A physical perception wind power prediction method of complex three-dimensional spatio-temporal heterogeneity

Abstract

This invention provides a physical perception-based wind power prediction method for complex three-dimensional spatiotemporal heterogeneity. The method includes the following steps: S101, abstracting the wind farm as a directed graph; S102, decoupling the representation in the vertical spatial dimension and mapping multi-height weather forecast data to a high-dimensional latent space using dense feature fusion technology; S103, decoupling the representation in the horizontal spatial dimension, constructing a wind turbine cluster topology graph, introducing an improved graph attention network and combining it with a physical perception masking mechanism; S104, expanding the receptive field using dilated causal convolution of a temporal convolutional network, extracting local high-frequency fluctuation features of the wind farm and filtering out random noise, and inputting the local high-frequency fluctuation features into an Informer model; S105, constructing an ST-GAT-Informer model, training the ST-GAT-Informer model, and using the ST-GAT-Informer model to predict wind power. This invention verifies the effectiveness of three-dimensional spatial information fusion and temporal multi-scale feature extraction in improving the accuracy of wind power prediction.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and in particular to a physical perception method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity. Background Technology

[0002] With a high proportion of renewable energy connected to the grid, accurate short-term wind power forecasting is crucial for the safe and stable operation of the power system. However, the strong randomness of wind speed and the complex three-dimensional spatiotemporal heterogeneity within wind farms mean that existing deep learning models, when processing multi-location, multi-dimensional data, suffer from insufficient representation of dynamic topological dependencies between units and vertical wind shear physical characteristics due to coarse-grained flattening processing.

[0003] The representation of three-dimensional spatial structure information is simplistic and lacks refined modeling of vertical-horizontal coupling characteristics. Most existing deep learning models treat multi-site wind turbine data as simple multi-channel time series or reduce it to a one-dimensional vector through flattening operations. This coarse-grained approach severely ignores the inherent three-dimensional spatial heterogeneity of wind farms.

[0004] In the vertical dimension, wind speed varies with altitude according to specific hydrodynamic laws. Ignoring wind shear characteristics across multiple altitude layers leads to the loss of crucial vertical energy transfer information in the model, thereby increasing prediction bias. In the horizontal dimension, the relative topological positions between wind turbines reflect the spatial distribution of local micro-meteorological systems and the evolution direction of weather systems. Coarse-grained feature reshaping disrupts the original manifold structure, resulting in the loss of spatial topological information, making it impossible for the model to effectively model the complex wake interactions among distributed turbines explicitly. Summary of the Invention

[0005] Therefore, the purpose of this invention is to provide a physical sensing wind power prediction method for complex three-dimensional spatiotemporal heterogeneity, so as to solve or at least partially solve the above-mentioned problems existing in the prior art.

[0006] To achieve the above objectives, this invention provides a physical sensing method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity, the method comprising the following steps: S101. Abstract the wind farm into a directed graph and use the directed graph to represent the heterogeneity within the wind farm. S102. Based on directed graphs, decoupled representation is performed in the vertical space dimension, and dense feature fusion technology is used to map weather forecast data from multiple altitude layers to a high-dimensional latent space. S103. Based on directed graphs, decoupled representation is performed in the horizontal spatial dimension to construct a wind turbine cluster topology graph. An improved graph attention network is introduced and combined with a physical perception masking mechanism to dynamically reconstruct the wind turbine cluster topology graph. S104. Expand the receptive field by using dilated causal convolution of temporal convolutional network, extract local high-frequency fluctuation features of wind farm and filter out random noise, input local high-frequency fluctuation features into Informer model, use the probabilistic sparse attention mechanism of Informer model to reduce computational complexity and capture the global time dependence of long sequence prediction. S105. Based on global time dependence and decoupled representation of vertical and horizontal spatial dimensions, construct the ST-GAT-Informer model, train the ST-GAT-Informer model, and use the ST-GAT-Informer model to predict wind power.

