Geographic constraint and space-time decoupling mechanism fused wind speed correction method applied to wind power prediction

Abstract

The invention provides a geographic constraint and space-time decoupling mechanism fused wind speed correction method applied to wind power prediction, and relates to the technical field of wind power prediction. The method comprises the following steps: clustering wind power cluster stations by adopting a flow field space clustering method fused with geographic constraints; based on a clustering result, adopting a conditional random field model with a space-time decoupling mechanism to carry out collaborative correction on the numerical weather forecast wind speed; and on the basis of a clustering result and the corrected wind speed, constructing a heterogeneous graph containing various types of nodes and edges, and performing wind power prediction by adopting a multi-feature fusion heterogeneous graph neural network model. According to the method, the description precision of regional wind speed evolution characteristics can be remarkably improved, multi-station wind speed collaborative correction is realized, and high robustness is kept in complex weather scenes such as cold tide and monsoon switching, so that the day-ahead wind power prediction precision is effectively improved. The method is suitable for dispatching operation and consumption analysis of a high-proportion new energy power grid, and has high applicability and engineering practicability.

Description

Technical Field

[0001] This application relates to the field of wind power prediction technology, and in particular to a wind speed correction method that integrates geographical constraints and spatiotemporal decoupling mechanisms for wind power prediction. Background Technology

[0002] With the depletion of fossil fuels and the increasing prominence of environmental issues, the proportion of new energy sources, represented by wind power, in the energy structure is increasing. As of the end of July 2025, China's cumulative installed power generation capacity reached 3.67 billion kilowatts, of which wind power capacity was 570 million kilowatts, representing a year-on-year increase of 22.1%. However, the increase in new energy installed capacity has had a significant impact on the stable operation of the power system. New energy output is significantly affected by weather, exhibiting not only significant randomness and volatility within a single day but also intermittent periods with different cycles depending on the season, thus altering the traditional power system balance mechanism. Wind power forecasting is a key technology for mitigating the adverse effects of the randomness of new energy output, playing a crucial role in maintaining the safe and stable operation of the power grid and the balance between power generation and load. Simultaneously, accurate day-ahead wind power forecasting is a vital foundation for supporting various operations such as grid dispatching, maintenance planning, and new energy consumption. Therefore, improving the accuracy of power forecasting for large-scale wind power clusters has become an urgent problem to be solved in the context of building a new type of power system.

[0003] With the increasing proportion of wind power in the power system in recent years, the demand for wind power forecasting on the dispatch side has gradually expanded from traditional site forecasting to regional-level wind power cluster forecasting. Therefore, research on power forecasting for large-scale wind power clusters has gradually become an important direction for both academia and engineering practice. Large-scale wind power cluster power forecasting faces the challenge of balancing accuracy and efficiency. Individual site forecasting methods are accurate but inefficient, while overall forecasting methods are efficient but inaccurate. Large-scale wind power clusters are affected by meteorological changes, wind farm characteristics, and spatial differences in geographical distribution, which poses new challenges to the spatiotemporal correlation mining capabilities of wind power forecasting models. Existing forecasting models have strong modeling capabilities in handling time series features, but they struggle to effectively capture the complex dynamic relationships between large-scale wind farms. Summary of the Invention

[0004] This application provides a wind speed correction method that integrates geographical constraints and spatiotemporal decoupling mechanisms for wind power prediction. It aims to construct a method for clustering wind power clusters and correcting the accuracy of numerical weather forecast wind speed characteristics in the model input, as well as a heterogeneous graph model suitable for large-scale wind power clusters. This fills a research gap in the field, thereby enhancing the accuracy of numerical weather forecast wind speed and improving the accuracy of large-scale wind power prediction.

[0005] This application provides a wind speed correction method that integrates geographical constraints and spatiotemporal decoupling mechanisms for wind power prediction, including: A flow field spatial clustering method incorporating geographical constraints is used to cluster wind power clusters. Based on the clustering results, a conditional random field model with a spatiotemporal decoupling mechanism is used to collaboratively correct the wind speed in numerical weather forecasts. Based on the clustering results and the corrected wind speed, a heterogeneous graph containing multiple types of nodes and edges is constructed, and a multi-feature fusion heterogeneous graph neural network model is used for wind power prediction.

[0006] Furthermore, the methods for clustering wind power clusters using flow field spatial clustering methods that incorporate geographical constraints include: Based on the gridded wind speed and wind direction data, the original wind speed data is decomposed into two wind speed components in the horizontal and vertical directions, which are represented as a two-dimensional wind speed field. The two-dimensional wind speed field is transformed into a continuous streamline representation using a bidirectional Euler algorithm. Slope and aspect are introduced as geographical constraints to modulate streamlines; Extract features of the modulated streamline, including spatial location features, local neighborhood features, and dominant direction features; Using the features of the modulated streamlines as input, the K-Means algorithm is used to cluster the streamlines, and the streamlines are mapped to the actual geographical area based on the clustering results to complete the clustering identification of wind farms.

[0007] Furthermore, the calculation formula for converting the two-dimensional wind speed field into a continuous streamline representation using the bidirectional Euler algorithm is expressed as follows: (1) (2) in It calculates the initial point. For the integration step size, It is the horizontal wind speed component at the initial point. It is the vertical wind speed component at the initial point. and The x and y coordinates of the second initial calculation point are obtained from the first calculation.

