Volume 20 • Number 3 • May/June 2022

magazine

Volume 20 • Number 03 • may/june 2022

www.ieee.org/power

on the 64
cover
INDUSTRIAL INTERNET OF THINGS—©SHUTTERSTOCK.COM/NADYA C
ARTIFICIAL INTELLIGENCE—©SHUTTERSTOCK.COM/OMELCHENKO

©SHUTTERSTOCK.COM/METAMORWORKS
features

contents
14 A
 rtificial Intelligence 54 Clustering Electricity Consumers
for Load Forecasting By Andrés M. Alonso, Eduardo Martín,
By Tao Hong and Pu Wang Alicia Mateo, Francisco J. Nogales,
Carlos Ruiz, and Andrea Veiga
24 Electricity Price Forecasting
By Arkadiusz Je˛drzejewski, Jesus Lago, 64 Data-Driven Dynamic Modeling
Grzegorz Marcjasz, and Rafał Weron in Power Systems
By Lingling Fan, Zhixin Miao, Shahil Shah,
32 Machine Learning in Power Systems Przemyslaw Koralewicz, Vahan Gevorgian,
By Spyros Chatzivasileiadis, Andreas Venzke, and Jian Fu
Jochen Stiasny, and Georgios Misyris

42 D
 emonstration of Intelligent HVAC
Load Management With Deep
Reinforcement Learning
By Yan Du, Fangxing Li, Kuldeep Kurte,
Jeffrey Munk, and Helia Zandi

columns &
departments
4 Editors’ Voice
8 Leader’s Corner
12 Guest Editorial
77 History

77 86
88
Book Review
Letters to the Editor
89 Calendar
92 In My View
Digital Object Identifier 10.1109/MPE.2022.3158445

may/june 2022 ieee power & energy magazine 1

 magazine

Editor-in-Chief Editorial Board IEEE Periodicals/Magazines Department

Steve Widergren L. Anderson, L. Barroso, H. Chen, A. Conejo, 445 Hoes Lane, Piscataway, NJ 08854 USA
[email protected] M. Correia de Barros, J. Feltes, N. Hatziargyriou, Kristin Falco LaFleur, Journals Production Manager
H. Holtinnen, T. Hong, S. Imai, B. Johnson, C.-H. Kim, Patrick J. Kempf, Senior Manager, Production
Past Editors-in-Chief B. Kroposki, D. Lew, N. Lu, J. Matevosyan, B. Mather,
Janet Dudar, Senior Art Director
Michael Henderson M. Miller, D. Novosel, M. O’Malley, N. Ochoa,
A. Orths, J. Paserba, C.E. Root, H. Rudnick, Gail A. Schnitzer, Associate Art Director
Melvin Olken Theresa L. Smith, Production Coordinator
E. Uzunovic, S.S. Venkata, J. Wang
Associate Editors Felicia Spagnoli, Advertising Production Manager
John Paserba, History Spanish Editorial Board Peter M. Tuohy, Production Director
Antonio Conejo, Ning Lu, Barry Mather, Issues Enrique Tejera, Editor-in-Chief Kevin Lisankie, Editorial Services Director
Jianhui Wang, Submissions Editors: M. Baquedano, J. Cerda, G. Gonzalez Dawn M. Melley, Senior Director, IEEE Publishing
Operations
Advertising
Assistant Editor Beverly Grimme, Naylor Association Solutions
Susan O’Bryan direct: +1 352 333 3367, cell: +1 904 881 0862, IEEE prohibits discrimination, harassment, and bullying. For more infor-
fax: +1 703 790 9199, [email protected] mation, visit http://www.ieee.org/web/aboutus/whatis/policies/p9-26.html.

IEEE POWER & ENERGY SOCIETY (PES)

The IEEE Power & Energy Society is an organization of IEEE members whose principal interest is the advancement of the science and practice of electric power generation,
­transmission, distribution, and utilization. All members of the IEEE are eligible for membership in the Society. Mission Statement: To be the leading provider of scientific and
­engineering information on electric power and energy for the betterment of society, and the preferred professional development source for our members.

Governing Board Technical Council IEEE Trans. on Energy Conversion, A. Tessarolo

J. Bian, President H. Chen, Chair; D. Watkins, Vice Chair IEEE Trans. on Power Delivery, F. DeLeon
S. Bahramirad, President-Elect J. McBride, Secretary; V. Vittal, Past-Chair IEEE Trans. on Power Systems, J. Milanovic
M. Armstrong, Vice President, Chapters IEEE Trans. on Smart Grid, C. Canizares
H. Chen, Vice President, Technical Activities Technical Committee Chairs IEEE Trans. on Sustainable Energy, B. Chowdhury
B. Enayati, Vice President, Education K. Cheung, Analytical Methods for Power Systems IEEE Open Access Journal of Power and Energy, F. Li
B. Pal, Vice President, Publications J. Lau, Electric Machinery IEEE Power & Energy Magazine, S. Widergren
W. Bishop, Vice President, Meetings R. Thornton-Jones, Energy Development & Power eNewsletter, S. Fattah
J. Romero Aguero, Vice President, Membership Generation Editor-in-Chief at Large, W. Xu
& Image B. Chalamala, Energy Storage & Stationary Battery Marketing, E. Batzelis
N. Hadjsaid, Vice President, New Initiatives/ Y. Wen, Insulated Conductors PES Representative to IEEE Press, K. Bhattacharya
Outreach J. White, Nuclear Power Engineering Website, R. Rana
J. Montero, Treasurer C. Preuss, Power System Communications
J. Sykes, Secretary & Cybersecurity Meetings
S. Rahman, Past-President L. Lima, Power System Dynamic Performance Committee Chairs
D. Diaz, Region Rep., U.S. & Canada E. Hanique, Power System Instrumentation General Meeting Steering, A. St. Leger
I. Kuzle, Region Rep., Europe, Middle East, & Measurements Joint Technical Committee Meeting Steering, D. Sabin
& Africa J. Wang, Power System Operation Planning Technically Cosponsored Conferences Steering,
S. Leon, Region Rep., Latin America & Economics A. Borghetti
D. Sharafi, Region Rep., Asia & Pacific M. Yalla, Power System Relaying & Control Transmission & Distribution Conference &
G. Gonzales, Member-at-Large J. Mathieu, Smart Building, Load & Exposition Committee for North America
R. Kappagantu, Member-at-Large Customer Systems Steering, C. Segneri
A. Sannino, Member-at-Large P. Fitzgerald, Substations Innovative Smart Grid Technologies North America,
J. Yu, Member-at-Large S. Hensley, Surge Protective Devices R. Melton
K. Flowers, Switchgear Innovative Smart Grid Technology Conference­­–
IEEE Division VII Director E. teNyenhuis, Transformers Europe, D. Van Hertem
C. Cañizares S. Santoso, Transmission & Distribution Innovative Smart Grid Technology Conference–
IEEE Division VII Director-Elect Asia, D. Sharafi
Open Technical Council Coordinating Committees PowerAfrica Steering Committee, H. Louie
H. Sun, Energy Internet Website, Open
PES Executive Director J. Follum, Intelligent Grid & Emerging
T. Licitra Technology Education
D. Alexander, Marine Systems Committee Chairs
Standing Committee Chairs A. Leon, Renewable Systems Integration Power & Energy Education, A. Srivastava
Open, Constitution & Bylaws PES Scholarship Plus, J. Hoffman, C. Root
M.T. Correia de Barros, Fellows Technical Council Standing Committees PES University, L. Hennebury
N. Hadjsaid, Finance V. Vittal, Awards Website, J. Rickson
A. Apostolov, History Open, Industry Education
J. Bian, Long-Range Planning J. McBride, Organization & Procedures New Initiatives and Outreach
S. Rahman, Nominations & Appointments T. Irwin, Standards Coordination Committee Chairs
D. Watkins, Technical Sessions Corporate Engagement Program, M. Vaiman
Region Representatives Open, Webmaster Diversity, J. McClanahan
D. Diaz, M. Gosalia, B. Gwyn, J. Khan, Executive Advisory Council, K. Curtis
I. Kuzle, S. Leon, J. Milanovic, M. Papic, V. Vittal, Entity Proposal Management Committee
IEEE Smart Cities, G. Zissis
D. Sharafi, C. Wong, X. Wu Publications IEEE Smart Grid, P. Wung
Chapter Committee Chairs Publications Board Chair, B. Pal IEEE Smart Village, J. Nelson
C. Diamond, Awards & Resources Editors-in-Chief Industry Tech Support Leadership, D. Novosel
Z. Wang, Chapters Website IEEE Electrification Magazine, L. Fan Media Engagement, D. Kushner
N. Nair, Distinguished Lecturer Program IEEE Power Engineering Letters, R. Jabr Website, K. Anastasopoulos
Z. Wang, Electronic Communications
T. Ribeiro de Alencar, Student Chapters Coordinator IEEE Power & Energy Magazine
Membership & Image Committee Chairs IEEE Power & Energy Magazine (ISSN 1540-7977) (IPEMCF) is published bimonthly by the Institute of Electrical and
A. Vaccaro, Awards & Recognition ­Electronics E­ ngineers, Inc. Headquarters: 3 Park Avenue, 17th Floor, New York, NY 10016-5997 USA. Responsibility for the
S. Fattah, Humanitarian Activities contents rests upon the authors and not upon the IEEE, the Society, or its members. IEEE Operations Center (for orders, sub-
J. Benedict, Marketing scriptions, address changes): 445 Hoes Lane, Piscataway, NJ 08854 USA. Telephone: +1 732 981 0060, +1 800 678 4333.
M. Seif, Membership Development Individual copies: IEEE members US$20.00 (first copy only), nonmembers US$89 per copy. Subscription Rates: Society members
included with membership dues. Subscription rates available upon request. Copyright and reprint permissions: Abstracting is permit-
L.F. Gaitán, Social Media
ted with credit to the source. Libraries are permitted to photocopy beyond the limits of U.S. Copyright law for the private use of
J. Cardoso, Web Site Development patrons 1) those post-1977 articles that carry a code at the bottom of the first page, provided the per-copy fee indicated in the code
R. Li, Women in Power is paid through the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923 USA; 2) pre-1978 articles without fee.
S. Akar, Young Professionals For other copying, reprint, or republication permission, write Copyrights and Permissions Department, IEEE Operations Center, 445
C. Naik, PES Day Hoes Lane, Piscataway, NJ 08854 USA. Copyright © 2022 by the Institute of Electrical and Electronics Engineers, Inc. All rights
reserved. Periodicals postage paid at New York, NY, and at additional mailing offices. Postmaster: Send address changes to IEEE
Digital Object Identifier 10.1109/MPE.2022.3158446 Power & Energy Magazine, IEEE Operations Center, 445 Hoes Lane, Piscataway, NJ 08854 USA. Canadian GST #125634188
 Printed in U.S.A.

2 ieee power & energy magazine may/june 2022

 ULTRA-FAST APPROXIMATION OF
LOAD FLOW VARIABLES USING
NEURAL NETWORKS
The new Artificial Intelligence module in PowerFactory offers an innovative method
of training neural networks to estimate the results of a large number of different load
flow scenarios, so speeding up load flow calculations considerably.
 Dataset generation, neural network training and use of the neural network are all realised
within PowerFactory
 Generation of training dataset using PowerFactory’s inbuilt Probabilistic Analysis
 Use of distribution and parameter characteristics for the dataset generation, allowing
historic data and other information about the electrical network to be taken into account
in a comprehensive dataset
 Explicit selection of training output variables, to focus on specific load-flow quantities
 Application of pre-trained neural networks in Quasi-Dynamic Simulations, for ultra-fast estimation
of network variables
 Training dataset and neural network are stored externally and so can easily be shared with
other users

For more information visit POWER SYSTEM SOLUTIONS

www.digsilent.de/pf-ai
MADE IN GERMANY
Serving more than 160 countries.
 editors’ voice
Ning Lu and Steve Widergren

a long road to singularity

making inroads with artificial intelligence

D
DID YOU HEAR THE ONE ABOUT In This Issue transform our society, and our way of
the self-driving car that wrote its auto- In the past decade, we have witnessed the living, working, and entertaining.
biography? I hear the groans, but in- rise of AI. Thanks to the rapid advance- However, thus far, when managing
troducing an issue devoted to artificial ment in data science, computing hard- and operating critical infrastructures,
intelligence (AI) and machine learn- ware, and the theory of computation, re- the acceptance of AI remains limited
ing (ML) needs an icebreaker. IEEE searchers have made a quantum leap in to applications in low-risk, low-im-
Power & Energy Magazine topics tend training AI applications to play complex pact, labor-intensive, decision-support
to focus on power system challenges board games, process images, and per- systems. For example, AI adoption in
and then the cool techniques or tech- form a wide range of natural language power grid operation has been limited
nologies being applied to address them. processing tasks such as translation, to planning and predictive applications
With the topics of AI and ML, this is- summarization, classi- such as load forecast-
sue inverts the paradigm to explore the fication, and extraction. ing, anomaly detection,
data science and methodology under- In recent years, with one
As professionals secu r ity assessment,
lying AI as a disciplinary toolbox for success after another, and members synthetic data genera-
power system applications. AI-powered software tion, and system model
As professionals and members of has demonstrated its of the digital identification. This is
the digital age, we cannot avoid ben- exceptional potential because the critical na-
efiting from AI in our daily lives and in social media, enter-
age, we ture of secure and reli-
being exasperated by its shortcom- tainment, health care, cannot avoid able operation of power
ings. Examples include search engines, education, finance, sur- grids demands trust in
digital assistants interpreting speech, veillance, robotics, and benefiting from decision making in all
call- and web-based chat services, and transportation. The list situations, whether pre-
autonomous, intelligent controls in all grows longer each year.
AI in our daily viously seen or complete-
sorts of devices and systems. Encouraged by the lives and being ly unexpected.
AI is beginning to penetrate power a c h ieve m e n t of A I Integrating AI into the
system operations and planning as programs with perfor- exasperated decision-making process
well. It is a complex field full of jargon mance comparable or requires human operators
such as feature selection, hidden lay- superior to their human
by its to understand the reasons
ers, learning rates, and test and train- counterparts in data-in- shortcomings. behind AI-generated de-
ing data. Although a single magazine tensive tasks, research- cisions. Awareness of the
issue cannot make one an expert in AI ers in many domains consequence and risk of
and ML, the contributors to this is- are rethinking standard techniques for each AI-driven action needs experiential
sue provide insights that peel back the data processing, information extraction, support to engender faith in the technol-
cover to explain a bit of how AI works, model development, and automation. ogy. Establishing trust between human
where techniques are being adapted to Optimists believe that the fourth techno- operators and AI-powered applications
power system applications, and where logical revolution, fueled by AI-infused requires proof of interpretability, general-
it struggles to gain trust. systems (i.e, systems with AI incorporat- ization, and robustness of their decisions.
ed into the decision-making process), is To introduce the status of AI ap-
Digital Object Identifier 10.1109/MPE.2022.3150815
underway. If true, the breadth and depth plications in the power industry, Guest
Date of current version: 19 April 2022 of upcoming changes will profoundly Editors Tao Hong and Fangxing Li

4 ieee power & energy magazine may/june 2022

 have organized six articles written by regulation, and adaptation. He then his leadership role. As the Roman god
experts around the world to cover the offers his predictions for the next Janus looks in two directions, Licitra
following topics: disruptive phase, fueled by the trio discusses this transition period reflect-
✔ A brief introduction of the his- of 5G networks, AI, and distributed ing on the Society’s future directions
tory of AI digital ledger technology. in context of the groundwork laid by
✔ Development of AI-based load Although the material in this issue past actions. Focusing on clear goals
forecasting tools cannot paint a complete picture of the for improvement, Licitra takes stock
✔ Illusions and opportunities for surging AI research and application of the Society’s accumulated assets, its
applying AI-based techniques in activities in the power engineering accomplishments, and strong support
load forecasting domain, it gives our readers glimps- from volunteers.
✔ Roadblocks and possible solu- es of successful stories and lessons
tions for ML adoption in grid learned. By sharing experts’ insights Letters
operation on what works and what does not, the A colleague recently found out the iden-
✔ Examples of AI applications in issue will hopefully spark discussions tity of the editor-in-chief of IEEE Power
• security assessment on AI-infused grid technology devel- & Energy Magazine. He opened the lat-
• load classification and custom- opment. By raising awareness of AI est issue and commented, “Sure enough,
er segmentation technologies, may the electric power there’s your name.” Despite the effort
• demand response profession better embrace AI and brace that goes into publishing this magazine
• dynamic model identification for its impacts. from so many contributors, one is never
✔ An outlook on AI technology. sure if it is being read. That is until a
The “In My View” column by Hami- A Word From the note arrives from a reader with a com-
dreza Zareipour reviews the history Executive Director ment. Good, bad, or indifferent, pub-
of disruptions through the evolution This month’s Leader’s Corner features lishing a letter to the editor gives voice
of North American power systems Executive Director Tim Licitra, who to the people this magazine serves. We
over the past 140 years as charac- has quickly come up to speed with the look forward to sharing more letters to
terized by technology innovation, IEEE Power & Energy Society and the editor in future issues.

Leaders in converter control

have something in common.

They use real-time simulation

to test their systems.

How do you validate novel control systems for power electronics?

The RTDS® Simulator is the world standard for real-time digital simulation, used globally for the closed-loop testing of
power system protection and control systems. The Simulator’s coonverter models allow the user to test PWM schemes
switching in the 150 kHz range with unprecedented accuracy and fidelity. Testing controls with a simulated converter
circuit throughout development and prior to deployment is efficient, safe, and reliable.

In the hands of innovators, the RTDS® Simulator is changing the grid as we know it.

YOUR WORLD IN REAL TIME.

RTDS.COM

IEEE Power and Energy - Half Page.indd 1 2022-01-27 9:50:00 AM

History—Energy rise of intelligent, communicating energy is O’Bryan’s last issue as assistant editor,
From the Sun devices is calling for the coordination of having joined the editorial team a year
Today’s solar power discussions focus operational flexibility for supply-demand and a half ago. Throughout her tenure,
on energy conversion to electricity from balance in wholesale market design. The she has trumpeted the cause of simple,
photovoltaic systems and large solar ra- book borrows on the concept of swing understandable language while endur-
diation collection farms. But the harness- options (a contract with the right to ad- ing the hard-to-understand terminology
ing of solar energy for human develop- just the quantity received) in commodity that pervades the various specializations
ment is truly a centuries-old tradition. trading. The author carefully applies the in our industry. We wish her well in her
This month’s “History” column exam- concept to the flexibility needs arising in next phase of life.
ines the utilization of solar energy in wholesale electricity markets, consider- Thanks as well to Associate Editor
early applications as context for the rise ing appropriate contract constraints. Brian Johnson for overseeing the book
of photovoltaic development in the pres- reviews, Associate Editor John Paserba
ent era. The authors present the evolution Curtain Call for his reliable and resilient management
of seminal physical concepts that led to Great thanks to Guest Editors Hong of the “History” column, and Kristin
the evolution of various types of photo- and Li for conceiving this issue and ar- LaFleur and her professional team at
voltaic cells, with their associated vir- ranging the articles. We are also grate- IEEE Publishing.
tues and shortcomings. ful to the authors who worked hard to Singularity represents the point where
explain the complex concepts behind intelligent machines become enlightened
Swinging Book Review AI and ML and educate us on the state to control their destiny and rival human
IEEE Power & Energy Magazine past of the art of the technology and expe- civilization. Contemplating such a con-
Guest Editor and article contributor Pier- riences in power system applications. cept is scary. As this issue shows, even
luigi Mancarella provides his insights on The editorial team earned recogni- with advancements in AI, there remains
the book, A New Swing-Contract Design tion for their efforts to improve the com- plenty of work to overcome the challeng-
for Wholesale Power Markets, written munication of the concepts to our broad es the field faces. For all the speed and
by Leigh Tesfatsion. The combination of power and energy audience. Assistant excitement of technological change, may
trends in variable renewable generation, Editor Susan O’Bryan deserves to take a we find some comfort in that fact.
p&e
greater electrification of load, and the bow for her language editing skills. This

Power System Security Assessment for Grids with High

Penetrations of Renewables
Dynamic Security Assessment Software
• Comprehensive capabilities for voltage, transient, small signal, and
Offering a simple, user-friendly interface
frequency stability assessment
TM
with extensive analysis options, DSAToolsTM • Supporting generic and user-defined dynamic models for accurate modeling
is a suite of software tools for power system of aggregate renewable plants and energy storage systems
analysis. The software is designed for • Integration with EMS/WAMS for on-line dynamic security assessment to determine
applications in both off-line studies and security of real-time system conditions as well as forecasted scenarios
on-line dynamic security assessment.
• Computation of secure operation regions (stability limits) and suggestions for
remedial actions to maximize renewable usage
www.dsatools.com
Experience and support
• Prominent provider of the on-line DSA technology with over 50 installations
worldwide including 8 out of 9 ISOs in North America
• Engineers highly experienced in dynamic stability studies and providing consulting
services for system studies, custom model development, and training

VSAT TSAT DSA

Manager

Contact us:
[email protected]
 leader’s corner
Tim Licitra

a brief look back

a big look forward

I
IN JUNE 2021, I WAS LUCKY ENO­UGH As we progress through the new ✔ ✔ Transparency: All of the ef-
to be selected as the next executive di- calendar year, we are excited for the forts to gather and document
rector of the IEEE Power & Energy successes that 2022 will bring for PES. our policies and procedures are
Society (PES). Coming from an organi- There is a renewed focus in many areas being done so that we can in-
zation outside IEEE and doing so dur- that you, the membership, will experi- crease transparency and allow
ing the COVID-19 pandemic certainly ence. Some of these items include our members and volunteers to
creates a different type of transition and ✔✔ Efficiency: We are reviewing our access this information. A major
onboarding experience. However, my operations to streamline and im- goal is not only to increase trans-
experience is with a standards devel- prove efficiency. These improve- parency but to do so in a way that
opment association within the lighting ments should not only benefit is beneficial to the membership.
engineering industry, so this familiarity PES staff but also our volunteers One by-product of transparency
has aided me during this time. and members. This would im- can be an increased sense of be-
prove the support that we can longing and community. There
Future Endeavors provide and allow for greater are several areas where we be-
We have navigated many issues dur- flexibility. lieve increasing trans-
ing this transition period and started ✔✔ Operations man- parency will improve
to lay the groundwork for our future ual: If we are go-
Our goal is interaction with our
successes. Working hand in hand with ing to improve ef- to create a customer base, which
Society leadership, we have worked ficiency, we must includes, the roles and
through several issues, such as the first understand GB-approved responsibilities of each
continued impact of the pandemic on our operations. GB member position;
conferences and meetings (and the So- Although we have
operations p ol ic ie s t h a t gove r n
ciety as a whole), the unfortunate pass- documented some manual that the PES Publications
ing of PES President Frank Lambert, of our policies and Board and surrounding
and more. Together, many improve- procedures, some contains the processes; nominations
ments have been made already but have remained and appointments pro-
more so internally. We have worked to as “histor ical
policies and cess; awa r d s , confer-
improve our bylaws, the roles and re- knowledge,” and procedures of ences, Chapters, oper-
sponsibilities of our Governing Board we certainly have ating committees, and
(GB) members, the onboarding pro- not compiled them all aspects of technical activities; and
cess for new GB members, alignment into one document joint IEEE programs.
(structurally) of our committees with accessible to our
the Society. To be effective with our
the GB, and started to gather more in- members. Our transparency, we must
formation from our members and cus- goal is to create a GB-approved also improve how we communi-
tomers to better understand how we operations manual that contains cate this information. We have to
can serve you all in the future, among the policies and procedures of all make this information accessible
other items. aspects of the Society. This also on our website and shareable
includes the rules around how when changes are made. As an
Digital Object Identifier 10.1109/MPE.2022.3150813
these policies can be amended to example, we recently spent time
Date of current version: 19 April 2022 remain relevant. going through improvements

8 ieee power & energy magazine may/june 2022

 made to our policies as approved ing improvements and changes website. Due to various reasons
by the Publications Board and that also positively impact our and interruptions, we have not
compiled those changes into a nominations and appointments completed this task. Over the
document that is process as well as de- last few months, we have re-
available online. A major goal veloping future lead-
ers for t he Societ y.
viewed and secured a vendor to
Proactive commu- conduct this work. We are re-en-
nication is criti- is not only S i g nificant progress gaging volunteers to assist with
cal to increasing has been made wit h the process, and we have a goal
transparency, and to increase regard to diversity and to launch a new website by the
we look forward
to providing this as
transparency i nclusion, a nd I a m
proud to join a Society
end of 2022. This is a very large
task and it might cause some dis-
we move forward. but to do so in t hat has made t hese ruption, but the long-term bene-
✔ Diversity and in- great advances. I think fits of improved communication,
clusion: PES has a way that is we would all agree, branding and image, transpar-
both a Diversity beneficial to the however, that this is an
area where the work is
ency, accessibility, and more will
be more than worth it.
& Inclusion Com-
mittee as well as membership. never done. Our volun- ✔ Entering into the new normal:
a GB posit ion teers have this mindset Although we will, hopefully soon,
named Member At Large Diversity as well, and I am excited about begin to progress more toward
& Inclusion. Together, they work the work and improvements that increased in-person meetings,
to identify how we can increase will be made. we still have to continue to iden-
diversity, inclusion, equity, and ✔ Website: Over the last few years, tify how we can incorporate
respect in all aspects of our So- there have been efforts to begin virtual participation into our
ciety. We look forward to mak- the process of redoing the PES committee and working group

World Leader in Developing

BIGWOOD SYSTEMS, INC. Powerful and Field
Field-Proven
www.bigwood-systems.com
+1.607.257.0915 On--line and Off
On Off--line Tools
[email protected] for Power Grid Modeling, Analysis, Assessment,
NY, USA
Head Office Branch Offices Enhancement and Optimization
Ithaca, NY Taipei, TWN | Beijing, CHN
Innovation prevails!

Network Applications for the Control Center

“Bigwood Systems’ Online Voltage • Real-time and Look-ahead Static and Dynamic Security
Stability product is simply the best in the Assessment (patented screening, ranking, and detailed
field.” - CAISO Project Leader analysis)
“BSI On-line Transient Stability • Reactive Power Reserve Monitoring, Management &
Assessment is a superb product.” Enhancement
- TEPCO Manager
• Actionable Control Recommendation engine to
“The BSI VAR Management System is a
eliminate violations and mitigate unstable contingencies
best practice tool for all utilities.” • New-Generation State Estimation for Transmission and
- SERC Audit Staff
Distribution Networks with a low number of measurements
using SCADA & PMU data
• Global Optimal Power Flow (for over 250,000 control variables with
AC/DC constraints)
• Distribution Network Management Tools
• Renewable Energy Forecasting, Uncertainty Modeling,
Representation & Management
• Advanced Measurement-Based Power System Modeling
(Generation, Transmission & Load)
World-Wide Customer Base
Over 40 electrical utilities served world-wide, including long-standing
customers such as California ISO, PJM Interconnection, Tokyo Electric
Power Company, Tennessee Valley Authority, and UK Power Networks,
among others. BSI’s advanced technology, innovation and satisfied
customers are key differentiators.

10 ieee power & energy magazine may/june 2022

 meetings, Chapter meetings, and There is no shortage of work to be is always willing to raise his or her
conferences. This is not some- done, but it is important to recognize hand to help. PES (staff, volunteers,
thing that we have to consider the amazing accomplishments over the and members) has the feeling of one
for the current disruption that we years. We have a dedicated staff, and big family, and I am thrilled to be a
continue to face but also for the the teams work well together. Our vol- part of it. I cannot wait to see all the
long-term success of our meet- unteer base is nothing short of remark- great things we can accomplish togeth-
ings and events. Beyond the pan- able, and I am continually impressed er for the benefit of the global power
demic, there must be a strategic by the quality of work done. Someone and energy community. p&e
review of our different in-person

Our volunteer
base is nothing ®
short of
remarkable,
and I am
continually
Intelligent • Intuitive • Instantaneous
impressed by Efficient Modeling
the quality of · Utility scale and Distributed Generation
· Transmission
work done. · Distribution
interactions to identify how we · Commercial / Industrial
can incorporate new technolo- Powerful Analysis
gies to increase global engage-
· ANSI and IEC Short Circuit
· Protective Device Coordination
ment and participation without
cannibalizing the value of in-
person attendance. Every meet- · Power Flow
ing and/or event is different and
· Arc Flash - OSHA Compliant Up to 800 kV
· Harmonics
serves a different purpose, there-
fore, a one-size-fits-all model
does not work. · Grounding
✔ Education: PES has a tremen- · And More!
dous amount of educational
content, but education and edu-
cational needs are different for As Easy As
every single member and institu-
· Easy to Learn
CURRENT IN AMPERES X 100 AT 480 VOLTS
.5 .6 .8 1 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 100 2 3 4 5 6 7 8 9 1000 2 3 4 5 6 7 8 9 10000
1000 1000
800 TX-2 800
FLA
600 600
500 500
0.15

400 400

300 TX-2 300

1 / 1.288 MVA
13.8 - 0.48 kV

tion with which we engage. We

200 200
6%
18.71

100 BL-1 100 Upstream Upstream Electrode Est Arc Flash Working Incident
GE AKR-50 Arc Fault Arc Fault Electrode
80
TX-2 80
Bus Name Bus kV Trip Device Trip Device Equip Type Configuration Gap Boundary Distance Energy
1000 / 1288 kVA 1600/1760 Name Function (mm) (inches) (inches) (cal/cm2)

· Easy to Use
60 60
50 6% 50
.13

40 40
BUS-3 13.8 R-6 51/50 Open Air VOA 152 32.1 + 26 1.7
24

30 BL-3 30 BUS-7 13.8 R-7 51/50 Open Air VOA 152 30.5 + 26 1.5
GE AKR-30H
20 800/400 20 M-1 0.48 BL-3 Other HCB 32 31.6 + 18 3.8
1.90

MAIN SWG 3.8 R-18 51/50 Switchgear VCB + HCB 152 189.3 + 18 59.3
BL-1
10 GE MVT
MV T-Plus
MVT-Plus 10 REFINER 2.4 R-7 51/50 Switchgear VCB + HCB 104 261.9 + 18 122.8
17.58

are working to gather informa-

8 Sensor = 1600 8
TIME IN SECONDS

TIME IN SECONDS

Plug = 1600 SWG-4 0.48 R-6 51/50 Switchgear VCB + HCB 32 213.2 + 18 135.1
6 6
Cur Set = 1.1 (1760A)
Cur
.48

5 5 MCC-1 0.48 BL-2 MCC VCB 25 53.6 + 18 6.9

LT
T Band = 1 M-1
19

4 4
R R 3
STPU = 2
STPU
ST Delay
ST D elay
ela
.5 (4400A)
2.5
y = IInt
nt
BL-5
3
MCC-2 0.48 BL-3 MCC VCB 25 53.1 + 18 6.8
C-H HFD
ST
ST Delay y II²t
²t = Out PNL-1 0.208 FS-2 Panel VCB + HBB 25 48.2 + 18 7.1
2 Override
Override = 50000A 225/150 2
PNL-2 0.208 PNL-1 Panel VCB 25 57.8 + 18 7.7

1 1
.8 .8

.6 BL-3 .6

· Fast Results
GE MVT
MV T-9
MVT-9
d
.5 .5

Risk Hazar
.4 Sensor = 800 .4

tion to better serve our constit-

Plug = 800

and Shock
BL-1

Arc Flash priate PPE Required

.3 C ur Set = 0.5
Cur 0.5
5 (400A) .3
18627A
LT Band = 1 C-6
Arc Flash and Shock
.2 IInst
nst = 4 (3200A) .2
1 - 400 kcm
kcmil
mil CU
CU
Appro
TX-2
1000 / 1288 kVA Appropriate PPE Risk Hazard
Incident Energy
Boundary
Required
.1 .1
.08 INRUSH .08 Flash Hazard inches - Arc Flash rated coverall
BL-5 4’ - 0” at 18 pants or arc
arc rated 4’ - 0”
2
.06
06
C tler Hammer
Cutler H Serie
Series C .06
cal/cm Flash Hazard Boundary
.05
HFD .05 6.0 shirt and 6.0
BL-3 Arc-rated cal/cm2 at 18 inches
cover is removed
.04 .04
Frame = 225A (150AT) - Arc Flash Incident
29249A
.03 Trip = 150 .03
Hazard when Energy
kV Shock
.02 .02 0.48 Limited Approach 0.48
BL-5 kV Shock Hazard
3’ - 6” Approach when cover is removed

uencies and to ensure that we

PANEL PANEL PANEL 20574A Restricted 3’ - 6” Limited Approach
1’ - 0” 27B)
(Fed By:1’ - 0” ONLYRestricted
t Name: MCC-23A CONFIGURATION
.01 .01
.5 .6 .8 1 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 100 2 3 4 5 6 7 8 9 1000 2 3 4 5 6 7 8 9 10000 Approach
CURRENT IN AMPERES X 100 AT 480 VOLTS Equipmen SYSTEM Equipment Name:
NORMAL
VALID FOR VALID FOR NORMAL
SWG-4 (Fed By:
18A)
SYSTEM CONFIGURATION
ONLY

are generating the content—and

disseminating it—in a way that Try instantly online or download a free demo copy at:
maximizes its value to every- www.EasyPower.com/demo
one. Our committees are hard
at work reviewing our body of
knowledge from a topic-matter ®
standpoint as well as the plat-
Power made easy.
forms that we are using to make
this content available.

