734
102. M. Abouhea, W. Gueaieb, A. Sharaf, Load frequency regulation for multi-area power system
using integral reinforcement learning. IET Gener. Transm. Distrib. 13 (2019). https://doi.org/
10.1049/iet-gtd.2019.0218
103. Z. Yan, Y. Xu, A multi-agent deep reinforcement learning method for cooperative load
frequency control of multi-area power systems. IEEE Trans. Power Syst. (2020). https://doi.
org/10.1109/tpwrs.2020.2999890
104. Q. Wang, F. Li, Y. Tang, Y. Xu, Integrating model-driven and data-driven methods for power
system frequency stability assessment and control. IEEE Trans. Power Syst. 34 (2019). https://
doi.org/10.1109/TPWRS.2019.2919522
105. Z. Hou, J. Folium, P. Etingov, F. Tuffner, D. Kosterev, G. Matthews, Machine learning of
factors influencing damping and frequency of dominant inter-area modes in the WECC
interconnect, in 2018 International Conference on Probabilistic Methods Applied to Power
Systems, PMAPS 2018 - Proceedings, (2018)
Chapter 6
Introduction and Literature Review
of the Application of Machine Learning/
Deep Learning to Load Forecasting
in Power System

Arash Moradzadeh, Amin Mansour-Saatloo, Morteza Nazari-Heris,

Behnam Mohammadi-Ivatloo, and Somayeh Asadi

Abbreviation

ACE Average coverage error

ANN Artificial neural network
APE Absolute percentage error
AR Auto-regressive
ARIMA Auto-regressive integrated moving average
ARMA Auto-regressive moving average
BGA Binary genetic algorithm
BPNN Back-propagation neural network
CNN Convolution neural network
CV Coefficient of variance
CWC Coverage width-based criterion
DA Direction accuracy
DAME Daily absolute maximum error
DBN Deep belief network
DBN Deep neural network
DC Directional change

A. Moradzadeh · A. Mansour-Saatloo · B. Mohammadi-Ivatloo (*)

Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
e-mail: [email protected]; [email protected];
[email protected]
M. Nazari-Heris · S. Asadi
Department of Architectural Engineering, Pennsylvania State University,
State College, PA, USA
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 119
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_6
120 A. Moradzadeh et al.

DMD Dynamic mode decomposition

EMAE Envelope-weighted mean absolute error
EMD Empirical mode decomposition
ENN Elman neural network
ESN Echo state network
FCRBM Factored conditional restricted Boltzmann machine
FFN Feed-forward neural network
GB Gradient boosting
GBA Gradient boosting machine
GELM Generalized extreme learning machine
GRU Gated recurrent unit
GWO Gray wolf optimizer
HR Heat rate
IA Index of agreement
IWNN Improved wavelet neural network
LSTM Long short-term memory
MAAPE Mean arctangent absolute percentage error
MAE Mean absolute error
MAPE Mean absolute percentage error
MedAE Median absolute error
MFFNN Multilayer feed-forward neural network
MLP Multilayer perceptron
MLR Multiple linear regression
MOD Mean outside distance
MWPI Mean width of prediction interval
nMAE Normalized mean absolute error
NRMSE Normalized root mean squared error
NYISO New York independent system operator
PCR Principal component regression
PICP Prediction interval coverage probability
PJM Pennsylvania New Jersey Maryland
PMSE Prognostication mean square error
QRF Quantile regression forest
RBFNN Radial basis function neural network
RF Random forest
RMSE Root mean square error
RMSLE Root mean square logarithmic error
RNN Recurrent neural network
RVM Relevance vector machine
VMD Variational mode decomposition
WMAE Weighted mean absolute error
WNN Wavelet neural network
WOA Whale optimization algorithm
WT Wavelet transform
6 Introduction and Literature Review of the Application of Machine. . . 121

6.1 Introduction

So far, many scholars have proposed various methods to improve the accuracy of
load forecasting. Electrical load forecasting approaches can be classified into three
categories: statistical methods, artificial intelligence methods, and hybrid methods.
Statistical methods include time series models such as auto-regressive (AR) [1],
auto-regressive moving average (ARMA) [2], auto-regressive integrated moving
average (ARIMA) [3], seasonal ARIMA (SARIMA) [4], linear regression methods
[5], multiple linear regression methods [6], and exponential smoothing methods
[7]. The accuracy of ARIMA-based methods depends on some external variables
and can be further improved using exogenous variables [8]. However, all these
methods have good performance in linear systems and are insufficient in nonlinear
systems. Since the real-world electrical load has nonlinear inherent, some
researchers have tried to introduce some models by extending Kalman filter methods
to handle this issue [9], but the extended models could not show high accuracy. To
overcome the mentioned challenges, artificial intelligence methods have been devel-
oped for usage in the electrical load forecasting.
In recent years, artificial intelligence, due to its ability in forecasting and dealing
with nonlinear data, has received great significance. The most popular techniques are
artificial neural network (ANN) methods such as generalized regression neural
network (GRNN) [10], multilayer perceptron (MLP) [11], radial bias function
(RBF) recurrent neural network [12], back-propagation neural network (BPNN)
[13], support vector machine (SVM) [14], support vector regression (SVR) methods
[15], fuzzy logic methods [16], and data clustering methods [17]. These methods are
used for electrical load forecasting because of their capacity to learn and handle
complex systems. However, artificial intelligence methods still have drawbacks,
including random selection of weight parameters, high execution time, overfitting,
etc. To this end, hybrid models were developed to enhance artificial intelligence
performance. For instance, in [18], linear extrapolation combined with fuzzy SVR,
and in [19], the season-specific similarity concept was combined with SVM and
firefly algorithm to forecast the seasonal electrical load. Furthermore, many of the
artificial intelligence methods are combined with optimization methods, e.g., the
SVM method was hybridized with gray wolf optimizer (GWO) in [20] or a hybrid
model of genetic algorithm, particle swarm optimization, and back-propagation
neural network (GA-PSO-BPNN) was proposed in [21].
This chapter has a review of the artificial intelligence methods in electrical load
forecasting. To this end, the existing scholars are scrutinized from two different
scopes. The first scope classifies the papers from a forecasting time horizon point of
view, i.e., short-term forecasting and long-term forecasting. Reviewing of the
utilized techniques and their performance analysis are the targets of the second
scope.
122 A. Moradzadeh et al.

6.2 Most Important Challenges in Power Systems Short-,

Medium-, and Long-Term Load Forecasting

6.2.1 Short-Term Load Forecasting

Short-term load forecasting (STLF) includes 1 h ahead up to 1 week ahead forecast-

ing, and it is important to the scheduling of the power system since it has a
considerable impact on power resources’ generation, limitations, usage constraints,
spinning reserve, security, and reliability. Underestimation of the load leads to power
shortage, and overestimation of the load causes power waste and an insufficient
number of power plants.
In [22], binary genetic algorithm (BGA) and Gaussian process regression were
used for the feature selection process and fitness score measurement of features,
respectively. Multi-kernel algorithm and SVR methods were proposed in [23] for
load forecasting application, where kernel functions of SVR machine are optimized
via the utilized algorithm. For the distribution system reconfiguration, genetic
algorithm was introduced in [24], in which wavelet transform-based ARIMA
was employed for wind, solar, and load forecasting. In [25], the deep belief network
was utilized for hourly load forecasting, in which the preprocessing of data was
conducted using Box-Cox transformation. A probabilistic load forecasting study
considering data noise uncertainty was performed in [26] using the wavelet neural
network improved by the generalized extreme learning machine. RNN based on
LSTM and gated recurrent unit (GRU) was proposed in [27] for STLF, where the
RNN consists of two different layers that combined with LSTM and GRU. In [28],
an STLF considering the uncertainty of PV, wind, and load was introduced, in which
deep wavelet transform was used to feature selection, and deep neural network
(DNN) consisting of autoencoder neural networks with a cascade layer was used
as the forecasting tool. In [29], wavelet decomposition was used to decompose the
data into different resolutions. Then gray and cubic exponential smoothing models
were applied to forecast the load. Wavelet transform for feature selection was
utilized in [30], and relevance vector machine was proposed to forecast the load.
Three different aggregation strategies, i.e., information aggregation, hierarchical
aggregation, and model aggregation, were analyzed in [15] for four machine learning
algorithms. Full wavelet transform instead of traditional wavelet transform along
with multilayer neural network were proposed in [31] for STLF. In [11], a feature
selection method, namely, RReliefF, was employed to choose the features, and then,
MLP was applied to forecast the load. Different DNN-based techniques using the
stacking ensemble method was developed in [32]. In this reference, combination of
various STLF methods outperforms single DNN methods. In [33], a supervised
Fuzzy-ARTMAP neural network was developed to forecast the various nodes’ load,
where a reverse training technique was introduced to forecast. Dynamic mode
decomposition was used in [34] to introduce an STLF method based on error
correction. The introduced algorithm includes three stages, i.e., data selection
stage, error forecasting stage, and error correction stage, in which the first stage
6 Introduction and Literature Review of the Application of Machine. . . 123

benefits from the gray relation analysis method. To consider the relation of cost and
load in the electrical grid, a hybrid of two-sided wavelet transform and SVM, along
with the revised mutual information-based feature selection method, was developed
in [14] for STLF. An ensemble framework based on WaveNet learners for STLF was
proposed in [35], where multiple techniques such as bootstrapping and stacked
generalization algorithms were used for the ensemble aggregation of WaveNet
learners. Gravitational search algorithm (GSA) and PSO were hybridized with single
forecasting methods such as SVR for STLF in [36]. An STLF of an individual
household was performed in [37], where the Bayesian networks were used as the
forecasting tool. Two of ANN-based algorithms, i.e., FFNN and echo state network,
were compared in [38] to investigate their performance in the commercial buildings
STLF. In the same context, in [39], nine different combinations of RNN along with
clustering were introduced to commercial buildings STLF. GRNN and GSA were
applied to the short-term load and price forecasting in [10], where variational mode
decomposition was used to improve the forecasting accuracy. In [40], cross-industry
standard process for data mining combined with ARIMA was proposed, where the
autocorrelation function was employed to compute the correlation of a series and its
time shift. In [20], to overcome the shortcoming of existing methods in the forecast-
ing of special days load, a hybrid of SVM and GWO were proposed. To conduct the
probabilistic load forecasting in [41], the probability distribution of loads was
discretized by dividing the load for multiple ranges, and then, a convolutional neural
network (CNN) was applied for STLF. A deep learning framework, namely, hybrid
deep meta-ensemble networks composed of four modules, was proposed in [42], in
which the first one is the local forecaster of each series, the second one is the global
forecaster, the third one is supervised feature learner, and the last one is combiner of
local and global forecasts based on obtained features. A hybrid factored conditional
restricted Boltzmann machine (FCRBM) and the genetic wind-driven optimization
algorithm was introduced in [43], where a modified mutual information approach
was used to feature selection. A deep-ensemble STLF method based on the LASSO
quantile forecast combination approach was introduced in [44], where the model is a
probabilistic load forecasting, and in this model, there is no need for feature
selection. A comparative study of three different STLF models, i.e., random forest,
multiple linear regression, and gradient boosting, were conducted in [45]. According
to the obtained results, the gradient boosting model outperforms the other two
models, and it shows good performance in both low and high load periods. In
[46], Takagi-Sugeno-Kang neuro-fuzzy forecasting technique was applied for
STLF, in which, via training this technique with a locally linear model tree method,
all the parameters such as the number of neurons and functions are adjusted by the
algorithm. In [47], to perform an STLF, a variational mode decomposition was used
for preprocessing of data. After then, quantile regression forest hybridized with
Bayesian optimization algorithm were applied to forecast the decomposed data
separately, and kernel density estimation was utilized to reconstruct the forecasting
results. A multi-objective deep belief network along with empirical mode decom-
position was proposed in [48]. A hybrid model composed of variational mode
decomposition and autoencoder methods to extract the sufficient subsignals along
124 A. Moradzadeh et al.

with LSTM network for STLF were proposed in [49]. In [50], decision tree and
weighted average methods were used to decompose the data based on hourly and
daily attributes, respectively. Then, a regression model based on time series method
and SVM method were applied for STLF. In [51], whale optimization method was
used to optimize the parameter of CNN, where the CNN was restructured via MLP
method to extract an efficient STLF algorithm. A hybrid Elman neural network and
PSO were proposed in [52] for STLF, where the PSO was used to search for optimal
rate of the network. To overcome the overfitting problem of ANN, a hybrid
algorithm composed of an evolutionary algorithm to search for the optimal weights
and ANN was proposed in [53], and the evolutionary algorithm was integrated with
a controlled Gaussian mutation technique to enhance the convergence solutions. In
[54], a methodology for time series analysis was introduced to extend feature
selection of SVR idea which is optimized by the Kernel-penalized iterative method.
In [55], the fuzzy clustering method was used to initially cluster raw data. Then
radial basis function neural network and CNN were applied to conduct the load
forecasting, in which both of them optimized via Adam optimization method.
One-dimensional CNN hybridized with RNN in [56] to overcome the incorrect
generation of the hidden state vector and also to calibrate the forecasting time. To
eliminate the outlier data effect from training data, differential evolutionary PSO was
proposed in [57], where ANN was applied for STLF. In [58], to enhance both
forecasting error and time, a multi-objective model combined with the ANN was
proposed. In [59], a hybrid STLF method for microgrids composed of SVR and
LSTM was introduced, in which the hybrid method outperforms single SVR and
LSTM. In [60], a Levenberg-Marquardt algorithm based on RNN was utilized to
forecast the maritime microgrids. An overview of the short-term literature is pro-
vided in Table 6.1.

6.2.2 Midterm Load Forecasting

Midterm load forecasting (MTLF) includes 1 week up to 12 months ahead of

forecasting. This type of load forecasting is important for the maintenance and
operation of the power system. In [61], a combination of three different models,
i.e., random forest regression (RFR), gradient boosting decision tree (GBDT), and
SVR, were proposed for probabilistic MTLF, where the kernel density function was
applied to achieve the probability density distribution of the load. To plan to produce
and purchase a hybrid ST-MTLF model was presented in [62], where the proposed
model employs the multilayer feed-forward neural network (MFFNN) and grass-
hopper optimization algorithm (GOA) to forecast the load. The transfer learning
method, along with the Pearson correlation coefficient to enhance the forecasting
performance, was introduced in [63], in which the aim of the algorithm is monthly
MTLF of a city. Empirical mode decomposition with the LSTM network was used in
[64] to forecast the seasonal and daily load demand of a city. A hybrid of PSO and
6 Introduction and Literature Review of the Application of Machine. . . 125

Table 6.1 Overview on short-term forecasting literature

Ref. Year Data Time horizon
[22] 2019 Occupancy, weather features, and electricity price 30-minute
[23] 2017 Highest temperature, lowest temperature, and wind Daily
[24] 2019 IEEE 34-bus and IEEE 123-bus Hourly
[25] 2019 Temperature, electricity price, humidity, barometric pressure, Day-ahead and
and wind speed week-ahead
[26] 2018 Ontario and Australian electricity markets data Hourly
[27] 2019 Daily weather data Daily
[28] 2020 Solar and wind generations Hourly
[29] 2019 Daily
[30] 2020 Highest and lowest temperature Hourly
[31] 2020 Temperature, humidity, hours, days, and day types Hourly
[15] 2020 Weather features Daily
[32] 2020 Weather features Daily
[11] 2020 New England independent system operator data Hourly
[33] 2020 New Zealand consumption data 30-minute
[14] 2020 NYISO, NSW zone in Australia’s market, and PJM data Hourly
[34] 2020 Weather features, renewable generation, and demand response Hourly
data
[35] 2019 Hourly load from Italy Hourly
[36] 2019 Queensland, Australia data Hourly
[37] 2020 Commission for Energy Regulation (CER) smart metering Hourly
project data
[38] 2020 University campus in Milan, Italy Hourly
[39] 2020 Weather data 15-minute
[10] 2020 PJM and Spanish market data Hourly
[40] 2020 Control signals, operating modes and disturbance signals, the 10-s
status of the auxiliary units, and physical properties
[20] 2020 Regional special event days and temperature Daily
[41] 2020 New England independent system operator data Hourly
[42] 2020 US utility data Hourly
[43] 2020 PJM data Hourly
[44] 2019 CER data Hourly
[45] 2019 Temperature, meteorological variable, holiday, and solar Hourly
capacity
[46] 2020 Climate data and holiday Hourly
[47] 2020 Weighted temperature and humidity index (WTHI) and day Daily
type
[48] 2020 Power load data and algorithmic parameters Weekly
[49] 2020 Hierarchical data Daily
[50] 2020 Dry-bulb temperature, dew point temperature, relative humid- Hourly
ity, moisture content, wind speed, and air enthalpy
[51] 2020 Electrical load time series 30-minute
[52] 2019 Particular power consumption 30-minute
(continued)
126 A. Moradzadeh et al.

Table 6.1 (continued)

Ref. Year Data Time horizon
[53] 2019 Electrical load data, temperature, rainfall, wind speed, and Hourly
holiday
[54] 2019 Methodological data Daily
[55] 2020 Temperature data Weekly
[56] 2019 Day type, temperature, humidity, wind speed, electric load, 30-minute
off-peak, mid-peak, and on-peak
[57] 2019 Weather data, day type, and flag Daily
[58] 2019 Dry-bulb temperature, dew point temperature, hour of the day, Hourly
day of the week, holiday/weekend indicator (0 or 1), 168-hr
(previous week) lagged load, 24-hr lagged load, the previous
24-hr average load
[59] 2020 Household and commercial load consumption Hourly
[60] 2020 Next three latest days for thrusting demand and environmental Hourly
distortion of 72 h

Table 6.2 Overview on medium-term forecasting literature

Ref. Year Data Time horizon
[61] 2019 Electrical load, natural gas load, and price, the average Monthly
retail price of electricity, natural gas consumed by the
electric power sector, and energy electric power sector
CO2 emissions
[62] 2020 The times per hour, day, and month and the temperature Hourly and monthly
[63] 2020 Calendar data, population data, and weather data Monthly
[64] 2018 Data of Chandigarh, India Season, day, and time
interval of a day
[16] 2013 Time and hourly load Hourly, daily, and
monthly

fuzzy neural network were used in [16] to conduct hourly, daily, and monthly load
forecasting. An overview of the medium-term literature is provided in Table 6.2.

6.2.3 Long-Term Load Forecasting

Long-term load forecasting (LTLF) includes 1 year up to the next 50 years ahead of
forecasting. LTLF has a significant impact on the planning of the power system and
expansion decisions. LTLF is complicated than STLF and MTLF due to the high
uncertainty of load and economic factors such as energy resources price values. The
Prophet and Holt-Winters methods were used in [65] for LTLF of Kuwait for the
next decade. Three different configuration methods were employed in [66] to
achieve a multi-timestep forecasting problem from a multiyear LTLF. Moreover,
an unsupervised learning technique was used to group the feeders. In [12], a fuzzy
6 Introduction and Literature Review of the Application of Machine. . . 127

Table 6.3 Overview on long-term forecasting literature

Ref. Year Data Time horizon
[65] 2020 Load data Yearly
[66] 2020 The urban distribution system of Canada Yearly
[12] 2020 Temperature series and holidays Yearly
[67] 2020 Iran National Grid data Yearly
[68] 2020 Brazilian energy utility (Cemig distribution) Yearly

neural network along with robust-type fuzzy rules was developed for LTLF, where a
selection method for weather conditions was proposed to improve the accuracy of
the forecasting. A hybrid model consisting of ARIMA, ANN, and SVR were
introduced in [67], where the PSO algorithm was used to find the optimal value of
parameters. In [68], a spatial load forecasting using the spatial convolution operator
was conducted, in which the utilized convolution method acts as a low-pass filter in
the frequency domain and in the space defines the relation of nodes. An overview of
the long-term literature is provided in Table 6.3.

6.3 Machine Learning and Deep Learning Applications

in Load Forecasting of Power System

Today, machine learning and deep learning methods have found a special place in
applications related to energy and power systems. Especially in issues of power
systems that are related to data processing, the effectiveness of the machine learning
and deep learning methods is multiplied. Meanwhile, as mentioned in the previous
sections, load forecasting is one of the issues that mainly deals with big, historical,
and time series data. So far, there are many methods for load forecasting in power
grids, each of which has provided different forecasts according to their capabilities.
The load used in power grids follows a variety of parameters such as temperature,
wind and solar information, date and time data of the week, etc., which are highly
interrelated. However, those methods that can accurately estimate and model the
relationship between input variables and load data can be highly effective. Most
machine learning and deep learning methods have this capability, and they have
been able to perform high-performance load forecasting in a variety of projects. It
should be noted that because the data related to the load are of continuous type,
regression applications are suitable and ideal for doing this process. In the contin-
uation of this section, different methods of machine learning and deep learning that
have been used to forecast the load will be introduced.
128 A. Moradzadeh et al.

6.3.1 Machine Learning and Deep Learning Algorithms Used

for Load Forecasting

So far, many methods of machine learning and deep learning have been developed as
a tool for processing and predicting load data. Among the machine learning algo-
rithms, the applications of SVR, GRNN, ELM, random forest, and decision tree in
load prediction can be mentioned. Each of these methods uses a specific pattern for
data processing. The SVR can be considered as a regression version of the support
vector machine (SVM) [59]. The ELM is also derived from feed-forward neural
networks that have a very high convergence performance with a structure similar to
the MLP neural network. In addition to regression applications, this method is used
for discrete data and classification. The ELM can be used to evaluate big data, and
this is one of the advantages of this method [69]. Meanwhile, the GRNN method is
also seen as one of the most widely used algorithms in solving regression problems.
The GRNN is an improved technique in neural networks based on nonparametric
regression. This method in particular can be used as a powerful tool to solve
problems related to online dynamic systems. High training speed and fast conver-
gence against big data are the prominent features of this technique [70]. As men-
tioned, most machine learning algorithms have a neural network-based structure and
often focus on improving the training process so that they can estimate the close
relationship between input and output parameters.
With the introduction of deep learning techniques, the use of machine learning
algorithms was somewhat reduced so that deep learning methods were able to
improve most of the problems associated with machine learning algorithms. By
focusing on the literature review, methods such as the LSTM, deep belief network
(DBN), Bi-LSTM, autoencoder, and CNN can be mentioned, which have been used
to forecast the load. Each of these methods has a unique structure that can have
different functions based on its original structure. The LSTM method has been
introduced as an algorithm for improving the structure of RNNs. This algorithm
was able to eliminate the vanishing gradient problem in the structure of the RNNs
[71]. The LSTM is a deep learning technique that is used as a powerful tool for time
series data. The proper performance of this method and its high ability to analyze the
big, high-dimension, and time series data has introduced this method as one of the
most suitable methods of load forecasting [59]. The DBN is one of the deep learning
networks that with one deep neural network in its structure, consisting of several
variable layers and with connections between layers, learns in-depth training of data
structure. This network can also be used without supervision, in which case it may be
able to reconstruct its inputs. High-processing speed, data reconstruction for better
output, and acceptable performance for different volumes of data are the obvious
advantages of this network [72]. Autoencoders are one of the deep learning appli-
cations used for unsupervised training of data. These networks have a structure
similar to FNNs and consist of two components, including an encoder and a decoder.
Minimizing the mean reconstruction loss is one of the most important issues to
consider in the training of the autoencoders. These networks are mainly used in
6 Introduction and Literature Review of the Application of Machine. . . 129

Table 6.4 Overview on load forecasting categories

Method Algorithm References
ANN MLP [11–13, 15, 16, 38, 40, 42, 51, 53, 58, 62, 67]
RBF [10, 12, 55]
BPNN [12, 16, 21, 36, 52]
Fuzzy [16, 18, 33, 46, 55, 76]
WNN [31, 35, 36, 42]
Machine learning SVR [12, 14, 15, 18, 50, 51, 59, 61, 67]–[20, 23, 30, 36, 49]
GRNN [10, 51, 52]
ELM [26]
Random forest [15, 45, 47, 61, 77]
Decision tree [61]
Deep learning LSTM [12, 27, 39, 42, 49, 59, 64, 66, 78]–[80]
RNN [49, 56]
DBN [25, 48, 49]
DNN [28, 32]
Autoencoder [39, 49]
CNN [39, 41, 55, 56, 68]

applications such as face recognition, fingerprinting, and image processing. The high
performance of these networks in processing large-scale data is one of the advan-
tages of this method. Autoencoders have so far shown significant results in load
forecasting applications [73, 74]. In addition to most of the deep learning algorithms
mentioned in the introduction and performance of each, the CNN is one of the most
powerful and widely used methods of deep learning, which is mainly used for
classification applications. Layer-to-layer structure and feature extraction from
input data in several stages are prominent features of the CNN [41, 75]. This
technique is mainly used in many industrial applications and scientific projects due
to its high ability to extract data features and detect behavioral patterns.
By reviewing most of the machine learning and deep learning techniques and
observing the structure of each, it is inferred that the first basis of most of these
methods is ANNs. In presenting each of the machine learning and deep learning
techniques, an attempt has been made to improve one of the ANN algorithms and to
present a new structure. Accordingly, a review of the literature reveals the dramatic
applications of various ANN methods in load forecasting programs. Among the
ANN algorithms, networks such as MLP, BPNN, RBF, WNN, and fuzzy network,
have been used mainly for load forecasting. Table 6.4 categorizes the load forecast-
ing studies based on the proposed method for each paper, which is based on the
ANNs, machine learning, and deep learning algorithms.
130 A. Moradzadeh et al.

6.3.2 Performance Assessment of Algorithms

Data processing and estimating the relationship between input variables and the
target can be done by different types of learning methods. Each method provides
performance based on its structure and formulation. Using machine learning and
deep learning applications for forecasting issues, a variety of results are obtained,
whereas the superiority and effectiveness of each method are achieved after evalu-
ating and comparing the results. Comparison of results related to learning methods is
done by various statistical evaluation metrics. Mean squared error (MSE), root mean
squared error (RMSE), mean absolute error (MAE), mean absolute percentage error
(MAPE), coefficient of variation of root mean squared error (CVRMSE), and
correlation coefficient (R) are the most widely used statistical metrics that have
been used to evaluate the results of load forecasting by deep learning and machine
learning methods [58, 78, 81]. Each of these metrics refers to a specific concept.
Thus, the lower the error values and the higher the correlation value, the more
accurate and close the prediction results are to the actual values. In recent years,
researchers have dramatically developed a combination of a variety of machine
learning and deep learning methods that primarily aim to increase prediction accu-
racy and reduce results-related errors. Therefore, it is observed that the evaluation of
results and the use of statistical performance evaluation metrics are of great impor-
tance. Table 6.5 categorizes the load forecasting studies based on the evaluation
metrics used in each paper.

6.4 Conclusions

This chapter aimed at providing an updated review on application of machine/deep

learning methods and artificial intelligence approaches in electrical load forecasting.
Accordingly, this study focused on two viewpoints of the application of machine
learning/deep learning to load forecasting in power system in terms of a: time
horizon and b: utilized techniques and their performance analysis. In the first

Table 6.5 Load forecasting categories based on various evaluation metrics

Statistical evaluation
metric References
MSE [59, 61, 79, 82]
RMSE [10, 14, 40, 47]–[24, 25, 27, 52, 58, 59, 62, 64, 66, 67, 78, 80]–[29, 36,
38, 39]
MAE [10–13, 15, 22, 27, 28, 30, 36, 46, 47, 50, 53, 58, 59, 62, 67, 78, 80]
MAPE [10, 11, 25, 27, 28, 30, 31, 33, 36, 38]–[12, 40, 46]–[14–16, 19, 20, 22,
24, 53, 58, 59, 62, 66, 67, 78, 80]
CVRMSE [78]
R [10, 15, 21, 35, 36, 39, 40, 45, 50, 58, 59, 66, 78, 80]
6 Introduction and Literature Review of the Application of Machine. . . 131

viewpoint, short-term, midterm, and long-term forecasting strategies were basically

defined and investigated in terms of the pros and cons of machine learning/deep
learning methods. Various methods used in the forecasting of electrical load, the data
used, and the exact time horizon of the selected papers from the literature showed
successful application and high performance of such methods. In the second part, the
applied techniques and their performance analysis were performed that demonstrated
the focus of most machine learning algorithms in improving the training process.
Also, various statistical evaluation metrics that are used for evaluating the perfor-
mance of the forecasting methodologies were discussed, and the popularity of each
evaluation metrics in the literature were analyzed. The current chapter can be useful
for researchers of energy systems with a focus on in the area of load forecasting and
machine/deep learning studies.

References

1. A. H. Vahabie, M. M. R. Yousefi, B. N. Araabi, C. Lucas, and S. Barghinia, Combination of

singular spectrum analysis and autoregressive model for short term load forecasting. 2007 IEEE
Lausanne Power Tech, (2007), pp. 1090–1093
2. S.-J. Huang, K.-R. Shih, Short-term load forecasting via ARMA model identification including
non-Gaussian process considerations. IEEE Trans. Power Syst. 18(2), 673–679 (2003)
3. X. Wang and Y. Liu, ARIMA time series application to employment forecasting. In 2009 4th
International Conference on Computer Science & Education, (2009), pp. 1124–1127
4. V. Debusschere, S. Bacha, One week hourly electricity load forecasting using neuro-fuzzy and
seasonal ARIMA models. IFAC Proceedings Volumes 45(21), 97–102 (2012)
5. K.-B. Song, Y.-S. Baek, D.H. Hong, G. Jang, Short-term load forecasting for the holidays using
fuzzy linear regression method. IEEE Trans. Power Syst. 20(1), 96–101 (2005)
6. N. Amral, C. S. Ozveren, and D. King, Short term load forecasting using multiple linear
regression. In 2007 42nd International universities power engineering conference, (2007),
pp. 1192–1198
7. J.W. Taylor, Short-term load forecasting with exponentially weighted methods. IEEE Trans.
Power Syst. 27(1), 458–464 (Feb. 2012). https://doi.org/10.1109/TPWRS.2011.2161780
8. N. Elamin, M. Fukushige, Modeling and forecasting hourly electricity demand by SARIMAX
with interactions. Energy 165, 257–268 (2018)
9. H. Takeda, Y. Tamura, S. Sato, Using the ensemble Kalman filter for electricity load forecasting
and analysis. Energy 104, 184–198 (2016)
10. A. Heydari, M.M. Nezhad, E. Pirshayan, D.A. Garcia, F. Keynia, L. De Santoli, Short-term
electricity price and load forecasting in isolated power grids based on composite neural network
and gravitational search optimization algorithm. Appl. Energy 277, 115503 (2020)
11. A. Rafati, M. Joorabian, and E. Mashhour, An efficient hour-ahead electrical load forecasting
method based on innovative features. Energy, p. 117511, 2020
12. Z. Wen, L. Xie, Q. Fan, H. Feng, Long term electric load forecasting based on TS-type recurrent
fuzzy neural network model. Electr. Pow. Syst. Res. 179, 106106 (2020)
13. A. Ganguly, K. Goswami, A. Mukherjee, and A. K. Sil, Short-term load forecasting for peak
load reduction using artificial neural network technique. In Advances in Computer, Communi-
cation and Control, Springer, 2019, pp. 551–559
14. E. Zhao, Z. Zhang, and N. Bohlooli, Cost and load forecasting by an integrated algorithm in
intelligent electricity supply network. Sustainable Cities and Society, p. 102243, 2020
132 A. Moradzadeh et al.

15. C. Feng, J. Zhang, Assessment of aggregation strategies for machine-learning based short-term
load forecasting. Electr. Pow. Syst. Res. 184, 106304 (2020)
16. R.-J. Wai, Y.-C. Huang, Y.-C. Chen, Y.-W. Lin, Performance comparisons of intelligent load
forecasting structures and its application to energy-saving load regulation. Soft. Comput. 17
(10), 1797–1815 (2013)
17. Z. Deng, B. Wang, Y. Xu, T. Xu, C. Liu, Z. Zhu, Multi-scale convolutional neural network with
time-cognition for multi-step short-term load forecasting. IEEE Access 7, 88058–88071 (2019)
18. C. Sun, J. Song, L. Li, P. Ju, Implementation of hybrid short-term load forecasting system with
analysis of temperature sensitivities. Soft. Comput. 12(7), 633–638 (2008)
19. M. Barman, N.B. Dev Choudhury, Season specific approach for short-term load forecasting
based on hybrid FA-SVM and similarity concept. Energy 174, 886–896 (2019). https://doi.org/
10.1016/j.energy.2019.03.010
20. M. Barman and N. B. D. Choudhury, A similarity based hybrid GWO-SVM method of power
system load forecasting for regional special event days in anomalous load situations in Assam,
India. Sustainable Cities and Society, p. 102311, 2020
21. Y. Hu et al., Short term electric load forecasting model and its verification for process industrial
enterprises based on hybrid GA-PSO-BPNN algorithm—A case study of papermaking process.
Energy 170, 1215–1227 (2019)
22. A.T. Eseye, M. Lehtonen, T. Tukia, S. Uimonen, R.J. Millar, Machine learning based integrated
feature selection approach for improved electricity demand forecasting in decentralized energy
systems. IEEE Access 7, 91463–91475 (2019)
23. L. Limei and H. Xuan, Study of electricity load forecasting based on multiple kernels learning
and weighted support vector regression machine. In 2017 29th Chinese control and decision
conference (CCDC), (2017), pp. 1421–1424
24. P. Gangwar, A. Mallick, S. Chakrabarti, S.N. Singh, Short-term forecasting-based network
reconfiguration for unbalanced distribution systems with distributed generators. IEEE Trans.
Indust. Inform 16(7), 4378–4389 (2019)
25. T. Ouyang, Y. He, H. Li, Z. Sun, S. Baek, Modeling and forecasting short-term power load with
copula model and deep belief network. IEEE Trans. Emerging Top. Comput. Intelligence 3(2),
127–136 (2019)
26. M. Rafiei, T. Niknam, J. Aghaei, M. Shafie-Khah, J.P.S. Catalão, Probabilistic load forecasting
using an improved wavelet neural network trained by generalized extreme learning machine.
IEEE Trans. Smart Grid 9(6), 6961–6971 (2018)
27. X. Tang, Y. Dai, T. Wang, Y. Chen, Short-term power load forecasting based on multi-layer
bidirectional recurrent neural network. IET Generation, Transmission & Distribution 13(17),
3847–3854 (2019)
28. M. Alipour, J. Aghaei, M. Norouzi, T. Niknam, S. Hashemi, and M. Lehtonen, A novel
electrical net-load forecasting model based on deep neural networks and wavelet transform
integration. Energy, p. 118106, 2020
29. H.-A. Li et al., Combined forecasting model of cloud computing resource load for energy-
efficient IoT system. IEEE Access 7, 149542–149553 (2019)
30. J. Ding, M. Wang, Z. Ping, D. Fu, and V. S. Vassiliadis, An integrated method based on
relevance vector machine for short-term load forecasting. Eur. J. Oper. Res. (2020)
31. M. El-Hendawi, Z. Wang, An ensemble method of full wavelet packet transform and neural
network for short term electrical load forecasting. Electr. Pow. Syst. Res. 182, 106265 (2020)
32. J. Moon, S. Jung, J. Rew, S. Rho, and E. Hwang, Combination of short-term load forecasting
models based on a stacking ensemble approach. Energy and Buildings, p. 109921, 2020
33. A.J. Amorim, T.A. Abreu, M.S. Tonelli-Neto, C.R. Minussi, A new formulation of multinodal
short-term load forecasting based on adaptive resonance theory with reverse training. Electr.
Pow. Syst. Res. 179, 106096 (2020)
34. X. Kong, C. Li, C. Wang, Y. Zhang, J. Zhang, Short-term electrical load forecasting based on
error correction using dynamic mode decomposition. Appl. Energy 261, 114368 (2020)
6 Introduction and Literature Review of the Application of Machine. . . 133

35. G.T. Ribeiro, V.C. Mariani, L. dos Santos Coelho, Enhanced ensemble structures using wavelet
neural networks applied to short-term load forecasting. Eng. Appl. Artif. Intel. 82, 272–281
(2019)
36. R. Wang, J. Wang, Y. Xu, A novel combined model based on hybrid optimization algorithm for
electrical load forecasting. Appl. Soft Comput. 82, 105548 (2019)
37. M. Bessani, J.A.D. Massignan, T.M.O. Santos, J.B.A. London Jr., C.D. Maciel, Multiple
households very short-term load forecasting using bayesian networks. Electr. Pow. Syst. Res.
189, 106733 (2020)
38. M. Mansoor, F. Grimaccia, S. Leva, and M. Mussetta, Comparison of echo state network and
feed-forward neural networks in electrical load forecasting for demand response programs.
Mathematics and Computers in Simulation, (2020)
39. G. Chitalia, M. Pipattanasomporn, V. Garg, S. Rahman, Robust short-term electrical load
forecasting framework for commercial buildings using deep recurrent neural networks. Appl.
Energy 278, 115410 (2020). https://doi.org/10.1016/j.apenergy.2020.115410
40. B. Dietrich, J. Walther, M. Weigold, E. Abele, Machine learning based very short term load
forecasting of machine tools. Appl. Energy 276, 115440 (2020)
41. Q. Huang, J. Li, and M. Zhu, An improved convolutional neural network with load range
discretization for probabilistic load forecasting. Energy, p. 117902, 2020
42. S. Ma, A hybrid deep meta-ensemble networks with application in electric utility industry load
forecasting. Inform. Sci. 544, 183–196
43. G. Hafeez, K.S. Alimgeer, I. Khan, Electric load forecasting based on deep learning and
optimized by heuristic algorithm in smart grid. Appl. Energy 269, 114915 (2020)
44. Y. Yang, W. Hong, S. Li, Deep ensemble learning based probabilistic load forecasting in smart
grids. Energy 189, 116324 (2019)
45. N. Zhang, Z. Li, X. Zou, S.M. Quiring, Comparison of three short-term load forecast models in
Southern California. Energy 189, 116358 (2019)
46. M. Malekizadeh, H. Karami, M. Karimi, A. Moshari, M.J. Sanjari, Short-term load forecast
using ensemble neuro-fuzzy model. Energy 196, 117127 (2020)
47. F. He, J. Zhou, L. Mo, K. Feng, G. Liu, Z. He, Day-ahead short-term load probability density
forecasting method with a decomposition-based quantile regression forest. Appl. Energy 262,
114396 (2020)
48. C. Fan, C. Ding, J. Zheng, L. Xiao, Z. Ai, Empirical mode decomposition based multi-objective
deep belief network for short-term power load forecasting. Neurocomputing 388, 110–123
(2020)
49. J. Bedi, D. Toshniwal, Energy load time-series forecast using decomposition and autoencoder
integrated memory network. Appl. Soft Comput. 93, 106390 (2020). https://doi.org/10.1016/j.
asoc.2020.106390
50. Y. Chu et al., Short-term metropolitan-scale electric load forecasting based on load decompo-
sition and ensemble algorithms. Energ. Buildings 225, 110343 (2020)
51. X. Ma and Y. Dong, An estimating combination method for interval forecasting of electrical
load time series. Expert Systems with Applications, p. 113498, 2020
52. K. Xie, H. Yi, G. Hu, L. Li, and Z. Fan, Short-term power load forecasting based on elman
neural network with particle swarm optimization. Neurocomputing, (2019)
53. P. Singh, P. Dwivedi, V. Kant, A hybrid method based on neural network and improved
environmental adaptation method using controlled Gaussian mutation with real parameter for
short-term load forecasting. Energy 174, 460–477 (2019)
54. S. Maldonado, A. González, S. Crone, Automatic time series analysis for electric load fore-
casting via support vector regression. Appl. Soft Comput. 83, 105616 (2019)
55. G. Sideratos, A. Ikonomopoulos, N.D. Hatziargyriou, A novel fuzzy-based ensemble model for
load forecasting using hybrid deep neural networks. Electr. Pow. Syst. Res. 178, 106025 (2020)
56. J. Kim, J. Moon, E. Hwang, P. Kang, Recurrent inception convolution neural network for multi
short-term load forecasting. Energ. Buildings 194, 328–341 (2019)
134 A. Moradzadeh et al.

57. D. Sakurai, Y. Fukuyama, T. Iizaka, T. Matsui, Daily peak load forecasting by artificial neural
network using differential evolutionary particle swarm optimization considering outliers. IFAC-
PapersOnLine 52(4), 389–394 (2019)
58. P. Singh, P. Dwivedi, A novel hybrid model based on neural network and multi-objective
optimization for effective load forecast. Energy 182, 606–622 (2019)
59. A. Moradzadeh, S. Zakeri, M. Shoaran, B. Mohammadi-Ivatloo, F. Mohamamdi, Short-term
load forecasting of microgrid via hybrid support vector regression and long short-term memory
algorithms. Sustainability (Switzerland) 12(17), 7076 (Aug. 2020). https://doi.org/10.3390/
su12177076
60. M. Mehrzadi et al., A deep learning method for short-term dynamic positioning load forecasting
in maritime microgrids. Applied Sciences 10(14), 4889 (2020)
61. S. Wang, S. Wang, D. Wang, Combined probability density model for medium term load
forecasting based on quantile regression and kernel density estimation. Energy Procedia 158,
6446–6451 (2019)
62. M. Talaat, M.A. Farahat, N. Mansour, A.Y. Hatata, Load forecasting based on grasshopper
optimization and a multilayer feed-forward neural network using regressive approach. Energy
196, 117087 (2020)
63. S.-M. Jung, S. Park, S.-W. Jung, E. Hwang, Monthly electric load forecasting using transfer
learning for Smart Cities. Sustainability 12(16), 6364 (2020)
64. J. Bedi, D. Toshniwal, Empirical mode decomposition based deep learning for electricity
demand forecasting. IEEE Access 6, 49144–49156 (2018)
65. A.I. Almazrouee, A.M. Almeshal, A.S. Almutairi, M.R. Alenezi, S.N. Alhajeri, Long-Term
Forecasting of Electrical Loads in Kuwait Using Prophet and Holt–Winters Models. Applied
Sciences 10(16), 5627 (2020)
66. M. Dong, J. Shi, Q. Shi, Multi-year long-term load forecast for area distribution feeders based
on selective sequence learning. Energy 206, 118209 (2020)
67. M.-R. Kazemzadeh, A. Amjadian, and T. Amraee, A hybrid data mining driven algorithm for
long term electric peak load and energy demand forecasting. Energy, p. 117948, 2020
68. D.A.G. Vieira, B.E. Silva, T.V. Menezes, A.C. Lisboa, Large scale spatial electric load
forecasting framework based on spatial convolution. International Journal of Electrical Power
& Energy Systems 117, 105582 (2020)
69. S. Kumar, S.K. Pal, R.P. Singh, A novel method based on extreme learning machine to predict
heating and cooling load through design and structural attributes. Energ. Buildings 176,
275–286 (2018). https://doi.org/10.1016/j.enbuild.2018.06.056
70. A. Mansour-Saatloo, A. Moradzadeh, B. Mohammadi-Ivatloo, A. Ahmadian, A. Elkamel,
Machine learning based PEVs load extraction and analysis. Electronics (Switzerland) 9(7),
1–15 (Jul. 2020). https://doi.org/10.3390/electronics9071150
71. W. Kong, Z. Y. Dong, Y. Jia, D. J. Hill, Y. Xu, and Y. Zhang, Short-term residential load
forecasting based on LSTM recurrent neural network. IEEE Transactions on Smart Grid,
(2019), doi: https://doi.org/10.1109/TSG.2017.2753802
72. A. Yu et al., Accurate fault location using deep belief network for optical Fronthaul networks in
5G and beyond. IEEE Access 7, 77932–77943 (2019). https://doi.org/10.1109/ACCESS.2019.
2921329
73. Z.A. Khan, S. Zubair, K. Imran, R. Ahmad, S.A. Butt, N.I. Chaudhary, A new users rating-trend
based collaborative Denoising auto-encoder for top-N recommender systems. IEEE Access 7,
141287–141310 (2019). https://doi.org/10.1109/ACCESS.2019.2940603
74. J. Han, S. Miao, Y. Li, W. Yang, H. Yin, A wind farm equivalent method based on multi-view
transfer clustering and stack sparse auto encoder. IEEE Access 8, 92827–92841 (2020). https://
doi.org/10.1109/ACCESS.2020.2993808
75. A. Moradzadeh and K. Pourhossein, Location of disk space variations in transformer winding
using convolutional neural networks. In 2019 54th International Universities Power Engineer-
ing Conference, UPEC 2019 - Proceedings, (2019), pp. 1–5, doi: https://doi.org/10.1109/
UPEC.2019.8893596
6 Introduction and Literature Review of the Application of Machine. . . 135

76. R.D. Rathor, A. Bharagava, Day ahead regional electrical load forecasting using ANFIS
techniques. J Instit. Engineers (India): Series B 101(5), 475–495 (2020). https://doi.org/10.
1007/s40031-020-00477-2
77. G. Dudek, Short-term load forecasting using random forests. Advances in Intelligent Systems
and Computing 323, 821–828 (2015)
78. N. Son, S. Yang, J. Na, Deep neural network and long short-term memory for electric power
load forecasting. Appl Sci (Switzerland) 10(18), 6489 (Sep. 2020). https://doi.org/10.3390/
APP10186489
79. M. Tan, S. Yuan, S. Li, Y. Su, H. Li, F.H. He, Ultra-short-term industrial power demand
forecasting using LSTM based hybrid ensemble learning. IEEE Trans. Power Syst. 35(4),
2937–2948 (Jul. 2020). https://doi.org/10.1109/TPWRS.2019.2963109
80. S. Pei, H. Qin, L. Yao, Y. Liu, C. Wang, J. Zhou, Multi-step ahead short-term load forecasting
using hybrid feature selection and improved long short-term memory network. Energies 13(6),
4121 (Aug. 2020). https://doi.org/10.3390/en13164121
81. A. Moradzadeh, A. Mansour-Saatloo, B. Mohammadi-Ivatloo, A. Anvari-Moghaddam, Perfor-
mance evaluation of two machine learning techniques in heating and cooling loads forecasting
of residential buildings. Applied Sciences (Switzerland) 10(11), 3829 (2020). https://doi.org/10.
3390/app10113829
82. S. Tzafestas, E. Tzafestas, Computational intelligence techniques for short-term electric load
forecasting. Journal of Intelligent and Robotic Systems: Theory and Applications 31(1–3), 7–68
(2001). https://doi.org/10.1023/A:1012402930055
Chapter 7
A Survey of Recent Particle Swarm
Optimization (PSO)-Based Clustering
Approaches to Energy Efficiency
in Wireless Sensor Networks

Emrah Hancer

7.1 Introduction

Thanks to the advanced developments in wireless communication and

microelectromechanical system (MEMS), wireless sensor networks (WSNs) have
gained wide popularity for various fields, such as disaster management, industrial
automation, military reconnaissance, smart buildings, etc. [1]. WSNs can therefore
establish the connection between the environment, the computational world, and the
human society. Typically, WSNs are built of low-price and low-energy sensor
nodes. The positions of nodes do not need to be specifically determined, i.e., their
positions can be randomly indicated. In a network, all the sensor nodes are able to
sense, process, and communicate. Through these characteristics, data acquisition and
distribution processes are carried out. The network is operated by a base station or
sink node that may be located close to the network framework.
The overall goal of all sensor nodes is to transfer data to the base station or sink
node which is the final destination for the acquired data. After the data transmission
process, the sink node carries out some required processes (e.g., data transfer to a
task manager over the Internet or satellite network). The main distinction between
standard and sink nodes is energy usage. While a sink node owns an unlimited power
battery, standard sensor nodes own in-rechargeable restricted power batteries.
Accordingly, sensor nodes should carefully consider energy consumption as soon
as possible.
The most challenging issue in WSNs to be taken into consideration is the
consumption of energy since it affects the sensor node lifetime and thereby the

E. Hancer (*)
Department of Software Engineering, Mehmet Akif Ersoy University, Burdur, Turkey
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 137
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_7
138 E. Hancer

whole network lifetime. There also exists a trade-off between the energy consump-
tion and the number of sensor nodes. When the number of sensors within the
network exceeds the capacity, the usual direct routing process requires higher energy
consumption and so inversely affects the network lifetime. The fundamental tech-
niques to enhance the network lifetime are as follows [2]: (1) energy efficiency
scheduling, (2) energy efficiency node transmission power tuning, (3) energy effi-
ciency routing, and (4) energy efficiency clustering. Among such techniques, clus-
tering and routing have been widely used for WSNs since bringing several
advantages, such as scalability, efficient communication, and fault tolerance [3]. In
this chapter, the motivation will be on clustering techniques.
Clustering helps WSNs to manage efficient energy usage by reducing the number
of sensor nodes that are far from the sink node and dividing current consumption
evenly among sensor nodes. Each separated group of sensor nodes is represented by
a cluster head node that gathers data within the group and then transfers it to the sink
node. Through this process, it is possible to reduce data transmission traffic, enhance
resource allocation, and improve bandwidth reusability. The most well-known
clustering approaches to address energy efficiency in WSNs are LEACH [4],
LEACH-C [5], SEP [6], EECS [7], and PEGASIS [8]. Although such clustering
approaches have obtained promising results, energy efficiency is still an open issue.
Various surveys have been published in the literature to outline the profile of
clustering approaches proposed for WSNs. Abbasi and Younis [9] described some
important conventional clustering approaches and then classified them in terms of
the convergence time. Mamalis et al. [10] considered clustering approaches in the
case of probability. While the most popular clustering approaches such as LEACH,
HEED, and EEHC were evaluated in the category of probabilistic approaches, other
approaches were categorized as weight-based, graph-based, and biologically
inspired. Liu [11] reviewed 16 well-known clustering methods and introduced a
taxonomy of the corresponding methods in the basis of cluster characteristics. Jiang
et al. [12] reviewed some popular clustering approaches in terms of selectivity, count
variability, and existence and then introduced a comparative study of the reviewed
approaches. Kumarawadu et al. [13] considered clustering approaches in four
groups: probabilistic, biologically inspired, neighborhood-based, and identity-
based. The criteria to analyze and compare approaches in this work were energy
efficiency, clock synchronization, and load balancing. Deosarkar et al. [14] intro-
duced a survey on cluster head selection, where all the related works were catego-
rized into four categories: deterministic, combined, adaptive, and hybrid. Aslam
et al. [15] considered LEACH and its variants and introduced a comparative study of
the related approaches. Afsar and Tayarani-N [3] considered a large number of
clustering approaches from different perspectives, such as load balancing, fault
tolerance, scalability, and connectivity. Sambo et al. [16] investigated
computational-based clustering methods in terms of 10 criteria, including scalability,
radio model, data aggregation, energy efficiency, and nature. Other related surveys
can be found in [17–19].
When considering the aforementioned works, it is not possible to find a work
except [16] that specifically focuses on particle swarm optimization (PSO)-based
7 A Survey of Recent Particle Swarm Optimization (PSO)-Based Clustering. . . 139

clustering approaches. Although [16] considered PSO-based approaches, most of the

works were missing. This issue motivated us to perform this comprehensive survey
on PSO-based clustering approaches.
The rest of the chapter is as follows: We first provide a general background on the
WSNs framework, PSO algorithm, and conventional clustering approaches. We then
review the related works with discussions from different perspectives. Finally, we
conclude the chapter with current drawbacks and future trends.

7.2 Background

In this section, we will first explain the overall structure of WSNs and particle swarm
optimization. We then define the problem.

7.2.1 Structure of WSNs

Sensitive cells in a sensor node can transform measured physical values into an
electrical signal to be used in many applications, such as temperature, noise levels,
and pressure. As shown in Fig. 7.1, a typical sensor node includes the following
components: (1) Power supply, also called battery unit, is built of a tiny battery and
AC-DC converter, which provides suitable voltages for the electronic circuits of a
sensor node. (2) Physical sensing device is used to measure or observe physical
events (e.g., sounds, temperature, earthquakes) in the environment. (3) AD-DC
converter transforms analog signals obtained by the sensing device to digital signals.
(4) RF communication unit transmits data to another sensor node through radio
communication. (5) Processor unit is a microprocessor system that controls and
monitors the whole activities of the sensor node. Sensor nodes are located in the
target area to gather data concerning related physical events and transmit such data to

Processor
RF radio
Sensor ADC
communication
Storage

Power supply

Fig. 7.1 Overall structure of a sensor node [1]

140 E. Hancer

the sink node. Accordingly, they require energy to properly carry out the tasks of
WSN. The communication task is treated as the main source of energy consumption.
In addition to communication task, there also exists energy loss inactive states due to
the following reasons [1]: (1) Idle listen: As it is possible to receive data from its
neighbors to a sensor, it must listen the messages over the medium. (2) Overhearing:
When a data is sent from a sensor node, all of its neighbors receive this data even
though it is sent to only one of them. (3) Interference: A sensor node receives a
packet between transmitting and interference ranges but cannot decode the packet.
(4) Collision: When a collusion exists during the transmission, the energy arranged
for this task is lost. (5) Control packet overhead: A small number of control packets
should be used to support transmission tasks. In summary, due to their small
batteries, sensor nodes need to address limited energy consumption. This case
become more harmful, especially when manual recharge is not possible for the
sensor nodes which are generally located in unreachable area. Accordingly, energy
efficiency has been considered as an important parameter in the design of WSNs to
extend the lifetime of network.

7.2.2 Particle Swarm Optimization

Swarm intelligence is the discipline of evolutionary computation in artificial intel-

ligence that investigates the collective behavior of decentralized and self-organized
systems. In particular, the source of collective behaviors is the interactions of
individuals within the swarm and with the environment. Ants and termites, schools
of fish, bird flocks, and honeybees are representative examples of the swarm
intelligence. It is not possible to consider all kinds of individuals in swarm intelli-
gence. To be considered as swarm intelligence, a swarm of individuals should have
the following properties:
1. Self-organization: This is an important characteristic of natural systems, which
means that it is possible to carry out required processes without the requirement
of any centralized authority. With the help of self-organization, swarm agents can
build the best work coordination with each other and so pursue their task at high
speed and fault tolerance. The main characteristics of self-organization are the
following: a) Positive feedback: More agents take part in the same work to
promote smart solutions. b) Negative feedback: Agents avoid getting involved
in the same state to counterbalance the effect of positive feedback. c) Fluctuation:
Randomness allows agents to occur random changes in the system. d) Multiple
interactions: Agents tend to interact with each other and so learn from each other
to create intelligent behaviors within the swarm.
2. Division of labor: Thanks to the cooperation of specialized agents, it is possible to
simultaneously perform a variety of tasks in the swarm rather than carrying out
sequential tasks. Accordingly, there exists a more effective and efficient task
management system without a doubt.
7 A Survey of Recent Particle Swarm Optimization (PSO)-Based Clustering. . . 141

Since first introduced in 1995, swarm intelligence algorithms have gained over-
whelming interest by researchers and so have been applied to a variety of fields from
numeric optimization to machine learning to power systems [20]. The earliest swarm
intelligence algorithm, particle swarm optimization (PSO) [21] has received a
growing-up interest by the researchers, resulting in a variety of applications in
various fields. PSO mimics the behaviors of social groups, like fish and birds. The
algorithm seeks the optimal solution through agents, called particles, each of which
is represented by its position in the possible solution space. The algorithm iteratively
updates the particle positions using the information of its local best and the swarm
best position in a randomized weighted manner by Eqs. 7.1 and 7.2. After the
process of updating the particle positions, the swarm best position is redefined for
the next generations. One of the advantages of PSO over evolutionary algorithms,
like genetic algorithms (GA) and evolutionary strategies, is a smaller number of
parameters to be tuned by a user. Furthermore, the algorithm does not tend to face
stagnation problems as fluently as GA. Besides a variety of successful applications in
various fields from numerical problems to energy systems [22, 23], PSO has also
been frequently used in WSNs.

vi ðt þ 1Þ ¼ wvi ðt Þ þ c1 r 1 ðbxi ðt Þ  xi ðt ÞÞ þ c2 r 2 ðgbest ðt Þ  xi ðt ÞÞ ð7:1Þ

x i ð t þ 1Þ ¼ x i ð t Þ þ v i ð t þ 1Þ ð7:2Þ

where xi(t) = {xi,1, xi,2, . . ., xi,D} is the position of the ith particle at time t, b
xi ðtÞ is its
best solution as of time t, gbest(t) represents the best position in the swarm as of time
t, w is the inertial weight which keeps the particle moving in the same direction it was
originally heading and is generally selected between 0.8 and 1.2, c1 and c2 are
cognitive and social coefficients which are generally set to 2, and r1 and r2 are
randomly generated numbers between 0 and 1.

7.2.3 Problem Definition

Efficient energy consumption is a crucial issue in WSNs to enhance the lifetime of a

network. The most intensive task for energy usage is the data transfer from sensor
nodes to the sink node. The required energy for the transmission process varies
according to the distance between the sink and sensor nodes. To be specific, energy
consumption exponentially increases proportionally to the transmission distance,
and so the lifetime of WSNs largely depends on the efficient energy management of
the transmission process. Clustering is one of the most widely applied techniques to
efficiently manage the transmission process from sensor nodes to the sink node. By
clustering, a set of sensor nodes within a WSN are grouped into clusters based on
predefined similarity metrics (e.g., distance). Each cluster is represented by its
cluster head, which is responsible for acquiring data from sensor nodes within its
142 E. Hancer

Fig. 7.2 Transmission process in WSNs

cluster and then forward the acquired data to the sink node as shown in Fig. 7.2. A
cluster head that is assigned for a long duration prematurely exhausts its power
supply. The selection of a cluster head therefore plays a crucial role in the perfor-
mance and longevity of the network. The well-known conventional clustering
approaches are presented as follows:
1. LEACH [4]: The selection of a cluster head is very crucial in LEACH. If the
cluster head is optimally selected, it is possible to increase the energy efficiency
and address data transmission problems. In the initial stage, all the sensors within
the cluster can be chosen as a cluster head, but LEACH introduces some
additional definitions for the cluster head selection. The protocol operation is
divided into rounds. The processes such as detection of all sensors, selection of
cluster head, and data transmission are carried out in a certain plan in each round.
Each sensor node is selected at least once due to the statistical formulation used
for the cluster head selection. This is because a sensor acting as a cluster head
consumes more energy, and so each sensor node must be selected at least once as
a cluster head to evenly divide energy consumption among sensor nodes.
2. LEACH-C [5]: Since the number of cluster heads is not clear and low-energy
sensor nodes are possible to be selected as a cluster head in LEACH, the
researchers developed LEACH-C, which carries out a two-layer cluster head
selection. In detail, all sensor nodes inform the sink node about their position
and energy level. Using this information, a threshold value is determined, and the
sensor nodes which are smaller than the determined threshold value are not
selected as the cluster head. LEACH-C outperforms LEACH in terms of the
cluster quality.
3. EECS [7]: The algorithm shows similar characteristics with LEACH. All the
clusters are combined to only one cluster head, and this cluster head directly
7 A Survey of Recent Particle Swarm Optimization (PSO)-Based Clustering. . . 143

makes a connection with the sink node. To detect the position of sensor nodes, the
sink node sends a “hello” packet, and the response time determines the approx-
imate position of sensor nodes. The selection of a cluster head is carried out in a
probabilistic manner for each cluster. The energy levels of sensor nodes are
evaluated using the COMPETE-HEAD-MSG message, and then the node with
the highest energy level is selected as the cluster head.
4. PEGASIS [8]: The algorithm is a hierarchical protocol which is based on a greedy
algorithm and a chain-based approach. Sensor nodes come together to build a
chain. If any sensor node dies within the chain, the chain is then rebuilt by
skipping the dead node. The requests transmitted from other nodes to the leader
node are then forwarded to the sink node. Using PEGASIS protocol in a network
with many sensor nodes may cause delays in data transmission. To alleviate this
drawback, one of the ways is to use multiple leader nodes in the network. When
compared to LEACH, PEGASIS was proved to be much more efficient.
5. H-PEGASIS [24]: The algorithm has been developed to address time delays in
PEGASIS due to data collisions. In H-PEGASIS, simultaneous data transfer is
introduced to prevent data collisions using CDMA encoding and spatially sepa-
rated sensors. A chain-based sensor based on CDMA, which acts like a tree
structure, transmits data from sublayers to the sink node. Accordingly, delays are
reduced while data is simultaneously transmitted.
Other popular conventional clustering approaches to address energy efficiency in
WSNs are SEP [6], DEEC [25], HEED [26], and EEHC [27]. More information
concerning conventional clustering approaches proposed for WSNs can be found
in [28].

7.3 PSO-Based Approaches

Clustering in WSNs is an NP-hard problem due to the selection of m optimal cluster

heads among n sensor nodes leading m  n possibilities. Thanks to their effective
search characteristics, swarm intelligence algorithms have strongly been treated as
good NP-hard problem-solving tools. Due to its historical and successful back-
ground, PSO is maybe the most applied swarm intelligence algorithm in WSNs.
The considered PSO-based clustering approaches in this work are as follows:
PSO-C [29]: The approach owns a centralized mechanism that is operated at the
sink node or base station. In the initial stage, the base station receives information
from sensor nodes concerning their energy level and locations. The base station then
calculates the average energy level of all nodes using the received information. The
nodes with sufficient energy levels are determined as candidate cluster heads. In the
second stage, PSO is applied to determine K best cluster heads among candidate
cluster heads by optimizing the objective function which is the weighted combina-
tion of the intra-cluster distance and the energy efficiency. After the determination of
cluster heads and cluster members, each cluster head builds a TDMA schedule to
144 E. Hancer

prevent collisions during data transmission. In this scheme, sensor nodes are only
activated during their transmission time to enhance the energy efficiency within the
network. The cluster heads gather data from its allocated sensor nodes within its
cluster and then forward the gathered data to the base station. According to the
results, PSO-C is far superior to LEACH and LEACH-C.
PSO-HC [30]: The approach is another centralized protocol that aims to enhance
the network lifetime and maximize the network scalability by minimizing the
average energy consumption and building two-hop communication within the clus-
ter. In the first stage, each sensor node sends a hello packet with its ID. When a
sensor node receives this packet, it updates its neighbor table with the RRSI value
included in the packet. After all sensors determine their neighbors, each sensor node
transmits the data including ID, residual energy, and its neighbor table to the base
station using the flooding method. The base station then calculates the average
energy level of all sensor nodes based on the received data. If any sensor in the
network owns a higher energy level than the average, it is determined as a candidate
cluster head. In the second stage, PSO is applied to select K optimal cluster heads
among the candidate ones. Each particle is represented as a sequence of cluster head
IDs. After obtaining the set of cluster heads from the particle, the base station
constructs two-tier clusters to improve the network scalability and lifetime. The
first tear clusters are constructed by assigning each sensor node to a cluster node
based on the RSSI value. The second tear clusters are constructed using all
non-clustered sensor nodes from the first tier. According to the results, PSO-HC is
more efficient than LEACH, LEACH-C, and PSO-C.
PSO-SD [31]: The approach is not based on a centralized mechanism, which
makes it a semi-distributed approach. The approach tries to find the locations of
cluster heads by optimizing the following objectives in a weighted manner: the intra-
cluster distance, the residual energy, the node degree, and the head count. Each
particle represents the locations of cluster heads. The impact of the packet
retransmission size along the estimated path to the cluster head is also considered
in the approach. According to the results, it performs better than PSO-C and
LEACH-C.
PECC [32]: The approach uses a centralized mechanism to form clusters. The
sink node sends info collection messages to sensor nodes. After receiving messages,
sensor nodes transmit information concerning the location, id, energy level, and
energy loss to the sink node. Then the sink node forms clusters by optimizing the
average distance and the average energy level in a weighted manner using PSO. For
each cluster, a cluster head is selected by optimizing the within-cluster distance and
the number of sensor nodes in a weighted manner using PSO. Finally, a multi-hop
communication protocol is applied to carry out the data transmission process from
sensor nodes to the cluster head and from the cluster head to the sink node. Data is
gathered by the cluster head in each cluster to save the residual energy. If the distance
from the cluster head to the sink node does not exceed the predefined threshold
value, the cluster head transmits data to the sink node through the single-hop
transmission. Otherwise, the cluster head seeks for new hop based on the residual
energy and distance.
7 A Survey of Recent Particle Swarm Optimization (PSO)-Based Clustering. . . 145

EMPS [33]: Like PSO-C, PSO-HC, and PECC, the approach is based on a
centralized mechanism. In this approach, sensor nodes share information with the
sink node concerning their position and energy level. Then PSO first splits the
network into subregions. Each particle is represented as a region boundary, includ-
ing the (x,y) coordinates of the point line and the angles of the point line to the X and
Y axes. After the split process of subregions, a cluster head is determined for each
region based on the distance from the sensor node to the center of gravity and the
residual energy. EMPS consists of three characteristic messages to carry out the
transmission process: a) The hello packet detects the cluster region that transmits
data to the sink node. b) The message-s packet transmits data to the sink node. c) The
message-h packet transmits data to the cluster head. From the results, it can be
indicated that it efficiently manages energy usage compared to the conventional
clustering approaches.
PSO-ECHS [34]: The approach aims to select the optimal cluster heads on the
network to improve the network efficiency using PSO. Each particle represents the
possible locations of the cluster heads in the network space. The objective function
handles the following objectives in a weighted manner: the average intra-cluster
distance and the average sink distance. Different from the approaches where sensor
nodes are assigned to a cluster based on the distance to the cluster head, the approach
assigns sensor nodes to the cluster heads based on a weighted function which
comprises of the residual energy, the distance from the cluster head to the sink
node, the distance from the sensor node to the cluster head, and the degree of the
cluster head. According to the various experiments on a variety of scenarios, the
approach outperforms a variety of well-known approaches, such as LEACH,
LEACH-C, and PSO-C.
PSO-HSA [35]: The approach follows a two-way hybridized methodology to
manage energy usage in the network. In the first stage, clustering is applied using
PSO, and 10% of the available sensor nodes are selected as the cluster heads. In the
second phase, harmony search [36] is applied to carry out the transmission process.
A gateway node is positioned between the sink node and the cluster head. The
gateway nodes are responsible for improving energy efficiency during the data
transmission process to the sink node. The cluster head first transmits the obtained
data to the gateway node based on the distance between the cluster head and the sink
node. The gateway nodes are selected based on the distance with the cluster head
using harmony search.
SCE-PSO [37]: The approach assumes that the cluster heads (called gateways)
and sensor nodes are randomly located in a given area. Any sensor node can connect
to any cluster head if the sensor node is within the communication range of the
cluster head. If the number of sensor nodes within a cluster exceeds the cluster
capacity, its load is divided between other clusters. A position of each particle
represents the assigned cluster of the corresponding sensor node. Thus, the dimen-
sionality of each particle is equal to the number of sensor nodes in the network. The
objective function consists of three fundamental components: the load of the cluster
heads, the intra-cluster distance, and the number of heavily loaded cluster heads. It
can be extracted from the objective function that the assignment process of sensor
146 E. Hancer

nodes depends on the cluster head load. This enhances the lifetime of the cluster
head and so maximizes the whole network lifetime. Once particles are evaluated
using the objective function, they are sorted according to the objective value and
partitioned into groups, named complexes. Then each complex is individually
evaluated using PSO and is updated using the newly generated particles.
PUDCRP [38]: The cluster heads closer to the sink node are more likely to
participate in data transmission processes in multi-hop routing. This leads to the
end of nodes near the sink node area, referred to as the hotspot problem. Different
from the aforementioned approaches, the approach considers the hotspot problem as
well as grouping sensor nodes. To alleviate the hotspot problem, the approach
divides the network into different-sized circles according to the distribution of sensor
nodes. The determination of cluster heads is then considered as a multi-objective
problem. It should be notified that the optimal number of cluster heads is dynami-
cally determined without the requirement of any user-specified parameters. From the
results, it can be revealed that PUDCRP performs better than a variety of recently
introduced approaches, including PSO-SD and PSO-ECHS.
GA-PSO [39]: The approach involves two fundamental stages. In the first stage,
the cluster heads are determined using GA based on the weighted objective function
which considers the distance from nodes to the cluster head, the distance from the
cluster head to the base station, the distance from a non-cluster head to the cluster
head, and the total system energy. After the determination of the cluster heads, the
second stage is carried out. In this stage, PSO is applied to manage the routing
process. The relay nodes are put into the transmission traffic to increase the effi-
ciency of the cluster head. In particular, the cluster head first search for the closest
relay nodes or base station to transfer the aggregated data. The method performed
better than LEACH variants, but it did not compare with recent PSO-based cluster-
ing approaches.
PSO-UFC [40]: The approach first determines the cluster heads by optimizing the
following objectives in a weighted manner: (1) average intra-cluster distance,
(2) average inter-cluster distance, and (3) residual energy. After the selection of
the cluster heads, clusters are formulated in a such way that the clusters near to base
station have a smaller size to keep their energy for the inter-cluster relay transmission
process. Then a multi-hop routing three is constructed among the selected cluster
head. To deal with the fault tolerance task, a surrogate cluster head is determined for
each cluster head. According to the experiments, PSO-UFC outperforms various
approaches such as LEACH, PSO-C, and EBUC.
PSO-ECSM [41]: The sink node is placed at the middle of the WSN for collecting
data from nodes. In other words, the approach is built on a heterogeneous model. The
approach selects the cluster heads by considering the following factors: residual
energy, node degree, energy distance, average energy, and energy consumption. The
approach also resolves the data traffic in a multi-hop network by applying sink
mobility. According to a number of experiments, the approach performs better than a
variety of clustering approaches in terms of stability, robustness, and network
lifetime.
7 A Survey of Recent Particle Swarm Optimization (PSO)-Based Clustering. . . 147

Table 7.1 Comparison of PSO-based approaches

Data Multi Multi
Method agg. Efficiency Scalability Nature Network hop path
PSO-C Yes Average Low Centralized Homogeneous No No
[29]
PSO-HC – Average High Centralized Homogeneous Yes –
[30]
PSO-SD Yes High High Distributed Homogeneous – Yes
[31]
PECC [32] Yes Average Medium Distributed Homogeneous Yes –
EMPS Yes High High Centralized Homogeneous – Yes
[33]
PSO- No High High Centralized Homogeneous – –
ECHS [34]
PSO-HSA – Average Medium Centralized Homogeneous No –
[35]
SCE-PSO – Average High Centralized Homogeneous No –
[37]
PUDCRP – High High Centralized Homogeneous Yes Yes
[38]
GA-PSO – Average Medium Centralized Homogeneous – –
[39]
PSO-UFC – Average Medium Centralized Homogeneous Yes Yes
[40]
PSO- Yes High High Centralized Heterogeneous Yes –
ECSM
[41]

Table 7.1 summarizes the presented PSO-based clustering approaches from

different perspectives. According to Table 7.1, a centralized mechanism is much
more preferred than a distributed mechanism by PSO-based approaches. It can also
be extracted from Table 7.1 that all the PSO-based approaches can improve the
network lifetime, thanks to their efficient energy management. Among the compared
approaches, PSO-SD, EMPS, PSO-ECHS, PUDCRP, and PSO-ECSM are high-
level energy-saving approaches. Furthermore, the scalability of all the approaches
can be treated as sufficient except PSO-C.

