Electric Power Systems Research 192 (2021) 106995

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: [Link]/locate/epsr

Short-term electricity load and price forecasting by a new optimal

LSTM-NN based prediction algorithm
Gholamreza Memarzadeh a, Farshid Keynia b, *
a
Department of Power and Control Engineering, Graduate University of Advanced Technology, Kerman, Iran
b
Department of Energy Management and Optimization, Institute of Science and High Technology and Environmental Sciences, Graduate University of Advanced
Technology, Kerman, Iran

A R T I C L E I N F O A B S T R A C T

Keywords: Nowadays, a basic commodity for a human being to lead a standard lifestyle with human comfort irrespective of
Wavelet transform the nature of environmental conditions is electric power. The electricity load demand increases tremendously
Feature selection especially for a metropolitan city due to climatic conditions, population growth, local area development, in­
LSTM
dustries expansion, air pollution, thermal device usage, etc. Hence, the accuracy of electricity load and its price
Neural networks
Load forecast
forecasting is a deciding factor for the power distribution network to retain as an efficient, sustainable, and
secure consumer-friendly network. On the other hand, based on the volatile, intermittent, and uncertain behavior
of electricity load and price, an accurate, and robust forecast model should be designed. In this paper, a new
hybrid forecast model for short-term electricity load and price prediction has been developed. The proposed
method includes three modules: wavelet transform that is used to eliminate fluctuation behaviors of the elec­
tricity load and price time series, feature selection based on entropy and mutual information has been proposed
to rank candidate inputs and eliminate redundant inputs according to their information value, and a new
learning algorithm. The proposed learning method consists of a deep learning algorithm with LSTM networks
which improves the accuracy of predictions. The performance of the proposed method has been validated suc­
cessfully on load and price data collected from the Pennsylvania-New Jersey-Maryland (PJM) and Spain elec­
tricity markets. Also, for further test, the load data in Iran have been used.

examine the medium-term and long-term energy prediction of

large-scale city-wide using the nonlinear autoregressive model (NARM),
1. Introduction
linear model using stepwise regression (LMSR) and random forest (least
square boosting) (LSBoost) approaches [4]. Wei et al. proposed the
Forecasting important parameters of the power system, including
ARIMA model for STLF and improve predictive precision [5]. These
load and price, are one of the most important issues in the optimal
methods can’t forecast the non-linear load and time series accurately,
operation of power systems in the deregulated power market. Accurate
but the can interpret the physical properties of the components. How­
prediction of the load and price help electricity producers to be more
ever, intelligent methods perform better for forecasting non-linear time
successful in the electricity market. Similarly, the network operator
series. For example, Karampelas et al. proposed artificial neural net­
needs to have accurate load and price information for reliable and secure
works (ANN) models to identify future Hellenic energy consumption [6].
network operation. Also, STLF and STPF play an important role in the
Hamed proposed a novel technique for short-term load forecasting based
management and operation of energy resources in the microgrid.
on combinations of Kalman filtering (KF), Wavelet, and Artificial Neural
Load time series are typically complex behavior. One of the com­
Network (WNN and ANN) schemes [7]. Vrablecová et al. online support
plexities of load time series is its non-linear and variable behavior.
vector regression (SVR) method to short term power load forecasting
Electricity Price is even a more complex signal. Due to the importance of
public Irish CER dataset [8]. In particular, ANN has a better capability
STLF and STPF and their time series complexity, much research has been
for load and price forecasting. So they have attracted a lot of attention
done in recent years [1, 2]. Among these researches are static methods
[9-12]. However, among neural networks, deep neural networks are
for STLF and STPF [3-5]. For example, Taylor proposed developed
more efficient in forecasting important power system parameters. For
exponentially weighted methods for load forecasting [3]. Ahmad et al.

* Corresponding author.
E-mail addresses: [Link]@[Link] (G. Memarzadeh), [Link]@[Link] (F. Keynia).

[Link]
Received 24 July 2020; Received in revised form 21 November 2020; Accepted 28 November 2020
Available online 4 December 2020
0378-7796/© 2020 Elsevier B.V. All rights reserved.
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Nomenclature P(x, y) Joint probability distribution

xt The input vector
STLF Short term load forecasting yt The output vector
STPF Short term price forecasting it The output of the input gate
WT Wavelet transform ft The output of the output gate
DWT Discrete wavelet transform ot The output of forget gate
IWT Inverse wavelet transform ct The finishing state in the memory block
LSTM Long Short Term Memory ct The temporary of the memory cell in the memory block
FS Feature selection σ The sigmoid function which determining the propagation
MI mutual information of the information
RNN Recurrent Neural Network Wxf , Wxi , Wxc and Wxo The input weight matrices
MAE Mean Absolute Error Whf , Whi , Whc and Who The recurrent weight matrices
MAPE Mean Absolute Percentage Error Why The hidden output weight matrix
RMSE Root Mean Square Error bf , bi , bc , bo and by The related bias vectors
VAR Variance SACT The actual value at time i
i
IG Interaction gain
SFOR
i Forecasted value at the time i
TH threshold
N The total number of data
P(x) The probability distribution of random variable x

