Wind power plant generation power prediction method based on space-time depth map network

Abstract

The invention discloses a wind power plant generation power prediction method based on a space-time depth map network, and the method comprises the steps: enabling a wind power plant to be divided into a map structure, and enabling each wind turbine to represent a node; an actual connection diagram is constructed according to the physical connection condition of the wind turbines in the wind power plant; correlation coefficients among the turbines are calculated through historical generated power data of the wind turbines, the correlation coefficients serve as edge weights of a full-connection graph where all the turbines are connected with one another, and an edge weight full-connection graph is established; establishing a multi-scale graph convolutional neural network, and extracting spatial features and more complex hidden information of the turbine through the established actual connection graph and the edge weight full connection graph; introducing an attention mechanism to fuse the extracted multi-scale spatial information; time features in the space information are extracted through a long-short-term memory network, and the generated power of the wind power plant is obtained through a full-connection network. The method can accurately predict the generated power of the wind power plant.

Description

Technical Field

[0001] The present invention relates to the technical field of wind farm power generation prediction, and particularly to a wind farm power generation prediction method based on a spatio-temporal depth graph network. Background Art

[0002] Wind energy has attracted people's attention due to its large storage capacity, renewable nature, low carbon, etc. However, since the wind power generation is affected by multiple factors such as wind speed and wind direction, the wind power generation is unstable and difficult to predict, and this instability brings great challenges to integrating wind power into the power grid system. To ensure the balance between power production and consumption, when the wind power generation fluctuates, other energy sources need to be used to suppress the impact on the power grid caused by the volatility of wind power. However, this alternative power cannot be quickly obtained in a short time. For example, it usually takes at least 8 hours to cold start a 1000MW thermal power plant. Therefore, accurate prediction of wind power is required to make decisions on supplementing wind power with thermal power in advance. It can be seen that wind power prediction is very important and also extremely challenging. However, most of the wind power prediction targets are only individual wind turbines, and adding up the predicted power values of all wind turbines in the wind farm will amplify the prediction error, resulting in a lower credibility of the predicted power of the wind farm. Therefore, it is necessary to change the power prediction target to the entire wind farm. Summary of the Invention

[0003] The purpose of the present invention is to provide a wind farm power generation prediction method based on a spatio-temporal depth graph network, which can accurately predict the wind farm power generation.

[0004] To achieve the above technical purpose, the technical solution adopted by the present invention is as follows:

[0005] The present invention discloses a wind farm power generation prediction method based on a spatio-temporal depth graph network, and the method includes the following steps:

[0006] S1, regarding the wind farm as a graph structure, and each wind turbine represents a node; constructing an actual connection graph through the physical connection situation of the wind turbines in the wind farm;

[0007] S2, calculating the correlation coefficient between each wind turbine through the historical power generation data of each wind turbine, and using this correlation coefficient as the edge weight of the fully connected graph of all turbines connected to each other, and establishing a fully connected graph with edge weights;

[0008] S3, establishing a multi-scale graph convolutional neural network, and extracting the spatial features and more complex hidden information of the turbines through the established actual connection graph and the fully connected graph with edge weights;

[0009] S4. Introduce an attention mechanism to fuse the extracted multi-scale spatial information to reduce the interference of redundant information on prediction;

[0010] S5. Extract the temporal features in the spatial information through a long short-term memory network and obtain the power generation of the wind farm through a fully connected network.

[0011] Step S2 further includes:

[0012] S21. Obtain the readings of each sensor of each wind turbine in the wind farm at fixed time intervals, including the sampling date d, the sampling timestamp t s , wind speed S W , the angle D between the wind direction and the turbine nacelle position W , the ambient temperature T around the turbine E , the internal temperature T of the turbine nacelle I , the yaw angle D of the turbine nacelle N , the pitch angle D of each of the three blades P1 , D P2 , D P3 , the reactive power P of the turbine r , the active power P of the turbine a ;

[0013] S22. Perform sine-cosine transformation on the sampling date and sampling timestamp; fill the null values in all data using linear interpolation based on time difference; calculate the Pearson correlation coefficient between each turbine according to the active power of each turbine;

[0014] S23. According to the actual connection situation of the wind turbines in the wind farm, construct the undirected actual connection graph data at each moment; establish the undirected fully connected graph data with edge weights at each moment, and use the Pearson correlation coefficient between two connected turbines as the edge weight between them.

