Applied Energy 87 (2010) 3226–3234

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

An enhanced radial basis function network for short-term electricity

price forecasting
Whei-Min Lin a, Hong-Jey Gow a, Ming-Tang Tsai b,*
a
Department of Electrical Engineering, National Sun Yat-Sen University, Kaohsiung 804, Taiwan, ROC
b
Department of Electrical Engineering, Cheng-Shiu University, Kaohsiung 833, Taiwan, ROC

a r t i c l e i n f o a b s t r a c t

Article history:
Received 18 January 2010
Received in revised form 24 March 2010
Accepted 4 April 2010
Available online 26 May 2010
Keywords:
Orthogonal Experimental Design (OED)
Locational Marginal Price (LMP)
Radial Basis Function Network
Electricity price forecasting
Stochastic Gradient Approach (SGA)
Factor analysis
This paper proposed a price forecasting system for electric market participants to reduce the risk of price
volatility. Combining the Radial Basis Function Network (RBFN) and Orthogonal Experimental Design
(OED), an Enhanced Radial Basis Function Network (ERBFN) has been proposed for the solving process.
The Locational Marginal Price (LMP), system load, transmission flow and temperature of the PJM system
were collected and the data clusters were embedded in the Excel Database according to the year, season,
workday and weekend. With the OED applied to learning rates in the ERBFN, the forecasting error can be
reduced during the training process to improve both accuracy and reliability. This would mean that even
the ‘‘spikes” could be tracked closely. The Back-propagation Neural Network (BPN), Probability Neural
Network (PNN), other algorithms, and the proposed ERBFN were all developed and compared to check
the performance. Simulation results demonstrated the effectiveness of the proposed ERBFN to provide
quality information in a price volatile environment.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction
Deregulating the power market creates competition and a trad-
ing mechanism for market players. It moves the market from a
cost-based operation to a bid-based operation [1,2]. Electricity
has become a commodity and the price could become volatile in
an energy market where sudden ‘‘spikes” could appear. Forecasting
prices accurately is an important task for producers, consumers
and retailers. In accordance with a price forecast, a balancing and
settlement system can be established for participants to maximize
profit with lower risks [3]. Price forecasting also helps investors in
better planning the Grids.
In the US, the Pennsylvania–New Jersey–Maryland (PJM) power
market [4] is commonly recognized as one of the most successful
market models. The PJM that operates the competitive market is
coordinated by an Independent System Operator (ISO) to deter-
mine the Locational Marginal Price (LMP) according to the status
of the system nodes. The LMP at each node then reflects not only
the price of voluntary bids but also the overhead of delivering en-
ergy to locations. Generally, the LMP includes three components:
an energy cost component, a transmission congestion component
and the marginal loss component. The LMP in a pool is different
for different locations while the energy cost can be identical for
* Corresponding author. Tel.: +886 7 7315367; fax: +886 7 7337390.
E-mail addresses: wmlin@ee.nsysu.edu.tw (W.-M. Lin), ghjey19@yahoo.com.tw
(H.-J. Gow), tsaymt@post.csu.edu.tw (M.-T. Tsai).
all the nodes. A feasible and practical method for LMP forecasting
would provide better risk management for all market participants.
Many factors, including historical prices, system load, transmis-
sion flow and temperature, could impact the LMP. The loads and
prices in the wholesale market are mutually intertwined activities
[5]. Loads are heavily affected by weather parameters, so prices are
strongly volatile with the changing weather. Another factor is use
at various times of the day, week, month, season and year. Prices
could rise to a hundred times the normal value in reflecting this
volatility. In the PJM, the LMP is introduced at nodes. Congestion
occurs when a transmission flow exceeds its limits. Line flow infor-
mation becomes an important factor in price forecasting. It is com-
plicated to perform LMP forecasting, especially when trying to find
the best strategy in a world of uncertainties. Reported techniques
for forecasting day-ahead prices include time series models [6],
weighted nearest neighbors techniques [7], auto regressive inte-
grated moving average models (ARIMA) [8], Mixed ARIMA models
[9,10], k-factor GIGARCH process [11], and Markov models [12].
These approaches can be very accurate given sufficient information
and computation time; however, no approach has shown a satis-
factory performance in dealing with the spikes.
Recently, the Artificial Neural Network (ANN) has been applied
to forecast prices in various markets [13–19]. The ANN is a simple,
powerful and flexible tool for forecasting, providing better
solutions to model complex non-linear relationships than the tra-
ditional linear models. ANNs have weaknesses in the determina-
tion of network architecture and network parameters. Running in

