Energy Conversion and Management 205 (2020) 112345

Contents lists available at ScienceDirect

Energy Conversion and Management

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

Application of hybrid model based on double decomposition, error T

correction and deep learning in short-term wind speed prediction
Zherui Maa, Hongwei Chena, Jiangjiang Wanga, Xin Yangb, , Rujing Yana, Jiandong Jiaa,
⁎

Wenliang Xuc
a
School of Energy, Power and Mechanical Engineering, North China Electric Power University, Baoding, Hebei Province 071003, China
b
School of Water Conservancy and Hydroelectric Power, Hebei University of Engineering, Handan, Hebei Province 056002, China
c
DaTang East China Electric Power Test & Research Institute, Hefei, Anhui Province 230000, China

ARTICLE INFO ABSTRACT

Keywords: As wind power accounts for an increasing proportion of the electricity market, the wind speed prediction plays a
Wind speed prediction vital role in the stable operation of the power grid. However, owing to the stochastic nature of wind speed,
Long short term memory neural network predicting wind speeds accurately is difficult. Aims at this challenge, a new short-term wind speed prediction
Hybrid model model based on double decomposition, error correction strategy and deep learning algorithm is proposed. The
Complete ensemble empirical mode
complete ensemble empirical mode decomposition with adaptive noise and variational mode decomposition are
decomposition with adaptive noise
applied to decompose the original wind speed series and error series, respectively. The deep learning algorithm
Variational mode decomposition
Error correction based on long short term memory neural network, is utilized to detect the long-term and short-term memory
characteristics and build the suitable prediction model for each sub-series. In the four real forecasting cases, nine
models were built to compare the performance of the proposed model. The experimental results show that the
proposed model performs better than all other considered models without double decomposition, and the var-
iational mode decomposition for error series can improve the effect of error correction strategy.

1. Introduction vector machine (SVM) [9], extreme learning machine (ELM) [10], and
deep learning [11]. Zhou et al. [12] employed least squares support
With the changes in the global energy structure, the proportion of vector machines (LSSVM) model for one-step ahead wind speed fore-
wind energy in the future energy is gradually increasing [1]. However, casting. Deep learning is more capable of mining the deep features of
due to the stochastic and intermittent characteristics of wind speed, the data than traditional neural networks, and thus has been widely applied
wind power is highly random, causing voltage and frequency fluctua- in many fields [13], which gradually penetrated into the field of wind
tions in the power system [2]. The high-precision wind speed prediction speed prediction [14]. Wang et al. [15] applied the k-means cluster
can reduce the occurrence of wind power failure, ensuring the security approach to process the numerical weather prediction (NWP) data as
and stability of the power system [3]. Therefore, accurate wind speed the input of deep belief network (DBN) model. The experimental results
prediction is crucial in the effective improvement of power grid safety showed the model can accurately follow the trend of wind speed
and the full utilization of wind energy resources. changes. Chen et al. [16] proposed a multifactor spatio-temporal cor-
At present, short-term wind speed prediction methods can be relation model combined with convolutional neural networks (CNN)
broadly classified into two categories: (1) physical model (2) statistical and long short term memory neural network (LSTM). The result showed
model [4]. The physical models are based on meteorological data, that the model can extract the spatio-temporal correlation features
which predict wind speed through thermodynamic equations and fluid while utilizing the advantages of individual models effectively. Chen
dynamic equations [5]. The statistical models predict wind speed by et al. [17] proposed a model by using nonlinear-learning ensemble of
historical data reflecting the relationship between input variables and deep learning time series prediction based on LSTMs, SVM and extremal
output variables [6]. In the past several years, statistical models have optimization (EO) algorithm. The proposed model had satisfactory
received more and more attention with the rapid development of data performance. Hu et al. [18] established three LSTM models and three
science [7]. The development of machine learning algorithms is parti- hysteretic ELM models, and the forecasting results of each predictor are
cularly rapid, such as artificial neural network (ANN) [8], support aggregated by a novel nonlinear combined mechanism composed of

⁎
Corresponding author.
E-mail address: [email protected] (X. Yang).

https://doi.org/10.1016/j.enconman.2019.112345
Received 23 September 2019; Received in revised form 18 November 2019; Accepted 25 November 2019
0196-8904/ © 2019 Elsevier Ltd. All rights reserved.
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Nomenclature MLP multi-layer perceptron

RNN recurrent neural network
ANN artificial neural network SE sample entropy
AWNN adaptive wavelet neural network SVM support vector machine
BPNN back-propagation neural network VMD variational mode decomposition
CEEMDAN complete ensemble empirical mode decomposition with WPD wavelet packet decomposition
adaptive noise WT wavelet transform
CNN convolutional neural network
DBN deep belief network Symbols
ENN Elman neural network
EEMD ensemble empirical mode decomposition MAE mean absolute error (m/s)
ELM extreme learning machine MAPE mean absolute percentage error (%)
EO extremal optimization PMAE promoting percentages of the MAE (%)
EWT empirical wavelet transform PMAPE promoting percentages of the MAPE (%)
GRNN general regression neural network PRMSE promoting percentages of the RMSE (%)
LSTM long short term memory neural network RMSE root mean square error (m/s)

