A wind farm wind power prediction method and system based on spatiotemporal correlation

By using the Pearson correlation coefficient method to screen characteristic factors and LSTM and GCN models in wind farms for spatiotemporal prediction, combined with DBSCAN and K-means algorithms to process abnormal data, the problem of insufficient power prediction accuracy of traditional wind power is solved, and more efficient wind power is achieved Power prediction.

CN118300089BActive Publication Date: 2024-11-29ZHEJIANG UNIV +1
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202410406819.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-07
Publication Date
2024-11-29
Estimated Expiration
2044-04-07

AI Technical Summary

Technical Problem

Traditional wind power power prediction methods fail to effectively utilize the spatial and temporal correlation of wind farms, resulting in insufficient prediction accuracy of wind power power and unable to accurately reflect the correlation between wind speed and power at different locations in the wind farm.

Method used

The Pearson correlation coefficient method is used to screen the characteristic factors of strong correlation between wind power, and the LSTM and GCN models are used to extract and predict the wind power in time and space. The abnormal data is processed by combining DBSCAN and K-means algorithms, and the variance-cooperative The variance-weighted combination prediction method combines space-time information to improve the prediction accuracy of wind power power.

Benefits of technology

Through the combination of spatiotemporal correlation analysis and deep learning models, the accuracy and efficiency of wind power prediction are significantly improved, and abnormal data is accurately identified and corrected, data redundancy is reduced, and prediction accuracy is improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118300089B_ABST
    Figure CN118300089B_ABST
Patent Text Reader

Abstract

The invention discloses a wind power prediction method and system for a wind farm based on spatiotemporal correlation, relates to the technical field of wind power prediction, obtains a historical wind power data set; identifies and corrects the historical wind power data set to obtain an optimized wind power data set; uses the Pearson correlation coefficient method to quantitatively analyze the relationship between the characteristic factors affecting wind power and wind power, and screens out characteristic factors with strong correlation with wind power; constructs a wind power combined prediction model based on LSTM and GCN, takes the characteristic factors with strong correlation with wind power as model input, uses LSTM and GCN to extract and predict the characteristics of wind power in time and space respectively, and then fuses the prediction results to obtain the wind power prediction result. Based on the spatiotemporal correlation between wind turbine powers, the invention reasonably uses the information of reference stations to effectively improve the wind power prediction accuracy of the target wind farm.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and more particularly to a method and system for predicting wind power in a wind farm based on time-space correlation. Background Art

[0002] As the global consensus on the development of renewable energy continues to grow, wind power will play a more important role in future energy and power systems. Real-time and accurate wind power forecasting can help power system operators effectively respond to drastic wind power fluctuations, quickly and accurately make decisions on dispatch control and power trading, and effectively alleviate the strong uncertainty brought about by large-scale wind power grid connection to power system operation. It has important economic significance and academic value.

[0003] Traditional wind power prediction methods are basically based on the data relationship of time series for modeling. Various methods are used to mine the time demonstration law from historical data to future data or the mapping law from meteorological data to power data, and then predict the future wind speed or power of the target wind farm. Few methods integrate the spatial correlation characteristics of wind turbines for prediction. In fact, the wind speed and power of wind farms are not statistically independent in space, but have a certain correlation. The wind speed and power values ​​at a specific location are spatially affected by other wind farms or wind turbines at different locations (such as wake effects, etc.), the topological structure of the space (such as the arrangement of wind turbines, the terrain of wind farms, etc.), and environmental conditions (such as air pressure, temperature, etc.).

[0004] Therefore, how to provide a wind farm wind power prediction method and system based on spatiotemporal correlation, based on the spatiotemporal correlation between wind turbine powers, rationally utilize the information of reference stations, and effectively improve the wind power prediction accuracy of the target wind farm is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] In view of this, the present invention provides a wind farm wind power prediction method and system based on spatiotemporal correlation, which rationally utilizes the information of reference stations based on the spatiotemporal correlation between wind turbine powers and effectively improves the wind power prediction accuracy of the target wind farm.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a wind farm wind power prediction method based on spatiotemporal correlation, comprising:

[0007] Get historical wind power datasets of wind farms;

[0008] Identifying and correcting the historical wind power data set to obtain an optimized wind power data set;

[0009] The Pearson correlation coefficient method is used to quantitatively analyze the relationship between the characteristic factors affecting wind power and wind power, and the characteristic factors with strong correlation with wind power are screened out;

[0010] A combined wind power prediction model is constructed based on LSTM and GCN. The characteristic factors that have a strong correlation with wind power are used as model inputs. LSTM and GCN are used to extract and predict wind power features in time and space respectively. The prediction results in time and space are then fused to obtain the wind power prediction result.