[0008] Methods of modulating streamlines by introducing slope and aspect as geographical constraints include: The slope and aspect are defined as follows: (3) (4) in Two-dimensional coordinates Digital elevation data at the location, For slope, Slope direction; When the modulation trigger condition is met and At that time, among them yes The wind direction at that location and These are the slope threshold and the slope aspect angle threshold. The wind speed vector is corrected using a terrain correction vector, which is expressed as: (5) in Use the Singmoid activation function; The terrain correction vector is normalized and then added to the wind speed to obtain the modulated wind speed vector; The spatial location features are obtained by calculating the geographical average value of each point on the streamline, and the local neighborhood features and dominant direction features are calculated using the following formulas: (6) (7) in It is the radius The number of streamlines within the range, It is a local neighborhood feature. It is the dominant directional characteristic. It is the first The direction of the streamline, N It is the number of points; Using the features of the modulated streamlines as input, the K-Means algorithm is used to cluster the streamlines, and the streamlines are mapped to actual geographical areas based on the clustering results. The methods for clustering and identifying wind farms include: Using spatial location features, local neighborhood features, and dominant direction features as input, the K-Means algorithm randomly initializes cluster centers and then determines the cluster centers based on the distance of each sample in the feature space to the cluster centers. Perform clustering: (8) in For feature data, For the first Cluster centers, For the dimensions of data objects, , For feature data and cluster center The Attribute values ​​in each dimension; The samples with the highest similarity to the cluster centers are assigned to their corresponding clusters, and the cluster centers are then updated using the mean features of each cluster, thereby minimizing the sum of squared errors in the dataset. SSE The calculation formula is as follows: (9) in, This represents the number of clusters.

[0009] Furthermore, based on the clustering results, a conditional random field model with a spatiotemporal decoupling mechanism is used to collaboratively correct the numerical weather forecast wind speed, including: Based on the clustering results, temporal adjacency matrices and spatial adjacency matrices are constructed for wind farms within the same cluster. Based on the aforementioned temporal adjacency matrix and spatial adjacency matrix, a temporal potential function and a spatial potential function are constructed. Combine the nodal potential function and the edge potential function to construct the total energy function, and then transform the total energy function into the form of a Laplace matrix; The optimal solution is obtained by minimizing the total energy function, which serves as the corrected wind speed.

[0010] Furthermore, the time potential function and the space potential function are expressed as follows: (10) (11) in This is the time adjacency matrix. It is a spatial adjacency matrix. Let be the time potential function. Let be the space potential function. For the first One wind farm, For the first One wind farm, For the first A wind farm in The eigenvalues ​​at time t, For the first A wind farm in The eigenvalues ​​at time t, For the first A wind farm in The characteristic value at time; The temporal and spatial adjacency matrices are constructed using the following formulas: (12) (13) in For time Gaussian kernel, For a spatial Gaussian kernel, It is the first A cluster, It is an exponential function. This is the time adjacency matrix. It is a spatial adjacency matrix. For the first A wind farm Wind speed at all times For the first A wind farm Wind speed at all times For the first A wind farm Wind speed at all times; After the potential function is constructed, the nodal potential functions and edge potential functions are further combined to construct the overall total energy function, which uniformly represents the data fitting and spatiotemporal smoothing constraints in the wind speed correction process. (14) The total energy function is expressed as: (15) in , , These are the weighting coefficients for each part of the energy function. Let be the node energy function. It is a time-energy function. For space energy function, The total energy function; The total energy function is expressed in Laplace matrix form as follows: (16) Among them, matrix The matrix represents the diagonal weight matrix corresponding to the nodal potential function. The Laplacian matrix constructed from the edge weights is expressed as: (17) in, This represents the edge weight value.

[0011] Furthermore, based on the clustering results and the corrected wind speed, a heterogeneous graph containing multiple types of nodes and edges is constructed as follows: Each wind farm in the wind power cluster is used as a node in the graph, with wind speed as the node feature, and three different types of edge relationships are introduced to construct a heterogeneous graph: the first type of edge relationship is geographical proximity, the second type of edge relationship is wind speed correlation, and the third type of edge relationship is meteorological similarity.

[0012] Furthermore, the geographical proximity edge is calculated as follows: the Havesing distance is calculated based on the latitude and longitude coordinates of the wind farm station. This is then transformed into edge weights of a geographic proximity matrix, and the calculation process is expressed as follows: (19) (20) (twenty one) in Indicates the latitude of the station. Indicates the longitude of the station. Due to the latitude difference of the station, Due to the longitude difference of the station, This is the geographic distance threshold. For Havesing expression, For the first The longitude of the wind farm For the first The longitude of the wind farm; The method for determining the wind speed-related edge is as follows: Calculate all station-to-station Correlation of historical wind speed sequences at different times : (twenty two) in, For the first A wind farm Wind speed sequence at any time For the first The mean of the wind speed sequence of each wind farm. For the first A wind farm Wind speed sequence at any time For the first Mean of wind speed sequence for each wind farm; Based on all stations Correlation of historical wind speed sequences at different times The wind speed correlation matrix is ​​constructed using the following formula. : (twenty three) The method for determining the meteorological similarity edges is as follows: Calculate the cosine similarity between feature vectors of numerical weather prediction data from wind farms and construct a meteorological similarity matrix. : (twenty four) (25) in, Cosine similarity among feature vectors of numerical weather prediction data from wind farms; For the first A wind farm Weather characteristics at any time For the first A wind farm Weather characteristics at any given time.