LSTM network. The results demonstrated the proposed hybrid model is heteroscedasticity cannot be ignored. Liang et al. [32] predicted the
feasible compared to the baseline models. wind speed series by SVM, and established error models to correct the
Signal processing algorithms can effectively extract the features of wind speed by SVM and ELM. The experimental results showed that the
the signal [19], which has been used by many scholars in combination model can significantly improve the accuracy of short-term wind speed
with machine learning algorithms in the field of wind speed prediction prediction. Huang et al. [33] employed general regression neural net-
[20]. Meng et al. [21] applied wavelet packet decomposition (WPD) to work model (GRNN) to predict the error series of sub-series for wind
decompose the original wind speed series and back-propagation neural speed data. The results showed the method can correct the error of sub-
network (BPNN) optimized by crisscross optimization algorithm is series. Li et al. [34] established error model base on regularized ELM to
employed to predict the each sub-series. A hybrid model for wind speed improve the performance of the hybrid model significantly. Mi et al.
prediction was proposed by Sun [22], which was applied complete [35] proposed a outlier correction method to detect and correct the
ensemble empirical mode decomposition (CEEMD) to divide the ori- wrong predictions. The experimental results demonstrated the effec-
ginal wind speed data into a finite set of components. Niu et al. [23] tiveness of the model and proved that the outlier correction can en-
adopted empirical mode decomposition (EMD) to process the wind hance the robustness of the model. The above researches showed that a
speed series. The experimental results showed the EMD algorithm can reasonable correction strategy can improve the performance of the
successfully preprocess the wind speed data to enhance the forecasting model. However, error series also have strong randomness and volati-
accuracy. Santhosh et al. [24] adopted ensemble empirical mode de- lity. The single prediction cannot accurately describe the fluctuation of
composition (EEMD) and adaptive wavelet neural network (AWNN) for error series. Data preprocessing is an effective way to promote the ac-
wind speed prediction. The experimental results showed the EEMD al- curacy of error series. In this study, a new wind speed prediction model
gorithm can successfully preprocess the wind speed data to enhance the combining double decomposition, error correction and deep learning is
forecasting accuracy. Zhang et al. [25] employed complete ensemble proposed. The contributions are as follows:
empirical mode decomposition with adaptive noise (CEEMDAN) to di-
vide the wind speed data, and each component was forecasted by dif- (1) The double decomposition strategy is applied to the wind speed
ferent neural networks. The result showed the combined model has prediction. The original wind speed series and error series are de-
higher stability and prediction precision compared with other single composed respectively to reduce the complexity and non-statio-
models. narity of each series.
In the past several years, research on combining signal processing (2) The LSTM network, a deep learning algorithm which has satisfac-
with deep learning for wind speed prediction is gradually increasing tory performance in long short term dependencies, is used to
[26]. Zhou et al. [27] employed variational mode decomposition complete the forecasting for the sub-series.
(VMD) to decompose the wind speed series and the CNN was used to (3) A novel wind speed prediction model is proposed by combing the
predict wind speed sub-series. Compared with the undecomposed CNN CEEMDAN, VMD and LSTM. The error model combined with VMD
model, the proposed model has better performance in both multi-step and LSTM can extract deeper features and improve the correction
prediction and single-step prediction. Wang et al. [28] combined the effect of the model.
wavelet transform (WT) and DBN to predict the wind speed. The results (4) The superiority of the proposed CEEMDAN-error-VMD-LSTM model
showed that the method can train deep nonlinear features well and is successfully demonstrated on four experimental cases. Nine
make the model have good performance. Liu et al. [29] decomposed the models are employed to verify the accuracy and validation of the
wind speed signal by empirical wavelet transform (EWT), and em- proposed hybrid model.
ployed LSTM network and Elman neural network (ENN) to predict the
low frequency and high frequency components respectively. The ex- This paper is organized as follows: Section 2 describes the methods
perimental results verified that the model has better prediction accu- used in the proposed model. Section 3 presents the framework of the
racy. However, although the prediction accuracy is greatly improved by proposed model. Section 4 introduces four experiment cases and some
the wind speed series decomposition, the error of each component will metrics to evaluate the performance of the model. Section 5 shows the
be superposed and the final error will be larger. results and discussions about the performance of the models. Section 6
To obtain better wind speed prediction performance, some scholars summarizes some conclusions.
introduced error correction mechanism in the hybrid model [30]. Wang
et al. [31] proposed an error correction strategy based on a generalized
2. Methods
auto-regressive conditionally heteroscedastic model to modify the
prediction errors of the LSTM model when the inherent correlation and
The CEEMDAN, VMD technique and LSTM network included in this

2
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

study, which are described in detail.

2.1. Complete ensemble empirical mode decomposition with adaptive noise

The empirical mode decomposition (EMD) [36] has been widely

studied as a common signal processing technique in recent years. The
EMD can process nonlinear signals, but the modal mixing problem is
often encountered in practical applications. Therefore, the ensemble
empirical mode decomposition (EEMD) was proposed by Wu [37]. The
Fig. 1. RNN expansion diagram.
EEMD eliminates modal mixing problems by adding noise on the basis
of EMD. However, the EEMD cannot completely eliminate white noise,
so the reconstruction error is large. Although the increase in the 2.2. Variational mode decomposition
number of ensemble can reduce the reconstruction error, it leads to a
badly increase in the amount of calculation. To this end, Torres [38] The variational mode decomposition (VMD) method is proposed by
proposed the complete ensemble empirical mode decomposition with Dragomiretskiy [40]. This method has the advantages of high compu-
adaptive noise (CEEMDAN) based on EEMD. By adding adaptive white tational efficiency and robustness, which can defect the modal mixing
noise, the CEEMDAN not only solves the modal mixing problem, but problem [41]. By applying VMD, the signal x(t) is decomposed into K
also eliminates the reconstruction error. sub-series or variational modes uk (k = 1, …, K), and the sum of the
The steps of CEEMDAN algorithm are the following [39]: bandwidth estimates of each mode is minimized. Therefore, the key is
how to construct and solve the variational problem. The construction
(1) The white noise of the standard normal distribution is added into steps of the variational problem are as below [42]:
the original signal x(t), then the signal is decomposed to obtain
intrinsic mode functions (IMF) by EMD. The first mode is calculated (1) Obtain a corresponding spectrum by performing a Hilbert transform
by the following equation: on each mode function uk;
N
(2) Move the spectrum of uk to the respective estimated center fre-
1 quencies ωk by an exponential hybrid modulation method;
IMF1 = Ei [x + i]
N i=1 (1) (3) Use Gauss smoothness and gradient squared criteria for the signal to
demodulate and estimate the bandwidth of uk.
where IMF1 is the first-order IMF component, N is the number of times
to add noise, Ei is the operator which produces the i-th mode obtained
The constrained variational problem can be presented as follow:
by EMD, ε is the ratio of the added noise, and ωi is the added white
noise. K
j j kt
2
min t ( (t ) + t
) uk (t ) e
{uk }{ k } k =1 2
(2) Calculate the first residual at the first stage. K
s. t . uk = x (t )
r1 = x (t ) IMF1 (2) k =1 (7)
where r1 is the first residual. where δ(t) is the Dirac distribution and ∗ denotes convolution.
In order to solve the optimal solution of the constrained variational
(3) The signal (r1 + εE1[ωi], i = 1,2,3,…,N) was decomposed, which problem, the Lagrangian multiplier λ(t) and the quadratic penalty term
continue to iteratively decompose the component that satisfy the α are introduced to transform the constrained problem into an un-
condition of IMF1. The second mode IMF2 is defined as: constrained problem, and it can be described as:
N K 2
1 j
IMF2 = Ei [r1 + Ei [ i]] L({uk },{ k }, )= t ( (t ) + ) uk (t ) e j kt
N i=1 (3) t
k =1 2
K 2 K
(4) Calculate the k-th residual. + f (t ) uk (t ) + (t ), f (t ) uk (t )
k =1 k=1
rk = rk 1 IMFk (4) 2
(8)
(5) Decompose realization the signal (rk + εEk[ωi], i = 1,2,3,…,N), and The saddle point of the above Lagrangian function is calculated by
the (k + 1)-th mode IMFk+1 is defined as: iteratively updating uk+1, ωk+1 and λk+1 by using the alternating di-
rection multiplier algorithm. The steps of the iterative process are as
N
IMFk + 1 =
1
Ei [rk + Ek [ i ]] follows [43]:
N i=1 (5)
(1) Initialize each mode component and center frequency
(6) Continue the above iterative screening until no further decom- 1
position is possible. The final residual component is calculated as The initial values of {uk1} 、{ k1} 、{ ^ } and n are set to 0, and K is set
follows: to a positive integer of the number of times to be resolved.
K n+1
R = x (t ) IMFk (2) Update the u^k and n+1
k by the following equations.
k=1 (6) ^( )
x^ ( ) u^i ( ) +
where R is the residual and K is the number of IMF components. n+1 i k
2

