You are on page 1of 10

IEEJ TRANSACTIONS ON ELECTRICAL AND ELECTRONIC ENGINEERING

IEEJ Trans 2019


Published online in Wiley Online Library (wileyonlinelibrary.com). DOI:10.1002/tee.22853

Paper

Prediction of Wind Power Generation Base on Neural Network


in Consideration of the Fault Time
Yunkun Cao, Non-member
Qingqing Hua , Non-member
Hui Shi, Non-member
Yanli Zhang, Non-member

Along with the rapid increase in wind power penetration into the power grid, wind power generation predicting is becoming
increasingly important to power system operators and electricity market participants. However, the random nature of the wind
power would increase the uncertainty of power systems. The influencing factor is one of the most important factors in the
quality of wind power prediction. In order to obtain a higher prediction accuracy, a two-stage prediction method combined with
meteorological factor and fault time is proposed. In the first stage, we present a detailed review of fault time of wind power
predicting with machine-learning methods and compare ten different models based on different influencing factors, and then,
the second stage is combined with the predicted results of first stage, coupled with meteorological factors for the final wind
power prediction. The results show good prediction accuracy in operation data of one wind farm from the Hu-Nan province with
different machine-learning algorithms. © 2019 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.

Keywords: wind power predictive models; support vector machines (SVM); machine learning; weather factors; fault time; radial basis function
(RBF) networks

Received 13 June 2018; Revised 5 August 2018

1. Introduction Statistics and machine learning methods to predict wind power


based on historical data, rather than physical models, simply add
With the growing popularity of the concept of ‘smart grid’,
complex variables to the model [6].
among various renewable resources, such as tidal, solar,
Statistical models include a variety of models, such as autore-
and geothermal energy, wind energy has the characteristics
gressive (AR), AR integrated moving average (ARIMA), and
of being clean, inexpensive, and indispensable [1,2]. Wind energy
exponential smoothing (ES) models [6]. Statistical models use his-
is an effective energy solution to a world that is threatened by an
torical data to tune the model parameters, and when the model
energy crisis and other related environmental disasters. However,
prediction result matches the historical data in a specific range, the
due to the intermittent and nonschedulable nature of wind energy,
minimum prediction error is obtained [13]. However, in the real
it causes great influences on and challenges to power system
situations, this requirement is rarely met because the main influ-
operation [3–5]. To make wind energy a reliable source of energy,
efficient and accurate models for monitoring and predicting of encing factors on wind speed are not always available or accurately
wind power are the most important and urgently required. measurable due to uncontrollable events, such as data transmission
At present, many methods have been proposed to predict wind errors or anemometer machine failures [14].
power generation, mainly including continuous prediction method, The statistical model of wind power predicting mainly uses
Kalman-filtering method, fuzzy logic method, random time series the method based on historical data [15,16]. Instead, the struc-
method, artificial neural network (ANN) method, spatial autocorre- ture of the machine learning model is to establish a relational
lation method, and support vector machines (SVM). The predicting network between the input data and the output data and learn
models widely used in wind power energy predictions can be the network through various algorithms [17–20]; the research
divided into three categories [6]: Physical, Statistical, and Machine direction of [21] is to promote the combination of model-driven
Learning-based models. and data-driven solutions. The machine learning model com-
The common physical modeling approach used in wind power bines a number of reasonable models together into a compre-
predicting is the numerical weather predict (NWP) model, which hensive approach that not only takes advantage of each single
utilizes a variety of weather data and operates by solving com- method but also avoids some drawbacks of each single approach
plex mathematical models [7–9]. In [10], the probabilistic wind [22,23].
power prediction models use meteorological integration achieved The influence of input data on various factors of wind power
through traditional wind-predicting time series. However, these generation, the Physical models, is more likely to follow certain
NWP errors may follow certain patterns and persist for a patterns, resulting in a larger error. Statistical models are more
long time [11]. efficient than physical models under certain conditions.
The Machine Learning model has the ability of autonomous
learning, which is not affected by the length of time. It has the
ability of learning, extracting features, and recognizing and can
a Correspondence to: Qingqing Hu. E-mail: huqq515151@163.com improve the accuracy of prediction. It is more efficient than the two
Shanghai Electric Power Design Institute Co., Ltd., No. 550, Xujiahui models due to the computational power of the machine learning
Road, 23rd Floor Shanghai, Huangpu District 200090, China mode [24].

© 2019 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.
Y. CAO ET AL.