[0015] Furthermore, in step S22, the process of performing sine-cosine transformation on the sampling date and sampling timestamp includes:

[0016] Taking one year as a cycle, perform sine-cosine transformation on the sampling date D:

[0017]

[0018] where d is the original sampling date, and d sin and d cos are the sine and cosine values after transformation of the original sampling date respectively;

[0019] Taking one day as a cycle, perform sine-cosine transformation on the sampling timestamp:

[0020]

[0021] Among them, t s is the original sampling timestamp, and t s,sin and t s,cos are the sine and cosine values after the transformation of the original sampling timestamp respectively.

[0022] Furthermore, in step S22, the process of filling the null values in all data by using the linear interpolation method based on the time difference includes the following steps:

[0023] Calculate the quantity change between the non-null values on both sides of each missing value, and calculate the missing value by using the linear interpolation formula based on the time difference between the non-null values on both sides. The calculation formula is as follows:

[0024]

[0025] Among them, y NaN is the missing value, y1 and y2 are the nearest valid observed values on both sides of the missing value respectively, t1 and t2 are the timestamps corresponding to the valid observed values y1 and y2 respectively, and t is the timestamp corresponding to the missing value.

[0026] Furthermore, in step S22, calculate the Pearson correlation coefficient between each turbine according to the active power of each turbine:

[0027]

[0028] Among them, r x,y is the Pearson correlation coefficient between turbine x and turbine y, and are the active powers of turbine x and turbine y at moment k respectively, k = 1, 2,..., K, and represent the average values of the historical active powers of turbine x and turbine y respectively.

[0029] Furthermore, in step S2, the process of establishing the edge-weighted fully connected graph includes the following steps:

[0030] Construct an acyclic undirected graph according to the geographical location distribution information of each wind turbine in the wind farm, where each wind turbine is represented as a node in the graph, and the connection relationship between the nodes is represented as an edge; represent the acyclic undirected graph through the node feature matrix and the edge matrix, and construct the edge-weighted fully connected graph; among them:

[0031] The node feature matrix is where n is the number of nodes, m is the feature dimension of each node, each row represents the feature vector of a node, the column represents different feature dimensions, and the features include the wind speed S W , the included angle D between the wind direction and the turbine nacelle positionW , the ambient temperature T around the turbine E , the internal temperature T of the turbine nacelle I , the yaw angle D of the turbine nacelle N , the pitch angle D of each of the three blades P1 , D P2 , D P3 , the reactive power P of the turbine r , the active power P of the turbine a ;

[0032] The edge matrix is where p is the number of edges in the graph, the first row represents the starting node of the edge, the second row represents the ending node of the edge, and the index of each node corresponds to the row index in the node feature matrix;

[0033] The edge weight matrix is Each element corresponds to the weight of each edge in the edge matrix E.

[0034] Furthermore, in step S3, the multi-scale graph convolutional neural network is established as:

[0035]

[0036] For the angle matrix the elements in represent the weighted degree, σ(·) represents the activation function, H (l+1) and H (l) represent the node feature matrices of the (l + 1)-th layer and the l-th layer respectively, and W (l) represents the weight matrix of the l-th layer; where, the adjacency matrix is:

[0037]

[0038] In the formula, w ij represents the edge weight from node i to node j;

[0039] The features extracted by GCN are aggregated through an average graph pooling layer for all node information to represent the global information h pooling :

[0040]

[0041] where, K represents the index of the state H (K) output by the last graph convolutional layer, represents the feature information of the i-th node in H (K) , i, j = 1, 2,..., n, and n is the total number of nodes.