Compared with EMD and EEMD, the CEEMDAN not only solves the u^k ( ) =
1+2 ( k)
2
(9)
modal mixing problem perfectly, but also has higher computational
efficiency and less reconstruction error. Therefore, the CEEMDAN
technique is more suitable for filtering and denoising process.

3
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 2. LSTM unit structure diagram.

K n+ 1 n
k=1
u^k u^k 2
2
n <
u^k 22 (12)

2.3. Long short term memory neural network

The standard recurrent neural network (RNN) consists of an input

layer, a hidden layer, and an output layer. The unfolded structure of
RNN is shown in Fig. 1. The input series x = [x1,x2,…,xt], when the
time t is, the output ht of the RNN and the output layer output yt are
calculated as follows:
ht = F (Whh ht 1 + Wxh xt + bh) (13)

yt = (Why ht + b y ) (14)
Fig. 3. LSTM expansion diagram. where ht is the output of the hidden layer; F() is the hidden layer ac-
tivation function; ht-1 is the previous hidden layer output; xt is the input
|u^k ( )|2 d at time t; Whh, Wxh are the weight matrix corresponding to the output of
the previous hidden layer and the weight matrix corresponding to the
n+1 0
k =
^
|uk ( )|2 d
0 (10) input of the time t respectively; yt is the output of the output layer; σ
where ukn+ 1 ( ) , ui ( ) , x ( ) and ( ) are the Fourier transform of represents the activation function of the output layer; Why is the weight
ukn + 1 (t ) , ui (t ) , x (t ) and (t ) respectively; n is the number of iterations. matrix corresponding to the output layer; bh, by are the bias corre-
sponding to the hidden layer and the output layer, respectively.
(3) Update the λ(ω) in the non-negative frequency interval, being ex- The RNN has the problem of vanishing gradients during back-
pressed as: propagation, which makes it difficult to continuously optimize network
parameters [44]. As a special kind of RNN, long short term memory
K
^n + 1 ( ) = ^ n ( ) + (f^ ( )
n+ 1
u^k ( )) neural network (LSTM) can effectively solve this problem, and the
k=1 (11) LSTM network controls the influence of instant information on histor-
ical information by adding a gating unit, so that the network model can
where τ is the iterative factor.
save and transmit information for a long time [45].
The LSTM network can be understood as a memory unit that attempts
(4) Set the evaluation accuracy ε. If the condition of Eq. (12) is sa-
to store information for a long time. The memory unit is protected by
tisfied, the iteration is stopped, otherwise it returns to step (2) to
three activation gates (forgotten gates, input gates, and output gates), so
perform the calculation again. Finally, k mode component dis-
that efficient information can be updated and utilized. The structure
tributions are obtained, that is, the variational mode distribution
diagram of the LSTM unit is shown in Fig. 2. The expressions of the forget
decomposition is completed.
gate, the input gate, and the output gate are as follows [46]:

4
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 4. Framework of proposed hybrid model.

gf = Whf ht 1 + Wxf x t + bf previous time; σ represents the sigmoid activation function.

ft = (gf ) The LSTM unit output ht is defined as follows:

(15) Ct = tanh(Whc ht 1 + Wxc xt + bc ) (18)

gi = Whi ht 1 + Wxi xt + bi
it = (gi ) Ct = ft Ct + it C^t (19)
(16) 1

go = Who ht 1 + Wxo xt + bo ht = ot tanh(Ct ) (20)

ot = ( go ) (17)
where Ct is the temporary state entered for time t; Ct is state of the
where xt is the input series at time t; ht-1 is the output of the LSTM unit at current cell; tanh is the tanh activation function; ht is output of the
time t-1; Wxf, Wxi, Wxo, Whf, Whi, and Who represent the corresponding current cell; ° denotes the Hadamard product.
weight matrices; bf, bi, bo are the bias of the forget gate, the input gate, The LSTM network transmits the information of the previous time
and the output gate respectively; ft, it, ot are the state of the forget gate, and the information of all the previous moments after filtering, so that
the input gate, and the output gate respectively; gf, gi, go are the linear the output of the previous time can balance the information of the
outputs of the series input at time t and the output of the LSTM at the previous time. Fig. 3 shows LSTM expansion diagram.

5
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 5. Original wind speed series (S1–S4).

Table 1 data is collected in the M2 measurement tower of the National Re-

Statistical information of wind speed series (S1–S4) (unit: m/s). newable Energy Laboratory (NREL) National Wind Energy Technology
Parameters Mean Standard deviation Maximum Minimum Median
Center (NWTC) [47]. The samples in S1 and S2 are separated by 15 min,
and every sample is the average of the 15 min observations. The two
S1 5.31 4.50 32.21 0.29 3.84 data sets have a total of 2880 points and the duration is 30 days. The
S2 3.77 2.10 14.41 0.30 3.32 samples in S3 and S4 are separated by 1 h and every sample is the
S3 4.98 3.83 25.01 0.34 3.90
average of the 1 h observations. The two datasets have a total of 2160
S4 4.03 2.44 16.84 0.34 3.46
points and the duration of 90 days. The last 20% samples of the four
wind speed data are used as the test sets. The statistics of the four wind
3. Framework of the proposed hybrid model speed series are shown in the Table 1.