Determination of influencing factors


10 neural network models with
different influence factors
The first stage of failure time prediction
RBF
Four influencing
Data
GRN factors were
preprocessing
N combined
Time series Forecast of
Original Input the fault model based on failure time
data time data BP neural for the next
network three days
Wind Atmospheric Tempe- Failure
speed pressure rature time

The second Stage of wind power generation


Neural network model of
wind power generation
Failure
time SVM
Wind speed GRNN
atmospheric pressure RBF
temperature
BP

Fig. 1. Flow chart of system

Above all, this paper will use the machine learning model to Hidden layer
predict the various experimental stages. X1
Input layer ω1
Output layer

2. Method and Framework


There are many factors that can affect wind power generation,
for example, many meteorological factors are mentioned in the
article [25] such as wind speed, temperature, and barometric
Xm
pressure. However, according to the actual operating experience
of wind farms and the situation, a wind turbine fault will affect u (k–1) ω 2

the speed of the fan and then affect the power generation.
Consequently, our purpose is to combine the meteorological factors X (k)
and wind generator time to predict the power generation. The
prediction is divided into two phases: Fault Time Prediction Based Xc (k)
on Neural Network and Time Series and using the generalized
regression neural network (GRNN) neural network for wind power
generation based on predicted fault time. There are two stages Receiver layer
of prediction using different neural network algorithms and the
specific flow chart, as shown in Fig. 1. Fig. 2. Elman neural network structure
Step 1 : Obtain a large number of original data from the wind
farm; model. The Elman neural network and BP neural network have
Step 2 : Collate the dataset and normalized the different dimen- the same multilayer feed-forward topology [26]. The difference is
sions of the data pretreatment; that the Elman neural network has a function of undertaking layer,
Step 3 : Select the required training set for fault time in the first which plays the role of delayed memory so that the Elman network
stage; has the ability to adapt to the incident characteristics, which can
Step 4 : The first prediction stage is carried out: time series-based directly reflect the dynamic process characteristics [27,28]. The
neural network for fault time prediction; Elman neural network is generally divided into four layers: input
Step 5 : According to the predicted fault time and the selected layer, hidden layer (intermediate layer), receiver layer, and output
factors, the training set of power generation prediction is deter- layer, as shown in Fig. 2 [28]. The input layer unit only acts as
mined. a signal transmission, and the output layer unit acts as a linear
Step 6 : Use SVM, GRNN neural network, and radial basis weighting. The transfer function of the hidden layer unit can be
function (RBF) for final power generation predict. linear or nonlinear, and the receiver layer is used to memorize the
Step 7 : Compare the mean square error (MSE) and mean output value of the hidden layer before the element and feedback
absolute error (MAE) values of the predicted results. to the input of the network. It can be regarded as a one-step delay
operator [28,29].
3. Processing Strategy The Elman neural network can store the output value of the
previous time and feedback to the input, so the output of the
Wind power generation prediction includes two stages; the first hidden layer can achieve the function of delayed storage [30].
stage is the prediction of the fault time of the fan. The Elman The structure of the Elman neural network has the ability to adapt
neural network and the time method based on the Error Back to the characteristics of the incident and can directly reflect the
Propagation (BP) neural network are used. The first stage is the dynamic characteristics of the process [31]. Nonlinear State Space
basis of the second stage. Combined with the characteristic data Expressions of Elman Networks:
and the results of the first stage, we can start the second stage of
the power generation prediction by using SVM, GRNN, and RBF, y(k ) = g(ω3 X (k )) (1)
three neural network models. In this chapter, each algorithm used X (k ) = f (ω1 Xc (k )) + ω2 (U (k − 1)) (2)
in the two stages is described in detailed.
XC (k ) = X (k − 1) (3)

3.1. Elman neural network The Elman recurrent neu- In the formula, y is a linear combination of hidden layer output
ral network, proposed by Elman, is a partial recurrent network of m-dimension; x is the n-dimensional hidden layer node unit

2 IEEJ Trans (2019)


NEURAL NETWORK, FAULT TIME, WIND POWER PREDICTION

vector; u is the input vector of R dimension; X C is the n- 3.3.1. Support vector machines SVM are a kind of super-
dimensional feedback state vector; ω1 , ω2 , and ω3 , respectively, vised learning models in which data need to be divided into
are the undertaking layer to the middle layer, the input layer to the two or more different categories. SVM was originally developed
middle layer, and the middle layer to the output layer weights; g(*) for solving classification problems in which data need to be divided
is the transfer function for output neurons; is a linear combination into two or more different categories and has been used to solve
of hidden layer output; f (*) is the transfer function of the hidden the regression problem (Ren yi, Yuhui) [31]. Using SVM to solve
layer neurons; and the s function is often used. the regression problem is called support vector machine regres-
Elman neural network also uses the BP algorithm to modify the sion, which has been widely used in regression problems such as
weights and learning index function using the error square sum wind energy [28] and solar energy prediction [33,34].
function [28]: SVR is a linear function of f (x ) = ωT x + b fitted to the map
to hyperspace φ(*) within the tolerance margin ε while allowing