[0042] Further, the graph convolutional neural network processes the actual connection graph and the edge weight fully-connected graph from four scales; among them:

[0043] The input of the first-scale module is the actual connection graph, which fully learns the spatial relationship and mutual physical influence between wind turbines through the physical actual connection graph; the input of the fourth-scale module is the correlation edge weight fully-connected graph, which captures more and more complex global information through the correlation edge weight fully-connected graph;

[0044] The inputs of the second-scale module and the third-scale module are the actual connection graph and the correlation edge weight fully-connected graph respectively. The second-scale module and the third-scale module jointly correct the parameters of the graph convolutional neural network by introducing a parameter sharing mechanism, so that the two graph information is fully fused.

[0045] Further, in step S4, the four graph pooling results at the same moment respectively pass through a first linear layer to obtain their respective weights and are normalized by the SoftMax function for the four weights; finally, the multiple weights are multiplied by the four graph pooling results and then added and fused to obtain the output result of the spatial convolution module; the process of the attention mechanism is as follows:

[0046]

[0047] where represents the weight corresponding to the pooling information of the i-th scale at time t; is the spatial feature information at time t in the output of the spatial convolution module, represents the global information obtained by average pooling of the i-th * th scale at time t, Linear is the linearization function, and i * = 1, 2, 3, 4.

[0048] Further, in step S5, a two-layer LSTM is used to extract the temporal feature information in the spatial features. According to the requirement of the output time step, the hidden state of the last time step obtained by the last layer of LSTM is passed to the second linear layer to output the prediction result

[0049]

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] The purpose of the present invention is to provide a wind farm power generation prediction method based on a spatio-temporal depth graph network, which realizes accurate prediction of the wind farm power generation. Description of the Drawings

[0052] Figure 1 It is the architecture block diagram of the wind farm power prediction method based on the spatio-temporal depth map network designed by the present invention;

[0053] Figure 2 It is the structural schematic diagram of the spatio-temporal depth map network designed by the present invention;

[0054] Figure 3 It is the schematic diagram of the spatial convolution module in the spatio-temporal depth map network designed by the present invention;

[0055] Figure 4 It is the schematic diagram of the temporal module in the spatio-temporal depth map network designed by the present invention. Detailed implementation manners

[0056] The following further describes the embodiments of the present invention in detail with reference to the accompanying drawings.

[0057] See Figure 1 , the present invention discloses a wind farm power generation prediction method based on a spatio-temporal depth map network, and the method includes the following steps:

[0058] S1. Consider the wind farm as a graph structure, and each wind turbine represents a node; construct an actual connection graph through the physical connection situation of the wind turbines in the wind farm.

[0059] S2. Calculate the correlation coefficient between each turbine through the historical power generation data of each wind turbine, and use this correlation coefficient as the edge weight of the fully connected graph of all turbines connected to each other, and establish a fully connected graph with edge weights.

[0060] S3. Establish a multi-scale graph convolutional neural network, and extract the spatial features and more complex hidden information of the turbines through the established actual connection graph and the fully connected graph with edge weights.

[0061] S4. Introduce an attention mechanism to fuse the extracted multi-scale spatial information to reduce the interference caused by redundant information to the prediction.

[0062] S5. Extract the temporal features in the spatial information through a long short-term memory network, and obtain the power generation of the wind farm through a fully connected network.

[0063] The present invention first obtains the readings of each sensor of each wind turbine in the wind farm at fixed time intervals, including the sampling date d, the sampling timestamp t s , the wind speed S W , the angle D between the wind direction and the turbine nacelle position W , the ambient temperature T around the turbine E , the temperature T inside the turbine nacelle I , the yaw angle D of the turbine nacelle N, the pitch angle D of each of the three blades P1 , D P2 , D P3 , the reactive power P of the turbine r , the active power P of the turbine a .

