A multi-unit wind power prediction method based on graph network

Abstract

The present invention discloses a multi-unit wind power prediction method based on a graph network, which belongs to the field of wind farm wind power prediction and includes the following steps: recording the relative positions of all wind turbines in the wind farm to characterize the spatial distribution, using a supervisory control and data acquisition system to collect monitoring data of the wind farm, obtaining the power data of each unit, and performing data preprocessing; designing a graph convolutional network feature extraction module to extract features from the spatial position information of the wind turbines and the inherent features of each unit, converting non-Euclidean data into Euclidean data; then extracting time features from the fused feature vector sequence through a time series feature learning layer; and finally predicting the wind power of each wind turbine through a regression layer. By combining a graph convolutional network and a two-dimensional time coefficient matrix, the present invention automatically selects and extracts spatial features and time series features from the original data, thereby improving the overall prediction accuracy, prediction efficiency, and model robustness, making the dispatch of the power system safer and more stable.

Description

Technical Field

[0001] The present invention relates to the field of wind farm wind power prediction, and in particular to a multi-unit wind power prediction method based on a graph network. Background Art

[0002] With increasing energy consumption and increasingly serious environmental pollution, the development and utilization of wind energy has attracted widespread attention. However, the high volatility and intermittent nature of wind energy make wind power generation uncertain, posing significant potential risks to power grids. To ensure the safe and stable operation of power systems, accurately predicting wind farm output power (Wind Power Forecasting, WPF) is a major issue that needs to be addressed.

[0003] Currently, research institutions and scholars at home and abroad have conducted extensive research on wind power forecasting. Generally speaking, these forecasting methods fall into two categories: physical model-based and data-driven. Physical model-based forecasting methods are significantly affected by wind farm topography and natural meteorological conditions, resulting in certain limitations in practical applications. Data-driven forecasting methods have become a hot topic of research in both academia and industry.

[0004] Most existing studies only predict power for a single wind turbine, without considering the interactions between surrounding turbines in different locations. This results in poor overall prediction results for multiple turbines. Geographic information from different turbines can provide more comprehensive information for multi-turbine wind power prediction in a wind farm. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a multi-unit wind power prediction method based on a graph network, which can effectively improve the accuracy of multi-unit wind power prediction, thereby making timely planning for power dispatching and improving the stability of the coordinated operation of wind farms and power systems.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A multi-unit wind power prediction method based on a graph network includes the following steps:

[0008] S1. Record the location information of all wind turbines in the wind farm, use the Supervisory Control and Data Acquisition (SCADA) system to collect monitoring data of the wind farm, obtain the power data of each unit as the node feature, and perform data preprocessing;

[0009] S2. Design a graph network feature extraction module based on the location information and power data to integrate the location information of the wind turbine with the node features of each unit;

[0010] S3, multiply the feature sequence obtained in S2 by the two-dimensional time coefficient matrix to further extract the features of the time dimension and output the multi-view spatiotemporal fusion features;

[0011] S4: Build a fully connected regression layer and input the multi-view spatiotemporal fusion features obtained in S3 into the fully connected regression layer to obtain the wind power prediction results of all wind turbines in the wind farm.

[0012] A further improvement of the technical solution of the present invention is that S1 specifically includes the following steps:

[0013] S11, obtain the unit's location information as X i , Y i , represents the position coordinates of the i-th wind turbine; the wind power data size is n×H, n represents the total number of units, and H represents the length of the read data;

[0014] S12. Based on the spatial position of the units, calculate the Euclidean distance between each unit and form an adjacency matrix to represent the positional relationship between them. The formula is as follows:

[0015] Among them, s ij represents the distance between points i and j, x i ,y i Represents the horizontal and vertical coordinates of point i, x j ,y j Represents the horizontal and vertical coordinates of point j;

[0016] S13. Power data is processed by sliding window to fully extract time series feature information. The window size is W. After further normalization of maximum and minimum values, the final data size is: n×H / W, where n is the number of units. The calculation formula is as follows:

[0017]

[0018] where y ij is the jth value of the i-th unit in the standardized time series, x ij is the jth power value of the i-th unit in the original time series, min(x i ) and max(x i ) are the minimum and maximum power values ​​in the time series of the i-th unit, respectively.

[0019] A further improvement of the technical solution of the present invention is that S2 specifically includes the following steps:

[0020] S21. Design a graph convolutional network feature learning module for position information and power data. The convolution layer uses ReLu as the activation function, and a batch normalization layer is added within the module to improve the training effect of the model, thereby obtaining the fusion features of the signal.

[0021] S22, learning through graph convolutional network, the calculation formula is as follows:

[0022]

[0023]

[0024]

[0025] Where W is the parameter to be learned, D is the degree matrix, A is the adjacency matrix, and I N is the identity matrix.