may/june 2022 ieee power & energy magazine 11

 guest editorial
Tao Hong and Fangxing Li

witnessing a new wave

artificial intelligence and machine learning

P
POWER ENGINEERING PROFES- collection of six articles and an “In G. Marcjasz, and R. Weron provide an
sionals are seeing another wave of ar- My View” column covering various overview of the main trends and elec-
tificial intelligence (AI), which is ar- aspects of AI and ML for several tricity price forecasting models. They
guably the second wave in the power applications, such as load and price also point out the areas to improve and
industry and the third wave in the AI forecasting, power system opera- offer a list of best practices for electric-
community since the 1940s. The tech- tions, building energy management, ity price forecasters.
nological advancement in AI has been customer segmentation, and dynamic Many ML techniques have been
making it a universal solution to just modeling. The issue covers a histori- considered as black-box techniques.
about everything. Compared to their cal perspective and lessons learned as Such characteristics prevent them from
human counterparts, in some applica- well as the latest thinking. being widely adopted for many power
tions, the AI-based solutions are more Load forecasting is one of the first systems applications in practice. In
accurate, faster in execution, and have power system applications of AI, dat- the third article, “Machine Learning
a higher tolerance level for tedious jobs ing back to the late 1980s. During the in Power Systems: Is It Time to Trust
and extreme environments. past three decades, people have seen a It?” S. Chatzivasileiadis, A. Venzke,
As people are getting used to AI in significant gap between the academic J. Stiasny, and G. Misyris attempt
their daily life, readers of this maga- literature and industry practice, which to encourage trust in artificial neural
zine may wonder how AI and machine leads to various misunderstandings networks by proposing two direc-
learning (ML) affect the present and and confusion about what AI can do tions: neural network verification and
future operation of the electric power to load forecasting models. The first physics-informed neural networks. The
grid. Putting all bluffs and marketing article, “Artificial Intelligence for former aims at providing guarantees
buzzwords aside, AI and ML tech- Load Forecasting: History, Illusions, about the performance of a trained
niques have already been adopted in and Opportunities,” by T. Hong and P. neural network, while the latter focuses
real-world power industry applications, Wang, examines five illusions associ- on how to take advantage of physical
though not as many as many people ated with developing AI-based models. power system models.
have thought or claimed. Some readers They also present clarifications to each Building energy management has been
may be familiar with AI-based com- illusion to help improve the efficiency a challenging problem due to possible
mercial solutions for load forecasting. of AI for load forecasting. measurement and prediction errors and
Many readers may be interested to un- Restructuring of the energy sector the lack of generalization. ML-based
derstand what other power systems ap- in the 1990s resulted in deregulated approaches have gained great interest in
plications AI can handle, whether we energy markets, which made electricity this area in the smart grid community.
can trust AI to run our power grid, and price forecasting a must-solve problem. In the fourth article, “Demonstration of
what future grids may look like as AI Electricity price forecasting models Intelligent HVAC Load Management
continues to evolve. have been evolving during the past 25 With Deep Reinforcement Learning,”
The maturity of AI and ML dif- years from relatively simple models, Y. Du, F. Li, K. Kurte, J. Munk, and H.
fers across various power engineer- such as linear regression models, to Zandi discuss building load manage-
ing applications. This issue of IEEE sophisticated ML models that rely on ment for heating, ventilation, and air
Power & Energy Magazine offers a more data and computing power. In conditioning (HVAC) systems using deep
the second article, “Electricity Price reinforcement learning (RL). The au-
Digital Object Identifier 10.1109/MPE.2022.3156688
Forecasting: The Dawn of Machine thors describe and evaluate two deep RL
Date of current version: 19 April 2022 Learning,” A. Je drzejewski, J. Lago, approaches for HVAC ­control. They
‘
12 ieee power & energy magazine may/june 2022
also discuss the deployment of a deep Data-driven approaches for devel- The issue concludes with an “In My
RL-based HVAC control approach in a oping dynamic system models have View” column by H. Zareipour. He re-
real-world residential house in Knox- become highly relevant today for power views a history of disruption in the power
ville, Tennessee, and then analyze the grids with high-penetration inverter- industry since the war of currents. Con-
deployment results, followed by con- based resources (IBRs) as IBR infor- sidering the push to decarbonize the
cluding remarks. mation is proprietary. In the sixth arti- grid, he predicts that three technologies
Millions of smart meters installed cle, “Data-Driven Dynamic Modeling will shape the next disruption: AI, 5G
worldwide are generating an enormous in Power Systems,” L. Fan, Z. Miao, networks, and distributed digital ledger.
amount of load profiles every day. Ac- S. Shah, P. Koralewicz, V. Gevorgian, The ideation of this special issue
cess to these data creates opportunities and J. Fu first review five applications dates back to 2018, when former Editor-
for power companies to better under- of data-driven power system dynamic in-Chief Mike Henderson was in full
stand their consumers’ consumption modeling. The review includes param- support. The issue would have not come
patterns. In the fifth article, “Cluster- eter identification for generators and to reality without the continuous support
ing Electricity Consumers: Challenges loads via data and oscillation mode from the current Editor-in-Chief Steve
and Applications for Operating Smart analysis via phasor-measurement data, Widergren and his editorial team. We
Grids,” A. Alonso, E. Martín, A. Ma- among other applications. Further- would like to thank the authors for con-
teo, F. Nogales, C. Ruiz, and A. Veiga more, the authors present frequency- tributing high-quality articles and mak-
connect clustering methods to power domain admittance IBR model iden- ing the deadlines during the pandemic
system operations. They specifically tification via data generated from period. We would also like to thank As-
address the aspects of efficiency, ro- controlled experiments and challenges sociate Editor Ning Lu for her insights
bustness, and scalability, considering in designing dynamic model struc- and constructive comments through the
the outliers, missing values, and the tures and deriving model parameters review process. We hope the readers en-
volume of the smart meter data. for IBRs. joy reading this special issue. p&e

may/june 2022 ieee power & energy magazine 13

 Artificial
Intelligence for
Load Forecasting
©SHUTTERSTOCK.COM/WAN WEI

A
History, Illusions, ARTIFICIAL INTELLIGENCE (AI) HAS FOUND MANY
applications in today’s world, such as computer vision for

and Opportunities self-driving cars, speech recognition for personal assistants,

and algorithm design for strategy gaming systems. Although
enjoying the convenience that AI has brought to our daily
lives, people may be wondering when and how it started
and evolved.
AI has gone through several waves since conceptualiza-
By Tao Hong tion in the 1940s. One of the first commercial AI applications
and Pu Wang was in the power industry, where artificial neural networks
(ANNs) were practically used for short-term load forecast-
ing in the 1990s. During the past three decades, the research
community has published thousands of load forecasting
Digital Object Identifier 10.1109/MPE.2022.3150808 papers that promote AI-based models, many of which have
Date of current version: 19 April 2022 originated from or resulted in illusions (misunderstandings,

14 ieee power & energy magazine 1540-7977/22©2022IEEE may/june 2022

 the construction of the Pearl Street Station in 1882, the first
power plant in the United States, serving an initial load of
400 lamps for 82 customers in lower Manhattan Island. The
load profiles during those early years were as simple as step
functions, while load forecasting was as easy as counting
the light bulbs turned on in the evening. The same method is
still being used today by many local distribution companies
to estimate the load of streetlights.
The first large-scale, alternating current (AC) electric
generating plant in the world, Niagara Falls Power Plant,
was built by Nikola Tesla and George Westinghouse in
1895. The AC systems made long-distance power transmis-
sion an economically viable solution, which helped power
companies realize economies of scale. With larger power
generators, electricity could be produced at a lower unit
cost. Meanwhile, power companies started to encourage
customers to use electricity, further stimulating the inven-
tion of electric appliances. As electricity end uses became
more diverse, people started to find more curvatures in the
load profiles driven by human and business activities, such
as turning lights on and off, using electric irons, listening
to radio broadcasts, and shopping during holiday seasons.
Most of these activities can be captured by calendar vari-
ables, just as in today’s load forecasting models.
The first modern air conditioner was invented in 1902 by
Willis Carrier. Air conditioners not only brought comfort to
our daily lives but presented challenges to calendar-based
load forecasting methods. As the market penetration of air
conditioners increased, the effect of weather started to play a
major role in load profiles. In 1944, Henry A. Dryar, a chief
load dispatcher at Philadelphia Electric Company, stated
“there is a variable component of the load that reflects the
effect of the weather.” He described three weather variables:
temperature, wind velocity (wind speed), and cloudiness
(cloud cover). Load forecasting then started to attract atten-
tion from the power industry and the research community.
Figure 1(a) summarizes the important years for load fore-
casting before the 1950s.
The Electronic Numerical Integrator and Computer
(ENIAC), the first programmable, electronic, general-pur-
confusion, exaggerated beliefs, and misleading ideas) of pose digital computer was made in 1945. Before the wide
what AI can do. adoption of personal computers (PCs), many methods were
Today, most of the AI-based load forecasting models pro- already being applied to load forecasting. Dryar’s work rep-
posed in academic literature are still at the theoretical level, resents some of them, assigning weights to different vari-
with few adopted in practice. In this article, we examine five ables to forecast the system load. The concept is like multiple
illusions associated with developing AI-based models. We linear regression, which uses several explanatory variables
also present solutions and clarifications to help improve the to predict the outcome of a response variable. Apparently, at
efficiency of AI for load forecasting, given opportunities in that time, the parameters were not estimated by computers
the big data era. running statistical software packages.
Some load forecasters kept a book of temperature profiles
Prepersonal Computer-Era matched with load profiles for different day types in differ-
Load Forecasting ent seasons. In the morning, they looked at the weather fore-
After inventing the incandescent light bulb in 1879, Thomas cast, picked the closest temperature profile from the book,
Edison began a series of efforts to commercialize his light- and found the matching load profile for that day. This is also
bulb and demonstrate his electric lighting systems. One was known as the similar day method, which is still used in many

may/june 2022 ieee power & energy magazine 15

 In 1956, a group of scientists participated in
the Dartmouth Summer Research Project on Artificial Intelligence,
widely recognized as the birth of AI as a research field.

control rooms today, either manually just like in the pre-PC computability, that is, the ability to solve a problem effec-
era, or in an automated fashion via computer programs with tively. In the 1940s, scientists started to investigate whether
sophisticated algorithms. machines could think like humans. A notable contributor was
Some load forecasters also figured out other creative Alan Turing, who published a series of papers on the topic of
ways. For example, a forecaster might take a bicycle ride intelligent machinery from the late 1940s to the early 1950s.
across the system territory every morning. Upon returning On the other hand, the foundation of biological neural net-
to the office, he weighed his shirt, which carried much sweat works was built in the late 19th century by philosophers and
during a hot and humid day or little sweat during a chilly psychologists such as Alexander Bain and William James. Bain
and windy day. A heavier shirt indicated a higher peak load and James independently proposed that interactions among
for the day and vice versa. This physical method relied on neurons within the brain result in thoughts and body activities.
a human body to incorporate the effect of weather on the The studies of mathematical logic and psychology eventually
system load. Today, we can capture weather effects using led to the inception of AI. In 1943, neurophysiologist Warren
comprehensive models with various weather variables. Sturgis McCulloch and logician Walter Pitts implemented the
first artificial neurons that modeled the basic logic units of a
AI’s Inception, Two Waves, brain. A collection of these artificial neurons is called an Arti-
and Two Winters ficial Neural Network (ANN), which is inspired by the bio-
Long before the Middle Ages, Chinese, Indian, and Greek logical neural networks that constitute animal brains. Figure 2
philosophers and mathematicians had already begun to shows a simple three-layer, feedforward ANN, which has two
investigate formal reasoning. In the 1930s, several mathema- neurons in the input layer, three in the hidden layer, and one in
ticians and logicians attempted to formalize the notion of the output layer. In 1956, a group of scientists participated in

1879: Thomas Edison invented incandescent lightbulb (a) Load Forecasting Events
1882: Construction of Pearl Street Station
1895: Niagara Falls Power Plant was built
1902: Willis Carrier invented air conditioner
1944: Henry Dryar’s paper “The Effect of Weather on
the System Load” was published
1945: ENIAC was made

1974:
1930s: Backpropagation from Paul Werbos’
Mathematicians and dissertation
logicians attempted to 1969: Perceptrons: An Introduction to Computational
formalize the notion Geometry, by Marvin Minsky and Seymour Papert
of computability was published
1940s:
Scientists started to 1956: Dartmouth Summer Research Project
investigate whether a 1948: Intelligent machinery, by Alan Turing
machines can think
like humans 1943: First artificial neuron
(b) AI Events

figure 1. The major events related to load forecasting and AI. ENIAC: The Electronic Numerical Integrator and Computer.

16 ieee power & energy magazine may/june 2022

the Dartmouth Summer Research Project on Artificial Intelli- XYZ might refer to any practical problem such as power qual-
gence, widely recognized as the birth of AI as a research field. ity, security assessment, fault diagnosis, or load forecasting,
Many theoretical breakthroughs and research advances the answer was often as simple as “Artificial intelligence.”
were made in the 1950s and 1960s. However, the field of One of the first papers discussing ANNs for load fore-
AI quickly experienced its first winter after Marvin Minsky casting was presented at the American Power Conference
and Seymour Papert discovered two drawbacks of ANNs: 1) in 1989. Since then, thousands of papers have adopted AI
single-layer NNs were incapable of processing the exclusive- and machine learning (ML) techniques for load forecast-
or circuit; and 2) computers were not sophisticated enough to ing. The topics included fuzzy expert systems, support vec-
effectively handle the long run time required by large ANNs. tor machines, and deep learning. Fuzzy expert systems are
The findings were published in Perceptrons: An Introduc- based on fuzzy logic, which models logical reasoning with
tion to Computational Geometry. vague or imprecise statements rather than straightforward
The field of AI faced even more challenges in the early true-or-false statements. Support vector machines are mod-
1970s. For example, people realized that many important els that can find a hyperplane in a high-dimensional space
applications of AI, such as vision and natural language pro- to classify observations into one category or the other. Deep
cesses, required too much information about the world to be learning is based on ANN, where the word deep refers to the
handled by a database at that time. Although some of these use of “many” hidden layers in the network.
challenges were overcome in later years, they slowed down Figure 3 depicts the number of journal articles through
the research progress seriously enough for funding agencies 2021 with “load forecasting” in the title and selected AI/
to end financial support.
The second AI wave started in the mid-1970s, when Paul
Werbos proposed a backpropagation algorithm to practi-
cally train multilayer ANNs. In the 1980s, advances in digi-
Input Layer Hidden Layer Output Layer
tal electronics and distributed computing enabled the use of
even larger networks. AI saw its first commercial bubble,
with hundreds of AI companies being incorporated and mil-
lions of dollars flooding into AI research. The bubble, how-
ever, burst rather quickly in the late 1980s and early 1990s
because commercial vendors failed to develop a wide variety
of practical solutions. That was widely considered the sec-
ond AI winter. Figure 1(b) summarizes the important years
for AI before the 21st century.

Applications of AI to Load Forecasting

Despite the second winter AI was facing, power engineers
found several AI applications in power systems. In the 1990s,
AI became a hot topic among the power engineering com-
munity. At power engineering society general meetings, when
people asked, “How do we solve ‘XYZ’ problem?” where figure 2. A simple three-layer, feedforward ANN.

160 250
140 Neural Network Fuzzy Support Vector Deep ALL
200
120
100 150
80
60 100

40
50
20
0 0
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021

figure 3. The journal papers with “load forecasting” in the title and “AI/ML techniques” in the abstract.

may/june 2022 ieee power & energy magazine 17

 ML techniques in the abstract. In chronological order, load ANN predicted the change in hourly load from yesterday to
forecasting literature, and journal articles in particular, today. The final load forecast for each hour was the linear
found ANNs in 1991, fuzzy expert systems in 1992, sup- combination of the outputs from these two ANNs. The com-
port vector machines in 2002, and deep learning in 2015. mercialized version of the system later became one of the
In total, 1,984 journal articles with “load forecasting” in major load forecasting solutions in the power industry, used
the title were published from 1991 to 2021, of which 1,245 by many power companies and energy trading firms.
mentioned AI/ML techniques in the abstract. ANNs are Load forecasting was widely recognized as a success-
mentioned in 796 of them. Indeed, ANNs have been the ful application of ANNs in the 1990s, but in the following
most popular technique in load forecasting literature dur- decade, the load forecasting community made little progress
ing the past three decades. in terms of methodological innovation and improvements
Most of these ANN models only stayed in academic on model accuracy and practicality. Meanwhile, other com-
papers or the laboratory environment. Few received much munities, such as computer vision and mathematical pro-
attention or were accepted by the industry. Nevertheless, one gramming, were making significant progress both method-
branch of research was able to make load forecasting one of ologically and practically. Looking back at load forecasting
the most successful ANN applications. In 1995, IEEE Trans- literature in the 1990s–2000s and comparing it with those
actions on Power Systems published an article describing a flourishing areas, we can find several reasons for the slow
load forecasting system based on ANNs, which was imple- progress:
mented at 20 United States utilities and used by several of ✔✔ The load forecasting literature had no benchmarking
them. The system was later named an ANN short-term load data or models until the 2010s.
forecaster (ANNSTLF). In 1997, IEEE Transactions on Neu- ✔✔ Many load forecasting papers were not reproducible.
ral Networks published an article describing ANNSTLF’s ✔✔ Many researchers failed to make direct comparisons
second generation, with 32 utilities across the United States with other state-of-the-art load forecasting models.
and Canada using the system. A year later, IEEE Transac- As a result, load forecasting literature has been filled with
tions on Power Systems published an article describing the papers of varying degrees of quality. Many have created
third generation of ANNSTLFs used by 35 utilities in the illusions among the load forecasting community, further dis-
United States and Canada. tracting it from effectively making the next breakthroughs.
Although all three generations of ANNSTLFs are based Some of the aforementioned issues have been or are in the
on ANNs, their structures are quite different. The first gen- process of being resolved, while others may still take some
eration included 38 ANNs and 24 combiners. Each ANN had time to overcome. The remainder of this article will focus on
a three-layer, feedforward structure, as shown in Figure 2. clearing up these illusions.
Some of them had nine input neurons, while others had 72.
The 38 ANNs were designed to capture load and weather Modeling Process
relationships. The outputs from these ANNs were fed to 24 Figure 4 presents a typical load forecasting process. At
combiners to generate a forecast for the 24 h of a day. The first, a load forecaster would gather load and weather his-
second generation eliminated the combiners and reduced the tory data to build a load forecasting model. By putting the
number of ANNs to 24, one for each hour of the day. The model and weather forecast together, the load forecaster
ANNs were separated into four groups for different periods can generate load forecasts. Some load forecasters may
of the day. The third generation consisted of two ANN load exclude weather data when the load is not weather-sensitive,
forecasters and one combiner. One ANN was trained to pre- the weather data are not reliable, or the modeling techniques
dict the regular (base) load of the next day, while the other do not require weather data. Different organizations may

Load Model
History

Forecasting Load
Modeling
Process Forecast
Process

Weather Weather
History Forecast

figure 4. A typical load forecasting process.

18 ieee power & energy magazine may/june 2022

execute different variants of the process. For example, nonlinear curvatures. The “linear” in linear models means
some may add a preprocessing step to cleanse the load and linearity in parameters. In other words, as long as the
weather history before sending them to the modeling pro- parameters being estimated are in a set of linear equations,
cess, while others may add a postprocessing step at the end the model is linear.
to fine-tune the load forecast. Looking at load forecast- The shape in Figure 5 can be modeled by a polyno-
ing literature, most papers have focused on the modeling mial regression model, which is in the family of linear
process by trying various techniques for load forecasting. regression models. Both second- (blue line in Figure 5)
For most load forecasting techniques, the modeling pro- and third-order polynomials (brown line in Figure 5) have
cesses are somewhat similar. A forecaster slices the histori- been used for load forecasting models. A piecewise linear
cal data into three periods. The first is for parameter estima- regression model can also model the shape.
tion (or training). The goodness of fit for this period is called
in-sample fit. An accurate model would result in a good in- Clarification 1
sample fit, but a good fit may not indicate an accurate model. Linear models can be used to model nonlinear relation-
Sometimes a model overfits the data by capturing random ships. The business needs designate which family of models
noises so that the prediction may be far away from the actual should be considered. The exact model to be used should
observation or normal energy consumption level. This is be determined based on the outcome of the model selec-
known as the overfitting issue. tion process.
To avoid the overfitting issue, the second period is set for
model selection (or validation). The goodness of fit for this Illusion 2: Techniques Are Siloed;
period is called a postsample fit. A forecaster may train sev- Either AI or Statistics, But Not Both
eral models, using each to forecast the second period. The AI and statistics (or econometrics) are often considered
model with the best postsample fit result can be selected as rivals. Many load forecasters in either camp rarely look at
the winner for forecasting. In practice, when a large number the work done by the other group. In academic literature,
of models are trained and fed to the validation period, the many AI papers compare proposed load forecasting mod-
winning model may be the result of overfitting the valida- els with other AI-based models. Even if comparisons were
tion period. made using statistical models, the models are often far less
To add another layer of insurance, the third period powerful than state of the art in that model family.
is reserved for an out-of-sample test. A forecaster tests ANNs and regression analysis, two seemingly differ-
the winning model from the postsample fit on the third ent techniques in AI and statistics, respectively, are con-
period, which is completely blind from parameter estima- nected in many ways. Both techniques ask for similar input
tion and model selection. When the selected model does variables such as weather and calendar information. Both
not give surprisingly bad results, it can be promoted require the estimation of parameters, which can be formu-
for forecasting. lated as optimization problems. The parameters of ANNs
Some variants of the aforementioned process have been can be estimated through backpropagation, while the
used in many business sectors, including load forecasting. parameters (or coefficients) for a regression model can be
For example, the process may be simplified by removing the estimated by minimizing the sum of squared errors. Such
out-of-sample test. In many case studies, forecasters execute regression analyses are called an ordinary least squares
the training and validation steps multiple times on different (OLS) regression.
periods of the historical data, known as cross validation. As
load forecasting is a time series forecasting problem, sliding
simulation is often used in place of training and validation,
6,000
mentioned previously.
5,000
Illusions 1: A Nonlinear Shape, So We
Demand (MW)

Have to Use Nonlinear Models 4,000

The relationship between load and temperature is nonlin-
ear, as depicted in Figure 5. The shape of that scatter plot 3,000
is commonly known as a hockey stick or Nike swoosh. On
the left side, the load increases as the temperature decreases 2,000
due to heating needs. On the right, the load increases as the
1,000
temperature increases due to cooling needs. –10 10 30 50 70 90
Many papers used the nonlinear relationship as motiva- Temperature (F)
tion for nonlinear models. However, a nonlinear relation-
ship does not require nonlinear models. In many situations, figure 5. Using regression models to capture the load
linear models are better than nonlinear models at modeling temperature relationship.

may/june 2022 ieee power & energy magazine 19

 Mathematically, the word norm can be used to describe variables could improve the load forecast accuracy of OLS
the distance of a vector from the origin. Minimizing the sum regression models, a similar benefit would be expected for
of squared errors is equivalent to minimizing the Euclid- L1 regression, SVR, ANN, and so forth. The natural next
ean distance of the error vector from the origin, which is step would simply be to test humidity variables for the other
called the Euclidean norm, 2-norm, or L2 norm. Similarly, techniques rather than having another group of people redis-
minimizing the sum of absolute errors is equivalent to mini- cover the same from scratch.
mizing the Manhattan distance of the error vector from the
origin, which is called the Manhattan norm, 1-norm, or L1 Clarification 2
norm. The regression analysis that minimizes the L1 norm Many different techniques are connected. Many significant
of the error vector is called L1 regression, or least absolute findings, useful insights, and methodological breakthroughs
deviation regression. are not technique specific; instead, they are portable among
Starting from least squares regression, we can also obtain different approaches.
several extensions that are well within the field of AI. For
example, in addition to minimizing the sum of squared Illusion 3: No One Else Has Used a
errors, adding a penalty (regularization) to reduce the magni- Combination of These Metaheuristics,
tude of the parameters being estimated would give us what is So That Is the Novelty
termed ridge regression. Using an extra penalty to reduce the In mathematical optimization, metaheuristic algorithms
number of variables would give us the least absolute shrink- aim at providing a sufficiently good solution to an optimi-
age and selection operator. Moving absolute errors from the zation problem. They are especially useful when the global
objective function to the constraints section while minimiz- optimal solution cannot be easily reached due to limited
ing the magnitude of the parameters would give us support computational resources. Many well-known optimization algo-
vector regression (SVR). Imposing a fuzzy membership func- rithms, such as simplex method and interior point method,
tion to the regression line would give us fuzzy regression. have specific and strong requirements for the formulation of
As listed in Table 1, by simply moving or tuning the the problems, so they are constrained for specific types of
norms of both the error vector and parameter vector, we optimization problems. On the other hand, metaheuristics
have already connected some of the most frequently used make few assumptions about the problem formulation, so
AI and statistics techniques in load forecasting literature. they can be widely adopted to solve many types of optimi-
We can treat these decision variables (e.g., which norm to zation problems.
use and where to place them) as hyperparameters and let Many metaheuristic algorithms are metaphor based. For
the model-building process determine the winner. Note example, genetic algorithms (GAs) are inspired by the pro-
that the hyperparameters here refer to the configurations of cess of natural selection, particle swarm optimization (PSO)
the modeling process, as opposed to the parameters (coef- is motivated by the motion of bird flocks and schooling fish,
ficients) of a specific mathematical function to be esti- simulated annealing is caused by the thermodynamic pro-
mated from the data. cess of metal annealing, ant colony optimization (ACO) is
There is no clear-cut difference between AI and statis- incited by the foraging behavior of natural ants, and gray
tics techniques. Practitioners do not have to separate the wolf optimization (GWO) is influenced by the hunting
two fields either. Connecting them makes the entire flow of behavior of gray wolves. As various formulations of mod-
research and development (R&D) much more efficient and eling processes lead to different forecasting techniques,
productive than what has been presented in the literature. practitioners can conveniently pick some metaheuristics for
For example, when someone discovered that using humidity parameter estimation rather than investigating the specific

table 1. Six variants of regression models.

Primary Objective to Minimize Regularization Constraints
OLS regression Sum of squared errors None None
L1 regression Sum of absolute errors None None
Ridge regression Sum of squared errors Reduce the magnitude of the None
parameters
LASSO regression Sum of squared errors Reduce the number of variables None
SVR Magnitude of the parameters None Upper bound for absolute errors
Fuzzy regression Spread in fuzzy membership None Threshold for fuzzy membership
LASSO: least absolute shrinkage and selection operator.

20 ieee power & energy magazine may/june 2022

formulation and its corresponding well-established problem- ✔✔ First, investigate novel methods, not a hybrid of differ-
solving routine, if there is any. ent metaheuristics, to efficiently and effectively tune
Today, when the metaheuristics community invents a the hyperparameters.
new algorithm, we often see it applied to estimate load fore- ✔✔ Second, find actionable insights, rules of thumb, and
casting models. In load forecasting literature, we can also practical guidelines to help forecasters better navigate
find many papers that present hybrid algorithms, such as and understand hyperparameter tuning.
GA + PSO + ANN, GWO + SVR, and GA + ACO + Fuzzy. ✔✔ Finally, explore new structures of ANNs and new
The claimed contribution is often twofold: 1) the combina- AI techniques that may benefit load forecasting. For
tion of these specific algorithms/techniques has never been instance, the structures of the three generations of
done before for load forecasting and 2) the accuracy from ANNSTLF are quite different, which led to an im-
the proposed hybrid algorithm is superior to its counterparts. provement in forecast accuracy.
However, in the authors’ experience, most, if not all of these
papers tend to be hard to read, impossible to reproduce, and Clarification 4
have no presence in the load forecasting tools used by power Given a network structure, such as three-layer, feedforward
companies. Although these papers may add load forecast- ANNs, the “optimal” number of hidden neurons is unique on
ing as an indirect application to corresponding metaheuristic a case-by-case basis. Attention should be paid to exploring
algorithms, their contributions to load forecasting literature the innovative design of ANNs structures, such as recurrent
appear to be marginal thus far. More about the accuracy and deep neural networks.
issue will be discussed in the “Illusion 5. The Highest Accu-
racy Reported in the Academic Literature Represents State Illusion 5: The Highest Accuracy
of the Art” section. Reported in Academic Literature
Represents State of the Art
Clarification 3 Many factors affect load forecast accuracy, such as data
For most of the models in load forecasting literature, param- quality, load composition, and size of the load. For example,
eter estimation can be formulated as an optimization prob- given all other factors set the same, the load forecast accu-
lem. Metaheuristics is one way to find sufficiently good racy of an independent system operator (ISO) is expected to
solutions to these optimization problems. However, combi- be higher than that of a power company within the ISO. The
nations or hybrid uses of different metaheuristics with other accuracy reported in one paper could be for a specific area
AI techniques appear to offer a marginal contribution to the during a certain period, which may not be generalized to
load forecasting practice thus far. other jurisdictions.
Even putting aside various uncontrollable factors, there
Illusion 4: A Three-Layer, Feedforward may be a tendency for some authors to exaggerate forecast
ANN With 35 Hidden Neurons Is Better accuracy as a result of flawed processes of model building
Than the One With 41 and forecast evaluation. For example, a paper may report
Researchers have tried many types of ANNs, such as feed- day-ahead forecast accuracy values in 2020 from a particu-
forward, recurrent, and deep ANNs. Among the various lar model M based on a proposed methodology, which beats
network structures with different setups, three-layer, feed- three other benchmark models based on existing techniques
forward ANNs, due to their relative simplicity, were most in the literature. All four models are built based on data from
frequently used before deep learning was brought to load 2017 to 2019. Everything may look legitimate, fair, and rig-
forecasting. Some papers argue that the superior perfor- orous on paper. However, behind the scenes, the authors may
mance of an ANN structure is due to some minor changes, have tried 10 other variants by tuning hyperparameters of
such as increasing the number of hidden neurons from 35 the proposed methodology, among which model M had the
to 41, increasing hidden layers from one to two, and chang- highest accuracy. The other variants may not be better than
ing activation functions to radial basis functions. All of the three benchmark models. In other words, the year 2020
these changes can be viewed as hyperparameter tuning. The is used for model selection without being disclosed to the
resulting setup of hyperparameters may work well for one readers. This is known as peeking into the future in forecast-
specific data set but fail for the others. The contribution to ing. In reality, a forecaster cannot pick a model based on
load forecasting from such results tailored for a specific case perfect information of future load values.
study is insignificant. Another approach used to promote a proposed model is
Nevertheless, this is by no means discouraging practi- cherry-picking. Authors may secretly manipulate data with-
tioners from trying various ANNs or other AI techniques. out disclosing details in the paper. Some observations may
Instead of devoting resources and efforts to fine-tuning be excluded from the model-building process, while others
hyperparameters, researchers have been encouraged to look may be replaced by different values. All of these changes
at the bigger picture. Here we list three research directions, to the raw data have implications for final forecasts. In
which are not meant to be exhaustive: other words, data preprocessing methods may have put the

may/june 2022 ieee power & energy magazine 21

 proposed approach at an unfair advantage. Meanwhile, read- developed load forecasting models. Many models had to be
ers who do not know these details can hardly make the pro- built offline and retrained every week or month. The meth-
posed technique work well in real-world data. ods at that time had to find a tradeoff between using the most
recent data and how fast the models could be estimated. Just
Clarification 5 like computer technologies pushed AI to its second hype in
The accuracy values reported by many academic papers the 1980s, today’s computing infrastructure enables applica-
are too good to be true. Some reported values have been tions of deep learning and other computationally intensive
significantly exaggerated due to the flawed process of techniques to load forecasting.
model building and forecast evaluation, such as peeking In Dryar’s 1944 paper, the weather was first identified as
and cherry-picking. Figure 6 lists a few checkpoints for a driving factor of electricity demand. In the 20th century,
readers to conduct a smoke test when they see astonish- the data used by load forecasters were quite limited. Weather
ingly good results. data came from a few weather stations in the service ter-
ritory. Other than some interval meters collecting hourly
Opportunities demand for selected households for load research, there were
Many opportunities are presented for today’s researchers no advanced communicating (smart) meters. Utilities had
and practitioners interested in load forecasting. Modern only monthly or bimonthly meter readings for most custom-
computers are more powerful than ever before. In the 1980s, ers. For long-term load forecasting, monthly or annual load
computing resources were a major concern when people data were used.
Today’s load forecasters can access a variety of data
sources and high granularity. Smart meters are collecting
load readings every hour, if not more frequently. Weather
r hybrid) o
f conditions are monitored by thousands of in situ weather sta-
p ro p o si n g a mix (o tions and satellites. In many places, power companies have
r is
• The pape
niques. ords. installed cameras to capture sky images. Meanwhile, demo-
a ny te ch new buzzw
m
is m e re ly catching
• The pap e r graphics and economic information are monitored and col-
ry.
is proprieta ith industry lected in great detail. Social media platforms also generate
• The data t co -a uthored w
a p e r is n o industry).
• Th e p
nsored by
th e useful data for load forecasters to better understand human
r n o t sp o e new
people (o s not utiliz behaviors and local environments.
p o se d m ethod doe
• The pro Forecasting competitions can significantly stimulate
wledge
variables t take kno
d m e th o d does no the advance of a forecasting field. A well-organized com-
ose
• The prop petition can recognize state-of-the-art methods, attract tal-
mains. ge
of other do s not levera
p o se d m ethod doe ents from other fields to solve a problem of interest, and
• The pro urces.
uting reso ating its release benchmark data for future researchers to continue
new comp d is domin
sed metho d) the investigation.
•T h e p ro p o ib metho
le
a rt s (a n other cred Table 2 lists five major load competitions during the
counterp
cts.
in all aspe last three decades. Before the 2010s, load forecasting
literature saw only two remarkable competitions. One was
supported by Puget Sound Power and Light Company in
figure 6. The smoke tests for too-good-to-be-true results. 1991–1992, and the other was organized by the EUNITE

table 2. Five notable load forecasting competitions from the 1990s to 2010s.
Time Competition Name Problem Affiliations of Winning Teams
1991–1992 Puget Sound Shootout Short-term load forecasting University of California, San Diego
Southern Methodist University
2001 EUNITE Competition Medium-term load forecasting National Taiwan University
2012 GEFCom2012 Hierarchical load forecasting CountingLab
Cambridge University
EDF R&D
2014 GEFCom2014 Probabilistic load forecasting EDF R&D
University of North Carolina at Charlotte
2017 GEFCom2017 Hierarchical probabilistic load forecasting Japan Meteorological Corporation
EDF: Électricité de France.