[0007] Furthermore, the wind farm is abstracted as a directed graph, represented as follows:

[0008] in, It is a directed graph. for A set of nodes consisting of typhoon turbines. This is the set of edges connecting the nodes.

[0009] Furthermore, the method of mapping multi-altitude weather forecast data to a high-dimensional latent space using dense feature fusion technology specifically includes the following steps: S21. Decoupling representation is performed in the vertical spatial dimension by introducing highly relevant physical prior vectors, as follows:

[0010] in, For physical prior vectors, For the logarithmic transformation term of height, Standardized height item, This is a transpose operation; S22. In the vertical spatial dimension, each node contains weather forecast data at different altitude levels. Any node in... The original input features at time step 1 are defined as tensors containing vertical profile information, as follows:

[0011] in, It is a tensor containing vertical profile information. The number of vertical height floors, For each height layer, the physical feature dimension, For the real number field; S23. Based on the physical prior vector and the original input features, the weather forecast data is mapped to a high-dimensional latent space using dense feature fusion technology, as shown below:

[0012] in, This is the initial embedding vector for nodes containing aerodynamic information. For activation function, and These are the learnable weight matrix and the bias term, respectively. This is a vector concatenation operation.

[0013] Furthermore, step S103 specifically includes the following steps: S31. Based on any node and its neighboring nodes, obtain the query, key, and value vectors through a linear transformation matrix mapping, and calculate the initial attention coefficients, as shown below:

[0014] in, The initial attention coefficient, For attention weight vectors, It is a linear transformation matrix. For nodes The input feature representation, For nodes Feature input representation; S32. Using a physical sensing masking mechanism, the azimuth angle calculated from the absolute coordinates of the wind turbine is matched with the real-time average wind direction, as shown below:

[0015] in, This is a physical sensing masking mechanism. This represents the real-time average wind direction. It is the azimuth angle; S33. Based on the initial attention coefficient and the physical perception masking mechanism, the final normalized spatial contribution weight is calculated. The normalized spatial contribution weight is used to dynamically measure the intensity of the wake influence inside the wind farm, as shown below:

[0016] in, Contribute weights to the normalized space. For nodes With nodes Initial attention coefficients between For nodes With nodes The physical sensing mask values ​​between them For nodes The set of domain nodes, For nodes The index of a certain neighboring node; S34. Based on the normalized spatial contribution weight, calculate the node update characteristics, and dynamically reconstruct the wind turbine cluster topology based on the node update characteristics, as shown below:

[0017] in, Update the features for the nodes.

[0018] Furthermore, the method of expanding the receptive field using dilated causal convolution of a temporal convolutional network, extracting local high-frequency fluctuation features of the wind farm, and filtering out random noise is expressed as follows:

[0019] in, For convolution kernel, It is the expansion factor. The kernel size is [size]. For time steps, For time step The output after dilated causal convolution, The index corresponding to the convolution kernel. For input time series features, The input features are the historical moments that participate in the convolution operation after dilation sampling.

[0020] Furthermore, the reduction of computational complexity using the probabilistic sparse attention mechanism of the Informer model is expressed as follows:

[0021] in, This is the sparse query matrix after KL divergence filtering. For querying the matrix, The key matrix, For value matrices, The feature dimension of the key vector. This is a transpose operation.

[0022] Furthermore, the Informer model is also used to address the error accumulation problem in long sequence prediction by introducing a generative decoder, as shown below:

[0023] in, To decode the input vector, a concatenation mechanism is used to construct it; The start token is a known historical segment extracted from the end of the input sequence to provide decoding context; Placeholder for the sequence to be predicted.