m
E (ω) = [Yk (ω) − Y 2
(4) some slack ξ (*) at the same time [32]:
K (ω)]
k =1 1 
min  ω2 + C (ξi + ξi∗ ) (5)
which is the target output vector. 2
i
The Elman model determines how the input and output matrices
subject to:
refer to the Algorithm 1 in the next section.
yi − ωT xi − b ≤ ε + ξi
3.2. Time series model based on BP NN The time
series prediction method is a statistical method based on dynamic ωT xi + b − yi ≤ ε + ξi∗
data to demonstrate the dynamic structure and regularity of the
system. From a statistical sense, the so called time series refers to
a different value in a different time in accordance with the time ξi∗ ≥ 0
sequence [29]. where C is the penalty factor, that is, the more tolerance of the
The time series method based on neural network avoids the error, the higher the value, indicating the possibility the error may
tendency of the time series method to predict the value of future be smaller.
time only by the trend of time variation. Using historical data as In order to solve the minimum value in (5), Lagrange multi-
the input of the neural network, it can make use of the ability plier method and Karnsh-Kuhn-Tucker (KKT) conditions can be
of self-learning of neural network and study the relationship from applied. Those calculations will not be described in detail here.
large amount of historical data, and it can better predict the actual Ultimately, the regression function f (x ) = ωT x + b can be
problem. BP is one of the most widely used neural network models; rewritten as:
its structure is simple, and the simulation ability is strong and 
easy to realize [32], so in the first phase of the prediction, we f (x ) = (αi − αi∗ )xiT x + b (6)
use BP neural network combined with time series to predict the i
fault time. Due to the fault time, data are arranged according The complexity of f (x ) does not depend on the dimension of x
to time, and neural network training requires the input matrix but only when(αi − αi∗ ) = 0 as set of support vectors x i .
and the corresponding output matrix; so, when using the method In addition, according to Mercer’s theorem, if kernel function
of combining BP with time series, first determine the training K (·) is symmetric and positively semidefinite and K (·)without the
sample set. knowledge of the mapping function φ(·), the inner product xiT x
For a time series with n values in all, the values of the previous can be replaced by the kernel function K (·) [36].
3 days are used to predict the value of the fourth day. Next, a SVM commonly uses the kernel function as the RBF, and it is
detailed method of determining the input matrix and the output also used in experiments in this paper. The RBF equation is as
matrix will be given (we are going to use the previous 3 days to follows [34]:
predict the value of the fourth day):  
 x − xi 2
K (x , xi ) = exp − = exp(γ  x − xi 2 ) (7)
Algorithm 1. Determination of input matrix 2σ 2
/*The value of the first 3 days is taken as input P = [a 1 , a 2 , a 3 ], where x is the testing data, x i is a support vector, and σ is a
and the value of the fourth day is predicted as output Y = a 4 . tunable kernel parameter.
Similarly, to predict the 5-day fault time, use the actual value of SVR parameter optimization selection can give more accurate
the second day to the fourth day of the prediction and so on*/ prediction results. The most commonly used method of interna-
Parameters: tional recognition is to define c and g in a certain range of values,
a—Fault time of the actual data or normalized data where c is a parameter of epsilon-SVR and g is gamma in kernel
n—The total amount of fault data function [38,39]. The original data as the training set using K-CV
P—Input matrix method to get in this group c and g training regression accuracy,
Y—Output matrix and ultimately take the training set to predict the highest accuracy
1) for i = 1:n of the set of c and g as the best parameters.
2) P(i,:) = [a(i),a(i + 1),a(i + 2)];
3) Y(i) = a(3 + i) 3.3.2. Generalized regression neural network The
4) end GRNN is a highly parallel radial basis network with strong
So, we get the input matrix P is a 3*n scale, and the output nonlinear mapping capability and flexible network structure,
matrix is a 1*a scale. high degree of fault tolerance, and robustness, which is suitable
for solving nonlinear problems.
The remarkable characteristic of the GRNN is that the number
3.3. Wind power generation model The prediction of neurons in the hidden layer and output layer is the same as the
of wind power generation is the second stage of this paper. We number of input sample vectors; GRNN has a middle layer and
use SVM, GRNN, and RBF to establish the model. summation layer.

3 IEEJ Trans (2019)


Y. CAO ET AL.

The calculation formula for GRNN’s summation layer is: GRNN VS RBF-comparison of 10 model accuracy
n   60
(X − Xi )T (X − Xi ) Prediction error of the GRNN model
SDj = Yi exp − (8) Prediction error of the RBF model
2σ 2
i =1 50

Prediction error MSE value



n  
(X − Xi )T (X − Xi )
SDj = Yi exp − (9) 40
2σ 2
i =1

The application of the GRNN model in MATLAB has an 30


adjustable parameter speed, which indicates the approximation
degree of smoothing function. Using of K-cross validation and
20
an iterative method to find the optimal speed value minimizes the
prediction error.
In the second stage of the wind power generation, we use the 10
BP and RBF to perform a comparison. In the first stage BP is
combined with time series predicting failure time prediction, and
0
RBF is used as the kernel function of SVM, so it is not described
1 2 3 4 5 6 7 8 9 10
here in detail.
Model number

