Multi-node wind speed prediction method based on static and dynamic graph space-time learning

Abstract

The invention provides a multi-node wind speed prediction method based on static and dynamic graph space-time learning. The method comprises the following steps: acquiring wind speed data of a plurality of monitoring points in different time periods; and constructing a wind speed prediction model, and predicting the wind speed by using the trained wind speed prediction model. The wind speed prediction model comprises a static graph learning module, a dynamic graph learning module, a graph convolution module and a time convolution module, static and dynamic graphs are constructed based on prior knowledge and real-time wind speed data, and a complex and unknown spatial dependency relationship between nodes is captured; through integration of time convolution and graph convolution, long-term and short-term network outputs are effectively combined so that accurate wind speed prediction is generated. The method effectively improves the prediction precision by fully capturing the spatial and temporal features of the wind speed data, and obtains an effective wind speed prediction result based on the wind direction features, the time features of the wind speed data and the spatial features by capturing the long-term static and short-term dynamic association among multiple nodes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind speed prediction, and particularly relates to a multi-node wind speed prediction method based on static and dynamic graph spatio-temporal learning. BACKGROUND

[0002] As an important form of renewable energy, wind energy plays an important role in addressing energy crises and reducing carbon emissions. In the process of developing and utilizing wind energy, accurate wind speed prediction can provide necessary decision support for wind farm site selection and wind energy scheduling. However, the flexibility and intermittency of wind speed make accurate wind speed prediction a challenging task.

[0003] The prior art mainly predicts wind speed through physical models and statistical models. Physical models such as numerical weather prediction (NWP) can simulate wind speed changes and are suitable for large-scale or long-term prediction, but have high computational complexity. Traditional statistical models such as autoregressive moving average (ARMA) and autoregressive integrated moving average (ARIMA) are good at handling short-term prediction tasks, but have limited performance when facing complex nonlinear and high-dimensional data.

[0004] To make up for these deficiencies, machine learning methods such as support vector regression (SVR), extreme learning machine (ELM), and echo state network (ESN) are widely used in wind speed prediction and further improve performance by combining with other technologies. For example, VMD is used for data preprocessing, and then ESN is used for prediction, or ELM is combined with AdaBoost algorithm to optimize the prediction effect. These methods significantly improve the accuracy of wind speed prediction, but machine learning still has difficulty in fully mining deep temporal features in wind speed data, which provides space for further application of deep learning.

[0005] To solve the limitations of traditional methods, deep learning methods such as LSTM, TCN, and GRU are applied to wind speed prediction and have achieved good results. However, multi-node wind speed prediction needs to consider both temporal dependence and spatial correlation. Graph convolution network (GCN) is widely used to build spatial correlation and improve wind speed prediction performance due to its advantage in processing irregular graph structure data. However, this graph structure based on prior knowledge ignores the hidden spatial relationship in the data, making it difficult to accurately predict wind speed. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a multi-node wind speed prediction method based on static and dynamic graph spatio-temporal learning to solve the problem that the performance of existing wind speed prediction models is limited and cannot effectively capture spatio-temporal dependence.

[0007] The technical scheme of the present application is: a multi-node wind speed prediction method based on static and dynamic graph space-time learning, comprising the following steps: S1), obtaining wind speed data of multiple monitoring points in different time periods; S2), constructing a wind speed prediction model and training it using the obtained wind speed data; S3), using the trained wind speed prediction model to predict wind speed.

[0008] As preferred, in step S2), the wind speed prediction model comprises a static graph learning module, a dynamic graph learning module, a graph convolution module and a time convolution module; the static graph learning module captures the long-term static association between wind speed nodes by optimally weighting the geographical distance and dynamic time warping distance of the graph nodes; the dynamic graph learning module constructs short-term dynamic association based on node embedding and multi-head attention mechanism stacked dynamic graph learning; the time convolution module is used to capture deep time correlation and fully consider the change of relationship between different time points caused by weather condition change; The graph convolution module is used to perform information propagation operation on the output results of the static graph learning module and the dynamic graph learning module, and finally predict the final result through a fully connected layer.

[0009] As preferred, in step S2), the static graph learning module quantifies the similarity between different time series by introducing dynamic time warping (DTW) to capture the compression or stretching relationship between different node wind speed time series.

[0010] As preferred, in step S2), for different time series And }, the dynamic time warping distance is expressed as: ; Wherein: ; In the formula, represents the Euclidean distance between nodes and ; represents the wind speed data of monitoring point at time ; represents the wind speed data of monitoring point at time .

