short-term load forecasting using generalized regression and probabilistic neural networks in the...
TRANSCRIPT
M.M. Tripathi is Senior DesignEngineer in DOEACC Society
Gorakhpur Centre, Gorakhpur (UttarPradesh), India. He has also workedas Engineer-SC with the Institute for
Plasma Research in Gandhinagar.His research interests include
artificial neural networks and fuzzy-neural applications in power system
problems, application of real-timecontrol in power systems, and powersystem restructuring. He received his
B.Tech. degree from M.M.M.Engineering College, Gorakhpur, in
1994. Presently he is pursuing Ph.D.work from U.P. Technical University
in Lucknow.
K.G. Upadhyay is a faculty memberin the Department of Electrical
Engineering of M.M.M. EngineeringCollege in Gorakhpur. His research
interests include power systemsoperation and control, FACTS, and
deregulation. He received hisM.Tech. and Ph.D. degrees from IITDelhi, and U.P. Technical University
in Lucknow in 1989 and 2002,respectively.
S.N. Singh is Professor in theDepartment of Electrical
Engineering of I.I.T. Kanpur. Hisresearch interests include power
system restructuring, FACTS, powersystem optimization and control,security analysis, artificial neural
networks and fuzzy/neuralapplications in power system
problems, and transient stability.
24 1040-6190/$–see front matter # 2008 Else
Short-Term Load ForecastingUsing Generalized Regressionand Probabilistic NeuralNetworks in the ElectricityMarket
For the economic and secure operation of power systems, aprecise short-term load forecasting technique is essential.Modern load forecasting techniques – especially artificialneural network methods – are particularly attractive, asthey have the ability to handle the non-linear relationshipsbetween load, weather temperature, and the factorsaffecting them directly. A test of two different ANN modelson data from Australia’s Victoria market is promising.
M.M. Tripathi, K.G. Upadhyay and S.N. Singh
I. Introduction
Short-term load forecasting,
which plays an important role in
the economic and secure
operation of the power system,
is always a concern of power
system operators.1 Reasonably
accurate short-term load
forecasting has become a
challenging issue in the ongoing
deregulation of electricity sector
vier Inc. All rights reserved., doi:/10.1016/j.
where load demand is greatly
influenced by electricity prices
which change hourly or half-
hourly depending on the market
structure and rules. With accurate
forecasted load, on the one hand,
the market operators can operate
the system in an economical
manner; on the other hand,
utilities can optimize their
resources for better profit. Load
forecasting in a power system can
tej.2008.09.016 The Electricity Journal
N
normally be segregated into three
categories:
� Short-term forecasting of up
to a few minutes ahead,
� Forecasting with a lead time
of up to a few days ahead,
� Long-term forecasting of the
power system.
Artificial neuralnetworks havebeen applied tomany areas of powersystem analysis,operation, andcontrol problems.
M any conventional
methods are available for
short-term forecasting, including
the time-of-day method,
regression methods, stochastic
time-series methods, and state-
space methods. Also available are
techniques based on artificial
intelligence (AI) such as expert
system-based methods. These
existing methods are capable of
forecasting short-term load but
suffer from several limitations.
Moreover, very few works
reported in the literature take care
of price as an input signal in
forecasting the load.
In recent years, artificial neural
networks (ANNs) have been
applied to many areas of power
system analysis, operation, and
control problems. These include
load forecasting,2 static and
dynamic security assessment,
dynamic load modeling, and
alarm processing and fault
diagnosis.3 The availability of
historical load data on the utility
databases makes this area highly
suitable for ANN implementation.
ANNs are able to learn the
relationship among past, current,
and future variables and loads
combining both time series and
regression approaches. As is the
case with the time series approach,
the ANN traces previous load
patterns and predicts (i.e.,
ovember 2008, Vol. 21, Issue 9 1040-6190/$–
extrapolates) a load pattern using
recent load data. It can also use
weather information for modeling.
The ANN is able to perform non-
linear modeling and adaptation. It
does not need assumption of any
functional relationship between
load and weather variables in
advance.