4. Case Analysis Fig. 3. GRNN VS RBF-comparison of 10 model accuracy

In the following, the details of the two-stage wind power


generation prediction will be presented; the performance of these Table I. Relative variance of RBF and GRNN
predictive models has been evaluated using the performance
metrics MAE and MSE, which are defined as follows: Model number 1 2 3 4 5

1 
N
| yn − yn | RBF/MSE 25.742 23.559 14.007 4.5674 65.951
EMAE = (10) GRNN/MSE 58.776 58.953 58.647 58.476 69.548
N yn
n=1 Model number 6 7 8 9 10
N 
 2 RBF/MSE 51.876 46.408 63.647 59.003 71.229
1 yn − yn
EMSE = (11) GRNN/MSE 69.100 68.618 68.845 68.424 68.740
N yn
n=1

where yn is the predicted value of fault time or wind power


generation, y n is the corresponding actual value, and n is the framework is the same as in the model 4 has the smallest MSE
sample size. value. The same conclusion can be drawn when compared with
the detailed data in Table I.
Model 4 combines wind speed, temperature, atmospheric pres-
4.1. Predict data analysis This article is based on a
sure, and the fault time; according to the above results, not only
large amount of original data generated by the actual operation
the meteorological factors of wind power generation but also fault
of a wind farm in the Hunan province. The time period of the data
time to a certain extent can affect wind power. Therefore, fault
is from February 2015 to November 2016. The main task of this
time of equipment cannot be ignored in order to obtain higher
article is to improve the prediction accuracy of wind power and to
prediction accuracy. The required characteristics of the prediction
extend the prediction time. The prediction time range is 48 or 72 h
are the prerequisites and basis for wind power prediction.
depending on the system operation.
As mentioned above, we have demonstrated the impact of
fault time on wind power forecasting results. There are four 4.2. Fault time prediction Once the fan is in trouble,
factors that may have impact on the power prediction: wind after the disappearance of the fault of fan, the fan has to reach a
speed, temperature, atmosphere pressure, and fault time. In order certain speed in order to produce power generation. As a result,
to determine which factors have the smallest impact on it, the fault time has a certain impact on the fan’s power generation.
these factors can be combined into ten models. The ten model The wind being too powerful and leading to an instantaneous
combinations are as follows: overload of the wind turbine can cause wind fan failure. Because
Model 1 : Wind speed; many causes of fan fault cannot be quantified, we use the time
Model 2 : Wind speed, temperature; series method to predict the fault time using historical fault data.
Model 3 : Wind speed, temperature, atmospheric pressure; The prediction results of the fault time are greatly influenced by the
Model 4 : Wind speed, temperature, atmospheric pressure, fault recent generation data. The fault times were predicted in December
time; 2016. The value of October and November 2016 is used to predict
Model 5 : Temperature; original data, and the algorithm mentioned in Algorithm 1 is used
Model 6 : Temperature, atmospheric pressure; to get the input matrix.
Model 7 : Temperature, atmospheric pressure, fault time; The prediction of fault time is based on two methods: ELMAN
Model 8 : Atmospheric pressure; neural network and time series based on BP neural network,
Model 9 : Atmospheric pressure, fault time; using the evaluation criteria mentioned in the previous section:
Model 10 : Fault time. MAE, MSE.
RBF and GRNN neural networks are used to test ten models of The number of neurons in the hidden layer of the ELMAN
MSE, and the test results are shown in Fig. 3. The specific values neural network is unknown. Before using the ELMAN neural
of MSE are shown in Table I. network to predict the fault time, the number of hidden neurons
According to Fig. 3, RBF for the framework of the predict in ELMAN is determined by the method of loop. The number of
model, the model 4 which has a very clear minimum MSE, GRNN different neurons is set to obtain the optimal prediction result. The

4 IEEJ Trans (2019)


NEURAL NETWORK, FAULT TIME, WIND POWER PREDICTION

Elman prediction error graph Best c = 0.57435 g = 0.57435 CVmse = 0.032807


0.8
7
0.6 11
14 1
0.4 18
0.8
0.2 0.2

MSE
Error

0 0.4

–0.2 0.2

0
–0.4
5
5
–0.6 0
0
–5 –5
–0.8 log2g
1 2 3 4 5 6 7 log2c
Time point
Fig. 5. Roughly select 3D view
Fig. 4. Elman prediction error graph
Result of parameter selection (Contour map) [Grid search method]
Best c = 0.57435 g = 0.57434 CVmse = 0.032807
number of neurons was set at 7, 11, 14, 18. The result of the case
8
is shown in Fig. 4.
Based on the results of this case, the average error of the 6
prediction of the four different numbers of hidden neurons is:
−0.0374, −0.1256, −0.1078, and −0.1552. When the number of 4
neurons in the hidden layer is 7, the best results are predicted, so
the hidden layer of Elman is 7. 2
The time series method based on BP neural network and the
log2g

