Electric Power Systems Research 208 (2022) 107886

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

A wind power forecasting method based on optimized decomposition

prediction and error correction
Jun Li a, Shuqing Zhang a, *, Zhenning Yang b
a
School of Electrical Engineering, Yanshan University, Qinhuangdao 066004, Hebei, China
b
School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China

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

Keywords: To reduce the effect of nonlinearity and volatility in the wind power time sequence, a two-stage short-term wind
Short-term wind power forecasting power forecasting method based on optimized decomposition prediction and error correction is proposed. In the
Variational mode decomposition first stage, in order to improve the decomposition effect of variational mode decomposition (VMD), the
Flower pollination algorithm
decomposition loss is defined as the evaluation criterion to guide the parameter setting of VMD, and flower
Bi-directional long short-term memory neural
network
pollination algorithm (FPA) is utilized to automatically optimize the parameters of VMD. Then the complex wind
Principal component analysis power sequence is decomposed into simple intrinsic mode functions (IMFs). Besides, bi-directional long short-
Error correction term memory (BiLSTM) neural network is built for each IMF to explore the deep time-series features of wind
power in both past and future directions. In the second stage, to reduce the correlation among meteorological
factors, principal component analysis (PCA) is employed to convert the multi-dimensional meteorological factors
into low-dimensional principal components. Then, with the input of IMFs and principal component, an error
correction model based on BiLSTM neural network is established to reduce the inherent error of the model. The
experimental results show that the proposed method has higher prediction accuracy than the traditional methods
in single-step and multi-step ahead forecasting.

1. Introduction forecasting methods can be divided into three main categories: physical
models, statistical models and artificial intelligence models [6]. Physical
The rapid growth of the global economy has caused massive energy models require weather forecast data to build and solve complex phys­
consumption and has resulted in a large number of greenhouse gas ical models to achieve wind power forecasting [7]. The advantage of
emissions, which makes global warming and other meteorological physical models is that they do not require the input of historical data,
problems more serious. Therefore, it is imperative to promote the but the difficulty lies in the acquisition of physical data. Statistical
structural reform of energy. Renewable energy sources such as wind, methods usually use historical data to build statistical models, and
solar and tidal energy have the advantages of clean, pollution-free, and classical algorithms include autoregressive moving average (ARMA) [8]
rich reserves, and are currently the key new energy sources for most and autoregressive integrated moving average (ARIMA) [9]. These
countries. Wind energy is easier to install on a large scale and is far less models can only deal with linear data, but cannot handle nonlinear wind
expensive than solar and tidal energy. It is one of the fastest growing power data with frequent changes [10]. Since the 21st century, the
renewable energy sources [1–2]. Precise short-term wind power fore­ performance of computers has been greatly improved, and artificial
casting provides a reliable scientific basis for grid integration and grid intelligence technology has been greatly developed. And many powerful
dispatching, which can improve the utilization of wind energy, reduce artificial intelligence models have been derived, such as support vector
the impact of wind power brought to the smart grid, and improve the machine (SVM) [11], backpropagation (BP) neural network [12],
operational efficiency of new energy power systems [3–5]. extreme learning machine (ELM) [13], deep belief network (DBN) [14],
Wind energy is highly intermittent and volatile, and this uncertainty and so on. Artificial intelligence (AI) models are rich in variety and high
brings great challenge to the management and operation of power sys­ in prediction accuracy. And most importantly, AI models can handle
tems. Wind power is affected by meteorological factors such as wind nonlinear relationships in the data, so it has become the main tool for
speed and temperature, so it is difficult to predict. Existing wind power short-term wind power forecasting [15].

* Corresponding author: Shuqing Zhang

E-mail address: [email protected] (S. Zhang).

https://doi.org/10.1016/j.epsr.2022.107886
Received 15 November 2021; Received in revised form 19 February 2022; Accepted 19 February 2022
Available online 25 February 2022
0378-7796/© 2022 Elsevier B.V. All rights reserved.
J. Li et al. Electric Power Systems Research 208 (2022) 107886

To further improve the prediction performance, researchers have bi-directional long short-term memory (BiLSTM) neural network is
started to propose hybrid models based on artificial intelligence algo­ proposed, and the contributions and innovations of this paper are as
rithms. For example, in the literature [16], a hybrid wind power pre­ follows:
diction model with the imperial competition algorithm optimized BP
neural network was proposed; in the literature [17], to better distinguish (1) An evaluation criterion to guide the parameter setting of VMD
the long-term and short-term trends of wind power, empirical mode applicable to the field of wind power forecasting is proposed. And
decomposition was used to decompose the wind power series into simple FPA is used to automatically find the optimal decomposition
sub-sequence, and prediction models were built separately for the parameters of VMD. FPA-VMD reduces the signal loss in the
sub-sequence; in the literature [18], principal component analysis (PCA) decomposition process and improves the decomposition effect of
is used to decrease the dimensionality of meteorological factors. Then VMD.
the integrated meteorological factors after dimension reduction are (2) The BiLSTM neural network which has the advantages of bi-
taken as the input of the long short-term memory (LSTM) neural directional processing and long-term memory is applied to the
network to achieve ultra-short-term wind power prediction. From the field of wind power forecasting, and can better mine the deep
above research, it can be seen that the hybrid models are capable to temporal features within the data.
increase the accuracy of the model from different aspects and have (3) With the input of IMFs and principal component, a novel error
higher forecast accuracy than the ordinary model. correction model is proposed to reduce the inherent error by
In the hybrid forecasting model, the combined forecasting method using the implicit information contained in the error.
based on signal decomposition uses signal decomposition algorithm to (4) Principal component analysis (PCA) was used to convert multi­
decompose the complex power wind power sequence into simple and dimensional meteorological factors into low-dimensional prin­
regular sub-series, followed by building the corresponding forecasting cipal components, which reduced the correlation among
model for each sub-series separately, so that the effects of volatility and meteorological factors and the computational burden of the
nonlinearity in the wind power sequence can be well reduced [19–20]. model.
Classical signal decomposition algorithms, including wavelet decom­ (5) Experimental analysis proves the effectiveness of this proposed
position (WD), empirical mode decomposition (EMD), and Variational method in short-term wind power forecasting. Finally, multi-step
mode decomposition (VMD), are widely used in the field of wind power ahead forecasting is employed, and the results show that the
prediction. The decomposition effect of WD is related to the mother method can better understand the future trend of wind power
wavelet and the number of sub-series, so WD lacks adaptivity [21]. The series and ensure the stable operation of the power system.
EMD can decompose the signal adaptively, but there is a problem of
mode mixing [22]. In recent years, VMD has been widely used in the The rest of this paper is designed as follows Section 2. illustrates the
field of short-term wind power prediction. VMD decomposes the original flower pollination algorithm optimized variational mode decomposition
signal into a given number of intrinsic mode functions (IMFs) by con­ (FPA-VMD), BiLSTM neural network and a novel error correction model.
structing and solving a variational model, which alleviates the mode The two-stage forecasting model is given in Section 3 Section 4. dis­
mixing [23]. However, VMD lacks evaluation criteria to guide the cusses the experimental results of wind power forecasting and gives the
setting of parameters. We often give the parameters of VMD based on corresponding analysis. Finally, Section 5 introduces the conclusions
experience, which hinders the development of VMD [24–25]. and future research directions.
Recently, deep learning (DL) has been widely developed [26]. LSTM
neural network is the most popular deep neural network in the field of 2. Methods
temporal prediction, which inherits the hidden layer structure of
sequential connection of recurrent neural network (RNN) and replaces 2.1. The FPA-VMD model
the simple implicit layer neurons with long short-term memory units,
thus solving the long-term dependency problem [27–28]. However, the 2.1.1. Variational mode decomposition
one-way LSTM neural network can only use the information of historical VMD is a fully non-recursive signal decomposition method proposed
moments and does not consider the information of future moments, thus by Dragomiretskiy and Zosso in 2014 [36], which can decompose the
it cannot learn the relationship between the two temporal directions original signal into intrinsic mode function (IMF) with finite bandwidth.
before and after the wind power sequence [29]. The steps of the method are as follows:
The prediction error of the model is gradually increasing with the
increase of nonlinearity and volatility in the wind power series. If the (1) Constructing the variational problem. The variational problem is
prediction errors are large enough, they can lead to significant economic to minimize the sum of the estimated bandwidths of the IMFs,
losses. Several studies have shown that considering the error factor can with the constraint that the sum of the IMFs is the original signal,
effectively improve the prediction performance [30–31]. In the litera­ and the equation is as follows:
ture [32], an error prediction model based on ELM optimized by the { [( ) ] 2
}
∑ K
multi-objective grey wolf algorithm has been developed; the literature min ‖ ∂t δ(t) +
j
∗ uk (t) e− jωk t
‖
[33] proposed an error prediction model based on the Markov process, {uk },{ωk }
k=1
πt 2