example, Sideratos et al. proposed a hybrid structure based on artificial

neural networks and deep learning architectures for week-ahead load
forecasting [13]. Wang et al. proposed a novel deep learning-based
ensemble approach based on wavelet transform and convolutional
neural network for probabilistic wind power forecasting [14]. Wen et al.
proposed a deep recurrent neural network with long short-term memory
units is developed to forecast aggregated power load and the photo­
voltaic (PV) power output in community microgrid [15]. Suryanarayana
et al. proposed a deep learning method for thermal load forecasting
[16]. Gou et al. utilized a deep feed-forward network for short-term
electricity load forecasting. The results are compared to random forest
and gradient boosting machine models [17]. Chen et al. presented a
model for forecasting short-term electric load based on deep residual
networks that can integrate domain knowledge and researchers’ un­
derstanding [18]. Now, the power of the forecasting method improved
by selecting the most efficient and effective inputs to remove a large
portion of redundant data. Several FS methods were presented for load
and price forecasting [19-24]. For example, Fan et al. proposed the
self-organized map method to cluster the input data set into several
subsets in an unsupervised manner [19]. Rana et al. proposed the
correlation-based feature selection method to select a small set of lag
variables [20]. Mandal et al. proposed numerical sensitivity analysis to
improve the performance of electricity price forecasting [22]. MI-based
FS methods will be more efficient [23]. This method calculates the rel­
evancy of input and output data. A modified version of MI analysis is
presented for FS of STPF for electricity markets [23]. Abedinia et al.

Fig. 1. The basic LSTM cell. Fig. 2. Flowchart of the hybrid forecasting model.

2
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Fig. 3. Framework of the proposed method for STLF and STPF.

3
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Fig. 4. The STLF result in PJM 2006 electricity market.

modeling interaction in addition to relevancy and redundancy, based on wavenet ensemble for short-term load forecasting [28]. Rafiei et al.
information-theoretic criteria, for FS [24]. Some researchers also pro­ proposed a novel hybrid method, including generalized extreme
posed a hybrid forecasting method to increase prediction accuracy learning machine (GELM), wavelet preprocessing, and bootstrapping for
[25-29]. For example, Ribero et al. proposed a framework to build a probabilistic electricity load forecasting [29].

Fig. 5. Forecasted and real electrical load and error of summer season in PJM 2006 electricity market.

4
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Table 1. Based on the entropy concept, one of the most advanced FS techniques is
Weekly MAPE (%) for the STLF of the PJM 2006 electricity market. MI. In the following, the proposed method to select the best inputs is
Method Winter Spring Summer Fall Average described.
The MI between output and input is used to calculate the similarity of
NSA [24] 1.69 1.43 1.74 1.57 1.61
PCA [24] 1.6 1.29 1.66 1.58 1.53 these features. The MI(x, y)between the two variables is defined as
CA [24] 1.45 1.27 1.58 1.40 1.43 following [23]:
MI [24] 1.34 1.24 1.63 1.40 1.40 ( ( ) )
MR [24] 1.28 1.19 1.53 1.44 1.36 ∑n ∑ m
( ) P xi , yj
CA + CA [24] 1.33 1.24 1.49 1.39 1.36 MI(x, y) = P xi , yj log2 ( ) (1)
P(xi )P yj
MI + CA [24] 1.32 1.02 1.36 1.17 1.22
i=1 j=1

