Wind power ultra-short-term prediction method based on multi-graph convolution and time sequence feature extraction

Abstract

A wind power ultra-short-term prediction method based on multi-graph convolution and time sequence feature extraction belongs to the field of regional wind power prediction, and comprises the following steps: extracting historical wind speed data and historical power data of each wind power station in a regional wind power cluster, and dividing the regional wind power cluster into a plurality of sub-regions; taking the sub-regions of the regional wind power cluster as graph nodes, and constructing a multi-graph convolutional neural network from different dimensions; and constructing a regional wind power ultra-short-term prediction model, dividing the extracted historical wind speed data and historical power data into a training set and a test set, training the model by the training set, and performing performance test and verification on the model on the test set. The wind power ultra-short-term prediction precision is improved through an algorithm of combining multi-graph convolution and time sequence feature extraction.

Description

Technical Field

[0001] This invention belongs to the field of regional wind power prediction, specifically involving a wind power ultra-short-term prediction method based on multiple graph convolution and temporal feature extraction. Background Technology

[0002] As one of the core clean power sources in the new power system, wind power has seen continuous expansion in installed capacity and a rising grid connection rate. Its output characteristics have become a key factor affecting the safe and stable operation of the power grid. However, wind power output is significantly constrained by meteorological conditions. Sudden increases or decreases in wind speed, significant shifts in wind direction, strong wind shear, and extreme weather (such as low-temperature icing, typhoons, and prolonged periods without wind) can all lead to strong fluctuations, intermittency, and uncertainty in wind power output. In scenarios with a high proportion of wind power connected to the grid, these output fluctuations directly impact the grid's operational balance: when wind speeds decrease sharply or a windless period arrives, wind power output may decline rapidly or even approach zero, creating a power gap on the generation side and causing a mismatch between grid supply and demand; when encountering strong gusts or sudden changes in wind direction, sudden increases or decreases in output can easily cause grid frequency deviations to exceed allowable ranges, exacerbating voltage fluctuations, reducing grid voltage stability margins, significantly increasing the pressure on grid peak shaving and frequency regulation, and potentially affecting the reliable operation of relay protection systems, posing multiple risks to the safe operation of the power grid.

[0003] While existing research has constructed a series of prediction models based on graph neural networks from the perspective of spatiotemporal correlation extraction, it still lacks a comprehensive consideration of the multidimensional influencing factors affecting wind power generation, thus limiting the improvement of prediction accuracy. Therefore, there is an urgent need to design a wind power ultra-short-term prediction method based on multiple graph convolutions and temporal feature extraction to address these issues, fill the technological gap in this field, and provide support for grid dispatching and operation under the new power system context. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a wind power ultra-short-term prediction method based on multi-graph convolution and temporal feature extraction. By improving the algorithm, the accuracy of wind power ultra-short-term prediction is improved.

[0005] The technical solution adopted in this invention is: a wind power ultra-short-term prediction method based on multiple graph convolution and temporal feature extraction, comprising the following steps: Step S10: Extract historical wind speed data and historical power data of each wind farm in the regional wind power cluster, and divide the regional wind power cluster into multiple sub-regions; Step S20: Using sub-regions of the regional wind power cluster as graph nodes, construct a multi-graph convolutional neural network from different dimensions; Step S30: Construct a regional wind power ultra-short-term prediction model; This model uses a weighted summation method to fuse the outputs of multiple graph convolutions. The fused output is then used as the input to a gated recurrent unit, which extracts temporal features to obtain the final prediction result. Step S40: The extracted historical wind speed data and historical power data are divided into training set and test set. The training set is used to train the model, and the test set is used to test and verify the model's performance.

[0006] The present invention also provides a storage medium for storing a computer program, which is executed by a processor to implement the wind power ultra-short-term prediction method based on multi-graph convolution and temporal feature extraction.

