IEEE - 45670
Load Forecasting
by Group Method of Data Handling
Vaibhav Vaishnav Jayashri Vajpai
Department of Electrical Engineering Department of Electrical Engineering
Indian Institute of Technology M. B. M. Engineering College
Jodhpur, India Jodhpur, India
[email protected] [email protected] Abstract— In today’s competitive deregulated market, and distribution tasks to different utilities and heralded the
forecasting load demands and electricity sales prices is an entry of private and small players in the electricity market.
important task for both, government and private electric utilities, Today, an electric utility can buy/sell power through trading,
as it helps them to take important decisions regarding load power exchange and bilateral exchange.
switching, load shedding, voltage control, network
reconfiguration, energy purchasing, contract evaluation, fuel The deregulation of energy markets has highlighted the
purchase and power infrastructure development. A variety of need for accurate forecasts. The total power required in a State
time series based techniques have been developed for accurate cannot be fully generated by the State Electricity Corporation
forecasting of load despite its nonlinear and dynamic nature. itself. Hence, it purchases power from the Central Grid/Private
With many hidden patterns and correlations in bulk volumes of Companies/other States having surplus power. For purchasing
data, modern load time series forms an important part of big power the Schedule has to given by the buyer in 96 time slots
data analytics. In this paper, a novel self organizing machine for upcoming day/month/year. Over drawl from the given
learning technique called Group Method of Data Handling schedule can be made at a lesser cost if the grid frequency is
(GMDH) has been used for time series based monthly peak load more than 50 Hz at that time whereas penalty has to be paid if
forecasting. The proposed GMDH based model is first it happens when the grid frequency is already lower than 50
standardised using a monthly peak load data time series of a Hz.
regional Chinese grid. Then, the model has been applied to
forecast the peak load demand of Rajasthan state whose results “The charges for the deviations for all the time-blocks are
have been further compared with the ones published by the payable for Over-drawl by the buyer or the beneficiary and
Central Electricity Authority of India to showcase the superiority under-injection by the generating station or the seller and
of the proposed technique. receivable for Under-drawl by the buyer or the beneficiary and
over-injection by the generating station or the seller” and is
Keywords—GMDH, Big Data Analytics, Data Mining, Monthly worked out on the average frequency of a time-block at the
Peak Load Forecasting, Artificial Intelligence. rates specified in the grid code [1]. Hence, a good forecast is
very important to efficiently and profitably purchase power
I. INTRODUCTION from external agents. Moreover, it helps utilities to take
Load forecasting refers to the prediction of the load important decisions regarding energy purchasing, load
behaviour for the future. As per Indian Electricity Grid code switching, load shedding, contract evaluation, fuel purchase,
[1], Electrical Load means MW, MWh, MVAR, MVARh commissioning of new generation plants, etc. The popular
consumed by an installation or utility. Words such as, power forecast techniques are often based on time series analysis.
demand and energy consumption are also used for electric This paper presents following sections. Section II describes
load. the State of Art, Section III briefly introduces the GMDH
technique, Section IV discusses the proposed methodology,
Ever since the development of electrical supply systems, Section V describes the standardisation of proposed model,
majority of governments practiced a regulated and Section VI explains the application of the model to forecast a
monopolistic power industry where generation, transmission load time series and finally, last Section provides the
and distribution were managed by a single utility company. conclusion.
But with time, improved materials reduced the machine
building cost and technological innovations improved the II. STATE OF ART
efficiency of small units. Computerized control systems
reduced the requirement of working personnel and off site A variety of forecasting methods have been developed
monitoring of units from remote locations became possible. over the past few decades based on forecasting horizon. Load
Consequently, high demand growth and inefficient system forecasting has been classified as short term (minutes to
management caused the electric power industry to be hours), medium term (days to weeks) and long term (months
deregulated in late 90’s. This led to allocation of generation to years). Load forecasting has been carried out in all these
10th ICCCNT 2019
July 6-8, 2019, IIT - Kanpur
Authorized licensed use limited to: Indian Institute of Technology - Jodhpur.Kanpur, Indiaon October 06,2021 at 11:53:32 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IEEE - 45670
categories using several methods, the most common being small groups of arguments rather than their complete set as in
multiple regression, exponential smoothing, iterative other models. Hence, the name Group Method of Data
reweighted least-squares, stochastic time series, autoregressive Handling. The input output relationship is represented by
integrated moving average (ARIMA) models and neural discrete time analogue of Volterra series as follows:
networks.