and the experiments showed that the method can effectively improve (1)
∑K
the accuracy of wind power forecasting. Different from the above error s.t. uk (t) = f (t)
correction models, the method in this paper also considers multiple k=1

meteorological factors. Meteorological factors such as wind speed, wind

direction, and humidity are key factors that directly affect the operation
of wind turbines, and their inclusion in the model can advance the
forecast accuracy of the model. In addition, the strong correlation where:uk is the IMF; ωk is the center frequency of uk ; Kis the number of
among meteorological factors may increase the complexity and training IMF; δ(t) is the Dirac function; f(t)is the original signal.
time of the model, and reduce the prediction accuracy of the model. It is
necessary to represent the high-dimensional meteorological factors with (2) Transforming the variational problem. The constrained varia­
low-dimensional principal components [34–35]. tional problem in Eq. (1) is transformed into an unconstrained
To solve the above problems, a novel two-stage wind power fore­ problem using the quadratic penalty factor αand Lagrange
casting method based on VMD, flower pollination algorithm (FPA), and

2
J. Li et al. Electric Power Systems Research 208 (2022) 107886

multiplier λ(t), which α guarantees the reconstruction accuracy of 2.1.3. Flower pollination algorithm optimized variational mode
the signal, and λ(t) maintains the stringency of the constraints. decomposition (FPA-VMD)
In the decomposition parameters of VMD, the K determines the
∑ [( j
) ] 2
L({uk }, {ωk }, λ) = α ‖ ∂t δ(t) + ∗ uk (t) e− jωk t ‖ number of IMF, and the appropriate K can completely separate the
k
π t 2 modes of the original signal and avoid the occurrence of mode mixing.
〈 〉 (2) The α affect the decomposition process of VMD to ensure the recon­
∑ ∑
struction accuracy of signals. Therefore, the optimization dimension is
2
+‖ f (t) − uk (t) ‖ + λ(t), f (t) − uk (t)
k 2 k set to 2, that is, Kandα, and the value range is as shown in reference [40].
The parameter selection problem of VMD is expressed as a constrained
optimization problem, and the formula is as follows:
(3) Solving the variational problem. The unconstrained problem in
∑T {
Eq. (2) is solved by the alternate direction method of multipliers |f (t) − f ∗ (t)| K ε[{2, 3, ...34}
min t=1 s.t. ] (8)
(ADMM), uk and ωk are iteratively calculated by the following K,α T αε 100, 105
formulas.
The flower pollination algorithm (FPA) is then used to solve the above
̂f (ω) − ∑̂ ̂
u i (ω) + λ(2ω) constrained optimization problem. The FPA simulates the flower polli­
u n+1
i∕
=k
(3) nation process in nature and updates the population individuals with
̂ k (ω) =
1 + 2α(ω − ωk )2 self-pollination and cross-pollination. FPA is a popular swarm intelli­
gence optimization algorithm with the advantages of fewer parameters,
∫∞ ⃒ ⃒2
ω⃒̂u n+1 (ω)⃒ dω faster and less likely to fall into locally optimal solutions [41–42]. The
ωn+1 = ∫0 ∞ k (4)
k 2 steps of FPA-VMD are as follows:
0
u n+1
|̂ k (ω)| d ω

(1) The parameters of the algorithm are given, which include the
maximum number of iterationsMaxite, the number of individuals
where: ̂f (ω), ̂
λ(ω) and ̂u k (ω) are respectively the Fourier transforms of in the populationn, the transition probabilityPc , and the optimi­
f(t), λ(t) and u(t), and n represents the number of iterations. zation dimensionD. It is worth noting thatD is 2, representing K
and α.
2.1.2. The new evaluation criterion to guide the parameter setting of VMD (2) The initial population P = {Xi1 } = {[x1i,1 , x1i,2 ]}(i = 1, 2, ...n) with
In practical applications, the optimal decomposition parameters of uniform distribution is randomly generated, and the fitness of per
the VMD are generally obtained by spectral analysis [37]. However, due individual in the population is calculated according to Eq. (7).
to the volatility and uncertainty of wind energy, there are many spectral Then the global optimal individual Xbest and the optimal fitness
components of wind power sequences, so the results of spectral analysis Fbest under the current population are recorded. It is worth noting
are poor. It is necessary to define an evaluation criterion to guide the that xi,1 stores the number of K and xi,2 stores the number of α.
parameter selection of VMD applicable to the field of wind power (3) A random number ε with uniform distribution is generated
forecasting [38]. We regard the VMD algorithm as a complex nonlinear randomly. If ε is less than the Pc , new individuals can be gener­
function fvmd (), and the decomposition and reconstruction process can ated through cross-pollination. Otherwise, new individuals are
be expressed as: produced by self-pollination.

{uk (t)|k = 1, 2, ..., K; t = 1, 2, ..., T}

(5) Cross-pollination is a global pollination process by Levy flight, and
= fvmd (f (t), K, α)
the formula is as follows:
( )
∑
K
f ∗ (t) = uk (t) (6) Xit+1 = Xit + γL Xjt − Xbest (9)
k=1
Self-pollination is a local pollination process with different flowers of
Where: uk (t) is the Kth IMF, T is the signal length; f ∗ (t) is the recon­ the current population, and the formula is as follows:
structed signal; f(t), K and α are the inputs of the function, which ( )
respectively represent the original wind power sequence, the number of Xit+1 = Xit + ε Xjt − Xit (10)
IMF and the quadratic penalty factor.
Ideally, the reconstructed signal f ∗ (t) should be the same as the where: Xit is the flower individual i at time t; Xit+1 is the updated flower
original signal f(t). However, there is a certain decomposition loss in the individual i at time t + 1; Xjt is another flower individual j of the same
process of decomposition. Decomposition loss comes from the residual population at time t; Xbest is the global optimal individual; γ is the scaling
signal which does not meet the definition of the IMF. The residual signal factor that controls the step size; L is levy flight.
is often not considered in the model because of its small amplitude and
rapid fluctuation [39]. However, if the decomposition loss is large, more (4) Since the K must be a positive integer, the xti,1 in the new indi­
information will be ignored. It is not conducive to the improvement of vidual is rounded off, and the approximate whole value is taken
prediction accuracy. Therefore, we take the decomposition loss as the as the new population individual. Since the α is a continuous
evaluation standard to guide the parameter setting of VMD, and the
variable so xti,2 can be directly taken as the new population
formula is:
individual.
∑T
|f (t) − f ∗ (t)| (5) Calculate the fitness of the new population and update the global
loss = t=1 (7) optimal individualXbest and the optimal fitnessFbest .
T
(6) Judge whether Maxite is reached. If not, repeat steps (3) to (5).
The loss is the average absolute error between the reconstructed
Otherwise, Xbest andFbest are output. Xbest and Fbest are respectively
signal and the original signal. The smaller the value, the smaller the loss
the optimal decomposition parameters and the minimum
in the decomposition process, which leads to more information IMFs
decomposition loss found by the FPA-VMD model.
contain, and consequently, a more accurate model.