MR + CA [24] 1.18 1.11 1.35 1.26 1.23

The set of inputs required to STLF and STPF, which includes their
CA + MI [24] 1.18 0.99 1.33 1.19 1.17
MI + MI [24] 0.97 0.88 0.98 0.98 0.95 historical data, is very large. Beside, these data may include ineffective
MR + MI [24] 1.01 0.77 0.96 0.95 0.92 inputs that complicate the problem of predicting power system param­
MI + IG [24] 0.98 0.74 0.89 0.93 0.88 eters such as electrical load and electricity energy prices and degrade its
MRMRMS [24] 0.95 0.72 0.86 0.90 0.86 performance. Also, with more input features, forecasters usually need
Proposed 0.35 0.26 0.80 0.19 0.40
more historical data, while there is usually a serious constraint on
available historical data for forecasting. Therefore, the candidate inputs
Table 2. should be modified using a feature selection method, so that at least a
VAR (× 10− 5) for the STLF of the PJM 2006 electricity market. subset of the most informative features are selected and other unim­
Method Winter Spring Summer Fall Average
portant candidates are filtered. In STLF and STPF problem, more MI
between input variable (xi ) and target variable (y) means that (xi ) is a
NSA [24] 6.98 8.53 14.40 11.97 1.61
more relevant feature for the forecast (y). To select the candidate input
PCA [24] 7.07 8.05 14.29 12.33 1.53
CA [24] 6.71 8.21 14.06 11.25 1.43 the MI between inputs and target is calculated. Then more relevant
MI [24] 6.63 8.14 13.70 11.07 1.40 feature is selected based on MI and other irrelevant features filters out. If
MR [24] 6.36 8.02 13.14 10.44 1.36 the selected input and target are similar, the MI value is large, and also if
CA + CA [24] 6.58 7.99 13.63 11.2 1.36
the MI becomes zero, the two random variables are unrelated or inde­
MI + CA [24] 6.44 6.49 12.92 9.83 1.22
MR + CA [24] 6.03 7.26 12.9 10.61 1.23
pendent. By applying this method, after calculating the MI value based
CA + MI [24] 5.93 6.21 12.74 9.92 1.17 on Eq. (1), the input features are ranked. More mutual information be­
MI + MI [24] 4.93 5.08 9.56 8.08 0.95 tween two candidate inputs (xi ) and (xj ), results in their more redundant
MR + MI [24] 5.15 3.63 9.4 7.65 0.92 information. When two inputs are very similar, one of these input fea­
MI + IG [24] 4.89 3.05 8.12 7.06 0.88
MRMRMS [24] 4.73 2.72 7.16 6.86 0.86
tures can be selected. In other words, using the concept of redundancy,
Proposed 0.52 0.61 4.08 0.19 1.35 the similarity of two data inputs is investigated. So if the similarity of the
two specified inputs is higher than a threshold value, deleting one of
these inputs will not only slow down the simulation process but also
This paper developed a hybrid forecasting model based on WT, FS speed up the convergence of the algorithm. For this purpose, the pro­
based on entropy and mutual information, and deep learning time series posed redundancy criterion for two candidate inputs xi and,xj is defined
prediction based on LSTM to STLF and STPF. The WT decomposes the as follow:
load and price fluctuation characteristics to some of the sub-series. The ( ) ⃒ ( )⃒
FS module is used to select efficient inputs for STLF and STPF. Finally, Redundancy xi , xj = ⃒IG xi ; xj ; y ⃒ (2)
the outputs of the two previous steps have been used as inputs of the ( ) [( ) ] ( )
LSTM module. The performance of the proposed model is successfully IG xi ; xj ; y = MI xi , xj ; y − MI(xi ; y) − MI xj ; y (3)
validated on load and price data collected from the PJM electricity
The detail of the proposed method for selecting inputs is given in
market, the price data of the Spanish electricity market, the load data of
[30].
Hormozgan province in Iran. Results have shown that the proposed
The FS algorithm used in this paper is implemented as follows:
hybrid model has an excellent performance in accurately predicting the
electrical load and the electricity price. The remaining parts of the paper
1- First, the relevancy between all inputs and outputs is calculated
are organized as follows. In Section 2, the FS method is described. In the
using (2). Candidate features with relevance factor higher than TH1
third and fourth section, the WT and LSTM model is proposed. Simula­
are selected.
tion and discussion of results and conclusions are presented in Sections 5
2- For each selected inputs in the previous step, the redundancy value
and 6.
based on the IG criterion is calculated using the Eq. (3). If redun­
dancy between two inputs is greater than TH2, the less relevant
2. Wavelet transform (WT)
feature is filtered out.

Generally, an electric load and price time series have non-linear and
The remaining feature after implementing of two above steps are
dynamic properties that make them difficult to predict. WT can extract
relevant and non-redundant candidate inputs. This proposed FS method
various aspects of data, such as breakpoints and discontinuities, which
is an accurate criterion for selecting efficient inputs.
other signal analysis methods may not display. The WT maps a time
series signal to a frequency space and shows it at different times and
scales. The DWT is used to decompose the load and price time series. The 4. LSTM model
wavelet filter bank used in this paper is the Daubechies orthogonal
wavelet. It divides the input data into 4 distinct levels (a3, d1, d2, d3). In recent years, deep learning has created a new and exciting trend in
machine learning. The basics of deep learning are rooted in the classic
3. Feature selection technique and algorithm neural network. But the difference with traditional neural networks is
that in deep learning of many hidden neurons and layers - usually more
FS is commonly used in machine learning, while a subset of existing than two - has been used as an architectural advantage and new teaching
features is selected from the data to use the learning algorithm [23]. patterns. Thus, deep learning is a machine learning method that uses
deep neural networks. As we know, a deep neural network is a

5
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Fig. 6. Coefficient of determination for the STLF of the PJM 2006 electricity market.

multilayer neural network that has several layers.

ct = tanh(Wxc .xt + Whc .ht− 1 + bc ) (6)
The original idea of the deep neural network dates back almost 30
years. But the reason it took so long to use it was that the proper learning
ct = ft ∗ ct− 1 + it ∗ ct (7)
rule was not found that this problem was solved by introducing the
backpropagation method. Deep learning has recently been considered ot = σ (Wxo .xt + Who .ht− 1 + bo ) (8)
because of the successful resolution of some of the challenges facing
artificial intelligence. Deep learning has shown remarkable performance ht = ot ∗ tanh(ct ) (9)
in some areas. Deep learning is a branch of machine learning has had
( )
considerable power and performance in areas such as image processing, yt = σ Why .ht + by (10)
text generation, and so on [31].
RNN is a type of deep neural network used for speech recognition, 1
σ (x) = (11)
machine translation, and text generation. But one of its major drawbacks 1 + e− x

is the problem of gradient vanishing. To solve this problem, the LSTM

network is introduced, a combination of short-term and long-term 5. Simulation and discussion of results
memory. The LSTM network is a special kind of RNN. It is stable and
powerful for solving long-term and short-term dependency problems As mentioned, forecasting the electrical load and electricity price are
[32]. One of the important features of the LSTM network is its ability to one of the most important issues in the optimal and economical opera­
maintain cell state. The key parameter of the LSTM network is the tion of power systems in the deregulated power market. On the other
memory cell. The basic LSTM cell is shown in Fig. 2. hand, the methods proposed in this field usually have low accuracy,
According to Fig. 1 in the structure of the LSTM cell, there exist three which can reduce the profit of electricity producers as one of the most
gates units called input gates, output gates, and forget gates in memory important pillars of the electricity market. In this paper, to improve the
blocks. The implementation of updating the state of the cell and calcu­ accuracy and efficiency of STLF and STPF, a hybrid forecasting model
lating the output of LSTMs can be followed below: based on WT, FS based on entropy and mutual information, and deep
( ) learning time series prediction based on LSTMs is proposed. The flow­
ft = σ Wxf .xt + Whf .ht− 1 + bf (4)
chart of the hybrid forecasting model is shown in Fig. 2.
For this purpose, the size of the training matrix is the 1200×400, and
it = σ(Wxi .xt + Whi .ht− 1 + bi ) (5)
the size of the target matrix is 1200×1. Training samples are considered
before the last week of each season. Data up to 50 days ago the forecast

6
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Fig. 7. The STPF result in PJM 2006 electricity market.

