Neural Computing and Applications

https://doi.org/10.1007/s00521-018-3544-8 (0123456789().,-volV)(0123456789().,-volV)

ORIGINAL ARTICLE

Short-term prediction of market-clearing price of electricity

in the presence of wind power plants by a hybrid intelligent system
Afshin Aghajani1 • Rasool Kazemzadeh1 • Afshin Ebrahimi1

Received: 12 May 2017 / Accepted: 11 May 2018

 The Natural Computing Applications Forum 2018

Abstract
This paper provides a new hybrid intelligent method for short-term prediction of the market-clearing price of electricity in
the presence of wind power plants. The proposed method uses a data filtering technique based on wavelet transform and a
radial basis function neural network, which is utilized for primary prediction. The main prediction engine comprises three
MLP neural networks with different learning algorithms. To get rid of local minimums and to optimize the all neural
networks, the meta-heuristic Imperialist Competitive Algorithm method is used. The input data for network training belong
to the Nord Pool power market. The information includes a complete set of the historical record on electricity price and
wind power generation. Moreover, the simultaneous impact of wind power generation is analyzed to predict the market-
clearing price. Besides, the correlation coefficient factor is provided to consider the impact of wind power in forecasting the
electricity price. Simulation results show the supremacy of the proposed method over other methods, to which it has been
compared in this study. Also, the prediction error decreases significantly.

Keywords Neural networks  Imperialist competitive algorithm  Power system market  Price forecasting

1 Introduction along with risk and fluctuations determined in the com-

petitive market change the behavior of the market partici-
Following the restructuring of the electricity market from a pants [6]. Hence, price forecasting is crucial for electricity
state monopoly market to a more competitive one, elec- producers and consumers to plan [7], invest, exploit and
tricity price is determined by market forces and is of also manage the risk resulting from intense price changes
paramount importance for the electricity market partici- [8]. Short-term price forecasting has the greatest impact on
pants [1, 2]. Electricity demand depends on the economic pricing strategy [9, 10]. Thermal pollution is increased and
activities and climate conditions [2, 3]. Therefore, insta- greenhouse gases are produced more, due to the growth of
bility in demand leads to sharp fluctuations in electricity electrical energy generation resulting from thermal power
prices [4]. In order to have a dependable prediction of plants. It causes more interest in power generation based on
electricity prices offered in the market, these features renewable energies [11]. Among renewable energy
(sharp fluctuations in electricity prices and instability in resources, the wind and Solar PV have more importance
demand) should be considered in the prediction model because of their prevalence and relatively low production
while using appropriate modeling tools [5]. Instant prices cost [12, 13]. According to predictions, 12% of the world’s
electricity would be generated from wind energy by 2020,
and it would have a high annual growth rate [14]. One of
& Rasool Kazemzadeh
r.kazemzadeh@sut.ac.ir the major setbacks of wind energy generation is that the
wind regeneration process is uncertain. Consequently, the
Afshin Aghajani
aghajani2006@gmail.com production capacity of wind power plants is intermittent
[15, 16]. With the presence of wind power plants in the
Afshin Ebrahimi
aebrahimi@sut.ac.ir electricity market, which creates an imbalance between
production and demand, volatility and uncertainty of the
1
Renewable Energy Research Center, Electrical Engineering electricity price would include new and complex
Faculty, Sahand University of Technology, Tabriz, Iran

123
 Neural Computing and Applications

dimensions [17–19]. Therefore, as wind power becomes 1. Providing a new hybrid intelligent approach for the
more important in electric power systems, it encompasses short-term prediction of electricity market-clearing
both types of uncertainty (uncertainty in the wind power price accurately.
and electricity price) and, as a result, price prediction 2. Examining the effect of wind generation capacity on
becomes more difficult [20, 21]. Reviewing literature predicting the electricity market-clearing price.
reveals a great number of research works conducted on 3. Providing correlation coefficient factor to consider the
forecasting the electricity price over the recent years. The impact of wind power on the accuracy of MCP
available studies aim to decrease the prediction error. To forecasts.
this end, further researches are required to improve fore- 4. Investigating the prediction accuracy of the proposed
casting accuracy [1, 22, 23]. Predictions are often short method compared to other proposed methods in the
term [24, 25] and are divided into two major categories: field of Soft Computing.
soft computing and hard computing [26]. AR, ARIMA and 5. A detailed analysis of a case study in the Nord Pool
GARCH methods and their combination with other algo- market.
rithms such as Wavelet-ARIMA are among the methods
The Nord pool power market is the biggest and the
used in the hard computing [27, 28]. These methods require
oldest power market in Europe since the 1990s. This
abundant physical data, detailed modeling of the system to
market is active in four countries on Norway, Sweden,
be predicted, and complex computations [1]. Regarding
Finland, and Denmark. Then, the sub-sections include:
soft computing, the applied methods are self-learning and
Sect. 2 describes proposed approach, imperialist competi-
use the earlier information to create the model [22]. They
tive algorithm, correlation coefficient, wavelet transform,
include neural networks (NNs) [29, 30], intelligent and
error measurement criteria and technical data. Section 3,
meta-heuristic methods (CNEA, SSA, FNN, SAR-NN) and
illustrates prediction engine and proposed algorithm’s
hybrid methods (WT ? PSO ? ANFIS, NN ? WT, HIS)
steps. Section 4, provides investigation and analysis of the
[26, 31–35]. The artificial neural network is one of the most
simulation results. Finally, Sect. 5 provides conclusion of
widely used methods for forecasting electricity prices in
the present work. The notation used throughout the paper is
soft computing [36, 37]. As compared to hard computing,
given in the Appendix.
the lack of complex mathematical equations to describe the
model and great learnability capacity are among the priv-
ileges of the neural networks [38]. Due to the highly
2 Proposed approach
nonlinear nature of the electricity price signal, those
approaches with their great potential in modeling nonlinear
The main objective of this study is to decrease the pre-
functions should be employed to give a more accurate
diction error of the electricity price in the competitive
prediction [26]. A neural network is an approach that can
market with the presence of wind power plants. Therefore,
create a nonlinear model through learning processes
some parameters including wind power, an effective
[39, 40]; however, they are not sufficient for modeling
combination of forecasting methods, information process-
highly nonlinear signals. Therefore, a combination of
ing tools, and meta-heuristic methods are used to improve
neural networks and intelligent systems is used to model
the prediction results. As statistical methods employ his-
such signals [41, 42]. The hybrid methods are used because
torical data, a data mining problem is faced [48]. Data
they improve prediction accuracy using the strengths of
mining is involved in three parts: data preparation, mod-
each method [43]. Given that each of the methods is sen-
eling, and model evaluation. In Fig. 1, the proposed algo-
sitive to certain situations, another advantage of hybrid
rithm structure is summarized.
methods is the reduced risk of unexpected events [38, 44].
The first stage in data preparation is to extract data from
Generally, in predicting the electricity price, the concerned
data sources. In this research, historical information set of
input component is the historical record of prices for the
electricity price signals is used as basic information, as
same period [45, 46], considering that increasing wind
well as information on the amount of wind power gener-
power generation reduces the electricity price [47]. In the
ation. Considering the relationship between climate con-
present study, in addition to the historical record of price,
ditions and wind power generation [49], and also the
information on the wind power plant is also considered as
impact of climate conditions on power consumption,
one of the input variables to predict the electricity price.
information on the climate conditions is also indirectly
The main motivation of this research can be mentioned in
considered through applying wind power information to
the following:
predict electricity prices. The second step is to pre-process
the extracted data. Hence, using different methods, the data
shall be prepared for training model. The wavelet transform

