Wind power prediction method based on cross-modal space-time attention fusion

Abstract

The invention relates to a wind power prediction method based on cross-modal space-time attention fusion. Accurate wind power prediction in different weather scenes is realized. The method comprises the following steps: preprocessing wind power numerical weather forecast data to obtain preprocessed data; performing feature extraction on the preprocessed data by adopting a Transform model to obtain a time feature; performing feature extraction on the preprocessed data by adopting a GAT model to obtain spatial features; performing dual-scale feature fusion on the time feature and the spatial feature by adopting a preset neural network to obtain a fused feature; and carrying out standardization processing on the fusion features to obtain standard wind power data.

Description

Technical Field

[0001] This application relates to the field of wind power prediction technology, and in particular to a wind power prediction method based on cross-modal spatiotemporal attention fusion. Background Technology

[0002] With the increasing proportion of wind power in the energy structure, wind power forecasting has become a key technology for ensuring the safe and stable operation of the power grid. Currently, mainstream wind power forecasting technologies are mainly divided into three categories. The first category is traditional statistical models, such as ARIMA and exponential smoothing, which model based on the statistical regularities of historical power data and are widely used in early-stage short-term forecasting scenarios. The second category is single deep learning models, including temporal modeling models (LSTM, GRU, Transformer) and spatial modeling models (GCN, GAT). Transformer is used for temporal forecasting due to its ability to capture long-sequence dependencies, while GAT is used for wind farm node topology modeling due to its adaptive spatial attention weight learning. The third category is simple hybrid models, which integrate temporal and spatial models through direct concatenation, fixed-weight summation, etc. (such as GAT+LSTM, Transformer+GCN), attempting to balance spatiotemporal feature modeling.

[0003] Although existing technologies have been applied in wind power prediction, the following key shortcomings still exist: incomplete capture of spatiotemporal features; traditional statistical models cannot model complex spatiotemporal coupling relationships, such as the lack of spatial topology awareness in a single Transformer model, making it difficult to utilize the physical connections between wind turbine nodes, and the neglect of temporal dependencies in a single GAT model, making it unable to adapt to short-term fluctuations and long-term trends in wind power; and low feature fusion efficiency, with existing hybrid models mostly using simple splicing or fixed-weight fusion, failing to address the modal differences between temporal and spatial features, resulting in insufficient feature interaction, poor fusion effect, and inability to dynamically adapt to changes in the importance of spatiotemporal features under different weather scenarios.

[0004] A search of existing technical literature revealed a patent application (application number 202411841746.5) entitled "A Method, System, Device, and Medium for Ultra-Short-Term Wind Power Prediction." This patent acquires historical wind power data, preprocesses it to obtain preprocessed wind power data, inputs the preprocessed wind power data into a GAT model to extract graph features, enhances the graph features using connection class labeling and position embedding methods, performs dropout regularization, and then inputs it into a Transformer model for further processing to obtain enhanced features. These enhanced features are then combined with meteorological data, and external factor features are extracted using a multilayer perceptron. The wind power prediction result is then obtained based on these external factor features. However, this patent has issues such as failing to consider temporal and spatial modal differences. Summary of the Invention

[0005] Therefore, it is necessary to provide a wind power prediction method based on cross-modal spatiotemporal attention fusion to address the above-mentioned technical problems and achieve accurate wind power prediction under different weather scenarios.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a wind power prediction method based on cross-modal spatiotemporal attention fusion, the method comprising: S1: Preprocess the wind power numerical weather forecast data to obtain the preprocessed data; S2: Use the Transformer model to extract features from the preprocessed data to obtain time features; S3: Use the GAT model to extract features from the preprocessed data to obtain spatial features; S4: A preset neural network is used to fuse the temporal features and the spatial features at two scales to obtain fused features; S5: Standardize the fusion features to obtain standard wind power data.

[0007] Preferably, the wind power numerical weather forecast data is preprocessed, including: The missing values ​​in the wind power numerical weather forecast data were filled in using a linear interpolation method to obtain complete wind power numerical weather forecast data. The complete wind power numerical weather forecast data is normalized using the min-max normalization method to obtain preprocessed data.

[0008] Preferably, the Transformer model includes: a multi-head attention mechanism module, a feedforward neural network module, and a residual connection module; The preprocessed data is sequentially passed through the multi-head attention mechanism module, the residual connection module, and the feedforward neural network module to obtain time features.

