An Integrated Machine Learning Model for

Day-Ahead Electricity Price Forecasting

Shu Fan, James R. Liao, Kazuhiro Kaneko, Member, IEEE, and Luonan Chen, Senior Member, IEEE


AbstractThis paper proposes a novel model for short-term
electricity price forecasting based on an integration of two
machine learning technologies: Bayesian Clustering by Dynamics
(BCD) and Support Vector Machine (SVM). The proposed
forecasting system adopts an integrated architecture. Firstly, a
BCD classifier is applied to cluster the input data set into several
subsets in an unsupervised manner. Then, groups of 24 SVMs for
the next days electricity price profile are used to fit the training
data of each subset in a supervised way. To demonstrate the
effectiveness, the proposed model has been trained and tested on
the data of the historical energy prices from the New England
electricity market.
Index TermsElectricity Price, Non-stationarity, Machine
Learning.

I. INTRODUCTION

lectricity price forecasting has become increasingly

important with the evolvement of the electric power
industry into an era of market economy due to deregulation and
free competition [1]. Generators and loads rely on price
forecasting information to develop their corresponding bidding
strategies. If a generator has an accurate forecasting of the
prices, it can develop a bidding strategy to maximize its profit.
On the other hand, a load can make a plan to minimize its own
electricity cost if an accurate price forecasting is available.
This paper focuses on forecasting the day-ahead hourly
prices associated with a pool. Several approaches have been
proposed in the recent years for such as problem, such as linear
transfer function [2], and Autoregressive Integrate Moving
Average (ARIMA) models [3]; Artificial Neural Networks
(ANNs) [4]-[8]; wavelet transform models [9]; machine
learning techniques, [10]-[12] and so on [1], [13].
As indicated by many researchers, one key characteristic of
electricity prices is their volatility and non-stationarity [1], [8],
[11], [12]. The dynamics and properties of price series that
frequently switch between different segments or regions are
mainly due to discrete changes in competitors strategies and
multiple seasonality that generally give rise to piece-wise
stationary time series. However, almost all the methods of time

S.Fan, K.Kaneko and L.Chen are with Osaka Sangyo University, 3-1-1
Nakagaito, Daito, Osaka, 574-0013, Japan (e-mail: fanshu@hotmail.com;
k-kaneko@eic.osaka-sandai.ac.jp, chen@eic.osaka-sandai.ac.jp).
James R.Liao is with the Western Farmers Electric Cooperative, Anadarko,
OK, 73005 USA (e-mail: j_liao@wfec.com)

142440178X/06/$20.002006IEEE

series analysis, both linear and nonlinear, assume some

stationarity of the system under investigation. Therefore, when
modeling time series of price, a key challenge is on how to
handle the non-stationarity.
This paper proposes a novel and effective forecasting model
by applying machine learning techniques with the emphasis on
tackling the non-stationarity problem. Specifically, we adopt an
integrated architecture based on two machine learning
techniques: Bayesian Clustering by Dynamics (BCD) and
Support Vector Machine (SVM). The basic procedure of the
proposed model is stated as follows: firstly, a BCD classifier is
used to identify the switching or piece-wise stationary
dynamics for the input data and to partition the dataset into
several subsets, so that the dynamics in each subset are similar;
then groups of 24 SVMs are applied to respectively fit the
hourly electricity price profile data in each partitioned subset
by taking advantage of all past information and similar dynamic
properties (e.g. piece-wise stationarity). After being trained, the
forecasting model can predict the next-day electricity price
with an acceptable level of accuracy on the specific subset in a
first past the post voting manner among the BCD and SVM,
where the output of only one SVM model is used for the final
forecast.
This forecasting model based on the integrative machine
learning technique possesses the following advantages:
1) It is well suited for capturing the dynamics of electricity
price time series. The BCD classifier models the process
generating each price series as an autoregressive model of order
p, say AR(p), and then clusters those price time series with a
high posterior probability of being generated by the same
AR(p) model [14],[15]. BCD is based on unsupervised learning,
which has the ability to partition the space of input training data
set into many subsets without prior knowledge about the
classifying criteria. Compared to other clustering methods such
as hierarchical clustering or Self Organizing Maps (SOM),
BCD identifies the set of clusters with maximum posterior
probability without requiring any prior input about the number
of clusters, thereby avoiding the risk of overfitting.
2) It can fit the data well due to multiple local models. For
each cluster of time series, we adapt 24 SVRs for 24-hour
predictions of a day. SVM or SVR (Support Vector
Regression) is a new and powerful machine learning technique
for data classification and regression based on recent advances
in statistical learning theory [16]. Established on the unique
theory of the structure risk minimization principle to estimate a

