Short-term wind power prediction method considering wave process division

Abstract

The application relates to the technical field of wind power, and is a short-term wind power prediction method considering wave fluctuation process division, which has the characteristics that wave fluctuation characteristics can be extracted according to wind turbine output characteristic analysis results and a wind speed-power physical conversion model, then an improved K-means clustering method for wind power prediction is proposed, and based on the classification results, an abnormal data processing model and a short-term wind power prediction model under different wave fluctuation processes are established. Simulation calculation verifies that the prediction method is scientific and reasonable, the prediction process is simple, the prediction precision is high, the physical meaning is clear, the prediction result is effective, and the practicability is high.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power prediction, and specifically relates to a short-term wind power prediction method considering fluctuation process division. BACKGROUND

[0002] As the most potential renewable energy, wind energy has become an important direction of energy development in China. However, the time-varying nature of wind energy leads to strong uncertainty of wind power generation, and large-scale grid-connected operation of wind power will seriously threaten the stability of the power system. Therefore, it is extremely important to improve the accuracy of wind power prediction. The power grid dispatching department can arrange the dispatching plan in advance according to the prediction results, avoid many adverse effects caused by grid-connected wind power, and effectively reduce the phenomenon of wind curtailment.

[0003] Short-term wind power prediction can provide reference for power system optimization of daily generation plan and maintenance plan. The short-term wind power prediction method mostly uses time series method or neural network method to construct a nonlinear conversion model between meteorological information (wind speed is the most important meteorological information) and wind power, and then uses the meteorological information of the prediction period such as numerical weather prediction (NWP) as the input of the model to predict the wind power. However, the wind speed-power conversion relationship under different wind speed variation characteristics has obvious differences, and a single wind speed-power conversion model is difficult to accurately describe the nonlinear relationship between wind speed and power.

[0004] The present application analyzes the influence of wind speed variation on wind speed-power conversion relationship according to the principle of wind power generation, and the wind speed-power conversion relationship corresponding to different wind speed variation characteristics has obvious differences. Based on the above theory, the present application performs short-term wind power combination prediction considering the wind speed variation characteristics. SUMMARY

[0005] The present application aims to overcome the shortcomings of the prior art and provide a short-term wind power prediction method considering fluctuation process division.

[0006] The technical scheme adopted to achieve the present application is as follows: a short-term wind power prediction method considering fluctuation process division, characterized in that it comprises the following steps:

[0007] 1) Abnormal data elimination and completion:

[0008] Firstly, the wind speed is filtered, if there is an abnormal value point of wind speed, the filtering error is large, then the filtering error abnormal value point is identified by the Shaw Weier method, and the corresponding wind speed abnormal data is removed, secondly, the grey correlation method is used to find the wind turbine with the strongest correlation with the wind speed data of the previous period of the missing data, and the wind speed of the wind turbine at the same period is used to fill in the missing data, if all wind turbines in the wind farm have abnormal output at the same period, the missing data is filled in by using the cubic spline interpolation method, thirdly, the Copula function is used to fit the wind power probability distribution, and the abnormal value point of suspicious probability is identified and removed, then the grey correlation method is used to find the wind turbine with the strongest correlation with the wind speed data of the same period of the missing data, and the wind turbine power data at the same period is used to fill in the missing data, if all wind turbines have abnormal output, the data is reconstructed by using support vector machine, finally, the processed data of each wind turbine is used to obtain the wind speed and the whole field power generation of the wind farm;

[0009] 2) Wind speed change feature recognition:

[0010] The influence of wind speed change feature on power is not only related to the trend of wind speed change, but also closely related to the change amplitude of wind speed, the greater the change amplitude of wind speed, the more the wind turbine output deviates from the standard value,

[0011] According to the principle of fluid mechanics, the actual output mechanical power P M of the wind turbine is related to the wind speed v as follows:

[0012]

[0013] Wherein, C P represents the wind energy utilization coefficient; v in represents the cut-in wind speed of the unit; v n represents the rated wind speed of the unit; v out represents the cut-out wind speed of the unit; p represents the air density; S represents the impeller swept area;