Modern load forecasting
techniques have been developed,
recently, showing encouraging
results. Among them, ANN
methods are particularly
attractive as they have the ability
to handle the non-linear
relationships between load and
the factors affecting it directly.4
ANN can perform better than
traditional methods, especially
during rapidly changing weather
conditions. Also the short time
required for their development
has made ANN-based load
forecasting models a very
attractive alternative.
T he most important input
variables, which affect load
forecasting, are weather
temperature and price, as there is
a strong correlation between them
and load. The Radial Basis
see front matter # 2008 Elsevier Inc. All rights
Function Network (RBFN) seems
to be very useful in short-term
load forecasting.5,6
In this article, two different
ANN models, the generalized
regression neural network
(GRNN) and probabilistic neural
network (PNN), have been used
for short-term load forecasting
using hour and day indicators,
weather temperature data, and
pricing signal as inputs. The
proposed approach is tested on
publicly available data of the
Victorian electricity market from
the Australian National
Electricity Market Management
Company (NEMMCO) Web site.
The results show that mean
absolute percentage error
(MAPE) in load forecasting is
reasonable; however, the results
obtained with GRNN are superior
to those using PNN.
II. Short-Term LoadForecasting Methods
The system load to be
forecasted is a random non-
stationary process composed of
thousands of individual
components. Therefore, the range
of possible approaches to the load
forecasting is wide. Some of the
most popular methods are
discussed below.
A. Time-of-day method
In the simplest form, a time-of-
day method takes the previous
week’s actual load pattern as a
model to predict the load of the
present week. Alternatively, a set
reserved., doi:/10.1016/j.tej.2008.09.016 25
26
of load patterns is stored for
typical weeks with different
weather conditions.7 These are
then heuristically combined to
create the forecast.
B. Regression method
The weaknessin the stochastic
models is in theiradaptability,
since loadbehavior can change
quite quickly.
Regression models, normally,
assume that the load can be
divided into a standard load
component and a component
linearly dependent on some
explanatory variables. The most
typical explanatory variables are
weather factors. A typical
regression model has been used by
Rasanen and Ruusunen.8 The load
is divided into a rhythm
component and a temperature-
dependent component. The
rhythm component corresponds
to the load of a certain hour in the
average temperature of the
modeling period. More
complicated model variations
have also been proposed. Some
models use earlier load values as
explanatory variables in addition
to external variables.9 Regression
methods are among the oldest
methods suggested for load
forecasting. They are quite
insensitive to occasional
disturbances in the measurements.
The easy implementation is the
strength of this method. The serial
correlation, which is typical when
regression models are used on
time series, can cause problems.
C. Stochastic time-series
methods
This is a very popular class of
dynamic forecasting models.10
1040-6190/$–see front matter # 2008 Els
There are many names
encountered in the literature for
the class, such as ARMA
(autoregressive–moving
average) models, ARIMA
(integrated autoregressive–
moving average) models, Box-
Jenkins method, and linear time-
series models. A general
treatment of the model type can
be found in Pindyck and
Rubenfeld.11 The basic principle
is that the load time series can
first be transformed into a
stationary time series (i.e.,
invariant with respect to time) by
a suitable differencing. Then the
remaining stationary series can
be filtered into white noise. The
methods assume that the
properties of the time series
remain unchanged for the period
used in model estimation, and all
disturbances are due to this
white noise component
contained in the identified
process.
T he basic ARIMA model is
not by itself suitable for
describing the load time series,
since the load series incorporates
seasonal variation. Therefore, the
evier Inc. All rights reserved., doi:/10.1016/j.
differencing with the period of
seasonal variation is required.
The model then obtained is called
a seasonal ARIMA (SARIMA)
model.
A n external input variable,
such as temperature in the
case of load time series, can also
be included in the model. Such a
variant of the ARIMA model is
called an ARIMAX model.
The ARIMA model includes
both, the seasonal variation and
external variable, and is
sometimes called a SARIMAX
model.
The stochastic time-series
models have many attractive
features. First, the theory of the
models is well known and
therefore it is easy to understand
how the forecast is composed.