123
Neural Computing and Applications

Pn   
Input Data i¼1 Xi  X Yi  Y
Primery predictor Primery r ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn   2 Pn ð1Þ
(Wind power & forecang of 2
Electricity price) ANN (RBF)+ICA electricity price i¼1 Xi  X i¼1 ðYi  Y Þ

1 X
n
X ¼ Xi ð2Þ
n i¼1
Details &
Add to input data Wavelet
Aproximaon of
(Electricity price) Transform
Price signal 1 X
n
Y ¼ Yi ð3Þ
n i¼1

 Y are the
where r is the sample correlation factor and X,
Add to input data Main predictor forecasng results
average rate of the two index correspondingly.
(Wind power) ANNs (MLP)+ ICA (combine of results)

2.1 Imperialist competition algorithm

Fig. 1 The proposed algorithm structure
Imperialist competitive algorithm (ICA) is one of the
is used in this study. Since the inputs of the neural network evolutionary algorithms that is inspired by the human’s
are not homogenous, data have been normalized in order to socio-political evolution. This algorithm was first offered
homogenize the inputs and to increase the model’s strength by Atashpaz-Gargari and Lucas [51]. Each species of the
and training speed. For modeling, a suitable method should population is named a country. All the countries are allo-
be used in a correct place. Therefore, a suitable preliminary cated to two categories, colonies, and imperialist states.
predictor should be used to improve the outcomes and Imperialistic competition is the central part of this algo-
assist the main predictor. Radial neural networks such as rithm and optimistically results in the convergence of
RBF can be used as the preliminary predictor. This is countries to the global minimum of the cost function. In
because these networks have a superior capability to model this competition, the weak empires collapse step by step
nonlinear relations and search local features of the input and lastly, there is only one imperialist that all the other
data. For the main predictor, one can use the combination countries are its colonies. ICA has shown its exceptional
of some neural networks, containing MLP which uses capability for the many problems [52–55]. In [11, 56], ICA
different techniques for learning. MLP Neural networks are pseudocode is defined.
worthy of catching global data trends and modeling non-
linear manners. The mixture of neural networks of MLP 2.2 Wavelet transform
and RBF leads to a consideration of a complete set of local
and global manners of the target variables. Meta-heuristic Wavelet Transform (WT) method is a mathematical tactic
algorithms have great exploration capabilities. For addi- for noise reduction without abolishing the dynamics of the
tional optimization of forecasting engine, the meta- original signal [57]. The price data series contains various
heuristic technique of ICA is used. Considering the above oscillations, spikes, and different types of nonstationarity.
explanations, the mixture of the RBF, wavelet, and series The WT can be considered as feature management tool to
neural networks containing three kinds of MLP with dif- separate these spikes. Therefore, price forecasting using the
ferent learning schemes along with meta-heuristic algo- WT can be used to decrease price forecast error. Wavelet
rithm of ICA for prediction is used. In the model evaluation transforms are generally separated into two groups: discrete
step, the forecasted value is evaluated with respect to the wavelet transforms (DWT) and continuous wavelet trans-
actual value based on error criteria. The proposed method forms (CWT) and [58, 59]. The CWT is given by [60, 61]:
is designed for day-ahead price forecasting. These com- Zþ1
ponents are described next. 1
CWTx ða; bÞ ¼ pffiffiffiffiffiffi w ðtÞxðtÞdx; a[0 ð4Þ
A correlation coefficient is a statistical measure of the jaj
1
degree to which deviations to the rate of one variable  
calculate the change in the rate of another. In positively 1 tb
wa;b ðtÞ ¼ pffiffiffi w ; a [ 0 and  1\b\ þ 1
correlated variables, the value rises or falls in tandem. In a a
negatively correlated variables, the value of one rises as the ð5Þ
value of the other falls. Correlation coefficients are stated
as values between ? 1 and - 1. The correlation coefficient where x(t) is the signal to be evaluated, wa,b(t) is the mother
of n-dimensional row vector X and Y is as follows [50]: wavelet scaled by a factor a and shifted by a translated
parameter b, and * denotes complex conjugate. CWT will
be very slow due to overlapping feature and duplicity of

