524
1, FEBRUARY 2010
A Hybrid ARIMA and Neural Network Model for
Short-Term Price Forecasting in Deregulated Market
Phatchakorn Areekul, Student Member, IEEE, Tomonobu Senjyu, Senior Member, IEEE, Hirofumi Toyama, and
Atsushi Yona, Member, IEEE
Abstract—In the framework of competitive electricity markets,
power producers and consumers need accurate price forecasting
tools. Price forecasts embody crucial information for producers
and consumers when planning bidding strategies in order to maxi-
mize their benefits and utilities, respectively. The choice of the fore-
casting model becomes the important influence factor on how to
improve price forecasting accuracy. This paper provides a hybrid
methodology that combines both autoregressive integrated moving
average (ARIMA) and artificial neural network (ANN) models for
predicting short-term electricity prices. This method is examined
by using the data of Australian national electricity market, New
South Wales, in the year 2006. Comparison of forecasting perfor-
mance with the proposed ARIMA, ANN, and hybrid models are
presented. Empirical results indicate that a hybrid ARIMA-ANN
model can improve the price forecasting accuracy.
Index Terms—Artificial neural networks (ANNs), autoregres-
sive integrated moving average (ARIMA), Australian national
electricity market (NEM), electricity, price forecasting.
I. INTRODUCTION
HE deregulated power market is an auction market, and
T the market clearing price (MCP) are volatile. On the other
hand, due to the upheaval of deregulation in electricity markets,
price forecasting has become a very valuable tool. The compa-
nies that trade in electricity markets make extensive use of price
prediction techniques either to bid or to hedge against volatility.
When bidding in a pool system, the market participants are re-
quested to express their bids in terms of prices and quantities.
Since the bids are accepted in order of increasing price until the
total demand is met, a company that is able to forecast the pool
price can adjust its own price/production schedule depending on
hourly pool prices and its own production costs. High-quality
MCP prediction and its confidence interval estimation can help
utilities and independent power producers submit effective bids
with low risks.
Price forecasting has become increasingly necessary and
more complex for all kinds of market participants, as evident
from the various approaches that exist today, partly because
ongoing market reforms create continuous changes in the
Manuscript received May 08, 2009; revised July 05, 2009. First published
December 15, 2009; current version published January 20, 2010. Paper no.
TPWRS-00080-2009.
The authors are with the Faculty of Engineering, Department of Electrical
and Electronics Engineering, University of the Ryukyus, Okinawa 903-0213,
Japan (e-mail: k088662@eve.u-ryukyu.ac.jp; phatokinawa@hotmail.com;
b985542@tec.u-ryukyu.ac.jp).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TPWRS.2009.2036488
0885-8950/$26.00 © 2009 IEEE
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 25, NO.
dynamics of prices, and partly because electricity prices exhibit
complex volatility patterns. A wide variety of models have been
proposed in the last two decades, such as linear regression,
time-series techniques based upon autoregressive integrated
moving average (ARIMA), Kalman filtering, data mining
approach, and state-space methods [1]–[7].
One of the most important and widely used time-se-
ries models is the ARIMA model. The popularity of the
ARIMA model is due to its statistical properties, as well as
the well-known Box-Jenkins methodology [8] in the model
building process. In addition, various exponential smoothing
models can be implemented by ARIMA models [9], i.e., a
linear correlation structure is assumed among the time-series
values, and therefore, no nonlinear patterns can be captured
by the ARIMA model. The approximation of linear models to
complex real-world problem is not always satisfactory.
Recently, artificial neural networks (ANNs) have been exten-
sively studied and used in time-series forecasting. Zhang et al.
[10] presented a recent review in this area. The major advantage
of neural networks is their flexible nonlinear modeling capa-
bility. With ANNs, there is no need to specify a particular model
form. Rather, the model is adaptively formed based on the fea-
tures presented from the data. This data-driven approach is suit-
able for many empirical datasets where no theoretical guidance
is available to suggest an appropriate data generating process.