0306-2619/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.apenergy.2010.04.006
 W.-M. Lin et al. / Applied Energy 87 (2010) 3226–3234 3227

a dynamic environment, especially for online applications, a tradi-

tional ANN can become a bottleneck in adaptive applications [20].
Among ANNs, the Radial Basis Function Network (RBFN) [21] can
function as a classifier and forecaster, and it has the advantages
of a simple construct. However, a RBFN probably causes over-
learning due to large adjustable parameters. In order to improve
the issue, this paper has used the Orthogonal Experimental Design
(OED) to enhance the traditional RBFN.
The OED is an effective tool for robust design and engineering
methodology in optimizing process conditions which are mini-
mally sensitive to various causes of variations. The characteristics Fig. 1. Illustration of the orthogonal array L4(23).
of an OED are: (1) results obtained through few experiments, (2)
good recurrence of results in the same experimental environment,
(3) simple construction of mathematical model with the applica- Table 1
tion of an Orthogonal Array and (4) simple analytical procedure. Tow factors x1, x2 experiments.

Combining the RBFN [21] and the OED [22] into an Enhanced RBFN Experiment no. Factors Function value yi
(ERBFN) has been proposed in this paper. The OED has been used to x1 x2
adjust the parameters in the RBFN training stage to improve the
1 2.4 60 5.4
forecasting ability, and a good performance with a close spike 2 2.4 80 5.2
tracking capability can be seen. This paper has developed day- 3 3.6 60 7.0
ahead forecasts for electricity price using an ERBFN based on a sim- 4 3.6 80 6.4
ilar day PJM market model. The Mean Absolute Percentage Error
(MAPE), Mean Absolute Error (MAE) and Root Mean Square Error
(RMSE) obtained from the forecasting results have demonstrated Table 2
that the ERBFN can efficiently forecast the price for any day of Tow factors A, B experiments.
the week.
Experiment no. Factors Function value yi
A B AB
2. Orthogonal Experimental Design
1 1 1 1 5.4
2 1 1 1 5.2
The Orthogonal experiment uses a small amount of experimen- 3 1 1 1 7.0
tal data to construct a mathematical model derived from the 4 1 1 1 6.4
orthogonal experiment to calculate experiment values. The OED
contains two main parts, Orthogonal Array and factor analysis. In
engineering experiments, the cause that influences the result is a x  ðxU 2þxL Þ
factor of the Orthogonal Array, while various states of the factor xN ¼ ðxU xL Þ
: ð2Þ
are called the ‘‘level number”. Factor analysis is carried out after 2

the Orthogonal Array set-up, and the ‘‘factor effects” of each factor For convenience of description, a transformation of variable is
can be obtained from the analytical results. The impact of each fac- made for x1 and x2 by:
tor on the experiment is deduced from those effects.
An orthogonal array with F factors and Q levels can be denoted x1  3:6þ2:4
2 x1  3:0
A¼ 3:62:4
¼ ð3Þ
as LB (QF), where L denotes a Latin square, and B is the chosen num- 2
0:6
ber of combinations of levels. The notion of using orthogonal arrays
has been associated with the Latin Square from the outset. We let x2  80þ60
2 x2  70
B¼ ¼ ð4Þ
LB (QF) = [ebf]BF, where the fth factor in the bth combination has le- 8060 5
2
vel value ebf and ebf 2 {1, 2, . . . , Q}. An example of an orthogonal ar-
ray can be seen as: A and B are used for Table 1, with another A  B added as shown in
Table 2, where 1 is called ‘‘Level 1”, and ‘‘+1” is Level 2. The re-
2 3
1 1 1 sponse value Rf,l is calculated from Eq. (5), and the effect of each fac-
6 7 tor Ef is calculated through Eq. (6). We have:
61 2 27
L4 ð23 Þ ¼ 6
6
7:
7 ð1Þ
42 1 25 Rf ;l ¼ Mean of yi for factor f at Level l ð5Þ
2 2 1 where f = A, B, and l ¼ 1; 2, and the effect:

Fig. 1 shows a three-factor solution space for the example. The Rf ;Level 2  Rf ;Level 1
Ef ¼ ð6Þ
four vertices of the cube cover presented in Eq. (1) are all facets of 2
the problem and are representative enough to calculate the opti- The related data are listed in Table 3.
mal solution. With the above arrangement, the relationship [22] between
Consider an experiment containing two factors, x1 and x2, function value yi and other factors A, B, A  B, Rf,l and Ef is can be
assuming each factor has two levels. The two levels of x1 are 2.4 described by:
and 3.6, the two levels of x2 are 60 and 80 and the experimental
yi ¼ Average þ EffectA  A þ EffectB  B þ EffectAB  A  B ð7Þ
output data y is obtained from the experiment. A sample of the
experimental data is shown in Table 1. A numerical example of Experiment 1 can be seen by:
Let Eq. (2) be defined as a transformation of the variable for any
y1 ¼ 5:4 ¼ 6 þ 0:7  ð1Þ  0:2  ð1Þ  0:1  ð1Þ  ð1Þ ð8Þ
factor with two levels, where x is the original variable, xU is the
high level and xL is low level. Then xN will be a new variable with Other experimental outputs can be obtained from Eq. (7) simi-
values +1 or 1, that is: larly. It shows that the function of an Orthogonal Array can use a
3228 W.-M. Lin et al. / Applied Energy 87 (2010) 3226–3234

Table 3
Tow factors effect analysis.

Experiment no. Factors Function value yi

A B AB
1 1 1 1 5.4
2 1 1 1 5.2
3 1 1 1 7.0
4 1 1 1 6.4
Response Level 1 Rf,1 5.3 6.2 6.1 Average of yi equal 6.0.
Level 2 Rf,2 6.7 5.8 5.9
Effect Ef 0.7 0.2 0.1

small amount of experimental data to construct a mathematic

model like Eq. (7).
In Eq. (7), A  B is the ‘‘interaction term”; if it is removed from
Eq. (7), the numerical value of experiment will have an error of 0.1
as:
y1 ¼ 6 þ 0:7  ð1Þ  0:2  ð1Þ ¼ 5:5 ð9Þ
In the OED, if the interaction term A  B is neglected, Eq. (7) is
regarded as an output equation affected only by the main effect A
Fig. 2. The ERBFN structure jth forecast.
and B, as shown in Eq. (10).
yi ¼ Average þ EffectA  A þ EffectB  B ð10Þ
3.2. Hidden layer
From Table 3, the optimal level of factors can be obtained based
on the response of each level. Take the minimum as an example. In the hidden layer, Cj = [cj1, . . . , cjk, . . . , cjK], is called the jth cen-
Level 1 response of Factor A is 5.3, which is less than the Level 2 ter of ERBFN. kxji  cjkk is the Euclidean distance between the ith
response 6.7. Therefore, Factor A selects Level 1. The Level 2 re- node of the input layer and the kth node of the hidden layer. The
sponse of Factor B is 5.8, which is less than the Level 1 response Euclidean distance is determined by Eq. (13). The kth hidden layer
6.2, so Factor B selects Level 2. The final experimental output is output is defined as Eq. (14):
the minimum value 5.2. The above process is called ‘‘factor analy- vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N
sis”. From Eq. (7), we can also see that: uX
kxji  cjk k ¼ t ðxji  cjk Þ2 ð12Þ
i¼1
1. The OED can approximate the input–output relationship
through a simple linear equation when neglecting the interac- 0v ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
u N
tion term. uX
2. We can predict the system output not shown in the Orthogonal Hjk ¼ ujk @t ðxji  cjk Þ2 A ð13Þ
i¼1
Array based on (7).
3. We can use the OED to search for the extreme values of the
2 2
system. By selecting a proper factor-level in linear Eq. (7), the uðxÞ ¼ ex =r ð14Þ
preferable system output can be obtained.
u() is the Gaussian distribution function, and r is the smooth-
ing parameter.