[0011] Preferably, the historical wind power data set is identified and corrected, including: using a DBSCAN algorithm to identify abnormal wind power data;

[0012] The K-means algorithm is used to cluster all wind turbines in the wind farm according to the measured power data, and the abnormal wind power data is replaced by the sample mean within the same cluster at the same time.

[0013] Preferably, the specific steps of using the DBSCAN algorithm to identify abnormal wind power data are as follows:

[0014] A sample point p is randomly selected from the historical wind power data set, and the neighborhood radius ε and the minimum number of sample points M are set. ε(p) represents the neighborhood with p as the center and ε as the radius.

[0015] If there are at least M data points in the neighborhood, then the sample point p is a core object;

[0016] If the number of data points contained in the neighborhood is less than M, and the sample point p is not located in other neighborhoods, then the sample point p is abnormal wind power data.

[0017] Preferably, the neighborhood radius ε and the minimum number of sample points M are determined by a k-dist graph method;

[0018] Calculate the value of each object p in the historical wind power dataset i With the rest of the objects p j The Euclidean distance between the two objects, as the i k-dist value of

[0019] Arrange the k-dist values ​​of all objects in ascending or descending order and draw the k-dist curve;

[0020] Get the neighborhood radius ε and the minimum number of sample points M.

[0021] Preferably, a K-means algorithm is used to cluster all wind turbines in the wind farm according to the measured power data, and the abnormal wind power data is replaced with the sample mean in the cluster at the same time, including:

[0022] Based on the historical wind power dataset, the measured power data of all wind turbines in the wind farm within the t time period is intercepted to construct a sample matrix; the sample matrix P is expressed as:

[0023]

[0024] where P i,j represents the power measured by the i-th wind turbine at the j-th moment, m represents the total number of wind turbines; k represents the number of samplings;

[0025] The sample matrix includes multiple sample groups; all the sample groups included in the sample matrix are divided into c classes, 1 < c < m, c ∈ N, and any c sample groups are selected as the initial clustering centers of the c classes Group a Calculate the Euclidean distance from any sample group to the a-th initial clustering center;

[0026] After determining the Euclidean distance from the sample group to the a-th initial clustering center, the sample group is assigned to the class Group a with the smallest Euclidean distance. Traverse all sample groups to complete the first division;

[0027] Calculate the sample mean of each class and the total sum of squared errors of clustering respectively;

[0028] Use the sample mean of each class as the new initial clustering center, and repeat the above operations until the total sum of squared errors of clustering converges and the clustering is completed;

[0029] Replace the abnormal wind power data with the sample mean within the class at the same time.

[0030] Preferably, the calculation formula of the Pearson correlation coefficient method is expressed as:

[0031]

[0032] where Y i is the wind power data sequence, X i is the characteristic factor data sequence affecting the wind power, n is the length of the sequence, ρ X,Y represents the degree of correlation between the characteristic factor variable X and the wind power Y;

[0033] According to the preset threshold, when ρ X,Y is greater than the preset threshold, it indicates that the characteristic factor variable X and the wind power Y have a strong correlation;

[0034] Select the characteristic factors with strong correlation as the input characteristic variables of the wind power combined prediction model. To reduce data redundancy and time consumption.

[0035] Preferably, LSTM and GCN are used to extract and predict wind power features in time and space respectively, including:

[0036] The LSTM model is used to predict the wind power at a certain moment in the future and obtain the wind power time prediction result; LSTM has the ability to retain the information of time nodes that are far away from the current time node, and can better process time series data with a long time span.