3
J. Li et al. Electric Power Systems Research 208 (2022) 107886

Fig. 1. Structure and connection of long short-term memory units.

The pseudo-code of the FPA-VMD model is as follows:

Begin
Step 1: FPA and VMD parameter initialization:Maxite, n, Pc , D
Set Eq. (7) as the fitness function
FPA population initialization X1i (i = 1, 2, ..., n)
Step 2: Calculate fitness of each agent and retain the optimal individual Xbest and fitness
Fbest
Step 3: while (t ≤ Maxite)
for each agent
if random numberε< Pc
Update position by Eq. (9)
else
Update position by Eq. (10)
end if
end for
Round off discrete variables xti,1
Create a new population X∗
Calculate fitness of X∗
UpdateXbest and Fbest if there is a better one
t = t+ 1 Fig. 2. Structure of BiLSTM neural network model.
End while
Step 4: ReturnXbest andFbest for best result
Obtain optimal decomposition parameters form Xbest and minimum decomposition The calculation formulas are as follows:
loss form Fbest
End Lt = σ (Dl ∗ [ht− 1 , xt ] + El ) (11)

Mt = σ (Dm ∗ [ht− 1 , xt ] + Em ) (12)

2.2. Bi-directional long short-term memory ( )
(13)
′
Qt = tanh Dq ∗ [ht− 1 , xt ] + Eq
The hidden layers of recurrent neural network (RNN) are connected
(14)
′
in sequence, so the influence of previous input can be retained and Qt = Lt ∗ Qt− 1 + Mt ∗ Qt
transmitted to the current moment in the training process [43]. This
unique structure of sequential connection of hidden layers makes RNN Nt = σ(Dn ∗ [ht− 1 , xt ] + En ) (15)
have natural advantages in dealing with time series data, but the
ht = Nt ∗ tanh(Qt ) (16)
following problems need to be solved in practical application [44].
Where Lt , Mt and Nt are respectively the outputs of the forget gate, input
(1) Due to the problems of gradient exploding and gradient vanish­
gate, and output gate at time t; Qt is cell state at time t; Qt is the inter­
′

ing, RNN cannot effectively deal with the problem of long-term

mediate cell state at time t; xt and ht are the input and output of the
dependence, and the ability to learn temporal correlations be­
hidden layer at time t; ht− 1 is the output of the hidden layer at time t − 1;
tween sequences is limited.
Dl , Dm , Dn , and Dq are the weights of forget gate, input gate, output gate
(2) In the process of training, RNN can only use the information of
the previous moment, and does not consider the information of and intermediate cell state; El , Em , En and Ec are the bias of forget gate,
the future moment, so it is unable to obtain the correlation be­ input gate, output gate, and intermediate cell state; σ is sigmoid func­
tween previous and future of the wind power series. tion; tanh is tanh function.
To solve problem (2), the bi-directional concept is introduced to
In order to solve the problem of long-term dependence of the prob­ obtain the information in two time directions, thus forming the BiLSTM
lem (1), the ordinary hidden layer neurons in RNN are replaced with neural network with the structure shown in Fig. 2. The BiLSTM neural
LSTM units, thus forming LSTM neural network Fig. 1. shows the network has two LSTM network layers with the same output but
structure of LSTM units and the connection mode from t − 1 time to t+1 different directions, so the model can be trained using the information of
time. The LSTM unit includes three parts: the forget gate, the input gate, both previous and future directions at the same time [46]. Compared
and the output gate. The forget gate can discard irrelevant information. with the LSTM neural network, the BiLSTM neural network has
The input gate determines the new information stored in the cell state. enhanced robustness to biased inputs and model uncertainty in addition
And the output gate controls the output of the hidden layer node [45]. to shorter training time, and has stronger nonlinear representation

4
J. Li et al. Electric Power Systems Research 208 (2022) 107886

Fig. 3. Two-stage short-term wind power forecasting model with error correction.

capability to better handle uncertainty in the data [47]. 2.3.1. The selection of input factor for error correction model
Data fluctuations of wind power are strongly influenced by meteo­
2.3. Error correction model rological factors such as wind speed, and this information is difficult to
be mined from wind power series. Therefore, the input of the error
The FPA-VMD-BiLSTM (FVBL) model uses only historical wind correction model not only includes the IMFs, but also adds multi-
power itself to predict future wind power, which greatly saves compu­ dimensional meteorological factors.
tational resources but limits the sensitivity of the model to fluctuations. The following will select the main meteorological factors related to
In addition, the use of the quadratic penalty factor in the VMD also data fluctuations of wind power by analyzing the physical model of the
weakens the sensitivity of the model to fluctuations. In summary, the wind turbine. Wind turbines capture wind energy and convert it into
prediction error of the FVBL model mainly comes from the random electricity. The output power is related to the kinetic energy of the air
fluctuations of wind power. passing through the effective area of the blade [51], and the calculation
The random fluctuations of wind power are largely influenced by formula is as follows:
meteorological factors, such as wind speed and wind direction. These
1
factors are difficult to be separated from the wind power series [48][38]. Pm = Cp ρAV 3 (17)
2
It has been shown that introducing more factors in the model can make
the model more sensitive to fluctuations and thus improve the prediction Where Pm is the output power of the wind turbine; Cp is the turbine
accuracy [49], so more meteorological factors are introduced in the power coefficient, also known as the utilization coefficient of wind en­
error correction model in this paper. The selected meteorological factors ergy, which is often related to temperature and humidity. According to
will be given in section 2.3.1. Betz’ Law, in an ideal state, the limit ratio of wind energy to kinetic
In addition, it is necessary to represent highly correlated multi- energy is 16/27, about 59.3%; ρ is the air density, which shows a pos­
dimensional meteorological factors by uncorrelated low-dimensional itive relationship with air pressure; A is the area swept by the turbine
principal components because of the strong correlation among meteo­ blades, which is related to the wind direction; V is the wind speed.
rological factors. For example, the relationship between temperature In summary, the error correction model needs to comprehensively
and humidity shows an inverse ratio in summer. In this paper, principal consider wind speed, wind direction, temperature, air pressure, and
component analysis (PCA) was used to reduce the dimension of mete­ humidity.
orological factors. PCA is a dimension reduction method based on
feature fusion that converts the original variables into linearly uncor­ 2.3.2. Multi-meteorological factor fusion based on PCA
related principal components by orthogonal transformation. The con­ PCA can sequentially get the maximum difference between samples,
version coefficients of the two are the eigenvectors of the covariance so as to ensure the minimum information loss. And the principal com­
matrix of the original variables [50], the principles of which are ponents after dimensionality reduction are uncorrelated and can be
described in Section 2.3.2. discussed alone without reference to others [52]. The specific principle
In the error correction model, a BiLSTM neural network is used to is as follows:
establish the mapping between the input factors (including the IMFs and Let the sample size be N, and each sample has Pdimensional features
meteorological principal component) and the prediction error, so as to to obtain the sample matrix AN×P .
tap the bidirectional implicit time-series relationship between the input
data and the prediction error. (1) To eliminate the effect of dimensional, each column of the sample
matrix is standardized by the following equation.