123
 Neural Computing and Applications

neighbor data. In addition, it will have extra and unusable A1 ¼ A2 þ D2 ð7Þ

data [62]. DWT uses scale and position values based on
A2 ¼ A3 þ D3 ð8Þ
powers of two, called dyadic dilation and translations,
which are achieved by discretizing the scaling and trans- S ¼ A1 þ D1 ¼ A2 þ D1 þ D2 ¼ A3 þ D1 þ D2 þ D3
lation factors, denoted as [60, 61]: ð9Þ
X
T 1  
t  n  2m The detail (D1, D2, D3) and approximation (A1, A2,
DWTx ðm; nÞ ¼ 2ð 2 Þ
m
xðtÞw ð6Þ
t¼0
2m A3) signals are achieved by downsampling and are only
half the size of the original signal. Hence, before repro-
where T is the length of the signal x(t). The translation and ducing the original signal, it is essential to reconstruct the
scaling parameters are functions of the integer variables m approximation and detail coefficients [62].
and n, where a = 2m and b = n2m, and t is the discrete time There are several kinds of wavelet basis functions that
index. Stephane Mallat’s multiresolution model is fre- can be used as a mother wavelet for WTs such as Morlet,
quently used to employ DWT in the related literature [63]. Mexican Hat, Haar, and Meyer. Collections of them are
This method is composed of two main steps: decomposi- called families [64–66]. Between these families, Daube-
tion and reconstruction. Figures 2 and 3 demonstrate chies have often superior results [61, 67, 68]. In this
related steps in decomposition and reconstruction in this research, a wavelet function of type Daubechies of order 4
method [62]. The reconstructed original signal S can be (abbreviated as db4) is used as mother wavelet. This
created with the following tactic: wavelet suggests an appropriate trade-off between wave-
length and smoothness, resulting in a suitable performance
for short-term electricity prices forecasting.

2.3 Error measurement criteria

Considering that all forecasts naturally have errors, appli-

cation procedure of error criteria for forecast purpose is very
important [69]. Different criteria have been presented in the
literature which can be categorized into two main groups
[69, 70]: first-order and second-order criteria. These criteria
are defined based on the vertical variance of actual values.
Based on the above description, the mean absolute percent-
age error (MAPE) criterion is defined as follows [47]:
 

N MCPðtÞ  MCP d ðtÞ
1 X
%MAPE ¼  100 ð10Þ
N t¼1 MCP
Fig. 2 Wavelet decomposition where N is the number of hours, MCP(t) is the actual
d
market-clearing price at the hour t, MCP(t) is the fore-
casted market-clearing price at the hour t, and MCP is the
average value of actual market-clearing price for all the N
test hours, which is given by [47]:

1X N
MCP ¼ MCPðtÞ ð11Þ
N t¼1

MAPE is connected to error’s first order and offers the

absolute mean of forecasting error.

2.4 Technical data

In Table 1, the parameters and properties of the employed

optimization algorithms have been brought. These values
are obtained by trial and error method.
Fig. 3 Wavelet reconstruction

123
Neural Computing and Applications

Table 1 The parameters and

ANN ICA
properties of the employed
optimization algorithms Parameters Value Parameters Value

Number of neurons in the hidden layer 10 Number of initial countries 40

Learning coefficient (g) 0.9 Number of initial imperialists 8
Momentum (a) 0.2 Revolution rate 0.3
Activation functions in the hidden layer TANSIG Assimilation coefficient (b) 2
Activation functions in the output layer PURELIN Assimilation angle coefficient (c) 0.5
Number of epochs 1000 f 0.02

3 Prediction engine 1. The process of selecting the appropriate data is one of

the influential factors in the training power of neural
Forecasting engine has two parts [48]: (1) Preliminary networks [72, 73]. Therefore, the first step is to prepare
predictor, (2) the main predictor. In this research to opti- data. Input data are of two categories (wind power and
mize the forecasting engine, the meta-heuristic ICA electricity price). Information on the electrical power
method is used. Preliminary predictor consists of a radial signal and wind power generation on the same days
basis function (RBF), which aims at finding local solutions and at the same hours close to the electricity price
and employs meta-heuristic ICA algorithm to optimize the prediction time significantly contributes to the accu-
weights and to determine the best centers of radial func- racy of predictions and has the greatest impact. In this
tions and distances. The main predictor is made of three research, the past price information, which is closely
MLP neural networks coupled with training algorithms of associated with the price of next hour, is determined by
LM, BFGS, and BR [59], which are associated with meta- feature selection method. This case has also been the
heuristic ICA algorithm to optimize the weights. MLP same in the references [58, 74]. In the studies that are
neural networks have a hidden layer and this hidden layer investigated and compared, this method has also been
consists of 10 neurons with a trial and error approach. used, which is followed by the similar results [47]. In
Figure 4 shows the main forecasting engine. the used historical record of the electricity prices
Figure 5 depicts a schematic diagram of the proposed signals, for the 3 past hours it was predicted and for the
method for predicting market-clearing price. Predictions other days, which are about a week; values are
are related to 1-h intervals and 24-h periods. The impact of considered for the 2 past hours. (The set of past
different inputs on simulation results is also examined. prices: Pt1, Pt2, Pt3, Pt24, Pt25, Pt48, Pt49, Pt72, Pt73, Pt96,
Information used is relevant to the western region of Pt97, Pt120, Pt121, Pt144, Pt145, Pt168 and Pt169). Like
Denmark (DK1) [71], which possesses high wind power electricity price, the background of wind generation
penetration in the power system and has nonlinear effects capacity, in addition to the predicted value for the
on the market-clearing price. Then, the sequential exami- concerned time of forecasting electricity prices, is also
nation of the proposed method is presented. used. Hence, as much as 35 data are considered as the
input to train the neural network. 48 days are consid-
ered as the training period.
2. Correlation analysis is done in the second step to filter
out irrelevant candidate inputs.
Weights
NN1 (LM) 3. The third step is related to the preliminary predictor.
This is performed by the RBF, which has been
Weights & Target optimized by the ICA algorithm.
4. In the fourth step, the output of the RBF ? ICA, which
Weights
NN2 (BFGS) ICA is a preliminary prediction of the electricity price, is
added to the electricity price information and wavelet
Weights & Target transform is conducted on this data set and the time
series for electricity prices are analyzed into four
Weights
NN3 (BR) signals A3, D1, D2, and D3. For each of which, the
fifth and the sixth steps must be replicated.
5. In the fifth step, due to the positive effect of
Final Target
normalizing data on the performance of neural
Fig. 4 Main forecasting engine