The hybrid model framework proposed in this paper is shown in 4.2. Evaluation indexes
Fig. 4. The main process is as follows:
The mean absolute error (MAE), root mean square error (RMSE) and
(1) In the processing I, the original wind speed series is decomposed mean absolute percentage error (MAPE) are adopted to reflect the
into several sub-series and a residual series by CEEMDAN algo- prediction performance of the models. The error indexes are defined as
rithm, and the LSTM network is applied to predict each sub-series below:
(prediction result IMF').
N
(2) In the processing II, the error series is constructed by the original 1
MAE = |yi yi |
series prediction result in step I and the original observation value. N i=1 (21)
(3) In the processing III, the VMD is applied to decompose the error
series to obtain a series of sub-series, and each sub-series is pre- 1
N
yi yi
dicted using LSTM network (prediction result MODE'). MAPE =
N yi (22)
(4) In the processing IV, the predicted error series (processing III) is i=1

used to correct the prediction result of the original series to obtain

N
the final predicted wind speed series. RMSE =
1
(yi yi ) 2
N i=1 (23)
4. Case study
where yi is the actual value, yi is the predicted value and N is the total
number of the data.
In this section, the data description and evaluation indexes are first
In order to compare the performance of the involved models, the
introduced. Then, the process analysis is carried out on one of the cases.
promoting percentages of the MAE (PMAE), the promoting percentages
All of the experiments are developed in Python version 3.6, and the
of the MAPE (PMAPE), and the promoting percentages of the RMSE
relevant deep learning algorithms are implemented using the Keras
(PRMSE) are executed in this study. The definitions of these evaluation
deep learning package. The core configuration of the computer includes
indexes are demonstrated as follows as:
a 64-bit 3.6 GHz Intel Core i7-4790U CPU and 16.0 GB of RAM.
MAE1 MAE2
PMAE =
4.1. Data description MAE1 (24)

RMSE1 RMSE2
Four wind speed series from S1 to S4 are taken as the experimental PRMSE =
cases to verify the performance of model, as shown in the Fig. 5. The RMSE1 (25)

6
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 6. CEEMDAN processing results for wind speed series S2.

Table 2 series. Taking the S2 case as an example, the original wind speed series
Sample entropy of the original wind speed series after CEEMDAN decomposi- is decomposed by CEEMDAN algorithm, and the result is shown in
tion. Fig. 6. Table 2 shows the sample entropy of the original wind speed
Parameters SE Parameters SE series and the sub-series. From Table 2, it can be clearly seen that the
original series has the biggest sample entropy. After CEEMDAN de-
Origin 2.6514 IMF7 0.2640 composition, the sample entropy of each sub-series is greatly reduced
IMF1 1.7897 IMF8 0.1048
(by 33% reduction), indicating the series can be better described.
IMF2 0.2978 IMF9 0.0395
IMF3 1.3564 IMF10 0.0087 The SE of the error series is 2.0485, which is slightly lower than the
IMF4 1.2976 IMF11 0.0048 original wind speed series, but much larger than its sub-series. It shows
IMF5 0.9630 Residual 0.0000 that the error series is as random and complex as the original wind
IMF6 0.5601 speed series. Therefore, the VMD algorithm is adopted to decompose
the error series. The error series is analyzed by the center frequency
method to determine the number of VMD components. Fig. 7 shows the
MAPE1 MAPE2
PMAPE = decomposition result of the error series, and Table 3 shows the corre-
MAPE1 (26)
sponding sample entropy after VMD decomposition. The results show
that the sample entropy of error sub-series is much lower than the
original error series. Therefore, it is shown that CEEMDAN and VMD
4.3. Process analysis
can effectively reduce the complexity of the original wind speed series
and the error series, respectively.
In the proposed model, the original wind speed series and error
series are respectively decomposed by CEEMDAN and VMD algorithms.
The sample entropy (SE) is applied to illustrate the complexity of the

7
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 7. VMD processing results for error series of S2.

Table 3 Table 4
Sample entropy of the error series after VMD decomposition. Error evaluation results of different models for S1.
Parameters SE Parameters SE Models RMSE (m/s) MAPE (%) MAE (m/s)

Error 2.0485 MODE6 0.8934 Persistence 0.9079 21.82% 0.6317

MODE1 0.1651 MODE7 0.9751 MLP 0.8858 23.42% 0.6232
MODE2 0.3721 MODE8 0.6512 SVR 0.8784 22.11% 0.6081
MODE3 0.6699 MODE9 0.3967 LSTM 0.8729 20.28% 0.5924
MODE4 0.9712 MODE10 0.1204 EEMD-LSTM 0.5212 14.83% 0.4287
MODE5 0.9870 CEEMDAN-LSTM 0.3427 10.52% 0.2570
EEMD-error-LSTM 0.3992 9.28% 0.2791
CEEMDAN-error-LSTM 0.3063 8.50% 0.2286
EEMD-error-VMD-LSTM 0.1568 3.63% 0.1050
5. Results and discussions
CEEMDAN-error-VMD-LSTM 0.1110 3.21% 0.0831

To verify the prediction performance of the proposed model, four

wind speed data were tested. In each experiment, the proposed hybrid (MLP) model, the support vector regression (SVR) model, the LSTM
model (CEEMDAN-error-VMD-LSTM model) as well as nine other pre- model, the EEMD-LSTM model, the CEEMDAN-LSTM model, the EEMD-
diction models is applied to the four wind speed series. The involved error-LSTM model, the CEEMDAN-error-LSTM model, the EEMD-error-
models consist of the persistence model, the multi-layer perceptron VMD-LSTM model and the CEEMDAN-error-VMD-LSTM model. The

8
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 8. Comparison of foresting results of proposed model and other nine models (S1).

Fig. 9. Forecasting results of proposed model and other related models for S1.

9
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 10. Scatter plot of forecast and observed values of ten models for S1.

Table 5 ahead prediction was applied in this paper.