0
ELMAN neural network are used to predict the fault time of the
next 3 days in order to verify the effectiveness of the method, and –2
the detailed prediction results are shown in Table II.
The prediction of the fault time is used as the basis for the –4
prediction of the next stage, so two different algorithms used in
this chapter are compared, and the two algorithms are accurately –6
compared in the selection of parameters in order to obtain a higher-
–8
precision fault time. –8 –6 –4 –2 0 2 4 6 8
log2c
4.3. Wind power generation prediction The data Fig. 6. Roughly select contour map
for wind power generation are collected from February 1, 2015
to November 31, 2016, and more than 600 days of data are from
the actual production data of a wind farm in Hunan, China. During The preliminary results of a rough selection: Best Cross
this period, due to the maintenance of wind turbines, the amount Validation MSE = 0.0343403, Best c = 0.5, Best g = 0.5. For
of wind power generation was zero. The data of these days is called Figs 5 and 6: the x -axis represents the value of the logarithm of
the abnormal value, and 589 days of data can be used in addition the base of 2, and the y-axis represents the value of the logarithm
to the data of these days. of 2.
When the SVR method is used, the K -fold cross-validation The contours indicate the accuracy of the corresponding K-CV
method is utilized to obtain a set of optimal c and g values in method after taking the corresponding c and g. It can be seen from
order to achieve the predicted MSE minimum. Because the c and the figure that the range of c can be reduced to between 2−2 and
g values must be discretized, and they are difficult to find, here, c 22 ; the range of g is reduced to between 20 and 22 .
and g are searched within the exponential range of 2. First, there Based on the above rough selection of parameters, a precise
is only a rough search, and an accurate search can be carried out selection used the same method was carried out. The final selection
based on it. The results of the rough search are shown in Figs 5 results are as follows: best cross validation MSE = 0.0335834,
and 6, contour lines and 3D views, respectively. best c = 1, best g = 0.707107. The exact result of the parameter

Table II. Predicted detailed result

BP ELMAN
Future days Original data Prediction data MAE (%) MSE (%) Prediction data MAE (%) MSE (%)
−3
1 52 51.809 0.367 1.349 × 10 53.207 2.321 5.388 × 10−2
2 64.83 61.790 4.689 0.222 66.819 3.068 9.412 × 10−2
3 76.50 79.718 1.36 1.855 × 10−1 73.776 3.565 12.707 × 10−2
Average value 2.139 0.136 2.985 9.169 × 10−2

5 IEEJ Trans (2019)


Y. CAO ET AL.

SVR parameter selection result chart (3D View Plot) [Grid search method]
Algorithm 2. Selection of the optimal parameters of GRNN
Best c = 0.70711 g = 0.5 CVmse = 0.032256
/*Through tenfold cross validation, each time to determine the
different training set and the parameters of speed were obtained
1 from the interval of 0.1–2 each time; we obtain the prediction
results of MSE value while the optimal training set and the value
0.8 of speed were obtained*/
0.2 for i = 1:10
/*Select a set of training sets for each cycle*/
MSE

0.4 1) p_cv_train = p_train(train,:);


0.2 2) t_cv_train = t_train(train,:);
3) p_cv_test = p_train(test,:);
0 4) for spread = 0.1:0.1:2;
4
2 4 /*The parameters speed of each set of training sets from 0.1 to
0 2 2*/
0
log2g
–2 –2 5) net = newgrnn (p_cv_train,t_cv_train,spread);
–4 –4 log2c 6) test_Out = sim(net,p_cv_test);
end
Fig. 7. Accurate select 3D view
end

Result of parameter selection (Contour map) [Grid search method]


Best c = 0.70711 g = 0 .5 CVmse = 0.032256 The optimal value of parameter speed is 0.2 by Algorithm 2. The
4 training set is determined by the GRNN neural network and the BP
neural network, and a decision was made to use this training set in
3 a test. The data of the next 20 days’ power generation is considered
a test set. In order to predict the test set, optimal training has been
2
used. The results are shown in Fig. 12, and the error and relative
1 error are shown in Figs 13 and 14, respectively.
log2g