[0064] Secondly, perform sine-cosine transformation on the sampling date and sampling timestamp; fill the null values in all data using linear interpolation method based on time difference. The specific process of sine-cosine transformation is as follows:

[0065] 1) Taking one year (365 days) as a cycle, perform sine-cosine transformation on the sampling date D:

[0066]

[0067] where d is the original sampling date, and d sin and d cos are the sine and cosine values after transformation of the original sampling date respectively;

[0068] 2) Since the sampling interval is 10 minutes, taking one day (1440 minutes) as a cycle, perform sine-cosine transformation on the sampling timestamp:

[0069]

[0070] where t s is the original sampling timestamp, and t s,sin and t s,cos are the sine and cosine values after transformation of the original sampling timestamp respectively.

[0071] The present invention uses a linear interpolation method based on time difference to fill the null values. By calculating the quantity change between the non-null values on both sides of each missing value, and then using the linear interpolation formula based on the time difference between the non-null values on both sides to calculate the missing value. The calculation formula is as follows:

[0072]

[0073] where y NaN is the missing value, y1 and y2 are the nearest valid observed values on both sides of the missing value, t1 and T2 are the timestamps corresponding to these two valid observed values respectively, and t is the timestamp corresponding to the missing value.

[0074] Calculate the Pearson correlation coefficient between each turbine based on the historical active power of each turbine. The process is as follows:

[0075]

[0076] where r x,y is the Pearson correlation coefficient between turbine x and turbine y, and are the active powers of turbine x and turbine y at time i, respectively, and represent the average values of the historical active powers of turbine x and turbine y, respectively.

[0077] According to the actual connection situation of the wind turbines in the wind farm, undirected actual connection graph data at each moment are constructed. An undirected fully connected graph data with edge weights is established at each moment, and the edge weight is the Pearson correlation coefficient between two connected turbines. Specifically, an acyclic undirected graph is constructed according to the geographical location distribution information of each wind turbine in the wind farm, where each wind turbine unit is represented as a node in the graph, and the connection relationship between nodes is represented as an edge; the acyclic undirected graph is represented by a node feature matrix and an edge matrix, where:

[0078] a) The node feature matrix is where n is the number of nodes, m is the feature dimension of each node, each row represents the feature vector of a node, and the columns represent different feature dimensions. The features include the wind speed S W , the angle D between the wind direction and the turbine nacelle position W , the ambient temperature T of the turbine E , the internal temperature T of the turbine nacelle I , the yaw angle D of the turbine nacelle N , the pitch angle D of each of the three blades P1 , D P2 , D P3 , the reactive power P of the turbine r , the active power P of the turbine a ; the rows of the node feature matrix X represent the wind turbine units, and the columns represent the multi-dimensional operation features of the wind turbine units, specifically as follows:

[0079]

[0080] where each x i,j represents the j-th dimension feature of node i.

[0081] b) The edge matrix is where p is the number of edges in the graph. The first row represents the starting node of the edge, and the second row represents the ending node of the edge. The index of each node corresponds to the row index in the node feature matrix; the edge matrix E represents the connection relationship between nodes, and its form is:

[0082]

[0083] where e 0,k and e 1,k represent the starting node and the ending node of the k-th edge, respectively.

[0084] c) The edge weight matrix is where p is the number of edges in the graph, and each element corresponds to the weight of each edge in the edge matrix E, which is the Pearson correlation coefficient calculated previously; the edge weight matrix W represents the weight of each edge:

[0085]

[0086] where each w k represents the weight of the k-th edge.