Using hybrid model or combining several models has become
a common practice to improve the forecasting accuracy, since
the well-known M-competition [11] in which combination of
forecasts from more than one model often leads to improved
forecasting performance. The basic idea of the model combina-
tion in forecasting is to use each model’s unique feature to cap-
ture different patterns in the data. Both theoretical and empirical
findings suggest that combining different methods can be an ef-
fective and efficient way to improve forecasts [12]–[14]. How-
ever, combination of the proposed techniques in the framework
of the hybrid forecast method and especially its application for
MCP prediction can be considered as the contribution of this
paper.
This paper presents a successful application of using a hybrid
approach based for time-series price forecasting considering the
historical prices of the Australian national electricity market
(NEM), New South Wales, in year 2006 for both ARIMA and
ANN models, i.e., this paper provides ARIMA and a three-lay-
ered feedforward ANN, trained by the Levenberg-Marquardt al-
gorithm. A case study presented with the historical data for hy-
brid model by real market data in the year 2006 is used for fore-
casting the next 168 h (one week) electricity prices.
AREEKUL et al.: HYBRID ARIMA AND NEURAL NETWORK MODEL
The rest of this paper is organized as follows. In Section II,
review the ARIMA and ANN modeling approaches to time-se-
ries forecasting is presented. The hybrid methodology is intro-
duced in Section III. Empirical results from three real datasets
are reported in Section IV. Section V contains the concluding
remarks.
II. TIME-SERIES FORECASTING MODELS
There are several different approaches to time-series mod-
eling. Traditional statistical models including moving average
(MA), exponential smoothing, and ARIMA are linear in that
predictions of the future values are constrained to be linear func-
tions of past observations. To overcome the restriction of the
linear models and account for certain nonlinear patterns ob-
served in real problems, several classes of nonlinear models
have been proposed in the literature. These include the bilinear
model [15], the threshold autoregressive (TAR) model [16], and
the autoregressive conditional heteroscedastic (ARCH) model
[17]. Although some improvement has been noticed with these
nonlinear models, the gain of using them for general forecasting
problems is limited [18]. More recently, ANNs have been sug-
gested as an alternative to time-series forecasting. The main
strength of the ANNs is their flexible nonlinear modeling ca-
pability. In this section, we focus on the basic principles and
modeling process of the ARIMA and ANN models.
A. ARIMA Model
In an ARIMA model, the future value of a variable is assumed
to be a linear function of several past observations and random
errors, i.e., the underlying process that generate the time series
has the form
(1)
where and are the actual value and random error
at time period , respectively, and
are model parameters, and and
are integers and often referred to as orders of the model.
Random errors are assumed to be independently and iden-
tically distributed with a mean of zero and a constant variance
of .
Equation (1) entails several important special cases of the
ARMA family of models. If , then (1) becomes an AR
model of order . When , the model reduces to a MA
model of order . One extension to the ARMA class of
processes, which greatly enhance their value as empirical de-
scriptors of nonstationary time series, is the class of ARIMA.
Nonstationary time series processes can be transformed, by dif-
ferencing the series once or more, to make them stationary. The
number of times that the integrated process must be differ-
enced to make a time series stationary is said to be the order
of the integrated process. In this case, the series is called
an autoregressive (AR) integrated MA process of order
and denoted by ARIMA . and output layer.
525
Based on the earlier work of Yule [19] and Wold [20], Box
and Jenkins [8] developed a practical approach to building
ARIMA models, which has the fundamental impact on the
time-series analysis and forecasting applications. This method-
ology includes three iterative steps of model identification,
parameter estimation, and diagnostic checking. The basic idea
of model identification is that if a time series is generated from
an ARIMA process, it should have theoretical autocorrelation
properties. By matching the empirical autocorrelation patterns
with the theoretical ones, it is often possible to identify one or
several potential models for the given time series.