[0014] When v<v in or v>v n , the output power of the wind turbine is a fixed value, so there is no difference in the influence of different wind speed change features on power in these two wind speed intervals, but for the whole wind farm, since the wind speed has spatial dispersion, there will be P>0 when v<v in , therefore the upper limit v small_max of the small wind period should be less than v in , and similarly, there will be P<P n when v>v n , therefore the lower limit v large_min of the large wind period should be greater than v n ;

[0015] From formula (1), compared with the first-order difference of wind speed Δv i , the first-order difference of wind speed cube Δv can better reflect the influence of different wind speed change characteristics on the power generation of wind farm. If the wind speeds corresponding to time i and time i-1 are v i and v i-1 respectively, then Δv i , The calculation formula is as follows:

[0016] Δv i = v i -v i-1 (2)

[0017]

[0018] The improved K-means clustering algorithm is used to identify the wind speed change characteristics:

[0019] ① Selection of initial clustering center

[0020] i. Based on the identification results of wind speed change characteristics, the change characteristics of different rising and falling amplitudes need to be analyzed, so (v w , 0) is set as the first initial clustering center X1, where v w is the wind speed value corresponding to the vertex of the Weibull function fitted to the probability distribution of wind speed samples;

[0021] ii. The Euclidean distance d ij between any two clustering samples x i and x j is calculated, and the two sample points with the largest distance are taken as the second and third initial clustering centers. Since the vector v i is smaller than the vector , the two sample points with the largest distance correspond to a positive value and a negative value of the vector . The sample point when is positive is selected as X2, and the sample point when is negative is selected as X3, where the calculation formula of d ij is as follows:

[0022]

[0023] iii. In the remaining clustering samples, the sample point with the largest distance product of X1 and X2 when is positive is selected as X4; the sample point with the largest distance product of X1 and X3 when is negative is selected as X5; and the remaining (K-5) initial clustering centers are sequentially derived. In order to maintain the symmetry of wind speed change characteristic identification, the value of K can only be an odd number;

[0024] ② Selection of the optimal K value

[0025] The Elbow method is an effective means of determining the optimal number of clusters K. When K is less than the optimal number of clusters, as K increases, the similarity of samples within each category increases rapidly, and the sum of squared errors (SSE) decreases significantly. However, as K increases further, the rate of decrease in SSE gradually decreases, and its change curve resembles the shape of an elbow. The optimal number of clusters K is the K value corresponding to the elbow. The expression for the sum of squared errors (SSE) is:

[0026]

[0027] Where C n For the nth cluster category; L n C n Number of internal samples; C n Cluster center; x i C n A specific sample within;

[0028] Since the characteristics of wind speed-power conversion can be reflected through the scatter distribution of wind speed-power, in a two-dimensional vector... After performing K-means clustering, it will be compared with wind speed v i Corresponding simultaneous power data P i Based on the clustering results, construct n groups of two-dimensional vectors (v i P i ), and use formula (5) to calculate its SSE;

[0029] Wind speed variation features identified based on the improved K-means clustering algorithm are as follows:

[0030] First, v i <v small_max The data is defined as small wind fluctuations, and v i >v large_min The data is defined as wind fluctuations, and changes in wind speed within two ranges have no impact on the wind speed-power conversion relationship.

[0031] Two-dimensional vector As clustering samples, v was analyzed using an improved K-means clustering algorithm. min <v i <v max The data is used to identify wind speed change characteristics, and wind speed change characteristics can be classified according to the wind speed-power conversion relationship.