[0011] As preferred, in step S2), the dynamic time warping distance is used to measure the optimal alignment path of two time series at different time steps, therefore, based on the dynamic time warping distance The optimal weighted graph of different nodes is constructed, and the DTW weighted adjacency matrix and the position weighted adjacency matrix of the optimal weighted graph are respectively: ; ; In the formula, DTW weighted adjacency matrix is represented; N represents the number of monitoring points; : is a parameter of the position weighted adjacency matrix; By combining with , we get: ; In the formula, representing a static graph structure representation; representing a learnable weight factor.

[0012] As preferred, in step S2), the dynamic graph learning module includes an information fusion module, a DGL module and a multi-layer perception machine MLP.

[0013] As preferred, in step S2), the information fusion module introduces a gating mechanism to adaptively fuse the graph structure and the dynamic input; for each input window of time steps, the information fusion is as follows: ; ; ; ; In the formula, reset gate is represented; update gate is represented; intermediate hidden state is represented; fused node information is represented; , reset gate and update gate weight matrix are represented respectively; , reset gate and update gate weight matrix are represented respectively; is the weight matrix of the intermediate hidden state; sigmoid activation function is represented; dynamic input is a matrix after FFN layer transformation; node embedding is represented; node number is represented; historical time series length is represented; denotes the embedding dimension; is a tangent hyperbolic function; denotes the Hadamard product.

[0014] As preferred, in step S2), the DGL module captures the feature dependency relationship between different wind speed node time series through the multi-head attention mechanism, the DGL module first performs LayerNorm on the fused node information After normalization processing, then through The scaled inner product is used to calculate The association matrix between nodes That is: ; In the formula, denotes the association matrix between nodes learned in the th space; ; , respectively denote the weight matrix of the query vector, the key vector of the space; denotes the scaling factor; denotes the scale factor; denotes the bias vector; is the mean of ; is the standard deviation of ; denotes the inner product.

[0015] As preferred, in step S2), the sum of the similarity results between nodes captured by each space of the spatio-temporal layer is taken as the final similarity measure between nodes, that is: ; ; In the formula, denotes the matrix after performing Dropout operation on the association matrix ; denotes a random regularization operation, which prevents overfitting by randomly discarding neurons, especially in the case of high feature dimension or small data volume; denotes the cumulative similarity result between nodes identified in the h th space of each spatio-temporal layer as the final similarity measure between nodes; denotes a learnable parameter matrix; denotes the association matrix between nodes learned in the ​The sum of all the inter-node correlation matrices learned by the space; Then smoothed by LayerNorm; that is: ; ; In the formula, represents the result obtained by using the inner product of skip connection; represents the result after passing through two FFN layers; represents the LayerNorm normalization function; , are the learnable parameter matrices of the first , heads, respectively; , , , represents the trainable parameters; represents the current layer inference dynamic graph structure representation after passing through the FeedForward module twice linear transformation and then passing through the RELU and SoftMax two activation functions; ; In the formula, represents the dynamic node embedding; is the transpose symbol; To ensure the relevance between the dynamic node embedding of the previous layer and the static node embedding , the momentum update form is as follows: ; In the formula, is the momentum coefficient; the parameters of the dynamic node embedding are updated by back propagation.

[0016] As preferred, in step S2), the time convolution module adopts a dilated causal convolution network TCN. For a given node 1D sequence input , the dilated convolution is defined as: ; In the formula, represents a 1D convolution filter with a kernel size of ; is the time; represents the expansion factor; represents the convolution operation; ​​​By extending the receptive field of the convolution kernel, the network can capture longer-range dependencies while reducing the number of required layers and computational complexity.

[0017] As preferred, in step S2), the temporal convolution module introduces a gating mechanism, which realizes dynamic selection and weighting of information through two activation functions, thereby effectively managing information memory and forgetting in sequence modeling tasks. For a given two extended initial layers 、 , the following is obtained , that is: In the formula, represents the amount of information processed by the activation function and adjusted by the gating mechanism, and represents the final output of the i th node; 、 are activation functions; 、 、 、 respectively represent the parameters of the two extended initial layers; represents the Hadamard product, which performs element-wise multiplication on the two outputs.

[0018] As preferred, in step S2), the spatial propagation mechanism of the random walk diffusion process is used to model the information propagation on the graph as a graph signal diffusion process with a limited number of steps. The input signal is diffused by a truncated diffusion convolution, and the result is expressed as: ; In the formula, represents the step length of diffusion, and represents the number of propagations; represents a weight matrix used to learn features in the diffusion process; represents a degree diagonal matrix; represents an adjacency matrix; represents a transition matrix of the diffusion process.