Box and Jenkins [8] proposed to use the autocorrelation func-
tion (ACF) and the partial ACF (PACF) of the sample data as
the basic tools to identify the order of the ARIMA best model;
data transformation is often needed to make the time series sta-
tionary. This involves selecting the most appropriate lags for the
AR and MA parts, as well as determining if the variable requires
first-differencing to induce stationarity.
Once a tentative model is specified, estimation of the model
parameters is straightforward. The parameters are estimated
such that an overall measure of errors is minimized. This
usually involves the use of a least-squares estimation process.
The last step of model building is the diagnostic checking of
model adequacy. This is basically to check if the model assump-
tions about the errors are satisfied. Several diagnostic statis-
tics and plots of the residuals can be used to examine the good-
ness of fit of the tentatively entertained model to the historical
data. If the model is not adequate, a new tentative model should
be identified, which is again followed by the steps of parameter
estimation and model verification. Diagnostic information may
help suggest alternative model(s).
This three-step model building process is typically repeated
several times until a satisfactory model is finally selected. The
final selected model can then be used for prediction purposes.
B. ANNs Approach to Time-Series Modeling
When the linear restriction of the model form is relaxed, the
possible number of nonlinear structures that can be used to de-
scribe and forecast a time series is enormous. A good non-
linear model should be “general enough to capture some of the
nonlinear phenomena in the data” [18]. ANNs are one of such
models that are able to approximate various nonlinearities in the
data.
ANNs are flexible computing frameworks for modeling a
broad range of nonlinear problems. One significant advantage
of the ANN models over other classes of nonlinear model is
that ANNs are universal approximators that can approximate a
large class of functions with a high degree of accuracy. Their
power comes from the parallel processing of the information
from the data. No prior assumption of the model form is
required in the model building process. Instead, the network
model is largely determined by the characteristics of the data.
Network architecture of ANNs consists of the three-layer fully
connected feedforward neural network, which is depicted in
Fig. 1 and is used here; it includes input layer, hidden layer,
526
Fig. 1. Neural network model.
Single hidden-layer feedforward network is the most
widely used models form for time-series modeling and fore-
casting [10]. The model is characterized by a network of
three layers of simple processing units connected by acyclic
links. The relationship between the output and the inputs
has the following mathematical repre-
sentation:
(2)
where and
are the model parameters, often called the con-
nection weights, is the number of input nodes, and is the
number of hidden nodes. The logistic function is often used as
the hidden-layer transfer function, i.e.
(3)
Hence, the ANN model of (2), in fact, performs a non-
linear functional mapping from the past observations
to the future value , i.e.
(4)
where a vector of all parameters, and is a function deter-
mined by the network structure and connection weights. Thus,
the neural network is equivalent to a nonlinear AR model. Note
that expression (2) implies one output node in the output layer
that is used for one-step-ahead forecasting. The simple network
given by (2) is surprisingly powerful in that it is able to ap-
proximate arbitrary function as the number of hidden nodes
is sufficiently large [21]. In practice, simple network structure
that has a small number of hidden nodes often works well in
out-of-sample forecasting. This may be due to the overfitting
effect typically found in neural network modeling process. An
overfitted model has a good fit to the sample used for model
building, but has poor generalization ability for data out of the
sample. The choice of is data-dependent and there is no sys-
tematic rule in deciding this parameter.
In addition to choosing an appropriate number of hidden
nodes, another important task of ANN modeling of a time
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 25, NO. 1, FEBRUARY 2010
series is the selection of the number of lagged observations
and the dimension of the input vector. This is perhaps the
most important parameter to be estimated in an ANN model
because it plays a major role in determining the (nonlinear)
autocorrelation structure of the time series. However, there is
no theory that can be used to guide the selection of . Hence,
experiments are often conducted to select an appropriate , as
well as .