The properties of the model are
easy to calculate; the estimate for
the variance of the white noise
component allows the confidence
intervals for the forecasts to be
created. The model identification
is also relatively easy.
Established methods for
diagnostic checks are available.
Moreover, the estimation of the
model parameters is quite
straightforward and the
implementation is not difficult.
The weakness in the stochastic
models is in their adaptability. In
reality, load behavior can change
quite quickly in certain parts of
the year. While in ARIMA
models the forecast for a certain
hour is in principle a function of
all earlier load values, the model
cannot adapt to the new
conditions very quickly, even if
model parameters are estimated
tej.2008.09.016 The Electricity Journal
N
recursively. A forgetting factor
can be used to give more weight
to the most recent behavior and
thereby improve the
adaptability. Another problem is
the handling of the anomalous
load conditions. If the load
behavior is abnormal on a certain
day, this deviation from the
normal conditions will be
reflected in forecasts into the
future. A possible solution to the
problem is to replace the
abnormal load values in the load
history by the corresponding
forecast values.
One possibilityis to classify D. State-space methodneural networkmodels on thebasis of thelearningprinciple.
There exist a number of
variations of the state-space
model. Some examples can be
found in Camp and Ruiz.12 In fact,
the basic state-space model can be
converted into an ARIMA model
and vice versa, so there is no
fundamental difference between
the properties of the two model
types. According to Gross and
Galiana, a potential advantage
over ARIMA models is the
possibility to use a priori
information in parameter
estimation via Bayesian
techniques.13 It is also pointed out
that the advantages are not very
clear and more experimental
comparisons are needed.
E. Expert systems
Expert systems are heuristic
models, which are usually able to
take both quantitative and
qualitative factors into account.
Many models of this type have
ovember 2008, Vol. 21, Issue 9 1040-6190/$–
been proposed since the mid-
1980s. A typical approach is to try
to imitate the reasoning of a
human operator. The idea is then
to reduce the analogical thinking
behind the intuitive forecasting to
formal steps of logic.14 A possible
method for a human expert to
create the forecast is to search in
history database for a day that
corresponds to the target day with
regard to the day type, social
factors, and weather factors. Then
the load values of this similar day
are taken as the basis for the
forecast. An expert system can
thereby be an automated version
of this kind of a search process.15
On the other hand, the expert
system can consist of a rule base
defining relationships between
external factors and daily load
shapes. Recently, a popular
approach has been used to
develop rules on the basis of
fuzzy logic.16 The heuristic
approach in arriving at solutions
makes the expert systems
attractive for system operators
where the system can provide the
user with the line of reasoning
followed by the model.17
see front matter # 2008 Elsevier Inc. All rights
III. Neural-Network-Based Load Forecasting
Neural networks, also known
as artificial neural networks
(ANNs), are inspired by
biological nervous systems.
ANNs are composed of many
computing elements, usually
denoted as neurons, working in
parallel. The elements are
connected by synaptic weights,
which are allowed to adapt
through a learning process.
Neural networks can be
interpreted as adaptive machines,
which can store knowledge
through the learning process. The
research in the field has a history
of many decades, but after a
diminishing interest in the 1970s,
a massive growth started in the
early 1980s. Today, neural
networks have applications, for
example, in pattern recognition,
identification, speech recognition,
vision, classification, and control
and power systems application
problems.
T here are many types of
neural network models,
which can be categorized in many
ways. One possibility is to classify
them on the basis of the learning
principle. A neural network uses
either supervised or
unsupervised learning. In
supervised learning, the network
is provided with example cases
and desired responses. The
network weights are then adapted
in order to minimize the
difference between network
outputs and desired outputs. In
unsupervised learning the
network is given only input
reserved., doi:/10.1016/j.tej.2008.09.016 27
Figure 1: Load and Price Changes during the Day
able 1: Correlation Factor betweenoad and Historical Price
pain Australia PJM
.8993 0.8517 0.9397
28
signals, and the network weights
change through a predefined
mechanism, which usually
groups the data into clusters
of similar data. The most
common network type using
supervised learning is a feed-
forward (signal transfer) network.