[0032] The wind speed fluctuation characteristic data is marked as 0, and the wind speed rising and falling characteristic data is marked as the mean value of the first-order difference of the wind speed in each data set;

[0033] 3) Short-term wind power prediction model based on extreme learning machine:

[0034] The extreme learning machine (ELM) can randomly generate initial weight values and hidden layer node parameters, and use the least square method to calculate the weight values of the output layer. However, in order to obtain the optimal solution, the ELM neural network needs to adjust the number of hidden layer neurons,

[0035] If there are N groups of training samples (x i ,t i ), wherein the input vector x i =[x i1 ,x i2 ,…,x in ] T ∈R n , the output vector t i =[t i1 ,t i2 ,…,t im ]∈R m , then the ELM neural network model with L hidden nodes and activation function g(x i ) is recorded as:

[0036]

[0037] wherein β i is the output layer node weight; ω i is the input layer node weight; b i is the bias of the i-th hidden layer node; y j is the final output of the network, and N represents the number of samples,

[0038] The training target of the ELM neural network is to obtain the minimum output error, that is:

[0039]

[0040] wherein t i represents the expectation in the training process of sample i;

[0041] According to formulas (6) and (7), the objective function of the ELM neural network is:

[0042]

[0043] If H represents the hidden layer output matrix, then the matrix form of formula (8) is:

[0044] βH=T (9)

[0045] ELM neural network maintains H unchanged during training, and beta can be obtained by solving the least square solution of the following formula:

[0046] min||βH-T|| (10)

[0047] The solution of formula (9) is :

[0048]

[0049] Where H + is the Moore-Penrose generalized inverse matrix of H, and the short-term wind power prediction result is obtained by taking ELM as a predictor.

[0050] The short-term wind power prediction method considering fluctuation process division provided by the application extracts fluctuation characteristics according to the wind turbine output characteristic analysis result and the wind speed-power physical conversion model, further proposes an improved K-means clustering method for wind power prediction, and establishes an abnormal data processing model and a short-term wind power prediction model under different fluctuation processes based on the classification result. The simulation calculation verifies that the prediction method is scientific and reasonable, the prediction process is simple, the prediction precision is high, the physical meaning is clear, the prediction result is effective, and the practicability is strong. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The flow chart of abnormal data elimination and filling;

[0052] Figure 2 The schematic diagram of ELM basic structure;

[0053] Figure 3 The wind speed change characteristic identification result curve diagram;

[0054] Figure 4 The wind speed-power scatter plot;

[0055] Figure 5 The wind power prediction result curve diagram;

[0056] Figure 6 The wind power combination prediction result block diagram. DETAILED DESCRIPTION

[0057] The short-term wind power prediction method considering fluctuation process division will be further described below in combination with the drawings and specific embodiments.

[0058] In combination with Figures 1-6 , the short-term wind power prediction method considering fluctuation process division comprises the following steps:

[0059] 1) Abnormal data removal and completion:

[0060] like Figure 1 As shown, firstly, wind speed is filtered. If outliers exist, the filtering error is large. The Schauville method is then used to identify these outliers and remove the corresponding abnormal wind speed data. Secondly, the grey relational analysis method is used to find the wind turbine with the strongest correlation to the wind speed data of the previous period of the missing data. The missing data is then filled in using the wind speed data of the wind turbine in the same period. If all wind turbines in the wind farm have anomalies in the same period, cubic spline interpolation is used to fill in the missing data. Thirdly, the Copula function is used to fit the wind power probability distribution, and outliers with suspicious probabilities are identified and removed. Then, the grey relational analysis method is used to find the wind turbine with the strongest correlation to the wind speed data of the same period of the missing data. The missing data is then filled in using the power data of the wind turbine in the same period. If all wind turbines have abnormal output, support vector machines are used to reconstruct the data. Finally, the wind speed and total power generation of the wind farm are obtained using the processed data of each wind turbine.

[0061] 2) Identification of wind speed change characteristics:

[0062] The impact of wind speed variation characteristics on power is not only related to the trend of wind speed variation, but also closely related to the magnitude of wind speed variation. The greater the magnitude of wind speed variation, the more the output of the wind turbine deviates from the standard value.