[0013] Furthermore, the multi-feature fusion heterogeneous graph neural network model includes an encoder and a decoder. Using this model for wind power prediction includes: In response to the heterogeneous input graph, during the encoder stage, an attention mechanism based on meta-relations is used to calculate the attention weights of different types of connections between the target node and its neighboring nodes, aggregate the messages of neighboring nodes, and update the node features through residual connections. In the decoder stage, based on the meta-relationships between nodes at the predicted time, mutual attention is calculated and node features are reconstructed. The encoder output is then fused again through residual connections to obtain the final node representation. The final node representation is input into the subsequent prediction network to output the wind power prediction result.

[0014] Furthermore, in the encoder stage, a meta-relation-based attention mechanism is used to calculate the attention weights of different types of connections between the target node and its neighboring nodes, aggregate the messages of neighboring nodes, and update node features through residual connections, including: For each edge of the heterogeneous graph of Attention weights for each attention head The following formula is used to calculate: (26) (27) (28) (29) in , , , , These are learnable parameters; , , These are the three degrees of correlation of features calculated through the attention mechanism. It is time. It is the initial node. It is the edge structure of a heterogeneous graph, and Softmax is the activation function. It is a priori tensor. These are characteristics of the initial node. It is a prediction of the station nodes. Characteristics of time, It is a query vector. It is the key vector. Indicates the number of attention heads. It is the vector dimension of each head. It is attention weight; Edge structure of a heterogeneous graph with a pair of nodes The multiple messages of its source node The calculation method is as follows: (30) (31) (32) in , , These are learnable parameters. It is a node with respect to the edge structure Value vector at the location Attention weights; Aggregation features of target nodes Represented as: (33) The updated node features of the encoder output are obtained through residual connections: ,in These are the node features output by the encoder. These are the original node features. It is a node aggregation feature.

[0015] Furthermore, the method also includes error analysis of the wind power prediction results using normalized root mean square error and normalized average absolute error.

[0016] This application provides a wind speed correction method that integrates geographical constraints and spatiotemporal decoupling mechanisms for wind power prediction, which has at least the following beneficial effects: The results, validated through numerical examples, demonstrate that the introduction of geographic information enables the proposed clustering method to more accurately capture regional wind speed evolution characteristics and achieve reasonable clustering of stations with similar wind speed evolution. The spatiotemporal decoupling mechanism enhances the modeling ability of relationships between stations. The heterogeneous graph model, leveraging an attention mechanism, efficiently integrates different types of nodes and edge information, significantly improving feature utilization efficiency. Overall, the proposed method outperforms existing models in prediction accuracy, with an average improvement of 3.91% in RMSE and 3.36% in MAE. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] Figure 1 A flowchart of a wind speed correction method for wind power prediction that integrates geographical constraints and spatiotemporal decoupling mechanisms, provided for embodiments of this application; Figure 2 A flowchart illustrating an overall wind speed correction method for wind power prediction that integrates geographical constraints and spatiotemporal decoupling mechanisms, as provided in this application embodiment; Figure 3 A flowchart illustrating the spatial clustering method for flow fields that incorporates geographical constraints, provided in this application embodiment; Figure 4 A schematic diagram illustrating the spatiotemporal decoupling mechanism provided in the embodiments of this application; Figure 5 Comparison curves of the correction results of various methods provided in the embodiments of this application; Figure 6 The structure diagram of the heterogeneous graph neural network model for day-ahead power prediction of large-scale wind power clusters provided in the embodiments of this application is shown. Figure 7 The wind speed input prediction results curves of each model are provided for the embodiments of this application.

[0019] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0021] The collection, storage, use, processing, transmission, provision, and disclosure of relevant data and information in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0022] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0023] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0024] Example 1: This application provides a wind speed correction method that integrates geographical constraints and spatiotemporal decoupling mechanisms for wind power prediction, such as... Figure 1 The diagram shows a flowchart of a wind speed correction method for wind power prediction that integrates geographical constraints and spatiotemporal decoupling mechanisms, provided by an embodiment of this application. The continuous learning model hot update method for industrial product defect detection includes the following steps S10~S30.

[0025] S10: Clustering of wind power clusters is performed using a flow field spatial clustering method that incorporates geographical constraints.

[0026] In this embodiment, firstly, based on the initial wind field, the original wind speed data is decomposed into two wind speed components, one horizontal and one vertical, using gridded wind speed and direction data, representing a two-dimensional wind speed field. The two-dimensional discrete wind field is then transformed into a continuous streamline representation using a two-way Euler algorithm. The two-way Euler streamline calculation algorithm is as follows: (1) (2) in It calculates the initial point. For the integration step size, It is the horizontal wind speed component at the initial point. It is the vertical wind speed component at the initial point. and The x and y coordinates of the second initial calculation point are obtained from the first calculation.

[0027] Slope characterizes the steepness of the earth's surface, reflecting the likelihood that wind is blocked or accelerated by the terrain in a local area; aspect describes the direction in which the slope faces. Slope aspect The mathematical definition of is as follows: (3) (4) in Two-dimensional coordinates Digital elevation data at the location.

[0028] Slope threshold This threshold is used to determine the modulating effect of topographic relief on streamlines. When the local slope exceeds this threshold, the topography is considered to have a significant impact on airflow; the slope aspect angle threshold is also included. This measure assesses the consistency between airflow direction and slope orientation; when the angle is less than a threshold, it indicates that streamlines are guided by the slope aspect. Modulation trigger bar. and ,in yes Wind direction at the location. The terrain influence factor is set as follows: Then the terrain correction vector as follows: (5) in The Singmoid activation function is used. The modulated vector is normalized and then added to the wind speed to obtain the modulated wind speed vector.

