A wind power prediction method across a wind farm

Abstract

This invention provides a method for predicting wind power across wind farms, applicable to the field of new energy technology. Based on the geographical coordinates, meteorological data, and historical power of the source wind farm cluster, this invention constructs a dynamic spatial topology and joint spatiotemporal features to train a prediction model. Utilizing a small amount of historical data from the target wind farms, the model is transferred and dynamically updated by combining real and predicted power data from continuous operation, achieving adaptive adjustment to the operating status of the target wind farms. This invention constructs a dynamic spatial topology by fusing static geographical and dynamic meteorological information, and further combines historical wind power data to generate joint spatiotemporal features, thereby accurately modeling the spatiotemporal coupling evolution of the wind farm cluster. Through transfer learning and dynamic updates, it achieves high-precision and high-stability cross-wind farm power prediction to meet the actual operational needs of the power system.

Description

Technical Field

[0001] This invention relates to the field of new energy technology, and in particular to a method for predicting wind power across wind farms. Background Technology

[0002] With the growing global demand for renewable energy, wind power, as an important clean energy source, is contributing to energy transition and climate change response. However, wind power output is greatly affected by weather conditions, exhibiting significant randomness and uncertainty. As the installed capacity and number of wind farms continue to increase, accurate wind power forecasting is crucial for ensuring the safe and stable operation of the power system.

[0003] However, traditional forecasts are mostly targeted at single wind farms, making it difficult to fully utilize the spatial correlation between multiple wind farms. Furthermore, due to differences in geographical location and operational characteristics, different wind farms exhibit heterogeneity and non-stationarity in spatial relationships and temporal characteristics, resulting in insufficient cross-scenario generalization ability of wind power prediction models.

[0004] Therefore, how to enhance the accuracy and stability of wind power prediction across wind farms to meet the actual power system operation requirements has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method for predicting wind power across wind farms to overcome or at least partially solve the above problems, the technical solution of which is as follows:

[0006] A method for predicting wind power across wind farms includes:

[0007] Obtain the geographic coordinates, meteorological observation data, and first historical wind power data of the source wind farm cluster;

[0008] Using the geographic coordinate information, the meteorological observation data, and the first historical wind power data, a dynamic spatial topology relationship of the source wind farm group is constructed;

[0009] Using the dynamic spatial topology and the first historical wind power data, the joint spatiotemporal characteristics of wind power of the source wind farm group are constructed;

[0010] The source wind farm prediction model is obtained by using the combined spatiotemporal characteristics of wind power and the first historical wind power data for model training.

[0011] Obtain the second historical wind power data of the target wind farm group;

[0012] The source wind farm prediction model is adjusted using the second historical wind power data to obtain the target wind farm prediction model, wherein the target wind farm prediction model is used to predict the wind power of the target wind farm group and output wind power prediction data.

[0013] By using the actual wind power data continuously accumulated during the operation of the target wind farm, and the matching wind power prediction data, the model parameters of the prediction model for the target wind farm are adaptively updated to adapt to changes in the operating status of the target wind farm group.

[0014] Optionally, the step of constructing the dynamic spatial topology of the source wind farm group using the geographic coordinate information, the meteorological observation data, and the first historical wind power data includes:

[0015] Using the geographic coordinate information, the static spatial distribution characteristics of the source wind farm group are obtained;

[0016] Using the meteorological observation data and the first historical wind power data, the dynamic correlation characteristics of the source wind farm group are obtained;

[0017] The static spatial distribution characteristics and the dynamic correlation characteristics are fused to generate the dynamic spatial topology of the source wind farm group.

[0018] Optionally, obtaining the static spatial distribution characteristics of the source wind farm group using the geographic coordinate information includes:

[0019] Obtain the Euclidean distance between any two wind farms in the source wind farm group;

[0020] Based on the Euclidean distance, spatial similarity is calculated using a normalization function;

[0021] Based on the spatial similarity between all wind farm pairs, the static spatial distribution characteristics of the source wind farm group are constructed.

[0022] Optionally, fusing the static spatial distribution features with the dynamic correlation features to generate the dynamic spatial topology of the source wind farm group includes:

[0023] The static spatial distribution features and the dynamic correlation features are weighted and summed using weighted fusion parameters.

[0024] The weighted summation result is processed by an activation function to generate the dynamic spatial topology of the source wind farm group.

[0025] Optionally, obtaining the dynamic correlation characteristics of the source wind farm cluster using the meteorological observation data and the first historical wind power data includes:

[0026] The meteorological observation data and the first historical wind power data are aligned and fused to construct a node feature vector for each wind farm in the source wind farm group;

[0027] The node feature vectors are mapped to a high-dimensional feature space to obtain a high-dimensional feature representation;

[0028] Based on the high-dimensional feature representation, the attention coefficient between any two wind farms is calculated using a graph attention network.

[0029] The attention coefficients are normalized to obtain the dynamic correlation characteristics of the source wind farm group.