Among all these models, the non-linear models based on
a ¦ a x ¦ ¦ a x x
yn
M M M
soft computing, particularly, Neural Networks (NNs) have 0 i 1 i i i 1 j 1 ij i j
(1)
gained attention over the past two decades. This is because ¦ ¦ ¦ a x x x ...
M M M
i 1 j 1 k 1 ijk i j k
researchers have achieved remarkable developments in NNs.
unsupervised NN, supervised feedforward NN, supervised
feedback NN and hybrid NN have been used by researchers where x1, x2, x3......xM denote the input variables and a0, ai, aij,
for load forecasting [2]. NNs are able to learn from past and aijk denote the weights of polynomials. Ivakhnenko [5]
generalize to new cases extracting useful information from demonstrated that a self organizing feed forward neural
input variables. They are capable of discovering nonlinear network structure can be modelled by cascading of second
relations in data which makes them a good choice for use in order polynomials generated using pairs of variables. The
electrical demand forecasting, where the linear models suffer weights of the polynomials are obtained using the ordinary
limitations. least squares fitting and the fitness of each polynomial output
is tested using a Regularity criteria. The regularity criteria
However, optimising the neural network models is a big calculates the mean squared error between the actual and
challenge. They cannot be applied blindly in ‘Black Box’ predicted outcome of each node in each layer. The smaller the
mode. Exceptional modelling skills and experience are regularity criteria, better the fit of node polynomial on the
required to select a good NN model. These include selection data.
of number of nodes, layers and proper lagged variables as the
inputs. In simple terms, a neural network is not fully self IV. PROPOSED METHODOLOGY
organized.
The input output relationship of system is modelled using
In modern power systems, addition of advanced metering a multilayer iterative algorithm with polynomials as partial
interface (AMI) and smart sensors have led to a drastic descriptions. The network consists of several layers and each
increase in electricity transmission and distribution data. AMIs layer has a set of nodes. Each node has two inputs and a single
have increased the data recording frequency from once per output [6]. As the polynomial based GMDH algorithm is
month to one reading per 15 minutes. It is estimated that, in being used, each node performs a nonlinear quadratic function
next few years electric utility industry will be supplied by 3 on its two inputs, which is a second order Ivakhnenko
petabytes of data merely from smart meters [3]. Hence, to polynomial given as follows:
handle and fully exploit the information contained in such
bulk amount of data, innovative big data algorithms and g(x1, x2) = a0 + a1x1 +a2x2 + a3x12 +a4x22 + a5x1x2 (2)
sophisticated self organized artificial intelligence based where x1 and x2 are inputs to node ‘g’ and (a0......a5) are
models are required, the models which are competent enough weights of polynomial. Although its structure is analogous to
to grasp the auto correlative relations in a time series, which that of multilayer feed forward neural networks, the number of
can deal with the nonlinear relationship between the effecting layers and the number of nodes in each layer is determined by
factors and the output and enhance the forecasting precision. an external criterion. The number of nodes is equal to the
This paper applies one such model, i.e., the Group Method number of inputs in the first layer and these number of hidden
of Data Handling. This is a self organizing data mining nodes change according the value of external criterion for the
approach to automatically generate multilayer polynomial subsequent layers.
structures, thus forming a network capable of extracting The coefficients are estimated with the least squares
knowledge about an object directly from data samples. The method and regression technique is used to minimise the
expected advantage of this method, in addition to its inherent difference between the predicted outcomes and the actual
capability of self organization is that the method has moderate outcomes. Data set is divided into training and testing sets
memory requirement as compared to the statistical time series sequentially and different training to testing set ratios are used
models. It is important to note that the forecasting error for the to analyze the performance of algorithm on different datasets.