22 ieee power & energy magazine may/june 2022

 The accuracy reported in one paper could be for a
specific area during a certain period, which may not
be generalized to other jurisdictions.

network in 2001. In the 2010s, a series of three energy fore- To accelerate the improvement of load forecasting
casting competitions, known as Global Energy Forecast- practices, talented people with experience beyond power
ing Competitions (GEFComs), was organized, attracting companies deserve consideration. As listed in Table 2, the
hundreds of participants worldwide. These competitions winning teams of those five competitions came from eight
also stimulated interest from the industry, resulting in different organizations. Only one of the eight organizations
many other smaller competitions organized by power is a power company. The other seven include a technology
companies. In the past, good models might be buried start-up, a weather company, and five universities. Load
by many other mediocre papers. Currently, good models forecasting is an interdisciplinary subject. Building indus-
have many opportunities to show up in the leaderboards try–academia partnerships as well as collaborations among
of various competitions. the energy and other communities, such as meteorology
Before the 2010s, load forecasting, or energy forecast- and data science, may provide new insights and lead to
ing in general, was a minority subject in power engineer- further breakthroughs. By working together, better models
ing. Energy forecasters did not have a professional group as may be realized.
their home. In 2011, the IEEE Working Group on Energy
Forecasting was formed by the IEEE Power & Energy Soci- For Further Reading
ety. The group actively promoted energy forecasting by G. Gross and F. D. Galiana, “Short-term load forecasting,”
organizing the three GEFComs, delivering panel sessions Proc. IEEE, vol. 75, no. 12, pp. 1558–1573, Dec. 1987, doi:
at major power and energy conferences, teaching tutorials, 10.1109/PROC.1987.13927.
and so forth. In 2019, the International Institute of Fore- R. Ramanathan, R. Engle, C. W. Granger, F. Vahid-
casters started the Section on Water, Energy, and Environ- Araghi, and C. Brace, “Short-run forecasts of electricity
ment (SWEET) to bring together forecasters interested in loads and peaks,” Int. J. Forecasting, vol. 13, no. 2, pp. 161–
issues around water, energy, and environment. Today, load 174, 1997, doi: 10.1016/S0169-2070(97)00015-0.
forecasters can easily find a welcoming community to net- A. Khotanzad, R. Afkhami-Rohani, and D. Maratukulam,
work with other forecasters. “ANNSTLF-artificial neural network short-term load fore-
caster-generation three,” IEEE Trans. Power Syst., vol. 13,
Together no. 4, pp. 1413–1422, Nov. 1998, doi: 10.1109/59.736285.
Traditionally, the load forecasting job is done by power B.-J. Chen, M.-W. Chang, and C.-J. lin, “Load forecast-
system operators and planners in power companies. How- ing using support vector Machines: A study on EUNITE
ever, the aforementioned load forecasting competitions indi- competition 2001,” IEEE Trans. Power Syst., vol. 19, no.
cate winning contributions from internationally recognized 4, pp. 1821–1830, Nov. 2004, doi: 10.1109/TPWRS.2004.
experts who are not power system operators or planners. 835679.
Their mastery of techniques from fields such as statistics, T. Hong, P. Pinson, S. Fan, H. Zareipour, A. Troccoli, and
econometrics, and computer science allowed them to develop R. J. Hyndman, “Probabilistic energy forecasting: Global
models that proved superior to the competition. energy forecasting competition 2014 and beyond,” Int. J.
Newcomers to the field of load forecasting often ask, Forecasting, vol. 32, no. 3, pp. 896–913, Jul./Sep. 2016, doi:
“What is the most accurate technique for load forecast- 10.1016/j.ijforecast.2016.02.001.
ing?” The short and sweet answer should be “None.” No T. Hong and S. Fan, “Probabilistic electric load forecast-
technique is superior to all other methods in load forecast- ing: A tutorial review,” Int. J. Forecasting, vol. 32, no. 3, pp.
ing. What matters in load forecast accuracy depends more 914–938, Jul./Sep. 2016, doi: 10.1016/j.ijforecast.2015.11.011.
on the mastery of applying the technique than the tech-
nique itself. For example, an ANN expert and OLS regres- Biographies
sion specialist in a competition may eventually obtain sim- Tao Hong is with the University of North Carolina at Char-
ilarly accurate forecasts, while their forecasts are likely to lotte, Charlotte, North Carolina, 28223, USA.
be better than the forecasts developed by people without Pu Wang is with Duke Energy Corporation, Charlotte,
in-depth experience applying any specific technique or North Carolina, 28202, USA.
p&e
tuning the models. 

may/june 2022 ieee power & energy magazine 23

 Electricity Price
Forecasting
The Dawn of Machine Learning

©SHUTTERSTOCK.COM/WALDEMARUS

By Arkadiusz Jędrzejewski, Jesus Lago,

Grzegorz Marcjasz, and Rafał Weron

Digital Object Identifier 10.1109/MPE.2022.3150809

Date of current version: 19 April 2022

24 ieee power & energy magazine 1540-7977/22©2022IEEE may/june 2022

E
ELECTRICITY PRICE FORECASTING (EPF) IS A with D Mon = 1 for Monday and zero otherwise, and so on.
branch of forecasting at the interface of electrical engineer- Arrows indicate the flow of information—there is a coef-
ing, statistics, computer science, and finance that focuses on ficient (or weight) assigned to each arrow. The output is just
predicting prices in wholesale electricity markets for a whole a weighted sum of all of the inputs.
spectrum of horizons. These range from a few minutes (real- The most relevant and, often, only used exogenous vari-
time/intraday auctions and continuous trading); through days able, even in more recent models, is the DA load forecast.
(day-ahead [DA] auctions); to weeks, months, or even years Interestingly, the actual load is not a better predictor. The
(exchange and over-the-counter traded futures and forward reason is that the bids in the DA market are placed based
contracts). DA markets are the workhorse of power trading, on DA load forecasts instead of the actual loads observed
particularly in Europe, and a commonly used proxy for “the a day later. Since the electricity price–load relationship is
electricity price.” The vast majority—up to 90%—of the nonlinear (see Figure 2) forecasters also played around with
EPF literature has focused on predicting DA prices. These (artificial) NNs.
are typically determined around noon during 24 uniform- An NN consists of layers of nodes. Each node is a neuron.
price auctions, one for each hour of the next day. This has Layers are connected by links between neurons, so a neuron
direct implications for the way EPF models are built. can transmit a signal to another neuron in the subsequent
Over the last 25 years, various methods and compu- layer, the structure of which resembles the synapses in a
tational tools have been applied to intraday and DA EPF. human brain. Thus, the output of a neuron is a weighted sum
Until the early 2010s, the field was dominated by relatively of all of the inputs—as in linear regression—transformed by
small linear regression models and (artificial) neural net- the so-called activation function.
works (NNs), typically with no more than two dozen inputs. The simplest NN with input nodes connected directly to
As time passed, more data and more computational power an output node with a linear activation function is equivalent
became available. The models grew larger to the extent to the linear regression model depicted in Figure 1(a). By
that expert knowledge was no longer enough to manage the using nonlinear activation functions and inserting additional
complex structures. This, in turn, led to the introduction of (so-called hidden) layers of nodes, the relationship between
machine learning (ML) techniques in this rapidly develop- the inputs (i.e., the predictors) and the output (i.e., the pre-
ing and fascinating area. Here, we provide an overview of dicted variable) becomes more complex. The NN models of
the main trends and EPF models as of 2022. Note that the the 1990s and 2000s were typically shallow structures with
article uses EPF as the abbreviation for both “electricity only one hidden layer, one output neuron (e.g., the electric-
price forecasting” and “electricity price forecast.” The plural ity price today at 6 p.m.), and a feedforward architecture
form, i.e., forecasts, is abbreviated as EPFs. [see Figure 1(b)]. The latter means that the information was
propagated in only one direction—from inputs to the out-
25 Years of Evolution put. In contrast to linear regression, where the output is just
The beginnings of EPF can be traced back to the 1990s. The a weighted sum of all of the inputs, in NN models, addi-
first attempt to predict electricity price dynamics used linear tional nonlinear transformations may be applied to this sum
regression techniques. Recall that such a model assumes a at each node.
linear relationship between the predicted variable (e.g., the
electricity price today at 6 p.m.) and inputs (e.g., past elec- Shrinking Redundant Features
tricity prices and the load forecast for today). The inputs are Uncovering the nonlinear relationship between the vari-
also called features, explanatory variables, regressors, or ables is a challenge for the linear regression-based approach.
predictors. More formally, the predicted variable is repre- Another difficulty is the handling of a large number of
sented as a weighted sum of the inputs. inputs. Even a few dozens of explanatory variables can lead
Early regression models were built on expert knowl- to unreliable estimates and predictions when using the clas-
edge. The inputs included past prices (typically, those from sical regression model and ordinary least squares for mini-
one, two, and seven days ago), exogenous variables, and a mizing the sum of squared errors between the forecasted
seasonal component. Seasonality was captured either by a and observed values. A working remedy provides so-called
sinusoidal function or—more commonly—a set of so-called regularization algorithms. These shrink (hence, the alterna-
dummy variables taking the value of one on a given day of tive name—shrinkage algorithms) coefficients, or weights
the week and zero otherwise. of the less important inputs, toward zero by adding a penalty
A sample regression EPF model is illustrated in Figure 1(a). for large coefficients to the sum of squared errors.
White squares represent individual explanatory variables One of the most widely used regularization algorithms
(also called regressors, inputs, or features), whereas the is the least absolute shrinkage and selection operator
purple circle represents an output variable: p d, h and X d, h, (LASSO), popularized by Robert Tibshirani in the mid-
respectively, stand for the electricity price and an exogenous 1990s. This ML (also called statistical learning) technique
variable (e.g., the DA load forecast) for day d and hour h, shrinks some coefficients to zero and, thus, performs semi-
while D Mon, f, D Sun are the so-called dummy variables, automatic, data-driven variable selection. EPF models that

may/june 2022 ieee power & energy magazine 25

 initially have hundreds of inputs can be reduced to structures proposed in 2021 as one of two challenging benchmarks for
with only a dozen or two relevant regressors. The selection contemporary EPF. The LEAR model is a parameter-rich
is “semiautomatic” because LASSO requires the setting of autoregressive st r uct u re withabout 250 explanatory
the tuning hyperparameter that regulates which features are variables. That is, to predict the price for one hour, the model
eliminated. This hyperparameter can be set using cross vali- utilizes hourly prices from one, two, three, and seven days
dation, a resampling method that uses different portions of ago; hourly DA forecasts of two exogenous variables (e.g.,
the data to test and train a model. the system load and wind power generation) for today, yes-
LASSO-estimated EPF models appeared in the mid- terday, and a week ago; and weekday dummies.
2010s and quickly revolutionized the field. Despite their The model structure is visualized in Figure 4 together
relative simplicity and inability to handle nonlinear depen- with a graphical summary of the importance of selected
dencies, it is not easy to outperform them, even using much inputs for Nord Pool market data. Four years of data (2013–
more complex models. In one of the first EPF applications, 2016) were used for calibration and two (2017–2018) for test-
LASSO was used to sparsify 24 regression models, one for ing; the coefficients (weights) were re-estimated every day.
each hour of the day and each with 100+ inputs, as shown The gray squares in the middle represent 24-dimensional
in Figure 3. vectors of explanatory variables: hourly prices one, two,
A more recent example is the LASSO-estimated autore- three, and seven days ago (p d - 1, p d - 2, p d - 3, and p d - 7) and
gressive (LEAR) model, an open source method that was hourly DA forecasts of two exogenous variables for today,

Model Inputs Model Output Model Inputs Hidden Layer Model Output

pd−1, h pd−1, h

pd−2, h pd−2, h

u1

pd−7, h pd−7, h

Xd, h pd, h Xd, h u2 pd, h

.
.
.
DMon DMon

un
. .
. .
. .

DSun DSun

(a) (b)

figure 1. The early EPF models: (a) linear regression and (b) a single-output shallow NN. The white squares represent the
inputs, the purple circles are the output variables, and the light blue circles show hidden neurons (or nodes). The arrows indi-
cate the flow of information (darker/lighter for better visibility). (Source: The human evolution graphic is a modification of the
“Human Evolution Scheme” by José-Manuel Benitos. Licensed under the terms of the GNU Free Documentation License.)

26 ieee power & energy magazine may/june 2022

yesterday, and a week ago (X id, X id - 1, X id - 7; i = 1 for system the field and models were named deep learning (DL) and
load, and i = 2 for wind power generation). The bottom gray deep NNs (DNNs), respectively, to stress the importance of
square represents weekday dummies (D). depth in the achieved improvements.
The purple circle is the predicted variable—here, it is While this success of DL models was initiated in com-
the price for hour 18, i.e., 6 p.m. A set of 24 such models puter science applications, e.g., image and speech recogni-
yields forecasts for all 24 h of day d. The “chessboards” to tion as well as machine translation, the benefits of DL also
the left illustrate the importance of selected explanatory spread in the late 2010s to energy-related applications, such
variables. The greener the square, the more often selected as EPF or wind power forecasting. Since then, deep mod-
the feature (vertical axis) for a given target hour (horizon- els have been heavily used to better exploit and model the
tal axis; hour 18 is emphasized) in the test period. Dark nonlinear relations between energy-based quantities (prices,
gray squares represent features that were (almost) never load, and generation) and their drivers (e.g., human behavior
selected. There is a correlation between yesterday’s load and calendar effects).
forecasts for hour h today and the price for this hour— In the context of EPF, early works were based on feed-
see the green diagonal in the middle “chessboard.” In forward architectures with features (i.e., inputs) modeled
addition, yesterday’s load forecasts for the morning peak as hyperparameters. The latter are settings of the models
(hours 8–9) and late evening (hours 22–24) also carry rel- (e.g., the number of neurons per layer, type of activation
evant information—see the green horizontal stripes in the
same “chessboard.”

Deeper and Deeper

In the past, the challenge of using NNs was computational
complexity. Training small structures—like the one in Fig-
ure 1(b)—was feasible, but calibrating larger models was
time-/resource consuming and, in some cases, even impos- pd−1,1 Model Inputs Model Output
sible. For instance, during the optimization phase, the gradi-
ent would often become either zero or infinity, leading to the
so-called vanishing gradient problem, effectively preventing pd−1,2
the weights from changing their values. pd−1
Over the last decade, advances in computational resources ..
.
(e.g., the massive usage of GPUs and optimization algo- pd−2
rithms) have made it possible to efficiently train complex pd−1,24
structures, including NNs with hundreds of inputs, multiple ..
.
outputs, many hidden layers, and links to earlier layers (see
Figure 5). Because networks whose depth (i.e., the number of pd−7 pd,h
layers) was not limited to just a single hidden layer systemi-
cally showed better results and generalization capabilities,
Xd
DMon
..
.
80
Today’s (Actual) Prices (€/MWh)

Winter DTue
70 Summer D
60 ..
.
50
40 DSun
30
20
10
figure 3. A LASSO-estimated regression, the next step in
0 the evolution of EPF models. Gray squares denote vec-
30,000 40,000 50,000 60,000 70,000
tors of variables, and dashed arrows reflect the fact that
Yesterday’s Load Forecasts for Today (MW)
LASSO may eliminate some of the links. Like in Figure 1,
white squares represent inputs, and the purple circle is the
figure 2. The relationship between electricity prices output variable. (Source: The human evolution graphic is
and the forecasted load is nonlinear and exhibits distinct a modification of the “Human Evolution Scheme” by José-
seasonal variations, as illustrated by the 2017 data from the Manuel Benitos. Licensed under the terms of the GNU Free
Scandinavian Nord Pool market. Documentation License.)

may/june 2022 ieee power & energy magazine 27

 function, dropout rate, learning rate, regularization coef- in Figure 5(b), except for the feedback connections (dotted
ficient, and choice of optimization algorithm) that need arrows). In 2021, its open source code was released, and
to be optimized and tuned separately from the main NN the DNN model itself was recommended as a challenging
parameters (i.e., the weights). Like hyperparameters, the benchmark—even more than the LEAR model—for con-
features need to be selected/optimized separately from temporary EPF.
the main parameters. Therefore, these early works pro- In terms of training and real-time usage, the DNN con-
posed DNNs where the hyperparameters and inputs were sists of two phases. In the first, the hyperparameters and
optimized together. input features are jointly optimized using historical data. To
The most prominent example of deep NN for EPF is do this, the inputs are modeled as binary hyperparameters,
probably the DNN model proposed in 2018, whose inputs which can be either selected or discarded. This step is per-
and hyperparameters are optimized and tailored for each formed periodically but not very often, e.g., once per month.
case study without the need for expert knowledge. It uses In the second phase, the DNN is recalibrated daily with the
one of the simplest deep architectures proposed for EPF: optimal set of inputs and hyperparameters to account for
feedforward with two hidden layers, 24 outputs (i.e., it the newest market data. Since each hyperparameter/fea-
jointly predicts 24 hourly prices), and the same inputs as theture selection yields a different local optimum, the DNN
LEAR model. The DNN’s structure is identical to the one can greatly benefit from averaging forecasts, that is, train-
ing multiple DNNs and building a
forecast as the arithmetic average
of the individual predictions of
the DNNs.
3 The first wave of DL models
6 was followed, in the last few years,
Yesterday’s Prices

by hybrid architectures utilizing

9
so-called long short-term memory
12 pd−1 and/or convolutional NNs. Unlike
15 feedforward networks, the long
18 pd−2 short-term memory architecture
21
has multiple feedback connections
and can process not only single
24 pd−3 data points but also entire time
3 6 9 12 15 18 21 24
series. Convolutional NNs are reg-
Today’s (Predicted) Prices pd−7 ularized—to prevent overfitting—
versions of multilayer feedfor-
Yesterday’s Load Forecasts for Today

1 ward networks that use convolution

3 Xd
(instead of matrix multiplication)
6 in at least one of their layers.
1
9 X d−1
p d,18 However, despite the recent
12 work on DL models, it is unclear
1
Xd−7 whether all of the extra complex-
15
ity brings improvements in fore-
18 casting accuracy. The existing
2
21 Xd evidence indicates that, although
24 simple DL models (like a DNN)
2
3 6 9 12 15 18 21 24 Xd−1 can, on average, improve upon
Today’s (Predicted) Prices LASSO-estimated regression
2 models (like LEAR), the per-
Xd−7
formance of more complex DL
Monday
Weekdays

models is generally unknown. A

Wednesday
D common pattern that many works
Friday
Sunday in DL share is the failure to com-
3 6 9 12 15 18 21 24 pare the proposed models with
Today’s (Predicted) Prices state-of-the-art statistical meth-
ods and/or employ long enough
figure 4. The variable selection in the LEAR model calibrated to data from the Nord data sets to derive statistically
Pool power market. sound conclusions.

28 ieee power & energy magazine may/june 2022

 Model Inputs Hidden Layer Model Outputs Model Inputs Hidden Layer 1 Hidden Layer 2 Model Outputs

pd −1 pd −1

pd −2 u1 pd,1 pd −2 u1,1 u2,1 pd,1

.. ..
. .

pd −7 u2 pd,2 pd −7 u1,2 u2,2 pd,2

.. .. .. .. ..
. . . . .
Xd Xd
un pd,24 u1,n u2,m pd,24
.. ..
. .

D D

(a) (b)

figure 5. The multioutput NN architectures for EPF: (a) shallow and (b) deep with two hidden layers. The notation is simi-
lar to that in Figures 1 and 3. Dotted arrows represent feedback connections; if present, the structure is called a recurrent
NN. (Source: The human evolution graphic is a modification of the “Human Evolution Scheme” by José-Manuel Benitos.
Licensed under the terms of the GNU Free Documentation License.)

Model Performance German EPEX market, by as much as 55% (rMAE = ~0.45).

Let us now compare the predictive accuracy of the two chal- Moreover, for all five electricity markets, the LEAR model
lenging benchmarks—LEAR and DNNs. Here, we report on is outperformed by the DNN. Clearly, there is a benefit to
a comprehensive study involving five major DA electricity using the more complex method. However, it comes at a
markets with test periods spanning two years each. Three
out of the five markets—Germany, France, and Belgium—
are operated by European Power Exchange (EPEX SPOT),
Relative MAE
the largest power exchange in Europe. In their case, the test PJM
periods range from 4 January 2016 to 31 December 2017 LEAR
for Germany and 4 January 2015 to 31 December 2016 for DNN
0.7
France and Belgium. The remaining two include Nord Pool,
one of the world’s oldest power markets, operating in Scan- 0.6
Nord EPEX
dinavia and the Baltic states, and the Pennsylvania–New Jer- 0.5
Pool Germany
sey–Maryland (PJM) Interconnection, the largest competi-
0.4
tive wholesale electricity market in the Americas. In their
case, the test periods range from 27 December 2016 to 24
December 2018.
Figure 6 plots the performance of a LEAR and DNN
model. The errors are computed with respect to a naive
forecast, which takes last week’s price for the same hour.
Both models provide significantly more accurate fore- EPEX EPEX
casts in terms of the mean absolute error (MAE) than the France Belgium
naive prediction; if plotted, the relative MAE (rMAE) of
the naive method would be one. For instance, for the PJM figure 6. A radar plot of the relative mean absolute errors
market, the MAE values of both models are lower by 40% (rMAEs) for the LEAR and DNN models and two-year test
(rMAE = ~0.6) than that of the naive approach while, for the periods from five power markets.

may/june 2022 ieee power & energy magazine 29

 price—the significantly (up to 100 times) higher compu- layer in Figure 7 is placed between the last hidden and the
tational cost. Is it worth the effort? Does it pay off? Answer- output layer. Nodes of the distribution layer contain esti-
ing these questions requires a much more thorough analysis, mates (or predictions) of the parameters of a distribution
involving diverse error metrics, trading strategies, realistic (here, Gaussian). The output is a predictive distribution, fh,
operational costs, and so on. one for each hour of the day.
The downside is that the distribution itself is a hyperpa-
Outlook for the Future rameter; i.e., it has to be estimated or set ex ante. A more
The methods presented so far yield point forecasts, like the flexible distribution can account for skewness and heavy
expected price today at 6 p.m. This knowledge is important tails, but this also comes at a price. In general, the more
for decision makers, but it does not convey all of the rel- parameters there are to be estimated, the harder it is to cali-
evant information. Probabilistic EPF is a possible remedy. brate the model to a given data set. Nevertheless, probabilis-
The output of such a model will no longer be a single value, tic NNs may be objects of interest in the future.
e.g., the expected price today at 6 p.m., but a distribution of
possible values. For instance, looking at a probabilistic (also The Good, the Bad, and the Ugly
called distributional) forecast, we will be able to conclude Like the main characters in the epic spaghetti Western film
that the price today at 6 p.m. will be between €100 and €110 directed by Sergio Leone, the practices we can observe in
with 90% probability. the EPF literature come in all shapes and sizes. They range
There are multiple approaches to probabilistic forecast- from highly undesirable (the bad, which lead to meaning-
ing. However, one that is particularly relevant in our case less results that cannot be compared) and undesirable
uses NNs to output the distributions directly. This can be (the ugly, which compromise research reproducibility) to
achieved by making only a small change in the DNN model worth recommending (the good). We conclude this article
presented earlier. Instead of returning the 24 hourly prices, by summarizing these practices—particularly common in
the network can return the parameter sets of 24 probabil- the ML EPF literature—and providing better alternatives.
ity distribution functions. Assuming that the distribution of
interest is Gaussian, the probabilistic NN will have 48 out- The Bad—Avoid at All Costs
puts with two parameters (i.e., the mean and standard devia- New models are rarely compared against established
tion) for each hour of a day, as illustrated in Figure 7. Com- state-of-the-art models and even less frequently against
pared to the DNN in Figure 5(b), the so-called distribution open access benchmarks. This leads to unjustified claims

Model Inputs Hidden Layer 1 Hidden Layer 2 Distribution Layer Model Output

µd,1
pd −1

f1
σd,1
pd −2 u1,1 u2,2

.
.
.
µd,2

pd −7 u1,2 u2,1
f2
σd,2
. .
. .
. .
Xd . .
. .
. .
u1,n u2,m
.
.
. µd,24

D f24
σd,24

figure 7. A probabilistic NN architecture for EPF. The output is a predictive distribution, fh, one for each hour of the day.

30 ieee power & energy magazine may/june 2022

regarding the “outstanding” performance of newly proposed ✔✔ to evaluate the models, several metrics are considered,
models. From a practitioner’s perspective, such results are and one of them is the relative MAE
meaningless. The cost of implementing a new predictive ✔✔ statistical testing is used to assess whether differences
model may be high, but its performance may be mediocre. in the predictive performance are significant
Another problem is the limited size and scope of the test ✔✔ the split and dates of the data set are explicitly stated,
data set. Results are usually reported for just one market and all inputs and parameters of the model are explic-
and often comprise a selection of only four weeks. This itly defined
favors selectively picking results and leads to inflated fore- ✔✔ the computational cost of new methods is evaluated
casting accuracy. and compared against that of existing approaches
The last issue is data contamination. The test data set is ✔✔ forecasting models are recalibrated daily, not sim-
not always selected as the last section of the full data set, nor ply estimated once, and evaluated in the full test
is the test data set completely independent of the validation/ data set.
training data sets. Moreover, model hyperparameters are
sometimes estimated in the test data set. Acknowledgment
Jesus Lago contributed to this work as an outside activity
The Ugly—Handle With Care and not as part of his role at Amazon.
The use of evaluation metrics is not standardized for the
price forecasting community. New methods are often com- For Further Reading
pared based on a single accuracy metric without considering R. J. Hyndman and G. Athanasopoulos, Forecasting: Prin-
its properties. Also, statistical testing is not always used to ciples and Practice, 3rd ed. Melbourne, Australia: OTexts,
measure whether the differences in accuracy are statistically 2021.
significant. This can lead to selectively picked results and G. James, D. Witten, T. Hastie, and R. Tibshirani, An In-
biased conclusions. troduction to Statistical Learning, 2nd ed. New York, NY,
Another issue is the lack of details when proposing USA: Springer Science & Business Media, 2021.
new methods. In particular, the split between training, J. Lago, G. Marcjasz, B. De Schutter, and R. Weron,
validation, and testing is often not reported nor are the “Forecasting day-ahead electricity prices: A review of
optimal input features of the model or optimal model state-of-the-art algorithms, best practices and an open-ac-
hyperparameters. This hinders reproducibility. Moreover, cess benchmark,” Appl. Energy, vol. 293, p. 116,983, Jul.
the computational costs of new methods are often ignored. 2021, doi: 10.1016/j.apenergy.2021.116983.
As short-term EPF often requires real-time forecasting, R. Weron, “Electricity price forecasting: A review of
the computational cost is an important metric that should the state-of-the-art with a look into the future,” Int. J. Fore-
be considered. casting, vol. 30, no. 4, pp. 1030–1081, 2014, doi: 10.1016/j.ij
Finally, models are not always recalibrated daily. This forecast.2014.08.008.
is particularly true for benchmarks. Often, a new proposed F. Ziel, “Forecasting electricity spot prices using LASSO:
model would be frequently recalibrated, but the benchmark On capturing the autoregressive intraday structure,” IEEE
is estimated only once and directly evaluated for the whole Trans. Power Syst., vol. 31, no. 6, pp. 4977–4987, 2016, doi:
data set. This leads to reporting worse accuracy and, when 10.1109/TPWRS.2016.2521545.
done for benchmarks, leads to artificial accuracy improve- I. Goodfellow, Y. Bengio, and A. Courville, Deep Learn-
ments of new models. ing. Cambridge, MA, USA: MIT Press, 2016. http://deep
learningbook.org
The Good—Best Practices
Based on the issues listed, new work in EPF should aim Biographies
to follow a series of good practices to ensure reproducibil- Arkadiusz Je˛drzejewski is with the Wrocław University of
ity and meaningful comparisons. In particular, it should Science and Technology, Wrocław, 50-370, Poland.
ensure that Jesus Lago is with Amazon, Amsterdam, 1097DP,
✔✔ the test data set comprises at least a year of data and is The Netherlands.
based on multiple markets Grzegorz Marcjasz is with the Wrocław University of
✔✔ the test data set is selected as the last section of the Science and Technology, Wrocław, 50-370, Poland.
available electricity price series, and hyperparameters Rafał Weron is with the Wrocław University of Science
are estimated using a validation data set that is differ- and Technology, Wrocław, 50-370, Poland, and DB Energy,
ent from the test data set Wrocław, 50-541, Poland.
✔✔ any new model is tested against well-known, state-
of-the-art, preferably open source models and well-
 p&e
known open access data sets

may/june 2022 ieee power & energy magazine 31

 By Spyros Chatzivasileiadis,
Andreas Venzke,
Jochen Stiasny,
and Georgios Misyris

Machine Learning
©SHUTTERSTOCK.COM/ANDREY SUSLOV

in Power Systems
W
WE EXPERIENCE THE POWER OF MACHINE LEARNING
(ML) in our everyday lives—be it picture and speech rec-
ognition, customized suggestions by virtual assistants, or just

Is It Time unlocking our phones. Its underlying mathematical princi-

ples have been applied since the middle of the last century

to Trust It? in what is known as statistical learning. However, the enor-

mous increase in computational power, even in devices as
small as a smartphone, has enabled significant advances and
wide adoption of ML in nearly every part of our lives and the
scientific world.
Still, despite the large body of academic literature on ML
approaches for power system operation during the past 30 years,
there is only an extremely limited set of approaches that has
Digital Object Identifier 10.1109/MPE.2022.3150810
found application in practice. From them, the vast majority is
Date of current version: 19 April 2022 related to load forecasting, such as the Artificial Neural Net

32 ieee power & energy magazine 1540-7977/22©2022IEEE may/june 2022

Short Term Load Forecaster, a tool developed by the Elec- computationally intensive, can predict well how voltage and
tric Power Research Institute in the United States; and a few current behave in many operating conditions. But compared
approaches have been related to decision trees for security to a few decades ago, power systems have become signifi-
assessment. The main reason for the slow adoption of ML in cantly more complex. The connection of millions of power
the power industry is that most ML approaches, such as neural electronic devices leads to rapidly evolving phenomena,
networks and others, are considered a black box. requiring the inclusion of more complex models and a faster
Who would trust a black box to avoid a blackout or to decision-making process. Distributed renewable genera-
find an optimal operating point that will not violate any line tion and electric vehicles add a lot of uncertainty and create
limits? Who would trust a black box for any safety-critical thousands of new injection points that continuously change
application, e.g., in energy, health care, or automobiles? The the balance between energy supply and demand. This
risk is just too high. all means one thing: if a system operator wants to make
What does it take for ML to get adopted? For insight into sure that no blackout happens (a procedure called security
the adoption of novel methods, not only related to ML, con- assessment), he or she needs to run not only more complex
sider the following two aspects: models but also for a lot more scenarios at faster paces than
1) To what extent does the industry already have solu- before. A security assessment that once took a few hours to
tions to a problem? carry out would now need a few days, which is unacceptable
2) What is the risk associated with using a novel method for meeting real-time grid-operational requirements.
instead of a conventional one? ML methods, on the other hand, can continuously learn
A novel method will only be seriously considered if it falls and adapt to their environment and are extremely fast when
below a risk threshold. If it satisfies the risk requirements, its computing an output. Such practices have been shown to
better performance, such as higher accuracy or computation outperform conventional techniques, e.g., predictive main-
speed, can lead to adoption. Figure 1 presents an illustration tenance of transmission lines and transformers, or smart
of these considerations. charging of electric vehicles. ML approaches are gradu-
ML, and artificial intelligence (AI) in general, has been ally being adopted for such applications because of their
particularly successful in areas where there are little to no low risk to power system operation. However, handling the
established solution methods. AlphaGo is an AI program sheer complexity of power system operation procedures
developed by DeepMind Technologies that harnessed the to avoid blackouts (security assessment) or determine an
complexity of the game Go and repeatedly beat the human optimal operating point without violating any single oper-
world champion and other software programs. By itself, this ational constraint is still too “safety critical” to accept an
feat is impressive if we consider that Go is an ancient Chi- ML-based solution. This, despite early promising results
nese game with an astronomical number of possible combi- reported in the research literature that ML can help. Power
nations of player moves (2.1 × 10170), vastly greater than the system operators find it difficult to trust methods they do
number of known atoms in the known, observable universe not understand, and which have thus far provided no perfor-
(1 × 1080). AlphaFold is an AI program that predicted, better mance guarantees.
than any other program, protein structures that were impossi- Researchers have recently been working to address
ble to predict for the past 50 years. Both AlphaGo and Alpha- this skepticism and remove barriers, allowing ML to enter
Fold yielded results never before possible. Because the risk of power system applications, exploiting its benefits. In the
catastrophe associated with their task was nonexistent, their first transition period of using ML tools for power system
adoption was extremely rapid. On the other hand, self-driving safety-critical operations, the authors view ML algorithms
cars, which heavily use ML algorithms to detect surround- as “assistants” to established procedures in the form of a
ings, are also attempting to solve
an unsolved problem, but in sev-
eral cases, and despite significant Risk Level
progress, we are still above the risk
threshold when it comes to their Unacceptable ML in Security
widespread deployment without Autonomous Assessment
Driving
the involvement of a human driver.
Threshold
Power systems are different.
We have spent the past 100 years AlphaGo Load Forecasting
trying to understand how electric- Acceptable AlphaFold Smart Charging Agent
ity flows along power lines and Predictive Maintenance of Grid Assets
Maturity
what happens in the voltage and No Methods Established Methods
current right after a disturbance.
And we have managed that quite figure 1. The adoption of novel methods (blue area) in light of the existence or non-
well. The models, despite being existence of alternative solution techniques, and the risks the novel processes pose.

may/june 2022 ieee power & energy magazine 33

 decision-support tool. The ultimate responsibility for the ral network training procedure. Instead of generating
final decision remains with the operator. ML algorithms sets of relevant data in a process to train the model,
can be extremely fast to estimate, for example, whether which can be computationally intensive, integrate a
operating points are safe, as ML methods can assess thou- physics-based model into the training process and let
sands of scenarios at the same time, whereas conventional the neural network learn from it.
tools can assess only a few. This allows the system operator Before diving into these approaches in the next sections, we
to screen an extremely large number of scenarios and deter- first introduce the type of ML tools that are the focus of this
mine the few critical ones in a very short period of time. article: neural networks and their training procedure. In the
Following this step, the operator can then use conventional rest of this article, we use the terms AI and ML interchange-
techniques to study, in more detail, those few scenarios and ably; in reality, AI is a more general term that includes all
determine what drives them to instability and when. ML methods and tools.
The approach is similar when we apply ML, and AI in
general, for optimization. Neural networks can determine How Do Neural Networks Work?
a very good estimate for an optimal operating point. We Neural networks belong to the most promising group of
can then assess, with conventional tools, whether the point AI tools, proving their ability to approximate any function
violates any constraints. This way of approaching AI can (universal approximator). They have received considerable
drive wider adoption of AI tools for system operation, at attention in recent years. The two main categories are neural
least in the first transition period. But even before this step, networks for classification and regression tasks.
the AI tools need to build trust. If operators and other users Classification neural networks are used in a variety of
do not trust them, they will not even be used for screening. applications. In self-driving cars, for example, neural net-
The rigorous methods this article describes aim to deliver works continuously receive pictures of their surroundings
the guarantees (through performance certificates) that will from car cameras and answer simple questions such as, “Is
allow users to trust AI tools for safety-critical operations. this a red traffic light?” A classification neural network
Besides the benefits these certificates give us in terms of shall return a yes or a no, and even if it is not quite sure,
trusting AI tools and determining the worst-case perfor- it must choose between the two. For power systems, such a
mance, they can help provide better insights into how the classification task can be “Will this combination of genera-
resulting AI model can be improved. The techniques dis- tor set points and loads—called an operating point—lead
cussed in this article fall into the broader spectrum of trans- to a blackout or not?” We can ask this question for many
parent AI (the results of the AI process are explainable) and possible operating points (see the “Possible Inputs” bubble
interpretable AI (the accuracy of a model to associate a cause in Figure 2), and through the neural network, we can obtain
to an effect), initiatives recently launched in the AI commu- an answer for each point extremely fast. We can label the
nity to bring a deeper understanding of the inner workings resulting regions “safe” (blue dot) or “not safe” (red dots), as
of ML algorithms. shown in the “Predicted Classifications” bubble in Figure 2.
Several algorithms under this umbrella deliver rigorous If we want to assess the predictions of our neural net-
approaches to understand how different features of the prob- work, we obviously cannot check for every single operat-
lem at hand affect the performance of the applied neural ing point as this requires immense computing resources.
network. The algorithms we discuss go a step further and Instead, we sample a fraction of operating points and per-
deliver performance guarantees. To enhance the perfor- form a conventional security assessment (e.g., run power
mance of AI tools, this article also presents ways to take flow, time-domain simulations, or an eigenvalue analysis)
advantage of available power system models developed for and associate a true label with each point, such as truly safe
decades based on first principles. Through that, neural net- or truly unsafe. A good neural network will match (i.e., pre-
work applications can better avoid overlooking critical cases dict) these true labels as closely as possible.
or giving predictions that are largely at odds with governing Instead of a binary response (yes/no), regression neural
physics. In short, these possibilities are enabled by the fol- networks use almost the same training process as the clas-
lowing two approaches: sification neural networks but yield a continuous function
1) Verification: This approach delivers certificates on value. These networks are widely used, for example, to fore-
how the neural network will behave, i.e., what its out- cast load demand or predict the power output of wind tur-
put will be for all possible inputs. Thus far, research- bines and solar photovoltaics based on weather conditions,
ers have been assessing the neural network perfor- including the movement of clouds, temperature variations,
mance purely statistically, naïvely hoping that testing and others. In a power system operation context, a regression
an application on a large set of random samples can neural network can predict a cost-optimal operating point
accurately capture the behavior of the neural network or how the value of frequency evolves after a line outage.
model over the whole input space. As presented in Figure 2, in the case of frequency evolu-
2) Physics-informed neural networks: These networks tion, a neural network predicts a continuous function of the
include the governing physical equations inside a neu- input to output variables over time (dashed orange line) that

34 ieee power & energy magazine may/june 2022

approximates the “true” trajectory (dashed blue line). A good whereas the nonlinear activation function in each neuron
neural network will do so accurately. remains unchanged.
Classification and regression neural networks have the In many neural network implementations, neurons act
same structure (see the right side of Figure 2). Neurons are as a switch, either allowing the signal to pass or restrict-
organized in layers, with each layer connected in series ing it from doing so. Because of these nonlinear activation
with each other. Each neuron contains a nonlinear acti- functions, neural networks can represent very complex, and
vation function. The neurons in each layer are connected sometimes not even explainable, nonlinear processes. The
through linear functions with some or all of the neurons of magnification of a single neuron in Figure 2 shows one
the previous layer and the next layer. In this simple struc- option for how such a nonlinear function can look, but there
ture, the linear functions contain parameters we can adjust, are many more. With the appropriate selection of linear

Classification Neural Network Output

Predicted
Possible Inputs Architecture
Classifications

Layer
Input

Neural Network

In
Out
In
Out
Regression In
Output
Prediction

Neural Network Nonlinear

Linear
Input Ground Truth Activation
Transformation
Function

Input

figure 2. The basic structure of a classification and regression task, and the fundamental architecture of a simple neural
network that forms the core of either task.