Once a network structure is specified, the network is
ready for training—a process of parameter estimation. As in
ARIMA model building, the parameters are estimated such that
an overall accuracy criterion such as the mean squared error
is minimized. This is again done with some efficient nonlinear
optimization algorithms other than the basic backpropagation
training algorithm. One of these is generalized reduced gradient
(GRG2), a general-purpose nonlinear optimizer [22]. A GRG2-
based training system is used in this study [23].
The estimated model is usually evaluated using a separate
hold-out sample that is not exposed to the training process. This
practice is different from that in ARIMA model building, where
one sample is typically used for model identification, estima-
tion, and evaluation. The reason lies in the fact that the general
(linear) form of the ARIMA model is prespecified, and then,
the order of the model is estimated from the data. The standard
statistical paradigm assumes that under stationary condition, the
model best fitted to the historical data is also the optimum model
for forecasting. With ANNs, the (nonlinear) model form, as well
as the order of the model, must be estimated from the data. It is,
therefore, more likely for an ANN model to overfit the data.
There are some similarities between ARIMA and ANN
models: both of them include a rich class of different models
with different model orders, data transformation is often
necessary to get best results, a relatively large sample is re-
quired in order to build a successful model, and the iterative
experimental is common to their modeling processes and the
subjective judgement is sometimes needed in implementing
the model. Because of the potential overfitting effect with both
models, parsimony is often a guiding principle in choosing an
appropriate model for forecasting.
III. HYBRID METHODOLOGY
Both ARIMA and ANN models have achieved successes in
their own linear or nonlinear domains. However, none of them
is a universal model that is suitable for all circumstances. The
approximation of ARIMA models to complex nonlinear prob-
lems may not be adequate. On the other hand, using ANNs to
model linear problems have yielded mixed results. For example,
using simulated data, Denton [24] showed that when there are
outliers or multicollinearity in the data, neural networks can sig-
nificantly outperform linear regression models. Markham and
Rakes [25] also found that the performance of ANNs for linear
regression problems depends on the sample size and noise level.
Hence, it is not wise to apply ANNs blindly to any type of data.
Since it is difficult to completely know the characteristics of the
data in a real problem, hybrid methodology that has both linear
and nonlinear modeling capabilities can be a good strategy for
practical use. By combining different models, different aspects
of the underlying patterns may be captured.
AREEKUL et al.: HYBRID ARIMA AND NEURAL NETWORK MODEL
It may be reasonable to consider a time series to be composed
of a linear autocorrelation structure and a nonlinear component,
i.e.
(5)
where denotes linear component and denotes the non-
linear component. These two components have to be estimated
from the data. First, we let ARIMA to model the linear compo-
nent, then the residuals from the linear model will contain only
the nonlinear relationship. Let denote the residuals at time
from the linear model, then
(6)
where is the forecast value for time from the estimated re-
lationship (2). Residuals are important in diagnosis of the suffi-
ciency of linear models. A linear model is not sufficient if there
are still linear correlation structures left in the residuals. How-
ever, residual analysis is not able to detect any nonlinear patterns
in the data. In fact, there is currently no general diagnostic statis-
tics for nonlinear autocorrelation relationships. Therefore, even
if a model has passed diagnostic checking, the model may still
not be adequate in that nonlinear relationships have not been
appropriately modeled. Any significant nonlinear pattern in the
residuals will indicate the limitation of the ARIMA. By mod-
eling residuals using ANNs, nonlinear relationships can be dis-
covered. With input nodes, the ANN model for the residuals
will be
(7)
where is a nonlinear function determined by neural network
and is the random error. If the model is not an appropriate
one, the error term is not necessarily random. Therefore, the cor-
rect model identification is critical. Denoting the forecast from
(7) as , the combined forecast will be
(8)
In summary, the proposed methodology of the hybrid system
consists of two steps. In the first step, an ARIMA model is
used to analyze the linear part of the problem. ARIMA model
building can be summarized in four steps, which are as follows.