[0024] Furthermore, the prediction of wind power using the ST-GAT-Informer model specifically includes the following steps: S51. Obtain the length of the historical observation sequence and use the ST-GAT-Informer model to learn the nonlinear mapping function; S52. Calculate the historical spatiotemporal state based on the length of the historical observation sequence. , means as follows:

[0025] in, The length of the historical observation sequence. The historical spatiotemporal state at time t; S53. Based on the nonlinear mapping function and historical spatiotemporal state, predict the wind power output for the next Q time steps, as shown below:

[0026] in, It is a nonlinear mapping function. Encoding the periodic time features of the corresponding time period. This is the wind power prediction result for the ST-GAT-Informer model for the next Q time steps.

[0027] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention proposes a physical perception wind power prediction method for complex three-dimensional spatiotemporal heterogeneity. It proposes a modeling strategy of "horizontal-vertical spatial orthogonal decoupling", which decouples and represents the wind power in the vertical space dimension. By using dense feature fusion technology, the weather forecast data of multiple altitude layers are mapped to a high-dimensional latent space, and the nonlinear fluid law of the vertical wind shear profile is explicitly captured. 2. Decoupled representation is performed in the horizontal spatial dimension. By introducing an improved graph attention network and combining it with a physical perception masking mechanism, the topology of the wind turbine cluster is dynamically reconstructed to adaptively capture the spatiotemporal evolution characteristics of the wake effect. 3. By combining temporal convolutional networks with the Informer model, a synergistic effect of "local denoising first, then global capture" is achieved, accurately capturing the global temporal dependencies of long sequence predictions. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a schematic diagram of a physical perception method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity, provided as an embodiment of the present invention. Detailed Implementation

[0030] The principles and features of the present invention are described below with reference to the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0031] Reference Figure 1 This embodiment provides a physical perception method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity. The method includes the following steps: S101. Abstract the wind farm into a directed graph and use the directed graph to represent the heterogeneity within the wind farm. S102. Based on directed graphs, decoupled representation is performed in the vertical space dimension, and dense feature fusion technology is used to map weather forecast data from multiple altitude layers to a high-dimensional latent space. S103. Based on directed graphs, decoupled representation is performed in the horizontal spatial dimension to construct a wind turbine cluster topology graph. An improved graph attention network is introduced and combined with a physical perception masking mechanism to dynamically reconstruct the wind turbine cluster topology graph. S104. Expand the receptive field by using dilated causal convolution of temporal convolutional network, extract local high-frequency fluctuation features of wind farm and filter out random noise, input local high-frequency fluctuation features into Informer model, use the probabilistic sparse attention mechanism of Informer model to reduce computational complexity and capture the global time dependence of long sequence prediction. S105. Based on global time dependence and decoupled representation of vertical and horizontal spatial dimensions, construct the ST-GAT-Informer model, train the ST-GAT-Informer model, and use the ST-GAT-Informer model to predict wind power.

[0032] This embodiment proposes a three-dimensional spatial modeling strategy of "horizontal-vertical spatial orthogonal decoupling," abandoning the traditional feature flattening approach. This embodiment reconstructs spatial information from two orthogonal dimensions to achieve comprehensive perception of the three-dimensional wind farm, including: effectively capturing the nonlinear characteristics of the vertical wind shear profile through decoupling representation in the vertical spatial dimension; and adaptively capturing the spatiotemporal evolution characteristics of the wake effect through decoupling representation in the horizontal spatial dimension. Furthermore, a serial processing flow of "local denoising first, then global capture" is designed using a temporal convolutional network and the Informer model to achieve deep deconstruction of complex temporal dynamic characteristics.

[0033] In step S101, in order to accurately characterize the heterogeneity within the wind farm, the wind farm is abstracted as a directed graph, represented as follows:

[0034] in, It is a directed graph. for A set of nodes consisting of typhoon turbines. This is the set of edges connecting nodes, used to model the spatial airflow coupling relationship between wind turbine units.