3. Enhanced Radial Basis Function Network 3.3. Output layer

The Enhanced Radial Basis Function Network (ERBFN) consists In the output layer, wjk is the weight between the hidden node
of the input, hidden and output layers. The jth input vector Xj = [xj1, Hjk and the output node yj. The jth output of output layer is formu-
x j2, . . . , x ji, . . . , x jN ], i = 1, 2, . . . , N, j = 1, 2, . . . ,. M, is connected to lated as Eq. (16).
the input layer. The number of output nodes yj, j = 1, 2, . . . , M, is
equal to the number of training input–output data pairs. That is, in- X
K
yj ¼ wjk Hjk ð15Þ
put nodes (a matrix) and output nodes (a vector) are paired. The jth
k¼1
hidden nodes vectors are: Hj = [Hj1, H j2, . . . , Hjk, . . . , HjK], k = 1,
2, . . . , K, j = 1, 2, . . . , M. The weights wjk connect the kth hidden
node with the jth output node. The ERBFN structure is shown in 3.4. Stochastic Gradient Approach (SGA) [20]
Fig. 2. In Fig. 2, N is the number of input variables, M is the number
of training sets and K is the number of centers. In this paper, K is set In order to adjust the three parameters, which are weights wjk,
20. the center Cj of ERBFN and the smoothing parameters rjk of func-
tion u(), the Stochastic Gradient Approach (SGA) is adopted in this
3.1. Input layer paper. The procedure of SGA is described as follows:

In this paper, xi is the ith variable of the expected output. For (1) Calculate Euclidean distance kxji  cjkk.
each of the training data pairs, the input matrix can be described (2) Calculate hidden layer output Hjk by Eq. (14).
as: (3) Calculate output layer output by Eq. (16).
(4) Calculate the error between output yj and its expected value
X ¼ ½xji MN ; j ¼ 1; 2; . . . ; M; i ¼ 1; 2; . . . ; N: ð11Þ
Tj by error function. Error function is defined as Eq. (17):
 W.-M. Lin et al. / Applied Energy 87 (2010) 3226–3234 3229

ej ðnÞ ¼ ½T j  yj ðnÞ2 Table 5

" !#2 Factor analysis of the L4(23) orthogonal array.
X K
kX j ðnÞ  C jk ðnÞk2
¼ Tj  wjk ðnÞexp  ð16Þ Experiment no. Factors Error function
k¼1
r2jk ðnÞ 1 2 3
value ej

where ej(n) and yj(n) are the jth error and the jth output of 1 ljw_L1 ljc_L1 ljr_L1 2.34
2 ljw_L1 ljc_L2 ljr_L2 3.35
nth epoch, respectively.
3 ljw_L2 ljc_L1 ljr_L2 4.67
4 ljw_L2 ljc_L2 ljr_L1 3.52
(5) The related parameters are updated by Eqs. (17)–(19). Response Level 1 2.85 3.51 2.93
Level 2 4.20 3.44 4.01
@ Optimal level 1 2 1
wjk ðn þ 1Þ ¼ wjkðnÞ  ljw ej ðnÞ ð17Þ Optimal learning rates ljw_L1 ljc_L2 ljr_L1
@wjk
@
cjk ðn þ 1Þ ¼ cjkðnÞ  ljc ej ðnÞ ð18Þ
@cjk
4. Implementation of ERBFN
@
rjk ðn þ 1Þ ¼ rjkðnÞ  ljr ej ðnÞ ð19Þ
@ rjk Daily LMP profiles are similar on weekdays but different on
weekends. The fluctuation of the LMP may be due to loads, trans-
mission flow and temperature changes. The ERBFN is capable of
3.5. The Orthogonal Experimental Design (OED) Process coping with complicated interactions between those factors and
LMPs; therefore, those factors can be taken into account as inputs
The OED is used to optimize learning rate parameters ljw, ljc, to the ERBFN.
and ljr. The accuracy of these optimal parameters would be re- The input layer contains three input variables – loads, transmis-
fined in the dynamic environment to reach a minimum forecast er- sion flow and temperature. The hidden layer contains K hidden
ror. The procedure of OED is used to adjust the learning rate nodes. The output layer only contains one output variable which
parameters as shown in following steps. is the LMP. ERBFN training is done to minimize the fitting error
for a sample training set. For a given training data, the objective
Step 1. Learning rate parameters, ljw, ljc, and ljr are regarded function is defined as in Eq. (17). After training procedure, input
as three factors of OED, and each ej is regarded as a function the tth hr forecasting data in the tth hr ERBFN network to forecast
value yi in Table 3. the LMPt.
Step 2. Construct a Two-level Orthogonal Array LB(2F), where F The price profile presents seasonal characteristics, usually day
is the number of control factors, and B is the experiment num- and week cycles. The training data are classified as weekdays from
ber. B is defined as Eq. (20). Monday to Friday and weekends as Saturday and Sunday. The
selection of similar days for each season is part of the training in
B ¼ 2½log2 ðFþ1Þ ð20Þ the ERBFN. For example, eight similar days are selected for train-
In this paper, F is set as 3 (ljw, ljc, and ljr) and B as 4 by calcu- ing to predict the LMP on Monday or Sunday etc. One day is taken
lating Eq. (21). The orthogonal array is described as Eq. (1).

