Internationa l Journal of Systems Science, 2002, volume 33, number 1, pages 23±34

Electric load forecasting: literature survey and classi®cation of

methods

Hesham K. Alfares* and Mohammad Nazeeruddin

A review and categorization of electric load forecasting techniques is presented. A wide

range of methodologies and models for forecasting are given in the literature. These
techniques are classi®ed here into nine categories: (1) multiple regression, (2) expo-
nential smoothing, (3) iterative reweighted least-squares, (4) adaptive load forecast-
ing, (5) stochastic time series, (6) ARMAX models based on genetic algorithms, (7)
fuzzy logic, (8) neural networks and (9) expert systems. The methodology for each
category is brie¯y described, the advantages and disadvantage s discussed, and the
pertinent literature reviewed. Conclusions and comments are made on future research
directions.

1. Introduction has a signi®cant e€ect on power system operations, as

Load forecasting is a central and integral process in the economy of operations and control of power systems
planning and operation of electric utilities. It involves may be quite sensitive to forecasting errors. Haida and
the accurate prediction of both the magnitudes and geo- Muto (1994) observed that both positive and negative
graphical locations of electric load over the di€erent forecasting errors resulted in increased operating costs.
periods (usually hours) of the planning horizon. The Hobbs et al. (1999) quanti®ed the dollar value of
basic quantity of interest in load forecasting is typically improved STLF for a typical utility; a 1% reduction in
the hourly total system load. However, according to the average forecast error can save hundreds of thou-
Gross and Galiana (1987), load forecasting is also con- sands or even millions of dollars.
cerned with the prediction of hourly, daily, weekly and The system load is a random non-stationar y process
monthly values of the system load, peak system load and composed of thousands of individual components. The
the system energy. Srinivasan and Lee (1995) classi®ed system load behaviour is in¯uenced by a number of
load forecasting in terms of the planning horizon’s dura- factors, which can be classi®ed as: economic factors,
tion: up to 1 day for short-term load forecasting time, day, season, weather and random e€ects.
(STLF), 1 day to 1 year for medium-term load fore- A wide variety of models, varying in the complexity of
casting (MTLF), and 1±10 years for long-term load functional form and estimation procedures, has been
forecasting (LTLF). proposed for the improvement of load forecasting accu-
Accurate load forecasting holds a great saving poten- racy. Matthewman and Nicholson (1968) conducted an
tial for electric utility corporations. According to Bunn early survey of electric load forecasting techniques. Abu
and Farmer (1985), these savings are realised when load El-Magd and Sinha (1982) , Bunn and Farmer (1985),
forecasting is used to control operations and decisions and Gross and Galiana (1987) also reviewed load
such as dispatch, unit commitment, fuel allocation and demand modelling and forecasting. Recently,
o€-line network analysis. The accuracy of load forecasts Moghram and Rahman (1989) surveyed electric load
forecasting techniques.
The aim of this paper is to survey and classify electric
load forecasting techniques published since the last com-
Received 17 February 2000. Revised 20 March 2001. Accepted 9 prehensive review of Moghram and Rahman (1989). In
April 2001. comparison with those earlier literature reviews, this
Systems Engineering Department, King Fahd University of
survey not only covers newer papers, but also includes
Petroleum & Minerals, Dhahran 31261, Saudi Arabia. new categories that re¯ect recent research trends. It also
*Author for correspondence. e-mail: hesham@ccse.kfupm.edu.sa provides up-to-date brief verbal and mathematical
International Journal of Systems Science ISSN 0020±7721 print/ISSN 1464±5319 online # 2002 Taylor & Francis Ltd
http://www.tandf.co.uk/journals
DOI: 10.1080/00207720110067421
24 H. K. Alfares and M. Nazeeruddin

descriptions of each category. Load forecasting tech- next step, it uses the maximum of the initial hourly
niques are classi®ed into nine categories. In subsequent forecast, the most recent initial peak forecast error,
sections, one section is devoted to each category, where and exponentially smoothed errors as variables in a
a brief description is given of the technique and a litera- regression model to produce an adjusted peak forecast.
ture review o€ers a representative selection of principal Haida and Muto (1994) presented a regression-based
publications in the given category. Arranged in roughly daily peak load forecasting method with a transforma-
chronological order, the nine categories of load fore- tion technique. Their method uses a regression model to
casting techniques to be discussed are: predict the nominal load and a learning method to pre-
. multiple regression; dict the residual load. Haida et al. (1998) expanded this
model by introducing two trend-processing techniques
. exponential smoothing;
designed to reduce errors in transitional seasons. Trend
. iterative reweighted least-squares; cancellation removes annual growth by subtraction or
. adaptive load forecasting; division, while trend estimation evaluates growth by the
. stochastic time series; variable transformatio n technique. Varadan and
Makram (1996) used a least-squares approach to iden-
. ARMAX models based on genetic algorithms;
tify and quantify the di€erent types of load at power
. fuzzy logic; lines and substations.
. neural networks; and Hyde and Hodnett (1997a) presented a weather-load
. knowledge-based expert systems. model to predict load demand for the Irish electricity
supply system. To include the e€ect of weather, the
model was developed using regression analysis of histor-
ical load and weather data. Hyde and Hodnett (1997b)
2. Multiple regression
later developed an adaptable regression model for 1-
Multiple regression analysis for load forecasting uses the day-ahead forecasts, which identi®es weather-insensitive
technique of weighted least-square s estimation. Based and -sensitive load components. Linear regression of
on this analysis, the statistical relationship between past data is used to estimate the parameters of the two
total load and weather conditions as well as the day components. Broadwater et al. (1997) used their new
type in¯uences can be calculated. The regression coe- regression-based method, Nonlinear Load Research
cients are computed by an equally or exponentially Estimator (NLRE), to forecast load for four substations
weighted least-squares estimation using the de®ned in Arkansas, USA. This method predicts load as a func-
amount of historical data. Mbamalu and El-Hawary tion of customer class, month and type of day.
(1993) used the following load model for applying this
Al-Garni et al. (1997) developed a regression model of
analysis:
electric energy consumption in Eastern Saudi Arabia as
Y t ˆ vt a t ‡ " t ; …1† a function of weather data, solar radiation, population
where and per capita gross domestic product. Variable selec-
tion is carried out using the stepping-regression method,
t sampling time, while model adequacy is evaluated by residual analysis.
Yt measured system total load, The non-parametric regression model of Charytoniuk et
vt vector of adapted variables such as time, tem- al. (1998) constructs a probability density function of
perature, light intensity, wind speed, humidity, the load and load e€ecting factors. The model produces
day type (workday, weekend), etc., the forecast as a conditional expectation of the load
at transposed vector of regression coecients, given the time, weather and other explanatory variables,
and such as the average of past actual loads and the size of
"t model error at time t. the neighbourhood.
The data analysis program allows the selection of the Alfares and Nazeeruddin (1999) presented a regres-
polynomial degree of in¯uence of the variables from 1 to sion-based daily peak load forecasting method for a
5. In most cases, linear dependency gives the best results. whole year including holidays. To forecast load precisely
Moghram and Rahman (1989) evaluated this model and throughout a year, di€erent seasonal factors that e€ect
compared it with other models for a 24-h load forecast. load di€erently in di€erent seasons are considered. In
Barakat et al. (1990) used the regression model to ®t the winter season, average wind chill factor is added as
data and check seasonal variations. The model devel- an explanatory variable in addition to the explanatory
oped by Papalexopulos and Hesterberg (1990) produces variables used in the summer model. In transitional sea-
an initial daily peak forecast and then uses this initial sons such as spring and Fall, the transformation tech-
peak forecast to produce initial hourly forecasts. In the nique is used. Finally for holidays, a holiday e€ect load
 Electric load forecasting 25