1643

PSCE2006

several particular properties with price. The hourly price curve

is widely divergent and fluctuates with a high frequency, and
there are also a high percentage of abrupt changes or spikes in
the price curve (mainly in periods of high demand). The price
presents high volatility and non-constant mean. The abrupt
changes and volatility of price can be reflected as a switch in
the price series dynamics due to the discrete behaviors in
competitors strategies. In other words, there exist different
regimes in the price time series, which generally give rise to
piece wise stationary dynamics.
200

NEPOOL's Hub Price ($/MWh)

function by minimizing an upper bound of the generalization

error, SVMs are shown to be very resistant to the over-fitting
problem, eventually achieving a high generalization
performance in solving forecasting problems of various time
series [17],[18].
3) It is strongly robust. The proposed forecasting model uses
different input variables for BCD and SVR respectively. All the
input variables are decomposed into two groups: only price and
load data are used for the SVMs; for BCD, besides price and
load data, other factors, such as day type, weather factors and
fuel price can be also used to capture the dynamics. Compared
with a unity model dealing with all variables together, the
approach will alleviate the sensitivity of the model to these
variables and the user can intentionally split input variables
using his knowledge and experience.
In the experiment, we adopt the price data of the New
England Independent System Operator (ISO) to demonstrate
the effectiveness of the learning and forecasting for the
proposed method. To compare the forecasting accuracy, we
also examine the model only with SVMs.

150

100

50

II. TASK DESCRIPTION AND PRICE DATA ANALYSIS

This paper uses the hourly electrical hub price series of New
England Power Pool (NEPOOL) as a test example of our
method [19]. In order to develop an appropriate model, we give
a brief description of test system and examine the main
characteristics of the hourly price series in this section.
The New England electricity market covers six-state region,
including Connecticut, Maine, Massachusetts, New
Hampshire, Rhode Island and Vermont. The competitive
electrical wholesale markets were implemented on December
1, 1998 [20]. All of New England has been treated as a single
power market, with price set by an hourly bid process
administered by the ISO New England. On March 1, 2003, the
New England power market successfully implemented the
Standard Market Design (SMD), a major redesign of the
wholesale electricity market. The day-ahead market consists of
24 hourly auctions that take place simultaneously one day in
advance. Hourly bids, expressed in $/MWh, are submitted on a
day-ahead basis for the next 24 hours by the generators and
loads. ISO New England then optimally schedules the
generating units that will run on the following day by
minimizing the total costs of the energy market. The market is
settled after the fact on an hourly basis. All transactions are
priced at the (ex-post) energy clearing price (ECP), which is
determined at the intersection of aggregate supply and demand
curves.
Fig.1 illustrates the latest electricity prices for the hub of
NEPOOL from March 1, 2003 to July 31, 2005. According to
Fig.1, it is clear that the price dynamics have multiple seasonal
patterns, corresponding to a daily and weekly periodicity
respectively, and are also influenced by calendar effect, i.e.
weekend and holidays. These properties are just similar to those
of load. However, in contrast to the load time series, there are

5000

10000

15000

20000

Hour

Fig.1. Hourly electricity price in NEPOOL from March 1, 2003 to July 31, 2005

III. METHOD AND THE LEARNING ALGORITHM

A. The Architecture of the Forecasting System
In this paper, a time series based nonlinear discrete-time
dynamical model, is represented by (1) for the price
forecasting,
y (t  1) f ( y (t ),..., y (t  m  1); D)
(1)
where y(t) is a vector representing the daily electricity price
profile at time t, and m is the order of the dynamical system,
which is a predetermined constant. D is a vector representing
the control parameters of the dynamical system, such as
weather conditions, day types, power plant mix, and fuel prices.
Then, the task in price forecasting is to extrapolate past price
behavior while taking into account the effect of other
influencing factors.
An integrated machine learning model is proposed to
reconstruct the dynamics of electric price using the time series
of its observables. The forecasting system is shown in Fig. 2.
This model is based on an integrated architecture. Firstly, a
BCD classifier is applied to cluster the input training data set
into several subsets with similar dynamical properties in an
unsupervised manner. Then, several groups of 24 SVMs (or
exactly SVRs) are used to fit the training data in each subset in
a supervised way. In other words, depending on their
stationarity as well as other dynamic features, this method
breaks the time series into different segments or subsets where

