Academic Journal of Science and Technology 
ISSN: 2771-3032 | Vol. 6, No. 3, 2023 

 

77 

Short	Term	Wind	Power	Prediction	Based	on	
CEEMDAN‐LSTM 

Congming Zhang, Zicheng Yang, Shaofei Gao 

Department of computer, North China Electric Power University, Baoding 071000, China 

 

Abstract: To improve the accuracy of wind power prediction, a wind power prediction method based on time series 
decomposition and error correction is proposed. Firstly, the maximum information coefficient (MIC) method was used to select 
the features that have strong correlation with wind power to reduce the complexity of the original data; Then, according to the 
non-stationary characteristics of wind power, complete ensemble empirical mode decomposition with adaptive 
noise(CEEMDAN) was used to decompose wind power into several stationary subsequences; Finally, the long short memory 
network (LSTM) was used to dynamically model the wind power multivariable time series; Add the predicted values of each 
subsequence to get the final predicted value. Combined with the measured data of a domestic wind farm, the simulation results 
showed that the proposed method had higher short-term wind power prediction accuracy compared with other prediction models. 

Keywords: Wind power prediction, CEEMDAN, Deep Learning, LSTM; MIC. 
 

1. Introduction 
By the end of 2020, China's installed wind power capacity 

was 281 million kilowatts, and the wind power generation 
capacity is 466.5 billion kilowatt hours, accounting for 6.1% 
of the total power generation [1]. Wind power gradually 
changes from 'auxiliary power supply' to 'main power supply' 
[2]. However, due to the time-varying characteristics of wind, 
wind power has strong randomness, intermittency and 
uncertainty [3]. Therefore, accurate prediction of wind power 
was of great significance to the safe and economic operation 
of power system [4]. 

Wind power time series data were non-stationary series, 
and it was difficult to capture their change characteristics and 
autocorrelation in time series. Therefore, the direct prediction 
of wind power data is not ideal[5]. In view of the random 
volatility of wind power data, the literature[6] used the 
Complex Ensemble Empirical Model Decomposition 
(CEEMD) to decompose the wind power sequence, and used 
the Satin Gardener algorithm to optimize the least squares 
support vector machine to predict the wind power. In 
reference[7], the serial information such as historical wind 
power was decomposed into modal components  of 
specified layers by using variational   mode 
decomposition(VMD).Different modal components represent 
features of different scales. Weight sharing gate recurrent unit 
(WSGRU) was used for modeling. Finally, ANN was used to 
correct and obtain the prediction results of wind power. In 
reference[8], for the low adaptability of VMD method, the 
Enhanced Colliding Bodies Optimization (ECBO) algorithm 
was used to optimize the core parameters of the variational 
mode. After decomposition, the wavelet kernel limit learning 
machine prediction model is used for prediction. 

In order to improve the accuracy of wind power prediction, 
a short-term wind power prediction method based on fully 
adaptive noise ensemble empirical decomposition 
(CEEMDAN) and long short memory network (LSTM) 
prediction was proposed in this paper. The CEEMDAN was 
used to decompose the wind power into several subsequences, 
establish the LSTM wind power prediction model of 
subsequences, and stack the predicted values of subsequences 

to obtain the final predicted value. The proposed method was 
compared with LSTM and EMD-LSTM to verify that the 
proposed model has higher accuracy. 

2. Data Preprocessing 

2.1. MIC feature selection 
Reshef et al. proposed the maximum information 

coefficient (MIC) method basedon mutual information [9]. 
Mutual information is a measure that describes the mutual 
relationship between two random variables x and y. The 
greater the mutual information value, the stronger the 
correlation between the two variables. Mutual information is 
defined as: 

( , )
( , ) ( , )log

( ) ( )x X y Y

p x y
I x y p x y

p x p y 

       (1) 

 
In the formula: p(x,y) is the joint probability density 

function of x, y; p(x) and p (y) are edge density functions of x 
and y, respectively. 