Error evaluation results of different models for S2.
Models RMSE (m/s) MAPE (%) MAE (m/s)
5.1. Forecasting results
Persistence 1.1172 25.43% 0.7423
MLP 1.0654 26.63% 0.7208 The prediction errors including RMSE, MAPE and MAE of the total
SVR 1.0823 26.28% 0.7252 ten models for S1 are presented in Table 4 and Fig. 8. The forecasting
LSTM 1.0186 24.80% 0.6960
EEMD-LSTM 0.5812 15.18% 0.3779
results of the aforementioned models for S1 case have been compared
CEEMDAN-LSTM 0.5031 11.99% 0.3107 and discussed. The detailed comparison results of S1 case are sum-
EEMD-error-LSTM 0.6092 14.61% 0.3892 marized below.
CEEMDAN-error-LSTM 0.5201 11.50% 0.3142
EEMD-error-VMD-LSTM 0.2161 5.21% 0.1391
(1) The LSTM model has lower error than the other single models,
CEEMDAN-error-VMD-LSTM 0.1855 4.85% 0.1238
especially the persistence model, indicating that the LSTM model is
better for time series processing than other single models. For ex-
Table 6 ample, the MAE of the LSTM model for S1 is 0.5924 m/s while those
Error evaluation results of different models for S3. of the Persistence, MLP and SVR models are 0.6317 m/s, 0.6232 m/
s, and 0.6081 m/s, respectively.
Models RMSE (m/s) MAPE (%) MAE (m/s)
(2) Comparing the results of other single models, the performance of
Persistence 1.5664 30.93% 1.1593 the hybrid model is better than the single models. For example, the
MLP 1.5099 30.81% 1.0999 best RMSE of single models is 0.8729 m/s in the S1 case, and the
SVR 1.5262 29.93% 1.1105 worst RMSE in the hybrid model has an RMSE of 0.5212 m/s. The
LSTM 1.4745 29.45% 1.0786
error is reduced by 67.42%, which is much lower than the single
EEMD-LSTM 0.8534 20.10% 0.6367
CEEMDAN-LSTM 0.7534 14.23% 0.5630 model.
EEMD-error-LSTM 0.8072 17.23% 0.5984 (3) Compared with the EEMD-error-LSTM and the CEEMDAN-error-
CEEMDAN-error-LSTM 0.6676 13.36% 0.4828 LSTM model, the models with error series decomposition have
EEMD-error-VMD-LSTM 0.2364 5.25% 0.1776 higher performance. This indicates that the error series decom-
CEEMDAN-error-VMD-LSTM 0.2250 4.45% 0.1660
position is conducive to raising the prediction accuracy of wind
speed.
(4) The CEEMDAN-error-VMD-LSTM model proposed in this paper has
Table 7
Error evaluation results of different models for S4.
the lowest error and the best prediction performance among the ten
models. It can be concluded that the double decomposition, deep
Models RMSE (m/s) MAPE (%) MAE (m/s) learning and error correction strategies are helpful to improve the
Persistence 1.7314 32.92% 1.1747 prediction accuracy.
MLP 1.6444 32.54% 1.1190
SVR 1.6181 31.29% 1.1032 To further exhibit the forecasting performance of different models,
LSTM 1.5821 29.17% 1.0843 the forecasting results of the all models for Dataset S1 have been illu-
EEMD-LSTM 1.0616 20.75% 0.6870
strated in Figs. 9 and 10. From the Fig. 10, it can be seen that the scatter
CEEMDAN-LSTM 0.8652 16.38% 0.5597
EEMD-error-LSTM 1.1048 21.42% 0.7153 points of the CEEMDAN-error-VMD-LSTM model distribute the most
CEEMDAN-error-LSTM 0.9549 17.86% 0.6240 uniformly around the regression line and also are the closest to the line,
EEMD-error-VMD-LSTM 0.4931 8.64% 0.2992 which demonstrates that the proposed model can produce a better
CEEMDAN-error-VMD-LSTM 0.2937 5.38% 0.1925
forecasting result compared with the other methods.
To further verify the performance of the proposed model, the ten
models developed in this study have been applied to datasets S2, S3 and
original wind speed series, the error series and their sub-series are
S4. The error indexes of the all models for are showed in Tables 5, 6, 7
predicted by the LSTM network in the above non-single models. For
and Figs. 11, 14, 17, respectively. Moreover, the Figs. 12, 15 and 18
better comparison of model performance, approximate parameters were
intuitively show the forecast effect of the proposed model and the
set during the test. The model input series length was 7, and one-step-
comparison results of the CEEMDAN-related models for three cases.

10
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 11. Comparison of foresting results of proposed model and other nine models (S2).

And the corresponding figures to illustrate the forecasting performance the error series can improve the prediction accuracy and model ap-
of the ten models are presented in Figs. 13, 16 and 19 for datasets S2, S3 plicability.
and S4, respectively. The conclusions obtained for other three datasets It can also be seen from Table 8 that the EEMD-error-LSTM model
are similar to those of datasets S1. It can be found from the above fig- has different performances in the S1 and S3 cases as in the S2 and S4
ures and Tables that the CEEMDAN-error-VMD-LSTM models demon- cases. In the S1 and S3 cases, the prediction accuracy of the EEMD-error-
strate the best performance while the Persistence model exhibit the LSTM model is better than the EEMD-LSTM model, but the EEMD-error-
worst. Compared with the model without series decomposition, the LSTM model is lower than that of the EEMD-LSTM model in the S2 and
model with double decomposition technology can significantly reduce S4 cases. The reason for this problem is that the error series inherits the
RMSE, MAPE and MAE. randomness and complexity of the original wind speed series, making
the series difficult to describe directly, thus reducing the performance
5.2. Improvements of proposed hybrid model of the model.
It can be shown from Table 9 that the CEEMDAN-LSTM model
To further compare the performance difference between models, the predicts better performance than the LSTM model. In the four cases,
effectiveness of the proposed model is demonstrated by the promoting compared to the LSTM model, the RMSE of the CEEMDAN-LSTM model
percentage of between models. Table 8 lists the comparison results of are reduced by 60.74%, 50.61%, 48.90% and 45.32%, respectively. It
the EEMD-related models. Table 9 lists the comparison results of the indicates that the CEEMDAN algorithm can effectively reduce the
CEEMDAN-related models. Table 10 lists the performance comparison complexity of the time series and greatly improve the predictability of
results of similar models of EEMD and CEEMDAN. The above tables the series.
contain the results of four cases, and the reliability of the conclusions is The CEEMDAN-error-VMD-LSTM model has better prediction ac-
enhanced through comparison of multiple cases. curacy than the CEEMDAN-LSTM model and the CEEMDAN-error-LSTM
It can be seen from Table 8 that the prediction accuracy of the model. In four cases, the MAPE of the CEMMD-error-VMD-LSTM model
EEMD-LSTM model is better than the single LSTM models. For four is at least 59.53% and 57.80% lower than the CEEMDAN-LSTM model
cases, compared to the LSTM model, the RMSE of the EEMD-LSTM and the CEEMDAN-error-LSTM model, respectively. It shows that the
model are reduced by 40.29%, 42.94%, 42.12% and 32.90%, respec- proposed model greatly reduces the error and improves the prediction
tively. It indicates that the EEMD can effectively reduce the prediction performance of the model.
error. As can be seen from Table 10, all the promoting performance in-
The prediction performance of the EEMD-error-VMD-LSTM model is dexes of error are positive. The phenomenon indicates that the
better than the EEMD-LSTM model and the EEMD-error-LSTM model. CEEMDAN-related model is superior to the same type of EEMD-related
For instance, in S3 case, the MAPE of the EEMD-error-VMD-LSTM model model, indicating that the CEEMDAN algorithm has better decom-
are reduced by 70.71% and 72.30%, respectively; the MAE are reduced position performance than the EEMD for wind speed series. In the four
by 69.54% and 73.89%, respectively; and the RMSE are reduced by cases, the prediction performance of the CEEMDAN-error-VMD-LSTM
70.31% and 72.10%, respectively. In the four cases, the prediction ac- model is significantly better than the EEMD-error-VMD-LSTM model. In
curacy of the EEMD-error-VMD-LSTM model is better than the EEMD- case S4, the reductions of RMSE, MAPE and MAE are 40.44%, 37.72%
error-LSTM model. This phenomenon shows that using VMD to process and 35.66%, respectively.