[0009] Preferably, step S3 includes: The preprocessed data is transformed using a dimension transformation function to obtain node features; The node features are linearly projected onto the learnable weight matrix initialized by Xavier to obtain enhanced node features. Bilinear projection is performed on the enhanced node features to obtain the node's left attention vector and right attention vector; The left attention vector and the right attention vector of the node are added together through a broadcast mechanism, and then activated using an activation function to obtain the spatial attention score. The spatial attention score is normalized using a normalization function to obtain the attention coefficient; The spatial features are obtained by weighted summation of the enhanced node features based on the attention coefficients.

[0010] Preferably, the calculation expressions for the node's left attention vector and the node's right attention vector are as follows:

[0011]

[0012] in, Represents the left attention vector of a node. Represents the right attention vector of a node. Represents a node Enhanced node features, Represents a node Enhanced node features, This represents the contribution of the capture node's own features. This indicates the contribution of the captured inter-node interaction features. This represents the dimension transformation function.

[0013] Preferably, the activation function is the LeakyReLU function, and the normalization function is the Softmax function.

[0014] Preferably, step S4 includes: Using the spatial features as the query vector and the temporal features as the key and value vectors, an attention mechanism is employed to obtain spatially enhanced temporal attention; using the temporal features as the query vector and the spatial features as the key and value vectors, an attention mechanism is employed to obtain temporally enhanced spatial attention. Global average pooling is performed on the spatially enhanced temporal attention and the temporally enhanced spatial attention to obtain the temporal global feature vector and the spatial global feature vector, respectively. A gating fusion mechanism is used to fuse the temporal global feature vector and the spatial global feature vector to obtain the first gating fusion feature; The spatial features and the temporal features are swapped in the first dimension using a dimension transformation function to obtain transformed spatial features and transformed temporal features; Using the transformation space features as query vectors and the transformation time features as key and value vectors, an attention mechanism is employed to obtain transformation space-enhanced temporal attention; using the transformation time features as query vectors and the transformation space features as key and value vectors, an attention mechanism is employed to obtain transformation time-enhanced spatial attention. Global average pooling is performed on the transformation space-enhanced temporal attention and the transformation time-enhanced spatial attention to obtain the first transformation time global feature vector and the first transformation space global feature vector. The transformation time global feature vector and the transformation space global feature vector are swapped in a second dimension using a dimension transformation function to obtain a second transformation time global feature vector and a second transformation space global feature vector. A gating fusion mechanism is used to fuse the second transformation time global feature vector and the second transformation spatial global feature vector to obtain the second gating fusion feature; The first gated fusion feature and the second gated fusion feature are normalized using the Softmax function to obtain the fusion feature.

[0015] Preferably, the Transformer model, the GAT model, and the preset neural network constitute a first neural network. The first neural network is connected to a second neural network through a first residual network, and the second neural network is connected to a third neural network through a second residual network. The first neural network, the second neural network, and the third neural network have the same structure. After step S4, the method further includes: The fusion feature and the preprocessed data are input together into the first residual network, and the output is the first fusion feature. The first fused feature is input into the second neural network, and the output is the second fused feature; The first fusion feature and the second fusion feature are input together into the second residual network to obtain the third fusion feature. After the third fusion feature is input into the third neural network, a fourth fusion feature is output and replaced.

[0016] Preferably, the first residual network and the second residual network have the same structure, and the first residual network includes: The fused feature and the preprocessed data are added together and then subjected to layer normalization to obtain the first fused feature.

[0017] Preferably, the fused features are standardized using a multilayer perceptron model.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention provides a wind power prediction method based on cross-modal spatiotemporal attention fusion. It uses a Transformer model to extract the temporal features of the data and a GAT model to extract the spatial features of the data. The parallel architecture of GAT spatial modeling and Transformer temporal modeling respectively captures spatial topological correlation and multi-scale temporal dependence. Furthermore, it adopts a dual-scale fusion that covers the dual-dimensional structural features of spatiotemporal data. The temporal scale focuses on the temporal evolution of nodes, while the spatial scale focuses on the correlation strength of nodes, resulting in stronger adaptability and higher quality of fused features. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a wind power prediction method based on cross-modal spatiotemporal attention fusion in one embodiment; Figure 2 This is a schematic diagram of the cross-modal attention pair fusion network structure of a wind power prediction method based on cross-modal spatiotemporal attention fusion in one embodiment; Figure 3 This is a simulation diagram comparing different models of wind power prediction based on cross-modal spatiotemporal attention fusion in one embodiment. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] Example 1 like Figure 1 As shown in the figure, this embodiment proposes a wind power prediction method based on cross-modal spatiotemporal attention fusion, the method comprising: S1: Preprocess the wind power numerical weather forecast data to obtain the preprocessed data; S2: Use the Transformer model to extract features from the preprocessed data to obtain time features; S3: Use the GAT model to extract features from the preprocessed data to obtain spatial features; S4: A preset neural network is used to fuse the temporal features and the spatial features at two scales to obtain fused features; S5: Standardize the fusion features to obtain standard wind power data.