1644

data in the same segment can be modeled by the same SVR due
to similar property. Notice that there are 24 SVRs for each
subset to train and predict 24-hour prices of next day
respectively.
Prior
probability

L, W, F

Subset 1
P1
L1

SVR1
SVR2
SVR24

Subset 2

BCD classifier

For day type classification

P0

P1
L1

.
.
..

SVR1
SVR2
SVR24

.
.
Subset n

P2
L2

.
.
..

Day-ahead price profile P(d+1)

.
.
.
.

Table II. List of input data of the SVM network for regular days
Input
Variable name
Lagged value (hours)
24,25,26,48,72,96,120,
1-9
Hourly price (L1)
144,168
-1,0,1,24,48,72,96,120,
10-19
Hourly load (T1)
144,168
The hour of load predication is assumed at 0, the lag 0 represents the target
instant, and the 24 lagged hours means the values that were measured 24 hours
earlier than the hour of predication.

SVR1
SVR2
SVR24

Fig .2. Integrated machine learning model for the electricity price forecasting

In Fig. 2, the input variables are different for the BCD

compared to the SVM networks. As stated above, many
exogenous variables besides the price itself should be taken
into account in the price forecasting. On the other hand, the
electricity price mainly depends on the evolution of balance
between the electricity demand and the available supply.
Therefore, in this paper, all the input variables are decomposed
into two groups: a group with only price and load series for the
SVMs; a group with price, load series and further other factors
(to capture the dynamics) for BCD. This arrangement alleviates
the sensitivity of the model to these variables without losing
useful information at the same time. And the user can
intentionally choose appropriate input variables for BCD
classifier using his knowledge and experience. In addition, the
prior probability provides an interface for the accommodation
of new variables into the forecasting model.
The input variables for BCD and SVR are shown in Tables I
and II.

Input

Table I. List of input data of the BCD classifier

Variable
Detail description

1-10

Price series P0

Average daily price series of previous ten days

11-13

Variables to
determine prior
probability

L: Maximal daily load

W: Maximal daily temperature
F: Daily fuel price

As indicated in Table I, the input data of BCD consist the

following elements: average daily price series of previous ten
days P0, maximal daily demand L and temperature W, and daily
fuel price F. Specifically, the price time series P0 are used as the
primary input for BCD to cluster the day by dynamics, the other
factors, including load, weather conditions and fuel prices, can
be used to determine the prior probability for the BCD. Other
factors can also be included, and the selection of input data
mainly depends on the system to be studied.
For the SVM, in addition to the forecasted and actual load
values, the input variables are the hourly price values of the last
day available and the similar hours in the previous days or
weeks. The input data of the SVM network is shown in Table
II.

The inputs of training data consist of past hourly electricity

price and the forecasted demand. In order to capture the time
series style in price, we include the electricity price of the
previous seven days around the predicting hour. Because our
experiment shows that using load of the forecasting day only
does not produce satisfactory forecasting results, load variables
are added for every time point at which price was included.
B. Learning Algorithm: The BCD Classifier
Dynamics of electricity prices are non-stationary or
approximately piece-wise stationary during a short period of
time due to the switching nature related to multiple seasonality
and discrete changes in participants strategies. In this paper,
the BCD classifier is used as a gating network to identify the
switching or piece-wise stationary dynamics for the input
training data set firstly.
The clustering method implemented in BCD is based on a
novel concept of similarity for time series: two time series are
similar when they are generated by the same stochastic process.
Therefore, the components of BCD are a model describing the
dynamics of time series, a metric to decide when two time
series are generated by the same stochastic process, and a
search procedure to efficiently explore the space of possible
clustering models.
BCD models time series by autoregressive equations [15].
Let Sj = {xj1, . . . , xjt, . . . , xjn} denote a stationary time series of
continuous values. The series follows an autoregressive model
of order p, say AR(p), if the value of the series at time t > p is a
linear function of the values observed in the previous p steps.
We can describe the model in a matrix form as
(2)
x j X jE j  H j