11
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 12. Forecasting results of proposed model and other related models for S2.

Fig. 13. Scatter plot of forecast and observed values of ten models for S2.

Through analysis of four cases, it is found that the CEEMDAN-error- However, the above four cases show that the VMD decomposition of
VMD-LSTM model proposed in this paper always has the lowest error in the error series can greatly reduce the error of the model. It shows that
all the above wind speed prediction models. The prediction accuracy of the VMD algorithm can effectively reduce the randomness and com-
the EEMD-error-LSTM model and the CEEMDAN-error-LSTM model is plexity of the error series, so that LSTM can better predict the error
sometimes lower than EEMD-LSTM model and CEEMDAN-LSTM modes, series and finally better correct the prediction results.
respectively. It shows that the error series generated by CEEMDAN and Wind power prediction can improve the stability of grid connection.
EEMD is highly unpredictable. Therefore, directly using error series The common method of wind power prediction is based on the wind
correction cannot improve the performance of the model stably and speed prediction and the wind power curve. The wind power curve is
effectively. applied to calculate the wind power, which the wind speed prediction

12
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 14. Comparison of foresting results of proposed model and other nine models (S3).

Fig. 15. Forecasting results of proposed model and other related models for S3.

13
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 16. Scatter plot of forecast and observed values of ten models for S3.

Fig. 17. Comparison of foresting results of proposed model and other nine models (S4).

value is the input. The model proposed in this paper can provide ac- (1) The hybrid models have much higher prediction accuracy than the
curate wind speed prediction and provide an important basis for wind single models. The CEEMDAN-LSTM model and the EEMD-LSTM
power prediction. model are superior to the single LSTM model, which indicates that
the decomposition strategy is an effective means to improve pre-
diction performance of wind speed.
6. Conclusions (2) The double decomposition strategy is better than the single de-
composition. Considering series decomposition in error correction
To improve the energy conversion efficiency, it is of great important models, the models have higher forecasting accuracy than the
to improve the accuracy of wind speed prediction. In this study, a new models without error series decomposition. In particular, the
hybrid model base on error correction, double decomposition and deep CEEMDAN-error-VMD-LSTM model reduced the MAPE by 62.20%,
learning is proposed. In the proposed CEEMDAN-error-VMD-LSTM 57.80%, 66.66% and 69.88% than the CEEMDAN-error-LSTM
model, the CEEMDAN and VMD are applied to decompose the original model in four experimental cases.
wind speed series and error series respectively, and the LSTM is adopted (3) Among all the models involved, the CEEMDAN-error-VMD-LSTM
to predict each sub-series. The proposed model is verified by four wind model has the best predictive performance. In proposed CEEMDAN-
speed cases and compared with other nine models. It can be concluded error-VMD-LSTM model, the VMD algorithm can effectively reduce
that:

14
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Fig. 18. Forecasting results of proposed model and other related models for S4.

Fig. 19. Scatter plot of forecast and observed values of ten models for S4.

the randomness and complexity of the error series, so that LSTM wind speed forecasting, the factors such as humidity, pressure and
can better predict the error series and finally correct the prediction temperature have great influence on wind speed. Therefore, future
results. studies should consider these influencing factors and incorporate them
into the wind speed prediction.
Overall, the proposed hybrid model can provide more reliable and
accurate short wind speed forecasting for wind power generation sys- CRediT authorship contribution statement
tems, thus enhancing the security and stability of the power system. On
the other hand, accurate wind speed prediction can improve the flex- Zherui Ma: Conceptualization, Methodology, Writing - original
ibility of wind turbine maintenance and operation arrangement. draft, Software. Hongwei Chen: Visualization, Data curation,
Although the proposed model has obvious advantages in short-term Investigation. Jiangjiang Wang: Supervision, Writing - review &

15
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

Table 8
Performance improvement percentage of among models about EEMD (S1–S4).
Models Indexes S1 S2 S3 S4

LSTM vs. EEMD-LSTM PRMSE 40.29% 42.94% 42.12% 32.90%

PMAPE 26.86% 38.81% 31.76% 28.88%
PMAE 27.64% 45.70% 40.97% 36.64%

EEMD-LSTM vs. EEMD-error-LSTM PRMSE 23.41% −4.82% 5.42% −4.06%

PMAPE 37.40% 3.74% 14.28% −3.24%
PMAE 34.89% −2.99% 6.02% −4.12%

EEMD-error-LSTM vs. EEMD-error-VMD-LSTM PRMSE 60.71% 64.52% 70.71% 55.36%

PMAPE 60.92% 64.37% 69.54% 59.68%
PMAE 62.36% 64.25% 70.31% 58.17%

EEMD-LSTM vs. EEMD-error-VMD-LSTM PRMSE 69.91% 62.81% 72.30% 53.55%

PMAPE 75.53% 65.71% 73.89% 58.37%
PMAE 75.50% 63.18% 72.10% 56.45%

Table 9
Performance improvement percentage of among models about CEEMDAN (S1–S4).
Models Indexes S1 S2 S3 S4

LSTM vs. CEEMDAN-LSTM PRMSE 60.74% 50.61% 48.90% 45.32%

PMAPE 48.10% 51.67% 51.69% 43.85%
PMAE 56.61% 55.36% 47.80% 48.39%

CEEMDAN-LSTM vs. CEEMDAN-error-LSTM PRMSE 10.60% −3.37% 11.39% −10.37%

PMAPE 19.21% 4.10% 6.10% −9.03%
PMAE 11.08% −1.11% 14.25% −11.50%

CEEMDAN-error-LSTM vs. CEEMDAN-error-VMD-LSTM PRMSE 63.78% 64.34% 66.30% 69.24%

PMAPE 62.20% 57.80% 66.66% 69.88%
PMAE 63.63% 60.58% 65.61% 69.15%

CEEMDAN-LSTM vs. CEEMDAN-error-VMD-LSTM PRMSE 67.62% 63.14% 70.14% 66.05%

PMAPE 69.47% 59.53% 68.69% 67.16%
PMAE 67.66% 60.14% 70.51% 65.60%

Table 10
Performance improvement percentage of among models about CEEMDAN and EEMD (S1–S4).
Models Indexes S1 S2 S3 S4