Fig. 8. Forecasted and real electricity price and error of summer season in PJM 2006 electricity market.

7
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Table 3. uncertainty of the forecaster model. The structure of the proposed

Weekly MAPE (%) for the STPF of the PJM 2006 electricity market. method for STLF and STPF is given in Fig. 3. In this paper, to evaluate the
Method Winter Spring Summer Fall Average performance of the proposed method for STLF and STPF, the real-world
electricity markets of PJM 2006 [33], Spain 2002 [34], and the electrical
NSA [24] 8.45 8.23 10.09 9.68 9.11
PCA [24] 7.31 8.12 10.28 9.59 8.83 load of Hormozgan province, Iran, in the year 2009 have been used. In
CA [24] 11.1 8.07 9.94 6.82 8.98 this paper, the proposed method is implemented on MATLAB R2018b.
MI [24] 6.46 6.34 6.87 6.93 6.38
MR [24] 6.20 6.41 6.38 6.35 6.33
5.1. Pennsylvania-New Jersey-Maryland (PJM) electricity market
CA + CA [24] 6.11 6.31 6.45 6.29 6.29
MI + CA [24] 5.65 6.10 6.23 5.95 5.89
MR + CA [24] 6.49 5.63 6.16 5.39 5.92 5.1.1. PJM 2006
CA + MI [24] 5.59 5.14 5.76 5.62 5.53 The PJM electricity market is one of the largest electricity markets in
MI + MI [24] 4.94 4.99 4.81 4.75 4.87 the world, located in the United States. In this paper, electricity price
MR + MI [24] 4.83 4.79 4.81 4.88 4.83
and load data from this market in 2006 have been used to show the
MI + IG [24] 4.71 4.47 4.70 4.43 4.58
MRMRMS [24] 4.62 4.33 4.62 4.38 4.49 capability of the proposed model. As mentioned, for the simulation of
NN [35] 6.35 4.74 4.76 7.00 5.71 STLF and STPF, the last 4 weeks of each season of the year are selected as
MLP + LM [36] 9.82 8.87 10.43 9.54 9.67 the test data. Available weekly in the PJM market for the winter, spring,
MLP + BR [36] 13.22 12.92 11.98 12.24 12.59
summer, and fall seasons from February 15 to 21, May 15 to 21, August
WT+ NN + EA [36] 4.44 4.31 4.78 4.75 4.57
Proposed 0.91 0.63 1.29 0.91 0.93
15 to 21, and November 15 to 21, respectively.
The results of the proposed model for STLF are presented in Fig. 4. As
can be seen from the results in Fig. 4, the model presented in this paper is
capable of load forecasting.
Table 4. According to the results of this table, the MAPE value for all seasons
VAR (× 10− 4) for the STPF of the PJM 2006 electricity market. is less than one. The lowest value of this index is for fall and the highest
Method Winter Spring Summer Fall Average of it for summer. However, the values of these indicators proposed for
NSA [24] 64 57 58 65 61
the summer are still within the acceptable range. This claim is confirmed
PCA [24] 62 52 60 61 58.75 by the real and forecasted electrical load diagram. This diagram is
CA [24] 69 53 57 43 55.5 shown in Fig. 5. Although this season has the most forecasting error, the
MI [24] 52 50 46 47 48.75 forecast curve has followed the real data well.
MR [24] 54 47 45 48 48.5
For comparison purposes, results are compared in Tables 1 and 2
CA + CA [24] 55 45 47 41 47
MI + CA [24] 50 39 42 40 42.75 with other load forecasting methods (with the same test weeks and
MR + CA [24] 47 34 37 41 39.75 criteria). The weekly MAPE of the presented method in this paper has
CA + MI [24] 44 35 33 39 37.75 better performance in forecasting electrical load than other methods
MI + MI [24] 34 28 28 32 30.5 proposed for this market.
MR + MI [24] 31 27 25 30 28.25
MI + IG [24] 31 27 24 28 27.5
The VAR value of the method proposed in this paper and the other
MRMRMS [24] 31 26 24 27 27 methods is given in Table 2. The method outlined in this paper is also
NN [35] NR1 NR NR NR NR more successful in this index. In other words, this method has a lower
MLP + LM [36] NR NR NR NR NR forecasting uncertainty than other methods.
MLP + BR [36] NR NR NR NR NR
The coefficient of determination (R2) was also proposed to show the
WT+ NN + EA [36] NR NR NR NR NR
Proposed 0.58 0.34 1.89 0.51 0.83 effectiveness of the proposed method for STLF. Fig. 6 shows the corre­
1
lation between the real and the forecasted electrical load values using
Not reported.
the proposed method. It has been observed that the value of R2 for all
seasons of the PJM 2006 electricity market using the proposed method is
day (results in 50 * 24 = 1200) are considered as training samples. The not less than 0.99.
80% of training samples selected as train set and the 20% that not seen The results of the proposed model for STPF are presented in Fig. 7. As
in train phase is selected as validation set. the last week of each season is can be seen, among the seasons, the best forecasting performance is
used as a test set. The designed LSTM network has two layers, and the related to the spring season, and also the worst forecasting performance
first layer has 200 hidden units, and the second layer has 100 hidden is in the summer season. However, in this season, the forecast error is
units. In this paper, four indices are proposed to validate the perfor­ still acceptable. For example, the weekly MAPE value for this season of
mance of the proposed method. The formula of the MAPE, RMSE, MAE, the year 2006 is 1.2860, that this amount of error is acceptable for STPF.
and VAR are described in Eqs. (12), 13, 14, and 15 respectively. To illustrate the STPF accuracy obtained proposed method, the
⃒
N ⃒ ACT
⃒ forecasted, real, and forecasted error for STPF in the summer season is
1 ∑ Si − SiFOR ⃒
MAPE = × 100% (12) shown in Fig. 8. As seen, although this season has the most forecasting
N i=1 SiACT
error, the forecast curve acceptably followed the real curve.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ The results of the proposed method are compared with other
√
√1 ∑ N
( ACT )2 methods for STPF in Tables 3 and 4. These tables show that both the
RMSE = √ S − SFOR (13) weekly MAPE and VAR values of the proposed method are considerably
N i=1 i i