[0063] According to the principles of fluid mechanics, the actual output mechanical power P of the wind turbine is... M The correspondence between wind speed v and wind speed v is as follows:

[0064]

[0065] Among them, C P Indicates the wind energy utilization coefficient; v in Indicates the unit's cut-in wind speed; v n Indicates the rated wind speed of the unit; v out ρ represents the cut-out wind speed of the unit; S represents the air density; and S represents the impeller swept area.

[0066] When v < v in Or v > v n At that time, the output power of the wind turbine is a fixed value. Therefore, the impact of different wind speed variations on the power is the same in these two wind speed ranges. However, for the wind farm as a whole, due to the spatial dispersion of wind speed, v < v in There are also cases where P > 0, therefore an upper limit v is set for periods of light wind. small_max It should be less than v in Similarly, v > v n Sometimes P < Pn The case occurs, the lower limit of the strong wind period v large_min The setting should be greater than v n .

[0067] From formula (1), compared to the first-order difference of wind speed Δv i , the first-order difference of wind speed cube can better reflect the influence of different wind speed change characteristics on the power generation of wind farm, if the wind speed corresponding to i time and i-1 time are v i and v i-1 , then Δv i , The calculation formula is:

[0068] Δv i = v i -v i-1 (13)

[0069]

[0070] The improved K-means clustering algorithm is used to identify the wind speed change characteristics:

[0071] ③ Selection of initial clustering center

[0072] i. Based on the wind speed change characteristic identification result, there is "fluctuation", and the change characteristics of different rising and falling amplitudes need to be analyzed, so set (v w , 0) as the first initial clustering center X1, wherein v w is the wind speed value corresponding to the vertex of the Weibull function fitted to the probability distribution of wind speed samples;

[0073] ii. Calculate the Euclidean distance d ij between any two clustering samples x i and x j , and the two sample points with the largest distance are taken as the second and third initial clustering centers. Since the vector v i and the vector are small in value, the two sample points with the largest distance correspond to a positive value and a negative value of the vector , and the sample point when is positive is taken as X2, and the sample point when is negative is taken as X3, wherein the calculation formula of d ij is:

[0074]

[0075] iii. In the remaining clustering samples, select the sample point when is positive and the product of the distance from X1 and X2 is the largest as X4; select the sample point when The sample point with the maximum distance product of X1 and X3 at negative time is taken as X5; the remaining (K-5) initial clustering centers are sequentially derived. To maintain the symmetry of wind speed variation feature recognition, the value of K can only be an odd number.

[0076] ④ Selection of optimal K value

[0077] Elbow method is an effective means to determine the optimal clustering number K. When the value of K is less than the optimal clustering number, with the increase of K value, the similarity of samples in each category rapidly increases, and SSE significantly decreases. However, with the increase of K value, the decrease amplitude of SSE gradually decreases, and the change curve is similar to the shape of elbow. The optimal clustering number K is the K value corresponding to the elbow, and the expression of SSE is:

[0078]

[0079] where C n is the n-th clustering category; L n is the number of samples in C n ; C n is the clustering center; and x i is a sample in C n .

[0080] The wind speed-power conversion characteristics can be reflected by the wind speed-power scatter distribution. After K-means clustering is performed on the two-dimensional vector , the simultaneous power data P i corresponding to the wind speed v i is divided according to the clustering results, n groups of two-dimensional vectors (v i , P i ) are constructed, and the SSE is calculated by using formula (5).

[0081] The wind speed variation feature recognition based on the improved K-means clustering algorithm is as follows:

[0082] First, the data of v i <v small_max are defined as small wind fluctuations, and the data of v i >v large_min are defined as large wind fluctuations. The variation of wind speed in the range of two segments has no effect on the wind speed-power conversion relationship.

[0083] The two-dimensional vector is taken as the clustering sample, and the improved K-means clustering algorithm is used to perform wind speed variation feature recognition on the data of v min <v i <v max . The wind speed variation feature can be classified according to the wind speed-power conversion relationship.​

[0084] The characteristic value of each wind speed variation feature is marked, the wind speed fluctuation feature data is marked as 0, and the wind speed rise and fall feature data is marked as the mean value of the first-order difference of the wind speed in each data set.