EEMD-LSTM vs. CEEMDAN-LSTM PRMSE 34.25% 13.44% 11.72% 18.51%

PMAPE 29.03% 21.02% 29.21% 21.06%
PMAE 40.04% 17.78% 11.57% 18.54%

EEMD-error-LSTM vs. CEEMDAN-error-LSTM PRMSE 23.26% 14.63% 17.30% 13.57%

PMAPE 8.41% 21.31% 22.45% 16.63%
PMAE 18.10% 19.27% 19.32% 12.76%

EEMD-error-LSTM vs. CEEMDAN-error-VMD-LSTM PRMSE 29.25% 14.18% 4.83% 40.44%

PMAPE 11.43% 6.79% 15.10% 37.72%
PMAE 20.85% 11.00% 6.55% 35.66%

editing. Xin Yang: Conceptualization, Project administration. Rujing References

Yan: Validation. Jiandong Jia: Software. Wenliang Xu: Writing - re-
view & editing. [1] Jung C, Taubert D, Schindler D. The temporal variability of global wind energy –
long-term trends and inter-annual variability. Energy Convers Manag
2019;188:462–72. https://doi.org/10.1016/j.enconman.2019.03.072.
Declaration of Competing Interest [2] Foley AM, Leahy PG, Marvuglia A, McKeogh EJ. Current methods and advances in
forecasting of wind power generation. Renewable Energy 2012;37(1):1–8. https://
doi.org/10.1016/j.renene.2011.05.033.
The authors declare that they have no known competing financial
[3] Qian Z, Pei Y, Zareipour H, Chen N. A review and discussion of decomposition-
interests or personal relationships that could have appeared to influ- based hybrid models for wind energy forecasting applications. Appl Energy
ence the work reported in this paper. 2019;235:939–53. https://doi.org/10.1016/j.apenergy.2018.10.080.
[4] Liu H, Chen C, Lv X, Wu X, Liu M. Deterministic wind energy forecasting: a review
of intelligent predictors and auxiliary methods. Energy Convers Manag
Acknowledgement 2019;195:328–45. https://doi.org/10.1016/j.enconman.2019.05.020.
[5] Tascikaraoglu A, Uzunoglu M. A review of combined approaches for prediction of
short-term wind speed and power. Renew Sust Energy Rev 2014;34:243–54.
This project is supported by the Fundamental Research Funds for https://doi.org/10.1016/j.rser.2014.03.033.
the Central Universities of China (Grant No. 2018QN085) and the au- [6] Ahmed A, Khalid M. An intelligent framework for short-term multi-step wind speed
thors acknowledge the good suggestions by the anonymous reviewers forecasting based on Functional Networks. Appl Energy 2018;225:902–11. https://
doi.org/10.1016/j.apenergy.2018.04.101.
that improved the paper.