Step 3. Define the levels of ljw, ljc, and ljr as: Observation Location
LMP Website
(
lj L1 ¼ lj þ Dl
ð21Þ
lj L2 ¼ lj  Dl
Data Processor
where lj_L1: Level 1 of a learning rate parameter; lj_L2: Level 2 Periodic Sampling
of a learning rate parameter; lj: original value of a learning rate
parameter; Dl: predetermined alteration of a learning rate
Yes
parameter. A Large Change
The level defined of three factors is shown in Table 4. on PJM? New
Reconstruct Training
No ERBFN
Step 4. Perform the factor analysis and get the optimal learning Excel Database Workspace
rates. During the training procedure, OED is tied together with Matlab Workspace
1) Data cluster and Analysis
Eqs. (18)–(20) to select the best learning rates for minimizing 1) ERBFN 2) Other ANN Link
2) Data Storage
error function ej in each epoch. For example, Table 5 is devel-
oped by assigning the factors of learning rates to the level
defined in Table 4. The error function ej is calculated for each Predicting Results
experiment i with Eqs. (17)–(20). After the factor analysis, the 1) daily LMP of the various seasons
optimal learning rates are ljw L1 , ljc L2 and ljr L1 in some epoch. 2) weekly LMP of the various seasons

Table 4
Level definition in L4(23) orthogonal array.
Stop ?
Level Factor No
1 2 3 Yes
ljw ljc ljr
end
Level 1 ljw_L1 ljc_L1 ljr_L1
Level 2 ljw_L2 ljc_L2 ljr_L2
Fig. 3. The forecasting procedure of ERBFN.
3230 W.-M. Lin et al. / Applied Energy 87 (2010) 3226–3234

Table 6
The data sets for all studies Cases.

Case Training data Training data (date) Test data (date)

1 Similar days of Monday in summer 2002.05.27 (Monday)–2002.07.15 (Monday) 2002.07.22 (Monday)
2 Similar days of Sunday in summer 2002.05.26 (Sunday)–2002.07.14 (Sunday) 2002.07.21 (Sunday)
2002.03.18 (Monday)
3 Similar days 2002.01.21 (Monday)–2002.03.17 (Sunday) |
of weekly price in spring 2002.03.24 (Sunday)
2002.06.24 (Monday)
4 Similar days 2002.04.29 (Monday)–2002.06.23 (Sunday) |
of weekly price in summer 2002.06.30 (Sunday)
2002.09.23 (Monday)
5 Similar days 2002.07.29 (Monday)–2002.09.22 (Sunday) |
of weekly price in autumn 2002.09.29 (Sunday)
2002.12.23 (Monday)
6 Similar days 2002.10.28 (Monday)–2002.12.22 (Sunday) |
of weekly price in winter 2002.12.29 (Sunday)

Actual and Forecast DA/LMP Price Monday, July 22,2002

110
Forecasted (ERBFN)
ERBFN: Actual
RMSE=2.2853
100 MAPE=6.9140 %
MAE=2.0411

90

80
LMP($/Mwh)

70

60

50

40

30

20

10
2 4 6 8 10 12 14 16 18 20 22 24
Hour

Fig. 4. Actual and forecast day-ahead PJM electricity price of Case 1.

from the similar days as test data. The sample data of PJM is con- (hourly) for daily and weekly sample data. Fig. 3 is the forecasting
structed in EXCEL Workspace. The data analysis and data storage procedure of ERBFN.
can be easily manipulated with this database. To evaluate the
accuracy of the ERBFN in forecasting, the Mean Absolute Percent-
5. Results and analysis
age Error (MAPE), Mean Absolute Error (MAE) and Root Mean
Square Error (RMSE) have all been used. The MAPE, MAE, and
The data from the PJM website [4] were used to train and test
RMSE are defined as:
the proposed method. For comparison purposes, ERPFN, BPN,
1X T
jPture
t  Ppredict
t j PNN and other algorithms [11,12,23] were also built for tests.
MAPE ¼  100% ð22Þ
T t¼1 Pture
t
The data set are divided into two parts; the training data and test-
ing data are shown in Table 6. The training data were used for
training and updating the biases and learning rates. The test data
1X T
MAE ¼ jPture  Ppredict j ð23Þ were used to test the proposed methods after training. The simula-
T t¼1 t t
tion was implemented with Matlab on a PIV-2.6GHZ computer
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi with 512 MB RAM.
u T
u1 X ture
RMSE ¼ t jP  Ppredict j2 ð24Þ
T t¼1 t t
5.1. Simulation results