1) Model identification: Using graphs, statistics, ACF, PACF,
transformations, etc., achieve stationary and tentatively
identify patterns and model components.
2) Parameter estimation: Determine the model coefficients
through the method of least squares, maximum likelihood
methods and other techniques.
3) Model diagnostics: Determine if the model is valid. If
valid, then use the model; otherwise, repeat identification,
estimation, and diagnostic steps.
4) Forecast verification and reasonableness: It is necessary
to revisit the question of identification to see if the selected
model can be improved. Several statistical techniques and
527
confidence intervals determine the validity of forecasts and
track model performance to detect out of control situations.
In the second step, a neural network model is developed to
model the residuals from the ARIMA model. Since the ARIMA
model cannot capture the nonlinear structure of the data, the
residuals of linear model will contain information about the non-
linearity.
The results from the neural network can be used as predic-
tions of the error terms for the ARIMA model. The hybrid model
exploits the unique feature and strength of ARIMA model, as
well as ANN model, in determining different patterns. Thus, it
could be advantageous to model linear and nonlinear patterns
separately by using different models, and then combine the fore-
casts to improve the overall modeling and forecasting perfor-
mance.
To evaluate the accuracy of forecasting electricity prices, dif-
ferent criterions are used. This accuracy is computed in func-
tion of the actual market prices that occurred. The average per-
centage error in the entire series is a general measure of fit useful
in comparing the fits of different models. This measure adds up
all of the percentage errors at each time point and divides them
by the number of time points. This measure is sometimes abbre-
viated mean percentage error (MPE):
% (9)
Therefore, we would compare its forecasts with those of alter-
native methods. There are different alternative method for this
purpose : the mean absolute percentage error (MAPE criterion),
mean absolute error (MAE criterion) and rms error (RMSE cri-
terion), which are defined as follows:
% (10)
(11)
(12)
where
actual value of price at hour;
forecasted value of price at hour;
number of hour for forecasting.
IV. EMPIRICAL RESULTS
A. Datasets
The proposed a hybrid ARIMA and neural network model is
performed using data of Australian NEM, New South Wales,
[26] in year 2006 is considered in this real-world case study.
The price curves is given in Fig. 2.
Moreover, models similar to the one reported in this paper
are routinely used by the power industry in Australia for price
528 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 25, NO. 1, FEBRUARY 2010

TABLE I
COMPARISON OF FORECASTING RESULTS WITH ARIMA MODEL

Fig. 2. Electricity prices : January to December 2006 in the Australian NEM.

forecasting. It should be noted that the electricity market of Aus- TABLE II

tralia is a duopoly with a dominant player. This results in price COMPARISON OF FORECASTING RESULTS WITH ARIMA-ANN MODEL
changes related to the strategic behavior of the dominant player,
which are hard to predict.
The data have been extensively studied with a vast variety
of linear and nonlinear time-series models, including ARIMA
and ANN. To assess the forecasting performance of different
models, each dataset is divided into two samples of training and
testing. The training dataset is used exclusively for model de-
velopment, and then the test sample is used to evaluate the es-
tablished model.
To illustrate the behavior of the proposed technique, results
comprising four weeks corresponding to the four seasons of year
2006 are presented. In this manner, representative results for the
whole year are provided.
For the sake of a fair comparison, the fourth week in each
of the seasons of January, May, August, and October are se-
lected, i.e., weeks with particularly good price behavior are pur-
posely not sought. To build the forecasting model for each one of
the considered weeks, the information available includes hourly
price historical data of the four weeks previous to the first day
of the week, whose prices are to be predicted.