The most popular of all neural
networks, the Multi-Layer
Perceptron network (MLP), is of
this type.
I nterest in using ANNs for
forecasting has led to a
tremendous surge in research
activities in the past decade. They
can achieve complicated input–
output mappings without explicit
programming and extract
relationships (both linear and
non-linear) between data sets
presented during a learning
process. ANNs are massively
parallel, so that, in principle, they
are able to respond with high
speed. Furthermore, the
redundancy of their
interconnections ensures
robustness and fault tolerance,
and they can be designed to self
adapt and learn.18 The ANN
models are used in many power
system applications, with short-
term forecasting being one of the
most typical areas. Most of the
suggested models use MLP
networks.19 MLP forecasters,
models based on unsupervised
learning, have been suggested for
load forecasting.20 The purpose of
these models can be the
classification of the days into
different day types, or choosing
the most appropriate days in the
history to be used as the basis for
the actual load forecasting.
1040-6190/$–see front matter # 2008 Els
IV. Input Selection forANN
The most important work in
building an ANN load forecasting
model is the selection of input
variables. There is no general rule
that can be followed in this
process. It depends on
engineering judgment and
experience and is carried out
almost entirely by trial and error.
However, some statistical
analysis can be very helpful in
determining which variables have
significant influence on the
system load.
The factors affecting the load
forecasting can be represented as
L ¼ fðday;weather; special;
price; randomÞ
where f(�) is a highly non-linear
function.
I n general, variables like
hour and day indicators,
weather-related inputs
(temperature), and historical
loads are used as inputs to the
neural network. Some new
evier Inc. All rights reserved., doi:/10.1016/j.
variables like price can be more
important input in load
forecasting. In load forecasting
modeling, interdependence
between price and load can be the
deciding factor. This will be
reflected in pricing patterns of the
market. The price–load
relationship is neither linear nor
stationary in time but price–load
relationship may be relatively
stable over shorter periods of
time. Volatile electricity prices in
power markets are a new
phenomenon and this needs a
reliable solution. Since the
relationship between electricity
price and load is complex and
dynamic, further research is
needed to study how different
customers’ price response
characteristics and locations affect
the load forecasting. It is clear
from Figure 1 and Table 1 that
TL
S
0
tej.2008.09.016 The Electricity Journal
N
there is a strong correlation
between load and price; hence,
price will be a major deciding
factor for load forecasting. This is
why in this analysis only price
with hour and day indicators has
been taken as input.
V. Proposed NeuralNetwork Models forLoad Forecasting
Figure 2: Generalized Regression Neural Network (GRNN)
The well-known Radial BasisFunction Network (RBFN) having
special features is used here as a
primary test in this application.
An RBFN consists of two layers, a
hidden layer with non-linear
neurons and an output layer with
linear neurons. Thus the
transformation from the input
space to the hidden unit space is
non-linear whereas the
transformation from the hidden
unit space to the output space is
linear. Two types of well-known
RBFNs, which have several merits
over other ANN models, have
been used in this work for load
forecasting.
A. Generalized Regression
Neural Network
G eneralized Regression
Neural Network (GRNN)
is a new kind of neural
network that Donald F. Specht
put forward in 1991. At present,
this neural network has found
applications in system
distinguishing, prediction, and
the like.
The architecture for the
GRNN is similar to the radial
ovember 2008, Vol. 21, Issue 9 1040-6190/$–
basis network, but has a
slightly different second layer.
The first layer has as many
neurons as there are input/target
vectors. Each neuron’s weighted
input is the distance between the
input vector and its weight
vector. Each neuron’s net input is
the product of its weighted input
with its bias. Each neuron’s
output is its net input passed
through radial basis layer. The
second layer also has as many
neurons as input/target vectors.
A larger spread (distance) leads
to a large area around the input
vector where layer 1 neurons will
respond with significant outputs.
Therefore, if the spread is small
the radial basis function is
very steep so that the neuron
with the weight vector closest
to the input will have a much
larger output than other
neurons. The network will
tend to respond with the target
vector associated with the
nearest design input vector.