[0030] Optionally, adjusting the source wind farm prediction model using the second historical wind power data to obtain the target wind farm prediction model includes:

[0031] Freeze the parameters of the graph convolutional network in the source wind farm prediction model;

[0032] Using the second historical wind power data, the temporal convolutional network, feature fusion parameters, and prediction layer parameters in the source wind farm prediction model are fine-tuned to obtain the target wind farm prediction model.

[0033] Optionally, the step of constructing the joint spatiotemporal characteristics of wind power of the source wind farm group using the dynamic spatial topology relationship and the first historical wind power data includes:

[0034] Based on the dynamic spatial topology relationship, the spatial topology features of the source wind farm group are extracted using a graph convolutional network;

[0035] The first historical wind power data is processed using a temporal convolutional network to extract the temporal evolution characteristics of wind power in the source wind farm group;

[0036] The spatial topological features and the temporal evolution features of wind power are fused to generate the joint spatiotemporal features of wind power of the source wind farm group.

[0037] Optionally, fusing the spatial topological features and the temporal evolution features of wind power to generate the joint spatiotemporal features of wind power in the source wind farm cluster includes:

[0038] The spatial topological features and the wind power time-series evolution features are aligned and stitched together along the feature dimension to obtain the fused features;

[0039] The fusion features are reduced in dimensionality and integrated to obtain the joint spatiotemporal features of wind power.

[0040] Optionally, the step of adaptively updating the model parameters of the prediction model of the target wind farm by using the actual wind power data continuously accumulated during the operation of the target wind farm and the matching wind power prediction data includes:

[0041] Obtain the actual wind power data continuously accumulated during the operation of the target wind farm group, and the matching wind power prediction data;

[0042] Calculate the prediction error between the actual wind power data and the predicted wind power data;

[0043] The prediction error is used as a feedback signal to construct a loss function;

[0044] The gradient of the graph attention network and the target wind farm prediction model is calculated simultaneously using the backpropagation algorithm and the loss function.

[0045] The gradient descent algorithm is used to collaboratively update the graph attention network and the target wind farm prediction model based on the calculated gradient.

[0046] Optionally, the step of fine-tuning the temporal convolutional network, feature fusion parameters, and prediction layer parameters in the source wind farm prediction model using the second historical wind power data includes:

[0047] Based on the final learning rate of the source wind farm prediction model in the pre-training stage, an initial fine-tuning learning rate is set, wherein the initial fine-tuning learning rate is less than the final learning rate;

[0048] During the fine-tuning iteration of the temporal convolutional network, feature fusion parameters, and prediction layer parameters in the source wind farm prediction model, the initial fine-tuning learning rate is exponentially decayed according to a preset number of cycles.

[0049] By employing the above technical solution, this invention provides a cross-wind farm wind power prediction method, which obtains the geographic coordinate information, meteorological observation data, and first historical wind power data of the source wind farm group; constructs the dynamic spatial topology of the source wind farm group using the geographic coordinate information, meteorological observation data, and first historical wind power data; constructs the joint spatiotemporal characteristics of wind power of the source wind farm group using the dynamic spatial topology and first historical wind power data; trains the model using the joint spatiotemporal characteristics of wind power and first historical wind power data to obtain the source wind farm prediction model; obtains the second historical wind power data of the target wind farm group; adjusts the source wind farm prediction model using the second historical wind power data to obtain the target wind farm prediction model, wherein the target wind farm prediction model is used to predict the wind power of the target wind farm group and outputs wind power prediction data; and adaptively updates the model parameters of the target wind farm prediction model using the actual wind power data continuously accumulated during the operation of the target wind farm and the matching wind power prediction data to adapt to changes in the operating status of the target wind farm group. This invention constructs a dynamic spatial topology by integrating static geographic information and dynamic meteorological information, and further combines historical wind power data to generate joint spatiotemporal features, thereby accurately modeling the spatiotemporal coupling evolution law of wind farm clusters. It utilizes transfer learning and dynamic updates to achieve high-precision and high-stability cross-wind farm power prediction to meet the actual operation requirements of the power system.

[0050] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0051] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0052] Figure 1 A flowchart illustrating one embodiment of the cross-wind farm wind power prediction method provided by this invention is shown.

[0053] Figure 2 This is a flowchart illustrating a specific implementation of step S110 in the cross-wind farm wind power prediction method provided by an embodiment of the present invention.

[0054] Figure 3This is a flowchart illustrating a specific implementation of step S200 in the cross-wind farm wind power prediction method provided by an embodiment of the present invention;

[0055] Figure 4 This is a flowchart illustrating a specific implementation of step S210 in the cross-wind farm wind power prediction method provided by an embodiment of the present invention.

[0056] Figure 5 The diagram shows a specific implementation of step S120 in the cross-wind farm wind power prediction method provided by an embodiment of the present invention. Detailed Implementation

[0057] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0058] With the continued growth of global demand for renewable energy, wind power, as an important form of clean energy, plays a crucial role in promoting energy structure transformation and addressing climate change. Wind power output is significantly affected by weather conditions, exhibiting strong randomness and uncertainty. As the installed capacity and number of wind farms continue to expand, accurate wind power forecasting has become a vital foundation for ensuring the safe, stable, and economical operation of the power system.