is deducted from normal load to estimate the actual where Y is an n 1 vector of observations, X is an n p
holiday load better. matrix of known coecients (based on previous load
data), ­ is a p 1 vector of the unknown parameters
and " is an n 1 vector of random errors. Results are
more accurate when the errors are not Gaussian. ­ can
3. Exponential smoothing
be obtained by iterative methods (Mbamalu and El-
Exponential smoothing is one of the classical methods Hawary 1992). Given an initial ­, one can apply the
used for load forecasting. The approach is ®rst to model Newton method. Alternatively, one can also use the
the load based on previous data, then to use this model Beaton±Turkey iterative reweighted least-square s algor-
to predict the future load. In exponential smoothing ithm (IRLS). In a similar work, Mbamalu and El-
models used by Moghram and Rahman (1989), the Hawary (1993) proposed an interactive approach
load at time t, y…t†, is modelled using a ®tting function employing least-squares and the IRLS procedure for
and is expressed in the form: estimating the parameters of a seasonal multiplicative
y…t† ˆ ­…t†T f …t† ‡ "…t†; …2† autoregressive model. The method was applied to pre-
dict load at the Nova Scotia Power Corporation.
where
f …t† ®tting function vector of the process,
5. Adaptive load forecasting
­…t† coecient vector,
"…t† white noise, and In this context, forecasting is adaptive in the sense
T transpose operator. that the model parameters are automatically corrected
to keep track of the changing load conditions. Adaptive
The Winter’s method is one of several exponential load forecasting can be used as an on-line software
smoothing methods that can analyse seasonal time package in the utilities control system. Regression analy-
series directly. This method is based on three smoothing sis based on the Kalman ®lter theory is used. The
constants for stationarity, trend and seasonality. Results Kalman ®lter normally uses the current prediction
of the analysis by Barakat et al. (1990) showed that the error and the current weather data acquisition programs
unique pattern of energy and demand pertaining to fast- to estimate the next state vector. The total historical
growing areas was dicult to analyse and predict by data set is analysed to determine the state vector, not
direct application of the Winter’s method. El-Keib et only the most recent measured load and weather data.
al. (1995) presented a hybrid approach in which expo- This mode of operation allows switching between mul-
nential smoothing was augmented with power spectrum tiple and adaptive regression analysis. The model used is
analysis and adaptive autoregressive modelling. A new the same as the one used in the multiple regression sec-
trend removal technique by In®eld and Hill (1998) was tion, as described by equation (1).
based on optimal smoothing. This technique has been Lu et al. (1989) developed an adaptive Hammerstein
shown to compare favourably with conventional model with an orthogonal escalator structure as well as
methods of load forecasting. a lattice structure for joint processes. Their method used
a joint Hammerstein non-linear time-varying functional
relationship between load and temperature. Their algor-
4. Iterative reweighted least-squares ithm performed better than the commonly used RLS
Mbamalu and El-Hawary (1992) used a procedure (Recursive Least-square) algorithm. Grady et al.
(1991) enhanced and applied the algorithm developed
referred to as the iteratively reweighted least-squares
by Lu et al. An improvement was obtained in the ability
to identify the model order and parameters. The
to forecast total system hourly load as far as 5 days.
method uses an operator that controls one variable at
McDonald et al. (1989) presented an adaptive-time-
a time. An optimal starting point is determined using the
operator. This method utilizes the autocorrelation func- series model, and simulated the e€ects of a direct load-
control strategy.
tion and the partial autocorrelation function of the
Park et al. (1991b) developed a composite model for
resulting di€erenced past load data in identifying a sub-
load prediction, composed of three components: nom-
optimal model of the load dynamics. The weighting
inal load, type load and residual load. The nominal load
function, the tuning constants and the weighted sum
is modelled such that the Kalman ®lter can be used and
of the squared residuals form a three-way decision vari-
able in identifying an optimal model and the subsequent the parameters of the model are adapted by the expo-
nentially weighted recursive least-squares method. Fan
parameter estimates. Consider the parameter estimation
and McDonald (1994) presented a practical real-time
problem involving the linear measurement equation:
implementation of weather adaptive STLF.
Y ˆ X­ ‡ "; …3† Implementation is performed by means of an ARMA
26 H. K. Alfares and M. Nazeeruddin

model, whose parameters are estimated and updated on- random component, removing subjective judgement,
line, using the WRLS (Weighted Recursive Least- and improving forecast accuracy. Zhao et al. (1997)
squares) algorithm. developed two periodical autoregressive (PAR) models
Paarmann and Najar’s (1995) adaptive online load- for hourly load forecasting.
forecasting approach automatically adjusts model par-
ameters according to changing conditions based on time
6.2. Autoregressive moving-averag e (ARMA) model
series analysis. This approach has two unique features:
autocorrelation optimization is used for handling cyclic In the ARMA model the current value of the time
patterns and, in addition to updating model parameters, series y…t† is expressed linearly in terms of its values at
the structure and order of the time series is adaptabl e to previous periods ‰y…t ¡ 1†; y…t ¡ 2†; . . .Š and in terms of
new conditions. An important feature of the regression previous values of a white noise ‰a…t†; a…t ¡ 1†; . . .Š. For
model of Hyde and Hodnett (1997b) is adaptability to an ARMA of order (p; q), the model is written as:
changing operational conditions. The load-forecastin g y…t† ˆ ¿1 y…t ¡ 1† ‡ ‡ ¿p y…t ¡ p† ‡ a…t† ¡ ³1 a…t ¡ 1†
software system is fully automated with a built-in pro-
cedure for updating the model. Zheng et al. (2000) ¡ ¡ ³q a…t ¡ q†: …5†
applied a wavelet transform-Kalma n ®lter method for The parameter identi®cation for a general ARMA
load forecasting. Two models are formed (weather sen- model can be done by a recursive scheme, or using a
sitive and insensitive) in which the wavelet coecients maximum-likelihoo d approach, which is basically a
are modelled and solved by the recursive Kalman ®lter non-linear regression algorithm. Barakat et al. (1992)
algorithm. presented a new time-temperature methodology for
load forecasting. In this method, the original time
series of monthly peak demands are decomposed into
6. Stochastic time series
deterministic and stochastic load components, the latter
It has been observed that unique patterns of energy and determined by an ARMA model. Fan and McDonald
demand pertaining to fast-growing areas are dicult to (1994) used the WRLS (Weighted Recursive Least-
analyse and predict by direct application of time-series Squares) algorithm to update the parameters of their
methods. However, these methods appear to be among adaptive ARMA model. Chen et al. (1995) used an
the most popular approaches that have been applied and adaptive ARMA model for load forecasting, in which
are still being applied to STLF. Using the time-series the available forecast errors are used to update the
approach, a model is ®rst developed based on the pre- model. Using minimum mean square error to derive
vious data, then future load is predicted based on this error learning coecients, the adaptive scheme outper-
model. The remainder of this section discusses some of formed conventional ARMA models.
the time series models used for load forecasting.