0
4.5. Model prediction application A prediction model
–1 of SVR, GRNN, BP, and RBF neural networks has been introduced
in detail. The four models are used to predict the power generation
–2 in the next 20 days. The detailed results are shown in Table III, and
the prediction of GRNN, BP, and RBF the results of the operation
–3 are shown in Fig. 13.
Predicting short-term wind power with higher accuracy rates is
–4
–4 –3 –2 –1 0 1 2 3 4 very important for power system operators as they face challenges
log2c with increased ripple wind power production and wind power
capacity. Based on the presented simulation results in Table III,
Fig. 8. Accurate select contour map we compare the MAPE and MSE values of the four models. We
can clearly see that the SVM model has the lowest MAPE and
selection is shown in Figs 7 and 8. Based on the values of the MSE values and shows the best prediction result.
optimal c and g that have been obtained, the next power generation
prediction can be made. 5. Conclusion
As mentioned above, we have obtained actual production data
from February 2015 to November 2016. We use these existing Based on the actual operation data of a wind farm in Hunan,
historical data as a training set to verify the rationality of the China, this paper presents a two-stage wind power generation pre-
obtained parameters of c and g. Figure 9 shows the final prediction diction that takes the time of fault, the conclusions, and test results
results. It is a comparison between the original data and the into account. The conclusions and test results are summarized
predicted data. Figures 10 and 11 are the error graph and the as follows:
relative error graph, respectively.
By predicting a large amount of original data and comparing 1 When determining the factors the affect power generation,
the predicted result with the existing data, the rationality of the the influence of the fault time on power generation is also
obtained parameters of c and g is verified. It can be seen from the taken into consideration according to the actual operation
comparison result of Fig. 11 that the maximum relative error, not situation of the wind farm; ten models including meteoro-
more than 8, proved to be the predictive model parameter of the logical factors and fault times were established, and the fault
final power generation. time was discovered to have an influence on the generation
capacity to a certain extent.
4.4. GRNN and BP and RBF The GRNN neural 2 Fault time cannot be obtained through the weather station.
network has an adjustable parameter speed. In order to improve Therefore, first, use the historical data to predict the fault
the accuracy of the prediction, the input matrix and the output are time, In order to avoid the singularity of time series
also taken into consideration to determine which contributes more prediction, the method based on neural network time series
to the best prediction result. is used to predict the fault time and make the prediction
A tenfold cross-validation method was used to train the GRNN result more accurate.
neural network and to find the best speed parameter. The detailed 3 Finally, the prediction results of BP, RBF, GRNN, and SVM
algorithm is given below. are compared, and the results show that the SVM has the

6 IEEJ Trans (2019)


NEURAL NETWORK, FAULT TIME, WIND POWER PREDICTION

Original data vs regression prediction data


350 Original data
Regression prediction data

300

Power generation/unit: kW
250

200

150

100

50

–50
0 100 200 300 400 500
Power generation date (2015.2.01–2016.11.31/unit: day)

Fig. 9. Prediction results

Error chart (prediction data – original data)


150

100

50
Error amount

–50

–100

–150
0 100 200 300 400 500
Power generation date (2015.2.01–2016.11.31/unit: day)

Fig. 10. Error chart

Relative error chart (predicted data – original data)/original data


8

4
Relative error

–2

–4
0 100 200 300 400 500
Power generation date (2016.12.01–2016.12.20/unit: day)

Fig. 11. Relative error char

7 IEEJ Trans (2019)


Y. CAO ET AL.

Original data vs regression perdiction data


300 Original data
GRNN prdiction data
BP prdiction data
RBF prdiction data
250

Power generation/unit: kW
200

150

100

50

–50
–0 2 4 6 8 10 12 14 16 18 20
Power generation date (2016.12.01–2016.12.20/unit: day)

Fig. 12. The Prediction Result of GRNN and BP and RBF

Error chart
100 GRNN prdiction error
BP prdiction error
80 RBF prdiction error

60

40
Error amount

20

–20

–40

–60

–80

–100
–0 2 4 6 8 10 12 14 16 18 20
Power generation date (2016.12.01–2016.12.20/unit: day)

Fig. 13. Prediction error of GRNN and BP and RBF

Relative error
2
GRNN relative error
BP relative error
RBP relative error

1.5
Relative error amount

0.5

–0.5

–1
0 2 4 6 8 10 12 14 16 18 20
Power generation date (2016.12.01–2016.12.20/unit: day)