Data Creation Neural Network Training Neural Network Testing

Simulating Statistical
Evaluation
Physical
Models Training Testing
Neural
Training Data Physical Physical
Models Network Models
Performance
Guarantees

Measuring Testing Data

Additional Elements

figure 3. The classical ML procedure, and the new elements for improving the procedure.

may/june 2022 ieee power & energy magazine 35

 function parameters and neural network size, in theory, the following sections, we explain how we adjust this process
network has the power to approximate any possible function to add the physical models inside the neural network train-
or process. ing in the form of continuous equations. But first, we focus
The process of determining these parameters is com- on how to obtain rigorous performance guarantees that are
monly referred to as the training of a neural network. It is a not based on statistical indices and are not dependent on the
part of a general ML algorithm, together with the creation of quality of the test set.
the training data set beforehand and the testing carried out
afterward. Figure 3 illustrates this procedure and highlights Guarantees of Neural Network Behavior
where new elements of neural network verification and phys- Thus far, conventional methods evaluate how well a neural
ics-informed neural networks enter the process. network performs by measuring its performance on a test set
Before neural network training, we start by putting together (e.g., accuracy and other statistical indices). However, this
a data set with both actual measurements and simulation process is purely statistical and cannot tell us with certainty
results (if the measurements are not enough). This data set what the prediction will be for any points not included in
is usually a representative scenario, yet it is a small frac- this set, which is crucial for any safety-critical applications.
tion of the possible scenarios that could occur. It contains The goal is to develop techniques that evaluate whole con-
both the inputs and the outputs, e.g., what is the operating tinuous regions instead of just discrete points when testing a
point and whether it is “safe” or “unsafe?” The goal is that neural network. This allows us to consider any possible point
through this data set, the neural network learns the rela- within an input region and guarantee the neural network per-
tionship between inputs and outputs so that when training formance for any point in this region. We achieve that using
is over, it can correctly estimate the output for any unseen optimization methods. As it turns out, we can rewrite the
input not included in the training data set. To assess how equations that define a neural network such that they can
well the neural network predicts that output, we use a sepa- be incorporated into an optimization problem. Due to this
rate test data set to statistically determine the neural net- reformulation, we can now analyze a neural network in a
work’s performance. The test data set is often a smaller completely new way.
part of the one generated for training but kept aside to use Methods such as the ones mentioned in this article can
only for testing. provide system operators with guarantees about how a neu-
The next step in the ML algorithm is to determine the ral network will behave for entire regions of power system
neural network’s parameters so that predicted outputs match operation. At the same time, they eliminate the dependency
the true ones from the training data set as closely as possible. on the quality of the test data set. Instead of having to sample
The procedure is an iterative optimization algorithm, where a potentially extremely large number of test data to cover all
we pass batches of data several times through the neural possible scenarios (and still not extract any guarantee or pro-
network until we arrive at a good match between network vide any certainty), we can now solve a single optimization
outputs and those in the training set. When we are satisfied problem for one continuous part of the input region.
with the performance on the training data set, or we reach a To illustrate this framework, we examine the following
predetermined time limit, we fix the parameters. two vital questions, one each for a classification and regres-
After training comes testing. Depending on how accu- sion task:
rate the neural network’s predictions are on a random set 1) How large is the continuous input region for which
of previously unseen test samples, we conclude how “good” the neural network classification remains the same?
the network is compared to others. But what happens if the The answer provides a guarantee that any input in
samples do not reflect reality? What if the neural network this region will be classified to a specific known
achieves 99% accuracy but the test set fails to include critical class. Let us consider a power system. Such a guar-
cases or even some high-risk outliers? How can an operator antee can state that the all operating points where
trust this performance index? We address this in the next generator one is between 0 and 200 MW, generator
section by describing methods that do not depend on the test two is between 0 and 100 MW, and the load is be-
data set to determine the network performance. tween 0 and 300 MW will be classified as “safe” by
The creation of a neural network is purely based on data the neural network.
and statistical learning. One could even say that training a 2) What is the largest prediction error of the neural net-
neural network is an advanced form of nonlinear regression. work across a continuous input region (regression)?
Physical models have thus far not been involved in training. Considering power systems, assume a simple exam-
In traditional ML algorithms, a physical model influences ple: we have two generators serving a load over two
only the data-generation stage, either because data were transmission lines; we train a neural network to output
generated through simulations using first-principles models the combination of generator set points that result in
or the data were collected from the actual physical process. the minimum cost 1) for a load that varies between
The physical models, in this case, are represented in a dis- 10 and 300 MW and 2) without violating the trans-
crete form through data points but not in equations. In the mission capacity of any line. A worst-case guarantee

36 ieee power & energy magazine may/june 2022

 can indicate that across all combinations of generator value analyses), we train a neural network to assess whether
set points (i.e., one combination for 10 MW, one for the operating point is secure or not secure in a fraction of
10.1 MW, and so on), there is a 5% maximum viola- the time (usually 100 to 1,000 times faster). As inputs, we
tion of the line transmission limit. provide the active power set points of generators and active
The methods we describe allow for such an assessment. If and reactive power consumption of loads. The boundary
the performance is unsatisfactory, we can take a step back, where the classification changes from secure to not secure
retrain the neural network on more data points, and re-evalu- is called the security boundary. The correct prediction of
ate the performance. This improves the accuracy and robust- that boundary is crucial because the neural network could
ness of the prediction, ultimately delivering better guaran- misclassify a “not secure” operating point as “secure,” pos-
tees and enabling the adoption of such techniques. ing a risk to power system operation and potentially leading
to a blackout.
Secure or Not Secure? Guaranteeing Using this guiding example, Figure 4(a) and (b) shows a
How Neural Networks Classify comparison of the standard procedure used to evaluate the
Let us consider a guiding example: we want to determine performance and the proposed methodology, respectively.
whether a specific operating point of a given power sys- The conventional procedure relies on a test data set of prob-
tem is safe under a combination of security criteria (e.g., able power system operating scenarios. For this test data set,
N-1 criterion, small-signal stability, and so on). Instead of the neural network classification is compared to the ground
using conventional methods (e.g., power flows and eigen- truth, assessing classification accuracy.

Conventional Performance Assessment (Statistical Evaluation)

Possible Power System Operating Scenarios Security Classification by Neural Network

Secure

Trained
Neural Network

Predicted Not Secure

Security Boundary
Ground Truth
Security Boundary
(a)

Obtaining Performance Guarantees

Possible Power System Operating Regions Guaranteed Classification by Neural Network

Secure

Trained
Neural Network

Predicted Not Secure

Security Boundary
Add Samples to Training
Ground Truth Data Set to Improve
Security Boundary Robustness
(b)

figure 4. A comparison of the standard evaluation of the neural network performance with the new proposed methodol-
ogy. (a) The conventional procedures use a test set, and (b) the proposed techniques obtain performance guarantees for
continuous input regions.

may/june 2022 ieee power & energy magazine 37

 The drawback of this approach is that it requires many robustness. The moment the operator has the performance cer-
samples to build confidence in the correctness of the tificates and is satisfied with the neural network performance
network prediction, including prediction of the security is when we will see neural networks applied in safety-critical
boundary. As illustrated in Figure 4, using only a small power system applications, such as security assessment.
number of samples could suggest that the boundary is pre-
dicted accurately, although there may still be an undiscov- Worst-Case Guarantees: What Is the Worst
ered mismatch. Related to this is that even for a larger num- Neural Network Prediction Error?
ber of samples, an effective sampling strategy is difficult Let us consider a guiding example where a neural network
to devise along the security boundary, especially if many acts as an optimal power flow algorithm. It predicts how
power system states are considered. A second shortcoming much controllable generators should produce for a given
of the traditional evaluation is that there are no guarantees load situation so that 1) the production cost is as inexpensive
the neural network classification will not falsely change as possible and 2) there is no violation of system constraints,
between discrete samples. such as transmission line loading limits. Existing optimal
Instead, we apply a new method that reformulates the power flow algorithms can perform this task quite well.
neural network to an optimization problem. By solving it, However, the optimal power flow problem in its full non-
we can certify how the neural network classifies entire input linear form is still a major challenge to solve in a reasonable
regions (grey rectangles), i.e., as secure or not secure [Fig- time. Well-trained neural networks can determine a solution
ure 4(b)]. The procedure works as follows: up to 1,000 times faster and consider constraints that are
✔✔ First we take a discrete operating point. This could impossible to include directly in conventional optimization
be a random sample from the training set or any other methods (e.g., differential equations representing stability
operating point for which we know it is secure (the constraints). Now imagine that we need to run thousands of
process is very similar for points that are not secure). scenarios through an optimal power flow algorithm, consid-
✔✔ Then we solve an optimization problem to compute ering different load profiles and highly varying injections
the closest input for which the neural network predic- from renewables, and determine the cost-optimal and safe
tion changes from secure to not secure. generation dispatch for all of them. That is when neural net-
The distance between the reference point and this input works can be quite helpful.
determines the size of the region around the reference oper- We have to keep in mind, however, that a neural network
ating point where the classification is guaranteed not to does not necessarily predict a solution that respects all sys-
change. In other words, the result of this optimization sup- tem constraints and is also the least expensive option. Its
plies us with the guarantee that all the points inside this strength is its extremely fast computation, not necessarily its
region will be classified by the neural network as secure. accuracy. When testing the network, we want to ensure that
This is represented as the colored rectangles in Figure 4 (i.e., predictions 1) do not significantly violate constraints (e.g.,
each rectangle corresponds to a different region we certify). do not breach more than 1% of the maximum limits) and 2)
Using that, the operator can start trusting the neural network remain close to the optimal solution from a cost perspective
as they can now anticipate how it will behave for any pos- (e.g., they determine a solution that does not cost more than
sible point in the certified regions. 1% of the optimal cost).
The neural network is no longer a black box. We can Similar to classification neural networks, the standard
repeat this process for many reference points until we approach used to assess network performance after train-
obtain several regions with certified mapping and unveil ing is to take all samples of the test data set, compare the
the neural network behavior across the input domain to sys- network prediction to the physical ground truth, and com-
tem operators. pute whether and to what extent constraint violations occur
This procedure can also be used to evaluate the neural for specific samples (see Figure 5(a)]. As an additional
network robustness illustrated in Figure 4. For the point we metric, we can also calculate the Euclidean distance of
determined through our optimization, we can also com- the neural network prediction to the ground-truth optimal
pute a ground-truth classification and assess whether the point for these points to calculate the “optimality” of the
neural network predicts the security boundary correctly. prediction. Reviewing all test data set samples, we can then
Thus far, existing approaches could only evaluate misclas- easily find the maximum of these violations and maximum
sification through random sampling. This is the first time that “optimality distance.” However, this process does not give
a systematic procedure can measure how well a neural net- any guarantee that the worst prediction that can occur was
work predicts a power system’s security boundary. If we are found. With conventional neural network assessment meth-
not satisfied with the network performance, we can add the ods, the only option for improving this is to evaluate the
misclassified points to the training data set with the correct metrics on even more data samples. Beyond a point, the
ground-truth labels and retrain the neural network to improve procedure becomes computationally expensive and still
its robustness. This is a systematic and well-defined proce- does not yield certainty that an even worse prediction has
dure that can be repeated until we reach a desired level of not been missed.

38 ieee power & energy magazine may/june 2022

 A recently introduced method, however, bypasses all or whether further training and evaluation are required to
these barriers as it eliminates the need to evaluate discrete reach an acceptable performance.
test samples. The trained neural network is reformulated to a
set of linear equations with continuous and binary variables Physics-Informed Neural Networks
to set up an optimization problem. Inside the optimization The previous sections focused on how to provide guarantees
problem, we also include the physical power system model- about the performance of a trained neural network, building
ing equations. In essence, these act as a representation of the trust of system operators and other users. In this section,
the ground truth. Note that this procedure is applied only we discuss how neural networks can take advantage of the
to extract worst-case guarantees. As soon as the user (e.g., decades-long development of physical power system models.
power system operator, trader, and so on) is satisfied with the Including this knowledge in neural network training yields
neural network performance, the trained neural network can significantly better performance with much fewer data.
be deployed for real application. These are so-called physics-informed networks.
The goal of this method is to find the maximum con- First let us look at a concrete example. Assume we want
straint violation, which can now run across the whole con- to determine the evolution of power system frequency under
tinuous input region [Figure 5(b)]. If we find that the maxi- a disturbance. The models that describe this evolution often
mum constraint violation is zero, we can then certify that for use differential equations that express how the system
the entire input region, the neural network predictions will changes from one instance to the next (the simplest of which
never violate the constraints. If the maximum constraint is a “swing equation” that many power engineers learn dur-
violation is not zero, we then obtain worst-case guarantees, ing their studies). To solve a system of differential equations,
i.e., we determine the maximum violation that the neural we usually employ computer software and numerical solv-
network prediction can result in for any possible input it ers which, in most cases, means that we track in small steps
can receive. This then becomes a powerful tool that can how the system state (e.g., the frequency) changes over time.
build the trust lacking from grid operators. When consid- The resulting trajectory can become a very complex func-
ered from a risk assessment perspective, these metrics can tion, possibly requiring significant computational resources.
decide whether the network’s performance is good enough And the rapidly increasing integration of power electronic

Conventional Performance Assessment (Statistical Evaluation)

Possible Power System Operating Scenarios Sample-Based Estimation of Worst-Case Performance

Trained
Neural Network

Dispatch Prediction by Neural Network Physical Ground Truth

Obtaining Performance Guarantees

Possible Power System Operating Region Worst-Case Performance Guarantees for Entire Region

Trained
Neural Network

Dispatch Prediction by Neural Network Physical Ground Truth

figure 5. A comparison of the (a) standard evaluation of neural network performance using a test set with (b) the
proposed methodology for obtaining guarantees of neural network behavior. The proposed method can determine the
largest deviation of the neural network prediction from the physical ground truth across the entire region (and not only for
random discrete samples): this offers a rigorous, worst-case guarantee.

may/june 2022 ieee power & energy magazine 39

 converters and renewables makes this computational effort is governed by the model’s differential equations, we see
significantly larger for the following two reasons: clear mismatches for all three examples. The prediction in
1) Power electronic converters (e.g., from wind turbines, Figure 6(c), in particular, highlights what is referred to as
solar photovoltaics, batteries, high-voltage dc lines, overfitting. The prediction matches all data points nearly
electric vehicles, and so on) have much more com- perfectly but shows large approximation errors between
plex models based on differential equations, with time the data points.
constants that interfere with the electromagnetic tran- Now imagine that we want our neural network to predict
sients of transmission lines. a series of different trajectories that correspond to different
2) Renewables add significant degrees of uncertainty disturbances and operating points. Poor fitting or overfitting
with their fluctuating production, meaning that we becomes a major issue when checking whether the under-
need to assess a much larger set of possible operating frequency or overfrequency thresholds have been violated.
points for each disturbance. One way to achieve a better fit is to use more data points
Neural networks could assist here by substituting traditional from more trajectories. However, this can be difficult if, for
numerical solvers and providing a fast estimate of the actual example, the data are obtained from measurement devices
solution (up to 1,000 times faster). This creates the following that do not sample the frequency to an adequate degree of
two major benefits: granularity. It also can be computationally expensive if we
1) A much faster estimate (within a few milliseconds) of need to generate more trajectories with simulation software
what the frequency will be 2 or 5s after the distur- to obtain more points.
bance, enabling the activation of necessary counterac- Recent advances in neural network training, however,
tions earlier. offer us a new way. We now can numerically compute the
2) The ability to assess up to 1,000 potentially critical derivatives of neural network outputs with respect to inputs
disturbances in the same time frame that conventional during the neural network process. If the frequency is a
software will calculate only a single disturbance. This neural network output, we can calculate, for example, the
allows us to screen a very large set of disturbances derivative of the frequency over time inside the network
for many different operating points and select only the training. This enables us to include differential equations,
most critical ones, which we can then assess in more such as the swing equation, inside the neural network train-
detail and with higher accuracy than with convention- ing to drive the training procedure to a neural network
al software. where the frequency (which is a neural network output) and
To train such a neural network with conventional meth- its derivative validate the swing equation. This additional
ods, we need to provide data points (the black dots in requirement leads to a much better approximation of the
Figure 6) and fit network parameters so they provide a true functions.
prediction that matches the data points. As the training Figure 7 (dashed black lines) shows the derivatives at
process progresses, the prediction function (orange line) selected points. Intuitively speaking, by adding physical
becomes more sophisticated and fits the data points better. equations inside the neural network training, the network
However, when comparing the prediction to the function prediction needs to cross a data point (black dot) and cre-
that produced the data points (i.e., ground truth), which ate a trajectory that fits the “shape” of its neighborhood.

Increases During Training

“Underfitting” “Overfitting”

Frequency Frequency Frequency

Time Time Time

(a) (b) (c)

figure 6. The training process of a simple regression neural network, showing the evolution of network output during
training (orange) versus ground truth (blue). (a) At the start of the training, the neural network output is “underfitting.”
(b) As the training progresses, the neural network output fits the data points better. (c) If the training continues,
the neural network output is “overfitting.”

40 ieee power & energy magazine may/june 2022

 Increases During Training
Prediction Contradicts Physics Prediction in Line With Physics
Frequency Frequency Frequency

NN
Physical Model
Derivative
Time Time Time
(a) (b) (c)

figure 7. The training process for physics-informed neural networks, showing the evolution of neural network output dur-
ing training (orange) versus ground truth (blue). NN: neural network. (a) At the start of training, the neural network output
contradicts the physical model. (b) As the training progresses, the neural network output fits the physical model better.
(c) After sufficient training, the neural network output is in line with the physical model.

Figure 7(b) and (c) show how the neural network adjusts its For Further Reading
parameters so that the derivatives of the predicted trajec- A. Venzke and S. Chatzivasileiadis, “Verification of neural
tory can fit what the differential equations describing the network behaviour: Formal guarantees for power system
system dictate. applications,” IEEE Trans. Smart Grid, vol. 12, no. 1, pp.
Physics-informed neural networks may lead to much bet- 383–397, Jan. 2021, doi: 10.1109/TSG.2020.3009401.
ter predictions while requiring much fewer training data. A. Venzke, G. Qu, S. Low, and S. Chatzivasileiadis,
This can remove significant barriers. Instead of creating “Learning optimal power flow: Worst-case guarantees for
large training databases by randomly sampling thousands of neural networks,” in Proc. 2020 IEEE Int. Conf. Commun.,
trajectories to train a network with sufficient accuracy, we Control, Comput. Technol. Smart Grids (SmartGridComm),
can now directly include governing equations in the training pp. 1–7, doi: 10.1109/SmartGridComm47815.2020.9302963.
process and let the neural network train until it minimizes G. S. Misyris, A. Venzke, and S. Chatzivasileiadis,
the prediction error. “Physics-informed neural networks for power systems,” 2020,
in Proc. IEEE Power & Energy Society General Meeting,
Where Does This Lead Us? 2020, pp. 1–5, doi: 10.1109/PESGM41954.2020.9282004.
The two directions we presented, neural network verification L. Duchesne, E. Karangelos, and L. Wehenkel, “Recent
and physics-informed neural networks, aim to build trust developments in machine learning for energy systems reli-
with grid operators, utilities, and other potential users and ability management,” Proc. IEEE, vol. 108, no. 9, pp. 1656–
improve network performance while requiring much fewer 1676, 2020, doi: 10.1109/JPROC.2020.2988715.
data. They arguably remove the most important barrier V. Tjeng, K. Xiao, and R. Tedrake, “Evaluating robust-
when it comes to the safety-critical operations of energy sys- ness of neural networks with mixed integer programming,”
tems: ML tools no longer need to be considered a black box; 2017, arXiv:1711.07356.
instead, they can be trusted. Such methods can also further M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-
help to develop systematic approaches that identify regions informed neural networks: A deep learning framework for
where the neural network demonstrates poor performance solving forward and inverse problems involving non-linear
and then systematically improve it through retraining. partial differential equations,” J. Comput. Phys., vol. 378,
There are still many steps needed until ML tools are pp. 686–707, Feb. 2019, doi: 10.1016/j.jcp.2018.10.045.
embraced inside the control room. For instance, scalable
algorithms that can apply to very large neural networks and Biographies
power grids need to be developed. Expertise in power sys- Spyros Chatzivasileiadis is with the Technical University of
tems, operations research, and ML is necessary to arrive Denmark, Kongens Lyngby, 2800, Denmark.
at production-level algorithms. Once achieved, gaining the Andreas Venzke was previously with the Technical Uni-
trust of system operators will be a deciding step toward the versity of Denmark, Kongens Lyngby, 2800, Denmark.
wider adoption of ML algorithms across energy systems. Jochen Stiasny is with the Technical University of Den-
This includes safety-critical operations, where ML can dras- mark, Kongens Lyngby, 2800, Denmark.
tically accelerate computation speed and enable the manage- Georgios Misyris was previously with the Technical Uni-
ment of millions of controllable appliances and converter- versity of Denmark, Kongens Lyngby, 2800, Denmark.
 p&e
connected devices.

may/june 2022 ieee power & energy magazine 41

 Demonstration of
Intelligent HVAC
Load Management
With Deep
Reinforcement
Learning
By Yan Du, Fangxing Li, Kuldeep Kurte,
Jeffrey Munk, and Helia Zandi

B
BUILDINGS ACCOUNT FOR 40% OF TOTAL PRIMARY thermal dynamics of the HVAC system and its interaction
energy consumption and 30% of all CO2 emissions world- with the ambient environment. The development of such a
wide. A large portion of building energy consumption is due model may introduce measurement and prediction errors,
to heating, ventilation, and air-conditioning (HVAC) sys- which may undermine the control performance. In addition,
tems. In the summer, for example, more than 50% of a build- models developed for one building may not generalize well for
ing’s electricity consumption is used for cooling. With proper other buildings or unseen operation environments.
energy management, buildings can provide load shifting, In contrast to the model-based approaches, model-free
peak shaving, frequency regulation, and many other demand algorithms require no prior knowledge of the physical
response services. model, such as the thermal-dynamic model in the HVAC
Many existing approaches for building energy manage- control case; rather, they learn the model through estimation
ment are model based and require the modeling of the complex and exploration. One representative model-free approach is
reinforcement learning (RL). As shown in Figure 1, an agent
Digital Object Identifier 10.1109/MPE.2022.3150825
(e.g., a demand response controller) interacts with an envi-
Date of current version: 19 April 2022 ronment (for instance, the building). At each control time

42 ieee power & energy magazine 1540-7977/22©2022IEEE may/june 2022

 Real-World Experience of Machine
Learning in Demand Control

interval, the agent observes the state of the environment and output either the estimated value of the actions or the optimal
takes an action, making the environment transit from one action (depending on the specific deep RL approach) at the
state to another. The environment can send feedback in the current state due to the strong feature extraction ability of its
form of a reward to the agent. The agent improves its action multiple hidden layers. Deep RL is model free since the DNN
based on that reward. is trained through interacting with and without prior knowl-
This process repeats until the environment reaches the edge of the environment. In addition, the deep RL approach
terminating state. The goal of the agent is to learn an action also has high generalization abilities for new environments.
policy that maximizes its cumulative reward from all states. A well-trained DNN can be regarded as a function with fine-
In this learning process, the agent does not rely on prior tuned parameters. Whether a state as input is seen before or
knowledge to make decisions, but it gradually formulates an not, a well-trained DNN can always generate an output.
optimal action policy through intelligent “trial-and-error.” In this article, we introduce an end-to-end workflow for
As a result, the model-free approach has greater flex- developing a deep RL-based residential HVAC controller
ibility in solving control and optimization problems with that can control multiple zones, where the zones represent
unknown models or partial observabilities. different floors in a house. In particular, we describe two
Deep RL, which is a combination of deep learning and deep RL approaches for HVAC control, present the evalua-
RL, is a more recently developed approach. The key idea behind tion of the deep RL-based HVAC control strategies through
deep RL is the adoption of a deep neural network simulation studies, and discuss the deployment of a deep
(DNN) to let the agent learn an optimal policy. As shown RL-based HVAC control approach in a real-world residen-
in Figure 1, the DNN is a neural network with multiple hid- tial house. Finally, we analyze the deployment results and
den layers. With the given input as the state, the DNN can provide conclusions.

©SHUTTERSTOCK.COM/IGORDEYKA

may/june 2022 ieee power & energy magazine 43

 By optimally designating the setpoint of the HVAC system,
the energy cost can be minimized while keeping the indoor
temperature within the user’s comfort range.

The Deep RL Approach ✔✔ the time of day

for HVAC Control ✔✔ the current indoor temperature
✔✔ the current outdoor temperature
Schematic Description of Deep RL ✔ ✔ a 6-h look-ahead outdoor temperature series for
for the HVAC Control Problem planning
A critical premise for applying deep RL approaches is ✔✔ the current retail price
that the problem under investigation is a time-sequential ✔✔ a 6-h look-ahead retail price series
decision-making process. At each time step, the current ✔✔ the lower bound of the user comfort level
state is only related to the previous state, and the opti- ✔✔ the maximum retail price within the next 6 h
mal decision can be made based only on the current state ✔✔ the length of time to reach the next price peak.
information. This is the case for the optimal control of a The action is the setting of the HVAC system setpoint.
multizone HVAC system. The indoor temperature at It can be either discretely or continuously adjusted within
the current state is only related to the parameters at a certain range. The reward is defined as the negative
the previous time interval, and it is not affected by the sum of the energy consumption cost and comfort viola-
indoor temperature at earlier intervals. By optimally des- tion cost for the control interval, and the comfort viola-
ignating the setpoint of the HVAC system, the energy cost tion cost is calculated based on how many degrees the
can be minimized while keeping the indoor temperature indoor temperature deviates from the user comfort level.
within the user’s comfort range. Without losing general- The environment is the entire building or house includ-
ity, in the following discussion, we assume that all HVAC ing the HVAC system.
zones need heating. Also, a zone can simply be regarded This process of applying deep RL for HVAC control
as a floor in a house in this study. is illustrated in Figure 1. Note that we do not define the
When applying deep RL approaches, four essential ele- state transition probability for the process. The prob-
ments should be first defined: the state (s), action (a), state ability refers to the probability of transitioning to a
transition probability (p), and reward (r). In the context of specific next state after taking an action at the current
a multizone residential HVAC control problem, the state state. If the state transition probability model is known,
includes the following factors: the HVAC control problem can be explicitly formulated
and solved analytically. How-
ever, obtaining an accurate state
transition probability model for
Agent the HVAC control problem is
(DNN-Based RL Controller) not a trivial task. This is because
the thermal-dynamic model of
buildings with HVAC systems
is related to a variety of param-
eters, including resistances and
Input Hidden
Layer Output capacitors from different build-
State: ing components; weather factors,
Reward:
Time,
such as outdoor temperature and
Energy Cost + Action:
Temperature, solar irradiance; and so on.
Comfort Violation (HVAC Setpoint)
Retail Price, and Cost Therefore, as previously men-
User Comfort Level tioned, a model-free approach like
RL is more suitable for solving
Environment: the HVAC control problem. Fur-
(Building With HVAC
ther, the building models can
System)
vary, and we need a more gener-
alized and robust HVAC control
figure 1. The RL-based approach for HVAC control. DNN: deep neural network. approach that can work efficiently

44 ieee power & energy magazine may/june 2022

in different building environments. Deep RL, with the pow- algorithm does not require the action space to be discretized.
erful embedded DNN providing more generalization and Given the setpoint range as [20 °C, 22 °C], the algorithm
adaptability, is an ideal approach for achieving a flexible and can generate one continuous number within the range. The
intelligent HVAC control strategy. More details of the deep ability to handle a continuous action space makes the DDPG
RL approach will be discussed in the next sections. algorithm more suitable for solving problems where the con-
trol action is a continuous variable.
Deep-Q Network for HVAC Control The DDPG algorithm can be regarded as an extension of
The deep-Q network (DQN) is a combination of Q-learn- the DQN algorithm. In the algorithm, there are two types of
ing and a DNN. The main idea behind the DQN is to build neural networks applied: the actor and critic networks. The
a neural network to estimate the optimal control action actor network outputs a deterministic control action based
in a discrete domain. The neural network functions like a on the given current state. The critic network outputs the
complex lookup table, with the input being a state and the Q-value based on both the state and action provided by the
output being the action values for all of the possible control actor network. The actor network is further updated by maxi-
actions at the current state. In the DQN, this action value is mizing the Q-value under the current policy, and the critic
called the Q-value, where a higher value indicates a better network is updated by minimizing the mean square error
or more effective action. This DQN approach selects the of the Q-value, which is the same as the DQN algorithm. In
action with the highest Q-value
evaluated by the neural network.
The DQN approach adopts a
DQN RL Agent
stabilization strategy to train two
neural networks simultaneously, Residential HVAC Update L = DQN Loss
System Behavior
namely, the target and behavior Network
networks. The function of the tar- System State Current Q-Values Target Q-Values
get network is to provide stable-
labeled samples for the behavior Soft
Update Target
network to learn. The DQN con- Setpoints Behavior
Network Network
verges when the outputs from the
Environment
two networks are close to each
other. An overview of the DQN Data
Collection <state and action> <next state> <reward>
approach for HVAC control is
shown in Figure 2(a).
Reply Buffer
Minibatch Training
Deep Deterministic Policy
Gradient for HVAC Control (a)
The deep deterministic policy gra-
dient (DDPG) for HVAC control
is specially designed for solving DDPG RL Agent
problems with continuous vari- Actor: Outputs the Actions
Residential HVAC Soft
ables. In the described DQN con- System
System Behavior Update Target
trol, the neural network outputs
State Network Network
all of the possible action values,
and the number of action values
generated is limited. As a result, Maximize
Action
Q-Value
the algorithm processes discrete Setpoints Soft
actions. In the case of HVAC con- Update
Environment
trol, given the control action as the Behavior Target
Data Network Network
setpoint and an example range for
Collection
the setpoint from 20 to 22 °C, we Minimize Difference
need to discretize the proposed Critic: Evaluates the Actions
range. For instance, if we set the Reply Buffer
step size as 1 °C, then the poten- Minibatch Training
tial action set becomes {20 °C, (b)
21 °C, 22 °C}, and the algorithm
learns to choose from the three figure 2. A multizone HVAC control framework with (a) a DQN and (b) a DDPG.
actions. In contrast, the DDPG DDPG: deep deterministic policy gradient.

may/june 2022 ieee power & energy magazine 45

 addition, the DDPG algorithm adopts a stabilization strat- an RL algorithm interacts with the environment (i.e.,
egy by implementing the behavior and target networks for the building or house) and learns from it based on
both the actor and critic networks, which is similar to the the reward (i.e., operation cost) from the environment.
DQN algorithm. An overview of the DDPG approach for Depending on the application, an RL algorithm might
HVAC control is shown in Figure 2(b). need long experience to learn how to behave optimally.
✔✔ During the initial learning phase, an RL algorithm tends
Evaluation of a Deep RL-Based HVAC to take random actions to explore and understand the
Control Strategy Through Simulation environment. However, homeowners will not be happy
Our main objective is to evaluate the performance of the deep if RL designates random setpoints on their thermostats.
RL-based HVAC control strategy in a real-world environ- Therefore, instead of directly deploying an RL algo-
ment. However, one cannot deploy an RL algorithm directly rithm from scratch in a real-world environment, like a
in a real-world environment and allow it to learn from scratch building or house, we train and validate it in a simula-
for two main reasons: tion environment as a starting point, allowing for faster
✔✔ RL-based approaches are essentially trial-and-­ development and overcoming the challenge of training
error methods (albeit with high intelligence) in which RL in a real-world situation. Once we are satisfied with
its performance in the simulation, we can deploy the
0 pretrained RL model (the trained and deployable RL
Episodic Cumulative Reward

algorithm) in a real house.