[0007] The beneficial effects of this invention are as follows: By dividing wind power regions using a hierarchical clustering algorithm, the similarity characteristics of wind speed and power output of wind farms within the region are accurately captured, laying a precise data foundation for subsequent predictions; a graph convolutional neural network is constructed from multiple dimensions, including linear correlation, curve shape similarity, and geographical distance, to comprehensively integrate the diverse influencing factors of wind power output, breaking through the limitations of single-dimensional considerations in existing technologies and significantly improving prediction accuracy; by using weighted summation to achieve effective fusion of multiple graph convolution outputs, and combining gated recurrent units to efficiently extract temporal features, it can accurately adapt to the volatility, intermittency, and uncertainty of wind power output, significantly enhancing the stability and reliability of prediction results; the finally constructed ultra-short-term prediction model can provide high-precision wind power output prediction basis for grid dispatch, effectively reducing the risks of supply and demand mismatch and frequency and voltage fluctuations caused by wind power grid connection, alleviating the pressure of grid peak shaving and frequency regulation, and ensuring the safe and stable operation of the grid. Attached Figure Description

[0008] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0010] This invention provides a wind power ultra-short-term prediction method based on multiple graph convolution and temporal feature extraction, comprising the following steps: Step S10: Extract historical wind speed data and historical power data of each wind farm in the regional wind power cluster, and divide the regional wind power cluster into multiple sub-regions; Step S20: Using sub-regions of the regional wind power cluster as graph nodes, construct a multi-graph convolutional neural network from different dimensions; Step S30: Construct a regional wind power ultra-short-term prediction model; This model uses a weighted summation method to fuse the outputs of multiple graph convolutions. The fused output is then used as the input to a gated recurrent unit, which extracts temporal features to obtain the final prediction result. Step S40: The extracted historical wind speed data and historical power data are divided into training set and test set. The training set is used to train the model, and the test set is used to test and verify the model's performance.

[0011] The historical wind speed data includes measured wind speed and predicted wind speed, with a time resolution of 15 minutes and a time span of N days.

[0012] Step S10 includes: Step S101: Obtain wind speed data for each wind farm within the regional wind power cluster; The measured wind speed of the i-th power station can be denoted as: The predicted wind speed of the i-th power station is denoted as . ; The time series of measured and predicted wind speeds are normalized. , , In the formula, , It is the normalized time series of measured and predicted wind speeds for the i-th power station, representing different wind farms within the wind power cluster; Step S102: Extract the distribution characteristics of the measured wind speed data for each wind farm and construct a feature vector. , In the formula, , , These are the maximum, mean, and variance of the normalized wind speed for wind farm station numbered i, respectively. Hierarchical clustering algorithm is used to cluster the wind farms within the regional wind power cluster, grouping wind farms with similar measured wind speed characteristics into the same sub-cluster. The number of sub-clusters is set to M. The steps include: 1) Calculate the distance matrix between wind farms: Let the eigenvectors of wind farms i and j be Feat i =[max i ,mean i std i Feat j =[max j ,mean j std j The Euclidean distance between the two wind farms is calculated using the Euclidean distance formula: , Based on the distances between all wind farm pairs, construct a distance matrix D´, where D´[i,j] = d i,j ; 2) Initialize clusters: Each wind farm is considered as an initial cluster, i.e., the cluster set C = {C1, C2, ..., C}. L}, where L is the total number of wind farms, and each cluster contains only one wind farm; 3) Iteratively merge the most similar clusters: Repeat the following steps until the number of clusters is reduced to the preset M: ① Find the cluster pair with the smallest distance: Traverse the distance matrix D' and find the two clusters C with the smallest distance. p and C q ,Right now , ②C p and C q Merge into a new cluster C mew= C p ∪C q And remove C from the cluster set C. p and C q , will C mew Add C.

[0013] ③ Update the distance matrix: Calculate the new cluster C mew Other remaining clusters C k Distance ∈ C, new distance d mew,k For all wind farms in the new cluster and cluster C k The average distance of all wind farms in the region: If C p ={i1,i2,…i m}, C q ={j1,j2,…j n},but , Update the distance matrix D' with the new distance.

[0014] 4) Obtain the final sub-cluster: The iteration stops when the number of clusters decreases to M. At this point, each cluster is a sub-cluster, and the wind farms in the clusters have similar measured wind speed characteristics.

[0015] Step S20 includes: Step S201: Based on the sub-region division results, obtain the normalized historical total power output of each sub-region. ,in ; Step S201: Starting from the linear correlation dimension of the total power output of the sub-regions, construct the adjacency matrix between different sub-regions to characterize the linear correlation of the total power generation of different sub-regions. , in, and Let m1 and m2 represent the historical power output of subregions m1 and m2 at time t, respectively. The average total power output of subregions m1 and m2 are respectively... and ; Step S203: Starting from the dimension of the similarity of the curve shape of the total power generation of the sub-regions, construct the adjacency matrix between different sub-regions to characterize the similarity of the curve shape of the total power generation of the different sub-regions. , Where α is the similarity control coefficient; This is the function for calculating the dynamic time-warped distance; Step S204: Starting from the geographical distance dimension between representative power stations in the sub-region, construct an adjacency matrix between different sub-regions to characterize the geographical distance correlation between representative power stations in different sub-regions. , in, σ represents the geographical distance between representative power stations in subregion m1 and subregion m2; σ represents the standard deviation of the geographical distance.