[0029] Spatial location characteristics represent the center position of each streamline. This is obtained by calculating the geographical average value of each point on the streamline, and is used to identify the spatial distribution pattern of the wind field; local neighborhood characteristics. By statistically analyzing the number of other streamlines within a given radius threshold, the clustering of streamline orientations within a local area can be reflected.

[0030] (6) in It is the radius Number of streamlines within the range. Dominant directional characteristics. Reflecting the dominant propagation path of the wind field, This represents the average direction of the i-th streamline.

[0031] (7) in, It is the first The direction of the streamline.

[0032] Using spatial location, local neighborhood, and dominant flow direction features as input, the K-Means algorithm randomly initializes cluster centers and then determines the cluster centers based on the distance of each sample in the feature space to the cluster centers. Perform clustering.

[0033] in For feature data, For the first Cluster centers, For the dimensions of data objects, , For feature data and cluster center The The dataset uses attribute values ​​in each dimension. Samples with the highest similarity to the cluster centers are assigned to their corresponding clusters. The cluster centers are then updated using the mean of the cluster features, thus minimizing the sum of squared errors in the dataset. The formula for the sum of squared errors is as follows: (9) by The value of the clustering effect is used as a measure. When the value tends to stabilize or converge, the iteration is terminated and the final streamline clustering result is obtained.

[0034] Based on the clustering results, streamlines are divided into several clusters, each cluster effectively revealing the evolution characteristics of the wind field and the flow path of wind speed. By mapping the streamline clusters to actual geographical areas, the spatial division of the wind field's influence area is achieved. Furthermore, based on the correspondence between the latitude and longitude coordinates of wind farm stations and the divided areas, the clustering identification of wind farm stations is completed.

[0035] S20: Based on the clustering results, a conditional random field model with a spatiotemporal decoupling mechanism is used to collaboratively correct the wind speed in numerical weather forecasts.

[0036] In step S20, a numerical weather prediction wind speed correction method is proposed based on a conditional random field model with a spatiotemporal decoupling modeling strategy. This model introduces a spatiotemporal decoupling modeling strategy on the basis of the traditional continuous conditional random field model, constructs the spatial adjacency relationship and temporal evolution relationship between stations, thereby more effectively representing the spatiotemporal characteristics of wind speed and realizing the collaborative correction of wind speed at multiple stations.

[0037] Wind farms are clustered to determine farm clusters. Within each cluster, temporal and spatial adjacency matrices are constructed based on spatiotemporal correlation. The structural information of nodes and edges is jointly optimized by minimizing the energy function to obtain the corrected wind speed sequence at each time step. The temporal and spatial potential functions are as follows: (10) (11) in This is the time adjacency matrix. It is a spatial adjacency matrix. Let be the time potential function. Let be the space potential function. For the first One wind farm, For the first One wind farm, For the first A wind farm in The eigenvalues ​​at time t, For the first A wind farm in The eigenvalues ​​at time t, For the first A wind farm in The eigenvalues ​​at each time point, and the adjacency matrix, are used to characterize the connectivity between wind farms in both time and space dimensions. The calculation method is as follows: (12) (13) in For time Gaussian kernel, For a spatial Gaussian kernel, It is the first A cluster, It is an exponential function. This is the time adjacency matrix. It is a spatial adjacency matrix. For the first A wind farm Wind speed at all times For the first A wind farm Wind speed at all times For the first A wind farm Wind speed at any given time. Adjacency between two wind farms is calculated only if they belong to the same cluster; if they are in different clusters, their corresponding positions in the spatial or temporal adjacency matrix are assigned a value of 0, indicating that there is no direct spatiotemporal connection between them.

[0038] After the potential function is constructed, the nodal potential functions and edge potential functions are further combined to construct the overall total energy function, which uniformly represents the data fitting and spatiotemporal smoothing constraints in the wind speed correction process. (14) Rewrite the energy function: (15) in , , These are the weighting coefficients for each part of the energy function, determined adaptively using the steepest descent method (gradient descent) during the solution process. Subsequently, the constructed energy function is transformed into a Laplace matrix form. (16) Among them, matrix This is the diagonal weight matrix corresponding to the nodal potential function. .matrix The Laplace matrix constructed from the edge weights is... (17) Because the energy function is related to the variable Since it is a quadratic and convex function, the optimal solution can be obtained by minimizing this energy function. This yields the corrected wind speed. Calculation as follows

[0039] For example, step S20 can be implemented through the following steps: ① Determine the cluster to which each wind farm belongs based on the clustering of the wind power cluster.

[0040] ②Evaluate the spatiotemporal and spatial relationships of stations within the same cluster using temporal Gaussian kernel and spatial Gaussian kernel methods.

[0041] ③ Based on the calculation results of step ②, construct the temporal adjacency matrix and the spatial adjacency matrix.

[0042] ④ Construct the time potential function and the space potential function.

[0043] ⑤ Construct the overall total energy function and transform it into the form of a Laplace matrix.

[0044] ⑥ Determine the relevant coefficients in the energy function using the steepest descent method.

[0045] ⑦ Solve for the minimum function value based on different inputs to obtain the corrected wind speed.

[0046] S30: Based on clustering results and corrected wind speed, a heterogeneous graph containing multiple types of nodes and edges is constructed, and a multi-feature fusion heterogeneous graph neural network model is used for wind power prediction.

