Research on Wind Power Prediction Modeling Based on Adaptive Feature Entropy Fuzzy Clustering

HUANG Haixin,KONG Chang

(1.Shenyang Ligong University,Shenyang 110159,China;2.Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110159,China)

Research on Wind Power Prediction Modeling Based on Adaptive Feature Entropy Fuzzy Clustering

HUANG Haixin1,2,KONG Chang1

(1.Shenyang Ligong University,Shenyang 110159,China;2.Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110159,China)

Wind farm power prediction is proposed based on adaptive feature weight entropy fuzzy clustering algorithm.According to the fuzzy clustering method,a large number of historical data of a wind farm in Inner Mongolia are analyzed and classified.Model of adaptive entropy weight for clustering is built.Wind power prediction model based on adaptive entropy fuzzy clustering feature weights is built.Simulation results show that the proposed method could distinguish the abnormal data and forecast more accurately and compute fastly.

fuzzyC-means clustering;adaptive feature weighted;entropy;wind power prediction

As the world energy crisis increasingly sharp contradictions,frequent extreme climate conditions,new energy and renewable energy research,development and utilization has become the focus of people deal with the situation.Wind power as one of the world's most important renewable clean energy sources that has been vigorously in our country.By the end of 2012,China′s wind power installed capacity of 75324.2 MW,rose 20.8%from a year than in 2011[1].And compared with other renewable energy,wind power is random,intermittent and controllability.All that resulted the volatility of wind power exists.With the increasing proportion of wind power grid,the instability of wind power not only brought great security hidden danger to power system,and greatly increases the reserve capacity and running cost of the system.So quickly and accurately predict wind power can reduce the economic costs colleagues to improve the stability and reliability of power generation,provide the basis for power generation plan for the grid arrangement and fault diagnosis.Existing mostly used in wind power forecasting method for days,when the interval of a small amount of data sample training and forecasting,this will most likely not enough comprehensive analysis of all the trends of wind power.In this paper,using a large number of historical data of a wind farm in Inner Mongolia to repeated data mining,to dig out as much as possible from wind power and the relationship between the related factors,to provide more accurate wind farm power prediction model is established.

The wind farm power prediction methods mainly include:kalman filtering method,the continuous method,time series method,spatial correlation method and artificial neural network,fuzzy clustering method[2].Kalman filtering method need to estimate the statistics of a large number of prediction errors in advance,the application more difficult.Time series method is the most used method,suitable for short-term prediction for several hours at a time.Continuous method need to take some time before sliding the historical average wind speed as the next time forecast,forecast time range is small,not apply for 24 h forecasts.Spatial correlation method based on correlation between adjacent point wind speed,weather forecast,need to deal with a large number of historical data.Artificial neural network is suitable for short-term wind speed forecasting of 1 d,but is highly affected by the input parameters and the sample selection[3].Traditional fuzzy clustering method using characteristic values of the weight of 1 Euclidean distance to evaluate the degree of similarity,while in the actual wind farm power prediction value the importance of each feature is different[4-7].Aiming at this problem,this paper presents an adaptive characteristics of entropy algorithm to optimize the weight of characteristic value,and analyze the different characteristics influence the degree of wind power,find out the most suitable for the clustering center.Calculate the future weather data to the center of the membership degree,the matrix power point will be multiplied by the membership degree and predictive value for power.Experiments prove that this method can effectively dig out from the data regularity between wind power and its related factors,improve the accuracy of wind power prediction.

1 THE FCM ALGORITHM

To set sample sets,X={x1,…,xN},the clustering center,V={v1,…vC},the samples of membership degree of cluster centerxj(j=1,…,N),then optimization model of clustering are as follows:

(1)

0≤uij≤1

Among them:m>1 is a fuzzy clustering results which can control the degree of constants;anddij=‖xj-vi‖ denotes the samplejwith the firsticluster center Euclidean distance.Using the Lagrange multiplier algorithm model iterative formula:

(2)

1≤j≤C

(3)

2 WIND POWER PREDICTION METHOD BASED ON FUZZY CLUSTERING OF FEATURE EXTRACTION

Quickly and accurately to predict the output power of the wind farm makes great difference to power system stability,security and power scheduling.Analysis the output characteristics of wind turbine and the factors affecting wind power output.We choosev3、wind direction Angle and wind farm output power p as adaptive characteristics of the characteristic value of the entropy weight fuzzy clustering algorithm.Selectv3of the wind speed and wind direction angle of the next moment w as a characteristic of the wind farm power prediction algorithm.