123
 Neural Computing and Applications

Fig. 5 Schematic diagram of

the proposed method
Market-Clearing Price Forecasng Flowchart

START

Set Historical Data series (Price & Wind power)

Correlaon Analysis

Primary Predictor
(Wind Power)
(RBF)+ICA

(Price)

Apply Wavelet Transform to create

detail and Approximaon series

A3,D1,D2,D3

Normalizaon

Forecaster engine Forecaster engine Forecaster engine Forecaster engine

Based HNN+ICA Based HNN+ICA Based HNN+ICA Based HNN+ICA

Reverse the Normalizaon

Inverse WT to
find price

MCP Forecasts & error

END

networks, fourth-step output data along with wind training phase, the weights of the neural network are
power data are normalized. regarded as meta-heuristic algorithm input. When the
6. The sixth step is related to the main predictor. The fifth meta-heuristic algorithm achieves optimal results, the
step output forms this step input. The main predictor final values will be considered as the final weights of
contains a combination of three MLP neural networks the neural network. These weights will also be
along with training algorithms, LM, BFGS, and BR. In considered as the initial weights of the next neural
each of these networks after finding weights in the network. This process is repeated for each of the three

123
Neural Computing and Applications

neural networks. In the prediction phase, the output of

each of these networks would be regarded as input
along with other neural network inputs. In each of
these three neural networks, at any stage, we will have
a better prediction of electricity price.
7. All four analyzed series are obtained at the fourth step
and investigated at the fifth and sixth steps. They
would then be converted to real values in the seventh
step.
8. By applying the inverse wavelet transform on the
results of the seventh step, electricity price signal
reconstruction is performed and the predicted value of
electricity price signal is obtained.
The above procedure is repeated for ten times. And,
each time by altering test data, validation and training are
performed at the rate (15, 15, 70). Finally, average pre-
dicted values are reserved as final results. Test data have Fig. 6 A sample of the training process, training errors, validation,
and testing in each epoch
not been used in the training phase.
12 DEC 2011
400.00

4 Evaluation and analysis of the simulation

Price (DKK/MWh)
350.00
results 300.00

250.00
Data in the neural network are divided into three major
categories; in this research, the composition of these data is 200.00

as follows: 150.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Hours
• Training (70% input data).
Actual Forecast
• Validation (15% input data).
• Test (15% input data). Fig. 7 The train results on December 12, 2011

In Fig. 6, a sample of the training process, training

Days were February 18, May 19, August 13, and Novem-
errors, validation, and testing is shown in each epoch. As
ber 12, respectively. The comparison criteria used in this
can be observed, increasing the number of epochs leads to
table are MAPE. Reviewing the table shows that the pro-
less training error. But after a point, the validation error is
posed method has superiority over other methods and
increased slightly; this is where the over-fitting occurs.
prediction error in all seasons is relatively less. In spite of
Therefore, the training process is stopped. Figure 7 shows a
severe shocks and extreme nonlinear changes in electricity
sample of the training results on December 12, 2011. The
price, using the wavelet transform has led to filtering these
information selected for validation and testing is not used
sharp fluctuations. The preliminary prediction engine has
in the training step. The results presented in the simulation
had an effective role in improving outcomes. In general,
section of this research are related to the third part (test
improved prediction results, compared to the other meth-
data) and have not been used in the training part.
ods, indicate the right position of the wavelet filter and
Simulations are performed in 1-h intervals for 24-h
combination of meta-heuristic algorithm and neural net-
periods. Moreover, to investigate the effect of input
works. Comparing the results in Table 2 shows that the
parameters, the simulation results are compared in two
combination of neural network and smart algorithms leads
conditions in the case of any change in the type and number
to an improvement in the prediction results.
of input data. Methods compared with the proposed
Figures 9 and 10 represent the simulation results of May
(RBF ? HNN ? WT ? ICA) method are as follows [75]:
and November in the proposed method. In this simulation
(ANN, ANN ? WT, ANN ? RBF, ANN ? ICA ? GA).
process, two types of data (wind power and electricity
Table 2 and Fig. 8 show the simulation results for 1-h
price) are used for the network training. The figures show
intervals for 24-h periods and 2 types of data (electricity
the predicted values, actual values, and MAPE value for
prices and wind power). In Table 2, the results are pre-
the concerned day.
sented for 4 different days of different seasons in 2012.

123
 Neural Computing and Applications

Table 2 MAPE results for the

Models Months Average
hybrid prediction neural
networks within 24 h (2 types of Feb May Aug Nov
data)
ANN 9.0690 5.9549 6.8972 5.9236 6.9611
ANN ? WT 6.7485 5.6438 6.3291 5.7354 6.1142
ANN ? RBF 6.5568 4.3479 5.3829 4.2176 5.1263
ANN ? ICA ? GA 6.4826 3.1225 4.9279 4.1119 4.6612
RBF ? HNN ? WT ? ICA 1.8886 2.9712 3.9284 2.9623 2.9376

10 analyzing and comparing the figures along with the find-

9
8
ings presented in Table 3. The analysis of the figures rele-
7 vant to February 18 reveals that the amount of wind power
MAPE (%)