As the spread gets larger, the
radial basis function’s slope
see front matter # 2008 Elsevier Inc. All rights
gets smoother and several
neurons may respond to an
input vector. The network
then acts like it is taking a
weighted average among target
vectors whose design input
vectors are closest to the new
input vector. As the spread gets
larger, more and more neurons
contribute to the average with
the result that the network
function becomes smoother
(Figure 2).
B. Probabilistic neural
network
Probabilistic neural network is
a kind of radial basis network
suitable for classification
problems.21 The probabilistic
neural network (PNN) constitutes
an alternative approach for class
conditional density estimation. It
is an RBF-like neural network
adapted to provide output values
corresponding to the class
conditional densities. Since the
network is RBF, the components
(hidden units) are shared among
reserved., doi:/10.1016/j.tej.2008.09.016 29
Figure 4: The ANN Model Used in this Application
Figure 3: Probabilistic Neural Network (PNN)
30
classes and each class conditional
density is evaluated using not
only the corresponding class data
points (as in the case of separate
mixtures) but also all the available
data points. Probabilistic neural
networks can be used for
classification problems. When an
input is presented, the first layer
computes distances from the
input vector to the training input
vectors, and produces a vector
whose elements indicate how
close the input is to a training
input. The second layer sums
these contributions for each class
of inputs to produce as its net
output vector of probabilities.
Finally, a complete transfer
function on the output of the
second layer picks the maximum
of these probabilities, and
produces a 1 for that
class and a zero for the other
classes.
A n example of a
probabilistic neural
network is shown in Figure 3.
It has three layers. The network
contains an input layer, which
has as many elements as there
are separable parameters needed
to describe the objects to be
classified. It has a pattern layer,
which organizes the training set
1040-6190/$–see front matter # 2008 Els
such that an individual
processing element represents
each input vector. It contains an
output layer, called the
summation layer, which has as
many processing elements as
there are classes to be
recognized. Each element in
this layer combines via
processing elements within the
pattern layer, which relate to the
same class and prepares that
category for output. The transfer
function is radial basis function
for the first layer and is
competitive function for the
evier Inc. All rights reserved., doi:/10.1016/j.
second layer. Only the first layer
has biases.
VI. ANN Training: Dataand Algorithm
GRNN and PNN, which were
used for forecasting short-term
load, in this article were tested
on publicly available data from
the NEMMCO Web site to
forecast electricity prices and
loads for the Victorian electricity
market in Australia. The
Australian National Electricity
Market (NEM) is the deregulated
electricity supply industry
covering Victoria, New South
Wales, Queensland, South
Australia, and the Australian
Capital Territory. The data of
2006 is divided into several
windows where half of them
(non-consecutive ones) are used
for training and the other half is
used for testing the ANN.
tej.2008.09.016 The Electricity Journal
Figure 5: Output Compared to Actual Load Using GRNN
N
More precisely, for each month,
the first week and the third
week are used for training,
while the second and fourth
weeks are left for testing the
ANN. Training was done for
all the data windows at the
same time; i.e., the same ANN
is trained to be used at any
time during the year. All
inputs and outputs are
normalized before training.