[0059] Traditional wind power forecasting often focuses on individual wind farms. However, with the coordinated development of regional power grids, joint forecasting across wind farms has gradually become a research hotspot. Comprehensively utilizing spatial correlation information among multiple wind farms can help improve the overall forecasting accuracy and reliability for the region. However, due to differences in geographical location, meteorological conditions, and operational characteristics among different wind farms, the spatial correlations and power time-series characteristics between them exhibit significant heterogeneity and non-stationarity. This makes it difficult to directly generalize forecasting models trained on a single wind farm to other wind farms, especially for the forecasting of newly added wind farms.

[0060] Existing methods for predicting wind power across wind farms mostly employ static spatial structure modeling and offline training, which are insufficient to effectively adapt to the dynamic changes in the spatial structure and operating status of wind farm clusters. Although some methods introduce online update strategies, they only update the parameters of the prediction model and fail to simultaneously adjust the construction method of the spatial structure model, thus limiting the model's ability to capture the dynamic evolution characteristics of wind farm clusters.

[0061] Based on this, this invention provides a method for predicting wind power across wind farms. It utilizes the geographical coordinates, meteorological observation data, and historical wind power data of the source wind farm cluster to construct a dynamic spatial topology relationship. Using this dynamic spatial topology relationship and historical wind power data, it constructs a joint spatiotemporal feature of wind power in the source wind farm cluster. Based on this spatiotemporal feature and historical power data, it trains a prediction model for the source wind farm cluster. Using a small amount of historical wind power data from the target wind farm, it migrates the prediction model from the source wind farm cluster to the target wind farm, accumulating real and predicted wind power data from the target wind farm during continuous operation. A dynamic update mechanism is introduced to enable the model to adaptively adjust to changes in operating status, achieving accurate prediction of wind power, improving the model's generalization ability and adaptability, thereby enhancing the accuracy and stability of cross-wind farm wind power prediction and meeting the operational needs of actual power systems.

[0062] like Figure 1 The diagram shows a flowchart of one embodiment of the cross-wind farm wind power prediction method provided by this invention. The method may include:

[0063] S100: Obtain the geographic coordinates, meteorological observation data, and first historical wind power data of the source wind farm group.

[0064] In this context, a wind farm cluster refers to a collection of multiple wind farms within a designated area, which exhibit spatial correlations in terms of geographical location, meteorological conditions, and power output. This invention uses the wind farm cluster as a whole to perform joint modeling and prediction of wind power across wind farms, reflecting the comprehensive characteristics of regional wind power resources.

[0065] Geographic coordinate information refers to spatial data used to describe the geographical location of a wind farm, usually expressed in longitude and latitude.

[0066] Meteorological observation data refers to various meteorological parameters of the wind farm's location, such as wind speed, wind direction, temperature, and air pressure.

[0067] Among them, the first historical wind power data refers to the actual active power output data of the source wind farm group recorded in chronological order over a period of time.

[0068] Specifically, embodiments of the present invention can collect the geographic coordinate information of each wind farm in the source wind farm cluster, which serves as the basis for constructing the spatial relationships of the wind farms. Meteorological observation data within the target area are collected simultaneously, along with historical wind power output sequences of each wind farm in the source wind farm cluster over a past period. These data can be time-aligned and preprocessed to form a unified multidimensional time series dataset, providing basic input for subsequent feature extraction and model construction.

[0069] As examples, embodiments of the present invention can obtain the geographic coordinates of all wind farms in the source wind farm cluster from their operation management systems or geographic information databases, forming the spatial location basis of the wind farm nodes. Secondly, meteorological observation data is collected from meteorological stations deployed within or near the wind farms, mainly including core meteorological parameters strongly correlated with power generation, such as wind speed and direction, and time synchronization and outlier processing are performed to ensure data quality. Simultaneously, first-historical wind power data is collected from the power grid dispatch center or wind farm energy management system, i.e., the active power output time series data of each wind farm in the source wind farm cluster over a past period (e.g., the past few hours to days). Finally, all data from all sources are aligned and normalized under a unified timestamp, forming a standardized dataset containing multi-station, multi-modal features, providing input for subsequent modeling.

[0070] S110. Using geographic coordinate information, meteorological observation data, and first historical wind power data, construct the dynamic spatial topology of the source wind farm group.

[0071] Among them, dynamic spatial topology refers to a graph structure that can characterize the time-varying spatial dependencies between wind farms by integrating geographical static adjacency and meteorological dynamic similarity and using graph attention network (GAT) adaptive learning. Specifically, it is embodied in dynamic adjacency matrix.