6
5
generation is greater than the common amount and data
4 correlation has a positive slope with a small value. In this
3 case, the amount of power generated by the wind not only
2
1 has little impact on the market-clearing price but also has a
0 demolishing effect on the prediction of market-clearing
Feb May Aug Nov Average
Month
price. In such cases, including the information relevant to
the wind power leads to deviation from the right prediction
ANN ANN+WT ANN+RBF ANN+ICA+GA RBF+HNN+WT+ICA
value. Therefore, the prediction results contain fewer pre-
Fig. 8 MAPE results for the hybrid prediction neural networks within diction errors if the information of electricity price is only
24 h (2 types of data) used.
Analysis of the figures on May 19 and November 12
Table 3 and Fig. 11 show the simulation results without shows that the wind power generation value is on an
correlation analysis in a 1-h interval for a 24-h period to acceptable scale and data correlation has a negative slope
verify the impact of the wind power generation on the with the proper amount. Hence, the wind power generation
accuracy of the results. has had an effect on electricity price. In such cases, using
Considering the impact of wind power generation on the wind power information in electricity price prediction
electricity price, it is observed that MAPE has decreased in will have a significant impact on reducing the prediction
some cases and had a relative increase in some other cases error.
compared to the condition associated with two types of On August 13, although the wind power generation is
input data. It may be because of the influencing amount of within an acceptable range, the amount of data correlation
the wind power on electricity price. has a positive slope and a small value. Therefore, the use of
Figures 12, 13, 14, 15, 16, 17, 18 and 19 show the data related to wind power enhances the prediction error.
amount of power produced by wind, electricity price, and Results for August 28 is the same as, due to a positive
the correlation between wind power generation and elec- correlation between wind power generation and electricity
tricity price in simulated days. The reason for differences price, the prediction results have worsened because of
in the results shown in Table 3 can be described by using the wind power information. Finally, it can be

Fig. 9 Short-term prediction of

electricity price on May 19

123
Neural Computing and Applications

Fig. 10 Short-term prediction

of electricity price on Nov 12

Table 3 MAPE results for the

RBF ? HNN ? WT ? ICA Months
hybrid prediction neural
networks within 24 h (wind Feb May Aug Nov
power effect)
MAPE with wind effect (without correlation analysis) 1.8886 2.9712 3.9284 2.9623
MAPE without wind effect 1.5806 5.4142 3.1002 4.4125

RBF+HNN+WT+ICA Feb 18 Correlaon= 0.205857

6 275.00
5
270.00
MAPE (%)

Price (DKK / MWh)

4
265.00
3
260.00
2
255.00
1
0 250.00
Feb May Aug Nov 245.00
Month 240.00
0.0 500.0 1000.0 1500.0 2000.0 2500.0
MAPE with wind effect (without correlaon analysis) MAPE without wind effect
Power (MW)

Fig. 11 MAPE results for the hybrid prediction neural networks

Fig. 13 Correlation between electricity price and wind power gener-
within 24 h (wind power effect)
ation on February 18

Feb 18 May 19
280.00 2500.0
400.00 900.0
Price (DKK/MWh)

270.00 2000.0
350.00 800.0
Power (MW)

260.00
Price (DKK / MWh)

1500.0 300.00 700.0

250.00

Power (MW)
1000.0 600.0
250.00
240.00 500.0
230.00 500.0 200.00
400.0
220.00 0.0 150.00
300.0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324 100.00
Hours 200.0
50.00 100.0
Price Wind Power
0.00 0.0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Fig. 12 Electricity price and wind power generation on February 18 Hours
Price Wind Power
concluded that in order to include the impact of wind
Fig. 14 Electricity price and wind power generation on May 19
power on the anticipation of the market-clearing price, the
impact range of the wind power on the market-clearing
negative slope, the positive impact of wind component on
price as well as the correlation between wind power and
the prediction results will be more remarkable. With
electricity price data should first be determined. Then, with
increasing the amount of correlation and the positive slope,
regard to such criteria, the information of wind power
the negative impact of wind component on the prediction
should be included and used to predict the electricity price.
results will be more. Generating wind power over or under
Investigating the correlation between wind power and
a certain range will have a little impact on the market-
electricity price data in time periods close to the prediction
clearing price.
time showed that with increasing the correlation and the

123
 Neural Computing and Applications

May 19 Correlaon= -0.86116 Aug 13 Correlaon= 0.362909

400.00 500.00

Price (DKK / MWh)

350.00 400.00

300.00 300.00
Price (DKK / MWh)

250.00 200.00

200.00 100.00

150.00 0.00
0.0 100.0 200.0 300.0 400.0 500.0 600.0 700.0 800.0 900.0
100.00 Power (MW)
50.00
Fig. 19 Correlation between electricity price and wind power gener-
0.00
0.0 100.0 200.0 300.0 400.0 500.0 600.0 700.0 800.0 900.0 ation on Aug 13
Power (MW)

Fig. 15 Correlation between electricity price and wind power gener- 5 Conclusion
ation on May 19

Electricity market participants have a strong tendency to

Nov 12
500.00 1200.0 know the exact price of electricity in the electricity market.
Price (DKK/MWh)

400.00 1000.0 Hence, many studies have been conducted to predict the

Power (MW)
800.0
300.00
600.0 exact price of electricity as well as reduce the prediction
200.00 400.0 error. In the proposed method, a logical combination of
100.00 200.0 various neural networks is proposed with regard to the
0.00 0.0
1 2 3 4 5 6 7 8 9 101112131415161718192021222324 turbulent signal conditions of electricity price. In this
Hours
research, a proper combination of neural networks, wavelet
Price Wind power
transform, and meta-heuristic algorithm is presented for
Fig. 16 Electricity price and wind power generation on Nov 12 predicting the electricity prices. The method proposed in
this research is a hybrid method that uses wavelet trans-
Correlaon= -0.7328
form to filter distortion and intense noise in electricity price
Nov 12
500.00 signal and the radial basis function (RBF) are used as a
preliminary predictor to find local solutions. The ICA
Price (DKK / MWh)

400.00
meta-heuristic algorithm is also used to optimize the radial
300.00
neural network parameter. Three MLP neural networks
200.00
with different training methods and the ICA meta-heuristic
100.00
algorithm are combined to have a more accurate final
0.00 prediction and create a better model for the nonlinear
0.0 200.0 400.0 600.0 800.0 1000.0 1200.0
Power (MW) behavior of the electricity price signal. The algorithm
inputs include wind power and electricity price. Compar-
Fig. 17 Correlation between electricity price and wind power gener-
ation on Nov 12 ison of the simulation results for the proposed method
along with two other methods proposed in the literature
Aug 13 showed that the appropriate placement of the meta-
500.00 1000.0
heuristic algorithm along with wavelet transform and
Price (DKK / MWh)