Cross-correlation between load
and price is found and only those
price inputs were considered
which are best correlated. The
inputs to the ANN as shown in
Figure 4 are:
� H(k) hour indicator
� D(k) day indicator
� P(k) Price at hour k
� P(k � 1/2) Price at 30 min-
utes before hour k
� P(k + 1/2) Price at 30 min-
utes after hour k
� P(k � 1) Price at 60 minutes
before hour k
� P(k � 1) Price at 60 minutes
before hour k
� T(max) Maximum tempera-
ture of the day
� T(min) Minimum tempera-
ture of the day
VII. Simulation andResults
Figure 6: Output Compared to Actual Load Using PNN
The algorithm resulted in a
very fast training due to its
special features as explained in
the previous sections and the
error was significantly reduced
to very low value. Then the
performance of the developed
ANN model for load profile
ovember 2008, Vol. 21, Issue 9 1040-6190/$–
forecasting was tested using
windows of data that were not
included in the training set. The
forecasted hourly load for the
several days was estimated but
due to limited space, only two
cases are presented here which
gave more error in the
see front matter # 2008 Elsevier Inc. All rights
prediction. Figures 5 and 6 show
the actual and forecasted load for
Jan. 17, 2006, with GRNN and
PNN. It is clear from the results
that forecasted load patterns are
similar to the actual one. The
percentage error in load
forecasting by GRNN and PNN
reserved., doi:/10.1016/j.tej.2008.09.016 31
Figure 7: Error in Load Forecasting Using GRNN
able 2: MAPE (%) in Case of BothRNN and PNN
MAPE (%)
ay PNN GRNN
unday 8.94 2.72
onday 9.51 2.63
uesday 2.96 2.34
ednesday 6.79 3.64
hursday 9.07 4.00
riday 2.23 1.80
aturday 3.98 2.86
32
methods is shown in Figures 7
and 8, respectively.
ANN performance is checked
with the mean absolute
percentage error (MAPE) as
Figure 8: Error in Load Forecasting Using PN
1040-6190/$–see front matter # 2008 Els
defined below:
MAPEð%Þ ¼ 1
N
XN
i¼1
LiF � Li
A
�� ��
LiA
� 100 (2)
N
evier Inc. All rights reserved., doi:/10.1016/j.
TG
D
S
M
T
W
T
F
S
tej.2008.09.016 The Electricity Journ
where LA and LF are the
actual and forecasted load,
respectively. N is the number of
hours.
T he maximum error in
load forecasting is 7.62
percent and 6.45 percent in the
GRNN and PNN models,
respectively, which can be
seen from Figures 7 and 8.
Comparison of mean absolute
percentage error (MAPE) in the
case of both GRNN and PNN is
given in Table 2. The maximum
value of MAPE is 4.0 percent in
the case of GRNN whereas it is
9.51 percent in the case of PNN
during different days of the week.
From Table 2, it can be seen that
the performance of GRNN is
better than PNN.
A comparison of different
methods used for load
forecasting in recent years has
been done as shown in Table 3.
The data has been collected
from various published
literatures.22 The market
considered in all the cases is
different. It is clear from the table
that average MAPE (avg. MAPE)
for GRNN is better than all the
methods.
al
Table 3: Comparison of MAPE (%) Using Different Methods for Load Forecasting inRecent Years
Methods Max. MAPE Min. MAPE Avg. MAPE
GRNN 3.14 0.14 1.85
PNN 11.30 4.27 8.24
MLP 3.15 0.80 2.43
Back Propagation 3.48 0.73 2.10
ANN-Fuzzy 4.28 1.00 2.00
Multi-stage ANN-STLF 6.39 2.81 4.85
SOM-SVM Hybrid 2.68 1.34 2.06
GA 2.75 0.7 2.43
LAVF 3.89 1.69 2.79
LS 11.5 3.76 7.63
Back Propagation 3.27 1.73 2.53
SVM 6.10 1.50 2.71
Dual-SVM Hybrid 3.62 1.21 2.10
ARMA 10.34 1.53 4.77
Recurrent ANN 4.10 1.39 2.08
N
VIII. Conclusion
Load forecasting in the
emerging electricity market
plays a very important role for
the economic and secure
operation of power systems.
ANN methods are particularly
attractive in the case of load
forecasting as they have the
ability to handle the non-linear
relationships between load and
the factors affecting it directly.
Volatile electricity price in power
markets is a major input for load
forecasting using ANN. In this
article, two types of neural
networks, known as generalized
regression neural network
(GRNN) and probabilistic neural
network (PNN), have been used.