[0022] Example 2 This embodiment further supplements the description of the wind power prediction method based on cross-modal spatiotemporal attention fusion proposed in the above embodiment.

[0023] Preprocessing of wind power numerical weather forecast data includes: The missing values ​​in the wind power numerical weather forecast data were filled in using a linear interpolation method to obtain complete wind power numerical weather forecast data. The complete wind power numerical weather forecast data is normalized using the min-max normalization method to obtain preprocessed data.

[0024] The specific implementation of this step is as follows: After using linear interpolation to fill in missing values ​​in the wind power numerical weather forecast data, the calculation expression for normalizing the complete wind power numerical weather forecast data using the minimum-maximum normalization method is as follows:

[0025] in, This refers to the preprocessed data, i.e., the normalized data. This represents complete numerical weather forecast data for wind power. This represents the maximum value in the dataset. This represents the minimum value in the dataset.

[0026] The Transformer model includes: a multi-head attention mechanism module, a feedforward neural network module, and a residual connection module; The preprocessed data is sequentially passed through the multi-head attention mechanism module, the residual connection module, and the feedforward neural network module to obtain time features.

[0027] The specific implementation of this step is as follows: Time information acquisition employs a Transformer encoder structure, including a multi-head attention mechanism module, a feedforward neural network module, and a residual connection module. The Transformer model introduces sine and cosine position encoding. Position encoding is generated using sine and cosine functions of different frequencies, providing information about the relative position of the input data within the wind power sequence. The position encoding is represented as:

[0028] in, The input data is positioned within the entire wind power series. The first characteristic One dimension, The dimension of the feature is denoted as . For the Transformer encoder model, its core is the multi-head attention module, and the multi-head attention mechanism formula is as follows:

[0029]

[0030]

[0031] in, For the first A self-attention function, Let be the linear mapping matrix of the multi-head attention function, and concat() be the concatenation function. , , For the first A self-attention weight matrix.

[0032] Step S3 includes: The preprocessed data is transformed using a dimension transformation function to obtain node features; The node features are linearly projected onto the learnable weight matrix initialized by Xavier to obtain enhanced node features. Bilinear projection is performed on the enhanced node features to obtain the node's left attention vector and right attention vector; The left attention vector and the right attention vector of the node are added together through a broadcast mechanism, and then activated using an activation function to obtain the spatial attention score. The spatial attention score is normalized using a normalization function to obtain the attention coefficient; The spatial features are obtained by weighted summation of the enhanced node features based on the attention coefficients.

[0033] The specific implementation of this step is as follows: In order to efficiently capture the spatial topology dependencies between wind turbine nodes, an efficient GAT architecture is designed, using bilinear projection to replace the traditional concat operation, reducing computational load and memory usage. The specific process is as follows: Dimensional flattening: Preprocessed data ∈R is first flattened by time step, converting the spatiotemporal dimension into "(batch × time) - node - feature", which is adapted to the spatial modeling paradigm of GAT:

[0034] in, This represents the dimension transformation function. Represents node characteristics.

[0035] Characteristic linear projection: through a learnable matrix ∈R maps node features to the same space, enhancing feature representation capabilities:

[0036] Among them, the weight matrix Xavier uniform initialization is used to ensure consistent signal variance during forward propagation. This indicates enhanced node features.

[0037] Attention score calculation: Traditional GAT calculates attention scores using the concat operation, which generates redundant temporary tensors. This embodiment uses bilinear projection instead.

[0038]

[0039] in, Represents the left attention vector of a node. Represents the right attention vector of a node. Represents a node Enhanced node features, Represents a node Enhanced node features, This represents the contribution of the capture node's own features. The transpose operation represents the contribution of inter-node interaction features. To achieve dimension alignment, the node's left attention vector and right attention vector are added via a broadcast mechanism, and then non-linearity is introduced through the LeakyReLU activation function.