[0037] Taking the wind turbine feature matrix and the weighted wind turbine adjacency matrix as input, the GCN model is used to predict the wind power at a certain moment in the future and obtain the wind power spatial prediction result;

[0038] The weighted wind turbine adjacency matrix is ​​expressed as:

[0039] A={R i1 ,R i2 ,…,R ij};

[0040] In the formula, R ij represents the spatial correlation between fan No. i and fan No. j, and the spatial correlation is calculated according to the Pearson correlation coefficient method;

[0041] The fan characteristic matrix is ​​expressed as:

[0042]

[0043] In the formula, represents the measured power value of wind turbine i at time t in the past. That is, each row of X represents a wind turbine, and each column represents the power value of different wind turbines at a certain moment in history.

[0044] Preferably, the prediction results in time and space are integrated to obtain the wind power prediction result, including:

[0045] The variance-covariance weighted combination prediction method is used to calculate the weight coefficients of LSTM and GCN respectively, and the power prediction value at a certain moment in the future is output according to the calculated weights;

[0046]

[0047] In the formula, Represents the sequence of power prediction values ​​of LSTM and GCN at a certain moment in the future; y t represents a sequence consisting of the weighted average forecast values ​​of the forecast results of two single forecasting methods; w T and w S Respectively and The weight coefficient ofT +w S =1.

[0048] Preferably, the variance-covariance weighted combination prediction method is used to calculate the weight coefficients of LSTM and GCN respectively, including:

[0049] set up is a sequence of measured power data at time t, then:

[0050]

[0051]

[0052]

[0053] In the formula, e t There are three prediction results: y t The error sequence of

[0054] get:

[0055]

[0056] Then t The variance of is expressed as:

[0057]

[0058] w T and w S The choice should make the prediction error e t minimum; so about w T Var(e t ) to find the minimum value, we get:

[0059]

[0060] w S =1-w T .

[0061] Preferably, a wind farm wind power prediction system based on spatiotemporal correlation comprises:

[0062] A data acquisition module, used to acquire historical wind power data sets of wind farms;

[0063] A data processing module, used for identifying and correcting the historical wind power data set to obtain an optimized wind power data set;

[0064] The feature selection module is used to quantitatively analyze the relationship between the characteristic factors affecting wind power and wind power using the Pearson correlation coefficient method, and screen out characteristic factors that have a strong correlation with wind power;

[0065] Model building module, used to build a wind power combination prediction model based on LSTM and GCN;

[0066] The power prediction module is used to take the characteristic factors that have a strong correlation with wind power as model input, use LSTM and GCN to extract and predict the characteristics of wind power in time and space respectively, and then fuse the prediction results in time and space to obtain the wind power prediction result.

[0067] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a wind farm wind power prediction method and system based on spatiotemporal correlation, which can effectively improve the wind power prediction accuracy by introducing spatiotemporal related information of the wind farm.

[0068] The density-based clustering method in noisy space (DBSCAN) can accurately identify high-density areas and separate them from other areas, making it relatively easy and accurate to identify abnormal wind power data; the abnormal data correction method based on K-means spatial clustering uses the measured power as a feature to cluster wind turbines, and replaces the abnormal feature value with the mean value within the same class at the same time, which can effectively integrate spatiotemporal features to improve the accuracy of abnormal wind power data correction.

[0069] The Pearson correlation coefficient method is used to quantitatively analyze the relationship between various characteristic factors that may affect wind power and wind power, and to screen out characteristics that have a strong correlation with wind power as the input of the model, which can effectively improve the prediction accuracy and efficiency while streamlining the model and data.

[0070] A combined prediction model based on LSTM and GCN is used to predict wind power with spatiotemporal fusion. LSTM is used to extract the time characteristics of power, GCN is used to extract the spatial characteristics between wind turbine powers, and then the spatiotemporal characteristics are integrated through the variance-covariance combined prediction method to complete the final wind power prediction, which can make full use of spatiotemporal information to improve the accuracy of wind power prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0072] Figure 1 A schematic flow chart of a method for predicting wind power in a wind farm based on spatiotemporal correlation is provided in an embodiment of the present invention.

[0073] Figure 2 A schematic diagram of the structure of a wind power combination prediction model based on LSTM and GCN provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0074] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] The embodiment of the present invention discloses a wind farm wind power prediction method and system based on spatiotemporal correlation, which effectively improves the wind power prediction accuracy of a target wind farm based on the spatiotemporal correlation between wind turbine powers and reasonable use of reference station information.