The results show the
effectiveness of the proposed
model, as the forecasted load is
very close to the actual load. The
maximum error in load
ovember 2008, Vol. 21, Issue 9 1040-6190/$–
forecasting is 7.62 percent and
6.45 percent in the GRNN and
PNN models, respectively,
whereas the maximum value of
MAPE is 4.0 percent in the case
of GRNN and 9.51 percent in the
case of PNN during different
days of the week. It is also clear
from the comparison of all other
methods used for load
forecasting in recent years that
avg. MAPE for GRNN is 1.85
percent, which is better than all
other methods.&
Endnotes:
1. G. Gross and F.D. Galiana, Short-Term Forecasting, Proceedings of IEEE,75(12), 1987, at 1558–73.
2. D.C. Park, M.A. El-Sharakawi andRi Marks II, Electric Load ForecastingUsing Artificial Neural Networks, IEEETRANS. POWER SYS., 6(2), 1991, at 442–449.
3. N. Kandil, V.K. Sood, K. Khorasaniand R.V. Patel, Fault Identification in
see front matter # 2008 Elsevier Inc. All rights
an AC–DC Transmission SystemUsing Neural Networks, IEEE TRANS.POWER SYS., 7(2), 1992, at 812–9.
4. Nahil Kandil, R. Wamkeue, M.Saad and S. Georges, An EfficientApproach for Short-Term LoadForecasting Using Artificial NeuralNetworks, J. ELEC. POWER & ENERGY
SYS., 28 (2006): 525-530.
5. D.K. Ranaweera, N.F. Hubele andA.D. Papalexopoulos, Application ofRadial Basis Function Neural NetworkModel for Short Term LoadForecasting, IEE PROC. GENERATION,TRANSMISSION & DISTRIBUTION, Vol. 142,No. 1, Jan. 1995.
6. C. Constantinopoulos and A. Likas,An Incremental Training Method for theProbabilistic RBF Network, IEEE TRANS.ON NEURAL NETWORKS, Vol. 17, No. 4,July 2006.
7. Gross and Galiana, supra note 1.
8. M Rasanen and J Ruusunen,Verkoston Tilan Seurantamittauksillaja Kuormitusmalleilla, ResearchReport B17, Systems AnalysisLaboratory, Helsinki Univ. ofTechnology (in Finnish).
9. Id.
10. M.T. Hagan and S.M. Behr, TheTime Series Approach to Short Term LoadForecasting, IEEE TRANS. ON POWER SYS.,Vol. PWRS-2, No. 3, Aug. 1987, at785-791.
11. R.S. PINDYCK AND D.L. RUBINFELD,
ECONOMETRIC MODELS AND ECONOMIC
FORECASTS (Singapore: McGraw-Hill,1981).
12. R. Campo and P. Ruiz, AdaptiveWeather-Sensitive Short-Term LoadForecast, IEEE TRANS. ON POWER SYS., Vol.PWRS-2, No. 3, Aug. 1987, at 592-600.
13. See Gross and Galiana, supranote 1.
14. S. Rahman, and R. Bhatnagar, AnExpert System Based Algorithm for Short-Term Load Forecast, IEEE TRANS. ON
POWER SYS., Vol. 3, No. 2, May 1988, at392-399.
15. K Jabbour, J.F.V. Riveros, D.Landsbergen and W. Meyer, ALFA,Automated Load ForecastingAssistant, IEEE TRANS. ON POWER SYS.,Vol. 2, No. 3, Aug. 1988, at 908-914.
reserved., doi:/10.1016/j.tej.2008.09.016 33
34
16. Y.-Y. Hsu, and K.L. Ho, FuzzyExpert Systems: An Application to Short-Term Load Forecasting, IEEEProceedings - C, Vol. 139, No. 6, Nov.1992, at 471-477.
17. A. Asar and J.R. McDonald, ASpecification of Neural NetworkApplications in the Load ForecastingProblem, IEEE Transactions on ControlSys. Tech., Vol. 2, No. 2, Nov. 1992, at135-141.
18. B. SOUCEK AND M SOUCEK, NEURAL
AND MASSIVELY PARALLEL COMPUTERS:THE SIXTH GENERATION (New York:Wiley, 1988); Z. Guoqiang, B. EddyPatuwo and M.Y. Hu, Forecasting withArtificial Neural Networks: The State ofthe Art, INTL. J. FORECASTING, 14, 1998, at35–62; and B. WIDROW AND D. STEMS,ADAPTIVE SIGNAL PROCESSING (NewYork: Prentice-Hall, 1985).