[0085] 3) Short-term wind power prediction model based on extreme learning machine:

[0086] As shown in Figure 2 , the extreme learning machine (ELM) is widely used in the prediction field, which can randomly generate initial weight and hidden layer node parameters, and calculate the weight of the output layer by using the least square method, which eliminates the iterative process of determining network parameters in traditional neural network, greatly shortens the time of adjusting network parameters, increases the learning speed, and effectively avoids the situation of falling into local optimal solution. However, in order to obtain the optimal solution, the number of hidden layer neurons of ELM neural network needs to be adjusted.

[0087] If there are N groups of training samples (x i ,t i ), the input vector x i = [x i1 ,x i2 ,…,x in ] T ∈R n , the output vector t i = [t i1 ,t i2 ,…,t im ] ∈R m , then the ELM neural network model with L hidden nodes and activation function g(x i ) is recorded as:

[0088]

[0089] Where β i is the output layer node weight; ω i is the input layer node weight; b i is the bias of the i-th hidden layer node; y j is the final output of the network, and N represents the number of samples.

[0090] The training target of ELM neural network is to obtain the minimum output error, that is:

[0091]

[0092] Where t i represents the expectation in the training process of sample i.

[0093] From formula (6), (7), the objective function of ELM neural network is:

[0094]

[0095] If H represents the output matrix of the hidden layer, the matrix form of formula (8) is:

[0096] βH = T (20)

[0097] ELM neural network maintains H unchanged during training, and β can be obtained by solving the least square solution of the following formula:

[0098] min||βH-T|| (21)

[0099] The solution of formula (9) is :

[0100]

[0101] Where H + is the Moore-Penrose generalized inverse matrix of H. Taking ELM as a predictor, the short-term wind power prediction result is obtained.

[0102] (4) Simulation calculation

[0103] As Figure 6 shown, in order to explore the influence of wind speed change feature recognition on short-term wind power prediction, a simulation experiment analysis was carried out on a wind farm with a capacity of 250 MW. The data resolution was 15 minutes. Based on the improved K-means clustering algorithm, the wind speed was divided into nine categories according to the change characteristics, including small wind fluctuation, medium wind fluctuation, large wind fluctuation, small amplitude rise, small amplitude decline, medium amplitude rise, medium amplitude decline, large amplitude rise, and large amplitude decline. After data anomaly processing, ELM was used to model short-term prediction under each type of wind speed fluctuation. The prediction results of each sub-model were reconstructed in time sequence to obtain the final prediction result. The prediction results were evaluated and analyzed according to the root mean square error (RMSE) and the mean absolute error (MAE).

[0104] Specific example analysis

[0105] As Figure 3 shown, the wind speed was divided into nine categories according to the change characteristics, including small wind fluctuation, medium wind fluctuation, large wind fluctuation, small amplitude rise, small amplitude decline, medium amplitude rise, medium amplitude decline, large amplitude rise, and large amplitude decline. Figure 4 The wind speed-power scatter plots corresponding to different wind speed change characteristics after data set division according to wind speed change feature recognition results are given. From the figure, it can be seen that the scatter distribution positions corresponding to different wind speed characteristics are obviously different, that is, the wind speed-power conversion relationship under different wind speed change characteristics is significantly different.

[0106] The basic prediction model is an ELM neural network, and a BP neural network is used as a comparative model. The prediction results of each model are the average of 10 times, and the number of neurons is gradually increased from 5 to 50, and the optimal result is shown. Based on the above analysis, a prediction model is established for each wind speed fluctuation characteristic data, and a feature value vector is added to the prediction model input for each wind speed fluctuation characteristic data. The final prediction result is recombined in time sequence. The improved combined prediction model is denoted as G-ELM and G-BP, and the prediction results are compared with the single prediction model ELM and BP. Figure 5 The prediction results of a certain day are given, and it can be seen from the figure that the combined prediction model is closer to the true value than the single prediction model.