3 ADAPTIVE FEATURE OF ENTROPY ALGORITHM

In the FCM algorithm,the degree of similarity between two points was evaluate by the Euclidean distance,and used the feature weight as constant 1.However,in practical problems,the importance of each attribute characteristics is not the same,based on this analysis will inevitably be distorted.Here have several methods to solve the problem:

3.1 Data standardization

(4)

Namely to compress data processing,make its numerical range of [0,1].Standardized processing to eliminate the numerical difference between the features,but still could not distinguish the importance of attributes.

3.2 Expert scoring,feature weighting method

This method is too subjective,for the increasingly complex manufacturing systems,subjective judgment is difficult,error is not easy to control.

3.3 The information entropy value added to the objective function[8]

(5)

The goal is not only allows optimization of the similarity between the same maximum,but also makes the total membership based on information entropy minimum.When the membership to take either 0 or 1,the definition of information entropy reaches a minimum,that is,its essence is a combination of hard clustering and fuzzy clustering.

3.4 Information entropy weighting algorithms

Characteristic properties of entropy value is:

(6)

The greater the entropy values and the more disorder of the features.So we need to reduce their weights,like below:

(7)

The algorithm before clustering need to determine the feature weights.

In fact,before carrying clustering,orderly feature properties importance is not equal to the clustering.After clustering,the more orderly feature properties the better could we distinguish the sample which without clustering.So it is of greater importance clustering.Therefore,this paper put forward an adaptive feature of entropy algorithm,clustering and the features of entropy used in the clustering algorithm to select feature weights.So the weight is not fixed,they changed after clustering iteration optimization.

The standardized sample data,xjk∈[0,1],the value range of the characteristic value is divided intolequal portions,

(8)

From the perspective of feature attributes,after clustering sample data in thelinterval falls in clustering centeriwith the specified probability:

(9)

Entropy of feature attributek:

(10)

According to the idea of entropy to build feature weights:

(11)

(12)

4 PREDICTED MODEL BASED ON FUZZY CLUSTERING ALGORITHM

Take the algorithm of adaptive feature weight entropy fuzzy clustering to train the sample data,then the optimal clustering center and feature weights are obtained.The clustering center of the forecast datavi′ equal the optimal clustering center which being removed its column of being forecast.And so does the feature weight.New membershipuij′ was built with equal(13).So the predictive models based on the fuzzy clustering ideas were:

(13)

k=the column of power

5 FORECASTING MODEL SELECTION OF EVALUATION INDEXES

According to the national energy administration management of the current wind farm wind power real-time prediction[9],we have selected the root mean square error as the predictive model for evaluation.

Root mean square error:

(14)

In the formula:nis the number of units target wind farm;PMiis the actual power value at timei;PPiis the predicted power value at timei;Cap for the boot Capacity.

6 STEPS OF WIND FARM POWER PREDICTION BASED ON ADAPTIVE FEATURE WEIGHT ENTROPY FUZZY CLUSTERING ALGORITHM

Step1:Set the number of clustersC,parameterm,loop iterations and accuracy;

Step2:Standardized the sample data and initial weightwk=1/K(k=1,…,K);

Step3:Initial membershipuij;

Step4:According to equation(2)and(3),calculate and update the cluster centers and membership;

Step5:If‖Unew-Uold‖<εis met then the loop terminates;

Step6:According to equation(8)～(12),calculate and update the feature weightWij,jump to step 4 until arrive the loop iterations;

Step7:Calculating the predicted data cluster centervij′,feature weightwk′ and membershipuij′,then according to equation(13)to calculate the predicted power.

7 THE EXPERIMENTAL RESULTS AND ANALYSIS

A large number measured data of a wind farms in Inner Mongolia to do experiment.Predict the next days wind power output.The total installed capacity of wind farm is 38MW.Before carrying adaptive feature entropy clustering analysis,first preprocess the data,excluding maintenance outage or machine caused by abnormal data.Select 2010,2011,2012 February historical data as the training data,data collection interval is 15 minutes.Choose the wind speed and wind direction in mid-February 2013 to predict the wind farm power.Here used two methods:They are Weighted Fuzzy C-Means Clustering and Adaptive Feature Entropy Fuzzy Clustering algorithm[10].In figure 1 and 3 the datas those were circled are coursed by the faulted equipments.The black stars are the centers.With many times experiments,after all,we set up the classification category number for 5,the number of iterations is 3,the intervallis 80,blurring constantmis 2,and the cut-off conditionεequals 10-5.Here in after referred to as WFCC and AFEFC.By the weighted fuzzyC-average clustering method to get the characteristics of the weight is:[0.1132 0.3321 0.5547],and the clustering center is:

According to the fifth part,we calculated the weight and the center of the forecast data respectively are as follows:

And the root mean square error obtained by the formula(14)was 6.92%.The simulation speed is 4.132652 seconds.Clustering results shown in Figure 1.And wind power prediction shown in Figure 2.With the adaptive characteristics of entropy fuzzy clustering method,we obtained the entropy and weights as follows table 1 for each attribute after every clustering.The bigger the entropy,the smaller the weight.So we get the optimal weights of the attributes were:

[0.3519 0.1873 0.4609];and the clustering center is:

As the same methods above obtained the weights and the center of the forecast data were:

With the same formula(14) obtained the root mean square error:5.79%.The simulation speed is 3.988436 seconds.Clustering results shown in Figure 3;Wind power prediction shown in Figure 4.

Figure 1 The clustering result of WFCC

Figure 2 Farm wind power prediction results WFCC

Figure 3 The clustering result of AFEFC

Figure 4 Wind farm power prediction results of AFEFC

Table 1 Entropy value and weight list

Proposed an adaptive feature weights FCM algorithm and applied to short-term wind power prediction.In the basic FCM algorithm based on adaptive feature weight added Entropy recursive algorithm,the algorithm for clustering results from the feature attributes described in the context and create a feature entropy increase clustering effect.Experimental results show that the algorithm can identify the equipment due to maintenance or other causes of abnormal data,and can be more accurate for predicting the power of the wind farm.

9 CONCLUSION

Feature weight algorithm has great impact on the classification results.However,traditional fuzzy clustering considers each feature attributes are equally important,while in the actual wind farm power prediction value the importance of each feature is different.Proposed and inside classes.The experimental results illustrate that the algorithm caneffection among and inside classes.The experimental results illustrate that the algorithm can effectively distinguish the features attributes on the importance of wind power.In addition to this,it can identify the abnormal points and in the wind power prediction higher accuracy can be obtained.

[1]Chinese wind energy association.The statistics of China′s wind power installed capacity in 2012 [EB/OL].http：//www.Cwea.Org.Cn/download/display_info.asp?id=44，2012-03-23.

[2]LIN Hai-tao,JIANG Chuan-wen,REN Bo-qiang,et al.Short Term Combined Prediction of W ind Speed Based on Fuzzy C lustering[J].East China Electric Power,2010,(38):295-299.

[3]M C Alexiadis,P S Dokopoulos,H S Sahsamanoglou.Wind Speed and Power Forecasting based on Spatial Correlation Models[J].IEEE Transactions on Energy Conversion,1999,(14):836-842.

[4]Shan Zeng.Research on Fuzzy clustering algorithm[D].WU Han:Huazhong University of Science and Technology,2012.

[5]GL Valentini,W Lassonde,SU Khan.An overview of energy efficiency techniques in cluster computing systems[J].Cluster Computi,Sep 2011,pp.3-15,doi:10.1007/s10586-011-0171-x.

[6]LIU Yu.Study on wind power prediction of large-scale wind farm based on real data analysis[J].Heilongjiang Electric Power,2011,(33):11-15.

[7]WANG Jian-cheng,YANG Ping,YANG Xi.Research on wind power prediction modeling based on numerical weather prediction[J].Renewable Energy Resources,2013,(31):34-38.

[8]XB Zhi,JL Fan,F Zhao.Fuzzy Linear Discriminant Analysis guided maximum entropy fuzzy clustering algorithm[J].Pattern Recognition,2013,(46):1604-1615.

[9]YANG Mao,XIONG Hao,YAN Gan-gui,et al.Real-time prediction of wind power based on data mining and fuzzy clustering[J].Power System Protection and Control,2013,(41):1-6.

[10]Bonian Li.Weighted Fuzzy C-Means Clustering[J].Fuzzy Systems and Mathematics,2007,(21):106-110.

马金发)

date： 2013-12-09

This work was supported by the Natural Science Foundation of China under contact(61233007)

Biography： HUANG Haixin(1972—),female,associate professor,Research direction:systems engineering,smart grids and fuzzy game.

1003-1251(2014)04-0075-06

TP206.1DocumentcodeA