1645

where xj is the vector (xj(p+1), . . . , xjn)T , Xj is the (np)(p+1)

regression matrix whose tth row is (1, xj(t1), . . . , xj(tp)) for t >p,
j is the vector of autoregressive coefficients and j is the vector
of uncorrelated errors that are assumed normally distributed
with expected value E(jt) = 0 and variance V(jt) = j2, for any
time point t. Given the data, the model parameters can be
estimated using standard Bayesian procedures, and details are
described in [15].
To select the set of clusters, BCD uses a novel model-based
Bayesian clustering procedure. A set of clusters C1, . , Ck., . ,Cc,
each consisting of mk time series, is represented as a model MC.
The time series assigned to each cluster are treated as
independent realizations of the dynamic process represented by
the cluster, which is described by an autoregressive equation.
Each cluster Ck can be jointly modeled as
xk X k E k  H k
where the vector xk and the matrix Xk are defined by stacking
the mk vectors xkj and regression matrices Xkj, one for each time
series, as follows
x k1
X k1

X k 

xk 

xkmk
X kmk
Given a set of possible clustering models, the task is to rank
them according to their posterior probability. The posterior
probability of the model MC is computed by Bayes Theorem as
(3)
P ( M C | x) v P ( M C ) f ( x | M C )
where P(MC) is the prior probability of MC and f (x| MC) is the
marginal likelihood. Assuming independent uniform prior
distributions on the model parameters and a symmetric
Dirichlet distribution on the cluster probability pk, the marginal
likelihood of each cluster model MC can be easily computed in a
closed form by solving the integral:
f (x | M C )

f (x | T

C)

f (T C )dT C

(4)

where C is the vector of parameters that describe the likelihood

function, conditional on a clustering model MC, and f(C) is the
prior density. In this way, each clustering model has an explicit
probabilistic score and the model with maximum score can be
found. In particular, f (x| MC) can be computed as

f (x | M c )
c

k 1

* (1)
* (1  m )

As the number of clusters or subsets grows exponentially

with the number of time series, BCD uses an agglomerative
search strategy, which iteratively merges time series into
clusters. The procedure starts by assuming that each of the m
electricity price time series is generated by a different process.
Thus, the initial model Mm consists of m clusters, one for each
time series, with score f(x|Mm). The next step is the computation
of the marginal likelihood of the m(m1) models in which two
of the m profiles are merged into one cluster. The model Mm1
with maximal marginal likelihood is chosen and the merging is
rejected if f ( x | M m ) t f ( x | M m1 ) and the procedure stops. If
f ( x | M m )  f ( x | M m1 ) , the merging is accepted and a cluster
Ck merging the two time series is created. In such a way, the
procedure is repeated on the new set of m1 time series that
consist of the remaining m2 time series and the cluster profile.
Although the agglomerative strategy makes the search
process feasible, the computational effort can be extremely
demanding when the number of time series is large. To further
reduce this effort, we use a heuristic strategy based on a
measure of similarity between time series. The intuition behind
this strategy is that the merging of two similar time series has
better chances of increasing the marginal likelihood. The
heuristic search starts by computing the m(m1) pair-wise
similarity measures of the time series and selects the model
Mm1 in which the two closest time series are merged into one
cluster. If f ( x | M m )  f ( x | M m1 ) , the two time series are
merged into a single cluster, a profile of this cluster is computed
by averaging the two observed time series, and the procedure is
repeated on the new set of m1 time series. If this merging is
rejected, the procedure is repeated on the two time series with
the second highest similarity until an acceptable merging is
found. If no acceptable merging is found, the procedure stops.
Note that the clustering procedure is actually performed on the
posterior probability of the model and the similarity measure is
only used to increase the speed of the search process and to
limit the risk of falling into local maxima.
Similarity measures of two time series implemented in BCD
include Euclidean distance, correlation and KullbackLeiber
distance. In the numerical experiments, we have tried different
distances and finally adopted the Euclidean distance of two
time series Sj = {xi1, . . . , xin} and Sj = {xj1, . . . , xjn}, computed
as
n

RSS k ( q  n k ) / 2 n k  q
*(
)
)
* (mk / m  mk )
2
2
( q  nk ) / 2
* (mk / m )
( 2S )
det( X kT X k ) (1 / 2 )

D( S i , S j )

(x

it

 x jt ) 2

(6)

t 1

(5)
where nk is the dimension of the vector xk, and
RSS k xkT ( I n  X k ( X kT X k ) 1 X kT ) xk is the residual sum of
squares in cluster Ck. When all clustering models are a priori
equally likely, the posterior probability P(MC| x) is proportional
to the marginal likelihood f(x|MC), which becomes our
probabilistic scoring metric.