[0107] Table 1 shows the wind power prediction error analysis results of each prediction period. As can be seen from the table, the prediction accuracy in June and September is higher, the RMSE of G-ELM is reduced by 0.86% and 0.89% compared with ELM, and the MAE is reduced by 0.76% and 1.08%. At the same time, the prediction accuracy of G-BP is also improved compared with BP. The prediction accuracy in March and December is low, the RMSE of G-ELM is reduced by 0.44% and 0.26% compared with ELM, and the MAE is reduced by 0.44% and 0.3%. The prediction accuracy of G-BP has not been significantly improved, but the MAE is reduced by 0.33% and 0.21%. Overall, the method of the present application can significantly improve the wind power prediction accuracy.

[0108] Table 1 shows the wind power prediction error analysis results of each prediction period. As can be seen from the table, the prediction accuracy in June and September is higher, the RMSE of G-ELM is reduced by 0.86% and 0.89% compared with ELM, and the MAE is reduced by 0.76% and 1.08%. At the same time, the prediction accuracy of G-BP is also improved compared with BP. The prediction accuracy in March and December is low, the RMSE of G-ELM is reduced by 0.44% and 0.26% compared with ELM, and the MAE is reduced by 0.44% and 0.3%. The prediction accuracy of G-BP has not been significantly improved, but the MAE is reduced by 0.33% and 0.21%. Overall, the method of the present application can significantly improve the wind power prediction accuracy.

[0109]

[0110] The embodiments of the present application are not exhaustive and do not limit the scope of the claims, and those skilled in the art can obtain other substantially equivalent alternatives based on the inspiration obtained from the examples of the present application without creative labor, which are within the scope of protection of the present application.