[0026] A further improvement of the technical solution of the present invention is that S3 specifically includes the following steps:

[0027] S31, such as Figure 4 As shown in the figure, a time coefficient matrix is ​​calculated based on the unit's historical power data. The least squares method is used to solve the multivariate regression problem. Corresponding coefficients are assigned to different time points in the historical data of different units. The time coefficient matrices of all units are then concatenated to obtain a two-dimensional time coefficient matrix. Temporal information is extracted by multiplying the spatial features extracted by the graph convolutional network with the values ​​at the corresponding positions in the two-dimensional time coefficient matrix, achieving the goal of spatiotemporal feature fusion.

[0028] A further improvement of the technical solution of the present invention is that S4 specifically includes the following steps:

[0029] S41. Define wind power prediction as a regression prediction problem;

[0030] S42, converting the final spatiotemporal features obtained in S31 into a two-dimensional matrix and inputting it into a regression layer with a mean absolute error loss function to evaluate the wind power of the unit; wherein the calculation formula of the mean absolute error is as follows:

[0031]

[0032] Among them, y i represents the true value, Represents the predicted value, and MAE is the mean absolute error between the predicted value and the true value after averaging.

[0033] Due to the adoption of the above technical solution, the technical advancements achieved by the present invention are:

[0034] 1. The present invention considers the spatial distribution between units, calculates the adjacency matrix through Euclidean distance, and uses the spatial distribution and node information as model input.

[0035] 2. The present invention automatically extracts spatial features by utilizing graph convolutional networks and retains the temporal nature of feature sequences.

[0036] 3. The present invention achieves the purpose of spatiotemporal feature fusion by learning time series features using a two-dimensional time coefficient matrix. This method is simple and efficient.

[0037] 4. The present invention improves the prediction accuracy of wind power of multiple units, thereby making timely planning for power dispatching and improving the stability of coordinated operation of wind farms and power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the prediction method of the present invention;

[0039] Figure 2 Schematic diagram of the spatiotemporal feature learning and fusion module in the present invention;

[0040] Figure 3 This is a schematic diagram of spatial feature extraction using a graph convolutional network in the present invention;

[0041] Figure 4 This is a schematic diagram of constructing the time coefficient matrix of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0043] like Figure 1 As shown, a multi-unit wind power prediction method based on a graph network includes the following steps:

[0044] S1. Record the location information of all wind turbines in the wind farm, use the Supervisory Control and Data Acquisition (SCADA) system to collect monitoring data of the wind farm, obtain the power data of each unit as the node feature, and perform data preprocessing. Specifically, it includes the following steps:

[0045] S11, obtain the unit's location information as X i , Y i , represents the position coordinates of the i-th wind turbine; the wind power data size is n×H, n represents the total number of units, and H represents the length of the read data.

[0046] S12. Based on the spatial position of the units, calculate the Euclidean distance between each unit and form an adjacency matrix to represent the positional relationship between them. The formula is as follows:

[0047] Among them, s ij represents the distance between points i and j, x i ,y i Represents the horizontal and vertical coordinates of point i, x j ,y j Represents the horizontal and vertical coordinates of point j;

[0048] S13. Power data is processed by sliding window to fully extract time series feature information. The window size is W. After further normalization of maximum and minimum values, the final data size is: n×H / W, where n is the number of units. The calculation formula is as follows:

[0049]

[0050] where y ij is the jth value of the i-th unit in the standardized time series, x ij is the jth power value of the i-th unit in the original time series, min(x i ) and max(x i ) are the minimum and maximum power values ​​in the time series of the i-th unit, respectively.

[0051] S2, such as Figure 2 As shown in the figure, a graph convolution module is used to extract features from location information and node information. For location information and power information, a graph network feature extraction module is designed to integrate the spatial location information of the wind turbine with the inherent characteristics of each unit. The specific steps include:

[0052] S21. Design a graph convolutional network feature learning module for position information and power data. The convolution layer uses ReLu as the activation function, and a batch normalization layer is added within the module to improve the training effect of the model, thereby obtaining the fusion features of the signal.

[0053] S22, learning through graph convolutional network, the calculation formula is as follows:

[0054]

[0055]

[0056]

[0057] Where W is the parameter to be learned, D is the degree matrix, A is the adjacency matrix, and I N is the identity matrix.

[0058] S3: Multiply the feature sequence obtained in S2 by the two-dimensional time coefficient matrix to further extract the features of the time dimension and output the multi-view spatiotemporal fusion features. This specifically includes the following steps:

[0059] S31, such as Figure 4As shown in the figure, a time coefficient matrix is ​​calculated based on the unit's historical power data. The least squares method is used to solve the multivariate regression problem. Corresponding coefficients are assigned to different time points in the historical data of different units. The time coefficient matrices of all units are then concatenated to obtain a two-dimensional time coefficient matrix. Temporal information is extracted by multiplying the spatial features extracted by the graph convolutional network with the values ​​at the corresponding positions in the two-dimensional time coefficient matrix, achieving the goal of spatiotemporal feature fusion.

[0060] S4: Build a fully connected regression layer and input the multi-view spatiotemporal fusion features obtained in S3 into the fully connected regression layer to obtain the wind power prediction results of all wind turbines in the wind farm. The specific steps are as follows:

[0061] S41. Define wind power prediction as a regression prediction problem;

[0062] S42, converting the final spatiotemporal features obtained in S31 into a two-dimensional matrix and inputting it into a regression layer with a mean absolute error loss function to evaluate the wind power of the unit; wherein the calculation formula of the mean absolute error is as follows:

[0063]

[0064] Among them, y i represents the true value, Represents the predicted value, and MAE is the mean absolute error between the predicted value and the true value after averaging.