[0040] in, Indicates the first The node is the first The spatial attention score of each node reflects the strength of the spatial association between the two nodes. To better allocate weights, the correlation (spatial attention score) calculated for all adjacent nodes is normalized using the Softmax function to obtain the attention coefficient. :

[0041] in, This represents the activation function LeakyReLU. This indicates that two nodes are being computed. and The relevance function. After obtaining the attention coefficients, the node can be obtained by following the weighted summation approach. The output feature vector:

[0042] in, This indicates the node at this layer. The new feature vector, i.e., spatial features; This represents the activation function.

[0043] Step S4 includes: Using the spatial features as the query vector and the temporal features as the key and value vectors, an attention mechanism is employed to obtain spatially enhanced temporal attention; using the temporal features as the query vector and the spatial features as the key and value vectors, an attention mechanism is employed to obtain temporally enhanced spatial attention. Global average pooling is performed on the spatially enhanced temporal attention and the temporally enhanced spatial attention to obtain the temporal global feature vector and the spatial global feature vector, respectively. A gating fusion mechanism is used to fuse the temporal global feature vector and the spatial global feature vector to obtain the first gating fusion feature; The spatial features and the temporal features are swapped in the first dimension using a dimension transformation function to obtain transformed spatial features and transformed temporal features; Using the transformation space features as query vectors and the transformation time features as key and value vectors, an attention mechanism is employed to obtain transformation space-enhanced temporal attention; using the transformation time features as query vectors and the transformation space features as key and value vectors, an attention mechanism is employed to obtain transformation time-enhanced spatial attention. Global average pooling is performed on the transformation space-enhanced temporal attention and the transformation time-enhanced spatial attention to obtain the first transformation time global feature vector and the first transformation space global feature vector. The transformation time global feature vector and the transformation space global feature vector are swapped in a second dimension using a dimension transformation function to obtain a second transformation time global feature vector and a second transformation space global feature vector. A gating fusion mechanism is used to fuse the second transformation time global feature vector and the second transformation space global feature vector to obtain the second gating fusion feature; The first gated fusion feature and the second gated fusion feature are normalized using the Softmax function to obtain the fusion feature.

[0044] The specific implementation of this step is as follows: temporal features and spatial features belong to different modalities of information, and their feature distribution and expression focus differ: temporal features focus on "temporal change patterns," while spatial features focus on "node association strength." To achieve effective alignment of the two types of features, and to ensure that the weights can adapt to the dynamic changes in the importance of spatiotemporal features under different scenarios, this embodiment uses a pre-defined neural network that is a cross-modal attention pair fusion network, with the structure as follows: Figure 2As shown, dependencies differ across different scales (temporal scale: T-dimensional; spatial scale: N-dimensional): the temporal scale focuses on the "historical-future evolution of the same node," while the spatial scale focuses on the "inter-node relationships at the same time step." The multi-scale fusion block enhances the model's ability to capture multi-scale information by fusing features at both scales.

[0045] At the temporal scale, spatially enhanced temporal attention is obtained by using the spatial feature G output by GAT as the query (Q) and the temporal feature T output by Transformer as the key (K) and value (V); temporally enhanced spatial attention is obtained by using the temporal feature T output by Transformer as the query (Q) and the spatial feature G output by GAT as the key (K) and value (V), and feature space alignment is achieved through independent linear layers.

[0046] in To ensure a learnable projection matrix , , The two attentions obtained, which are in the same feature space, are:

[0047] in, This indicates spatially enhanced temporal attention. This represents temporally enhanced spatial attention, which performs spatiotemporal global pooling on the enhanced features to obtain the global feature vector for each sample:

[0048] in, Represents the global feature vector over time. Representing the spatial global feature vector, a multi-head adaptive gating is introduced into the spatiotemporal global feature vector, fusing different attention head features at each spatiotemporal location separately to improve the fusion precision:

[0049] in, , For adaptive parameters, , As a bias, the first gated fused feature is obtained after multi-head feature fusion. .