[0076] Specifically, the embodiment of the present invention discloses a wind farm wind power prediction method based on spatiotemporal correlation, such as Figure 1 As shown, including:

[0077] S100, obtaining a historical wind power data set of the wind farm;

[0078] S200, identifying and correcting the historical wind power data set to obtain an optimized wind power data set;

[0079] S300, using the Pearson correlation coefficient method to quantitatively analyze the relationship between the characteristic factors affecting wind power and wind power, and screen out characteristic factors that have a strong correlation with wind power;

[0080] S400. Construct a combined wind power prediction model based on LSTM and GCN, take the characteristic factors that have a strong correlation with wind power as model input, use LSTM and GCN to extract and predict wind power features in time and space respectively, and then fuse the prediction results in time and space to obtain a wind power prediction result.

[0081] Specifically, the historical wind power data set is identified and corrected, including: using a DBSCAN algorithm to identify abnormal wind power data;

[0082] The K-means algorithm is used to cluster all wind turbines in the wind farm according to the measured power data, and the abnormal wind power data is replaced by the sample mean within the same cluster at the same time.

[0083] Specifically, the embodiment of the present invention adopts the DBSCAN algorithm to identify abnormal wind power data. During the operation of the wind turbine, the normal data recorded by the SCADA (Supervisory Control And Data Acquisition) system will be more closely distributed around the power curve, while the abnormal data is usually generated by a unit failure or abnormal state that cannot be instantly restored, and has uncertainty and randomness, presenting a dispersed distribution form. Therefore, the embodiment of the present invention adopts a density-based noisy space clustering method (DBSCAN) to accurately identify high-density areas and separate them from other areas, thereby relatively easily identifying abnormal wind power data. The specific steps of using the DBSCAN algorithm to identify abnormal wind power data are as follows:

[0084] S211. For a given historical wind power data set D, set an appropriate neighborhood radius ε and a minimum number of sample points M, where ε(p) represents a neighborhood with p as the center and ε as the radius, and M represents the minimum number of samples contained in ε(p).

[0085] S212. Randomly select a sample point p from the historical wind power data set. If there are at least M data points in the neighborhood, the sample point p is a core object; if the neighborhood contains less than M data points and the sample point p is not located in other neighborhoods, the sample point p is a noise point, that is, abnormal wind power data.

[0086] S213, randomly select unprocessed points, and repeat the process of S212 until all sample points in the historical wind power data set D are processed.

[0087] In general, the neighborhood radius ε and the minimum number of sample points M are determined by the k-dist graph method.

[0088] Specifically, S2111, calculate the p of each object in the historical wind power data set i With the rest of the objects p j The Euclidean distance between the two objects, as the i k-dist value of

[0089]

[0090] Among them, n represents the data dimension; that is, the characteristic number (including unit status data and meteorological data, such as temperature, wind speed, pitch angle, etc.).

[0091] S2112. Arrange the k-dist values ​​of all objects in ascending or descending order and draw a k-dist curve.

[0092] S2113. Obtain the neighborhood radius ε and the minimum number of samples M.

[0093] Specifically, take the distance corresponding to the first "inflection point" of the k-dist curve as the neighborhood radius ε, and take k + 1 as the minimum number of samples M. (To reduce the computational amount, generally default k = 2n - 1).

[0094] Specifically, use the K-means algorithm to cluster and divide all wind turbines in the wind farm according to the measured power data, and replace the abnormal wind power data with the sample mean within the class at the same time, including:

[0095] S221. Based on the historical wind power data set, intercept the measured power data of all wind turbines in the wind farm within the t time period to construct a sample matrix; the sample matrix P is expressed as: (assuming that all wind turbines in the wind farm have been operating normally in the past t time period, and the total number of units is m; the sampling interval of the power data in the past t time period is fixed, and the number of samples is k):

[0096]

[0097] where P i,j represents the power measured by the i-th wind turbine at the j-th moment, m represents the total number of wind turbines; k represents the number of samples; regarding the vector P i as a sample group in the space P, then the number of samples in the sample matrix P is m, and the dimension is k.