The maximum information coefficient is defined as follows: 
assuming that D(x,y) is a finite two-dimensional data set, the 
current two-dimensional space was divided into m intervals 
and n intervals in the x and y directions, respectively, to form 
a grid of m * n. There are multiple ways to divide grids with 
the same interval [9]. Under grids of different scales, the 
maximum mutual information value was selected for 
calculation, and the maximum information coefficient is 
defined as: 

( , )
( , ) max

log(min( , ))m n B

I x y
MIC x y

m n 
       (2) 

 
In the formula, m*n<B represents the constraint condition 

of the total number of grid divisions. It can be seen from 
literature[10] that the optimal effect can be achieved when it 
is set to the 0.6th power of the total amount of data. Therefore, 
this setting method was also used in this paper. 

2.2. Normalization 
When using multivariate time series to predict wind power, 



 

78 

the dimensions of different variables are different, and the 
numerical differences are also large [11]. In orderto make the 
prediction model consider the effect of variables on wind 
power equally, the wind speed, temperature,wind direction 
angle and active power in the wind power data are normalized 
to the [0,1] range. 

min

max min
norm

X X
X

X X





                   (3) 

 
In the formula: Xmin and Xmax are the minimum and 

maximum values of variables respectively. 

3. Prediction Model Analysis 

3.1. CEEMDAN principle 
CEEMDAN is based on EMD decomposition. In the 

process of decomposing wind power signals, it adaptively 
adds white noise to weaken the mode aliasing problem, and 
the decomposition process is characterized by integrity and 
almost no reconstruction error [12]. 

The CEEMDAN decomposition algorithm is implemented 
as follows: 

1) CEEMDAN decomposition algorithm makes use of 
EMD for signal X(t)+εi

0n(t)is decomposed repeatedly for N 
times, and the first modal component is obtained through 
mean value calculation: 

 

1 1
1

1
(t) ( )

N
i

i

IMF IMF t
N 

             (4) 

 
2) Calculate the first residual signal r1(t) decomposed by 

CEEMDAN as: 
 

1 1( ) ( ) ( )r t X t IMF t               (5) 

 
3) The signal 1 1 1( ) ( ( ))ir t E n t   is decomposed repeatedly 

for N times to obtain the second modal component: 
 

2 1 1 1 1
1

1
( ) ( ( ) ( ( )))

N
i

i

IMF t E r t E n t
N




         (6) 

 
4) For k=2,...,K, calculate the kth residual signal: 
 

1 ( )k k Kr r IM F t            (7) 

 
5) Repeat the calculation process in step 3) to obtain the 

k+1 modal function: 
 

1 1
1

1
( ) ( ( ) ( ( )))

N
i

k k k k
i

IMF t E r t E n t
N




            (8) 

 
6) Repeat steps 4) and 5) until the residual signal meets the 

termination condition of decomposition, and finally obtain k 
modal components. The final residual signal decomposed is: 

 

1

( ) ( ) ( )
k

k
k

R t X t IMF t


              (9) 

 
The original signal can be decomposed into: 

 

1

( ) ( ) ( )
k

K
k

X t R t IMF t


         (10) 

 

3.2. LSTM network model 
LSTM is a variant of the cyclic neural network, which can 

effectively capture the dependencies in the time series, and is 
suitable for modeling the wind power time series data[13]. Its 
basic unit structure is shown in Figure 1. 

 

 
Figure 1. LSTM basic unit structure diagram 

 
As shown in the figure above, the calculation formula is as 

follows: 
 

1( )t fx t fh t ff W x W h b               (11) 

 

1( )t ix t ih t ii W x W h b               (12) 

 

1( )t gx t gh t gg W x W h b            (13) 

 

1( )t ox t oh t oo W x W h b              (14) 

 

1t t t t tS g i S f � �             (15) 

 
( )t t th S o  �               (16) 

 
In the formula: ft、it、gt、ot、St、ht are the states of forgetting 

gate, input gate, input node, output gate, intermediate output 
and state unit respectively; s is the activation function 
sigmoid; Wfx、Wtx、Wgx、Wox are weight matrices; bf、bi、bg、
bo offset terms; F Is the change of tanh function. 