[0035] In step S102, the process of mapping multi-height-layer weather forecast data to a high-dimensional latent space using dense feature fusion technology specifically includes the following steps: S21. Decoupling representation is performed in the vertical spatial dimension by introducing highly relevant physical prior vectors, as follows:

[0036] in, For physical prior vectors, For the logarithmic transformation term of height, Standardized height item, This is a transpose operation used to combine the components into a physical prior vector in column vector form; S22. In the vertical spatial dimension, each node contains weather forecast data at different altitude levels. Any node in... The original input features at time step 1 are defined as tensors containing vertical profile information, as follows:

[0037] in, It is a tensor containing vertical profile information. The number of vertical height floors, The physical characteristic dimensions (such as wind speed and wind direction) for each altitude layer. It is the field of real numbers.

[0038] The entire wind farm is constantly global input tensor It is expressed as follows:

[0039] in, This represents the number of wind turbine nodes in the entire wind farm.

[0040] S23. Based on the physical prior vector and the original input features, the weather forecast data is mapped to a high-dimensional latent space using dense feature fusion technology, as shown below:

[0041] in, This is the initial embedding vector for nodes containing aerodynamic information. For activation function, and These are the learnable weight matrix and the bias term, respectively. This is a vector concatenation operation.

[0042] In this embodiment, in the vertical spatial dimension, wind speed variation with altitude typically follows fluid dynamic laws such as logarithmic or exponential laws, and there is a strong and stable nonlinear correlation between different altitude layers. Treating different altitude layers as graph nodes would introduce significant computational redundancy. Therefore, considering the strict physical adjacency and stable correlation between vertical altitude layers, this embodiment reduces computational redundancy while preserving the aerodynamic profile through steps S21-S23. Furthermore, the ST-GAT-Informer model automatically learns the distribution characteristics of the vertical wind profile and outputs initial node embedding vectors containing aerodynamic information.

[0043] Step S103 specifically includes the following steps: S31, Based on arbitrary nodes and its neighboring nodes The query, key, and value vectors are obtained through linear transformation matrix mapping, and the initial attention coefficients are calculated, as follows:

[0044] in, The initial attention coefficient, For attention weight vectors, It is a linear transformation matrix. For nodes The input feature representation, For nodes Feature input representation; S32. To enable the ST-GAT-Informer model to sense changes in wind direction and wake, a physical sensing mask mechanism is introduced. This mechanism matches the azimuth angle calculated from the absolute coordinates of the wind turbine with the real-time average wind direction. When the wind direction causes the wind turbine (node) to... ) is located in the wind turbine (node) If the connection weight is located upwind of the target sector and the included angle is within the effective physical sector, the connection weight is retained; otherwise, it is masked, as shown below:

[0045] in, This is a physical sensing masking mechanism. This represents the real-time average wind direction. It is the azimuth angle; S33. Based on the constraints of the physical perception masking mechanism and the initial attention coefficient, calculate the final normalized spatial contribution weight. When wind flows past upstream wind turbines, the wake effect it generates will delay and affect the power output of downstream wind turbines. The intensity of the wake effect within the wind farm is dynamically measured by the normalized spatial contribution weight, as shown below:

[0046] in, Contribute weights to the normalized space. For nodes With nodes Initial attention coefficients between For nodes With nodes The physical sensing mask values ​​between them For nodes The set of domain nodes, For nodes The index of a certain neighboring node; S34. Based on the normalized spatial contribution weight, calculate the node update characteristics, and dynamically reconstruct the wind turbine cluster topology based on the node update characteristics, as shown below:

[0047] in, Update the features for the nodes.