[0098] S222. The sample matrix P includes multiple sample groups; divide all the sample groups included in the sample matrix P into c classes, that is, c clusters, 1 < c < m, c ∈ N, and arbitrarily select c sample groups as the initial clustering centers P a of the c classes Group a , and calculate the Euclidean distance from any sample group P i to the a-th initial clustering center;

[0099]

[0100] In the formula, P i represents a sample group of the sample matrix P; P a represents the initial clustering center; a represents the a-th initial clustering center, a = 1, 2,..., c;

[0101] After determining the Euclidean distance d(P i ,P i ,P a ) of the sample group P i to the a-th initial clustering center, divide the sample group P i ,Pa )The smallest class Group a In , all sample groups are traversed to complete the first division.

[0102] S223, respectively calculating the sample mean of each class and the sum of square errors of the clusters;

[0103] Calculating the sample mean of each class is expressed as:

[0104]

[0105] in, is the sample mean of the ath class; N a is the total number of samples of the ath class;

[0106] The sum of squared errors for clustering is calculated as:

[0107]

[0108] Where E is the sum of squared errors of clustering.

[0109] S224, using the sample mean of each class as the new initial cluster center, repeating steps S222-S223 until the sum of square errors of clustering converges and clustering is completed; replacing the abnormal wind power data with the sample mean within the class at the same time, the specific formula is as follows:

[0110]

[0111] in, represents the measured value of the nth characteristic of the i-th wind turbine at the k-th moment; Indicates the correction value of abnormal wind power data; N a Represents the total number of samples in the a-th class.

[0112] Specifically, the embodiment of the present invention adopts the Pearson correlation coefficient method to quantitatively analyze the relationship between various characteristic factors that may affect wind power (such as unit status data, meteorological data, etc.) and wind power, and selects characteristics that have a strong correlation with wind power as the input of the model. The Pearson correlation coefficient can accurately quantify the degree of correlation between two variables.

[0113] Specifically, the calculation formula of the Pearson correlation coefficient method is expressed as:

[0114]

[0115] Among them, Y i is the wind power data series, X i is the data sequence of characteristic factors affecting wind power, n is the length of the sequence, ρX,Y Indicates the correlation between characteristic factor variable X and wind power Y;

[0116] According to the preset threshold, when ρ X,Y When it is greater than the preset threshold, it means that the characteristic factor variable X has a strong correlation with the wind power Y; the preset threshold is 0.6;

[0117] Characteristic factors with strong correlation are selected as input characteristic variables of the wind power combination prediction model.

[0118] When 0.6<ρ X,Y ≤1.0, indicating that the variable has a strong correlation with wind power; when 0.2<ρ X,Y When ρ ≤0.6, it indicates that the variable is moderately correlated with wind power; when ρ X,Y ≤0.2, indicating that the variable is weakly correlated or has no correlation with wind power; the embodiment of the present invention selects ρ X,Y When the value is greater than 0.6, the characteristic variables with strong correlation with wind power are used as the input characteristic variables of the final wind power prediction model to reduce data redundancy and time consumption.

[0119] Specifically, the wind power combination prediction model mainly includes three modules: a time feature extraction prediction module, a space feature extraction prediction module and a time-space feature fusion prediction module. The model structure is as follows: Figure 2 shown.

[0120] (1) Time feature extraction and prediction module: This module is mainly composed of LSTM, which has the ability to retain information about time nodes that are relatively long from the current time node, and can better process time series data with a relatively long time span. In this module, the embodiment of the present invention uses the historical wind power data of the wind turbine H time steps obtained after preprocessing as input (including multivariate sequences such as unit status data and meteorological data), and uses the LSTM basic model to predict the power at p moments in the future, and outputs

[0121] (2) Spatial feature extraction prediction module: This module is mainly composed of GCN, which is a deep learning method for processing non-Euclidean spatial graph data. It aggregates the features of its own and neighboring nodes through convolution operations in the adjacency matrix to generate feature vectors of new nodes. In this module, the embodiment of the present invention takes the weighted wind turbine adjacency matrix A and the wind turbine power matrix X as input, and uses the GCN basic model to output power prediction for the next p moments.