[0019] As preferred, in step S2), the graph convolution module is used to perform information propagation operations on the output results of the static graph learning module and the dynamic graph learning module, and outputs through the output layer, specifically: ; In the formula, represents a feature matrix calculated by GCN, which is the comprehensive result of static graph and dynamic graph convolution operations; represents the The weight matrix of the static graph in diffusion steps; Represents a static image; represents the output of the temporal convolution module; Indicates the The weight matrix of the dynamic graph in diffusion steps; Represents a dynamic graph; The output layer accumulates the information captured by the spatiotemporal layer through skip connections, and is processed by RELU and two 1×1 convolution layers before predicting the final result through a fully connected layer.

[0020] The beneficial effects of the present invention are: 1. This paper constructs static and dynamic graphs based on prior knowledge and real-time wind speed data to capture the complex and unknown spatial dependencies between nodes. By integrating temporal convolution and graph convolution, it effectively combines long-term and short-term network outputs to generate accurate wind speed forecasts. 2. The present invention can fully capture the spatiotemporal characteristics of wind speed data, thereby effectively improving the prediction accuracy. Moreover, the present invention captures the long-term static and short-term dynamic associations between multiple nodes, and comprehensively considers the nonlinear time dependence and dynamic spatial association in the wind speed data, and obtains effective wind speed prediction results based on wind direction characteristics, temporal characteristics of wind speed data, and spatial characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the process of the present invention; Figure 2 A structural framework diagram of the wind speed prediction model constructed for the present invention; Figure 3 This is a structural framework diagram of the static graph learning module and the dynamic graph learning module of the present invention; Figure 4 This is a structural framework diagram of the information fusion module and the DGL module of the present invention; Figure 5 Comparison diagram of wind speed sequences at 4 steps in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings: like Figure 1 As shown, this embodiment provides a multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning, including the following steps: S1) Obtain wind speed data at multiple monitoring points in different time periods; S2), building a wind speed prediction model and training it using the acquired wind speed data; like Figure 2As shown, the wind speed prediction model comprises a static graph learning module, a dynamic graph learning module, a graph convolution module and a time convolution module; the static graph learning module captures long-term static association between wind speed nodes by optimally weighting the geographical distance and dynamic time warping distance of graph nodes; the dynamic graph learning module constructs short-term dynamic association based on node embedding and multi-head attention mechanism stacked dynamic graph learning; the time convolution module is used to capture deep time correlation and fully consider the change of relationship between different time points caused by weather condition change; The graph convolution module is used to perform information propagation operation on the output results of the static graph learning module and the dynamic graph learning module, and finally predict the final result through a fully connected layer.

[0023] As Figure 3 shown, the static graph learning module quantifies the similarity between different time series by introducing dynamic time warping (DTW) to capture the compression or stretching relationship between different node wind speed time series. For different time series and , the dynamic time warping distance is expressed as: ; wherein: ; In the formula, , represents the Euclidean distance between nodes and ; represents the wind speed data of time monitoring point ; represents the wind speed data of time monitoring point .

[0024] The dynamic time warping distance is used to measure the optimal alignment path of two time series at different time steps, therefore, based on the dynamic time warping distance , the best weighted graph of different nodes is constructed, and the DTW weighted adjacency matrix and the position weighted adjacency matrix of the best weighted graph are respectively: ; ; In the formula, , represents the DTW weighted adjacency matrix; , represents the number of monitoring points; : is a parameter of the position weighted adjacency matrix; by combining with Combining them, we get: ; Where, Represents static graph structure representation; represents the learnable weight factor.

[0025] In this embodiment, Figure 3 As shown, the dynamic graph learning module includes an information fusion module, a DGL module and a multi-layer perceptron MLP. Figure 4 As shown, the information fusion module introduces a gate mechanism to adaptively fuse graph structure and dynamic input; for each input window The information fusion in the time step of is as follows: ; ; ; ; Where, Reset gate. represents the update gate; represents the intermediate hidden state; Indicates the fused node information; 、 Represent the weight matrices of the reset gate and update gate respectively; 、 Represent the weight matrices of the reset gate and update gate respectively; is the weight matrix of the intermediate hidden state; is the sigmoid activation function; Represents dynamic input The matrix transformed by the FFN layer; represents node embedding; Indicates the number of nodes; Indicates the length of the historical time series; represents the dimension of embedding; is the tangent hyperbolic function; represents the Hadamard product.