[0048] In this embodiment, in the horizontal spatial dimension, the spatial dependencies between distributed wind turbines are mainly dominated by local micro-meteorological systems and wake effects, exhibiting significant dynamism and anisotropy. Traditional graph convolutional networks (GCNs) typically rely on predefined static adjacency matrices, making it difficult to adapt to topologies that change in real time with the meteorological environment. Therefore, this embodiment, through steps S31-S34, endows the ST-GAT-Informer model with the ability to perceive wind direction and wake changes, enabling the ST-GAT-Informer model to accurately correct the power prediction of downstream wind turbines using the state information of upstream wind turbines, thus achieving explicit modeling of the complex aerodynamic interactions between distributed wind turbines. For example, when the wind direction causes the wind turbine (node) to... It is located right at the wind turbine (node) When the wind direction is upwind, the ST-GAT-Informer model will automatically assign a wind turbine (node) to the wind turbine. A greater weight is given to upstream wind turbines, thereby using the state information of upstream wind turbines to correct the power forecast of downstream wind turbines.

[0049] In step S104, the dilated causal convolution of the temporal convolutional network (TCN) effectively expands the receptive field through an exponentially increased dilation factor without destroying the temporal causal relationship, extracting local high-frequency fluctuation features of the wind farm and filtering out random noise, as shown below:

[0050] in, For convolution kernel, It is the expansion factor. The kernel size is [size]. For time steps, For time step The output after dilated causal convolution, The index corresponding to the convolution kernel. For input time series features, This provides the historical time-series input features after dilation sampling for convolution operations. This embodiment can effectively extract local high-frequency fluctuation features and smooth random interference, providing a more robust feature representation for subsequent long sequence modeling.

[0051] In step S104, to address the computational bottleneck faced by standard self-attention mechanisms when processing long-term time-series data, the probabilistic sparse attention mechanism of the Informer model is used to filter out query vectors that significantly contribute to the attention distribution, thereby reducing computational complexity. Optimized to The optimized computational logic is expressed as follows:

[0052] in, This is the sparse query matrix after KL divergence filtering. For querying the matrix, The key matrix, For value matrices, The feature dimension of the key vector. This is a transpose operation.

[0053] In step S104, the Informer model is further used to address the error accumulation problem in long sequence prediction by introducing a generative decoder. The generative decoder directly outputs the predicted values ​​for all future time steps in parallel through a one-time forward propagation, as shown below:

[0054] in, To decode the input vector, a concatenation mechanism is used to construct it; The start token is a known historical segment extracted from the end of the input sequence to provide decoding context; This is a placeholder for the sequence to be predicted, usually initialized to zero or the mean of the target sequence.

[0055] In step S105, during the ST-GAT-Informer model training phase, the selection of the objective function directly determines the ST-GAT-Informer model's ability to capture power mutation features. Given that wind power data may contain outliers introduced by sensor inaccuracies or acquisition failures, the mean square error is easily affected by extreme values, leading to gradient instability. Therefore, this embodiment selects Smooth L1 Loss as the basic objective function to enhance the robustness of the training process and maintain the smooth convergence characteristic of the squared term when the error is small, while switching to linear loss when the error is large. This effectively limits the negative impact of outliers on the ST-GAT-Informer model weights, as shown below:

[0056] in, The Smooth L1 loss function is used. This represents the actual wind power output. The wind power value predicted by the ST-GAT-Informer model. This refers to the piecewise threshold parameter in the Smooth L1 loss function; To address the common problems of "prediction phase lag" and inaccurate capture of power change rate in long sequence prediction, this embodiment introduces a variation loss. This variation loss term minimizes the mean square error of the first-order difference between the predicted and true sequences, thus forcing the ST-GAT-Informer model to learn the dynamic evolution of power fluctuations. Its mathematical definition is as follows:

[0057] in, To predict sequence length, The loss function is constrained by the rate of change. For the first The actual wind power value at each time step For the first The actual wind power value at each time step For the first The predicted wind power value at each time step. For the first The predicted wind power value at each time step; In this embodiment, the two parts of the loss are finally weighted and summed using hyperparameters to construct a joint optimization hybrid loss function, as shown below:

[0058] in, For a mixed loss function, For hyperparameters. Through a collaborative optimization strategy using hybrid loss functions, the ST-GAT-Informer model can more accurately fit the slope changes of the power curve while maintaining numerical accuracy, thereby improving the prediction accuracy of power ramp-up events.