[0087] As Figure 2 shown, the input of the graph convolutional neural network model is divided into two parts. The upper part is the actual connection graph constructed according to the actual connection situation of the wind turbines; the lower part is the fully connected graph in which all wind turbines are interconnected. Based on the historical data of the power generation of all units, the correlation between each wind turbine is calculated, and the correlation coefficient is used as the edge weight of the fully connected graph. Each of these graphs represents a time point. Since the wind farm units are fixed, the structure of all graphs will not change, and only the feature data on each node changes with time. The present invention uses a sliding time window method to input these graph data into the model. Each window has win_size time points, that is, win_size graphs. To improve the model training efficiency and reduce the time complexity of model training, the win_size graphs in the same time window are merged into a larger window graph (win_graph). Each sub-graph in the window graph represents the data of a moment, and there is no connection between the sub-graphs between moments. Using these two graphs as the input aims to provide the model with the real physical and spatial relationships of the wind turbines, facilitating the model to learn more real and effective complex logic between the data.

[0088] In the spatial convolution module, in order to fully learn the internal relationship of the data provided by the two graphs, it is processed through four scales:

[0089] 1) Scale one and scale four: The input of scale one is the actual connection graph, and the input of scale four is the fully connected graph of correlation edge weights. Both pass through a graph convolutional neural network composed of two graph convolutional layers (GCNConv) and one average graph pooling layer respectively, and extract the features of each moment of the two graphs. There is no interleaving between the two scales at this stage, and each independently learns one kind of graph. Scale one fully learns the spatial relationship and mutual physical influence between wind turbines through the physical actual connection graph; scale four captures more and more complex global information through the fully connected graph of correlation edge weights, improving the robustness of the model and being able to accurately predict the power of wind turbines even in the case of missing node information.

[0090] 2) Scales Two and Three: The inputs of Scales Two and Three are the actual connection graph and the correlation edge weight fully connected graph respectively. By introducing a parameter sharing mechanism, both of them jointly pass through the same graph convolutional neural network composed of two graph convolutional layers (GCNConv) and one average graph pooling layer. The two scales jointly correct the parameters of this graph convolutional neural network, which not only achieves the purpose of fully fusing the information of the two graphs, improves the generalization ability of the model, but also improves the computational efficiency of the model to a certain extent and reduces the time cost of model calculation.

[0091] The present invention inputs two types of graph data into a spatial convolutional module including four scales to obtain the spatial features of each scale; weights and sums the four groups of features through an attention mechanism to obtain the spatial features of the wind farm at each moment; inputs the spatial features of the wind farm at each moment into a temporal module, and finally obtains the wind farm power output at a future moment.

[0092] Specifically, for the spatial feature information extracted by the four scales, the present invention introduces an attention mechanism to fuse them. This step enables the model to dynamically adjust the weights of different scales and enhances the model's attention to important information. Then, the spatial features at each moment obtained by the spatial convolutional module are passed to the temporal module composed of two layers of LSTM to learn the long-term and short-term dependencies of the time series, and finally the prediction results are output through a fully connected layer (FCN).

[0093] As Figure 3 shown, the GCN used in the present invention is a spatial graph convolution that simplifies and approximates the spectral domain; assuming that the convolution filter can be represented by a polynomial function, the convolution operation in the spectral domain can transform the signal to the frequency domain through the graph Fourier transform, then multiply the frequency domain signal with the filter function g θ (Λ), and finally transform the result back to the node domain through the inverse graph Fourier transform; that is:

[0094] g θ (Λ) = diag(θ)

[0095] y = g θ (Λ)x = Ug θ (Λ)U T x

[0096] where, Λ is the eigenvalue diagonal matrix of the graph Laplacian matrix, θ is the vector of learnable parameters, U is the eigenvector matrix of the graph Laplacian matrix, x is the input signal, and y is the output signal;

[0097] To reduce the computational complexity, the Chebyshev polynomial is used to approximate the filter function, and we get:

[0098]

[0099] Among them, T k is the kth order of the Chebyshev polynomial, λ max is the largest eigenvalue of Λ, I is the identity matrix;