[0050] At the spatial scale, a dimension transformation function is used to swap the dimensions of spatial and temporal features, resulting in transformed spatial and temporal features. The temporal and node dimensions are also swapped. The transformed spatial features are used as the query vector, and the transformed temporal features as the key and value vectors. An attention mechanism is employed to obtain transformed spatial enhanced temporal attention. The transformed temporal features are used as the query vector, and the transformed spatial features as the key and value vectors. An attention mechanism is then employed to obtain transformed temporal enhanced spatial attention. Global average pooling is applied to both the transformed spatial enhanced temporal attention and the transformed temporal enhanced spatial attention to obtain a first transformed temporal global feature vector and a first transformed spatial global feature vector. A dimension transformation function is then used to swap the dimensions of the transformed temporal global feature vector and the transformed spatial global feature vector, resulting in a second transformed temporal global feature vector and a second transformed spatial global feature vector. A gated fusion mechanism is then used to fuse the second transformed temporal global feature vector and the second transformed spatial global feature vector to obtain a second gated fused feature.

[0051] Indicates cross-integration, The dimension transformation function treats the spatial dimension as a "pseudo-temporal" dimension for fusion, simulating the "global correlation between nodes at the same time step." The original dimension is then restored. The first and second gated fusion features are normalized using the Softmax function, and then subjected to multi-scale weighted fusion to obtain the fused features.

[0052] in, , As learnable parameters, the model can automatically adjust the importance of temporal and spatial scales based on task characteristics.

[0053] The fused features are standardized using a multilayer perceptron model.

[0054] The specific implementation of this step is as follows: a multilayer perceptron (MLP) model is used to map the fused features to the predicted values, and the fused features ( * , , ), It's a batch. It is a node. It is all time steps. It is an advanced spatiotemporal feature that merges the dimensions of batches and nodes, allowing each node to be processed independently in each batch, and finally outputting standard wind power data.

[0055] Example 3 The Transformer model, the GAT model, and the preset neural network constitute a first neural network. The first neural network is connected to a second neural network through a first residual network, and the second neural network is connected to a third neural network through a second residual network. The first neural network, the second neural network, and the third neural network have the same structure. After step S4, the following is also included: The fusion feature and the preprocessed data are input together into the first residual network, and the output is the first fusion feature. The first fused feature is input into the second neural network, and the output is the second fused feature; The first fusion feature and the second fusion feature are input together into the second residual network to obtain the third fusion feature. After the third fusion feature is input into the third neural network, a fourth fusion feature is output and replaced.

[0056] The specific implementation of this step is as follows: To achieve optimal prediction results, the first neural network, composed of the Transformer model, the GAT model, and the cross-modal attention network, is stacked in three layers. Each layer includes two parts: directly performing cross-modal attention processing on temporal features T and spatial features G, and performing cross-modal attention processing on spatiotemporal features through dimension exchange. Furthermore, the layers are connected through a residual network. Figure 3 As shown, commonly used classic prediction models CNN_LSTM, BiGRU, and MSTGCN were selected to conduct prediction comparison experiments with the CMA fusion model of this embodiment. To more intuitively demonstrate the prediction capabilities of each comparison model, the error evaluation results of each model are shown in Table 1. The MAE error of the model in this embodiment is 0.21, and the RMSE error is 0.40. Compared with the CNN_LSTM model, the MAE and RMSE errors are reduced by 0.15 and 0.04 MW, respectively; compared with the BiGRU model, the MAE and RMSE errors are reduced by 0.08 and 0.05, respectively; and compared with the MSTGCN model, the MAE and RMSE errors are reduced by 0.04 and 0.02, respectively. The above prediction error results show that, under the premise of the same prediction step size, the fusion model proposed in this embodiment has higher prediction accuracy than the comparison models. Compared with the comparison models, the R2 of this model is 0.975, which is closer to 1, indicating that the model has better prediction capabilities.

[0057] Table 1

[0058] The first residual network and the second residual network have the same structure. The first residual network includes: The fused feature and the preprocessed data are added together and then subjected to layer normalization to obtain the first fused feature.

Claims

1. A wind power prediction method based on cross-modal spatiotemporal attention fusion, characterized in that, include: S1: Preprocess the wind power numerical weather forecast data to obtain the preprocessed data; S2: Use the Transformer model to extract features from the preprocessed data to obtain time features; S3: Use the GAT model to extract features from the preprocessed data to obtain spatial features; S4: A preset neural network is used to fuse the temporal features and the spatial features at two scales to obtain fused features; S5: Standardize the fusion features to obtain standard wind power data.

2. The wind power prediction method based on cross-modal spatiotemporal attention fusion according to claim 1, characterized in that, Preprocessing of wind power numerical weather forecast data includes: The missing values ​​in the wind power numerical weather forecast data were filled in using a linear interpolation method to obtain complete wind power numerical weather forecast data. The complete wind power numerical weather forecast data is normalized using the min-max normalization method to obtain preprocessed data.