[0059] In order to verify the superiority of the ST-GAT-Informer model in the multidimensional spatiotemporal wind power prediction task, this embodiment selects a representative comparative model as the benchmark, covering mainstream methods from traditional deep learning to the latest spatiotemporal architecture. The key parameters of the comparative model are shown in Table 1 below.

[0060] Table 1 Key Parameters of the Comparison Model

[0061] In this embodiment, experiments were conducted based on the comparative model and the ST-GAT-Informer model. The quantitative performance of the comparative model and the ST-GAT-Informer model on four key indicators—MAE, RMSE, R2, and MAPE—over one day in the experiment is detailed in Table 2 below.

[0062] Table 2 Evaluation Indicators for Comparative Experiments

[0063] In this embodiment, based on the experimental results shown in Table 2, the ST-GAT-Informer model significantly outperforms existing mainstream benchmark models (i.e., the comparison models mentioned above) across all prediction step lengths and evaluation metrics. In the 24-step (6-hour) short-term prediction task, the MAE of the ST-GAT-Informer model is only 0.6359, a reduction of approximately 44.9% compared to the second-best performing spatiotemporal model, GAT-Crossformer (1.1537), and a reduction of over 80% compared to the traditional Bi-LSTM model (3.1927). This demonstrates the necessity of capturing the heterogeneity of three-dimensional space through vertical-horizontal decoupling modeling.

[0064] In step S105, the prediction of wind power using the ST-GAT-Informer model specifically includes the following steps: S51. Obtain the length of the historical observation sequence and use the ST-GAT-Informer model to learn the nonlinear mapping function; S52. Calculate the historical spatiotemporal state based on the length of the historical observation sequence. , means as follows:

[0065] in, The length of the historical observation sequence. The historical spatiotemporal state at time t; S53. Based on the nonlinear mapping function and historical spatiotemporal state, predict the wind power output for the next Q time steps, as shown below:

[0066] in, It is a nonlinear mapping function. Encoding the periodic time features of the corresponding time period. This is the wind power prediction result for the ST-GAT-Informer model for the next Q time steps.

[0067] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity using physical sensing, characterized in that, The method includes the following steps: S101. Abstract the wind farm into a directed graph and use the directed graph to represent the heterogeneity within the wind farm. S102. Based on directed graphs, decoupled representation is performed in the vertical space dimension, and dense feature fusion technology is used to map weather forecast data from multiple altitude layers to a high-dimensional latent space. S103. Based on directed graphs, decoupled representation is performed in the horizontal spatial dimension to construct a wind turbine cluster topology graph. An improved graph attention network is introduced and combined with a physical perception masking mechanism to dynamically reconstruct the wind turbine cluster topology graph. S104. Expand the receptive field by using dilated causal convolution of temporal convolutional network, extract local high-frequency fluctuation features of wind farm and filter out random noise, input local high-frequency fluctuation features into Informer model, use the probabilistic sparse attention mechanism of Informer model to reduce computational complexity and capture the global time dependence of long sequence prediction. S105. Based on global time dependence and decoupled representation of vertical and horizontal spatial dimensions, construct the ST-GAT-Informer model, train the ST-GAT-Informer model, and use the ST-GAT-Informer model to predict wind power.

2. The method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity using physical sensing, as described in claim 1, is characterized in that... The wind farm is abstracted as a directed graph, represented as follows: in, It is a directed graph. for A set of nodes consisting of typhoon turbines. This is the set of edges connecting the nodes.