[0100] Furthermore, in order to make the model more efficient, the polynomial order K is limited to 1 to approximate the graph Laplacian matrix and obtain the final GCN layer formula:

[0101]

[0102] Among them, H (l) is the node feature matrix of the lth layer, W (l) is the weight matrix of the lth layer, σ(·) is the activation function (the activation function used in this paper is ReLU(·)=max(0,·)), A represents the adjacency matrix of the graph, and D and They are all angle matrices, that is, D ii =∑ j A ij ,

[0103] When edge weights are introduced into the graph, the original adjacency matrix A needs to be converted into a weighted normalized adjacency matrix The specific form is as follows:

[0104]

[0105] Among them, w ij represents the edge weight from node i to node j;

[0106] At this time, the angle matrix Elements in represents the weighted degree; therefore, the final GCN layer transfer formula with edge weights becomes:

[0107]

[0108] Then, the features extracted by GCN are aggregated through an average graph pooling layer to represent the global information:

[0109]

[0110] Among them, K represents the state H output by the last layer of graph convolutional layer (K) The index of Indicates H (K) The feature information of the i-th node in is the i-th row vector; because each moment is represented as a graph, and there are four scales to convolve the graph at each moment, this paper uses Represents the state of the output of the last graph convolutional layer at the i-th scale at time t, using It represents the global information obtained by average pooling of the i-th scale at time t.

[0111] The present invention introduces an attention mechanism to dynamically weight and fuse the information of four scales; for the pooling results of four graphs at the same moment respectively pass through a linear layer to obtain their respective weights and normalize the four weights through the SoftMax function; finally, the four weights are respectively multiplied by the four graph pooling results and then added and fused to obtain the output result of the spatial convolution module; the process of the attention mechanism is as follows:

[0112]

[0113] where represents the weight corresponding to the pooling information of the i-th scale at time t; is the spatial feature information at time t in the output of the spatial convolution module.

[0114] As Figure 4 shown, the present invention uses two layers of LSTM to extract the temporal feature information in the spatial features, and finally uses one layer of linear layer to output the predicted power value;

[0115] LSTM controls the flow of information by introducing memory cells and three gating mechanisms (input gate, forget gate, and output gate), so as to be able to more effectively handle the long-term dependence relationship in sequence data;

[0116] The forget gate determines which information should be discarded from the memory cell; it receives the input x at the current time step t and the hidden state h at the previous time step t-1 , and outputs a weight vector f between 0 and 1 t ; 1 means retaining all information, and 0 means completely forgetting; the calculation formula of the forget gate is:

[0117] f t =σ(W f [h t-1 ,x t +b f )

[0118] where, W f is the weight matrix, b f is the bias term, and σ represents the sigmoid function;

[0119] The input gate determines which new information should be stored in the memory cell; it first calculates the candidate memory cell through a dot product operation and then calculates the weight i of the input gate through another sigmoid function t, finally multiply these two to obtain the new information that should actually be added to the memory unit; the calculation formula is:

[0120] i t = σ(W i [h t-1 , x t +b i )

[0121]

[0122] The output gate determines which information to output based on the current state of the memory unit; it first calculates the weight o of the output gate t , then uses the tanh function to calculate the state c of the memory unit t , and finally multiplies the two to obtain the output h at the current time step t ; the calculation formula is:

[0123] o t = σ(W o [h t-1 , x t +b o )

[0124] h t = o t ·tanh(c t )

[0125] In this way, the LSTM can dynamically and selectively retain, update, and output information at each time step, thereby achieving effective long-term dependence learning; finally, according to the requirements of the output time step length, the timing module passes the hidden state of the last time step obtained by the last layer of LSTM to a specific linear layer for output to obtain the prediction result:

[0126]

[0127] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages. For example, object-oriented programming languages such as Java and interpreted scripting languages such as JavaScript, etc.