400.00 800.0
Power (MW)

neural networks in the structure of the proposed algorithm

300.00 600.0
improves the prediction accuracy. Adding information
200.00 400.0
100.00 200.0
related to wind power generation implies that this approach
0.00 0.0 has positive effects on improving the simulation results in
1 2 3 4 5 6 7 8 9 101112131415161718192021222324
Hours
some cases and has adverse effects in some other cases.
Price Wind Power Therefore, in order to include the impact of wind power on
the anticipation of the market-clearing price, the impact
Fig. 18 Electricity price and wind power generation on Aug 13 range of the wind power on the market-clearing price as
well as the correlation between wind power and electricity
The average processing time required by the proposed price data should be determined first. Then, with regard to
method is less than 2 min using MATLAB software on a such criteria, the information of wind power should be
PC with 4 GB of RAM and a 2.3-GHz-based processor included and used to predict the electricity price. Consid-
[76]. ering the correlation between wind power and electricity
price data in similar hours showed that if the correlation

123
Neural Computing and Applications

coefficient is negative, increasing the correlation enhances PSO Particle swarm optimization
the positive impact of the wind component on the predic- r correlation coefficient
tion results. On the other hand, if the correlation coefficient RBF Radial basis function
is positive, increasing the correlation enhances the negative SAR Seasonal auto-regressive neural network
impact of the wind component on the prediction results. SCM Soft computing model
The results of the simulations and a comparison with SSA Singular spectrum analysis
methods in the field of soft computing show that the pro- WPF Wind power forecasting
posed method is a proper one for more accurate prediction WT Wavelet transform
of electricity prices. The possible future work would be to
expand the proposed method to the special case of price
spike forecasting.
References
1. Osório G, Matias J, Catalão J (2014) Electricity prices forecasting
Compliance with ethical standards by a hybrid evolutionary-adaptive methodology. Energy Convers
Manag 80:363–373
Conflict of interest The authors certify that they have no affiliations 2. Shahidehpour M, Yamin H, Li Z (2002) Market operations in
with or involvement in any organization or entity with any financial electric power systems. IEEE. Wiley-Interscience, Wiley, New
interest or non-financial interest in the subject matter or materials York
discussed in this manuscript. 3. Shafiee S, Zamani-Dehkordi P, Zareipour H, Knight AM (2016)
Economic assessment of a price-maker energy storage facility in
the Alberta electricity market. Energy 111:537–547
4. Nogales FJ, Contreras J, Conejo AJ, Espı́nola R (2002) Fore-
Appendix: Terms and definitions casting next-day electricity prices by time series models. IEEE
Trans Power Syst 17(2):342–348
Here, all terms mentioned in this paper and their definitions 5. Bourbonnais R, Meritet S (2007) Electricity spot price modelling:
are listed in alphabetical order: univariate time series approach. In: The econometrics of energy
systems. Springer, pp 51–74
A Approximation 6. Dukpa A, Duggal I, Venkatesh B, Chang L-Y (2010) Optimal
ANFIS Adaptive network-based fuzzy inference participation and risk mitigation of wind generators in an elec-
tricity market. Renew Power Gener IET 4(2):165–175
system 7. Vahidinasab V, Jadid S (2010) Bayesian neural network model to
AR Auto-regressive predict day-ahead electricity prices. Eur Trans Electr Power
ARIMA Auto-regressive integrated moving average 20(2):231–246
BFGS Broyden–Fletcher–Goldfarb–Shanno 8. Aggarwal SK, Saini LM, Kumar A (2009) Electricity price
forecasting in deregulated markets: a review and evaluation. Int J
BPNN Back propagation neural network Electr Power Energy Syst 31(1):13–22
BR Bayesian regularization 9. Sandhu HS, Fang L, Guan L (2016) Forecasting day-ahead price
CNEA Cascaded neuro-evolutionary algorithms spikes for the Ontario electricity market. Electr Power Syst Res
CWT Continuous wavelet transform 141:450–459
10. Catalao J, Pousinho H, Mendes V (2011) Hybrid wavelet-PSO-
D Detail ANFIS approach for short-term electricity prices forecasting.
db4 Daubechies of order 4 IEEE Trans Power Syst 26(1):137–144
DKK Danish krones 11. Bigdeli N, Afshar K, Gazafroudi AS, Ramandi MY (2013) A
DWT Discrete wavelet transform comparative study of optimal hybrid methods for wind power
prediction in wind farm of Alberta, Canada. Renew Sustain
FNN Fuzzy neural networks Energy Rev 27:20–29
GARCH Generalized auto-regressive conditional 12. Montoya FG, Manzano-Agugliaro F, López-Márquez S, Her-
heteroskedastic nández-Escobedo Q, Gil C (2014) Wind turbine selection for
HPF High pass filter wind farm layout using multi-objective evolutionary algorithms.
Expert Syst Appl 41(15):6585–6595
HNN Hybrid neural network 13. Manzano-Agugliaro F, Alcayde A, Montoya F, Zapata-Sierra A,
HIS Hybrid intelligent system Gil C (2013) Scientific production of renewable energies world-
ICA Imperialist competitive algorithm wide: an overview. Renew Sustain Energy Rev 18:134–143
LPF low pass filter 14. Fan S, Liao JR, Yokoyama R, Chen L, Lee W-J (2009) Fore-
casting the wind generation using a two-stage network based on
LM Levenberg–Marquardt meteorological information. IEEE Trans Energy Convers
MAPE Mean absolute percentage error 24(2):474–482
MCP Market-clearing price 15. Hernandez-Escobedo Q, Manzano-Agugliaro F, Gazquez-Parra
MLP Multilayer perceptron JA, Zapata-Sierra A (2011) Is the wind a periodical phenomenon?
The case of Mexico. Renew Sustain Energy Rev 15(1):721–728
NN Neural network