5
J. Li et al. Electric Power Systems Research 208 (2022) 107886

ai − a
a∗i = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (18)
∑N
1
N
(ai − a)2
i=1

Where, ai and a∗i represent the sequence before and after standardization
respectively, a is the mean of the column, and N is the number of sam­
ples in each column.
∑ Fig. 4. The neural network structure of the first stage in single step
(2) The covariance matrix of A is calculated as follows: ahead prediction.
⎡ ⎤
∑ σ (a1 , a1 ) ... σ (aP , a1 )
=⎣ ... ... ... ⎦ (19) sequence is removed from the wind power prediction sequence, which
σ (a1 , aP ) ... σ (aP , aP ) results in the final wind power prediction sequence after error correc­
tion.
In addition, trial and error method [12] is introduced in the first
stage to determine the dimensionality of the input data so as to ensure
Where, σ (a1 , a1 ) is the variance between a1 and a1 . the reasonableness of the experiment. The forecast accuracy of the
model is superior when the input dimension is 3. The neural network
(3) The eigenvalues λi (i = 1, 2, ⋯, p) and eigenvectors structure of the first stage in single-step prediction is shown in Fig. 4,
∑
vi (i = 1, 2, ⋯, P) of are solved by singular value decomposition which f represents the intrinsic mode function. Furthermore, multi-step
(SVD). ahead forecasting is employed in the paper. Unlike the single-step ahead
(4) The principal components are calculated according to Eq. (20), prediction of f1 , f2 and f3 predicting f4 in Fig. 4, the step length of the
and the variance contribution of each principal component is multi-step ahead prediction increases, but the inputs do not change. For
calculated according to Eq. (21). In addition, the principal com­ example, in the 3-step prediction, the inputs are still f1 , f2 and f3 , but the
ponents are ranked according to the magnitude of the outputs are changed to f4 , f5 and f6 .
eigenvaluesλi . It is worth noting that only the first principal
component is considered in the subsequent experiments because 4. Experiment and analysis
of the large percentage of variance contribution of the first
principal component. 4.1. Data set
Yi = Avi (i = 1, 2...P) (20)
The experimental data were obtained from the No. 1 wind farm in
λi Mahuangshan, Ningxia. And the time was selected as January, March,
Ci = P (i = 1, 2, ..., P) (21)
∑ July, and October in 2017, representing winter, spring, summer, and
λk
k=1
autumn in Ningxia, respectively. The fluctuation trend of wind power in
different months is shown in Fig. 5, and the data characteristics are
shown in Table 1. It is worth noting that the sampling interval of this
data set is 15 minutes, with a total of 2976 sampling points for each
3. Two-stage short-term wind power forecasting with error month. We use the data of the last week of each month (672 points) as
correction the test set and the data of the rest of the time as the training set, and
analyze the effect of different models by the results of the test set. In this
In order to improve the prediction accuracy, a two-stage wind power study, all experiments were implemented in MATLAB 2020a. The
forecasting model with error correction is proposed. The primary workstation used was configured with an Intel Core i5-4590 3.30 GHz
structure of the developed forecasting model is illustrated in Fig. 3. CPU and 16 GB RAM.
In the first stage, several component forecasting models based on Through the analysis in Section 2.3.1, the data set contains not only
FPA-VMD-BiLSTM (FVBL) neural networks are constructed. The process wind power, but also wind speed, wind direction, temperature, baro­
is as follows: firstly, the signal loss in the decomposition process is metric pressure, and humidity are taken into account. Therefore, the
defined as Lossand used as the evaluation criterion to guide the input data of the neural network need to be normalized. To eliminate the
parameter setting of VMD; secondly, the FPA is used to find the optimal effect of magnitude, the data set is normalized to the interval of [-1,1]
decomposition parameters containing K and α, so as to decompose the and the test set is anti-normalized with reference to the maximum and
wind power sequence into simple intrinsic mode functions IMF1 ,IMF2 ,..., minimum values of the training set. The formulas for normalization and
IMFK ; then the BiLSTM neural network is built separately for each IMFs anti-normalization are:
to obtain the forecasting values IMF1∗ , IMF2∗ , ..., IMFK∗ ; finally, the wind g − gmin
power prediction sequence is obtained by linearly summing the pre­ g∗ = (kmax − kmin )⋅ + kmin (22)
gmax − gmin
diction values of IMFs.
In the second stage, a new error correction model with multi- g ∗ − kmin
̂
g = (gmax − gmin )⋅
̂ + gmin (23)
dimensional meteorological factors is proposed. Considering the corre­ kmax − kmin
lation between multi-dimensional meteorological factors, PCA was used
to convert multidimensional meteorological factors into low- where g denotes the data before normalization; g∗ denotes the data after
dimensional principal components. Then considering that the error normalization; ̂ g denotes the normalized prediction data; ̂g denotes the
∗

mainly comes from the random fluctuations of wind power, the low- prediction data after anti-normalization; gmax denotes the maximum
dimensional principal component and IMF1 , IMF2 , ..., IMFK are selected value in the training set; gmin denotes the minimum value in the training
as the input of the error correction model. Then the BiLSTM neural set; kmax is denoted as the upper limit of the normalized target interval,
network is used to establish the nonlinear relationship between the two here it is 1; kmin is denoted as the lower limit of the normalized target
to obtain the prediction error sequence. Finally, the prediction error interval, here it is -1.

6
J. Li et al. Electric Power Systems Research 208 (2022) 107886

Fig. 5. Wind power profile.

4.2. Evaluation Criteria

Table 1
Wind power generation statistics for four months in different seasons. The evaluation criteria are selected as mean absolute error (MAE),
January March July October mean absolute percentage error (MAPE) and root mean square error
Mean (GW) 26.3753 23.9186 25.8747 26.0305 (RMSE) [53]. MAE is a basic measure, which is the average of absolute
Minimum (GW) 0 0 0 0 errors, and its unit is MW. In addition, since the wind power series
Maximum (GW) 97.8050 96.7517 94.2558 91.3662 contains many zero values, the calculation method of MAPE is modified,
Standard Deviation (GW) 25.5470 24.8187 22.0236 25.5282 and the common MAPE formula is shown in (25), and the modified
Total Number 2976 2976 2976 2976
MAPE formula is shown in (26). The modified MAPE is the average
absolute error divided by the average wind power, which can avoid the
influence of zero values in the wind power series. In addition, consid­
ering the large influence of outliers on the accuracy of the model, RMSE
is introduced, and the calculation formulas are shown in (27).
m ⃒ ⃒
1 ∑ ⃒∧ ⃒
MAE = ⃒yi − yi ⃒
⃒ ⃒ (24)
m i=1

1 ∑m
y i − yi |
|̂
MAPE = × 100% (25)
m i=1 yi

∑
m
y i − yi |
|̂
MAPE = i=1 ∑
m × 100% (26)
yi
i=1

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑
√m
√ (̂ y − yi )2
√i=1 i
RMSE = (27)
m

Fig. 6. MAE of six single-input basic models.