standard Box & Jenkins airline passenger time series has been Training set is used to determine coefficients of polynomial
found to be significantly low in comparison to the prior and testing set is used to select the fittest nodes using
published results [4]. Moreover, the GMDH model has been regularity criterion. As mentioned in previous section, in this
applied on a practical Indian load time series data for medium criterion, the mean square error is determined between actual
term load forecasting for the first time in this paper. and predicted output of each node. If it is less than a threshold
value then node is passed to the next layer. It is expressed as:
III. GMDH TECHNIQUE
GMDH is an inductive learning algorithm where model
2
under investigation is represented using data pairs of multiple N y g
input and single output. The main idea behind GMDH is that R ¦ i i
(3)
¦ y
N 2
i 1
here the combination producers receive as their inputs only i 1 i
10th ICCCNT 2019
July 6-8, 2019, IIT - Kanpur
Authorized licensed use limited to: Indian Institute of Technology - Jodhpur.Kanpur, Indiaon October 06,2021 at 11:53:32 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IEEE - 45670
It also tells the relevance of each input connection to V. MODEL STANDARDISATION
the network and fitness capability of a node polynomial to Dongxiao et. al. [7] have performed the monthly peak load
required system output. The entire procedure is repeated for forecasting on the data set of a regional power grid as shown
each layer unless the stopping criterion is achieved, i.e., unless in Table I. The data set consists of 84 values denoting the
the value of the lowest Regularity Criteria of a layer is not maximum load demand from January 2002 to December 2008.
greater than lowest Regularity Criteria of previous layer. Thus, The time series exhibits an upward trend superimposed with
the polynomial with the minimum error criterion is chosen as seasonal nature as shown in Fig. 2. Peak load has been
the final optimized model. The algorithm for implementing the projected by fitting a suitable time series model and the data
proposed methodology is as follows: set is a good candidate for the task of medium term load
forecasting.
1. Divide the data sample into training set and
testing set comprising of definite percentage of TABLE I. MONTHLY PEAK LOAD DEMANDS OF REGIONAL GRID
(2002- 2008) (MW)
samples.
2. Generate functions for partial models using the Month/
Volterra polynomial. Year 2002 2003 2004 2005 2006 2007 2008
Jan 12675 14118 14409 17053 18524 21609 24538
3. Estimate the weights of partial models using the
least squares fitting and training set Feb 11982 13426 13323 16101 18300 20752 24022
4. Calculate the value of external criterion for partial March 12270 13482 14008 16474 18919 21494 23758
models using testing set.
April 11687 12284 13734 15476 17554 19752 23003
5. Calculate minimum value of external criterion to
reach the best model. May 12045 14004 14778 14918 17453 20030 23049
The proposed algorithm is shown as a flow chart in Fig. 1. June 13566 14151 15695 16817 19972 24368 27014
As per the proposed GMDH based methodology, the next
section discusses one such application of GMDH algorithm in July 15144 15578 18161 19103 21373 26043 26788
the field of load forecasting. Aug 14472 15586 17842 19450 21737 25465 28495
Sep 12812 13469 16479 16517 19202 22238 25459
Start Oct 12398 12693 16055 17482 19556 21096 24952
Nov 13891 14597 16636 19301 21411 23762 27598
Initialize
layer = 1 Dec 13821 15026 16932 19218 22020 25144 28779
Estimate the weights
in the kth layer
using training data
Calculate the outputs
of the nodes in the
kth layer using
the checking data
Yes
Increment Is kth layer
layer set final ?