123
 Neural Computing and Applications

16. Hu P, Karki R, Billinton R (2009) Reliability evaluation of 36. Hong YY, Liu CY, Chen SJ, Huang WC, Yu TH (2015) Short-
generating systems containing wind power and energy storage. term LMP forecasting using an artificial neural network incor-
IET Gener Transm Distrib 3(8):783–791 porating empirical mode decomposition. Int Trans Electr Energy
17. Arjmand R, Rahimiyan M (2016) Impact of spatio-temporal Syst 25(9):1952–1964
correlation of wind production on clearing outcomes of a com- 37. Panapakidis IP, Dagoumas AS (2016) Day-ahead electricity price
petitive pool market. Renew Energy 86:216–227 forecasting via the application of artificial neural network based
18. Exizidis L, Kazempour SJ, Pinson P, de Greve Z, Vallée F (2016) models. Appl Energy 172:132–151
Sharing wind power forecasts in electricity markets: a numerical 38. Jung J, Broadwater RP (2014) Current status and future advances
analysis. Appl Energy 176:65–73 for wind speed and power forecasting. Renew Sustain Energy
19. MacCormack J, Hollis A, Zareipour H, Rosehart W (2010) The Rev 31:762–777
large-scale integration of wind generation: impacts on price, 39. González AM, Roque AMS, Garcı́a-González J (2005) Modeling
reliability and dispatchable conventional suppliers. Energy Policy and forecasting electricity prices with input/output hidden Mar-
38(7):3837–3846 kov models. IEEE Trans Power Syst 20(1):13–24
20. Kabouris J, Kanellos F (2010) Impacts of large-scale wind pen- 40. Rodriguez CP, Anders GJ (2004) Energy price forecasting in the
etration on designing and operation of electric power systems. Ontario competitive power system market. IEEE Trans Power
IEEE Trans Sustain Energy 1(2):107–114 Syst 19(1):366–374
21. Negnevitsky M, Johnson P, Santoso S. Short term wind power 41. Turner AJ, Miller JF (2014) NeuroEvolution: evolving hetero-
forecasting using hybrid intelligent systems. In: 2007 IEEE geneous artificial neural networks. Evol Intell 7(3):135–154
Power Engineering Society General Meeting 42. Koppejan R, Whiteson S (2011) Neuroevolutionary reinforce-
22. Catalão J, Pousinho H, Mendes V (2011) Short-term electricity ment learning for generalized control of simulated helicopters.
prices forecasting in a competitive market by a hybrid intelligent Evol Intell 4(4):219–241
approach. Energy Convers Manag 52(2):1061–1065 43. Soman SS, Zareipour H, Malik O, Mandal P (2010) A review of
23. Abedinia O, Amjady N, Shafie-Khah M, Catalão J (2015) Elec- wind power and wind speed forecasting methods with different
tricity price forecast using combinatorial neural network trained time horizons. In: North American power symposium (NAPS),
by a new stochastic search method. Energy Convers Manag 2010. IEEE, pp 1–8
105:642–654 44. Hibon M, Evgeniou T (2005) To combine or not to combine:
24. Kaur A, Nonnenmacher L, Pedro HT, Coimbra CF (2016) Ben- selecting among forecasts and their combinations. Int J Forecast
efits of solar forecasting for energy imbalance markets. Renew 21(1):15–24
Energy 86:819–830 45. Abedinia O, Amjady N (2015) Day-ahead price forecasting of
25. Jin CH, Pok G, Lee Y, Park H-W, Kim KD, Yun U, Ryu KH electricity markets by a new hybrid forecast method. Model
(2015) A SOM clustering pattern sequence-based next symbol Simul Electr Electron Eng 1(1):1–7
prediction method for day-ahead direct electricity load and price 46. Tripathi M, Upadhyay K, Singh S (2014) Electricity price fore-
forecasting. Energy Convers Manag 90:84–92 casting using generalized regression neural network (GRNN) for
26. Anbazhagan S, Kumarappan N (2014) Day-ahead deregulated PJM electricity market. Int J Manag Theory Appl (IREMAN)
electricity market price forecasting using neural network input 2(4):137–142
featured by DCT. Energy Convers Manag 78:711–719 47. Kakhki IN, Taherian H, Aghaebrahimi MR (2013) Short-term price
27. Voronin S, Partanen J, Kauranne T (2014) A hybrid electricity forecasting under high penetration of wind generation units in smart
price forecasting model for the Nordic electricity spot market. Int grid environment. In: Computer and knowledge engineering
Trans Electr Energy Syst 24(5):736–760 (ICCKE), 2013 3th international econference on. IEEE, pp 158–163
28. Amjady N, Keynia F (2009) Day-ahead price forecasting of 48. Amjady N, Keynia F, Zareipour H (2011) Wind power prediction
electricity markets by mutual information technique and cascaded by a new forecast engine composed of modified hybrid neural
neuro-evolutionary algorithm. IEEE Trans Power Syst network and enhanced particle swarm optimization. IEEE Trans
24(1):306–318 Sustain Energy 2(3):265–276
29. Azizi N, Rezakazemi M, Zarei MM (2017) An intelligent 49. Aghajani A, Kazemzadeh R, Ebrahimi A (2016) A novel hybrid
approach to predict gas compressibility factor using neural net- approach for predicting wind farm power production based on
work model. Neural Comput Appl. https://doi.org/10.1007/ wavelet transform, hybrid neural networks and imperialist com-
s00521-017-2979-7 petitive algorithm. Energy Convers Manag 121:232–240
30. Senjyu T, Toyama H, Areekul P, Chakraborty S, Yona A, Urasaki 50. Yongjun W, Gang Y, Yanying H (2015) Research on Correlation
N, Funabashi T (2009) Next-day peak electricity price forecasting analysis of industry electricity quantity. International Conference
using NN based on rough sets theory. IEEJ Trans Electr Electron on Information Technology and Management Innovation
Eng 4(5):618–624 (ICITMI 2015), pp 906–912
31. Saâdaoui F (2017) A seasonal feedforward neural network to 51. Atashpaz-Gargari E, Lucas C (2007) Imperialist competitive
forecast electricity prices. Neural Comput Appl 28(4):835–847 algorithm: an algorithm for optimization inspired by imperialistic
32. Liu Z, Li W, Sun W (2013) A novel method of short-term load competition. In: Evolutionary computation, 2007. CEC 2007.
forecasting based on multiwavelet transform and multiple neural IEEE Congress on. IEEE, pp 4661–4667
networks. Neural Comput Appl 22(2):271–277 52. Abdechiri M, Faez K, Bahrami H (2010) Neural network learning
33. Amjady N, Hemmati M (2009) Day-ahead price forecasting of based on chaotic imperialist competitive algorithm. In: Intelligent
electricity markets by a hybrid intelligent system. Eur Trans systems and applications (ISA), 2010 2nd international workshop
Electr Power 19(1):89–102 on. IEEE, pp 1–5
34. He Y, Xu Q, Wan J, Yang S (2016) Short-term power load 53. Niknam T, Fard ET, Pourjafarian N, Rousta A (2011) An efficient
probability density forecasting based on quantile regression hybrid algorithm based on modified imperialist competitive
neural network and triangle kernel function. Energy 114:498–512 algorithm and K-means for data clustering. Eng Appl Artif Intell
35. Shayeghi H, Ghasemi A, Moradzadeh M, Nooshyar M (2015) 24(2):306–317
Simultaneous day-ahead forecasting of electricity price and load 54. Jolai F, Sangari MS, Babaie M (2010) Pareto simulated annealing
in smart grids. Energy Convers Manag 95:371–384 and colonial competitive algorithm to solve an offline scheduling