[0047] Heterogeneous graph neural networks can process graph structure data containing multiple types of nodes and edges, and have significant advantages in modeling complex heterogeneous relationships. This embodiment proposes a multi-feature fusion heterogeneous graph neural network model.

[0048] In wind power prediction, each wind farm within a wind power cluster is treated as a node in a graph, with wind speed as the node feature. Three different types of edge relationships are introduced to construct a heterogeneous graph. The first type of edge relationship is the geographical proximity edge, calculated based on the Havelsing distance between wind farms using their latitude and longitude coordinates. The calculation is as follows:

[0049] in Indicates the latitude of the station. Indicates the longitude of the station. Due to the latitude difference of the station, Due to the longitude difference of the station, This is the geographic distance threshold. For Havesing expression, For the first The longitude of the wind farm For the first The longitude of each wind farm. Convert the Havelsing distance into a geographic proximity matrix. The edge weights,

[0050] in The geographic distance threshold is set to 50km in this paper.

[0051] The second type of edge relationship is the wind speed correlation edge, used to measure the historical wind speed relationship between different stations. Time, station Historical wind speed sequences of 96 time steps were used to calculate the wind speeds between all stations. The correlation of historical wind speed sequences at different times was calculated using the Pearson correlation coefficient method.

[0052] in For the first A wind farm Wind speed sequence at any time For the first The mean of the wind speed sequence of each wind farm. For the first A wind farm Wind speed sequence at any time For the first Mean of wind speed sequence for each wind farm.

[0053] Constructing a wind speed correlation matrix The edge weights are as follows:

[0054] The third type of edge relationship is the meteorological similarity edge, where stations... The corresponding numerical weather forecast data feature vector is By calculating the cosine similarity between meteorological data (excluding wind speed) from wind farms. ,

[0055] The cosine values ​​are then converted into a meteorological similarity matrix. Boundary weights:

[0056] Using the heterogeneous graph constructed above as input, the model calculates three degrees of association between various features based on an attention mechanism, focusing on the interaction relationships between wind farm nodes and their neighboring nodes. , , Predicting the station nodes Characteristics of a moment Mapping to query vector , the initial node Features Mapping to key vector ,in Indicates the number of attention heads. This is the vector dimension of each head. Each edge of Attention weights for each attention head The following formula can be used to calculate it:

[0057] in It is attention weight. , , , , These are learnable parameters; , , These are the three degrees of correlation of features calculated through the attention mechanism. It is time. It is the initial node. It represents the edge structure of a heterogeneous graph. Softmax is the activation function. Prior tensor. Since different relations contribute differently to the target node, a priori tensor is introduced. The adaptive scaling factor, serving as the attention mechanism, represents the global importance of each type of meta-relation. After calculating the attention value, the encoder adjusts the scaling factor according to the different meta-relations. Extract the multi-head message from the source node. The encoder first extracts the features of the source node. Mapped to value vector In order to alleviate the distribution differences in characteristics between different stations. For the edge structure of a heterogeneous graph with a pair of nodes. The multiple messages of its source node The calculation method is as follows:

[0058] in , , , is a learnable parameter. It is a node with respect to the edge structure place Attention weights. Similar to the calculation of attention values, the encoder assigns an independent weight matrix to each edge type. By multiplying and aggregating the meta-relation attention weights with node messages, the encoder extracts feature-differentiated relation information, including the aggregated features of the target node. Represented as:

[0059] To further preserve the target node's own feature information, the encoder uses residual connections. ,in These are the node features output by the encoder. These are the original node features. It is a node aggregation feature.

[0060] The decoder predicts the field nodes. The meta-relationships between time-series data and their neighboring nodes are used to calculate their mutual attention. Each edge is obtained. of Head attention value After completing the attention calculation, the decoder extracts source node features based on different meta-relations and reconstructs the multi-head message. The decoder aggregates information from neighboring nodes to reconstruct node features. After reconstructing node features based on meta-relations, the decoder combines the reconstructed node features with the encoder output through residual connections to obtain the final representation. . This involves transforming the heterogeneous graph information constructed above into representation information consistent with the node feature dimensions. Wind power is used as the feature input for a multi-layer graph convolutional neural network model.

[0061] After step S30, step S40, simulation calculation, is also included.

[0062] Simulation inputs: Analyze the measured data of the wind farm to determine the total installed capacity of the wind farm; the rated capacity of each wind farm; numerical weather forecasts for 100m wind speed, temperature, and dew point temperature; data sampling interval is 15 minutes; according to steps S10 to S30, obtain the real-time prediction results of the wind power of the entire wind farm.

[0063] After step S40, step S50, error analysis, is also included.

[0064] set up for iThe actual average power over the period for i Predicted average power for the time period The total time period for daily assessment Let be the operating capacity of the photovoltaic field. Then, the normalized root mean square error is defined as equation (34):

[0065] The normalized mean absolute error is defined as Equation (35):

[0066] The predicted power calculated by the model is compared with the measured power using the error evaluation criteria in step 5) to calculate the error, and the root mean square error and mean absolute error of the prediction are obtained.