model in each season to estimate nonlinear composite. We

B. Results
In this study, the approach for analysis and building a time-se-
ries model is a method of finding, for a model (ARIMA) that
adequately represents a data generating process for a given set
of data. The test time-series data was processed by taking the
first-order regular difference and the first seasonal difference in
order to remove the growth trend and the seasonality charac-
teristics. We checked stationary of time series of each of the
seasons, identified ARIMA (p, d, and q) model by comparing
characteristic of ACF and PACF (Correlogram testing), esti-
mated parameters by ordinary least-squares method (OLS), and
checked hypothesis of the model are validated by Akaike infor-
mation criterion (AIC), which was used to determine the best
model and final forecasted 168 h (one week) in each of the sea-
sons. Table I presented the values of RMSE, MAE, and MAPE
from price forecasting performance of the proposed ARIMA
model.
In this section, artificial neural network (ANN) model is used
to model the residuals obtained from forecasting by ARIMA
propose an ANN and hybrid (ARIMA-ANN) models to price
forecast and use MATLAB for training the ANN. The proposes
ANN model is a three-layered feedforward neural network
which has the feature of memory and learning is constructed,
trained by the Levenberg-Marquardt algorithm its proposed
for forecasting the next 168 hour (one week) electricity prices.
The hidden layer has 5, 10, 15, and 20 neurals for finding best
suitable accuracy and output layer has one unit.
The same experience was repeated about testing and finding
the best number of neurals in hidden layer. Table II shows the
result of forecasting 168 h (one week), presented the values of
RMSE, MAE, and MAPE from price forecasting performance
of the proposed hybrid (ARIMA-ANN) model in each of the
seasons.
From Table II, the performance of 5, 10, 15, and 20 neurals
in hidden layer are similar in values, in which 15 neurals in the
hidden layer have best suitable accuracy. Therefore, numerical
results for actual and forecasting prices in each of the seasons
with hybrid (ARIMA-ANN) proposed approach are shown in
Figs. 3–6 for the summer, fall, winter, and spring, respectively.
AREEKUL et al.: HYBRID ARIMA AND NEURAL NETWORK MODEL
Fig. 3. Forecasting results of January 22–28, 2006.
Fig. 4. Forecasting results of May 21–27, 2006.
Fig. 5. Forecasting results of August 20–26, 2006.
Fig. 6. Forecasting results of October 22–28, 2006.
Table III summarizes the numerical results, where the com-
parison of forecasting performance of the proposed ARIMA,
ANN, and hybrid models is presented.
From Table III, it can be observed that the hourly MAPE,
MAE, and RMSE of both neural network and hybrid models
529
TABLE III
COMPARISON OF FORECASTING RESULTS WITH
ARIMA, ANN, AND HYBRID MODELS
are better in accuracy than ARIMA model. The MAPE values
obtained in every season for ARIMA model range from 10.46%
to 16.06%, ANN model range from 10.03% to 15.62%, and hy-
brid model range from 9.98% to 15.57%. The results of the hy-
brid (ARIMA-ANN) model show that by combining two models
together, the overall forecasting errors can be significantly re-
duced, which in terms of MAPE, the percentage improvements
of the hybrid model over the ARIMA and ANN, are 4.5% and
0.5%, respectively. Therefore, the hybrid model gives better pre-
dictions than either ARIMA or ANN forecasts; its overall fore-
casting capability is improved.
V. CONCLUSION
This paper presented an approach for short-term price fore-
cast problem based on hybrid correction method, which is a
combination of ARIMA and ANN. The choice of forecasting
model becomes the important influence factors on how to im-
prove price forecasting accuracy. The linear ARIMA model and
the nonlinear ANN model are used jointly, aiming to capture dif-
ferent forms of relationship in the time-series data. Hence, we
propose the price correction method for the generation of new
method in which the new price data that are suitable to train
the neural network are generated by correcting the historical
days data with the help of price correction rates. For learning
the neural network, back propagation algorithm along with on-
line learning are adopted that consist of correction and forecast
errors. To verify the predictive ability of the proposed method,
we performed simulations for three different cases as price fore-
casting by ARIMA, ANN, and hybrid model approach. The test
results showed that the proposed forecasting method could pro-
vide a considerable improvement of the price forecasting accu-
racy, especially the hybrid model gives better predictions than
either ARIMA or ANN forecasts, and its overall forecasting ca-
pability is improved.