[0016] Step S205: Construct a multi-graph convolutional neural network and perform graph convolution operations for different graph structures. The node input features F of each graph structure include: the total measured power generation of each sub-region at 96 historical points, the average measured wind speed of all wind farms within each sub-region at 96 historical points, and the average predicted wind speed of all wind farms within each sub-region at 16 future points.

[0017] Graph neural networks are defined as convolutional transformations UΘU over the Fourier domain. T X, where U is the feature matrix of the graph Laplacian matrix, X is the input feature, and Θ is the graph convolution kernel; By using Chebyshev polynomials to construct graph convolution kernels to simplify the computation process, we can obtain: , In the formula: It is a diagonal matrix containing the eigenvalues ​​of the Graph Laplacian matrix; Let θ be the order of the Chebyshev polynomial; q λ is the q-th order Chebyshev coefficient; mxax I represents the maximum value of the eigenvalues ​​of the graph Laplacian matrix.identity Let be the identity matrix. Based on this, considering the first-order Chebyshev polynomial, the formula can be simplified further, leading to the basic form of spatial graph convolution: , In the formula, Where A is the adjacency matrix; D is... The degree matrix; represents the learnable parameters.

[0018] Method for selecting representative power plants in a sub-region: 1) Calculate the time series of total power generation in sub-region m. The calculation formula is: , In the formula, P i,t Let P be the power sequence of the i-th power station within the sub-region. m,t Let m be the time series of total power generation in subregion m.

[0019] 2) For each power station i in the sub-region, calculate its power generation time series P. i,t Time series P of total power generation in sub-region m,t Pearson correlation coefficient i : , In the formula, Cov(P) i,t P m,t ) is P i,t With P m,t covariance, σ Pi,t It is P i,t Standard deviation, σ Pm,t It is P m,t Standard deviation; 3) Select representative power stations The Pearson correlation coefficient calculated for each power station. i Comparison, select Perrson i The largest power station serves as the representative power station for this sub-region.

[0020] Step S30 includes: Step S301: The outputs of the multi-graph convolutional neural network are fused using a weighted summation method; The formula for fusing the outputs of a multi-graph convolutional neural network at time step t. , In the formula, x tIt is both the fusion result of multi-graph convolution at time step t and the input of the subsequent gated recurrent unit neural network at time step t; G1, G2 and G3 represent the outputs obtained by graph convolution operations on the linear correlation graph, the curve shape similarity graph and the geographic distance graph, respectively.

[0021] Step S302: Establish a gated recurrent unit neural network and extract the temporal features of the multi-graph fusion output; , In the formula, z t The update gate output at time t determines the fusion ratio of historical information and current candidate information; r t The reset gate output at time t is used to control the degree of forgetting of the hidden state from the previous time step; Let h be the candidate hidden state at time t; t Let W be the final hidden state at time t; sigmoid is the activation function, and W is the final hidden state at time t. z W r W and b are parameter matrices. z b r b is the bias matrix, x t It is the input for each time step.

[0022] Step S40 includes: Step S401, dividing all historical data into a training set and a test set in an 8:2 ratio; the training set data is used to train the proposed model, and the test set is used to verify the model's predictive performance; Step S402, performing end-to-end training of the model based on the training set. The model is trained using node input features F and adjacency matrices A of three different graph structures as input, and the total regional wind power output at the next 16 time points as output, to obtain the regional wind power ultra-short-term prediction model model(F, A); Step S403, testing the model performance on the test set based on the trained model model(F, A).

[0023] The present invention also provides a storage medium for storing a computer program, which is executed by a processor to implement the wind power ultra-short-term prediction method based on multi-graph convolution and temporal feature extraction.

[0024] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A wind power ultra-short-term prediction method based on multi-graph convolution and temporal feature extraction, characterized in that, Includes the following steps: Step S10: Extract historical wind speed data and historical power data of each wind farm in the regional wind power cluster, and divide the regional wind power cluster into multiple sub-regions; Step S20: Using sub-regions of the regional wind power cluster as graph nodes, construct a multi-graph convolutional neural network from different dimensions; Step S30: Construct a regional wind power ultra-short-term prediction model; This model uses a weighted summation method to fuse the outputs of multiple graph convolutions. The fused output is then used as the input to a gated recurrent unit, which extracts temporal features to obtain the final prediction result. Step S40: The extracted historical wind speed data and historical power data are divided into training set and test set. The training set is used to train the model, and the test set is used to test and verify the model's performance.