Claims

1.A short-term wind power prediction method considering wave process division, characterized in that: It comprises the following steps: 1) Abnormal data rejection and completion: First, filter the wind speed, if there are abnormal value points in the wind speed, the filtering error is large, then use the Shawiller method to identify the filtering error abnormal value points, and reject the corresponding wind speed abnormal data, secondly, use the grey correlation method to find the wind turbine with the strongest correlation with the wind speed data of the previous period of the missing data, and use the wind speed of the same period of the wind turbine to complete the missing data, if all wind turbines in the wind farm have abnormal output at the same period, use the cubic spline interpolation method to complete the missing data, thirdly, use the Copula function to fit the wind power probability distribution, and identify and reject the abnormal value points of the suspicious probability, then use the grey correlation method to find the wind turbine with the strongest correlation with the wind speed data of the same period of the missing data, and use the power data of the same period of the wind turbine to complete the missing data, if all wind turbines have abnormal output, use support vector machine to reconstruct the data, finally, use the processed data of each wind turbine to obtain the wind speed and overall power generation of the wind farm; 2) Wind speed change characteristic identification: The influence of wind speed change characteristics on power is not only related to the trend of wind speed change, but also closely related to the change amplitude of wind speed, the larger the change amplitude of wind speed, the more the wind turbine output deviates from the standard value, According to the principle of fluid mechanics, the actual output mechanical power P of the wind turbine M The corresponding relationship with the wind speed v is: where C P represents the wind energy utilization coefficient; v in represents the cut-in wind speed of the unit; v n represents the rated wind speed of the unit; v out represents the cut-out wind speed of the unit; p represents the air density; and S represents the impeller swept area. When v < v in or v > v n , the wind turbine output power is a fixed value, so there is no difference in the influence of different wind speed variation characteristics on power in these two wind speed intervals, but for the whole wind farm, since the wind speed has spatial dispersion, there will also be P > 0 when v < v in , so the upper limit v small_max of the small wind period should be less than v in ; similarly, there will also be P < P n when v > v n , so the lower limit v large_min of the large wind period should be greater than v n . From formula (1), compared with the first-order difference of wind speed Δv i , the first-order difference of wind speed cubed can better reflect the influence of different wind speed change characteristics on the power generation of the wind farm. If the wind speeds corresponding to the i time and the i-1 time are v i and v i-1 , respectively, then Δv i , The calculation formula is: An improved K-means clustering algorithm is used to identify the wind speed change characteristics: ① Selection of initial clustering center i. Based on the wind speed change feature recognition result exists "fluctuation", need to different lifting amplitude change feature analysis, therefore set (v w , 0) is the first initial clustering center X1, wherein v w is the Weibull function vertex corresponding to the wind speed value fitted to the wind speed sample probability distribution; ii. Calculate the clustering of any two cluster samples x i and x j The Euclidean distance d between them ij And the two sample points with the largest distance are used as the second and third initial cluster centers, since vector v i sum vector Compared to smaller numerical values, the vectors corresponding to the two farthest sample points are... Given one positive value and one negative value, select... The sample points that are positive are taken as X2. The sample points when d is negative are taken as X3, where d ij The calculation formula is: iii.In the remaining cluster samples, select X4 as the sample point with the largest distance product of X1 and X2; select X5 as the sample point with the largest distance product of X1 and X3; and sequentially derive the remaining (K-5) initial cluster centers. To maintain the symmetry of the wind speed variation feature recognition, the K value can only be an odd number; ② Selection of optimal K value Elbow method is an effective means to determine the optimal clustering number K, when K value is less than the optimal clustering number, with the increase of K value, the similarity of samples in each category improves rapidly, SSE decreases significantly, but with the increase of K value, the decrease amplitude of SSE gradually decreases, the change curve is similar to the shape of elbow, the optimal clustering number K is the K value corresponding to the elbow, and the expression of error square sum SSE is: where C n is the nth cluster class; L n is the C n number of samples within the cluster; is the C n cluster center; x i is the C n ith sample within the cluster; The wind speed-power conversion characteristics can be reflected by the wind speed-power scatter distribution, and the wind speed-power scatter distribution is a two-dimensional vector After K-means clustering, the wind speed v i The corresponding simultaneous power data P i According to the clustering results, n groups of two-dimensional vectors (v i , P i ) are constructed, and the SSE is calculated by formula (5). The wind speed change characteristic identification based on the improved K-means clustering algorithm is: First, v i <v small_max Data defined as small fluctuations in the wind, v i >v large_min Data defined as large fluctuations in the wind, changes in wind speed within the two ranges have no effect on the wind speed-power conversion relationship; in two-dimensional vectors As clustering samples, the improved K-means clustering algorithm is used to cluster v min <v i <v max The data is subjected to wind speed change feature recognition, and the wind speed change features can be classified according to the wind speed-power conversion relationship; The characteristic value is marked for each wind speed variation characteristic, the wind speed fluctuation characteristic data is marked as 0, and the wind speed rise and fall characteristic data is marked as the mean value of the first-order difference of wind speed in each data set. 3) Short-term wind power prediction model based on extreme learning machine: Extreme learning machine (ELM) can randomly generate initial weight and hidden layer node parameters, and use least square method to calculate the weight of output layer, but in order to obtain the optimal solution, ELM neural network needs to adjust the number of hidden layer neurons, If there are N groups of training samples (x i ,t i ), where the input vector x i =[x i1 ,x i2 ,…,x in ] T ∈R n , the output vector t i =[t i1 ,t i2 ,…,t im ]∈R m , and there are L hidden nodes, the ELM neural network model with the activation function g(x i ) is denoted as: where β i is the output layer node weight; ω i is the input layer node weight; b i is the i-th hidden layer node bias; y j is the network final output, N represents the number of samples, The training target of ELM neural network is to obtain the minimum output error, that is: where t i represents the expectation during the training process for sample i; From formulas (6) and (7), the objective function of ELM neural network is: If H represents the hidden layer output matrix, the matrix form of formula (8) is: βH=T (9) ELM neural network maintains H unchanged during training, β can be obtained by solving the least square solution of the following formula: min||βH-T|| (10) The result of solving equation (9) is: where H + is the Moore-Penrose generalized inverse matrix of H, and the short-term wind power prediction result is obtained by taking the ELM as a predictor.

Images