where P ture
t is the actual price at hour t and P predict
t is the forecasting Fig. 4 shows the day-ahead price forecasting for Case 1. The
price at hour t. T is the number of testing data. T is 24 and 168 forecasting results are compared with the actual LMP value. The
 W.-M. Lin et al. / Applied Energy 87 (2010) 3226–3234 3231

Actual and Forecast DA/LMP Price Sunday, July 21,2002

50
Forecasted (ERBFN)
ERBFN: Actual
RMSE=1.2310
MAPE=5.0163 %
45 MAE=1.0659

40
LMP($/Mwh)

35

30

25

20

15

10
2 4 6 8 10 12 14 16 18 20 22 24
Hour
Fig. 5. Actual and forecast day-ahead PJM electricity price of Case 2.

Actual and Forecast DA/LMP Price, March 18 2002 ~ March 24 2002

60
ERBFN: Forecasted (ERBFN)
RMSE=2.5905 Actual
55 MAPE=7.1546 %
MAE=2.0307

50

45
LMP($/Mwh)

40

35

30

25

20

15

10
20 40 60 80 100 120 140 160
Hour
Fig. 6. Actual and forecast day-ahead PJM electricity price of Case 3.

MAPE of ERBFN is 6.9140%. The forecasting results of ERBFN can
track very closely the actual LMP with OED applied. Note that
the algorithm could even track the first spike nicely; this is not eas-
ily attainable by other algorithms. A common neural network such
as PNN won’t have this capability.
Similarly, day-ahead price forecasting for Case 2 is shown in Fig. 5,
where the MAPE value is about 5% for the ERBFN Method. According
to the above results, the ERBFN has the ability to find better solutions.
Figs. 6 and 7 show the weekly price forecast for Case 3 and Case
4, respectively. Case 3 is a typical spring week and the forecasting
price of ERBFN is close to the actual LMP values. From Fig. 7, it can
be seen that ERBFN has the capability of following the spike as
shown in the 25th–48th (Tuesday) hour of this week. MAPE from
the ERBFN method is 6.0817%.
Similarly, Figs. 8 and 9 show the weekly price forecasting re-
sults for Case 5 and Case 6, which are a typical autumn and winter
3232 W.-M. Lin et al. / Applied Energy 87 (2010) 3226–3234

Actual and Forecast DA/LMP Price, June 24 2002 ~ June 30 2002

140
Forecasted (ERBFN)
ERBFN: Actual
RMSE=3.0230
MAPE=6.0817 %
120 MAE=2.4610

100
LMP($/Mwh)

80

60

40

20

0
20 40 60 80 100 120 140 160
Hour
Fig. 7. Actual and forecast day-ahead PJM electricity price of Case 4.

Actual and Forecast DA/LMP Price, September 23 2002 ~ September 29 2002

60
Forecasted (ERBFN)
ERBFN: Actual
RMSE=1.6126
MAPE=5.5522 %
MAE=1.3485
50

40
LMP($/Mwh)

30

20

10

0
20 40 60 80 100 120 140 160
Hour
Fig. 8. Actual and forecast day-ahead PJM electricity price of Case 5.

week, respectively. In Case 5, the weekly MAPE value of ERBFN is
lowest among Cases 3–6.
From Case 6, the MAPE of ERBFN is 7.0867%, which is very close
to the MAPE of Case 3. As shown in Figs. 6 and 9, the bumpier the
LMP characteristic is; the larger the MAPE of forecasting is.
5.2. Performance test
Table 7 shows the training performances of the ERBFN, PNN and
BPN for comparison in Case 6. The PNN has a fast training time,
needing no iteration for updating weights. However, with the reg-
ulation of learning rates lw, lc and lr from the OED, the ERBFN has
a greater accuracy than the other methods. The convergent time of
ERBFN is about 57.4538 s, which is faster than the BPN. The BPN
needs more than 3000 iterations and is time consuming. With
the same training data, the proposed ERBFN shows an overall bet-
ter performance than the PNN and BPN.
5.3. Accuracy test for EPNN
Fig. 10 shows the curves of the daily RMSE in the summer (Case
3). For these tests, the number of sets of training data increases from
 W.-M. Lin et al. / Applied Energy 87 (2010) 3226–3234 3233