REFERENCES
[1] A. D. Papalexopoulos and T. C. Hesterberg, “A regression-based ap-
proach to short-term load forecasting,” IEEE Trans. Power Syst., vol.
5, no. 4, pp. 1535–1550, Nov. 1990.
530
[2] H. S. Hippert, C. E. Pedreira, and R. C. Souza, “Neural networks for
short-term load forecasting: A review and evaluation,” IEEE Trans.
Power Syst., vol. 16, no. 1, pp. 44–55, Feb. 2001.
[3] S. Rahman and O. Hazim, “A generalized knowledge-based short term
load-forecasting technique,” IEEE Trans. Power Syst., vol. 8, no. 2, pp.
508–511, May 1993.
[4] S. J. Huang and K. R. Shih, “Short-term load forecasting via ARMA
model identification including non-Gaussian process considerations,”
IEEE Trans. Power Syst., vol. 18, no. 2, pp. 673–679, May 2003.
[5] H. Wu and C. Lu, “A data mining approach for spatial modeling in
small area load forecast,” IEEE Trans. Power Syst., vol. 17, no. 2, pp.
516–521, May 2003.
[6] T. Senjyu, H. Sakihara, Y. Tamaki, and K. Uezato, “Next day peak
load forecasting using neural network with adaptive learning algo-
rithm based on similarity,” Elect. Mach. Power Syst., vol. 28, no. 7,
pp. 613–624, 2000.
[7] D. Srinivasan, S. S. Tan, C. S. Cheng, and E. K. Chan, “Parallel neural
network fuzzy expert system strategy for short-term load forecasting:
System implementation and performance evaluation,” IEEE Trans.
Power Syst., vol. 14, no. 3, pp. 1100–1105, Aug. 1999.
[8] G. E. P. Box and G. Jenkins, Time Series Analysis, Forecasting and
Control. San Francisco, CA: Holden-Day, 1970.
[9] E. D. McKenzie, “General exponential smoothing and the equivalent
ARMA process,” J. Forecasting, vol. 3, pp. 333–344, 1984.
[10] G. Zhang, E. B. Patuwo, and M. Y. Hu, “Forecasting with artificial
neural networks: The state of the art,” J. Forecasting, vol. 14, pp. 35–62,
1998.
[11] S. Makridakis, A. Anderson, R. Carbone, R. Fildes, M. Hibdon, R.
Lewandowski, J. Newton, E. Parzen, and R. Winkler, “The accuracy
of extrapolation (time series) methods: Results of a forecasting com-
petition,” J. Forecasting, vol. 1, pp. 111–153, 1982.
[12] R. Winkler, “Combining forecasts: A philosophical basis and some cur-
rent issues,” J. Forecasting, vol. 5, pp. 605–609, 1989.
[13] N. Kohzadi, M. S. Boyd, B. Kermanshahi, and I. Kaastra, “A compar-
ison of artificial neural network and time series models for forecasting
commodity prices,” Neurocomputing, vol. 10, pp. 169–181, 1996.
[14] G. P. Zhang, “Time series forecasting using a hybrid ARIMA and
neural network model,” Neurocomputing, vol. 50, pp. 159–175, 2003.
[15] C. W. J. Granger and A. P. Anderson, An introduction to bilinear time
series models. Göttingen, Germany: Vandenhoeck and Ruprecht,
1978.
[16] H. Tong, Threshold Models in Non-Linear Time Series Analysis. New
York: Springer-Verlag, 1983.
[17] R. F. Engle, “Autoregressive conditional heteroscedasticity with esti-
mates of the variance of U.K.,” Econometrica, vol. 50, pp. 987–1008,
1982.