–500 In this section, we introduce the training and validation
–1,000 of a deep RL-based multizone residential HVAC control
strategy on a simulation testbed with real-world data. Per-
–1,500 formance comparisons with benchmark control strategies
–2,000 demonstrate the efficiency and generalization ability of
model-free deep RL approaches.
–2,500

–3,000 Simulation Setup

0 20 40 60 80 The simulated HVAC building model requires weather-
Episode
related and price data. The weather data are taken from
(a)
typical meteorological year (TMY) data from 2019 to 2020
380 in Knoxville recorded by the National Renewable Energy
360 Laboratory. For price data, the time-of-use price signals with
340
320 23% Cost a peak price at US$0.25/kWh and an off-peak price at
300 Reduction US$0.05/kWh are applied. These input data sets are applied to
Cost

280 a building simulation software testbed for the following training

260
and validation of deep RL-based HVAC control strategies.
240
220
200 Training and Validation of
0 20 40 60 80
Episode
the DQN for HVAC Control
We first present the simulation result of the DQN algorithm
(b)
for multizone residential HVAC control. For training the algo-
7,000
rithm, the Knoxville TMY data from 21 December 2019 to 10
Minutes Outside Comfort

6,000 March 2020 were utilized. The simulation step of the HVAC
5,000 thermal dynamics is 1 min, and for every 5 min, the algorithm
4,000 provides a setpoint control action. The user c­ omfort level is
3,000 set to 20–22.22 °C (i.e., 68–72 °F). The state information used
2,000 in the DQN control includes all of the elements as listed in the
1,000 “Deep RL Approach for HVAC Control” section. The DQN
0 approach generates a discrete control action by adjusting the
0 20 40 60 80
Episode
setpoint with a fixed step within the user comfort level.
(c)
The DQN algorithm was trained for 75 episodes with
these settings. As mentioned earlier, we used approximately
figure 3. The training performance of the DQN algorithm three months of data from 21 December 2019 to 10 March
with the discrete control strategy: the (a) episodic cumula- 2020 for training the DQN algorithm. In this training pro-
tive reward, (b) cost of operation, and (c) minutes outside the cess, a single training episode is considered complete after
comfort level. the DQN algorithm has explored the three-month data. Next,

46 ieee power & energy magazine may/june 2022

we repeated the same process for several episodes until the stage. A fixed-setpoint control strategy is designed as a
DQN algorithm converged. Within each episode, at every baseline case for a comparison with the DQN-based HVAC
control step, the DQN algorithm observes the environment control strategy. In the baseline case, the heating setpoint is
through various features (called states) and then takes actions always kept at 20 °C.
that are implemented in the environment. For example, our For brevity, we present the indoor temperature variations
DQN algorithm provides a setpoint as the action, which for operating the pretrained DQN-based HVAC controller on
leads the environment to evolve into the next state, and then 14–15 December in Figure 4. An important observation here
the same procedure repeats. At the end of each control step, is that the DQN control has learned a preheating strategy.
the environment provides the reward in response to the DQN The DQN control preheated the zones before the peak price.
algorithm’s action. This reward could be the electricity cost This was beneficial during the initial hours of the peak period
incurred due to the DQN algorithm’s action. Based on this where the power consumption was zero, as shown by the
reward, the DQN algorithm then adjusts its Q-table. Readers power consumption graph in Figure 4. This is how the DQN
may refer to Figure 2 and the relevant description for more control achieved cost savings. The associated daywise energy
details on DQN algorithm training. cost comparison between the DQN control and baseline cases
It is important to evaluate the DQN algorithm’s training for 1–20 December is shown in Figure 5. We observed that
performance. Throughout the training, we keep track of the the DQN approach achieved a >32% cost savings over the
episodic average reward, total electricity cost of operating fixed-setpoint baseline for both the 1–20 December and 11–31
HVAC using the DQN for three months, and comfort viola- March data. Full details of the entire duration of 1–20 Decem-
tions that count the minutes during which the indoor tempera- ber and 11–31 March are not plotted due to space limits.
ture violated the user comfort level. The average reward, cost
of operation, and minutes outside of the user comfort level dur- Training and Validation of the
ing the training are shown in Figure 3. In Figure 3(a), the episodic DDPG for HVAC Control
cumulative reward gradually increases as the training proceeds For the DDPG algorithm, the Knoxville TMY data from
and stabilizes in the end. A 23% cost reduction is observed by 1–30 November 2019 are utilized for training. The c­ ontrol
the end of the training session in Figure 3(b). The minutes out interval of the DDPG algorithm is 60 min. The state
of comfort also show a decreasing trend and remain at zero by ­information input to the DDPG algorithm includes the cur-
the end of the training in Figure 3(c). These observations con- rent indoor temperature for each zone, outdoor temperature,
firm the convergence of the DQN-based HAVC control. and retail price as well as the lower bound of the user com-
The control performance of the pretrained DQN algorithm fort level (Table 1). The DDPG algorithm directly generates a
is further validated in unseen scenarios: two half-month deterministic, continuous setpoint for each zone of the build-
Knoxville TMY data sets from 1–20 December 2019 and ing’s HVAC system. The range of the setpoint is designated
11–31 March 2020, which were not used during the training to be the same as the user comfort temperature zone.

Outdoor Temperature RL Action (Zone 2)

Indoor Temperature (Zone 1) Electricity Price
Indoor Temperature (Zone 2) Power Consumption
RL Action (Zone 1)
Preheating

20
Temperature (°C)

15

10
Preheating
5

–5
19,000 19,500 20,000 20,500 21,000
Minutes
December 14 December 15

figure 4. The validation of indoor temperature variation for 14 and 15 December with the pretrained DQN model.

may/june 2022 ieee power & energy magazine 47

 We may conclude that the DDPG control can effectively solve
unseen physical environments and provide an efficient and flexible
HVAC control strategy after its offline training.

The DDPG algorithm is trained for 300 episodes. After benchmark control strategies: 1) a rule-based case, where the
the training, the pretrained DDPG algorithm is further vali- temperature is set at the lowest during the peak price hours
dated in two unseen scenarios: 1) with the Knoxville TMY and the highest during the off-peak price hours to achieve the
data from 1–20 January 2020 and 2) with the same weather preheating effect to reduce energy costs, and 2) a fixed-set-
data as the first scenario but with 10 building models that point case, where the setpoint is always at the highest value
have different thermal mass parameters from the simula- of the setpoint range to avoid violation of user’s comfort level.
tion testbed. The DDPG approach is compared with two In the first scenario, the final optimized results of the
DDPG algorithm and benchmark cases are shown in
Table 2 in which the total energy cost is the accumulated
Cost Comparison Daywise energy cost over the 10 test days. The average comfort
8 violation shows the average value, in degrees, by which
7 Baseline Cost
the indoor temperature is lower than the setpoint. Table 2
RL Cost
Daily Cost (US$)

6 shows that the rule-based case has the lowest total energy
5 cost because of its temperature setting logic based on
4 the peak or off-peak price. However, this control strat-
3 egy may result in a severe comfort violation because it
2 always designates the temperature setpoint to the lowest
1 value at peak price hours. In contrast, since the tempera-
0 ture is always set at the highest value in the fixed-setpoint
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

case, there is no temperature violation of the user’s com-

Day
fort level. Meanwhile, the energy cost is also the highest
Baseline RL among the three control strategies.
Since the temperature is always set at the highest value in
US$60.24 US$40.81 (32.25%↓)
the fixed setpoint case, there is no temperature violation of
user’s comfort level. Meanwhile, the energy cost is also the
figure 5. A daywise cost comparison of the DQN-based highest among the three control strategies. The setpoint set-
HVAC control with the fixed-setpoint baseline for 1–20
tings and the associated indoor temperature variations based
December 2019.
on the three approaches are illustrated in Figure 6. For each
approach, its control strategies for zones 1 and 2 share simi-
table 1. The daily user comfort level. lar patterns. Therefore, for the sake of simplicity, we plot
0:00– 6:00– 12:00– 18:00– only the control strategies of each approach in zone 1 in the
Time Period 6:00 12:00 18:00 24:00 figure and the control results in the first five days as repre-
User comfort level: 18 17 18 19 sentative values.
lower bound (°C) In all parts of Figure 6, the hour-by-hour yellow rect-
User comfort level: 20 19 20 21 angular bars represent the user comfort level as acceptable
upper bound (°C) temperature ranges, which correspond to Table 1. In Fig-
ure 6(a), it can be observed that the DDPG-based control
will designate the setpoint at a relatively low value during
table 2. A performance comparison of three the peak price hours and at a relatively high value during
HVAC control approaches.
the off-peak hours. As such, the DDPG-based control can
Rule Fixed achieve the preheating effect and reduce energy costs dur-
Control Approach DDPG Based Setpoint ing the winter.
Total energy cost (US$) 55.21 39.08 71.48 Figure 6(b) shows the results of the rule-based case, in
which the control strategy designates the setpoint at the
Minutes out of user comfort 48 2,617 0
level lowest value during peak price hours and the highest value
during off-peak hours. When the outdoor temperature is
Average comfort violation (°C) 0.13 1.85 0
extremely low, this control strategy results in severe indoor

48 ieee power & energy magazine may/june 2022

temperature violation (i.e., between
5,000 and 6,000 min in the figure),
25
mainly due to the low setpoint. In
contrast, the DDPG-based control 20
strategy does not place the setpoint
at the lowest possible value even 15
ac Status: On Tlower

Temperature (°C)
ac Status: Off
during peak price hours to avoid
comfort violations during the time 10
of extremely low outdoor tempera- Peak Price: US$0.25/kWh
5
tures. These comparisons show that,
once well trained, the DDPG control 0
has learned the impacts of the price Off-Peak Price: US$0.05/kWh
signal and outdoor temperature on –5
the reward, and it develops an intel-
ligent setpoint control strategy to –10
0 1,000 2,000 3,000 4,000 5,000 6,000 7,000
accommodate both the price peak Time (min)
and low outdoor temperature. (a)
Figure 6(c) shows the fixed-set-
point case, so-called because this 25
control strategy always sets the
20
temperature at the highest value.
Thus, the indoor temperature also 15 Temperature Violation
Temperature (°C)

remains at the highest level among

the three control strategies. Conse- 10
quently, this fixed-setpoint control
leads to the highest energy cost. 5
Note that, when calculating tem-
0
perature or comfort violations, only
the time when the indoor tempera- –5
ture is below the lower bound is
counted. The reason is that this –10
is a heating scenario, and a low 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000
Time (min)
indoor temperature is considered an
(b)
unbearable violation, while a high
indoor temperature is acceptable to 25
residential HVAC users.
In the second scenario, the 20
pretrained DDPG control strategy
15
is validated with 10 unseen build-
Temperature (°C)

ing models with different thermal 10

mass parameters to demonstrate
its generalization ability. Table 3 5
shows a comparison of the energy
costs and temperature violations 0
for the DDPG control and two
aforementioned benchmark con- –5
trols. As the table shows, like in
–10
scenario 1, the rule-based con- 0 1,000 2,000 3,000 4,000 5,000 6,000 7,000
trol gives the lowest energy cost, Time (min)
while the fixed-setpoint con- (c)
trol gives the fewest violations. Tin,1 Tout Setpoint1 Price ac Status
The pretrained DDPG control
achieves a balanced HVAC con-
trol strategy that results in a rela- figure 6. A comparison of the setpoint control strategies for the first five test days
tively lower energy cost and fewer (zone 1): (a) DDPG, (b) rule based, and (c) fixed setpoint.

may/june 2022 ieee power & energy magazine 49

 violations. Thus, we may conclude that the DDPG control Since the DQN-based RL approach achieves slightly
can effectively solve unseen physical environments and pro- higher energy savings (e.g., >32% energy cost savings) com-
vide an efficient and flexible HVAC control strategy after pared to the fixed-setpoint control, we selected the DQN-
its offline training. based RL algorithm as the HVAC control strategy to deploy
From these simulation studies, both the DQN and in the real house. The detailed deployment process and
DDPG controls demonstrate a better HVAC control per- experiment results are introduced in the next section.
formance compared with conventional approaches, like a
fixed-setpoint or simple rule-based control strategy, which Deployment of a Deep RL-Based
implies their considerable potential for online deployment. HVAC Control Strategy
As shown in Table 3, the DDPG-based RL algorithm To further validate the control performance of the deep RL
shows energy cost savings in the range of 19–29% (aver- approach in real-world scenarios, the pretrained DQN control
age = 25.9%) compared to the fixed-setpoint control in for controlling the multizone HVAC system was deployed at
the simulation. the Yarnell Station research house in Knoxville, Tennessee,
as shown in Figure 7. The house
is equipped with a two-stage heat
table 3. A comparison of results for different control strategies pump, two-zone control system,
DDPG Rule-Based Fixed Setpoint and two smart thermostats. The
two-story house is zoned by floor
Comfort Comfort Comfort
Building Violation Violation Cost Violation with a smart thermostat located
Index Cost (US$) (min) Cost (US$) (min) (US$) (min) centrally on each floor. Supply air
1 42.22 31 27.78 1,296 57.98 0
dampers are used to control the
delivery of conditioned air from
2 44.13 41 29.22 1,586 60.13 0 the heat pump to the appropriate
3 52.14 45 36.51 2,347 68.52 0 zone(s) based on the call for condi-
tioning from the thermostats.
4 59.66 101 43.94 3,364 75.61 0
The zone controller manages the
5 45.84 41 31.3 1,879 62.91 0 staging of the heat pump and indoor
6 42.49 39 27.68 1,398 59.06 0 airflow rate, which is adjusted based
on the number of zones calling for
7 37.47 24 23.51 1,012 53.42 0
conditioning and the “size” settings
8 61.21 81 45.42 3,520 76.44 0 of those zones (set using jumpers
9 35.34 25 21.98 818 49.9 0 on the control board during the
commissioning of the system).
10 43.19 59 28.41 1,323 58.46 0
The staging of the heat pump is
controlled based on a supply air
temperature sensor located in the
duct downstream of the air han-
dler. The staging is controlled to
maintain a heating mode supply
air temperature of at least 32.2 °C
(90 °F). The power consumption
of the heat pump, therefore, is de­­
pendent on the outdoor air tem-
(a) perature, indoor temperature, and
combination of zones calling for
conditioning. During some con-
dition combinations, it may only
be possible to elicit a single-stage
response from the two-stage heat
pump. With these many different
conditions affecting the achievable
(b) (c)
power response, the system is dif-
figure 7. The Yarnell Station research house in Knoxville, Tennessee, United States: the ficult to accurately model, making
(a) front view, (b) back view, and (c) data acquisition system. (Source: Oak Ridge National it a challenging real-world applica-
Laboratory, https://www.ornl.gov/content/smart-buildings; used with permission.) tion for the deep RL approach.

50 ieee power & energy magazine may/june 2022

 The DQN control strategy was
deployed at the research house 25 March 2021 to 05 April 2021, Simulation With RL Setpoints
from 25 March to 5 April 2021. 25

Temperature (°C)
Since this period often has mild
24
temperatures, the user comfort
level was shifted higher than 23
usual, i.e., 22.2–24.4 °C (72–76 °F), 22 Tin1_sim_DQN
to ensure the home would ex­­ Tin1_Tstat
21
perience adequate heating load.
Simulated Cost = US$20.9, Measured Cost = US$20.64
The fixed-setpoint baseline case 26

Temperature (°C)
is used to compare the DQN con-
trol strategy’s performance dur- 25
ing the deployment phase. In the 24
baseline case, the setpoint is fix­­ 23
Tin2_sim_DQN
ed to 22.2 °C (72 °F) to minimize Tin2_Tstat
22
the energy cost while maintaining
the user comfort level. However, Simulated Energy = 212,915 Wh, Measured Energy = 218,488
it is not feasible to deploy the Power_sim_DQN
6,000
fixed-setpoint control under the Power_measured
Power (W)

same weather conditions in the 4,000

same research house as the DQN
deployment, so it is not straight- 2,000
forward to have a truly fair com-
0
parison between the DQN and
25 March 27 March 29 March 31 March 1 April 3 April 5 April
baseline controls. Time
Instead, we can create a “sim- (a)
ulated DQN” case using the re­­ 25 March 2021 to 05 April 2021, Baseline Simulation
corded thermostat setpoints and 25
Temperature (°C)

measured weather data during the

DQN deployment. This allows 24
Tin1_sim_BL
us to evaluate the accuracy of the 23 Tin1_Tstat
DQN simulation by comparing
22
the measured to simulated results.
More importantly, this makes a 21
more direct comparison between Baseline Simulated Cost = US$23.67, RL Measured Cost = US$20.64
the “simulated DQN” case and the 26 Tin2_sim_BL
Temperature (°C)

baseline (fixed-setpoint) simulation 25 Tin2_Tstat

case. The model parameters were
24
adjusted based on the comparison
between the measured data from 23
the DQN deployment and “simu- 22
lated DQN” control to increase the Baseline Simulated Energy = 188,389 Wh, RL Measured Energy = 218,488
accuracy of the simulation. In the
Power_sim_BL
following discussion, we present 6,000
Power_measured
Power (W)

the details of this procedure and

4,000
performance comparison.
The procedure of the compari- 2,000
son study, which is based on the
DQN deployment, is as follows: 0
✔ ✔ Fi r st , we f i n e -t u n e t h e 25 March 27 March 29 March 31 March 1 April 3 April 5 April
Time
building simulation model (b)
used for DQN training with
the new data collected dur- figure 8. A comparison of the simulation and deployment results: (a) a simulated
ing the DQN deployment DQN case versus measured data for DQN control and (b) a simulated baseline case
period. versus measured data for DQN control. BL: baseline.

may/june 2022 ieee power & energy magazine 51

 With these many different conditions affecting the achievable
power response, the system is difficult to accurately model, making
it a challenging real-world application for the deep RL approach.

✔✔ Next, we re-create the DQN’s deployment (in simu- Tin1_sim_DQN curve (in blue) shows the temperatures of
lation) by running the DQN’s setpoints through this the “simulated DQN” case of the first floor in the research
building simulation for the duration of 1 March–5 house, while Tin1_sim_BL (in blue) in Figure 8(b) shows
April 2021. This is referred to as the “simulated DQN” the temperatures of the simulated baseline case (i.e., fixed-
case, stemming from the actual DQN deployment. setpoint control) on the first floor. Note that, in both top
✔✔ Then, we use this fine-tuned building simulation for plots of Figure 8(a) and (b), the red Tin1_Tstat curve, which
the baseline case by using a fixed setpoint of 22.2 °C represents the temperatures measured by thermostats dur-
for the same duration. This later run mimics the de- ing the DQN deployment, is plotted as a reference to signify
ployment of the fixed-setpoint baseline control in the the difference between the “simulated DQN” and baseline
research house during the deployment period. cases. The “simulated DQN” case closely resembles the
It should be noted that, since we do not have any mea- DQN deployment because the two curves in the top plot of
sured data for the baseline case to use for setting the initial Figure 8(a) are very close. Thus, the comparison between the
temperatures in the simulation, we instead simulate an addi- “simulated DQN” and baseline is highly credible. It is also
tional 3.5 weeks (i.e., 1–24 March 2021) of baseline opera- evident that the fixed-setpoint control gives quite different
tion before the period of interest. Simulating this additional results from the DQN deployment.
time allows the modeled temperatures to stabilize to realistic The mid plots in Figure 8(a) and (b) represent the second
values and minimizes the effect of any error associated with floor of the research house and show similar patterns to the
the selected initial temperatures. first floor. The bottom plots show the simulated and mea-
The results from the “simulated DQN” case using the sured power use of the HVAC system. We may observe in
retrained model and baseline simulation case employ- the bottom plot of Figure 8(a) that the measured power usage
ing the fixed-setpoint control are shown in Figure 8(a) during the DQN deployment (in red) matches well with the
and (b), respectively. In the top plot of Figure 8(a), the power usage in the simulated DQN case (in blue). This also

table 4. The daily electricity cost and energy use comparison between DQN/RL and the fixed-setpoint baseline.
Baseline DQN/RL Cost Reduction
DQN/RL DQN/RL
Simulated Simulated Measured Simulated Measured
Simulated Energy Simulated Measured Energy Energy Versus Baseline Versus Baseline
Date Cost (US$) (Wh) Cost (US$) Cost (US$) (Wh) (Wh) Simulated (%) Simulated (%)
25 March 1.19 10,475 1.18 1.47 16,338 19,602 0.8 –23.5
26 March 0.84 5,005 0.64 0.73 5,670 6,824 23.8 13.1
27 March 1.63 13,886 1.29 1.35 18,094 18,892 20.9 17.2
28 March 1.06 9,766 0.97 1.17 12,874 15,650 8.5 –10.4
29 March 2.69 19,316 2.23 2.09 21,025 21,506 17.1 22.3
30 March 2.18 14,412 1.59 1.57 14,311 14,871 27.1 28
31 March 1.11 13,453 1.08 1.11 15,884 16,089 2.7 0
1 April 3.38 29,741 3.02 2.82 31,657 29,552 10.7 16.6
2 April 3.79 31,228 3.7 3.47 33,272 31,867 2.4 8.4
3 April 3.31 24,237 3.31 3.04 26,208 25,751 0 8.2
4 April 2.48 16,872 1.87 1.82 17,582 17,884 24.6 26.6
All 23.67 188,389 20.9 20.64 212,915 218,488 11.7 12.8

52 ieee power & energy magazine may/june 2022

 The results signify that the deep RL approach is of considerable
potential for online applications in solving complex control and
optimization problems like residential demand management.

verifies the accuracy of the fine-tuned building simulation approach is of considerable potential for online applications
model. In addition, the bottom plot of Figure 8(b) shows in solving complex control and optimization problems like
the power consumption from the baseline case (in blue) and residential demand management.
measured power (in red); the difference between the base- For future directions, an interesting topic would be to
line power consumption and measured power in Figure 8(b) further investigate the generalization ability of deep RL
is greater than that in Figure 8(a), which demonstrates the approaches, for example, to make the control workable for
considerable difference between the fixed-setpoint baseline various scenarios, including both cooling and heating as
case and “simulated DQN” case. well as idle scenarios, without additional retraining efforts.
Table 4 shows a breakdown of the daily energy cost and Another promising direction would be physics-informed
consumption comparisons for the DQN-based RL control deep RL, which would introduce physical laws to guide the
approach and fixed-setpoint baseline case. The daily cost exploration of the approach and further improve learning
savings from the DQN control (either simulated or deployed) efficiency. Further, similar approaches based on machine
range from 0 to 28% (except for a couple of outlier days with learning may be extended from HVAC control to other
negative cost savings), depending on the day. While there are building energy controls.
some fluctuations in the day-to-day comparison of energy
use as well as the cost of the simulated and measured DQN For Further Reading
cases, the overall energy use and cost over the 11-day deploy- V. Mnih et al., “Human-level control through deep rein-
ment have a difference of less than 3%. This indicates that the forcement learning,” Nature, vol. 518, no. 7540, pp. 529–533,
building simulation is well calibrated. The significant obser- 2015, doi: 10.1038/nature14236.
vation from Table 4 is that the DQN cases (either simulated T. P. Lillicrap et al., “Continuous control with deep rein-
or deployed) consumed more energy than the baseline while forcement learning,” 2015, arXiv:1509.02971.
still managing to reduce the total cost by 11.7% and 12.8%, Y. Du et al., “Intelligent multi-zone residential HVAC
respectively, in comparison with the baseline case. This is control strategy based on deep reinforcement learning,”
because the DQN control preheats the home to a higher tem- Appl. Energy, vol. 281, p. 116,117, Jan. 2021, doi: 10.1016/j.
perature during low-price periods such that the total cost is apenergy.2020.116117.
decreased, while more energy is consumed, and the comfort Y. Du et al., “Multi-task deep reinforcement learning
level is better. for intelligent multi-zone residential HVAC control,” Elec-
Note that the minor difference between the simulated tric Power Syst. Res., vol. 192, p. 106,959, Mar. 2021, doi:
DQN case and DQN deployment is likely due to a combi- 10.1016/j.epsr.2020.106959.
nation of factors, including small inaccuracies associated K. Kurte et al., “Evaluating the adaptability of reinforce-
with the building model compared to the real-world building ment learning based HVAC control for residential houses,”
response and difference in HVAC response over the decision Sustainability, vol. 12, no. 18, p. 7727, Sep. 2020, doi:
interval of 5 min. Future performance improvements dur- 10.3390/su12187727.
ing deployment could be achieved by decreasing the deci-
sion interval of the DQN control to a shorter interval and the Biographies
inclusion of online learning to fine-tune the decisions of the Yan Du is with the University of Tennessee, Knoxville, Ten-
DQN control over a longer operational period. nessee, 37996, USA.
Fangxing Li is with the University of Tennessee, Knox-
Summary ville, Tennessee, 37996, USA.
This article explores the application of deep RL approaches Kuldeep Kurte is with Oak Ridge National Laboratory,
to implement energy management in a multizone residential Oak Ridge, Tennessee, 37830, USA.
HVAC system to minimize energy costs and maintain user Jeffrey Munk is with Oak Ridge National Laboratory,
comfort. Both simulation and real-world deployment results Oak Ridge, Tennessee, 37830, USA.
demonstrate that the deep RL approach can learn an HVAC Helia Zandi is with Oak Ridge National Laboratory, Oak
control strategy that is more economical, generalized, and Ridge, Tennessee, 37830, USA.
adaptive than either the rule-based or simple fixed-set-
point control strategy. The results signify that the deep RL  p&e

may/june 2022 ieee power & energy magazine 53

 Clustering
Electricity
Consumers
Challenges and Applications
for Operating Smart Grids

©SHUTTERSTOCK.COM/SOLARSEVEN

By Andrés M. Alonso, Eduardo Martín, Alicia Mateo,

Francisco J. Nogales, Carlos Ruiz, and Andrea Veiga
Digital Object Identifier 10.1109/MPE.2022.3150826
Date of current version: 19 April 2022

54 ieee power & energy magazine 1540-7977/22©2022IEEE may/june 2022

W
WITH THE PROGRESSIVE ADVANCEMENT OF THE to 266 million (a 92% penetration rate) by 2030. In particu-
smart grid paradigm, electricity systems worldwide are lar, Spain, one of the European leaders, achieved a penetra-
deploying advanced metering technologies in residential tion rate above 99% in 2020.
sectors. Smart meters can monitor and process the electric- In the past, aggregated household consumption data
ity use of households at very high sampling rates. The avail- were collected every one or two months. The installation of
ability of hundreds of thousands of time-series load profiles smart meters allows utilities to monitor, record, and trans-
opens the possibility of applying clustering methods for mit household consumption data at a high resolution (typi-
grouping consumers with similar consumption patterns. cally from every minute to every hour), leading to a signifi-
In this article, we describe the main characteristics of cant increase in data volume. This technology shift brings
load profiles and explore several clustering methods for cus- new and important technical and analytical challenges to
tomer classification and how they can improve the operation process, store, and interpret massive electricity time series.
of electricity systems. We also review the most widely used Thanks to other advanced metering infrastructure technolo-
clustering techniques for smart meter data and present some gies, the smart meter data can now be combined with the
recent advances in this field. Specifically, we illustrate how other relevant variables (e.g., weather, mobility, presence,
smart meter clustering techniques need to be fast and effi- demographics, economics, and so on) to enrich applications
cient, robust against outliers and missing data, and scalable of advanced data analytics in power grid operation.
to hundreds of thousands of time series. Indeed, there is a growing interest in how these new data
sets can be exploited from an analytical point of view—to
Smart Grids Revolutionize the gain knowledge by extracting valuable information. On one
Energy Analytics Paradigm hand, market/system operators seek to improve system oper-
Power systems are being upgraded worldwide as part of ation reliability and efficiency to reduce CO2 emissions and
a transition toward climate-neutral systems. The trend lower market prices. On the other hand, retailing/generation
is to depart from traditional centralized and carbon-based utilities are interested in staying competitive while increasing
energy systems to affordable, renewable, and reliable their profits and customer satisfaction. The electricity indus-
frameworks. In this regard, several initiatives are being try now has the opportunity to update and improve its tradi-
considered that affect both the regulatory and technical tional analytical techniques, which until recently were mostly
operations of these systems. based on small data sets and basic descriptive and predictive
One of the main drivers of this transition is the need for approaches. In this regard, prescriptive analytics methods
a full digitalization of the electricity supply chain. In par- may need to be implemented to aid in the decision-making
ticular, the integration of modern information and telecom- process and cope with complex real-time and massive data
munication technologies plays a key role in securing the grid inputs from the smart grid.
operation while fostering an active role of electricity con- This new opportunity is being tackled from the perspec-
sumers. This is aligned with the concept of a “smart grid,” tive of big data analytics. Not only is data volume an issue, but
an electrical network that can integrate the actions of all of also its variety (structured and unstructured data, e.g., energy
its users to achieve a sustainable, secure, safe, and economi- use and weather time series, digital images, sensor data, social
cally efficient power system. media, voice recordings, and so on) and the need for processing
These technologies open new business and technical pos- velocity present challenges. Recent developments in machine
sibilities (and big challenges, too) in the operation of power learning and artificial intelligence may foster data mining and
systems. For instance, we can highlight new concepts like dis- feature selection processes for designing accurate and efficient
tributed energy resources, prosumers, demand-side manage- prescriptive decision-making tools.
ment, peer-to-peer energy trading, virtual power plants, electric
vehicles, and home batteries—all of which need to be deeply Main Characteristics of Individual
studied and capitalized by energy utilities and system operators. Household Time Series
More specifically, the locally managed electricity systems In this section, we illustrate the main characteristics of a real-
known as microgrids are probably undertaking the major world smart meter data set. Endesa, one of the main electric
technical transformations in power systems. The installation utilities in Spain, has provided a sample of 10,000 household
of new power electronics that allow bidirectional power flows, consumptions on an hourly basis for one year. For each house-
together with advanced control and protection technologies, hold, a time series containing 8,760 electricity energy con-
is greatly improving the microgrid operation and reliability. sumptions intervals is provided. Specifically, these households
One of the main game changers in the efficient operation have power consumptions below 10 kW, and they are distrib-
of microgrids is the deployment of advanced metering infra- uted throughout the geography of Spain. The data were anony-
structures. For example, around 107 million smart meters mized in compliance with all data protection regulations.
are estimated to be currently installed in the United States Figure 1 shows the consumption of three clients over
(more than 60% penetration rate) and 123 million in Europe the year. They have been selected to illustrate their differ-
(a 43% penetration rate), which plans to increase this number ent consumption behavior. Figure 1(a) represents the average

may/june 2022 ieee power & energy magazine 55

 client in our sample. The mean consumption throughout the in a year for a household (x-axis) and its corresponding stan-
year in the household sample is around 330 Wh, while the dard deviation (y-axis). We have highlighted in red the three
corresponding standard deviation is around 400 Wh, and customers selected for illustrative purposes.
these are roughly the same numbers for the first selected cli- There is a clear relation between the hourly average con-
ent. In some periods [Figure 1(a)], the consumption drops sumptions and their corresponding dispersion. On average,
while in others the consumption is clearly above the mean. the household consumptions are around 300 Wh with a stan-
On the other hand, the second client has a mean consump- dard deviation of around 400 Wh. That means the dispersion
tion of around 1,000 Wh and a standard deviation of around is higher than the mean consumption level. Note also that
1,000 Wh [Figure 1(b)]. This client has a consumption pat- the distribution of the hourly average consumption and the
tern more volatile than that of the first one. Finally, the third standard deviation is highly asymmetric. For instance, 75%
client [Figure 1(c)] has the highest consumption in our sam- of the clients have an hourly average consumption of less
ple. This client has a different consumption trend (more in than 440 Wh and a standard deviation of less than 525 Wh.
the summer and less during the winter). For these reasons, Finally, there are several households with an almost-constant
an appropriate clustering tool should be able to classify these and close-to-zero consumption along the year (standard devia-
customers into different groups. tion smaller than 1 Wh).
To illustrate the behavior pattern for all of the household The individual time series contains the consumption
consumptions in the sample, there are several visualization behavior of the clients across time, so we can infer many
alternatives. Figure 1(d) summarizes each time series by a interesting aspects, such as static quantile hourly profiles
point representing the hourly average consumption (in Wh) (i.e., how consumption is statistically distributed within the

Consumption, Customer Number: 1,768 Consumption, Customer Number: 3,143

6,000
Electricity Consumption (Wh)

Electricity Consumption (Wh)

4,000

3,000
4,000

2,000
2,000
1,000

0 0
Apr. 2018 July 2018 Oct. 2018 Jan. 2019 Apr. 2019 Apr. 2018 July 2018 Oct. 2018 Jan. 2019 Apr. 2019
(a) (b)

Consumption, Customer Number: 2,738

Customer 2,738
4,000
Electricity Consumption (Wh)

15,000
Standard Deviation (Wh)

3,000

10,000
2,000
Customer 3,143

5,000
1,000

0 0 Customer 1,768
Apr. 2018 July 2018 Oct. 2018 Jan. 2019 Apr. 2019 0 2,000 4,000
Average (Wh)
(c) (d)

figure 1. (a)–(c) Hourly consumption throughout a year for three clients selected for illustrative purposes. (d) Scatterplot
of the average consumption for each household versus the standard deviation for 10,000 households.