[0128] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, as well as the combination of flows and / or blocks in the flowchart and / or block diagram. These computer program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to generate a machine, such that the instructions run by the processor of the computer or other programmable data processing device generate means for implementing the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0129] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0130] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operating steps are run on the computer or other programmable device to generate a computer-implemented process, so that the instructions running on the computer or other programmable device provide steps for implementing the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 one or more blocks.

[0131] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications falling within the scope of the present application.

[0132] Obviously, those skilled in the art can make various changes and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

Claims

1. A method for predicting the power generation of a wind farm based on a spatio-temporal depth map network, characterized in that, The method includes the following steps: S1. Consider the wind farm as a graph structure, where each wind turbine represents a node; construct an actual connection graph based on the physical connection situation of the wind turbines in the wind farm; S2. Calculate the correlation coefficient between each turbine through the historical power generation data of each wind turbine, and use this correlation coefficient as the edge weight of the fully connected graph where all turbines are interconnected, to establish a fully connected graph with edge weights; S3. Establish a multi-scale graph convolutional neural network, and extract the spatial features and more complex hidden information of the turbines through the established actual connection graph and the fully connected graph with edge weights; S4. Introduce an attention mechanism to fuse the extracted multi-scale spatial information to reduce the interference caused by redundant information to the prediction; S5. Extract the temporal features in the spatial information through a long short-term memory network, and obtain the power generation of the wind farm through a fully connected network.

2. The wind farm power generation prediction method based on the spatio-temporal depth map network according to claim 1, wherein, Step S2 further includes: S21, obtain the readings of each sensor of each wind turbine in the wind farm at fixed time intervals, including the sampling date d, the sampling timestamp t s , wind speed S E , the angle D between the wind direction and the turbine nacelle position W , the ambient temperature T around the turbine E , the temperature T inside the turbine nacelle I , the yaw angle D of the turbine nacelle N , the pitch angle D of each of the three blades P1 , D P2 , D P3 , the reactive power P of the turbine r , the active power P of the turbine a ; S22. Perform sine-cosine transformation on the sampling date and sampling timestamp; fill the null values in all data using linear interpolation based on the time difference; calculate the Pearson correlation coefficient between each turbine according to the active power of each turbine; S23. According to the actual connection situation of the wind turbines in the wind farm, construct the undirected actual connection graph data at each moment; establish the undirected fully connected graph data with edge weights at each moment, and use the Pearson correlation coefficient between two connected turbines as the edge weight between them.

3. The wind farm power generation prediction method based on the spatio-temporal depth map network according to claim 1, wherein In step S22, the process of performing sine-cosine transformation on the sampling date and sampling timestamp includes: Taking one year as a cycle, perform sine-cosine transformation on the sampling date D: where d is the original sampling date, d sun and d cos are respectively the sine and cosine values after transformation of the original sampling date; Taking one day as a cycle, perform sine-cosine transformation on the sampling timestamp: where t s is the original sampling timestamp, and t s,sin and t s,cos are the sine and cosine values after transformation of the original sampling timestamp, respectively.

4. The method for predicting the power generation of a wind farm based on a spatio-temporal depth map network according to claim 1, characterized in that, In step S22, the process of filling the null values in all data using linear interpolation based on the time difference includes the following steps: Calculate the quantity change between the non-null values on both sides of each missing value, and calculate the missing value using the linear interpolation formula based on the time difference between the non-null values on both sides. The calculation formula is as follows: where y NaN is a missing value, y1 and y2 are the nearest valid observations on both sides of the missing value, t1 and t2 are the timestamps corresponding to the valid observations y1 and y2 respectively, and t is the timestamp corresponding to the missing value.

5. The wind farm power generation prediction method based on the spatio-temporal depth map network according to claim 1, wherein In step S22, calculate the Pearson correlation coefficient between each turbine according to the active power of each turbine: where r x,y is the Pearson correlation coefficient between turbine x and turbine y, and are the active powers of turbine x and turbine y at time k, respectively, where k = 1, 2, ..., K, and represent the average values of the historical active powers of turbine x and turbine y, respectively.