[0026] The DGL module captures the feature dependency between different wind speed node time series through the multi-head attention mechanism. The DGL module first uses LayerNorm to fused node information. By normalizing, and then by The number of heads is calculated using the scaled inner product The connection matrix between nodes ,Right now: ; Where, Indicates in Nodes learned in space The correlation matrix between 、 Respectively The query vector and key vector weight matrix of the space; represents the scaling factor; represents the scale factor; represents the bias vector; for The mean of for The standard deviation of represents the inner product.

[0027] By combining the various time and space layers The sum of the similarity results between the nodes captured in the space is used as the final similarity measure between the nodes, that is: ; ; Where, Represents the incidence matrix The matrix after Dropout operation; Represents a random regularization operation that prevents overfitting by randomly dropping neurons, especially when the feature dimension is high or the amount of data is small; Indicates that the first h The cumulative similarity results between the nodes identified in the space are used as the final similarity measure between the nodes; represents the learnable parameter matrix; Indicates in The sum of the association matrices between all nodes learned in the space; Then smooth it out using LayerNorm; that is: ; ; Where, Indicates use Inner product to obtain the result of skip connection; express The result after passing through two FFN layers; Represents the LayerNorm normalization function; 、 Respectively 、 a learnable parameter matrix of a size of a head; 、 、 、 represents a trainable parameter; will be representing the dynamic graph structure representation inferred by the current layer after two linear transformations through the FeedForward module and two activation functions of RELU and SoftMax , namely: ; In the formula, represents a dynamic node embedding; is a transpose symbol; To ensure the relevance between the dynamic node embedding of the previous layer and the static node embedding , the momentum update is as follows: ; In the formula, is a momentum coefficient; the parameters of the dynamic node embedding are updated through back propagation.

[0028] The time convolution module adopts an expanded causal convolution network TCN, and for a given 1D sequence input of a node , the expanded convolution is defined as: ; In the formula, represents a 1D convolution filter with a kernel size of ; is time; represents an expansion factor; represents a convolution operation; By expanding the receptive field of the convolution kernel, the network can capture longer-range dependencies while reducing the number of required layers and computational complexity.

[0029] The time convolution module introduces a gating mechanism, which dynamically selects and weights information through two activation functions, thereby effectively managing information memory and forgetting in sequence modeling tasks. For a given two expanded initial layers 、 , the following formula is obtained , namely: In the formula, represents the amount of information processed through the activation function and adjusted through the gating mechanism, representing the first ithe final output of the node; 、 are activation functions; 、 、 、 respectively represent the parameters of the two extended initial layers; represents the Hadamard product, and the element-wise multiplication operation is performed on the two outputs.

[0030] By adopting the spatial propagation mechanism of the random walk diffusion process, the information propagation on the graph is modeled as a s-step graph signal diffusion process, and the input signal is diffused by truncated diffusion convolution The result obtained is represented as: ; In the formula, represents the step length of diffusion, and represents the number of propagation times; represents a weight matrix used to learn the features in the diffusion process; represents a degree diagonal matrix; represents an adjacency matrix; represents a transition matrix of the diffusion process.

[0031] The graph convolution module is used to perform information propagation operations on the output results of the static graph learning module and the dynamic graph learning module, and outputs through the output layer, specifically: ; In the formula, represents a feature matrix calculated by GCN, which is the comprehensive result of static graph and dynamic graph convolution operations; represents the weight matrix of the static graph in the th diffusion step; represents the static graph; represents the output of the time convolution module; represents the weight matrix of the dynamic graph in the th diffusion step; represents the dynamic graph; The output layer accumulates the information captured by the space-time layer through the skip connection, and then processes it through RELU and two 1x1 convolution layers, and finally predicts the final result through the fully connected layer.

[0032] S3), predicting the wind speed by using the trained wind speed prediction model.

[0033] ​The embodiment adopts a certain wind farm in the United States and 25 observation stations around it. These nodes are located in several important wind power regions in the United States, including Portland, San Francisco, Denver, Chicago, and are affected by the marine climate and winter storms, with frequent fluctuations in wind speed, showing shorter period changes. The data set contains hourly environmental temperature, wind speed, wind direction and barometric pressure measurements. The data from October 2012 to December 2016 is selected for training and verification, and the sample from January to October 2017 is used as the test set. Before inputting the model, the data set is normalized based on the training set using the min-max method. Each sample contains the history of selected meteorological variables of all nodes at multiple time steps.