56 ieee power & energy magazine may/june 2022

 A partitioning method divides the observations into
a number, k, of nonoverlapping groups such that each
group contains at least one observation.

24 h of a day), peak and valley hours, and autodependen- In a model-based approach, a time-series model is selected
cies (how consumption in the previous hours is related to the for each series, and the difference (in model parameter) is
consumption at the current hour). These static and dynamic computed between each pair of models. A time-series model
features can be used to characterize, analyze, and forecast is a mathematical formulation that relates the value at time
the main consumers’ consumption patterns and to group t with the values at the previous moments, usually involving
those households with similar profiles. a set of parameters that must be estimated. Finally, depen-
dence-based clustering considers the cross dependency among
Clustering Households Time Series the series (cross correlations) and is based on computing a
dissimilarity measure.
Clustering Methods for Time-Series Data In all of these approaches, a dissimilarity (distance)
Clustering is one of the most important tools in unsuper- function is needed. There are convenient dissimilarity func-
vised machine learning for grouping unlabeled data sets. tions, for instance, Euclidean and weighted Euclidean,
It is a mathematical process to find groups of observations Fréchet, or Mahalanobis distances using the extracted fea-
(in our case, end-use load profiles) in such a way that the tures. Figure 2 illustrates the calculation of the Euclidean
observations within the same cluster have similar character- and Fréchet distances. The Fréchet distance is the mini-
istics, while the observations in different clusters vary from mum–maximum distance between the two sets of points
each other. In that way, we expect households in a group to maintaining an increasing order, while the Euclidean dis-
have a similar consumption pattern but a very different pat- tance is the sum of the squared distances maintaining a
tern from households in other groups. The similarity among strict order that is point to point (the dashed lines in the
households involves measures of distance between the attri- graph). The weighted Euclidean and Mahalanobis distances
butes or characteristic values that describe the time series of give different weights to the point-to-point distances or dif-
electrical consumption. ferences, respectively. They are especially useful for giv-
There are two main clustering tools: partitioning meth- ing more importance to some of the features or for taking
ods and hierarchical methods. A partitioning method divides into account the variability and/or covariability between
the observations into a number, k, of nonoverlapping groups the extracted features. Those extracted features can be, for
such that each group contains at least one observation. A instance, percentile profiles, autocorrelations, periodo-
hierarchical method starts with singleton clusters (clusters withgram values, estimated model parameters, and so on, as
raw vectors.
a single observation) and joins the clusters at each step in a hier-
archical way.
For time-series data, there are
three approaches to build the clus-
ters: features based, model based,
and dependence based. The features
can be statistics that summarize the
raw data or the raw data themselves Fréchet
(in our case, 8,760 features that Customer A
Distance
Features
correspond to the hourly consump-
tion of a household). The following
summary features could be used in
place of the original 8,760 raw ones:
Customer B
✔✔ mean or median (total or by the Features
hour, day, month, and so on)
✔✔ variance or some represen-
tative percentiles figure 2. An illustration of Euclidean and Fréchet distance assuming five features.
✔✔ autocorrelation function and The Fréchet distance (double arrow) is the min–max distance between the two sets
partial autocorrelation func- of points maintaining an increasing order. The Euclidean distance is the sum of the
tion (for a given set of lags). squared distances (dashed lines).

may/june 2022 ieee power & energy magazine 57

 Exploring Household Clusters: hours. This is valuable information that can be used to
An Illustrative Proof of Concept design appropriate retail tariffs.
Next, we propose and analyze an effective clustering pro- When large numbers of electricity consumers are
cedure to segment household consumptions. The procedure considered, the following practical issues must be taken
is based on two consecutive clustering tools. The first tool into account:
is applied to obtain major divisions among consumers and ✔✔ Data normalization: The hourly consumption for each
capture their static behavior. The second tool is applied to household must be normalized to better compare the
each of the previous divisions to capture specific features consumption levels between households. In particular,
describing dynamic behavior. for the normalization, we have divided the hourly con-
In particular, the first division is intended to group con- sumption for each client by its maximum consumption
sumers with similar consumption levels and volatility. It can throughout the year.
be used to differentiate among large, medium, and small ✔✔ Seasonal impacts: Household consumption is sea-
(households) consumers. The second division is designed sonal. To make the time series more homogeneous,
to identify groups with similar consumption profiles within the clustering tools should be trained using data for a

Cluster Centers for Each Group

Percentage of Normalized Consumption

60 Group Group 4
(Respect to the Maximum)

1 2 3 4
50

40

30 Group 3

20 Group 2
10
Group 1
0
10 20 30 40 50 60 70 80 90
Deciles (%)
(a)

Hourly Consumption Profile by Cluster With 80% Confidence Intervals

Clustering by Consumption Deciles
Group 1: 3,525 Clients
80
60
40
20
0
Normalized Consumption (%)

Group 2: 4,412 Clients

80
60
40
20
0
Group 3: 1,596 Clients
80
60
40
20
0
Group 4: 223 Clients
80
60
40
20
0
–1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
hour
(b)

figure 3. Major divisions: (a) centers and (b) hourly consumption profile.

58 ieee power & energy magazine may/june 2022

 given season (instead of a complete year). In our case, divided again based on their dynamic behavior. Group 4 will be
we have selected the spring season with 2,184 hours. excluded from the analysis as it contains only 233 households.
✔✔ Scalability: An efficient algorithm should scale well
for large data sets. Partitioning methods and, in par- Clustering for Finer Divisions
ticular, k-means clustering are usually good choices in To capture the dynamic behavior of electricity consumers,
terms of computational time. we can consider the following three approaches:
✔✔ Number of clusters: In partitioning methods, a critical deci- ✔✔ An approach based on median hourly profiles for
sion is the number of groups. This number should be deter- each consumer: For each household, the 24-h medi-
mined by combining 1) knowledge from domain experts: ans (features) are computed, and a k-means algorithm
a number large enough to properly separate consumer be- is applied to each macrogroup using these 24 features.
haviors, but small enough to include it in possible actions as In this case, the number of subgroups is three.
an additional feature (e.g., marketing or tariff campaigns); ✔ An approach based on the 95th-percentile hourly pro-
and 2) technical criteria: based on a specialized statistical files for each consumer: This method is like the preced-
analysis. To illustrate our case study, we have considered ing approach, but it uses 95th percentiles instead of me-
k = 10 groups to segment 10,000 households. dians. This approach may be useful when the focus is
✔✔ Outliers: Note that there may be several consumers with on the peak demand.
no (or very low) consumption. These consumers should ✔ An approach based on autoconsumption: For each house-
not be considered by a clustering tool but, instead, be hold, the autocorrelation function of the corresponding
included in a particular and extra group. In our case, time series is computed to capture individual dynamic
there are 244 households with no consumption; they are behavior. In this case, we have considered the first 168
assigned to group 0. autocorrelations, but to reduce the dimensionality, we
have selected the 19 most relevant ones: 1, 2, 3, 4, 5, 6, 8,
Clustering for Major Divisions 9, 10, 11, 12, 13, 24, 25, 48, 49, 72, 73, and 168. Note that
The clustering for major divisions is based on the static behav- they capture dynamics in the very short term but also in
ior of the households to capture global properties given by the the first, second, third, and seventh previous days.
nine consumption deciles in a season (spring in our case). Hence, Figure 4 shows the cluster centers for three finer divisions
the input for the k-means tool is a matrix containing 9,756 house- of each macrogroup computed in the previous subsection. The
holds (10,000 minus 244 with no consumption) and nine deciles. finer division is based on the first approach (median hourly
Figure 3(a) shows the cluster centers for each group. In profiles). The sizes for each minor division of the first mac-
particular, the y-axis contains, for each of the four groups, rogroup are 1,613; 1,066; and 846 households. For the second
the normalized consumption for all hourly consumptions macrogroup, the sizes are 1,938; 1,891; and 583. For the third
in the corresponding group and a given decile probability macrogroup, the sizes are 998, 354, and 244.
(x-axis). The clustering divides the clients into groups based In Figure 4 (macrogroup 1), note that the division of the
on total seasonal electricity consumption computed by the first macrogroup can isolate the third minor group: consum-
nine deciles. There is a large group (group 1) with small and ers who increase their loads during the evening. For macro-
occasionally moderate consumption. Another small group group 2, the clustering can identify in the third minor group
(group 4) has high consumption and is occasionally close to those consumers with a higher consumption during morn-
the maximum. The sizes for macrogroups 1, 2, 3, and 4 are ings, while the other two minor groups consume more in the
3,525, 4,412, 1,596, and 223 households, respectively. evenings. Finally, for macrogroup 3, the clustering can iden-
Figure 3(b) shows the hourly consumption profile for tify in the third minor group those consumers with a higher
households in each group for 24 h. The black line repre- consumption load in the morning. Those in the first and sec-
sents the average consumption for all households in a given ond minor groups prefer to consume during evenings.
group for each hour. The corresponding 80% confidence On the other hand, Figure 5 shows the cluster centers for
intervals are shown in shaded colors. Note that the higher the three finer divisions of each of the macrogroups comput-
the consumption, the higher the volatility. Moreover, group 4 ed in the previous section. The finer division is based on the
has a high peak demand for all hours in a day. For the other second approach: 95th-percentile hourly profiles. The sizes
groups, peak demand is concentrated on the usual peak for each minor division of the first macrogroup are 946,
hours but with different magnitudes. 1,525, and 1,054 households. For the second macrogroup,
In particular, groups 1, 2, and 3 capture those consum- the sizes are 1,699, 1,267, and 1,446. Finally, for the third
ers with two daily consumption peaks: morning and evening. macrogroup, the sizes are 691, 612, and 293.
Group 2 presents an afternoon peak at around 2 p.m., while In Figure 5 (macrogroup 1), note that, for the first mac-
group 3’s is around 11 a.m. The consumption pattern is more rogroup, the clustering can identify in the second and third
stable for group 4. In summary, previous groups characterize minor groups consumers with two daily consumption peaks:
customers based on their total amount of electricity consumed morning and evening. In the first minor group, the consump-
given by their percentiles. Next, each of the three groups will be tion is more constant with a small peak in the mornings.

may/june 2022 ieee power & energy magazine 59

 For the other two macrogroups,
the clustering identifies two con-
Subgroups With 1,613, 1,066, and 846 Households, Respectively sumption peaks and two valleys,
Consumption (Respect to the Maximum)

Subgroup
but at different times and with dif-
40
1 2 3 ferent sizes.
Percentage of Normalized

Finally, Figure 6 shows the

30 cluster centers for the three finer
divisions of each metagroup based
on the third approach: autocorre-
20
lations. In particular, for a given
h-lagged hour, each vertical line
10 represents the hth autocorrelation
for the hourly consumption in
each group, that is, the correlation
0
between the consumption at hour t
–1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
in a group and the same consump-
hour
tion but at hour t – h.
(a)
The sizes for each minor divi-
Subgroups With 1,938, 1,891, and 583 Households, Respectively sion of the first macrogroup are
1,373, 732, and 1,420 households.
Consumption (Respect to the Maximum)

40 Subgroup
For the second macrogroup, the
1 2 3
sizes are 2,749, 1,145, and 518. The
Percentage of Normalized

30 third macrogroup’s sizes are 992,

451, and 153.
The autocorrelation clustering
20 provides information about the dif-
ferent dynamics of the time series,
10 not only in the short term (lags
1–13), but also on a daily (lags 24,
25, 48, 49, 72, and 73) and weekly
0 basis (lag 168). Although the me-
–1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

dian and the 95th-percentile hourly

hour profiles take into account the dy-
(b)
namics, these profiles do not mea-
Subgroups With 998, 354, and 244 Households, Respectively sure the strength of the relationship
Consumption (Respect to the Maximum)

between the consumption of con-

40
tiguous hours, which is some-
thing that we can measure using
Percentage of Normalized

30 autocorrelation functions. In this

framework, autocorrelations allows
us to separate clients with highly
20 persistent dependence in the short
term from those with smaller de-
10 pendencies. The second group can
Subgroup be divided further into two groups:
1 2 3 those with small but positive cor-
0 relations and those with small and
–1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

negative values at lags 8–13. This

hour reveals a clear difference in con-
(c)
sumption behavior.
In Figure 6 (macrogroup 1),
figure 4. Subgroups in metagroups 1, 2, and 3 clustered by median (50th the first and second minor groups
­p ercentile). (a) Cluster centers for each division of metagroup 1. (b) Cluster in the clustering identify those
centers for each division of metagroup 2. (c) Cluster centers for each division consumers with a steep consump-
of metagroup 3. tion peak in the morning (10 a.m.)

60 ieee power & energy magazine may/june 2022

and a more sustained increment of
consumption during the evening
(8 p.m.–11 p.m.), although lower Subgroups With 946, 1,525, and 1,054 Households, Respectively

Consumption (Respect to the Maximum)

than in the morning. However, the Subgroup
highest consumption for the first 1 2 3

Percentage of Normalized
division during the evening is at 60
8 p.m., while the second division
peaks at 11 p.m. The third minor
group postpones the morning peak 40
to 3 p.m. and presents a second
peak (higher than the morning
one) during the evening at 10 p.m. 20
For macrogroup 2, the third
minor division captures consum-
0
ers with a morning peak a bit ear- –1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
lier and a more sustained evening
hour
peak. The difference between the
(a)
first and third divisions is that, in
the first, the morning peak and Subgroups With 1,699, 1,267, and 1,446 Households, Respectively
the afternoon consumption valley
Consumption (Respect to the Maximum)

Subgroup
are slightly lower.
1 2 3
For macrogroup 3, the third
Percentage of Normalized

60
minor division presents a more
stable consumption pattern (the
night and afternoon consumption 40
valleys are not so significant). For
the other two minor groups, house-
holds consume more during morn- 20
ings and evenings, respectively.
These minor divisions differ in the
hours for which the morning and 0
–1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
afternoon peaks occur.
hour
(b)
Potential Applications
The clustering methods can be Subgroups With 691, 612, and 293 Households, Respectively
Consumption (Respect to the Maximum)

used in several applications within

the activities of an electricity re­­
tailer. First, the resulting groups
Percentage of Normalized

can be used for the development 60

of personalized products (e.g.,
electricity tariffs) in each cluster
of clients, allowing savings to the 40
final customer and increasing the
competitiveness of the retailer. 20
Moreover, this is closely related Subgroup
to the successful implementation 1 2 3
of demand-response programs. 0
–1
0
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

Clustering methods can help charac-

hour
terize and quantify those groups of
(c)
consumers who have higher price
elasticity and are prone to dyna­
mically adapt their consumption. figure 5. Subgroups in metagroups 1, 2, and 3 clustered by 95th percentile.
This can help modify the consump- (a) Cluster centers for each division of metagroup 1. (b) Cluster centers for
tion profile of a given area, region, each division of metagroup 2. (c) Cluster centers for each division of meta-
or country to avoid consumption group 3.

may/june 2022 ieee power & energy magazine 61

62
Subgroups With 1,373, 732, and 1,420 Households, Respectively Subgroups With 2,749, 1,145, and 518 Households, Respectively

Subgroup: 1 Subgroup: 1
0.6 0.6
0.4 0.4
0.2 0.2
0 0
Subgroup: 2 Subgroup: 2
0.6 0.6

ieee power & energy magazine

0.4 0.4
0.2 0.2
0 0

Auto-Correlations

Auto-Correlations
Subgroup: 3 Subgroup: 3
0.6 0.6
0.4 0.4
0.2 0.2
0 0
Lagged Hour Lagged Hour
Subgroup 1 2 3 Subgroup 1 2 3
(a) (b)

Subgroups With 992, 451, and 153 Households, Respectively

Subgroup: 1
0.6
0.4
0.2
0
Subgroup: 2
0.6
0.4
0.2
0

Auto-Correlations
Subgroup: 3
0.6
0.4
0.2
0
Lagged Hour
Subgroup 1 2 3
(c)

figure 6. Subgroups in metagroups 1, 2, and 3 clustered by autocorrelation. (a) Cluster centers for each division of metagroup 1. (b) Cluster centers for each division of meta-

may/june 2022
group 2. (c) Cluster centers for each division of metagroup 3.
 Tailor-made forecasting models can now be designed
and trained to characterize the different dynamic
behaviors of each type of customer.

peaks or to force them to follow the availability of renewable in this field are expected to combine traditional statistical tech-
resources to improve the system efficiency. niques with machine learning approaches for capturing both
In the same vein, this information is useful to perform ex- static and dynamic electricity consumption patterns.
post analysis and customer diagnosis. An electrical utility, or a
consumer itself, can compare its historical use of energy with its Acknowledgments
cluster peers. By applying different statistical techniques, we can The authors gratefully acknowledge the support from the
identify, with a prespecified confidence level, possible anom- Spanish government and the Madrid government (Comuni-
alies derived from defective appliances or even detect fraud. dad de Madrid).
Clustering tools can also help to design more reliable electri-
cal systems. The progressive adoption of new technologies, For Further Reading
such as household solar panels, electric vehicles, home bat- S. Aghabozorgi, A. S. Shirkhorshidi, and T. Wah, “Time-
teries, and so on, has a direct impact on customers’ consump- series clustering: A decade review,” Inf. Syst., vol. 53, pp.
tion patterns that can be unveiled and characterized by clus- 16–38, Oct./Nov. 2015, doi: 10.1016/j.is.2015.04.007.
tering tools. This information can be extrapolated into the A. M. Alonso, F. J. Nogales, and C. Ruiz, “Hierarchical
future, under different integration level scenarios, to quan- clustering for smart meter electricity loads based on quantile
tify the technical requirements of future generation systems autocovariances,” IEEE Trans. Smart Grid, vol. 11, no. 5, pp.
and transmission and distribution networks. 4522–4530, 2020, doi: 10.1109/TSG.2020.2991316.
Finally, these tools can be used in demand forecasting A. M. Alonso, P. Galeano, and D. Peña, “A robust pro-
to make a more precise purchase in the wholesale electric- cedure to build dynamic factor models with cluster struc-
ity market. Tailor-made forecasting models can now be ture,” J. Econometrics, vol. 216, no. 1, pp. 35–52, 2020, doi:
designed and trained to characterize the different dynamic 10.1016/j.jeconom.2020.01.004.
behaviors of each type of customer. J. Caiado, E. Maharaj, and P. D’Urso, “Time-series clus-
tering,” in Handbook of Cluster Analysis, C. Hennig, M.
Conclusions Meila, F. Murtagh, and R. Rocci, Eds. London, U.K.: Chap-
Most energy systems worldwide, including the European man and Hall, 2015, pp. 262–285.
energy sector, are undergoing a major transformation toward R. J. Hyndman, X. Liu, and P. Pinson, “Visualizing big
a new model where energy efficiency and customer empow- energy data: Solutions for this crucial component of data
erment are increasingly important. analysis,” IEEE Power Energy Mag., vol. 16, no. 3, pp. 18–
How customers consume electricity is constantly chang- 25, 2018, doi: 10.1109/MPE.2018.2801441.
ing. The electrification of the demand, distributed renewable T. Liao, “Clustering of time series data: A survey,” Pat-
generation, electric cars, and telecommuting are having a tern Recognit., vol. 38, no. 11, pp. 1857–1874, 2005, doi:
major impact on consumption profiles, which are growing 10.1016/j.patcog.2005.01.025.
increasingly heterogeneous. Thus, the customer segmentation
needs to be updated frequently and adapted to the new data Biographies
received from smart meters. The frequency of the measure- Andrés M. Alonso is with Universidad Carlos III de Madrid,
ments and the number of monitored customers are expected Getafe, 28903, Spain.
to continuously increase, creating opportunities for extracting Eduardo Martín is with Endesa Energía, Madrid, 28042,
more information to gain insights into consumer behavior. Spain.
This new paradigm entails major challenges and opportu- Alicia Mateo is with Endesa Energía, Madrid, 28042, Spain.
nities for both consumers and companies participating in the Francisco J. Nogales is with Universidad Carlos III de
electricity sector. There is a clear need to develop new and Madrid, Leganés, 28911, Spain.
efficient analytical clustering techniques to cope with larger Carlos Ruiz is with Universidad Carlos III de Madrid,
data volumes that include both structured and unstructured Leganés, 28911, Spain.
data types (consumption and weather time series combined Andrea Veiga is with Endesa Energía, Madrid, 28042,
with digital images, sensor data, social media, voice record- Spain.
 p&e
ings, and so on). In this regard, emerging clustering techniques

may/june 2022 ieee power & energy magazine 63

 ©SHUTTERSTOCK.COM/METAMORWORKS

Data-Driven
Dynamic Modeling
in Power Systems
A Fresh Look on Inverter-Based Resource Modeling

By Lingling Fan, Zhixin Miao, Shahil Shah,

Przemyslaw Koralewicz, Vahan Gevorgian, and Jian Fu
Digital Object Identifier 10.1109/MPE.2022.3150827
Date of current version: 19 April 2022

64 ieee power & energy magazine 1540-7977/22©2022IEEE may/june 2022

T
THIS ARTICLE INTRODUCES WAYS TO IDENTIFY Next, we discuss IBR model identification. The current
dynamic system models using measurement data. In power practice of which focuses mainly on obtaining frequency-
system analysis, a static model represents the time-invari- domain admittance/impedance measurements using fre-
ant input–output relationship of a system, while a dynamic quency scans. The resulting models are black-box models
model describes the behavior of the system over time. For that reflect input–output relations. There are many ways to
example, how will a system transit from one steady-state structure IBR dynamic models to map the same input–out-
operation point to another? put relation. A more challenging question is: Can we guess
In the control community, learning dynamic models is a model structures and figure out the model parameters based
system identification problem. Essentially, machine learn- on measurements? The follow-up discussions attempt to
ing and system identification are all about inferring models address those questions. Figure 1 summarizes the six appli-
from data. Both rely on optimization. The exact processes cations discussed in this article.
of the inference may vary from statistical modeling to deep
learning neural networks. This article focuses on presenting A Brief Classification of
unique applications for deriving power system dynamic mod- Measurement Data and Models
els from measurement data. Measurement data can be expressed in the time domain or
Dynamic behaviors are difficult to capture, especially for frequency domain. In power grids, digital fault recorders
applications that lack analytic models. That is where data- and PMUs capture time-stamped (time-domain) dynamic
driven/machine learning techniques can play a critical role. response data. Frequency-domain data are usually produced
Indeed, there is a long history of power system engineers build- via frequency scans; this is also known as the harmonic
ing dynamics models using data-driven approaches, well injection method. To measure the admittance of a device,
before machine learning was a popular term. The integra- a test circuit is first built to connect the device to a control-
tion of inverter-based resources (IBRs) adds more complexity lable voltage source. A sinusoidal perturbation is injected
to the existing modeling framework because of the underly- into the input portal: the voltage source. The output port (the
ing complex physics of IBR systems and the strict nondisclo- current)’s steady-state time-domain responses are processed
sure requirements from original equipment manufacturers via Fourier transform to extract the frequency components.
(OEMs). Thus, data-driven-based system identification meth- Thus, the frequency response of the input–output system is
ods are playing an increasingly important role, especially measured at that frequency. This experiment can be repeated
for systems where physical models are elusive. For example, for a varying frequency.
when representing aggregated distributed energy resources We use a simple example of a series-connected resistor-
at a transmission and distribution interface, one popular inductor-capacitor (RLC) circuit to illustrate the types of mea-
approach is to use the interface measurement data to produce surement data and the identified models. Figure 2 presents the
a distributed energy resource model that can map the inputs procedure of estimating the parameters of the resistor R, the
(e.g., the voltage and frequency) to the outputs (real and reac- inductor L, and the capacitor C from time-domain dynamic
tive power generation). response data. The time-domain dynamic response data are
We begin by introducing the classification of measurement generated by a step change in the source voltage with the
data and models. We then review five commonly used data- capacitor voltage measured at a sampling period of 0.001 s.
driven dynamic modeling applications in power systems:
1) synchronous generator model parameter identification
2) aggregated load model parameter identification
3) reduced-order model identification for control design Aggregated
4) admittance model identification for subsynchronous Load Model
Identification
resonance (SSR) screening Synchronous
Reduced-Order
5) electromechanical oscillation mode identification from Model
Generator
Identification
phasor measurement unit (PMU) data. Model for Control
Identification Data-Driven
The first two applications are different from the last Design
Dynamic
three in terms of the outcomes of the estimation. The first Modeling Admittance
two applications estimate model parameters. This means Model
that the model structure is prior knowledge, and the estima- IBR Model Identification
Identification for SSR
tion leads to model parameters. Compared to the last three
PMU-Based Screening
applications, partial information of the estimation model is
Oscillation
known; i.e., the model is a gray-box model. Thus, dynamic Mode Analysis
model parameter estimation problems are indeed examples
of gray-box model identification. On the other hand, if the
model structure is not imposed, the estimation leads to figure 1. Measurement-based dynamic modeling: six
black-box models. applications.

may/june 2022 ieee power & energy magazine 65

 White noise is imposed in the capacitor voltage measurement estimated output with the measurement output. Figure 2 also
data to emulate the effect of noise in the measurement sensor. presents the measured output data versus the estimated out-
The basic procedure is to first build a dynamic model to put. The red line represents the estimated output based on an
represent the RLC circuit, with R, L, and C as the model initial guess of the RLC parameters, while the blue line rep-
parameters. The parameters are then tuned to match the resents the estimated output based on the RLC parameters

vs Simulated Response Comparison

1.8
1 Measurement
+ 1.6
t Output vc Error sys: 88.66%
+ R L + 1.4 init_sys: 62.97%
C vc
vs 1.2
–

Amplitude
–
– 1

y1
Vary R, L, C 0.8
0.6
Estimation 0.4 Initial Guess: [0.9*R; 0.8*L; 1.5*C]
vc 1
= 0.2 Final Estimation: [0.275*R; 0.238*L; 4.146*C]
vs LCs 2 +RCs + 1
0
0 0.02 0.04 0.06 0.08 0.1
Time (s)
(a) (b)

figure 2. Estimate model parameters (R, L, and C) from time-domain measurement data. This application is an example
of gray-box model identification. (a) The estimation procedure. (b) The comparison of outputs from the estimated models
with the measurement.

+ R L +

Perturbation in C vc Capacitor
vs
the Source Voltage
– – Measurement
Voltage

Input–Output
Time-Series Data

Fourier Transform

Data Fitting
Input–Output Transfer Function
Frequency
Response Data vc –0.01054 s2 + 4.706 s + 56,420 s
= 3
vs s + 75.4 s2 + 56,420 s – 1,030

figure 3. The procedure of extracting a transfer function from the frequency-domain measurements. This application is
an example of black-box model identification.

66 ieee power & energy magazine may/june 2022

identified through a least-squared error minimization proce- frequency-domain response measurements. Curve fitting of fre-
dure. The parameter set after optimization leads to a much quency-domain data also may be used to find the transfer func-
better degree of matching with the original data. tion. The parameters of the identified transfer function may be
This procedure of data-driven parameter estimation is further mapped to the parameters with physical meaning.
an example of gray-box model identification. Note that the Generating time-domain short circuit transient response
estimation model structure has been given as a second- data and frequency-domain reactance data requires testing a
order transfer function with its numerator and denominator generator offline. For example, to obtain frequency responses,
coefficients associated with the RLC parameters. the rotor shaft of a generator is kept standing still, while the ac
Figure 3 presents the procedure of extracting an input– side is connected to a voltage source with a varying frequency.
output model represented by a Laplace transform transfer The current phasor at that frequency must be extracted to gen-
function from the frequency-domain data. First, the source erate a frequency response plot of the reactance. This probing
voltage is perturbed with a sinusoidal signal at a frequency method uses frequency scans. More recently, methods using
with a known magnitude. The capacitor voltage is measured. online measurement data for model parameter identification
A Fourier transform (a technique to transform a function of have also been proposed and implemented.
time to a function of frequency) is conducted on the mea-
surement data to extract the complex Fourier coefficient, or Application 2: Aggregated Load
the phasor, of that frequency. The ratio of the output pha- Model Identification
sor and the input phasor is then obtained. This experiment At a bulk power system level, load modeling deals with
is repeated for a varying frequency. Figure 3 also presents aggregated load modeling so that the performance of the
the resulting frequency response data. Data fitting of the fre- computer model matches the field measurements. This appli-
quency response measurements leads to a third-order trans- cation of load model parameter identification is also nor-
fer function describing the input–output relationship. mally a gray-box model identification problem. Before 1990,
This procedure of directly fitting frequency-domain loads were represented by static models in computer software
measurements to a third-order transfer function is an exam- packages. For example, the ZIP model assumes the total real
ple of black-box model identification. The resulting trans- power consumption of aggregated loads is a combination of
fer function does not give explicit information about the constant-impedance (Z), constant-current (I), and constant-
dynamic model structure. The coefficients of the numera- power (P) components. Starting from the late 1980s, dynamic
tor and the denominator of the transfer function also do load models were developed to improve the system model-
not associate with the physical parameters of the RLC cir- ing accuracy. Time-domain data have been used to identify
cuit. In summary, the measurement data used for dynamic load models. In the 1990s, a research project carried out at
model identification can be in either the time domain or the the Panchiao substation in the Taiwan power system demon-
frequency domain, and the model structure can be either strated that using the first-order representation of the voltage
known (a gray-box model) or unknown (a black-box model). and real-power relationship cannot provide a good match to
the field measurement data before/after a single-phase fault.
Application 1: Synchronous Generator Thus, higher-order models (second and third) were used, and
Model Identification the matching accuracy was significantly improved. The model
This application of generator model parameter identifi- identification process included three key steps:
cation takes the gray-box model identification approach 1) Determine the transfer function structure based on
because the parameters identified are associated with a prior knowledge.
known model structure. 2) Convert the transfer function to its discrete-time pre-
After the invention of synchronous generators in the diction error model.
1880s, a synchronous generator model relying on Park’s trans- 3) Obtain measurement data and conduct curve fitting to
formation (a technology developed by R.H. Park to transform find the model parameters.
variables in an ABC-frame to variables expressed in a rotating Other load model representations are also available, e.g.,
rotor frame) was developed in the 1920s. In this representa- an induction motor parallel with a resistor–capacitor circuit.
tion, a solid rotor is represented by the rotor circuits in the Measurement data are then used to find the induction motor
direct and quadrature (dq)-axes. Test procedures to obtain parameters and the resistor–capacitor circuit parameters.
the dq-axis circuit parameters were designed after the 1920s.
As a result, IEEE published IEEE Standard 115, “Test Pro- Application 3: Reduced-Order Model
cedures for Synchronous Machines,” in 1965 and revised the Identification for Control Design
standard in 1983. Both transient response measurements from Besides modeling, another category of application of mea-
short circuit tests and frequency response measurements have surement data is the development of a reduced-order dynamic
been used to find reactance and time constants. For example, model for control design. The control design problem could
a procedure relying on asymptotic approximation can be be, e.g., devising a power system stabilizer for a synchro-
used to find the parameters of a transfer function from its nous generator’s excitation system or a damping control for

may/june 2022 ieee power & energy magazine 67

 a flexible ac transmission system device. The application of work resonance is complementary to the frequency of a tor-
reduced-order model identification normally uses the black- sional model, i.e., the sum of the two frequencies is 60 Hz,
box model identification approach because the internal model torsional oscillations may become severe.
structure is usually unknown. Compared to dc circuit analysis dealing with a single-
For any control design problem, a plant model describ- input and single-output system, stability analysis in power
ing the input–output relationship is a necessity. Usually, a systems usually deals with three-phase systems. Modeling a
reduced-order plant model is desired. How do we find the three-phase system in a rotating dq-frame can greatly sim-
plant model? The measurement-based approach is to perturb plify the resulting model. Indeed, one of the most influential
the original system’s input and record the output data. From modeling technologies of power systems is Park’s transfor-
either the time-domain data or frequency-domain data, the mation, which converts variables in an ABC-frame to those
input–output plant model can be found. in a rotor dq-frame. As a result, a synchronous generator
For example, in a power system stabilizer design, the model is expressed from the perspective of the rotating rotor
plant has the reference order of the voltage regulator as an frame. Besides generators, other components of power sys-
input and the generator speed as the output. The input can be tems may also be expressed in a dq-frame for simplicity.
perturbed with an impulse signal, and the output response Thus, dq-frame models are preferred in stability analysis.
will be recorded. Subspace methods, e.g., the eigensystem For SSR analysis, a circuit of a generator with a series-
realization algorithm, may be used to process the output data compensated interconnection can be interpreted as a two-
and lead to a reduced-order plant model. Based on the plant input and two-output feedback system. The forward unit is the
model, a controller modulating the voltage regulator’s refer- line’s admittance; the feedback unit is the generator’s imped-
ence with the generator speed as the input can be designed ance. A stability analysis can then be carried out by examin-
and tested for the closed-loop system performance. ing the feedback system via well-established multi-input and
Subspace methods such as the eigensystem realization multioutput frequency-domain system analysis theories.
algorithm can be traced back to the seminal state-space To obtain the generator and network impedance from
model realization theory established by Ho and Kalman computer simulations, frequency scans have since been
in the 1960s. The core message of the theory is that time- popularly used in SSR studies. Recently, frequency scans
domain dynamic response data can be stacked properly to have been used in wind farm SSR screening in electromag-
form a large-dimension Hankel data matrix, and this Hankel netic transient simulation software environments by the grid
data matrix is associated with the state-space model’s system operating industry. In the power electronics field, obtaining
matrices. Factorizing the Hankel matrix via singular value dq impedance/admittance frequency-domain measurement
decomposition leads to the system matrices. From there, a through hardware setup, perturbation signal injection, and
state-space model that matches the input–output relation- measurement processing has been a research topic.
ship can be recovered. Furthermore, with the system matri- Frequency scans lead to frequency-domain measure-
ces known, the eigenvalues of the system can also be found. ment. A benefit of frequency-domain measurement is its use
Subspace methods have been used in other applications, e.g., for stability analysis. Either open-loop system Bode plots or
PMU-data-based oscillation mode identification. They may Nyquist plots can be plotted, and stability predictions can
be viewed as inference methods that relate data to a model. be made. However, those diagrams have disadvantages com-
These techniques also have the capability of differentiating pared to closed-loop system eigenvalues. They may not lead
noise versus meaningful information, a feature of unsuper- to accurate stability prediction, or they may not be straight-
vised clustering learning. forward for interpretation.
An eigenvalue, in the form of a complex number, gives a
Application 4: Admittance Model direct and accurate prediction of the stability of a dynamic
Identification for SSR Screening system. The real part of an eigenvalue must be less than zero
Stability analysis via frequency-domain models has a his- for a system to be stable, and the imaginary part of an eigen-
tory dating back to the 1970s in both the power electronics value implicates the oscillation frequency. Thus, eigenvalues
and power systems communities. In the power electronics directly tell if the system is stable or not and what the system’s
community, the initial use of impedance models for dc cir- oscillation modes are. For SSR stability analysis, a generator or
cuit stability analysis started in 1976. In the power systems transmission system’s frequency-domain admittance/imped-
community, dq admittance-based SSR stability analysis ance measurements must be fitted into a model in the form of
also started in the 1970s, after the Mohave power plant SSR a transfer function matrix. From there, eigenvalue calculation
events in Nevada. For these events, a synchronous generator is possible. Though it seems trivial to arrive at eigenvalues
radially connected to a series-compensated line experienced after obtaining the admittance measurements of a subsystem,
oscillations in its torque shaft, causing shaft damage. The note that the frequency-domain data-fitting technology was
torsional oscillations were triggered by the electric network not available in the 1970s. This technology became available
resonance due to the interaction of the series capacitor and only after 2000. Without frequency-domain data fitting, it is
the line inductance. When the frequency of the electric net- difficult to identify models and compute eigenvalues.