7.4 Conclusions

In this chapter, we introduced a survey of recently proposed PSO-based clustering

approaches to address energy efficiency in WSNs. Among a variety of works in this
field, there exists no such work in the literature that tries to reflect the profile of recent
PSO-based approaches proposed to deal with energy consumption in WSNs.
According to the results, PSO-based approaches can better carry out energy
148 E. Hancer

management of the whole network than conventional clustering protocols. In our

opinion, the following works can be considered by researchers in the future. First,
hybridized frameworks may be an option to improve the efficiency of PSO-based
approaches. Second, it is not possible to find a comprehensive comparative analysis
of PSO-based approaches in the literature from different perspectives. Third, most of
the PSO-based approaches are based on a centralized mechanism, i.e., the impact of
distributed mechanism on energy efficiency may be investigated deeply in the future.

References

1. M. A. H. Hussein, Energy Efficiency in Wireless Sensor Networks. Master’s thesis, (The

Graduate School of Natural and Applied Sciences of Cankaya University, 2015)
2. S. Mahfoudh, Energy Efficiency in Wireless Ad Hoc and Sensor Networks: Routing, Node
Activity Scheduling and Cross-Layering. PhD theses, (Universite Pierre et Marie Curie - Paris
VI, 2010)
3. M. Mehdi Afsar, T.-N. Mohammad-H, Clustering in sensor networks: A literature
survey. J. Netw. Comput. Appl. 46, 198–226 (2014)
4. W.B. Heinzelman, A.P. Chandrakasan, H. Balakrishnan, An application-specific protocol
architecture for wireless microsensor networks. IEEE Trans. Wirel. Commun. 1(4), 660–670
(2012)
5. M. Tripathi, M. S. Gaur, V. Laxmi, R. B. Battula, Energy efficient LEACH-C protocol for
Wireless Sensor Network. In Proc. Third International Conference on Computational Intelli-
gence and Information Technology (CIIT), 2013
6. G. Smaragdakis, I. Matta, S.E.P. Bestavros, A stable election protocol for clustered heteroge-
neous wireless sensor networks. In Proc. Second International Workshop on Sensor and Actor
Network Protocols and Applications (SANPA), 2004
7. V. Saranya, S. Shankar, G.R. Kanagachidambaresan, Energy efficient clustering scheme
(EECS) for wireless sensor network with Mobile sink. Wireless Personal Communications:
An Int. J. 100(4), 1553–1567 (2018)
8. S. Lindsey, C.S. Raghavendra, PEGASIS: power-efficient gathering in sensor information
systems. In Proc. IEEE Aerospace Conference, (2002)
9. A. Abbasi, M. Younis, A survey on clustering algorithms for wireless sensor networks. Comput.
Commun. 30(14), 2826–2841 (2007)
10. B. Mamalis, D. Gavalas, C. Konstantopoulos, G. Pantziou, Clustering in Wireless Sensor
Networks. RFID and Sensor Networks: Architectures, Protocols, Security, and Integrations
(2009), pp. 323–354
11. X. Liu, A survey on clustering routing protocols in wireless sensor networks. Sensors 12(8),
11113–11153 (2012)
12. C. Jiang, D. Yuan, Y. Zhao, Towards clustering algorithms in wireless sensor networks-a
survey. In Proc. IEEE Wireless Communications and Networking Conference, (2009)
13. P. Kumarawadu, D.J. Dechene, M. Luccini, A. Sauer, Algorithms for node clustering in
wireless sensor networks: a survey. In Proc. 4th International Conference on Information
and Automation for Sustainability, (2008)
14. B.P. Deosarkar, N.S. Yadav, R.P. Yadav, Cluster head selection in clustering algorithms for
wireless sensor networks: a survey. In Proc. International Conference on Computing, Commu-
nication and Networking, (2008)
15. M. Aslam, N. Javaid, A. Rahim, U. Nazir, A. Bibi, Z. Khan, A survey of extended LEACH-
based clustering routing protocols for wireless sensor networks. In Proc. 9th IEEE International
Conference on Embedded Software and Systems, (2012)
7 A Survey of Recent Particle Swarm Optimization (PSO)-Based Clustering. . . 149

16. D. Wohwe Sambo, B.O. Yenke, A. Forster, P. Dayang, Optimized clustering algorithms for
large wireless sensor networks: A review. Sensors 19, 1–27 (2019)
17. R.V. Kulkarni, A. Förster, G.K. Venayagamoorthy, Computational intelligence in wireless
sensor networks: A survey. IEEE Communications Surveys Tutorials 13(1), 68–96 (2011)
18. P. Kumari, M.P. Singh, P. Kumar, Survey of clustering algorithms using fuzzy logic in wireless
sensor network. In Proc. International Conference on Energy Efficient Technologies for
Sustainability, (2013)
19. S. Sirsikar, K. Wankhede, Comparison of clustering algorithms to design new clustering
approach. In Proc. 4th International Conference on Advances in Computing, Communication
and Control, (2015)
20. E. Hancer, D. Karaboga, A comprehensive survey of traditional, merge-split and evolutionary
approaches proposed for determination of cluster number. Swarm and Evolutionary Computa-
tion 32, 49–67 (2017)
21. J. Kennedy, R. Eberhart, Particle swarm optimization. In Proc. International Conference on
Neural Networks, (1995)
22. P.H. Mahmoud, N.-H. Morteza, M.-I. Behnam, S. Heresh, A hybrid genetic particle swarm
optimization for distributed generation allocation in power distribution networks. Energy 209,
118218 (2020)
23. N.-H. Morteza, S. Madadi, P.H. Mahmoud, M.-I. Behnam, Optimal distributed generation
allocation using quantum inspired particle swarm optimization, in Quantum Computing: An
Environment for Intelligent Large Scale Real Application, (Springer, Cham, 2018),
pp. 419–432
24. S. Lindsey, C. Raghavendra, K.M. Sivalingam, Data gathering algorithms in sensor networks
using energy metrics. IEEE Trans. Parallel Distributed Syst. 13(9), 924–935 (2002)
25. L. Qing, Q. Zhu, M. Wang, Design of a distributed energy-efficient clustering algorithm for
heterogeneous wireless sensor networks. Comput. Commun. 29(12), 2230–2237 (2006)
26. O. Younis, S. Fahmy, HEED: A hybrid, energy-efficient, distributed clustering approach for ad
hoc sensor networks. IEEE Trans. Mob. Comput. 3(4), 366–379 (2004)
27. D. Kumar, T.C. Aseri, R.B. Patel, EEHC: Energy efficient heterogeneous clustered scheme for
wireless sensor networks. Comput. Commun. 32(4), 662–667 (2009)
28. B. Jan, H. Farman, H. Javed, B. Montrucchio, M. Khan, S. Ali, Energy efficient hierarchical
clustering approaches in wireless sensor networks: A survey. Wirel. Commun. Mob. Comput.
6457942 (2017)
29. N.M.A. Latiff, T.C. Simonides, B.S. Sharif, Energy-aware clustering for wireless sensor
networks using particle swarm optimization. In Proc. 18th IEEE International Symposium on
Personal, Indoor and Mobile Radio Communications, (2007)
30. R.S. Elhabyan, M.C.E. Yagoub, PSO-HC: Particle swarm optimization protocol for hierarchical
clustering in wireless sensor networks. In Proc. 10th IEEE International Conference on
Collaborative Computing: Networking, Applications and Worksharing, (2014)
31. B. Singh, D.K. Lobiyal, A novel energy-aware cluster head selection based on particle swarm
optimization for wireless sensor networks. HCIS 2(1), 2–13 (2012)
32. C. Vimalarani, R. Subramanian, S.N. Sivanandam, An enhanced PSO-based clustering energy
optimization algorithm for wireless sensor network. Scientific World J. (2016)
33. J. Wang, Y. Cao, B. Li, H. Kim, S. Lee, Particle swarm optimization based clustering algorithm
with mobile sink for WSNs. Futur. Gener. Comput. Syst. 76, 452–457 (2017)
34. P.C.S. Rao, P.K. Jana, H. Banka, A particle swarm optimization based energy efficient cluster
head selection algorithm for wireless sensor networks. Wirel. Netw 23(7), 2005–2020 (2017)
35. V. Anand, S. Pandey, Particle swarm optimization and harmony search based clustering and
routing in wireless sensor networks. Int. J. Comput. Intelligence Syst. 10(1), 1252–1262 (2017)
36. Z. Woo, J. Hoon, G.V. Loganathan, A new heuristic optimization algorithm: Harmony search.
Simulation 76(2), 60–68 (2001)
37. D.R. Edla, M.C. Kongara, R. Cheruku, SCE-PSO based clustering approach for load balancing
of gateways in wireless sensor networks. Wirel. Netw 25(3), 1067–1081 (2019)
150 E. Hancer

38. D. Ruan, J. Huang, A PSO-based uneven dynamic clustering multi-hop routing protocol for
wireless sensor networks. Sensor Networks 19, 1835 (2019)
39. D. Anand, S. Pandey, New approach of GA-PSO based clustering and routing in wireless sensor
networks. Int. J. Commun. Syst. 33, e4571 (2020)
40. T. Kaur, D. Kumar, Particle swarm optimization-based unequal and fault tolerant clustering
protocol for wireless sensor networks. IEEE Sensors J. 18, 4614–4622 (2018)
41. B.M. Sahoo, T. Amgoth, H.M. Pandey, Particle swarm optimization based energy efficient
clustering and sink mobility in heterogeneous wireless sensor network. Ad Hoc Netw. 106,
102237 (2020)
Chapter 8
Clustering in Power Systems Using
Innovative Machine Learning/Deep
Learning Methods

Mohammad Hossein Rezaeian Koochi, Mohammad Hasan Hemmatpour,

and Payman Dehghanian

8.1 Introduction

Clustering is the task of dividing a dataset into groups with similar characteristics.
These datasets can be made of different types of data such as a set of time-series
variations of variables or a set of single data points placed in a two- or more-
dimensional space. The choice of the mechanism for assessing such similarities
depends on the type of the data. For the case of single data points placed in an p-
dimensional space, the Euclidean distance could be a promising approach. For
example, Fig. 8.1 shows that a typical set of data points in a 2-D plane can be
clustered according to the Euclidean distances between the data points. However, in
order to assess the similarity of time-series variations, other criteria and
preprocessing tasks rather than Euclidian distance, such as calculating correlation
coefficients, applying dynamic time warping (DTW), or using feature extraction
techniques such as principle component analysis (PCA) or independent component
analysis (ICA), may be needed. Figure 8.2 illustrates how these preprocessing steps
are applied to the raw time-series data prior to clustering.

M. H. Rezaeian Koochi (*)

Department of Electrical Engineering, Shahid Bahonar University of Kerman, Kerman, Iran
e-mail: [email protected]
M. H. Hemmatpour
Department of Electrical Engineering, Jahrom University, Jahrom, Iran
e-mail: [email protected]
P. Dehghanian
Department of Electrical and Computer Engineering, School of Engineering and Applied
Sciences, George Washington University, Washington, DC, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 151
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_8
152 M. H. Rezaeian Koochi et al.

Fig. 8.1 Clustering data points in a 2-D space according to the Euclidean distances

Directly calculate the Euclidian distance or the

correlation coefficient between each pair of
time-series variations

Time-series variations Performing pre-processing Calculate the Euclidian distance or the

Apply clustering
steps such as applying correlation coefficient between each
algorithm
dynamic time warping pair of amended time-series variations

Applying feature extracting techniques

such as PCA or ICA to transform the
time-series variations into data points

Fig. 8.2 Steps of clustering a dataset of time-series variations

The deployment of advanced metering devices such as smart meters and phasor
measurement units (PMUs) in both distribution and transmission systems has
broaden the horizons of power system monitoring, operation, and control. In modern
power systems, advanced monitoring systems provide a huge amount of raw data
including time-series variations of various signals. The scale of these time-series
variations can range from less than a minute (related to the post-disturbance varia-
tions) to a one-day time frame (related to the daily bids offered by consumers) or
even monthly patterns of end users’ load variations. In this context, assessing the
similarities between the variations and clustering similar signals can be helpful in
better operation and planning of the power system. However, since the number of
these signals is high and their variation patterns are complicated, the use of advanced
feature extraction and clustering techniques is inevitable. Therefore, in the literature,
the use of clustering techniques in power system studies including both transmission
and distribution networks has been proposed.
Among the works in the related literature, several studies have been dedicated to
the use of clustering techniques for grouping the consumers. In [1], the use of
clustering techniques for dividing the end users into different categories according
to their 24-h bid-offers in a smart grid is proposed. This task could be very helpful in
demising flexible incentive rate strategies for demand response programs in power
distribution systems. Another class of research in this field relates to the use of
clustering techniques for dividing the loads in a smart grid or a micro grid. For
example, in [2, 3], the electricity consumption patterns across the households are
clustered in order to select the most appropriate household to be incentivized in an
incentive-based household demand response program. In doing so, thousands of
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 153

households are clustered into groups according to the similarity of their energy
consumption behaviors. Clustering has also been proposed in [4] for demising
local control of distributed energy resources (DERs). In doing so, clustering tech-
niques are used to cluster similar DER characteristic curves based on voltage
variations. Furthermore, the use of clustering techniques has been proposed in
other applications such as phase identification of smart meters in distribution sys-
tems [5] and forecasting of the energy demand [6].
Apart from the above examples on the application of clustering techniques in
power system studies, majority of works have been devoted to cluster or, in other
words, partition the bulk power system into areas so that the dynamic response of
elements in each area is similar. Such power system partitioning is based on a
concept called coherency, which is defined as the similarity of post-disturbance
dynamic response of the system elements [7]. In this regard, those elements reveal-
ing similar response to a disturbance are placed in the same cluster and are called
coherent. Similar to other applications of clustering in power system studies, here the
clustering is carried out on the basis of similarities between time-series signals,
which are the post-disturbance observations of the generators’ speeds or rotor angles
or phase angle of the voltages at all buses across the network.
In general, coherency-based power system partitioning methods can be catego-
rized in two groups, i.e., model-based and measurement-based methods [8]. Model-
based methods mostly rely on the slow coherency concept and use the linearized
model of the system to distinguish the coherent generators. Such methods are mainly
suitable for applications such as dynamic equivalencing or control actions to miti-
gate the low-frequency oscillations in the system. Figure 8.3 shows how a power
system is clustered into areas for dynamic equivalencing. Measurement-based
methods, which are also known as data-driven methods, use the data measured by
PMUs gathered from all over the system to find the coherent generators. Such
methods can also be extended to the buses to determine the boundary of clusters.
In such methods, it is assumed that the number of coherent groups and their
boundaries may vary for different disturbances with different characteristics. There-
fore, measurement-based methods are suitable for online applications such as con-
trolled islanding and special protection systems. Due to the large scale of

Areas (clusters) identified Power system after area

Original system
in the system reduction

G
G

G G

Fig. 8.3 Power system clustering for dynamic equivalencing

154 M. H. Rezaeian Koochi et al.

interconnected power systems, it is difficult to find the areas in the system quickly
following disturbances. Moreover, each bus or generator in the system can be treated
as a data point. As a result, the use of machine learning for clustering a power system
has been addressed vastly in the literature. In this regard, feature extraction and
supervised and unsupervised learning techniques have been used to partition the
power system into different areas.
From a measurement-based point of view, the number of areas and their bound-
aries are not fixed and may change for different disturbances. In fact, factors such as
the type of the disturbance and its location, as well as the power system condition at
the time of the disturbance occurrence (e.g., the system load level), may affect the
dynamic response of the system to the disturbance and cause different power system
partitioning schemes. Therefore, the clustering algorithm to be used for power
system partitioning is expected to have the following features:
• It should have no dependency on the prior assumption on the number of clusters.
• It should be capable to work well for imbalanced datasets. An imbalanced dataset
is a dataset in which both very small and large clusters exist. In power system
studies, it is probable that in some cases, a small area along with large areas are
formed in the system. Moreover, in unstable cases, usually a single generator
starts to lose synchronism and, therefore, forms a small group with one member.
The clustering algorithm should be capable to distinguish such small clusters.
• It shouldn’t be dependent on the random selection in its procedure (such random
selections will be discussed in Sect. 8.4 where unsupervised learning techniques
are described). This is necessary to ensure a deterministic solution.
• It is better for the clustering algorithm to feature a few parameters to be tuned.
While this is not a necessity, the lower the number of parameters, the simpler the
algorithm implementation.
It is noteworthy that similarity is a relative term meaning that two objects that
seem to be highly similar from one’s view may not be seen as similar objects from
the view of someone else. In this regard, the degree of similarity is a better term,
especially when dealing with clustering problems. Therefore, it is the user who
determines the desired degree of similarity according to the requirements of his
clustering problem to set the parameters of the chosen clustering technique. On the
other hand, various similarity evaluation criteria are available in the literature for
time-series signals. Examples of such criteria are Euclidean distance between the
samples of the two signals, Pearson coefficient and semblance coefficient, respec-
tively, defined in (8.1)–(8.3).
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u1 X N
cx,y ¼t ðxðk Þ  yðk ÞÞ2 ð8:1Þ
N k¼1
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 155

P
N
xðk ÞyðkÞ
cx,y ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
k¼1
ð8:2Þ
P N PN
x2 ð k Þ y2 ðk Þ
k¼1 k¼1

P
N
ðxðk Þ þ yðk ÞÞ2
k¼1
cx,y ¼ ð8:3Þ
P
N
2 ð x2 ð k Þ þ y2 ð k Þ Þ
k¼1

where x and y are the two signals whose similarity is to be assessed, and N is the
number of samples.
Along with techniques that have been proposed in the literature for power system
clustering and will be discussed in the rest of this chapter, a test case will also be used
and simulated to numerically evaluate the effectiveness of some of these techniques.
The test system used here is the 16-machine, 68-bus test system, which has been
introduced for dynamic studies and has been used widely for coherency evaluation
and its applications. This system includes two large areas named as New England
Test System (NETS) and New York Power System (NYPS) and three reduced areas
represented by G14, G15, and G16. The one-line diagram of this system is shown in
Fig. 8.4, details of which can be found in [9]. The test case used in this chapter,
which is adopted from [10], is characterized by a single line to ground fault applied
on the line connecting buses 1 and 2 and close to bus 1 and is cleared after 0.06 s
without any line tripping. Note that before applying this fault, the line connecting

G14 New York Power System (NYPS) G1 G8 New England Test System (NETS)
66 53 60

41 40 48 47
2 25 26 28 29
1
27 61
31 3
G11 17
30 18 G9
46 38 62 63
32 22
15
42 G10 33 24 58
49 9
4 16
67 34 12 14
5 21 G6
35 8 6 11
51 36
G15 45 13 23
64
7 10 19
50 G12 54 20 56 59
55
52 44 37
68 G2
39 65 G3 57 G4
G16 G7
43 G13
G5

Fig. 8.4 One-line diagram of the 16-machine, 68-bus test system

156 M. H. Rezaeian Koochi et al.

buses 8 and 9 is removed in order to increase the electrical distance between the
NETS and the rest of the system.
In the related literature, different data mining and pattern recognition techniques
have been proposed for power system partitioning. Although most effort is on using
unsupervised learning approaches, feature extraction techniques and supervised
learning approaches have been also proposed in the literature. Therefore, in the
rest of this chapter, the use of feature extraction techniques is first discussed. After
describing the application of supervised learning approaches, the use of
unsupervised learning methods will be then addressed. In addition, the advantages
and disadvantages of these techniques will be explored in power system clustering,
helping system planners and operators’ decision-making in real-world settings.

8.2 Power System Clustering Using Feature Extraction

Methods

Apart from various supervised and unsupervised learning methods examined by

researchers for finding groups of coherent generators or buses, there are also feature
extraction methods which have been proposed for coherency evaluation in power
systems. Feature extraction methods are mainly introduced as a preprocessing task
for other data mining techniques. To be specific, these methods are mainly intro-
duced for transforming the huge volumes of raw data into more appropriate datasets
with lower dimensions. Among such methods are PCA [11, 12] and ICA [13] which
have been applied to generators’ speed variations to transform these data into a 3-D
space for grouping.

8.2.1 Principal Component Analysis

PCA is a multivariate analysis technique which is capable of extracting uncorrelated

components from the input signals. PCA has been introduced mainly for reducing
the dimension of the problems. Assume that there are N observations of M variables
where M  N. By applying PCA, these N variables, which are in fact correlated
variables, are transformed into a set of uncorrelated variables. These uncorrelated
variables (principal components), which are the result of applying the p largest
eigenvectors of the original matrix to the original matrix, are then mapped to a p-
dimensional space. The idea behind PCA is to investigate if the first few components
account for the variations in the original variables. Mathematically speaking, con-
sider a matrix X with dimension M  N in which rows are the variables and columns
are the observations. By applying PCA, X will be reconstructed as the sum of the
orthonormal functions as follows [14]:
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 157

0 1 0 1 0 1
d 1,1 d1,2 d1,m
B C B C B C
X ¼ @ ⋮ Aω01 þ @ ⋮ Aω02 þ ⋯ þ @ ⋮ Aω0m ð8:4Þ
dm,1 dm,2 d m,m

where ω0i is the ith normalized right eigenvector of matrix XTX. In a p-dimensional
space ( p  M), the ith component is mapped to a point with coordinates di, 1, di, 2,
. . ., di, p. In this regard, since similar components have similar coordinates in the new
space, a group of data points corresponding to the original variables with similar
variations will form a cluster in the new space.
In power system studies, the use of PCA for power system partitioning has been
proposed in several research works. In [11], the efficiency of PCA on clustering the
system generators using the post-disturbance variations in the generators’ speed
signals has been examined. Moreover, PCA has been also applied in [11] for
power system partitioning by investigating the similarities between post-disturbance
phase angle variations of the voltages at all buses. However, it is not possible in all
cases to visually cluster the data points mapped to the reduced p-dimensional space,
meaning that further analysis would be required to find the clusters. For example, in
[15], after applying PCA, the use of hierarchical clustering for grouping the com-
ponents in the three-dimensional space has been proposed. One advantage of PCA is
that if it works well, there would be no need to have a prior assumption on the
number of clusters. This is because, as stated in the Introduction, from a
measurement-based point of view, a power system can be partitioned in different
number of areas with different boundaries for different disturbances occurring in the
system.
Figure 8.5 shows the locations of coefficients of the first three components for
each bus in the 3-D space obtained for the test case. As it can be seen from Fig. 8.5,
area 1 (cluster 1) has been formed in a far distance from the other areas and,
therefore, can be easily distinguished. However, although it seems that the other

Fig. 8.5 Clusters of buses

obtained from applying
PCA in a 3-D space
Cluster 2

Cluster 1
Cluster 3
158 M. H. Rezaeian Koochi et al.

G14 New York Power System (NYPS) G1 G8 New England Test System (NETS)
66 53 60

41 40 48 47
2 25 26 28 29
1
27 61
31 3
G11 17
30 18 G9
46 38 62 63
32 22
15
42 G10 33 24
49 58
9
4 16
67 12 14
34 5 G6
35 21
51 8 6 11
G15 36 13
45 23
64
7 10 19
50 G12 54 20 56 59
55
52 44 37
68 G2
39 65 G3 57 G4
G16 G7
43 G13
G5
Cluster 2 Cluster 3 Cluster 1

Fig. 8.6 Clusters (areas) obtained for the test system

two clusters can be discriminated visually, in some cases, it will be hard to find the
clusters in the 3-D space, and thus, using a clustering technique would be essential.
Note that the clustering scheme in this figure is the same as the one obtained by
subtractive clustering in [10]. A graphical illustration of these areas is shown in
Fig. 8.6.

8.2.2 Independent Component Analysis

ICA is another multivariate technique which is aimed to extract hidden features from
raw data [16]. Unlike PCA in which a set of uncorrelated components are extracted
from the raw data, ICA tries to transform the data into a set of independent
components. More specifically, ICA tries to find components that are independent
and non-Gaussian. According to [13], being independent can be an implication of
being uncorrelated, but being uncorrelated does not guarantee an independence.
Considering a set of N observations of M variables stored in a matrix X with
dimension M  N, ICA will decompose X into independent components as shown in
(5) [16].
0
1 0 1 0 1
d1,1 d1,2 d1,m
B C B C B C
X ¼ @ ⋮ Ac01 þ @ ⋮ Ac02 þ ⋯ þ @ ⋮ Ac0m ð8:5Þ
dm,1 dm,2 dm,m
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 159

Fig. 8.7 Clusters obtained by ICA for the test case: (a) 3-D space and (b) 2-D space

Similar to PCA, independent components have to be sorted according to their

dominance. Then dominant components, which can be the first two for the 2-D
visualization or the first three for the 3-D visualization, will be used for clustering.
For example, considering the first three dominant independent components and
plotting the first three elements of rows of d-vectors, i.e. di, 1, di, 2, and di, 3, it will
be seen that the data points corresponding to coherent signals will take place close
together in a 3-D space. However, as described for PCA, it is not always possible for
ICA to cluster the data points visually or graphically. Therefore, although ICA can
extract useful information from raw data, at last, it may be needed to use a clustering
technique to find the final clusters. Figure 8.7 shows the results of applying ICA
(here fast ICA) to the test case used in this chapter. As it can be seen, ICA partitions
the system into two clusters, where the second cluster is formed by combining the
clusters 2 and 3 in Fig. 8.5.

8.3 Power System Clustering Using Supervised Learning

Methods

Supervised learning approaches, mostly known as classifiers, are another type of

learning methods, which are used in various scientific fields such as engineering,
medicine, and economics. In these methods, a set of data with already known group
labels are used for training the classifiers. In fact, supervised methods try to find the
rules that determine the relationships between the already observed data and their
associated known labels and then use them to predict the label of the next observa-
tion. Therefore, a dataset consisting several attributes and a target is needed for
training and testing the classifier where each of these attributes and targets can be of
any types of numerical, nominal, or Boolean variables.
In the literature, supervised learning approaches such as decision trees (DTs) and
artificial neural networks (ANNs) have been developed to build classifiers for
160 M. H. Rezaeian Koochi et al.

Data measured at first cycles

Final grouping

Fig. 8.8 Data window for generators grouping prediction using generators’ speed variations

predicting the stable or unstable groups of generators using the first cycles after fault
occurrence. In coherency analysis, these approaches are aimed to replace the timely
coherency identification methods with a multi-class classifier to predict the coherent
groups of generators. Among classification methods, the use of neural networks and
decision trees for coherency prediction has been suggested. A typical representation
of generators’ speed oscillations has been shown in Fig. 8.8. According to this figure,
a classifier is intended to predict the final grouping that is going to be formed using
the information and patterns hidden in the data measured at the first cycles following
the disturbance.

8.3.1 Artificial Neural Networks

In ANNs, several layers, each consists of neurons, are put together to build a
classifier. ANNs are vastly used in different fields for predicting and regression
applications. In the literature related to power system coherency analysis, the use of
ANN has been addressed in few studies. The primary work has been published in
[17], where authors have suggested the use of ANN for finding clusters using three
samples of speed signals of generators. ANNs have also been proposed to serve as a
preprocessing task in generators clustering. In a recent study, Siddiqui et al. [18]
used a neural network structure to predict the time-series signals which is the
post-disturbance rotor angle trajectories. In their work, they have used the first six
cycles’ variations following a disturbance to predict the future rotor swing trajecto-
ries and then have clustered the generators by evaluating the similarities of these
predicted trajectories.
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 161

8.3.2 Decision Trees

DTs are a sort of classifiers with a treelike shape forming from a set of decision rules
extracted from a training dataset. To extract the rules, every observation in the data
must be accompanied with a known target content. A DT consists of two types of
nodes, namely, decision and terminal nodes. The input of a DT includes several
attributes, which should be tested in their related decision nodes. For example, in the
root node, which is the first decision node of a DT, a test on the associated attribute is
carried out, and based on the output value of the first test, the next decision node is
then determined. This procedure will continue until it reaches to one of the terminal
nodes or leaves of the DT. Similar to ANNs, a training dataset along with a testing
dataset are used to train and evaluate the accuracy of a DT.
In the literature, several studies have used different types of DTs for generator
coherency prediction. For example, authors in [19, 20] have shown that using
appropriate DT training techniques, it is possible to build a power system model to
predict the clusters of coherent generators using the data measured by PMUs in a
very short time period following a disturbance. In another work presented in [21], it
has been proposed to use a set of simple DTs to find the most coherent generators
following a disturbance. Quantile regression forest is another type of DTs, which has
been proposed in [22] to be used for predicting unstable group of coherent genera-
tors, which is very helpful for early identification of the necessary remedial actions in
the emergency scenarios following disturbances.
In these studies, it is assumed that having a good knowledge of the uncertainties
in a power system, it can be possible to find almost all possible clustering schemes
that would happen in the system. Then a training dataset is generated in which the
target is any of these probable clustering schemes while the attributes are demised
characteristics appropriately extracted from the synchrophasor measurements.
Finally, a training and testing procedure is carried out to discover the hidden patterns
and rules between the values of the attributes and the target. However, one notable
issue arising in these studies is that even in these methods, a clustering technique
may be needed to establish the target input, if the number of data points (generators
or buses) is high.

8.4 Power System Clustering Using Unsupervised Learning

Methods

The vast majority of works in the field of power system clustering is, as expected,
done using unsupervised learning approaches, which are called as clustering type of
data mining techniques. Various clustering methods have been proposed for power
systems clustering. However, these methods have their advantages and disadvan-
tages, which should be examined and addressed before being used for power system
clustering.
162 M. H. Rezaeian Koochi et al.

With respect to the nature of their procedures, unsupervised clustering techniques

can be divided into two categories. In the first one, which includes algorithms such as
k-means (KM) [23] and fuzzy c-means (FCM) [24], the desired number of clusters
(NC) is known, and therefore, the task is to divide the dataset into NC groups
according to the degree of similarities between each pair of data points. However,
in the second category are placed algorithms, such as subtractive clustering
(SC) [25–27] and density-based spatial clustering of applications with noise
(DBSCAN) [28], in which instead of having a prior assumption on the desired
number of clusters, a radius indicating the desired highest degree of dissimilarity
between each pair of data points in a cluster would be determined by the user, and
thus, the only task would be to let the algorithm find the clusters. In other words, in
the second category, it would be the algorithm itself that determines the number of
clusters. Figure 8.9 illustrates how the clustering works in these two categories.
Subsects. 8.4.1–8.4.7 will describe these methods as well as their cons and pros.
In addition, the literature related to the use of each clustering method in the field of
power system clustering will be introduced in their respective subsections.

8.4.1 K-Means Clustering Algorithm

K-means clustering is a popular clustering algorithm which uses an iterative proce-

dure to cluster the data points into a predefined number of clusters. In fact, KM is a
hard clustering algorithm meaning that in its final solution the membership value of
each data point to one cluster is found to be 1 while for other clusters, it is 0. This
algorithm searches for the clusters in the dataset so that a cost function is minimized.
The cost function, which is defined as the sum of dissimilarities between the member
of NC clusters, is as follows:

Data set Data set

Clustering the data Clustering the data Finding clusters Finding clusters
points into 2 groups points into 3 groups according to a according to a
small radius large radius

Fig. 8.9 Two types of clustering procedure by clustering techniques

8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 163

!
X
NC X
J¼ di,j 2
ð8:6Þ
j¼1 i2Gj

For any data point i in the dataset, a membership function will be defined as in
(8.7).

1 if jth datapoint belongs to ith cluster
ui,j ¼ ð8:7Þ
0 otherwise

Note that a data point i is considered to belong to the jth cluster if among all
centers, the jth center has the lowest dissimilarity with the data point. At each
iteration, cluster centers are updated according to (8.8).

1 X
z j ¼   xi ð8:8Þ
G j xi 2G j

In (8.8), cj is the center of the jth cluster, Gj is the set of data points in the jth
cluster, and |Gj| represents its size, while xi is the ith data point in Gj. It should be
noted that as (8.8) shows, in KM algorithm, a cluster center in the final solution may
not essentially be a data point. Considering (8.6)–(8.8), the iterative procedure for
clustering a dataset will be as follows:
• Step 1: Initialize NC number of cluster centers selected from the dataset.
• Step 2: Find membership values for all data points using (8.7).
• Step 3: Compute the cost function defined in (8.6).
• Step 4: Update cluster centers using (8.8) and return to step 2.
Two stopping criteria can be used here. In the first one, algorithm will stop if the
difference between the value of the cost function obtained in two consecutive
iterations becomes less than a threshold, while in the second, algorithm will stop if
it has been run for a sufficient number of time. However, the drawbacks of KM are
twofold. Firstly, KM needs the number of clusters to be predefined. Secondly, its
performance highly depends on the random selection of initial cluster centers.
Therefore, it may happen that KM finds a wrong solution especially for a dataset
in which both very small and very large clusters exist. To overcome this challenge,
one approach is to run KM several times to find better solutions or use a subtractive
method to select the initial cluster centers appropriately.
In the related literature, several studies have used KM to find coherent generators
and coherent areas. For example, authors in [29, 30] have used KM to cluster
generators in predefined number of groups. In so doing, they define a row vector
for each generator representing the modal response of the generator and then use the
angle between the vectors of two generators as a criterion for obtaining dissimilar-
ities. In another work presented in [31], singular value decomposition (SVD) is first
applied to the rotor angle variation signals of generators to construct a matrix called
164 M. H. Rezaeian Koochi et al.

the characteristic coefficient matrix with dimension k  NG where NG is the number

of generators, and k is selected to retain the largest k singular values during SVD
process. Next, KM is applied on this matrix to cluster NG generators. It is worthy to
note that the characteristic coefficient matrix is considerably smaller than the matrix
of the original samples, resulting in KM to be computationally more attractive.
However, there are other types of studies in which a combination of KM and
evolutionary algorithms is proposed for better partitioning of the power systems
[32, 33]. For example, in [33], KM is applied on the generators’ rotor samples in
conjunction with PSO to contrive a PSO-KM method for better clustering.
In this subsection, the KM algorithm is applied to cluster the buses of the test
case. To do that, the Pearson coefficient is first used to assess the similarity of phase
angle variations of voltages at buses. Then the distance between the value of Pearson
coefficient between two buses and unity value is considered as the dissimilarity
between the two buses. Finally, KM algorithm uses these dissimilarity values to
cluster the buses. Note that the number of clusters has been set to 3. However, as
stated above, the solution obtained by KM depends on the centers selected at the
initial iteration. Therefore, KM algorithm is applied to the dataset 1000 times to see
how many times the correct solution shown in Fig. 8.6 will be obtained. Results
showed that KM algorithm succeeded to find the clusters in 95 runs of the algorithm.

8.4.2 Partitioning around Medoid Algorithm

PAM clustering algorithm, which is known also as k-medoids clustering algorithm,

is a clustering algorithm basically similar to KM. However, unlike KM in which
cluster centers are not the real data points, in PAM algorithm, cluster centers
(medoids) are essentially selected in every iteration from data points in the dataset.
In fact, a medoid is a data point whose average dissimilarity to other data points in
the cluster is minimal. PAM algorithm needs the initial selection of medoids. In
doing so, a building phase, which consists of an iteration-based procedure, has been
added for selecting the initial medoids wisely [34]. However, it is essential in PAM
to predefine the desired number of clusters. Briefly speaking, the implementation
steps of the PAM algorithm are as follows (more details can be found in [35, 36]):
• Step 1: Initialize NC number of medoids (cluster centers) selected from the
dataset.
• Step 2: Find membership values for all data points.
• Step 3: Update medoids according to the dissimilarities of the data points to the
medoids selected previously.
• Step 4: Stop updating medoids if the stopping criterion holds. Otherwise, return to
step 2.
• Step 5: Assign each data point to the cluster centered by the nearest medoid.
In the field of coherency evaluation, authors in [34, 37] have proposed the use of
PAM algorithm. In their works, they have applied PAM algorithm on dissimilarities
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 165

between rotor angle variations of generators to cluster them for dynamic

equivalencing. In order to examine the effectiveness of PAM algorithm, it has
been applied on the test case defined in Sect. 8.1. In doing this examination, the
building phase has not been considered, and the aim has been to examine the
capability of the algorithm in finding clusters. Moreover, the number of clusters
has been set to three, and the algorithm has applied 1000 times on the dataset. It has
been found that in 274 runs, the PAM algorithm has succeeded to find the correct
clusters.

8.4.3 Fuzzy C-Means Clustering Algorithm

Unlike KM and PAM algorithms, in FCM algorithm, an object is allowed to belong

to several clusters with different degrees of membership. FCM is an iteration-based
clustering algorithm, and its solution is dependent on the initial selection of the
membership values. This is because cluster centers in each iteration are determined
on the basis of membership values. Moreover, the final solution of the algorithm is
dependent on the random selection of the initial memberships and the number of
clusters. FCM algorithm utilizes a cost function more general than (8.6), in which
membership values are included. In this cost function, which is defined in (8.9) and
should be minimized, a weighting exponent (m  1) is used to control the fuzziness
of the clustering process [24].
!
X
NC X
J¼ um
i,j d i,j
2
ð8:9Þ
j¼1 i2Gj

Moreover, though a data point belongs to all clusters with different degrees of
memberships, the sum of degrees must equal to 1, or:

X
NC
ui,j ¼ 1 8i ¼ 1, . . . , n ð8:10Þ
j¼1

Similar to (8.8), a set of cluster centers will be updated in each iteration except
that in FCM, the fuzziness of the membership values controls the updating process.
In addition to updating the centers, membership values should be updated at each
iteration as well. In doing so, (8.11) and (8.12) are used, respectively.

P
n
um
i,j xi
i¼1
zj ¼ Pn ð8:11Þ
um
i,j
i¼1
166 M. H. Rezaeian Koochi et al.

1
ui,j ¼ N C  2=ðm1Þ
ð8:12Þ
P di,j
di,k
k¼1

As stated, FCM algorithm is an iteration-based algorithm in which centers and

membership values are updated at each iteration so that the cost function is mini-
mized. The implementation steps of this algorithm are as follows:
• Step 1: Initialize the algorithm by random selection of membership values.
• Step 2: Find cluster centers using (8.11).
• Step 3: Compute the cost function defined in (8.9).
• Step 4: Update the membership values using (8.12) and return to step 2.
Among significant works which have used FCM, we can refer to [38] where FCM
has been proposed for grouping the power system generators into coherent areas. To
do that, a coherency measure is first defined to assess the similarity of rotor angle
variations of generators. Then, FCM algorithm is applied to cluster the generators on
the basis of such similarities. In this subsection, FCM algorithm is applied on the test
case defined in this chapter. Similar to Subsections 8.4.1 and 8.4.2, the number of
clusters has been set to 3, and algorithm has been run for 1000 times. Moreover, in
the FCM algorithm, the number of iterations is set to 1000. It has been found that
FCM could obtain the correct results in 995 out of 1000 runs.

8.4.4 Fuzzy C-Medoids Clustering Algorithm

FCMd is another soft clustering algorithm which, similar to FCM, simultaneously

assigns the data points to different clusters with different values of memberships.
FCMd is again an iterative-based clustering algorithm, and the difference between
FCMd and the FCM is that FCM focuses on updating the membership values, while
in FCMd, the focus is on appropriately upgrading the medoids (cluster centers)
[39]. Therefore, instead of initializing the algorithm with random selection of
membership values, FCMd starts with random selection of medoids. In its algorithm,
FCMd uses a cost function similar to (9) except that di, j will be the dissimilarity
between the ith data point in the jth cluster and its medoid. However, here the
membership values will be obtained as follows:
 1=ðm1Þ
1
di,j
ui,j ¼ N C  1=ðm1Þ
ð8:13Þ
P 1
di,k
k¼1
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 167

Moreover, medoids are updated in each iteration according to the dissimilarities

of the data points to the medoids selected in the previous iteration. Mathematically
speaking, medoids are obtained using the membership values as follows [40]:

X
n
q ¼ arg min um
i,j d k,i ð8:14Þ
1kn
i¼1

z j ¼ xq ð8:15Þ

Accordingly, the implementation steps of the FCMd algorithm are as below:

• Step 1: Initialize NC number of medoids (cluster centers) selected from the
dataset.
• Step 2: Compute membership values for all data points using (8.13).
• Step 3: Update medoids according to (8.14) and (8.15).
• Step 4: Stop updating medoids if the stopping criterion holds. Otherwise, return to
step 2.
• Step 5: Assign each data point to the cluster centered by the nearest medoid.
The stopping criteria is that whether for all clusters the new medoid obtained in
the current iteration is equal to the medoids obtained in the previous iteration: If the
answer is yes, the algorithm will stop. It is noteworthy that FCMd suffers from those
shortcomings associated to KM and FCM, but its performance can be better than that
of FC, if the initial centers are selected appropriately. In the literature, FCMd has
been proposed by the authors in [40, 41] to provide a better solution for large power
system partitioning. The reason is that in case where the number of data points is
high, using FCM clustering algorithm can be time-consuming. To evaluate the
performance of the FDMd algorithm, it has been applied to the test case used in
this chapter for 1000 times, and it was seen that FCMd algorithm could find the
clustering scheme shown in Fig. 8.6 in 94 runs.

8.4.5 Subtractive Clustering Algorithm

SC is another clustering algorithm that has been primarily introduced as a

pre-processing tool for other clustering algorithms such as KM and FCMd. SC is
primarily aimed to find cluster centers using an iterative procedure. However, it can
itself be used for clustering as well. In this algorithm, a density value Di is first
assigned to a data point xi using the density measure defined in (8.16).
h i
X
NB

ðdi,j Þ2
Di ¼ e ðr a =2Þ2
ð8:16Þ
j¼1
168 M. H. Rezaeian Koochi et al.

where di,j is the dissimilarity between data points i and j, and ra is a positive constant
used to represent the neighboring radius. According to (8.16), the lower the value of
di,j is (the more data point j is similar to data point i), the higher the value of Di will
be. In other words, since the density measure in (8.16) is defined on the basis of sum
of similarities between all data points, the data point with maximum density value
will be chosen as the first cluster center which is because it has the highest similarity
with the highest number of data points. In the next step of the algorithm, all density
values will be revised with respect to the density value of the cluster center
determined in the previous step. This subtractive revision is done through using
the following equation:
 
ðdi,cp1 Þ
2


ðrb =2Þ2
Dpi ¼ Dp1
i  Dp1
c e i ¼ 1, 2, . . . , N B ð8:17Þ

where
n o
p1
Dp1
c ¼ max D i ð8:18Þ
i

In (8.17), rb is a positive constant aimed to define a neighborhood around the

previous center whose members experience a measurable reduction in their density
values. Moreover, Dp1 c is the density value of the cluster center selected at iteration
p-1, di,cp1 is the dissimilarity between data point i and the center determined at
iteration p-1. Now, the cluster center at iteration p will be selected as the one having
the highest density value among the revised values obtained from (8.17). The
process of finding cluster centers will be repeated until the rejection criterion rejects
the candidate center. Note that various rejection criteria can be defined ranging from
a simple one as in [10] to a more complicated criterion. Once the cluster centers are
determined, the similarity of each data point with each cluster center should be
assessed. Accordingly, a matrix with dimensions NC  ND will be formed as follows
[10]:
2 3
r 1,1 ... r 1,N B
6 7
R¼4 ⋮ ⋱ ⋮ 5 ð8:19Þ
r N C ,1 . . . r N C ,N B

where
 
ðd Þ
2
j,ci

ðrb =2Þ2
r i,j ¼ D1j  D1ci e ð8:20Þ

Finally, assuming that the ith element of the jth column of R has the lowest value in
the column, it can be said that the jth data point should be assigned to a cluster whose
center is the ith cluster center determined by the algorithm. The performance of SC
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 169

algorithm for power system partitioning and its effectiveness over FCMd and KM
algorithms has been demonstrated in [10]. Particularly, the effectiveness of the SC
algorithm in finding the areas for the test case used in this chapter has been shown in
[10]. Note that SC is mainly featured with no random selection operation, and
therefore, it has to be run once, and its results are deterministic. According to the
SC algorithm, two parameters have to be predefined, i.e., a neighborhood radius (ra)
and the squash factor which is defined as the ratio of rb to ra. In addition, one or more
parameters have to be predefined as well according to the rejection criteria selected
by the user.

8.4.6 Density-Based Spatial Clustering of Applications

with Noise

DBSCAN is another density-based clustering technique which has been used in

different areas of science. In a way similar to subtractive clustering, the basics of
DBSCAN rely on the concept of density in an n-dimensional space. Considering a
2-D space shown in Fig. 8.10, the following definitions hold [28]:
• A data point pr is a core data point if at least Pmin data points including pr are
within the distance ε of data point pr.
• If data point qr is within distance ε of data point pr, it is said that data point qr is
directly reachable from data point pr.
• A data point pr is a noise if it is not directly reachable from any core data point.
• A data point qr is reachable from data point pr if there is a path pr1, . . ., prn with
pr1 ¼ pr and prn ¼ qr, where each pr(i + 1) is directly reachable from pri.
According to the above definitions, a cluster is a set of at least Pmin data points so
that each pair of them are directly reachable. For example, considering the example
data points shown in Fig. 8.9, if Pmin is set to 3, a cluster with eight members
(shown in green) will be formed, while the other three data points are found to be
noises (shown in yellow).

Fig. 8.10 Example of

clustering the data points in
a 2-D space (Pmin ¼ 3)
170 M. H. Rezaeian Koochi et al.

G14 New York Power System (NYPS) G1 G8 New England Test System (NETS)
66 53 60

41 40 48 47
2 25 26 28 29
1
27 61
31 3
G11 17
30 18 G9
46 38 62 63
32 22
15
42 G10 33 24 58
49 9
4 16
67 34 12 14
5 21 G6
35 8 6 11
G15 51 36
45 13 23
64
7 10 19
50 G12 54 20 56 59
55
52 44 37
68 G2
39 65 G3 57 G4
G16 G7
43 G13
G5

Cluster 2 Cluster 1

Fig. 8.11 Clusters (areas) obtained for the test system using DBSCAN

In SubSect. 8.4.5, the use of subtractive clustering and its advantages for power
system coherency-based clustering are discussed. However, DBSCAN is a cluster-
ing algorithm simpler than SC algorithm and is more suitable for power system
partitioning due to the following features [42].
• It doesn’t need a prior assumption on the number of clusters. This is a very
important feature since from a measurement point of view, groups of coherent
generators are not fixed for different disturbances.
• It needs fewer parameters to be set.
• It doesn’t use any random selection operation, and therefore, its solution is
deterministic.
• It works well for imbalanced datasets. It can easily find small clusters and noises
which is necessary for online measurement-based power system partitioning.
In [42], the use of DBSCAN for clustering the buses in a power system is
proposed and examined. Here, this algorithm is applied to the test case defined in
this chapter, and its clustering result is shown in Fig. 8.11. Note that to obtain the
clustering scheme, ε is set to 0.1. As it can be seen from this figure, DBSCAN has
clustered the buses into two areas instead of the three areas shown in Fig. 8.6. The
reason backs to the nature of the two categories of clustering techniques defined in
the beginning of Sect. 8.4.

8.4.7 Support Vector Clustering Algorithm

Support vector clustering (SVC) is derived from support vector machine on the basis
of the fact that when data points in the original dataset are mapped into a new space
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 171

Fig. 8.12 Transforming the data points from the original space into the feature space in the SVC
algorithm

with higher dimensions, the minimum sphere in the new space that encloses a group
of similar data points is found, and after transforming back into the original space, it
can be considered as the cluster boundary (see Fig. 8.12). SVC doesn’t need to have
a pre-assumption on the number of clusters since it solves an optimization problem
to find the optimal clustering. In other words, it can be said that SVC can cluster
datasets with arbitrary shapes and is able to estimate the optimum number of clusters.
However, SVC needs its input to be in the form of data points in a p-dimension space
since it uses the Euclidean distances between data points to transform them into the
feature space. Therefore, using PCA or other embedding techniques is necessary to
transform the similarity of the time-series data into the Euclidean distances.
A brief mathematical expression of SVC procedure is as follows. In order to
prepare the data to be used in this algorithm, SVC employs a nonlinear kernel
transformation for mapping the data points to the higher dimensional feature
space. In order to find the minimum sphere, an optimization problem in the Lagrang-
ian form is first defined as in (8.21).
X 2
 X X
L ¼ R2  R2 þ ξ j  ϕ x j  a βj  ξ jμ j þ C ξ j ð8:21Þ
j j j

where X ¼ x1, x2, . . . ., xN is a dataset of N data points, ϕ is a nonlinear transformation

function, and R is the smallest enclosing sphere.Pβj and μj are the Lagrangian
multipliers. C is the user defined constant, and C ξ j denotes the penalty term.
j
Moreover, R2 is the distance of each data point x from the center of the shape which
is calculated by (8.22).
X X
R2 ðxÞ ¼ K ðx, xÞ  2 K x j, x β j þ β i β j K xi , x j ð8:22Þ
j i, j

In (8.22), K xi , x j ¼ ϕðxi Þ:ϕ x j ¼ eqkxi x j k is the Gaussian kernel transfor-

2

mation, and q is the width parameter which controls the boundaries of clusters. The
optimization problem defined in (8.21) is then converted to its dual quadratic form to
obtain (8.23). Now, the goal would be to obtain the Lagrangian multiplier βj.
172 M. H. Rezaeian Koochi et al.

X X
W¼ K x j, x j β j  β i β j K xi , x j ð8:23Þ
j i, j

subject to
X
0  β j  C, βj ¼ 1 ð8:24Þ
j

In the literature, the use of SVC for finding groups of coherent generators is
presented in [43, 44] to show its effectiveness over KM and FCM techniques.

8.5 A Discussion on the Use of Machine Learning for Power

System Clustering and Possible Future Works

As discussed in Sect. 8.3, several classifiers have been proposed for power system
clustering and its predictions. However, a survey of the related literature shows that
the number of works in this field is limited. One can note that in addition to the
concerns regarding the fact that a clustering technique is still needed to form the
schemes in the target variable in the dataset, another important issue that should be
taken into account is the sufficiency of the training dataset. This must be addressed as
it can affect the efficiency of a classifier for an observation made out of the range of
those in the training dataset.
On the other hand, among clustering methods that have been proposed for
clustering the generators or buses in a power system, it can be said that each of
them suffers from one or more of the following shortcomings, particularly when
being considered for online applications; some of these algorithms have more than
one parameters to be set, and therefore, their final results depend on the setting of
these parameters. Moreover, some of them such as SVC needs more calculation and
are time-consuming. In addition, algorithms including KM, FCM, and FCMd use
random selection processes in their initial iteration, meaning that their results may
change as the initial selection changes. Also, for almost all these algorithms, the
number of clusters needs to be predefined.
In order to provide a better vision on the cons and pros of the unsupervised
learning clustering techniques described in Sect. 8.4, more evaluations and simula-
tions will be given here. At first, KM, PAM, FCM, and FCMd algorithms have been
again applied to the test case for a larger number of runs, which is 10,000 runs in this
section. Table 8.1 shows the number of times each algorithm has found the correct
clustering scheme among the 1000 and 10,000 runs. It can be seen that the best
performance is obtained by FCM clustering algorithm.
As stated in Sect. 8.1, clustering techniques have been proposed to be used in
various fields of power system studies. The suitability of a clustering technique for a
specific application depends on the requirements of that application. For example, in
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 173

Table 8.1 Performance of different clustering algorithms on the test case under different number of
simulation runs
Total no. of runs KM PAM FCM FCMd
10,000 1001 2647 9998 873
1000 95 274 995 94

applications where the number of clusters are given or at least can be easily
determined, the use of KM or FCM algorithms could be advisable. In the field of
power system clustering, an ideal algorithm for online grouping of coherent gener-
ators or buses should be characterized with the following features:
• It has less parameters to be set, i.e., at most one parameter. This is very important
particularly if a measurement-based online clustering is desired.
• It uses no random operation to make the results deterministic. It is essential to run
the algorithm once and obtain the correct solution. This feature relates to the
proper selection of initial values in the first iteration of algorithms such as KM or
FCM. In the literature, there have been techniques proposed to be used as a
preprocessing for preparing the initial values. However, these preprocessing can
add burden in online applications.
• The algorithm itself determines the number of clusters. From a measurement-
based point of view, in power system post-disturbance clustering, the number of
areas and their boundaries are not fixed and may change for different distur-
bances. Therefore, it would not be easy to have a prior assumption on the number
of areas (clusters).
• It is suitable for both stable and unstable situations since in the latter, one
generator starts to lose its synchronization, and therefore, the algorithm must
detect it as an individual cluster.
According to the above features needed for the ideal algorithm to be used for
power system clustering and considering the features of most well-known clustering
techniques proposed in the literature, it can be concluded that SC and DBSCAN
algorithms can be the most proper ones. The reasons are that they use no random
operation; therefore, their solutions are deterministic, they have low number of
parameters to be set, they can find very small clusters in a dataset, and they can
themselves find the number of clusters.

8.6 Conclusions

This chapter discussed the use of data mining and machine learning methods in
power system clustering applications. Particularly, the use of feature extraction
methods as well as supervised and unsupervised learning approaches for power
system clustering was discussed. For a better understanding of the performance of
174 M. H. Rezaeian Koochi et al.

the unsupervised approaches, these methods were applied on a test case, and their
solutions were examined. These simulations also made it possible to have a more
clear vision on the advantages and disadvantages of different clustering methods.
Nevertheless, future research is needed to widen the application of advanced clus-
tering techniques, particularly due to the evolving interests in the use of artificial
intelligence and deep learning methods in power system applications.

References

1. Z. Luo, S. Hong, Y. Ding, A data mining-driven incentive-based demand response scheme for a
virtual power plant. Appl. Energy 239, 549–559 (2019)
2. T. Teeraratkul, D. O’Neill, S. Lall, Shape-based approach to household electric load curve
clustering and prediction. IEEE Trans. Smart Grid. 9(5), 5196–5206 (2017)
3. S. Dasgupta, A. Srivastava, J. Cordova, R. Arghandeh, Clustering household electrical load
profiles using elastic shape analysis. In. 2019 IEEE Milan PowerTech. (2019), pp. 1-6
4. S. Karagiannopoulos, G. Valverde, P. Aristidou, G. Hug, Clustering data-driven local control
schemes in active distribution grids. IEEE Syst. J., 1–10 (2020). https://doi.org/10.1109/JSYST.
2020.3004277
5. F. Olivier, A. Sutera, P. Geurts, R. Fonteneau, D. Ernst, Phase identification of smart meters by
clustering voltage measurements. In. 2018 Power Systems Computation Conference (PSCC).
(2018), pp. 1-8
6. C. Bock, Forecasting energy demand by clustering smart metering time series. In. International
conference on information processing and Management of Uncertainty in knowledge-based
systems. (Springer, 2018). pp. 431-42
7. M.H.R. Koochi, S. Esmaeili, P. Dehghanian, Coherency detection and network partitioning
supported by wide area measurement system. In 2018 IEEE Texas power and energy confer-
ence (TPEC). (2018), pp. 1-6
8. M.H.R. Koochi, S. Esmaeili, G. Ledwich, Taxonomy of coherency detection and coherency-
based methods for generators grouping and power system partitioning. IET Generation, Trans-
mission & Distribution. 13(12), 2597–2610 (2019)
9. G. Rogers, Power system oscillations. (Springer Science & Business Media, 2012)
10. M.H. Rezaeian, S. Esmaeili, R. Fadaeinedjad, Generator coherency and network partitioning for
dynamic equivalencing using subtractive clustering algorithm. IEEE Syst. J. 12(4), 3085–3095
(2017)
11. K.K. Anaparthi, B. Chaudhuri, N.F. Thornhill, B.C. Pal, Coherency identification in power
systems through principal component analysis. IEEE Trans. Power Syst. 20(3), 1658–1660
(2005)
12. K. Mandadi, B.K. Kumar, Generator coherency using Zolotarev polynomial based filter bank
and principal component analysis. Int. J. Emerging Electric Power Syst. 19(2), 1–10 (2018)
13. M. Ariff, B.C. Pal, Coherency identification in interconnected power system—An independent
component analysis approach. IEEE Trans. Power Syst. 28(2), 1747–1755 (2012)
14. C. Chatfield, A. Collins, Introduction to Multivariate Analysis (CRC Press, Thousand Oaks,
1981)
15. A.M. Almutairi, S.K. Yee, J. Milanovic, Identification of coherent generators using PCA and
cluster analysis. In Proc. Power Systems Computation Conference, 1–10 (2008)
16. A. Hyvärinen, J. Karhunen, E. Oja, Independent component analysis, adaptive and learning
systems for signal processing, communications, and control. John Wiley & Sons, Inc. 1, 11–14
(2001)
8 Clustering in Power Systems Using Innovative Machine Learning/Deep. . . 175

17. M.-H. Wang, H.-C. Chang, Novel clustering method for coherency identification using an
artificial neural network. IEEE Trans. Power Syst. 9(4), 2056–2062 (1994)
18. S.A. Siddiqui, K. Verma, K. Niazi, M. Fozdar, Real-time monitoring of post-fault scenario for
determining generator coherency and transient stability through ANN. IEEE Trans. Ind. Appl.
54(1), 685–692 (2017)
19. M.H.R. Koochi, S. Esmaeili, R. Fadaeinedjad, New phasor-based approach for online and fast
prediction of generators grouping using decision tree. IET Generation, Transmission & Distri-
bution. 11(6), 1566–1574 (2017)
20. T. Guo, J.V. Milanović, Online identification of power system dynamic signature using PMU
measurements and data mining. IEEE Trans. Power Syst. 31(3), 1760–1768 (2015)
21. M.H.R. Koochi, P. Dehghanian, S. Esmaeili, P. Dehghanian, S. Wang, A synchrophasor-based
decision tree approach for identification of most coherent generating units. In: IECON 2018-
44th Annual Conference of the IEEE Industrial Electronics Society, (2018), pp. 71–6
22. S.M. Mazhari, N. Safari, C. Chung, I. Kamwa, A quantile regression-based approach for online
probabilistic prediction of unstable groups of coherent generators in power systems. IEEE
Trans. Power Syst. 34(3), 2240–2250 (2018)
23. J.A. Hartigan, M.A. Wong, Algorithm AS 136: A k-means clustering algorithm. J Royal
Statistical Society Series C (Applied Statistics) 28(1), 100–108 (1979)
24. J.C. Bezdek, Pattern Recognition with Fuzzy Objective Function Algorithms. (Springer Science
& Business Media, 2013)
25. S. Chiu, Method and software for extracting fuzzy classification rules by subtractive clustering.
In: Proceedings of North American Fuzzy Information Processing. (IEEE, 1996), pp. 461-5
26. S.L. Chin, An efficient method for extracting fuzzy classification rules from high dimensional
data. JACIII. 1(1), 31–36 (1997)
27. S.L. Chiu, Fuzzy model identification based on cluster estimation. Journal of Intelligent &
Fuzzy Systems. 2(3), 267–278 (1994)
28. M. Ester, H.P. Kriegel, J. Sander, X. Xu, A density-based algorithm for discovering clusters in
large spatial databases with noise. In: KDD Conference. (1996), pp. 226–31
29. S.K. Joo, C.C. Liu, J.W. Choe, 2001 Power Engineering Society Summer Meeting. Conference
proceedings (cat. No. 01CH37262). (2001), pp. 1811-6
30. S.-K. Joo, C.-C. Liu, L.E. Jones, J.-W. Choe, Coherency and aggregation techniques incorpo-
rating rotor and voltage dynamics. IEEE Trans. Power Syst. 19(2), 1068–1075 (2004)
31. Q. Zhu, J. Chen, X. Duan, X. Sun, Y. Li, D. Shi, A method for coherency identification based on
singular value decomposition. In. 2016 IEEE Power and Energy Society General Meeting
(PESGM). (2016), pp. 1-5
32. E. Cotilla-Sanchez, P.D. Hines, C. Barrows, S. Blumsack, M. Patel, Multi-attribute partitioning
of power networks based on electrical distance. IEEE Trans. Power Syst. 28(4), 4979–4987
(2013)
33. M. Davodi, H. Modares, E. Reihani, M. Davodi, A. Sarikhani, Coherency approach by hybrid
PSO, K-means clustering method in power system. In 2008 IEEE 2nd International Power and
Energy Conference. (2008), pp. 1203-7
34. G.-C. Pyo, J.-W. Park, S.-I. Moon, Coherency identification of generators using a PAM
algorithm for dynamic reduction of power systems. Energies 5(11), 4417–4429 (2012)
35. L. Kaufman, P. Rousseeuw, Clustering by means of Medoids in statistical data analysis based
on the L1–norm and related methods. (Y. Dodge, Dü.) reports of the Faculty of Mathematics
and Informatics. (Delft University of Technology, 1987)
36. H. Spath, Cluster Analysis Algorithms for Data Reduction and Classification of Objects. (Ellis
Horwood Chichester, 1980)
37. G. Pyo, J. Park, S. Moon, A new method for dynamic reduction of power system using PAM
algorithm. In. IEEE PES General Meeting. (2012), pp. 1–7
38. S.C. Wang, P.H. Huang, Fuzzy c-means clustering for power system coherency. In. 2005 IEEE
International Conference On Systems, Man and Cybernetics. (2005). pp. 2850-5
39. R. Krishnapuram, A. Joshi, O. Nasraoui, L. Yi, Low-complexity fuzzy relational clustering
algorithms for web mining. IEEE Trans. Fuzzy Syst. 9(4), 595–607 (2001)
176 M. H. Rezaeian Koochi et al.

40. I. Kamwa, A.K. Pradhan, G. Joós, Automatic segmentation of large power systems into fuzzy
coherent areas for dynamic vulnerability assessment. IEEE Trans. Power Syst. 22(4),
1974–1985 (2007)
41. I. Kamwa, A.K. Pradhan, G. Joos, S. Samantaray, Fuzzy partitioning of a real power system for
dynamic vulnerability assessment. IEEE Trans. Power Syst. 24(3), 1356–1365 (2009)
42. F. Znidi, H. Davarikia, M. Arani, M. Barati, Coherency detection and network partitioning
based on hierarchical DBSCAN. In. 2020 IEEE Texas Power and Energy Conference (TPEC).
(2020), pp. 1-5
43. R. Agrawal, D. Thukaram, Support vector clustering-based direct coherency identification of
generators in a multi-machine power system. IET Generation, Transmission & Distribution. 7
(12), 1357–1366 (2013)
44. R. Agrawal, D. Thukaram, Identification of coherent synchronous generators in a multi-
machine power system using support vector clustering. In. 2011 International Conference on
Power and Energy Systems. (2011), pp. 1-6
Chapter 9
Voltage Stability Assessment in Power
Grids Using Novel Machine Learning-Based
Methods

Ali Mollaiee, Sasan Azad, Mohammad Taghi Ameli,

and Morteza Nazari-Heris

9.1 Introduction

Voltage instability is a crucial issue that has been challenging power system engi-
neers over the past two decades [1, 2]. This issue typically occurs in power systems
that suffer increasing load demand and lagged development of the transmission
system. As a result, power is delivered near power system limits, and the potential
of instability is significantly increased. The voltage collapse can lead to power grid
blackouts, which cause significant economic losses and affecting unpredictable
impacts on residents’ lives and industrial production [3–5]. To mitigate the risk of
voltage collapse, stability analysis should be considered during both planning and
online operating of power systems. In contrast to planning, online analysis requires
online voltage stability assessment to take remedial actions for preventing possible
voltage collapse.

A. Mollaiee · S. Azad · M. T. Ameli (*)

Department of Electrical Engineering, Shahid Beheshti University, Tehran, Iran
e-mail: [email protected]; [email protected]; [email protected]
M. Nazari-Heris
Department of Architectural Engineering, Pennsylvania State University,
State College, PA, USA
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 177
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_9
178 A. Mollaiee et al.