[0072] Specifically, embodiments of the present invention can construct an initial static spatial adjacency matrix based on the geographical coordinates of wind farms using Euclidean distance calculations, reflecting the spatial distribution relationships between wind farms. Meteorological observation data and first historical wind power data are used as node feature inputs to a graph attention network to dynamically learn the nonlinear spatial dependencies between wind farms. By calculating the attention weights between nodes, time-varying spatial correlations are captured. Combining the static adjacency matrix and the dynamic feature matrix, a learnable weighted fusion mechanism is used to generate a dynamic adjacency matrix, characterizing the dynamic spatial topology relationships of the wind farm cluster over time.

[0073] As examples, embodiments of the present invention can utilize geographic coordinate information to calculate the pairwise distances between wind farms using Euclidean distance, constructing a static spatial adjacency matrix to characterize their inherent geographic proximity. To capture complex relationships beyond static distances, meteorological observation data (such as wind speed and direction) and historical power sequences (such as recent power values) for each wind farm at time t are fused into a node feature vector. Then, a graph attention network is introduced, using its attention mechanism to dynamically calculate the nonlinear dependencies between the node features of any two wind farms, generating a dynamic feature similarity matrix. Finally, a weighted fusion mechanism with learnable parameters is designed to adaptively fuse the static adjacency matrix and the dynamic feature similarity matrix, generating the final dynamic adjacency matrix. This dynamic adjacency matrix, as a representation of the dynamic graph structure, allows its edge weights to change in real time with meteorological conditions and power output status, thereby accurately characterizing the dynamic evolution of spatial correlations among wind farm groups.

[0074] S120. Using dynamic spatial topology and first historical wind power data, construct the joint spatiotemporal characteristics of wind power in the source wind farm group.

[0075] Among them, the joint spatiotemporal features of wind power refer to the unified feature representation formed by extracting spatial features through Graph Convolutional Network (GCN) and temporal features through Temporal Convolutional Network (TCN) based on dynamic spatial topology, and fusing the two to comprehensively characterize the spatiotemporal evolution of wind power.

[0076] Specifically, embodiments of the present invention can extract the spatial topological features of wind farm clusters based on a dynamic adjacency matrix using graph convolutional networks, capturing the mutual influences and spatial dependencies between wind farms. Simultaneously, temporal convolutional networks are used to extract temporal features from the first historical wind power data of each wind farm, uncovering the temporal evolution patterns of power output and multi-scale temporal dependencies. By aligning and fusing spatial and temporal features, a unified high-dimensional spatiotemporal feature representation is constructed, providing rich input information for the prediction model and improving the accuracy and robustness of predictions.

[0077] As examples, embodiments of the present invention can utilize a dynamic adjacency matrix as input to a graph structure, employing a graph convolutional network to propagate and aggregate information across the entire wind farm cluster. This effectively extracts the spatial topological features of each wind farm node under the influence of its dynamic neighbor nodes, capturing spatial correlation patterns between wind farms. Simultaneously, the first historical wind power data for each wind farm is processed using a temporal convolutional network. The temporal convolutional network, utilizing an dilated causal convolutional structure, can efficiently capture the long-term dependencies, periodicity, and trends of power data over time. Finally, the spatial feature matrix reflecting inter-farm correlations extracted by the graph convolutional network and the temporal feature matrix reflecting the individual changes of each farm, extracted by the temporal convolutional network, are aligned and deeply fused to form a unified joint spatiotemporal feature of wind power. This joint spatiotemporal feature of wind power integrates the spatiotemporal driving factors of power changes, providing a high-dimensional information representation for the final prediction.

[0078] S130. The model is trained using the spatiotemporal characteristics of wind power and the first historical wind power data to obtain the source wind farm prediction model.

[0079] Among them, the source wind farm prediction model refers to a machine learning model that has been pre-trained, can receive the joint spatiotemporal features of wind power as input, and output the predicted value of wind power of the source wind farm in the future period. Its core may include a spatiotemporal feature extraction module and a prediction layer.

[0080] Specifically, this embodiment of the invention can use historical data of the source wind farm cluster over a long period of time (including the geographical coordinates of the wind farms, meteorological observation data at each time step, and the first historical wind power data) as input. First, a static spatial distribution feature is constructed based on the geographical coordinates of the source wind farms, and a dynamic adjacency matrix is ​​dynamically generated at each time step in combination with real-time data. Then, on each batch of data, dynamic spatial topological features are extracted through a graph convolutional network, and temporal evolution features of each wind farm are extracted through a temporal convolutional network. The two are then fused and used for power prediction through a fully connected layer. The goal of model training is to minimize the mean square error between the predicted power and the actual power. Through the backpropagation algorithm, all learnable parameters involved in cross-wind farm wind power prediction are jointly optimized, including parameters for generating dynamic spatial topological relationships, graph convolutional network parameters, temporal convolutional network parameters, feature fusion, and prediction layer parameters. After sufficient iteration, the model learns the parameter configuration that captures universal spatiotemporal correlation patterns from the source wind farm cluster data, completes pre-training, and obtains a trained source wind farm prediction model.

[0081] S140, Obtain the second historical wind power data of the target wind farm group.

[0082] The second historical wind power data refers to the actual active power output data of the target wind farm group recorded in chronological order over a period of time.