lower than other forecasting methods, which indicates that the proposed
STPF forecasting method capable than other methods.
1 ∑N ⃒ ⃒
MAE = ⃒SACT − SFOR ⃒ (14) Fig. 9 shows the correlation between the real and the forecasted
N i=1 i i
electrical price values using the proposed method. It has been observed
⃒ ⃒ that the value of R2 for all seasons of the PJM 2006 electricity market
N (⃒ ACT )2
1 ∑ Si − SiFOR ⃒ using the proposed method is not less than 0.99.
VAR = ACT
− MAPE (15)
N i=1 Si
5.1.2. PJM 2018
N is the total number of data which has a value of 168. In Eq. (15), The second case study used to implement the method presented in
the weekly error variance is proposed that is used to measure the this paper is the electricity price of the PJM electricity market in the year

8
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Fig. 9. Coefficient of determination for the STPF of the PJM 2006 electricity market.

5.2. Spanish electricity market

Table 5.
The STPF result in PJM electricity market in the year 2018. The third case study is the Spanish electricity market that is located
Index Winter Spring Summer Fall in Europe. In this paper, electricity price data from this market in 2002
MAPE 4.0456 4.9326 5.8849 2.6430
has been used to show the capability of the proposed model. The fourth
MAE 1.1903 1.4896 1.6127 0.6209 week of February, May, August, and November are selected for winter,
RMSE 1.8809 2.2413 2.2297 0.8375 spring, summer, and fall seasons, respectively, as the test data.
VAR 0.0014 0.0028 0.0031 4.5733e-04 The results of the proposed model for STPF are presented in Fig. 11.
As can be seen from the results in Fig. 11, the model presented in this
paper is capable of electricity price forecasting. However, the electricity
2018. The price of electricity in this case studies have a lot of spikes,
price in this market is variable and fluctuating. That makes the price
which is one of the most challenges to implement the STPF. This has
predictions are difficult and complicated. However, the method pre­
made it difficult to predict the electricity price for this case study. But
sented in this paper is capable of forecasting the electricity price.
the method presented in this paper can predict the electricity price well.
As can be seen, among the seasons, the best forecasting performance
As with the previous case study, the last weeks of all seasons are
is related to the winter season, and also the worst forecasting perfor­
considered test data.
mance is in the fall season. For example, the weekly MAPE value for the
The results of the proposed model for STPF of the PJM electricity
fall season is 3.4467, that this amount of error is acceptable for STPF.
market in the year 2018 are presented in Table 5. As can be seen, among
The forecasted, real, and forecasted error for STPF in this season is
the seasons, the best forecasting performance is related to the fall sea­
shown in Fig. 12. Despite the high volatility of prices in this season, the
son, and also the worst forecasting performance is in the summer season.
forecasted price curve by very little difference followed the real price
For example, the weekly MAPE value for the summer season is 5.8849.
curve.
However, the proposed method is capable of predicting the electricity
Table 6 presents the results of weekly MAPE values for the seasons
price with considerable accuracy.
with the method described in this paper and the other methods pre­
The forecasted, real, and forecasted error for STPF for the summer
sented in other papers. As can be seen from the results in Table 6, the
season of the year 2018 is shown in Fig. 10. It is observed that the
method presented in this paper has a very good ability to predict elec­
predicted electricity price is followed by the real electricity price very
tricity prices than other methods proposed for this market. The
well. Also, the amount of prediction error is small.
remarkable point is that the methods proposed for comparison have a
wide variety. However, the proposed method is more successful in STPF
than other methods.

9
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Fig. 10. Forecasted and real electricity price and error of summer season in PJM 2018 electricity market.

Fig. 11. The STPF result in Spanish 2002 electricity market.

10
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Fig. 12. Forecasted and real electricity price and error of fall season in Spanish 2002 electricity market.

load with considerable accuracy.