6.3. Autoregressive integrated moving-averag e

6.1. Autoregressive (AR) model (ARIMA) model
If the load is assumed to be a linear combination of If the process is non-stationary, then transformatio n
previous loads, then the autoregressive (AR) model can of the series to the stationary form has to be done ®rst.
be used to model the load pro®le, which is given by Liu This transformation can be performed by the di€eren-
et al. (1996) as: cing process. By introducing the r operator, the series
X
m ry…t† ˆ …1 ¡ B†y…t†. For a series that needs to be di€er-
L^k ˆ ¡ ¬ik Lk¡i ‡ wk …4† enced d times and has orders p and q for the AR and
iˆ1
MA components, i.e. ARIMA(p; d; q), the model is
where L^k is the predicted load at time k (min), wk is a written as:
random load disturbance, ¬i ; i ˆ 1; . . . ; m are unknown
¿…B†rd y…t† ˆ ³…B†a…t†: …6†
coecients, and (4) is the AR model of order m. The
unknown coecients in (4) can be tuned on-line using The procedure proposed by Elrazaz and Mazi (1989)
the well-known least mean square (LMS) algorithm of used the trend component to forecast the growth in
Mbamalu and El-Hawary (1993). The algorithm pre- the system load, the weather parameters to forecast
sented by El-Keib et al. (1995) includes an adaptive the weather sensitive load component, and the
autoregressive modelling technique enhanced with par- ARIMA model to produce the non-weather cyclic com-
tial autocorrelation analysis. Huang (1997) proposed an ponent of the weekly peak load. Barakat et al. (1990)
autoregressive model with an optimum threshold strati- used a seasonal ARIMA model on historical data to
®cation algorithm. This algorithm determines the mini- predict the load with seasonal variations . Juberias et
mum number of parameters required to represent the al. (1999) developed a real time load forecasting
 Electric load forecasting 27

ARIMA model that includes the meteorological in¯u- and evolutionary programming. Ma et al. (1995) used a
ence as an explanatory variable. genetic algorithm with a newly developed knowledge-
augmented mutation-like operator called the forced
mutation. Lee et al. (1997) used genetic algorithms for
long-term load forecasting, assuming di€erent func-
7. ARMAX Model based on genetic algorithms tional forms and comparing results with regression.
The genetic algorithm (GA) or evolutionary program-
ming (EP) approach is used to identify the autoregres-
sive moving average with exogenous variable (ARMAX)
model for load demand forecasts. By simulating natural 8. Fuzzy logic
evolutionary process, the algorithm o€ers the capability It is well known that a fuzzy logic system with centroid
of converging towards the global extremum of a com- defuzzi®cation can identify and approximate any
plex error surface. It is a global search technique that unknown dynamic system (here load) on the compact
simulates the natural evolution process and constitutes a set to arbitrary accuracy. Liu et al. (1996) observed that
stochastic optimization algorithm. Since the GA simul- a fuzzy logic system has great capability in drawing
taneously evaluates many points in the search space and similarities from huge data. The similarities in input
need not assume the search space is di€erentiable or data (L¡i ¡ L0 ) can be identi®ed by di€erent ®rst-
unimodal, it is capable of asymptotically converging order di€erences (Vk ) and second-order di€erences
towards the global optimal solution, and thus can (Ak ), which are de®ned as:
improve the ®tting accuracy of the model. Vk ˆ …Lk ¡ Lk¡1 †=T; Ak ˆ …Vk ¡ Vk¡1 †=T: …8†
The general scheme of the GA process is brie¯y
described here. The integer or real valued variables to The fuzzy logic-based forecaster works in two stages:
be determined in the genetic algorithm are represented training and on-line forecasting. In the training stages,
as a D-dimensional vector P for which a ®tness f …p† is the metered historical load data are used to train a 2m-
assigned. The initial population of k parent vectors Pi , input, 2n-output fuzzy-logic based forecaster to generate
i ˆ 1; . . . ; k, is generated from a randomly generated patterns database and a fuzzy rule base by using ®rst-
range in each dimension. Each parent vector then gen- and second-order di€erences of the data. After enough
erates an o€spring by merging (crossover) or modifying training, it will be linked with a controller to predict the
(mutation) individuals in the current population. load change online. If a most probably matching pattern
Consequently, 2k new individuals are obtained. Of with the highest possibility is found, then an output
these, k individuals are selected randomly, with higher pattern will be generated through a centroid defuzzi®er.
probability of choosing those with the best ®tness Several techniques have been developed to represent
values, to become the new parents for the next genera- load models by fuzzy conditional statements. Hsu (1992)
tion. This process is repeated until f is not improved or presented an expert system using fuzzy set theory for
the maximum number of generations is reached. STLF. The expert system was used to do the updating
Yang et al. (1996) described the system load model in function. Short-term forecasting was performed and
the following ARMAX form: evaluated on the Taiwan power system. Later, Liang
and Hsu (1994) formulated a fuzzy linear programming
A…q†y…t† ˆ B…q†u…t† ‡ C…q†e…t†; …7†
model of the electric generation scheduling problem,
where representing uncertainties in forecast and input data
using fuzzy set notation. Al-Anbuky et al. (1995) dis-
y…t† load at time t,
cussed the implementation of a fuzzy-logic approach
u…t† exogenous temperature input at time t,
to provide a structural framework for the representa-
e…t† white noise at time t, and
tion, manipulation and utilization of data and informa-
q¡1 back-shift operator.
tion concerning the prediction of power commitments.
and A…q†, B…q†, and C…q† are parameters of the auto- Neural networks are used to accommodate and manip-
regressive (AR), exogenous (X), and moving average ulate the large amount of sensor data.
(MA) parts, respectively. Yang et al. (1996) chose the Srinivasan et al. (1992) used the hybrid fuzzy-neural
solution(s) with the best ®tness as the tentative model(s) technique to forecast load. This technique combines the
that should further pass diagnostic checking for future neural network modelling and techniques from fuzzy
load forecasting. Yang and Huang (1998) presented a logic and fuzzy set theory. The models were later
fuzzy autoregressive moving average with exogenous enhanced by Dash et al. (1995a, b). This hybrid
variable (FARMAX) model for load demand forecasts. approach can accurately forecast on weekdays, public
The model is formulated as a combinatorial optimiza- holidays, and days before and after public holidays.
tion problem, then solved by a combination of heuristics Based on the work of Srinivasan et al., Dash et al.
28 H. K. Alfares and M. Nazeeruddin