C. Learning Algorithm: The SVM Network

In this stage, several groups of SVRs are used for price time
series learning and prediction in each subset, and each SVR
corresponds to one time instant. Suppose that we are given
training data (x1,y1),( xi,yi),...( xn,yn), where xi are input
patterns and yi are the associated output value of xi, the support
vector regression solves an optimization problem

1646

min
Z [[
,b, ,

1 T
Z Z  C ([ i  [ i* )
2
i 1

Subject to yi  (Z T I ( xi )  b) d H  [ i* ,

(7)

(Z I ( xi )  b)  yi d H  [ i ,

[ i , [ i* t 0, i=1,,n
where xi is mapped to a higher dimensional space by the
function , and i* is slack variables of the upper training error
(i is the lower) subject to the -insensitive tube
(Z T I ( xi )  b)  yi d H . The constant C>0 determines the trade
off between the flatness and losses. The parameters which
control regression quality are the cost of error C, the width of
the tube , and the mapping function .
The constraints of (7) imply that we put most data xi in the
tube . If xi is not in the tube, there is an error i or i* which we
tend to minimize in the objective function. SVR avoids
under-fitting and over-fitting of the training data by minimizing
n

the training error C

([

 [ i* ) as well as the regularization

i 1

term Z Z / 2 . For traditional least-square regression, is

always zero and data are not mapped into higher dimensional
spaces. Hence, SVR is a more general and flexible treatment on
regression problems.
Since might map xi to a high or infinite dimensional space,
instead of solving for (7) in a high dimension, we deal with
its dual problem, which can be derived using the Lagrange
theory.

properties (e.g. piece-wise stationarity). The next-day

electricity price forecasting is conducted by the trained network
with an acceptable level of accuracy in a voting manner in the
BCD and SVRs. For numerical experiments in this paper, we
use the software LIBSVM [21], which is a library for support
vector machines with an efficient implementation of solving
(8).

IV. NUMERICAL EXPERIMENTS

A. Data Collection and Preprocess
The hourly electrical prices and load data in New England
have been considered for the study [19]. The SMD market
period of 3/2003-present is studied in the experiments. We use
the hub price of New England market as forecasting object.
Two months have been selected to forecast and validate the
performance of the proposed model. The first month
corresponds to May 2005, and the latter month corresponds to
August 2005. The hourly data used to forecast the three months
are from March 1, 2003 to April 30, 2005; and March 1, 2003 to
July 31, 2005.
The test sets are completely separate from the training sets
and are not used during the learning procedure. Depending on
the data variability for the different time periods, a larger
forecasting lead-time does not necessarily imply a larger
forecasting error.

B. Learning Procedure and Implementation

The learning procedure of the proposed architecture is
n
n
n
1
T
outlined
as follows.

(D i  D i *) Q(D j  D j *)  H
(D i  D i *) 
(D i  D i *)
2
1)
Classify
the entire training data set into two groups: a
Di ,Di *
i 1
i 1
i, j 1
training
set
used
for updating the network parameters and a
n
verification
set
for
testing the performance.
Subject to,
(8)
(D i  D i *) 0
2)
Determine
prior
probability of the BCD classifier
i 1
according
to
the
control
variables;
0 d D i , D i * d C , i=1,,n
3) Classify the day type using BCD according to the
where Qij I ( xi )T I ( x j ) , i and i* are the Lagrange dynamics of price time series;
4) In each subset of the input space, train the SVM to fit the
multipliers. However, this inner product may be expensive to
compute because I (x) has too many elements. Hence, we data subset according to (8);
5) Forecast the price profile using the data set for verification
apply kernel trick to do the mapping implicitly. That is, to
and
calculate the MAPE;
employ some special forms, inner products in a higher space
6)
Tune the parameters of the model and repeat step 1) to
yet can be calculated in the original space. Typical examples for
step
5)
until satisfactory results are obtained;
the
kernel
functions
are
polynomial
kernel
7)
Select
the network parameters at the minimum of the
I ( xi )T I ( x j ) (Jx1T x2  c0 ) d
and
RBF
MAPE as the final ones.
2
After the training procedure is finished, the following test
kernel I ( xi )T I ( x j ) e J ( x1  x 2 ) . Here, , c0, and d are kernel
process is applied to verify the proposed model.
parameters. They are inner products in a very high dimensional
1) Identify the type of the test day according to the
space (or infinite dimensional space) but can be computed information of the test day and previous days, using the BCD
efficiently by the kernel trick even without knowing I (x ) .
classifier.
As each data subset classified from the BCD is considered to
2) Use the corresponding SVM network to output the
be approximately stationary, 24 SVRs are applied to forecasting price.
respectively fit the hourly electricity price profile data by taking
advantage of all past information and similar dynamic