[0034] In the embodiment, the history time is set to T = 30, and the prediction target is the wind speed prediction result of all nodes in the future T = 2, 4, 6 and 8. The model training uses the Adam optimizer with an initial learning rate of lr = 0.001. Four spatio-temporal layers with an initial inflation exponent factor of 1 are stacked. The starting 1x1 convolution has 1 input channel and 32 output channels. The node embedding dimension d is 10. The information propagation is modeled as s finite steps. All skip connection layers have 256 output channels. The output channels of the TCN module and the GCN module are both 32. The learning rate lr is 0.001. The training epoch is 100, and the batch size is set to 64. Table 1 compares the method of the embodiment with existing common wind speed prediction models. Among them, 2D-CNN, 2D-CNN + Attention, 2D-CNN + Upscaling represent capturing local spatio-temporal dependencies by applying convolution kernels in spatial and temporal dimensions; Multidimensional fuses 2D and 3D convolution features to realize multi-dimensional feature fusion through a fully connected layer; WeatherGCNet and WeatherGCNet with γ represent GCN-based models and their extended versions; Graph WaveNet represents a spatio-temporal graph modeling network based on diffusion convolution; MTGNN represents a multivariate spatio-temporal modeling network. MAE represents the mean absolute error, and MAE represents the mean square error.

[0035] Table 1 Comparison results of the method of the embodiment and existing common wind speed prediction methods As can be seen from Table 1, the method of the embodiment shows lower error values in MAE and MAE indicators, and the prediction effect is optimal at multiple prediction time steps. For example, Figure 5 Figure 1 shows the wind speed sequence comparison of different models at different steps, where Figure 5 (a), Figure 5 (b), Figure 5 (c),Figure 5 (d) prediction results of 2, 4, 6, 8 time lengths respectively. From Figure 5 (a), Figure 5 (b), Figure 5 (c), ​ (d) It can be seen that the embodiment can more accurately predict the wind speed and has stronger ability to capture the change rule of the wind speed.

[0036] The above embodiments and descriptions are only to illustrate the principles and the best embodiments of the present application, and various changes and improvements can be made without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application.

Claims

1. A multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning, characterized in that: The steps include: S1) Obtain wind speed data at multiple monitoring points in different time periods; S2), building a wind speed prediction model and training it using the acquired wind speed data; The wind speed prediction model includes a static graph learning module, a dynamic graph learning module, a graph convolution module, and a temporal convolution module. The static graph learning module captures the long-term static correlation between wind speed nodes by optimally weighting the geographic distance and dynamic time-warped distance of graph nodes. The dynamic graph learning module builds short-term dynamic associations based on stacked dynamic graph learning using node embedding and multi-head attention mechanisms. The temporal convolution module is used to capture deep temporal correlations and fully consider the changes in the relationship between different time points caused by changing weather conditions. The graph convolution module is used to perform information propagation operations on the output results of the static graph learning module and the dynamic graph learning module, and finally predict the final result through the fully connected layer; S3) Use the trained wind speed prediction model to predict the wind speed.

2. The multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning according to claim 1 is characterized by: In step S2), the static graph learning module introduces dynamic time warping (DTW) to quantify the similarity between different time series and capture the compression or stretching relationship between wind speed time series at different nodes. For different time series and }, its dynamic time warping distance Expressed as: ； in: ； Where, Representation node and The Euclidean distance between Indicates time Monitoring points Wind speed data; Indicates time Monitoring points Wind speed data; The dynamic time warping distance It is used to measure the optimal alignment path of two time series at different time steps. Therefore, it is based on the dynamic time warping distance. Construct the optimal weighted graph of different nodes. The DTW weighted adjacency matrix and position weighted adjacency matrix of the optimal weighted graph are: ； ； Where, represents the DTW weighted adjacency matrix; Indicates the number of monitoring points; : are the parameters of the position-weighted adjacency matrix; By and Combining them, we get: ； Where, Represents static graph structure representation; represents the learnable weight factor.

3. The multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning according to claim 1 is characterized by: In step S2), the dynamic graph learning module includes an information fusion module, a DGL module and a multi-layer perceptron MLP.

4. The multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning according to claim 3 is characterized by: In step S2), the information fusion module introduces a gate mechanism to adaptively fuse the graph structure and dynamic input; for each input window The information fusion in the time step of is as follows: ； ； ； ； Where, Reset gate. represents the update gate; represents the intermediate hidden state; Indicates the fused node information; 、 Represent the weight matrices of the reset gate and update gate respectively; 、 Represent the weight matrices of the reset gate and update gate respectively; is the weight matrix of the intermediate hidden state; is the sigmoid activation function; Represents dynamic input The matrix transformed by the FFN layer; represents node embedding; Indicates the number of nodes; Indicates the length of the historical time series; represents the dimension of embedding; is the tangent hyperbolic function; represents the Hadamard product.