(1995a) presented two fuzzy neural network (NN) highly automated hybrid STLF approach with un-
models capable of fuzzy classi®cation of patterns. The supervised learning.
®rst network uses the membership values of the linguistic
properties of the past load and weather parameters, where
the output of the network is de®ned as the fuzzy class 9. Neural networks
membership values of the forecasted load. The second Neural networks (NN) or arti®cial neural networks
network is based on the fact that any expert system can (ANN) have very wide applications because of their
be represented as a feedforward NN. ability to learn. According to Damborg et al., (1990),
Mori and Kobayashi (1996) used fuzzy inference neural networks o€er the potential to overcome the re-
methods to develop a non-linear optimization model liance on a functional form of a forecasting model.
of STLF, whose objective is to minimize model errors. There are many types of neural networks: multilayer
The search for the optimum solution is performed by perceptron network, self-organizing network, etc.
simulated annealing and the steepest descent method. There are multiple hidden layers in the network. In
Dash et al. (1996) used a hybrid scheme combining each hidden layer there are many neurons. Inputs are
fuzzy logic with both neural networks and expert multiplied by weights !i , and are added to a threshold ³
systems for load forecasting. Fuzzy load values are to form an inner product number called the net func-
inputs to the neural network, and the output is corrected tion. The net function NET used by Ho et al. (1992), for
by a fuzzy rule inference mechanism. Ramirez-Rosado example, is put through the activation function y, to
and Dominguez-Navarr o (1996) formulated a fuzzy produce the unit’s ®nal output, y…NET †.
model of the optimal planning problem of electric The main advantage here is that most of the fore-
energy. Computer tests indicated that this approach out- casting methods seen in the literature do not require a
performs classical deterministic models because it is able load model. However, training usually takes a lot of
to represent the intrinsic uncertainty of the process. time. Here we describe the method discussed by Liu et
al. (1996), using fully connected feed-forward type
Chow and Tram (1997) presented a fuzzy logic meth-
neural networks. The network outputs are linear func-
odology for combining information used in spatial load
tions of the weights that connect inputs and hidden units
forecasting, which predicts both the magnitudes and
to output units. Therefore, linear equations can be
locations of future electric loads. The load growth in
solved for these output weights. In each iteration
di€erent locations depends on multiple, con¯icting fac-
through the training data (epoch), the output weight
tors, such as distance to highway, distance to electric
optimization training method uses conventional back-
poles, and costs. Therefore, Chow et al. (1998) applied
propagation to improve hidden unit weights, then
a fuzzy, multi-objective model to spatial load fore-
solves linear equations for the output weights using
casting. The fuzzy logic approach proposed by Senjyu
the conjugate gradient approach.
et al. (1998) for next-day load forecasting o€ers three Srinivasan and Lee (1995) surveyed hybrid fuzzy
advantages . These are namely the ability to (1) handle neural approaches to load forecasting. Park and
non-linear curves, (2) forecast irrespective of day type Osama (1991) used a NN approach for forecasting
and (3) provide accurate forecasts in hard-to-model which, compared to regression methods, gave more ¯ex-
situations. ible relations between temperature and load patterns.
Mori et al. (1999) presented a fuzzy inference model Extending this work, Park et al. (1991a) presented a
for STLF in power systems. Their method uses tabu NN algorithm that combines time series and regression
search with supervised learning to optimize the inference approaches. Park et al. proposed an improved training
structure (i.e. number and location of fuzzy membership procedure for training the ANN. Atlas et al. (1989) ear-
functions) to minimize forecast errors. Wu and Lu lier compared a similar technique with other regression
(1999) proposed an alternative to the traditional trial methods. Hsu and Yang (1992) estimated the load pat-
and error method for determining of fuzzy membership tern of the day under study by averaging the load pat-
functions. An automatic model identi®cation is used, terns of several past days, which are of the same day
that utilizes analysis of variance, cluster estimation, type (ANN being used for the classi®cation). To predict
and recursive least-squares. Mastorocostas et al. (1999) the daily peak load, a feed-forward multilayer neural
applied a two-phase STLF methodology that also uses network was designed.
orthogonal least-squares (OSL) in fuzzy model identi®- Peng et al. (1992) used a minimum distance measure-
cation. Padmakumari et al. (1999) combined fuzzy ment to identify the appropriate historical patterns of
logic with neural networks in a technique that load and temperature weights to be used to ®nd the
reduces both errors and computational time. network weights. They also proposed an improved
Srinivasan et al. (1999) combined three techniquesÐ algorithm that combined linear and non-linear terms
fuzzy logic, neural networks and expert systemsÐin a to map past load and temperature inputs to the load
 Electric load forecasting 29

forecast output. This work was an extension to a Anbuky et al. (1995) presented fuzzy logic based neural
strategy by Peng et al. (1990) which was applied on networks for load forecasting. Dash et al. (1995a, b,
daily load. The major di€erence lies in the alternate 1996) also used fuzzy logic in combination with neural
method for the selection of the training cases. Later, networks for load forecasting. Their work has been dis-
Peng et al. (1993) applied a neural network approach cussed in the previous section.
to one-week ahead load forecasting based on an adap- Chen et al. (1996) applied a supervisory functional
tive linear combiner called the adaline. ANN technique to forecast load for three substations
Ho and Hsu (1992) designed a multilayer ANN with a in Taiwan. To enhance forecasting accuracy, the load
new adaptive learning algorithm for short term load was correlated with temperature as well as the type of
forecasting. In this algorithm the momentum is auto- customers served, which is classi®ed as residential, com-
matically adapted in the training process. Lee and mercial or industrial. Al-Fuhaid et al. (1997) incorpo-
Park (1992) proposed a non-linear load model and sev- rated temperature and humidity e€ects in an ANN
eral structures of ANNs were tested. Inputs to the ANN approach for STLF in Kuwait. Vermaak and Botha
include past load values, and the output is the forecast (1998) proposed a recurrent NN to model the STLF
for a given day. Lee and Park demonstrated that the of the South African utility. They utilized the inherent
ANN could be successfully used in STLF with accepted non-linear dynamic nature of NN to represent the load
accuracy. Chen et al. (1992) presented an ANN, which is as the output of some dynamic system, in¯uenced by
not fully connected, to forecast weather sensitive loads weather, time and environmental variables.
for a week. Their model could di€erentiate between the McMenamin and Monforte (1998) used an econo-
weekday loads and the weekend loads. Lu et al. (1993) metric and statistical approach to NN-based load fore-
conducted a computationa l investigation to evaluate the casting. Considering NN models as ¯exible non-linear
performance of the ANN methodology. equations, they used non-linear least-square s to estimate
Djukanovic et al. (1993) proposed an algorithm using parameters, and simple statistics such as MAPE to
an unsupervised/supervised learning concept and histor- determine the number of nodes. Papadakis et al.
ical relationship between the load and temperature for a (1998) developed a three-step fuzzy ANN approach,
given season, day type and hour of the day. They used involving the prediction of load curve peaks and valleys
this algorithm to forecast hourly electric load with a lead and mapping them to forecasted peak values. Dash et al.
time of 24 h. Papalexopoulos et al. (1994) developed and (1998) presented a fuzzy NN load forecasting system
implemented the ANN based model for the energy con- that accounts for seasonal and daily changes, as well
trol centre of the Paci®c Gas and Electric Company. as holidays and special situations. An adaptive
Attention was paid to accurately model special events, mechanism is used to train the system on line, providing
such as holidays, heat waves, cold snaps and other con- accurate results when tested with actual data of the
ditions that disturb the normal pattern of the load. Ho Virginia Utility. Another adaptive NN technique,
et al. (1992) extended the three-layered feedforward employing genetic algorithms in the design and training
adaptive neural networks to multilayers. Dillon et al. phase, was used by Kung et al. (1998) on the Taiwan
(1991) proposed a multilayer feedforward neural net- power system.
work, using a learning algorithm for adaptive training ANN have been integrated with several other tech-
of neural networks. niques to improve their accuracy. Chow and Leung
Srinivasan et al. (1991) used an ANN based on back (1996), for example, combined ANN with stochastic
propagatio n for forecasting, and showed its superiority time-series methods, in the form of non-linear autore-
to traditional methods. Liu et al. (1991) compared an gressive integrated (NARI) model. They implemented
econometric model and a neural network model, an ANN capable of weather compensation, based on
through a case study on electricity consumption fore- NARI, to forecast electric load in Hong Kong.
casting in Singapore. Their results show that a fully Choueiki et al. (1997) used weighted least-squares pro-
trained NN model with a good ®tting performance for cedure in the training phase of developing an ANN for
the past may not give a good forecasting performance load forecasting. Several other hybrid methods invol-
for the future. Kalra et al. (1992) demonstrated how ving ANNs in combination with fuzzy logic and expert
present methods for solving such problems could be systems are discussed in Sections 6 and 10, respectively.
converted to NN approaches. It is very hard to keep track of all publications on load
Azzam-ul-Asar and McDonald (1994) trained a forecasting using NN, which is currently a very active
family of ANNs and then used them in line with a super- area of research. Niebur (1995) and Czernichow et al.
visory expert system to form an expert network. They (1996) surveyed methods and applications of electrical
also investigated the e€ectiveness of the ANN approach load forecasting with ANNs.
to short term load forecasting, where the networks were Oonsivilan et al. (1999) presented an approach for
trained on actual load data using back-propagation . Al- predicting electric power system commercial load using
30 H. K. Alfares and M. Nazeeruddin