Table 6. The forecasted, real, and forecasted error for STLF for all season of
Weekly MAPE (%) for the STPF of the Spanish 2002 electricity market. the year 2009 is shown in Fig. 14. It is observed that the predicted
Method Winter Spring Summer Fall Average electrical load is followed by the real electrical load very well. Also, the
Mixed Model [37] 6.15 4.46 14.90 11.68 9.30 amount of prediction error is small.
MLP [38] 5.23 5.36 11.40 13.65 8.91
FNN [39] 4.62 5.30 9.84 10.32 7.52
HIS [40] 6.06 7.07 7.47 7.30 6.97
5.4. Discussion
NNWT [41] 3.61 4.22 9.50 9.28 6.65
CNEA [24] 4.88 4.65 5.79 5.96 5.32 In this paper, a hybrid method for STLF and STPF is implemented on
WPA [42] 3.37 3.91 6.50 6.51 5.07 the PJM electricity market, the price data of the Spanish electricity
MRMRMS + RBFNN [24] 3.56 3.90 6.24 6.06 4.94
market, and the load data of Hormozgan province in Iran. The proposed
MRMRMS + MLP [24] 3.60 3.94 6.12 5.83 4.87
MRMRMS + WNN [24] 3.36 3.84 5.96 5.74 4.72 method has an excellent performance in accurately predicting the
Proposed 1.33 1.90 2.13 3.45 2.20 electrical load and the electricity price in all case studies. For example in
the PJM electricity market in the year 2006 the proposed method
compared to other methods, has the lowest amount of error. The results
The graph of the correlation between predicted, and real price values in Tables 1 to 4 confirms this claim. For example, the average MAPE in
are used to show the exact fit of the predicted and real data on a line. this case study for STLF and STPF is respectively 0.4 and 0.93. These
Fig. 13 shows the graph for all seasons of the Spanish electricity market. results confirm that the proposed method has a good capability in STLF
These figures illustrate that the algorithm fits the real and forecasted and STPF. And also the results are presented in Fig. 6 and Fig. 9 shows
price well. the correlation between the real and the forecasted electrical load and
electricity price values using the proposed method for all seasons of the
5.3. Hormozgan province electrical load PJM 2006 electricity market is not less than 0.99. These results
emphasize that using the proposed method, the forecasted values follow
The fourth case study used to implement the method presented in the real values well. Also, in the PJM electricity market in the year 2018,
this paper is the electrical load of Hormozgan province in the year 2009. the worst forecasting performance is in the summer season. For example,
Hormozgan is located in southern Iran. Temperature changes in the area the weekly MAPE value for the summer season is 5.8849. However, the
have led to variations in the amount of electricity consumed. This has proposed method is capable of predicting the electricity price with
made it difficult to predict the electrical load for this area. But the considerable accuracy. In the Spanish electricity market in the year
method presented in this paper can predict the electrical load well. As 2002, the results of the proposed method show its capability in STPF. For
with previous study cases, the last weeks of all seasons are considered as example, the average MAPE in this case study for STPF is 2.20. This
test data. result indicated that compared to other methods, it has better perfor­
The results of the proposed model for STLF of Hormozgan province mance in STPF. All results of this case study is presented in Figs. 11 to 13
are presented in Table 7. According to the above table, the worst forecast and Table 6. The other case study is the electrical load of Hormozgan
result is for the fall. This is due to heavy rainfall and critical conditions province in the year 2009. The geographical location of this region
on the fourth and fifth days of the last week of fall. Due to the conditions causes its temperature and climate changes. This point directly affects
mentioned above, the electrical load has been decreased significantly. the amount of electricity consumption. Also, in 2009 there was heavy
However, the proposed method is capable of predicting the electrical rain and critical conditions in the last week of autumn. All of this has

11
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Fig. 13. Coefficient of determination for all season of Spanish 2002 electricity market.

price fluctuation characteristics to some of the sub-series. The FS module

Table 7. is used to select efficient inputs for STLF and STPF. Finally, the outputs
The STLF result in Hormozgan province 2009. of the two previous steps have been used as inputs of the LSTM module.
Index Winter Spring Summer Fall The performance of the proposed method was evaluated using real data
MAPE 1.2510 0.9873 1.3145 7.1456 from load and price data of the PJM electricity market in the year 2006
MAE 1.5197 1.8751 2.4205 4.8619 and 2018, price data of the Spanish electricity market in the year 2002,
RMSE 1.9059 2.4027 2.9275 8.5137 and load data of Hormozgan province in the year 2009. For example in
VAR 7.8941E-05 6.1722 E-05 9.1367E-05 0.0185 the PJM electricity market in the year 2006, the average MAPE in this
case study for STLF and STPF are respectively 0.4 and 0.93. And also the
complicated the STLF in this area. However, the proposed method is correlation between the real and the forecasted electrical load and
capable of predicting the electrical load with considerable accuracy. So electricity price values using the proposed method for all seasons of the
that the lowest amount of MAPE is related to the winter season at 1.9059 PJM 2006 electricity market is not less than 0.99. In the PJM electricity
and the highest amount is related to the fall season at 8.5137. The market in the year 2018, the average MAPE in this case study for STPF is
forecasted, real, and forecasted error for STLF for all season of the year 4.3765. In the Spanish electricity market in the year 2002, the average
2009 is shown in Fig. 14. These result indicated that the proposed MAPE in this case study for STPF is 2.20. The result of STLF in Hor­
method in this case study for STLF has good precision. mozgan province in the year 2009 also confirms that the proposed
method has a good capability in forecasting. The results of this method
6. Conclusion also have been compared with newly published papers in this field.
These results of the proposed method have shown its capability in
STLF and STPF play a significant role in electrical energy markets. forecasting electric load and electricity price.
Electricity market participants can have a more prosperous and reliable
performance in the electricity market based on accurate forecasting of Credit author statement
electricity load and price. In this paper, a new price and electricity
method is proposed, composed of WT, FS technique, and LSTM network Gholamreza Memarzadeh: Investigation, Methodology, Software
as a forecasting part. For this purpose, The WT decomposes the load and Development, Software, Data Acquisition and Curation, Writing-

12
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

Fig. 14. Forecasted and real electrical load and error for all season of the year 2009.