Fig. 14. Prediction relative error of GRNN and BP and RBF

8 IEEJ Trans (2019)


NEURAL NETWORK, FAULT TIME, WIND POWER PREDICTION

Table III. Prediction result (10) Hong J-S. Evaluation of the high-resolution model forecasts over
the Taiwan area during GIMEX. Weather and Forecasting 2003;
Prediction data 18:836–846.
(11) Taylor J, McSharry P, Buizza R. Wind power density forecasting
Original data SVM GRNN BP RBF
using ensemble predictions and time series models. IEEE Transac-
268.62117 233.456 198.799 209.275 192.538 tions on Energy Conversion 2009; 24(3):775–782.
215.56123 204.862 174.115 193.738 182.093 (12) Taylor J, McSharry P. Short-term load forecasting methods: An
237.04620 218.618 196.075 225.147 211.147 evaluation based on European data. IEEE Transactions on Power
113.75210 101.140 79.325 97.526 96.763 Apparatus and Systems 2007; 22(4):2213–2219.
34.681000 48.297 50.281 40.665 40.205 (13) Conejo A, Plazas M, Espinola R, Molina A. Day-ahead electric-
118.26130 68.755 65.001 67.933 65.215 ity price forecasting using the wavelet transform and ARIMA
153.55000 132.738 120.900 160.103 123.331 models. IEEE Transactions on Power Apparatus and Systems 2005;
60.054000 93.2018 75.8550 93.922 100.436 20(2):1035–1042.
(14) Deng J, Jirutitijaroen P. Short-term load forecasting using time series
86.832000 83.471 67.8200 70.476 82.326
analysis: A case study for Singapore. 2010 IEEE Conference on
101.05320 131.724 116.652 132.239 121.643
Cybernetics and Intelligent Systems 2010; 231–236.
221.22280 165.800 149.653 169.307 157.520
(15) Lin Q, Wang J. Vertically correlated echelon model for the interpo-
159.28843 147.890 134.518 174.284 137.025
lation of missing wind speed data. IEEE Transactions on Sustainable
30.498210 31.256 39.312 35.475 36.131
Energy 2014; 5:804–812.
83.683820 70.171 62.275 69.140 72.493 (16) Douak F, Melgani F, Benoudjit N. Kernel ridge regression with
203.94568 146.115 127.825 174.709 131.354 active learning for wind speed prediction. Applied Energy 2013;
114.20718 93.640 78.018 89.842 86.854 103:328–340.
104.25735 98.200 83.825 89.675 86.258 (17) Liu H, Erdem E, Shi J. Comprehensive evaluation of ARMA-
81.079834 79.688 71.645 56.520 53.085 GARCH(−M) approaches for modeling the mean and volatility of
68.828400 67.274 61.554 44.552 33.357 wind speed. Applied Energy 2011; 88:724–732.
23.984000 22.445 24.070 21.714 −19.375 (18) SA K. Artificial neural networks in renewable energy systems
Performance applications: a review. Renewable and Sustainable Energy Reviews
MAE 0.166 0.2406 0.2064 0.3306 2001; 5:373–401.
MSE 0.0304 0.0699 0.0587 0.2461 (19) Hu CHC, Zhao FZF. Improved methods of BP neural network
algorithm and its limitation. 2010 International Forum on Information
Technology and Applications, 2010.
(20) Khosravi A, Nahavandi S, Creighton D. Prediction intervals for short-
lowest error in predicting the next 20 days. The correlation term wind farm power generation forecasts. IEEE Transactions on
analysis between fault and external meteorological factors Sustainable Energy 2013; 4(3):602–610.
(temperature, wind speed, atmospheric pressure) is proposed (21) Zeng J, Qiao W. Short-term wind power prediction using a wavelet
by using the SVM method. The prediction accuracy is support vector machine. IEEE Transactions on Sustainable Energy
2012; 3(2):255–264.
improved to verify the validity and feasibility of this
(22) Amjady N, Keynia F, Zareipour H. Wind power prediction by a
prediction method. new forecast engine composed of modified hybrid neural network
and enhanced particle swarm optimization. IEEE Transactions on
Sustainable Energy 2011; 2(3):265–276.
References (23) Fan S, Liao J, Yokoyama R, Chen L, Lee W-J. Forecasting the
wind generation using a two-stage network based on meteorological
information. IEEE Transactions on Energy Conversion Jun. 2009;
(1) Mohandes MA, Rehman S. Short term wind speed estimation in Saudi 24(2):474–482.
Arabia. Journal of Wind Engineering and Industrial Aerodynamics (24) Costa A, Crespo A, Navarro J, Lizcano G, Madsen H, Feitosa E. A
2014; 128:37–53. review on the young history of the wind power short-term prediction.
(2) Wang J, Zhang W, Wang J, Han T, Kong L. A novel hybrid approach Renewable and Sustainable Energy Reviews 2008; 12(6):1725–1744.
for wind speed prediction. Information Sciences 2014; 273:304–318. (25) Razavi S, Tolson BA. A new formulation for feedforward neural net-
(3) Porter K, Rogers J. Status of centralized wind power forecasting in works. IEEE Transactions on Neural Networks 2011; 22:1588–1598.
North America. National Renewable Energy Laboratory, Golden, CO, (26) Han S, Liu Y, Yang Y, Li J. Ultra-short term wind power prediction
Technical Report NREL/SR-550-47853, 2010. and uncertainty assessment. Acta Energiae Solaris Sinica 2011;
(4) Zhao SZS, Zhao JZJ, Zhao GZG, Zhang WZW, Guo ZGZ. Effective 32(8):1251–1256.
wind power density prediction based on neural networks. 2010 (27) Chandra R. Competition and collaboration in cooperative coevolu-
International Conference on Multimedia Technology 2010. tion of Elman recurrent neural networks for time-series prediction.
(5) Sheela KG, Deepa SN. Neurocomputing neural network based hybrid IEEE Transactions on Neural Networks and Learning Systems 2015;
computing model for wind speed prediction. Neurocomputing 2013; 26(12):3123–3136.
122:425–429. (28) Ding Y, Gao J, Wang X. Ad Hoc network traffic prediction based on
(6) Quan H, Srinivasan D, Khosravi A. Short-term load and wind the Elman neural network. International Conference on Machinery
power forecasting using neural network-based prediction intervals. 2015.
IEEE Transactions on Neural Networks and Learning Systems 2014; (29) Shouping C, Rui D, Xiaoli L. Analysis of 30 Cases of MATLAB Neural
25:303–315. Network . Beihang University Press: Beijing; 2010; 73–74.
(7) Lydia M, Selvakumar AI, Kumar SS, Kumar GEP. Advanced algo- (30) Pin Z. MATLAB Neural Network Design and Application [M] .
rithms for wind turbine power curve modeling. IEEE Transactions Tsinghua University Press: Beijing; 2013.
on Sustainable Energy 2013; 4:827–835. (31) Renyi Z, Yuhui Z. Discussion on time series forecasting. Science &
(8) Watson SJ, Landberg L, Halliday JA. Application of wind speed Technology Information 2011; 15:198–199.
forecasting to the integration of wind energy into a large scale (32) Ge Z-F, Sun Z-Q. Neural Network Theory and MATLAB R2007
power system. IEE Proceedings—Generation, Transmission and Implementation. ElectronicsIndustry Press: Beijing; 2007.
Distribution 1994; 141:357–362. (33) Li G, Shi J. On comparing three artificial neural networks for wind
(9) Haque AU, Nehrir MH, Mandal P. A hybrid intelligent model speed forecasting. Applied Energy 2010; 87:2313–2320.
for deterministic and quantile regression approach for probabilistic (34) Tagliaferri F, Viola IM, Flay RGJ. Wind direction forecasting
wind power forecasting. IEEE Transactions on Power Systems 2014; with artificial neural networks and support vector machines. Ocean
29:1663–1672. Engineering 2015; 97:65–73.