2. The wind power ultra-short-term forecasting method according to claim 1, characterized in that, The historical wind speed data includes measured wind speed and predicted wind speed, with a time resolution of 15 minutes and a time span of N days.

3. The wind power ultra-short-term forecasting method according to claim 1, characterized in that, Step S10 includes: Step S101: Obtain wind speed data for each wind farm within the regional wind power cluster; The measured wind speed of the i-th power station can be denoted as: The predicted wind speed of the i-th power station is denoted as . ; The time series of measured and predicted wind speeds are normalized. ， ， In the formula, , It is the normalized time series of measured and predicted wind speeds for the i-th power station, representing different wind farms within the wind power cluster; Step S102: Extract the distribution characteristics of the measured wind speed data of each wind farm and construct a feature vector. , ， In the formula, , , These are the maximum, mean, and variance of the normalized wind speed for wind farm station numbered i, respectively. Hierarchical clustering algorithm is used to cluster the wind farms in the regional wind power cluster. Wind farms with similar measured wind speed characteristics are grouped into the same sub-cluster, and the number of sub-clusters is set to M.

4. The wind power ultra-short-term forecasting method according to claim 1, characterized in that, Step S20 includes: Step S201: Based on the sub-region division results, obtain the normalized historical total power output of each sub-region. ,in ; Step S201: Starting from the linear correlation dimension of the total power output of the sub-regions, construct the adjacency matrix between different sub-regions to characterize the linear correlation of the total power generation of different sub-regions. ， in, and Let m1 and m2 represent the historical power output of subregions m1 and m2 at time t, respectively. The average total power output of subregions m1 and m2 are respectively... and ; Step S203: Starting from the dimension of the similarity of the curve shape of the total power generation of the sub-regions, construct the adjacency matrix between different sub-regions to characterize the similarity of the curve shape of the total power generation of the different sub-regions. ， Where α is the similarity control coefficient; This is the function for calculating the dynamic time-warped distance; Step S204: Starting from the geographical distance dimension between representative power stations in the sub-region, construct an adjacency matrix between different sub-regions to characterize the geographical distance correlation between representative power stations in different sub-regions. ， in, σ represents the geographical distance between representative power stations in subregion m1 and subregion m2; σ represents the standard deviation of the geographical distance. Step S205: Construct a multi-graph convolutional neural network and perform graph convolution operations for different graph structures.

5. The wind power ultra-short-term forecasting method according to claim 4, characterized in that, Method for selecting representative power plants in a sub-region: 1) Calculate the Pearson correlation coefficient between the total power generation of the sub-region and the power generation of each power station within the sub-region; 2) Select the power station with the largest Pearson correlation coefficient as the representative power station of the sub-region.

6. The wind power ultra-short-term forecasting method according to claim 1, characterized in that, Step S30 includes: Step S301: The outputs of the multi-graph convolutional neural network are fused using a weighted summation method; The formula for fusing the output time step t of a multi-graph convolutional neural network. ， In the formula, x t It is both the fusion result of multi-graph convolution at time step t and the input of the subsequent gated recurrent unit neural network at time step t; G1, G2 and G3 represent the outputs obtained by graph convolution operations on the linear correlation graph, the curve shape similarity graph and the geographic distance graph, respectively; Step S302: Establish a gated recurrent unit neural network and extract the temporal features of the multi-graph fusion output; In the formula, z t The update gate output at time t determines the fusion ratio of historical information and current candidate information; r t The reset gate output at time t is used to control the degree of forgetting of the hidden state from the previous time step; Let h be the candidate hidden state at time t; t Let W be the final hidden state at time t; sigmoid is the activation function, and W is the final hidden state at time t. z W r W and b are parameter matrices. z b r b is the bias matrix, x t It is the input for each time step.

7. The wind power ultra-short-term forecasting method according to claim 1, characterized in that, The extracted historical wind speed and historical power data were divided into training and testing sets in an 8:2 ratio.

8. A storage medium for storing computer programs, characterized in that, The computer program is executed by a processor to implement the wind power ultra-short-term prediction method based on multiple graph convolution and temporal feature extraction as described in any one of claims 1-7.

Images