Original draft, Formal analysis. 27 (2 December) (2011) 596–603. [Link]

74659.
Farshid Keynia: Conceptualization, Validation, Writing- Reviewing
[11] M.S. Nazar, A.E. Fard, A. Heidari, M. Shafie-khah, J.P Catalao, Hybrid model using
and Editing, Supervision, Visualization, Funding acquisition, Project three-stage algorithm for simultaneous load and price forecasting, Electr. Power
administration, Resources. Syst. Res. 165 (December) (2018) 214–228. [Link]
9.004.
[12] L. Ekonomou, C.A. Christodoulou, V Mladenov, A short-term load forecasting
Declaration of Competing Interest method using artificial neural networks and wavelet analysis, Int. J. Power Syst. 1
(2016) 64–68.
[13] G. Sideratos, A. Ikonomopoulos, N.D Hatziargyriou, A novel fuzzy-based ensemble
The authors declare that there is no conflict of interest. model for load forecasting using hybrid deep neural networks, Electr. Power Syst.
Res. 178 (January) (2020), 106025. [Link]
[14] H.Z. Wang, G.Q. Li, G.B. Wang, J.C. Peng, H. Jiang, Y.T Liu, Deep learning based
Reference ensemble approach for probabilistic wind power forecasting, Appl. Energy 188
(February) (2017) 56–70. [Link]
[1] R Weron, Modeling and Forecasting Electricity Loads and Prices: a Statistical [15] L. Wen, K. Zhou, S. Yang, X Lu, Optimal load dispatch of community microgrid
Approach, John Wiley & Sons, 2007. Jan 30. with deep learning based solar power and load forecasting, Energy 171 (March)
[2] R Weron, Electricity price forecasting: a review of the state-of-the-art with a look (2019) 1053–1065. [Link]
into the future, Int. J. Forecast. 30 (4 October) (2014) 1030–1081. [Link] [16] G. Suryanarayana, J. Lago, D. Geysen, P. Aleksiejuk, C Johansson, Thermal load
org/10.1016/[Link].2014.08.008. forecasting in district heating networks using deep learning and advanced feature
[3] J.W. Taylor, Short-term load forecasting with exponentially weighted methods, selection methods, Energy 157 (August) (2018) 141–149. [Link]
IEEE Trans. Power Syst. 27 (1 August) (2011) 458–464. [Link] 16/[Link].2018.05.111.
/TPWRS.2011.2161780. [17] Z. Guo, K. Zhou, X. Zhang, S Yang, A deep learning model for short-term power
[4] T. Ahmad, H. Chen, Nonlinear autoregressive and random forest approaches to load and probability density forecasting, Energy 160 (October) (2018) 1186–1200.
forecasting electricity load for utility energy management systems, Sustain. Cities [Link]
Soc. 45 (February) (2019) 460–473. [Link] [18] K. Chen, K. Chen, Q. Wang, Z. He, J. Hu, J He, Short-term load forecasting with
[5] L. Wei, Z. Zhen-gang, Based on time sequence of ARIMA model in the application of deep residual networks, IEEE Trans. Smart Grid (June) (2018). [Link]
short-term electricity load forecasting, in: 2009 International Conference on 0.1109/TSG.2018.2844307.
Research Challenges in Computer Science, IEEE, 2009, pp. 11–14. Dec 28, [19] S. Fan, L. Chen, Short-term load forecasting based on an adaptive hybrid method,
[Link] IEEE Trans. Power Syst. 21 (1 January) (2006) 392–401. [Link]
[6] P. Karampelas, V. Vita, C. Pavlatos, V. Mladenov, L Ekonomou, Design of artificial 9/TPWRS.2005.860944.
neural network models for the prediction of the Hellenic energy consumption, in: [20] M. Rana, I. Koprinska, A Khosravi, Feature selection for interval forecasting of
In10th Symposium on Neural Network Applications in Electrical Engineering, electricity demand time series data, in: In Artificial Neural Networks, Springer,
IEEE, 2010, pp. 41–44. Sep 23, [Link] Cham, 2015, pp. 445–462. [Link]
[7] H.H. Aly, A proposed intelligent short-term load forecasting hybrid models of ANN, [21] S. Kouhi, F. Keynia, S.N Ravadanegh, A new short-term load forecast method based
WNN and KF based on clustering techniques for smart grid, Electr. Power Syst. Res. on neuro-evolutionary algorithm and chaotic feature selection, Int. J. Electr. Power
182 (May) (2020), 106191. [Link] Energy Syst. 62 (November) (2014) 862–867. [Link]
[8] P. Vrablecová, A.B. Ezzeddine, V. Rozinajová, S. Šárik, A.K Sangaiah, Smart grid 14.05.036.
load forecasting using online support vector regression, Comput. Electr. Eng. 65 [22] P. Mandal, A.K. Srivastava, M. Negnevitsky, J.W Park, Sensitivity analysis of neural
(January) (2018) 102–117. [Link] network parameters to improve the performance of electricity price forecasting,
[9] M. López, S. Valero, C. Senabre, J. Aparicio, A Gabaldon, Application of SOM Int. J. Energy Res. 33 (1 January) (2009) 38–51. [Link]
neural networks to short-term load forecasting: the Spanish electricity market case [23] N. Amjady, F. Keynia, Day-ahead price forecasting of electricity markets by mutual
study, Electr. Power Syst. Res. 91 (October) (2012) 18–27. [Link] information technique and cascaded neuro-evolutionary algorithm, IEEE Trans.
16/[Link].2012.04.009. Power Syst. 24 (1 December) (2008) 306–318. [Link]
[10] Y.M. Wi, S.K. Joo, K.B Song, Holiday load forecasting using fuzzy polynomial /TPWRS.2008.2006997.
regression with weather feature selection and adjustment, IEEE Trans. Power Syst.