9.1.1 Literature Review

The classic voltage stability assessment methods utilize static analysis based on
power flow methods such as the Gauss-Seidel or Newton-Raphson method. In
reference [6–10], various voltage stability indexes using conventional power flow
have been proposed. The main drawback of these techniques is the singularity of the
Jacobian matrix at the maximum loading point. To address this issue, the continu-
ation power flow (CPF) method is employed to obtain the voltage stability margin
(VSM) [11]. Lee et al. [12] proposed a P-Q-V curve base technique for VSM
assessment, which indicates the maximum limits of power demands. Due to the
increasing size and complexity of the modern power system, classic power system
stability analysis is becoming highly computational and time-consuming. Therefore,
the aforementioned techniques are not effective for online applications.
In the last few years, machine learning (ML)-based techniques such as artificial
neural networks (ANNs), decision trees (DTs), support vector machines (SVMs),
and ensemble methods have been considered in various studies due to their capa-
bility to solve nonlinear and complex problems independent of system modeling
with desired speed and accuracy [13, 14]. The multilayered perceptron (MLP) neural
network trained using the back-propagation algorithm was first employed in [15, 16]
to obtain the VSM that utilizes the energy method. Arya et al. [17] proposed
applying a radial basis function (RBF) network to approximate the probabilistic
risk of voltage instability for several operating conditions. In the proposed work, the
database has been generated using the Monte-Carlo simulation. In [18], input feature
selection based on neural networks using mutual information is employed to esti-
mate the voltage stability level for several scenarios according to load condition and
contingency. Debbie et al. [19] proposed an ANN-based method to estimate the
VSM of the power system under normal and N-1 contingency operating conditions
(OC). Devaraj et al. [20] proposed a new online monitoring technique using ANN to
estimate the VSM of the power system based on synchrophasor measurement under
normal and under N-1 contingencies OCs. In [21], various input feature reduction
techniques are applied to the RBF network to estimate the voltage stability level with
enhanced prediction accuracy. Hashemi et al. [22] use a multi-resolution wavelet
transform and principal component analysis for feature extraction of voltage profile
along with RBF network to approximate VSM. In [23], a multilayer feedforward
artificial neural network (MLFFN) and radial basis function network (RBFN) are
employed to assess power system static security assess. This paper uses a composite
security index for contingency ranking and security classification. Bahmanyar et al.
[24] proposed a new approach to obtain the optimal input variables required to
estimate the VSM using ANN. In [25], a Z-score-based bad data processing algo-
rithm is employed to improve the estimation accuracy of the feedforward ANNs.
Walter M et al. [26] presented a novel approach for VSM estimation that combines a
kernel extreme learning machine (KELM) with a mean-variance mapping optimiza-
tion (MVMO) algorithm. In [27], the association rules (AR) technique is used to
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 179

select the most effective loading parameters for the main input of the adaptive neuro-
fuzzy inference system (ANFIS).
Generally, ANNs are known as a powerful and flexible method for carrying out
nonlinear regression; nevertheless, they have some issues with training time and
overfitting. Support vector machine (SVM) is another powerful machine learning
technique that uses various kernel functions to perform classification and regression
problems. In [28], a multi-class SVM method is applied to classify the power
system’s security level, either normal, alert, emergency_1, or emergency_2. More-
over, the enhanced multi-class SVM has been employed using the pattern recogni-
tion approach for security assessment [29, 30]. Suganyadevi et al. [31] proposed a
support vector regression (SVR) model to assess the voltage stability of the power
system consolidating flexible alternating current transmission systems (FACTS)
devices. In [32], the v-SVR and ε-SVR models with RBF and polynomial kernel
functions have been applied to estimate VSM. Sajan et al. [33] proposed a genetic
algorithm-based support vector machine (GA-SVM) approach to estimate the volt-
age stability margin index (VSMI). In this work, the optimal values of SVM
parameters are obtained using the genetic algorithm. In [34], a new least-square
SVM using synchrophasor measurements is employed to estimate voltage stability
based on online learning.
DT is considered a fast and accurate ML technique for the development of
classification and regression models. Furthermore, DT training results’ interpretabil-
ity is a significant advantage over other ML techniques such as SVM and ANN.
Firstly, in the subject of power systems, Wehenkel employed DT to assess the
transient stability of power systems [35]. Further, DT has been applied for security
assessment applications [36, 37]. In [38], DT has been employed for online voltage
stability assessment using wide-area measurements. Zheng et al. [39] employed a
regression tree to predict the power system stability margin based on the VSM and
oscillatory stability margin (OSM). DT can be consolidated with other algorithms
such as principal component analysis (PCA) or fuzzy logic (FL) to enhance the
training performance. Mohammadi et al. [40] proposed a hybrid model for the online
voltage security assessment using a reduced predictor set extracted by PCA. In [41],
a contingency grouping method was developed for deriving DTs to assess the power
system security considering multiple contingencies. Meng et al. [42, 43] employed
participation factor analysis and relief algorithm to select attributes for DT. Recently,
ensemble methods have been introduced to improve the accuracy of DTs, and they
combine several base models to produce one optimal predictive model. Beiraghi
et al. [44] proposed a DT-based method for online voltage security assessment using
wide-area measurements. In the proposed work, the ensemble methods such as
bagging and adaptive boosting (AdaBoost) are employed to improve the voltage
security assessment of the power system performance. Su and Liu [45] proposed a
novel online learning framework for monitoring voltage stability using wide-area
measurements. In the proposed study, a new enhanced online random forest (EORF)
model based on the drift detection and online bagging techniques is implemented,
enabling online update of the trained model instead of reconstructing an entire
model. In [46], a novel methodology is presented for real-time assessment of
180 A. Mollaiee et al.

short-term voltage stability (STVS) based on multivariate time series. In this study,
random forest (RF) and symbolic representation technique is employed to classify
the power system stability status in multiple class using the maximal Lyapunov
exponent and various dynamic voltage indices. Dharmapala et al. [47] proposed an
ML approach to predict the long-term VSM based on loadability margin (LM). In the
proposed study, random forest regression (RFS) is employed to estimate LM using
different voltage stability indices (VSI).

9.1.2 Contributions and Novelties

In this chapter, the voltage stability assessment using ML technique’s voltage

stability assessment has been studied. As mentioned above, classic methods utilize
CPF to calculate VSM and determine whether the corresponding operating point
(OP) is stable or not. The obtained results using analytical methods are deterministic,
and CPF is highly computational and can’t be applied for online assessment. On the
other hand, ML techniques are widely applied in voltage stability assessment
according to their potential to train a model for complex systems. ML techniques
have a major drawback: Predictive accuracy is related to input data, which can affect
actual performance in practical application. Besides, a variety of ML techniques and
several approaches in performance enhancement complicated the implementation of
an optimal model. To overcome these problems, an online voltage stability assess-
ment framework using ML techniques is proposed to develop the most appropriate
ML model for the online voltage stability assessment based on phasor measurement
unit (PMU) data.

9.1.3 Structure of Chapter

The rest of the proposed chapter is presented as follows: The problem statement of
static voltage stability and some mathematical preliminaries are introduced in Sect.
9.2. Section 9.3 explains the details of the proposed framework for voltage stability
assessment. Section 9.4 presents the numerical results of case studies. Finally, the
conclusion appears in Sect. 9.5.

9.2 Voltage Stability Problem Statement

The steady-state power system model can be formulated in (9.1). Where x is system
state-variable, which is formed by a vector of power grid voltage magnitudes and
phase angles. And λ is the loading factor related to load and generator power as
follows [11]:
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 181

Fig. 9.1 P-V curve for an

operation state

F ðx, λÞ ¼ 0 ð9:1Þ
PL ¼ P0L þ λPdL ð9:2Þ

QL ¼ Q0L þ λQdL ð9:3Þ

PG ¼ P0G þ λPdG ð9:4Þ

In the above equation, P0L , Q0L , and P0G are the base load and generator powers,
whereas PdL , QdL , and PdG are the load and generator power increasing directions. In
practice, PdL , QdL , and PdG can be obtained from load forecasting and generation
dispatch, respectively [44].
Figure 9.1 represents a P-V curve for the specific operating condition. P-V curves
are employed to visually illustrate VSM, which corresponds to the distance from the
current OP to the critical point. For each OC, a P-V curve can be obtained using CPF.
The CPF program calculates the maximum loading level starting from a specified
initial operating point. The VSM is calculated through (9.5); therefore, the voltage
stability index (VSI) can be defined as the percentage of the maximum loading level
at the critical point.

VSM ¼ Pmax  P0 ð9:5Þ

VSM
VSI ¼ ð9:6Þ
Pmax

In the proposed study, the long-term voltage stability margin for the power
system is defined using the voltage stability criterion [48]. According to the
WECC criteria, the voltage stability index criteria is set at 7% for normal operating
182 A. Mollaiee et al.

conditions, while s is set at 5% for single-contingency conditions. Subsequently,

each operating point is labeled as “Stable” (S) if VSI < VSIc; otherwise, it is labeled
as “Unstable” (U).

9.3 The Proposed Framework

This chapter proposes a novel methodology in the online voltage stability assess-
ment using the data mining and machine learning approach. The proposed method-
ology demonstrates the performance and effectiveness of ML techniques in the
online voltage stability assessment. As shown in Fig. 9.2, the framework involves
4 stages: (1) database generation, (2) ML techniques training, (3) performance
evaluation, and (4) online application. The OCs are characterized using nominated
topology scenarios and load variation patterns (e.g., residential, industrial, or agri-
cultural load) to cover contingency situations and various load types. Various

Fig. 9.2 Flowchart of the proposed framework

9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 183

predictor sets are obtained using feature selection and feature extraction methods
based on power system variables. Furthermore, an importance sampling approach is
employed to enhance the speed and the accuracy of training. The utilized ML
technique is optimized using hyperparameters tuning based on the grid search.
Several metrics are conducted to evaluate the performance of the implemented
model. In addition to numerical metrics, the confusion matrix is also employed to
visually describe the classification model’s performance. Also, the PMU measure-
ments infected with noisy and missing data are employed to assess the robustness of
the framework in the presence of PMU uncertainty. It is remarkable to mention that
all ML-implemented models validated using the K-Fold cross-validation. These
stages are explained in detail in the following sections.

9.3.1 Database Generation

ML techniques employ datasets to build a model and predict results. Therefore,

achieving a comprehensive database is crucial to obtain an efficient and accurate
model. Either historical data or synthetic data can be applied to build a database of
the power system OPs for ML applications. Notwithstanding the historical data
involving actual information of the power system, it can be lacked details to
represent system status, precisely. Therefore, the synthetic data based on prior
knowledge can be employed to obtain a sufficient database. In the proposed frame-
work, various topology scenarios, load variation patterns, and the random sampling
method are applied to generate an eligible database, which spots all possible
operating points. Various combinations of the power system’s primary variables
(V, δ, P, Q) are considered predictors to indicate the power system’s state in each
OP. Eventually, importance sampling is employed to remove useless samples and
enhance training performance.

9.3.1.1 Sampling

As illustrated in Fig. 9.3, database generation approach consists of three loops to

obtain required samples for training:
1. Load Variation Patterns: There are various load types in realistic power
systems that each type is indicated with specified increasing characteristics.
Therefore, loads of the whole system can’t increase at the same rate. To address
this issue, the system is divided into some areas to define the load variation
pattern [44]. In each load variation pattern, the loads in different areas increase at
different rates, and all the generators balance the incremental load. The dividing
approach can be based on the similarity of load types (e.g., residential and
industrial) or the same geographical region.
184 A. Mollaiee et al.

Fig. 9.3 Proposed approach for generating samples

2. Topology Scenarios: To contemplate both normal and contingency situations,

single transmission lineouts are considered as a contingency. In contingencies
that power flow never converge is neglected. Hence, the number of total topology
scenarios and the overall number of OCs can be calculated as:
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 185


N scenario,total ¼ N lineout,total  N lineout,deverged þ 1 ð9:7Þ
N OC,total ¼ N scenario,total  N load pattern ð9:8Þ

3. Random Sampling: For determined OCs in the last steps, CPF is conducted
to calculate VSM. Furthermore, to obtain a more extensive database, a
random sampling approach is applied for each OC. Final load and genera-
tion variation can be represented as:


j
PL,i ¼ 1 þ λK L,i 1 þ k j PL0,i 1 < i < N Area , 1 < j < N random ð9:9Þ



j
QL,i ¼ 1 þ λK L,i 1 þ k j QL0,i 1 < i < N Area , 1 < j < N random ð9:10Þ


j
PG,i ¼ 1 þ λK Gj PG0,i 1 < j < N random ð9:11Þ
P

K L,i 1 þ k j PL0,i
i2fAreag
K Gj ¼ P ð9:12Þ
PG0,n
n2fPV g

j j
where PL,i and QL,i are the load active power and reactive power in each O. Also,
j
PG,i is the generated active power, respectively. KL, i is indicate the load variation
coefficient for each area, and kj is a random number. K Gj is a factor that determines
the rate of increase in the production in generators. Finally, for each generated
sample, the input variable X is constructed by combining various power system
variables such as line-injected power or voltage of bus that are obtained from power
flow calculation. Also, the target variable is labeled as “Stable” or “Unstable”
according to WECC criteria for corresponding OC mentioned in Sect 9.2.

9.3.1.2 Importance Sampling

Importance sampling (IM) is a useful technique to shrink the database without the
risk of losing information [44, 49]. Also, this technique reduces training time and
improves accuracy. In the proposed framework, IM is employed to remove
extremely heavy or light-loaded operating points and keep samples within the
vicinity of VSM. To discriminate samples more comprehensively, the boundary
for each topology scenario is calculated distinctly as:

PiL, min < PiL,total < PiL, max 1 < i < N scenario,total ð9:13Þ
186 A. Mollaiee et al.

PiL, max ¼ ð1 þ Lmax ÞVSM i, max 1 < i < N scenario,total ð9:14Þ

PiL, min ¼ ð1 þ Lmin ÞVSM i, min 1 < i < N scenario,total ð9:15Þ

where Lmin and Lmax is criteria for the minimum and maximum limit of the boundary,
respectively. Further, VSMi, max is the highest, and VSMi, min is the lowest stability
margin for the determined topology scenario.

9.3.2 Training the Machine Learning Technique

In order to build an optimal model, the ML technique must be chosen carefully with
regard to state of the problem. The online voltage stability assessment is a classifi-
cation problem using supervised learning to determine whether system status is
stable or unstable. As discussed previously, SVM, DT, and ensemble methods are
capable techniques to apply in the online voltage stability assessment. Even though
the proposed framework was adopted to utilize various techniques to train data,
particularly, the ensemble method is employed to implement the ML model in the
voltage stability assessment with respect to the impressive performance [44–46]. Fur-
thermore, dimensionality reduction approaches such as feature selection [50, 51] and
feature extraction [22, 27, 52, 53] are augmented to model to enhance training
efficiency. Eventually, the implemented model can be optimized using
hyperparameter tuning to obtain the online voltage stability assessment’s most
satisfactory performance.

9.3.2.1 Ensemble Method

The ensemble methods aim to combine multiple classification models with a given
learning algorithm to improve performance over a single classifier. Various classi-
fication models can be utilized to compose the ensemble model. But DT is the base
model to combine and construct a set of classifiers generally. Learning algorithms
are applied to ensemble methods categorized into two major groups, the first known
as averaging methods such as bagging trees and random forest. In these methods,
several classifiers are combined independently, and the final result is obtained
through averaging single classifiers’ predictions. The second group is called
boosting. The boosting methods compose classifiers sequentially to combine the
outputs from the weak learner and create a powerful ensemble learner that eventually
enhances the base learner’s performance. Bagged tree and AdaBoost are considered
as primary techniques utilized in the proposed framework. Details of which are given
in the following subsections.
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 187

Bagged Tree

Bootstrap aggregating, often abbreviated as “bagging” is a type of ensemble learn-

ing algorithm employed to reduce the error generated by small fluctuations in the
dataset and defined as the variance. To bag a weak learner such as a DT on a dataset,
multiple bootstrapped subsamples of the dataset are generated. A DT is trained on
each bootstrapped subsample. After subsamples are trained using DT, a voting
algorithm is applied to aggregate DTs and form the most efficient classifier. Besides,
independent DTs can be trained using parallel processing that improves speed
significantly. As shown in Fig. 9.4, bootstrapping conducted random selection
with a replacement, which means a sample can appear multiple times for each
subset. Therefore, some of the observations can be excluded from the subset called
“out of bag” samples. Usually, the voting approach is applied using equal weights
although in online applications, weights can be updated instead of training a new
model. Random forest (RF) is a bagging technique that combines the random
decision trees with the bagging algorithm to achieve high-classification accuracy.
As mentioned, the bagging technique reduces variance and can avoid overfitting
issues. Thus, this technique is an appropriate choice for applying higher dimension-
ality data. Also, it can maintain performance for missing data.

AdaBoost

Boosting methods work in the same way as bagging methods: Combine a group of
aggregated models to obtain an ensemble model with better performance. However,
unlike bagging, which aims at reducing variance and avoiding overfitting, the
boosting algorithm’s purpose is to reduce bias. Boosting, like bagging, can be
used for r both regression and classification applications. Adaptive boosting
(AdaBoost) and gradient boosting are well known as boosting algorithms. AdaBoost
involves training a sequence of weak learners on repeatedly modified versions of the

Fig. 9.4 Bootstrap-

aggregating algorithms for
training multiple DT
188 A. Mollaiee et al.

data. Further, all DTs’ predictions are combined using a weighted voting approach to
produce the final prediction. Eventually, an ensemble learner with lower bias and
even better variance is obtained. The main procedure of the AdaBoost algorithm
considering a DT as the base classifiers can be represented in major steps as follows
[44]:
• Step 1) Input data:
1-. Load training set with N samples.
2-. Determination of the boosting iteration T and base DT specification.
3-. Training samples weight set equally as

1
wn ¼ ð9:16Þ
N

• Step 2) Boosting approach:

1-. Start first iteration: t ¼ 1.
2-. Fit DT using the initial training set.
3-. Calculate the learning error as

X
N
εt ¼ wtn IðDT t ðX n Þ 6¼ yn Þ ð9:17Þ
n¼1

where εt indicates the weighted classification error, and I is a logical function that
returns 1 when the corresponding condition was satisfied. Otherwise, it’s 0.
4-. Calculate the weight for the trained DT

 
1 1  εt
α ¼ log
t
ð9:18Þ
2 εt

5-. Check iteration t < T.

6-. Update the weight of samples.

wtn
wtþ1
n ¼ exp ðθn αt Þ ð9:19Þ
Zt

where θn is an indicator function that is θn ¼  1 when DTt(Xn) ¼ yn to decrease

correctly classified samples weights and θn ¼ 1 when DTt(Xn) 6¼ yn to increase
correctly misclassified samples weight.
7-. Go to next iteration t ¼ t + 1.
• Step 3) Build the final model:
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 189

Fig. 9.5 Booting Algorithm for update weights

After the boosting iterations are completed, based on calculated weights, an

ensemble model is obtained. Finally, the AdaBoost model computes prediction
for new data using:

X
T
DT ensemble ðX new Þ ¼ αt DT t ðX new Þ ð9:20Þ
t¼1

As shown in Fig. 9.5, when a DT misclassifies a sample in the sequence.

Corresponding weight is increased so that the next DT take more efforts to predict
inconvenient sample correctly. Weak models that are only slightly better than
random guessing is often employed for boosting to reduce the bias error, such as
small decision trees. Another important motivation to use weak classifiers with low
variance for boosting is that fitting these models requires less computational effort.
Indeed, as the boosting algorithm can’t be done in parallel (unlike bagging), it could
become time-consuming to obtain the final model using complex learners. Another
drawback of using AdaBoost is the risk of overfitting according to higher variance.

9.3.2.2 Dimensionality Reduction

Fitting a predictive model to the dataset with numerous features is computationally

expensive. Also, inattentive hyperparameter tuning increases the risk of overfitting.
Dimensionality reduction is a useful technique that can be employed in the ML
model to transform training data from a high-dimensional space into a
low-dimensional space so that the low-dimensional representation keeps important
190 A. Mollaiee et al.

essential information of the original data. According to the power system, it is a

large-scale and complex system, and the power system’s state can be represented by
various features and measurements, so the voltage stability assessment is considered
a high-dimensional space problem. Besides, utilizing PMU wide-area measurements
in the voltage stability assessment requires online frameworks with desirable speed
and accuracy. Consequently, augmenting dimensionality reduction techniques to
ML-based framework is crucial to reduce training data size, speed up the calculation,
improve accuracy, and avoid overfitting. Dimensionality reduction approaches can
also be divided into feature selection and feature extraction.

Feature Selection

Feature selection is the process of identifying and selecting relevant features to the
target variable. Generally, feature selection employed an evaluation process that
scores the features. Further, a filtering criterion is applied to pick the most relevant
feature for the ML technique input predictors. There are various methods for scoring
features, and these methods are divided into three main approaches: knowledge-
based, ML-based, and statistical-based. In the knowledge-based approach, practical
experiences and technical knowledge are employed to indicate the most important
features associated with targets. In the ML-based approach, after the model is
trained, the relative rank represented by the model is utilized to arrange features
concerning the target values’ predictive ability. Also, the statistical-based methods
select variables regardless of the model and prior knowledge. These methods use a
statistical test such as correlation to determine the most interesting features. Univar-
iate feature selection is a statistical-based method that scores each feature individu-
ally to evaluate the feature’s relevance with the response variable. In the proposed
framework, all three approaches are applied as a preprocessing module to choose
optimal features set.

Feature Extraction

Unlike feature selection, feature extraction rebuilds whole new features intended to
be informative as original features. Brand-new features have reduced the dimen-
sionality of input data and also remove redundant features. Thus, training ML
techniques using extracted features improve the classification performance.
Extracted features can’t be as interpretable as original features, so they can’t be
applied to corresponding preventive control. Besides, training ML techniques using
feature extraction diminish the robustness of the model for noisy or missing data.
There are various feature extraction algorithms such as principal component analysis
(PCA), linear discriminant analysis (LDA), and autoencoder. The proposed frame-
work uses PCA as a feature extraction method to reconstruct new and enhanced
predictors.
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 191

The principal component analysis is defined as a linear transformation that maps

data to a lower-dimensional space to maximize the variance of the data.
Low-dimensional space is a new coordinated system in which each axis represents
one of the principal components. The covariance matrix of the data is calculated to
perform PCA, and then eigenvalues and corresponding eigenvectors of this covari-
ance matrix are obtained. The eigenvectors that correspond to the largest eigenvalues
can be interpreted as the principal components. Finally, these eigenvectors are
applied to obtain new features set with a large fraction of the variance. Transforma-
tion of original data to low-dimensional space is performed as [54, 55]

t ki ¼ X i W k k ¼ 1, 2, . . . L i ¼ 1, 2, . . . N ð9:21Þ

where L is the number of extracted space dimensional, and N is the number of input
data dimensional, respectively. X is input features data that is subtracted the mean of
each feature. W is vectors of coefficients to map data, and t is the principal
component scores representing extracted features.

9.3.2.3 Hyperparameter’s Tuning

A hyperparameter is a parameter of the ML model whose value is used to control the

learning process such as the number of trees in the bagged tree technique.
Hyperparameter tuning is an approach that finds the optimal ML model with respect
to input data and the state of the problem. In the proposed framework,
hyperparameter tuning is utilized to optimize ML model’s performance using grid
search. Grid search is a simple approach to find the best values for hyperparameters.
It works by searching exhaustively through a determined boundary for
hyperparameters.

9.3.3 Performance Evaluation

Performance evaluation of the implemented ML model is crucial to determine the

model’s efficiency in the online voltage stability assessment using wide-area mea-
surements. However, accuracy and CPU time are employed as primary metrics in
performance evaluation. They can’t represent comprehensive detail of training
performance. When training data are distributed unbalanced, the accuracy is inade-
quate to demonstrate classification performance. The training set for voltage stability
assessment may be imbalanced according to the number of stable operating points in
the power system that is much more than the number of unstable points. Therefore, a
new assessment approach is necessary to be employed in the ML model evaluation.
The confusion matrix useful tool is employed to demonstrate the performance of
classification.
192 A. Mollaiee et al.

Fig. 9.6 Confusion matrix

for the voltage stability
assessment problem

As illustrated in Fig. 9.6, confusion matrix for voltage stability assessment is 2*2
table, in which table rows represent true labels, and columns represent predicted
labels. Each cell of the table demonstrates a portion of data based on the
corresponding labels. The confusion matrix describes the classification performance
visually. Also, it can be employed to calculate several measures as [44, 56]

N TrueStable þ N TrueUnstable
AccuracyðAccuÞ ¼ ð9:22Þ
N Stable þ N Unstable
N FalseStable
MisdetectionðMisÞ ¼ ð9:23Þ
N Stable þ N Unstable
N FalseUnstable
Fasle alarmðfalÞ ¼ ð9:24Þ
N Stable þ N Unstable
N TrueUnstable
ReliabilityðRelÞ ¼ ð9:25Þ
N TrueUnstable þ N FalseStable
N TrueStable
SecurityðSecÞ ¼ ð9:26Þ
N TrueStable þ N FalseUnstable
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
N TrueStable N TrueUnstable ð9:27Þ
G mean ¼ :
N TrueStable þ N FalseUnstable N TrueUnstable þ N FalseStable

Misdetection (Mis) represents the ratio of misclassified unstable cases, while

false-alarm (Fal) indicates the proportion of stable samples predicted unstable.
Besides, security and reliability stand for the proportion of secure cases that are
correctly identified and the proportion of insecure cases correctly identified. G-mean
(G) acts as a geometric mean of the security and reliability that evaluates learning’s
overall performance. Subsequently, K-Fold cross-validation is applied in the perfor-
mance evaluation process to assess the effectiveness model with respect to an
anonymous dataset. K-Fold cross-validation is a validation procedure that utilizes
random splitting data to guarantee every sample of the original dataset has the
chance of involving in training and test set.
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 193

9.3.3.1 PMU Uncertainly

As PMU measurements can be infected with noise or even missing, the ML model’s
performance should evaluate in noisy and missing data conditions to determine the
robustness and efficiency. Therefore, two scenarios are considered for noisy data, in
which the first scenario involves noise to all input data, and in the second scenario,
noise is only added to the test data. Further, to produce synthetic noisy data, total
vector error (TVE) criteria are employed. According to the IEEE standard [57], the
TVE criteria for PMU data uncertainly should be under 1%. The TVE for the voltage
measurements is represented in (9.28). For the missing data scenario, some PMUs
are picked randomly, and then corresponding measurements in the testing samples
for selected PMU are set to NaN.
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u  
ujV real,measured  V real,ideal j2 þ V imag,measured  Vimag,ideal 2
TVE ¼ t ð9:28Þ
V real,ideal 2 þ V imag,ideal 2

9.3.4 Online Application

In the last stage, after the developed model was approved in terms of accuracy,
speed, and robustness using the performance evaluation process, the final model can
be employed in online voltage stability assessment based on wide-area PMU
measurements.

9.4 Simulations and Results

The proposed framework has been tested on two different-size case study. The first
case is the new England 39-bus test system [58], and the second one is a bigger grid,
the IEEE 118-bus test system [59]. In the case studies, the simulation programs
including database generation, model training, and testing are performed within
MATLAB® environment. Furthermore, all the experiments are executed on a
system with Intel® Core™ i7 2.6 GHz processor and 16 GB of RAM.
194 A. Mollaiee et al.

9.4.1 IEEE 39-Bus Test System

9.4.1.1 Database Generation

This system involves 46 lines and 10 generators. In order to generate the database,
the grid is divided into three areas based on the topographical location of each bus.
Further, for each area, different confidences are picked to construct 20 load variation
patterns. The buses included in each area, along with some instances of the load
variation pattern, are presented in Table 9.1. Besides, in the topology scenario’s
definition, line 13 outage is excluded from contingencies according to power flow
that can’t converge in this scenario, and 45 + 1 topology scenarios are considered.
The database is eventually generated using the proposed sampling approach, which
involves 8155 stable samples and 7492 unstable. Figure 9.7 illustrates the distribu-
tion of the samples based on voltage magnitude.

Table 9.1 Existing buses in Area 1 buses Area 2 buses Area 3 buses
each area along sample
1–13 14–26 27–39
coefficients
K area 1 K area 2 K area 3
0 0 0
0 0 0.2
0 0.2 0
0.2 0 0

Fig. 9.7 Dataset sample distribution based on the voltage magnitude

9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 195

9.4.1.2 Simple ML Model Training

In order to illustrate the effectiveness of the proposed framework, a simple ML

model is essentially employed. This model employed the bagged tree techniques for
training and voltage magnitude of buses as predictors. The bagged tree model
involves 30 DTs with a depth of 15,646. The dataset was randomly split into two
subsets, where the training set includes 80% of the data and the remaining 20%
allocated to the test set. After training the ML model using the training set, the testing
set’s classification accuracy is evaluated as 80.70%. The predicted label’s distribu-
tion for the testing set samples is shown in Fig. 9.8.

9.4.1.3 Optimal Dataset for Training

To represent the state of each sample, four combinations of the predictors are chosen
based on the power system primary variables as follows:
Vm: Represent the voltage magnitude of buses
Vm-δ: Represent the voltage magnitude and phase angle of buses
Pactive-reactive: Represent the injected active and reactive power of lines
Vm-δ-Pactive-reactive: Represent the voltage magnitude and phase angle of buses and
the injected active and reactive power of lines together
The importance sampling approach is applied to improve the classification
performance by removing irrelevant samples. The dataset size is reduced by 41%,
and the new dataset involves 4252 stable samples and 5260 unstable samples. Also,
random selecting is conducted to construct a dataset with the same size as the
reduced dataset and compare it with it. Four combinations of the predictors along
three types of samples set are utilized to form 12 different training sets presented in

Fig. 9.8 Distribution of the predicted labels

 196

Table 9.2 The number of samples and predicators for each dataset
Vm Vm-δ Pactive-reactive Pactive-reactive
Dataset Predictor Sample Predictor Sample Predictor Sample Predictor Sample
Original dataset 39 15,647 78 15,647 92 15,647 178 15,647
Importance sampling 39 9123 78 9123 92 9123 178 9123
Random selecting 39 9123 78 9123 92 9123 178 9123
A. Mollaiee et al.
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 197

Table 9.2. The performance of training various datasets by the bagged tree in the
previous section is demonstrated in Table 9.3. Table 9.3 indicates that a larger set of
predictors may give better performance. Also, reducing the dataset with the impor-
tance sampling approach will decrease computational time without significantly
improving accuracy.

9.4.1.4 ML Techniques Performance

As previously discussed, DT and SVM are powerful ML techniques employed in the

voltage stability assessment frequently. Besides, the bagged tree and the AdaBoost
are presented as ensemble methods that combine several single models to improve
classification performance. To compare these techniques and choose the finest ML
technique, four different models are employed. The specification of these ML
models is summarized in Table 9.4. Table 9.5 demonstrates the performance of
training models using the original dataset. The superiority bagged tree technique is
obvious in terms of accuracy and computational time. The DT is a fast technique but
suffers a lack of classification accuracy. Further, SVM presents the desired accuracy,
but it is heavily computational and may not be appropriate in the online voltage
stability assessment. AdaBoost, due to its higher variance overfitted, the perfor-
mance of classification is corrupted.

9.4.1.5 Dimensionally Reduction

In the proposed framework, three different approaches for feature selection tech-
niques are conducted as follows:
Knowledge-base: In this approach, buses with the most voltage deviation and
lines with the most loading are chosen as critical elements of the power system.
Further, the most important input features are selected based on critical element
variables.
ML-based: The bagged tree model trained in the previous section is employed to
score features concerning the predictor importance that is evaluated by the
bagged tree.
Statistical-base: In this approach, the relevance of the features with the target
variable is evaluated using the chi-square test.
Figure 9.9 represents feature scores that are evaluated using the ML-based and
statistical-based approaches. Also, PCA is employed to rebuild new input features
with higher variance. As shown in Fig. 9.10, the distribution of samples using
principal components has a greater distinction than the distribution based on selected
variables. For a better comparison of the different methods, the new feature set’s size
is considered equal to 23 of the number of original features. Therefore, 25,50,60 and
110 features are picked up for the Vm,Vm-δ, Pactive-reactive, and Vm-δ,-Pactive-reactive
predictors set, respectively. The performance of various dimension reduction
 198

Table 9.3 Classification accuracies and training CPU time for each dataset
Vm Vm-δ Pactive-reactive Pactive-reactive
Dataset Accuracy (%) Time (s) Accuracy (%) Time (s) Accuracy (%) Time (s) Accuracy (%) Time (s)
Original dataset 80.70 65 85.36 59.75 95.16 64.53 94.87 80.84
Importance sampling 80 31.35 85 40.03 95.5 30.79 95 49
Random selecting 78 39.46 82 38.79 92 43.28 91 52.85
A. Mollaiee et al.
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 199

Table 9.4 Specification of employed ML models

ML technique Specification
DT Algorithm Max number of splits Split criterion
Cart 100 Gdi
SVM Kernel Solver
Radial basis function Iterative single data algorithm
AdaBoost Learning rate Number of trees Max number of splits
0.1 30 30
Bagged tree Learning rate Number of trees Max number of splits
0.1 30 15,646

methods is demonstrated in Table 9.6. From the results shown in Table 9.6, it is
observed that the training speed of all proposed methods is improved according to
input predictors reduction. Further, there is a minor downgrade in classification
accuracy for knowledge-based and statistical-based methods. The results show that
PCA proposes the best dimensionally reduction method, which enhances speed and
accuracy simultaneously.

9.4.1.6 Hyperparameters Tuning

All trained models in the previous section are optimized using hyperparameters
tuning. The tuning approach is applied with the grid search to find the best param-
eters for the model. Figure 9.11 illustrates the hyperparameters searching process for
the bagged tree with the Vm predictors that involve three primary parameters: the
number of DT, the max depth in each DT, and the number of principal components
as the input features. Furthermore, a trade-off between speed and accuracy is
required to pick optimum values with respect to the problem state. The chosen
optimal values for the bagged three parameters are indicated with stars in Fig. 9.11.

9.4.1.7 Performance Evaluation

As mentioned in Sect. 9.3.3, the confusion matrix is an effective tool to demonstrate

the classification performance. The confusion matrix for the DT and the bagged tree,
which are trained using the Vm predictors set, is shown in Fig. 9.11. Further,
calculated measures are based on the confusion matrix presented in Table 9.7. The
table clearly shows that the bagged tree has better classification performance than the
DT, especially in classification reliability, which means the bagged tree can separate
stable samples from unstable samples more accurately (Fig. 9.12).
In addition to confusion matrix-based metrics, the robustness of the proposed
approach is also evaluated. For the considered simulation, all buses of the power grid
are assumed to be equipped with PMU. Therefore, the Vm-δ predictors set is
considered as the input features. Let p be the percentage of the PMUs with noisy
 200

Table 9.5 Classification accuracy and training CPU time for each technique
Vm Vm-δ Pactive-reactive Pactive-reactive
ML technique Accuracy (%) Time (s) Accuracy (%) Time (s) Accuracy (%) Time (s) Accuracy (%) Time (s)
DT 69 5.26 78.06 9.28 83.40 11 83.33 21.31
SVM 80.53 368.6 83.91 404.85 89.39 331.34 89.88 494.56
AdaBoost 69.27 105.57 77.08 219.53 84.37 230.70 87.63 497.82
Bagged tree 80.70 65.10 85.35 59.75 95.16 64.53 96.02 80.84
A. Mollaiee et al.
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 201

Fig. 9.9 Scores for the voltage-related features. (a) Univariate feature selection. (b) Predictor
importance for bagged tree

data. So synthetic noisy data can be generated based on the PMU TVE criterion and
the number of PMUs that are infected with noise. Classification accuracy using the
bagged tree for the first scenario in which the noise included in the whole dataset is
demonstrated in Table 9.8. Further, results for the second scenario that involves
noisy data for the testing set, presented in Table 9.8 and Table 9.9, indicates that
adding noise to the whole dataset can improve the robustness of the training ML
model. Also, PCA is vulnerable to noise due to the high variance of the new
extracted space. It is worth noting that the DT has shown the most robustness against
noisy data according to its simple approach for learning.
Training performance in the presence of PMU missing data for given missing
PMU ratio is shown in Table 9.10. It can be seen from the table that a high rate of
missing data can ruin classification performance.
202 A. Mollaiee et al.

Fig. 9.10 Dataset samples distribution based on (a) selected voltage magnitude and (b) principal
component

9.4.1.8 Online Application

In order to demonstrate the effectiveness of the proposed method in online applica-

tions, a synthesized 24-hour load curve is considered. The load curve involves
100 district points to represent various electric energy demands during the day.
The trained model in previous sections is conducted to online voltage stability
assessment based on the daily load curve. As shown in Fig. 9.13, the trained
model has been capable to accurately detect instability during peak hours.
Table 9.6 Classification accuracy and training CPU time for each method
Vm Vm-δ Pactive-reactive Pactive-reactive
ML technique Accuracy (%) Time (s) Accuracy (%) Time (s) Accuracy (%) Time (s) Accuracy (%) Time (s)
None 80.70 65.1 85.35 59.75 95.15 64.53 94.87 80.84
Knowledge-base 79.03 48.43 84.69 53.81 94.74 45.37 94.31 60.21
Statistical-base 79.65 52.54 83.83 56.1 93.72 49.62 92.30 64.54
Bagged tree-base 80.95 50.18 86.54 55.79 95.29 46.28 95.52 62.46
PCA 93.12 31.14 94.82 38.84 96.12 36.19 96.02 47.88
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . .
203
204 A. Mollaiee et al.

Fig. 9.11 Time and accuracy for corresponding parameter values. (a) Max number of splits. (b)
Number of PCA components. (c) Number of trees
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 205

Table 9.7 Evaluation measures for each dataset

Vm Vm-δ Pactive-reactive Pactive-reactive
ML Bagged Bagged Bagged Bagged
technique DT tree DT tree DT tree DT tree
Accuracy 68.99 80.70 78.06 85.36 83.40 95.16 83.33 94.87
(%)
Reliability 57.92 80.31 74.53 84.82 81.57 94.00 82.12 94.03
(%)
Security 81.06 81.11 81.90 85.93 85.40 96.38 85.80 95.82
(%)
G-mean (%) 68.52 80.71 78.13 85.37 83.46 95.20 83.94 94.91

Fig. 9.12 Confusion matrix. (Left) decision tree. (Right) bagged tree

Table 9.8 Accuracy for given noisy PMU ratio (p) – scenario 1
Accuracy (%)
p ¼ 0% p ¼ 10% p ¼ 50% p ¼ 90%
ML With With With With
technique Base PCA Base PCA Base PCA Base PCA
Tree 78.06 77.40 78.01 77.78 77.70 76.38 76.65 76.46
SVM 83.91 88.24 81.47 86.48 78.80 85.91 76.40 85.06
AdaBoost 77.08 80.76 77.31 77.45 76.68 76.79 76.35 77.30
Bagged tree 85.35 94.82 85 91.77 83.70 88.28 81.84 86.59

9.4.2 118-Bus Test System

The IEEE 118-bus test case is employed to demonstrate the effectiveness of the
proposed framework for a larger power system. The simulation details for this case
are summarized in Table 9.12. As shown in Fig. 9.14, the distribution of stable and
unstable samples is imbalanced, which means the number of stable OPs is signifi-
cantly greater than unstable in contract with the case 39-bus. The performance of
206 A. Mollaiee et al.

Table 9.9 Accuracy for given noisy PMU ratio (p) – scenario 2
Accuracy (%)
p ¼ 0% p ¼ 10% p ¼ 50% p ¼ 90%
ML With With With With
technique Base PCA Base PCA Base PCA Base PCA
Tree 78.06 77.40 72.66 65.87 73.69 59.77 72.95 56.95
SVM 83.91 88.24 62.36 79.13 47.82 47.89 47.88 47.74
AdaBoost 77.08 80.76 76.95 76.67 75.59 69.53 74.23 67.72
Bagged tree 85.35 94.82 84.46 75.64 82.31 64.69 79.12 59.03

Table 9.10 Accuracy for Accuracy (%)

given missing PMU ratio (p)
ML technique p ¼ 0% p ¼ 10% p ¼ 25% p ¼ 50%
Tree 78.06 74.33 74.33 66.03
SVM 83.91 52.12 52.12 52.12
AdaBoost 77.08 72.40 72.40 59.76
Bagged tree 85.35 72.29 72.29 58.76

Fig. 9.13 Online voltage stability assessment based on the daily load curve

classification using the bagged tree for the Vm predictors set is demonstrated in
Table 9.12. From the Table 9.11, it is observed that despite better accuracy than the
previous case, it has a lower G-mean due to the unbalanced distribution of samples.
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 207

Fig. 9.14 Dataset sample distribution based on the voltage magnitude

Table 9.11 Given details generate 118-bus system database

Specification
Powe system Bus Line Generator
118 186 54
Load variation Area Load variation pattern
3 12
Topology scenario Contingency Divergent contingencies
186 lineout +1 normal condition 7–9–113-116-133-134-176-177-183-
184
Dataset Total samples Stable samples Unstable samples
62,135 45,978 16,157

Table 9.12 Evaluation measures for given predictors set

Vm Vm-δ Pactive-reactive Pactive-reactive
ML With With With With
technique Base PCA Base PCA Base PCA Base PCA
Accuracy 82.74 91.30 84.81 95.36 93.78 96.50 93.31 96.90
Reliability 60.14 78.91 64.25 89.04 85.89 91.66 86.64 92.63
Security 91.76 96.26 92.48 97.89 97.04 98.44 96.76 98.62
G-mean 74.30 87.15 77.08 93.36 91.30 94.99 90.50 95.56
208 A. Mollaiee et al.

9.5 Conclusion

A framework for the development of the ML application in the online voltage

stability assessment is presented. The framework provides several options to imple-
ment in the optimal model. These options can be utilized in the database generation
like importance sampling or the training stage, i.e., feature selection, feature extrac-
tion, and hyperparameters tuning. The proposed framework was tested on the IEEE
39-bus and IEEE 118-bus test systems. The results of case studies demonstrate that
the bagged tree model with PCA presents the best classification performance in
terms of accuracy and speed the robustness of the framework in the presence of PMU
uncertainty is evaluated. So it indicates that a high rate of noisy or missing data can
ruin the voltage stability assessment using the ML model.
Relevant future works aspire to augment the online weighted majority voting in
the bagged tree technique. Weighted voting enables online model updating. There-
fore, the bagged tree can be adapted to unforeseen changes in system conditions.
Also, robust feature extraction approaches can be employed to ensure desirable
classification performance in the presence of noise and missing data.

References

1. Begovic, M., Voltage collapse mitigation. IEEE Power System Relaying Committee, Working
Group K12, 1997. IEEE Publ. No. 93, THO596-7PWR
2. V. Ajjarapu, B. Lee, Bibliography on voltage stability. IEEE Trans. Power Syst. 13(1), 115–125
(1998)
3. Z. Bo et al., An analysis of previous blackouts in the world: Lessons for China0 s power industry.
Renew. Sust. Energ. Rev. 42, 1151–1163 (2015)
4. M.Z. El-Sadek, Preventive measures for voltage collapses and voltage failures in the Egyptian
power system. Electr. Power Syst. Res. 44(3), 203–211 (1998)
5. O.P. Veloza, F. Santamaria, Analysis of major blackouts from 2003 to 2015: Classification of
incidents and review of main causes. Electr. J. 29(7), 42–49 (2016)
6. P. Kessel, H. Glavitsch, Estimating the voltage stability of a power system. IEEE Transactions
on Power Delivery 1(3), 346–354 (1986)
7. A. Tiranuchit, R.J. Thomas, A posturing strategy against voltage instabilities in electric power
systems. IEEE Trans. Power Syst. 3(1), 87–93 (1988)
8. C.W. Taylor, Power System Voltage Stability. (McGraw-Hill, 1994)
9. J. Hongjie, Y. Xiaodan, Y. Yixin, An improved voltage stability index and its application.
Int. J. Electr. Power Energy Syst. 27(8), 567–574 (2005)
10. R. Tiwari, K.R. Niazi, V. Gupta, Line collapse proximity index for prediction of voltage
collapse in power systems. Int. J. Electr. Power Energy Syst. 41, 105–111 (2012)
11. V. Ajjarapu, C. Christy, The continuation power flow: A tool for steady state voltage stability
analysis. IEEE Trans. Power Syst. 7(1), 416–423 (1992)
12. C.-Y. Lee, S.-H. Tsai, Y.-K. Wu, A new approach to the assessment of steady-state voltage
stability margins using the P– Q– V curve. Int J Elec Power Energ Syst 32, 1091–1098 (2010)
13. O.A. Alimi, K. Ouahada, A.M. Abu-Mahfouz, A review of machine learning approaches to
power system security and stability. IEEE Access 8, 113512–113531 (2020)
14. M. Amroune, Machine learning techniques applied to on-line voltage stability assessment: A
review. Archives of Computational Methods in Engineering (2019)
9 Voltage Stability Assessment in Power Grids Using Novel Machine. . . 209

15. B. Jeyasurya, Artificial neural networks for power system steady-state voltage instability
evaluation. Electr. Power Syst. Res. 29(2), 85–90 (1994)
16. A.A. El-Keib, X. Ma, Application of artificial neural networks in voltage stability assessment.
IEEE Trans. Power Syst. 10(4), 1890–1896 (1995)
17. L.D. Arya, L.S. Titare, D.P. Kothari, Determination of probabilistic risk of voltage collapse
using radial basis function (RBF) network. Electr. Power Syst. Res. 76, 426–434 (2006)
18. D. Devaraj, J.P. Roselyn, R.U. Rani, Artificial neural network model for voltage security based
contingency ranking. Appl. Soft Comput. 7(3), 722–727 (2007)
19. D.Q. Zhou, U.D. Annakkage, A.D. Rajapakse, Online monitoring of voltage stability margin
using an artificial neural network. IEEE Trans. Power Syst. 25(3), 1566–1574 (2010)
20. D. Devaraj, J. Preetha Roselyn, On-line voltage stability assessment using radial basis function
network model with reduced input features. Int. J. Electr. Power Energy Syst. 33(9), 1550–1555
(2011)
21. M. Moghavvemi, S. Yang, ANN application techniques for power system stability estimation.
Electric Power Components and Systems 28, 167–177 (2000)
22. S. Hashemi, M.R. Aghamohammadi, Wavelet based feature extraction of voltage profile for
online voltage stability assessment using RBF neural network. Int. J. Electr. Power Energy Syst.
49, 86–94 (2013)
23. S. Rajan, S. Kumar, D. Mathew, Online static security assessment module using artificial neural
networks. Power Systems, IEEE Transactions on 28, 4328–4335 (2013)
24. A. Bahmanyar, F. Karami, Power system voltage stability monitoring using artificial neural
networks with a reduced set of inputs. Int. J. Electr. Power Energy Syst. 58, 246–256 (2014)
25. S. Ashraf et al., Voltage stability monitoring of power systems using reduced network and
artificial neural network. Int. J. Electr. Power Energy Syst. 87, 43–51 (2017)
26. W.M.L.-L., J.M. Villa-Acevedo, D.G. Colomé, Voltage Stability Margin Index Estimation
Using a Hybrid Kernel Extreme Learning Machine Approach. Energies, 2020. 13 (857)
27. A. Ghaghishpour, A. Koochaki, An intelligent method for online voltage stability margin
assessment using optimized ANFIS and associated rules technique. ISA Trans. 102, 91–104
(2020)
28. B. Gharehpetian, Power system on-line static security assessment by using multi-class support
vector machines. J. Appl. Sci. 12, 8 (2008)
29. H. Mohammadi et al., Voltage stability assessment using multi-objective biogeography-based
subset selection. Int. J. Electr. Power Energy Syst. 103, 525–536 (2018)
30. K. Sundaram, S. Swarup, Classification of static security status using multi-class support vector
machines. Journal of Engineering Research 9, 21–30 (2012)
31. M.V, S. and B. C.K, Fast assessment of voltage stability margin of a power system. Journal of
Electrical Systems, 2014. 10: p. 305–316
32. M.V. Suganyadevi, C.K. Babulal, Support vector regression model for the prediction of
Loadability margin of a power system. Appl. Soft Comput. 24, 304–315 (2014)
33. K.S. Sajan, V. Kumar, B. Tyagi, Genetic algorithm based support vector machine for on-line
voltage stability monitoring. Int. J. Electr. Power Energy Syst. 73, 200–208 (2015)
34. H. Yang et al., PMU-based voltage stability prediction using least square support vector
machine with online learning. Electr. Power Syst. Res. 160, 234–242 (2018)
35. L. Wehenkel, M. Pavella, Decision tree approach to power systems security assessment.
Int. J. Electr. Power Energy Syst. 15(1), 13–36 (1993)
36. R. Diao et al., Decision tree-based online voltage security assessment using PMU measure-
ments. IEEE Trans. Power Syst. 24(2), 832–839 (2009)
37. R.F. Nuqui, et al. Fast on-line voltage security monitoring using synchronized phasor measure-
ments and decision trees. In 2001 IEEE Power Engineering Society Winter Meeting. Confer-
ence Proceedings (Cat. No.01CH37194). (2001)
38. Z. Li, W. Wu, Phasor measurements-aided decision trees for power system security assessment.
In 2009 Second International Conference on Information and Computing Science. (2009)
210 A. Mollaiee et al.

39. C. Zheng, V. Malbasa, M. Kezunovic, Regression tree for stability margin prediction using
synchrophasor measurements. IEEE Trans. Power Syst. 28(2), 1978–1987 (2013)
40. H. Mohammadi, M. Dehghani, PMU based voltage security assessment of power systems
exploiting principal component analysis and decision trees. Int. J. Electr. Power Energy Syst.
64, 655–663 (2015)
41. V. Krishnan, J.D. McCalley, Progressive entropy based contingency grouping for deriving
decision trees for multiple contingencies. Int. J. Electr. Power Energy Syst. 45(1), 35–41 (2013)
42. X. Meng et al., Construction of decision tree based on C4.5 algorithm for online voltage
stability assessment. Int. J. Electr. Power Energy Syst. 118, 105793 (2020)
43. X.Z. Meng, P. Zhang, Decision tree for online voltage stability margin assessment using C4.5
and relief-F algorithms. Energies 13, 3824 (2020)
44. M. Beiraghi, A.M. Ranjbar, Online voltage security assessment based on wide-area measure-
ments. IEEE Transactions on Power Delivery 28(2), 989–997 (2013)
45. H. Su, T. Liu, Enhanced-online-random-Forest model for static voltage stability assessment
using wide area measurements. IEEE Trans. Power Syst. 33(6), 6696–6704 (2018)
46. J.D. Pinzón, D.G. Colomé, Real-time multi-state classification of short-term voltage stability
based on multivariate time series machine learning. Int. J. Electr. Power Energy Syst. 108,
402–414 (2019)
47. K.D. Dharmapala et al., Machine learning based real-time monitoring of long-term voltage
stability using voltage stability indices. IEEE Access 8, 222544–222555 (2020)
48. A. Abed, WECC voltage stability criteria, load shedding strategy, and reactive power reserve
monitor methodology. IEEE Power Eng. Soc. 1, 191–197 (1999)
49. V. Krishnan et al., Efficient database generation for decision tree based power system security
assessment. IEEE Trans. Power Syst. 26(4), 2319–2327 (2011)
50. C.A. Jensen, M.A. El-Sharkawi, R.J. Marks, Power system security assessment using neural
networks: Feature selection using fisher discrimination. IEEE Trans. Power Syst. 16(4),
757–763 (2001)
51. K. Verma, K.R. Niazi, Supervised learning approach to online contingency screening and
ranking in power systems. Int. J. Electr. Power Energy Syst. 38, 97–104 (2012)
52. M. Sun, I. Konstantelos, G. Strbac, A deep learning-based feature extraction framework for
system security assessment. IEEE Transactions on Smart Grid 10(5), 5007–5020 (2019)
53. S. Varshney, L. Srivastava, M. Pandit, ANN based integrated security assessment of power
system using parallel computing. Int. J. Electr. Power Energy Syst. 42(1), 49–59 (2012)
54. W.J. Ridgman, Principles of Multivariate Analysis. By W. J. Krzanowski. 563 pages.
(Clarendon Press, Oxford, 1988). (Price £65.00 (hard covers). ISBN 0 19 852211 8. The Journal
of Agricultural Science, 1989. 112(1): p. 141–142)
55. J.E. Jackson, A User's Guide to Principal Components. (Wiley, 2003)
56. L. Zhu et al., Imbalance learning machine-based power system short-term voltage stability
assessment. IEEE Transactions on Industrial Informatics 13(5), 2533–2543 (2017)
57. IEEE Standard for Synchrophasors for Power Systems. IEEE Std C37.118–2005 (Revision of
IEEE Std 1344–1995), 2006: p. 1–65
58. A. Pai, Energy Function Analysis for Power System Stability. (Springer, New York, 1989)
59. Gan, R.Z.a.D., MATPOWER: A MATLAB Power System Simulation Package. [Online]. www.
pserc.cornell.edu/matpower/
Chapter 10
Evaluation and Classification of Cascading
Failure Occurrence Potential Due to Line
Outage

Morteza Abedi, Mohammad Reza Aghamohammadi,

and Mohammad Taghi Ameli

10.1 Introduction

The modern power systems have developed a continuous topology to increase

reliability and economic performance. Increasing the dimensions of the power
system on the one hand and the need for continuous monitoring of the network for
secure operation on the other hand has caused the issue of assessing the security of
the power system to face challenges. In such a condition, considering economic
factors and the fast growth of consumers in a power system that should be supplied,
the power systems are operated close to their allowed operating limit [1]. Since a
power system consists of a large number of transmission lines, the location of the
lines in the power system and their multiplicity compared to other equipment such as
transformers and generators make the lines more vulnerable. Therefore, the sudden
loss of one or more transmission lines can lead to violations of network operation
limits and endanger the stability margins of the power system and affect the
performance of other equipment and the power system in general [2].
Cascading failure usually occurs as a result of an initial failure like outage of a
transmission line due to heavy loading caused by a short-circuit fault. In this
situation, cascading failure in the power system may lead to uncontrolled separation
and islanding and eventually collapse and blackout on each island [3]. In the
blackout event on 14 August 2003, the cascading failure had an important role.
Due to cascading failure, the power system is separated into independent islands and
finally leads to collapse and blackout in each island [4]. Therefore, if the potential of
cascading failure and the resulting blackout due to the outage of the line can be

M. Abedi · M. R. Aghamohammadi (*) · M. T. Ameli

Faculty of Electrical Engineering, Shahid Beheshti University, Tehran, Iran
e-mail: [email protected]; [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 211
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_10
212 M. Abedi et al.

evaluated and predicted, respectively, a set of preventive actions can be taken to

reduce the vulnerability of the power system and prevent cascading failure caused by
line outages in the power systems.
The development of power systems has increased the number of operating vari-
ables of the power system. Extracting information of operating variables provides
knowledge for the PSCC that can make fast and effective decisions in threatening
situations and prevent failures [5]. When a specific line is disconnected, a limited
number of operating variables of the power system would be effective. Also, using
the information on all operating variables of the power systems for evaluating the
potential of cascading failure resulting from line outage is not cost-effective, because
receiving information of all operating variables by the PSCC requires a large number
of measurement and communication devices, and storing this large volume of
information increases the computational costs and the cost of providing devices.
Thus, to evaluate the potential of cascading failure resulting from the line outage
requires reducing the dimensions of the operating variables and identifying a limited
number of dominant variables with rich information regarding the operation
conditions.
In recent scientific studies, identification and extraction of DOVs in various
contexts of power systems, including evaluating the security of the power system
in terms of voltage stability [6], rotor fault detection [7], effective fault location [8],
short-term price and load prediction [9], and transient stability analysis [10], have
been investigated. However, extracting information from variables and identifying
DOVs for evaluating the potential of cascading failure due to a line outage in power
systems have not been studied. In [11], two variance reduction methods based on
meta-modeling have been presented for quick approximation of cascading failure
probability. In [12], factors affecting the probability of power system equipment
failure as a function of system load level have been studied and used in a simple
model to assess the probable risk of cascading failure. In [13], considering the effect
of the hidden fault on a protection system, the probability of cascading failure is
evaluated. In [14], an improved OPA model has been presented for cascading failure
in which the issues regarding line outages, simulation of cascading failure, and the
probability distribution of outage severity have been studied. In [15], the total
number of line outages has been predicted in cascading failure by using a branching
process of Galton-Watson. In [16], the graph theory has been used to analyze the
interaction among the devices undergoing failure to specify the general propagation
pattern of the cascading failure. In [17], the Random Chemistry search algorithm has
been used to present a model for estimating the risk of large cascading outages as a
result of multiple failures. In [18], two different branch process models have been
used to simulate the probability distribution of load shedding using OPA and
TRELSS in cascading blackout. In [19], a tool has been presented to evaluate the
risk of cascading failure for measuring the power system reliability. The proposed
method looks for the initial N-x event that has the maximum impact on cascading
failure. In [20], a new random model of cascading failure has been presented. In this
model, AC power flow and decentralized transformation have been used to analyze
the vulnerability of the power system, and the random probability distribution has
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 213

been used to model load shedding relays. Also, the effect of wind turbine penetration
of cascading failure has been studied. In [21], an approach based on sample-induced
semi-analytic has been proposed to determine the effect of the failure of power
system devices on the blackout risk during cascading failure and identifying the
relationship between the failure of devices and blackout risk. The Markov sequence
model has been used to develop an accurate relationship between the failure of
devices and blackout risk. In [22], an approach based on data mining has been
presented for predicting the vulnerability after cascading failure in a power system in
real time. In this method, the singular value decomposition and post-events data are
used to determine the vulnerable areas. The method employed in [23] can predict the
blackout probability online by analyzing the transmission line event along with the
probabilistic framework using SVM as the learning tool. In [24], the propagation and
development of transmission lines’ outage are estimated, and the parameters of the
probabilistic branch process model resulting from cascading failure are obtained.
Then, the branch process model is used to predict the total number of outages for one
initial outage. In [25], a new intelligent machine method has been presented for
protection relays based on SVM, SCADA, and communications among protection
relays, which is generally called intelligent protection relays. By making intelligent
decisions about the trip and its time, and using SCADA information, intelligent
relays can reduce cascading failure and global blackouts to a great extent. In [26],
using the decision tree and calculating the brittleness index at each instant, the
blackout of the power system during a cascading failure is predicted.
Many of the reviewed studies are evaluated based on the offline data obtained
from the system model. The studies that have employed the online model cannot be
used for online evaluation of the potential of cascading failure and prediction of the
resulting blackout severity using the information before the failure occurrence. The
reviewed studies look to find the propagation pattern of the cascading failure, and
none of them have evaluated the potential of cascading failure before the occurrence
of the initial failure. Also, the introduced indices represent the power system
vulnerability against cascading failure and do not estimate the blackout severity
using the information before the occurrence of the cascading failure. To this end, a
proper intelligent scheme (like DT) should be presented to predict the blackout
severity if there is a potential of cascading failure using the information before the
occurrence of the cascading failure.
In this study, a new approach is presented for early prediction of CFOP and the
resulting blackout severity using the DT technique and the DOVs for each line of the
system. To this end, a three-step intelligent method using three DTs predicting the
CSPDT and DOVs is presented. In the proposed method, the DOVs are identified
using a method based on entropy, and mutual information theory between the
operating variables and their information is given to the CSPDTs instantaneously
using WAMS/PMU as an input vector. At each specific operating point and in real
time, using the information of the DOVs before line outage and using the result of
CSPDT performance, the CFOP resulting from line outage and the blackout severity
are estimated. The proposed scheme determines the CFOP and predicts and classifies
the blackout severity resulting from cascading failure by the CSPDTs in three levels.
214 M. Abedi et al.

Therefore, in the presented scheme, the CFOP in the power system before line
outage is evaluated instantaneously and online. Also, if the CFOP is existed, the
blackout severity resulting from the outage of each line is also predicted. Therefore,
the PSCC can take preventive actions to reduce the potential. To construct CSPDT,
the C4.5 algorithm, which is more accurate and robust against noisy data [27, 28]
is used.
The rest of this paper is organized as follows. Section 10.2 presents the general
framework of the proposed scheme. Section 10.3 introduces the employed
DT. Section 10.4 describes the identification of DOVs for a specific line #L.
Section 10.5 presents the three-step scheme completely. In Section 10.6, the pro-
posed scheme is applied to sample networks, and the results are analyzed and
validated. Finally, the paper is concluded in Sect. 10.7.

10.2 Overview of the Proposed Method

The occurrence of cascading failure is a complicated process in which cascading

outages weaken the power system and increase the potential of small or large
blackouts. Evaluating the potential of cascading failure helps the operator to decide
what preventive actions to take.
In the proposed scheme, operational information of the power system before line
outage is used to evaluate and predict the potential of cascading failure and blackout
severity, respectively. Also, the occurrence of cascading failure due to outage of line
#L, at the operating point (i), is detected by the outage of at least two lines after the
outage of #L line with more than 5% load shedding.
In the proposed scheme, the line that its outage results in cascading failure and
blackout in the power system is known as the critical line. In other words, the outage
of the critical line changes some of the operating variables, resulting in cascading
outage of other devices and cascading blackout. Therefore, the operating point at
which the #L line is disconnected from the power system due to a fault occurrence
and causes cascading failure is known as a critical scenario.
Figure 10.1 shows the conceptual structure of the proposed three-step predictor
scheme. First, the proposed scheme is designed, and then it is implemented in PSCC.
The design of the proposed scheme includes identifying DOVs among all
monitorable operating variables of the power system and evaluating the potential
of cascading failure and predicting the severity of the resulting blackout through
training CSPDTs. Information of DOVs before the failure of #L line is collected
continuously and instantaneously using WAMS/PMU and given to the CSPDTs. To
functionalize the correlation of potential of cascading failure and the blackout
severity with DOVs, the intelligent decision tree method is used.
The three-step DT is responsible for evaluating the potential and predicting the
magnitude and severity of the blackout; the blackout magnitude is defined as the
percentage of the total load shedding to the total initial load of the power system. To
discriminate the severity of the system’s cascading failure in terms of the magnitude
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 215

Fig. 10.1 Conceptual

structure of the proposed
scheme

Table 10.1 Output of CSPDTs for predicting severity of blackout

System vulnerability CSPDT1 CSPDT2 CSPDT3 Severity of predicted blackout
Non-critical 0 0 0 B.S < 5%
Critical 1 0 0 5%
B.S < 20%
Severely critical 1 1 0 20%
B.S < 40%
Extremely critical 1 1 1 40%
B.S
100%
B.S¼Blackout severity

of blackouts, three decision trees CSPDT1, CSPDT2, and CSPDT3 are used to
predict the specific magnitude of blackouts.
Considering the potential of cascading failure and blackout severity, the system
vulnerability is divided into four states of non-critical, critical, severely critical, and
extremely critical regarding the reasonable combination of the CSPDTs’ outputs.
Table 10.1 represents the reasonable combinations of the CSPDTs’ outputs for
deciding about the blackout severity resulting from cascading failure. For each
vulnerability state of the power system, the combination of CSPDT outputs should
be as shown in Table 10.1. According to Table 10.1, CSPDT1 is responsible for
evaluating the potential for cascading failure so that its zero output indicates that
there is no potential for cascading failure to occur due to specified L # line outage.
216 M. Abedi et al.

Conversely, output 1 indicates the potential of cascading failure occurrence due to

the outage of the specified L # line and the specific magnitude of a blackout.
CSPDT2 and CSPDT3 are responsible to predict the blackout severity resulting
from cascading failure due to outage of #L line. In this study, the blackout severity
conceptual is the same as the percentage of load shedding in the power system. For
implementing the proposed scheme, the identified DOVs and the trained CSPDTs
for each line are given to the PSCC. Thus, the PSCC can evaluate the potential of
cascading failure and predict the resulting blackout severity for each line and in each
operating point of the power system.

10.3 Cascading Failure Severity Predictor DT

In this study, DT is used to evaluate the potential of cascading failure and predict the
blackout severity using DOVs of the power system. To train and build the DT,
various algorithms like CHAID [29], CART [30], and C4.5 are used. In this paper,
the C4.5 algorithm is employed for the proposed CSPDT.

10.3.1 C4.5 Algorithm [31]

In the C4.5 DT, the statistical values called entropy and information gain are used to
determine how much the feature can split the training samples based on their
classification.

10.3.1.1 Entropy

The entropy determines the purity of a set of examples. If the set Z includes positive
and negative examples of a concept or objective, the entropy of Z regarding this
Boolean class is defined as follows:

EntropyðZ Þ ¼ pL log 2 pL  pΘ log 2 pΘ ð10:1Þ

where pL is the ratio of positive examples to total examples and pΘ is the ratio of
negative examples to total examples.

10.3.1.2 Information Gain

The information gain of a feature is the reduction of entropy resulting from splitting
the examples using this feature. In other words, the information gain of Z and A
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 217

which shown as Gain (Z, A) for a feature like A with respect to the set of examples Z
is defined as follows:

X jZ v j
GainðZ, AÞ ¼ EntropyðZ Þ  EntropyðZ v Þ: ð10:2Þ
jZ j
v2ValuesðAÞ

10.3.2 Considering Continuous Attributes

Algorithm C5 also includes features with continuous values. For a continuous

feature such as A, a Boolean class, such as Ac, is defined so that Ac is true if
A < C and otherwise false.

10.3.2.1 Gain Ratio

Gain ratio demonstrates the uniformity and extensiveness of the data split by a
feature. It is defined as in Eq. (10.3):

GainðZ, AÞ
Gain RatioðZ, AÞ ¼ ð10:3Þ
Split InformationðZ, AÞ

The split information of Z and A is known as SI and defined by Eq. (10.4):

Xc
jZ i j jZ j
Split InformationðZ, AÞ ¼  log 2 i ð10:4Þ
i¼1
j Z j jZ j

In Eq. (10.4), C is the number of subsets of Z.

10.4 Dominant Operating Variables

The outage potential of a line followed by cascading failure depends on the operating
condition of the power system. The operating condition of the power systems can be
described using the operation information before line outage like the voltage of
buses, power flow of lines, loads, and generation of the generators.
In real power systems, there is a large volume of operational information on each
operating point. However, considering the criticality of a specific line #L, all
operational information does not significantly depend on the line condition. In
other words, among all operating variables, only a limited number of the variables
218 M. Abedi et al.

known as DOVs are associated with line criticality and depend on it significantly. In
fact, the DOVs associated with the outage of each line are the minimum number of
operating variables that can be used to evaluate the criticality of a specific line #L.
Therefore, dimension reduction methods are used to identify the DOVs among a
large number of operating variables. These methods are usually used to process data
and select the dominant features. The main purpose of reducing the data dimension
and the number of features to the minimum possible value is to increase the data
classification accuracy [32]. The feature selection and identification methods reduce
the dimensions by selecting a subset of features that minimize a specific cost
function. Feature identification is used in many applications of expert and intelligent
systems like data mining, learning machine, image processing, anomaly detection,
and bioinformatics [33]. Also, the feature identification process is known as the
variable identification, feature reduction, or detection of a subset of variables.
The MIM method is one of the DOV variable dimension reduction and DOV
identification methods, which is defined based on the mutual information theory. In
this method, the relationship and dependency among all variables are studied
[34]. Then, the operating variables with rich information are identified, and the
operating variables with low or same information are neglected.
As mentioned before, for evaluating the line criticality, the identified DOVs
should be used. To this end, various operating points with maximum coverage of
the operating space should be developed in the operational space of the power
system, such that the line status is evaluated as critical for some operating points
and non-critical for the others.

10.4.1 DOVs’ Identification Algorithm

In power systems, the system status in a specific operating point is defined by a set of
operating variables {VAR} such as bus voltage, power flow, load, and generation of
the generators. At each operating point, various behavioral characteristics can be
attributed to the power system using a causal relationship between the operating
variables and the behavioral characteristics. However, considering the behavioral
characteristic B, all sets of operating variables do not have an equal contribution in
the formation of the behavioral characteristic. Some variables are strongly correlated
with the behavioral characteristic, while others have a weak correlation. For exam-
ple, the overloading after an outage of #L line is considered as a behavioral
characteristic that is associated with a number of operating variables before a line
outage. Therefore, for a set of operating variables of the power system {VAR},
considering a specific behavioral characteristic of the power system B, a subset of the
variables {VARD} with a strong correlation with B is known as the DOVs associated
with B. It should be mentioned that for each set of operating variables {VAR}, there
are various DOVs considering various behavioral characteristics.
In this study, the total set of operating variables before line outage is considered
as {VAR}. The line (#L) outage potential for initiating of cascading failure in the
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 219

power system is known as the behavioral characteristic B of the power system.

Considering the behavioral characteristic B, the goal is to identify the subsets of
operating variables {VAR} associated with B. To identify the DOVs compatible
with {VARD} and associated with the behavioral characteristic B among all oper-
ating variables of the set {VAR}, the mutual information theory and entropy can be
used as follows.
If an operating variable in N operating points is considered as a set like X ¼ {x1,
x2,. . . . . ., xN}, such that xi represents the values that the operating variable can adopt
at different operating points, a random variable xi with probability p(xi) is defined for
each operating variable. For the dataset X, an entropy function H(x) is obtained
using Eq. (10.5), which can be used to obtain useful information about the operating
variable X:
X
H ðxÞ ¼  pðxÞ: log 2 pðxÞ ð10:5Þ
x2X

such that H(X) is the entropy function and known as the positive Shannon function
[35]. Considering two independent random variables (X, Y), the bivariate entropy
function is presented as in Eq. (10.6):
X
H ðX, Y Þ ¼  pðx, yÞ: log 2 pðx, yÞ ð10:6Þ
x, y

where p(x,y) is the bivariate probability function, and the value of H(X,Y) represents
the information hidden in variable X to describe the status of variable Y [35]. If the
value of H(X,Y) is calculated for Y ¼ y0, the bivariate entropy is known as the
conditional entropy of variable X for Y ¼ y0, and it is described using Eq. (10.7).
This conditional entropy represents the information of variable X for a specific value
of Y:
X
H ðXjy0 Þ ¼  Pðxjy0 Þ: log 2 Pðxjy0 Þ ¼ H ðX, Y Þ  H ðY Þ ð10:7Þ
x

where P(x| y0) p(x| y0) is the random bivariate conditional probability function and H
(X|yo) is the conditional entropy of variable X for Y ¼ y0 [35]. Using Eqs. (10.6) and
(10.7), (10.8) is obtained as the mutual information between two variables of X and
Y. Mutual information between two variables shows the amount of common infor-
mation between the two variables:

XX Pðx, yÞ
I ðX; Y Þ ¼ H ðXÞ  HðXjYÞ ¼ Pðx, yÞ: log ð10:8Þ
x y pðxÞ:pðyÞ

In other words, mutual information represents the information of a variable about

the other variable [35]. In this study, the variable X is considered as an operating
220 M. Abedi et al.