4. Wind Power Prediction Model 

4.1. Forecast process 
The flow of short-term wind power prediction method 

proposed in this paper is as follows: 
(1) MIC method was used to select the features of wind 

power data, and the features with strong correlation with wind 
power were selected to accelerate network training; 

(2) Based on CEEMDAN decomposition algorithm, wind 
power time series data was decomposed to obtain several 
stable sub sequences; 

(3) LSTM network prediction model is established for each 
subsequence. After training, the prediction results of each 



 

79 

subsequence were obtained, and the prediction values were 
obtained after superposition. 

The flow chart was shown in Figure 2: 
 

 
Figure 2. Algorithm flow chart 

4.2. Evaluation index 
In this paper, three evaluation indicators, mean absolute 

error (MAE), mean relative error (MRE) and root mean 
square error (RMSE), are used to analyze the wind power 
prediction results. The evaluation indicators are as follows: 

 

1

ˆn
t t

t N

p p
MAE

np


                (17) 

 

1

ˆn
t t

t t

p p
MRE

np


                (18) 

 

 2

1

ˆ1 n
t t

tN

p p
RMSE

p n


            (19) 

 
In the formula: n is the total number of forecasts; PN is the 

capacity of wind turbine; Pt  is the Actual power of wind 
turbine; ˆ tp  is the predictive value of wind turbine power. 

5. Example Analysis 

5.1. Data preparation 
The measured data of a domestic wind farm was used for 

simulation experiments. Two adjacent wind turbines in the 
wind farm are selected and divided into wind turbine A and 
wind turbine B. The data collected in the wind farm include 
wind turbine ID, wind speed, wind direction angle and 
ambient temperature, with a sampling period of 10 minutes. 

The 15 day continuous SCADA data of wind turbine in 
September 2015 was used to predict the wind power of wind 
turbine A on the 15th day. The raw data of wind turbine A was 
shown in Figure 3. 

 

 
Figure 3. Raw data 

 

5.2. MIC feature selection 
There are many types of parameters of wind power 

SCADA data setted recording units. If all dimensions of the 
data setted are input into the prediction model, a large number 
of parameters will be generated in the network model, and the 
training speed will be slow, affecting the prediction accuracy. 
The maximum information coefficient method was used for 
feature selection of wind turbine multivariable time seriesto 
reduce the data scale and complexity. 

 
Table 1. MIC feature selection results 

influence factor WS GS RS WD GT T 
MIC coefficient 0.982 0.965 0.971 0.250 0.810 0.143 

 
 

In Table 1, six variables with the strongest correlation with 
output power were selected by using the maximum 
information coefficient method for feature selection. WS 
represented wind speed; GS represented the generator speed; 
RS stands for impeller speed; WD represented wind direction 
angle; GT standed for gearbox temperature; T were the 
ambient temperature. 

5.3. CEEMDAN decomposition results 
In this paper, CEEMDAN algorithm was used to 

decompose the wind power time series data. The white noise 
group number NR=100, the noise standard deviation 
Nstd=0.05, and the maximum iteration number MaxIter=300 
were added. Nine subsequences were obtained from the 
decomposition shown in Figure 4. 
 



 

80 

 
Figure 4. CEEMDAN decomposition sequence 

 

5.4. Results 
Figure 5 showed the power change curve of wind turbine A 

predicted by each model on September 15, with a prediction 
resolution of 10 minutes and 144 data points in total. 

 

 
Figure 5. The effect of each prediction model 

 
The MAE, MRE and RMSE indexes of each model were 

calculated for evaluation and analysis, as shown in Table 3. 
 