19. K.L. Ho, Y.Y. Hsu and C.C. Yang,Short-Term Load Forecasting Using aMultilayer Neural Network with anAdaptive Learning Algorithm, IEEETRANSACTIONS ON POWER SYS., Vol.7, No.1, Feb. 1992, at 141-148; T.M. Peng,N.F. Hubele and G.G. Karady,Advancement in the Application of NeuralNetworks for Short-Term LoadForecasting, IEEE TRANSACTIONS ON
POWER SYS., Vol. 7, No. 1, Feb. 1992, at
Load forecasting in
1040-6190/$–see front matter # 2008 Els
250-256; and O. Mohammed, D. Park,R. Merchant, T. Dinh, C. Tong, A.Azeem, J. Farah and C. Drake,Practical Experiences with anAdaptive Neural Network Short-TermLoad Forecasting System, IEEETRANSACTIONS ON POWER SYS., Vol. 10,No. 1, Feb. 1995, at 254-265.
20. M. Djukanovic, B. Babic, D.J.Sobajic and Y.H. Pao, Unsupervised/Supervised Learning Concept for 24-HourLoad Forecasting, IEE Proceedings-C,July 1993, Vol. 140, No. 4, at 311-318.
21. C. Constantinopoulos and A.Likas, An Incremental Training Methodfor the Probabilistic RBF Network, IEEETRANSACTIONS ON NEURAL NETWORKS,Vol. 17, No. 4, July 2006.
22. These include P. Mandal, T.Senjyu, N. Urasaki and T. Funabashi,A Neural Network Based Several-Hour-Ahead Electric Load Forecasting UsingSimilar Days Approach, ELEC. POWER &ENERGY SYS. 28 (2006) at 373; P. Mandal,T. Senjyu and T. Funabashi, NeuralNetworks Approach to Forecast SeveralHour Ahead Electricity Prices and Loadsin a Deregulated Market, ENERGY
CONVERSION & MGMT. 47 (2006) at 2128–2142; P.K. Dash, A.C. Liew andS.Rahman, Fuzzy Neural Network andFuzzy Expert System for Loads
the emerging electricity market plays a very i
evier Inc. All rights reserved., doi:/10.1016/j.
Forecasting, 114 IEEE PROCEEDINGS –GENERATION, TRANSMISSION &DISTRIBUTION, Vol. 143, No. 1, Jan. 1996;K. Methaprayoon, W.J. Lee, S.Rasmiddatta, J. Liao and R. Ross,Multi-Stage Artificial Neural NetworkShort-Term Load Forecasting Engine withFront-End Weather Forecast, 1-4244-0336-7/2006 IEEE; S. Fan and L. Chen,Short-Term Load Forecasting Based on anAdaptive Hybrid Method, 392 IEEETRANSACTIONS POWER SYS., Vol. 21, No.1, Feb. 2006; K.M. EL-Naggar and K.A.Al-Rumaih, Electric Load ForecastingUsing Genetic Based Algorithm, OptimalFilter Estimator and Least Error SquaresTechnique: Comparative Study,PROCEEDINGS OF WORLD ACAD. OF SCI.,ENG’G. & TECH., Vol. 6, June 2005; G.A.Adepoju, S.O.A. Ogunjuyigbe andK.O. Alawode, Application of NeuralNetwork to Load Forecasting in NigerianElectrical Power System, PACIFIC J. OF
SCI. & TECH. 8(1) 2007, at 68-72; L. Ao,Y. Wang, Q. Zhang, Application of aHybrid Model on Short-Term LoadForecasting Based on Support VectorMachines (SVM), NEW ZEALAND J.AGRICULTURAL RES. 2007, Vol. 50, at567–572; and A.K. Topalli, I. Erkmenand I. Topalli, Intelligent Short-TermLoad Forecasting in Turkey, ELEC.POWER & ENERGY SYS. 28 (2006),at 437–447.
mportant role.
tej.2008.09.016 The Electricity Journal