Fig. 10.2 Venn diagram for

the mutual information

variable of the set {VAR}, and variable Y is considered as the behavioral charac-
teristic B of the power system. Figure 10.2 shows the Venn diagram of Eq. (10.8).
To employ the mutual information theory method in a power system for identi-
fying the DOVs associated with critical condition of each line, several operating
points should be selected in the operating space of the power system. At each
operating point i, the information of operating variables of the set {VAR} is unique.
Thus, at each operating point i, considering the critical condition of #L line, for the
set of operating variables {VARi}, there is a behavioral characteristic Bi that
describes the #L line being either critical or non-critical. Therefore, for N operating
points, there are N sets of operating variables {VAR} that constitute a data matrix
along with a matrix including N behavioral characteristics. The set of operating
variables includes an N*m data matrix, where m is the total number of operating
variables of the power system. N behavioral characteristics constitute an N*1 vector.
To identify the DOVs regarding the outage of each line, the MIM [36] method
based on information theory is presented. The mathematical formulation of MIM is
described by Eq. (10.9). In Eq. (10.9), X and Y are the operating variable of {VAR}
and the behavioral characteristic B of the potential of cascading failure, respectively.
In this method, the mutual information between all operating variables of {VAR}
and the behavioral characteristic B is calculated. Then, the results of mutual infor-
mation calculation are sorted maximum to minimum. In Eq. (10.9), Xk is an
operating variable, k is the number of operating variables, and VARK(Xk) is the
set of identifying k variables with the highest mutual information. I(Xk; Y) is the
mutual information between Xk and Y, H(Xk) and H(Y) are single variable entropy
functions, and H(Xk; Y) is a bivariate entropy function:

VARk ðX K Þ ¼ arg max I ðX k ; Y Þ ¼ H ðX k Þ þ H ðY Þ  H ðX k , Y Þ ð10:9Þ

Considering Eq. (10.9), the number of operating variable k can be changed. In

other words, Eq. (10.9) can identify a set of k dominant variables among all
operating variables considering the various values of k. Figure 10.3 shows the
structural relationship between sets of operating variables with the corresponding
behavioral characteristics for N different operating points.
Thus, the result of MIM is examined by the learning machine for different values
of k. A set of operating variables that has the maximum accuracy in evaluating the
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 221

Fig. 10.3 relationship between operating variables with corresponding behavioral characteristics

potential of cascading failure and predicting the blackout severity is introduced as

the set of dominant operating variables known as DOVs. The k parameter
corresponding to the set of DOVs is the number of dominant operating variables
and represented by D. Thus, using the proposed MIM method, the set of dominant
operating variables {VARD} called DOVs and dimension of D < <m, which is
strongly correlated with behavioral characteristic B, is identified.

10.5 The Proposed Method

Using the proposed MIM method and various disturbance and perturbation scenar-
ios, a set of unique DOVs are detected for each line of the system. Also, using the
proposed CSPDTs and the detected DOVs, a relationship is created between the
system operating condition and CFOP evaluation and prediction of blackout sever-
ity. Therefore, the main advantage of CSPDTs is the evaluation of CFOP and
blackout prediction for unobserved disturbances. This capability of the CSPDTs is
based on interpolation capability, which is obtained considering various disturbance
and perturbation scenarios and covering the whole disturbance space of the power
system. Figure 10.4 shows the general structure of the proposed method for evalu-
ating the CFOP and blackout severity prediction using DOVs, CSPDT1, CSPDT2,
and CSPDT3.
Since the aim of the proposed scheme is to evaluate the CFOP in the power
system before the first failure, the measured information of the identified DOVs is
given to the CSPDTs by the WAMS at each instant. The instantaneous information
of the DOVs for evaluating the CFOP and blackout severity prediction resulting
from outage of line #L as the vector VARLD ðt Þ with D members is used as input at time
t to the CSPDTs as given in Eq. (10.10):

VARD ðt Þ ¼ ½var1 ðt Þ, var1 ðt Þ, var1 ðt Þ, . . . , varD ðt Þ ð10:10Þ

Asynchronous data of WAMS and temporary loss of some WAMS data might
affect CFOP evaluation. In this study, asynchronous or lost data of VARLD ðt Þ can be
222 M. Abedi et al.

Fig. 10.4 general structure

of the proposed scheme

replaced with the measured DOV of the previous sample (t-Δt). Δt is the time step of
data transmission by PMU and WAMS to the CSPDTs.
According to Fig. 10.4, the blackout severity procedure is classified in three
levels, and each level is associated to a unique CSPDT. At each time t, by presenting
the vector VARLD ðt Þ to each CSPDT simultaneously, each CSPDT returns 0 or 1 at the
output; by combining the outputs of all CSPDTs, the CFOP and blackout severity
resulting from outage of line #L are predicted.
The zero output for CSPDT1 indicates that if the line #L in the power system has
experienced outage as a result of a failure, the predicted blackout severity is smaller
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 223

than 5% of the power system loading level. Also, the cascading failure is not caused
by the result of outage of line #L; in other words, the CFOP does not exist at
the operating point. This state of the power system is called non-critical in which the
outputs of CSPDT2 and CSPDT3 are zero. Output 1 for CSPDT1 indicates that if the
line #L experiences outage as a result of a failure, the predicted blackout severity
would be more than 5% of the power system loading level. Also, the cascading
failure is caused by the result of outage of line #L; in other words, there is the
potential of cascading failure at the operating point.
The zero output for CSPDT2 indicates that if the line #L experiences outage as a
result of failure, the predicted blackout severity would be 5% to 20% of the power
system loading level. Also, cascading failure is caused by the result of outage of line
#L; in other words, there is the potential of cascading failure at the operating point.
This state is called critical in which the outputs of CSPDT1 and CSPDT3 are 1 and
0, respectively. The output of the CSDPT2 being 1 indicates that if line #L of the
power system experiences outage as a result of a failure, the predicted blackout
severity would be more than 20% of the power system loading. Also, the cascading
failure is the result of outage of line #L, indicating that there is the CFOP at the
operating point.
The zero output for CSPDT3 being zero indicates that in line #L which experi-
ences outage as a result of failure, the predicted blackout severity would be 20% to
40% of the power system loading level. Also, the cascading failure is the result of
outage of line #L, indicating that there is the potential of cascading failure at the
operating point. This state is called severe critical in which the outputs of CSPDT1
and CSPTS2 are one. The output of CSPDT3 being one indicates that if the line #L
experiences outage as a result of a failure, the predicted blackout severity would be
40% to 100% of the power system loading level. Also, the cascading failure is the
result of outage of line #L, indicating that there is the potential of cascading failure at
the operating point. This state is called extremely critical in which the outputs of
CSPDT1 and CSPDT2 are one.
After evaluating the potential and blackout severity prediction in the power
system considering the output of the designed CSPDTs, to reduce the system stress
and CFOP, regarding the vulnerability of the system and impact priority, preventive
actions might be taken. Among these actions, disconnecting the load/generation
[37], emergency line disconnection [38], and island control [39] can be mentioned.
Also, to prevent and stop excess loading of the line, which is the most effective
mechanism for power system outage, proper load removal can be adopted as a
preventive action [40]. Also, to prevent undesired performance of the distance
relay resulting from power fluctuations, zone 3 of the distance relay can be
blocked [41].
224 M. Abedi et al.

10.6 Simulation Studies

In this section, the proposed scheme is applied to lines 17–18 and 21–22 of the
IEEE39-bus system, and the results are studied. In Fig. 10.5, the IEEE39-bus power
system is shown. This network has 10 generators, 34 transmission lines, and
19 loads [42].
Table 10.2 shows various monitorable operating variables of the IEEE39-bus
power system, where the total number of the operating variables considering the
characteristics of the IEEE39-bus power system is 405. In the first and third columns
of Table 10.2, the type of the operating variable and its unit are given. In the second
and fourth columns, the complete name of the operating variables is given. Among
these variables, a number of variables like the voltage magnitude of the generators,
speed of the generators, and the rotor phase angle of the slack generator are always
constant in all operating points of the power system. Therefore, considering the

G8

37
W
G10 25 26 28 29

30
W
38
W
2 27 G9
18
17 24
1
G6
16
3
G1 W 35
15 21 22
W 4
39 14

6 12
19 23
WW
7 W
13 W
20 36
11
8 W W
34 G7
33
W
9 31
G5 G4
10
G2 W
32

G3

Fig. 10.5 IEEE39-bus power system

10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 225

Table 10.2 IEEE39-bus power system monitorable operating variables

Var. Var. name Var. Var. name
PLine(p.u) Line active power Vphase(deg) Bus voltage angle
QLine(p.u) Line reactive power δ Phase(deg) Rotor angle(estimated)
Line % Line loading PG(p.u) Gen. active power
PLoad(p.u) Load active power QG(p.u) Gen. reactive power
QLoad(p.u) Load reactive power Gen % Gen. loading
V(p.u) Bus voltage Tr % Trans. loading

characteristics of the power system, 383 operating variables are not constant, which
describe the operating point of the power system.

10.6.1 Preparing Scenarios

The modeling and simulation used to implement the cascading failure and the
proposed scheme are based on time-domain calculations. Determining the violation
of the devices from the operating limit using the time-domain method requires a
precise dynamic modeling of the power system devices. Also, the dynamic model of
the power system devices has an effective role in the formation of the cascading
failure. Thus, dynamic modeling of the power system is essential for simulating the
cascading failure. In the studied power system, sixth order of the generator, O.S
protection, generator primary actuator, generator voltage control system,
low-frequency generator protection, line distance protection, load shedding under
frequency/voltage protection and tap changer of the transformers connected to a
load, frequency and voltage load removal relays, and overcurrent relays are
modeled.
In this study, DIgSILENT Power Factory 15.1 is used for modeling of the
IEEE39-bus power system. The training scenarios resulting from the outage of line
#L are provided through offline simulation using this software. Also, MATLAB is
used to identify the DOVs and train the CSPDTs; the proposed three-step scheme is
also implemented in MATLAB. All three CSPDTs are trained offline using
MATLAB. Therefore, all scenarios which prepared offline by DIgSILENT are
provided to MATLAB software in order to train CSPDTs.
When the training of CSPDTs is finished, the identified DOVs and the trained
CSPDTs are provided to the PSCC for online evaluating the potential of cascading
failure and predicting the blackout severity caused by the outage of #L line.
Considering the above models, 770 operating points are studied for developing
various training scenarios. All scenarios are formed considering the basic load of
6000 MW at 11 load levels (0.7–0.75-0.8-0.85-0.9-0.95-1-1.05-1.1-1.15-1.2). Such
that 70 operating points are considered at each load level, and the operating points
are created such that the generators, lines, and the transformers are at maximum
loading at different load and generation levels especially high loading levels. Thus,
226 M. Abedi et al.

the power system is always under stress. Also, to predict the blackout severity in the
prepared training scenarios, the load shedding is determined after the occurrence of
each scenario.

10.6.2 Evaluating the Criticality of the Lines

The proposed scheme is used to evaluate the potential of cascading failure and
predicting blackout severity for the outage of lines 17–18 and 21–22 as a result of a
three-phase short circuit, and the DOVs are identified for each line.

10.6.2.1 Evaluating Criticality of Line 17–18

748 operating points are created for the outage of lines 17–18, and 383 operating
variables are sampled at each operating point, according to Table 10.2 before the
outage of lines 17–18. The specifications of the operating points are given in
Table 10.3 for identifying the DOVs and providing the test and training scenarios.
A three-phase short circuit fault is applied to the lines 17–18 at all of the
748 operating points. Thus, 266 critical scenarios and 482 non-critical scenarios
are obtained. Among the 748 scenarios, 523 scenarios (80% of the all scenarios) are
randomly selected as training scenarios and are provided to the MIM method.
Among the 523 selected scenarios, the ratio of the scenarios that result in cascading
failure to the ones that do not result in cascading failure is 30% to 70%.
The MIM method identifies a set of operating variables based on the mathemat-
ical relationships of mutual information theory and entropy and a value of K. Then,
the identified operating variables and the training scenarios are used to train the
three-step trees of evaluating potential of cascading failure and predicting blackout.
After training the CSPDTs, each CSPDT is validated using test scenarios to deter-
mine the accuracy and performance of the CSPDTs.

10.6.2.2 Identifying the DOVs and Training the CSPDTs for Lines
17–18

The MIM method identifies a set of operating variables with K members based on
mutual information theory and considering the value of K. In this study, the value of

Table 10.3 Number of different scenarios created as a result of outage of lines 17–18
Operating point Operating point
Fault Operating Critical Non- critical Training Testing
Line type point scenarios scenarios scenarios scenarios
17–18 3φ 748 266 482 523 225
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 227

CSPDTs Result for Line17-18

100

Acc. of CSPDTs
90
80
70
60
50
40
30
0 10 20 30 40 50 60 70 80 90 100
Different Combination of Variables
CSPDT1 CSPDT2

Fig. 10.6 The accuracy of CSPDTs for different Ks due to outage of lines 17–18

Table 10.4 The best accuracy of CSPDTs due to outage of lines 17–18
CSPDT1 CSPDT2 CSPDT3
Acc. Num. of test Acc. Num. of test Acc. Num. of test Num. of
% scenarios % scenarios % scenarios DOVs
97.959 225 94.285 225 86.142 225 50

K is selected from the set {2, 5, 10, 20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 75,
80, 85, 90, 95, 100}. Figure 10.6 shows the accuracy of CSPDTs for different K and,
consequently, the different number of identified operating variables to evaluate the
potential of cascading failure and to predict blackout. Therefore, the sets of operating
variables with different numbers of members are identified using MIM. CSPDTs are
trained using the identified operating variables and the provided training scenarios.
Finally, their accuracy and performance are evaluated using the test scenarios. A set
of operating variables identified by MIM for which the CSPDTs have maximum
accuracy in evaluating the potential of cascading failure and predicting blackout
severity due to outage of lines 17–18 is identified as the set of DOVs.
In Fig. 10.6, a diagram is represented for the evaluation accuracy of each CSPDT.
According to Fig. 10.6, the maximum accuracy of CSPDTs with the minimum
number of operating variables can be identified. Table 10.4 shows the best accuracy
of CSPDTs in evaluating the potential of cascading failure and predicting blackout
severity for test scenarios due to outage of lines 17–18.
Considering the results of Table 10.4, it is seen that the maximum accuracy in
evaluating the potential of cascading failure and predicting blackout severity is
obtained for 50 operating variables. The combination of these operating variables
identified by the MIM method is according to Table 10.5. The numbers after the
operating variable’s name are the bus number.
Based on Fig. 10.6 and Table 10.5, the combination of 50 operating variables
identified using MIM has the maximum accuracy in evaluating the potential of the
cascading failure and predicting blackout severity for the outage of lines 17–18.
Thus, these identified operating variables are called DOVs of lines 17–18. For online
evaluation of the power system vulnerability regarding the outage of lines 17–18, it
228 M. Abedi et al.

Table 10.5 Dominant oper- Number Operating Var Num. Operating Var
ating variables for lines 17–18
1 Qline_19–16 26 Qline_27–17
2 Qline_13–10 27 Qline_6–5
3 Qline_26–25 28 Pline_15–14
4 Qline_24–23 29 Qline_2–1
5 Qline_3–2 30 Qline_39–1
6 QG_38 31 Qline_18–17
7 Qline_5–4 32 Pline_18–17
8 Qline_14–13 33 Qline_22–21
9 QG_37 34 Line_27–26 (%)
10 QG_33 35 QG_39
11 Qline_11–10 36 Vphase _28
12 QG_32 37 Line_7–6 (%)
13 Pline_23–22 38 PG_36 (%)
14 Qline_29–28 39 Qline_27–26
15 Qline_8–7 40 Pline_3–2
16 Pline_27–17 41 Qline_16–15
17 Qline_23–22 42 Tr_25–37 (%)
18 QG_31 43 Tr_22–35 (%)
19 Line_3–2 (%) 44 Line_25–2 (%)
20 Pline_24–16 45 Qline_15–14
21 Pline_9–8 46 Tr_ 23–36 (%)
22 Vphase_29 47 PG_38 (%)
23 Pline_39–9 48 Tr_6–31 (%)
24 Qline_8–5 49 Tr_2–30 (%)
25 Qline_14–4 50 Vphase _23

is sufficient to provide the DOVs of Table 10.5 along with the trained CSPDTs to the
PSCC. In PSCC, the information of DOVs at each instant and each operating point is
received by WAMS and provides them as input to the trained CSPDTs. Then,
considering the output of each CSPDT, the potential of cascading failure and
blackout severity for the outage of lines 17–18 is determined with high accuracy.

10.6.2.3 Evaluating the Criticality of Lines 21–22

The proposed scheme is applied to lines 21–22 similar to lines 17–18. The specifi-
cations of the operating points are given in Table 10.6 for identifying the DOVs and
providing the test and training scenarios.
According to Table 10.6, a three-phase short circuit fault is applied to the lines
21–22 at all of the 751 operating points. Thus, 580 critical scenarios and 171 non-
critical scenarios are obtained. Among the 751 scenarios, 526 scenarios (80% of the
all scenarios) are randomly selected as training scenarios and are provided to the
MIM method.
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 229

Table 10.6 Number of different scenarios created as a result of outage of lines 21–22
Fault Operating
Line type point Operation point Operating point
21–22 3φ 751 Non-critical Critical Testing Training
scenarios scenarios scenarios scenarios
171 580 225 526

CSPDTs Result for Line21-22

100
95
90
Acc. of CSPDTs

85
80
75
70
65
60
55
0 10 20 30 40 50 60 70 80 90 100
Different combination of Variables
CSPDT1 CSPDT2 CSPDT3

Fig. 10.7 The accuracy of CSPDTs for different K’s due to outage of lines 21–22

Table 10.7 The best accuracy of CSPDTs due to outage of lines 21–22
CSPDT1 CSPDT2 CSPDT3
Acc. Num. of test Acc. Num. of test Acc. Num. of test Num. of
% scenarios % scenarios % scenarios DOVs
94.949 225 96.334 225 88.872 225 50

10.6.2.4 Identifying the DOVs and Training the CSPDTs for Lines
21–22

Figure 10.7 shows the accuracy of CSPDTs for different K and, consequently, the
different number of identified operating variables to evaluate the potential of cas-
cading failure and to predict blackouts. Similar to lines 17–18, for lines 21–22, the
sets of operating variables with different numbers of members are identified using
MIM. CSPDTs are trained using the identified operating variables and the provided
training scenarios. Finally, their accuracy and performance are evaluated using the
test scenarios. A set of operating variables for which the CSPDTs have maximum
accuracy in evaluating the potential of cascading failure and predicting blackout
severity resulting from the outage of lines 21–22 is identified as the set of DOVs.
According to Fig. 10.7, the maximum accuracy of CSPDTs can be identified
considering the minimum number of operating variables. Table 10.7 represents the
230 M. Abedi et al.

Table 10.8 Dominant oper- Num. Operating var. Num. Operating var.
ating variables for lines 21–22
1 QG_30 26 Pline_23–22
2 QG_32 27 Qline_15–14
3 QG_31 28 Qline_8–5
4 Qline_6–5 29 Qline_14–4
5 QG_33 30 Qline_27–26
6 Qline_14–13 31 Vphase _29
7 Qline_13–10 32 Qline_7–6
8 Qline_22–21 33 Pline_15–14
9 Qline_19–16 34 Pline_9–8
10 Qline_24–23 35 Qline_16–15
11 Qline_26–25 36 Pline_39–9
12 Qline_5–4 37 Tr_12–13(%)
13 Qline_11–10 38 Vphase _22
14 Qline_18–3 39 Pline_18–3
15 QG_37 40 Pline_26–25
16 Pline_27–17 41 load_24
17 Pline_18–17 42 Line_22–21(%)
18 QG_36 43 Qline_27–17
19 QG_35 44 Tr_2–30 (%)
20 Qline_8–7 45 Line_27–26 (%)
21 Qline_3–2 46 Tr_23–36 (%)
22 QG_39 47 Qline_18–17
23 Pline_24–16 48 Tr_6–31 (%)
24 Pline_22–21 49 Tr_20–34 (%)
25 QG_38 50 Tr_25–37 (%)

best performance accuracy of the CSPDTs in evaluating the potential of cascading

failure and predicting blackout severity for test scenarios due to outage of lines
21–22.
According to the results of Table 10.7 and Fig. 10.7, it is seen that the maximum
accuracy in evaluating the potential of cascading failure and predicting the blackout
severity is obtained for 50 operating variables. The combination of these operating
variables identified by the MIM method is according to Table 10.8.
The combination of 50 operating variables identified using MIM according to
Table 10.8 provides the maximum accuracy in evaluating the potential of cascading
failure and predicting the blackout severity. Thus, the identified operating variables
according to Table 10.8 are called the DOVs of lines 21–22. Considering the results,
PSCC can determine the potential of cascading failure and blackout severity for lines
17–18 and 21–22, independently at each operating point of the system using the
related DOVs and the trained CSPDTs.
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 231

10.6.3 Performance Validation

The performance of the proposed scheme is validated for four different scenarios.
Specifications of the four scenarios are given in Table 10.9. None of the above
scenarios are used to identify the operating variables and train the CSPDTs. In
Table 10.9, the amount and percentage of load shedding for the four scenarios are
determined. Based on the proposed scheme, information of the DOVs of each line is
received from the WAMS of the power system before the fault occurrence and
provide to the trained CSPDTs of each line. According to the output of the CSPDTs,
the potential of cascading failure and blackout severity of each scenario is predicted
and represented in Table 10.10. To validate the proposed scheme, the output of each
CSPDT should be examined considering the percentage of load shedding as blackout
severity. According to the results in Table 10.10, the proposed three-step scheme has
performed well in evaluating the potential of cascading failure and predicting
blackout severity for four different scenarios. According to the percentage of load
shedding for scenarios 1–4 represented in Table 10.9, it is expected that the system
vulnerability should be non-critical, critical, extremely critical, and severely critical,
respectively. According to Table 10.10, it is seen that the expected results are
achieved correctly. The numbers after the operating variable’s name are the bus
number.

Table 10.9 Specification of fault scenarios for performance validation

Load Cascading Fault location Load Load

Fault level failure Line shedding shedding/load
scenario (MW) status Fault type Line % (MW) level (%)
1 6214 Non-critical 3 phase 17–18 20 205 3.29
2 5557 Critical 3 phase 21–22 50 598 10.76
3 7089 Extremely 1 phase 17–18 70 5500 77.58
critical
4 7124 Sever Resistive 21–22 50 2841 39.87
critical fault 10 Ω

Table 10.10 Result of performance validation

Fault Actual system CSPDT1 CSPDT2 CSPDT3 Prediction of system
scenario vulnerability output output output vulnerability
1 Non-critical 0 0 0 Non-critical
2 Critical 1 0 0 Critical
3 Extremely critical 1 1 1 Extremely critical
4 Severely critical 1 1 0 Sever critical
232 M. Abedi et al.

10.6.4 Global Performance

In the proposed scheme, three CSPDTs should operate based on a logical and
synchronous process to evaluate the potential of cascading failure and predict the
resulting blackout severity. Therefore, the illogical performance of the three
CSPDTs concerning each other might be considered as incorrect performance.
Table 10.11 represents the eight possible combinations of the CSPDTs’ outputs,
and their logical performance can be interpreted as follows:
1. Combinations 1–4 are logical and the decision made by CSPDTs as the initial
decision is verified compared to the logical decision.
2. Combinations 5–6 are illogical and the decision made by CSPDTs as the initial
decision is verified compared to the logical decision.
3. Combinations 7–8 are illogical and the decision made by CSPDTs as the initial
decision is not verified, and it is modified and verified by the logical decision.

10.7 Conclusion

In this study, the DOVs and DT technique are used to present a three-step predictor
scheme for evaluating the potential of cascading failure and predicting the blackout
severity resulting from the outage of each line of the power system. This intelligent
scheme includes three CSPDTs that the logical combination of their outputs makes

Table 10.11 Logical combinations of CSPDTs outputs

State Logic/ Initial Logical
num. CSPDT1 CSPDT2 CSPDT3 illogical Malfunction decision decision
1 0 0 0 Logical – Non- Non-critical
critical
2 1 0 0 Logical Critical Critical
3 1 1 0 Logical – Severely Severely
critical critical
4 1 1 1 Logical – Extremely Extremely
critical critical
5 1 0 1 Illogical CSPDT2 or Extremely Extremely
CSPDT3 critical or critical or
critical critical
6 0 1 1 Illogical CSPDT1 Extremely Extremely
critical critical
7 0 1 0 Illogical CSPDT1 or Critical or Non-critical
CSPDT2 severely or severely
critical critical
8 0 0 1 Illogical CSPDT3 Extremely Non-critical
critical
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 233

the proposed scheme able to evaluate the potential of cascading failure and predict
the blackout severity is different sizes.
In the proposed scheme, the DOVs are identified using MIM based on mutual
information theory and entropy to evaluate the potential of cascading failure and
predict blackout severity. Then, the CSPDTs are trained using DOVs and various
disturbance scenarios. According to the results, the CSPDTs can evaluate the
potential of cascading failure and predict the resulting blackout severity with high
accuracy. The DOVs and trained CSPDTs of each line are provided to the PSCC to
evaluate the potential of cascading failure. In PSCC, the DOVs information is
prepared online at each instant by the WAMS and provide to the trained CSPDTs
of each line. The zero output of all CSPDTs at each operating point of the system
describes the secure operation of the system without CFOP.
The proposed scheme can be used as a warning system to enable emergency
actions and preventive control as a prediction of the severe blackout. In such a case,
PSCC can determine the potential of cascading failure and the resulting blackout
severity at each operating point, knowing the DOVs and the corresponding CSPDTs
of each line. Then, the PSCC can transfer the operating point of the power system
from critical to non-critical, if necessary, so that the power system always operates in
the secure area without cascading failure occurrence.

10.8 Suggestion for Future Research

In this study, the power system vulnerability after evaluating CFOP is examined by
predicting the blackout severity. The blackout severity index is used as a static
vulnerability measure of the power system. In future studies, in addition to the
blackout severity prediction as a measure of the power system vulnerability, other
indices like frequency and rotor angle can be used to predict the dynamic stability of
the power system.

References

1. S. Henry et al., New trends for the assessment of power system security under uncertainty. IEEE
Power Systems Conference and Exposition 3, 1380–1385 (2004)
2. D. Ajendra, Vulnerability Analysis and Fault Location in Power Systems Using Complex
Network Theory. PhD thesis, (College of Science, Engineering and Health RMIT University,
2011)
3. L. Zhou, Multi-Agent System Based Special Protection and Emergency Control Scheme against
Cascading Events in Power System. PhD thesis, (The Faculty of Engineering, Science and
Medicine, Aalborg University, 2013)
4. NERC. Protection System Response to Power Swings. http://www.nerc.com (2013)
5. A. Abedi et al., Review of major approaches to analyze vulnerability in power system Reli-
ability Engineering & System Safety. Reliability Engineering & System Safety 183, 153–172
(2019)
234 M. Abedi et al.

6. D. Seyed Javan et al., Information extraction from effective descriptor variables in reconstruc-
tion of power system security region by considering correlation between loads. International
Transaction Electrical Energy System 13, 145–181 (2017)
7. R. Casimir, E. Boutleux, G. Clerc, A. Yahoui, The use of features selection and nearest
neighbors rule for faults diagnostic in induction motors. Eng. Appl. Artif. Intell. 19(2),
169–177 (2006)
8. Chang H and Viet L.N (2017) Statistical feature extraction for fault locations in nonintrusive
fault detection of low voltage distribution systems. Energies 10(5):611
9. H. Muhammad Faisal et al., Prediction of building energy consumption using enhance
convolutional neural network. Web, Artificial Intelligence and Network Applications 927,
1157–1168 (2019)
10. Saunders C.S et al. (2019) Feature extraction-based real-time transient stability analysis.
Techno Econ Smart Grids Sustain Energy 4(15)
11. F. Cadini et al., Estimation of rare event probabilities in power transmission networks subject to
cascading failures. Reliability Engineering & System Safety 158, 9–20 (2017)
12. P. Henneaux, Probability of failure of overloaded lines in cascading failures. Int. J. Electr.
Power Energy Syst. 73, 141–148 (2015)
13. S.N. Ashida et al., Risk assessment of dynamic system cascading collapse for determining the
sensitive transmission lines and severity of total loading conditions. Reliability Engineering &
System Safety 157, 113–128 (2017)
14. M. Shengwei et al., An improved OPA model and blackout risk assessment. IEEE Trans. Power
Syst. 24(2), 814–823 (2009)
15. H. Ren et al., Using transmission line outage data to estimate cascading failure propagation in an
electric power system. IEEE Transactions on Circuits and systems 55(9), 927–931 (2008)
16. J. Qi et al., An interaction model for simulation and mitigation of cascading failures. IEEE
Trans. Power Systems 30(2), 804–819 (2015)
17. P. Rezaei et al., Estimating cascading failure risk with random chemistry. IEEE Trans. Power
Syst. 30(5), 2726–2735 (2015)
18. J. Kim et al., Estimating propagation and distribution of load shed in simulations of cascading
blackouts. IEEE Syst. J. 6(3), 548–557 (2012)
19. L. Che et al., Identification of cascading failure initiated by hidden multiple-branch contin-
gency. IEEE Trans. Reliab. 68(1), 149–160 (2019)
20. M.H. Athari, Z. Wang, Stochastic cascading failure model with uncertain generation using
unscented transform. IEEE Transactions on Sustainable Energy 11(2), 1067–1077 (2019)
21. J. Guo et al., Quantifying the influence of component failure probability on cascading blackout
risk. IEEE Trans. Power Syst. 33(5), 5671–5681 (2018)
22. J. Cepeda et al., Data-mining-based approach for predicting the power system post-contingency
dynamic vulnerability status. International Transactions on Electrical Energy Systems 25(10),
2515–2546 (2015)
23. S. Gupta et al., Support-vector-machine-based proactive Cascade prediction in smart grid using
probabilistic framework. IEEE Trans. Ind. Electron. 62(4), 2478–2486 (2015)
24. I. Dobson, Estimating the propagation and extent of cascading line outages from utility data
with a branching process. IEEE Trans. Power Syst. 27(4), 2146–2155 (2012)
25. Y. Zhang et al., Mitigating blackouts via smart relays: A machine learning approach. Proc. IEEE
99(1), 94–118 (2011)
26. M.R. Salimian, M.R. Aghamohammadi, A three stages decision tree-based intelligent blackout
predictor for power systems using brittleness indices. IEEE Transactions on Smart Grid 9(5),
5123–5131 (2018)
27. M. Abedi, M.R. Aghamohammadi, DT based intelligent predictor for out of step condition of
generator by using PMU data. Electr. Power Energy Syst. 99, 95–106 (2018)
28. N. Senroy et al., Decision tree assisted controlled islanding. IEEE trans. Power System 21,
1790–1797 (2006)
10 Evaluation and Classification of Cascading Failure Occurrence Potential. . . 235

29. G.V. Kass, An exploratory technique for investigating large quantities of categorical data.
Application Statistic 29(2), 119–127 (1980)
30. L. Breiman et al., Classification and Regression Trees (Wadsworth International Group,
Belmont, CA, 1984)
31. K. Max, K. Johnson, Applied Predictive Modeling (Springer, New York, 2013)
32. H. Liu et al., Feature selection with dynamic mutual information. Elsevier, Pattern Recognition
42(7), 1330–1339 (2009)
33. B. Remeseiro, V. Bolon-Canedo, A review of feature selection methods in medical applications.
Comput. Biol. Med. 112, 1–9 (2019)
34. T. Cover, Thomas, Elements of Information Theory (Wiley, New York, 2006)
35. M. Bennasar et al., Feature selection using joint mutual information maximization. Expert Syst.
Appl. 42, 8520–8532 (2015)
36. B. Zhang et al., Mutual information maximization-based collaborative data collection with
calibration constraint. IEEE Access 7, 21188–21200 (2019)
37. M. Khaji, M.R. Aghamohammadi, Online emergency damping controller to suppress power
system inter-area oscillation using load-generation tripping. Electrical Power System Research
140, 806–820 (2016)
38. Khaji M and Aghamohammadi M.R (2017) Emergency transmission line switching to suppress
power system inter-area oscillation. Int. J. Electr. Power Energy Syst. 87:52–64
39. M.R. Salimian, M.R. Aghamohammadi, Intelligent out of step predictor for inter area oscilla-
tions using speed-acceleration criterion as a time matching for controlled islanding. IEEE
Transactions on Smart Grid 9(4), 2488–2497 (2016)
40. M. Majidi, M.R. Aghamohammadi, New design of intelligent load shedding algorithm based on
critical line overloads to reduce network cascading failure risks. Turkish Journal of Electrical
Engineering & Computer Sciences 22(6), 1395–1409 (2014)
41. M.R. Aghamohammadi et al., A new approach for mitigating blackout risk by blocking
minimum critical distance relays. Int. J. Elect. Power Energy Syst. 75, 162–172 (2016)
42. Z. Shuangxi et al., Power System Voltage Stability and its Control (China Electric Power Press,
Beijing, 2003)
Chapter 11
LSTM-Assisted Heating Energy Demand
Management in Residential Buildings

Amin Mansour-Saatloo, Arash Moradzadeh, Sahar Zakeri,

and Behnam Mohammadi-Ivatloo

11.1 Introduction

The proliferation of urbanization in recent years requires a great deal of multiple

energy sources to supply citizens. However, lack of energy sources and limitation of
fossil fuel-based sources usage owing to climate change and global warming push
researchers to find optimal energy management methods to handle this issue
[1, 2]. Load prediction is one of the promising solutions, which can assist buildings’
energy systems to operate in an optimal and reasonable way. In the case of residen-
tial buildings, thermal energy is the foremost used energy. Heating load prediction is
the premise for heating operation and dispatching [3].
According to research, commercial and residential buildings today have the
highest daily energy consumption in people’s lives. Therefore, intelligent control
of building energy consumption can be considered as the most important tool in
energy saving and management. Many solutions have been introduced in various
studies for energy management of buildings, each of which has its own advantages
and disadvantages. Meanwhile, forecasting the heating load of a building can be
considered as one of the most basic management solutions in the energy of build-
ings, which can have better advantages than other methods [3, 4].

A. Mansour-Saatloo · A. Moradzadeh · S. Zakeri

Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
e-mail: [email protected]; [email protected];
[email protected]
B. Mohammadi-Ivatloo (*)
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
Department of Energy Technology, Aalborg University, Aalborg, Denmark
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 237
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_11
238 A. Mansour-Saatloo et al.

Prediction of heat load has a complicated procedure that requires dealing with a
nonlinear optimization problem. So far, many scholars were studied heat load
prediction deploying various methods. Some literature was deployed regression
analysis algorithms, e.g., a nonlinear autoregressive exogenous model (NARX)
with external input was used in [5] to commercial buildings heat load prediction.
In [6], a real-time prediction based on nonlinear predictive control besides advanced
machine learning methods was developed to heating and cooling load prediction.
The performance of the linear and nonlinear autoregressive model with exogenous
inputs was discussed in [7] to different load prediction. In addition, in [8, 9]
autoregressive moving average (ARMA) and in [10, 11] autoregressive integrated
moving average (ARMIA) scheme were deployed to load prediction. In [12],
seasonal autoregressive integrated moving average (SARIMA) was deployed to
heat load prediction. However, owing to multiple dimensions of heat load charac-
teristics, regression-based methods cannot be further developed. In addition, regres-
sion methods require time series and cannot utilize long-term historical data.
Realization of the data-driven modeling and prediction with neural network and
machine learning development cause to resolve the aforementioned problems. An
online sequential extreme learning machine (OSELM) was developed in [13] to
heating and cooling load prediction. Multilayer perceptron (MLP) and support
vector machine (SVM) performance for heat load prediction were analyzed in
[14]. Nine different extreme machine learning methods were developed for heating
load forecasting in [15]. A dynamic neural network was deployed in [16] to heating
load prediction of an apartment building. Various machine learning techniques
including SVM, regression tree, etc. were applied in [17] to residential and com-
mercial buildings load prediction. One of the deep learning applications called deep
neural network (DNN) has been employed in [18] to predict the heating and cooling
loads of a residential building. The proposed scheme in this study is trained based on
the structural features of the building as network input to predict the heating and
cooling loads as the network output. In [19], the firefly searching algorithm to
optimize SVM parameters was used in the case of heating load prediction based
on the SVM algorithm. In a similar context, the least-square SVM (LSSVM) was
used to multiple load prediction in [20]. Support vector regression (SVR), DNN, and
extreme gradient boosting were utilized to heating load prediction in [21]. In [22], a
temporal conventional neural network (CNN) was deployed to heating load predic-
tion. In [23], strand-based long short-term memory (LSTM) was used to heating load
prediction. A review of some valuable literature related to heating load forecasting
was done. Table 11.1 lists the literature in question by utilized method, evaluation
metric, data type, and purpose.
Some of the above reviewed studies suffer from the inability of the proposed
method to analyze continuous data and others from the lack of correlation between
input data. In this paper, prediction of heating load in building is performed using the
LSTM method as a suitable method for time series data analysis. The LSTM predicts
the heating load by applying on data related to basic factors in the structure of the
building.
11 LSTM-Assisted Heating Energy Demand Management in Residential Buildings 239

Table 11.1 Review of studies related to building heating load predicting

Refs. Year Data type Method Evaluation metric Propose
[5] 2017 Real NARX Mean squared nor- Estimating heat
malized error demand of com-
mercial buildings
[6] 2014 Real Adaptive Gaussian Root mean square Estimating local
process, hidden Mar- error (RMSE) weather condition
kov model, episode to reduce the
discovery and semi- heating and cooling
Markov model energy
consumption
[7] 2014 Real Linear ARX, NARX Mean absolute per- Estimating of
centage error heating, cooling
(MAPE) and load
forecasting
[8] 2003 Real ARMA Relative error Estimating electri-
cal load
[9] 2013 ARMA, ARIMA, Root mean square Comparison of dif-
autoregressive ANN error (RMSE), mean ferent methods in
bias error (MBE) forecasting of
inflow of Dez dam
reservoir
[11] 2018 Real ARMIA Mean absolute per- Estimating electri-
centage error cal load of residen-
(MAPE) tial and industrial
buildings
[12] 2016 Real SARMIA Adapted MAPE Estimating heat
(AMAPE) demand
[13] 2018 Simulated OSELM Mean absolute error Estimating energy
(MAE) consumption and
energy-efficient
building design
[14] 2018 Real MLP, SVM Root mean square Estimating heat
error (RMSE) demand
[15] 2016 Real Multi-step ahead Root means square Estimating heat
predictive model error (RMSE) demand
based on extreme
leaning machine
[16] 2016 Real Dynamic neural net- Root mean square Estimating heating
work (DNN) error (RMSE), the and cooling
mean absolute per- demand of apart-
centage error ment buildings
(MAPE), and the
absolute fraction of
variation (R2)
[17] 2016 Real SVM, regression Root mean square Estimating of heat
tree, feed forward error (RMSE) demand of residen-
neural network, mul- tial and commercial
tiple linear regression buildings
[18] 2020 Simulated DNN
(continued)
240 A. Mansour-Saatloo et al.

Table 11.1 (continued)

Refs. Year Data type Method Evaluation metric Propose
Variance accounted Estimating energy
for (VAF) consumption and
energy-efficient
building design
[19] 2016 Real Optimized SVM RMSE, Pearson cor- Estimating heat
relation coefficient, demand
coefficient of
determination
[20] 2020 Real LSSVM MAPE, RMSE Estimating of elec-
tricity, heating,
cooling, and gas
demand of indus-
trial park
[21] 2019 Real SVR, DNN, RMSE Estimating heat
XGBoost demand
[22] 2020 Real TCN MAE, MAPE, RMSE Estimating heat
demand of residen-
tial buildings
[23] 2020 Real Strand-based LSTM Mean absolute per- Heating operation
centage error management and
(MAPE) dispatching

LSTM falls into the recurrent neural network (RNN) and shows excellent per-
formance in different applications. Indeed, the LSTM is the improved version of the
RNN that covers disadvantages of it. In [24], LSTM was utilized to wind energy
prediction. Extreme Gradient Boost (XGBoost) and LSTM as shallow learning and
deep learning tools were compared in [25] in the building short-term (i.e., 1 h ahead)
load prediction area. Long-term energy consumption was predicted using LSTM in
[26]. In [27, 28], photovoltaic energy production was predicted using LSTM. In [29],
the LSTM network was trained using pinball loss instead of mean square error
(MSE) for individual consumers load prediction. In addition to the power system
applications, LSTM was used in medical diagnosis [30, 31], air pollution prediction
[32], human action recognition [33], etc.
According to the reviewed literature and to the best of the author’s knowledge,
there is no application of the LSTM in the residential buildings heating load
prediction. To fill this research gap, this chapter focuses on the heating load
prediction of residential buildings using LSTM as a deep learning technique. As it
is known, heating consumption of buildings is mainly depending on weather and
environmental conditions. In doing so, this chapter aims to conduct the prediction
based on weather and environmental conditions. Environmental and weather char-
acteristics along with buildings structure characteristics are the input of the LSTM
and heating value is the target.
The rest of this chapter is organized as follows: Section 11.2 describes the detail
of the utilized dataset. Section 11.3 provides the LSTM mathematical model.
11 LSTM-Assisted Heating Energy Demand Management in Residential Buildings 241

Section 11.4 discusses the obtained results from the simulation. Finally, the last
section concludes the chapter.

11.2 Dataset

The utilized dataset in this study contains 768 samples with 8 attributes and
1 decision variable for each sample. z1, z2, . . ., z8 are the attributes, which refer to
eight different aspects, i.e., relative compactness, surface area, wall area, roof area,
overall height, orientation, glazing area, and glazing area distribution, that aim to
predict the heat value w as a decision variable. All the simulations were conducted at
Ecotect software for 12 residential buildings by Tsanas and Xifara [34]. Each
building was made up using 18 elementary cubes (3.5
3.5
3.5) with identical
materials for all buildings. It was assumed that buildings are located in Greece,
Athens, and 60% humidity, 0.6 clothing, 0.3 m/s air speed, and 300 lx lightning level
were considered as environmental factors. Moreover, infiltration of 0.5 for air
change with the sensitivity of 0.25 air changer per hour was applied during simula-
tions. Table 11.1 reviews the features and response of the utilized dataset.

11.3 LSTM Mathematical Model

Today, deep learning plays an increasing role in most parts of people’s daily lives. It
has been able to make significant progress in scientific and industrial applications
such as diagnosing a variety of diseases in medicine, face recognition and finger-
printing in security issues, fault detection in power systems, and energy forecasting
issues around the world. Deep learning has a variety of techniques, each of which
can play a role in different applications [35, 36]. Among the various types of deep
learning techniques, the LSTM is used as a powerful tool with a very high ability to
process time series data. The LSTM is the improved algorithm of the RNN proce-
dure, which can overcome RNN’s drawbacks such as the vanishing gradient prob-
lem with keeping memory in their channels [37]. Adding and removing information
are possible in the LSTM neural network by using cell state through the following
gates that make the LSTM superior to the RNN [23]. The mathematical structure of
LSTM is formulated as follows [23, 37]:

it ¼ σ ðW ii zt þ bii þ W hi ht1 þ bhi Þ ð11:1Þ

f t ¼ σ W if zt þ bif þ W hf ht1 þ bhf ð11:2Þ


gt ¼ tanh W ig zt þ big þ W hg ht1 þ bhg ð11:3Þ
242 A. Mansour-Saatloo et al.

Fig. 11.1 Main block diagram corresponding to the one unit of LSTM

ot ¼ σ ðW io zt þ bio þ W ho ht1 þ bho Þ ð11:4Þ

ct ¼ f t  ct1 þ it  gt ð11:5Þ
ht ¼ ot  tanh ðct Þ ð11:6Þ
1
σ ðzÞ ¼ ð11:7Þ
1 þ ez

where, it, ft, gt, ot 2 ℝh are the input gate, forget gate, cell gate, and output gate,
respectively; ct, ht 2 ℝh are the cell state and hidden state; zt 2 ℝd is the input vector;
σ is the sigmoid function that is given in Eq. (11.7); Wii, Wif, Wig, Wio 2 ℝh
d are the
trainable parameters of input vector for input, forget, cell, and output gates, respec-
tively, as well as Whi, Whf, Whg, Who 2 ℝh
h for output vector; bii, bhi, bif, bhf, big,
bhg, bio, bho 2 ℝh are trainable biases; and * is the scalar product and tanh is the
active function. Figure 11.1 shows the LSTM block diagram based on the equations
provided.

11.4 Simulation and Numerical Results

In this chapter, predicting the heating load of a building by the LSTM is done by
considering the influential factors in the structure of the building as input data and
the amount of heating load related to each condition as a target. When the network is
11 LSTM-Assisted Heating Energy Demand Management in Residential Buildings 243

Fig. 11.2 Regression results for heating load prediction by LSTM in training stage

trained in the amount of heating load used for each building condition, it will be able
to determine the amount of consumption for each structure. In the training phase, the
network learns data-related behavioral patterns so that it can predict system behavior
for the future. Thus, when the LSTM is trained based on some basic parameters
affecting the structure of the building and the heating load associated with them, it
will be possible to predict the heating load of other buildings based on structural
parameters. Accordingly, an ideal approach to managing heating energy consump-
tion in buildings is presented to consumers. Once designed, the LSTM is trained with
70% of the data, and the remaining 30% of the data is used for network test.
Figure 11.2 shows the heating load prediction results for the training data. This
figure provides a correlation coefficient (CC) for LSTM in predicting the heating
load by training data. It can be seen that the prediction was done with high accuracy
(CC ¼ 0.9984) and the network was able to train well the behavioral patterns of the
building heating load in various conditions. Figure 11.3 evaluates the error values
related to the forecast results for the training stage in mean squared error (MSE) and
mean absolute error (MAE) formats.
The network learns data-related behavioral patterns during the training stage and
after completing this stage can be saved as a black box to be used for future
predictions. The heating load prediction results for the test data are presented in
Fig. 11.4. It can be seen that when the network is trained with high accuracy, it is
able to provide a favorable prediction for test data. The evaluation of the network
error values for predicting test data by MSE and RMSE indicators is shown in
Fig. 11.5. After training and testing the network and based on the accuracy and error
values obtained, a prediction error in the form of a histogram can be obtained.
Figure 11.6 shows the test error in the form of a histogram.
From the results presented in Fig. 11.4, it is clear that the LSTM is able to predict
the heating load of the building for test data with high accuracy (0.9980) with
244 A. Mansour-Saatloo et al.

Fig. 11.3 Error values for training stage of heating load prediction by LSTM in the forms of MSE
and RMSE

Fig. 11.4 Heating load prediction by LSTM for test data

previous appropriate training. Also, the error values presented in Figs. 11.5 and 11.6
indicate the proper performance of the LSTM network during the test stage.
According to the results presented in the training and test of the LSTM network to
predict the heating load of the building, it is now possible to predict the amount of
heating load by considering the new input data for the LSTM network. So, it is
understandable that deep learning techniques in addition to the ability in other
sciences and industrial applications can be considered as a powerful tool in the
energy management of buildings or energy networks. It should be noted that the
11 LSTM-Assisted Heating Energy Demand Management in Residential Buildings 245

Fig. 11.5 LSTM error values for predicting heating load by test data in the forms of MSE and
RMSE

Fig. 11.6 LSTM error values for predicting heating load by test data in the form of histogram

proposed method can provide acceptable performance for real-world data, especially
time-series data.
246 A. Mansour-Saatloo et al.

11.5 Conclusion

Due to the amount of energy consumption by commercial and residential buildings,

energy management in these buildings has created many challenges. In this paper,
energy management in a residential building was performed by predicting the
heating load of the building. To make an accurate prediction, one of the deep
learning techniques called LSTM was introduced and employed. To use the pro-
posed technique, data related to a simulated building in Ecotect software were
utilized. The basic parameters in the structure of the building were considered as
the LSTM input variables to predict the heating load of the building as an output
variable. After training and test the network, the results were analyzed using
statistical evaluation indicators. The results showed the ability of the LSTM method
with accuracy CC ¼ 99.84% and CC ¼ 99.80% for the training and test stages of
heating load forecasting of a residential building, respectively.

References

1. A. Mansour-Saatloo, A. Moradzadeh, B. Mohammadi-Ivatloo, A. Ahmadian, A. Elkamel,

Machine learning based PEVs load extraction and analysis. Electronics (Switzerland) 9(7),
1–15 (2020). https://doi.org/10.3390/electronics9071150
2. M.Z. Oskouei, B. Mohammadi-Ivatloo, M. Abapour, A. Ahmadian, M.J. Piran, A novel
economic structure to improve the energy label in smart residential buildings under energy
efficiency programs. J. Clean. Prod. 260, 121059 (2020). https://doi.org/10.1016/j.jclepro.2020.
121059
3. A. Moradzadeh, A. Mansour-Saatloo, B. Mohammadi-Ivatloo, A. Anvari-Moghaddam, Perfor-
mance evaluation of two machine learning techniques in heating and cooling loads forecasting
of residential buildings. Appl. Sci. (Switzerland) 10(11), 3829 (2020). https://doi.org/10.3390/
app10113829
4. A. Moradzadeh, O. Sadeghian, K. Pourhossein, B. Mohammadi-Ivatloo, A. Anvari-
Moghaddam, Improving residential load disaggregation for sustainable development of energy
via principal component analysis. Sustainability (Switzerland) 12(8), 3158 (2020). https://doi.
org/10.3390/SU12083158
5. A. Sandberg, F. Wallin, H. Li, M. Azaza, An analyze of Long-term Hourly District heat demand
forecasting of a commercial building using neural networks. Energy Procedia 105, 3784–3790
(2017). https://doi.org/10.1016/j.egypro.2017.03.884
6. B. Dong, K.P. Lam, A real-time model predictive control for building heating and cooling
systems based on the occupancy behavior pattern detection and local weather forecasting.
Building Simulation 7(1), 89–106 (2014). https://doi.org/10.1007/s12273-013-0142-7
7. K.M. Powell, A. Sriprasad, W.J. Cole, T.F. Edgar, Heating, cooling, and electrical load
forecasting for a large-scale district energy system. Energy 74, 877–885 (2014)
8. S.-J. Huang, K.-R. Shih, Short-term load forecasting via ARMA model identification including
non-Gaussian process considerations. IEEE Trans. Power Syst. 18(2), 673–679 (2003)
9. M. Valipour, M.E. Banihabib, S.M.R. Behbahani, Comparison of the ARMA, ARIMA, and the
autoregressive artificial neural network models in forecasting the monthly inflow of Dez dam
reservoir. J. Hydrol. 476, 433–441 (2013)
10. J. Contreras, R. Espinola, F.J. Nogales, A.J. Conejo, ARIMA models to predict next-day
electricity prices. IEEE Power Engg. Rev. 22(9), 57 (2002)
11 LSTM-Assisted Heating Energy Demand Management in Residential Buildings 247

11. D. Alberg, M. Last, Short-term load forecasting in smart meters with sliding window-based
ARIMA algorithms. Vietnam J. Comput. Sci. 5(3–4), 241–249 (2018)
12. T. Fang, R. Lahdelma, Evaluation of a multiple linear regression model and SARIMA model in
forecasting heat demand for district heating system. Appl. Energy 179, 544–552 (2016)
13. S. Kumar, S.K. Pal, R.P. Singh, A novel method based on extreme learning machine to predict
heating and cooling load through design and structural attributes. Energ. Buildings 176,
275–286 (2018). https://doi.org/10.1016/j.enbuild.2018.06.056
14. M. Dahl, A. Brun, O.S. Kirsebom, G.B. Andresen, Improving short-term heat load forecasts
with calendar and holiday data. Energies 11(7), 1678 (2018). https://doi.org/10.3390/
en11071678
15. S. Sajjadi et al., Extreme learning machine for prediction of heat load in district heating systems.
Energ. Buildings 122, 222–227 (2016)
16. S. Sholahudin, H. Han, Simplified dynamic neural network model to predict heating load of a
building using Taguchi method. Energy 115, 1672–1678 (2016)
17. S. Idowu, S. Saguna, C. Åhlund, O. Schelén, Applied machine learning: Forecasting heat load
in district heating system. Energ. Buildings 133, 478–488 (2016). https://doi.org/10.1016/j.
enbuild.2016.09.068
18. S.S. Roy, P. Samui, I. Nagtode, H. Jain, V. Shivaramakrishnan, B. Mohammadi-ivatloo,
Forecasting heating and cooling loads of buildings: A comparative performance
analysis. J. Ambient. Intell. Humaniz. Comput. 11(3), 1253–1264 (2020). https://doi.org/10.
1007/s12652-019-01317-y
19. E.T. Al-Shammari et al., Prediction of heat load in district heating systems by support vector
machine with firefly searching algorithm. Energy 95, 266–273 (2016). https://doi.org/10.1016/j.
energy.2015.11.079
20. Z. Tan et al., Combined electricity-heat-cooling-gas load forecasting model for integrated
energy system based on multi-task learning and least square support vector
machine. J. Clean. Prod. 248, 119252 (2020)
21. P. Xue, Y. Jiang, Z. Zhou, X. Chen, X. Fang, J. Liu, Multi-step ahead forecasting of heat load in
district heating systems using machine learning algorithms. Energy 188, 116085 (2019). https://
doi.org/10.1016/j.energy.2019.116085
22. J. Song, G. Xue, X. Pan, Y. Ma, H. Li, Hourly heat load prediction model based on temporal
convolutional neural network. IEEE Access 8, 16726–16741 (2020). https://doi.org/10.1109/
ACCESS.2020.2968536
23. J. Liu, X. Wang, Y. Zhao, B. Dong, K. Lu, R. Wang, Heating load forecasting for combined
heat and power plants via Strand-based LSTM. IEEE Access 8, 33360–33369 (2020). https://
doi.org/10.1109/ACCESS.2020.2972303
24. F. Shahid, A. Zameer, A. Mehmood, M.A.Z. Raja, A novel wavenets long short term memory
paradigm for wind power prediction. Appl. Energy 269, 115098 (2020)
25. Z. Wang, T. Hong, M.A. Piette, Building thermal load prediction through shallow machine
learning and deep learning. Appl. Energy 263, 114683 (2020)
26. J.Q. Wang, Y. Du, J. Wang, LSTM based long-term energy consumption prediction with
periodicity. Energy 197, 117197 (2020)
27. K. Wang, X. Qi, H. Liu, Photovoltaic power forecasting based LSTM-Convolutional Network.
Energy 189 (2019). https://doi.org/10.1016/j.energy.2019.116225
28. M. Gao, J. Li, F. Hong, D. Long, Day-ahead power forecasting in a large-scale photovoltaic
plant based on weather classification using LSTM. Energy 187, 115838 (2019)
29. Y. Wang, D. Gan, M. Sun, N. Zhang, Z. Lu, C. Kang, Probabilistic individual load forecasting
using pinball loss guided LSTM. Appl. Energy 235, 10–20 (2019). https://doi.org/10.1016/j.
apenergy.2018.10.078
30. M. Rahman, I. Saha, D. Islam, R.J. Mukti, A deep learning approach based on convolutional
LSTM for detecting diabetes. Comput. Biol. Chem., 107329 (2020)
248 A. Mansour-Saatloo et al.

31. İ. Kırbaş, A. Sözen, A.D. Tuncer, F.Ş. Kazancıoğlu, Comperative analysis and forecasting of
COVID-19 cases in various European countries with ARIMA, NARNN and LSTM approaches.
Chaos, Solitons & Fractals, 110015 (2020)
32. Y.-S. Chang, H.-T. Chiao, S. Abimannan, Y.-P. Huang, Y.-T. Tsai, K.-M. Lin, An LSTM-Based
Aggregated Model for Air Pollution Forecasting (Atmospheric Pollution Research, New York,
2020)
33. Z. Zhang, Z. Lv, C. Gan, and Q. Zhu, Human Action Recognition Using Convolutional LSTM
and Fully-Connected LSTM with Different Attentions. (Neurocomputing, 2020)
34. A. Tsanas, A. Xifara, Accurate quantitative estimation of energy performance of residential
buildings using statistical machine learning tools. Energ. Buildings 49, 560–567 (2012). https://
doi.org/10.1016/j.enbuild.2012.03.003
35. A. Moradzadeh and K. Pourhossein, Short circuit location in transformer winding using deep
learning of its frequency responses. In Proceedings 2019 International Aegean Conference on
Electrical Machines and Power Electronics, ACEMP 2019 and 2019 International Conference
on Optimization of Electrical and Electronic Equipment, OPTIM 2019, (2019), pp. 268–273,
doi: https://doi.org/10.1109/ACEMP-OPTIM44294.2019.9007176
36. A. Shrestha, A. Mahmood, Review of deep learning algorithms and architectures. IEEE Access
7, 53040–53065 (2019). https://doi.org/10.1109/ACCESS.2019.2912200
37. A. Moradzadeh, S. Zakeri, M. Shoaran, B. Mohammadi-Ivatloo, F. Mohamamdi, Short-term
load forecasting of microgrid via hybrid support vector regression and long short-term memory
algorithms. Sustainability 12(17), 7076 (2020)
Chapter 12
Wind Speed Forecasting Using Innovative
Regression Applications of Machine
Learning Techniques

Arash Moradzadeh, Amin Mansour-Saatloo, Morteza Nazari-Heris,

Behnam Mohammadi-Ivatloo, and Somayeh Asadi

12.1 Introduction

The increasing energy users and the widespread expansion of renewable energy
sources (RESs) have significantly expanded the world’s distribution systems. Con-
sumers and especially users of sensitive loads tend to have access to a reliable and
sustainable power supply [1]. Therefore, power producers need short-term planning
for sustainable investment, production, and operation. On the other hand, the
exorbitance use of fossil fuels, including oil, natural gas, and coal, paves the way
for climate change and global warming by releasing a high amount of greenhouse
gases, especially carbon dioxide [2, 3]. For a better future, it is needed to limit the
usage of fossil fuels. One promising solution is utilizing renewable energy sources

A. Moradzadeh · A. Mansour-Saatloo
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
e-mail: [email protected]; [email protected]
M. Nazari-Heris
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
Department of Architectural Engineering, Pennsylvania State University,
State College, PA, USA
e-mail: [email protected]
B. Mohammadi-Ivatloo (*)
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
Department of Energy Technology, Aalborg University, Aalborg, Denmark
e-mail: [email protected]
S. Asadi
Department of Architectural Engineering, Pennsylvania State University,
State College, PA, USA

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 249
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_12
250 A. Moradzadeh et al.

(RESs) in the energy sector instead of fossil fuel-based plants. Wind, solar, hydro,
biomass, tidal, and geothermal are the main RESs worldwide [4]. However, wind
energy owing to its potentials in implementing and operating has a significant place
among these alternatives. So, a viable and secure forecasting mechanism can assist
the power system’s operation more steadily and reliably [5, 6]. The wind energy is
mostly depending on wind speed, so it is stochastic and volatile, which causes an
adverse impact on the power system [7]. Wind speed is one of the weather variables
with a stochastic nature and is complicated to forecast. Generally, from the forecast-
ing time interval point of view, three different forecasting methods can be defined as
short-term forecasting with less than 6 h intervals, midterm forecasting with 6 h to
1-day intervals, and long-term forecasting with more than 1 day [8, 9]. Since short-
term forecasting is an appropriate method for scheduling and operation of the power
system, this work is focused on this method.
So far, many scholars attempted to introduce models to forecast wind energy
[10]. Some of them utilized physical methods, including atmosphere physical
conditions such as mass, momentum, etc. by applying physic laws, e.g., in [11],
Kalman filter, and in [12], flower pollination method were used to wind speed and
wind power forecasting. Some other studies were applied statistical methods, e.g., a
hybrid wavelet transform (WT) based on repeated auto-regressive integrated moving
average (ARIMA) was deployed in [13] to wind speed forecasting, which in order to
enhance the accuracy a new model, namely, RWT-ARIMA, was introduced. In a
similar context, nested ARIMA was deployed in [14] to wind speed forecasting. In
[15], auto-regressive moving average (ARMA) was applied to wind speed and wind
direction forecasting. The ARMA based algorithm for wind speed forecasting was
introduced in [16], in which variational mode decomposition was applied to decom-
pose the wind speed to linear, nonlinear, and noise parts. In [17], wind speed has
been forecasted via several solutions such as simple average strategy, traditional
combination forecasting method (TCFM) with multi-objective grasshopper optimi-
zation algorithm (MOGOA), and optimized extreme learning machine (ELM). In
this study, the results of comparisons emphasize the optimized ELM’s better per-
formance than other methods. In [18], for wind speed forecasting, the general
regression neural network to deal with the system’s nonlinearity and the ARIMA
model was developed. In [19], statistical methods were applied to forecast wind
power loss due to ice growth. However, statistical models provide a linear correla-
tion, and they are not suitable for nonlinear relationships.
In recent years artificial intelligence and machine learning techniques such as
artificial neural network (ANN) [20, 21], generalized neural network (GRNN) [22],
fuzzy logic models [23], support vector machine (SVM) [24], multilayer perceptron
(MLP) [25], group method of data handling (GMDH) [26], etc. were developed
rapidly, in which nonlinear structures can be effectively captured. In doing so, many
kinds of literature have been deployed machine learning techniques to forecast the
wind speed. For example, in [27], support vector regression (SVR), and in [28],
SVM, were applied to forecast the wind speed. A probabilistic extreme learning
machine model to wind generation forecasting was introduced in [29], where
bootstrap methods were used to consider the forecasting model’s uncertainty.
12 Wind Speed Forecasting Using Innovative Regression Applications of Machine. . . 251

Three different machine learning methods, i.e., SVR, ANN, and Gaussian process,
were utilized in [29] to capture the correlation between numerical weather prediction
and wind output. Extreme machine learning and deep learning methods were applied
in [30]. To reach a more accurate model, generalized correntropy was used instead of
the Euclidean norm of the mean square error. LSTM network was utilized in [31] to
extract the temporal features of different nodes in deep learning from nearby wind
forms wind speed and wind direction. A tree-based machine learning method was
proposed in [32] for short-term wind speed forecasting. In [33], a short-term wind
power forecasting using a hybrid model, decomposition and extreme learning
machine was proposed. The authors in [34] have proposed a deep learning neural
network for forecasting wind power based on the databases of supervisory control
and data acquisition (SCADA) with a high frequency rate. In this reference, the
proposed model’s main objectives are mentioned as reduction of the computational
efforts and costs in the forecasting process and attaining high accurate results for the
forecasted power. In [35], a numerical weather forecasting model for wind speed
error correction based on gated recurrent unit neural networks has been presented to
obtain a short-term prediction of wind power. A mathematical modelling is for wind
power prediction is proposed in [36], where genetic algorithm as heuristic optimi-
zation method is applied for dealing with the forecasting time delay. The authors in
[37] have proposed a combined model for forecasting of the wind speed for different
time horizons in various locations, where three various approaches including ANN,
ARIMA, and the integration of these methods have been used for evaluating the
results of the forecasting process. In this reference, numerical error evaluation
techniques are used for prediction of the model accuracy.
According to the reviewed scholars, there is no focus on applying the MLP and
GMDH methods of machine learning in the wind power forecasting. The current
work aims to investigate the performance of these methods, i.e., MLP and GMDH, in
the short-term wind power forecasting. MLP is a multi-input multi-output (MIMO)
system with a backpropagation learning algorithm. The specific advantages of the
nonlinear multilayer structure and fast convergence make the MLP favorable for the
regression problems [38, 39]. GMDH is another data-driven method that benefits
from subtractive clustering for data selection and is easy to implement in modeling
complex systems [26].
The rest of the chapter is organized as follows: Section 12.2 describes the utilized
data set. Section 12.3 provides methods and mathematical theories. The next section
discusses the simulation results, and, finally, the last section concludes the chapter.

12.2 Data Set

The selected wind farm for this study is located in the Khaf area with a longitude of
34.567
north and 60.148
east. The utilized data includes air temperature, relative
humidity, solar radiation, and wind speed at a 30-m height measured with a 10-min
resolution for 15 months from July 2007 to September 2008. The target is
252 A. Moradzadeh et al.

50
Air temperature (°C) 40

30

20

10

-10

-20
0 1 2 3 4 5 6
× 104
Time (10-minute interval)

Fig. 12.1 Air temperature

100
90
Relative humidity (%)

80
70
60
50
40
30
20
10
0
0 1 2 3 4 5 6
× 104
Time (10-minute interval)

Fig. 12.2 Relative humidity

forecasting the wind speed, so the deployed machine learning techniques find the
correlation between air temperature, humidity, solar radiation, and wind direction as
inputs and wind speed as output. The air temperature for the studied location is
shown in Fig. 12.1, which is based on a 10-min resolution. Figure 12.2 shows the
utilized data set of the network for humidity. The solar radiation for the selected
location is demonstrated in Fig. 12.3, which shows that the radiation has lower
values at the second and third months considering the sun’s location compared to the
earth. Finally, the wind direction and wind speed, which are used for defining the
power output of the wind turbines, are shown in Figs. 12.4 and 12.5, respectively.
Because wind speed is based on many variables consisting of temperature and
climate parameters such as air temperature and humidity, solar radiation, and wind
direction are needed for performing the forecasting process.
12 Wind Speed Forecasting Using Innovative Regression Applications of Machine. . . 253

1200

Radiation (W/m2) 1000

800

600

400

200

0
0 1 2 3 4 5 6
× 104
Time (10-minute interval)

Fig. 12.3 Solar radiation

400
350
300
Wind direction (°)

250
200
150
100
50
0
0 1 2 3 4 5 6
× 104
Time (10-minute interval)

Fig. 12.4 Wind direction

12.3 Methods

Data mining was introduced as an emerging and growing technology [36]. Data
mining has many techniques, each of which is used in some way to process data and
a tool to discover hidden knowledge of large and complex data. Hidden knowledge
in data can be considered as intrinsic patterns, features, correlations, and relation-
ships between data. The most widely used data mining techniques are artificial
neural networks and machine learning [37]. Artificial neural network (ANN) tech-
niques, derived from human intelligence, have long been used as tools for pattern
recognition, classification, feature extraction, and prediction applications. These
techniques can predict future events by learning and supervising past behavior
[38]. During the training process, the ANN techniques model the correlation
between input variables, and then the prediction operation is performed based on
254 A. Moradzadeh et al.

35
Wind speed (m/s) 30

25

20

15

10

0
0 1 2 3 4 5 6
× 104
Time (10-minute interval)

Fig. 12.5 Wind speed

the estimated model [39]. The ANNs have a variety of algorithms, one of the most
widely used being multilayer perceptron (MLP) which is used in this paper for wind
speed forecasting. Machine learning is one of the applications of data mining that
aims to develop computer programs and artificial intelligence. Machine learning
algorithms are able to identify data patterns, develop predictive models, and estimate
in regression and classification modes. Learning techniques in machine learning
algorithms are performed in three modes: supervised, semi-supervised, and
unsupervised. High-dimensional data processing is one of the most important
features of machine learning techniques [37]. Among all machine learning algo-
rithms, the group method of data handling (GMDH) is selected as a powerful tool in
big data processing in this paper for wind speed forecasting.

12.3.1 Multilayer Perceptron (MLP)

The MLP, as a neural network with a layer-to-layer and feed-forward structure, has a
special position in solving regression problems and predicting continuous data. The
MLP consists of one input layer, one output layer, and one or multiple hidden layers,
which uses a supervised learning algorithm, namely, backpropagation. Each layer’s
neurons are fully connected to the previous and next layers’ neurons through
weighting elements. The number of layers and parameters related to the MLP
structure is directly related to the problem under consideration. While the number
of hidden layer neurons is mainly adjusted and selected by trial and error method.
The training procedure in the MLP method is based on the minimization of the
selected cost function and weight functions are defined during this procedure
[25]. Each perceptron maps the inputs to the outputs using a nonlinear activation
function and output signals generated via a nonlinear transfer function. The typical
architecture of an MLP network is shown in Fig. 12.6 [40].
12 Wind Speed Forecasting Using Innovative Regression Applications of Machine. . . 255

Fig. 12.6 Typical Hidden layer

architecture of the MLP
network Input layer Output layer
1
w11 f11 2
w11
2
z1 1
w12 w12
f12
f21

z2 f22
f31

zN fM2
1
wnm 2
fP1 wpm

The input layer consists of N dimension and the hidden layer includes P neurons
with n weight elements for each neuron. The matrix IN  P denotes the input weight
as Eq. (12.1), and the matrix JP  M denotes the output weight as Eq. (12.2), where

T
T
M is the number of output signals. f 1 ¼ f 11 , f 12 , ⋯, f 1P and f 2 ¼ f 21 , f 22 , ⋯, f 2P
are the transfer function vectors of the hidden layer and output layer, respectively, as

T
Eqs. (12.3)–(12.4); z ¼ [z1, z2, ⋯, zN]T is the input vector; b1 ¼ b11 , b12 , ⋯, b1N

T
b2 ¼ b21 , b22 , ⋯, b2N are the bias vectors of the hidden layer and output layer,

T
respectively; and y ¼ y21 , y22 , ⋯, y2M is the output vector.
2

2 3
w111 ⋯ w11P
6 7
INP ¼ 4 ⋮ ⋱
⋮ 5 ð12:1Þ
w1N1 . . . w1NP
2 2 3
w11 ⋯ w21M
6 7
JPM ¼ 4 ⋮ ⋱ ⋮ 5 ð12:2Þ
w2P1 . . . w2PM
 
y1 ¼ f 1 b1 þ I:z ð12:3Þ
 
y2 ¼ f 2 b2 þ J:y1 ð12:4Þ

1.1. Group Method of Data Handling.

GMDH is a machine learning method that introduced by Alexey Grigorevich
Ivakhnenko [41]. The GMDH is also known as a polynomial neural network and is a
256 A. Moradzadeh et al.