max

1647

C. Numerical Results
The criteria to compare the performance are the mean
absolute error (MAE) and mean absolute percentage error
(MAPE) in this paper, which indicate the accuracy of recall.
Numerical results with the proposed method are presented.
To demonstrate the effectiveness of the integrated structure, a
SVM network without the BCD classifier is also adopted for
forecasting. The MAEs and MAPEs for the two test months are
shown in Tables III.

REFERENCES

Table III
Forecasting results of May and August 2005
Single SVM
Integrated architecture
MAE
MAPE
MAE
MAPE
May 2005
3.78
6.83
2.80
5.35
August 2005
8.04
8.11
6.58
6.82
Average
5.91
7.47
4.69
6.08

[1]
[2]

[3]

Clearly the integrated machine learning model outperforms

single SVM in all the situations. The variety of error indices for
the 24-hour is smooth, and indicate stability of the proposed
model. Especially, the average MAPE for the two months in the
new SMD market is 6.08%, which is considered high in terms
of accuracy.
Fig. 3 shows the forecasting and the actual price for ten days
in August. It can be seen that the proposed model well predicts
the trend of the price curve generally. Since the SMD market
has been implemented for two and a half years, only limited
data have been available for learning so far. When more
training samples become available and are included in training
in the future, it is possible to get more accurate forecasting.
150

Actual
Forecast

125

Then, in each subset, different SVMs have been used to fit the
input data belonging to different market state in a supervised
manner. The proposed method has been applied to the
prediction of New England spot market price and demonstrates
a high degree of effectiveness and efficiency in learning and
prediction compared to other methods. As shown in the
numerical experiments, the proposed method predicts the
prices with a high degree of accuracy.

[4]

[5]

[6]

[7]

[8]

[9]

[10]

Price ($/MWh)

[11]
100

[12]
[13]

75

[14]
50

[15]
25

350

400

450

500

550

Hour

[16]
[17]

Fig. 3 Forecasting results of August 14-24, 2005

V. CONCLUSION

[18]

In this paper, an integrated machine learning forecasting

system based on BCD and SVR has been developed to predict
electricity price series. The BCD classifier in the proposed
method can identify the state of electricity price and cluster the
input data set into several subsets in an unsupervised manner.

[19]
[20]

[21]