123
Neural Computing and Applications

problem with rejection. Proc Inst Mech Eng Part B J Eng Manuf 65. Pandey AS, Singh D, Sinha SK (2010) Intelligent hybrid wavelet
224(7):1119–1131 models for short-term load forecasting. IEEE Trans Power Syst
55. Lucas C, Nasiri-Gheidari Z, Tootoonchian F (2010) Applica- 25(3):1266–1273
tion of an imperialist competitive algorithm to the design of a 66. Chuanan Y, Yongchang Y (2011) A hybrid model to forecast
linear induction motor. Energy Convers Manag wind speed based on wavelet and neural network. In: Control,
51(7):1407–1411 automation and systems engineering (CASE), 2011 international
56. Mahmoudi MT, Forouzideh N, Lucas C, Taghiyareh F (2009) conference on. IEEE, pp 1–4
Artificial neural network weights optimization based on imperi- 67. Damousis IG, Alexiadis MC, Theocharis JB, Dokopoulos PS
alist competitive algorithm. In: 7th international conference on (2004) A fuzzy model for wind speed prediction and power
computer science and information technologies (CSIT’09), generation in wind parks using spatial correlation. IEEE Trans
Yerevan. pp 244–247 Energy Convers 19(2):352–361
57. Li S, Wang P, Goel L (2015) Short-term load forecasting by 68. Haque AU, Mandal P, Meng J, Srivastava AK, Tseng T-L, Senjyu
wavelet transform and evolutionary extreme learning machine. T (2013) A novel hybrid approach based on wavelet transform
Electr Power Syst Res 122:96–103 and fuzzy ARTMAP networks for predicting wind farm power
58. Conejo AJ, Plazas M, Espinola R, Molina AB (2005) Day-ahead production. IEEE Trans Ind Appl 49(5):2253–2261
electricity price forecasting using the wavelet transform and 69. Shaker H, Zareipour H, Wood D (2013) On error measures in
ARIMA models. IEEE Trans Power Syst 20(2):1035–1042 wind forecasting evaluations. In: Electrical and computer engi-
59. Amjady N, Daraeepour A, Keynia F (2010) Day-ahead electricity neering (CCECE), 2013 26th annual IEEE Canadian conference
price forecasting by modified relief algorithm and hybrid neural on. IEEE, pp 1–6
network. IET Gener Transm Distrib 4(3):432–444 70. Madsen H, Pinson P, Kariniotakis G, Nielsen HA, Nielsen T
60. Amjady N, Keynia F (2009) Short-term load forecasting of power (2005) Standardizing the performance evaluation of shortterm
systems by combination of wavelet transform and neuro-evolu- wind power prediction models. Wind Eng 29(6):475–489
tionary algorithm. Energy 34(1):46–57 71. Nord Pool market. Available online at the following website:
61. Catalao J, Pousinho H, Mendes V (2011) Hybrid wavelet-PSO- http://www.nordpoolspot.com/
ANFIS approach for short-term wind power forecasting in Por- 72. Camastra F, Filippone M (2009) A comparative evaluation of
tugal. IEEE Trans Sustain Energy 2(1):50–59 nonlinear dynamics methods for time series prediction. Neural
62. Mandal P, Zareipour H, Rosehart WD (2014) Forecasting Comput Appl 18(8):1021
aggregated wind power production of multiple wind farms using 73. Camastra F, Colla AM (1999) Neural short-term prediction based
hybrid wavelet-PSO-NNs. Int J Energy Res 38(13):1654–1666 on dynamics reconstruction. Neural Process Lett 9(1):45–52
63. Mallat SG (1989) A theory for multiresolution signal decompo- 74. Amjady N (2006) Day-ahead price forecasting of electricity
sition: the wavelet representation. IEEE Trans Pattern Anal Mach markets by a new fuzzy neural network. IEEE Trans Power Syst
Intell 11(7):674–693 21(2):887–896. https://doi.org/10.1109/TPWRS.2006.873409
64. Lei C, Ran L (2008) Short-term wind speed forecasting model for 75. Taherian H, Nazer-Kakhki I, Aghaebrahimi MR, Farshad M,
wind farm based on wavelet decomposition. In: Electric utility Goldani SR (2014) Short term price forecasting in electricity
deregulation and restructuring and power technologies, 2008. market considering the effect of wind units’ generation. Comput
DRPT 2008. Third international conference on. IEEE, Intell Electr Eng 5(1):105–122
pp 2525–2529 76. Mathworks. URL http://www.mathworks.com

123