Input variables First generation

Second generation
x1

x2
kth generation

Optimal output
x3
y*

x3

x4

x5

Selected block Eliminated block

Fig. 12.7 The GMDH structure

nonlinear regression algorithm. The GMDH aims to find a precise function as much
as possible that predicts the output of b
y a given input vector x ¼ [x1, x2, . . ., xN] in
complex systems based on natural selection to handle the system’s complexity. The
GMDH method uses the Kolmogorov-Gabor polynomial function as following [42]:

X
m m X
X m m X
X m X
m
by ¼ c0 þ ci xi þ cij xi x j þ cijg xi x j xg þ . . . ð12:5Þ
i¼1 i¼1 j¼1 i¼1 j¼1 g¼1

where, c ¼ [c0, ci, cij, . . .] is the vector of weights. To determine the model’s structure
and find the optimal model to generate the output, the GMDH algorithm uses the
lower order of the Kolmogorov-Gabor polynomial function. The first order of the
polynomial is given in Eq. (12.6), which is used in each generation, as shown in
Fig. 12.7. The generation of models continues until the process shows over-fitting.

by ¼ c0 þ c1 x1 þ c2 x2 þ . . . cN xN ð12:6Þ
12 Wind Speed Forecasting Using Innovative Regression Applications of Machine. . . 257

12.4 Simulation and Numerical Results

Forecasting wind speed based on the required data as input variables and influencing
the wind speed curve formation requires tools and methods of processing big data. In
this chapter, machine learning applications called GMDH and MLP have been
selected to forecast wind speed. Using these methods requires a database as input
and network formation. Given that wind speed is a parameter dependent on temper-
ature and climatic conditions, accordingly in this chapter, the parameters of air
temperature, relative humidity, solar radiation, and wind direction are considered
as input variables to form the input data set. After forming the input data set and the
design of the networks, each network is trained using 70% of the data. The rest of the
data is used as test data to predict wind speed. In the training phase, each network
analyzes the input data and, by determining the weights and applying the bias to the
input data, predict the output value. After each of the training and test stages, each
stage’s results should be evaluated using acceptable evaluation metrics. It should be
noted that the evaluation of results can be considered the most important part of
research work. In this chapter, the results of the training and test stages to predict
wind speed are evaluated using three statistical performance metrics such as corre-
lation coefficient (R), mean square error (MSE), root mean square error (RMSE), and
mean absolute error (MAE) for each method. Details of each of these indicators are
provided in [8, 43]. Figure 12.8 shows each network’s result in the training phase for
predicting and recognizing patterns related to input data.
It can be seen that in the training phase, the GMDH model with 99.91% of
R-value had better efficiency than the MLP method. Based on this, it is expected that
this network will provide better results for the test stage and forecasting of wind
energy than the MLP method. After training, test data is employed as network input
to predict the wind speed. Each network provides forecasting the wind speed based
on the done training. Figure 12.9 shows the results of wind speed prediction by the
test data. Figure 12.10 presents the evaluation of wind speed forecast results using
MSE and RMSE metrics for GMDH and MLP networks.
It can be seen that Fig. 12.10 shows the amount of prediction error values for each
test sample (each input data for 10 min) and finally a mean value of errors is
presented for the whole test step. It should be noted that the error values obtained
for each model shown in Fig. 12.10 are in m/s unit. The results of the evaluations
compare the accuracy and efficiency of each of the methods used. It can be seen that,
as the training phase, in this phase, the GMDH method was able to provide better
predictive results with lower error values. However, in this chapter, a comparative
approach is presented to emphasize the effectiveness of the GMDH method in big
data processing as well as to provide appropriate wind speed forecasts. In this
approach, the MLP and GMDH methods’ results are compared with the results of
other similar studies. It should be noted that the comparison of the methods’
performance should be done so that the data used for all methods are the same.
This comparison is performed so that in each of these studies, the input variables
selected for each method are the same. The compared methods used the same
258 A. Moradzadeh et al.

Fig. 12.8 The results of R value for each network in the training stage; (a) MLP (b) GMDH

meteorological data with the same time interval for sampling to predict wind speed.
Table 12.1 makes this comparison.
After comparing the results and evaluating each method’s performance, it is
observed that the GMDH method can be used as an ideal method for processing
large data due to its layer-to-layer structure. It was observed that the methods used in
this chapter were able to predict wind speed at any point in time as well as in
different weather conditions. Accurate wind speed forecasts can provide proper
planning to wind power plants to have a proper performance for generating electric-
ity through wind energy. Besides, other machine learning techniques can be used as
an accurate tool to perform the types of predictions needed in power system issues.
12 Wind Speed Forecasting Using Innovative Regression Applications of Machine. . . 259

Fig. 12.9 Results of wind speed prediction in the test phase; (a) MLP (b) GMDH

12.5 Conclusion

Today, wind power plants are considered as one of the essential RESs around the
world. Wind speed is one of the most effective factors in the production of electrical
energy by wind turbines, the accurate prediction of which can be used as an
impressive approach in the performance of wind power plants. In this chapter, two
machine learning applications called the MLP and GMDH were employed to
forecast the wind speed. To express the effectiveness of the suggested methods,
data related to the Khaf region in Iran were utilized. Each of the MLP and GMDH
techniques predicted the wind speed by training with input variables such as air
temperature, humidity, solar radiation, and wind direction. The results of the pre-
dictions were evaluated with different performance evaluation metrics that it was
260 A. Moradzadeh et al.

Fig. 12.10 Rate of MSE and RMSE errors to forecast wind speed; (a) MLP (b) GMDH

Table 12.1 Performance Method R MSE RMSE MAE

evaluation of various methods
MLP in this chapter 0.9969 0.0061 0.0782 0.0445
used to forecast wind speed
GMDH in this chapter 0.9982 0.0046 0.0684 0.0439
WT-ARIMA [13] – – 0.4328 0.3369
RWT-ARIMA [13] – – 0.3204 0.2269
Nested ARIMA [14] – – – 0.0446
TCFM-MOGOA [17] – – 0.4026 0.2046
Optimized ELM [17] – – 0.2531 0.1845

observed that the GMDH method with high accuracy (R ¼ 99.82%) than the MLP
(R ¼ 99.62%) presented better performance. Finally, in a comparative approach with
different methods used to forecast wind speed, the high capability of the GMDH
12 Wind Speed Forecasting Using Innovative Regression Applications of Machine. . . 261

technique was emphasized. It should be noted that the GMDH method can be used as
a powerful tool to solve other power system problems that depend on big data.
Wind power forecasting based on wind speed predictions and the improvement of
machine learning algorithms for time series data predictions can be considered as
future work. In addition, using statistical regression to evaluate and analyze the
sensitivity of each of the input variables in the forecast model can significantly
improve the forecast results in future work.

References

1. O. Sadeghian, A. Moradzadeh, B. Mohammadi-Ivatloo, B. Mohammadi-Ivatloo, M. Abapour,

F.P.G. Marquez, Generation units maintenance in combined heat and power integrated systems
using the mixed integer quadratic programming approach. Energies 13(11), 2840 (2020).
https://doi.org/10.3390/en13112840
2. R. K. Pachauri et al., Climate change 2014: synthesis report. Contribution of Working Groups I,
II and III to the fifth assessment report of the Intergovernmental Panel on Climate Change.
(IPCC, 2014)
3. A. Moradzadeh, O. Sadeghian, K. Pourhossein, B. Mohammadi-Ivatloo, A. Anvari-
Moghaddam, Improving residential load disaggregation for sustainable development of energy
via principal component analysis. Sustainability (Switzerland) 12(8), 3158 (2020). https://doi.
org/10.3390/SU12083158
4. Ren21. 2019. Renewables 2019 Global Status Report. Paris: Ren21Secretariat. Accessed
16 Nov 2019
5. S. Madadi, B. Mohammadi-Ivatloo, S. Tohidi, Dynamic line rating forecasting based on
integrated factorized Ornstein–Uhlenbeck processes. IEEE Trans Power Deliv 35(2),
851–860 (2019)
6. S. Madadi, B. Mohammadi-Ivatloo, and S. Tohidi, Probabilistic Real-Time Dynamic Line
Rating Forecasting Based on Dynamic Stochastic General Equilibrium with Stochastic Vola-
tility (IEEE Transactions on Power Delivery, 2020)
7. Y. Hao, C. Tian, A novel two-stage forecasting model based on error factor and ensemble
method for multi-step wind power forecasting. Appl. Energy 238, 368–383 (2019)
8. D.B. Alencar, C.M. Affonso, R.C.L. Oliveira, C.R. Jose Filho, Hybrid approach combining
SARIMA and neural networks for multi-step ahead wind speed forecasting in Brazil. IEEE
Access 6, 55986–55994 (2018)
9. A. Moradzadeh, S. Zakeri, M. Shoaran, B. Mohammadi-Ivatloo, F. Mohamamdi, Short-term
load forecasting of microgrid via hybrid support vector regression and long short-term memory
algorithms. Sustainability (Switzerland) 12(17), 7076 (2020). https://doi.org/10.3390/
su12177076
10. M. Lei, L. Shiyan, J. Chuanwen, L. Hongling, Z. Yan, A review on the forecasting of wind
speed and generated power. Renew. Sust. Energ. Rev. 13(4), 915–920 (2009)
11. F. Cassola, M. Burlando, Wind speed and wind energy forecast through Kalman filtering of
numerical weather prediction model output. Appl. Energy 99, 154–166 (2012)
12. W. Zhang, Z. Qu, K. Zhang, W. Mao, Y. Ma, X. Fan, A combined model based on CEEMDAN
and modified flower pollination algorithm for wind speed forecasting. Energy Convers. Manag.
136, 439–451 (2017)
13. S.N. Singh, A. Mohapatra, Repeated wavelet transform based ARIMA model for very short-
term wind speed forecasting. Renew. Energy 136, 758–768 (2019)
14. S.-K. Sim, P. Maass, P.G. Lind, Wind speed modeling by nested ARIMA processes. Energies
12(1), 69 (2019)
262 A. Moradzadeh et al.

15. E. Erdem, J. Shi, ARMA based approaches for forecasting the tuple of wind speed and
direction. Appl. Energy 88(4), 1405–1414 (2011)
16. S. Smyl, A hybrid method of exponential smoothing and recurrent neural networks for time
series forecasting. Int. J. Forecast. 36(1), 75–85 (2020)
17. P. Jiang, Z. Liu, Variable weights combined model based on multi-objective optimization for
short-term wind speed forecasting. Appl. Soft Comput. 82, 105587 (2019)
18. W. Zhao, Y.-M. Wei, Z. Su, One day ahead wind speed forecasting: A resampling-based
approach. Appl. Energy 178, 886–901 (2016)
19. S. Scher, J. Molinder, Machine learning-based prediction of icing-related wind power produc-
tion loss. IEEE Access 7, 129421–129429 (2019)
20. H. Li, J. Wang, R. Li, H. Lu, Novel analysis–forecast system based on multi-objective
optimization for air quality index. J. Clean. Prod. 208, 1365–1383 (2019)
21. P. Du, J. Wang, W. Yang, T. Niu, Container throughput forecasting using a novel hybrid
learning method with error correction strategy. Knowl.-Based Syst. 182, 104853 (2019)
22. A. Mansour-Saatloo, A. Moradzadeh, B. Mohammadi-Ivatloo, A. Ahmadian, A. Elkamel,
Machine learning based PEVs load extraction and analysis. Electronics (Switzerland) 9(7),
1–15 (2020). https://doi.org/10.3390/electronics9071150
23. H. Yang, Z. Jiang, H. Lu, A hybrid wind speed forecasting system based on a ‘decomposition
and ensemble’strategy and fuzzy time series. Energies 10(9), 1422 (2017)
24. X. Kong, X. Liu, R. Shi, K.Y. Lee, Wind speed prediction using reduced support vector
machines with feature selection. Neurocomputing 169, 449–456 (2015)
25. A. Moradzadeh, A. Mansour-Saatloo, B. Mohammadi-Ivatloo, A. Anvari-Moghaddam, Perfor-
mance evaluation of two machine learning techniques in heating and cooling loads forecasting
of residential buildings. Appl. Sci. (Switzerland) 10(11), 3829 (2020). https://doi.org/10.3390/
app10113829
26. D.H. Lim, S.H. Lee, M.G. Na, Smart soft-sensing for the feedwater flowrate at PWRs using a
GMDH algorithm. IEEE Trans. Nucl. Sci. 57(1), 340–347 (2010)
27. G.-R. Ji, P. Han, and Y.-J. Zhai, Wind speed forecasting based on support vector machine with
forecasting error estimation. In 2007 International Conference On Machine Learning and
Cybernetics, (2007), vol. 5, pp. 2735–2739
28. N. Shabbir, R. AhmadiAhangar, L. Kütt, M. N. Iqbal, and A. Rosin, Forecasting short term
wind energy generation using machine learning. In 2019 IEEE 60th International Scientific
Conference on Power and Electrical Engineering of Riga Technical University (RTUCON),
2019, pp. 1–4
29. C. Wan, Z. Xu, P. Pinson, Z.Y. Dong, K.P. Wong, Probabilistic forecasting of wind power
generation using extreme learning machine. IEEE Trans. Power Syst. 29(3), 1033–1044 (2013)
30. X. Luo et al., Short-term wind speed forecasting via stacked extreme learning machine with
generalized correntropy. IEEE Trans. Indust. Inform. 14(11), 4963–4971 (2018)
31. M. Khodayar, J. Wang, Spatio-temporal graph deep neural network for short-term wind speed
forecasting. IEEE Trans. Sustain. Energy 10(2), 670–681 (2018)
32. A. Ahmadi, M. Nabipour, B. Mohammadi-Ivatloo, A.M. Amani, S. Rho, M.J. Piran, Long-term
wind power forecasting using tree-based learning algorithms. IEEE Access 8, 151511–151522
(2020). https://doi.org/10.1109/ACCESS.2020.3017442
33. D. Zhang, X. Peng, K. Pan, Y. Liu, A novel wind speed forecasting based on hybrid decom-
position and online sequential outlier robust extreme learning machine. Energy Convers.
Manag. 180, 338–357 (2019)
34. L. Zhang, F. Tian, Performance study of multilayer perceptrons in a low-cost electronic nose.
IEEE Trans. Instrum. Meas. 63(7), 1670–1679 (2014)
35. A. Moradzadeh, K. Khaffafi, Comparison and evaluation of the performance of various types of
neural networks for planning issues related to optimal management of charging and discharging
electric cars in intelligent power grids. Emerging Sci. J. 1(4), 201–207 (2017). https://doi.org/
10.28991/ijse-01123
12 Wind Speed Forecasting Using Innovative Regression Applications of Machine. . . 263

36. A. Moradzadeh, K. Pourhossein, B. Mohammadi-Ivatloo, F. Mohammadi, Locating inter-turn

faults in transformer windings using isometric feature mapping of frequency response traces.
IEEE Trans. Indust. Inform, 1–1 (2020). https://doi.org/10.1109/tii.2020.3016966
37. I. H. Witten, E. Frank, and Mark A. Hall, Data Mining: Practical Machine Learning (2011)
38. A. Moradzadeh and K. Pourhossein, Early detection of turn-to-turn faults in power transformer
winding: an experimental study. In Proceedings 2019 International Aegean Conference on
Electrical Machines and Power Electronics, ACEMP 2019 and 2019 International Conference
on Optimization of Electrical and Electronic Equipment, OPTIM 2019 (2019), pp. 199–204,
doi: https://doi.org/10.1109/ACEMP-OPTIM44294.2019.9007169
39. S. Souahlia, K. Bacha, A. Chaari, MLP neural network-based decision for power transformers
fault diagnosis using an improved combination of Rogers and Doernenburg ratios DGA.
Int. J. Electr. Power Energy Syst. 43(1), 1346–1353 (2012). https://doi.org/10.1016/j.ijepes.
2012.05.067
40. A. Moradzadeh and K. Pourhossein, Early detection of turn-to-turn faults in power transformer
winding: an experimental study. In Proceedings 2019 International Aegean Conference on
Electrical Machines and Power Electronics, ACEMP 2019 and 2019 International Conference
on Optimization of Electrical and Electronic Equipment, OPTIM 2019, (2019), pp. 199–204,
doi: https://doi.org/10.1109/ACEMP-OPTIM44294.2019.9007169
41. A.G. Ivakhnenko, New methods of control-system investigation. Control 3(30), 96–99 (1960)
42. A. G. Ivakhnenko, Polynomial Theory of Complex Systems: IEEE Transactions on Systems,
Man and Cybernetics. (ISYMAW, 1971)
43. A. Moradzadeh and K. Pourhossein, Application of support vector machines to locate minor
short circuits in transformer windings. In 2019 54th International Universities Power Engi-
neering Conference (UPEC), (2019), pp. 1–6
Chapter 13
Effective Load Pattern Classification by
Processing the Smart Meter Data Based
on Event-Driven Processing and Machine
Learning

Saeed Mian Qaisar and Futoon Alsharif

13.1 Introduction

At the present, the demand for intelligent energy management is increasing. It is due
to the growing world population, declining oil reserves, and rising energy consum-
mation, happening because of increased usage of modern gadgets. In this frame-
work, the smart grid idea has arisen from attempts to make power grids more
efficient, cleaner, environmentally sustainable, and functional by creating technical
possibilities spanning all processes from electricity generation to delivery and use
[1]. Smart grid analyzes and monitors the transfer of electricity from the generation
unit to the customer. Thanks to current power networks and advanced metering
network, this scenario is more realistic for future technology deployment [2]. The
smart grid includes the advanced metering infrastructure (AMI). Smart meters are
the most important aspect of modern grid infrastructures.
All have become digital with the major advances in the world of the Internet and
technology. The Internet has been a huge part of our lives. In this picture, a novel
technology, namely, the Internet of Things (IoT), is developed. This is a network that
contains several electronic devices and sensors that are wired together to share such

S. Mian Qaisar (*)

Electrical and Computer Engineering Department, College of Engineering, Effat University,
Jeddah, Saudi Arabia
Communication & Signal Processing Lab, Energy & Technology Center, Effat University,
Jeddah, Saudi Arabia
e-mail: [email protected]
F. Alsharif
Electrical and Computer Engineering Department, College of Engineering, Effat University,
Jeddah, Saudi Arabia

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 265
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_13
266 S. Mian Qaisar and F. Alsharif

Data Acquision, Storage Data Processing Data Interpretaon

and Transmission and Analysis and Applicaon

Fig. 13.1 Components of the smart meter data intelligence chain [5]

information across the Internet. IoT-based apps tend to converse and exchange data
with each other. As sophisticated infrastructures, they link a variety of electrical
devices that can communicate with each other. The IoT is effectively incorporated to
develop the AMIs. Communication among devices is realized via a variety of
techniques. Home area network is used to connect the smart meter to the smart
devices. It monitors and manages and the micro generation network. It offers unified
control, resources, and facility maintenance. The linkage between the smart meter
and appliances can be realized via a wired or a wireless protocol such as Wi-Fi.
AMI distinguishes automatic smart reading (AMR) schemes. AMRs work with
less evolved technology such as manually switching off equipment, gathering data
offline, and so on. However, AMIs include real-time processes, customer/pricing/
usage options, remote-controlled and automated maintenance systems, etc. With
recent developments, the incorporation of AMI is crucial in smart grids. Neighbor-
hood area network is used for data sharing between smart meters in the neighbor-
hood. This supports communications for diagnosis, system updates, and real-time
notifications. ZigBee protocol is commonly used in NAN thanks to fast data transfer
rates and low cost. Many smart meters are attached to a central server over the wide
area network. The networking systems GSM, GPRS, 3G, and WiMAX will be used
to link these meters to the server.
Home smart meters measure fine granular energy usage in real time and are seen
as the basis for a future smart power grid. Technological developments have
increased the use of smart meters as a replacement of the traditional ones
[3, 4]. These meters are key components of smart grids that provide significant
civic, environmental, and economic benefits to various stakeholders [5]. The mas-
sive smart meter installations require a huge amount of data collection with the
desired granularity [6]. Automated data acquisition, storage, processing, and inter-
pretation are the main factors behind the performance of smart meters. The process is
demonstrated in the block diagram in Fig. 13.1.
A fine-grained metering data is required to attain realistic benefits in terms of
performance and resilience to multiple smart grid stakeholders [5, 6]. Every stake-
holder has a different objective from the other: suppliers, which are companies
buying electricity from the wholesale market and selling it back to consumers in
the retail market, want to lower the operational overheads associated with manual
meter reading and enhance customer loyalty potentially. The transmission system
and distribution network operators are expecting to get benefited from a more
flexible demand side to allow more low-carbon technology penetration. Govern-
ments are expecting the enhancements of energy efficiency in the end-use side,
which are provided by smart meters, and will help to reach the goals of binding
carbon reduction. End users are hoping to get benefited from the lower bills of
13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 267

electricity as they become more aware of energy. With such expectations, it is no

surprise that smart meters are experiencing a period of rapid growth.
During peak hours, the local consumption of electricity is about 40% of the total
residential sector consumption and more than a third of the total power demand
[7]. Consumer awareness will help in predicting the demand and adapt future
investments. Giving regular feedback on electricity usage is helpful for the end
users to have more details about their consumption of electricity which supports the
measures for effective use of energy [5, 7, 8].
In this framework, several automated appliance classification techniques have
been reported [9–11]. In [9], authors have used zero-normalization consumption
time series with delta and delta-delta coefficients as attributes of intended appliances.
The classification is performed by using the Gaussian mixture model (GMM)
classifier. In [10], authors have used entropy of consumption time series with delta
and delta-delta coefficients as attributes of intended appliances. The classification is
performed with hidden Markov models (HMM). In [11], authors have used zero-
normalization of real power, reactive power, and root mean square (RMS) current
time series with delta and delta-delta coefficients as attributes of intended appliances.
The classification is performed with hidden Markov models (HMM).
Conventionally, the smart meters data is collected in a time-invariant fixed-rate
manner [9–11], resulting in the collection, storage, processing, and classification of a
significant amount of unnecessary data [12]. In the suggested metering framework:
• Event-driven sensing is incorporated to attain significant real-time data
compression.
• A novel adaptive rate concept is presented for data collection, processing, and
classification.
• It improves the designed solution’s compression and processing effectiveness
compared to classical fix-rate counterparts.

13.2 Background and Literature Review

13.2.1 Smart Metering

Nowadays, the traditional grids of power are being replaced by smart grids around
the world [13]. Solar and wind are renewable energy sources which the smart grid
includes [13]. The smart grid allows the two-way communication and exchange of
data between the electricity providers and end users [13]. The smart grid allows the
two-way communication and exchange of data between the electricity providers and
end users. They are devices that monitor energy usage to the consumer for con-
sumption reduction purposes [14]. In the residential area, these meters determine the
consumption of power at fine granularities in real time and are being extensively
deployed worldwide because of their significant advantages to the industry of
electricity supply and its customers [5, 7, 8]. There are many smart metering
268 S. Mian Qaisar and F. Alsharif

Fig. 13.2 Applications and advantages of smart meters

applications such as quality of power and reliability monitoring, analysis, modeling

and forecasting of loads, consumer energy management, and much more [6]. The
following block diagram provided by the authors of [15] shows some examples of
the applications and advantages of smart meters:
The following points represent the characteristics and advantages of smart
metering mentioned in the above Fig. 13.2 [15]:
• Automatic collection, processing, transmitting, monitoring, and usage of
metering data.
• Automatic meter control.
• Two-way communication.
• Gives accurate and up-to-the-minute consumption data to the related people.
• Provides help to the services which enhance the efficiency of energy for the
consumption and billing.
The smart meter database represents the consumption data of typical home
appliances provided by a smart meter. Many studies have been done using smart
meter data, such as [16, 17]. The authors of [16] addressed aggregated demand
decomposition methodology which was based on sub metering that is allowing smart
meter measurements and artificial neural network. The approach in general results in
an estimate of the shares of various load classes and a combination of controllable
and uncontrollable load within the overall expected load, with the implementation of
forecast in different demand response systems [16]. Figure 13.3 shows its load
disaggregation flow chart. It represents the demand decomposition technique in
the system of the smart meter with just a few consumers who can monitor each
appliance they have, as a relatively practical situation for the potential smart distri-
bution grid. The authors of [16] had made two assumptions in this regard.
13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 269

Fig. 13.3 Flow chart for load disaggregation

The real power of each appliance can be recorded in the smart meters, whereas
reactive load is extracted based on probabilities [16].Total consumption (actual and
reactive power) is already estimated in the substation which is in block 5 [16]. In
block 1, the smart meter data is being preprocessed and collected at a concentrator
stage [16]. In block 2, the section of the consumption that is submetered can be
broken down into groups or manageable and unmanageable load by directly adding
up the consumption of the devices which belong to the same group [16]. In the
following stage, this sub-metric data trains the artificial neural network (ANN)
classifier, so that it gets the ability to identify the load classification using the active
and reactive load graph of the controlled consumers [16]. When the ANN is trained,
in block 4, the ANN utilizes the prediction of the total active and reactive loads at the
concentrator stage in block 5 as input and then provides the associated load classi-
fication, that is the weighting factors of each load group as output in block 6 [16].

13.2.2 Smart Meter Data Acquisition

In [17], the researchers presented a three-phase model that was beneficial for the
suppliers of electricity because of its demand management flexibility and energy
control effectiveness. The solution is an ensemble of unsupervised clustering and
supervised classification [17]. It can classify the consumers of electricity, recogniz-
ing their characteristics of consumption, and identifying recent customers [17]. In
the presented model, to perform load pattern clustering and characteristic recogni-
tion, all load patterns obtained from the extraction of load patterns are used in
customer categorization [17].
To extract its relevant features, the metering data is processed and analyzed.
Authors have reported the methods of extraction of the features in [6, 16]. The
270 S. Mian Qaisar and F. Alsharif

principle is to construct a collection of features that can reflect the information that is
essential for interpretation and classification, in the most practical way. There are
several tactics for feature extraction of the load profiles of appliances. Certain
examples are the short-time Fourier transform (STFT), wavelet transformation,
and K-means algorithm [18–20]. The derived features are peak values, average
values, the consumption’s root mean square (RMS) values, and their harmonics.

13.2.3 Feature Extraction

Feature selection techniques can be classified into two categories: classical tech-
niques and biologically oriented techniques as the authors of [21] had divided them.
The classical techniques are represented in statistics or syntactic nature, and the
biologically oriented techniques can be represented in a neural or genetic-based
algorithm [21]. A feature is a pattern with reduced-dimension representation [21]. To
lower the pattern representation dimensions, feature selection, and extraction in
pattern recognition depends on obtaining mathematical techniques [21]. To achieve
the reduction of dimensions, either feature selection is performed or feature extrac-
tion [21]. Some factors can be affected by the choice of features, attributes, or
measurements [21]. One is the class satisfaction accuracy, the second is the classi-
fication required to tie them, the third is the required number of examples for
learning, and the last is classification performing price [21]. For a feature to be
good, it needs to meet two conditions, and one should not change if any differences
happen within a class [21]. The second is that it should show the significant
differences when distinguishing between patterns of different classes [21]. There
are many feature extraction methods such as non-transformed signal characteristics,
transformed signal characteristics, structural descriptors, etc., and the following
Fig. 13.4 shows the main feature extraction methods that were mentioned in [21].

Fig. 13.4 Different methods of feature extraction

13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 271

13.2.4 Pattern Recognition

High-resolution smart meter data offer a variety of knowledge on customer habits

and preferences of energy use. Besides this, deregulation of the power market,
especially on the supply side, in many countries around the world continues to
push forward. Exploiting the massive smart meter data can promote and boost
efficiency and profitability on the demand side worldwide. In this sense, in recent
years, the electricity industry has experienced significant breakthroughs in machine
learning [6].
Recognition of concerned appliance patterns is attained by using the mined
features from their consumption time series. In [22], the classification of patterns
has been described by authors as a study of how machines can detect, learn to
discern, and make logical decisions about patterns. In [18], the authors used load
disaggregation pattern recognition. The idea is to collect the consumption informa-
tion of each concerned device separately without interference. The prepared database
is composed of the attributes, which can be explored and processed to classify the
concerned devices [18]. In [19], authors employed the automated pattern detection,
to exploit data which is collected from smart meters, to disaggregate overall energy
usage. The precession of pattern identification is dependent on the performance of
data collection, extraction of features, and classification. A tool used to apply a class
description to some given instances in the function space is called a classifier [22].
In [23], the authors proposed a fast approach for electrical appliance recognition
based on analyzing the signatures of the smart home load. These signatures are
defined by active and reactive power and fast Fourier transform of the signals of the
current [23]. The recognition algorithm they used to extract a decision tree was C4.5
[23]. One of the advantages of their approach was the automated decision tree
generation which enables the rapid deployment of the detection algorithm
irrespective of the smart home [23]. They tested 13 appliances, which were toaster,
electric kettle, two different stove burners, microwave, vacuum, oven, coffee maker,
blender, two different dryers with different modes, range hood fan, and mixer. The
classification rate achieved was 93.4397% of true positive [23].
The authors of [19] had proposed an approach of automatically identifying the
devices according to the distributed power measurement and actuation units. They
had two contributions to their paper. One is introducing their trace-base repository of
actual power consumption traces, and the other is evaluating the electrical appli-
ances’ automatic recognition based on the traces collected [19]. They extracted
517 features from the traces collected to prepare them for classifier training
[19]. They used different classification algorithms such as bagging, Bayesian net-
work, Naïve Bayes, random committee, etc. The highest classification accuracy they
got was 95.5% using the random committee classification method [19].
Bad power quality detection addresses may be lost data or irregular trends
triggered by unplanned occurrences or data collection, communication, or entry
failure. Detection of bad power quality can be broken down into predictive, machine
learning, and probabilistic approaches [6]. These methods and tools can be used
272 S. Mian Qaisar and F. Alsharif

directly on the collected smart meter data. Such studies are summed up as time
series-based methods, low-rank matrix technique, and time-window-based approach
as a function of the modeling techniques. Data on smart meters are simply time
series. An optimally weighted average approach for data denoising and classification
has been suggested. It can be incorporated for offline or online power issues
detection [6]. An ensemble of autoregressive with exogenous inputs (ARX) and
artificial neural network (ANN) is also used for the identification of power supply
issues [6]. The use of energy is correlated with both spatially and temporally.
Exploring the spatiotemporal connection may help to recognize and recuperate
outliers. A low-rank matrix fitting-based approach for performing data cleaning
and imputation was proposed in [6]. Because smart meter data is obtained in real-
time or near real-time fashion, an online mechanism of identification of anomalies
using the Lambda architecture has been developed in [6]. The suggested method of
online identification can be performed in parallel, providing high performance when
dealing with big datasets.
Under other techniques, energy theft will change the smart meter information.
Detection of energy theft can be achieved using smart meter data and status
information of the power grid, such as voltages of nodes. Supervised classifiers are
efficient energy theft detection techniques. They usually consist of two stages,
namely, extraction of features and classification. In [6], an approach is reported,
where the nontechnical failure was first measured to train a theft identification
classifier. The load profiles were clustered using K-means clustering. Various
potential destructive samples are used to train the classifier. Following a series of
suspicious detections, an energy theft warning is issued. The approach suggested
could also classify the category of energy theft. In addition to clustering-based
extraction of features, an encoding technique is first conducted on the load data in
[24], which served as classifier inputs such as SVM and a rule engine-based
algorithm for detecting energy theft. In [25], the authors have introduced an inde-
pendent variable that is implemented as a top-down structure. It is founded on the
principles of the decision tree and the SVM algorithms. The decision tree estimated
the expected energy consumption based on the number of appliances, the individ-
uals, and the ambient temperature. Then, the decision tree production was fed into
the SVM to decide whether the user is ordinary or abusive.
Load profiling corresponds to the characterization of the appliances’ consumption
manners or consumers as a function of the energy utilization manner. Specific
cluster-based approaches, such as K-means, hierarchical clustering algorithms, and
self-organizing maps (SOM), are utilized in this framework. They are applied
directly to the time series of consumption behavior, collected from the smart meters
[6]. In [26], authors investigated how the resolution of energy consumption time
series, collected via using smart meters, influences clustering results. The well-
known clustering techniques were conducted on the smart meter dataset. In this
case, the dataset is composed of a time series, collected at various sampling rates,
ranging from 1 min to 2 h. In [27], deep learning-based stacked sparse auto-encoders
were implemented to compress the load profile and extract the feature. A locality
sensitive hashing approach is further suggested to identify the load profiles and
13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 273

extract the descriptive load profiles based on the decreased and encoded load profile.
A profound analysis of the time series of consumption patterns, collected via smart
metes, is required to extract the pertinent local and global classifiable features. It
enhances the precision of the classification of load profiles. Three novel categories of
local and global features are suggested in this framework. These features are
produced via the effective application of conditional filters on the time series of
energy consumption. Moreover, the calibration and normalization, profile errors, and
shape signatures are also employed to realize an effective load profiling [28].
Using existing statistical information, the load forecasting methodology is often
used by electricity utilities for grid planning, generating capacity, demand control,
financial modeling, and potential energy use. Precise forecasting of electricity
charges is important for electricity providers to reduce financial risk, design power
grids, and maximize operating performance. The process of load forecasting could
be categorized into short-term, medium-term, and long-term predictions. Here the
short term refers to a period between 1 day and 1 week. The medium term refers to a
period from 1-week to 1-month, and the long term refers to the duration from months
to several years [6]. In this context, time series models with neural networks have
been used for short-term forecasting. Artificial neural networks have also been
employed for short-term forecasting [6]. Clustering approaches with support vector
machines have been used for medium-term forecasting [6].

13.3 Materials and Methods

Figure 13.5 shows the principle of the proposed system. The different system
modules are being described in the followed subsections.

13.3.1 Smart Meter Database

In this chapter, the evaluation results of the suggested solution are prepared while
studying its performance on the dataset ACS-F2 [9]. This dataset comprises of time
series of energy consumption parameters of 15 major home appliances [9]. In total
225 different appliances from 15 classes are considered. Two 1-h data collection and
recording sessions are conducted for each intended appliance [9]. Six different time
series are recorded for each recording session and from each considered appliance.
These are, respectively, real power, reactive power, RMS current, RMS voltage,
frequency, and power factor. Each time series is recorded in a disaggregated manner

Fig. 13.5 The system block diagram

274 S. Mian Qaisar and F. Alsharif

at an acquisition rate of 0.1 Hz. The recordings are made in a disaggregated fashion
[29]. In this study, only six categories of appliances are considered. It includes
kettles, fridges, freezers, microwave ovens, fans, monitors, and televisions. For each
intended device, real and reactive power consumptions are taken into account.

13.3.2 Signal Reconstruction

The reconstruction process is the process of producing a continuous-time signal that

goes along with the points of the discrete-time signal (i.e., it simply takes a group of
samples and converts them back to a continuous function) [30, 31]. The first step to
understand the reconstruction process is to generate a train of continuous-time
impulse from a sampled signal, which can be represented mathematically as follows:

X
1
ximp ðt Þ ¼ xs ðnÞδðt  nT s Þ: ð13:1Þ
n¼1

where xs(n) is the sampled signal and Ts is the sampling period. A low-pass filter
with specific conditions is then applied to generate an output signal exðt Þ. The result
of the construction process is as follows:

X
1
exðt Þ ¼ xs ðnÞgðt  nT s Þ: ð13:2Þ
n¼1

The considered consumption parameters waveforms are up-sampled with a factor

of 10,000. It is performed to evaluate the event-driven sensing module. A cascaded
arrangement of four stages of cubic-spline interpolators and anti-imaging filters is
utilized to conduct the up-sampling. In this way, the quasi-analog version of the
incoming signal y(tn) is obtained. It is named exðt Þ. The process of obtaining exðt Þ
up-sampling of y(tn) is given by Eq. (13.3), where U is the up-sampling factor.
   
!
exðt Þ ¼ y t Un x ðtÞ ¼ y tUn : ð13:3Þ

13.3.3 Event-Driven Sensing (EDS)

In the proposed case, the event-driven ADCs (EDADCs) are utilized for the acqui-
sition of the intended appliances’ energy consumption patterns [20, 32]. These data
converters are realized by using the principle of event-driven sensing (EDS). They
can change the rate of acquisition as a function of the incoming signal
13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 275

time-variations [33, 34]. In this case, a time series sample is recorded only when the
intended analog signal traverses one of the predefined threshold levels, placed across
the amplitude of the signal. Therefore, samples are divided in time in a nonuniform
!
manner. The size of the samples taken depends on the variations of the exðt Þ x ðtÞ
[33, 34]. The procedure can be represented mathematically with Eq. (13.4), where tn
represents the current sampling instance, tn  1 is the previous sampling instance, and
dtn represents the time distance between these two instances:

t n ¼ t n1 þ dt n : ð13:4Þ

The EDADC acquires only the relevant or active portions of the incoming analog
signal. On other hand, the remaining portion of the signal-like baseline is
disregarded. Therefore, relative to traditional counterparts, the obtained number of
samples is dramatically diminished. It renders a significant decrease in the amount of
collected information and could produce a notable real-time compression. It con-
tributes to the reduction of post-processing operations and improves the performance
of the system in terms of the time of execution and the usage of overhead power and
energy [33–36].

13.3.4 Event-Driven Segmentation

The segmentation operation separates the signal into several fixed-length portions
[37]. It permits the relevant parameters of the incoming signal to be identified and
mined effectively. The EDS output is placed in a nonuniform time-amplitude plane.
Therefore, the traditional segmentation mechanisms cannot be utilized in this case
[35–37]. In this context, an original event-driven segmentation mechanism is
employed. It is named the activity selection algorithm (ASA) [35, 36]. This parti-
tions the output of EDADC into portions of variable lengths. The ASA is conducted
by exploiting the nonuniformity of the event-driven acquired signal. It preserves the
valuable details, such as repartitioning among the consecutive sampling instants,
count of samples, etc. [34]. The EDADC provides the nonuniform data to the ASA
which selects the appropriate parts. Within the whole length of the signal, a series of
selected windows would be defined [36]. The activity selection is extremely impor-
tant to minimize the processing operation and hence the power consumption. It
permits the realization of post-adaptive rate feature extraction in an adaptive rate
manner. In this way, the useful features are mined solely in the time domain and are
utilized afterward by the classifiers for an automated load profiling.
276 S. Mian Qaisar and F. Alsharif

13.3.5 Feature Extraction

Pertinent features are mined from each portion. Thanks to the event-driven acquisi-
tion we can extract interesting and pertinent information regarding the frequency
content of the signal in the time domain [38]. In this way, the used technique of
feature extraction permits the extraction of important information from the power
consumption time series without requiring the complicated frequency domain trans-
formation. On another hand, the conventional counterparts are based on the fixed-
rate acquisition and processing methods, and therefore they are obliged to employ
the time-frequency analysis while performing the feature extraction. It confirms a
notable processing load effectiveness of the suggested solution in contrast to the
counterparts [37]. Therefore, the necessary classifiable features are mined by using
the nonuniformly recorded power consumption time series in only time domain.
Instantaneous real power consumption and instantaneous RMS current are con-
sidered for each designated appliance. Each considered time series is reconstructed
by using the suggested technique. Onward, the quasi-analog waveforms are recorded
by using the EDADC and segmented by using the ASA. Afterward, four different
parameters for each segment are extracted. Let i is indexing the ith selected segment,
Wi. Then, CiCi, ΔAi ‚aUA€ i , Ai Ai , and dt i
max max mean are respectively the extracted
number of threshold crossings, the peak-to-peak amplitude, the maximum ampli-
tude, and the average sampling step for WiWi. The EDS delivered time series of real
and reactive power are also considered. For each instance the collected version of the
incoming instances features are used to strengthen and improve the system precision.

13.3.6 Classification Techniques

The derived characteristics are used to distinguish targeted appliances. In this

context, we utilized the known algorithms of classification such as k-nearest neigh-
bor (KNN), artificial neural network (ANN), and Naïve Bayes.

13.3.6.1 K-Nearest Neighbor (KNN)

The KNN is regarded as an easy but powerful classifier and its ability to deliver high-
quality results even for applications that are known for their complexity [39]. It
depends on the nearest feature space training examples [22]. It classifies the
concerned object according to its majority vote by its neighbors [22]. In a dataset,
the features’ distance is used by KNN to decide which data belongs to what class.
When the distance in the data is near, a group is formed, and when the distance in the
data is far, other groups are formed. A category membership might be the output of
the KNN classifier. The categorization of an object is done through the majority vote
by its neighbors, the object being assigned to the most common classification among
13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 277

its k nearest neighbors (k could generally be a small positive whole number). The
object is assigned solely to the nearest neighbor’s single classification if the k equals
one [39].

13.3.6.2 Artificial Neural Network (ANN)

Processing data full of features is what the ANN classification method is known for
[39]. ANN is a model of neural network which is part of artificial intelligence.
Instead of programming the procedure system to do a certain number of tasks, this
classification method teaches the system to perform tasks [18]. Artificial intelligence
system (AI) is developed to execute these tasks. It is a useful model that can
understand the patterns hidden in the data that duplicate useful information fast
and accurately. Neural networks are one case of these AI models. AI systems should
be constantly discovered from knowledge [39]. The most available approaches are
artificial intelligence techniques in the fields of evaluation in relationships with
dissimilar information. A man-made neural network consists of several artificial
neurons that, according to requirements, are correlative together. The neural net-
work’s goal is to turn the inputs into critical outputs [6]. The teaching mode may be
monitored or not controlled. In the presence of noise, neural networks learn.

13.3.6.3 Naïve Bayes

The Naïve Bayes classification method is used. Naïve Bayes can be described as a
short version of the Bayes theorem [5]. It functions in a way that the probability of
one attribute is not affected by the probability of another attribute. It develops
independent assumptions of 2Q for a series of Q attributes. Authors of [6] examined
some of the Naïve Bayes classification method key performance criteria. They
concluded that the classifier accuracy is a function of training data noise, bias, and
variability. Only by choosing good training data, the noise can be reduced. Bias is
the error that is caused by very large groupings in the training data. Variance is the
error caused by the fact that these groupings are too small.

13.4 Results

The categories of appliances include kettles, fridges and freezers, microwave ovens,
fans, monitors, and televisions. Fifteen appliances of different brands from each
category are considered. Therefore a total of 90 appliances are taken into consider-
ation. For each appliance two 2-h retrieval sessions are held. It results in 180 multi-
dimensional time series of the power consumption-related characteristics of the
considered appliances. The classification techniques used in this case are KNN,
SVM, ANN, and Naïve Bayes. Examples of the intended instances of real power
 278 S. Mian Qaisar and F. Alsharif

Fig. 13.6 The real power instances digitized with a 4-bit resolution EDADC for kettles, fridges and
freezers, microwaves, and fans
InstanteniousRealPower(W)

InstanteniousRealPower(W)

Monitors (e) Televisions (g)

30
50
20

10 0
0

-10 -50
0 20 40 60 0 20 40 60
Time (Minutes) Time (Minutes)

Fig. 13.7 The real power instances digitized with a 4-bit resolution EDADC for monitors and
televisions

consumption and RMS current, obtained with the EDS mechanism, are shown
respectively in Figs. 13.6, 13.7, 13.8, and 13.9.
The current cannot be measured explicitly in actual application and requires the
current to voltage conversion. This is because most real-life data converters are built
on the principle of acquiring voltage waveforms. It could be realized by using charge
resistance or transimpedance mountings. However, for this simulation-based study,
the realization and design of this module are not considered.
 13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 279

Fig. 13.8 The RMS current instances digitized with a 4-bit resolution EDADC for kettles, fridges
and freezers, microwaves, and fans

Monitors (e) Televisions (g)

0.15 0.2
RMS Current(A)
RMS Current(A)

0.1
0.1
0
0.05
-0.1
0
-0.2
0 20 40 60 0 20 40 60
Time (Minutes) Time (Minutes)

Fig. 13.9 The RMS current instances digitized with a 4-bit resolution EDADC for monitors and
televisions

The obtained compression gains are summed up in Table 13.1. These are
computed as ratios between the collected number of samples, obtained in the
conventional and sugegsted approaches. It demonstrates that the suggested method
achieves a complete 3.1 times, 6.6 times, 1.6 times, 3.4 times, 2.2 times, and 6.0
times compression gains, respectively, for the case of kettles, fridges and freezers,
microwaves, fans, monitors, and televisions. The compression gain shows how
much we are going to gain in terms of the reduction in the amount of the collected
information and in the amount of information that is going to be processed and
transmitted. It aptitudes a noticeable reduction in the arithmetic complexity and the
280 S. Mian Qaisar and F. Alsharif

Table 13.1 Summary of the compression gains

Appliances Compression gain Average compression gain
Kettles 3.1 3.8
Fridges and freezers 6.6
Microwaves 1.6
Fans 3.4
Monitors 2.2
Televisions 6.0

Table 13.2 Accuracy for the six-class appliances consumption pattern recognition (KNN)
Appliances Classification accuracy (%age) Average classification accuracy (%age)
Kettles 91.5 90.9
Fridges and freezers 91.4
Microwaves 90.6
Fans 91.8
Monitors 89.9
Televisions 90.3

Table 13.3 Accuracy for the six-class appliances consumption pattern recognition (ANN)
Appliances Classification accuracy (%age) Average classification accuracy (%age)
Kettles 94.6 94.4
Fridges and freezers 95.5
Microwaves 93.3
Fans 95.2
Monitors 94.2
Televisions 93.8

power consumption of the proposed solution compared to the conventional

approach. Also, embedding the EDS in smart meters significantly decreases the
activity of data storage and collection compared to standard methods. It additionally
enhances the effectiveness of the designed strategy over the concurrent ones.
This case yields an overall 3.8-fold compression ratio. Compared to the tradi-
tional method, the proposed tactic has greatly reduced arithmetic complexity and
power consumption. Additionally, embedding the EDS into smart meters greatly
decreases data collection and transfer operation relative to conventional approaches.
This further increases the efficacy of the designed approach over the
conventional ones.
180 instances are considered for the 6 categories of appliances. For equal
representation, 30 instances are taken from each category. The limited dataset
could lead to biased classification evaluation. To prevent any bias in performance,
the tenfold cross-validation is utilized [22]. The accuracies in terms of the percentage
of correct recognition are listed respectively in Table 13.2, Table 13.3, and
Table 13.4 for the KNN, ANN, and Naïve Bayes classifiers.
13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 281

Table 13.4 Accuracy for the six-class appliances consumption pattern recognition (Naïve Bayes)
Appliances Classification accuracy (%age) Average classification accuracy (%age)
Kettles 87.6 86.9
Fridges and freezers 88.2
Microwaves 85.9
Fans 88.4
Monitors 84.8
Televisions 86.3

Table 13.5 Summary for the average classification techniques accuracy

Appliances KNN classifier ANN classifier Naïve Bayes classifier
Kettles 90.9 94.4 86.9
Fridges and freezers
Microwaves
Fans
Monitors
Televisions

Table 13.6 Summary for the average F-measure values

Appliances KNN classifier ANN classifier Naïve Bayes classifier
Kettles 0.907 0.937 0.865
Fridges and freezers
Microwaves
Fans
Monitors
Televisions

To overall summaries of results for different classifiers are presented by

Table 13.5 and Table 13.6, respectively, for the accuracy scores and F-measure
values.
From the above table, it is clearly shown that the ANN classifier achieves the best
performance. The best consumption pattern recognition of appliances is achieved by
employing the combination of EDADC, ASA, event-driven feature extraction, and
the ANN classification method.
Consequently, the suggested solution reduces considerably the complexity and
power consumption, lowers data transmission operation compared to the conven-
tional approach, and increases system efficiency.
282 S. Mian Qaisar and F. Alsharif

13.5 Discussion

Home smart meters are used in real time for calculating power usage at the finest
granularities and are seen as the bases of the future smart grid technological
developments have evolved the utilization of smart meters, instead of
traditional ones.
Such meters are the essential elements of smart grids and provide important
advantages for various stakeholders. These multitudes of advantages can be catego-
rized as social, environmental, and economic. The massive installations of smart
meters are producing a large volume of data collection with the desired granularity.
Automated information collection, storage, processing, and classification are the
main factors behind the performance of smart meters.
The collection of fine-grained metering data is important to give practical advan-
tages to several stakeholders of the smart grid in terms of performance and sustain-
ability. There are different goals for different stakeholders: vendors for example
would like to reduce the operating expenses involved with traditional meter reading
and significantly increase the loyalty of the consumer. The operators of the trans-
mission systems and distribution networks want to take advantage of a more robust
demand side that enables lower carbon technology penetration. Governments aim to
achieve the objectives of reducing carbon emissions by increasing the energy
efficiency in the consumer side, which is given by smart meters. Consumers are
having a better awareness of energy. Therefore, they are expecting to benefit from
the reduced electricity bills. With these goals, it is not surprising that smart meters
are having a faster growth period.
From a generalized viewpoint, identifying the type of device is a difficult task
because of several reasons. Firstly, there are possible overlaps among a wide variety
of types, for example, laptops and tablets. Secondly, there is a wide range of devices
that fall under the same group because of their unlike operating mechanisms and
technical changes that occur among appliances. Generally, the appliance identifica-
tion must guarantee the precise ability to simplify without exceeding a certain
number of appliances.
In many uses, the data collected through smart meters are analyzed. A smart
meter’s performance is dependent on algorithms that can run and perform real-time
intelligent operations. A smart meter chipset module can perform the appropriate
function by programming within hardware limitations. The smart meter processor
typically performs multiple tasks such as measuring electricity, showing electrical
parameters, reading smart cards, handling data and power, detecting malfunctions,
and interacting with other devices.
To consumers, one of the key benefits of smart metering is to help them conserve
more money. Comprehensive knowledge of usage would allow consumers to make
better decisions on their schedules for energy use. Despite the recent increases in
electricity prices, consumers should by moving their high-load household appliances
to off-peak hours to reduce their energy costs and billing. One feature that would
help the consumers directly is the applications that can offer details on intermittent
13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 283

electricity and overall energy consumption. In this context, the data can be collected
on an hourly basis, daily basis, or weekly basis. Another useful use of energy
disaggregation is to break down the gross power consumption of a household into
individual appliances. In this way, the consumer will be aware of how much energy
every appliance uses to help them control their consumption better.
Smart metering services would allow the distribution system operators (DSOs) to
properly control and sustain its network and more effectively deliver the electricity,
while reducing its running costs. Automatic billing systems will decrease billing-
related concerns and visits to the site. It helps the dealer to remotely read the
consumer meter and to submit correct, timely bills without daily on-site meter
readings. This also offers a two-way link that allows the detection and management
of the remote issue possible. Energy stealing is one of the DSOs’ most critical issues,
which causes a major profit deficit. A variety of methods for the prevention of energy
theft have been suggested by incorporating the AMI into electrical power
systems [40].
The installation of smart meters often poses many obstacles, while adding
important benefits to the community. One of the challenges in deploying modern
metering technology is from an economic standpoint. The AMI’s construction,
implementation, and servicing entail several problems and include budgets of
many billion dollars for deployment and servicing. Therefore, a cost-benefit study
will be a fair starting point for potential smart metering infrastructures to develop.
The advantages can be distinguished between those who are primary and secondary.
Primary benefits are that which will specifically influence the bills of consumers,
while indirect benefits are in terms of efficiency and changes in environmental
standards and would have potential economic impacts. In [41], the current value of
potential profits concerning costs is analyzed while presuming a project life of
13 years, including 3 years of execution and 10 years of service. It is reported that,
by considering both primary and secondary economic advantages, smart grid pro-
vides a favorable economic benefit if they obtained profit to cost ratio is between 1.5
and 2.6. There are also protection and privacy issues. Customers collaborate with the
utilities through the installation of smart meters and two-way communication capa-
bility to control electricity consumption. The details they exchange show consumer
preferences and behaviors, how they use electricity, the number of people in their
homes, and the devices in use that subject them to privacy breaches. In a smart grid
context, a stable and scalable distributed computing network is required. In this
viewpoint, one of the big challenges is to make real-time data accessible from smart
meters to all the stakeholders that use these details to meet those requirements.
The classical way of acquisition of data is time-invariant. The data is collected at a
rate of Nyquist, regardless of their rate of information, which results in a substan-
tially large amount of unwanted data being collected, transmitted, processed, and
analyzed. The classical analog-to-digital converters sample and process the data
based on the principle of Nyquist. The parameters of design of these classical
analog-to-digital converters are therefore chosen for the most unfavorable scenario.
This implies that these analog-to-digital converters are not efficient for arbitrary
signals like appliance usage parameters.
284 S. Mian Qaisar and F. Alsharif

Table 13.7 Comparison with state-of-the-art methods

Accuracy
Study Feature extraction Classification (%)
[9] Time series of power consumption and its derivated GMM 89.8
coefficients
[10] Entropy of time series of power consumption and its HMM 93.6
derivated coefficients
[11] Time series of multiple consumption parameters and their HMM 93.9
derivated coefficients
This EDS delivered time series of real and reactive powers ANN 94.4
study +
Ci, ΔAi, Aimax , and dt imean

The event-driven sensing mechanism is employed in this framework at the stage

of data acquisition to attain real-time data compression. In the following step, the
essential adaptive data processing, segmentation, and extraction techniques are
suggested. Compared with classical counterparts, it confirms a significant compres-
sion and computational efficiency of the method proposed. In this study, EDADCs
were used. They work based on event-driven sampling, and they can adjust their
frequency of sampling according to the variations of the incoming signal. Based on
the results, previously described in the section of results, it is obvious that the
different compression gains for the various studied cases guarantee a significant
diminishing in the computational complexity of and the power consumption of the
suggested solution in contrast with the traditional fixed-rate methods. Additionally,
embedding the EDS into smart meters greatly decreases data collection and transfer
operation relative to conventional approaches. It additionally enhances the effec-
tiveness of the designed strategy over the concurrent ones. To conclude, the pro-
posed solution results in a significant reduction in the complexity and consumption
of power, lowering the data transmission operation compared with the classical
method, and boosting the system performance while securing high automatic appli-
ance identification precision.
Implementing the event-driven signal acquisition and processing technique in
smart meters is still quite a new idea. It is not evident to make a comparison between
the suggested technique and the state-of-the-art techniques as they are based on
classical sampling and processing. Nevertheless, a comparison has been made
among the main previous researches using the same database of appliances [9–
11]. The following table represents the comparison among the classification accu-
racies for all the researches considered.
Table 13.7 shows that the proposed technique ensures greater accuracy compared
to the previous ones [6, 9, 11]. The key benefit of the proposed system over the
counter equivalents is the significant compression gain. Moreover, it aptitudes a
significant efficiency in the processing and power consumption of the proposed
solution over the equivalents. Also, a significant reduction in data storage and
transmission activities is ensured. On the other hand, the disadvantage of this method
13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 285

is that it is still a new method; it is not tested on a variety of appliances or extended

databases of appliances.

13.6 Conclusion

A new approach for automatically profiling the power consumption time series of
major household appliances is presented. It is based on event-driven processing,
extraction of features, and classification. In comparison to conventional methods, the
proposed solution does not need the complex computation of frequency domain-
based feature extraction. The final results showed that the implementation of
EDADC and ASA considerably decreased the count of samples to be processed. A
case of six categories of appliances has been studied. Different classification tech-
niques have been performed for the six appliances. It has been found that a 3.8-fold
reduction in the count of collected samples is achieved over the classical approach.
In comparison to classical techniques, the suggested method affirms a significant
decrease in the complexity of the system. The suggested methodology has been
approved to acquire an average of 94.4% precision of appliance consumption pattern
identification using the ANN classifier. It assures the desire in using the method
proposed in current dynamic load management and detailed electricity bills.
A potential expansion of this work is to explore the current solution when
considering broader appliance categories. The performance of the devised tactic
depends on the selected system parameters such as resolution, reference segment
length, quantization scheme, feature extraction, and classification algorithms. The
development of an automatic mechanism to choose the optimal system parameters
for a targeted application is another future task. Other research angles are the usage
of higher resolution converters, adaptive quantization, and other robust classifiers
such as rotation forest and random forest.

Acknowledgments This project is funded by the Effat University of Jeddah, under the grant
number UC#9/29 April.2020/7.1-22(2)2.

References

1. R. Ullah, Y. Faheem, B.-S. Kim, Energy and congestion-aware routing metric for smart grid
AMI networks in smart city. IEEE Access 5, 13799–13810 (2017)
2. H. Mohammed, S. Tonyali, K. Rabieh, M. Mahmoud, and K. Akkaya, Efficient privacy-
preserving data collection scheme for smart grid AMI networks, (2016), pp. 1–6
3. S. Darby, Smart metering: What potential for householder engagement? Build. Res. Inf. 38(5),
442–457 (2010)
4. Q. Sun et al., A comprehensive review of smart energy meters in intelligent energy networks.
IEEE Internet Things J. 3(4), 464–479 (2015)
286 S. Mian Qaisar and F. Alsharif

5. D. Alahakoon, X. Yu, Smart electricity meter data intelligence for future energy systems: A
survey. IEEE Trans. Ind. Inform. 12(1), 425–436 (2015)
6. Y. Wang, Q. Chen, T. Hong, C. Kang, Review of smart meter data analytics: Applications,
methodologies, and challenges. IEEE Trans. Smart Grid 10(3), 3125–3148 (2018)
7. G. Barnicoat, M. Danson, The ageing population and smart metering: A field study of
householders’ attitudes and behaviours towards energy use in Scotland. Energy Res. Soc. Sci.
9, 107–115 (2015)
8. N. Uribe-Pérez, L. Hernández, D. De la Vega, I. Angulo, State of the art and trends review of
smart metering in electricity grids. Appl. Sci. 6(3), 68 (2016)
9. A. Ridi, C. Gisler, and J. Hennebert, ACS-F2—A new database of appliance consumption
signatures. In 2014 6th International Conference of Soft Computing and Pattern Recognition
(SoCPaR), 2014, pp. 145–150
10. A. Ridi, C. Gisler, and J. Hennebert, Appliance and state recognition using Hidden Markov
Models. In 2014 International Conference on Data Science and Advanced Analytics (DSAA),
2014, pp. 270–276, doi: https://doi.org/10.1109/DSAA.2014.7058084
11. “Processing smart plug signals using machine learning - IEEE Conference Publication.” https://
ieeexplore.ieee.org/document/7122532. Accessed 2 June 2020
12. M. Verhelst, A. Bahai, Where analog meets digital: Analog? To? Information conversion and
beyond. IEEE Solid-State Circuits Mag. 7(3), 67–80 (2015)
13. X. Liu, L. Golab, W. Golab, I.F. Ilyas, S. Jin, Smart meter data analytics: Systems, algorithms,
and benchmarking. ACM Trans. Database Syst. TODS 42(1), 1–39 (2016)
14. N. Mogles et al., How smart do smart meters need to be? Build. Environ. 125, 439–450 (2017).
https://doi.org/10.1016/j.buildenv.2017.09.008
15. P. Koponen et al., Definition of Smart Metering and Applications and Identification of Benefits.
(2008)
16. J. Ponoćko, J.V. Milanović, Forecasting demand flexibility of aggregated residential load using
smart meter data. IEEE Trans. Power Syst. 33(5), 5446–5455 (2018). https://doi.org/10.1109/
TPWRS.2018.2799903
17. Z. Jiang, R. Lin, F. Yang, A hybrid machine learning model for electricity consumer catego-
rization using smart meter data. Energ. Basel 11(9), 2235 (2018). http://dx.doi.org.
effatuniversity.idm.oclc.org/10.3390/en11092235
18. S. Biansoongnern, B. Plungklang, Non-intrusive appliances load monitoring (nilm) for energy
conservation in household with low sampling rate. Procedia Comput. Sci. 86, 172–175 (2016)
19. A. Reinhardt et al., On the accuracy of appliance identification based on distributed load
metering data. In 2012 Sustainable Internet and ICT for Sustainability (SustainIT), (2012),
pp. 1–9
20. M. Weiss, A. Helfenstein, F. Mattern, and T. Staake, Leveraging smart meter data to recognize
home appliances. In 2012 IEEE International Conference on Pervasive Computing and Com-
munications, (2012), pp. 190–197
21. Pattern Recognition and Signal Analysis in Medical Imaging. (Elsevier, 2014)
22. K.-S. Fu, Applications of Pattern Recognition (CRC, New York, 2019)
23. C. Belley, S. Gaboury, B. Bouchard, A. Bouzouane, An efficient and inexpensive method for
activity recognition within a smart home based on load signatures of appliances. Pervasive
Mob. Comput. 12, 58–78 (2014)
24. S.S.S.R. Depuru, L. Wang, V. Devabhaktuni, R.C. Green, High performance computing for
detection of electricity theft. Int. J. Electr. Power Energy Syst. 47, 21–30 (2013)
25. A. Jindal, A. Dua, K. Kaur, M. Singh, N. Kumar, S. Mishra, Decision tree and SVM-based data
analytics for theft detection in smart grid. IEEE Trans. Ind. Inform. 12(3), 1005–1016 (2016)
26. R. Granell, C.J. Axon, D.C. Wallom, Impacts of raw data temporal resolution using selected
clustering methods on residential electricity load profiles. IEEE Trans. Power Syst. 30(6),
3217–3224 (2014)
13 Effective Load Pattern Classification by Processing the Smart Meter Data. . . 287

27. E.D. Varga, S.F. Beretka, C. Noce, G. Sapienza, Robust real-time load profile encoding and
classification framework for efficient power systems operation. IEEE Trans. Power Syst. 30(4),
1897–1904 (2014)
28. R. Al-Otaibi, N. Jin, T. Wilcox, P. Flach, Feature construction and calibration for clustering
daily load curves from smart-meter data. IEEE Trans. Ind. Inform. 12(2), 645–654 (2016)
29. S. Kiran, H. A. Khattak, H. I. Butt, and A. Ahmed, Towards Efficient Energy Monitoring Using
IoT. In 2018 IEEE 21st International Multi-Topic Conference (INMIC), (2018), pp. 1–4
30. Signal Reconstruction. http://pilot.cnxproject.org/content/collection/col10064/latest/module/
m10788/latest. Accessed 05 Dec 2019
31. Physically Based Rendering - 3rd Edition. https://www.elsevier.com/books/physically-based-
rendering/pharr/978-0-12-800645-0. Accessed 05 Dec 2019
32. Y. Hou et al., A 61-nW level-crossing ADC with adaptive sampling for biomedical applica-
tions. IEEE Trans. Circuits Syst. II Express Briefs 66(1), 56–60 (2018)
33. S. M. Qaisar, D. Dallet, S. Benjamin, P. Desprez, and R. Yahiaoui, Power efficient analog to
digital conversion for the Li-ion battery voltage monitoring and measurement. In 2013 IEEE
International Instrumentation and Measurement Technology Conference (I2MTC), (2013),
pp. 1522–1525
34. S. M. Qaisar, R. Yahiaoui, and D. Dominique, A smart power management system monitoring
and measurement approach based on a signal driven data acquisition. In 2015 Saudi Arabia
Smart Grid (SASG), (2015), pp. 1–4
35. S. M. Qaisar, A Computationally Efficient EEG Signals Segmentation and De-noising Based on
an Adaptive Rate Acquisition and Processing. In 2018 IEEE 3rd International Conference on
Signal and Image Processing (ICSIP), (2018), pp. 182–186
36. S.M. Qaisar, L. Fesquet, M. Renaudin, Adaptive rate filtering a computationally efficient signal
processing approach. Signal Process. 94, 620–630 (2014)
37. V.K. Ingle, J.G. Proakis, Digital Signal Processing Using Matlab: A Problem Solving Com-
panion (Cengage Learning, Boston, 2016)
38. B.A. Moser, Similarity recovery from threshold-based sampling under general conditions. IEEE
Trans. Signal Process. 65(17), 4645–4654 (2017)
39. M. Paluszek, S. Thomas, MATLAB Machine Learning (Apress, New York, 2016)
40. S.K. Singh, R. Bose, A. Joshi, Energy theft detection for AMI using principal component
analysis based reconstructed data. IET Cyber-Phys. Syst. Theory Appl. 4(2), 179–185 (2019)
41. S. Pawar and B. Momin, Smart electricity meter data analytics: A brief review. (2017), pp. 1–5
Chapter 14
Prediction of Out-of-Step Condition
for Synchronous Generators Using Decision
Tree Based on the Dynamic Data by
WAMS/PMU

Morteza Abedi, Mohammad Reza Aghamohammadi, Sasan Azad,

Morteza Nazari-Heris, and Somayeh Asadi

Nomenclature

ANN Artificial neural network

AR Autoregressive
CCT Critical clearance time
CDDT Clearance detector decision tree
DT Decision tree
FDDT Fault detector decision tree
FDT Fault duration time
IPDT Instability predictor decision tree
HV High voltage
L-G Line to ground
LLG Line to line to ground
LLL 3 line
MLPANN Multilayer perceptron artificial neural network
NW Number of moving time windows
O.S Out of step
PMU Phasor measurement unit
R5-LLL 3 line with resistive fault 5 ohm

M. Abedi (*) · M. R. Aghamohammadi · S. Azad

Department of Electrical Engineering, Shahid Beheshti University, Tehran, Iran
e-mail: [email protected]
M. Nazari-Heris · S. Asadi
Department of Architectural Engineering, Pennsylvania State University,
University Park, PA, USA
e-mail: [email protected]; [email protected]; [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 289
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_14
290 M. Abedi et al.

R15-LLL 3 line with resistive fault 15 ohm

SNR Signal to noise ratio
SVD Sampling vector data
SVM Support vector machine
TSTM Transient stability time margin
WAMS Wide-area management system

14.1 Introduction

Power systems are designed to be compatible with various disturbances such as

faults, significant and sudden changes in loads, and losing the generators. They
should keep system’s crucial parameters, like the voltage and frequency, within the
permissible range. However, a system is not always secure against all the conse-
quences, and some unexpected events may occur that lead to instability in the rotor
angle, frequency, and voltage, which may cause the instability of the whole system
[1]. By looking at power systems from a historical point of view, it is evident that
researchers have been working on the issue of the rotor angle, and they have tried to
tackle this problem. Also, in recent years with increasing the number of consumers,
the flowing power in the lines has increased; however, due to economic and
environmental situations, the number of transmission lines remained constant.
Therefore, it is evident that an increase in the flowing power in low-voltage lines
will threaten the power system security. The instability of the whole system may
occur due to a little disturbance [2, 3].
Rotor angle stability is an essential issue in power system stability problems that
plays a considerable role in the power system’s planning and operation. The transient
stability of the rotor angle is the system’s capability to maintain synchronism in the
event of a severe disturbance, such as a short circuit appearance in the transmission
line. The system’s response to such disturbances causes substantial changes in the
rotor angle of the generators, flowing power, and other variables. If the angular
distance between the machines of the system remains within certain boundaries, the
system will maintain its synchronism. Nevertheless, if synchronism is lost due to
transient instability, the infinite angular deviation between the system’s machines
will appear 2–3 s after the initial disturbance occurs. Transient stability is based on
the system’s initial state and disturbance severity. The most crucial form of distur-
bance usually studied in the stability research topics contains faults that bring severe
kinetic energy on the system at once, consequently affecting the system [4]. A power
system is a vast interconnected nonlinear system in which failure or instability
occurs due to the toleration of severe stresses to the system. Therefore, early
detection of instability, preventive actions, and on-time control are the power system
operator’s goals. Unless a proper response is made to the occurrence of a significant
disturbance in the system, the generators’ unintentional outage may happen,
followed by the subsequent outage of the system elements, and eventually to a
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 291

complete blackout. Procedures such as the controlled outage of generators, load

shedding, and isolating a section of the network are handled to maintain the system’s
synchronization and stability. Therefore, a fast and accurate method that can identify
the transient steady state from instability is required [5]. The instability of the rotor
angle of synchronous generators in the event of severe disturbance is considered a
significant threat to system security and reliability. In the event of a rotor angle
transient instability, the generator will get out of synchronism, which results in
thermal damage and mechanical on the generator. In order to prevent any damage
to the generator, the generator must be disconnected from the grid immediately after
the detection of the O.S conditions. Therefore, rapid detection and identifying of O.S
conditions of an unstable generator is essential to maintain the safety of the gener-
ator. In order to achieve this goal, it is very necessary to use equipment such as
WAMS and PMU in monitoring the whole system and receiving operating variables
in order to predict the transient instability [6]. Using the PMU and receiving data in
telecommunications and processing it in a control center are one of the tasks that can
improve data transferring and processing speed, hence the speed of the decision-
making process. Transient stability monitoring is shown in Fig.14.1.
In order to maintain the safety of the generator, the O.S relay is installed on all
synchronous generators of power system, which is responsible for detecting the
status of O.S after its occurrence and rapid interruption of the generator from the
power system. Therefore, early generator instability prediction is necessary to
maintain the generator’s safety and prevent possible damage to the generator.
Hence, using intelligent machine learning such as SVM, ANN, and DT compared
to the conventional analytical methods can improve the predictive ability.
The protection issue of the generator against unstable conditions can be investi-
gated in two parts: identifying and predicting the O.S conditions of a generator. In
recent years, in the field of research and industry, the evaluation, identification, and
prediction of the rotor angle’s transient stability have been considered an essential
issue in the power system [8]. In [9], the energy function-based method is used to
determine the stable operating area for the synchronous generator using the data after
a fault in the power system. In [10], using the measurement of rotor angle and the
Prony analysis-Hilbert transform method, a scheme for online monitoring and
prediction O.S status of the power system generator is presented. In [11], the
fuzzy logic theory is utilized to create a protection framework to detect O.S condi-
tions of generators. In [12], the use of a nonlinear classifier based on SVM learning is
proposed to detect the transient instability in a high-dimensional power system. In
[13], a new strategy for improving the interpretability of traditional SVM results,
avoiding the problem of false alarms and missed alarms, and solving the shortcom-
ings of SVM is proposed. In this strategy, two improved and consecutive SVMs are
used, which ensure the stability or instability of the power system in most cases. In
[14], a method for online detection of generator O.S is presented based on measuring
the voltage and current in a line and the energy function. In [15], the transient
stability is monitored and predicted by a relative angle and predefined thresholds of
the relative angle obtained with online data from PMUs and off-line simulations,
respectively; in [16], for fast transient stability evaluation, fuzzy logic classifiers
292 M. Abedi et al.