1648

D. W. Bunn, Modelling Prices in Competitive Electricity Markets.

England: John Willy & Sons, 2004.
F. J. Nogales, J. Contreras, A. J. Conejo, and R. Espnola, Forecasting
next-day electricity prices by time series models, IEEE Trans. Power
Syst., vol. 17, pp. 342348, May 2002.
J. Contreras, R. Espinola, F. J. Nogales and A. J. Conejo, ARIMA
models to predict next-day electricity prices, IEEE Trans. Power Syst.,
Vol. 18, pp. 10141020, Aug. 2003.
C. P. Rodriguez, G. J. Anders, Energy price forecasting in the Ontario
competitive power system market, IEEE Trans. Power Syst., Vol. 19, pp.
366374, Feb. 2004.
B. R. Szkuta, L. A. Sanabria, and T. S. Dillon, Electricity price
short-term forecasting using artificial neural networks, IEEE Trans.
Power Syst., vol. 14, pp. 851857, Aug. 1999.
J. J. Guo and P. B. Luh, Selecting input factors for clusters of gaussian
radial basis function network to improve market clearing price
prediction, IEEE Trans. Power Syst., vol. 18, pp. 665672, May 2003.
L. Zhang, P. B. Luh, Neural Network-Based Market Clearing Price
Prediction and Confidence Interval Estimation With an Improved
Extended Kalman Filter Method, IEEE Trans. Power Syst., vol. 20, pp.
59-66, Feb. 2005.
J. Guo, P. B. Luh, Improving market clearing price prediction by using a
committee machine of neural networks, IEEE Trans. Power Syst.,Vol.
19, pp. 18671876, Nov. 2004.
A. J. Conejo, M. A. Plazas, R. Espnola, and A. B. Molina, Day-Ahead
Electricity Price Forecasting Using the Wavelet Transform and ARIMA
Models, IEEE Trans. Power Syst., vol. 20, pp. 1035-1042, May 2005.
D. Zhou; F. Gao; X. Guan, Application of accurate online support vector
regression in energy price forecast, in Fifth World Congress on
Intelligent Control and Automation, June 2004, pp. 1838 - 1842.
A. M. Gonzalez, A. M. S. Roque and J. Garcia-Gonzalez, Modeling and
forecasting electricity prices with input/output hidden Markov models,
IEEE Trans. Power Syst., Vol. 20, pp. 13 24, Feb. 2005.
S. Fan, C. Mao, L. Chen, Next-Day Electricity Price Forecasting Using a
Hybrid Network, IEEE Trans. Power Syst., 2005. In press.
R. C. Garcia, J. Contreras, M. Akkeren, and J. C. Garcia. A GARCH
Forecasting Model to Predict Day-Ahead Electricity Prices, IEEE Trans.
Power Syst., vol. 20, pp. 867-874, May 2005.
M. Ramoni, P. Sebastiani, P. Cohen, Bayesian Clustering by Dynamics,
Mach. Learn., vol. 47, pp. 91-121, 2002.
P. Sebastiani, M. Ramoni, Clustering continuous time series, Proc
Eighteenth Intl Conf on Machine Learning (ICML-2001) pp 497-504.
Morgan Kaufmann, San Francisco, CA, 2001.
C. Cortes and V. Vapnik, Support-vector network, Mach. Learn., vol.
20, pp. 273-297, 1995.
K.-R. Mller, A. Smola, G. Rtsch, B. Schlkopf, J. Kohlmorgen, and V.
Vapnik, Predicting time series with support vector machines, in
Advances in Kernel MethodsSupport Vector Learning, B. Schlkopf, C.
J. C. Burges, and A. J. Smola, Eds. Cambridge, MA:MIT Press, 1999, pp.
243254.
B. J. Chen, M. W. Chang, C.-J. Lin, Load forecasting using support
vector machines: a study on EUNITE competition 2001, IEEE Trans.
Power Syst., vol. 19, pp. 1821-1830, Nov. 2004.
ISO, Overview of ISO New England, March, 2004.
C.-C. Chang and C.-J. Lin. (2001) LIBSVM: A Library for Support
Vector
Machines.
[Online].
Available:
http://www.csie.ntu.edu.tw/~cjlin/libsvm.
wwwISO-NE.com [Online].

Shu Fan received his B.S. M.S. and Ph.D. degrees in Department of Electrical
Engineering, from Huazhong University of Science and Technology (HUST),
in 1995, 2000 and 2004 respectively. Presently he is carrying out postdoctoral
research in Osaka Sangyo University. His research interests are power system
control, high-power power electronics and hybrid intelligent forecasting
system.
James R. Liao (M89) received the M.S. degree from the University of
Missouri, Rolla, in 1980, and the Ph.D. degree from the University of
Oklahoma, Norman, in 1992, both in electrical engineering. Since 1980, he has
been with the Western Farmers Electric Cooperative, Anadarko, OK. He was a
Transmission/Generation Systems Analyst from 1980 to 1985 and an EMS
System Software Engineer from 1985 to 1999. Since 1999, he has been
Principal Operations Engineer. Dr. Liao is a Registered Professional Engineer
in the State of Oklahoma.
Luonan Chen (M94, SM98) received the B.E. degree from HUST, Wuhan,
China in 1984, and the M.E. and Ph.D. degrees from Tohoku University,
Sendai, Japan, in 1988 and 1991, respectively. Since 1997, he has been a faculty
of Osaka Sangyo University, Osaka, Japan, where he is currently a Professor
with the department of Electrical Engineering and Electronics. His fields of
interest are nonlinear dynamics and optimization in power systems.

1649