[0122] Among them, LSTM and GCN are used to extract and predict wind power features in time and space respectively, including:

[0123] Use the LSTM model to predict the wind power at a certain moment in the future and obtain the wind power time prediction result;

[0124] Taking the wind turbine feature matrix and the weighted wind turbine adjacency matrix as input, the GCN model is used to predict the wind power at a certain moment in the future and obtain the wind power spatial prediction result;

[0125] The weighted wind turbine adjacency matrix is ​​expressed as:

[0126] A={R i1 ,R i2 ,…,R ij};

[0127] In the formula, R ij represents the spatial correlation between fan No. i and fan No. j, which is calculated according to the Pearson correlation coefficient method and is expressed as:

[0128]

[0129] Among them, P i is the wind power data sequence of wind turbine i; P j is the wind power data sequence of wind turbine j, and n is the sequence length.

[0130] The fan characteristic matrix is ​​expressed as:

[0131]

[0132] In the formula, represents the measured power value of wind turbine i at time t in the past. That is, each row of X represents a wind turbine, and each column represents the power value of different wind turbines at a certain moment in history.

[0133] (3) Temporal and spatial feature fusion prediction module: The prediction results in time and space are integrated to obtain the wind power prediction results, including:

[0134] The variance-covariance weighted combined prediction method is used to calculate the weight coefficients of LSTM and GCN respectively. According to the calculated weights, the power prediction values ​​at the next p moments are output. The established combined prediction model is expressed as:

[0135]

[0136] In the formula, Represents the sequence of power prediction values ​​of the next p moments by two single prediction methods: LSTM and GCN; y t represents a sequence consisting of the weighted average forecast values ​​of the forecast results of two single forecasting methods; w T and wS Respectively and The weight coefficient of T +w S =1.

[0137] Among them, the variance-covariance weighted combination prediction method is used to calculate the weight coefficients of LSTM and GCN respectively, including:

[0138] set up is a sequence of measured power data at time t, then:

[0139]

[0140]

[0141]

[0142] In the formula, e t There are three prediction results: y t The error sequence of

[0143] get:

[0144]

[0145] Then t The variance of is expressed as:

[0146]

[0147] w T and w S The choice should make the prediction error e t minimum; so about w T Var(e t ) to find the minimum value, we get:

[0148]

[0149] w S =1-w T .

[0150] The embodiment of the present invention discloses a wind farm wind power prediction method based on spatiotemporal correlation, which can effectively improve the wind power prediction accuracy by introducing the spatiotemporal related information of the wind farm. First, the DBSCAN algorithm is used to identify abnormal wind power data, and the wind turbines are clustered by the K-means algorithm, and the abnormal feature values ​​are replaced with the mean values ​​in the same class at the same time to obtain an optimized wind power data set; then, the Pearson correlation coefficient method is used to quantitatively analyze the relationship between various characteristic factors that may affect wind power and wind power, and the features with strong correlation with wind power are selected as the input of the model; finally, LSTM and GCN are used to extract and predict features in time and space respectively, and then the single prediction results are fused through the variance-covariance combined prediction method to realize the wind power prediction of the wind farm in a certain period of time in the future.

[0151] The DBSCAN and K-means algorithms are used to identify and correct abnormal wind power data. The density-based clustering method with noise (DBSCAN) can accurately identify high-density areas and separate them from other areas, so that abnormal wind power data can be identified relatively easily and accurately. The abnormal data correction method based on K-means spatial clustering clusters the wind turbines with measured power as a feature, and replaces the abnormal feature value with the mean value within the same class at the same time, which can effectively integrate the spatiotemporal features to improve the accuracy of abnormal wind power data correction.

[0152] The Pearson correlation coefficient method is used for sparse modeling and feature selection. The Pearson correlation coefficient method is used to quantitatively analyze the relationship between various characteristic factors that may affect wind power and wind power, and to select features that have a strong correlation with wind power, that is, features with a Pearson correlation coefficient greater than 0.6, as the input of the model, which can effectively improve the prediction accuracy and efficiency while streamlining the model and data.

[0153] A combined prediction model based on LSTM and GCN is used to predict wind power with spatiotemporal fusion. LSTM is used to extract the time characteristics of power, GCN is used to extract the spatial characteristics between wind turbine powers, and then the spatiotemporal characteristics are integrated through the variance-covariance combined prediction model to complete the final wind power prediction, which can make full use of spatiotemporal information to improve the accuracy of wind power prediction.