Application Real-Time Proposed

Servers Data Server Algorithm

Ethernet

Data Concentrator

State Stimator
Control Center

Communication
Links

GPS Satellite

PMU

PMU PMU

Fig. 14.1 Transient stability monitoring by WAMS/PMU [7]

obtained strategic bus information based on WAMS information in real-time condi-

tions, and they brought it to the DT. DT can decide about the O.S status with an
interval of 1–2 s from the time of clearance. In [17], a DT-based approach is
proposed to detect inter-area instability by WAMS/PMU data. The inter-area insta-
bility is detected using WAMS data, including voltage and voltage phase angle
measurements obtained by PMU. In [18], a rapid real-time detection method is
proposed to detect transient instability using WAMS, where trajectory characteristics
and transient energy are combined for improving the fast and accuracy of transient
instability detection. In [19], CCT and TSTM have been utilized as transient stability
identification criteria. Two multilayer perceptron artificial neural networks are used
to estimate CCT and TSTM, and generator stability status assessment has been
suggested. In [20], a model based on radial basis function and MLPANN is
presented in order to detect the transient instability, in which transient stability is
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 293

evaluated by the prediction of rotor angles and angular velocities of generators. In

[21], to estimate the generator transient stability state, a normalized transient stability
margin is presented by a multilayer perceptron neural network. The presented neural
network uses pre-fault operation information for predicting the generator stability
state. In [22], a real-time dynamic security assessment scheme is presented using
PMUs and DTs. DTs perform online security assessment operations based on real-
time measurements. In [7], a method for determining power fluctuations resulting in
power system instability is proposed for real-time monitoring of rotor angle stability,
utilizing WAMS/PMU data and based on Lyapunov exponentials. In [23], a real-
time cluster prediction and identification algorithm have been developed without
depending on the network model. The proposed method classifies generators as
unstable generators by dividing them into coordinated groups behaviorally. The
authors in [24] proposed an SVM classifier to predict the generator stability based on
the generator’s dynamic trajectories such as voltage, rotor angle, electromagnetic
power, speed, and other electrical parameters. In [25], O.S prediction logic is
presented based on the AR model. The proposed method is based on the time
domain, using the mth order AR model and the voltage phase difference between
the buses as input variables of the autoregressive model. The power fluctuations after
clearance of a severe fault are modeled to predict out of step in the proposed method.
In [5], a hybrid classifier is presented to predict the transient instability of the power
system following a disturbance. An index is made as input data for classification by
the relative rotor angles of the generators in the center of inertia scheme. In [2], the
authors proposed a method to predict transient stability after the fault occurrence
utilizing a post-fault voltage measurement at the generator bus of a multi-machine
power network and the SVM classifier. The proposed method always initiates the
fault event that causes the transient stability state’s evaluation process after the
voltage drop. In [3], an angular instability predictor is presented based on the bus
voltage’s measured trajectories after the occurrence of a severe disturbance. In [26],
an intelligent wide-area synchrophasor system with ANN is introduced for real-time
prediction of transient instability. In [27], a generator rotor angle instability predictor
is proposed by a DT-based method to predict the O.S of synchronous generators. The
proposed method can distinguish between stability state and O.S based on input
characteristics measured before the moment of out of step. In [28], using an adaptive
ANN, an instability rotor angle predictor is proposed. In [29], a new algorithm is
proposed to predict the generator’s O.S status using the measured information of
rotor acceleration and speed by PMU, based on the change rate of speed and
acceleration. In all the proposed methods, assuming that the moment of occurrence
of the fault and clearance of the fault has already been detected, they rely mainly on
the information after clearance of the fault. Some methods predict instability only for
a group of generators instead of a single generator. The performance of the proposed
methods due to noise, bad data, or miss data, changes in the grid configuration, and
asymmetric faults have not been investigated and remain unanswered.
In this chapter, to quickly predict the O.S state of an unstable generator, an
intelligent O.S prediction approach is proposed that includes a FDDT, a CDDT,
and an IPDT. In this three-step scheme, using local online operating data obtained by
294 M. Abedi et al.

PMU, which is installed on the generator terminal, in the first step, FDDT detects the
fault occurrence; in the second step, CDDT detects the fault clearance; and finally, in
the third step, IPDT predicts O.S. The input vector data for all DTs is a moving time
window containing several sampled vectors data measured in consecutive time steps.
Every SVD may contain operating variables such as voltage magnitude and phase
angle, active and reactive power, current magnitude and phase angle, generator rotor
speed, angle, and acceleration. In this method, first, the time of occurrence and
clearance is detected based on the variables of the generator’s electrical operation,
and then using IPDT and electromechanical operating variables after clearance, the
stability of the generator is predicted.

14.2 The Framework of the Proposed Method

In this section, the initial framework of the presented algorithm for predicting
transient stability is presented. The proposed scheme’s main structure consists of
three learning machines, which are used to classify the data. The machines apply the
measured data before, during, and after the disturbance as FDDT, CDDT, and IPDT
inputs, respectively, to precisely predict the system’s transient stability. The pro-
posed scheme’s primary attitude is based on the fact that the instability of a generator
could be distinguished from its transient behavior after-fault clearance condition. It
is worth noting based on the generator’s dynamic attributes; its stability status is
mostly visible in the performance after the fault clearance. Therefore, the behavior of
the generator after the fault clearance is categorized between stability and instability
status. In other words, the dynamic behavior of generator after-fault clearance
contains precious information about its stability nature that can be used to predict
the instability state.
Indeed, it should be noted that the use of a learning machine reduces the time in
determining the status of the system and the system operator has more chances to
achieve control duties. The studies performed to determine the transient stability of
the power system are based on the basic concept of fault CCT such that for the
duration of the fault in the system more than CCT, the generated scenario leads to
instability. Also, the generated scenarios are not concentrated on exact faults, and
they contain faults at different places and different loading levels. Significant and
evident signs in determining the state of power system instability for a synchronous
generator are the variables of rotor angle and speed.
The behavior of the generator rotor angle for different conditions is shown in
Fig. 14.2. According to Fig. 14.2, the typical trajectory of generator rotor angle is
shown for different fault clearance times according to the fault CCT ¼ 317 ms. In
Fig. 14.2, the rotor angle variations are investigated for four different scenarios
containing two unstable scenarios (tclear3 ¼ 320 ms > CCT, tclear4 ¼ 325 ms > CCT)
and two stable scenarios (tclear1 ¼ 310 ms < CCT, tclear2 ¼ 315 ms < CCT). As it is
observable, the change in generator rotor angle is relatively similar and slightly
different for stable and unstable scenarios until the fault is cleared. However, after
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 295

200
Stable_tclear=310ms Fault clearance
Stable_tclear=315ms
150 Unstable_tclear=320ms
Rotor angle(deg)

Unstable_tclear=325ms
100
Fault (CCT=317ms)
50

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time(Sec)

Fig. 14.2 Generator rotor angle fluctuation for four different scenarios

the clearance of the fault, the generator’s rotor angle starts to increase the distance
from the stable scenario and makes more distinction. The distinction made between
stable and unstable scenarios after clearance can be appropriate in discriminating
stable from unstable scenarios. It is worth noting that earlier studies and research
assume that the occurrence of a fault and its clearance is known by generator and
predictor of instability using generator information after the fault clearance. How-
ever, in the actual operation status, the fault occurrence and clearance affecting the
generator’s stability cannot be detected by the generator’s local protection.
Fault occurrence and clearance outside the generator protection zone cannot
activate the intelligent O.S relay, which operates according to post-fault information.
Therefore, for an intelligent O.S prediction relay that operates based on data, which
sampled after-fault clearance, detecting the fault occurrence and clearance using
local generator information are necessary. To this end, the motivation of proposing
an intelligent three-step scheme consisting of three DTs is to address all issues
related to fault occurrence and clearance detection and instability prediction in the
form of an O.S protection strategy.
The conceptual structure of the introduced intelligent O.S relay is shown in
Fig. 14.2. Three DTs operate based on actual local information obtained from the
PMU installed in the generator terminal in the proposed structure.
In the first step, the FDDT is used for detecting faults occurrence in the generator.
Suppose the FDDT detects any faults or abnormal conditions similar to the fault that
could be disadvantageous to the generator’s stability. In that case, it sends output
signal 1 to the CDDT to activate it to verify that the fault has been cleared or still
exists. When the protection relays clear the fault, the CDDT detects this circum-
stance and sending output signal 1 to IPDT to activate it.
As shown in Fig. 14.3, during actual operation, a time window consisting of n
data is provided to the FDDT at any given time by the fluctuating behavior of the
generator tasted by the PMU. If the FDDT detects any faults or abnormalities, it
sends output signal 1 to activate the CDDT. The next time window of the sampled
data goes into the CDDT; otherwise, the time window containing the sampled data
will still be available to the FDDT until it detects a fault or abnormal conditions.
296 M. Abedi et al.

V, I, P , Q , δ, ω
G2

PMU
Gi
G3

G4

Fault No Fault
Detector
0
FDDT
1
Clear No Clear
Detector
0
CDDT
1
Instability Stable
Predictor
IPDT

Out of step

Fig. 14.3 The proposed 3step scheme for predicting O.S

Instantly after detecting any faults, the sampling window information is trans-
ferred to the CDDT. If the fault clearance detects by CDDT, the next time window of
the sampling data will be transferred to the IPDT; otherwise, the time window
information will be provided to the CDDT until the CDDT detects a fault clearance.
All three DTs, which are introduced above, work in a design called an intelligent O.S
prediction relay scheme that works for a generator.

14.3 The Proposed Method

The basis of the presented algorithm is that the instability of a generator is reflected
in the electromechanical behavior after clearance of its fault, including the generator
rotor angle, speed, and acceleration. Following the occurrence of a fault, the rotor
angle and speed of a generator are modeled during the fault and after the fault is
cleared by Eq. (14.1), which is called oscillation equations:

dω ω0 ω
¼ ðP  Pe Þ ¼ 0 Pacc
dt 2H m 2H ð14:1Þ
dδ
¼ ω  ω0
dt
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 297

where ω, δ, and H are the speed and angle of the rotor and the constant inertia of the
generator before the fault occurs and is the accelerator power. By merging the swing
equations during the fault and after the fault clearance, the generator’s rotor angle
and speed after the fault clearance can be estimated by Eq. (14.2):

ω0 f ω
δpos ¼ δ0 þ P  t 2 þ 0 Ppos  t 2
4H acc f 4H acc p ð14:2Þ
ω ω
ωpos ¼ ω0 þ 0 Pfacc  t f þ 0 Ppos t
2H 2H acc p

which δ0and ω0 are the rotor angle and speed before the fault, respectively. δpos and
ωpos are the rotor angle and speed after the fault clearance, respectively. Pfacc and Ppos
acc
are the average accelerator power during fault and after the fault clearance, respec-
tively. tf and tpare the duration of the fault and the moment after the fault clearance,
respectively, which indicates the severity of the disturbance. In Eq. (14.2), the fault
duration tf and the average accelerator power Pfacc indicate the severity of the
disturbance, which affects the stability of the generator. Equation (14.2) shows
that the severity of the fault is directly affected by changes in the rotor angle and
speed of the generator after-fault clearance. In fact, in power system stability studies,
the trajectory trend of changing the rotor angle after-fault clearance is used as a
criterion for evaluating the stability of the generator. Therefore, the generator’s
behavioral information after clearance of the fault is prosperous and useful to
distinguish the generator’s stable state from the unstable state. In other words, the
information after-fault clearance in the generator behavior has properties that can be
used to predict the instability of the generator. Thus, intelligent methods play a
prominent role in extracting the characteristics of such precious information. The DT
is an intelligent, powerful tool for extracting features from complex data that can be
used to predict the synchronous generator’s instability in power systems.

14.3.1 Decision Tree

In this study, the DT is used to evaluate the system’s transient stability. Using the DT
learning machine is simple, fast, and appropriate for classifying complex issues.
These intelligent machines are suitable for problems with a low output. One of the
most prominent features of the DT compared to other learning machines is its
decision-making process algorithm, which is well-defined and transparent. On the
other hand, the decision-making process in other learning methods is not transparent
and is like a black box.
The advantages of the DT over other learning machines, such as the neural
network, are:
• The DT explains its prediction in the form of a set of rules, while in neural
networks, the only prediction is expressed and how it is hidden in the network.
298 M. Abedi et al.

• Also, in the DT, unlike neural networks, the data do not need to be numerical. It
can support numeric variables and can work with categorical variables such as
type and gender.
• Also, DT can determine outstanding operating variables concerning output. In
other words, by placing a prominent variable in the central node, the significance
of this variable in classification can be estimated.
In this chapter, for predicting the generator’s rotor angle instability based on fault
clearance data, the DT learning machine is used as an O.S predictor.
The introduced intelligent predictor scheme includes three consecutive DTs that
use local online data from the generator oscillation produced by the PMU. The
presented predictor scheme is a combination program of the DT learning machine as
the intelligent predictor and the PMU as the local dynamic data generator. In this
method, the task of fault occurrence and clearance detection and instability predic-
tion is performed by three consecutive DTs. IPDT has been used for performing the
main task of predicting O.S, for which the entrance data is provided by the PMU and
after-fault clearance. Therefore, there needs to be a difference between the faulty
data and after the fault clearance. For this purpose, in the previous two steps, the two
DTs are used as fault occurrence detector and the fault clearance detector to
distinguish the data after the fault clearance and during the fault.

14.3.2 Fault Detection by FDDT

The FDDT is the first step in predicting the presented O.S method. The main task of
FDDT is detecting faults or any abnormal conditions from the generator’s point of
view. FDDT uses electrical operating variables provided online by the PMU. Input
data of FDDT is a moving time window including six samples of a generator’s
electrical operating variables such as current and voltage magnitude, and their phase
angle, active and reactive power. The FDDT ability to detect a fault is mainly due to
the sudden change that occurs between the electrical operating variable values before
the fault occurrence and during the fault (e.g., I, V, θI, θV, P, Q). SVD at time
t includes some sampled operating variables (e.g., I, V, θI, θV, P, Q). In Eq. (14.3) the
sampled operating electrical variables at time t are shown:

SVDðt Þ ¼ ½I ðt Þ, V ðt Þ, θI ðt Þ, θV ðt Þ, Pðt Þ, Qðt Þ ð14:3Þ

Figure 14.4a represents a time window structure, including six consecutive SVDs
created with a sampling interval of ΔT. A time window is defined as a precious time
window for detecting the occurrence of fault when it contains a combination of
pre-fault and during-fault sampled operating data. Whenever all the sampled oper-
ating data of a time window is completely about before the fault or entirely during
the fault, that time window is defined as a weak time window with no information
about the fault. Thus, the FDDT cannot detect the fault occurrence.
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 299

Time Sample 1 Sample 2 Sample 3 Sample 4 Sample 5 Sample 6

Window
SVD (Tf )
TWf 1 SVD (Tf -5Δt) SVD (Tf -4Δt) SVD (Tf -3Δt) SVD (Tf -2Δt) SVD (Tf -Δt)
(fault )

SVD (Tf )
TWf 2 SVD (Tf -4Δt) SVD (Tf -3Δt) SVD (Tf -2Δt) SVD (Tf -Δt) SVD (Tf +Δt)
(fault )

SVD (Tf )
TWf 3 SVD (Tf -3Δt) SVD (Tf -2Δt) SVD (Tf -Δt) SVD (Tf +Δt) SVD (Tf +2Δt)
(fault )

SVD (Tf )
TWf 4 SVD (Tf -2Δt) SVD (Tf -Δt) SVD (Tf +Δt) SVD (Tf +2Δt) SVD (Tf +3Δt)
(fault )

SVD (Tf )
TWf 5 SVD (Tf -Δt) SVD (Tf +Δt) SVD (Tf +2Δt) SVD (Tf +3Δt) SVD (Tf +4Δt)
(fault )

SVD (Tf )
TWf 6 SVD (Tf +Δt) SVD (Tf +2Δt) SVD (Tf +3Δt) SVD (Tf +4Δt) SVD (Tf +5Δt)
(fault )

a. six precious time windows including pre-fault and during fault occurrence data for FDDT

Time Sample 1 Sample 2 Sample 3 Sample 4 Sample 5 Sample 6

Window
SVD (Tc)
TWC1 SVD (Tc-5Δt) SVD (Tc-4Δt) SVD (Tc-3Δt) SVD (Tc-2Δt) SVD (Tc-Δt)
(clear)

SVD (Tc)
TWC2 SVD (Tc-4Δt) SVD (Tc-3Δt) SVD (Tc-2Δt) SVD (Tc-Δt) SVD (Tc+Δt)
(clear)

SVD (Tc) SVD

TWC3 SVD (Tc-3Δt) SVD (Tc-2Δt) SVD (Tc-Δt) SVD (Tc+Δt)
(clear) (Tc+2Δt)

SVD (Tc) SVD SVD

TWC4 SVD (Tc-2Δt) SVD (Tc-Δt) SVD (Tc+Δt)
(clear) (Tc+2Δt) (Tc+3Δt)

SVD (Tc) SVD SVD SVD

TWC5 SVD (Tc-Δt) SVD (Tc+Δt)
(clear) (Tc+2Δt) (Tc+3Δt) (Tc+4Δt)

SVD (Tc) SVD SVD SVD SVD

TWC6 SVD (Tc+Δt)
(clear) (Tc+2Δt) (Tc+3Δt) (Tc+4Δt) (Tc+5Δt)

b. 6 precious time windows including during fault and post fault clearance data for CDDT

Time
Window Sample 1 Sample 2 Sample 3 Sample 4 Sample 5 Sample 6

SVD (Tc) SVD SVD SVD SVD SVD

SVD (Tc+Δt)
(clear) (Tc+2Δt) (Tc+3Δt) (Tc+4Δt) (Tc+5Δt) (Tc+6Δt)

C. Time window including post fault clearance data for IPDT

Fig. 14.4 The moving time windows structure for DTs. (a) Six precious time windows including
pre-fault and during-fault occurrence data for FDDT. (b) Six precious time windows including
during-fault and post-fault clearance data for CDDT. (c) Time window including post-fault clear-
ance data for IPDT
300 M. Abedi et al.

During the generator’s operation, the moving time windows are continuously
updated with new sampled operating variables. Thus, due to the occurrence of a
fault, six consecutive time windows have precious information about the moment of
the fault occurrence and the moment before the fault occurrence.
Figure 14.4a indicates the structure of six consecutive precious time windows,
each including six sampled operating data, in which the yellow samples belong to
the during-fault data, while the blue samples belong to pre-fault data. According to
Fig. 14.4a, SVD(Tf) (red) shows the operating data sampled at the moment of fault
occurrence, SVD(Tf-kΔT ) belongs to k sample data before fault occurrence, and
SVD(Tf + kΔT ) refers to the k sample data after occurrence of the fault.
As can be seen, after the fault occurrence time, the number of SVDs containing
the fault data in the time window increases. Adopting six SVDs per time window
enables FDDT to view precious information in six consecutive time windows that
increase FDDT detection reliability. Equation (14.4) displays the conceptual perfor-
mance of FDDT for the precious and weak time window:
8

< 0 ¼ f
SVDpre
>

: weak data
FDDT : 1 ¼ f SVDpre , SVDf : precious data ð14:4Þ
>
:
0 ¼ f ðSVDf Þ : weak data

After detection of fault occurrence by FDDT, it sends output signal 1 to the

CDDT to activate it.

14.3.3 Fault Clearance Detection by CDDT

After the FDDT detects the fault, the CDDT is activated for detecting the fault
clearance. The CDDT uses electrical operating variables provided online by the
PMU. Input information for CDDT is a moving time window including six samples
of generator electrical operating variables such as current and voltage magnitude,
and their phase angle, active and reactive power. The CDDT ability to detect a fault’s
clearance is mainly due to the sudden change that occurs between the electrical
operating variables values during fault and after the fault clearance (e.g., I, V, θI, θV,
P, Q).
Figure 14.4b demonstrates the structure of a time window that includes six
consecutive SVDs produced with the same time step ΔT. A time window is a
precious time window for detecting the clearance of fault when it contains a
combination of during-fault and after-fault clearance sampled operational data.
The window is a weak time window when all sampled operating data are entirely
gathered during the fault or after the fault clearance. This time window does not have
any clearance information, and CDDT cannot detect clearance. During the generator
operating, the moving time windows are continuously updated with newly sampled
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 301

operating data. Therefore, due to the fault clearance, six consecutive time windows
have precious information.
Figure 14.4b demonstrates the structure of six consecutive precious time win-
dows, each including six sampled operating data. In these time windows, the yellow
samples belong to the during-fault data, while the blue samples belong to the data
which is sampled after-fault clearance.
According to Fig. 14.4b, SVD(TC) (green) shows the sampled data at the time of
fault clearance, SVD(TC-kΔT ) belongs to k sample data before fault clearance, and
SVD(TC + kΔT ) refers to the k sample data after clearance the fault. According to
the time windows, it is observed that after the fault clearance time, the number of
SVDs with data that are sampled after-fault clearance increases.
Adopting six SVDs per time window makes CDDT capable of viewing precious
data through six consecutive time windows. This design enhances the reliability of
the CDDT detection performance. Equation (14.5) displays the conceptual perfor-
mance of CDDT concerning the precious and weak time window:
8
< 0 ¼ f ð
SVDf Þ
>

: weak data
CDDT : 1 ¼ f SVD f , SVD pos : precious data ð14:5Þ
>
:

0 ¼ f SVDpos : weak data

After detecting fault clearance by CDDT, it sends output signal 1 to the IPDT to
activate it.

14.3.4 Prediction of Instability by IPDT

The final step of the proposed intelligent relay is to predict the O.S status of the
generator by IPDT that is the presented scheme’s major objective. The IPDT works
based on the generator’s online data. IPDT input vector data is a moving time
window containing six sampled operating data, including generator oscillation and
electromechanical data such as rotor angle, speed, and acceleration.
Figure 14.4c demonstrates the structure of a time window, including six consec-
utive SVDs produced with the same time step ΔT. The time window is continuously
created and submitted to the IPDT. However, IPDT is only activated by CDDT when
the fault is cleared.
If the IPDT predicts the generator’s potential to O.S, it will send a “1” signal on
the output as a warning of an unstable status of the generator. It is necessary to
mention; if the IPDT fails for predicting the generator stability by the first time
window, the next time window will continue to enter the IPDT until the prediction of
the generator stability status is made even with a delay.
The proposed intelligent predictor relay is considered as a backup for the main O.
S relay. Equation (14.6) displays the conceptual performance of IPDT concerning
data that are sampled after-fault clearance:
302 M. Abedi et al.

(

0 ¼ f SVDpos : Stable
IPDT :
 ð14:6Þ
1 ¼ f SVDpos : Unstable

The input data of DTs which consists of six consecutive time windows provides
strong performance for DTs against any disturbances may occur in samplings such
as miss data or bad data.

14.3.5 Nature of Operating Variables Used by DT

Usually, in order to train DTs, various operating variables of generators can be used,
such as active and reactive output power, voltage magnitude and phase angle, rotor
angle, and speed. In the moment of fault occurrence and fault clearance, variations in
operating variables such as active power, voltage, and current are more precise
compared to variations in operating mechanical variables such as speed and rotor
angles.
Therefore, electrical operating variables that change suddenly at the moment of
occurrence and clearance of the fault are the most appropriate and precious variables
for FDDT and CDDT input data. The electromechanical variables such as rotor
angle, speed, and acceleration, which indicate the generator rotor angle’s stability,
are the most appropriate and precious variables for using as input data of IPDT.
Since the magnitude of changes in the electrical operating variables in the time
step ΔT is better than the magnitude of the variables, changes in the specified time
step’s electrical variables are used as the input vectors to the DTs.

14.4 Simulation

To evaluate the presented model’s effectiveness to predict the O.S of synchronous

generators following a fault occurrence, the proposed protection scheme has been
implemented on the IEEE 39-bus power system. The presented model for generator
#7 which is connected to bus #36 will be investigated.
In order to select the appropriate operating variables of the power system for
training FDDT, CDDT, and IPDT, all electrical and electromechanical operational
variables such as active and reactive power, voltage phasor, current phasor, gener-
ator rotor angle, speed, and acceleration are considered and have thoroughly been
reviewed. Finally, an appropriate number of operating variables are selected based
on DTs performance and detection/prediction. In this research, a set of specific
operating variables is introduced for each DT.
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 303

14.4.1 Producing Training Scenarios of DT

In this subsection, training scenarios from generator G7, located in bus #36 in the
IEEE New England 39-bus power system, have been produced. Since the proposed
O.S scheme is built to predict a particular generator’s stability status, the diversity of
events in which the generator experience unstable condition is not very wide. That is,
the generator may become unstable in many scenarios. However, the oscillator
behavior pattern of the generator is relatively limited for all of the instability
scenarios. This means that by adequately selecting different fault scenarios, the
main feature of the generator oscillation behaviors that lead to instability can be
obtained. Therefore, to prepare appropriate training scenarios, six different load
levels (4000 MW, 5000 MW, 6000 MW, 7500 MW, 9000 MW, 10500 MW) have
been selected.
Concerning transient stability, the criterion should be used to determine stability
in order to identify the status of the power system in terms of transient stability and
instability. The proposed method criterion is based on CCT to determine a stable and
unstable scenario. It is assumed that a fault occurs in the network and is cleared after
a certain period (relay operation time). This fault has its own CCT attribute that
indicates how long the network has time to clear the fault. CCT determines the
significance of the fault. The smaller the CCT related to a fault, the fault has more
significance, and the less time available to the network to clear it. As a result of this
fault, network instability is more likely to occur. Therefore, for the network to
remain stable after-fault clearance, the clearance time (relay operation) must be
less than the CCT.
In order to produce various stable/unstable scenarios for training three DTs in
each operating load level, the symmetrical LLL-faults with different fault clearance
times, according to generator G7 CCT (i.e., 0.1*CCT, 0.5*CCT, 0.8*CCT,
0.95*CCT, 1.2*CCT, 1.4*CCT), is applied to the power system under study.
Scenarios with less fault clearance time than CCT are defined as stable scenarios,
while scenarios with fault clearance duration higher than CCT are defined as
unstable scenarios. These faults are applied to buses No. 21, No. 22, No. 23, and
24 and lines 16–24, 21–22, 22–23, 21–16, 16–24, and 23–24 around generator G7.
This method creates a total of 324 disturbance scenarios.
The transient stability simulation is performed using DIgSILENT Power Fac-
tory® software, in which the generators are modeled in full detail. The stability
status of generator G7 is obtained for all scenarios. Among the designed scenarios,
220 scenarios, 70 stable scenarios, and 150 unstable scenarios are acceptable.
The generator operating variables transfer rate, which are provided by PMU for is
10 ms. Therefore, the sampling data’s sampling step is updated in the moving time
windows which is considered equal to 10 ms (Δt ¼ 10 ms). For each kind of stable/
unstable scenario, according to the FDT, the NW is determined by the speed of
movement of a sample based on Eq. (14.7):
304 M. Abedi et al.

Table 14.1 Number of time FDDT and CDDT IPDT

windows generated for train-
Time window Training Test Training Test
ing/testing of DTs
Stable state 2510 1124 503 209
Unstable state 5211 2108 1034 421
Total 7721 3232 1537 630

FDT
NW ¼ þ 10 ð14:7Þ
0:01 |{z}
|ffl{zffl} II
I

The first part of Eq. (14.7) defined by (I) is the number of time windows that
belongs to the fault duration time used for FDDT and CDDT training/testing
process. In the second part of Eq. (14.7) defined by (II), “10” is the number of
time windows after the fault is cleared for IPDT training/testing process.
According to Table 14.1, by applying Eq. (14.7) to 220 scenarios, a total of
13,390 time windows for stable/unstable training and testing are generated, of which
70% and 30% of the time windows for training and testing DTs are used,
respectively.

14.4.2 Time Windows Data

In preparing training scenarios, the number of samples that should be gathered after
clearance of the fault as the IPDT input time window to predict O.S is significant.
The duration time between clearance of the fault and O.S depends on the generator,
controllers, and location in the power system. In most severe instability conditions,
this duration time is short. There is a minimum duration for each generator, which is
related to the most severe instability conditions for the generator. The IPDT sam-
pling time window should not exceed this minimum duration time because the time
window may contain after O.S data, which is practically worthless data. Therefore,
for each generator, this minimum duration time must be specified first. Then, the
sampling time window is determined based on it.

14.4.3 FDDT Training

Table 14.2 presents the FDDT performance of training concerning different input
operating variables that provide the maximum accuracy for the detection of the fault
occurrence.
As seen in Table 14.2, the set with the highest accuracy is selected as the FDDT
input variables, which include all three variables of Δθv, ΔP, and ΔQ. The FDDT
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 305

Table 14.2 FDDT performance results for different input variables

Variables Train (7721) Test (3232)
Acc. Error detection Acc. Error detection
% % No. of % % No. of
misclassification misclassification
ΔP, Δθv, 99.71 0.29 19 99.75 0.25 8
ΔQ

input vector consisting of three selected variables along with samples taken at
specified time steps is shown in Eq. (14.8):

Sampling vector data : SVDðt Þ ¼ ½ΔθV ðt Þ, ΔPðt Þ, ΔQðt Þ

ΔθV ðt Þ ¼ ½ΔθV ðt  kΔt Þ, k ¼ 1, . . . , 5
ð14:8Þ
ΔPðt Þ ¼ ½ΔPðt  kΔt Þ, k ¼ 1, . . . , 5
ΔQðt Þ ¼ ½ΔQðt  kΔt Þ, k ¼ 1, . . . , 5

The structure of trained FDDT is shown in Fig. 14.5a. The numbers in parenthe-
ses refer to the sample number of Fig. 14.4a.

14.4.4 CDDT Training

Table 14.3 presents the CDDT performance of training concerning different input
operating variables, which provide maximum accuracy for detecting the fault
clearance.
Different combinations of variables are examined. As observed in Table 14.3, the
set has been selected as the CDDT input variables with the highest accuracy, which
includes all three variables of ΔθI, ΔP, ΔQ. The CDDT input vector consisting of
three selected variables, along with samples taken at specified time intervals, is
presented in Eq. (14.9):

Sampling vector data : SVDðt Þ ¼ ½ΔθI ðt Þ, ΔPðt Þ, ΔV ðt Þ

ΔθI ðt Þ ¼ ½ΔθI ðt  kΔt Þ, k ¼ 1, . . . , 5
ð14:9Þ
ΔPðt Þ ¼ ½ΔPðt  kΔt Þ, k ¼ 1, . . . , 5
ΔV ðt Þ ¼ ½ΔV ðt  kΔt Þ, k ¼ 1, . . . , 5

The structure of trained CDDT is shown in Fig. 14.5b. The numbers in paren-
theses refer to the sample number of Fig. 14.4b.
306 M. Abedi et al.

ΔQV(5) > 1.065

No yes

ΔP(2) > 0.232

No Fault
No yes

ΔQ V(2) > 1.095 No Fault

No yes

ΔQ(3) > 30.53

Fault Detected
No yes

No Fault Fault Detected

a. FDDT structure

ΔV(5) > 0.0

No yes

ΔP(5) > 3.321 ΔQ I(2) > 0.68

No yes No yes

Clear Detected ΔV(3) > 0.0

No Clear ΔV(3) > -0.011
No yes
No yes

Clear Detected No Clear

No Clear Clear Detected

b. CDDT structure

Δα (5) > 1.065

No yes

Δδ (5) > 1.065 Δδ (5) > 1.065

No yes No yes

Δω (3) > 1.065 Δδ (5) > 1.065 Stable Unstable

No yes No yes

Stable Δω (2) > 1.065 Stable Unstable

No yes

Stable Unstable

c. IPDT structure

Fig. 14.5 DTs structure (a) FDDT structure, (b) CDDT structure, (c) IPDT structure
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 307

Table 14.3 CDDT performance results for different input variables

Variables Train (7828) Test (3995)
Acc. Error detection Acc. Error detection
% % No. of % % No. of
misclassification misclassification
ΔP, ΔθI, 99.71 0.29 22 98.95 1.05 34
ΔQ

Table 14.4 IPDT performance results for different input variables

Variables Training (1537) Testing (630)
Acc. Error detection Acc. Error detection
% % No. of % % No. of
misclassification misclassification
Δα, Δw, 98.73 1.27 20 89.96 10.03 63
Δδ

14.4.5 IPDT Training

This is the last step of the proposed method, which is responsible for predicting the
instability. The stability status of the generator is predicted based on data that are
sampled after-fault clearance. For this purpose, changes in rotor angle, speed, and
acceleration magnitude are used as input variables. The phase angle of the genera-
tor’s terminal voltage is estimated as the generator rotor angle [30]. Therefore, the
rotor speed and rotor acceleration are calculated by the generator rotor angle using
Eqs. (14.10) and (14.11):

δ ð t þ 1Þ  δ ð t Þ
ωð t þ 1 Þ ¼ ð14:10Þ
Δt
ωð t þ 2 Þ  ωð t þ 1 Þ
αðt þ 1Þ ¼ ð14:11Þ
Δt

Due to the angle, speed, and acceleration of the rotor changes as input variables, a
different combination of these variables is used as the IPDT sampled data vector.
The performance of IPDT training concerning different combinations of variables
Δα, Δw, and Δδ are examined.
As shown in Table 14.4, the set has been selected as the IPDT input variables with
the highest accuracy, including all three variables of Δα, Δw, and Δδ. The IPDT
input vector consisting of three selected variables along with samples taken at
specified time steps is shown in Eq. (14.12).
308 M. Abedi et al.

Fig. 14.6 Operating variables of FDDT. (a) FDT of ΔθV. (b) FDT of ΔP, (c) FDT of ΔQ

Sampling vector data : SVDðt Þ ¼ ½Δδðt Þ, Δωðt Þ, ΔV ðt Þ

Δδðt Þ ¼ ½Δδðt  kΔt Þ, k ¼ 1, . . . , 5
ð14:12Þ
Δωðt Þ ¼ ½Δωðt  kΔt Þ, k ¼ 1, . . . , 5
Δαðt Þ ¼ ½Δαðt  kΔt Þ, k ¼ 1, . . . , 5

The structure of trained IPDT is shown in Fig. 14.5c. The numbers in parentheses
refer to the sample number of Fig. 14.4c. According to Fig. 14.5, the threshold
values of each DT are determined during the training process.
In Fig. 14.6a–c, a three-phase short circuit fault at t ¼ 1 s is applied to the
generator’s HV bus under study. As can be seen, the operating variables used in
FDDT have significant changes at the time of the fault.
In Fig. 14.7a–c, a three-phase short circuit fault at t ¼ 1 s is applied to the HV bus
of the generator under study and is cleared at t ¼ 1.27 s. As can be seen, the operating
variables used in CDDT have significant changes at the time of clearance.
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 309

Fig. 14.7 Operating variables of CDDT. (a) CDT of ΔV, (b) CDT of ΔθI, (c) CDT of ΔP

14.4.6 Over-/Under-Fitting Evaluation of IPDT

Since DTs’ performance is threatened by over-/under-fitting in the training proce-

dure, the number of training scenarios is gradually decreased/increased. The train-
ing/test performance of IPDT is also evaluated. According to Table 14.5, the number
of training/testing scenarios uses in this study is optimal and provides the best
training/testing accomplishment for IPDT. According to the optimal number of
scenarios, any increase/decrease in the number of scenarios causes over-/under-
fitting in the training process, which is a factor to reduce the accuracy of training.
It should be noted that since the intelligent O.S predictor relay is designed
individually for each generator with respect to the generator behavior, so the training
process includes the number of training scenario and training time independent of
the size of the power system.
310 M. Abedi et al.

Table 14.5 IPDT performance concerning over-/under-fitting

Testing
Training result
No. of training result
Training cases

scenario

Misclassifi-

Misclassifi-
Acc. %

Acc. %
No. of

No. of
cation

cation
1 4985 96.3 184 87.9 110
2 7846 95.8 329 88.7 93
3 10320 97.35 273 88.81 96
4 13390 98.73 170 89.96 63
5 14890 97.21 564 81.02 162
6 16281 94.05 968 85.73 138

Table 14.6 Comparison of DTs performance with ANN and SVM

DT SVM MLPANN
performance performance performance
Acc.% Train Test Train Test Train Test
Fault occurrence detection 99.71 99.75 99.34 98.96 99.08 99.54
Fault clearance detection 99.71 99.75 99.34 98.83 99.19 99.1
Instability occurrence prediction 98.73 89.96 94.93 85.72 96.65 85.9

14.4.7 Comparison of DT Performance with SVM and ANN

The proposed intelligent relay performance is compared with other learning

machines. Two learning machines consist of MLPANN and SVM trained with the
same data and then compared with the result of the DT-based algorithm. In this
comparison, ANN with ten neurons in the hidden layer and a Levenberg-Marquardt
algorithm is used. The other intelligent method is SVM with a standard deviation of
0.1 and a regulating parameter of 10. The accuracy of training and testing for the
presented scheme compared to other intelligent schemes is shown in Table 14.6.
According to Table 14.6, the ANN and SVM performances are almost the same,
while the proposed intelligent predictor relay based on DT performs better.

14.5 Validation of the Proposed Scheme Performance

To validate the proposed intelligent relay scheme, its performance is examined for
seven stable/unstable scenarios that were not used in training. The characteristics of
the seven scenarios are shown in Table 14.7. In addition to the 8 developed scenar-
ios, 57 other stable and unstable scenarios have been developed to evaluate IPDT
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 311

Table 14.7 Characteristics of scenarios for validation of DTs performance

Location of fault
Load level Fault Line CCT FDT
Scenario (MW) Stability status type Bus/line % (ms) (ms)
1 6832 Unstable LLL Lines 50 352 400
23–24
2 5040 Stable LLL Bus 24 – 495 300
3 6350 Unstable L-G Lines 40 428 550
22–23
5 6350 Unstable LLG Lines 60 318 450
22–23
5 6350 Unstable R5- Bus 24 – 363 450
LLL
6 6350 Stable R15- Bus 24 – 624 550
LLL
7 6350 Unstable (change LLL Lines 40 276 350
configuration) 23–24

Table 14.8 DTs validation results

Actual time FDDT CDDT IPDT
Scenario Fault Fault O.S Fault Clear O.S
(status) occurrence (s) clearance (s) (s) detection (s) detection (s) (s)
1 (U) 1.00 1.4 1.49 1.0 1.4 1.45
2 (S) 1.00 1.3 – 1.0 1.3 –
3 (U) 1.00 1.55 1.81 1.0 1.55 1.6
4 (U) 1.00 1.45 1.71 1.0 1.45 1.5
5 (U) 1.00 1.45 1.68 1.0 1.45 1.5
6 (S) 1.00 1.55 – 1.0 1.55 –
7 (U) 1.00 1.35 1.49 1.0 1.35 1.4

performance. According to each of the scenarios, the relevant CCT is calculated

using a time domain simulation to show the scenario’s stability and compare it with
the fault clearance time. The validation results of FDDT, CDDT, and IPDT for seven
scenarios compared to the real-time fault occurrence/clearance and O.S obtained by
time domain simulation studies are shown in Table 14.8. According to Table 14.8,
for all instability scenarios, IPDT predicts instability before the actual O.S time.
Figure 14.8 shows the analytical results of IPDT performance for 44 unstable
scenarios among 57 scenarios. Figure 14.8 shows the bar chart which the horizontal
axis is the length of time that the IPDT predicted O.S earlier than real-time. For
example, in 13 unstable scenarios, an O.S is predicted with a duration of 150–200 ms
earlier than the real time.
Scenario 1: In this scenario, the generator is unstable, and IPDT predicts an
unstable state. Figure 14.9a shows the changes in the active power of the generator
from the time before the fault occurrence to the time of after-fault clearance. Before
312 M. Abedi et al.

the fault occurrence time to the time of after-fault clearance, consecutive time
windows of continuously sampled data enter FDDT, CDDT, and IPDT.
According to Fig. 14.9b, in FDDT, the sudden change at the moment of the fault
occurrence in the sampled data is known as a fault occurrence and generates a “1”
activation signal to activate CDDT at its output. As long as this sudden change of

No. of Unstable Scenarios Predicted by IPDT

14
13
12
11
Num. of Scenario

10
9
8
7
6
5
4
3
2
1
0
10-100 100-150 150-200 200-250 250-300 300-500
Time duration

Fig. 14.8 IPDT statistical performance concerning different scenarios

Fig. 14.9 scenario 1—Time windows and output of DTs. (a) The fluctuation of generator electrical
active power following a fault occurrence and clearance in an unstable scenario, (b) time windows
for FDDT and CDDT and their output for the unstable scenario, (c) the fluctuation of rotor angle and
output of IPDT in an unstable scenario
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 313

Fig. 14.9 (continued)

314 M. Abedi et al.

data is seen in the sampled time window, the FDDT output remains “1” for six
consecutive time windows. At the moment of fault clearance, a similar sudden
change of data (blue) occurs in the sampling window. The CDDT knows this sudden
change in data as a fault clearance and generates a “1” signal at its output to activate
IPDT. So long as the sudden change of data is in the sampling time window, the
CDDT output remains “1” for six consecutive windows.
The result of IPDT performance of this scenario is shown in Fig. 14.9c. The IPDT
performance starts when the CDDT performance is finished. In other words, the
IPDT performance starts when all time window samples contain post-clearance data,
depending on the length of the time window, the duration it takes 50 ms.
Table 14.7 presents the time performance of DTs compared to actual time. In this
scenario, IPDT predicts the generator’s instability, 50 ms earlier than the actual time
of the generator O.S.
Scenario 2: The studied scenario is a stable case in which, according to
Table 14.7, IPDT predicts a stable state of the generator. In this scenario, the
IPDT continuously reflects the “0” signal as no instability at the output. FDDT
and CDDT clearly show the fault occurrence and fault clearance at t ¼ 1.0 and
t ¼ 1.3 s, respectively, which are the same as the time’s actual values. Figure 14.10a
shows the changes in the generator’s active power from the time before the fault
occurs after the time of the fault clearance. In this period, consecutive time windows
of continuously sampled data enter FDDT, CDDT, and IPDT.
As shown in Fig. 14.10b, in the moment of the fault occurrence owing to a sudden
change in electrical power, a sample of faulty data (yellow) enters the sampling

Fig. 14.10 scenario 2—Time windows and output of DTs. (a) The fluctuation of generator
electrical active power following a fault occurrence and clearance in stable scenario. (b) Time
windows for FDDT and CDDT and their output for the stable scenario. (c) The fluctuation of rotor
angle and output of IPDT
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 315

Fig. 14.10 (continued)

316 M. Abedi et al.

window. In FDDT, this change in the sampled data is known as a fault occurrence
and generates a “1” activation signal to start CDDT activity at its output. As long as
this sudden change of data is present in the sampled time window, the FDDT output
remains “1” for six consecutive time windows. At the moment of fault clearance, a
similar sudden change of data (blue) occurs in the sampling time window. The
CDDT knows this sudden change in data as a fault clearance that generates a “1”
signal at its output to activate IPDT. So long as the sudden change of data is in the
sampling time window, the CDDT output remains “1” for six consecutive windows.
In Fig. 14.10c, the IPDT performance result to predict the generator’s stable state is
demonstrated.
Scenario 3: This scenario shows an unstable case, and the performance of the DTs
is evaluated for L-G in lines 22–23 with the fault characteristics listed in Table 14.7.
As shown in Table 14.8, all DTs detect and predict the fault occurrence/fault
clearance and instability of the generator accurately.
Scenario 4: This scenario shows an unstable case, and the performance of the DTs
is evaluated for a LLG in lines 22–23 with the fault characteristics listed in
Table 14.7. All DTs accurately detect and predict the fault occurrence/fault clearance
and O.S of the generator. Based on results of Scenarios 3 and 4 based on Table 14.8,
the DTs can perform their tasks properly against asymmetric LG and LLG faults,
though DTs training with three-phase short circuit fault scenarios.
Scenario 5/6: These scenarios show unstable/stable cases, and the performance of
DTs is evaluated for the occurrence of a fault with resistance (RF ¼ 5 Ω/RF ¼ 15 Ω).
FDDT and CDDT detected fault occurrence and fault clearance for both stable and
unstable states, respectively. IPDT predicts the stable case for scenario 5(RF ¼ 15 Ω)
and then shows a “0” signal at the output and also predicts the unstable case for
scenario 6(RF ¼ 5 Ω) at t ¼ 1.5 s correctly.
Scenario 7: Regarding the change in the network configuration, the performance
of the DTs for an unstable fault occurred in the lines 23–24 is evaluated. All DTs
detect occurrence/clearance and correctly predict O.S, respectively.

14.5.1 Validation with Noisy Data

In this section, FDDT, CDDT, and IPDT training/testing is evaluated for noise-free
data in training and testing. Therefore, four noise levels (SNR ¼ 20, 30, 40, 50 dB)
are examined in the following conditions:
1. Train with noise-free data and test combined with noise
2. Test and train data combined with noise
According to Table 14.9, the performance of DTs accuracy in the case where DTs
are tested with noise data results in a relatively high prediction error in IPDT. In
Table 14.10, DTs are both tested and trained data combined with noise. The IPDT
prediction error is lower in this case, indicating that the presence of noise in the test
and train data can improve and make IPDT performance more robust. Also, by
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 317

Table 14.9 Performance of FDDT, CDDT, IPDT with noisy data in test
SNR(dB) No-noise ¼ 0 20 30 40 50
FDDT acc. (%) 99.75 99.45 99.47 99.47 99.52
CDDT acc. (%) 99.75 99.45 99.47 99.47 99.52
IPDT acc. (%) 89.96 73.78 79.46 85.78 88.9

Table 14.10 Performance of FDDT, CDDT, IPDT with noisy data in train and test
SNR(dB) No-noise ¼ 0 20 30 40 50
FDDT acc. (%) 99.75 99.79 99.8 99.79 99.81
CDDT acc. (%) 99.75 77.79 99.8 99.79 99.81
IPDT acc. (%) 70 76.52 76.52 81.37 89.85

examining the performance of FDDT and CDDT, it is determined that the error rates
are negligible because the noise cannot prevent the data from changing abruptly in
the time windows of the fault occurrence or fault clearance.

14.6 Conclusion

In this study, to predict a synchronous generator’s transient instability, an intelligent

relay scheme based on DT is presented. In the proposed scheme, fault occurrence,
fault clearance, and generator O.S are detected and predicted by FDDT, CDDT, and
IPDT. According to the results, IPDT relies on post-clearance data generated by
FDDT and CDDT in the previous steps to predict instability and O.S status. It was
discovered that electrical operating variables are the most appropriate variables for
fault occurrence/clearance detection. In contrast, electromechanical operational vari-
ables are the most relevant variables for predicting O.S. Adopting six data samples in
each time window, including data of fault occurrence and fault clearance as input
vector data, leads to increased reliability and provides the opportunity for DTs to
detect fault conditions and clearance in six consecutive time windows. In addition,
this type of time window ensures the proper performance of DTs against missed data
and bad data that may be present in the sampled time windows. According to results
it can be concluded that the FDDT, CDDT, and IPDT performance accuracy for
prediction and detection is 99.71, 99.71, and 98.73, respectively. The performance
of the presented framework has been assessed by various scenarios not observed in
the DT learning process. It is necessary to mention that the presented intelligent relay
can be used as a supportive relay for the main O.S relay. In other words, the
presented intelligent O.S relay can operate individually or in association with
prevalent relays. In cooperation status, its function can be utilized as a support for
a prevalent relay, so that in the final decision about the generator stability status, if
the presented relay predicts instability, it will act accordingly or if the output
impedance relay synchronous also detects instability and acts following its detection.
318 M. Abedi et al.

References

1. F.R. Gomez Lezama, Prediction and Control of Transient Instability Using Wide Area Phasor
Measurements, Thesis, Doctor of Philosophy, Department of Electrical and Computer Engi-
neering University of Manitoba (2011)
2. A.D. Rajapakse, F. Gomez, Rotor angle instability prediction using post-disturbance voltage
trajectories. IEEE Trans. Power Syst. 25(2), 947–956 (2009)
3. F.R. Gomez, A.D. Rajapakse, U.D. Annakkage, I.T. Fernando, Support vector machine-based
algorithm for post-fault transient stability status prediction using synchronized measurements.
IEEE Trans. Power Syst. 26(3), 1474–1483 (2010)
4. N. Amjady, S.A. Banihashemi, Transient stability prediction of power systems by a new
synchronism status index and hybrid classifier. IET Gener. Transm. Distrib. 4(4), 509–518
(2010)
5. A.G. Bahbah, Power System Transient Stability Assessment for Real-Time Applications, Thesis,
Doctor of Philosophy, Department of Electrical and Computer Engineering University of
Clemson (2000)
6. W.Xiaochen, et al., Review on transient stability prediction methods based on real time wide-
area phasor measurements, in 2011 4th International Conference on Electric Utility Deregula-
tion and Restructuring and Power Technologies (DRPT) (2011)
7. Y. Jie et al., PMU-based monitoring of rotor angle dynamics. IEEE Trans. Power Syst. 26(4),
2125–2133 (2011)
8. C.-X. Dou, J. Yang, X. Li, T. Gui, Y. Bi, Decentralized coordinated control for large power
system based on transient stability assessment. Int. J. Electr. Power Energy Syst. 46(March),
153–162 (2013)
9. D.P. Wadduwage, C.Q. Wu, U.D. Annakkage, Power system transient stability analysis via the
concept of Lyapunov exponents. Electr. Power Syst. Res. 104(November), 183–192 (2013)
10. F. Shi, H. Zhang, G. Xue, Instability prediction of the inter-connected power grids based on
rotor angle measurement. Int. J. Electr. Power Energy Syst. 88, 21–32 (2017)
11. H. Talaat, Predictive O.S relaying using fuzzy rule-based classification. Electr. Power Syst. Res.
48, 143–149 (1999)
12. L.S. Moulin, A.P.A. da Silva, M.A. El-Sharkawi, R.J. Marks II, Support vector machines for
transient stability analysis of large-scale power systems. IEEE Trans. Power Syst. 19(2),
818–825 (2004)
13. W. Hu, Z. Lu, S. Wu, et al., Real-time transient stability assessment in power system based on
improved SVM. J. Mod. Power Syst. Clean Energy 7, 26–37 (2019)
14. K.R. Padiyar, S. Krishna, Online detection of loss of synchronism using energy function
criterion. IEEE Trans. Power Deliv. 21(1), 46–55 (2006)
15. A.F. Diaz-Alzate, J.E. Candelo-Becerra, J.F. Villa Sierra, Transient stability prediction for real-
time operation by monitoring the relative angle with predefined thresholds. Energies MDPI
Open Access J. 12(5), 1–17 (2019)
16. I. Kamwa, S.R. Samantaray, G. Joos, Development of rule-based classifiers for rapid stability
assessment of wide-area post-disturbance records. IEEE Trans. Power Syst. 24(1), 258–270
(2009)
17. Q. Gao, S. Rovnyak, Decision trees using synchronized phasor measurements for wide-area
response-based control. IEEE Trans. Power Syst. 26(2), 855–861 (2011)
18. J.Z. Hui Deng, X. Wu, K. Men, Real time transient instability detection based on trajectory
characteristics and transient energy, in Power and Energy Society General Meeting, (2012),
pp. 1–7
19. A. Karimi, S.Z. Esmaili, Transient stability assessment of power systems described with
detailed models using neural networks. Int. J. Electr. Power Energy Syst. 45(1), 279–292 (2013)
20. Bahbah, A.A. Girgis, New method for generators’ angles and angular velocities prediction for
transient stability assessment of multimachine power systems using recurrent artificial neural
network. IEEE Trans. Power Syst. 19(2), 1015–1022 (2004)
14 Prediction of Out-of-Step Condition for Synchronous Generators Using. . . 319

21. A. Karimi, Power system transient stability margin estimation using neural networks.
Int. J. Electr. Power Energy Syst. 33(4), 983–991 (2011)
22. S. Kai, S. Likhate, V. Vittal, V.S. Kolluri, S. Mandal, An online dynamic security assessment
scheme using phasor measurements and decision trees. IEEE Trans. Power Syst. 22(4),
1935–1943 (2007)
23. Y. Wang, J. Yu, Real time transient stability prediction of multi-machine system based on wide
area measurement, in Power and Energy Society General Meeting, (2009), pp. 1–4
24. D. You, K. Wang, L. Ye, J. Wu, R. Huang, Transient stability assessment of power system
using support vector machine with generator combinatorial trajectories inputs. Int. J. Electr.
Power Energy Syst. 44(1), 318–325 (2013)
25. K. Yamashita, H. Kameda, O.S prediction logic for wide area protection based on an
autoregressive model, in Proceedings of IEEE PES Power Systems Conference Expo.
New York, 10–13 Oct 2004
26. F. Hashiesh, H.E. Mostafa, A.-R. Khatib, M. Ibrahim Helal, M.M. Mansour, An intelligent wide
area Synchrophasor based system for predicting and mitigating transient instabilities. IEEE
Trans. Smart Grid 3(2), 645–652 (2012)
27. T. Amraee, S. Ranjbar, Transient instability prediction using decision tree technique. IEEE
Trans. Power Syst. 28(3), 3028–3037 (2013)
28. A.N. Al-Masri et al., A novel implementation for generator rotor angle stability prediction using
an adaptive artificial neural network application for dynamic security assessment. IEEE Trans.
Power Syst. 23(3), 2516–2525 (March 2013)
29. A.R. Sobbouhi, M.R. Aghamohammadi, A new algorithm for predicting O.S condition in large-
scale power systems using rotor speed–acceleration. Int. Trans. Electr. Energy Syst. 26,
486–508 (2016)
30. A.R. Sobbouhi, M.R. Aghamohammadi, A new algorithm for predicting O.S using rotor speed-
acceleration based on phasor measurement units (PMU) data. Electr. Power Components Syst.
43(13), 1478–1486 (2015)
Chapter 15
The Adaptive Neuro-Fuzzy Inference
System Model for Short-Term Load, Price,
and Topology Forecasting of Distribution
System

Mehrdad Setayesh Nazar and Ashkan Eslami Fard

Nomenclature

Abbreviations
ANFIS Adaptive neuro-fuzzy inference system
ANN Artificial neural network
ARIMA Autoregressive integrated moving average
ARMA Autoregressive moving average
CANFIS Co-active neuro-fuzzy inference system
CSI Contingency severity index
DG Distributed generation
GA Genetic algorithm
GO Gravitational search optimization
LMP Locational marginal prices
LPF Load and price forecasting
MAPE Mean absolute percentage error
MLP Multilayer perceptron
SVM Support vector machine
SVR Support vector regression
WT Wavelet transform

Indices
l Decomposition index of wavelet decomposition
k Scaling index of wavelet decomposition

M. Setayesh Nazar (*) · A. Eslami Fard

Faculty of Electrical Engineering, Shahid Beheshti University, Tehran, Iran
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 321
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_15
322 M. Setayesh Nazar and A. Eslami Fard

Parameters
NB Number of buses
NCC Number of critical contingencies
NL Number of lines
Pmax Maximum active power flow of line
Vmin Minimum voltage of bus
α, β, χ Coarse and fine-scale coefficients of wavelet decomposition machine
W Weighting factor

Variables
P Active power flow of line
V Voltage of bus

15.1 Introduction

The load and price-forecasting (LPF) procedure is one of the most important tools
for the distribution system operational processes. The utilization of fossil-fueled
distributed generation (DG) units can highly change the load and price of system
patterns with respect to the cases that the DGs are not committed. The distributed
energy generation can change the nodal energy generation/consumption values.
Further, the DGs electricity transactions with the upward system can change the
hourly load and price of the distribution system buses.
In recent years, many kinds of research have carried out to optimize the LPF
process that can be categorized into different groups. The first group of LPF process
employs statistical algorithms such as autoregressive integrated moving average
(ARIMA) method [1]. The second group utilizes feature selection algorithms and
clustering machines such as Kalman filter [2], Box-Jenkins models [3], support
vector regression (SVR) [4], support vector machine (SVM) [5], fuzzy models [6],
artificial neural network (ANN) [7], and expert systems [8]. The third group com-
bines the forecasting machines of the first and second groups to increase the speed
and accuracy of the forecasting process. The wavelet transform (WT), co-active
neuro-fuzzy inference system (CANFIS), and adaptive neuro-fuzzy inference system
(ANFIS) are examples of the categories methods, which are used for price and load
forecasting [9, 10]. Ref. [11] presented a hybrid LPF process that utilized a feature
selection method and gravitational search optimization (GSO) algorithm. The ANN-
based LPF method maximum mean absolute percentage error (MAPE) of PJM
market was about 1.1849% for the worst case. Ref. [12] proposed a multistage
ANFIS-based LPF process that used feature selection, WT machine, and Kalman
filter. The first stage of LPF used the WT and Kalman machines to decompose the
input data into three frequency components and predict the decomposed signals.
Then, a Kohonen-based network found similar days, and the ANFIS and multilayer
perceptron (MLP) neural networks carried out the second stage of LPF process.
15 The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . . 323

Finally, in the third stage, the ANFIS and MLP neural networks were employed to
forecast the price and load curves. The average MAPE value of the proposed method
for the Spanish electricity market was about 3.88%. Ref. [13] introduced a LPF
model that forecasting procedures were carried out using SVM and WT machines,
and the optimization of learning parameter was performed by the GSO method. The
simulation of the algorithm showed that the proposed method increased the accuracy
of prediction by about 15.21% for the PJM market. Ref. [14] introduced a time series
decomposition and sin-cosine optimization process to predict the day-ahead price of
the market. The method used a deterministic and probabilistic forecasting algorithm
to increase the accuracy of the procedure. The MAPE of the introduced method took
on a value of 21.471% for the Australian electricity market price. Ref. [15] evaluated
a four-stage price-forecasting machine that consisted of preprocessing, optimization,
forecasting, and evaluation machines. The first stage used the signal-processing
machine to detect and remove white noises. The second stage optimization was
carried out by the grey wolf meta-heuristic approach to optimize the forecasting
parameters. The third stage utilized Elman neural network model for price forecast-
ing, and the fourth-stage machine evaluated the effectiveness of forecasting
machines. The MAPE of the forecasting method for the data of New South Wales
market, Australia, was about 3.331%. Ref. [16] proposed a two-stage price-forecast-
ing machine that correlated the wind power generation with the market price. The
direct method and rerouting method assessed the real-time prices and wind electric-
ity generation correlation. The rerouted method utilized SVM, ANN, and regression
machines, and the mean square error value for the ERCOT market was about
22.25%. Ref. [17] introduced a load-forecasting algorithm that used the relevance
vector machine to predict the load of the electricity market. The feature selection and
WT were carried out as preprocessing procedures, and the method was assessed for
New York and New England electricity markets. The values of the MAPE for the
hour-ahead and day-ahead price forecasting were about 0.86% and 1.58%, respec-
tively. Ref. [18] proposed a stochastic autoregressive moving average (ARMA)
model for price forecasting that was assessed for the Australian electricity market.
The results showed that the developed model predicted the spark values of prices and
considered the nonlinearity of regional markets. Ref. [19] presented an SVM-based
load-forecasting algorithm that used the grey wolf optimization process to optimize
the parameters of the forecasting machine. The proposed algorithm MAPE for the
New South Wales electricity load forecast was about 0.5204%. Ref. [20] assessed a
gradient boosting machine for load forecasting of Chinese electric power systems.
The inputs of forecasting machines were the daily, average, maximum, and mini-
mum temperature parameters, and the day of the week and the calculated value of
MAPE were about 0.032%. Ref. [21] introduced a dynamic decomposition process
for load forecasting that decomposed the error of forecasting. The error snapshot
matrices were calculated and the series of errors were generated. Then, the load
sequences were formulated based on the estimated values of errors. The MAPE of
the proposed method was about 4.452% for the Chinese power system. Ref. [22]
proposed a feature selection-based price-forecasting algorithm that clustered the
price data using a noncooperative game model. The game theory was used to
324 M. Setayesh Nazar and A. Eslami Fard

determine the winning neurons, and the forecasting was carried out using Bayesian
recurrent ANN. The MAPE of forecasting for the New York power system was
about 4.14%. Ref. [23] introduced a load-forecasting algorithm that employed
Boltzmann and mutual information machines. The forecasting parameter optimiza-
tion was carried out using genetic wind driven method. The case study was
performed for the PJM electricity market, and the MAPE value for the proposed
method was about 0.492%. Ref. [24] proposed a three-part algorithm that used WT,
mutual information machine, ARIMA machine, and SVM process. The SVM
machine parameter selection was determined using a meta-heuristic optimization
procedure. The MAPE values for load and price forecasting of New South Wales
electricity market were about 2.11% and 7.86%, respectively.
The described references do not consider the optimal topology forecasting of the
distribution system for the day-ahead horizon. This book chapter is about the
forecasting algorithm that considers the load, locational marginal prices (LMPs),
and topology forecasting of the distribution system for day-ahead scheduling
horizon.

15.2 Problem Modeling and Formulation

The distribution system operator (DSO) forecasts the day-ahead values of distribu-
tion system aggregated nodal load and the LMPs of buses. As shown in Fig. 15.1, the
DSO transacts energy with distributed generation (DG) units that are the nonutility
energy generation facilities.
The LMPs have nonconstant average and variance and depend on the marginal
costs of electricity generation/load reduction, congestion, and loss of lines. However,
the load of buses may have similar patterns. The LMPs are dependent on the loads
and vice versa, based on the fact that the higher values of loads may increase the
congestion, loss, and energy procurement costs of the system. Further, the higher
values of LMPs may reduce the volume of responsive loads based on their price
elasticity [25].
An iterative multistage LPF process is introduced, which its block diagram is
presented in Fig. 15.2.
At first, a primary LPF is carried out using a load, LMP, and topology historical
database. Then, the similar days clustering for load forecasting is performed using
Kohonen machine. The second stage load forecasting is processed using ANN
machine and daily temperature database. The feature selection procedure processes
the price data to find their relevancy, and the second stage price forecasting is carried
out using ANFIS process.
The topology-forecasting process uses the contingency analysis database and
determines the hourly available decision control variables. Then, the optimal topol-
ogy of the system is determined using a genetic algorithm (GA) optimization
procedure.
15 The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . . 325

Fig. 15.1 The schematic diagram of the distribution system

The first stage carries out WT process to decompose the data into different
frequencies. The forecasting processes of decomposed frequencies are easier than
the undecomposed data. The WT decomposition process is carried out in two levels,
as shown in Fig. 15.3. The high-frequency components (D2 and D1) and
low-frequency component (A1) are decomposed from the original data.
The WT formulation can be presented as Eq. (15.1):
326 M. Setayesh Nazar and A. Eslami Fard

Fig. 15.2 The proposed

iterative multistage LPF
process

Fig. 15.3 The

decomposition of the
original signal by wavelet
transform
15 The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . . 327

X XX l  
f ðt Þ ¼ cl0,k α j0,k ðt Þ þ βl,k 22 χ 2l t  k ð15:1Þ
k l<l0 k

where, l, k are decomposition and scaling index, respectively. Further, α, β, and χ are
coarse and fine-scale coefficients, respectively.
Thus, six components of the original load and price data are available. Then, a
Kalman forecasting machine is utilized for the first-stage LPF that can be formulated
as Eq. (15.2):

X
t X
j
loadðνÞ ¼ φi  loadðν  νi Þ þ ψ i  priceðν  θi Þ ð15:2Þ
i¼1 i¼1

where, price(ν  θi) and load(ν  νi) are the past price and load parameters,
respectively. Further, φ and ψ are calculated from transition matrices, and θi and νi
are the indices of previous prices and loads, respectively.
The similar day clustering process is carried out by the Kohonen machine and the
processed data are delivered to the ANN-based load-forecasting procedure. The
decomposed frequency components of the original load signal are forecasted by
the ANN machines in the second stage of the forecasting process. Further, the mutual
information machine finds the dependency of LMPs. The second stage LMP fore-
casting utilizes the ANFIS machine. The detailed formulations of mutual informa-
tion and ANFIS machines are presented in [12]. The forecasted values of A1, A2, and
D2 signals are reconstructed as shown in Fig. 15.4.
The optimal topology-forecasting process utilizes the off-line calculation engine
and consists of the following stages:
1. Off-line contingency analysis and ranking for different load patterns are
carried out.

Fig. 15.4 The restructuring

of decomposed signals
328 M. Setayesh Nazar and A. Eslami Fard

2. Optimal switching of the distribution system using the GA is performed. The

objective function of GA optimization process is defined as Eq. (15.3):

XV  V 2 XP  P 2
min Z ¼ W 1  min
þ W2  max
ð15:3Þ
NB
V min NL
Pmax

where, W is the weighting factor. Equation (15.3) is subjected to the AC load

flow and distribution system device loading constraints. The details of GA are
available in [26].
3. The homogeneous topologies are determined for different contingency and load
conditions.
4. The learning of CANFIS machine is performed. The detailed formulation of
CANFIS machine is presented in [27].
The online optimal topology-forecasting procedure consists of the following
steps:
1. Contingency Severity Index (CSI) is calculated using Eq. (15.4)
" 2 #
X XV  V 2 X
P  Pmax
CSI ¼ normal
þ ð15:4Þ
NCC NB
V normal NL
Pmax

2. The CANFIS machine forecasts the optimal topology of the distribution system
based on the CSI, LMPs, and load of the system.
The CANFIS machine is the extension of ANFIS machine that consists of multi-
outputs that are shown in Fig. 15.5.
The number of CANFIS machine outputs is equal to the number of distribution
topologies that are determined in the off-line studies.