[0067] Example 2: This application provides a wind speed correction method that integrates geographical constraints and spatiotemporal decoupling mechanisms for wind power prediction, such as... Figure 2 As shown, it includes the following steps: 1) A spatial clustering method for flow fields that integrates geographical constraints is proposed to achieve clustering of wind farms with similar spatiotemporal evolution characteristics; 2) Construct a conditional random field model with a spatiotemporal decoupling mechanism to achieve coordinated correction of wind speeds at multiple stations; 3) Based on the above clustering results and the corrected wind speed, and combined with numerical weather prediction data, various heterogeneous graphs are constructed. The proposed multi-feature fusion heterogeneous graph neural network model suitable for day-ahead power prediction of large-scale wind power clusters is used to predict wind power.

[0068] It should be noted that steps 1) to 3) above specifically correspond to steps S10-S30 of Embodiment 1. This embodiment aims to further illustrate the feasibility and advancement of the method of this application by combining the above steps with specific experimental data.

[0069] The specific implementation method is as follows: A wind power cluster was selected as a case study. This cluster comprises 70 wind farms with a total installed capacity of 9094MW. The data used included wind speed observations and meteorological elements such as wind speed, humidity, and temperature from numerical weather prediction (NWP) data, with a time resolution of 15 minutes, covering the period from March 1, 2020 to December 31, 2020, totaling 29376 time points. In the wind speed correction stage, the first 14688 time points were used as the training set, and the remaining 14688 time points were used as the test set to build and validate the wind speed correction model. In the wind power prediction stage, the results from the aforementioned wind speed correction test set were used as the wind speed input to the prediction model, and the samples were divided according to an 8:2 ratio of training set to test set for model training and performance evaluation.

[0070] Step 1: Clustering of wind farm sites. For example... Figure 3 As shown, a two-dimensional wind field is first constructed based on gridded wind speed data. Then, the wind field streamlines within the region are calculated using the Eulerian algorithm, and the calculated streamlines are modulated by introducing geographical constraints. Features are constructed from the modulated streamlines, and these features are used as input to the K-means clustering algorithm to achieve clustering of modulated streamlines with similar trends. Finally, based on the streamline clustering results and the geographical range of the streamline distribution, a mapping relationship is established from the streamline clustering results to the wind power cluster clustering, thus completing the wind power cluster clustering.

[0071] Step 2: Numerical Weather Forecast Wind Speed ​​Correction. Based on the above wind power cluster clustering results, determine the cluster to which each wind farm belongs. Evaluate the temporal and spatial relationships between wind farms within the same cluster using wind speed data from each farm and temporal and spatial Gaussian kernel functions, and generate temporal and spatial adjacency matrices, such as... Figure 4 As shown, a potential function is constructed using time and spatial adjacency matrices, and a total energy function is constructed. The correlation coefficient is determined using the steepest descent method. With different wind speeds as input, the optimal solution, i.e., the corrected wind speed, is obtained.

[0072] The comparison methods are set as follows: Method 1: Convolutional Long Short-Term Memory Network model; Method 2: Transformer; Method 3: Conditional Random Field (CRF); Method 4: CRF and a proposed spatial clustering method for flow fields incorporating geographical constraints; Method 5: CRF model with spatiotemporal decoupling mechanism and mean clustering method based on geographical distance; Method 6: CRF model with spatiotemporal decoupling mechanism and hierarchical clustering method based on modulated streamlines; Method 7: CRF model with spatiotemporal decoupling mechanism and fuzzy mean clustering method based on wind speed; Method 8: Only CRF model with spatiotemporal decoupling mechanism; and the modified method of this paper: CRF model with spatiotemporal decoupling mechanism and the proposed spatial clustering method for flow fields incorporating geographical constraints.

[0073] The RMSE and MAE indices of each correction method are shown in Table 1. The proposed method improves the RMSE by 0.2068 m / s and the MAE by 0.1195 m / s compared to the comparison model. Table 1 shows that the accuracy of conventional machine learning correction methods is lower than that of the original wind speed data. For the training phase of large-scale wind speed correction tasks, Method 1 and Method 2 typically aim for optimal overall accuracy. However, the wind speed differences between stations are significant, so the overall optimal training strategy often weakens the characterization of individual station characteristics. The correction result curves are shown in the figure below. Figure 5 As shown.

[0074] Table 1. Wind speed correction comparison model setup and correction result evaluation

[0075] Step 3: Wind Power Prediction. Using the clustering results from Step 1 and the corrected wind speed from Step 2, several heterogeneous graphs are established. These graphs mainly include Pearson correlation coefficient side graphs, meteorological cosine similarity side graphs, and geographical distance adjacency side graphs. Using these three heterogeneous graphs as data input and the corrected wind speed as a node feature, the proposed multi-feature fusion heterogeneous graph neural network model suitable for day-ahead power prediction of large-scale wind power clusters is used for prediction. Finally, the wind power cluster power prediction result is obtained, as shown in the specific structure below. Figure 6 As shown.

[0076] In this embodiment, the wind speed data input for the prediction model is selected from the numerical weather prediction correction results, specifically the wind speed corrected by three methods: Method 4 (wind speed), Method 5 (wind speed), and the wind speed corrected by the method proposed in this paper. Simultaneously, the original numerical weather prediction wind speed data is used as the baseline input to establish a comparative prediction experiment. For the comparison models, bidirectional long short-term memory networks, iTransformer, spatiotemporal graph convolutional networks, hypergraph neural networks, and dynamic evolution graph convolutional networks are selected as reference models, and compared and analyzed with the method proposed in this paper.