(k = k+1)
No
Fig. 2. Peak Loads (2002-2008)
The outputs are
set of the inputs
of the nodes in
As shown in Fig.2, the data consists of ascending trend
(k+1)th layer
with a seasonal component. Dongxiao et. al. [7] in their work
have applied the Time Series Seasonal Multiplicative Model
End (TSSMM) for forecasting future demand. They have divided
the data set of 84 samples into two sets. The data from Jan
Fig.1 GMDH Algorithm 2002 – Dec 2007 has been chosen to do fitting and that from
10th ICCCNT 2019
July 6-8, 2019, IIT - Kanpur
Authorized licensed use limited to: Indian Institute of Technology - Jodhpur.Kanpur, Indiaon October 06,2021 at 11:53:32 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IEEE - 45670
Jan 2008 – Dec 2008 for testing results. Following the same VI. APPLICATION
principles, the data set has been divided into Training and
Testing set with 72 and 12 observations respectively, for
The Central Electricity Authority of India (CEA) publishes
GMDH modelling. Mean error (ME), Mean Absolute error
Load Generation Balance Report (LGBR) [8] which provides
(MAE) and Mean Percentage Error (MPE) have been used for
the data for national as well as state wise availability of peak
a thorough comparison. Different ratios of training set to
demand and power (in MW). It also covers monthly
testing set have been used in GMDH modelling to showcase
anticipated energy demand and availability (in MU) thereby
their effect on accuracy of forecasted values. It can be
setting generation targets for upcoming year after proper
observed from Table II, that the results obtained from the
analysis of available power with State Electricity Boards and
GMDH model are far superior than the reported results. The
private companies. It also gives national and state wise
model where 80% of the data was used for training of the
comparison of actual power demand with forecasted power
model and 20% was used for testing purpose, appears to be the
demand.
best among these in terms of all ME, MAE and MAPE due to
appropriate balance of learning parameters.
TABLE III. Monthly Peak Load Demand for Rajasthan State (MW)
Further, it can be seen that the increase or decrease in the
Training to Test Set ratio from 80:20 to 90:10 or 70:30 Month 09-10 10-11 11-12 12-13 13-14 14-15 15-16 16-17
respectively leads to increase in error. Hence, 80:20 is the
April 5364 6159 6452 7143 7512 7977 7798 9027
most optimal and efficient ratio here for obtaining the best
load predictions. May 5971 6821 6800 7511 7799 8844 8577 9690
June 5968 6215 7054 7686 7626 9131 9010 9906
TABLE II. Comparison of TSSMM & GMDH Models
Training to July 6487 6171 6736 7765 7306 9403 8431 9288
S. Model Testing Set ME MAE MPE Aug 6240 5211 6443 7044 7206 10087 10080 7807
No. Ratio (MW) (MW)
Sep 5453 5426 7536 7464 8929 10188 10961 9816
1. TSSMM - -446.42 640.00 -1.71 %
Oct 6087 6275 7627 7454 7899 9339 9453 9046
2. GMDH 1 60/40 -231.47 615.80 -0.82 %
Nov 6143 6361 8188 7671 8627 9525 9780 9902
3. GMDH 2 70/30 -251.18 638.28 -0.90%
Dec 6771 7116 7556 8511 9735 10642 10123 10613
4. GMDH 3 80/20 -122.49 599.90 -0.40 %
Jan 6798 7582 7555 8940 9977 10179 10720 10348
5. GMDH 4 90/10 -231.75 -619.59 -0.83 %
Feb 6859 7729 7779 8396 10047 10095 10190 10527
Mar 6567 7549 7827 8433 9354 8199 9677 9859
It is important to note that the TSSMM model proposed by
Dongxiao et. al. required decomposition of time series into its
trend and seasonal component. It required the calculation of
seasonal factors and the 12 period Moving Average. The trend
and cyclic components were then separated from the moving
average sequence using least square estimation and finally the
forecasting model was obtained by multiplying the trend and
seasonal factors.
The GMDH model overcomes this drawback as no such
exercise is required in its case. With inductive self
organization, the GMDH network structure evolves with the
estimation process and does not demand any prior knowledge
or experimentation to evaluate parameters of a time series.
GMDH structure is in the form of flexible multilayered
network which is determined automatically in the process of
training and is competent enough to grasp the auto correlative
relations in a time series. To effectively prove the above Fig. 3 Peak Loads for Rajasthan State (2009-2016)
statement, the GMDH model has been applied on the recent
Indian load time series data in next section.