a wavelet neural network. Their results showed that The logical and syntactical relationships between
wavelet NNs may outperform traditional architectures weather load and the prevailing daily load shapes have
in approximation. Drenza et al. (1999) presented a new been widely examined to develop di€erent rules for dif-
ANN-based technique for STLF. The technique imple- ferent approaches. The typical variables in the process
mented active selection of training data employing k- are the season under consideration, day of the week, the
nearest neighbours concept. Excellent results were temperature and the change in this temperature.
reported using this technique. Yoo and Pimmel (1999) Illustrations of this method can be found in Rahman
developed a self-supervised adaptive NN to perform (1990, 1993) and Ho et al. (1990). The algorithms of
STLF for a large power system. They used the self- Rahman and Shreshta (1991) and Rahman and Hazim
supervised network to extract correlational features (1993) combine features from knowledge-based and sta-
from temperature and load data. Their results showed tistical techniques, using the pairwise comparison tech-
low forecasting errors. Kandil et al. (1999) used multi- nique to prioritize categorical variables. Rahman and
layer perceptron (MLP) type ANN for STLF using real Hazim (1996) developed a site-independent expert
load and weather data. Leyan and Chen (1999) used system for STLF. This system was tested using data
variable learning rate method combined with quasi- from several sites around the USA, and the errors
Newton method to expedite the learning process of were negligible. Brown et al. (1999) used a knowledge-
ANN for STLF. based load-forecasting approach that combines existing
Nazarko and Styczynski (1999) presented load-mod- system knowledge, load growth patterns, and horizon
elling methods useful for long term planning of power year data to develop multiple load growth scenarios.
distribution systems using statistical clustering and NN Several hybrid methods combine expert systems with
approach. Ijumba and Hunsley (1999) applied ANN other load-forecasting approaches. Dash et al. (1993,
model to predict hourly peak demands of loads in a 1996) combined fuzzy logic with expert systems. Kim
newly electri®ed area. Sinha and Mandal (1999) pre- et al. (1995) used a two-step approach in forecasting
sented an ANN-based model for bus-load prediction load for Korea Electric Power Corporation. First, an
and dynamic state estimation in power systems. ANN is trained to obtain an initial load prediction,
Drezga and Rahman (1999a) used phase-space concepts then a fuzzy expert system modi®es the forecast to
to embed electric load parameters, including tempera- accommodate temperature changes and holidays.
ture and cycle variables, into ANN-based STLF. Mohamad et al. (1996) applied a combination of
Drezga and Rahman (1999b) applied another ANN- expert systems and NN for hourly load forecasting in
based technique that features the following characteris- Egypt. Bataineh et al. (1996) used neural networks and
tics: (1) selection of training data by the k-nearest neigh- fuzzy logic for data representation and manipulation to
bours concept, (2) pilot simulation to determine the construct the expert system’s rule base. Chiu et al.
number of ANN units and (3) iterative forecasting by (1997a, b) determined that a combined expert system-
simple moving average to combine local ANN predic- NN approach is faster and more accurate than either
tions. one of the two methods alone. Chandrashekara et al.
(1999) applied a combined expert system-NN procedure
divided into three modules: location planning, fore-
casting and expansion planning.
10. Knowledge-based expert systems
Expert systems are new techniques that have emerged as
a result of advances in the ®eld of arti®cial intelligence. 11. Comparison of approaches
An expert system is a computer program that has the In addition to classifying load-forecasting approaches, it
ability to reason, explain, and have its knowledge base is important to compare di€erent categories and indi-
expanded as new information becomes available to it. vidual techniques. A number of researchers have
To build the model, the `knowledge engineer’ extracts attempted to empirically compare some of the methods
load forecasting knowledge from an expert in the ®eld used in load forecasting. One of the earliest and most
by what is called the knowledge base component of the comprehensive comparisons is made by Willis and
expert system. This knowledge is represented as facts Northcote-Green (1984), who performed comparison
and IF-THEN rules, and consists of the set of relation- tests on 14 load forecasting methods. Atlas et al.
ships between the changes in the system load and (1989) compared the performance of di€erent structures
changes in natural and forced condition factors that of neural networks with regression models. Dash et al.
e€ect the use of electricity. This rule base is used daily (1995a) also compared several fuzzy neural network-
to generate the forecasts. Some of the rules do not based methods. Another comparison between neural
change over time, while others have to be updated con- networks and econometric models of forecasting electri-
tinually. city consumption was performed by Liu et al. (1991).
 Electric load forecasting 31

Girgis et al. (1995) used actual load data to compare Although the time series approach is still widely used,
estimation errors of one-hour ahead and one-day- newer techniques o€er a lot of promise for this devel-
ahead forecasts associated with three self-learning fore- oping and rapidly changing ®eld. The rapidly increasing
casting techniques. These techniques are: adaptive power of the personal computer is making it possible to
Kalman ®lter, neural networks, and expert systems. apply more complicated solution techniques. New load
On the basis of a simulation study, Liu et al. (1996) forecasting methods based on fuzzy logic, genetic algor-
compared three other techniquesÐfuzzy logic (FL), ithms, expert systems, and neural networks o€er new
neural networks (NN) and autoregressive models hopes in this direction of research. Over the last few
(AR)Ðconcluding that NN and FL are much superior years, the most active research area has been neural net-
to AR models of STLF. work based load forecasting.
Other limited comparative data exist, provided by
many researchers to establish the superiority of their
proposed forecasting methods over a limited number
of previously published methods. For example, References
AB U E L-M AGD, M. A., and SIN HA, N. K., 1982, Short term load
Mbamalu and El-Hawary (1993) compared their inter-
demand modeling and forecasting. IEEE Transactions on Systems
active autoregressive model to the Box-Jenkins method. Man and Cybernetics, 12, 370±382.
Willis et al. (1995) compared their simulation-based AL -A NBUKY, A., BAT AINEH, S., and AL -AQTA SH, S., 1995, Power
method to two other simulation methods. Wu and Lu demand prediction using fuzzy logic. Control Engineering Practice,
(1999) compared their fuzzy modelling method to Box- 3, 1291±1298.
Jenkins transfer functions and ANN. Srinivasan et al. AL -FUHAID, A. S., E L-SAYED, M. A., and MAH MOUD, M. S., 1997,
Neuro-short-term forecast of the power system in Kuwait. Applied
(1999) compared their NN-fuzzy expert system method-
Mathematical Modeling, 21, 215±219.
ology to a regression-base d model, showing signi®cant AL -GARNI, A. Z., AH MED, Z., A L-NASSAR , Y. N., ZU BAIR, S. M., and
improvement in forecasting accuracy. The need for up- A L-SHE HRI, A., 1997, Model for electric energy consumption in
to-date comprehensive comparisons of the di€erent load Eastern Saudi Arabia. Energy Sources, 19, 325Ð334.
forecasting methods provides a challenging opportunity AL FARES, H. K., and N AZEE RU DDIN, M., 1999, Regression-based
methodology for daily peak load forecasting. Proceedings of the
for future research, given the wide variety of objectives
2nd Internationa l Conference on Operations and Quantitative
and assumptions , and the unlimited possibility of mixing Management, Ahmedabad, India, 3±6 January, pp. 468±471.
and matching di€erent components of various methods. AT LAS, L., C ONNOR, J., P ARK , D., MAR KS, R. J., LIPPMAN, A., and
M UT HUSAMY, Y., 1989, A performance comparison of trained multi-
layer perceptrons and trained classi®cation trees. Proceedings of the
IEEE Internationa l Conference on Systems, Man, and Cybernetics,
12. Conclusions pp. 915±920.
Di€erent techniques have been applied to load fore- AZ ZAM-UL-A SAR , A., and M C DONALD, J. R., 1994, A Speci®cation of
neural networks in the load forecasting problem. IEEE Transactions
casting. Nine approaches have been reviewed in this
on Control System Technology, 2, 135±141.
paper: (1) multiple regression, (2) exponential BARAKAT, E. H., A L-Q ASSIM , J. M., and AL-R ASHED, S. A., 1992,
smoothing, (3) iterative reweighted least-squares, (4) New model for peak demand forecasting applied to highly complex
adaptive load forecasting, (5) stochastic time series, (6) load characteristics of a fast developing area. IEE Proceedings - C,
ARMAX models based on genetic algorithms, (7) fuzzy 139, 136±149.
logic, (8) neural networks and (9) expert systems. After BARAKAT, E. H., Q AYYUM , M. A., H AM ED, M. N., and A L-R ASHE D,
S. A., 1990, Short-term peak demand forecasting in fast developing
surveying all these approaches, we can observe a clear utility with inherent dynamic load characteristics. IEEE
trend toward new, stochastic, and dynamic forecasting Transactions on Power Systems, 5, 813±824.
techniques. It seems a lot of current research e€ort is BATAINEH, S., A L-AN BUKY, A., and A L-AQ TASH, S., 1996, Expert
focused on three such methods: fuzzy logic, expert system for unit commitment and power demand prediction using
systems and particularly neural networks. There is also fuzzy logic and neural networks. Expert Systems, 13, 29±40.
BROADWATER, R. P., SARG ENT, A., YARALI, A., SHAALAN, H. E., and
a clear move towards hybrid methods, which combine
NAZ ARKO, J., 1997, Estimating substation peaks from research data.
two or more of these techniques. IEEE Transactions on Power Delivery, 12, 451±456.
Over the years, the direction of research has shifted, BROWN , R. E., H ANSON, A. P., and HAG AN , D. L., 1999, Long range
replacing old approaches with newer and more ecient spatial load forecasting using non-uniform areas. IEEE
ones. Apparently due to their limited success, a number Transmission and Distribution Conference Proceedings, 1, 369±373.
of old approache s seem to be out of favour nowadays. BUNN, D. W., and FARME R, E. D., 1985, Review of Short-term
Forecasting Methods in the Electric Power Industry (New York:
These include such methods as state space and Kalman
Wiley), pp. 13±30.
®lter modelling, on-line load forecasting, and fore- C HANDRASHEKA RA , A. S., A NANTHA PADMAN ABHA , T., and
casting by pattern recognition. There is also consider- KU LKARNI, A. D., 1999, A neuro-expert system for planning and
ably less emphasis on methods such as iterative load forecasting of distribution systems. Electrical Power and Energy
reweighted least-squares and adaptive load forecasting. Systems Research, 21, 309±314.
32 H. K. Alfares and M. Nazeeruddin