Actual and Forecast DA/LMP Price, December 23 2002 ~ December 29 2002

90
Forecasted (ERBFN)
ERBFN: Actual
RMSE=2.3730
80 MAPE=7.0867 %
MAE=1.9145

70
LMP($/Mwh)

60

50

40

30

20

10
20 40 60 80 100 120 140 160
Hour
Fig. 9. Actual and forecast day-ahead PJM electricity price of Case 6.

Table 7
Comparison with ERBFN, PNN and BPN in Case 6.

Method Number of sets of training data Learning parameter Training iteration Training time (s) MAPE (%) MAE RMSE
ERBFN 1344 lw = OED 30 57.4538 7.0867 1.9145 2.3730
lc = OED
lr = OED
PNN 1344 r = 0.1 – 3.4680 21.6921 5.1050 6.0296
r = 0.2 – 3.5780 29.1141 6.4257 7.4709
r = 0.3 – 3.4530 33.0441 7.1755 8.3694
BPN 1344 g = 0.65 >3000 >780.2415 14.3677 5.5401 7.9461
g = 0.22 >3000 >775.5627 30.6072 12.7365 17.1425
g = 0.10 >3000 >778.9645 21.1673 13.0457 17.3297

Table 8
Daily Accuracy Test Comparison with other algorithms in Case 5.
7
Mon Method MAPE (%) MAE RMSE
Tue
Wed ERBFN 5.5622 1.3485 1.6126
6 Thu k-factor GIGARCH [11] 6.4943 1.5943 1.9761
Fri ARMA-1 [12] 7.5416 1.7922 2.0777
Sat ARMA-2 [12] 6.7495 1.6352 2.0024
5 Sun ARMA-3 [12] 6.9011 1.6672 2.0338
ARMA-4 [12] 7.2656 1.7238 2.0187
PSO [23] 6.1520 1.5033 1.8665
(RMSE)

4 k-factor GIGARCH: k-factor GIGARCH process.

ARMA-1: ARMA, varying intercept, jumps, in logs, significant.
ARMA-2: crossed ARMA, varying intercept, jumps, significant.
3 ARMA-3: crossed ARMA, varying intercept, jumps, in logs,
ARMA-4: crossed ARMA, varying intercept, jumps, in logs, significant.
PSO: Particle Swarm Algorithm.