16
Z. Ma, et al. Energy Conversion and Management 205 (2020) 112345

[7] Wang SX, Zhang XY, Ge LJ, Wu L. 2-D wind speed statistical model for reliability [27] Zhou J, Liu H, Xu Y, Jiang W. A hybrid framework for short term multi-step wind
assessment of microgrid. IEEE Trans Sustain Energy 2016;7:1159–69. https://doi. speed forecasting based on variational model decomposition and convolutional
org/10.1109/tste.2015.2512608. neural network. Energies 2018;11:2292. https://doi.org/10.3390/en11092292.
[8] Noorollahi Y, Jokar MA, Kalhor A. Using artificial neural networks for temporal and [28] Wang H, Wang G, Li G, Peng J, Liu Y. Deep belief network based deterministic and
spatial wind speed forecasting in Iran. Energy Convers Manag 2016;115:17–25. probabilistic wind speed forecasting approach. Appl Energy 2016;182:80–93.
https://doi.org/10.1016/j.enconman.2016.02.041. https://doi.org/10.1016/j.apenergy.2016.08.108.
[9] Santamaría-Bonfil G, Reyes-Ballesteros A, Gershenson C. Wind speed forecasting for [29] Liu H, Mi X, Li Y. Wind speed forecasting method based on deep learning strategy
wind farms: a method based on support vector regression. Renew Energy using empirical wavelet transform, long short term memory neural network and
2016;85:790–809. https://doi.org/10.1016/j.renene.2015.07.004. Elman neural network. Energy Convers Manag 2018;156:498–514. https://doi.org/
[10] Liu H, Mi X, Li Y. An experimental investigation of three new hybrid wind speed 10.1016/j.enconman.2017.11.053.
forecasting models using multi-decomposing strategy and ELM algorithm. Renew [30] Pearre NS, Swan LG. Statistical approach for improved wind speed forecasting for
Energy 2018;123:694–705. https://doi.org/10.1016/j.renene.2018.02.092. wind power production. Sustain Energy Technol Assess 2018;27:180–91. https://
[11] Hu Q, Zhang R, Zhou Y. Transfer learning for short-term wind speed prediction with doi.org/10.1016/j.seta.2018.04.010.
deep neural networks. Renew Energy 2016;85:83–95. https://doi.org/10.1016/j. [31] Wang J, Li Y. Multi-step ahead wind speed prediction based on optimal feature
renene.2015.06.034. extraction, long short term memory neural network and error correction strategy.
[12] Zhou J, Shi J, Li G. Fine tuning support vector machines for short-term wind speed Appl Energy 2018;230:429–43. https://doi.org/10.1016/j.apenergy.2018.08.114.
forecasting. Energy Convers Manage 2011;52(4):1990–8. https://doi.org/10.1016/ [32] Liang Z, Liang J, Wang C, Dong X, Miao X. Short-term wind power combined
j.enconman.2010.11.007. forecasting based on error forecast correction. Energy Convers Manag
[13] Yang H, Chen Y. Hybrid deep learning and empirical mode decomposition model 2016;119:215–26. https://doi.org/10.1016/j.enconman.2016.04.036.
for time series applications. Expert Syst Appl 120 2019;128–38. https://doi.org/10. [33] Huang Y, Yang L, Liu S, Wang G. Multi-step wind speed forecasting based on en-
1016/j.eswa.2018.11.019. semble empirical mode decomposition, long short term memory network and error
[14] Zhang Z, Qin H, Liu Y, Wang Y, Yao L, Li Q, et al. Long short-term memory network correction strategy. Energies 2019;12:22. https://doi.org/10.3390/en12101822.
based on neighborhood gates for processing complex causality in wind speed pre- [34] Li Y, Wu H, Liu H. Multi-step wind speed forecasting using EWT decomposition,
diction. Energy Convers Manag 2019;192:37–51. https://doi.org/10.1016/j. LSTM principal computing, RELM subordinate computing and IEWT reconstruction.
enconman.2019.04.006. Energy Convers Manage 2018;167:203–19. https://doi.org/10.1016/j.enconman.
[15] Wang K, Qi X, Liu H, Song J. Deep belief network based k-means cluster approach 2018.04.082.
for short-term wind power forecasting. Energy 2018;165:840–52. https://doi.org/ [35] Mi X, Liu H, Li Y. Wind speed forecasting method using wavelet, extreme learning
10.1016/j.energy.2018.09.118. machine and outlier correction algorithm. Energy Convers Manag
[16] Chen Y, Zhang S, Zhang W, Peng J, Cai Y. Multifactor spatio-temporal correlation 2017;151:709–22. https://doi.org/10.1016/j.enconman.2017.09.034.
model based on a combination of convolutional neural network and long short-term [36] Huang NE, Shen Z, Long SR, Wu MC, Shih HH, Zheng Q, et al. The empirical mode
memory neural network for wind speed forecasting. Energy Convers Manag decomposition and the hilbert spectrum for nonlinear and non-stationary time
2019;185:783–99. https://doi.org/10.1016/j.enconman.2019.02. 018. series analysis. Proc: Mathem, Phys Eng Sci: R Soc 1998:903–95.
[17] Chen J, Zeng G-Q, Zhou W, Du W, Lu K-D. Wind speed forecasting using nonlinear- [37] Zhaohua W, Huang NE. Ensemble empirical mode decomposition: a noise-assisted
learning ensemble of deep learning time series prediction and extremal optimiza- data analysis method. Adv Adapt Data Anal 2009;1:1–41. https://doi.org/10.1142/
tion. Energy Convers Manage 2018;165:681–95. https://doi.org/10.1016/j. S1793536909000047.
enconman.2018.03.098. [38] Torres ME, Colominas MA, Schlotthauer G, Flandrin P. A complete ensemble em-
[18] Hu Y, Chen L. A nonlinear hybrid wind speed forecasting model using LSTM net- pirical mode decomposition with adaptive noise. 2011 IEEE International
work, hysteretic ELM and differential evolution algorithm. Energy Convers Manag Conference on Acoustics, Speech and Signal Processing (ICASSP); 2011. p. 4144–7.
2018;173:123–42. https://doi.org/10.1016/j.enconman.2018.07.070. [39] Zhou Z, Lin L, Li S. International stock market contagion: a CEEMDAN wavelet
[19] Bokde N, Feijoo A, Villanueva D, Kulat K. A review on hybrid empirical mode de- analysis. Econ Model 2018;72:333–52. https://doi.org/10.1016/j. econmod. 2018.
composition models for wind speed and wind power prediction. Energies 02.010.
2019;12:254. https://doi.org/10.3390/en12020254. [40] Dragomiretskiy K, Zosso D. Variational mode decomposition. IEEE Trans Signal
[20] Wang C, Zhang H, Fan W, Ma P. A new chaotic time series hybrid prediction method Process 2014;62:531–44. https://doi.org/10.1109/TSP.2013.2288675.
of wind power based on EEMD-SE and full-parameters continued fraction. Energy [41] Wang Y, Markert R, Xiang J, Zheng W. Research on variational mode decomposition
2017;138:977–90. https://doi.org/10.1016/j.energy.2017.07.112. and its application in detecting rub-impact fault of the rotor system. Mech Syst
[21] Meng A, Ge J, Yin H, Chen S. Wind speed forecasting based on wavelet packet Signal Pr 2015;60–61:243–51. https://doi.org/10.1016/j.ymssp.2015.02.020.
decomposition and artificial neural networks trained by crisscross optimization [42] Zhang Y, Liu K, Qin L, An X. Deterministic and probabilistic interval prediction for
algorithm. Energy Convers Manag 2016;114:75–88. https://doi.org/10.1016/j. short-term wind power generation based on variational mode decomposition and
enconman.2016.02.013. machine learning methods. Energy Convers Manag 2016;112:208–19. https://doi.
[22] Du P, Wang J, Guo Z, Yang W. Research and application of a novel hybrid fore- org/10.1016/j.enconman.2016.01.023.
casting system based on multi-objective optimization for wind speed forecasting. [43] Jiang X, Wang J, Shi J, Shen C, Huang W, Zhu Z. A coarse-to-fine decomposing
Energy Convers Manag 2017;150:90–107. https://doi.org/10.1016/j.enconman. strategy of VMD for extraction of weak repetitive transients in fault diagnosis of
2017.17.065. rotating machines. Mech Syst Signal Pr 2019;116:668–92. https://doi.org/10.
[23] Niu D, Liang Y, Hong W-C. Wind speed forecasting based on EMD and GRNN op- 1016/j.ymssp.2018.07.014.
timized by FOA. Energies 2017;10:2001. https://doi.org/10.3390/en10122 001. [44] Balderas D, Ponce P, Molina A. Convolutional long short term memory deep neural
[24] Santhosh M, Venkaiah C, Kumar DMV. Ensemble empirical mode decomposition networks for image sequence prediction. Expert Syst Appl 2019;122:152–62.
based adaptive wavelet neural network method for wind speed prediction. Energy https://doi.org/10.1016/j.eswa.2018.12.055.
Convers Manag 2018;168:482–93. https://doi.org/10.1016/j.enconman.2018.04. [45] Kong W, Dong ZY, Jia Y, Hill DJ, Xu Y, Zhang Y. Short-term residential load fore-
099. casting based on LSTM recurrent neural network. IEEE Trans Smart Grid
[25] Zhang W, Qu Z, Zhang K, Mao W, Ma Y, Fan X. A combined model based on 2019;10:841–51. https://doi.org/10.1109/TSG.2017.2753802.
CEEMDAN and modified flower pollination algorithm for wind speed forecasting. [46] Han S, Qiao Y-h, Yan J, Liu Y-q, Li L, Wang Z. Mid-to-long term wind and photo-
Energy Convers Manag 2017;136:439–51. https://doi.org/10.1016/j.enconman. voltaic power generation prediction based on copula function and long short term
2017.01.022. memory network. Appl Energy 2019;239:181–91. https://doi.org/10.1016/j.
[26] Wang H, Lei Z, Zhang X, Zhou B, Peng J. A review of deep learning for renewable apenergy.2019.01.193.
energy forecasting. Energy Convers Manag 2019;198:111799https://doi.org/10. [47] Jager D, Andreas A. NREL National Wind Technology Center (NWTC): M2 Tower;
1016/j.enconman.2019.111799. Boulder, Colorado (Data). 1996. https://doi.org/10.5439/1052222.

17