C HARYTON IU K, W., C HEN, M. S., and VAN O LIN DA, P., 1998, DJU KANOVIC, M., BAB IC, B., SOB AJIC, O. J., and PAO , Y. H., 1993, 24-
Nonparametric regression based short-term load forecasting. IEEE Hour load forecasting. IEEE Proceedings Ð C, 140, 311±318.
Transactions on Power Systems, 13, 725±730. DRE ZGA, I., and R AHMAN, S., 1999a, Phase-space based short-term
C HEN, C. S., TZ ENG, Y. M., and H WANG, J. C., 1996, Application of load forecasting for deregulated electric power industry.
arti®cial neural networks to substation load forecasting. Electric Proceedings of the Internationa l Joint Conference on Neural
Power Systems Research, 38, 153±160. Networks, 5, 3405±3409.
C HEN, J.-F., WAN G, W.-M., and HUANG , C.-M., 1995, Analysis of an DRE ZGA, I., and RAHM AN , S., 1999b, Short term load forecasting with
adaptive time-series autoregressive moving-average (ARMA) model local ANN predictors. IEEE Transactions on Power Systems, 14,
for short-term load forecasting. Electric Power Systems Research, 844±850.
34, 187±196. EL-KE IB, A. A., M A , X., and M A , H., 1995, Advancement of statistical
C HEN, S.-T., YU , D. C., and MOG HADDAMJO, A. R., 1992, Weather based modeling for short-term load forecasting. Electric Power
sensitive short-term load forecasting using nonfully connected arti- Systems Research, 35, 51±58.
®cial neural network. IEEE Transactions on Power Systems, 7, 1098± ELRAZAZ, Z. S., and M AZI, A. A., 1989, Uni®ed weekly peak load
1105. forecasting for fast growing power system. IEE Proceedings Ð C,
C HIU, C.-C., C OOK , D. F., KAO , J.-L., and C HOU, Y.-C., 1997, 136, 29±41.
Combining a neural network and a rule-based expert system for FAN , J. Y., and M C DONALD , J. D., 1994, A Real time implementation
short-term load forecasting. Computers and Industria l Engineering, of short-term load forecasting for distribution power system. IEEE
32, 787±797. Transactions on Power Systems, 9, 988±993.
C HIU, C.-C., K AO , L.-J., and C OOK , D. F., 1997, Combining a neural GIRGIS, A. A., VA RAD AN, S., EL -DIN, A. K., and ZH U, J., 1995,
network with a rule-based expert system approach for short-term Comparison of di€erent approaches to short-term load forecasting.
power load forecasting in Taiwan. Expert Systems with Applications, International Journa l of Engineering Intelligent Systems for Electrical
13, 299±305. Engineering and Communications, 3, 205±210.
C HOUE IKI, M. H., M OUNT -C AMPBELL, C. A., and A HALT, S., 1997, GRADY , W. M., GROCE, L. A., HUE BNER, T. M., LU , Q. C., and
Implementing a weighted least squares procedure in training a C RAWFORD, M. M., 1991, Enhancement, implementation and per-
neural network to solve the short-term load forecasting problem. formance of an adaptive load forecasting technique. IEEE
IEEE Transactions on Power Systems, 12, 1689±1694. Transactions on Power Systems, 6, 450±456.
C HOW , M., and TRAM , H., 1997, Application of fuzzy logic technology GROSS, G., and GAL IANA, F. D., 1987, Short term load forecasting.
for spatial load forecasting. IEEE Transactions on Power Systems, Proceedings of the IEEE, 75, 1558±1573.
12, 1360±1366. HAIDA, T., and M U TO, S., 1994, Regression based peak load forecast-
C HOW , M., Z HU, J., and TRAM , H., 1998, Application of fuzzy multi- ing using a transformation technique. IEEE Transaction s on Power
objective decision making in spatial load forecasting. IEEE Systems, 9, 1788±1794.
Transactions on Power Systems, 13, 1185±1190. HAIDA, T., MUT O, S., TAKAHASHI, Y., and I SHI , Y., 1998, Peak load
C HOW , T. W. S., and LEUNG, C.-T., 1996, Nonlinear autoregressive forecasting using multiple-year data with trend data processing tech-
neural network model for short-term load forecasting. IEE Proceed- niques. Electrical Engineering in Japan, 124, 7±16.
ings: Generation, Transmission, and Distribution, 143, 500±506. HO , K., HSU , Y., C HE N, C., LEE , T., L IANG, C., LAI , T., and C HEN,
C ZE RNICH OW, T., P IRAS, A., I MHOF , K., C AIRE , P., JACCARD , Y., K., 1990, Short term load forecasting of Taiwan power system
D ORI ZZI, B., and G ERMOND, A., 1996, Short term electrical load using a knowledge based expert system. IEEE Transactions on
forecasting with arti®cial neural networks. Internationa l Journal of Power Systems, 5, 1214±1221.
Engineering Intelligent Systems for Electrical Engineering and HO , K. L., HSU , Y. Y., and Y ANG, C. C., 1992, Short term load fore-
Communications, 4, 85±99. casting using a multi layer neural network with an adaptive learning
DAMB ORG, M. J., EL-SH ARK AWI, M. A., AG GOUNE, M. E., and algorithm. IEEE Transactions on Power Systems, 7, 141±149.
M ARKS , R. J., II, 1990, Potential of arti®cial neural network to HOB BS, B, F., J ITPRAPAIK ULSARN , S., KON DA, S., C HANKONG, V.,
power system operation. Proceedings of the IEEE Internationa l LOPARO , K. A., and M ARATUKU LAM, D. J., 1999, Analysis of unit
Symposium on Circuits and Systems, New Orleans, LA, pp. 2933± commitment of improved load forecasts. IEEE Transactions on
2937. Power Systems, 14, 1342±1348.
DASH , P. K., DASH , S., R AMA K RISHN A, G., and R AH MAN, S., 1993, HSU , Y. Y., 1992, Fuzzy expert systems: an application to short-term
Forecasting of a load time series using a fuzzy expert system and load forecasting. IEEE Proceedings Ð C, 139, 471±477.
fuzzy neural networks. Internationa l Journa l of Engineering HSU , Y. Y., and YANG , L.-C., 1992, Design of arti®cial neural net-
Intelligent Systems, 1, 103±118. works for short term load forecasting. IEEE Proceedings Ð C, 138,
DASH , P. K., L IEW, A. C., and R AHM AN, S., 1995a, Comparison of 407±413.
fuzzy neural networks for the generation of daily average and peak HUAN G, S. R., 1997, Short-term load forecasting using threshold auto-
load pro®les. Internationa l Journal of System Science, 26, 2091±2106. regressive models. IEE Proceedings: Generation, Transmission, and
DASH , P. K., LIEW, A. C., and R AHMAN, S., 1996, Fuzzy neural net- Distribution, 144, 477±481.
work and fuzzy expert system for load forecasting. IEE Proceedings: HYDE , O., and H ODNET T, P. F., 1997a, Modeling the e€ect of weather
Generation, Transmission, and Distribution, 143, 106±114. in short-term electricity load forecasting. Mathematical Engineering
DASH , P. K., L IEW, A. C., R AHMAN, S., and DASH , S., 1995b, Fuzzy in Industry, 6, 155±169.
and neuro-fuzzy computing models for electric load forecasting. HYDE , O., and HOD NETT, P. F., 1997b, Adaptable automated pro-
Engineering Applications of Arti®cial Intelligence, 8, 423±433. cedure for short-term electricity load forecasting. IEEE
DASH , P. K., SATPATHY, H. P., and L IEW, A. C., 1998, Real-time Transactions on Power Systems, 12, 84±94.
short-term peak and average load forecasting system using a self- I NFIE LD, D. G., and HIL L, D. C., 1998, Optimal smoothing for trend
organizing fuzzy neural network. Engineering Applications of removal in short term electricity demand forecasting. IEEE
Arti®cial Intelligence, 11, 307±316. Transactions on Power Systems, 13, 1115±1120.
DIL LON, T. S., SE ST ITO, S., and LEUNG, S., 1991, Short term load I JUMB A, N. M., and H UNSLEY, J. P., 1999, Improved load forecasting
forecasting using an adaptive neural network. Electric Power and techniques for newly electri®ed areas. Africon 1999 IEEE, 2, 989±
Energy Systems, 13, 186±192. 994.
 Electric load forecasting 33