[0077] The evaluation indicators for the prediction results are shown in Table 2. The prediction results of each model for wind speed using the correction method proposed in this paper are as follows: Figure 7As shown, the prediction results of the bidirectional long short-term memory network, iTransformer, and the three graph neural network models are generally similar. The proposed method improves the average results of the comparison models by 3.02% and 2.59% in NRMSE and MAE, respectively. Compared with the hypergraph neural network and the dynamic evolution graph convolutional network, the proposed heterogeneous graph prediction model reduces the RMSE by 0.62% and 2.56%, and the MAE by 0.63% and 2.38%, respectively. The proposed method not only effectively reduces the overall prediction error amplitude and narrows the error distribution range, but also overcomes the limitations of traditional isomorphic graph models in terms of structural representation and information interaction, demonstrating the best performance in prediction accuracy.

[0078] Table 2 Comparison of Evaluation Indicators for Prediction Results of Different Models

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A wind speed correction method for wind power prediction that integrates geographical constraints and spatiotemporal decoupling mechanisms, characterized in that, include: A flow field spatial clustering method incorporating geographical constraints is used to cluster wind power clusters. Based on the clustering results, a conditional random field model with a spatiotemporal decoupling mechanism is used to collaboratively correct the wind speed in numerical weather forecasts. Based on the clustering results and the corrected wind speed, a heterogeneous graph containing multiple types of nodes and edges is constructed, and a multi-feature fusion heterogeneous graph neural network model is used for wind power prediction.

2. The wind speed correction method for wind power prediction based on the fusion of geographical constraints and spatiotemporal decoupling mechanisms, as described in claim 1, is characterized in that... Methods for clustering wind power clusters using flow field spatial clustering that incorporates geographical constraints include: Based on the gridded wind speed and wind direction data, the original wind speed data is decomposed into two wind speed components in the horizontal and vertical directions, which are represented as a two-dimensional wind speed field. The two-dimensional wind speed field is transformed into a continuous streamline representation using a bidirectional Euler algorithm. Slope and aspect are introduced as geographical constraints to modulate streamlines; Extract features of the modulated streamline, including spatial location features, local neighborhood features, and dominant direction features; Using the features of the modulated streamlines as input, the K-Means algorithm is used to cluster the streamlines, and the streamlines are mapped to the actual geographical area based on the clustering results to complete the clustering identification of wind farms.

3. The wind speed correction method for wind power prediction based on the fusion of geographical constraints and spatiotemporal decoupling mechanisms, as described in claim 2, is characterized in that... The calculation formula for converting the two-dimensional wind speed field into a continuous streamline representation using the bidirectional Euler algorithm is as follows: （1） （2） in It calculates the initial point. The integration step size is... It is the horizontal wind speed component at the initial point. It is the vertical wind speed component at the initial point. and The x and y coordinates of the second initial calculation point are obtained from the first calculation. Methods of modulating streamlines by introducing slope and aspect as geographical constraints include: The slope and aspect are defined as follows: （3） （4） in Two-dimensional coordinates Digital elevation data at the location, For slope, Slope direction; When the modulation trigger condition is met and At that time, among them yes The wind direction at that location and These are the slope threshold and the slope aspect angle threshold. The wind speed vector is corrected using a terrain correction vector, which is expressed as: （5） in Use the Singmoid activation function; The terrain correction vector is normalized and then added to the wind speed to obtain the modulated wind speed vector; The spatial location features are obtained by calculating the geographical average value of each point on the streamline, and the local neighborhood features and dominant direction features are calculated using the following formulas: （6） （7） in It is the radius The number of streamlines within the range, It is a local neighborhood feature. It is the dominant directional characteristic. It is the first The direction of the streamline, N It is the number of points; Using the features of the modulated streamlines as input, the K-Means algorithm is used to cluster the streamlines, and the streamlines are mapped to actual geographical areas based on the clustering results. The methods for clustering and identifying wind farms include: Using spatial location features, local neighborhood features, and dominant direction features as input, the K-Means algorithm randomly initializes cluster centers and then determines the cluster centers based on the distance of each sample in the feature space to the cluster centers. Perform clustering: （8） in For feature data, For the first Cluster centers, For the dimensions of data objects, , For feature data and cluster center The Attribute values ​​in each dimension; The samples with the highest similarity to the cluster centers are assigned to their corresponding clusters, and the cluster centers are then updated using the mean features of each cluster, thereby minimizing the sum of squared errors in the dataset. SSE The calculation formula is as follows: （9） in, This represents the number of clusters.

4. The wind speed correction method for wind power prediction based on the fusion of geographical constraints and spatiotemporal decoupling mechanisms, as described in claim 1, is characterized in that... Based on the clustering results, a conditional random field model with a spatiotemporal decoupling mechanism is used to collaboratively correct the wind speed in numerical weather forecasts, including: Based on the clustering results, temporal adjacency matrices and spatial adjacency matrices are constructed for wind farms within the same cluster. Based on the aforementioned temporal adjacency matrix and spatial adjacency matrix, a temporal potential function and a spatial potential function are constructed. Combine the nodal potential function and the edge potential function to construct the total energy function, and then transform the total energy function into the form of a Laplace matrix; The optimal solution is obtained by minimizing the total energy function, which serves as the corrected wind speed.