7
J. Li et al. Electric Power Systems Research 208 (2022) 107886

neural network and ELM in January is 14.1000% and 13.4500%,

respectively, and the latter has improved 4.6099% compared with the
former. In addition, the MAE and MAPE of ELM in January are smaller
than those of the BP neural network, but the RMSE is slightly larger than
that of BP neural network, which indicates that the overall error of ELM
is smaller and the single point error is larger, thus proving the effec­
tiveness of multiple evaluation criteria. Deep neural networks such as
GRU, LSTM neural networks, and BiLSTM neural networks have a

Table 3
Parameters of the VMD in the four cases generated by FPA.
Parameters loss
K α
Fig. 7. MAPE of six single-input basic models.
January 29 100 0.0034
March 34 100 0.0018
July 33 100 0.0019
October 31 100 0.0016

Fig. 8. RMSE of six single-input basic models.

where: ̂
y i is the predicted value of the model, yi is the true value, and m is
the number of predicted points.

4.3. Experiment on the single-input prediction model

The prediction model that only takes historical wind power as input
Fig. 9. MAE of selected the parameter of VMD in different ways.
is known as the single-input model. By comparing the prediction accu­
racy of different single-input base models, the non-linear fitting ability
of each model can be analyzed. Six commonly used and powerful pre­
diction models are selected, which are autoregressive integrated moving
average (ARIMA), backpropagation (BP) neural network, extreme
learning machine (ELM), gated recurrent unit (GRU), long short-term
memory (LSTM) neural network, and bi-directional long short-term
memory (BiLSTM) neural network. The network structure settings of
BP, ELM, GRU, LSTM, and BiLSTM are as follows: the number of hidden
layers is 1 and the number of neurons is 9. The MAE, MAPE, and RMSE of
the six single-input base models are shown in Figs. 6-8, and Table 2.
As known from Figs. 6-8 and Table 2, the statistical method ARIMA
performs poorly. For example, the MAPE of ARIMA in January, July, and
October are 14.1000%, 9.3502%, and 10.3603%, respectively. It reveals
that ARIMA, as a linear method, cannot handle nonlinear wind power
data well. Shallow neural networks such as BP neural network and ELM
have less stable prediction results because they tend to fall into locally
optimal solution. However, in general, the prediction accuracy of ELM is
Fig. 10. MAPE of selected the parameter of VMD in different ways.
higher than that of the BP neural network. For example, the MAPE of BP

Table 2
Experimental results of six single-input basic models.
Model January March July October
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE

ARIMA 2.1653 14.1000 3.8090 2.2144 10.1800 4.1642 2.4040 9.3502 3.9670 1.8874 10.3603 2.9875
BP 2.0659 13.4500 3.1281 2.3036 10.5922 4.2093 2.3980 9.3266 3.6565 1.6509 9.0619 2.6839
ELM 1.9399 12.6300 3.1881 2.5541 11.7440 4.3200 2.3413 9.1061 3.7092 1.7865 9.8063 2.7793
GRU 1.8557 12.0800 3.1405 2.2235 10.2200 4.0585 2.3185 9.0176 3.5796 1.6294 8.9441 2.6099
LSTM 1.8001 11.7200 2.9844 2.1972 10.1028 4.0199 2.3119 8.9918 3.6029 1.6412 9.0089 2.6115
BiLSTM 1.6730 10.8900 2.1965 1.7808 8.1900 2.3521 2.0407 7.9370 2.7585 1.5369 8.4364 1.9266

8
J. Li et al. Electric Power Systems Research 208 (2022) 107886

high-frequency sub-series such as IMF1, IMF2 is small, but the fluctua­

tion is large. Therefore, the error of the prediction model established for
high-frequency sub-series is large. The low-frequency sub-series such as
IMF27, IMF28 have large amplitudes, but small fluctuations, so the error
of the prediction model is small. In addition, the error at the end of
IMF29 suddenly increased, In addition, the error at the end of IMF29
suddenly increased, which may be caused by insufficient model training.
So we need to retrain the model and update the parameters when the
model performance is poor.

4.5. Experiment on error correction

Error correction is used to improve the prediction accuracy of the

model by building a model to simulate the nonlinear relationship be­
tween the prediction error and the input data. The output of the model is
the prediction error, which is calculated by subtracting the true value
Fig. 11. RMSE of selected the parameter of VMD in different ways.
from the predicted value of the FPA-VMD-BiLSTM model. Two error
prediction models were developed, and the difference between these
unique hidden layer structure, which is very suitable for handling two models lies in the model inputs. The input of the error correction
nonlinear time series data, and have higher prediction accuracy. Finally,
model 1 (FPA-VMD-BiLSTM-EC1) is the IMFs and the five meteorolog­
the experiments show that BiLSTM neural network has the best accuracy ical factors in Section 3.4.1, and the input of the error correction model 2
in all months. For example, the MAE, MAPE, and RMSE of the BiLSTM
(FPA-VMD-BiLSTM-EC2, the proposed method in this paper) is the IMFs
neural network in March are respectively 1.7808, 8.1900, and 2.3521, and the low-dimensional principal component after dimensionality
which are the smallest among the six models. The BiLSTM neural
reduction by PCA.
network has better prediction accuracy because it can learn the corre­ The variance contribution rate of the five principal components ob­
lation between previous and future of the wind power series.
tained by PCA is shown in Fig. 13. Since the variance contribution rate of
the first principal components accounts for 88%, it suggests that the first
4.4. Experiment on parameter selection of VMD principal component contains most of the meteorological information.
So only the first principal component is considered in the error correc­
The parameter setting of VMD has always been a complicated issue. tion model 2 (FPA-VMD-BiLSTM-EC2). The wind power prediction re­
In this section, the FPA is used to find the optimal parameters under the sults before and after error correction in January, March, July, and
criteria for minimizing decomposition losses, and the results are shown October are shown in Figs. 14-17 and table 4.
in Table 3. In addition, we test both the frequently used parameters As can be seen from Figs. 14-17 and Table 4, compared with the
(K=5, 6, 7, 8, 9, 10, 11) and the parameters obtained by FPA. Firstly, model without error correction, the single-point error of the error
BiLSTM neural network was built for each IMF obtained from the correction model becomes smaller. The prediction accuracy is further
decomposition. Then, the predicted IMFs were summed to obtain the improved and the prediction results are more closely matched to the true
predicted wind power. Finally, the MAE, MAPE, and RMSE were used to values. In addition, the experiment also proves that PCA can improve the
evaluate the model. Experimental results are shown in Figs. 9-11. quality of meteorological factors and reduce the correlation between
As shown in Table 3, the decomposition loss of the parameters ob­ input factors, which improves the effect of error correction. For
tained by FPA is generally in 3 decimal places. For example, the example, the MAPE of the proposed method in October is only 1.0526%,
decomposition loss of the parameters obtained by FPA in January is only which is 67.20% better than that of the FPA-VMD-BiLSTM-EC1 model.
0.0034. FPA-VMD greatly reduces the loss in the decomposition process In summary, the error correction model that takes into account the low-
of VMD and makes the input data of the prediction model more com­ dimensional principal component can deeply explore the implied in­
plete. It is conducive to the improvement of prediction accuracy. formation contained in the errors and reduce the inherent errors of the
As shown in Figs. 9-11, we test both the frequently used parameters model.
and the parameter obtained by FPA. Among the frequently used pa­
rameters, the model has the best prediction accuracy when the number 4.6. Experiment on multi-step ahead prediction
of decompositions K is 11. For example, the MAE, MAPE and RMSE of
the model with K are 11 in January are respectively 0.9290, 6.3056%, To compare the performance of the forecasting models under
and 1.3982. But the model with the parameters obtained by FPA is different environments, multi-step ahead forecasting experiments are
better, and the MAE, MAPE, and RMSE in January are respectively conducted. multi-step ahead forecasting is more effective than single-
0.5358, 3.4894%, and 1.0751, which are 42.32%, 44.66%, and 23.11% step ahead forecasting in capturing the future dynamic performance of
higher than those of the model with K is 11. However, it is worth noting wind power. To emphasize the validity of the proposed method, ARIMA,
that the prediction accuracy of the model does not improve with the LSTM neural network, BiLSTM neural network, VMD-BiLSTM neural
increase of K. For example, the model with K is 10 performed slightly network, FPA-VMD-BiLSTM neural network, and the proposed method
worse than the model with K is 9 in January. Therefore, we cannot are selected Tables 5-8. show the prediction results of these six predic­
simply increase the number of K to improve the prediction accuracy. tion models at different steps in different months.
This experiment demonstrates that the method of obtaining VMD pa­ The conclusions are as follows.
rameters by FPA has higher prediction accuracy than the frequently used
parameters. (1) The error of most of the models gradually increases as the number
In addition, we will present the prediction details of the FPA-VMD- of forecasting steps increases, which is because multi-step ahead
BiLSTM model forJanuary as an example. From Table 3, the optimal forecasting is more difficult than single-step ahead forecasting.
number of K in January is 29. The predicted and true values of these 29 For example, the MAPE of ARIMA in 1-3 step prediction for
IMFs are given in Fig. 12, and also the predicted values of wind power October were 10.3603%, 11.2762%, and 11.5909%, respectively.
obtained by superimposing the predicted values of these 29 IMFs are (2) The BiLSTM neural network has higher prediction accuracy than
given at the end of the figure. As shown in Fig. 12, the amplitude of the the LSTM neural network because it can process the time-series