The monthly peak load demand data and plot for Rajasthan
state over the period of 8 years as shown in Table III & Fig. 3,
and as mentioned in LGBR (2009-2017) [8] has been collected
10th ICCCNT 2019
July 6-8, 2019, IIT - Kanpur
Authorized licensed use limited to: Indian Institute of Technology - Jodhpur.Kanpur, Indiaon October 06,2021 at 11:53:32 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IEEE - 45670
for case study and a real time application of the developed However, training to testing set ratio of 85:15 and 90:10 have
GMDH model. Out of 96 data samples, 84 samples from given the best results for this data set as compared to the other
2009-2016 have been used as training set and the last 12 have ones. The higher error values are because the time series
been used as testing set for forecasting the peak load demand exhibits different trend and seasonality components. The
for financial year 2016-2017. The results have been compared GMDH forecasts for next six years have also been tabulated in
with the values as forecasted by CEA mentioned in LGBR Table VI. The main advantage of GMDH is that its structure is
2017 as shown in Table 4 in terms of Peak demand (MW) and determined automatically in the process of training and it is
in terms of accuracy criterion in Table V. capable enough to automatically grasp the different relations
in a time series and hence produce best results even when the
substratal laws are too complex to define. It can deal with the
nonlinear relationship between the affecting factors and the
TABLE IV. Actual vs Forecasted peak Demands for 2016-17 (MW)
output, due to which the forecasting precision is increased
Month Actual CEA GMDH 1 GMDH 2 GMDH 4 GMDH 5 manifold.
April 16 9027 8900 9168.81 9171.93 9452.4 10011.6
May 16 9690 9000 9148.38 9185.08 9406.94 9643.86
June 16 9906 9300 9544.72 9561.28 8788.48 9630.68
July 16 9288 8700 10142.8 9567.44 9734.5 9972.05
Aug 16 7807 10500 10319.5 9892.8 9821.95 9961.37
Sep 16 9816 11500 10208.2 10432.7 10022.5 10303.1
Oct 16 9046 10000 9378.36 9470.17 10445.9 9935.43
Nov 16 9902 10300 9711.71 9363.98 10956.1 10277.2
Dec 16 10613 10700 9821.45 9040.66 10214.3 9909.5
Jan 17 10348 11300 9722.23 8572.52 10038.1 9896.32
Feb 17 10527 11100 8832.72 8915.87 9300.33 10238.1
Fig. 4. Plot of CEA Forecaster vs GMDH Model
Mar 17 9859 10600 9057.48 9316.15 9698.94 9159.67
TABLE VI. GMDH Forecasts for next six years
Month 2016-17 2017-18 2018-19 2019-20 2020-21 2021-22
TABLE V. Comparison of CEA Forecaster and GMDH Model
April 9452.40 9416.77 10574.8 11811.6 12010.0 12907.3
Model T/T ME MAE MPE MAPE SSE
May 9406.94 9291.77 10213.1 11243.9 11653.6 12550.8
Ratio (MW) (MW)
CEA - -505.91 841.08 -5.68% 9.18% 0.191 June 8788.48 9162.25 11261.2 11552.1 11955.5 12416.9
GMDH 1 60/40 64.386 769.99 -0.14% 8.29% 0.161 July 9734.50 10626.3 11244.0 11612.3 13467.4 14343.0
GMDH 2 65/35 278.00 870.00 2.05% 9.10% 0.163 Aug 9821.95 10118.5 12314.6 11900.4 12342.8 13150.2
GMDH 3 80/20 207.06 988.42 1.29% 10.72% 0.240 Sep 10022.5 10897.5 11538.1 12752.3 13087.0 14284.7
GMDH 4 85/15 -170.00 753.60 -2.40% 8.14% 0.136 Oct 10445.9 11597.7 11799.5 13439.5 14276.4 15407.0
GMDH 5 90/10 -259.15 669.00 -3.36% 7.37% 0.119
Nov 10956.1 11954.1 12182.3 13148.5 14961.8 15866.6
Dec 10214.3 11918.5 12386.4 12916.5 15077.8 16286.4
Hence, it is concluded that GMDH has outperformed CEA Jan 10038.1 11436.8 11695.2 11990.1 14108.2 15665.8
forecaster. It is observed that GMDH has produced better
results in comparison to the ones published by Central Feb 9300.33 10093.0 11923.4 12003.9 12794.5 14037.4
Electrical Authority (CEA) on this recent and practical data
March 9698.94 10774.1 11737.2 11934.9 13228.9 14822.8
set. Different validation criteria, i.e., Mean Error (ME), Mean
Absolute Error (MAE), Mean Percentage Error (MPE) and
Mean Absolute Percentage Error (MAPE) have been used to
test the accuracy of the GMDH algorithm and the results in
Table IV & Table V show that the proposed model is a winner VII. CONCLUSION
on all grounds. Plot of CEA and two best GMDH models has In this paper, GMDH model has been applied on the data
also been shown in Fig. 4 which clearly depicts that the set of Chinese regional Grid over a period of 7 years studied
GMDH models trace the actual variation of peak load by Dongxiao. et. al and the lower MPE values show the
demands much closely as compared to CEA forecaster. accuracy of our technique. It had also been applied on the
10th ICCCNT 2019
July 6-8, 2019, IIT - Kanpur
Authorized licensed use limited to: Indian Institute of Technology - Jodhpur.Kanpur, Indiaon October 06,2021 at 11:53:32 UTC from IEEE Xplore. Restrictions apply.