[0083] Specifically, embodiments of the present invention can obtain the historical wind power output sequence of each wind farm in the target wind farm group over a past period. This data can be time-aligned and preprocessed to form a unified multidimensional time series dataset.

[0084] As some examples, embodiments of the present invention can collect second historical wind power data from the power grid dispatch center or the wind farm energy management system, that is, the active power output time series data of each wind farm in the target wind farm group over a period of time (e.g., the past few hours to several days).

[0085] S150. Adjust the source wind farm prediction model using the second historical wind power data to obtain the target wind farm prediction model. The target wind farm prediction model is used to predict the wind power of the target wind farm group and output wind power prediction data.

[0086] Specifically, embodiments of the present invention can migrate the general spatiotemporal feature representations, data, and computational resources learned in the source wind farm prediction model to the target wind farm cluster. During the migration phase, the graph structure-related parameters in the source wind farm prediction model are frozen, and then the temporal feature extraction branch of the source wind farm prediction model is fine-tuned using the second historical wind power data of the target wind farm cluster. This optimizes the relevant feature fusion parameters and prediction layer parameters, resulting in a target wind farm prediction model adapted to the target wind farm cluster, achieving efficient cross-wind farm power prediction.

[0087] Furthermore, in this embodiment of the invention, the joint spatiotemporal characteristics of wind power of the target wind farm group can be input into the target wind farm prediction model to obtain the wind power prediction result of the target wind farm group. The wind power prediction result refers to the numerical prediction output of the active power output of the target wind farm group in a specific future period based on the input joint spatiotemporal characteristics, i.e., wind power prediction data.

[0088] It is understandable that the process of obtaining the joint spatiotemporal characteristics of wind power in the target wind farm cluster can refer to the description of the joint spatiotemporal characteristics of wind power in the source wind farm cluster in steps S100 to S120, and will not be repeated here. This target wind farm prediction model, through a fully connected layer, can fully utilize the spatiotemporal dependency information of the target wind farm cluster to predict power, outputting predicted wind power values ​​for multiple future time steps. The wind power prediction results can be used for power system dispatch and operation decisions, and can be continuously optimized during the model's online operation phase by incorporating an online learning mechanism to improve the real-time performance and stability of the prediction.

[0089] As examples, embodiments of the present invention can load a target wind farm prediction model. This target wind farm prediction model is a regression network with a fully connected layer at its core. Its parameters have been fully optimized on a large amount of historical data from the source wind farm cluster through a training mechanism involving spatiotemporal feature extraction and fusion, and online learning optimization. Then, the joint spatiotemporal features of wind power representing the current and recent state of the target wind farm cluster are input into the target wind farm prediction model. The fully connected layer (prediction layer) inside the target wind farm prediction model performs nonlinear transformation and mapping on the complex spatiotemporal features of the input, and finally outputs the wind power prediction results for each wind farm in the target wind farm cluster for a specific future time period (next hour or next 24 hours), i.e., wind power prediction data. Through the transfer learning mechanism, when this model is applied to a new target wind farm cluster, it only needs to be fine-tuned using a small amount of data from the target wind farms to quickly adapt and efficiently generate high-precision prediction results, meeting actual scheduling needs.

[0090] S160. Using the actual wind power data continuously accumulated during the operation of the target wind farm, and the matching wind power prediction data, the model parameters of the target wind farm prediction model are adaptively updated to adapt to changes in the operating status of the target wind farm group.

[0091] Specifically, the embodiments of the present invention can continuously monitor the prediction error between the actual wind power data accumulated during the operation of the target wind farm and the matching wind power prediction data. Using the prediction error as a feedback signal, the model parameters of the target wind farm prediction model are updated online, so as to realize the adaptive adjustment of the target wind farm prediction model to the changes in the operating status of the target wind farm group.

[0092] This invention provides a method for predicting wind power across wind farms. The method includes: obtaining geographic coordinate information, meteorological observation data, and first historical wind power data of a source wind farm group; constructing a dynamic spatial topology of the source wind farm group using the geographic coordinate information, meteorological observation data, and first historical wind power data; constructing a joint spatiotemporal feature of wind power for the source wind farm group using the dynamic spatial topology and the first historical wind power data; training a model using the joint spatiotemporal feature and the first historical wind power data to obtain a source wind farm prediction model; obtaining second historical wind power data of a target wind farm group; adjusting the source wind farm prediction model using the second historical wind power data to obtain a target wind farm prediction model, wherein the target wind farm prediction model is used to predict the wind power of the target wind farm group and output wind power prediction data; and adaptively updating the model parameters of the target wind farm prediction model using actual wind power data continuously accumulated during the operation of the target wind farm and matching wind power prediction data to adapt to changes in the operating status of the target wind farm group. This invention constructs a dynamic spatial topology by integrating static geographic information and dynamic meteorological information, and further combines historical wind power data to generate joint spatiotemporal features, thereby accurately modeling the spatiotemporal coupling evolution law of wind farm clusters. It utilizes transfer learning and dynamic updates to achieve high-precision and high-stability cross-wind farm power prediction to meet the actual operation requirements of the power system.