68 ieee power & energy magazine may/june 2022

 This application of admittance model identification is to parameters. This leads to ill-conditioned estimation problems.
identify a black-box model describing the terminal voltage If this is the case, the estimation problem can be formulated
and current relationship only. The resulting model does not to estimate a subset of the parameters. Algorithm-wise, con-
lead to further information on the generator model structure. vergence and local optimum are the main issues for nonlin-
ear optimization problems. For parameter estimation, a local
Application 5: PMU Data-Basedt optimum means the identified parameters may be far from the
Oscillation Mode Identification true parameters. The resulting estimated output may have a
PMU data have been used for oscillation mode identification poor matching degree with the measured output. Therefore,
where the black-box modeling approach is usually applied. much effort has been devoted to refining the optimization
In 2012, the IEEE Power and Energy Task Force of Iden- problem formulation.
tification of Electromechanical Modes published a report, Optimization is one of the key technologies in gray-box
“Identification of Electromechanical Modes in Power Sys- identification. A significant achievement in recent years is
tems.” The electromechanical modes are in the range of a the adoption of convex programming techniques into opti-
fraction of to several hertz. This report presents a perspec- mization problem formulation and solving. A benefit is that
tive on using different identification methods for finding the the solution to a convex optimization problem is the global
oscillation mode frequency, damping ratio, and mode shape optimum, which ensures that the identified parameters pro-
based on PMU data with a sampling rate of 30–60 Hz. duce the best matching results.
If time-domain dynamic response data are viewed as
the impulse responses of a dynamic system describing the IBR Modeling: A Forward Look
entire power grid, the Laplace transform of the time-domain The sixth application is data-driven IBR modeling. We use a
data represents the resulting input–output model. Oscillation 2.3-MVA inverter as an example to demonstrate the state-of-
modes are associated with the poles of the Laplace domain the-art technology in black-box IBR model identification,
expression, or the eigenvalues. Subspace methods, e.g., the and we give our perspective on challenges to be tackled in
eigensystem realization algorithm, may be used to form a IBR modeling.
Hankel data matrix, further extract the system matrices for
eigenvalue computing. From the eigenvalues, the oscillation dq Admittance Model Identification
modes, damping ratio, and frequencies are found. PMU- For IBRs, dq admittance measurement technology is a
based oscillation mode identification is a mature technology. mature technology. The measurement capability can be real-
For example, several utilities have real-time mode analyzers ized in software as well as hardware experiments with the
available to process PMU data. availability of advanced high-power converters and medium-
voltage sensors.
Summary of the Five Applications Figure 4 shows a measurement testbed set up at the National
and Recent Progress in Gray-Box Renewable Energy Laboratory’s Flatirons Campus in Col-
Model Identification orado, United States. A critical component of the testbed
To summarize, for the five applications, three of them, i.e., is the 7-MVA, 13.2-kV controllable grid interface (CGI).
finding reduced-order models for control design, finding dq The CGI essentially works as a grid-forming converter. It
admittance for SSR stability analysis, and PMU-based elec- draws electricity from a utility grid and acts as a control-
tromechanical oscillation mode identification, are related to lable voltage source. When an IBR is connected through
identifying black-box models from measurement data. Those a step-up transformer to a CGI, it can be configured to
models describe the input–output relationship only. The inter- operate at a certain operating condition. The CGI can pro-
nal structure and parameters of the system under investiga- duce a harmonic voltage source superimposed to its 60-Hz
tion are not imposed as for a gray-box model. The technology voltage source. This harmonic voltage source’s frequency
of black-box model identification is mature as we have seen may vary. Thus, frequency scans can be conducted using
real-world applications in these areas. The black-box models the CGI.
are all linear models. For this testbed, the model to be identified describes
On the other hand, gray-box model identification is actively the relationship between the two inputs (the dq-axis volt-
under investigation. The first two applications—identifying ages) and the two outputs (the dq-axis currents), as shown
generator reactance and time constants and identifying param- in Figure 4. This model is called dq admittance and it
eters for load modeling—belong to the category of gray-box is a two-by-two matrix in the Laplace domain. The four
model identification. For those applications, a prior knowl- components of the dq admittance matrix are Ydd , Ydq , Yqd ,
edge of internal physics must be combined with measure- and Yqq. Figure 5 presents photos of the CGI and medium-
ment-based learning to achieve the goal of model identifica- voltage sensing equipment at the National Renewable
tion. The models can be nonlinear. Energy Laboratory.
The main issue of gray-box identification is that the mea- The resulting dq admittance of the IBR viewed at the
surement data may not contain sufficient information on the measurement point is shown in Figure 6. The measurement

may/june 2022 ieee power & energy magazine 69

 testbed is configured so that the IBR works in four operating Case 2: r eal power at 500 kW or 0.5 p.u. (–6 dB), reactive
conditions expressed by real power and reactive power. For power at 0 Mvar
per-unit (p.u.) values, the base power is 1 MVA. Case 3: real power at 0 MW and reactive power at 500 kvar
Case 1: real power at 0 MW and reactive power at 0 Mvar or 0.5 p.u. (–6 dB)

Converter PCC Measurement Point

GSC
V I1 VPCC Vg
CGI
T1 P
L1 R1 Q
Battery

1 MVA 13.2 kV
C1 400 V: 13.2 kV Grid
vabc
RC
P Filter
GSC
Q Control

abc/dq dq/abc
i1,abc vPCC,abc
id iq vd vq

100
1

50 0.5

0 0

–0.5
–50
–1
–100
0 0.01 0.02 0.03 0.04 0 0.01 0.02 0.03 0.04
Response in Current Perturbed Voltage

figure 4. The measurement testbed for a 2.3-MVA inverter. GSC: grid-side converter; PCC: point of common coupling;
CGI: controllable grid interface.

Grid-Side Output Voltage Source (b) (c)

Transformer Transformer Converters
(a)

figure 5. The CGI and sensors. (a) The 7-MVA grid simulator. (b) and (c) Medium-voltage sensing. (Source: Mark McDade,
National Renewable Energy Laboratory; used with permission.)

70 ieee power & energy magazine may/june 2022

 Case 4: real power at 1 MW (0 dB) and reactive power frequency-domain measurement, we may apply frequency-
at 0 Mvar. domain data-fitting methods and obtain a black-box model.
At each operating condition, about 40 sinusoidal injec- This step is necessary if we aim to have an s-domain admit-
tion experiments are conducted. For each sinusoidal injection tance for eigenvalue analysis, which can lead to an overall pic-
experiment, injection in the d-axis voltage is first conducted, ture of the system modes.
and the resulting dq-axis current measurements are collected. The data in Figure 6 have been fitted using a frequency-
A Fourier transform is then applied to the steady-state time- domain data-fitting package. A comparison of the frequency
domain data to find the phasors. From there, the first column responses of the model versus the measurements are shown
admittance components Ydd(f) and Yqd(f) are found. Next, injec- in Figure 7(a). The figure shows that data fitting leads to a
tion in the q-axis voltage is conducted, and the second column high matching degree in the studied frequency spectrum.
admittance components Ydq(f) and Yqq(f) are found. From the One more comparison can be made: the step responses of

40 40
Ydd Ydq
–6.28
20 20
Mag (dB)

Mag (dB)
0 0

–20 –20

–40 –40
10–1 100 101 102 103 10–1 100 101 102 103
Frequency (Hz) Frequency (Hz)
(a) (b)
200 200
Phase (Degrees)

Phase (Degrees)

0 0

–200 –200
10–1 100 101 102 103 10–1 100 101 102 103
Frequency (Hz) Frequency (Hz)
(c) (d)
40 40
Yqd –8 Yqq
20 20
Mag (dB)

Mag (dB)

0 0

–20 –20

–40 –40
10–1 100 101 102 103 10–1 100 101 102 103
Frequency (Hz) Frequency (Hz)
(e) (f)
Phase (Degrees)

Phase (Degrees)

200 200

0 0

–200 –200
10–1 100 101 102 103 10–1 100 101 102 103
Frequency (Hz) Frequency (Hz)
(g) (h)

figure 6. dq admittance of a 2.3-MVA inverter under four operating conditions: black: case 1; blue: case 2; green:
case 3; red: case 4. Case 1: real power at 0 MW and reactive power at 0 Mvar. Case 2: real power at 500 kW or 0.5 p.u.
(–6 dB), reactive power at 0 Mvar. Case 3: real power at 0 MW and reactive power at 500 kVar or 0.5 p.u. (–6 dB).
Case 4: real power at 1 MW (0 dB) and reactive power at 0 Mvar. (a) Ydd’s magnitude. (b) Ydq’s magnitude. (c) Ydd’s
angle. (d) Ydq’s angle. (e) Yqd’s magnitude. (f) Yqq’s magnitude. (g) Yqd’s angle. (h) Yqq’s angle.

may/june 2022 ieee power & energy magazine 71

 the physical device versus the model. Figure 7(b) presents the based on the first principles and prior knowledge while esti-
comparison under two step changes in the dq-axis voltages: mating the model parameters using measurement data. On
10% or 20%. The physical device and the model have very the other hand, IBR gray-box model identification is a more
similar step responses. challenging problem from both the IBR dynamic model-
Another admittance model identification technology is to building perspective and mathematic optimization problem-
use time-domain step responses. Converting the step responses solving perspective.
into s-domain expressions and assembling lead directly to an
s-domain dq admittance model. In real-world applications, Challenges in Gray-Box Model Identification
step response data are polluted with noise. The resulting Using the 2.3-MVA inverter as an example, we will further
model may not be accurate in the high-frequency range or demonstrate how to use dq admittance measurement data to
when the measurement data have small values. speculate the inverter control structure and parameters. As the
Remark: Frequency scans and frequency-domain data fit- first step, the control structure needs to be specified. In this
ting are two mature technologies and can be employed for case, two types of popular control structures are examined.
IBR’s dq admittance model identification. Figure 8 presents the converter control structures and a com-
The admittance model identified is a linear model associ- parison of the frequency responses of the two models versus
ated with an operating condition. An IBR may have a variety the measurements.
of operating conditions. Thus, one challenge is how to find In both models, the converter controls have the same
any admittance model associated with a random operating goal of following the real power and reactive power orders.
condition. A straightforward solution is to build a nonlinear Both employ a cascaded control structure: outer controls
model that can reflect the operating condition. This approach to track real and reactive power orders and inner controls
is the gray-box modeling approach: building the model structure to track current orders. The two models differ in the inner

40 40
Ydd Ydq
20 20
Mag (dB)

Mag (dB)

0 0
–20 –20
–40 –40
100 102 100 102
200 200
Phase (Degrees)

Phase (Degrees)

0 0

–200 –200
100 102 100 102

40 40
Yqd Yqq
20 20
Mag (dB)

Mag (dB)

0 0
–20 –20
–40 –40
100 102 100 102

200 200
Phase (Degrees)

Phase (Degrees)

0 0

–200 –200
100 102 100 102
Frequency (Hz) Frequency (Hz)
(a)

figure 7. (a) Frequency-domain data-fitting results. Solid blue line: model; black crosses: measurements. Operating condi-
tion: Case 3 where P = 0 MW, Q = 500 kVar. (Continued)

72 ieee power & energy magazine may/june 2022

current control implementation frames. The dq-frame con- are employed to ensure the error between the measurement
trol has its inner current control implemented in a dq-frame; and the order achieving zero at the nominal frequency.
that is, the phase current measurements are first converted A comparison of the dq admittance of the 2.3-MW
to a dq-frame. This dq-frame has the rotating speed of the inverter versus the two models shows that the second model
nominal frequency at steady state. When projected to the results in better matching for the diagonal components Ydd
dq-frame, periodic signals with the nominal frequency be­­ and Yqq. Specifically, for the dd component, a large mismatch
come dc signals. is observed in the range of 1–100 Hz if the dq-frame control
The proportional-integral control units are known for is assumed. On the other hand, the 2.3-MW inverter and the
their capability of tracking dc signals. They can enforce the model match well in the range of 0.1–100 Hz if the station-
dq-axis current measurements to follow the current orders ary-frame control is assumed.
generated by the outer controls. The stationary-frame con- Refining the model structure and parameter tuning are the
trol has its inner current control implemented in a station- next steps. Specifically, parameter tuning can be achieved
ary frame. In this frame, currents are still periodic. To using an automatic procedure instead of manual tuning. To
track a current order, proportional-resonant control units achieve automatic tuning, optimization problem formulation

Measurement Model
Event 1: vd Step Down
1 (1) 1 (1)
id 10% Step Change id 10% Step Change
0.5 20% Step Change 0.5 20% Step Change
0 0

–0.5 –0.5

–1 –1
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
0 (1)
0
(1)
iq iq
–0.5 –0.5

–1 –1

–1.5 –1.5

–2 –2
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
Time (s) Time (s)
Event 2: vq Step Up
2 2
(2) (2)
id id
1 1

0 0

–1 –1

–2 –2
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
2 2
(2) (2)
iq iq
1 1

0 0

–1 –1

–2 –2
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
Time (s) Time (s)
(b)

figure 7. (Continued) (b) Comparison of the step response of the identified black-box model versus the time-domain
measurements.

may/june 2022 ieee power & energy magazine 73

 and solving becomes an immediate task. From this 2.3-MW Challenge 1: IBR Model Structure Design
inverter example, we may see that there are two main chal- In the past decade, a set of generic models for IBRs has
lenges: model design and customized model parameter esti- been developed by the Western Electricity Coordinating
mation algorithm design. Council’s model validation subcommittee for grid dynamic

VPLL
P
– +
P∗ + id∗ + ud +
vdc
PI PI
dq c
va
– –
c
id ωL
vb

c
iq ωL vc
– +
Q∗ – i∗
q uq vqc abc
PI + +
PI
+
Q

θPLL
(a)

40 40
20 CGI Measurements 20 Measurement
Mag (dB)

Mag (dB)

0 0
–20 dq-Frame Model –20 Model
–40 –40
100 102 100 102
Phase (Degrees)

Phase (Degrees)

200 200

0 0

–200 –200
100 102 100 102
40 40
20 20
Mag (dB)

Mag (dB)

0 0
–20 –20
–40 –40
100 102 100 102
Phase (Degrees)

Phase (Degrees)

200 200

0 0

–200 –200
100 102 100 102
Frequency (Hz) Frequency (Hz)

(c)

figure 8. Two types of control structures differ in current controls: (a) dq-based current control and (b) stationary-frame-
based current control. Comparison of frequency responses and measurements: (c) dq-frame control and (d) stationary-
frame control.

74 ieee power & energy magazine may/june 2022

a­ssessment. These models are suitable for power system not include fast electromagnetic transient dynamics and fast
transient simulation studies with numerical integration time inverter current controls.
steps in the range of 1–5 ms. Such models are based on To have models accommodating a wide range of oper-
quasi-steady-state positive-sequence phasors and usually do ating conditions, including unbalance, fast dynamics, and

P∗ + vα
PI LPF dq PR
αβ
va
–
P–
iα vb
Q
iβ vc
+
αβ
abc
Q∗ –
– vβ
PI LPF PR

θPLL

(b)

40 40
Mag (dB)

Mag (dB)

20 Measurement 20
0 0
–20 Model –20
–40 –40
100 102 100 102
Phase (Degrees)

Phase (Degrees)

200 200

0 0

–200 –200
100 102 100 102
40 40
Black: Measurements
Mag (dB)

Mag (dB)

20 20
0 0
–20 Red: Model –20
–40 –40
100 102 100 102
Phase (Degrees)

Phase (Degrees)

200 200

0 0

–200 –200
100 102 100 102
Frequency (Hz) Frequency (Hz)

(d)

may/june 2022 ieee power & energy magazine 75

 weak grid conditions, IBR models including electromag- Acknowledgments
netic transients and fast controls are desired. For the power We wish to recognize the support of the United States
industry, this is an ongoing effort. For example, the CIGRE Department of Energy and the National Science Foundation
C4.60 Working Group aims to design generic electromag- for some of the materials reported.
netic transient models of IBRs with transparent IBR con-
trol structures. For Further Reading
Because state variables are time varying at the fundamen- L. Ljung, C. Andersson, K. Tiels, and T. B. Schön, “Deep
tal frequency in the ABC domain, it is very difficult to derive learning and system identification,” IFAC-PapersOnLine,
linear models in the ABC-frame. Linear time-invariant mod- vol. 53, no. 2, pp. 1175–1181, Jan. 1, 2020, doi: 10.1016/j.ifa-
els are preferred since they are suitable for small-signal anal- col.2020.12.1329.
ysis. Therefore, modeling efforts are required to convert a C.-J. Lin et al., “Dynamic load models in power systems
model in the ABC-frame to a model in different coordinates so using the measurement approach,” IEEE Trans. Power Syst.,
that its state variables are constant at steady state. The result- vol. 8, no. 1, pp. 309–315, 1993, doi: 10.1109/59.221226.
ing nonlinear models can be easily linearized via numerical “Wind energy systems sub-synchronous oscillations:
methods for linear time-invariant model extraction. Events and modeling,” IEEE Power & Energy Society, Pis-
Besides the technical challenges in modeling, another cataway, NJ, USA, Wind SSO Task Force, Tech. Rep. PES-
significant technical gap in designing transparent mod- TR80, 2020. [Online]. Available: https://resourcecenter.
els is that the IBR controls are the proprietary infor- ieee-pes.org/publications/technical-reports/PES_TP_TR80_
mation of the OEM. Strict nondisclosure requirements AMPS_WSSO_070920.html
are imposed by OEMs, which makes any model design a L. Fan and Z. Miao, “Admittance-based stability analy-
challenging task. Thus, efforts must be made to standard- sis: Bode plots, Nyquist diagrams or eigenvalue analysis?”
ize IBR control to better define their dynamics and support IEEE Trans. Power Syst., vol. 35, no. 4, pp. 3312–3315, Jul.
gray-box modeling. Currently, there are ongoing endeav- 2020, doi: 10.1109/TPWRS.2020.2996014.
ors in the grid industry; e.g., the IEEE P2800 Working L. Fan, Z. Miao, P. Koralewicz, S. Shah, and V. Gevor-
Group aims to set up the minimum technical require- gian, “Identifying DQ-domain admittance models of a 2.3-
ments for IBRs. MVA commercial grid-following inverter via frequency-do-
main and time-domain data,” IEEE Trans. Energy Convers.,
Challenge 2: Customized Model Parameter vol. 36, no. 3, pp. 2463–2472, Sep. 2021, doi: 10.1109/
Estimation Algorithm Design TEC.2020.3048389.
The second challenge lies in the design of the model param- C. Yu, L. Ljung, A. Wills, and M. Verhaegen, “Con-
eter estimation algorithm. This requires familiarity with strained subspace method for the identification of struc-
the domain knowledge of IBR power electronic converter tured state-space models (COSMOS),” IEEE Trans. Autom.
control and various mathematical methods relying on lin- Control, vol. 65, no. 10, pp. 4201–4214, Dec. 6, 2019, doi:
ear algebra and optimization. To this end, power system 10.1109/TAC.2019.2957703.
engineers can leverage many recent advances in comput- S. Shah, P. Koralewicz, V. Gevorgian, and R. Wallen,
ing as well as free optimization solvers and platforms, such “Sequence impedance measurement of utility-scale wind
as nonlinear optimization solvers (e.g., the Interior Point turbines and inverters reference frame, frequency coupling,
OPTimizer) and optimization problem formulation inter- and MIMO/SISO forms,” IEEE Trans. Energy Convers.,
faces (e.g., YALMIP and CVX for MATLAB and JuMP early access, Jun. 2021, doi: 10.1109/TEC.2021.3093516.
for Julia).
Biographies
Conclusion Lingling Fan is with the University of South Florida, T
­ ampa,
In summary, data-driven dynamic model building has a Florida, 33620, USA.
long history of applications in power systems. As early as Zhixin Miao is with the University of South Florida,
the 1960s, measurements were used to identify a synchro- Tampa, Florida, 33620, USA.
nous generator’s dq reactance and time constants. For the Shahil Shah is with the National Renewable Energy
current power grids with high penetrations of IBRs, the Laboratory, Golden, Colorado, 80401, USA.
power system community once again is examining data- Przemyslaw Koralewicz is with the National Renewable
driven dynamic modeling for IBRs. Compared to the pre- Energy Laboratory, Golden, Colorado, 80401, USA.
vious century, we now have better hardware equipment to Vahan Gevorgian is with the National Renewable En-
conduct experiments, thanks to the advances in power elec- ergy Laboratory, Golden, Colorado, 80401, USA.
tronics. We also have better computing tools because of the Jian Fu is with the Wind Energy Technologies Office,
advances in operations research, system identification, and Department of Energy, Washington, D.C., USA.
 p&e
machine learning.

76 ieee power & energy magazine may/june 2022

history
Sayan Kumar Nag and Tarun Kumar Gangopadhyay

solar photovoltaics
a brief history of technologies

I
IN THE PRESENT CENTURY, SOLAR
energy has emerged as an important In this issue’s “History” column, we examine the utilization of solar energy from
source of nonconventional energy to a variety of historic perspectives and the evolution of different types of photo-
meet the energy demand for overall de- voltaic (PV) cells. We welcome Sayan Kumar Nag and Tarun Kumar Gangopad-
velopment of a nation. The use of solar hyay of the Department of Electrical Engineering, Techno Main Salt Lake Kolkata,
energy for human development is not a West Bengal, India, who join us for their first time to our “History” pages.
new discovery but instead is a century-
John Paserba
old tradition. As the demand for clean
Associate Editor, “History”
energy sources increases, the impor-
tance of the development of efficient
photovoltaic (PV) cells is in demand.
Here we examine the utilization of so- such as Native Americans, Babylonians, of free energy. The first known way to
lar energy in the initial stage, the rise Persians, ancient Hindus, and Egyp- capture the Sun’s rays occurred around
of PV development in the present era, tians, held the Sun in high regard to the 600 BCE, when wood was ignited by
and different kinds of PV cells with point of worshipping it. However, it was focusing the rays onto it through mag-
their merits and demerits. actually the Greeks and Romans who nifying lenses. An illustration of this is
first realized the Sun’s power as a source shown in Figure 1.
Solar Energy—
The Initial Years
It is a common belief among many
people that the use of solar energy is a
phenomenon of modern times either as
a repercussion of the 1970s oil embargo,
a byproduct of the 20th-century space
race, or both. Nevertheless, it isn’t all the
truth. When considering energy sources
in the modern world, the first thing that
generally comes to mind is traditional
energy sources such as coal, oil, and gas.
But biomass from wood, dried animal
manure, and peat has traditionally been
used as fuel for ancient cultures. Obvi-
ously, like other forms of renewable
energy, both traditional and modern bio-
fuels originally generated from the Sun.
The biofuels act like early storage units
of solar energy and are only useful dur-
ing daylight hours. Ancient civilizations,
Digital Object Identifier 10.1109/MPE.2022.3150814 figure 1. An illustration depicting the Sun’s rays being focused to start a fire in 1658.
Date of current version: 19 April 2022 (Source: J.A. Comenius, Orbis Sensualium Pictus, Michael Endter, 1658.)

may/june 2022 ieee power & energy magazine 77

 Few civilizations focused on develop- and started with the idea of providing for reasons other than religious worship,
ing new ideas or technologies unless the heat and light since it was very impor- as early as 400 BCE, implementing pas-
need for development grew. Greece ex- tant for the survival of the community. sive solar design into homes.
perienced a shortage of fuel in the fourth Greece was the first civilization to con- As shown in Figure 2, passive so-
century BCE, so it became innovative sciously use solar energy in its dwellings lar design ensures protecting the north

Summer Solar Irradiation

Control

Aperture
Winter Solar Radiation

Distribution

Absorber
Insulator

figure 2. A passive solar design to protect north side from wind and rain.

figure 3. A photo of ancient Roman baths.

78 ieee power & energy magazine may/june 2022

face of a building from external ele- In the 1800s, there was further sup- take decades before efficiency improved
ments (wind, rain, and so on). On the port for this discovery, but it took until and a true understanding of the causes
south side, the solar radiation during the mid-20th century for proof to be of early low power is reached.
the winter was open, but was shaded in made. A publication by William G. A lack of interest, a lack of knowl-
the summer. This simple principle was Adams and Richard E. Day considered edge, or the low price of fossil fuels may
widely accepted and implemented not the effects of sunlight on selenium, and be behind the little development that oc-
only in private houses but also in pub- Charles Edgar Fritts developed a very curred in the half-century after the ini-
lic buildings and the world famous Ro- inefficient prototype cell (1%–2% ef- tial discovery. The only notable event in
man baths. A picture of a Roman bath ficiency) during the late 1880s that the first half of the 20th century that fol-
is shown in Figure 3. looked like a typical cell in use today. lowed the same trend was Max Planck’s
A typical diagram of a modern PV cell new idea, his light quantum hypothesis.
Photovoltaic is shown in Figure 4. An explanation of the photoelectric
Development—A New Era To calculate the efficiency of PV effect was published by Albert Ein-
For centuries, we have used the power cells, one must measure the amount of stein in March 1905. Five years earlier,
of the Sun in several ways. But it wasn’t electricity generated using the total po- Planck had shown that atoms in cavi-
until the 1800s that scientific break- tential energy received by the PV mod- ties can only absorb or emit radiation
throughs allowed us to fully exploit the ule. In many ways, the prototype solar from discrete “quantum,” which is why
potential of this free and abundant fuel cell resembles today’s modern cells. Be- the energy of each proton is an integer
source. A French physicist named Ed- neath it was a glass sheath with a fine multiple: frequency times new default
mund Becquerel (Nobel laureate) pub- chunk of gold wire sandwiched between constant. Planck thus solved the prob-
lished his experiments conducted with the glass and a thin layer of selenium. lem of blackbody radiation.
wet cell batteries in 1839, which led to This was the first model for further im- Planck believed that his conception
the transition to solar energy use. provement by scientists, but it would of quantum was nothing more than a

making life visibly safer

BIRD DIVERTERS POWER LINE MARKERS LOW LINE MARKERS

FireFly® & BirdMark® SpanGuard TM & SpanLite TM QuikMark TM & SpanMark TM

The most effective bird diverters Mark any line. Anywhere. Saving lines & saving lives.
in independent studies. Standard & self-illuminated markers. Snaps directly onto live lines.

800-722-8078 • pr-tech.com
6874

may/june 2022 ieee power & energy magazine 79

 Light Energy

+
Nickel Plating
Antireflecting
Coating

N-Type Si

+ + Hole
+
Load
P-Type Si

– – – Electron

figure 4. A schematic of a PV cell. Si: silicon.

mathematical “trick” to fit the theory the metal. If the photon’s frequency is The first of the many solar cells that
into experimentation. But Einstein ex- sufficient to drop the electrons, the col- powered the space machinery were
tended Planck’s quantum into light itself lision causes the photoelectric effect. As deployed on the Vanguard I space sat-
(Planck assumed that only the vibrations a particle, light carries energy propor- ellite (Figure 5) in 1958 and operated
of atoms were quantized). Einstein said tional to the frequency of the wave. As until decommissioned in 1964. As an
that light is a beam of particles whose a wave, it has a frequency determined endless pollution-free power supply,
energy is related to frequency accord- by the energy of the particle. This work the tremendous success of PV cells
ing to Planck’s formula. A beam of light earned Einstein the 1921 Nobel Prize has secured a place for these cells. In
collides with atoms upon impact with in Physics. spite of their efficiency and cost, PV
Researchers at Bell Laboratory found cells have become a ubiquitous power
that by mixing arsenic with silicon and source in the aerospace industry due to
a thin coating on the cell with boron, their success as pollution-free infinite
cell efficiency of 6% (US$250 per watt sources of electricity.
versus US$3 per watt for coal) can be After the French discovery in the
achieved. Perhaps even if their findings early 19th century, the United States
were separate from transistor technol- led the development of solar cells. Until
ogy and expertise, it would not have 1990, it was a leader in the market, re-
been known to scientists who started the search and development, and implemen-
energy supply revolution that would be tation. Nevertheless, at the end of the
needed globally within 20 years. Despite 20th century, this domination changed
advances, the current product was too as it moved and was divided between
figure 5. The space satellite expensive for use on the ground. How- Europe and Japan. Over the past decade,
Vanguard 1 with its six solar cells ever, for space power applications, it was the world has reached a tremendous
attached. (Source: National Space Sci- a perfect solution as there were no other milestone integrating solar power into 1
ence Data Center, NASA, Vanguard 1 alternatives and, despite the demerits, it million households. During this period,
with its six solar cells attached, 2007.) was still the best solution. Japan and Europe, especially Germany,

80 ieee power & energy magazine may/june 2022

 table 1. Summary highlights of solar PV technologies of multijunction cells, crystalline modules, thin films, and some new solar modules.
Efficiency

may/june 2022
Solar Technology Characteristics in 2019 (%) Advantages Drawbacks
Monocrystalline Monocrystalline silicon is a pure material 26.7 1) Longevity 1) High cost
with a continuous, unbroken crystalline 2) Low installation cost 2) Fragility
lattice structure that has minimal or no 3) Nonhazardous 3) Silicon wasted in the Czochralski process
impurities.
Polycrystalline Numerous monocrystalline silicon grains 22.3 1) Simple production process 1) Generally less efficient than
are used to manufacture polycrystalline 2) More heat-tolerant than silicon-based monocrystalline
silicon. panels 2) Space efficiency is lower
CIGS Solar energy is converted into electricity by 23.4 1) The material is less or nontoxic, unlike 1) A less efficient alternative to
semiconductor layers in CIGS. Cadmium telluride (CdTe) monocrystalline panels
2) Better at resisting heat compared to
Si-based panels
CdTe It is based on a superstrate comprising of 21.0 1) Solar energy is absorbed at shorter 1) Cadmium is highly toxic
successive layers of soda lime glass, indium- wavelengths 2) Old CdTe panels may pose a disposal
tin oxide, Cadmium sulfide, and CdTe back 2) Cadmium is abundant, resulting in lower problem
contact. manufacturing costs 3) Less efficient compared with crystalline
silicon
Amorphous It is made of silica atoms are arranged in 13.4 1) Multijunction device capacity 1) Low efficiency
Silicon thin, homogeneous layers. 2) Easy fabrication
3) High optical absorption coefficient
4) The material is less toxic compared to CIGS
or CdTe
5) Flexible and less susceptible to crack
Dye-sensitized It includes the electrode film layer, the 11.1 1) Has a lower cost 1) Volatile organic solvents make up
solar cells conductive transparent conductive oxide, 2) Can operate in dim light and from a wider electrolyte liquid
the counter electrode layer, and redox angle 2) Cannot be used with large-scale cases with
electrolyte layer. 3) Durable and mechanically robust high cost and efficiency requirements
Perovskite It uses active light harvesting layer based on 20.9 1) Cost-effective fabrication 1) Material degradation occurs quickly when
solar cell the ABX3 crystal structure, also known as 2) Inexpensive to scale up exposed to heat, moisture, or snow
perovskite. 3) Less space required for installation 2) The substance is toxic by nature
4) High efficiency
Polymer It contains a layered structure consisting of a 3–8 1) Easy to size and shape due to its light 1) Extremely low efficiency
transparent front electrode, an active layer, weight and flexibility 2) Degradation is faster when exposed to
and a back electrode. 2) Easily storable and transportable outdoor conditions
3) Ecologically friendly due to being organic
in nature
Source: T.M. Razykov, C.S. Ferekides, D. Morel, E. Stefanakos, H.S. Ullal, H.M. Upadhyaya, “Solar photovoltaic electricity: Current status and future prospects,” Solar Energy,
vol. 85, no. 8, pp. 1580–1608, Aug. 2011. (https://doi.org/10.1016/j.solener.2010.12.002)

ieee power & energy magazine

81
 50
47.1%
45 III-V Multijunction
Concentrator Solar Cells
40 III-V on Si (Two-Terminal)
35.9%
Solar Cell Efficiency (%)

35 Perovskite on Si
29.5%
30 26.7% Monocrystalline Si
25.5%
25 24.4% Multicrystalline Si
23.4%
20 21.0%
CIGS
15 15.2% CdTe
10 Perovskite
5 Organic
0
1993 1995 1997 1999 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 2021
Year

figure 6. The development of laboratory solar efficiencies of multijunction cells, crystalline modules, thin films, and
some new solar modules over time. CIGS: copper indium gallium selenide; CdTe: cadmium telluride. (Source: Fraunhofer
Inst. for Solar Energy, “Photovoltaics report,” Jul. 27, 2021; used with permission.)