13
G. Memarzadeh and F. Keynia Electric Power Systems Research 192 (2021) 106995

[24] O. Abedinia, N. Amjady, H Zareipour, A new feature selection technique for load network and Elman neural network, Energy Convers. Manag. 156 (January) (2018)
and price forecast of electrical power systems, IEEE Trans. Power Syst. 32 (1 Apr) 498–514. [Link]
(2016) 62–74. [Link] [33] PJM Web Site. Available at: <[Link]
[25] M.Q. Raza, M. Nadarajah, D.Q. Hung, Z Baharudin, An intelligent hybrid short- [34] OMIE web site. Available at: <[Link]
term load forecasting model for smart power grids, Sustain. Cities Soc. 31 (May) [35] M. Gholipour Khajeh, A. Maleki, M.A. Rosen, M.H Ahmadi, Electricity price
(2017) 264–275. [Link] forecasting using neural networks with an improved iterative training algorithm,
[26] R. Li, P. Jiang, H. Yang, C Li, A novel hybrid forecasting scheme for electricity Int. J. Ambient Energy 39 (2 February) (2018) 147–158. [Link]
demand time series, Sustain. Cities Soc. (January) (2020), 102036. [Link] 01430750.2016.1269674.
org/10.1016/[Link].2020.102036. [36] N. Amjady, F. Keynia, Day ahead price forecasting of electricity markets by a mixed
[27] N. Amjady, F. Keynia, Short-term load forecasting of power systems by data model and hybrid forecast method, Int. J. Electr. Power Energy Syst. 30 (9
combination of wavelet transform and neuro-evolutionary algorithm, Energy 34 (1 November) (2008) 533–546. [Link]
January) (2009) 46–57. [Link] [37] C. García-Martos, J. Rodriguez, M.J Sanchez, Mixed models for short-run
[28] G.T. Ribeiro, V.C. Mariani, L dos Santos Coelho, Enhanced ensemble structures forecasting of electricity prices: application for the Spanish market, IEEE Trans.
using wavelet neural networks applied to short-term load forecasting, Eng. Appl. Power Syst. 22 (2 Apr) (2007) 544–552. [Link]
Artif. Intell. 82 (June) (2019) 272–281. [Link] 894857.
9.03.012. [38] New York Independent System Operator. [Online]. Available: [Link]
[29] M. Rafiei, T. Niknam, J. Aghaei, M. Shafie-Khah, J.P Catalão, Probabilistic load com, accessed 2002.
forecasting using an improved wavelet neural network trained by generalized [39] editorJP Catalão (Ed.), Electric Power Systems: Advanced Forecasting Techniques
extreme learning machine, IEEE Trans. Smart Grid 9 (6 February) (2018) and Optimal Generation Scheduling, CRC Press, 2012. Mar 7.
6961–6971. [Link] [40] N. Amjady, M. Hemmati, Day-ahead price forecasting of electricity markets by a
[30] N. Amjady, F. Keynia, A new prediction strategy for price spike forecasting of day- hybrid intelligent system, Eur. Trans. Electr. Power 19 (1 January) (2009) 89–102.
ahead electricity markets, Appl. Soft Comput. 11 (6 Sep) (2011) 4246–4256. htt [Link]
ps://[Link]/10.1016/[Link].2011.03.024. [41] J.P. Catalão, H.M. Pousinho, V.M Mendes, Neural networks and wavelet transform
[31] G. Memarzadeh, F. Keynia, A new short-term wind speed forecasting method based for short-term electricity prices forecasting, in: 2009 15th International Conference
on fine-tuned LSTM neural network and optimal input sets, Energy Convers. on Intelligent System Applications to Power Systems, IEEE, 2009, pp. 1–5. htt
Manag. 213 (June) (2020), 112824. [Link] ps://[Link]/10.1109/ISAP.2009.5352834.
2824. [42] J.P. Catalão, H.M. Pousinho, V.M Mendes, Hybrid wavelet-PSO-ANFIS approach
[32] H. Liu, X.W. Mi, Y.F Li, Wind speed forecasting method based on deep learning for short-term electricity prices forecasting, IEEE Trans. Power Syst. 26 (1 May)
strategy using empirical wavelet transform, long short term memory neural (2010) 137–144. [Link]

14