[0093] Optional, based on Figure 1 The method shown is as follows: Figure 2 The diagram shows a specific implementation of step S110 in the cross-wind farm wind power prediction method provided by this invention. Step S110 may specifically include:

[0094] S200. Using geographic coordinate information, obtain the static spatial distribution characteristics of the source wind farm group.

[0095] Among them, static spatial distribution characteristics refer to fixed features that reflect the relative spatial positions and distances between wind farms, constructed using spatial distance or adjacency relationships based on the geographic spatial location relationships of wind farms. Static spatial distribution characteristics are represented by a static spatial adjacency weight matrix, used to describe the spatial structure and topological distribution of wind farm clusters.

[0096] Specifically, this embodiment of the invention can calculate the spatial distance between any two wind farms using the Euclidean distance metric based on the geographical coordinates of each wind farm in the target wind farm cluster. The distance is then converted into spatial similarity using a normalization function, constructing a static spatial adjacency weight matrix. This static spatial adjacency weight matrix reflects the fixed spatial distribution characteristics of wind farms based on their geographical locations, forming a static graph structure G=(V, E), where V represents the set of nodes and E represents the set of edges. Each wind farm corresponds to one node, and the edge weights are determined by spatial similarity, providing the basic spatial topology for the subsequent construction of a dynamic adjacency matrix.

[0097] S210. Using meteorological observation data and first historical wind power data, obtain the dynamic correlation characteristics of the source wind farm group.

[0098] Among them, dynamic correlation features refer to the capture of nonlinear dependencies and mutual influences between wind farms over time based on real-time meteorological observation data and historical power sequences of wind farms, through dynamic learning mechanisms such as graph attention networks, reflecting the time-varying characteristics of the spatial topology of wind farm groups, and manifested as a dynamic feature similarity matrix.

[0099] Specifically, embodiments of the present invention can align and fuse meteorological observation data of source wind farms with historical wind power time series to construct a time-dynamic node feature vector. This node feature vector is then input into a graph attention network, which automatically captures the nonlinear time-varying dependencies between wind farms by learning the attention weights between nodes, resulting in a feature similarity matrix reflecting the dynamic spatial correlation of the wind farm cluster. This feature similarity matrix changes over time and can adaptively characterize the dynamic evolution of the spatial topology among wind farms caused by changes in meteorological and operational states.

[0100] S220. The static spatial distribution characteristics and dynamic correlation characteristics are fused to generate the dynamic spatial topology of the source wind farm group.

[0101] Specifically, embodiments of the present invention can utilize a learnable weighted fusion mechanism to weight and fuse a static spatial adjacency weight matrix and a feature similarity matrix: through parameterized weight matrices and activation functions, element-wise weighting and nonlinear mapping are applied to the two types of features to construct a fused dynamic adjacency matrix. This dynamic adjacency matrix comprehensively reflects the geographic spatial constraints and time-varying nonlinear relationships of wind farm clusters, accurately characterizing the dynamic spatial topological relationships between wind farms, and laying a solid foundation for subsequent spatiotemporal feature extraction using graph convolutional networks and temporal convolutional networks.

[0102] This invention utilizes geographic coordinate information to extract the stable static spatial distribution characteristics of wind farm clusters, and combines meteorological observation data and first historical wind power data to dynamically capture the time-varying correlations between wind farms, thereby comprehensively reflecting the spatial and temporal characteristics of wind farm clusters. By fusing static and dynamic features, the constructed dynamic spatial topology can more accurately describe the complex interactions and time-varying dependencies between wind farms, significantly improving the accuracy and generalization ability of wind power prediction models.

[0103] Optional, based on Figure 2 The method shown is as follows: Figure 3 The diagram shows a specific implementation of step S200 in the cross-wind farm wind power prediction method provided by this invention. Step S200 may specifically include:

[0104] S300: Obtain the Euclidean distance between any two wind farms in the source wind farm group.

[0105] Specifically, embodiments of the present invention can calculate the Euclidean distance between any two wind farms i and j based on the geographical coordinates (p, q) of each wind farm in the target wind farm cluster. As a basis for measuring their geographical proximity: .

[0106] S310. Based on Euclidean distance, spatial similarity is calculated using a normalization function.

[0107] Specifically, embodiments of the present invention can utilize Euclidean distance. Substitute into the formula ,in, This is a normalization parameter to control the similarity decay rate. This operation maps the distance to a value between 0 and 1; the closer the distance, the higher the spatial similarity between wind farm i and wind farm j. The closer a value is to 1, the stronger the static spatial correlation.

[0108] S320. Based on the spatial similarity between all wind farm pairs, construct the static spatial distribution characteristics of the source wind farm group.

[0109] Specifically, in this embodiment of the invention, the spatial similarity between all wind farm pairs can be used to construct a static spatial adjacency weight matrix. ,in, This represents the number of wind farms. This matrix is ​​a symmetric matrix, and its elements... The quantitative description of the static similarity between wind farm i and wind farm j based on geographical location does not change over time, providing a basic spatial structure framework for the subsequent construction of dynamic relationships.