Si Monocrystalline Cell (79 cm2) 26.7

Crystalline Si

Si Monocrystalline Module (13,177 cm2) 24.4

Si Multicrystalline Cell (268 cm2) 24.4

Si Multicrystalline Module (14,818 cm2) 20.4

CIGS Cell (1 cm2) 23.4

Thin Film

CIGS Module (841 cm2) 19.2

CdTe Cell (1 cm2) 21

CdTe Module (23,573 cm2) 19

New Concept

Perovskite Cell (1 cm2) 22.6

Perovskite Module (804 cm2) 17.9

0 5 10 15 20 25 30
Efficiency η (%)

figure 7. The solar PV efficiencies of multijunction cells, crystalline modules, thin films, and some new solar modules.
(Source: Fraunhofer Inst. for Solar Energy, “Photovoltaics report,” Jul. 27, 2021; used with permission.)

82 ieee power & energy magazine may/june 2022

introduced government subsidies, raised Mathematically, the fill factor is de- mum power generated by the panel
public awareness, and invested in re- fined as the ratio of the product of the under standard test conditions) and Pin
search and development. In the 1990s, maximum current value (I m) and the (the solar radiation, i.e., radiant energy
Japan saw a ten-fold increase in the mar- maximum voltage value (Vm) divided emitted from the Sun).
ket as Germany initially weakened, but by the product of the short circuit cur- According to the data obtained from
as a result of modifying subsidies, pro- rent (I SC) and the open circuit voltage the solar efficiency (Table 1), the ef-
duction increased forty-fold, surpassing (VOC) . The value of the conversion ef- ficiencies of multijunction cells and
Japan’s success. Their example has been ficiency is the ratio of Pm (the maxi- crystalline modules outperform those
followed by other European countries,
such as Spain, and have achieved simi-
lar growth rate.
800
There have been a considerable num-
Rest of World
ber of levels of development in the PV
Cumulative Installed PV Capacity (GWp)
700 India
market over the past several decades, Japan
including research and product devel- 600 North America
opment for tasks like lighting, desalina- Germany
tion, and pumping. Thus, it has become 500 Rest of Europe
interesting not only for scientists and China
400
engineers but also for the general pub-
lic. Solar energy is now an inexpensive, 300
highly efficient, and clean alternative to
fossil fuels as well as more easily ac- 200
cessible power supplies to meet an ever-
increasing demand. 100

0
Solar PV Modules— 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
Performance and Year
Efficiency
Solar cell performance depends on the figure 8. The global cumulative PV installation by region. (Source: Fraunhofer
fill factor and conversion efficiency. Inst. for Solar Energy, “Photovoltaics report,” Jul. 27, 2021; used with permission.)

100
Percentage of Total Production (MWp)

80

60

40

20

0
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020

Year

Rest of World North America Europe Asia

figure 9. PV module production by region, 1997–2020. (Source: Fraunhofer Inst. for Solar Energy, “Photovoltaics report,”
Jul. 27, 2021; used with permission.)

may/june 2022 ieee power & energy magazine 83

 of thin films and some new solar mod-
160 ules. Figures 6 and 7 illustrate the data.
140 Rest of World North America Europe Asia
Monocrystalline solar cell efficiency is
Global Annual Production (GWp)

about 4.3% higher than that of polycrys-

120 talline, while copper indium gallium
selenide (CIGS) is about 2.4% higher
100
than that of cadmium telluride (CdTe).
80 Table 1 highlights the key features and
efficiency levels of different solar PV
60 technologies as well as the pros and
40 cons of each implementation.

20 Solar PV Modules—
0 Production and
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 Installation
Year At the end of 2020, China had the most
installed solar PV power, followed by
figure 10. The global annual PV module production by region. (Source: Europe and North America, as shown
Fraunhofer Inst. for Solar Energy, “Photovoltaics report,” Jul. 27, 2021; used with in Figure 8. The chart in Figure 9
permission.) shows the global production rate of

Calling PES
Members!
Save $130 now!
During early bird registration
for the 2022 IEEE PES General Meeting
Early bird rates end 10 June 2022

pes-gm.org
Digital Object Identifier 10.1109/MPE.2022.3161855

84 ieee power & energy magazine may/june 2022

 table 2. Annual gross generation of power by source in India.
Fuel Category 2013 2014 2015 2016 2017 2018 2019 2020 2021****
Coal* 145,273 164,636 185,173 192,163 197,171 200,704 204,224 205,135 209,810
Gas 21,782 23,062 24,509 25,329 24,897 24,937 24,937 24,955 24,924
Diesel 1,199 1,199 993 838 838 638 510 510 510
Nuclear 4,780 5,780 5,780 6,780 6,780 6,780 6,780 6,780 6,780
Hydro** 44,335 45,323 47,057 48,858 49,779 49,992 50,047 50,382 51,351
Wind 21,043 23,354 26,777 32,280 34,046 35,626 37,279 37,694 40,083
Solar 2,632 3,744 6,763 12,289 21,651 28,181 32,578 34,628 49,347
Biomass*** 7,510 7,805 8,110 8,296 8,839 9,242 9,946 10,023 10,610
Total 248,554 274,903 305,162 326,833 344,001 356,100 365,891 370,127 393,415
* Includes lignite. ** Includes small hydropower. *** Includes biomass power/cogeneration and energy-from-waste.
**** Preliminary data based on 31 December 2021.
[Source: Central Electricity Authority, CEA (2019a), “All India installed capacity,” January 12, 2022 (https://www.cea.nic.in).]

solar modules over 20 years. Europe, is increasing and prices are coming G. Boyle, Renewable Energy: Pow-
Asia, and North America have made down. As discussed earlier, solar cells er for a Sustainable Future, 2nd ed.
significant investments in the produc- are being developed that have almost Oxford: Oxford Univ. Press, 2004.
tion of solar modules, while China has 40% efficiency. We also know that Fraunhofer Inst. for Solar Energy,
made a significant increase in produc- when demand and production go up, “Photovoltaics report,” Jul. 27, 2021.
tion. Data in the rest of the world indi- prices come down. Therefore, we can p&e

cates a gradual increase in production conclude that in the coming future, so-
compared to 2010. Figure 10 shows lar cells can emerge as the single larg-
that over the past 10 years, Asia has est supply system of energy just as coal
continued to increase solar PV pro- emerged in the 19th century during the
duction toward the goal of reducing its Industrial Revolution. The future looks
annual carbon footprint and environ- bright for solar PV technologies.
mental pollution. It has been followed
by other Asian countries like India and Acknowledgements
Japan, as shown in Figure 8. The authors are thankful to the Head
India’s installed generation capacity of Electrical Engineering Department,
is comprised of 365 GW, coal (55.8%), Techno Main Salt Lake, for providing
hydro (13.7%), wind (10.1%), solar PV the necessary opportunity for this re-
(8.8%), natural gas (6.8%), bioenergy search project. They are also thankful
and waste (2.7%), nuclear (2%), and to all faculty members of the depart-
oil (0.1%). India has rapidly increased its ment for their constant support.
installed solar capacity from 2.63 GW
in 2013 to 32.5 GW in 2019. Even For Further Reading
though renewable sources (wind, so- C. F. Naff, Solar Power, 1st ed. Farm-
lar, and bioenergy) accounted for 8% ington Hills, MI, USA: Thomson Gale,
of generation in 2017, they accounted 2007.
for 21.8% of installed capacity. Annual L. A. Lamont, History of Photovol-
gross generation of power by source in taics. Abu Dhabi, UAE: Elsevier, 2012.
India is shown in Table 2. There is a J. Schaeffer, Solar Living Source
huge increase in solar power plants as Book. Hopland, CA, USA: Gaiam Real
concluded from Table 2. Goods, 2008.
National Centre for Photovoltaic
Conclusion Research and Education (NCPRE).
As the research and development of PV [Online]. Available: http://www.ncpre.
cells are in progress, their efficiency iitb.ac.in/ncpre/about/profile.html
 book review

markets for flexibility

it don’t mean a thing if it ain’t got that swing

T
THIS ISSUE’S “BOOK REVIEW” economic dispatch mechanisms. How-
column discusses A New Swing-Con- ever, the result is higher flexibility for
tract Design for Wholesale Power providing supply-demand balance,
Markets, written by Leigh Tesfatsion. given the more flexible form of swing
This reviewer writes, “this book is contracts relative to current markets.
timely and topical, given the increas- The cleared contracts can be re-
ing amount of variable and distributed munerated based on a two-component
energy sources penetrating markets model, namely, an offer price (to recover
worldwide.” ex ante avoidable fixed costs to ensure
the availability of relevant power paths
A New Swing-Contract when required) and a performance
Design for Wholesale payment (to recover ex post and after
Power Markets actual performance verification of op-
By Leigh Tesfatsion erational costs to deliver relevant power
Leigh Tesfatsion’s latest book proposes a paths). This could more readily ensure
new wholesale electricity market design market players’ recovery of their dif-
paradigm, with a focus on North Amer- ferent types of costs, from investment
ica, for centrally managed electricity (e.g., instead of, but aligned with capac-
markets that are based on the novel con- ity mechanisms and capacity markets)
cept of “swing contracts.” Tesfatsion, a to operational, quasi-fixed costs (e.g.,
professor at Iowa State University, has fied physical attributes (e.g., location, instead of make-whole payments for
been working on electricity market de- exercise time, ramp rate limits, and binary-commitment cost recovery).
sign and agent-based computational minimum/maximum production lev- The book is timely and topical, given
economics for many years. Her exten- els) and economic attributes (for pric- the increasing amount of variable and
sive academic, research, and industry ing purposes). Dispatchable resources distributed energy sources penetrating
experiences, as well as her training in offer such power paths into multi- markets worldwide. The proposed con-
both economics and mathematics, make ple, centrally managed cepts and paradigm are
her highly qualified to provide fresh markets linked across extremely refreshing
thinking in this important area. different timescales The book and novel. At the same
The book proposes the adoption (from years ahead, for time, they build on the
of swing contracts between the sys- adequacy purposes, to contains rigorous status quo of current
tem operator and relevant resources
to provide flexibility (“swing”). This
minutes ahead, for re-
al-time balancing). The
mathematical electricity markets in
the United States. This
flexibility is represented in the form system operator clears descriptions of presents clear oppor-
of future availability of dispatchable different markets with tunities and p­ athways
power paths (“reserve”) with prespeci- a process similar to and the proposed for a smooth transition
consistent with current
security-constrained
mechanism and toward centrally man-
aged electricity markets
implementations.
Digital Object Identifier 10.1109/MPE.2022.3150811
Date of current version: 19 April 2022 unit commitment and with high penetration

86 ieee power & energy magazine may/june 2022

of variable renewable energy sources, mechanism and implementations. It
storage, and flexible demand. assumes a good knowledge of power
Although the focus is on centrally system operation (particularly optimal
managed markets with unit-commit- power flow and unit commitment), funda-
ment functions, the general concepts mentals of power system economics and
presented may also be extended to other electricity markets, and basics of linear s Meets FAA Specifications!
markets. The core thinking behind the and mixed-integer linear programming. Color – Size – Shape!
concept of swing contract is the valuation As such, the book may be most appeal- – International Orange
s Tested and approved by
of different types of flexibility that dif- ing to knowledgeable postgraduate major power companies!
ferent resources could of- students, researchers, and s Thousands still in service
after 40 years
fer into different markets. instructors as well as mar-
This is a key topic of dis-
The proposed ket designers and regula-
s Universal attaching! Fits any
wire .1" to 1"!
cussion around the world. concepts and tory support staff. It could s Installs in 5 minutes!
s Withstands hail!
In a swing contract, the be adopted in advanced s No maintenance! Does not
key product to be sold paradigm are courses of electricity mar- slip, oscillate, chafe, cause
electrolysis or harmonic
is a power paths-based kets, although some tutori-
“reserve,” rather than
extremely al-like chapters could also
vibration.
s Ships in halves nested.
multiple products such as refreshing be used for introductory 9, 12, 20, 24, 30, and 36" balls
& special sizes available
energy or capacity. Given courses to power system
this outlook, the physical and novel. operation and econom- Call now 573-796-3812
and economic construc- ics. Despite being United ext. 2001
tion principles of a swing contract rep- States centric, general concepts may be
Fax 573-796-3770
resent a solid framework for evaluating suitable for a more universal, international www.tanawiremarker.com
flexibility in different contexts. audience interested in the topic. TANA WIRE MARKERS
The book is composed of 20 chapters I commend the author for accom- P.O. Box 370, California, MO 65018
and an appendix that comprehensively plishing such a challenging task. It is
cover the subject matter. The topics range a book that proposes a coherent and
from the motivation to change current systematic redesign of well-established 634276_Tana.indd 1 23/03/13 10:38 AM

We want
market structures to modeling differ- market mechanisms while dealing with
ent types of swing contracts, potential the pressing issue of reliable and afford-
implementations, and examples of tran-
sitions from the current situation. As-
able integration of variable, renewable
and distributed energy resources. Final- to hear
pects such as the inclusion of distributed
energy resources and platforms to vali-
ly, students and researchers would ben-
efit if more worked examples, data, and
from you!
date the proposed mechanisms are also sample code could be made available on
discussed. The general tutorial chapters the book’s website.
on wholesale electricity markets provide
a helpful context. —Pierluigi Mancarella
The book contains rigorous math-
 p&e
ematical descriptions of the proposed

IMAGE LICENSED BY GRAPHIC STOCK

Do you like what you’re reading?

Your feedback is important.
Let us know—
send the editor-in-chief an e-mail!

may/june 2022 ieee power & energy magazine 87

 letters to the editor

share your thoughts

send comments to [email protected]

R
READERS ARE ENCOURAGED TO
share their views on issues affecting Editor’s Note
the electric power and energy profes- For more discussion of system reliability and potential problems with high-renew-
sion. Send your letters to Steve Wider-
able penetration, please see the March/April 2021 issue of IEEE P&E ­Magazine de-
gren: [email protected]. Letters may
voted to the topic of energy insecurity.
be edited for publication.
Steven Widergren,
Electric System Reliability Editor-in-Chief
I read with interest the many good ar-
ticles about renewable energy integra-
tion in the November/December issue
of IEEE Power & Energy Magazine Because such a wind drought happened not save the system from an extensive
(P&E). As a Member of IEEE with last year, a similar occurrence is likely wind drought as there would likely
more than 40 years of experience in to happen again. A future wind drought be insufficient reserve capacity in the
utility planning in the Midwest, I have could lead to the simultaneous loss of Eastern Interconnection to make up for
dedicated my career to electric system several tens of thousands of megawatts this tremendous loss of capacity.
reliability. During 28–30 January 2020, of wind capacity. At the same time, so- IEEE should also present articles
the Midcontinent Independent System lar capacity may also be reduced due to that focus on system reliability and re-
Operator (MISO) experienced a wind less available solar energy during the al-world potential problems associated
drought, when wind output dropped winter and snow-covered solar panels with high renewables penetration.
to less than 1% of nameplate capacity in at least the northern tier of states. All
for 39 h. The question is, how would this may occur at a time when winter —Karl Kohlrus
MISO supply its load under a future demand may be high due to additional
scenario if such a wind drought were to heating load and, in the future, addi- Guest Editor’s Response
occur again? Consider the following (in tional electric vehicle charging. Yes, there may be times when there are
addition to reduced wind output): Such a wind drought could not only lower levels of wind or solar generation
✔✔ little or no solar output (shorter affect MISO but also the Southwest than demand. Demand will be higher
daylight hours, low sun angle, Power Pool and PJM (an independent in the future with increased electrifica-
snow-covered panels, and cloudy system operator in the United States) tion, and we need to study future sce-
weather) as well. The loss of this much capacity narios with weather data to understand
✔✔ exhausted battery storage over an extended period of time could the impacts and potential mitigations.
✔✔ increased load due to additional cause energy storage to be exhausted A new National Renewable Energy
electric heat load and lead to a widespread blackout of Laboratory study examines this. Weath-
✔✔ reduced imports from other re- a large portion of the Eastern Inter- er systems tend to be approximately
gions facing similar conditions. connection. Such a blackout would be 1,000–3,000 km in size. So whether it is
A winter wind drought would be the catastrophic, not only in terms of loss an extreme weather event or an extended
Achilles’ heel of electric system reli- of life but also economic impact. period of low wind and solar, the trans-
ability with high renewable penetration. I realize the political winds are to mission needed to ­provide ­increased
Digital Object Identifier 10.1109/MPE.2022.3150812
advance renewables, but even a “cop-
Date of current version: 19 April 2022 per sheet” transmission system could (continued on p. 90)

88 ieee power & energy magazine may/june 2022

calendar

PES meetings
for more information, www.ieee-pes.org

T
THE IEEE P OW E R & E N E RGY September 2022 IEEE PES Asia-Pacific Power and
­Society’s (PES’s) website (http://www. IEEE International Conference on Energy Engineering Conference
ieee-pes.org) features a meetings section, Power Systems Technology (PowerCon (APPEEC 2022), 20–23 November,
which includes calls for papers and 2 0 2 2) , 14 –16 Sept emb er, Kua la , Melbourne, Australia, contact Shama
additional information about each of Lumpur, Malaysia, contact Zuhaina Islam, [email protected]
the PES-sponsored meetings. Please Zakaria, [email protected], https://
check the conference website for the attend.ieee.org/powercon-2022/ December 2022
most current information. IEEE Sustainable Power and En-
IEEE International Smart Cities Con- ergy Conference (iSPEC 2022), 4–7
May 2022 ference (ICS2 2022), 26–29 September, December, Perth, Australia, contact
IEEE PES Transactive Energy Sys- Pafos, Cyprus, contact Georges Zissis, Julius Susanto, julius.susanto@gmail.
tems Conference 2022 (TESC 2022), [email protected], com
3–5 May, virtual event, contact Ron https://attend.ieee.org/isc2-2022/
Melton, [email protected], https:// January 2023
ieee-tesc.org/ October 2022 IEEE PES Joint Technical Com-
IEEE PES Innovative Smart Grid mittee Meeting (JTCM 2023), 8–12 Jan-
June 2022 Technologies Europe (ISGT Europe uary, Jacksonville, Florida, United States,
IEEE Transportation Electrification 2022), 10–12 October, Novi Sad, Ser- contact Dan Sabin, [email protected]
Conference & Expo (ITEC 2022), bia, contact Bane Popadic, bane@uns.
15–17 June, Anaheim, California, United ac.rs, https://ieee-isgt-europe.org/ IEEE PES Innovative Smart Grid
States, contact Rebecca Krishnamurthy, Technologies (ISGT 2023), 15–19
[email protected], https:// IEEE PES Generation Transmis- January, Washington, DC, United States,
itec-conf.com sion and Distribution Conference and contact Kathy Heilman, kathy.heilman@
Expo Latin America (GTD LA 2022), ieee.org
July 2022 20–22 October, virtual event, contact Is-
IEEE PES General Meeting (GM rael Troncoso, [email protected], April 2023
2022), 17–21 July, Denver, Colorado, https://ieee-tdla2022.org/ IEEE PES Grid Edge Technologies
United States, contact Roseanne Jones, Conference and Expo (Gride Edge),
[email protected], https://pes-gm. November 2022 24–27 April, San Diego, California,
org/ IEEE PES Innovative Smart Grid United States, contact Kathy Heilman,
Technologies Conference Asia (ISGT [email protected]
August 2022 Asia), 2–6 November, Singapore, con-
IEEE PES/IAS PowerAfrica Con- tact Naayagi Ramasamy, Naayagi. May 2023
ference (PowerAfrica 2022), 22–26 [email protected] IEEE PES International Confer-
August, Kigali, Rwanda, contact Saman- ence and Exposition (GT&D Tur-
tha Niyoyita, niyoyitasamantha@gmail. IEEE Electrical Energy Storage Ap- key), 22–25 May, Istanbul, Turkey,
com, https://ieee-powerafrica.org/ plication and Technologies Conference contact Omer Usta, [email protected],
(EESAT 2022), 8–9 November, Austin, https://ieee-gtd.org/
Digital Object Identifier 10.1109/MPE.2022.3150816
Texas, United States, contact Christo-
Date of current version: 19 April 2022 pher Searles, [email protected] (continued on p. 90)

may/june 2022 ieee power & energy magazine 89

 in my view (continued from p. 92)

Into the Future Alexey Ivakhnenko. Advancements in However, a few developments in the
If the historical cycle holds, the power electronics led to better computation early years of the new century led to
industry is set for another major dis- power and memory in the deep learning revo-
ruption within the next 40 years as the the 1980s, paving the lution as we know it
push to decarbonization ramps up. I way for the develop-
If the historical today. Geoffrey Hinton
would argue that three key technolo- ment of practical neu- cycle holds, the and his team invented
gies will enable the paradigm shift in ral networks. Powered methods to train mul-
the power industry toward 100% re- by the backpropaga- power industry tilayer deep networks,
newable power systems. tion algorithm, neu- one layer at a time, cut-
ral networks reached
is set for another ting training times sig-
AI a milestone in 1986 major disruption nificantly. Moore’s law
AI as we know it today was put on the by predicting the next of exponential growth
map by Alan Turing in his seminal arti- word in a sequence. within the in computational per-
cle in the 1950s, in which he argued that By the end of the 20th formance and memory
machines could think and reason like a century, several practi-
next 40 years began to economically
human. The first proof of concept for AI cal research areas were as the push to deliver enough comput-
was showcased when Logic Theorist, a established that fo- ing power and memory
computer program designed to solve cused on solving prac- decarbonization to enable the spread of
problems, was presented at a 1956 con- tical problems and not AI applications. Col-
ference cohosted by John McCarthy, an necessarily building a
ramps up. lected and consumed
American computer and cognitive sci- machine “that could data became truly big
entist. Research programs over the fol- reason.” This is how new terminology, and passed the zettabyte threshold by
lowing years gradually established AI such as machine learning, deep learn- 2010. Andrew Ng figured out how to
in general and artificial neural networks ing, data mining, and data science was use GPUs to speed up model training
in particular as a scientific discipline. established in the day-to-day vocabu- by orders of 100. These advancements
The first functional, and arguably lary. Despite their sometimes-distinct led to the deep learning “big bang”
deep, neural network was published in differences, the terms are often used in 2011 when, for the first time, a fast
1965 by the Ukrainian mathematician interchangeably. GPU implementation of convolutional

calendar (continued from p. 89)

July 2023 For more information on additional tech- meetings-and-conferences/conference
IEEE PES General Meeting (GM nical committee meetings, webinars, -calendar.
2023), 16–20 July, Orlando, Florida, and events, please visit our IEEE PES  p&e
United States, contact Roseanne Jones, calendar: https://www.ieee-pes.org/
[email protected]

letters to the editor (continued from p. 88)

i­ nterconnectivity with resources across instead of 100% clean electricity) and/ future grid powered by wind and solar
the entire United States can offer sig- or using renewably generated fuels or may not be extreme weather, but simply
nificant benefits to make up for genera- other long-duration storage resources. a calm, cloudy day.
tion losses in one or more regions. There are many challenges to overcome —Debra Lew
The other options could include to reach clean, affordable, and reliable Nov./Dec. 2021 Guest Editor
continuing to use gas plants for capac- electricity, but there are also many po-
p&e
ity needs (for example, targeting 90 tential solutions. The biggest risk to a 

90 ieee power & energy magazine may/june 2022

neural networks outperformed humans 5G Networks DDL. I characterize the future grid as
in visual pattern recognition. Humanity’s obsession with connec- the “digital grid,” a physical electrical
tivity is nothing new. From the Royal grid that is massively interconnected,
DDL Road system developed by the Persian is decentralized, and grows smarter as
The most recognized application of King Darius I in the fifth century BC it integrates with 5G networks, DDL
DDL today is cryptocurrency. How- to connect his vast empire, to Samuel and AI.
ever, DDL can potentially be applied Morse’s first wired telegraph message 5G networks make it technically
for other far-reaching uses. DDL is a sent in 1844, and from the invention of possible and economically feasible to
ledger that is distributed (many iden- the cellular mobile phone in 1973 by push the IoT to a new level. 5G adds
tical copies exist across a network), Martin Cooper to the introduction of near-real-time (with a delay as short
decentralized (no single entity is in the World Wide Web by Tim Berners- as 1 ms) visibility and control to every
control), trustless (there is no need to Lee in 1991, this obsession has fueled “thing” connected to the “last mile” of
trust others), and immutable (it is very phenomenal social and economic a digital grid. The grid finally becomes
difficult to tamper with records). The change. Society’s pursuit for connec- inclusive. AI will make it possible to
1991 creation of a secure digital led- tivity was extended to man-made ma- extract actionable knowledge from
ger whose time stamps cannot easily chines when two computers “talked” to massive data collected from loads,
be tampered with is credited to Stuart each other in a Massachusetts Institute generators, and energy storage units.
Haber and Scott Stornetta, who at the of Technology lab in 1965. Data will finally become an asset and
time were working at Bell Communi- The turning point for human and not a liability. DDL will enable count-
cations Research. However, a white machine connectivity, however, was in less decentralized, automated, peer-
paper published on the Bitcoin block- the 2010s when the speed and power to-peer trustless transactions at every
chain in 2009 in response to the 2008 of long-term evolution cellular net- layer of a digital grid. You will not
financial crisis under the pseudonym works connected the masses, and the need to be a professional to take part in
of Satoshi Nakamoto kick-started the IoT vision became mainstream. On the energy trading.
DDL revolution. shoulders of several net- To hit the 100% renewable electri-
In 2010, the first ap- working enhancements cal energy target, the grid must fun-
plication of DDL was Today, in 2019, the first 5G damentally change to rely on mostly
born in the form of Bit- network was launched. variable renewable resources while
coin cryptocurrency. decentralized A 5G network enhances maintaining a high standard of reliabil-
With the introduction applications the capacity, speed, la- ity. To reach this goal, energy transition
of smart contracts in tency, and reliability entrepreneurs, early adapters, and for-
blockchains in 2015, that are built of preceding networks ward-thinking grid operators around
the era of decentral- by significant orders of the world are pushing the boundaries
ized trustless applica- on DDL magnitude at a reason- by experimenting with unconventional
tions started. In 2017, platforms able cost. Supplemen- products and services. Peer-to-peer mi-
new protocols were tary protocols, like nar- crotransactions of energy, cloud energy
created to bridge real- have started rowband IoT, have made storage, renewable energy as a service,
world data sources off it possible to bring low virtual power plants, and retail renew-
chain with on-chain to emerge in power, low latency, low able energy contracts with no limits
applications, allowing sectors other costs, high range, and on energy use (similar to unlimited
DDL to be expanded reliability to the con- data plans for mobile phones) are a
beyond the financial than finance, nectivity of things. The few examples. Such innovations have a
sector. Today, decen- advent of the 5G New much higher chance of being adopted
tralized applications including Radio protocol means beyond experimentation and will in-
that are built on DDL the electrical that realizing the mas- fluence widespread change if a digital
platforms have started sive interconnectivity- grid materializes. The reason is that a
to emerge in sectors energy sector. of-things vision is on digital grid is capable of facilitating a
other than finance, in- the horizon as it sup- massive number of micro- and macro-
cluding the electrical energy sector ports a wide range of requirements for transactions locally and regionally, in a
in particular (such as microtransac- data transfer, power, coverage, latency, decentralized, automated, secure, and
tions for local energy market trad- reliability, cost, and capacity. intelligent manner, among distributed
ing) and the Internet of Things (IoT) variable renewable generators, energy
in general (e.g., manufacturing au- A Digital Grid storage units, and active loads to de-
tomation, e-mobility, and supply I think the next disruption will be fu- liver reliable electricity.
chain management). p&e
eled by the trio of 5G networks, AI, and 

may/june 2022 ieee power & energy magazine 91

 in my view
Hamidreza Zareipour

digital power grid

lessons from history to imagine the future

M
MUCH HAS CHANGED IN THE U.S. Federal Power Commission, the By the late 1970s, the power indus-
power industry since the inception of predecessor to the U.S. Federal Energy try was set for another disruption driv-
commercial electrical power systems. Regulatory Commission, were born in en by energy security and air pollution
What has not changed is the cycle of this era to oversee and regulate the util- concerns. The latter led to the Clean
disruption and adaptation empowered ity business. Technological innovations Air Act of 1963 and put thermal power
by innovation. The next disruption is led to the “kingdom of coal” as ac pow- plants in the spotlight. The former,
on the horizon and is fueled by the push er plants became safer and bigger. Coal rooted in the war of 1973 in the Middle
to decarbonize the grid. In this article, was so loved that even hydropower was East, created the U.S. Public Utility
I argue that the next disruption will be called the white coal. Regulatory Policies Act (PURPA) of
shaped by three technologies: artificial The next major disruption, in my 1978. PURPA had two key elements
intelligence (AI), 5G networks, and opinion, occurred when the “kingdom that, in my opinion, shaped the power
distributed digital ledger (DDL). Even of coal” was shaken by two new power industry for the following 40 years.
though historical elements in this ar- generation technologies: gas and nucle- First, it allowed qualified nonutility
ticle are drawn from the evolution of ar. Both were born out- generators to enter the
the power industry in the United States, side the power industry. power business. This
parallels can be found elsewhere. Gas-fired power gener- The legend was a key step toward
ation, a gift of the avia- the introduction of com-
A History of Disruption tion industry, was de-
is that, in petition in electricity
The first major disruption in the power veloped in 1939 to help 1908, Thomas generation and the birth
industry was shaped by the so-called win the war in the air. of competitive electric-
war of currents. Within 20 years since Nuclear-powered elec- Edison sent a ity markets, disrupting
the commissioning of the first dc power tricity, experimented Insull’s monopoly utili-
station in New York, ac dominated the with in 1951, came from
message to the ty model. Second, PUR-
industry. The legend is that, in 1908, redirecting knowledge son of William PA explicitly qualified
Thomas Edison sent a message to the from weapons research wind- and solar-powered
son of William Stanley, the inventor of to peaceful uses. The Stanley, the electricity generation to
transformers, saying “I was wrong!” boom of electricity was enter electricity markets.
As the 20th century began, the first now in full swing, and
inventor of This marked the begin-
40 years could be characterized by in- society adapted to this transformers, ning of the wind and
novation in business, regulation, and new form of energy. solar industries as we
technology to adapt to and spread ac Within 40 years, elec- saying “I was know them today. Fast
power. The regulated monopoly busi- tricity reached almost forward to today, and
ness model was born in 1907 when every neighborhood in
wrong!” thanks to many inno-
Samuel Insull, a Chicago businessman, America, as did air pol- vations in technology,
convinced regulators that a vertically lution and heavy dependence on fossil business, and regulatory fronts, we
integrated power utility was the ef- fuels. Perhaps the most recognizable have adapted to the variability of wind
fective way to bring electricity to the television advertisement in the 1960s and solar and have learned how to trade
masses. Regulatory bodies, such as the was the “Live Better Electrically” cam- electrons like tomatoes!
Digital Object Identifier 10.1109/MPE.2022.3150817
paign, featuring Ronald Reagan, who
Date of current version: 19 April 2022 later became the 40th U.S. president. (continued on p. 90)

92 ieee power & energy magazine may/june 2022

advertisers index
The Advertisers Index contained in this issue is compiled as a service to our readers and advertisers: The
publisher is not liable for errors or omissions although every effort is made to ensure its accuracy. Be sure
to let our advertisers know you found them through IEEE Power & Energy Magazine.

Company page# URL Phone

ASPEN CVR4 www.aspeninc.com +1 650 347 3997

Bigwood Systems, Inc 10 www.bigwood-systems.com

DIgSILENT GmbH 3 www.digsilent.de

EasyPower LLC 11 www.EasyPower.com/demo

Neplan AG 9 www.neplan.ch

Manitoba Hydro International ltd 7 www.pscad.com +1 204 989 1240

P&R Technologies 79 pr-tech.com +1 800 722 8078

PowerSys/EMTP 85 www.emtp.com/ieeepes

PowerTech Labs 6 www.dsatools.com

PowerWorld Corporation 13 www.powerworld.com +1 217 384 6330

RTDS Technologies 5 www.rtds.com +1 204 989 9700

SKM Systems Analysis, Inc. CVR2 www.skm.com +1 800 500 4SKM

Tana Wire Markers 87 www.tanawiremarker.com +1 573 796 3812

445 Hoes Lane, Piscataway, NJ 08854

IEEE power & energy magazine representative

Beverly Grimme
Naylor Association Solutions
direct: +1 352 333 3367, cell: +1 904 881 0862, fax: +1 703 790 9199
[email protected]

Digital Object Identifier 10.1109/MPE.2021.3133043