3. The method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity using physical sensing, as described in claim 2, is characterized in that... The method of mapping multi-height-layer weather forecast data to a high-dimensional latent space using dense feature fusion technology specifically includes the following steps: S21. Decoupling representation is performed in the vertical spatial dimension by introducing highly relevant physical prior vectors, as follows: in, For physical prior vectors, For the logarithmic transformation term of height, Standardized height item, This is a transpose operation; S22. In the vertical spatial dimension, each node contains weather forecast data at different altitude levels. Any node in... The original input features at time step 1 are defined as tensors containing vertical profile information, as follows: in, It is a tensor containing vertical profile information. The number of vertical height floors, For each height layer, the physical feature dimension, For the real number field; S23. Based on the physical prior vector and the original input features, the weather forecast data is mapped to a high-dimensional latent space using dense feature fusion technology, as shown below: in, This is the initial embedding vector for nodes containing aerodynamic information. For activation function, and These are the learnable weight matrix and the bias term, respectively. This is a vector concatenation operation.

4. The method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity using physical sensing, as described in claim 3, is characterized in that... Step S103 specifically includes the following steps: S31. Based on any node and its neighboring nodes, obtain the query, key, and value vectors through a linear transformation matrix mapping, and calculate the initial attention coefficients, as shown below: in, The initial attention coefficient, For attention weight vectors, It is a linear transformation matrix. For nodes The input feature representation, For nodes Feature input representation; S32. Using a physical sensing masking mechanism, the azimuth angle calculated from the absolute coordinates of the wind turbine is matched with the real-time average wind direction, as shown below: in, This is a physical sensing masking mechanism. This represents the real-time average wind direction. It is the azimuth angle; S33. Based on the initial attention coefficient and the physical perception masking mechanism, the final normalized spatial contribution weight is calculated. The normalized spatial contribution weight is used to dynamically measure the intensity of the wake influence inside the wind farm, as shown below: in, Contribute weights to the normalized space. For nodes With nodes Initial attention coefficients between For nodes With nodes The physical sensing mask values ​​between them For nodes The set of domain nodes, For nodes The index of a certain neighboring node; S34. Based on the normalized spatial contribution weight, calculate the node update characteristics, and dynamically reconstruct the wind turbine cluster topology based on the node update characteristics, as shown below: in, Update the features for the nodes.

5. The method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity using physical sensing, as described in claim 1, is characterized in that... The method of using dilated causal convolution of a temporal convolutional network to expand the receptive field, extract local high-frequency fluctuation features of the wind farm, and filter out random noise is as follows: in, For convolution kernel, It is the expansion factor. The kernel size is [size]. For time steps, For time step The output after dilated causal convolution, The index corresponding to the convolution kernel. For input time series features, The input features are the historical moments that participate in the convolution operation after dilation sampling.

6. The method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity using physical sensing, as described in claim 5, is characterized in that... The reduction of computational complexity using the probabilistic sparse attention mechanism of the Informer model is expressed as follows: in, This is the sparse query matrix after KL divergence filtering. For querying the matrix, The key matrix, For value matrices, The feature dimension of the key vector. This is a transpose operation.

7. The method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity using physical sensing, as described in claim 1, is characterized in that... The Informer model is also used to address the error accumulation problem in long sequence prediction by introducing a generative decoder, as shown below: in, To decode the input vector, a concatenation mechanism is used to construct it; The start token is a known historical segment extracted from the end of the input sequence to provide decoding context; Placeholder for the sequence to be predicted.

8. The method for predicting wind power in complex three-dimensional spatiotemporal heterogeneity using physical sensing, as described in claim 1, is characterized in that... The method of predicting wind power using the ST-GAT-Informer model specifically includes the following steps: S51. Obtain the length of the historical observation sequence and use the ST-GAT-Informer model to learn the nonlinear mapping function; S52. Calculate the historical spatiotemporal state based on the length of the historical observation sequence. , means as follows: in, The length of the historical observation sequence. The historical spatiotemporal state at time t; S53. Based on the nonlinear mapping function and historical spatiotemporal state, predict the wind power output for the next Q time steps, as shown below: in, It is a nonlinear mapping function. Encoding the periodic time features of the corresponding time period. This is the wind power prediction result for the ST-GAT-Informer model for the next Q time steps.

Images