J UBERIAS, G., YUNT A, R., G ARCIA M ORINO, J., and M EN DIVIL, C., management for the control of electric load. Transactions of the
1999, A new ARIMA model for hourly load forecasting. IEEE Institute of Measurement and Control, 11, 79±91.
Transmission and Distribution Conference Proceedings, 1, 314±319. M C ME NAMIN, J. S., and M ONFORTE, F. A., 1998, Short term energy
KAL RA, P. K., SRIVAST AV , A., and C HAT URVEDI, D. K., 1992, Possible forecasting with neural networks. Energy Journa l, 19, 43±62.
application of neural nets to power system operation and control. M OGHRAM, I., and RAHMAN , S., 1989, Analysis and evaluation of ®ve
Electric Power Systems Research, 25, 83±90. short-term load forecasting techniques. IEEE Transactions on Power
KAN DIL, N., SO OD , V., and SAAD , M., 1999, Use of ANNs for short- Systems, 4, 1484±1491.
term load forecasting. 1999 Canadian Conference on Electrical and M OHAMAD, E. A., M ANSOUR, M. M., EL -DEB EIKY, S., M OHAMAD, K.
Computer Engineering, 2, 1057±1061. G., RAO , N. D., and R AMAKRISHNA, G., 1996, Results of
KIM , K.-H., P ARK , J.-K., HWANG, K.-J., and KIM , S.-H., 1995, Egyptian uni®ed grid hourly load forecasting using an arti®cial
Implementation of hybrid short-term load forecasting system using neural network with expert system interfaces. Electric Power
arti®cial neural networks and fuzzy expert systems. IEEE Systems Research, 39, 171±177.
Transactions on Power Systems, 10, 1534±1539. M ORI , H., and KOB AYASHI, H., 1996, Optimal fuzzy inference for
KU NG, C.-H., DEVANEY, M. J., H UANG, C.-M., and KUNG, C.-M., short-term load forecasting. IEEE Transactions on Power Systems,
1998, Adaptive power system load forecasting scheme using a 11, 390±396.
genetic algorithm embedded neural network. Proceedings of the M ORI , H., SON E, Y., M ORI DERA, D., and KONDO, T., 1999, Fuzzy
IEEE Instrumentation and Measurement Technology Conference, inference models for short-term load forecasting with tabu search.
Part 1, St Paul, MN, 18±21 May, pp. 308±311. IEEE Systems, Man, and Cybernetics Conference Proceedings, 6,
LEE, D. G., LEE , B. W., and C HANG , S. H., 1997, Genetic programming 551±556.
model for long-term forecasting of electric power demand. Electric Nazarko J., and Styczynski, Z. A., 1999, Application of statistical
Power Systems Research, 40, 17±22. and neural approaches to the daily load pro®les modelling in power
LEE, K. Y., and PARK , J. H., 1992, Short-term load forecasting using distribution systems. IEEE: Transmission and Distribution
an arti®cial neural network. IEEE Transactions on Power Systems, 7, Conference, 1, 320±325.
124±130. NIE BUR, D., 1995, Arti®cial neural networks in the power industry,
LEYAN, X., and C HEN , W. J., 1999, Arti®cial neural network short- survey and applications. Neural Network World, 5, 951±964.
term electrical load forecasting techniques. TENCON 99: OON SIVILAN, A., and EL -HAWARY , M. E., 1999, Wavelet neural net-
Proceedings of the IEEE Region 10 Conference, 2, pp. 1458±1461. work based short term load forecasting of electric power system
LIANG, R. H., and HSU , Y. Y., 1994, Fuzzy linear programming: an commercial load. 1999 Canadian Conference on Electrical and
application to hydroelectric generation scheduling. IEE Proceedings: Computer Engineering, 3, 1023±1028.
Generation, Transmission and Distribution, 141, 568±574. P AARMANN, L. D., and NAJAR, M. D., 1995, Adaptive online load
LIU, K., SUBBARAYAN, S., SHOULT S, R. R., M ANRY , M. T., KWAN , C., forecasting via time series modeling. Electric Power Systems
L EWI S, F. L., and NACCARINO , J., 1996, Comparison of very Research, 32, 219±225.
short-term load forecasting. IEEE Transactions on Power Systems, P ADMAKUMARI, K., M OHAND AS, K. P., and THERUVEN GADAM , S.,
11, 877±882. 1999, Long term distribution demand forecasting using neuro
LIU, X. O., AN G, B. W., and GOH , T. N., 1991, Forecasting of elec- fuzzy computations. Electrical Power and Energy Systems
tricity consumption: a comparison between an econometric model Research, 21, 315±322.
and a neural network model. IEEE International Joint Conference on P APA DAK IS, S. E., T HE OCHARIS, J. B., KIARTZIS, S. J., and BAKIRTZ IS,
Neural Networks, Singapore, 18±21 November, pp. 1254±1259. A. G., 1998, Novel approach to short-term load forecasting using
LU , C. N., W U , H. T., and V EM URI, S., 1993, Neural network based fuzzy neural networks. IEEE Transactions on Power Systems, 13,
short term load forecasting. IEEE Transactions on Power Systems, 8, 480±489.
336±341. P APA LEXOPOULOS , A. D., H AO , S., and P EN G, T.-M., 1994, An imple-
LU , Q. C., GRADY , W. M., CRAWF ORD, M. M., and A NDERSON, G. mentation of a neural network based load forecasting model for the
M., 1989, An adaptive non-linear predictor with orthogonal escala- EMS. IEEE Transactions on Power Systems, 9, 1956±1962.
tor structure for short-term load forecasting. IEEE Transactions on P APA LEXOPULOS , A. D., and HE ST ERB ERG, T. C., 1990, A regression-
Power Systems, 4, 158±164. based approach to short-term load forecasting. IEEE Transactions
M A , X., EL -KEI B, A. A., SM IT H, R. E., and M A , H., 1995, Genetic on Power Systems, 5, 1214±1221.
algorithm based approach to thermal unit commitment of electric P ARK , D. C., EL-SHARKAWI, M. A., M ARKS , R. J., A TLAS, L. E., and
power systems. Electric Power Systems Research, 34, 29±36. DAMB ORG, M. J., 1991a, Electric load forecasting using an Arti®cial
M ASTOROCOSTAS , P. A. T HEOCHARIS, J. B., and BAKI RT ZIS, A. G., Neural Network. IEEE Transactions on Power Systems, 6, 442±449.
1999, Fuzzy modeling for short term load forecasting using the P ARK , D. C., and O SAMA , M., 1991, Arti®cial neural network based
orthogonal least squares method. IEEE Transactions on Power peak load forecasting. IEEE Proceedings of the SOUTHEASTCON
Systems, 14, 29±36. `91, 1, pp. 225±228.
M ATTH EWMAN, P. D., and NICHOLSON, H., 1968, Techniques for load P ARK , J. H., PARK , Y. M., and LEE, K. Y., 1991b, Composite modeling
prediction in electricity supply industry. Proceedings of the IEEE, for adaptive short-term load forecasting. IEEE Transactions on
115, 1451±1457. Power Systems, 6, 450±456.
M BAMALU, G. A. N., and E L-H AWARY , M. E., 1992, Load forecasting P ENG, T. M., H UBEL E, N. F., and KARADY , G. G., 1990, An conceptual
via suboptimal seasonal autoregressive models and iteratively approach to the application of neural networks for short term load
reweighted least squares estimation. IEEE Transactions on Power forecasting. Proceedings of the IEEE International Symposium on
Systems, 8, 343±348. Circuits and Systems, New Orleans, LA, pp. 2942±2945.
M BAMALU, G. A. N., and E L-H AWARY , M. E., 1993, Load forecasting P ENG, T. M., HU BEL E, N. F., and KARADY , G. G., 1992,
via suboptimal autoregressive models and iteratively recursive least Advancement in the application of neural networks for short-term
squares estimation. IEEE Transactions on Power Systems, 8, 343± load forecasting. IEEE Transactions on Power Systems, 7, 250±257.
348. P ENG, T. M., HUBE LE, N. F., and KARADY , G. G., 1993, An adaptive
M C DONALD , J. R., LO , K. L., and SHERWOOD, P. M., 1989, Neural Network approach to one-week ahead load forecasting.
Application of short-term adaptive forecasting techniques in energy IEEE Transactions on Power Systems, 8, 1195±1203.
34 Electric load forecasting