[0154] In another specific embodiment of the present invention, a wind farm wind power prediction system based on spatiotemporal correlation includes:

[0155] A data acquisition module, used to acquire historical wind power data sets of wind farms;

[0156] The data processing module is used to identify and correct the historical wind power data set to obtain an optimized wind power data set; by analyzing the data characteristics, a data-driven combination method is used to identify and correct abnormal wind power data, effectively improving the quality of wind farm operation data and providing good data support for the development of subsequent power prediction models.

[0157] The feature selection module is used to quantitatively analyze the relationship between the characteristic factors affecting wind power and wind power using the Pearson correlation coefficient method, and screen out characteristic factors that have a strong correlation with wind power; through the analysis of characteristic factors, the Pearson coefficient is used to quantitatively analyze wind power and characteristic factors, and the characteristic factors that have a greater impact on wind power prediction are screened out as the input of the prediction model, thereby improving the prediction accuracy and efficiency.

[0158] Model building module, used to build a wind power combination prediction model based on LSTM and GCN;

[0159] The power prediction module is used to take the characteristic factors that have a strong correlation with wind power as model input, use LSTM and GCN to extract and predict the characteristics of wind power in time and space respectively, and then fuse the prediction results in time and space to obtain the wind power prediction result.

[0160] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0161] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A wind farm wind power prediction method based on spatiotemporal correlation, It is characterized in that include: Get historical wind power datasets of wind farms; Identifying and correcting the historical wind power data set to obtain an optimized wind power data set; The Pearson correlation coefficient method is used to quantitatively analyze the relationship between the characteristic factors affecting wind power and wind power, and the characteristic factors with strong correlation with wind power are screened out; A wind power combined prediction model is constructed based on LSTM and GCN, the characteristic factors having a strong correlation with wind power are used as model input, LSTM and GCN are used to extract and predict wind power features in time and space respectively, and then the prediction results in time and space are integrated to obtain the wind power prediction result; Among them, the prediction results in time and space are integrated to obtain the wind power prediction results, including: The variance-covariance weighted combination prediction method is used to calculate the weight coefficients of LSTM and GCN respectively, and the power prediction value at a certain moment in the future is output according to the calculated weights; In the formula, Represents the sequence of power prediction values ​​of LSTM and GCN at a certain moment in the future; y t represents a sequence consisting of the weighted average forecast values ​​of the forecast results of two single forecasting methods; w T and w S Respectively and The weight coefficient of T +w S =1; The variance-covariance weighted combination prediction method is used to calculate the weight coefficients of LSTM and GCN respectively, including: set up is a sequence of measured power data at time t, then: In the formula, e t There are three prediction results: y t The error sequence of get: Then t The variance of is expressed as: About w T Var(e t ) to find the minimum value, we get: In S =1-in T 。 2. A wind farm wind power prediction method based on spatiotemporal correlation according to claim 1, It is characterized in that Identifying and correcting the historical wind power data set includes: using a DBSCAN algorithm to identify abnormal wind power data; The K-means algorithm is used to cluster all wind turbines in the wind farm according to the measured power data, and the abnormal wind power data is replaced by the sample mean within the same cluster at the same time.

3. A wind farm wind power prediction method based on spatiotemporal correlation according to claim 2, It is characterized in that The specific steps of using the DBSCAN algorithm to identify abnormal wind power data are as follows: A sample point p is randomly selected from the historical wind power data set, and the neighborhood radius ε and the minimum number of sample points M are set. ε(p) represents the neighborhood with p as the center and ε as the radius. If there are at least M data points in the neighborhood, then the sample point p is a core object; If the number of data points contained in the neighborhood is less than M, and the sample point p is not located in other neighborhoods, then the sample point p is abnormal wind power data.

4. A wind farm wind power prediction method based on spatiotemporal correlation according to claim 3, It is characterized in that The neighborhood radius ε and the minimum number of sample points M are determined by the k-dist graph method; Calculate the value of each object p in the historical wind power dataset i With the rest of the objects p j The Euclidean distance between the two objects, as the i k-dist value of Arrange the k-dist values ​​of all objects in ascending or descending order and draw the k-dist curve; Get the neighborhood radius ε and the minimum number of sample points M.