3. The wind power prediction method based on cross-modal spatiotemporal attention fusion according to claim 1, characterized in that, The Transformer model includes: a multi-head attention mechanism module, a feedforward neural network module, and a residual connection module; The preprocessed data is sequentially passed through the multi-head attention mechanism module, the residual connection module, and the feedforward neural network module to obtain time features.

4. The wind power prediction method based on cross-modal spatiotemporal attention fusion according to claim 1, characterized in that, Step S3 includes: The preprocessed data is transformed using a dimension transformation function to obtain node features; The node features are linearly projected onto the learnable weight matrix initialized by Xavier to obtain enhanced node features. Bilinear projection is performed on the enhanced node features to obtain the node's left attention vector and right attention vector; The left attention vector and the right attention vector of the node are added together through a broadcast mechanism, and then activated using an activation function to obtain the spatial attention score. The spatial attention score is normalized using a normalization function to obtain the attention coefficient; The spatial features are obtained by weighted summation of the enhanced node features based on the attention coefficients.

5. The wind power prediction method based on cross-modal spatiotemporal attention fusion according to claim 4, characterized in that, The calculation expressions for the node's left attention vector and the node's right attention vector are as follows: in, Represents the left attention vector of a node. Represents the right attention vector of a node. Represents a node Enhanced node features, Represents a node Enhanced node features, This represents the contribution of the capture node's own features. This indicates the contribution of the captured inter-node interaction features. This represents the dimension transformation function.

6. The wind power prediction method based on cross-modal spatiotemporal attention fusion according to claim 4, characterized in that, The activation function is the LeakyReLU function, and the normalization function is the Softmax function.

7. The wind power prediction method based on cross-modal spatiotemporal attention fusion according to claim 1, characterized in that, Step S4 includes: Using the spatial features as the query vector and the temporal features as the key and value vectors, an attention mechanism is employed to obtain spatially enhanced temporal attention; using the temporal features as the query vector and the spatial features as the key and value vectors, an attention mechanism is employed to obtain temporally enhanced spatial attention. Global average pooling is performed on the spatially enhanced temporal attention and the temporally enhanced spatial attention to obtain the temporal global feature vector and the spatial global feature vector, respectively. A gating fusion mechanism is used to fuse the temporal global feature vector and the spatial global feature vector to obtain the first gating fusion feature; The spatial features and the temporal features are swapped in the first dimension using a dimension transformation function to obtain transformed spatial features and transformed temporal features; Using the transformation space features as query vectors and the transformation time features as key and value vectors, an attention mechanism is employed to obtain transformation space-enhanced temporal attention; using the transformation time features as query vectors and the transformation space features as key and value vectors, an attention mechanism is employed to obtain transformation time-enhanced spatial attention. Global average pooling is performed on the transformation space-enhanced temporal attention and the transformation time-enhanced spatial attention to obtain the first transformation time global feature vector and the first transformation space global feature vector. The transformation time global feature vector and the transformation space global feature vector are swapped in a second dimension using a dimension transformation function to obtain a second transformation time global feature vector and a second transformation space global feature vector. A gating fusion mechanism is used to fuse the second transformation time global feature vector and the second transformation space global feature vector to obtain the second gating fusion feature; The first gated fusion feature and the second gated fusion feature are normalized using the Softmax function to obtain the fusion feature.

8. The wind power prediction method based on cross-modal spatiotemporal attention fusion according to claim 7, characterized in that, The Transformer model, the GAT model, and the preset neural network constitute a first neural network. The first neural network is connected to a second neural network through a first residual network, and the second neural network is connected to a third neural network through a second residual network. The first neural network, the second neural network, and the third neural network have the same structure. After step S4, the following is also included: The fusion feature and the preprocessed data are input together into the first residual network, and the output is the first fusion feature. The first fused feature is input into the second neural network, and the output is the second fused feature; The first fusion feature and the second fusion feature are input together into the second residual network to obtain the third fusion feature. After the third fusion feature is input into the third neural network, a fourth fusion feature is output and replaced.

9. The wind power prediction method based on cross-modal spatiotemporal attention fusion according to claim 8, characterized in that, The first residual network and the second residual network have the same structure. The first residual network includes: The fused feature and the preprocessed data are added together and then subjected to layer normalization to obtain the first fused feature.

10. The wind power prediction method based on cross-modal spatiotemporal attention fusion according to claim 1, characterized in that, The fused features are standardized using a multilayer perceptron model.

Images