6. The wind farm power generation prediction method based on a spatio-temporal depth map network according to claim 1, wherein In step S2, the process of establishing a fully connected graph with edge weights includes the following steps: Construct an acyclic undirected graph according to the geographical location distribution information of each wind turbine in the wind farm, where each wind turbine unit is represented as a node in the graph, and the connection relationship between nodes is represented as an edge; represent the acyclic undirected graph through a node feature matrix and an edge matrix, and construct a fully connected graph with edge weights; where: The node feature matrix is where n is the number of nodes, m is the feature dimension of each node, each row represents the feature vector of a node, and the columns represent different feature dimensions. The features include the wind speed S W , the angle D between the wind direction and the position of the turbine nacelle W , the ambient temperature T around the turbine E , the internal temperature T of the turbine nacelle I , the yaw angle D of the turbine nacelle N , the pitch angles D of the three blades respectively P1 , D P2 , D P3 , the reactive power P of the turbine r , the active power P of the turbine a ; The edge matrix is where p is the number of edges in the graph. The first row represents the starting nodes of the edges, and the second row represents the ending nodes of the edges. The index of each node corresponds to the row index in the node feature matrix; The edge weight matrix is Each element corresponds to the weight of each edge in the edge matrix E.

7. The wind farm power generation prediction method based on the spatio-temporal depth map network according to claim 1, wherein In step S3, the multi-scale graph convolutional neural network established is: Angle matrix The elements in represent the weighted degree, σ(·) represents the activation function, H (l+1) and H (l) represent the node feature matrices of the (l + 1)-th layer and the l-th layer respectively, and W (l) represents the weight matrix of the l-th layer; among them, the adjacency matrix is: where w ij represents the edge weight from node i to node j; The features extracted by GCN are aggregated through an average graph pooling layer for all node information to represent the global information h pooling : Among them, K represents the index of the state H output by the last layer of graph convolutional layer (K) of, representing the feature information of the i-th node in H (K) where i, j = 1, 2,..., n, and n is the total number of nodes 8. The method for predicting the power generation of a wind farm based on a spatio-temporal depth map network according to claim 1, characterized in that The graph convolutional neural network processes the actual connection graph and the fully connected graph with edge weights from four scales; where: The input of the first-scale module is the actual connection graph, and it fully learns the spatial relationship and mutual physical influence between wind turbine units through the physical actual connection graph; the input of the fourth-scale module is the correlation fully connected graph with edge weights, and it captures more and more complex global information through the correlation fully connected graph with edge weights; The inputs of the second-scale module and the third-scale module are the actual connection graph and the correlation edge weight fully-connected graph respectively. By introducing a parameter sharing mechanism, the second-scale module and the third-scale module jointly correct the parameters of the graph convolutional neural network to fully integrate the two graph information.

9. The method for predicting the power generation of a wind farm based on a spatio-temporal depth map network according to claim 8, wherein, In step S4, for the four graph pooling results at the same moment respectively pass through a first linear layer to obtain their respective weights and normalize the four weights through the SoftMax function; finally, multiply the multiple weights by the four graph pooling results respectively and then sum and fuse them to obtain the output result of the spatial convolution module; the process of the attention mechanism is as follows: Among them represents the weight corresponding to the i-th scale pooling information at time t; is the spatial feature information at time t in the output of the spatial convolution module, represents the global information obtained by average pooling of the i-th * scale at time t, Linear is the linearization function, and i * = 1, 2, 3, 4.

10. The method for predicting the power generation of a wind farm based on a spatio-temporal depth map network according to claim 1, wherein In step S5, a two-layer LSTM is used to extract the temporal feature information in the spatial features, and according to the requirement of the output time step, the hidden state of the last time step obtained by the last layer of LSTM is passed to the second linear layer to output the prediction result

Images