[0065] From the above analysis we can see that:

[0066] This example proposes an effective method for predicting wind power for multiple units at different locations within a wind farm by performing spatiotemporal feature fusion learning. By collecting the location and power information of each unit, the Euclidean distance of the location information of all wind turbines is calculated to form an adjacency matrix. The power data is preprocessed to make it of the same size in the time dimension. Then, a graph network is used to fuse the overall location information and its own power information. A feature extraction module is constructed to convert non-Euclidean data into Euclidean data. The time series features of the data are further extracted through a two-dimensional time coefficient matrix to achieve the purpose of spatiotemporal feature fusion, fully extract historical information, and effectively improve the accuracy of multi-unit wind power prediction, so that scheduling plans can be formulated in a timely manner to ensure the stable operation of the power system.

[0067] Principle and process:

[0068] The present invention is based on the spatiotemporal feature extraction and fusion of a multi-perspective deep fusion network, thereby realizing the prediction of wind power of multiple wind turbines in a wind farm. First, the location information and power data of the wind turbine are obtained, and data preprocessing is performed. Then, the graph convolution module is used to extract features from the location information and node information. The feature sequence is then multiplied by the two-dimensional time coefficient matrix to further extract the features of the time dimension and output the multi-perspective spatiotemporal fusion features. A fully connected regression layer is constructed, and the multi-perspective spatiotemporal fusion features are input into the fully connected regression layer to obtain the wind power prediction results of all wind turbines in the wind farm.

[0069] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A multi-unit wind power prediction method based on a graph network, characterized by: The following steps are involved: S1. Record the location information of all wind turbines in the wind farm, use the supervisory control and data acquisition system to collect monitoring data of the wind farm, obtain the power data of each unit as the node feature, and perform data preprocessing; S2. Based on the location information and power data, a graph convolutional network feature extraction module is designed to extract the location information of the wind turbine and the node features of each unit; S3, multiplying the feature sequence obtained in step S2 by the two-dimensional time coefficient matrix to further extract the features of the time dimension and output the multi-view spatiotemporal fusion features; S3 specifically includes the following steps: S31. Calculate the time coefficient matrix based on the historical power data of the units, use the least squares method to solve the multivariate regression problem, assign corresponding coefficients to different time points of the historical data of different units, and concatenate the time coefficient matrices of all units to obtain a two-dimensional time coefficient matrix. Extract time series information by multiplying the spatial features extracted by the graph convolutional network with the values ​​at the corresponding positions of the two-dimensional time coefficient matrix, thereby achieving the purpose of spatiotemporal feature fusion. S4. Construct a fully connected regression layer and input the multi-view spatiotemporal fusion features obtained in step S3 into the fully connected regression layer to obtain the wind power prediction results of all wind turbines in the wind farm; S4 specifically includes the following steps: S41. Define wind power prediction as a regression prediction problem; S42, converting the final spatiotemporal features obtained in S31 into a two-dimensional matrix and inputting it into a regression layer with a mean absolute error loss function to evaluate the wind power of the unit; wherein the calculation formula of the mean absolute error is as follows: Among them, y i represents the true value, Represents the predicted value, and MAE is the mean absolute error between the predicted value and the true value after averaging.

2. The multi-unit wind power prediction method based on graph network according to claim 1, characterized in that: S1 specifically includes the following steps: S11, obtain the unit's location information as X i , Y i , represents the position coordinates of the i-th wind turbine; the wind power data size is n×H, n represents the total number of units, and H represents the length of the read data; S12. Based on the spatial position of the units, calculate the Euclidean distance between each unit and form an adjacency matrix to represent the positional relationship between them. The formula is as follows: Among them, s ij represents the distance between points i and j, x i ,y i Represents the horizontal and vertical coordinates of point i, x j ,y j Represents the horizontal and vertical coordinates of point j; S13. The power data is processed by sliding window to fully extract the time series feature information. The window size is W. After further maximum and minimum value normalization processing, the final data size is: n×H / W, where n is the number of units. The maximum and minimum value normalization calculation formula is as follows: where y ij is the jth value of the i-th unit in the standardized time series, x ij is the jth power value of the i-th unit in the original time series, min(x i ) and max(x i ) are the minimum and maximum power values ​​in the time series of the i-th unit, respectively.

3. The multi-unit wind power prediction method based on graph network according to claim 1, characterized in that: S2 specifically includes the following steps: S21. Design a graph convolutional network feature learning module for position information and power data. The convolution layer uses ReLu as the activation function, and a batch normalization layer is added within the module to improve the training effect of the model, thereby obtaining the fusion features of the signal. S22, learning through graph convolutional network, the calculation formula is as follows: Where W is the parameter to be learned, D is the degree matrix, A is the adjacency matrix, and I N is the identity matrix.

Images