9
J. Li et al. Electric Power Systems Research 208 (2022) 107886

Fig. 12. Forecasting results of IMF and wind power in January.

data in both directions. For example, the RMSE of the BiLSTM in the 2-step prediction for July was 1.4983, while the BiLSTM
neural network is 2.9570 in the 3-step prediction for January, neural network was as high as 1.7729.
which is 54.00% better than that of the LSTM neural network of (4) FPA-VMD algorithm solves the parameter selection problem of
6.4283. VMD. The model prediction accuracy was further improved after
(3) VMD can improve the quality of the input data and the accuracy the parameters of VMD were obtained by FPA with the objective
of the prediction method based on signal decomposition is higher of minimizing the decomposition loss, which proved the superi­
than that of direct prediction, which proves the effectiveness of ority of the evaluation criteria and FPA. For example, the MAPE
VMD. For example, the MAE of the VMD-BiLSTM neural network of the FPA-VMD-BiLSTM model in the 1-step prediction for March

10
J. Li et al. Electric Power Systems Research 208 (2022) 107886

Fig. 16. Forecasting results before and after error correction in July.
Fig. 13. The variance contribution rate of five principal components.

Fig. 14. Forecasting results before and after error correction in January. Fig. 17. Forecasting results before and after error correction in October.

was only 3.4988%, which was 24.58% better than that of the
VMD-BiLSTM model.
(5) After the error correction, the hysteresis of the prediction process
is reduced and the accuracy of the model is further improved.
Moreover, PCA can reduce the correlation between multi-
dimensional meteorological factors and improve the quality of
the model input. It is worth noting that the error of the proposed
method is minimized in both single-step ahead and multi-step
ahead prediction for all months. For example, MAE, MAPE and
RMSE of the proposed method in the 3-step prediction for
October are all the smallest, which are 0.3016, 1.6557, and
0.4708 respectively.

5. Conclusion

Accurate short-term wind power forecasting is necessary to improve

the operational efficiency of new energy power systems, reduce the
impact of wind power brought to the smart grid, and provide a reliable
scientific basis for grid dispatching. In order to reduce the influence of
volatility and nonlinearity of wind power time sequence, a wind power
Fig. 15. Forecasting results before and after error correction in March.
forecasting method based on optimized decomposition prediction and

11
J. Li et al. Electric Power Systems Research 208 (2022) 107886

Table 4
Experimental results before and after error correction.
Model January March July October
MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE MAE MAPE RMSE
FPA-VMD-BiLSTM 0.5358 3.4894 1.0751 0.7609 3.4988 1.2957 0.9125 3.5491 1.1770 0.5847 3.2095 0.8430
FPA-VMD-BiLSTM-RC1 0.4406 2.8694 1.0978 0.3509 1.6137 0.9444 0.4643 1.8060 0.9700 0.2524 1.3856 0.5305
Proposed Method 0.3954 2.5746 1.0491 0.3161 1.4534 0.8595 0.3387 1.3173 0.8453 0.1918 1.0526 0.4687

Table 5
Experimental results of the proposed model and other benchmark models in January.
Model MAE(MW) MAPE (%) RMSE(MW)
1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step

ARIMA 2.1653 1.9431 1.9826 14.1000 12.6532 12.9105 3.8090 3.7540 3.9879
LSTM 1.8001 3.0175 3.9012 11.7200 19.6498 25.4044 2.9844 5.1534 6.4283
BiLSTM 1.6730 1.6477 1.5325 10.8900 10.7295 9.9793 2.1965 2.6287 2.9570
VMD-BiLSTM 0.9290 0.8707 1.2811 6.3056 5.6701 8.3423 1.3982 1.5662 1.8024
FPA-VMD-BiLSTM 0.5358 0.7292 0.5635 3.4894 4.7484 3.6694 1.0751 1.4647 1.1946
Proposed Method 0.3954 0.4056 0.5232 2.5746 2.6413 3.4074 1.0491 1.0729 1.2375

Table 6
Experimental results of the proposed model and other benchmark models in March.
Model MAE(MW) MAPE (%) RMSE(MW)
1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step