R AHMAN, S., 1990, Formulation and analysis of a rule based short forecasting: system implementation and performance evaluation.
term load forecasting algorithm. Proceedings of IEEE, 78, 805±816. IEEE Transactions on Power Systems, 14, 1100±1106.
R AHMAN, S., 1993, Generalized knowledge-based short-term load fore- V ARADAN, S., and MA KRA M, E. B., 1996, Harmonic load identi®cation
casting technique. IEEE Transactions on Power Systems, 8, 508±514. and determination of load composition using a least squares
R AHMAN, S., and HAZI M, O., 1993, A generalized knowledge-based method. Electric Power Systems Research, 37, 203±208.
short term load forecasting technique. IEEE Transactions on V ERMAAK, J., and BOT HA, E. C., 1998, Recurrent neural networks for
Power Systems, 8, 509±514. short-term load forecasting. IEEE Transactions on Power Systems,
R AHMAN, S., and HAZ IM , O., 1996, Load forecasting for multiple sites: 13, 126±132.
development of an expert system-based technique. Electric Power W ILLIS, H. L., FIN LEY, L. A., and BU RI, M. J., 1995, Forecasting
Systems Research, 39, 161±169. electric demand of distribution system planning in rural and sparsely
R AHMAN, S., and SHRESHTA, G., 1991, A priority vector based tech- populated regions. IEEE Transactions on Power Systems, 10, 2008±
nique for load forecasting. IEEE Transactions on Power Systems, 6, 2013.
1459±1465. W ILLIS, H. L., and NORTHCOTE-GREEN, J. E. D., 1984, Comparison of
R AMIREZ-R OSAD O , I. J., and DOMINGUE Z-NAVARRO , J. A., 1996, fourteen distribution load forecasting methods. IEEE Transactions
Distribution planning of electric energy using fuzzy models. on Power Apparatu s and Systems, PAS-103, 1190±1197.
Internationa l Journal of Power and Energy Systems, 16, 49±55. W U , H.-C., and LU , C., 1999, Automatic fuzzy model identi®cation
SENJ YU, T., HIGA, S., and UE ZATO, K., 1998, Future load curve shap-
for short-term load forecast. Generation, Transmission and
ing based on similarity using fuzzy logic approach. IEE Proceedings:
Distribution, IEE Proceedings, 146, 477±482.
Generation, Transmission and Distribution, 145, 375±380.
YAN G, H.-T., and H UANG, C.-M., 1998, New short term load-fore-
SINHA, A. K., and MAN DAL, J. K., 1999, Hierarchical dynamic state
casting approach using self-organizing fuzzy ARMAX models.
estimator using ANN-based dynamic load prediction. IEE
IEEE Transactions on Power Systems, 13, 217±225.
Proceedings: Generation, Transmission and Distribution, 146, 541±
YAN G, H.-T., H UANG, C.-M., and HUANG , C.-L., 1996, Identi®cation
549.
of ARMAX model for short term load forecasting: an evolutionary
SRINIVASA N, D., CH ANG, C. S., and LIE W, A. C., 1992, Demand fore-
programming approach. IEEE Transactions on Power Systems, 11,
casting using fuzzy neural computation, with special emphasis on
403±408.
weekend and public holiday forecasting. IEEE Transactions on
Power Systems, 8, 343±348. YOO , H., and P IMMEL, R. L., 1999, Short term load forecasting using
SRINIVASA N, D., and LEE, M. A., 1995, Survey of hybrid fuzzy neural self-supervised adaptive neural network. IEEE Transactions on
approaches to electric load forecasting. Proceedings of the IEEE Power Systems, 14, 779±784.
Internationa l Conference on Systems, Man and Cybernetics, Part 5, ZHAO , H., RE N, Z., and H UANG, W., 1997, Short-term load forecasting
Vancouver, BC, pp. 4004±4008. considering weekly period based on periodical auto regression.
SRINIVASA N, D., LIEW, A. C., and C HE N, S. P. J., 1991, A novel Proceedings of the Chinese Society of Electrical Engineers, 17, pp.
approach to electric load forecasting based on neural networks. 211±213.
IEEE Internationa l Joint Conference on Neural Networks, ZHENG, T., GI RGIS, A. A., and M AKRAM , E. B., 2000, A hybrid wave-
Singapore, 18±21 November, pp. 1172±1177. let±Kalman ®lter method for load forecasting. Electric Power
SRINIVASA N, D., T AN , S. S., CH ANG, C. S., and C HAN, E. K., 1999, Systems Research, 54, 11±17.
Parallel neural network-fuzzy expert system for short-term load