[0110] This invention calculates the Euclidean distance between wind farms based on geographic coordinates and obtains spatial similarity using a normalization function. The constructed static spatial distribution features can accurately reflect the geographic spatial structure of wind farm clusters, providing a stable and effective spatial basis for constructing dynamic spatial topological relationships, and improving the accuracy and robustness of wind power prediction.

[0111] Optional, based on Figure 2 The method shown is as follows: Figure 4 The diagram shows a specific implementation of step S210 in the cross-wind farm wind power prediction method provided by this invention. Step S210 may specifically include:

[0112] S400: Align and fuse meteorological observation data and first historical wind power data to construct a node feature vector for each wind farm in the source wind farm group.

[0113] Specifically, in this embodiment of the invention, for each wind farm, at time t, its meteorological observation data and first historical wind power data are feature-aligned and stitched together to form the original feature vector h of each node. i (t) .

[0114] S410. Map the node feature vectors to a high-dimensional feature space to obtain a high-dimensional feature representation.

[0115] Specifically, embodiments of the present invention can utilize a learnable weight matrix W. s Regarding the feature h i (t) By performing nonlinear transformations and dimensionality increases, high-dimensional node feature representations are obtained. This is to enhance the expressive power of features.

[0116] S420. Based on high-dimensional feature representation, the attention coefficient between any two wind farms is calculated using a graph attention network.

[0117] Specifically, embodiments of the present invention can introduce a graph attention network mechanism, whereby for any pair of nodes (i,j), the dynamic attention coefficient of node j to node i is calculated. This is achieved by concatenating the high-dimensional features of the two nodes with a learnable attention vector. Implement by performing a dot product and then applying ReLU activation: ,in," " indicates vector concatenation.

[0118] S430. Normalize the attention coefficients to obtain the dynamic correlation characteristics of the source wind farm group.

[0119] Specifically, in embodiments of the present invention, the attention coefficients of all neighboring nodes j of each node i can be obtained using the Softmax function. Normalization is performed to obtain normalized attention weights. All node pairs Constructing a dynamic feature similarity matrix It quantifies the dynamic influence of wind farm j on wind farm i at time t. This matrix changes in real time with meteorological and power data, reflecting complex nonlinear spatial dependencies.

[0120] This invention, through the alignment and fusion of meteorological observation data and first historical wind power data, and the extraction of high-dimensional dynamic correlation features using graph attention networks, can effectively capture the nonlinear and time-varying dependencies between wind farm clusters, thereby improving the accuracy of spatial topology construction and the performance of wind power prediction models.

[0121] Optionally, in the above Figure 2 Based on one or more corresponding embodiments, in another optional embodiment provided by the present invention, step S220 may specifically include:

[0122] By using weighted fusion parameters, the static spatial distribution characteristics and dynamic correlation characteristics are weighted and summed; the weighted summation result is then processed by an activation function to generate the dynamic spatial topology of the source wind farm group.

[0123] The weighted fusion parameter refers to the trainable weight coefficients used to adjust the relative contributions of the two types of features during the fusion of static spatial distribution features and dynamic correlation features. The weighted fusion parameter can include a weight matrix, learnable parameters, bias terms, and activation functions.

[0124] Specifically, embodiments of the present invention can design a learnable gating parameter. (Usually constrained to the [0,1] interval by the Sigmoid function), used to dynamically determine the fusion weight of static and dynamic features. Then, a learnable weight matrix is ​​used respectively. and For the static space adjacency weight matrix respectively and node feature vectors Perform the transformation and add a bias term. ,Right now: ,in, This is the Sigmoid activation function. The fusion process is achieved through element-wise multiplication and weighted summation, as shown in the formula. ,in, This is a dynamic adjacency matrix used to characterize the dynamic spatial topological relationships of a wind farm group.

[0125] This invention employs weighted fusion parameters to adaptively weight and sum static spatial distribution features and dynamic correlation features, followed by activation function processing. This effectively integrates the time-varying and static spatial information of wind farm clusters, improves the accuracy of expressing dynamic spatial topological relationships and the model's ability to characterize complex correlation patterns of wind farm clusters, thereby enhancing the accuracy and robustness of wind power prediction.

[0126] Optionally, in the above Figure 1 Based on one or more corresponding embodiments, in another optional embodiment provided by the present invention, step S150 may specifically include:

[0127] The parameters of the graph convolutional network in the source wind farm prediction model are frozen; using the second historical wind power data, the temporal convolutional network, feature fusion parameters, and prediction layer parameters in the source wind farm prediction model are fine-tuned to obtain the target wind farm prediction model.

[0128] This invention, after obtaining the source wind farm prediction model, applies its overall architecture to the target wind farm cluster. First, the input data is replaced with data from the target wind farm cluster, including its geographical coordinates, meteorological data, and power series. The key operation is parameter freezing: based on... All parameters used to calculate the attention weights between nodes in the graph attention network. Parameters in graph convolutional networks used to extract spatial features from graph-structured data These parameters are fixed and will no longer participate in subsequent gradient updates. Freezing these parameters means that the model still uses the general patterns and prior knowledge learned from the source domain regarding the spatial relationships between wind farms in the target domain, providing a stable and informative basis for spatial feature extraction for subsequent fine-tuning.