Table 2. Evaluation index of each prediction model 
Model name MAE/% MRE/% RMSE/% 

RNN 7.94 8.99 9.96 
LSTM 7.13 8.63 9.37 

EMD-LSTM 6.94 7.44 8.21 
CEEMDAN-LSTM 5.54 5.88 5.53 
 
After training the model with the same data, it can be seen 

from Table 2 that the time series decomposition algorithm can 
effectively reduce the impact of wind power uncertainty. 
After decomposition, the prediction accuracy becomes higher. 
With LSTM as the benchmark model, CEEMDAN has a 
better prediction effect than EMD. MAE, MRE and RMSE 
have decreased by 1.40%, 1.56% and 2.68% respectively. The 
evaluation index showed that the method in this paper has 
higher prediction accuracy compared with LSTM, RNN and 
EMD-LSTM models. 

6. Conclusion 
A wind power prediction model based on CEEMDAN-

LSTM was proposed. Through the verification of the 
measured data of a domestic wind farm, the proposed model 
had a more accurate prediction effect. The main conclusions 
were as follows: 

(1) According to the fluctuation characteristics of wind 
power, CEEMDAN algorithm was used to decompose the 
wind power time series data into relatively stable 
subsequences, which can effectively improve the stability of 
prediction data and reduce the difficulty of prediction. 

(2) The forgetting gate in LSTM can update the effective 
information in the time series, enhanced the ability to capture 
the nonlinear dependencies in the wind power time series data, 
and had higher prediction accuracy than RNN. 

(3) The time series modeling method based on 
CEEMDAN-LSTM proposed in this paper had higher 
prediction accuracy, had some room for improvement, and 
can be applied to other time series prediction scenarios. 

References 
[1] National Renewable Energy Power Development Monitoring 

and Evaluation Report of the National Energy Administration 
in 2020 [2021-06-20] 
http://zfxxgk.nea.gov.cn/202106/20/c_1310039970.htm 



 

81 

[2] Sun Rongfu, Zhang Tao, et al. Overview of key technologies 
and applications of wind power prediction[J]. High Voltage 
Technology,2021,47(04):1129-1143. 

[3] Zhang Yachao, Liu Kaipei,et al. Study on Multi time Scale 
Source Load Coordinated Dispatching Model of Power System 
with Large Scale Wind Power[J]. High Voltage Technology, 
2019,45(02):600-608. 

[4] Tao Yubo, Chen Hao, et al. Short term wind power prediction 
concept, model and method[J]. Power Engineering Technology, 
2018, 37(05):7-13. 

[5] Zhao Dongmei, Du Gang, et al. Wind power combination 
prediction model based on time series decomposition and 
machine learning[J/OL]. Modern Power:1-11[2021-10-21] 

[6] Zhou Xiaolin, Tong Xiaoyang. Ultra short term wind power 
combination prediction based on CEEMD-SBO-LSSVR[J]. 
Power Grid Technology,2021,45(03):855-864. 

[7] Sheng Siqing, Jin Hang, et al. Short and medium term 
prediction of wind farm power generation based on VMD-
WSGRU[J/OL]. Grid technology:1-8[2021-10-27] 

[8] Li Qing, Zhang Xinyan, et al. Ultra short term multi-step 
prediction of wind power based on ECBO-VMD-
WKELM[J/OL]. Power Grid Technology,{3},{4}{5}:1-14 
[2021-07-14] 

[9] Reshef D N, Reshef Y A, et al. Detecting Novel Associations 
in Large Data Sets[J].Sicence,2011,334(6062):1518-1524. 

[10] Zheng Ruicheng, Gu Jie, et al. Research on input variable 
selection method of short-term load forecasting based on data 
driven and prediction error driven integration[J].Chinese 
Journal of Electrical Engineering,2020,40(02):487-500. 

[11] Power Grid Technology,2017,41 (12):3797-3802. 

[12] Shu Chang, Jin Xiao, et al. CEEMDAN based noise diagnosis 
method for distribution transformer discharge fault [J]. High 
Voltage Technology,2018,44(08):2603-2611. 

[13] Lu Jixiang, Zhang Qipei, et al. Short term load forecasting 
method based on CNN-LSTM hybrid neural network model[J]. 
Power System Automation, 2019,43(08):131-137.