ARIMA 2.2144 2.2403 2.2617 10.1800 10.3010 10.3995 4.1642 4.2023 4.2428
LSTM 2.1972 4.0744 5.5287 10.1028 18.7343 25.4214 4.0199 7.0833 9.1923
BiLSTM 1.7808 1.7660 2.5169 8.1900 8.1202 11.5727 2.3521 2.6296 3.5729
VMD-BiLSTM 1.0117 1.3578 1.1837 4.6393 6.2434 5.4426 1.3301 1.6476 1.6062
FPA-VMD-BiLSTM 0.7609 0.8828 1.0387 3.4988 4.0591 4.7758 1.2957 1.2348 1.3666
Proposed Method 0.3161 0.3739 0.4592 1.4534 1.7191 2.1115 0.8595 0.9572 0.9491

error correction is proposed. In the first stage, a hybrid forecasting (3) To reduce the correlation among meteorological factors and the
model based on FPA-VMD-BiLSTM is built to complete the initial pre­ computational burden of the model, PCA is employed to convert
diction of the original wind power sequence. In the second stage, with the multi-dimensional meteorological factors into low-
IMFs and low-dimensional principal component as input, an error dimensional principal components.
correction model based on BiLSTM neural network is constructed. In this (4) With the input of IMFs and principal component, An error
paper, several experiments and analyses are conducted using datasets correction model based on BiLSTM neural network is established
from four months in different seasons of Mahuangshan. The results show to reduce the inherent error of the model.
that the prediction accuracy of the proposed method is higher than that (5) Compared with other benchmark models, the proposed two-stage
of other benchmark models in 1-3 steps forecasting. forecasting method performs optimally in both single-step and
In addition, the specific contributions of the methodology in this multi-step ahead forecasting. For example, the performance of
paper are as follows. this proposed model is optimal for single-step ahead forecasting
in January, March, July, and October, with MAPE of 2.5745%,
(1) To guide the parameter setting of VMD, an evaluation criterion 1.4535%, 1.3173%, and 1.0526%, respectively. It indicates that
applicable to the field of wind power forecasting is proposed. the proposed method can better understand the future trend of
Besides, FPA is used to automatically find the optimal decom­ wind power series and guarantee the stable operation of the
position parameters of VMD. FPA-VMD reduces the signal loss in power system.
the decomposition process and improves the decomposition ef­
fect of VMD. So far, the decomposition-based wind power forecasting method is
(2) The forecasting model based on BiLSTM neural network has the still a hot research topic. However, the sequence obtained by decom­
function of bidirectional processing and long-term memory, position is not stable. For example, the sub-sequences decomposed by
which can better explore the deep temporal features inside the the same method with different window lengths are not the same.
wind power. Therefore, reducing the inconsistency problem in the decomposition
process will be the focus of future research.

Table 7
Experimental results of the proposed model and other benchmark models in July.
Model MAE(MW) MAPE (%) RMSE(MW)
1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step

ARIMA 2.4040 2.3706 2.4052 9.3502 9.2199 9.3546 3.9670 3.9715 4.1001
LSTM 2.3119 4.3018 5.8331 8.9918 16.7313 22.6870 3.6029 6.4966 8.4197
BiLSTM 2.0407 1.7729 2.3907 7.9370 6.8956 9.2983 2.7585 2.6068 3.8049
VMD-BiLSTM 1.1506 1.4983 1.2515 4.3908 5.8273 4.8675 1.6883 2.1293 1.8829
FPA-VMD-BiLSTM 0.9125 0.9279 0.8232 3.5491 3.6091 3.2015 1.1770 1.5704 1.5040
Proposed Method 0.3387 0.4060 0.4639 1.3173 1.5792 1.8041 0.8453 0.9213 0.8987

12
J. Li et al. Electric Power Systems Research 208 (2022) 107886

Table 8
Experimental results of the proposed model and other benchmark models in October.
Model MAE(MW) MAPE (%) RMSE(MW)
1-step 2-step 3-step 1-step 2-step 3-step 1-step 2-step 3-step

ARIMA 1.8874 2.0543 2.1116 10.3603 11.2762 11.5909 2.9875 3.5823 3.9180
LSTM 1.6412 3.2211 4.2271 9.0089 17.6811 23.2032 2.6115 5.0845 6.6293
BiLSTM 1.5369 1.4783 1.8144 8.4364 8.1147 9.9593 1.9266 1.9996 2.4662
VMD-BiLSTM 0.7839 0.6482 1.0523 3.9878 3.5583 5.7762 1.0369 0.9485 1.2863
FPA-VMD-BiLSTM 0.5847 1.2176 0.8851 3.2095 6.6835 4.8584 0.8430 1.4599 1.1253
Proposed Method 0.1918 0.2590 0.3016 1.0526 1.4219 1.6557 0.4687 0.6093 0.4708

[17] O Abedinia, M Lotfi, M Bagheri, et al., Improved EMD-based complex prediction