2
48 to 286 (2 to 12 days), and the execution time increases from
10.7854 s to 85.1708 s. The RMSE of all tests is well below 4% with
1 192 training data (8 days). It can be seen that it is accurate enough
2 4 6 8 10 12
Data sets in Days for ERBFN with 192 training data sets. This characteristic can be uti-
lized to greatly reduce the training time, and data storage can be re-
Fig. 10. The RMSE of ERBFN for various data sets. duced without losing originality. We can minimize the data storage,
3234 W.-M. Lin et al. / Applied Energy 87 (2010) 3226–3234
shorten the preprocessing needs and reduce the network size. From
Fig. 7, we can also see that the actual LMP on Wednesday (hours 48–
72) are more volatile than it is on Sunday. The RMSE on Wednesday
is larger, as expected from the training data.
5.4. Comparison with other algorithms
The simulation result of the case 5 is selected to compare with
other algorithms in Table 8 [11,12,23]. As shown in Table 8, the
proposed algorithm has the advantages of high accuracy and expli-
cit reasoning. It also shows that the proposed algorithm did have a
more robust structure and a better capability to improve the prob-
lem of predicting prices spikes than other algorithms.
6. Conclusions
The ERBFN integrated the RBFN and OED to forecast LMPs based
on similar days. The RBFN has the capability of dealing with varied
and complicated relationships between input and output data, and
the OED helps with the appropriated regulation of learning rates
lw, lc and lr to improve the forecasting results. The actual data
of the PJM were used to demonstrate the performance of the ERB-
FN. For the selected days and weeks under study, the daily and
weekly RMSE values are calculated and shows that the ERBFN
can even track the spikes closely, which is not easily attainable
through other methods. In comparison with the PNN and BP, the
results demonstrate that the ERBFN is robust, efficient and accu-
rate. It proves that the ERBFN has the capability to produce better
results for volatile price forecasting.
References
[1] Ma X, Sun DI, Cheung KW. Evolution toward standardized market design. IEEE
Trans Power Syst 2003;18(2):PP460–469.
[2] Shahidehpour M, Yamin H, Li Z. Market operations in electric power
systems. John Wiley & Sons Inc.; 2002.
[3] Erdogdu E. A paper on the unsettled question of Turkish electricity market:
Balancing and settlement system (part I). Appl Energy 2010;87:251–8.
[4] PJM energy management: <http://www.PJM.com/index.jsp>.
[5] Inglesi R. Aggregate electricity demand in South Africa:Condition forecasts to
2030. Appl Energy 2010;87:197–204.
[6] Cuaresma JC, Hlouskova J, Kossmeier S, Obersteimer M. Forecasting electricity
spot-prices using linear univariate time-series models. Appl Energy
2004;77:87–106.
[7] Lora AT, Santos JMR, Expósito AG, Ramos JLM, Santos JCR. Electricity market
price forecasting based on weighted nearest neighbors techniques. IEEE Trans
Power Syst 2007;22(3):1294–301.
[8] Contreras J, Nogales FJ, Conejo AJ. ARIMA models to predict next-day electricity
prices. IEEE Trans Power Syst 2003;18(3):1014–20.
[9] Zhou M, Yan Z, Ni YX, Li G, Nile Y. Electricity price forecasting with confidence-
interval estimation through an extended ARIMA approach. IEE Proc-Gener
Transm Distrib 2006;153(2):187–95.
[10] Conejo AJ, Plazas MA, Espínola R, Molina AB. Day-ahead electricity price
forecasting using the wavelet transform and ARIMA models. IEEE Trans Power
Syst 2005;20(2):1035–42.
[11] Diongue AK, Guegan D, Vignal B. Forecasting electricity spot market prices
with a k-factor process. Appl Energy 2009;86:505–10.
[12] Kosater P, Mosler K. Can Markov regime-switching models improve power-
price forecasts? Evidence from German daily power prices. Appl Energy
2006;83:943–58.
[13] González AM, Roque AMS, García-González J. Modeling and forecasting
electricity prices with input/output hidden Markov models. IEEE Trans
Power Syst 2005;20(1):13–24.
[14] Mandal P, Senjyu T, Urasaki N, Funabshi T, Srivastava AK. A novel approach to
forecast electricity price for PJM using neural network and similar days
method. IEEE Trans Power Syst 2007;22(4):2058–65.
[15] Catalão JPS, Mariano SJPS, Mendes VMF, Ferreira LAFM. An artificial neural
network approach for short-term electricity prices forecasting. In: The 14th
international conference on intelligent system applications to power systems.
ISPA, Kaohsiung, Taiwan; November 2007. p. 430–5.
[16] Zhang L, Luh PB, Kasiviswanathan K. Energy clearing price prediction and
confidence interval estimation with cascaded neural networks. IEEE Trans
Power Syst 2003;18(1):99–105.
[17] Catalão JPS, Mariano SJPS, Mendes VMF, Ferreira LAFM. Short-term electricity
prices forecasting in a competitive market: a neural network approach. Electr
Power Syst Res 2007;77:1297–304.
[18] Pao HT. Forecasting electricity market pricing using artificial neural networks.
Energy Convers Manage 2007;48:907–12.
[19] Pino R, Parreno J, Gomez A, Priore P. Forecasting next-day price of electricity in
the spanish energy market using artificial neural networks. Eng Appl Artif
Intell 2008;21:53–62.
[20] Lin CH, Tsao MC. Power quality detection with classification enhancible
wavelet-probabilistic network in a power system. IEE Proc-Gener Transm
Distrib 2005;152(6):969–76.
[21] Ham FM, Kostanic I. Principal of neurocomputing for science and
engineering. McGraw-Hill Companies, Inc.; 2001.
[22] Ross PJ. Taguchi techniques for quality engineering. Section ed. The McGraw-
Hill Companies Inc.; 1988. p. 203–43.
[23] Alrashidi MR, EL-Naggar KM. Long term electric load forecasting based on
particle swarm optimization. Appl Energy 2010;87:320–6.