5. A wind farm wind power prediction method based on spatiotemporal correlation according to claim 2, It is characterized in that The K-means algorithm is used to cluster all wind turbines in the wind farm according to the measured power data, and the abnormal wind power data is replaced with the sample mean in the same class at the same time, including: Based on the historical wind power data set, the measured power data of all wind turbines in the wind farm within the time period t are intercepted to construct a sample matrix; the sample matrix P is expressed as: Among them, P i,j represents the power measured by the i-th wind turbine at the j-th moment, m represents the total number of wind turbines; k represents the number of samples; The sample matrix includes multiple sample groups; all the sample groups included in the sample matrix are divided into c classes, where 1 < c < m and c ∈ N. Arbitrarily select c sample groups as the initial clustering centers of the c classes Group a and calculate the Euclidean distance from any one sample group to the a-th initial clustering center; After determining the Euclidean distance from the sample group to the ath initial cluster center, divide the sample group into the group with the smallest Euclidean distance. a In , traverse all sample groups and complete the first division; Calculate the sample mean of each class and the sum of squared errors of the clusters separately; The sample mean of each class is used as the new initial cluster center, and the above operation is repeated until the sum of square errors of clustering converges and clustering is completed; The abnormal wind power data is replaced by the sample mean within the class at the same moment.

6. A wind farm wind power prediction method based on spatiotemporal correlation according to claim 1, It is characterized in that The calculation formula of the Pearson correlation coefficient method is expressed as: Among them, Y i is the wind power data series, X i is the data sequence of characteristic factors affecting wind power, n is the length of the sequence, ρ X,Y Indicates the correlation between characteristic factor variable X and wind power Y; According to the preset threshold, when ρ X,Y When it is greater than the preset threshold, it means that the characteristic factor variable X has a strong correlation with the wind power Y; Characteristic factors with strong correlation are selected as input characteristic variables of the wind power combination prediction model.

7. A wind farm wind power prediction method based on spatiotemporal correlation according to claim 1, It is characterized in that LSTM and GCN are used to extract and predict wind power features in time and space, including: Use the LSTM model to predict the wind power at a certain moment in the future and obtain the wind power time prediction result; Taking the wind turbine feature matrix and the weighted wind turbine adjacency matrix as input, the GCN model is used to predict the wind power at a certain moment in the future and obtain the wind power spatial prediction result; The weighted wind turbine adjacency matrix is ​​expressed as: A={R i1 ,R i2 ,...,R ij }; In the formula, R ij represents the spatial correlation between fan No. i and fan No. j; the spatial correlation is calculated according to the Pearson correlation coefficient method; The fan characteristic matrix is ​​expressed as: In the formula, Represents the measured power value of wind turbine No. i at time t in the past.

8. A wind farm wind power prediction system based on spatiotemporal correlation, It is characterized in that include: A data acquisition module, used to acquire historical wind power data sets of wind farms; A data processing module, used for identifying and correcting the historical wind power data set to obtain an optimized wind power data set; The feature selection module is used to quantitatively analyze the relationship between the characteristic factors affecting wind power and wind power using the Pearson correlation coefficient method, and screen out characteristic factors that have a strong correlation with wind power; Model building module, used to build a wind power combination prediction model based on LSTM and GCN; A power prediction module is used to use the characteristic factors that have a strong correlation with wind power as model input, use LSTM and GCN to extract and predict wind power features in time and space respectively, and then fuse the prediction results in time and space to obtain a wind power prediction result; Among them, the prediction results in time and space are integrated to obtain the wind power prediction results, including: The variance-covariance weighted combination prediction method is used to calculate the weight coefficients of LSTM and GCN respectively, and the power prediction value at a certain moment in the future is output according to the calculated weights; In the formula, Represents the sequence of power prediction values ​​of LSTM and GCN at a certain moment in the future; y t represents a sequence consisting of the weighted average forecast values ​​of the forecast results of two single forecasting methods; w T and w S Respectively and The weight coefficient of T +w S =1; The variance-covariance weighted combination prediction method is used to calculate the weight coefficients of LSTM and GCN respectively, including: set up is a sequence of measured power data at time t, then: In the formula, e t There are three prediction results: y t The error sequence of ; we get: Then t The variance of is expressed as: About w T Var(e t ) to find the minimum value, we get: In S =1-in T 。

Citation Information

Patent Citations

  • Big data feature cleaning method and system for neural network training

    CN116166960A