9 IEEJ Trans (2019)


Y. CAO ET AL.

(35) Gala Y, Fernandez A, Dı́az J, Dorronsoro JR. Support vector Qingqing Hu (Non-member) obtained a master degree in 2018
forecasting of solar radiation values. In HAIS 2013. LNCS (LNAI), from Shanghai University of Electric Power
vol. 8073. Pan J-S, Polycarpou MM, Wozniak M, de Carvalho and is currently working at Shanghai Electric
ACPLF, Quinti’an H, Corchado E (eds). Springer: Heidelberg; 2013;
Power Design Institute Co. Ltd. The main
51–60.
(36) Cristianini N, Shawe-Taylor J. An Introduction to Support Vector
research direction is the neural network of
Machines and Other Kernel-Based Learning Methods. Cambridge: wind power grid and considering the econ-
The Press Syndicate of Cambridge University; (2000). http://dx.doi omy of carbon dioxide emissions in order to
.org/10.1017/CBO9780511801389. obtain the best wind–thermal power gener-
(37) Smola A, Schölkopf B. A tutorial on support vector regression. ation.
Technical Report TR-98-030, 2003.
(38) Basak D, Pal S, Patranabis DC. Support vector regression. Hui Shi (Non-member) State Grid Ningxia Province Elec-
Neural Information Processing—Letters and Reviews 2007; tric Power Co., Ltd. Ningdong City Power
11(10):203–224. Supply Company, Responsible for profes-
(39) Ren Y, Suganthan PN, Srikanth N. A novel empirical mode decom- sional management of power regulation, dis-
position with support vector regression for wind speed forecasting. patching plan, operation mode, hydropower
IEEE Transaction on Neural Network and Learning System 2016; and new energy, relay protection, dispatch
27(8):1793–1798.
automation and power communication in the
jurisdiction; participate in the formulation of
Yunkun Cao (Non-member) obtained a Ph.D. in 2006 from secondary system planning.
Chongqing University School of computer Yanli Zhang (Non-member) Lanzhou City in Gansu Province
science and technology. From 2007 to 2009, Xigu thermoelectric limited liability com-
she conducted postdoctoral research in the pany. Mainly engaged in electric power,
Ministry of Science and Technology Infor- heat production and supply; design related
mation Technology of Jiangsu Province. to electric power and heating engineering.
At the same time, she has been engaged
in power enterprise information architec-
ture planning, power enterprise master data
model modeling and so on.

10 IEEJ Trans (2019)

You might also like