Downloaded
IEEE - 45670
local and most recent dataset of 2010-2017, to forecast the REERENCES
monthly peak load demand of Rajasthan state for year 2016-17 [1] Indian Electricity Grid Code: Central Electricity Regulatory
and the results show that GMDH model has better output as Commission, New Delhi, homepage, www.cercind.gov.in.
compared the ones forecasted by the Central Electricity [2] Hernandez L., Baladron C., Aguiar J., Carro B., Sanchez-Esguevillas A.,
Authority (CEA) of India. Thus, it can be inferred that Lloret J. & Massana J.: A Survey on Electric Power Demand
GMDH has an upper edge over the general NN models, Forecasting: Future Trends in Smart Grids, Microgrids and Smart
mostly used for short and medium term load forecasting, Buildings. IEEE Communications Surveys & Tutorials, vol.16, no.3,
pp.1460-1495, 2014
which often give irrelevant forecasts in the absence of prior
[3] Zhang, Y., Huang, T. & Bompard, E.F., Energy Informatics, vol.1, no.8,
information of proper architecture, activation function and pp.1-24, 2018
initial values of weights. The ability of GMDH to successfully [4] Vaishnav, V. & Vajpai, J.: Seasonal Time Series Forecasting by Group
overcome most of the NN drawbacks has established it as an Method of Data Handling, Proceedings of 2018 IEEE International
efficient modelling procedure for real world systems. In this Students' Conference on Electrical, Electronics and Computer Science
paper, it was observed that one of the five GMDH models (SCEECS), Bhopal, India, 2018
selected by hit and trial method gave MAPE more than the [5] Ivakhnenko A.G.: Polynomial theory of complex systems, IEEE Trans.
CEA forecast. Hence, in future, the performance of GMDH Syst. Man Cyber, vol.SMC-, no.4, pp.364-378, 1971
model can be improved by combining it with standard [6] Onwubolu G. C.: GMDH Methodology and Implementation in
MATLAB, 1st edn., Imperial College Press., London, 2016
optimisation techniques like genetic algorithm, particle swarm
optimization, radial basis functions, etc. With the smart metre [7] Dongxiao N., Yunyun Z. & Jinpeng L.: The Application of Time Series
Seasonal Multiplicative Model and GARCH Error Amending Model on
network expected to cross 1 billion worldwide by 2022, Forecasting the Monthly Peak Load, International Forum on Computer
GMDH can prove to be an effective approach to process this Science-Technology and Applications, pp. 135-138, IEEE, Chongqing,
highly heterogeneous and dimensional data and convert it into China, 2009
operational intelligence for efficient working of future Smart [8] Load Generation Balance Report [2009-2017], published annually by
Grids. Central Electricity Authority, Ministry of Power, Government of India:
CEA homepage, www.cea.nic.in.
ACKNOWLEDGMENT
The authors are thankful to the GMDH LLC for making
available demonstration version of GMDH Shell 2.6.4.
10th ICCCNT 2019
July 6-8, 2019, IIT - Kanpur
Authorized licensed use limited to: Indian Institute of Technology - Jodhpur.Kanpur, Indiaon October 06,2021 at 11:53:32 UTC from IEEE Xplore. Restrictions apply.
Downloaded