5. The wind speed correction method for wind power prediction based on the fusion of geographical constraints and spatiotemporal decoupling mechanisms, as described in claim 4, is characterized in that... The time potential function and the space potential function are expressed as follows: （10） （11） in This is the time adjacency matrix. It is a spatial adjacency matrix. Let be the time potential function. Let be the space potential function. For the first One wind farm, For the first One wind farm, For the first A wind farm in The eigenvalues ​​at time t, For the first A wind farm in The eigenvalues ​​at time t, For the first A wind farm in The characteristic value at time; The temporal adjacency matrix and spatial adjacency matrix are constructed using the following formulas: （12） （13） in For time Gaussian kernel, For a spatial Gaussian kernel, It is the first A cluster, It is an exponential function. This is the time adjacency matrix. It is a spatial adjacency matrix. For the first A wind farm Wind speed at all times For the first A wind farm Wind speed at all times For the first A wind farm Wind speed at all times; After the potential function is constructed, the nodal potential functions and edge potential functions are further combined to construct the overall total energy function, which uniformly represents the data fitting and spatiotemporal smoothing constraints in the wind speed correction process. （14） The total energy function is expressed as: （15） in , , These are the weighting coefficients for each part of the energy function. Let be the node energy function. It is a time-energy function. For space energy function, The total energy function; The total energy function is expressed in Laplace matrix form as follows: （16） Among them, matrix The matrix represents the diagonal weight matrix corresponding to the nodal potential function. The Laplacian matrix constructed from the edge weights is expressed as: （17） in, This represents the edge weight value.

6. The wind speed correction method for wind power prediction based on the fusion of geographical constraints and spatiotemporal decoupling mechanisms, as described in claim 1, is characterized in that... Based on the clustering results and the corrected wind speed, a heterogeneous graph containing multiple types of nodes and edges is constructed as follows: Each wind farm in the wind power cluster is used as a node in the graph, with wind speed as the node feature, and three different types of edge relationships are introduced to construct a heterogeneous graph: the first type of edge relationship is geographical proximity, the second type of edge relationship is wind speed correlation, and the third type of edge relationship is meteorological similarity.

7. The wind speed correction method for wind power prediction based on the fusion of geographical constraints and spatiotemporal decoupling mechanisms, as described in claim 6, is characterized in that... The method for calculating the geographic proximity edge is as follows: the Havesing distance is calculated based on the latitude and longitude coordinates of the wind farm station. This is then transformed into edge weights of a geographic proximity matrix, and the calculation process is expressed as follows: （19） （20） （21） in Indicates the latitude of the station. Indicates the longitude of the station. Due to the latitude difference of the station, Due to the difference in longitude of the station, This is the geographic distance threshold. For Havesing expression, For the first The longitude of the wind farm For the first The longitude of the wind farm; The method for determining the wind speed-related edge is as follows: Calculate all station-to-station Correlation of historical wind speed sequences at different times : （22） in, For the first A wind farm Wind speed sequence at any time For the first The mean of the wind speed sequence of each wind farm. For the first A wind farm Wind speed sequence at any time For the first Mean of wind speed sequence for each wind farm; Based on all stations Correlation of historical wind speed sequences at different times The wind speed correlation matrix is ​​constructed using the following formula. : （23） The method for determining the meteorological similarity edges is as follows: Calculate the cosine similarity between feature vectors of numerical weather prediction data from wind farms and construct a meteorological similarity matrix. : （24） （25） in, Cosine similarity among feature vectors of numerical weather prediction data from wind farms; For the first A wind farm Weather characteristics at any time For the first A wind farm Weather characteristics at any given time.

8. The wind speed correction method for wind power prediction based on the fusion of geographical constraints and spatiotemporal decoupling mechanisms, as described in claim 1, is characterized in that... The multi-feature fusion heterogeneous graph neural network model includes an encoder and a decoder. Wind power prediction using this model includes: In response to the heterogeneous input graph, during the encoder stage, an attention mechanism based on meta-relations is used to calculate the attention weights of different types of connections between the target node and its neighboring nodes, aggregate the messages of neighboring nodes, and update the node features through residual connections. In the decoder stage, based on the meta-relationships between nodes at the predicted time, mutual attention is calculated and node features are reconstructed. The encoder output is then fused again through residual connections to obtain the final node representation. The final node representation is input into the subsequent prediction network to output the wind power prediction result.

9. The wind speed correction method for wind power prediction based on the fusion of geographical constraints and spatiotemporal decoupling mechanisms, as described in claim 8, is characterized in that... In the encoder stage, a meta-relation-based attention mechanism is used to calculate the attention weights of different types of connections between the target node and its neighboring nodes, aggregate the messages of neighboring nodes, and update node features through residual connections, including: For each edge of the heterogeneous graph of Attention weights for each attention head The following formula is used to calculate: （26） （27） （28） （29） in , , , , These are learnable parameters; , , These are the three degrees of correlation of features calculated through the attention mechanism. It is time. It is the initial node. It is the edge structure of a heterogeneous graph. It is an activation function. It is a priori tensor. These are characteristics of the initial node. It is a prediction of the station nodes. Characteristics of time, It is a query vector. It is the key vector. Indicates the number of attention heads. It is the vector dimension of each head. It is attention weight; Edge structure of a heterogeneous graph with a pair of nodes The multiple messages of its source node The calculation method is as follows: （30） （31） （32） in , , These are learnable parameters. It is a node with respect to the edge structure Value vector at the location Attention weights; Aggregation features of target nodes Represented as: （33） The updated node features of the encoder output are obtained through residual connections: ,in These are the node features output by the encoder. These are the original node features. It is a node aggregation feature.

10. The wind speed correction method for wind power prediction based on the fusion of geographical constraints and spatiotemporal decoupling mechanisms, as described in claim 1, is characterized in that... The method also includes error analysis of wind power prediction results using normalized root mean square error and normalized mean absolute error.

Images