5. The multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning according to claim 4 is characterized by: In step S2), the DGL module captures the feature dependency between different wind speed node time series through the multi-head attention mechanism. The DGL module first uses LayerNorm to fused node information. By normalizing, and then by The number of heads is calculated using the scaled inner product The connection matrix between nodes ,Right now: ； Where, Indicates in Nodes learned in space The correlation matrix between 、 Respectively The query vector and key vector weight matrix of the space; represents the scaling factor; represents the scale factor; represents the bias vector; for The mean of for The standard deviation of represents the inner product.

6. The multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning according to claim 5, characterized in that: In step S2), by The sum of the similarity results between the nodes captured in the space is used as the final similarity measure between the nodes, that is: ； ； Where, Represents the incidence matrix The matrix after the Dropout operation; Represents a random regularization operation that prevents overfitting by randomly dropping neurons, especially when the feature dimension is high or the amount of data is small; Indicates that the first h The cumulative similarity results between the nodes identified in the space are used as the final similarity measure between the nodes; represents the learnable parameter matrix; Indicates in The sum of the association matrices between all nodes learned in the space; Then smooth it out using LayerNorm; that is: ； ； Where, Indicates use Inner product to obtain the result of skip connection; express The result after passing through two FFN layers; Represents the LayerNorm normalization function; 、 Respectively 、 The learnable parameter matrix of the head; 、 、 、 Represents trainable parameters; Will It means that after two linear transformations in the FeedForward module and then two activation functions of RELU and SoftMax, the dynamic graph structure representation of the current layer is obtained. ;Right now: ； Where, represents dynamic node embedding; is the transposition character; To ensure the dynamic node embedding of the previous layer Embedded with static nodes The momentum update is as follows: ； Where, is the momentum coefficient; dynamic node embedding The parameters of are updated through back propagation.

7. The multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning according to claim 6, characterized in that: In step S2), the temporal convolution module uses an expanded causal convolutional network TCN. For a given node 1D sequence input , the dilated convolution is defined as: ； Where, The kernel size is 1D convolution filter; For time; represents the expansion factor; Represents the convolution operation; By expanding the receptive field of the convolution kernel, the network can capture longer-range dependencies while reducing the number of required layers and computational effort.

8. The multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning according to claim 7, characterized in that: In step S2), the temporal convolution module introduces a gating mechanism, which uses two activation functions to achieve dynamic selection and weighting of information, thereby effectively managing the memory and forgetting of information in the sequence modeling task. For the given two extended initial layers 、 , obtained by the following formula ,Right now: Where, It represents the amount of information processed by the activation function and regulated by the gating mechanism, representing the i The final output of each node; 、 Both are activation functions; 、 、 、 Represent the parameters of the two extended initial layers respectively; Represents the Hadamard product, which performs an element-wise multiplication of two outputs.

9. The multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning according to claim 8, characterized in that: In step S2), the information propagation on the graph is modeled as a graph signal diffusion process with a finite step length of s by adopting the spatial propagation mechanism of the random walk diffusion process, and the input signal is transformed into a graph signal diffusion process with a truncated diffusion convolution. Diffusion, the result Expressed as: ； Where, It represents the diffusion step length and the number of propagations; Represents the weight matrix, which is used to learn the features in the diffusion process; represents the degree diagonal matrix; represents the adjacency matrix; The transfer matrix representing the diffusion process.

10. The multi-node wind speed prediction method based on static and dynamic graph spatiotemporal learning according to claim 9, characterized in that: In step S2), the graph convolution module is used to perform information propagation operations on the output results of the static graph learning module and the dynamic graph learning module, and output them through the output layer. , specifically: ； Where, Represents the feature matrix calculated by GCN, which is the comprehensive result of the convolution operation of static graph and dynamic graph; Indicates the The weight matrix of the static graph in diffusion steps; Represents a static image; represents the output of the temporal convolution module; Indicates the The weight matrix of the dynamic graph in diffusion steps; Represents a dynamic graph; The output layer accumulates the information captured by the spatiotemporal layer through skip connections, and is processed by RELU and two 1×1 convolution layers before predicting the final result through a fully connected layer.

Images