[18] J. G. De Gooijer and K. Kumar, “Some recent developments in non-
linear time series modelling,” J. Forecasting, vol. 8, pp. 135–156, 1992.
[19] G. U. Yule, “Why do we sometimes get nonsense-correlations between
time series? A study in sampling and the nature of time series,” J. R.
Statist. Soc., vol. 89, pp. 1–64, 1926.
[20] H. Wold, A Study in the Analysis of Stationary Time Series. Stock-
holm, Sweden: Almgrist & Wiksell, 1938.
[21] K. Hornik, M. Stinchicombe, and H. White, “Using multi-layer feed-
forward networks for universal approximation,” Neural Netw., vol. 3,
pp. 551–560, 1990.
[22] M. S. Hung and J. W. Denton, “Training neural networks with the
GRG2 nonlinear optimizer,” Eur. J. Oper. Res., vol. 69, pp. 83–91,
1993.
[23] V. Subramanian and M. S. Hung, “A GRG2-based system for training
neural networks: Design and computational experience,” ORSA J.
Comput., vol. 5, pp. 386–394, 1993.
[24] J. W. Denton, “How good are neural networks for causal forecasting
?,” J. Forecasting, vol. 14, pp. 17–20, 1995.
[25] I. S. Markham and T. R. Rakes, “The effect of sample size and vari-
ability of data on the comparative performance of artificial neural net-
works and regression,” Comput. Oper. Res., vol. 25, pp. 251–263, 1998.
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 25, NO. 1, FEBRUARY 2010
[26] Australian National Electricity Market. [Online]. Available:
http://www.nemmco.com.au.
Phatchakorn Areekul (S’08) was born in Trang,
Thailand, on January 15, 1973. He received the B.S.
degree in electrical engineering from Rajamangala
University of Technology Thanyaburi, Pathumthani,
Thailand, in 1996 and the M.S. degree in electrical
engineering from Prince of Songkla, Songkla, Thai-
land, in 2003. He is currently pursuing the Ph.D.
degree in electrical and electronic engineering from
the University of the Ryukyus, Okinawa, Japan.
His research interests include electric power
system economics, power quality, security, and
application of intelligent systems.
Mr. Phatchakorn is a student member of the Institute of Electrical Engineers
of Japan and the Institute of Electronic, Information and Communication
Engineers.
Tomonobu Senjyu (M’02–SM’06) received the B.S.
and M.S. degrees in electrical engineering from the
University of the Ryukyus, Okinawa, Japan, in 1986
and 1988, respectively, and the Ph.D. degree in elec-
trical engineering from Nagoya University, Nagoya,
Japan, in 1994.
In 1988, he joined the Department of Electrical and
Electronics Engineering, University of the Ryukyus,
where he is currently a Professor. His research inter-
ests include stability of ac machines, advanced con-
trol of electrical machines, power electronics, appli-
cation of artificial intelligence techniques in power system, renewable energy,
and power economics.
Prof. Senjyu is a member of the Institute of Electrical Engineers of Japan.
Hirofumi Toyama was born in Okinawa, Japan, in
1984. He received the B.S. and M.S. degrees in elec-
trical and electronics engineering from the University
of the Ryukyus, Okinawa, Japan, in 2007 and 2009,
respectively.
He is currently with the University of the Ryukyus.
His research interests include electricity price fore-
casting, neural network, and unit commitment
problem.
Mr. Toyama is a student member of the Institute of
Electrical Engineers of Japan.
Atsushi Yona (M’08) was born in Okinawa, Japan, in
1982. He received the B.S. and M.S. degrees in elec-
trical and electronics engineering from the University
of the Ryukyus, Okinawa, Japan, in 2006 and 2008,
respectively.
He is currently an Assistant Professor with the Uni-
versity of the Ryukyus. His research interests include
wind generator, photovoltaic, and neural network.
Mr. Yona is a member of the Institute of Electrical
Engineers of Japan and the Institute of Electronic, In-
formation and Communication Engineers.