15.3 Simulation Results

The 33-bus IEEE test system was considered for simulating the proposed method.
Figure 15.6 presents the topology of the 33-bus IEEE test system, and its data is
available in [28].
Numerous electricity consumption scenarios were generated, and the LMPs were
calculated using the proposed method of [29]. The Nord Pool per unit load patterns
were considered for generating of the 33-bus IEEE system hourly load pattern for
different days and seasons. Then, the system load and LMP data were delivered to
the proposed LPF machines.
Figure 15.7 depicts the WT decomposition levels. The A2 signal has the highest
contribution value. However, the D2 and D1 signals have values lower than 0.045
that are very low with respect to the A2 signal value.
15 The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . . 329

Fig. 15.5 The schematic diagram of CANFIS machine

Fig. 15.6 The single line diagram of the 33-bus IEEE test system

Figure 15.8a, b present the correlation between load and the historical load
database. There is a linear correlation between the load of forecasting day and the
load of the previous day. Further, a linear correlation between the load of forecasting
day and the load of the 7 days before can be concluded.
The first-stage load-forecasting input data were selected as the following set:
330 M. Setayesh Nazar and A. Eslami Fard

Fig. 15.7 The wavelet transform decomposition levels of the original signal

fLðd, t  1Þ, Lðd, t  2Þ, Lðd, t  3Þ, Lðd, t  4Þ, Lðd, t  5Þ, Lðd  1, t  1Þ, Lðd  1, t Þ,
Lðd  1, t þ 1Þ, Lðd  7, t  1Þ, Lðd  7, t Þ, Lðd  7, t þ 1Þg
ð15:5Þ

Further, the first-stage price-forecasting input data were selected as the following
set:

fPðd, t Þ, Pðd, t  1Þ, Pðd, t  2Þ, Pðd, t  3Þ, Pðd, t  4Þ, Pðd, t  5Þ, Pðd  1, t  1Þ,
Pðd  1, t Þ, Pðd  1, t þ 1Þ, Pðd  7, t  1Þ, Pðd  7, t Þ, Pðd  7, t þ 1Þg
ð15:6Þ

The described data were considered as the input of Kalman filters, and the A2, D2
and D1 signals were forecasted by the Kalman filters as shown in Fig. 15.9.
An average LMP (ALMP) was defined, which was the hourly average of LMP for
the specified day-ahead horizon.
Figure 15.10 presents the ALMP A2, D2, and D1 signals and the final value of
ALMP for the day-ahead horizon.
Then, the Kalman forecasted LMPs, and loads were fed to the Kohonen machine
to cluster the input data. Figure 15.11 presents the clustered load data for the A2
signal, which was determined by the Kohonen machine. More details of parameter
selection are presented in [12].
The Kohonen process found the following patterns for the A2, D2, and D1 signals:
Similar days D2 component for load data:
{109‘122‘125‘230‘235‘236‘237‘244‘411‘594‘599}.
Similar days A2 component for load data:
{116‘117‘125‘228‘229‘234‘238‘475‘481‘287‘501‘591‘592‘593‘594‘598}
Similar days D1 component for load data:
{131‘159‘178‘250‘279‘496‘549‘556}.
15 The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . . 331

Fig. 15.8 (a) The correlation of load and the historical load database for d  1 and day of
forecasting, (b) the correlation of load and the historical load database for d  7 and day of
forecasting
332 M. Setayesh Nazar and A. Eslami Fard

Fig. 15.9 (a) The Kalman filter forecasted load D1 signal, (b) the Kalman filter forecasted load A2
signal, (c) the Kalman filter forecasted load D2 signal
15 The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . . 333

The second stage of load forecasting was performed by the multilayer perceptron
neural network. Figure 15.12 presents the second stage load forecasting of A2 signal
that was carried out by the neural network. Figure 15.13 depicts the maximum and
minimum values of MAPE for neural network load-forecasting process. As shown in
Fig. 15.13, the maximum and minimum values of MAPE took on a value of 3.88%
and 1.31%, respectively.
The ANFIS forecasting engine forecasted LMPs in the second stage of LPF.
Figure 15.14 presents the forecasted ALMP.

Fig. 15.10 (a) The Kalman filter forecasted ALMP D1 signal, (b) the Kalman filter forecasted
ALMP A2 signal, (c) the Kalman filter forecasted ALMP D2 signal, (d) the Kalman filter forecasted
ALMP signal
334 M. Setayesh Nazar and A. Eslami Fard

Fig. 15.10 (continued)

Figure 15.15 depicts the GA objective function values, which was calculated
from Eq. (15.3). The maximum values of the GA objective function took on a value
of 1242.432, 1358.784, and 1426.458 for the load factor 0.6, 0.8, and 1, respectively.
Tables 15.1 and 15.2 depict the CANFIS pattern recognition and GA optimiza-
tion outputs for different load factor (LF) and ALMP values. The total number of the
CANFIS machine pattern recognitions was 444 cases, and the number of recognition
errors was six cases or 1.35%. The proposed forecasting machines were successfully
determined the day-ahead load, price, and topology. More details of the justification
process are presented in [12].
15 The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . . 335

Fig. 15.11 The Kohonen machine clustered load data for the A2 signal

15.4 Conclusion

A day-ahead load, price, and topology-forecasting process was reviewed in this

chapter. The introduced framework employed feature selection and Kohonen clus-
tering machines. Two-level load and locational marginal price-forecasting process
were carried out. The wavelet machine decomposed the original signals into differ-
ent frequencies, and the Kalman filter predicted the decomposed signals for the
day-ahead horizon. Then, the Kohonen process clustered the first-level forecasted
signals, and the multilayer perceptron neural network machine forecasted the
decomposed frequencies of the load. The adaptive neuro-fuzzy inference system
was utilized to forecast the decomposed signals of locational marginal prices.
Finally, the load and price signals were restructured.
The topology-forecasting machine considered the contingency severity index,
locational marginal price, and load of system as inputs, and a genetic algorithm
optimized the topology of the system considering system constraints. Then, a
co-active neuro-fuzzy inference system was used to predict the optimal topology
of the system. The proposed method was successfully assessed for the 33-bus IEEE
test system.
The authors are working on the new structure of the proposed algorithm that
utilizes meta-heuristic optimization algorithm to speed up the process.
336 M. Setayesh Nazar and A. Eslami Fard

Fig. 15.12 (a) The MLP forecasted load D1 signal, (b) the MLP forecasted load A2 signal, (c) the
MLP forecasted load D2 signal
15 The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . . 337

Fig. 15.13 The neural network MAPE for the forecasted load data

Fig. 15.14 The ANFIS machine forecasted ALMP

Fig. 15.15 The GA objective functions for different load levels

Table 15.1 Comparison of the GA outputs and the CANFIS machine for optimal topology recognition and different ALMP and LF values
338

GA (ALMP ¼ 1) CANFIS (ALMP ¼ 1) GA (ALMP ¼ 0.7) CANFIS (ALMP ¼ 0.7) GA (ALMP ¼ 0.5) CANFIS (ALMP ¼ 0.5)
Contingency LF ¼ 0.95 LF ¼ 0.95 LF ¼ 0.9 LF ¼ 0.9 LF ¼ 0.85 LF ¼ 0.85
1 1 1 1 1 6 6
2 2 2 2 2 5 5
3 4 4 4 4 3 3
4 3 3 3 3 4 4
5 1 1 1 1 1 1
6 1 1 1 1 1 1
7 4 4 1 1 1 1
8 1 1 1 3 1 1
9 4 4 4 4 4 4
10 1 3 1 1 1 1
11 1 1 1 1 1 1
12 1 1 1 1 1 1
13 6 6 6 1 1 1
14 4 4 4 4 1 1
15 1 1 1 3 1 1
16 1 1 3 1 3 3
17 6 6 1 1 1 1
18 2 2 1 1 1 1
19 1 1 1 1 1 1
20 1 1 1 2 1 1
21 1 1 1 1 1 1
22 1 1 1 1 1 1
23 1 1 1 1 1 1
24 1 1 1 1 1 1
25 1 1 1 1 1 1
M. Setayesh Nazar and A. Eslami Fard
 15

26 1 1 1 1 1 4
27 1 1 1 1 1 1
28 2 2 1 1 1 1
29 1 1 1 1 1 1
30 2 2 1 1 1 1
31 1 1 1 1 1 1
32 1 1 1 1 1 1
33 1 1 1 1 1 1
34 1 1 1 1 1 1
35 1 1 1 1 1 1
36 1 1 1 1 1 1
37 1 1 1 1 1 1
The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . .
339
Table 15.2 Comparison of the GA outputs and the CANFIS machine for optimal topology recognition and different ALMP and LF values
340

GA (ALMP ¼ 1) CANFIS (ALMP ¼ 1) GA (ALMP ¼ 0.7) CANFIS (ALMP ¼ 0.7) GA (ALMP ¼ 0.5) CANFIS (ALMP ¼ 0.5)
Contingency LF ¼ 0.8 LF ¼ 0.8 LF ¼ 0.75 LF ¼ 0.75 LF ¼ 0.7 LF ¼ 0.7
1 7 7 1 1 1 7
2 5 5 1 1 1 1
3 1 1 1 1 1 1
4 1 1 1 1 1 1
5 2 2 2 2 2 2
6 1 1 1 1 1 1
7 3 3 1 1 3 3
8 1 1 1 1 1 1
9 1 1 1 1 1 1
10 4 4 4 4 3 3
11 4 4 1 1 4 4
12 1 1 1 1 1 1
13 3 3 3 3 3 3
14 1 1 1 1 1 1
15 3 3 1 1 3 3
16 7 7 7 7 7 7
17 2 2 2 2 1 1
18 6 6 2 2 2 2
19 1 1 1 1 1 1
20 1 1 1 1 1 1
21 1 1 1 1 1 1
22 6 6 1 1 6 6
23 1 1 1 1 1 1
24 1 1 1 1 1 1
25 1 1 1 1 1 1
M. Setayesh Nazar and A. Eslami Fard
 15

26 1 1 1 1 1 1
27 1 1 1 1 1 1
28 2 2 1 1 1 1
29 1 1 1 1 1 1
30 1 1 1 1 1 1
31 1 1 1 1 1 1
32 1 1 1 1 1 1
33 1 1 1 1 1 1
34 3 3 3 3 3 3
35 3 3 3 3 2 2
36 1 1 1 1 1 1
37 1 1 1 1 1 1
The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . .
341
342 M. Setayesh Nazar and A. Eslami Fard

References

1. M.H. Amini, A. Kargarian, O. Karabasoglu, ARIMA-based decoupled time series forecasting

of electric vehicle charging demand for stochastic power system operation. Electr. Power Syst.
Res. 140, 378–390 (2016)
2. P. Gupta, K. Yamada, Adaptive short-term load forecasting of hourly load using weather
information. IEEE Trans. Power Syst. 91, 2085–2094 (2007)
3. A.J. Amorim, T.A. Abreu, M.S. Tonelli-Neto, C.R. Minussi, A new formulation of multinodal
short-term load forecasting based on adaptive resonance theory with reverse training. Electr.
Power Syst. Res. 179, 106096 (2020)
4. J. Xing Che, J. Zhou Wang, Y. Juan Tang, Optimal training subset in a support vector
regression electric load forecasting model. Appl. Soft Comput. 12, 1523–1531 (2012)
5. P.F. Pai, W. Chiang Hong, Support vector machines with simulated annealing algorithms in
electricity load forecasting. Energy Convers. Manage. 46, 2669–2688 (2005)
6. R. Mamlook, O. Badran, E. Abdulhadi, A fuzzy inference model for short-term load forecast-
ing. Energy Policy 37(4), 1239–1248 (2009)
7. S. Kouhi, F. Keynia, A new cascade NN based method to short-term load forecast in
deregulated electricity market. Energy Convers. Manage. 71, 76–83 (2013)
8. A. Ghanbari, S.M.R. Kazemi, F. Mehmanpazir, M.M. Nakhostin, A Cooperative Ant Colony
Optimization-Genetic Algorithm approach for construction of energy demand forecasting
knowledge-based expert systems. Knowl. Based Syst. 39, 194–206 (2013)
9. J.P.S. Catalão, H.M.I. Pousinho, V.M.F. Mendes, Short-term electricity prices forecasting in a
competitive market by a hybrid intelligent approach. Energy Convers. Manage. 52, 1061–1065
(2011)
10. A.S. Pandey, D. Singh, S.K. Sinha, Intelligent hybrid models for short-term load forecasting.
IEEE Trans. Power Syst. 25, 1266–1273 (2010)
11. A. Heydari, M. Majidi Nezhad, E. Pirshayan, D. Astiaso Garcia, F. Keynia, L.D. Santoli, Short-
term electricity price and load forecasting in isolated power grids based on composite neural
network and gravitational search optimization algorithm. Appl. Energy 277, 115503 (2020)
12. M.S. Nazar, A.E. Fard, A. Heidari, M. Shafie-khah, J.P.S. Catalão, Hybrid model using three-
stage algorithm for simultaneous load and price forecasting. Electr. Power Syst. Res. 165,
214–228 (2018)
13. Y. Zhang, C. Deng, R. Zhao, S. Leto, A novel integrated price and load forecasting method in
smart grid environment based on multi-level structure. Eng. Appl. Artif. Intell. 95, 103852
(2020)
14. W. Yang, J. Wang, T. Niu, P. Du, A novel system for multi-step electricity price forecasting for
electricity market management. Appl. Soft Comput. J. 88, 106029 (2020)
15. W. Yang, J. Wang, T. Niu, P. Du, A hybrid forecasting system based on a dual decomposition
strategy and multi-objective optimization for electricity price forecasting. Appl. Energy 235,
1205–1225 (2019)
16. S. Luo, Y. Weng, A two-stage supervised learning approach for electricity price forecasting by
leveraging different data sources. Appl. Energy 242, 1497–1512 (2019)
17. J. Ding, M. Wang, Z. Ping, D. Fu, V.S. Vassiliadis, An integrated method based on relevance
vector machine for short-term load forecasting. Eur. J. Oper. Res. 287, 497–510 (2020)
18. H. Manner, F. Alavi Fard, A. Pourkhanali, L. Tafakori, Forecasting the joint distribution of
Australian electricity prices using dynamic vine copulae. Energy Econ. 78, 143–164 (2019)
19. A. Yang, W. Li, X. Yang, Short-term electricity load forecasting based on feature selection and
least squares support vector machines. Knowl.-Based Syst. 163, 159–173 (2019)
20. Z. Guo, K. Zhou, X. Zhang, S. Yang, A deep learning model for short-term power load and
probability density forecasting. Energy 160, 1186–1200 (2018)
21. X. Kong, C. Li, C. Wang, Y. Zhang, J. Zhang, Short-term electrical load forecasting based on
error correction using dynamic mode decomposition. Appl. Energy 261, 114368 (2020)
15 The Adaptive Neuro-Fuzzy Inference System Model for Short-Term Load. . . 343

22. M. Ghayekhloo, R. Azimi, R. Ghofrani, M.B. Menhaj, E. Shekari, A combination approach

based on a novel data clustering method and Bayesian recurrent neural network for day-ahead
price forecasting of electricity markets. Electr. Power Syst. Res. 168, 184–199 (2019)
23. G. Hafeez, K. Saleem Alimgeer, I. Khan, Electric load forecasting based on deep learning and
optimized by heuristic algorithm in smart grid. Appl. Energy 269, 114915 (2020)
24. A. Ghasemi, H. Shayeghi, M. Moradzadeh, M. Nooshyar, A novel hybrid algorithm for
electricity price and load forecasting in smart grids with demand-side management. Appl.
Energy 177, 40–59 (2016)
25. S. Salarkheili, M. Setayesh Nazar, Capacity withholding analysis in transmission constrained
electricity markets. IET Gener. Transm. Distrib. 10, 487–495 (2016)
26. S. Madadi, M. Nazari-Heris, B. Mohammadi-Ivatloo, S. Tohidi, Implementation of genetic-
algorithm-based forecasting model to power system problems, in Handbook of Research on
Predictive Modeling and Optimization Methods in Science and Engineering, (IGI Global,
Hershey, 2018), pp. 140–155
27. H. Singh, Y.A. Lone, Deep Neuro-Fuzzy Systems with Python: With Case Studies and Appli-
cations from the Industry (Apress Book, Berkeley, 2020)
28. A. Bostan, M. Setayesh Nazar, M. Shafie-khah, J.P.S. Catalão, Optimal scheduling of distri-
bution systems considering multiple downward energy hubs and demand response programs.
Energy 190, 116349 (2020)
29. S. Salarkheili, M. Setayesh Nazar, New indices of capacity withholding in power markets. Int.
Trans. Electr. Energy Syst. 25, 180–196 (2015)
Chapter 16
Application of Machine Learning
for Predicting User Preferences in Optimal
Scheduling of Smart Appliances

Milad Sadat-Mohammadi, Morteza Nazari-Heris, Alireza Ameli,

Somayeh Asadi, Behnam Mohammadi-Ivatloo, and Houtan Jebelli

16.1 Introduction

The presence of innovative pricing strategies in the residential sector for electrical
energy substantiates the necessity of home energy management systems [1]. Home
energy management systems optimally operate the flexible appliances, based on
predefined constraints by residents, to reduce the overall electricity bill [2, 3]. These
systems consist of hardware and software layers [4, 5]. The hardware layer consists
of smart meters and actuators to connect and establish the communication between
appliances and the central home energy management unit. Designing the actuators
and meters modules is a challenging issue since each appliance has its own control
circuit, which should be modified to receive the control commands from the home

M. Sadat-Mohammadi
Department of Electrical Engineering and Computer Science, Pennsylvania State University,
University Park, PA, USA
e-mail: [email protected]
M. Nazari-Heris · S. Asadi (*) · H. Jebelli
Department of Architectural Engineering, Pennsylvania State University, University Park, PA,
USA
e-mail: [email protected]; [email protected]; [email protected];
[email protected]; [email protected]; [email protected]
A. Ameli
Department of Civil Engineering, Malard Branch, Islamic Azad University, Tehran, Iran
e-mail: [email protected]
B. Mohammadi-Ivatloo
Department of Energy Technology, Aalborg University, Aalborg, Denmark
Faculty of Electrical and Computer Engineering, University of Tabriz, Tabriz, Iran
e-mail: [email protected]; [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 345
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_16
346 M. Sadat-Mohammadi et al.

energy management unit and implement the control commands. Recently, smart
appliances have been introduced which have embedded communication capability.
The smart appliance can receive external signals for controlling its operation from
the user, utility, or central home energy management unit without any need for
modification of its control circuit. Consequently, the central control unit can send
on/off commands to the appliance or receive information about the status of the
appliance. The control commands are being determined by running and optimization
algorithm embedded in the home energy management unit.
There are several heuristics algorithms proposed in the literature, such as particle
swarm optimization, genetic algorithm, etc. to optimize the residential load sched-
uling models [6, 7]. These models consist of optimization parameters, constraints,
decision variables, and an objective function. Constraints are defined to satisfy the
technical requirement of operating the appliances. The objective function aims to
reduce the monthly electricity bill, considering the constraints and hourly electricity
rates. Moreover, some studies also consider the user’s comfort level in the objective
function [8]. The user comfort level is measured as the total waiting time to get the
tasks done after running the scheduling unit. There are two types of parameters
present in the scheduling models: fixed and dynamic parameters. Fixed parameters
are total energy required by each appliance for completing the tasks, cycle times, etc.
that are defined in the central home energy management unit once and are not
required to be updated regularly. However, dynamic parameters such as start time
and length of the desired scheduling window for each appliance should be updated at
the beginning of the scheduling process. Updating the dynamic parameters is a
challenging task and considered as one of the factors that makes the implementation
of home energy management systems less attractive for residential customers. It is
clear that users’ lifestyle and activity patterns significantly affect the dynamic
parameters. Consequently, analyzing the activity pattern of the residents can provide
useful information in addressing the mentioned challenge in implementing the home
energy management systems.
The activity prediction of the residents based on the sensor network data in the
smart home has been studied in the literature. For instance, data motion sensors and
door sensors, installed on the doors and cabinets, can be analyzed and used to predict
the order of the activities, start times, as well as duration [9]. Studies have investi-
gated the possibility of discovering the daily routines of the residents from the usage
pattern of the appliances. The smart meters that are connected to the appliances
record their energy consumption profile; then, the on/off status of the appliance is
determined from the recorded profile. The results proved that human activity can be
predicted with an acceptable accuracy using Bayesian network [10]. Not only human
activities can be predicted using sensory data, but also several studies have reported
the possibility of human preference detection. For instance, environmental informa-
tion and historical data of the operation of appliances were used to predict the
comfort preferences of the residents in operating the air conditioner and water heater
[11]. Moreover, the GPS data from smart mobile phones and social application data,
in addition to weather data, can be used to predict the arrival and living time of the
residents and control the thermostatic loads in a way to not only reduce the electricity
16 Application of Machine Learning for Predicting User Preferences in Optimal. . . 347

bill but also provide the required thermal comfort for the residents [12]. Inspiring
from proposed approaches in the literature for the prediction of the residents’
preferences in the operation of the thermostatic loads, this study aims to study the
possibility of prediction of user preferences in scheduling the flexible loads, such as
washing machine and dishwasher, based on the recorded energy consumption
pattern by smart applicants. Our hypothesis is that users’ preferences in operating
the smart appliances can be learned from historical data collected by smart appli-
ances; then, the user’s preferences can be predicted for future observations. Smart
appliances can provide a comprehensive database for this purpose without any need
for installing additional hardware such as smart meters [13].
The rest of the study is organized as follows: Sect. 16.2 introduces the factors
affecting the dynamic parameters of the residential load scheduling problem.
Section 16.3 elaborates on supervised learning and data labeling. Section 16.4
investigates the accuracy of predicting the dynamic parameters with machine learn-
ing algorithms. Finally, the concluding remarks and some directions for future
research are provided in Sect. 16.5.

16.2 Key Factors Affecting Dynamic Parameters

Dynamic parameters in scheduling problems are affected by several factors, such as

the start time of the task, environmental factors, the weekly schedule of the residents,
etc. [14]. However, it should be mentioned that there is some level of randomness in
dynamic parameters that lie outside the scope of this study. Some of the mentioned
factors can be detected by the energy consumption profile of the appliances, while
others can be acquired from developed online databases, including environmental
factors such as temperature. Investigating the relation between the factors and the
dynamic parameters can provide insight into the feasibility of predicting the dynamic
parameters.

16.2.1 The Start Time of the Task

The start time of the task is defined as the moment when an appliance becomes ready
to start its operation cycle. This time can be detected from recorded data by smart
appliances. By analyzing the energy consumption profile of the appliance, the
beginning of its operation cycle can be detected whenever the energy consumption
increases from the standby energy consumption limit. Recorded data of appliances
of a case study for a 2-year period shows that the start time of the tasks can affect the
desired length of scheduling interval. For instance, the end of the scheduling interval
of the dishwasher is the typical time for serving the dinner or lunch, while the start
time is variable and depends on the user to get the dishwasher ready to start.
Consequently, whenever the user fills the dishwasher close to the meal serving
348 M. Sadat-Mohammadi et al.

time, the desired scheduling window is shorter in comparison to the scenario in

which the dishwasher is filled earlier. In this study, the start time of the task is
considered as a feature in training the machine learning algorithms.

16.2.2 Day of the Week

Analyzing the hourly electricity consumption pattern of 60 low-energy houses for

12 months showed that the weekend peak electricity consumption occurs at 19:00,
while there is a small peak in electricity consumption at 8:00 in addition to the main
peak during the weekdays [15]. The difference between weekday and weekend
energy consumption profile is reasonable as people typically left the house in the
morning during the weekdays, which causes a reduction in electricity consumption
after 8:00. It can be concluded that weekdays and weekends significantly affect the
activities of the residents. Consequently, the change in the activity at weekends can
result in a change in the desired length of the scheduling window of the appliances.
Moreover, each user has specific working hours and tasks, which should be done on
a daily basis. Due to the possible importance of the day of the week in the length of
the desired scheduling window of the appliances, this factor is considered as an input
for training machine learning algorithms.

16.2.3 Temperature

Outdoor temperature is also an important factor in the daily activity of the residents.
People spend more time doing outdoor activities whenever the temperature is
pleasant while they prefer to stay at home whenever outside is snowy or rainy.
Spending less time indoors limits the available time for getting daily tasks done;
consequently, the desired scheduling window for scheduling the appliances may be
preferred to be as short as possible. However, it should be noted that depending on
the lifestyle of the residents, the impact of temperature on the desired scheduling
window can be notably different. Therefore, the relationship between the outdoor
temperature and the desired scheduling window for each appliance can be detected if
this factor is considered as a feature for training the machine learning algorithm in
the next steps.

16.3 Supervised Learning

Supervised learning refers to learning the relationship between input and output from
labeled examples. The labeled examples are divided into two sets as the training and
test sets. The training set is used during training the machine learning algorithm; the
16 Application of Machine Learning for Predicting User Preferences in Optimal. . . 349

training process involves tuning the model parameters. Weights, support vectors,
and coefficients are model parameters of artificial neural network, support vector
machine, and linear regression machine learning algorithms, respectively
[16]. Model parameters are often optimally determined using an optimization algo-
rithm whose objective function is the training accuracy. Although optimization
algorithms increase the accuracy of training, the trained model may not be general-
ized well on unseen dataset; this issue is referred to as overfitting. Overfitting occurs
whenever the model learns the hidden pattern specific to the training dataset, which
cannot completely represent the overall pattern. The test dataset is used to have an
unbiased evaluation of the performance of the model and the accuracy of general-
izing the model. The accuracy of the trained model in predicting future examples is
investigated, comparing the predicted labels with the expected labels for the test set.
One practical approach to avoid overfitting is to consider both training and test errors
as the objective function of the optimization algorithm; by doing so, there will be a
balance between the ability of the machine learning algorithm to learn from available
data and its ability to predict the future observations.
Supervised machine learning algorithms such as support vector machine (SVM),
artificial neural networks (ANN), and random forest (RF) have been used in the
literature for a variety of prediction problems. Each algorithm has a specific basic
principle, which makes it suitable for a particular problem. SVM algorithm seeks to
determine a set of hyper plains in high dimensional space that can separate the
nearest training classes with a higher margin. A higher margin can increase the
accuracy of the generalization of the model to future observations. SVM has broad
application in face detection, handwriting recognition, generalized predictive con-
trol, etc. ANN algorithm is a well-known machine learning algorithm that is inspired
by the simplification of neurons in the human brain. The model consists of an
interconnected artificial neuron. Each artificial neuron has weighted input, bias,
transfer function, and one output [17]. The model parameters, weights, and biases
are adjusted by a training algorithm during the learning process to find the optimum
values which result in the higher accuracy of the prediction. Gradient descent
algorithm, backpropagation algorithm, and heuristic algorithms such as particle
swarm optimization (PSO) or genetic algorithms are some examples of training
algorithms. There are several architectures for ANN, such as feed-forward neural
networks, multilayer perceptron (MLP), and convolutional neural networks. RF
consists of several decision trees that are trained on a bootstrapped sample
[18]. Each decision tree is built in a way to reduce the correlation of the tress and
achieve high accuracy in prediction than that of each individual tree. The random
forest has been used widely in the literature due to being insensitive to high
dimensional features, performing fast out of sample prediction, requiring less
parameter tuning, and having feature ranking capability [19]. Each machine learning
algorithm is the best fit for a specific problem; however, there is not any rule to
determine which algorithm is the best choice for a specific problem without
implementing and analyzing the accuracy of the results. In this study, we
implemented ANN, SVM, and RF algorithms; then, the results were analyzed to
determine the algorithm with the highest accuracy.
350 M. Sadat-Mohammadi et al.

16.3.1 Data Labeling

Extracting the desired scheduling window from recorded data by smart appliances is
the primary step for labeling the input data of machine learning algorithms. The
desired scheduling window of appliances can be detected with the proceeding
activity recorded in the dataset. For instance, the end of the preferred scheduling
window for a washing machine can be determined using the recorded data by the
dryer. The start time for the operation of the dryer represents the end of the
scheduling window for the washing machine. Moreover, the end of the scheduling
window of the dishwasher is typically the closest meal serving time. Instead of
considering the continuous values for the desired scheduling window, each extracted
value is assigned to a cluster with a specific length. For instance, if the desired
scheduling window for the washing machine for an operation cycle is 2 h and
15 min, it is rounded to the nearest hour, which is 2 h. Then, the labeled data for
each operation cycle and the corresponding value for the factors mentioned in the
previous are used to train and test machine learning algorithms.

16.4 Simulation Results

The possibility of predicting the desired scheduling window based on the introduced
factors using machine learning algorithms is elaborated in this section. A dataset
including 1-min readings from meters connected to several flexible appliances such
as dishwasher, dryer, washing machine, etc. is used. The dataset was collected for
2 years interval, from 2012 to 2014, and includes weather data [20]. There is not any
dataset in the literature collected by smart appliances; consequently, we use this
dataset to investigate the feasibility of the proposed approach. However, the result of
this study can be generalized to any house with smart home appliances. In this study,
we only investigate the accuracy of the predicting desired scheduling interval of the
washing machine by collected data. Consequently, collected data is divided into two
sets: training and test sets to learn the relation between input and labels; then, analyze
the accuracy of prediction. Twelve labels are extracted where each label represents
the length of the desired scheduling window of the washing machine. Figure 16.1
illustrates the relative frequency of each label.
Three machine learning algorithms, ANN, SVM, and RF are used to learn the
hidden relations between the affecting factors and labels from the training set; then
predict the respective labels of the test set. The results showed that the accuracy of
prediction is significantly low for 12 labels, while including only three labels with
the highest probability of occurrence results in higher prediction accuracy. The
underlying reason for this outcome is the week relationship between some of the
labels with the inputs. The labels with the scheduling window higher than 3 h do not
represent the overall trend of using the washing machine by the residents. For
instance, they may relate to operating the washing machine in some circumstances
16 Application of Machine Learning for Predicting User Preferences in Optimal. . . 351

Fig. 16.1 Extracted length of scheduling window for washing machine

which are not part of daily routine such as special gatherings, events, parties, etc.;
consequently, the pattern of these labels cannot be learned and predicted as they do
not correlate with the daily life of the residents.
The results of predicting the three labels using ANN, SVM, and RF were
analyzed to determine which of the mentioned algorithms is the best fit for this
problem. The results showed that the KNN algorithm has higher accuracy in
predicting the labels while being less prone to overfitting. The accuracy of prediction
was 77.6% for the test set. In order to visualize the result of prediction, principal
component analysis (PCA) is used to reduce the dimension of the input data
[21]. PCA cuts down the dimension of the input data into two unitless dimensions,
as shown in Fig. 16.2. Labels 1, 2, and 3 refer to the desired scheduling window with
a length of 1, 2, and 3 h, respectively.
The background area in Fig. 16.2 illustrates the decision boundaries determined
by the KNN algorithm in the training process, and the data points refer to the test set
data. It can be noted that the KNN algorithm has better efficiency in predicting label
2 in comparison with the labels 1 and 3; one possible reason for this outcome is the
frequency of occurrence for label 2. Label 2 has the highest frequency of the
occurrence, which means the desired scheduling window with a length of 2 h is
the dominant trend in using the washing machine; consequently, the algorithm can
learn and predict this trend with higher accuracy.
Considering the results, it can be concluded that the proposed approach can be
used for predicting the desired length of the scheduling window for residential load
scheduling problems. The main contribution of this approach is in using machine
learning algorithms for learning the hidden relationship between factors such as the
start time of the task, day of the week and outdoor temperature, and the length of the
desired scheduling window. Although this study only reports the result of applying
the proposed approach for the washing machine as a highly used flexible home
appliance, the proposed concept can be used for other flexible appliances such as
dryer, dishwasher, etc. The proposed approach can be used in real-time applications
where the smart applicants can report the start command by the user to the central
352 M. Sadat-Mohammadi et al.

Fig. 16.2 Performance of the proposed approach

home energy management system. Then, the trained model with historical data can
predict the desired length of the scheduling window. The predicted parameter can be
fed into the optimal load schedule model to determine the optimal operation interval
of the appliance, considering the price of the electricity and operational constraints.
Finally, the on/off command can be sent to the smart appliance to finish its operation
cycle within the defined time interval. Consequently, the proposed approach can
make the implementation and operation of the home energy management systems
less demanding and nonintrusive in residential sectors by reducing the amount of the
data which should be provided to this system at the beginning of the scheduling
process.
Despite the effectiveness of the proposed approach in predicting the length of the
scheduling window, some challenges should be addressed properly to make this
approach applicable in real systems. The training set should be dynamically updated
after each operation of the appliances to record any change in the lifestyle of the
residents. The model may be retrained in specific time intervals with the updated
train set to keep the trained model updated. However, there may be another arising
issue with including the new observation in the training set, such as a reduction in the
accuracy of prediction. If there is a change in the lifestyle of the residents, some data
of the training set will be related to the previous preferences of the residents, while
the updated data will represent the current lifestyle. Consequently, the accuracy of
the prediction will be decreased as there will be a weak relationship between input
and output if the whole training set is fed to the model. Therefore, while adding new
observations to the training set, some historical data may be removed to increase the
accuracy of the prediction.
16 Application of Machine Learning for Predicting User Preferences in Optimal. . . 353

Moreover, future studies may identify the important factors which affect the
length of the scheduling window of the other flexible appliances such as dishwasher,
dryer, etc. Also, possible approaches in extracting the desired scheduling window of
the flexible appliances may be investigated for labeling the input data. The applica-
tion of wearable devices in providing a comprehensive dataset for implementing this
approach may be studied in the future. Nowadays, wearable devices such as
smartwatches are pervasively used by people. In addition to the health applications
of these devices, they can provide data about the lifestyle of the residents and their
location, which can be used for identifying the residents’ preferences besides the
data supplied by smart appliances and stationary sensors installed in houses for
monitoring the temperature, humidity, etc. Considering more features and increasing
the dimension of the input dataset can result in higher accuracy if the relevant
features are selected and used for training the machine learning algorithms.

16.5 Conclusion

This study proposes an approach to predict the length of the desired scheduling
window of flexible smart appliances. Smart appliances can collect the energy
consumption profile and receive on/off command from external sources such as
home energy management systems. The proposed approach was implanted on a
dataset to predict the desired scheduling window of the washing machine in a
residential building. The start time of the task, day of the week, and temperature
are introduced as features in learning and predicting the length of the desired
scheduling window. Three machine learning algorithms, ANN, RF, and SVM,
were implemented on the labeled data to analyze the accuracy of the prediction.
The results showed that KNN outperforms other algorithms by achieving 77.6%
accuracy in predicting the desired scheduling windows with the length of 1, 2, and
3 h. Although there were scheduling windows with lengths of higher than 3 h, their
frequency of occurrence was low. Consequently, they cannot represent the common
trend in using the washing machine and cannot be predicted. The results demon-
strated the possibility of predicting the desired scheduling window of the flexible
appliance that reduces the difficulty of using home energy management systems in
the residential sector. Including the proposed approach in the home energy manage-
ment systems will eliminate the need for updating the scheduling window of each
flexible appliance by the user at the beginning of the scheduling process. Future
studies can be followed in three directions: (1) increasing the accuracy of the
prediction by introducing and including other important factors besides what men-
tioned in this study; (2) implementing the proposed approach on a complete set of the
flexible loads and charger of the plug-in hybrid electric vehicle and analyzing the
results of prediction for all appliances; and (3) integrating the proposed approach in
the residential load scheduling algorithms to make scheduling process less demand-
ing for users.
354 M. Sadat-Mohammadi et al.

References

1. S. Althaher, P. Mancarella, J. Mutale, Automated demand response from home energy man-
agement system under dynamic pricing and power and comfort constraints. IEEE Trans. Smart
Grid 6(4), 1874–1883 (2015)
2. B. Zhou et al., Smart home energy management systems: concept, configurations, and sched-
uling strategies. Renew. Sust. Energ. Rev. 61, 30–40 (2016)
3. A. Anvari-Moghaddam, H. Monsef, A. Rahimi-Kian, Optimal smart home energy management
considering energy saving and a comfortable lifestyle. IEEE Trans. Smart Grid 6(1), 324–332
(2014)
4. F. Baig, A. Mahmood, N. Javaid, S. Razzaq, N. Khan, Z. Saleem, Smart home energy
management system for monitoring and scheduling of home appliances using
zigbee. J. Basic. Appl. Sci. Res 3(5), 880–891 (2013)
5. J. Han, C.-S. Choi, W.-K. Park, I. Lee, S.-H. Kim, Smart home energy management system
including renewable energy based on ZigBee and PLC. IEEE Trans. Consum. Electron. 60(2),
198–202 (2014)
6. M. Daneshvar, M. Pesaran, B. Mohammadi-Ivatloo, 7 - Transactive energy in future smart
homes, in The Energy Internet, ed. by W. Su, A. Q. Huang, (Woodhead Publishing, Sawston,
2019), pp. 153–179
7. M.S. Ahmed, A. Mohamed, T. Khatib, H. Shareef, R.Z. Homod, J. Abd Ali, Real time optimal
schedule controller for home energy management system using new binary backtracking search
algorithm. Energ. Buildings 138, 215–227 (2017)
8. A.-H.H. Mohsenian-Rad, A. Leon-Garcia, Optimal residential load control with price prediction
in real-time electricity pricing environments. IEEE Trans. Smart Grid 1(2), 120–133 (2010)
9. E. Nazerfard, P. Rashidi, D.J. Cook, Discovering temporal features and relations of activity
patterns, in Proceedings - IEEE International Conference on Data Mining, ICDM, (2010),
pp. 1069–1075
10. A. Yassine, S. Singh, A. Alamri, Mining human activity patterns from smart home big data for
healthcare applications. IEEE Access 99, 1 (2017)
11. S. Chen, F. Gao, X. Guan, T. Liu, Y. Che, Y. Liu, A residential load scheduling approach based
on load behavior analysis, in 2014 IEEE International Conference on Automation Science and
Engineering (CASE), (2014), pp. 954–959
12. S. Chen et al., She: smart home energy management system based on social and motion
behavior cognition, in 2015 IEEE International Conference on Smart Grid Communications
(SmartGridComm), (2015), pp. 859–864
13. A. Moradzadeh, O. Sadeghian, K. Pourhossein, B. Mohammadi-Ivatloo, A. Anvari-
Moghaddam, Improving residential load disaggregation for sustainable development of energy
via principal component analysis. Sustainability 12(8), 3158 (2020)
14. M. Sadat-Mohammadi, M. Nazari-Heris, E. Nazerfard, M. Abedi, S. Asadi, H. Jebelli, Intelli-
gent approach for residential load scheduling. IET Gener. Transm. Distrib. 14(21), 4738 (2020)
15. S. Lee, D. Whaley, W. Saman, Electricity demand profile of Australian low energy houses.
Energy Procedia 62(4), 91–100 (2014)
16. M.A. Nielsen, Neural Networks and Deep Learning, vol 2018 (Determination Press, San
Francisco, 2015)
17. S. Agatonovic-Kustrin, R. Beresford, Basic concepts of artificial neural network (ANN)
modeling and its application in pharmaceutical research. J. Pharm. Biomed. Anal. 22(5),
717–727 (2000)
16 Application of Machine Learning for Predicting User Preferences in Optimal. . . 355

18. G. Biau, E. Scornet, A random forest guided tour. TEST 25(2), 197–227 (2016)
19. J. Xia, P. Ghamisi, N. Yokoya, A. Iwasaki, Random forest ensembles and extended
multiextinction profiles for hyperspectral image classification. IEEE Trans. Geosci. Remote
Sens. 56(1), 202–216 (2018)
20. S. Makonin, B. Ellert, I.V. Bajić, F. Popowich, Electricity, water, and natural gas consumption
of a residential house in Canada from 2012 to 2014. Sci. Data 3(1), 1–12 (2016)
21. S. Karamizadeh, S.M. Abdullah, A.A. Manaf, M. Zamani, A. Hooman, An overview of
principal component analysis. J. Signal Inf. Process. 4(3B), 173 (2013)
Chapter 17
Machine Learning Approaches in a Real
Power System and Power Markets

Peyman Razmi and Mahdi Ghaemi Asl

17.1 Introduction

An electric power grid is a network of electrical elements employed to supply,

transfer, and utilize electric power that are often referred to as the largest and the
most complex industry in the world. A model of a power system is the network
which feeds a region’s considerable areas such as houses and industry. The power
system is widely divided into the generators (feeds the power), the transmission
system (transfers the power from the generating stations to the load hubs), and the
distribution system (supplies the power to close houses and industries). Thus, the
mixture of these parts that work systematically and seamlessly together is known as
power grid. Small power grids are also existing in industry, commercial buildings,
hospitals, and houses. Electrical energy is not able to be kept in large quantities and
therefore should be utilized the moment it is generated, so there should always be an
electrical connection between the generator and the load. This means that generation,
transmission, and distribution parts of power system should carry out their work with
complete consistency and safety. To maintain the safety and continuity of the
system, these parts need to be utilized and operated in an optimal fashion. Here,
operating means monitoring, controlling, and coordinating all components of the
power system, and it is a task that is heavily influenced by quality of human
resources (operator) and equipment.
But the growing demand for electricity and the advent of concepts such as
microgrids have added to complexity and nonlinearity of power systems, which in

P. Razmi (*)
Faculty of Electrical Engineering, University of Ferdowsi, Mashhad, Iran
M. Ghaemi Asl
Faculty of Economics, Kharazmi University, Tehran, Iran
e-mail: [email protected]

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 357
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1_17
358 P. Razmi and M. Ghaemi Asl

turn has caused the conventional and inflexible method of system safety to become
ineffective in a broad array of performing points. Power system is a dynamic system
with a wide range of different working conditions and uncertainties, which make it
susceptible to a wide range of problems that could emerge at any moment. The
quality of the stability is one of the main issues in the power system. Each power
system that after an interruption fails to return to its pre-fault operating condition
(or a close proximate of that condition) is called an unstable power system. In such
circumstances, the interconnected power system either loses its integration and starts
to operate as a segmented system or fails to supply the electrical energy from the
generator to the point of consumption. It is clear that operating such a complex
system needs advanced studies and extensively developed computer programs and
algorithms to be used across planning, control, and protection cycles for technical
investigation of safety as well as economic viability of electricity generation and
supply mechanism. In addition to stability, power systems also have a number of
other uncertainties which include:
1. Change in the system’s functional conditions: change in the magnitude of load or
output of power system.
2. Faults in the power system: type, location, and time of the faults can never be
predicted.
3. Uncertainties originating from the present-day problems of the power system,
e.g., electricity marketing and deregulation issues.
4. Inaccurate information of operator about control, function, management, and
planning of power systems.
Many articles have investigated the applications of ML in power systems, and
some have achieved very desirable results. The aim of this paper is to review the
works accomplished in the context of ML within the framework of the power
systems and express the importance and possible roles of ML in power networks.

17.2 Control-Themed Studies

The majority of studies that devoted to control-themed applications of ML in power

systems are emphasized on model of power system stabilizers (PSSs), control of
synchronous machine (SM), power system frequency, and flexible alternating cur-
rent transmission system (FACTS) components. On this basis, the concepts and
applications of ML in dealing with the uncertainties of power systems can be
classified into four major categories as shown in Table 17.1. With automatic voltage
regulators (AVRs) approaching in late 1950s, establishment of AVRs on power
generation units turns into a usual procedure [1]. Regrettably, the high performance
of AVRs brought about instability in the power system. The majority of the
challenges are related to the low-frequency fluctuation in interconnected power
systems, particularly in the deregulated models. Small magnitude and
low-frequency fluctuation usually remained a problem for a prolonged time. To
17 Machine Learning Approaches in a Real Power System and Power Markets 359

Table 17.1 Four major categories of machine learning application in power system
Control Detection Function Optimization
Design of PSS Detection of fault’s exact location Evaluation and Distribution
stability system
Control of synchro- Detection type of fault in power Load prediction Production
nous machine system location State estimation planning
Control of AGG System monitoring Load Reliability
Control of FACTS Protection of transmission lines distribution System
elements protection

enhance the dynamic performance of the system by providing a quick damping, an

extra control sign in excitation system or/and the governor system of a generation
block could be utilized. Being the most economical damping controller, PSS has
been extensively used to restrain the low-frequency fluctuation and improve the
system dynamic stability. PSSs provide a reliable performance to retain the power
system’s stability by adding a supplementary signal to the excitation system [1].
Over the recent years, artificial neural network controllers (ANNCs) [2] and fuzzy
logic controllers (FLCs) [3] have been applied and examined as PSSs. In contrast to
other conventional control approaches, ANNCs and FLCs are model-free control-
lers, i.e., they do not need an exact mathematical design of the controlled system. In
addition, their rapidity and strength are the most important features compared with
the other conventional methods. Since these controllers applied to the Fuzzy PSS
design, the laws and the participation function of the controller are tuned indepen-
dently, causing the design process to be arduous and time-consuming. Regarding
ANNCs, they have the ability to learn and adapt, but they perform like a “black box,”
and it is complicated to comprehend the action of the network. The examination has
showed that the adequate operation during the system upsets is connected with the
proper selection of the conventional power system stabilizer (CPSS) parameters. The
PSS parameters’ tuning difficulty is one of the optimization challenges in power
system stability. The artificial neural network (ANN), when adequately trained,
could be utilized as a controller in place of the CPSS.
Most of the ANN functions in the power systems utilize a multilayer feed-forward
network. ANNs are practical tools exerted for numerous years for recognition and
control of intricate systems because of their nonlinear mapping features. To attain the
finest operation, the ANN should be prepared for various operating circumstances to
adapt the CPSS parameters [1]. The learning strategies lead to interference by the
typical back propagation network under diverse circumstances. To solve this hin-
drance, a modular ANN was introduced instead of a back propagation network
[4]. This design comprises three local proficient networks and one alone gate
network which has three layers for each one of them. The ANN was disciplined
straightly through the output and input of the CPSS. The model outcomes validated
that the modular PSS was more capable of damping oscillations and bringing about
proper performances. To advance a neural adaptive PSS, a feed-forward neural
network with a single hidden layer was examined in [5–7]. The recommended neural
adaptive PSS consist of two sub-networks: adaptive neuro-identifier and adaptive
360 P. Razmi and M. Ghaemi Asl

neuro-controller which tracked the dynamic features of the plant and damp the
low-frequency oscillations, respectively. These two sub-networks were worked out
in an online mode using the back propagation approach.
An innovative method for online adaptive tuning of PSS parameters employing
radial basis function neural network (RBFNN), was cited by Abido and Abdel-
Magid [8]. The suggested RBFN was operated over a broad range of circumstances
and system parameter alterations for readjusting PSS parameters online on the basis
of real-time calibrations of machine loading situations. The mentioned RBFN-based
PSS was studied for diverse operating circumstances and system parameter alterna-
tions. To enhance the ephemeral stability of power systems, a recurrent neural
network (RNN) stabilization controller was offered in [9]. The suggested technique
was employed for both the AVR and governor. The weights of introduced controller
were modified online. For excitation control, the signal output of the first RNN was
added to the PSS signal output. In addition, to have a stabilization signal for the
governor unit, the signal output of the second RNN was utilized. To enhance the
damping performance, a probabilistic PSS was presented by Ping et al. employing a
single-neuron model [10]. The alteration of system performing conditions could be
recorded by adding ancillary self-adjusted gains to PSSs. In an analogous paper,
adaptive critics and ANNs were offered to plan a real-time digital signal processor
(DSP) to accomplish an optimal wide-area control system (WACS) for a power
system [11].
In 1964, fuzzy logic (FL) was developed by Lotfi Zadeh to give an attention to
inaccuracy and uncertainty that usually occur in engineering problems [12]. In the
mentioned method, FL is able to be utilized as a general technique to integrate
knowledge, heuristics, or theory into controllers and decision-makers [1]. The pros
are as follows: (1) precise description of the functional restrictions of the power
systems; and (2) fuzzified restrictions will be softer than traditional limitations
[13]. In 1979, for the first time FL was presented to solve power system conundrums.
To make classic controllers, it is prerequisite to linearize nonlinear systems. As a
result, control regulations are determined on the basis of the new design. These kinds
of controllers are used to monitor the system. Fuzzy logic controllers (FLCs) are
nonlinear, do not require controlled device models, and are not delicate to changes of
the device parameters. Reference [14] proposed a self-constructing fuzzy neural
network (SCFNN)-based static synchronous series control (SSSC) to soften the
inter-area fluctuations in interconnected power grids that consist of an online
equipped fuzzy neural network (FNN) controller with adaptive learning rates
(ALRs) and self-constructing structure. The Lyapunov paradigm is used to gain
the adaptive learning rates. Reference [15] offers fuzzy particle swarm optimization
of PID controller (PSO-FPIDC) utilized as a CPSS to enhance the dynamic stability
efficiency of generating unit during low-frequency fluctuations. Reference [16]
presents an adaptive fuzzy sliding-mode controller (AFSMC) with a PI switching
surface to damp power system fluctuations. A wavelet neural network (WNN) and
sliding mode as a monittor unit can be used to surmount the complications of
designing a sliding-mode controller, which are the conjecture of known uncertainty
restricts and the chattering incident in the control effort. In [17] the author proposed
17 Machine Learning Approaches in a Real Power System and Power Markets 361

an adaptive fuzzy logic PSS (AFLPSS) model to damp electromechanical forms and
improving the first-swing synchronized stability margins. The plan was to arrange a
multi-zonal PID formation and fuzzy logic variable gain to make the damping
operation more optimal. Ramirez and Malik introduced a self-tuned FLPSS to
improve the damping of power grid fluctuations [18]. The model of self-tuned
FLPSS was consisted of a simple FL controller which has a considerably simple
architecture and few tuning parameters which was implemented through a basic
control algorithm. The fuzzy tuner was utilized to nonlinearly and online adjust the
reactivity of the simple FL controller to its input variable, which obliquely alters the
relative sensitivity of areas of related input membership functions.
The tuning of traditional power system stabilizer for single-machine infinite bus
power system is presented in [19]. With applying the bat algorithm (BA) an
eigenvalue-based objective function, the author optimized the time and gain con-
stants (pole and zero) of CPSS. The speed feedback of the generators for the single-
machine infinite bus system (SMIB) excluding PSS was presented in [20]. With
PSO-CPSS and employing the BA-based advanced CPSS, power grid schemes are
examined in various operation circumstances. It is revealed the feedback with
PSO-CPSS is able to stabilize the grid during unwanted situations. But by using
continued settling time over decreased settling time and applying BA-CPSS can
stabilize the system for whole plants situations. The admirable work of BA-CPSS is
more demonstrated with utilizing performance indexes as the value is least in
comparison to the grid with PSO-CPSS and excluding PSS. The little signal stability
is assured with applying BA-CPSS since all electromechanical mode eigenvalue (for
133 plants) is shifted to left-hand side (LHS) of s-plane under the D-shape sector. On
the other hand, by using PSO-CPSS, only nine plant situations are confirmed to be
stable.
In [21], an intelligent power system stabilizer on the basis of a new system centric
control device is presented. The presented approach utilized hybrid structure of two
algorithms: the first one is a neural network (NN)-based control device with specific
neuro-identifier, and the second one is an adaptive control device employed as a
model reference adaptive controller (MRAC). An identifier is utilized to estimate the
nonlinear functional dynamics of the electrical grid. The MRAC control device
modifies when electrical grid (plant) parametric set varies. The notable uniqueness
and benefit of the presented model are the controller’s capability to supplement one
another during functional and parametric uncertainty and develop in the case of
unstable system dynamics. Work [22] demonstrates an improvement in a smart
neighborhood signs on the basis of a control device in order to damp
low-frequency fluctuations in electrical grids. The control device is developed
off-line for functioning adequately during a broad range of electrical grids function-
ing sites, permitting it to manage the intricate, haphazard, and time-varying charac-
teristic of electrical grids. NN-based unite recognition erases the demand for
developing precise designs through primary rules for control model, culminating a
methodology which is thoroughly drive by data.
A unified control device was replaced by the traditional PSS, and AVR for
adjusting the generators’ terminal voltage and lessening the power system
362 P. Razmi and M. Ghaemi Asl

fluctuations was presented in [23]. This brilliant control device is termed as online
trained self-recurrent wavelet neural network controller (OTSRWNNC). For
attaining the previous purposes, two control faults were decreased at the same time
with modifying the characteristic of OTSRWNNC. Since the mentioned control
device has great capability to learn, it doesn’t need an identifier for estimating the
dynamic of monitored electrical grid. In [24], a model for a grid with a real-time
closed-loop wide-area decentralized power system stabilizer (WD-PSSs) is exam-
ined. For this purpose, real-time wide-area measurement information are refined and
used to plan a series of stability factors on the basis of a reinforcement learning
(RL) approach. Current technical developments in wide-area measurement system
(WAMS) cause the ability of using the system-wide signals in constructing electrical
grid control devices. The primary purpose of such control devices is to maintain the
grid after serious disruptions and mitigate the fluctuations afterward.
An adaptive neuro-fuzzy inference system (ANFIS) approach on the basis of
ANN for planning a SSSC-based control device to enhance the transient stability
was demonstrated in [25]. The offered ANFIS control device merges the benefits of a
fuzzy control device along with the fast feedback and flexibility characteristic of
ANN. ANFIS design was trained by utilizing the created database from fuzzy control
device in SSSC. It’s illustrated that the introduced SSSC control device enhances the
system’s voltage profile markedly during serious disruptions.
ML algorithms have been lately deemed as a method for monitoring SMs.
Authors in [26] introduce a function of fuzzy logic for controlling a SM’s speed.
In accordance with the examination of the SM transient feedback and fuzzy logic,
the fuzzy control device was advanced. The fuzzy control device produces the
alterations of the reference current vector of the SM speed control on the basis of
the speed error and its variation. Authors in [27] examine a wide practical range of
the synchronous generator (SG), its intricate dynamics, transient operation, non-
linearities, and an altering unit formation. Hence this generator is not able to exactly
be designed as a linear device; the mentioned work offers a new optimal neuro-
controller which substitutes the conventional controller (CONVC). A combination
of the AVR and turbine governer can be used, for controlling a SG in an electrical
grid by utilizing a multilayer perceptron neural network (MLPN) and a RBFNN.
In the last years, developments in magnetic materials, semiconductor power
devices, and control theories caused permanent magnet synchronous motor
(PMSM) drives to perform significant role in motion-control applications
[28]. They own admirable characteristics, namely, small size, immense torque to
weight ratio, as well as lack of rotor losses. Moreover, higher power of the developed
magnetic materials causes broader applications of PMSM [29–32]. In Reference [28]
a strong speed control approach of a PMSM drive employing an adaptive neural
network model following speed controller (ANNMFC) is presented. The robust
speed control device is a compound of a feedback neural network controller
(NNC) along with an online trained NN model following controller (NNMFC) to
enhance the dynamic operation of the drive unit. Outputs of NNMFC are added to
the NN speed control device output to offset the difference among the reference
design and the PMSM drive unit output during load disruptions as well as parameters
17 Machine Learning Approaches in a Real Power System and Power Markets 363

Fig. 17.1 The topological structure of the RBFNN

uncertainties. A speed control device for PMSM during the field-oriented control
(FOC) approach is explained in [33]. A new adaptive neuro-control design, single
artificial neuron goal representation heuristic dynamic programming (SANGRHDP)
for speed adjusting of PMSMs, is discussed. Paper [34] proposes a smart speed
control device for the interior PMSM, in accordance with a single artificial neuron.
Reference [35] presents a new decoupling paradigm for a bearing-less PMSM
(BPMSM) to obtain quick response and high accuracy performances. The mentioned
control design includes the NN inverse (NNI) technique and two-degree-of-freedom
(DOF) internal design control devices. With presenting the NNI schemes into the
original BPMSM structure, a decoupled pseudo-linear arrangement could be formed.
In [36] a new internal model control (IMC) method on the basis of the back
propagation neural network inverse (BPNNI) control plan was suggested for the
purpose of effectively decoupling the PMIWM. Reference [37] proposes a new
nonlinear decoupling monitor plan for a permanent magnet in-wheel motor
(PMIWM), in which the radial basis function neural network inverse (RBFNNI)
and the state feedback robust pole placement (RPP) are both utilized. The offered
NNI is settled on the left side of the PMIWM shown in Fig. 17.1, which gives rise to
a pseudo-linear scheme for the system decoupling. The NNI control method is able
to enhance the robustness and reject the disruption.

17.3 Detection and Protection-Themed Studies

To sustain the efficiency and reliability of the power systems, it is undoubtedly

essential for the TL faults to be identified and settled in a reliable and precise
approach [38]. Furthermore, by noticeable growing of smart grids, the significance
of designing a smart fault controlling and diagnosis unit is to classify and locate
various kinds of errors is indisputable [39]. Procedures for error detecting, catego-
rizing, and locating in TLs and DSs have been extensively researched during the
recent years. A novel method on the basis of generalized neural network (GNN) with
364 P. Razmi and M. Ghaemi Asl

wavelet transform (WT) is proposed in [38] for fault locating evaluation. Acquired
qualities are employed as an input to the GNN design for approximating the zone of
the error in a provided TL. The differences between the results attained from GNN
scheme are examined with ANN which leads to establish mathematical paradigms
with more accuracy [40].
Advance TL fault location methods consists of stationary WT (SWT), determi-
nant function feature (DFF), support vector machine (SVM), and support vector
regression (SVR) which are discussed in [41]. Different kinds of errors at various
zones such as fault impedance and fault inception angles on a 400 kV, 361.297 km
TL are examined [41]. The system only uses single-end measurements. DFF is
utilized to separate specific fault characteristics from 1/4 cycle of post-fault signals
after noise, and the declining DC offset has been wiped out by filtering method on
the basis of SWT. Afterward, a classifier (SVM) and a regression (SVR) approaches
are prepared with characteristics attained from DFF, which is employed at the fault
zone on the TL. The outcomes of the discussed paper illustrate that fault zone on TLs
is able to be detected quickly and accurately regardless of fault impedance. Refer-
ence [42] presents k-nearest neighbor (k-NN)-based approach for fault zone estima-
tion of all kinds of errors in parallel lines with utilizing one-terminal measurement.
Discrete Fourier transform (DFT) is applied in the mentioned work for pre-analyzing
the signals. After that, the standard alteration of one cycle of pre-fault and one cycle
of post-fault examples is utilized as inputs to k-NN algorithm. Reference [43]
mentions the significance of fault location within the context of smart grids and
microgrids that can be investigated with all the capabilities of this new paradigm.
This reference proposes a multi-stage technique that consists of decision trees (DT)
and ANNs for fault location aims. Authors in [44] discuss the significance of ANNs
for fault location and compare it with other soft computing methods such as fuzzy
logic technique, wavelet approach, SVM, and other traditional procedures.
In [45], a unified framework with combination of error categorization and
locating is presented by employing a novel ML algorithm: the summation wavelet
extreme learning machine (SW-ELM) which incorporates quality derivation during
the learning procedure. Another essential subject requires further analysis in the
transmission and DS in categorizing and detecting the fault’s type. Fault designation
is important for dependable and quick response protective relays followed by digital
distance protection. Categorization of power system errors is the initial step for
enhancing power quality and guaranteeing the system protection. For this reason,
equipping a unique method in identifying the class of the faults is needed. Reference
[46] offers a unique plain and efficient approach for faulty feeder detecting in
resonant grounding DSs in accordance with the continuous WT (CWT) and
convolutional NN (CNN). At the same time, the characteristics derivation for error
signals and the defective feeder detection are performed by the instructed CNN.
In [47], categorizing the electrical grid errors by utilizing the empirical mode
decomposition (EMD) and SVMs is discussed. EMD is utilized to decompose the
TL voltages into intrinsic mode functions (IMFs). A multiple SVM design is
presented to classify the fault status in a group of ten power system faults. Reference
[48] introduces an approach employing discrete WT (DWT) and SVM to identify
17 Machine Learning Approaches in a Real Power System and Power Markets 365

and categorize faults through a TL. DWT is applied to find the high-frequency
components within a fault signal range. In this paper, various faults (short circuits)
on disparate positions of a transmission line are formed and simulated. In [49], a
brilliant method for fault categorization in a TL is presented. Ten various types of
faults (LAG, LBG, LCG, LABG, LBCG, LCAG, LAB, LBC, LCA, and LABC)
have been studied as one normal status on a simulated TL system. Post-fault current
signs have been utilized for characteristic derivation for additional research. EMD
approach is employed for decomposing post-fault current signals into IMFs. These
IMFs are utilized as input variables to an ANN on the basis of intelligent fault
classification method. Reference [50] presents a novel fault identifier algorithm for
photovoltaic (PV) systems in accordance with ANN and fuzzy logic system inter-
face. Considering the small number of examples of ML methods used in fault
detecting algorithms in PV systems, the major focus of the mentioned paper is to
design a system with an ability to find possible faults in PV systems by utilizing
radial basis function (RBF), ANN, and both Mamdani and Sugeno fuzzy logic
systems interface.
A new fault detection and categorization method for high-voltage DC transmis-
sion lines employing K-nearest neighboring was offered in [51]. In [52], problems of
monitored learning, namely, necessity for historical information and incapability of
classification of further errors precisely, are solved by a novel method utilizing
unmonitored learning for quick execution of predictive maintaining activity. This
method consists of error kind and forecasting for familiar and unfamiliar errors by
employing density approximation through Gaussian method. Combination of design
clustering and K-means algorithm is employed and compared their outcomes with an
actual case vibration information. A new approach to trace the ephemeral errors and
to categorize the error kind in DSs employing WTs and ANFIS has been evolved in
[53]. It implements on developed methods of signal processing in accordance with
WTs, utilizing data sampled from the main feeder current to collect major features
and dynamic qualities of the fault signal. In [40] a fault categorization method on the
basis of Haar-WT (HWT) and ANN is introduced for six-phase TL in opposition to
phase-to-phase faults. Reference [54] attempts to describe four multiwavelet packet
entropies to elicit the qualities of various TL faults and employs RBFNN to detect
and categorize ten fault types of transmission lines. In [55] DWT is utilized to elicit
transient data from the measured voltages. Then SVM classifiers are used to detect
the faulty section and faulty half. Moreover, writers in [56, 57] employed charac-
teristics elicited by DWT as inputs to SVMs. Authors in [58, 59] applied principal
component analysis (PCA) to diminish the dimensionality of the wavelet coefficients
in advance of sending the coefficients to the SVMs for fault-type categorization.
In [60] a novel categorization approach on the basis of NN is presented to
decrease the training time and dimensions of an ANN. Employing the mentioned
technique, high precision of fault categorization is obtained. In the mentioned
reference, primary elements of pre-fault and post-fault as well as positive sequence
elements of currents and voltages of three phases have been applied as inputs to
presented ANN. The output of the ANN is the approximated fault zone. Reference
[61] proposes a novel fault categorization method for high-speed relaying utilizing
366 P. Razmi and M. Ghaemi Asl

minimal RBFNN. This novel method reduces the training time noticeably and
presents a systematic framework for choosing the number of neurons in the hidden
layer. In addition, the minimal radial basis function network provides an exact error
kind categorization of a TL even in the existence of high fault impedance in the fault
line. The suggested method in [62] is a compound of a preprocessing block in
accordance with DWT and probabilistic neural network (PNN). The DWT operates
as extractor of distinct qualities in the input current signal, which are accumulated at
source end. The data is then fed into PNN to categorize the faults. Analysis of the
possible and obscure faults within the electric transformers is the means of
guaranteeing a steady power supply for costumers [63].
SVM is a modern ML approach on the basis of statistical learning theory, which is
a robust mechanism to overcome the challenges of nonlinearity, small sampling, and
high dimension. Selecting the SVM characteristics has a notable effect on the
categorization exactness of SVM. In [63] SVM with genetic algorithm (SVMG) is
employed to fault analysis of a power transformer, in which genetic algorithm (GA)
is applied for selecting the proper free characteristics of SVM. Empirical outcomes
illustrate that the SVMG approach is able to attain higher diagnostic precision than
International Electrotechnical Commission (IEC) three ratios, normal SVM arranger,
and ANN. Paper [64] proposes a brilliant fault categorization method to power
transformer dissolved gas analysis (DGA), handling highly adaptable or noise-
corrupted information. Bootstrap and genetic programming (GP) are applied to
enhance the interpretation precision for DGA of power transformers. Bootstrap
preprocessing is used to roughly adjust the sample numbers for various fault classes
to enhance subsequent fault categorization with GP aspect extraction. GP is utilized
to set up categorization qualities for each class in accordance with the collected gas
information. The qualities elicited from GP are then applied to the inputs of ANN,
SVM, and K-nearest neighbor (KNN) arrangers for error categorization. The cate-
gorization exactness of the mixed GP-ANN, GP-SVM, and GP-KNN arrangers is
compared sequentially with the ones extracted from ANN, SVM, and KNN
arrangers [65]. presents a novel differential protection design in accordance with
the SVM, which yields efficient distinction between internal faults in a power
transformer with the other disruptions, namely, different types of over excitation
and inrush currents states. The characteristic derivation is accomplished employing
WT, which will be given as input to the SVM classifier.
A novel approach to diagnose the transformer fault on the basis of relevance
vector machine (RVM) is presented in [66]. A Bayesian estimator is implemented to
SVM in the unique algorithm, which made the performance of the fault analysis
system more efficient. In paper [66], a diagnosis paradigm is offered in which the
resolutions of RVM have the characteristics of sparseness and RVM is able to
achieve global resolutions under confined samples. The outcomes proved that this
approach has discernible benefits of diagnosis time and precision in comparison with
back propagation (BP) NNs and general SVM models. Reference [67] presents a
fault direction separator that utilizes an ANN for protecting TLs. The separator
employs different features to achieve a decision and tends to imitate the traditional
pattern categorization challenge. In reference [67] an equation of the boundary
17 Machine Learning Approaches in a Real Power System and Power Markets 367

explaining the categorization is embedded in the multilayer feed-forward neural

network (MFNN) by training and employing a proper learning algorithm. The
separator utilizes immediate values of the line voltages and line currents to make
decisions. Consequences of the ANN-based separator operation in [67] illustrates
that it is quick, powerful, and precise. Reference [68] offers an anomaly detection
method utilizing ANNs and the WT for the condition monitoring of wind turbines. In
this method, nonlinear autoregressive NNs are applied to evaluate the temperature
signals of the gearbox.
Author in [69] proposes an ANN-based method control the voltage stability
margin (VSM) in electrical grids online. In Reference [70] ANN-based approach
is modified for rapidly approximating the long-term VSM. The applied examinations
in the mentioned work demonstrated that phase angles and node voltage magnitudes
are the most excellent forecasting systems of VSM. Moreover, the mentioned article
illustrates that the presented ANN-based approach is able to effectively evaluate the
VSM under normal situations and under N-1 contingency conditions as well. The
distance relays which are used to protect the TLs are commonly planned based on set
settings [71]. The operation of these relays will be affected by any alterations to
network situation. Appling a pattern discriminator for the power system diagnosis is
able to make major improvements in the protection field. Paper [71] illustrates the
utilization of an ANN as a model discriminator for a distance relay operation. The
method uses the magnitudes of three phase voltage and current phasors as inputs. In
Reference [72] authors discuss the potential ways to improve the digital power
transformer preservation. Establishing the inrush in power transformers is turning
into unreliability in existent numerical preservation. An ANN is implemented to
inrush sensing. Unless with an appropriate dimensioning, the saturation of protective
current transformers (CT) could not be completely eradicated. ANN has been
utilized to reconstruct the deformed secondary CT currents as a result of saturation.
In each case, an ANN has been provided in the preservation scheme as a develop-
ment of the existent designs, which enhanced the preservation system’s
trustworthiness.