model for wind power forecasting[J], IEEE Transactions on Sustainable Energy 11
(4) (2020) 2790–2802.
CRediT authorship contribution statement [18] Y Zhang, B Chen, G Pan, et al., A novel hybrid model based on VMD-WT and PCA-
BP-RBF neural network for short-term wind speed forecasting[J], Energy
Conversion and Management 195 (2019) 180–197.
Jun Li: Conceptualization, Methodology, Software, Writing – orig­
[19] H Liu, Z Duan, H Wu, et al., Wind speed forecasting models based on data
inal draft, Visualization, Validation. Shuqing Zhang: Methodology, decomposition, feature selection and group method of data handling network[J],
Supervision, Funding acquisition, Project administration. Zhenning Measurement 148 (2019), 106971.
Yang: Visualization, Data curation, Software, Writing – original draft. [20] Y Dong, H Zhang, C Wang, et al., Wind power forecasting based on stacking
ensemble model, decomposition and intelligent optimization algorithm[J],
Neurocomputing 462 (2021) 169–184.
Declaration of Competing Interest [21] B Li, J Zhang, Y He, et al., Short-term load-forecasting method based on wavelet
decomposition with second-order gray neural network model combined with ADF
test[J], 5, IEEE Access, 2017, pp. 16324–16331.
The authors declare that they have no known competing financial [22] M Yang, X Chen, J Du, et al., Ultra-short-term multistep wind power prediction
interests or personal relationships that could have appeared to influence based on improved EMD and reconstruction method using run-length analysis[J],
IEEE Access 6 (2018) 31908–31917.
the work reported in this paper. [23] L Han, H Jing, R Zhang, et al., Wind power forecast based on improved Long Short-
Term Memory network[J], Energy 189 (2019), 116300.
Acknowledgments [24] J Duan, P Wang, W Ma, et al., Short-term wind power forecasting using the hybrid
model of improved variational mode decomposition and Correntropy Long Short-
term memory neural network[J], Energy 214 (2021), 118980.
This research was supported by the National Natural Science Foun­ [25] Y Wang, L. Wu, On practical challenges of decomposition-based hybrid forecasting
dation of China under Grant 51875498, the Natural Science Foundation algorithms for wind speed and solar irradiation[J], Energy 112 (2016) 208–220.
[26] Y Wang, R Zou, F Liu, et al., A review of wind speed and wind power forecasting
of Hebei Province under Grant F2020203058, and the Key R&D projects with deep neural networks[J], Applied Energy 304 (2021), 117766.
of Hebei Province under Grant 18211833D. Besides, the authors are also [27] B Gu, T Zhang, H Meng, et al., Short-term forecasting and uncertainty analysis of
grateful to all the viewers and the editor for their valuable comments. wind power based on long short-term memory, cloud model and non-parametric
kernel density estimation[J], Renewable Energy 164 (2021) 687–708.
[28] F Shahid, A Zameer, M. Muneeb, A novel genetic LSTM model for wind power
References forecast[J], Energy 223 (2021), 120069.
[29] A Dolatabadi, H Abdeltawab, Y A R I Mohamed, Hybrid Deep Learning-Based
[1] A Kisvari, Z Lin, X. Liu, Wind power forecasting–A data-driven method along with Model for Wind Speed Forecasting Based on DWPT and Bidirectional LSTM
gated recurrent neural network[J], Renewable Energy, 163 (2021) 1895–1909. Network[J], IEEE Access 8 (2020) 229219–229232.
[2] J Yang, Q Liu, X Li, et al., Overview of wind power in China: status and future[J], [30] M Xu, Z Lu, Y Qiao, et al., Modelling of wind power forecasting errors based on
Sustainability 9 (8) (2017) 1454. kernel recursive least-squares method[J], Journal of Modern Power Systems and
[3] U Singh, M Rizwan, M Alaraj, et al., A Machine Learning-Based Gradient Boosting Clean Energy 5 (5) (2017) 735–745.
Regression Approach for Wind Power Production Forecasting: A Step towards [31] W Ko, D Hur, J K Park, Correction of wind power forecasting by considering wind
Smart Grid Environments[J], Energies 14 (16) (2021) 5196. speed forecast error[J], Journal of International Council on Electrical Engineering
[4] M A Hossain, R K Chakrabortty, S Elsawah, et al., Very short-term forecasting of 5 (1) (2015) 47–50.
wind power generation using hybrid deep learning model[J], Journal of Cleaner [32] Y Hao, C. Tian, A novel two-stage forecasting model based on error factor and
Production 296 (2021), 126564. ensemble method for multi-step wind power forecasting[J], Applied energy 238
[5] P Du, J Wang, W Yang, et al., A novel hybrid model for short-term wind power (2019) 368–383.
forecasting[J], Applied Soft Computing 80 (2019) 93–106. [33] Y Zhang, J Han, G Pan, et al., A multi-stage predicting methodology based on data
[6] M Ahmadi, M. Khashei, Current status of hybrid structures in wind forecasting[J], decomposition and error correction for ultra-short-term wind energy prediction[J],
Engineering Applications of Artificial Intelligence 99 (2021), 104133. Journal of Cleaner Production 292 (2021), 125981.
[7] H Wang, S Han, Y Liu, et al., Sequence transfer correction algorithm for numerical [34] Y Zhang, B Chen, G Pan, et al., A novel hybrid model based on VMD-WT and PCA-
weather prediction wind speed and its application in a wind power forecasting BP-RBF neural network for short-term wind speed forecasting[J], Energy
system[J], Applied Energy 237 (2019) 1–10. Conversion and Management 195 (2019) 180–197.
[8] J Lu, B Wang, H Ren, et al., Two-tier reactive power and voltage control strategy [35] X Kong, X Liu, R Shi, et al., Wind speed prediction using reduced support vector
based on ARMA renewable power forecasting models[J], Energies 10 (10) (2017) machines with feature selection[J], Neurocomputing 169 (2015) 449–456.
1518. [36] K Dragomiretskiy, D. Zosso, Variational mode decomposition[J], IEEE transactions
[9] S N Singh, A. Mohapatra, Repeated wavelet transform based ARIMA model for very on signal processing 62 (3) (2013) 531–544.
short-term wind speed forecasting[J], Renewable energy 136 (2019) 758–768. [37] H MA, Q TONG, Y ZHANG, Applications of optimization parameters VMD to fault
[10] Y Wang, Q Hu, D Meng, et al., Deterministic and probabilistic wind power diagnosis of rolling bearings[J], China Mechanical Engineering 29 (04) (2018)
forecasting using a variational Bayesian-based adaptive robust multi-kernel 390.
regression model[J], Applied energy 208 (2017) 1097–1112. [38] Y Huang, Y Gao, Y Gan, et al., A new financial data forecasting model using genetic
[11] M Ding, H Zhou, H Xie, et al., A time series model based on hybrid-kernel least- algorithm and long short-term memory network[J], Neurocomputing 425 (2021)
squares support vector machine for short-term wind power forecasting[J], ISA 207–218.
transactions 108 (2021) 58–68. [39] Y Hao, C. Tian, A novel two-stage forecasting model based on error factor and
[12] Z Wang, B Wang, C Liu, et al., Improved BP neural network algorithm to wind ensemble method for multi-step wind power forecasting[J], Applied energy 238
power forecast[J], The Journal of Engineering 2017 (13) (2017) 940–943. (2019) 368–383.
[13] A A Abdoos, A new intelligent method based on combination of VMD and ELM for [40] Y Huang, Y Gao, Y Gan, et al., A new financial data forecasting model using genetic
short term wind power forecasting[J], Neurocomputing 203 (2016) 111–120. algorithm and long short-term memory network[J], Neurocomputing 425 (2021)
[14] S Hu, Y Xiang, D Huo, et al., An improved deep belief network based hybrid 207–218.
forecasting method for wind power[J], Energy 224 (2021), 120185. [41] P Lu, L Ye, Y Zhao, et al., Review of meta-heuristic algorithms for wind power
[15] M S H Lipu, M S Miah, M A Hannan, et al., Artificial Intelligence Based Hybrid prediction: Methodologies, applications and challenges[J], Applied Energy 301
Forecasting Approaches for Wind Power Generation: Progress, Challenges and (2021), 117446.
Prospects[J], IEEE Access 9 (2021) 102460–102489. [42] M Abdel-Basset, L A Shawky, Flower pollination algorithm: a comprehensive
[16] C Lan, Y B Zhang, J Zhang, et al., Short-Term Wind Power Forecasting Model based review[J], Artificial Intelligence Review 52 (4) (2019) 2533–2557.
on ICA-BP Neural Network[J], Computer and Information Science 8 (1) (2015) 1.

13
J. Li et al. Electric Power Systems Research 208 (2022) 107886

[43] B Gu, T Zhang, H Meng, et al., Short-term forecasting and uncertainty analysis of [49] Y Zhang, B Chen, G Pan, et al., A novel hybrid model based on VMD-WT and PCA-
wind power based on long short-term memory, cloud model and non-parametric BP-RBF neural network for short-term wind speed forecasting[J], Energy
kernel density estimation[J], Renewable Energy 164 (2021) 687–708. Conversion and Management 195 (2019) 180–197.
[44] F Shahid, A Zameer, M. Muneeb, A novel genetic LSTM model for wind power [50] S. Wu, Research on Wind Power Ultra-short-term Forecasting Method Based on, in:
forecast[J], Energy 223 (2021), 120069. PCA-LSTM[C]//IOP Conference Series: Earth and Environmental Science 508, IOP
[45] E López, C Valle, H Allende, et al., Wind power forecasting based on echo state Publishing, 2020, 012068.
networks and long short-term memory[J], Energies 11 (3) (2018) 526. [51] M Ding, H Zhou, H Xie, et al., A gated recurrent unit neural networks based wind
[46] A Dolatabadi, H Abdeltawab, Y A R I Mohamed, Hybrid Deep Learning-Based speed error correction model for short-term wind power forecasting[J],
Model for Wind Speed Forecasting Based on DWPT and Bidirectional LSTM Neurocomputing 365 (2019) 54–61.
Network[J], IEEE Access 8 (2020) 229219–229232. [52] B E Moore, C Gao, R.R. Nadakuditi, Panoramic robust pca for
[47] H Zhen, D Niu, M Yu, et al., A hybrid deep learning model and comparison for wind foreground–background separation on noisy, free-motion camera video[J], IEEE
power forecasting considering temporal-spatial feature extraction[J], Transactions on Computational Imaging 5 (2) (2019) 195–211.
Sustainability 12 (22) (2020) 9490. [53] J M González-Sopeña, V Pakrashi, B. Ghosh, An overview of performance
[48] H Liu, R Yang, Z Duan, Wind speed forecasting using a new multi-factor fusion and evaluation metrics for short-term statistical wind power forecasting[J], Renewable
multi-resolution ensemble model with real-time decomposition and adaptive error and Sustainable Energy Reviews (2020), 110515.
correction[J], Energy Conversion and Management 217 (2020), 112995.

14