17.4 Function and Protection-Themed Studies

Electricity carries a significant value for national economies to grow. As a result,

electric services attempt to provide a balance between power generation and con-
sumption with intention of granting a satisfying service at a reasonable price
[72]. Consequently, these services require electric load prediction at the most precise
level. Since electric load is influenced by numerous factors (such as day of the week,
month of the year, etc.), load predicting becomes a very involved process which
demands something greater than statistical approaches. Authors in [72] offer a load
prediction structural paradigm on the basis of an ANN that operates short-term load
forecasting (STLF). Reference [72] illustrates significant outcomes as well as a plain,
efficient design. In [73] authors offer a hybrid scheme that consists of SVR, RBFNN,
368 P. Razmi and M. Ghaemi Asl

Fig. 17.2 The architecture

of a SVMG model

and dual extended Kalman filter (DEKF) to create a forecast method (SVR-DEKF-
RBFNN) for short-term load prediction. In the presented design, at first the SVR
method is applied to ascertain the form and primary characteristics of the RBFNN.
After initializing, the DEKF is utilized as the learning method to advance the features
of the RBFNN. At last, the optimized RBFNN method is applied to forecast short-
term load. In [74] authors tackle the trouble of forecasting hourly load demand by
employing adaptive ANNs. A PSO algorithm is used to adapt the network’s weights
in the training phase of the ANNs.
In Reference [75], recurrent support vector machines with genetic algorithms
(RSVMG) are presented to predict electricity load. Moreover, GAs are applied to
identify free characteristics of SVMs. The experimental outcomes from Reference
[75] illustrate that RSVMG paradigm offers an auspicious alternative for predicting
electricity load in power system. Figure 17.2 illustrates the structure of the presented
SVMG design. GAs are applied to provide a smaller mean absolute percentage error
(MAPE) by probing for better mixture of three parameters in SVMs. Authors in
Reference [76] proposed a model which combines WT, adaptive GA, and fuzzy
system with GNN and employed to predict the short-term weekday electrical load. In
[77], a new hybrid developmental fuzzy design with characteristic advancement is
presented. Considering the fact that determining optimal values for the fuzzy laws
and weights is a very involved work, the process of optimization is undertook by a
bio-inspired optimizer, named group evolutionary strategy (GES), which develops
from a mixture between two heuristic methods, namely, the evolution strategies and
the greedy randomized adaptive search procedure (GRASP). Valid information from
electric services obtained from the literature is utilized to justify the presented
approach. In Reference [78], V. L. Paucar utilized multilayer perceptron NNs trained
with the second order Levenberg/Marquardt design to compute voltage magnitudes
and angles of the power flow (PF) problem on account of the admirable speed of
17 Machine Learning Approaches in a Real Power System and Power Markets 369

ANN over traditional PF approaches. The discussed ANN procedure has been
triumphantly examined by applying the IEEE-30 bus system.
Arnagiriet in [79] has suggested an application of ANNs to identify bus voltages
of a radial DS for any allotted load without performing the load-flow algorithm.
Reference [80, 81] presented a novel, satisfying approach to perform stochastic load-
flow analysis by employing the standard back propagation method for training the
NN. In [82] Karami and Mohammadi suggested an approach for dealing with the
load-flow issue of the electric power systems by utilizing RBFNN with a rapid
hybrid training design. Accordingly, they tackled the function approximation prob-
lem by using RBFNN in order to solve the load-flow. State estimation (SE) is an
indispensable part in overall monitoring and controlling of the transmission net-
works. It is chiefly employed to provide a reliable approximation of the system
voltages. The data from the state estimator streams into control centers and database
servers within the network [83]. Authors in [84] propose a novel method in accor-
dance with ANN for power system network topology identification and static SE. In
the mentioned reference, the state estimator design uses the dynamic alteration in
network topology and bad data processing. The performance of RBFNN for SE is
examined in [85] by analyzing its pertinence on IEEE 14 bus system, and the
suggested estimator is compared with traditional weighted least squares (WLS)
state estimator based on time, exactness, and robustness. Reference [86] proposed
a fuzzy clustering and pattern matching for topology determination, bad data iden-
tification, and SE. In this approach, a fuzzy pattern vector (FPV) is produced on the
basis of the obtained analog measurement vector (telemetry data). The topology
determination and gross errors are identified from the difference between the analog
measurement data and FPV. The gross errors and topology errors in the measure-
ment data are determined and modified by utilizing the FPV which can be utilized
directly as an efficient data measurement for SE.
Voltage stability is a crucial factor that should be considered during the planning
and controlling the power systems for the purpose of avoiding voltage collapse and
partial or full system blackout afterward. Studying the voltage collapse incident may
yield an approach to avert this event from occurring [87]. Authors in [88] propose an
approach to monitor the voltage uncertainty in a power system with a hybrid artificial
NN which is a compound of a multilayer perceptron and the Kohonen NN. In the
presented approach, the Kohonen network is applied to categorize the system
operating status; the Kohonen produced patterns that are utilized as inputs to train
a multilayer perceptron for determination of alarm states which are threatening for
the system security [88]. In [69] authors studied the performance of real-time system
monitoring models which can ensure a timely warning ahead of the voltage collapse
happened in the electrical grid. In the mentioned paper, various kinds of line voltage
stability indices (LVSI) are distinguished to determine their capability to find the
weakest lines in the electrical grid. The suggested real-time voltage stability moni-
toring is implemented by applying ANN. Authors in [89] modified the ANN-based
approach to rapidly evaluate the long-term VSM. The discussed examination in [89]
illustrates that the phase angles and node voltage magnitudes are the most effective
estimators for determining the VSM. Moreover, the paper indicates that the
370 P. Razmi and M. Ghaemi Asl

presented ANN-based approach is able to satisfactorily evaluate the VSM under

normal operation as well as under N-1 contingency conditions. In [90] authors offer
a SVM regression network to assess voltage stability for both normal situation and
contingency cases. SVM receives real and reactive power from all buses in the
system and provides the loading margin. Eventually the outcomes of the presented
approach are compared with NN, ELM, online sequential ELM, and extreme SVM
regression models.

17.5 Optimization and Protection-Themed Studies

Power system optimization has developed with advancements in computation and

optimizing theories [91]. The optimal power flow (OPF) is an optimization method
in power system which is used for monitoring the generation-consumption to
optimize definite aims, namely, reducing the generation cost or the system power
losses. Much research has dealt with OPF problem by applying AI and ML algo-
rithms [92]. In paper [93] two kinds of neuron transfer functions are utilized to
examine the robustness and exactness of the presented method by considering the
OPF. Nguyen in [93] modified the NN which executes the Newton-Raphson-based
OPF calculations to reduce the real power losses. Hartati and El-Hawary proposed an
application of applying an adapted Hopfield NN to resolve the OPF problem with the
aim being the incremental fuel cost [94]. Reference [95] has offered an artificial NN
method on the basis of Lagrangian multiplier to deal with the economic load-flow
problem as well as increasing the convergence speed by applying the momentum
technique in the power system.
A novel approach to resolve the economic power dispatch problem (EPDP) with
piecewise quadratic cost function by employing the Hopfield NN was proposed in
[96]. Authors in [97] propose an Improved Hopfield Neural Network (IHNN) to
solve the economic dispatch problem (EDP). The presented IHNN method has quick
convergence and moves effectively toward possible equilibrium points. Paper [98]
presents an approach on the basis of quadratic programming (QP) and augmented
Lagrange Hopfield network (ALHN) for solving EDP through piecewise quadratic
cost functions and prohibited locations. In the mentioned approach, the QP model is
firstly utilized to identify the fuel cost curve for each of the units and set initial
parameters for the ALHN model; after that a heuristic search is applied to repair
prohibited location violations, and the ALHN approach is finally used to solve the
problem if any violations identified.
An automatic learning structure in order to control the dynamic security of an
electrical grid was presented in [99]. The presented approach uses RBFNN, which
serves to determine the dynamic security conditions of the electrical grid and to
evaluate the impact of a corrective control action employed in the event where
disruption happens. PSO is used to detect the optimum control action, where the
objective function is provided by the RBFNN for optimization. In Reference [100], a
novel dynamic security evaluation and generation rescheduling approach applying
17 Machine Learning Approaches in a Real Power System and Power Markets 371

GAs which are unified with PNNs and ANFISs are presented for the precautionary
control of vast electrical grids in the event of fleeting disruptions. In the mentioned
method, PNNs are applied in a practical manner to measure the security zones
precisely during the evaluation and control. The security confined generation
rescheduling is applied using a GA which improves the generation shifting or
amount of the fuel cost during the precautionary control.
Authors in Reference [100] presented DT-based preventive/corrective control
(PCC) approaches to advance the dynamic security of electrical grids against the
possible and unforeseen circumstances leading to fleeting disruptions. PCCs,
namely, generation rescheduling as well as load shedding models, are advanced on
the basis of security zones and boundaries, respectively. These two PCCs could be
computed in the space of proper decision variables. The security zones and bound-
aries are decided in accordance with laws of DTs which are modified by the
generated knowledge principles. A novel approach is proposed in [101] for
security-constrained corrective rescheduling of real power by utilizing the Hopfield
NN. The minimum aberrations in real power generations and loads at power system
buses are merged to form the objective function for optimization.
In order to decrease the energy losses in DSs, one of the approaches is demon-
strated in [102]. It is in accordance with heuristic methods, namely, branch exchange
method and fuzzy logic method. This algorithm is developed by taking the loads as
fuzzy numbers, which deals better with instabilities in the power demand in DSs.
Reference [103] presents a computational framework for the voltage adjustment of
DSs supplied by dispersed generation systems (DGS). The purpose is to determine
an efficient solution for the optimal regulation problem by unifying a traditional
nonlinear programming algorithm with an adaptive local learning algorithm (LLA).
The explanation for the method is that the LLA is able to quickly learn based on a
confined number of historical observations.
Reference [104] proposes a method to combine the feed-forward NN and the
simulated annealing approach for solving the problem of unit commitment, which is
a mixed integer combinational optimization issue in power system. ANN is applied
to identify the discrete variables in accordance with the status of each unit at its time
interval. The simulated annealing approach is utilized to produce the continuous
variables in accordance with the output power of each unit and the production cost.
The outcomes in [104] proved that the presented method is able to deal with unit
commitment in a less computational time with an optimal generation plan. H. Sasaki
[105] examines the practicability of using the Hopfield NN for combinational
optimization issues in power systems, specifically for unit commitment. Lots of
inequality constraints contained in unit commitment could be dealt with
dedicated NNs.
The monitoring exactness and speed are major parameters for the fixed step
perturb-and-observe (P&O) maximum power point tracking (MPPT) approach.
Paper [106] presents a new method to balance the trade-off between operation and
cost based on a PV system. The disturbance step size is identified off-line for a
certain zone on the basis of local irradiance information. The SVM is utilized to
automatically categorize the desert or coastal zones by applying historical irradiance
372 P. Razmi and M. Ghaemi Asl

information. The disturbance step size is modified for better system operation
without enhancing the control intricacies. Assessing the solar energy and thermal
transmission mechanism, efficiency needs a great deal of time, financial charge, and
human resources. Concurrently, straight forecasting the efficiency of these mecha-
nisms is an onerous task since they have involved internal designs. Luckily, a
knowledge-based ML approach is able to yield an auspicious forecasting and
improvement procedure for the efficiency of energy networks. In Reference [107],
authors illustrate how they apply the ML methods to train an immense experimental
database to achieve exact forecast for an enhancement on a solar water heater (SWH)
model.
A novel energy system improvement approach on the basis of a high-throughput
screening (HTS) procedure is suggested in [107]. The mentioned paper’s structure is
a compound of (1) comparative research on various ML methods: ANNs, SVM, and
ELM to forecast the efficiencies of SWHs; (2) advancement of an ANN-based
program to support the fast forecasting; and (3) an explanation for a computational
HTS approach for planning a SWH model with a high efficiency. In paper [108],
authors researched the advantages of applying deep reinforcement learning (DRL) in
the smart grid context, a hybrid type of approaches which unites RL with deep
learning (DL), in order to execute an online optimization for building energy
management systems. The learning process was examined by applying two
approaches, deep Q-learning and deep policy gradient; both of these methods were
being expanded to complete multiple actions at the same time. In Reference [109],
the improved chicken swarm optimization algorithm support vector machine (ICSO-
SVM) design is presented for forecasting the wind power. Regarding the conven-
tional chicken swarm optimization method (CSO) features, it usually falls into a
local optimal while dealing with high-dimensional matters. As a result, the CSO
approach is modified and the ICSO technique is suggested. Paper [110] proposes a
developed non-dominated sorting genetic algorithm II (NSGA-II) method including
a parameter-free self-tuning RL approach named NSGA-RL in order to deal with
environmental and EDPs with several purposes.

17.6 Conclusion

In this chapter a broad range of ML methods in order to solve timely matters in the
fields of generation, transmission, and distribution of recent electrical grids have
been discussed. Scope of ANNs, kernel regression, decision trees, fuzzy networks,
and GAs have been presented for security assessment, control, EDP, PF computa-
tions, prediction, rehabilitation, plant supervising, transformer construction, faulted
TL issues, and inductive inference matters. At every instance, the applications of ML
methods have demonstrated their effectiveness for tackling complicated electrical
grid operations, design, and construction problems. The discussed applications
showed the thriving activity in electrical grid applied study, as illustrated through
the considerable number of relevant publications. To sum up, ML approaches along
17 Machine Learning Approaches in a Real Power System and Power Markets 373

with developed systematic methods guarantee a great supportive effect for future
modern electrical grids.

References

1. M.S. Eslami, Application of artificial intelligent techniques in PSS design: A survey of the
state-of-the-art methods. Przegld Elektrotechniczny (Electr. Rev.) 87(4), 188–197 (2011)
2. R.S. Segal, A self-tuning power system stabilizer based on artificial neural network.
Int. J. Electr. Power Energy Syst. 26(6), 423–430 (2004)
3. S. Pillutla, Power system stabilization based on modular neural network architecture.
Int. J. Electr. Power Energy Syst. 19(6), 411–418 (1997)
4. N. Hosseinzadeh, A. Kalam, A hierarchical neural network adaptive power system stabilizer.
Int. J. Electr. Power Energy 19, 28–33 (1999)
5. P. Shamsollahi, Design of a neural adaptive power system stabilizer using dynamic back-
propagation method. Int. J. Electr. Power Energy Syst 22, 29–34 (2000)
6. P. Shamsollahi, An adaptive power system stabilizer using online trained neural networks.
IEEE Trans Energy Convers. 12, 382–387 (1997)
7. M.A.-M. Abido, Adaptive tuning of power system stabilizers using radial basis function
networks. Electr. Power Syst. Res. 49, 21–29 (1999)
8. T. Senjyu, Recurrent neural network supplementary stabilization controller for automatic
voltage regulator and governor. Electr. Power Components Syst. 31, 693–707 (2003)
9. H.K. Ping, Studies of the improvement of probabilistic PSSs by using the single neuron model.
Int. J. Electr. Power Energy Syst. 29(3), 217–221 (2007)
10. S. Ray, A wide area measurement based neuro control for generation excitation systems. Eng.
Appl. Artif. Intell. 22(3), 473–481 (2009)
11. L.A. Zadeh, Fuzzy sets. Inf. Control. 8(3), 338–353 (1956)
12. S. Pal, D. Mandal, Fuzzy logic and approximate reasoning: an overview. IETE J. Res. 37,
548–559 (1991)
13. A.R. Tavakoli, A.R. Seifi, M.M. Arefi, Designing a selfconstructing fuzzy neural network
controller for damping power system oscillations. Fuzzy Sets Syst. 356, 63–76 (2019)
14. K. Eltag, M.S. Aslamx, R. Ullah, Dynamic stability enhancement using fuzzy PID control
technology for power system. Int. J. Control. Autom. Syst. 17(1), 234–242 (2019)
15. M. Farahani, S. Ganjefar, Intelligent power system stabilizer design using adaptive fuzzy
sliding mode controller. Neurocomputing 226, 135–144 (2017)
16. T.T. Lie, An adaptive fuzzy logic power system stabilizer. Electr. Power Syst. Res. 38(1),
75–81 (1996)
17. M. Ramirez-Gonzalez, Self-tuned power system stabilizer based on a simple fuzzy logic
controller. Electr. Power Components Syst. 38(4), 407–423 (2010)
18. D.K. Sambariya, R. Prasad, Robust tuning of power system stabilizer for small signal stability
enhancement using metaheuristic bat algorithm. Int. J. Electr. Power Energy Syst. 61, 229–238
(2014)
19. H.M. Soliman, E.H.E. Bayoumi, M.F. Hassan, Power system stabilizer design for minimal
overshoot and control constraint using swarm optimization. Electr. Power Components Syst.
37(1), 111–126 (2008)
20. S. Kamalasadan, G.D. Swann, R. Yousefian, A novel systemcentric intelligent adaptive
control architecture for power system stabilizer based on adaptive neural networks. IEEE
Syst. J. 8(4), 1074–1085 (2014)
21. D. Molina, G.K. Venayagamoorthy, J. Liang, R.G. Harley, Intelligent local area signals based
damping of power system oscillations using virtual generators and approximate dynamic
programming. IEEE Trans. Smart Grid 4(1), 498–508 (2013)
374 P. Razmi and M. Ghaemi Asl

22. M. Farahani, A multi-objective power system stabilizer. IEEE Trans. Power Syst. 28(3),
2700–2707 (2013)
23. R. Hadidi, B. Jeyasurya, Reinforcement learning based real-time wide-area stabilizing control
agents to enhance power system stability. IEEE Trans. Smart Grid 4(1), 489–497 (2013)
24. S.R. Khuntia, S. Panda, ANFIS approach for SSSC controller design for the improvement of
transient stability performance. Math. Comput. Model. 57(1), 289–300 (2013)
25. A.G. Aissaoui, M. Abid, H. Abid, A. Tahour, A.K. Zeblah, A fuzzy logic controller for
synchronous machine. J. Electr. Eng. Bratislava 58(5), 285 (2007)
26. J.W. Park, R.G. Harley, G.K. Venayagamoorthy, Adaptivecritic-based optimal neurocontrol
for synchronous generators in a power system using MLP/RBF neural networks. IEEE Trans.
Ind. Appl. 39(5), 1529–1540 (2003)
27. F.M.E.S. Fayez, Robust adaptive wavelet-neural-network sliding-mode speed control for a
DSP-based PMSM drive system. J Power Electr (JPE) 10(5), 50517 (2010)
28. K.C. Yu, S.P. Hsu, Y.H. Hung, Optimization of fuzzy controller of permanent magne
synchronous motor. J. Appl. Sci. 7(19), 272535 (2007)
29. E. Cetin, U. Oguz, H.S. Hasan, A neuro-fuzzy controller for speed control of a permanent
magnet synchronous motor drive. Expert Syst. Appl. 34(1), 65764 (2008)
30. J. Faiz, A. Azami, A. Keyhani, A. Proca, Closed-loop control stability for permanent magnet
synchronous motor. Int. J. Electr. Power Energy Syst. 19(5), 3317 (1997)
31. K. Murat, I.E. Hasan, Speed and current regulation of a permanent magnet synchronous motor
via nonlinear and adaptive backstepping control. Math. Comput. Model. 53(910), 201530
(2011)
32. Q. Wang, H. Yu, M. Wang, X. Qi, A novel adaptive NeuroControl approach for permanent
magnet synchronous motor speed control. Energies 11(9), 2355 (2018)
33. C.B. Butt, M.A. Rahman, Untrained artificial neuron-based speed control of interior
permanent-magnet motor drives over extended operating speed range. IEEE Trans. Ind.
Appl. 49(3), 1146–1153 (2013)
34. X. Sun, L. Chen, H. Jiang, Z. Yang, J. Chen, W. Zhang, High performance control for a
bearingless permanent-magnet synchronous motor using neural network inverse scheme plus
internal model controllers. IEEE Trans. Ind. Electron. 63(6), 3479–3488 (2016)
35. Y. Li, B. Zhang, X. Xu, Decoupling control for permanent magnet in-wheel motor using
internal model control based on back-propagation neural network inverse system. Bull. Pol.
Acad. Sci. Tech. Sci. 66(6), 961 (2018)
36. Y. Li, B. Li, X. Xu, X. Sun, A nonlinear decoupling control approach using RBFNNI-based
robust pole placement for a permanent magnet in-wheel motor. IEEE Access 6, 1844–1854
(2018)
37. Z. Frijet, A. Zribi, M. Chtourou, Adaptive neural network internal model control for PMSM
speed regulation. J. Electr. Syst. 14(2), 118–126 (2018)
38. K. Chen, C. Huang, J. He, Fault detection, classification and location for transmission lines and
distribution systems: a review on the methods. High Volt. 1(1), 25–33 (2016)
39. A.A. Yusuff, A.A. Jimoh, J.L. Munda, Fault location in transmission lines based on stationary
wavelet transform, determinant function feature and support vector regression. Electr. Power
Syst. Res. 110, 73–83 (2014)
40. Z. Liu, Z. Han, Y. Zhang, Q. Zhang, Multiwavelet packet entropy and its application in
transmission line fault recognition and classification. IEEE Trans. Neural Netw. Learn. Syst.
25(11), 2043–2052 (2014)
41. A. Swetapadma, A. Yadav, A novel single-ended fault location scheme for parallel transmis-
sion lines using k-nearest neighbor algorithm. Comput. Electr. Eng. 69, 41–53 (2018)
42. A.L. da Silva Pessoa, M. Oleskovicz, P.E.T. Martins, A multi-stage methodology for fault
location in radial distribution systems, in 2018 18th International Conference on Harmonics
and Quality of Power (ICHQP) (IEEE, 2018), pp. 1–6
17 Machine Learning Approaches in a Real Power System and Power Markets 375

43. A. Prasad, J.B. Edward, Importance of artificial neural networks for location of faults in
transmission systems: a survey, in 2017 11th International Conference on Intelligent Systems
and Control (ISCO) (IEEE, 2017), pp. 357–362
44. Y.Q. Chen, O. Fink, G. Sansavini, Combined fault location and classification for power
transmission lines fault diagnosis with integrated feature extraction. IEEE Trans. Ind. Electron.
65(1), 561–569 (2018)
45. M.F. Guo, X.D. Zeng, D.Y. Chen, N.C. Yang, Deep learning-based earth fault detection using
continuous wavelet transform and convolutional neural network in resonant grounding distri-
bution systems. IEEE Sens. J. 18(3), 1291–1300 (2018)
46. N.R. Babu, B.J. Mohan, Fault classification in power systems using EMD and SVM. Ain
Shams Eng. J. 8(2), 103–111 (2017)
47. F.I. Lozada, F.L. Quilumba, F.E. Prez, Fault detection and classification in transmission lines
using wavelet transform and support vector machines. Revista Tecnica Energia 14 (2018)
48. H. Malik, R. Sharma, EMD and ANN based intelligent fault diagnosis model for transmission
line. J. Intell. Fuzzy Syst. 32(4), 3043–3050 (2017)
49. M. Dhimish, V. Holmes, B. Mehrdadi, M. Dales, Comparing Mamdani Sugeno fuzzy logic
and RBF ANN network for PV fault detection. Renew. Energy 117, 257–274 (2018)
50. J.M. Johnson, A. Yadav, Fault detection and classification technique for HVDC transmission
lines using KNN, in Information and Communication Technology for Sustainable Develop-
ment, (Springer, Singapore, 2018), pp. 245–253
51. N. Amruthnath, T. Gupta, Fault class prediction in unsupervised learning using model-based
clustering approach, in 2018 International Conference on Information and Computer Tech-
nologies (ICICT) (IEEE, 2018), pp. 5–12
52. A. Khaleghi, M.O. Sadegh, M. Ghazizadeh-Ahsaee, A.M. Rabori, Transient fault area location
and fault classification for distribution systems based on wavelet transform and adaptive
NeuroFuzzy inference system (ANFIS). Adv. Electr. Electron. Eng. 16(2), 155–166 (2018)
53. R. Kumar, E. Koley, A. Yadav, A.S. Thoke, Fault classification of phase to phase fault in six
phase transmission line using Haar wavelet and ANN, in 2014 International Conference on
Signal Processing and Integrated Networks (SPIN), (IEEE, 2014), pp. 5–8
54. H. Livani, C.Y. Evrenosoglu, A machine learning and waveletbased fault location method for
hybrid transmission lines. IEEE Trans. Smart Grid 5(1), 51–59 (2014)
55. V. Malathi, N.S. Marimuthu, S. Baskar, Intelligent approaches using support vector machine
and extreme learning machine for transmission line protection. Neurocomputing 73(10),
2160–2167 (2010)
56. S.R. Samantaray, P.K. Dash, G. Panda, Distance relaying for transmission line using support
vector machine and radial basis function neural network. Int. J. Electr. Power Energy Syst. 29,
551556 (2007)
57. J.A. Jiang, C.L. Chuang, Y.C. Wang, C.H. Hung, J.Y. Wang, C.H. Lee, Y.T. Hsiao, A hybrid
framework for fault detection, classification, and location. Part I: concept, structure, and
methodology. IEEE Trans. Power Deliv. 26(3), 1988–1998 (2011)
58. J.-A. Jiang, C.-L. Chuang, Y.-C. Wang, C.-H. Hung, J.-Y. Wang, C.-H. Lee, Y.-T. Hsiao, A
hybrid framework for fault detection, classification, and locationpart II: implementation and
test results. IEEE Trans. Power Deliv. 26(3), 1999–2008 (2011)
59. M.T. Hagh, K. Razi, H. Taghizadeh, Fault classification and location of power transmission
lines using artificial neural network, in IPEC 2007. International Power Engineering Confer-
ence (IEEE, 2007), pp. 1109–1114
60. P.K. Dash, S.R. Samantaray, An accurate fault classification algorithm using a minimal radial
basis function neural network. Eng. Intell. Syst. 4, 205–210 (2004)
61. J. Upendar, C.P. Gupta, G.K. Singh, Discrete wavelet transform and probabilistic neural
network based algorithm for classification of fault on transmission systems, in INDICON
2008. Annual IEEE India Conference, vol. 1 (IEEE, 2008), pp. 206–211
62. S.W. Fei, X.B. Zhang, Fault diagnosis of power transformer based on support vector machine
with genetic algorithm. Expert Syst. Appl. 36(8), 11352–11357 (2009)
376 P. Razmi and M. Ghaemi Asl

63. A. Shintemirov, W. Tang, Q.H. Wu, Power transformer fault classification based on dissolved
gas analysis by implementing bootstrap and genetic programming. IEEE Trans. Syst. Man
Cybernet. C Appl. Rev. 39(1), 69–79 (2009)
64. A.M. Shah, B.R. Bhalja, Discrimination between internal faults and other disturbances in
transformer using the support vector machine based protection scheme. IEEE Trans. Power
Deliv. 28(3), 1508–1515 (2013)
65. L. Liu, Z. Ding, Modeling analysis of power transformer fault diagnosis based on improved
relevance vector machine. Math. Prob. Eng. 2013, 636374 (2013)
66. T.S. Sidhu, H. Singh, M.S. Sachdev, Design, implementation and testing of an artificial neural
network based fault direction discriminator for protecting transmission lines. IEEE Trans.
Power Deliv. 10(2), 697–706 (1995)
67. Y. Cui, P. Bangalore, L.B. Tjernberg, An anomaly detection approach using wavelet transform
and artificial neural networks for condition monitoring of wind turbines’ gearboxes, in 2018
Power Systems Computation Conference (PSCC) (IEEE, 2018), pp. 1–7
68. A.R. Bahmanyar, A. Karami, Power system voltage stability monitoring using artificial neural
networks with a reduced set of inputs. Int. J. Electr. Power Energy Syst. 58, 246–256 (2014)
69. D.Q. Zhou, U.D. Annakkage, A.D. Rajapakse, Online monitoring of voltage stability margin
using an artificial neural network. IEEE Trans. Power Syst. 25(3), 1566–1574 (2010)
70. D.V. Coury, D.C. Jorge, Artificial neural network approach to distance protection of trans-
mission lines. IEEE Trans. Power Deliv. 13(1), 102–108 (1998)
71. J. Pihler, B. Grar, D. Dolinar, Improved operation of power transformer protection using
artificial neural network. IEEE Trans. Power Deliv. 12(3), 1128–1136 (1997)
72. L. Hernandez, C. Baladrn, J.M. Aguiar, B. Carro, A.J. SanchezEsguevillas, J. Lloret, Short-
term load forecasting for microgrids based on artificial neural networks. Energies 6(3),
1385–1408 (2013)
73. C.N. Ko, C.M. Lee, Short-term load forecasting using SVR (support vector regression)-based
radial basis function neural network with dual extended Kalman filter. Energy 49, 413–422
(2013)
74. Z.A. Bashir, M.E. El-Hawary, Applying wavelets to short-term load forecasting using
PSO-based neural networks. IEEE Trans. Power Syst. 24(1), 20–27 (2009)
75. P.F. Pai, W.C. Hong, Forecasting regional electricity load based on recurrent support vector
machines with genetic algorithms. Electr. Power Syst. Res. 74(3), 417–425 (2005)
76. D.K. Chaturvedi, A.P. Sinha, O.P. Malik, Short term load forecast using fuzzy logic and
wavelet transform integrated generalized neural network. Int. J. Electr. Power Energy Syst. 67,
230–237 (2015)
77. V.N. Coelho, I.M. Coelho, B.N. Coelho, A.J. Reis, R. Enayatifar, M.J. Souza, F.G. Guimares,
A self-adaptive evolutionary fuzzy model for load forecasting problems on smart grid envi-
ronment. Appl. Energy 169, 567–584 (2016)
78. V.L. Paucar, M.J. Rider, Artificial neural networks for solving the power flow problem in
electric power systems. Electr. Power Syst. Res. 62(2), 139–144 (2002)
79. A. Arunagiri, B. Venkatesh, K. Ramasamy, Artificial neural network approach-an application
to radial loadflow algorithm. IEICE Electron. Exp. 3(14), 353–360 (2006)
80. A. Jain, S.C. Tripathy, R. Balasubramanian, K. Grag, Y. Kawazoe, Neural network based
stochastic load flow analysis, in PowerCon 2004. 2004 International Conference on Power
System Technology, 2004, vol. 2 (IEEE, 2004), pp. 1845–1850
81. A. Karami, M.S. Mohammadi, Radial basis function neural network for power system load-
flow. Int. J. Electr. Power Energy Syst. 30(1), 60–66 (2008)
82. Y.F. Huang, S. Werner, J. Huang, N. Kashyap, V. Gupta, State estimation in electric power
grids: meeting new challenges presented by the requirements of the future grid. IEEE Sign.
Process. Mag. 29(5), 33–43 (2012)
83. D.V. Kumar, S.C. Srivastava, S. Shah, S. Mathur, Topology processing and static state
estimation using artificial neural networks, in IEE Proceedings-Generation, Transmission
and Distribution, vol. 143, no. 1 (IET, 1996), pp. 99–105
17 Machine Learning Approaches in a Real Power System and Power Markets 377

84. D. Singh, J.P. Pandey, D.S. Chauhan, Radial basis neural network state estimation of electric
power networks, in Proceedings of the 2004 IEEE International Conference on Electric Utility
Deregulation, Restructuring and Power Technologies, 2004. (DRPT 2004), vol. 1 (IEEE.
2004), pp. 90–95
85. D. Singh, J.P. Pandey, D.S. Chauhan, Topology identification, bad data processing, and state
estimation using fuzzy pattern matching. IEEE Trans. Power Syst. 20(3), 1570–1579 (2005)
86. F. Larki, M. Joorabian, H.M. Kelk, M. Pishvaei, Voltage stability evaluation of the
Khouzestan power system in Iran using CPF method and modal analysis, In 2010 Asia-
Pacific Power and Energy Engineering Conference (APPEEC) (IEEE, 2010), pp. 1–5
87. A. Zhukov, N. Tomin, D. Sidorov, D. Panasetsky, V. Spirayev, A hybrid artificial neural
network for voltage security evaluation in a power system, in 2015 5th International Youth
Conference on Energy (IYCE) (IEEE. 2015), pp. 1–8
88. H.H. Goh, Q.S. Chua, S.W. Lee, B.C. Kok, K.C. Goh, K.T.K. Teo, Evaluation for voltage
stability indices in power system using artificial neural network. Proc. Eng. 118, 1127–1136
(2015)
89. M.V. Suganyadevi, C.K. Babulal, S. Kalyani, Assessment of voltage stability margin by
comparing various support vector regression models. Soft. Comput. 20(2), 807–818 (2016)
90. M.B. Cain, R.P. Oneill, A. Castillo, History of Optimal Power Flow and Formulations
(Federal Energy Regulatory Commission, Washington, DC, 2012)
91. L. Gan, N. Li, U. Topcu, S.H. Low, Optimal power flow in distribution networks, in
Proceeding of 52nd IEEE conference on decision and control, December 2013
92. T.T. Nguyen, Neural network optimal-power-flow, in Fourth International Conference on
Advances in Power System Control, Operation and Management, 1997, pp. 266–271
93. R.S. Hartati, M.E. El-Hawary, Optimal active power flow solutions using a modified Hopfield
neural network, in 2001 Canadian Conference on Electrical and Computer Engineering, vol.
1 (IEEE. 2001), pp. 189–194
94. M. Mohatram, P. Tewari, N. Latanath, 2011, April Economic load flow using Lagrange neural
network, in 2011 Saudi International Electronics, Communications and Photonics Conference
(SIECPC), (IEEE, 2011), pp. 1–7
95. J.H. Park, Y.S. Kim, I.K. Eom, K.Y. Lee, Economic load dispatch for piecewise quadratic cost
function using Hopfield neural network. IEEE Trans. Power Syst. 8(3), 1030–1038 (1993)
96. S.S. Reddy, J.A. Momoh, Economic dispatch using improved Hopfield neural network, in
2015 North American Power Symposium (NAPS), (IEEE, Chicago, 2015), pp. 1–5
97. V.N. Dieu, P. Schegner, Augmented Lagrange Hopfield network initialized by quadratic
programming for economic dispatch with piecewise quadratic cost functions and prohibited
zones. Appl. Soft Comput. 13(1), 292–301 (2013)
98. E.M. Voumvoulakis, N.D. Hatziargyriou, A particle swarm optimization method for power
system dynamic security control. IEEE Trans. Power Syst. 25(2), 1032–1041 (2010)
99. C.F. Kucuktezcan, V.I. Genc, A new dynamic security enhancement method via genetic
algorithms integrated with neural network based tools. Electr. Power Syst. Res. 83(1), 1–8
(2012)
100. I. Genc, R. Diao, V. Vittal, S. Kolluri, S. Mandal, Decision tree-based preventive and
corrective control applications for dynamic security enhancement in power systems. IEEE
Trans. Power Syst. 25(3), 1611–1619 (2010)
101. S. Ghosh, B.H. Chowdhury, Security-constrained optimal rescheduling of real power using
Hopfield neural network. IEEE Trans. Power Syst. 11(4), 1743–1748 (1996)
102. G. Crtina, C. Bonciu, M. Musat, Z. Zisman, Application of fuzzy logic for energy loss
reduction in distribution networks, in 9th Mediterranean Electrotechnical Conference, 1998.
MELECON 98, vol. 2 (IEEE, 1998), pp. 974–977
103. D. Villacci, G. Bontempi, A. Vaccaro, An adaptive local learning-based methodology for
voltage regulation in distribution networks with dispersed generation. IEEE Trans. Power
Syst. 21(3), 1131–1140 (2006)
378 P. Razmi and M. Ghaemi Asl

104. R. Nayak, J.D. Sharma, A hybrid neural network and simulated annealing approach to the unit
commitment problem. Comput. Electr. Eng. 26(6), 461–477 (2000)
105. H. Sasaki, M. Watanabe, D. Kubokawa, N. Yorino, R. Yokoyama, A solution method of unit
commitment by artificial neural networks. IEEE Trans. Power Syst. 7(3), 974–981 (1992)
106. K. Yan, Y. Du, Z. Ren, MPPT perturbation optimization of photovoltaic power systems based
on solar irradiance data classification. IEEE Trans. Sustain. Energy 10, 514–521 (2018)
107. H. Li, Z. Liu, Performance prediction and optimization of solar water heater via a knowledge-
based machine learning method, in Handbook of Research on Power and Energy System
Optimization, (IGI Global, Hershey, 2018), pp. 55–74
108. E. Mocanu, D.C. Mocanu, P.H. Nguyen, A. Liotta, M.E. Webber, M. Gibescu, J.G. Slootweg,
On-line building energy optimization using deep reinforcement learning. IEEE Trans. Smart
Grid 10(4), 3698–3708 (2018)
109. C. Fu, G.Q. Li, K.P. Lin, H.J. Zhang, Short-term wind power prediction based on improved
chicken algorithm optimization support vector machine. Sustainability 11(2), 512 (2019)
110. T.C. Bora, V.C. Mariani, L. dos Santos Coelho, Multiobjective optimization of the
environmental-economic dispatch with reinforcement learning based on non-dominate sorting
genetic algorithm. Appl. Therm. Eng. 146, 688–700 (2019)
Index

A Artificial neural network (ANN), 121, 250, 251,

Activity selection algorithm (ASA), 275, 276, 253, 268, 269, 272, 273, 277, 280,
281, 285 281, 285, 322–324, 327, 349–351,
Adaptive fuzzy logic PSS (AFLPSS), 361 353, 359, 360, 362, 364–369, 372
Adaptive fuzzy sliding-mode controller Artificial neural network controllers
(AFSMC), 360 (ANNCs), 359
Adaptive learning rates (ALRs), 360 Augmented Lagrange Hopfield network
Adaptive neural network model following (ALHN), 370
speed controller (ANNMFC), 362 Autoencoders, 72
Adaptive neuro-fuzzy inference system Automatic billing systems, 283
(ANFIS), 29, 362 Automatic generation control (AGC), 30
CSI, 328–338, 340 Automatic smart reading (AMR) schemes, 266
DSO, 324, 325 Automatic voltage regulators (AVRs), 358
Kalman forecasting machine, 327 Autoregressive integrated moving average
LMPs (see Locational marginal prices (ARIMA), 238, 250, 251, 322
(LMPs)) Autoregressive moving average (ARMA), 238,
LPF (see Load and price-forecasting (LPF)) 250, 323
MAPE, 322–324 Autoregressive with exogenous inputs
and MLP, 322 (ARX), 272
online optimal topology-forecasting Average classification techniques accuracy, 281
procedure, 328 Average F-measure values, 281
optimal topology, 324, 327 Average LMP (ALMP), 330, 333, 334, 337,
restructuring of decomposed signals, 327 338, 340
topology-forecasting process, 324
WT, 325, 326
Advanced metering infrastructures (AMIs), B
265, 266, 283 Back propagation (BP) NNs, 366
Air conditioner, 346 Back propagation neural network inverse
Air temperature, 251, 252, 257, 259 (BPNNI), 363
ANN-based method, 367 Backpropagation neural network (BPNN), 68
Arithmetic complexity, 279, 280 BA-CPSS, 361
Artificial intelligence (AI), 250, 254, 370 Bad power quality, 271

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 379
M. Nazari-Heris et al. (eds.), Application of Machine Learning and Deep Learning
Methods to Power System Problems, Power Systems,
https://doi.org/10.1007/978-3-030-77696-1
380 Index

Bat algorithm (BA), 361 DT (see Decision tree (DT))

Bayesian network, 346 general structure, 221, 222
Bayesian recurrent ANN, 324 IEEE39-bus system (see IEEE39-bus power
Bearing-less PMSM (BPMSM), 363 system)
Binary genetic algorithm (BGA), 122 predicts and classifies, 213
Binding carbon reduction, 266 propagation pattern, 213
Biologically oriented techniques, 270 singular value decomposition and post-
Blackout classification events data, 213
CSPDTs, 222, 223 WAMS/PMU, 214
Boolean class, 216, 217 CDDT, 294, 317
Bootstrap-aggregating algorithms, 187 fault clearance detection by, 300, 301
Box-Cox transformation, 122 training, 305
Box-Jenkins models, 322 Charge resistance, 278
Branch process models, 212, 213 Classification algorithms, 285
Brittleness index, 213 Classifier, 271
Building energy management system Cluster-based approaches, 272
(BEMS), 237 Clustering techniques, 151, 272
68-bus test system, 155
118-bus test system, 205, 206
C coherency-based power system, 153
Cascading failure, 211 DER characteristics, 153
CFOP (see Cascading failures occurrence dynamic equivalencing, 153
potential (CFOP)) Euclidean distances, 152
CSPDTs (see Cascading failure severity feature extraction methods, 156
predictor DTs (CSPDTs)) ICA, 158
data mining, 213 PCA, 156
probability, 212 ideal algorithm, 173
wind turbine, 213 performance of, 173
Cascading failure severity predictor DTs supervised learning approaches, 159
(CSPDTs) in ANNs, 160
advantage, 221 decision trees, 161
blackout classification, 222, 223 time-series variations, 152
blackout severity, 215, 216 unsupervised learning approaches, 161
C4.5 algorithm, 216, 217 DBSCAN, 169
capability, 221 FCM algorithm, 165
continuous attributes, 217 K-means clustering algorithm, 162
global performance, 232 PAM clustering algorithm, 164
lines 17–18, 226–228 subtractive clustering algorithm, 167
lines 21–22, 229, 230 support vector clustering, 170
MATLAB, 225 Co-active neuro-fuzzy inference system
performance validation, 231 (CANFIS), 322, 328, 329, 334,
training, 214, 225 338, 340
WAMS, 221 Compressed-air energy storage (CAES), 6
WAMS/PMU, 213 Compression
Cascading failures occurrence potential (CFOP) designed solution’s, 267
asynchronous/lost data, 221 fold ratio, 280
and blackout prediction, 221 real-time, 275
and blackout severity, 222 real-time data, 267, 284
conceptual structure, 214, 215 Conditional entropy, 219
critical line, 214 Consumer awareness, 267
critical scenario, 214 Consumption patterns, 273, 274, 280, 281, 285
and DOVs (see Dominant operating Contingency severity index (CSI), 328–338,
variables (DOVs)) 340
Index 381

Continuous WT (CWT), 364 hidden knowledge, 253

Control modes, 94, 99 MLP, 254
Control-themed applications of ML in power Data motion sensors, 346
systems Database generation approach
AFLPSS, 361 importance sampling, 185
AFSMC, 360 sampling, 183
ALRs, 360 load variation pattern, 183
ANFIS, 362 proposed approach, 184
ANNCs, 359 random sampling, 185
AVRs, 358 topology scenarios, 184
BA-CPSS, 361 Data-driven modeling, 238
categories, 358, 359 Decision trees (DTs), 69, 297, 298, 364
CPSS, 359 blackout magnitude, 214, 215
FACTS, 358 and brittleness index, 213
FL, 360 CSPDTs (see Cascading failure severity
FLCs, 359, 360 predictor DTs (CSPDTs))
MRAC, 361 nature of operating variables used by, 302
neural network (NN)-based control producing training scenarios of, 303, 304
device, 361 Deep learning, 251
PMIWM, 363 applications, 241
PMSM, 362, 363 DNN (see Deep neural network (DNN))
power system frequency, 358 in energy engineering
PSO-CPSS, 361 application, formulation and
PSO-FPIDC, 360 structure, 70
PSS, 358, 359 autoencoders, 72
RBFNN, 360 convolutional neural network, 71
RL approach, 362 recurrent neural network, 72, 73
RNN, 360 restricted Boltzmann machine, 73, 74
self-constructing structure, 360 LSTM (see Long short-term memory
self-tuned FLPSS, 361 (LSTM))
SG, 362 power control systems, 89, 90
single-machine infinite bus power categorization of, 90, 91
system, 361 regression and classification applications,
SM, 358 74–76
SMIB excluding PSS, 361 and shallow learning, 240
WAMS, 362 types, 241
WD-PSSs, 362 Deep Neural Networks, 89
Conventional neural network (CNN), 238 Deep Q-learning, 372
Conventional power system stabilizer Deep regression and stump tree-based
(CPSS), 359 ensembles models (DRTSEM), 49
Convolutional neural network (CNN), 26, 71 Deep reinforcement learning (DRL), 372
Convolutional neural networks, 349, 364 Deployed regression analysis algorithms, 238
Correlation coefficient (CC), 243, 257 Detection and protection applications of ML in
Cyberattacks, 110 power systems
AGC, 30 ANN-based method, 367
LTC transformer, 31 DSs, 363
state estimation, 30 DWT, 364–366
EMD, 364
error detecting, 363
D GP, 366
Data labeling, 350 PCA, 365
Data mining, 62, 213, 218 RVM, 366
ANN, 253 SVM, 366
382 Index

Detection and protection applications of ML in start time of the task, 347, 348
power systems (cont.) temperature, 348
SW-ELM, 364 Dynamic Security Assessment (DSA), 85
TL, 363–367 Dynamic stability, 86
types of faults, 365
Determinant function feature (DFF), 364
DIgSILENT, 225 E
DIgSILENT Power Factory® software, 303 Economic dispatch problem (EDP), 370
Dimension reduction methods, 218 Economic power dispatch problem
Dimensionality reduction (EPDP), 370
feature extraction, 190 Electric power systems, 369
feature selection, 190 Electrical appliance recognition, 271
hyperparameter tuning, 191 Electrical energy, 345
Discrete Fourier transform (DFT), 364 Electricity industry, 271
Discrete WT (DWT), 364–366 Electricity usage, 267
Dispersed generation systems (DGS), 371 Elman neural network model, 323
Dissolved gas analysis (DGA), 366 Empirical mode decomposition (EMD), 364
Distributed energy resources (DERs), 153 Energy consummation, 265
Distributed generation (DG), 322, 324 Energy engineering
Distribution system, 357 deep learning in
Distribution system operators (DSOs), 283, application, formulation and
324, 325 structure, 70
Dominant operating variables (DOVs) autoencoders, 72
criticality, 217 convolutional neural network, 71
dimension reduction methods, 218 recurrent neural network, 72, 73
and DT, 213, 232 restricted Boltzmann machine, 73, 74
entropy, 213 machine learning methods in
feature reduction, 218 application, formulation and
identification structure, 64
algorithm, 218 data mining and applications, 62, 63
and extraction, 212 decision tree, 69
instantaneous information, 221 general regression neural network,
lines 17–18, 226–228 68, 69
lines 21–22, 229–231 group method data handling, 66, 67
MIM, 218, 220, 221 k-means, 69, 70
mutual information theory, 213 support vector machine, 64, 65
operation information, 217 support vector regression, 67, 68
power system, 216 Energy management, 237, 244, 246, 265
relationship between operating variables Energy system problems, 77
with corresponding behavioral Ensemble methods, 186
characteristics, 220, 221 AdaBoost, 187, 189
and trained CSPDTs, 225 bagged tree, 187
variable identification, 218 Entropy, 213, 216, 219, 226, 233
Door sensors, 346 ERCOT market, 323
DT-based preventive/corrective control Error values, 243–245
(PCC), 371 Event-driven acquisition, 276
Dual extended Kalman filter (DEKF), 368 Event-driven ADCs (EDADCs), 274–276, 278,
Dynamic neural network (DNN), 238 279, 281, 284, 285
Dynamic parameters Event-driven feature extraction, 281
day of the week, 348 Event-driven processing, 267, 274, 285
developed online databases, 347 Event-driven segmentation, 275
in scheduling problems, 347
Index 383

Event-driven sensing (EDS), 274–276, 278, Generator rotor angle fluctuation, 295
280, 284 Genetic algorithm (GA), 324, 328, 334, 335,
Event-driven sensing mechanism, 284 337, 338, 340, 346, 349, 366
Evolution strategies, 368 Genetic programming (GP), 366
Expert systems, 322 Gradient Boosting (XGBoost) method, 57
Extreme gradient boosting, 238 Graph theory, 212
Extremely critical, 215, 223, 231 Gravitational search optimization (GSO),
322, 323
Gray wolf optimizer (GWO), 121
F Greedy randomized adaptive search procedure
Fast Fourier transform, 271 (GRASP), 368
Fault clearance detection Greenhouse gas emissions, 4
by CDDT, 300, 301 Grid planning, 273
Fault detection, 63 Group evolutionary strategy (GES), 368
by FDDT, 298, 300 Group method data handling (GMDH), 66, 67,
FDDT, 294, 295, 317 250, 251, 254–260
fault detection by, 298, 300
training, 304, 305
Feature extraction, 270, 275, 276, 285 H
Feature reduction, 218 Haar-WT (HWT), 365
Feature selection techniques, 270 Heating load prediction
Feed-forward neural network, 349, 359 ARMA, 238
Fine-grained metering data, 266, 282 ARMIA, 238
Flexible alternating current transmission system CNN, 238
(FACTS), 84, 358 commercial and residential buildings, 237
Flower pollination method, 250 and cooling, 238
Fossil fuel-based plants, 250 and data-driven modeling, 238
Fossil fuel-based sources, 237 DNN, 238
Fossil fuels, 249 forecasting, 237
Frequency stability assessment heating operation and dispatching, 237
power control systems, machine learning LSTM (see Long short-term memory
methods, 106–108 (LSTM))
Fuzzy logic (FL), 360 MLP, 238
Fuzzy logic controllers (FLCs), 359, 360 nonlinear optimization problem, 238
Fuzzy logic models, 250 OSELM, 238
Fuzzy models, 322 SARIMA, 238
Fuzzy neural network (FNN), 360 SVM, 238
Fuzzy particle swarm optimization of PID SVM algorithm, 238
controller (PSO-FPIDC), 360 test data, 243, 244
Fuzzy pattern vector (FPV), 369 training data, 243
Fuzzy-ARTMAP neural network, 122 training stage, 244
utilized method, evaluation metric, data type
and purpose, 238–240
G Hidden Markov models (HMM), 267
Gain ratio, 217 Hierarchical clustering algorithms, 272
Galton-Watson High Voltage Direct Current (HVDC), 84
branching process, 212 High-dimensional data processing, 254
Gaussian method, 365 High-throughput screening (HTS)
Gaussian mixture model (GMM), 267 procedure, 372
General regression neural network (GRNN), Home energy management systems, 345, 346,
68, 69 352, 353
Generalized neural network (GNN), 250, Home energy management unit, 346
363, 368
384 Index

I K-means algorithm, 270

IEEE 118-bus test system K-nearest neighbor (KNN), 276, 277, 280,
accuracy and training classification, 203 365, 366
bagged tree model, 197 k-nearest neighbor (k-NN)-based approach, 364
confusion matrix, 205 Knowledge-based ML approach, 372
database generation, 194 Kohonen-based network, 322
hyperparameters tuning, 199 Kohonen network, 369
ML model training, 195 Kolmogorov-Gabor polynomial function, 256
ML techniques, 197
online application, 202
optimal dataset, 195–197 L
performance evaluation, 199, 201 Lagrangian coefficient, 68
IEEE 14 bus system, 369 Lambda architecture, 272
IEEE39-bus power system Least-square SVM (LSSVM), 238
characteristics, 224 Levenberg-Marquardt algorithm, 310
description, 224 Line voltage stability indices (LVSI), 369
DIgSILENT, 225 Linear regression machine learning
global performance, 232 algorithms, 349
lines 17–18 Load and price-forecasting (LPF)
criticality, 226 ANFIS, 322
CSPDTs, 226–228 ANFIS forecasting engine forecasted
DOVs, 226–228 LMPs, 333
lines 21–22 ARIMA, 322
criticality, 228, 229 block diagram, 324, 326
CSPDTs, 229, 230 distribution system operational
DOVs, 229, 230 processes, 322
monitorable operating variables, 224, 225 GSO, 322
operating points, 225 Kalman forecasting machine, 327
performance validation, 231 Kalman machines, 322
time-domain calculations, 225 LMP, 324
Improved chicken swarm optimization MAPE, 322
algorithm support vector machine MLP, 322
(ICSO-SVM) design, 372 WT, 322
Improved hopfield neural network (IHNN), 370 Load forecasting, 63
Intelligent protection relays, 213 Chinese electric power systems, 323
Internal model control (IMC) method, 363 dynamic decomposition process, 323
International Electrotechnical Commission Kohonen machine, 324
(IEC) three ratios, 366 MLP, 333
Internet, 265, 266 Load forecasting methodology, 273
Internet of Things (IoT), 265 Load pattern classification
Intrinsic mode functions (IMFs), 364, 365 fine-grained metering data, 266
IPDT GMM, 267
over-/under-fitting evaluation of, 309, 310 HMM, 267
prediction of instability by, 299, 301, 302 smart grid, 265
training, 306–308 smart meters (see Smart meter data)
transmission system and distribution
network operators, 266
K zero-normalization consumption time
Kalman filters, 250, 322, 330, 332, 333, 335 series, 267
Kalman forecasting machine, 327 Load prediction, 237
Kalman machines, 322 Load shedding, 212–214, 216, 225, 226, 231
Khaf area, 251 Load-forecasting algorithm, 323
K-means, 69, 70, 272 Load Tap Changing (LTC), 31
Index 385

Local learning algorithm (LLA), 371 LSTM, 52–55

Locality sensitive hashing approach, 272 results, 55
Locational marginal prices (LMPs) SVM, 51
ANFIS machine, 327 MLP (see Multilayer perceptron (MLP))
average, 330, 333, 334, 337, 338, 340 NN (see Neural network (NN))
buses, 324 power system (see Power system)
dependency, 327 power system planning, 56
nonconstant average and variance, 324 ARIMA, 56
and system load, 328 EnergyScope model, 57
and topology historical database, 324 PSO method, 56
Long short-term memory (LSTM) robust optimization method, 56
applications, 240, 241 stochastic programming method, 56
block diagram, 242 XGBoost method, 57
CC, 243 regression and classification applications,
dataset, 241 74–76
error values, 243–245 reinforcement learning, 47
fossil fuel-based sources, 237 supervised learning, 47
heating consumption of buildings, 240 supervised, semi-supervised and
heating load prediction results, training unsupervised, 254
data, 243 support vector machine, 48
heating load prediction, test data, 243, 244 SVR, 251
mathematical structure, 241 tree-based, 251
MSE, 243, 245 unsupervised learning, 47
parameters, 243 Machine learning algorithm
photovoltaic energy production, 240 ANN, 350, 351
pinball loss, 240 day of the week, 351
RMSE, 243, 245 dynamic parameters (see Dynamic
RNN, 240, 241 parameters)
time series data, 241 outdoor temperature, 351
training phase, 243 RF, 350, 351
wind energy prediction, 240 scheduling window, 350
Long-term load forecasting (LTLF), 126 start time of the task, 351
Low-rank matrix fitting-based approach, 272 supervised learning (see Supervised
Low-rank matrix technique, 272 learning)
Lyapunov method, 86 SVM, 350, 351
Lyapunov paradigm, 360 user preferences (see User preferences)
Machine learning and deep learning
methods, 127
M load forecasting categories, 129
Machine learning (ML) LSTM method, 128
ANN, 251 statistical evaluation metrics., 130
applications, 13 Machine learning methods
and artificial intelligence, 250 in energy engineering
challenges, 14 application, formulation and
data patterns, 254 structure, 64
and deep learning, 251 data mining and applications, 62, 63
Gaussian process, 251 decision tree, 69
GMDH (see Group method of data handling general regression neural network,
(GMDH)) 68, 69
high-dimensional data processing, 254 group method data handling, 66, 67
load forecasting, 50 k-means, 69, 70
hidden connection mechanism, 51 support vector machine, 64, 65
input attention mechanism, 51 support vector regression, 67, 68
386 Index

Machine learning methods (cont.) Metering data, 269

power control systems Micro-energy networks, 84
application of, 100, 101 Micro generation network, 266
emergency mode, 98, 99 Microgrids, 84, 357
frequency stability assessment, 106–108 Midterm load forecasting (MTLF), 124
normal mode, 97, 98 MIM
operating modes, evaluation of, 94, 95 definition, 218
power quality disturbances assessment, learning machine, 220
105, 106 mathematical formulation, 220
preventive mode, 96 mutual information theory, 233
restoration mode, 98–100 operating variables, 226, 227, 230
transient stability assessment, 101, 102 Model reference adaptive controller
voltage stability assessment, 102–104 (MRAC), 361
Machine learning/deep learning Morphological median filter (MMF), 26
power control systems Multi-input multi-output (MIMO) system, 251
energy storage technologies, 83 Multilayer feed-forward neural network
Flexible Alternating Current (MFFNN), 124, 367
Transmission Systems, 84 Multilayer perceptron (MLP), 238, 250, 251,
loads of, 84 254, 255, 257–260, 349
micro-energy networks, 84 and ANFIS, 323
modernization of, 84 forecasted load, 336
reliability and security, 85, 86 neural networks, 322, 333, 335
stability, 86 Multilayer perceptron neural network (MLPN),
Machine learning technique 362
power control systems Multi-objective grasshopper optimization
categorization of, 90, 91 algorithm (MOGOA), 250
exegesis and validation, 88 Mutual information theory, 213, 218–220,
feature selection, 88 226, 233
model selection, 88
model using, 88
practical limitations, 89 N
Markov sequence model, 213 Naïve Bayes, 271, 276, 277, 280, 281
Maximum power point tracking (MPPT), 97 Neighborhood area network, 266
Mean absolute error (MAE), 243, 244, 257 Networking systems, 266
Mean absolute percentage error (MAPE), 368 Neural network (NN)
ANN-based LPF, 322 ANN, 359, 360, 362, 364–369, 372
Australian electricity market, 323 ANNCs, 359
Chinese power system, 323 ANNMFC, 362
hour-ahead and day-ahead price BP, 366
forecasting, 323 BPNNI, 363
load and price forecasting, 324 feed-forward, 359
maximum and minimum values, 333 FNN, 360
neural network, 337 GNN, 363, 368
New South Wales electricity load IHNN, 370
forecast, 323 MFNN, 367
New South Wales market, 323 MLPN, 362
New York power system, 324 NNC, 362
PJM market, 322 OTSRWNNC, 362
Spanish electricity market, 323 PNN, 366, 371
Mean bias error (MBE), 75 RBFNN, 360, 362, 363, 365–367, 369, 370
Mean square error (MSE), 75, 240, 243–245, RBFNNI, 363
251, 257, 260, 323 RNN, 360
Measuring electricity, 282
Index 387

SCFNN-based static synchronous series Polynomial neural network, 255

control (SSSC), 360, 362 Positive Shannon function, 219
WNN, 360 Power consumption, 279–281
Neural network controller (NNC), 362 Power consumption-related characteristics, 277
Neural/genetic-based algorithm, 270 Power control systems
New England Test System (NETS), 155 challenges and comparative analysis,
New York Power System (NYPS), 155 108–110
Newton-Raphson-based OPF, 370 deep learning, 89–91
NN inverse (NNI) technique, 363 machine learning methods
Non-critical, 218, 220, 223, 226, 228, 231, 233 application of, 100, 101
Nonlinear autoregressive exogenous model emergency mode, 98, 99
(NARX), 238 frequency stability assessment, 106–108
Nonlinear regression algorithm, 255 normal mode, 97, 98
operating modes, evaluation of, 94, 95
power quality disturbances assessment,
O 105, 106
O.S method, 298 preventive mode, 96
O.S prediction, 291, 293, 295, 296, 304 restoration mode, 98–100
Oil reserves, 265 transient stability assessment, 101, 102
Online optimal topology-forecasting voltage stability assessment, 102–104
procedure, 328 machine learning technique
Online sequential extreme learning machine exegesis and validation, 88
(OSELM), 238 feature selection, 88
Online trained self-recurrent wavelet neural model selection, 88
network controller model using, 88
(OTSRWNNC), 362 practical limitations, 89
Optimal power flow (OPF), 370 machine learning/deep learning
Optimal scheduling energy storage technologies, 83
smart appliances (see Smart appliances) Flexible Alternating Current
Optimal topology, 324, 327, 328, 335, 338, 340 Transmission Systems, 84
Optimized ELM, 250 loads of, 84
Oscillatory stability margin (OSM), 179 micro-energy networks, 84
modernization of, 84
reliability and security, 85, 86
P stability, 86
Particle swarm optimization (PSO), 56, problems, 91, 93, 94
346, 349 Power grid, 357
Pattern identification, 271 Power quality disturbances assessment
Pattern recognition, 271–273 power control systems
Performance validation, 231 machine learning methods, 105, 106
Permanent magnet in-wheel motor Power system, 290
(PMIWM), 363 control, 31, 34
Permanent magnet synchronous motor control-themed applications, 358–363
(PMSM), 362, 363 detection and protection, 363–367
Perturb-and-observe (P&O) maximum power distribution system, 357
point tracking (MPPT) approach, electricity generation, 358
371 function and protection, 367–370
Phasor Measurement Units (PMU), 9, 84, 95, future challenges
152, 292, 293 big data, 11
Photovoltaic (PV) systems, 365 electric vehicles, 7
Photovoltaic energy production, 240 energy storage technologies, 6
Pinball loss, 240 greenhouse gas emissions, 3, 4
PJM market, 322–324 load forecasting techniques, 11
388 Index

Power system (cont.) Radial basis function neural network inverse

renewable energies, 5 (RBFNNI), 363
smart contract structures, 8 Random Chemistry search algorithm, 212
storage benefits, 7 Random forest (RF), 349–351, 353
transient stability assessment, 9 Real-life data converters, 278
generators, 357 Real-time applications, 351
goal of, 20 Real-time closed-loop wide-area decentralized
houses and industry, 357 power system stabilizer
microgrids, 357 (WD-PSSs), 362
operation challenges Real-time system monitoring models, 369
climate changes, 25 Recurrent neural network (RNN), 72, 73, 240,
cyberattack, 29, 31 241, 360
unexpected events, 26 Recurrent support vector machines with genetic
optimization, 370–372 algorithms (RSVMG), 368
planning Regression machines, 323
distributed generation, 23 Reinforcement learning, 90
generation expansion planning, 21 Reinforcement learning (RL) approach, 362
line expansion planning, 21 Relative Absolute Error (RAE), 75
network expansion planning, 23 Relative humidity, 251, 252, 257
substation expansion planning, 22 Relevance vector machine (RVM), 366
power grid Remote Terminal Units (RTU), 9
extreme contingencies, 35 Renewable energy sources (RESs)
frequency control, 35 distribution systems, 249
regular contingencies, 34 energy sector, 250
voltage and angle control, 37, 38 wind energy, 250
protection, 370–372 wind speed, 250
safety and continuity, 357 Residential load scheduling problems, 351
supply mechanism, 358 Restricted Boltzmann machine (RBM), 73, 74
transmission system, 357 Robust pole placement (RPP), 363
unstable, 358 Root mean square error (RMSE), 243–245,
working conditions and uncertainties, 358 257, 260
Power system frequency, 358 Rotor angle stability, 290
Power system stabilizers (PSSs), 358
Power systems, 1, 219, 290
development, 212 S
dimensions, 211 Seasonal autoregressive integrated moving
operating variables, 212 average (SARIMA), 238
transmission lines, 211 Self-constructing fuzzy neural network
Principal component analysis (PCA), 351, 365 (SCFNN)-based static synchronous
Probabilistic neural network (PNN), 366, 371 series control (SSSC), 360, 362
Prony analysis-Hilbert transform method, 291 Self-organizing maps (SOM), 272
PSCC, 212, 214, 216, 225, 228, 230, 233 Self-tuned FLPSS, 361
Shallow learning, 240
Short-term forecasting, 250
Q Short-term forecasting literature, 125
Quadratic programming (QP), 370 Short-term load forecasting (STLF), 122, 367
Quantization scheme, 285 Short-term wind power forecasting, 251
Short-term wind speed forecasting, 251
Short-time Fourier transform (STFT), 270
R Signal reconstruction, 274
Radial basic function (RBF) networks, 68, 365 Simulation
Radial basis function neural network (RBFNN), synchronous generators, out-of-step
360, 362, 363, 365–367, 369, 370 condition, 302
Index 389

CDDT, 305 energy disaggregation, 283

DT performance with SVM and event-driven sensing mechanism, 284
ANN, 310 event-driven signal acquisition and
DT, producing training scenarios of, processing technique, 284
303, 304 framework, 267
FDDT training, 304, 305 granular energy usage, 266
IPDT training, 307, 308 home appliances, 268
IPDT, over-/under-fitting evaluation of, installations, 282, 283
309, 310 load disaggregation flow chart, 268, 269
time windows data, 304 measurements, 268
Sin-cosine optimization, 323 measuring electricity, 282
Single-machine infinite bus power system, 361 principle, 273
Single-machine infinite bus system (SMIB) processing and power consumption, 284
excluding PSS, 361 residential area, 267
Single variable entropy functions, 220 solar and wind, 267
Smart appliances stakeholders, 282
activity prediction of residents, 346 state-of-the-art methods, 284
central control unit, 346 technological developments, 266
comfort level, 346 time-invariant, 283
communication capability, 346 time-invariant fixed-rate manner, 267
constraints, 346 wide area network, 266
data motion sensors, 346 Smart meters, 346, 347
dataset, 350 Solar radiation, 251–253, 257, 259
decision boundaries, 351 Solar water heater (SWH) model, 372
door sensors, 346 Split information (SI), 217
dynamic parameters, 346 Stability
energy consumption profile, 353 power control systems, 86
fixed parameters, 346 Start time of the task, 347, 348, 353
hardware layer, 345 Static Security Assessment (SSA), 85
heuristics algorithms, 346 Stationary WT (SWT), 364
home energy management systems, 345 Statistical evaluation, 76
on/off command, 352, 353 Summation wavelet extreme learning machine
PCA, 351 (SW-ELM), 364
scheduling window, 350, 351 Supervised learning, 90, 348–350
smart meters, 346 Supervisory control and data acquisition
thermostatic loads, 347 (SCADA), 9, 251
training and test sets, 350 Support vector machine (SVM), 64, 65, 213,
training set, 352 238, 250, 272, 277, 291, 293, 310,
users’ preferences, 347 322–324, 349–351, 353, 364, 366,
Smart grid, 265, 266 370, 372
Smart meter data Support vector regression (SVR), 67, 68, 238,
advantages, 267, 268, 282 250, 251, 322, 364, 367, 368
ANN, 268, 269 SVM with genetic algorithm (SVMG), 366
applications, 267, 268 Synchronous generators (SG), 362
benefits, 282, 283 out-of-step condition, prediction of
characteristics, 268 CDDT, 305
components, 266 CDDT, fault clearance detection by,
consumption, 269 300, 301
data acquisition, 269, 270 decision tree, 297, 298
data collection, 266, 282 DT performance with SVM and
database, 273, 274 ANN, 310
DSOs, 283 DT, nature of operating variables used
elements, 282 by, 302
390 Index

Synchronous generators (SG) (cont.) V

DT, producing training scenarios of, Variable identification, 218
303, 304 Venn diagram, 220
FDDT training, 304, 305 Voltage and angle control
FDDT, fault detection by, 298, 300 AVR controls, 37
framework of, 294–296 FACTS, 38
IPDT training, 307, 308 grid parameters, 38
IPDT, over-/under-fitting evaluation of, new control systems, 38
309, 310 novel security idea, 38
IPDT, prediction of instability by, 299, PSS control, 37
301, 302 set points and input signals, 38
proposed method, 296, 297 traditional linear contingency, 38
scheme performance, validation of, Voltage stability, 369
310–312, 314, 316 Voltage stability assessment
simulation, 302 analytical methods, 180
time windows data, 304 decision trees, 179
validation with noisy data, 316, 317 FACTS devices, 179
Synchronous machine (SM), 358 ML techniques
database generation (see Database
generation approach)
T dimensionality reduction, 189–191
Test set, 349 ensemble methods (see Ensemble
Thermostatic loads, 347 methods)
Three-step DT, 293 PMU measurements, 183
Time series data analysis, 238 proposed framework, 182
Time series decomposition, 323 online application, 193
Time series models, 273 performance evaluation
Time series of energy consumption, 273 confusion matrix, 192
Time-domain calculations, 225 misdetection, 192
Time-invariant, 283 PMU measurements, 193
Time-window-based approach, 272 PMU data, 180
Topology-forecasting process, 324 P-Q-V curve base technique, 178
Trace-base repository, 271 proposed framework, 180
Traditional combination forecasting method P-V curve operation state, 181
(TCFM), 250 steady-state power system model, 180
Training process, 348 support vector machine, 179
Training set, 348 Voltage stability evaluation models, 104
Transient stability, 86, 290–294, 303 Voltage stability margin (VSM), 104, 178, 367
Transient stability assessment Volterra series, 66
power control systems, machine learning
methods, 101, 102
Transimpedance mountings, 278 W
Transmission system, 357 Wavelet neural network (WNN), 360
Tree-based machine learning method, 251 Wavelet transform (WT), 250, 322–325,
328, 363
Wavelet transformation, 270
U Weather forecasting model, 251
Unstable power system, 358 Weighted least squares (WLS), 369
Unsupervised learning, 90 Wide area measurement system (WAMS), 101
User preferences Wide-area control system (WACS), 360
flexible loads, 347 Wide-area measurement system (WAMS), 221,
in optimal scheduling (see Optimal 222, 228, 231, 233, 292, 362
scheduling) Wind direction, 251–253, 257, 259
Index 391

Wind energy, 250, 257, 258 clustering approach, 138

Wind energy prediction, 240 clustering approaches
Wind speed, 250, 254 EECS algorithm, 142
Wind speed forecasting H-PEGASIS algorithm, 143
air temperature, 251, 252 LEACH, 142
ANN, 251 LEACH-C, 142
ARIMA, 250, 251 PEGASIS algorithm, 143
ARMA, 250 collision, 140
data mining (see Data mining) control packets, 140
definition, 250 interference, 140
Khaf area, 251 physical sensing device, 139
mathematical modelling, 251 processor unit, 139
MLP, 254, 255 PSO-based approach
MOGOA, 250 EMPS approach, 145
optimized ELM, 250 GA-PSO approach, 146
parameter, 257 PECC approach, 144
performance evaluation, 258, 260 PSO-C approach, 143
physical methods, 250 PSO-ECHS approach, 145
power producers, 249 PSO-ECSM approach, 146
relative humidity, 251, 252 PSO-HC approach, 144
RESs (see Renewable energy sources PSO-HSA approach, 145
(RESs)) PSO-SD approach, 144
RWT-ARIMA, 250 PSO-UFC approach, 146
SCADA, 251 PUDCRP, 146
short-term, 251 SCE-PSO approach, 145
short-term wind power forecasting, 251 sensor node, 139
simple average strategy, 250 swarm intelligence
solar radiation, 251–253 labor division, 140
system’s nonlinearity, 250 self-organization, 140
TCFM, 250 transmission process, 142
training and test stages, 257
training phase, 257
weather forecasting model, 251 Z
wind direction, 252, 253 Zero-normalization consumption time
Wind turbine, 213 series, 267
Wireless sensor networks (WSNs), 137 ZigBee protocol, 266
challenging issue, 137