[0129] After freezing the graph structure parameters, only the temporal feature extraction branch is fine-tuned. Based on... The focus of fine-tuning is on the convolutional kernel weights of the temporal feature extraction branch. Connection weights of the feature fusion layer and the fully connected parameters of the prediction layer For example, embodiments of the present invention can utilize the second historical wind power data (such as data from the past 24 hours) of the target wind farm cluster as the main input, and calculate the gradient of the loss function with respect to the following trainable parameters through backpropagation: the parameters of the TCN ( , This allows it to learn the power fluctuation patterns and time-dependent modes specific to the target wind farm; the parameters of the feature fusion layer (if the fusion method involves learnable weights); and the parameters of the prediction layer ( , ).

[0130] This invention first pre-trains a wind power prediction model on historical data of the source wind farm group, then transfers the pre-trained model to the target wind farm group while freezing the graph convolutional network parameters. Combined with the fine-tuning of the temporal convolutional network and fusion parameters using historical data of the target wind farm, the model achieves efficient transfer and rapid adaptation to the spatiotemporal characteristics of different wind farms, significantly improving prediction accuracy and generalization ability.

[0131] Optionally, in embodiments of the present invention, an initial fine-tuning learning rate can be set based on the final learning rate of the source wind farm prediction model in the pre-training stage, wherein the initial fine-tuning learning rate is less than the final learning rate; during the fine-tuning iteration of the temporal convolutional network, feature fusion parameters, and prediction layer parameters in the source wind farm prediction model, the initial fine-tuning learning rate is exponentially decayed according to a preset number of cycles.

[0132] Specifically, embodiments of the present invention may employ a smaller initial learning rate (e.g., ...) during the training process. Based on pre-training learning rate (Proportional settings) and periodic decay strategies (such as...) Where e is the preset number of periods (epochs). (where is the learning rate for the e-th training round) to stably and finely adjust the parameters. Through a small number of iterations, the model quickly learns the temporal behavior of the target field while retaining general spatial knowledge, thereby achieving efficient and accurate prediction of the power of the target wind farm group.

[0133] This invention employs an initial fine-tuning learning rate lower than the final learning rate in the pre-training stage, combined with an exponential decay strategy with a preset period, to effectively avoid model oscillations and overfitting caused by excessive parameter updates during fine-tuning, thereby improving the convergence stability and prediction performance of the model on the target wind farm.

[0134] Optional, based on Figure 1 The method shown is as follows: Figure 5 The diagram shows a specific implementation of step S120 in the cross-wind farm wind power prediction method provided by this invention. Step S120 may specifically include:

[0135] S500: Based on dynamic spatial topology relationships, it uses graph convolutional networks to extract the spatial topology features of source wind farm clusters.

[0136] Among them, spatial topological features refer to the representation of the dynamic spatial association patterns within a wind farm cluster extracted by using graph convolutional networks to propagate and aggregate information about dynamic spatial topological relationships.

[0137] Specifically, embodiments of the present invention can utilize graph convolutional networks to represent the high-dimensional features of wind farm nodes by applying dynamic adjacency matrices. Through the product of the adjacency matrix and the feature matrix, along with parameterized linear transformations and activation functions, the spatial topological features of each wind farm under dynamic spatial relationship constraints are extracted.

[0138] ,

[0139] Where l represents the graph convolutional layer index in the graph convolutional network; i represents the node index in the dynamic adjacency matrix, i.e., the wind farm number; Let L be the weight parameters of the l-th layer in the graph convolutional network; This is the bias term of the l-th layer in the graph convolutional network. The final output of the graph convolutional network is the spatial topological features of the target wind farm cluster. .

[0140] S510. Use a temporal convolutional network to process the first historical wind power data and extract the temporal evolution characteristics of wind power in the source wind farm group.

[0141] Among them, the wind power time series evolution characteristics refer to the characterization that reflects the inherent law of power change over time, which is extracted by deep processing of the historical power sequence of a single wind farm through a time convolutional network.

[0142] Specifically, embodiments of the present invention can utilize a temporal convolutional network to capture the multi-scale dependence and evolutionary characteristics of wind power sequences over time through multiple layers of convolutional operations with dilated kernels. Dilated convolution can effectively expand the receptive field and learn the temporal variation patterns from short-term to long-term. After a series of convolutional layers and activation function processing, the temporal convolutional network outputs high-dimensional temporal features reflecting the dynamic evolution of wind power sequences:

[0143] ,

[0144] in, This represents the weight parameters of the convolution kernel at the k-th position; Given a segment of the input sequence, denoted as the wind power value at the k-th time step m forward from the t-th time point; This is the bias term for the temporal convolutional layer. The final output of the temporal convolutional network is the temporal evolution characteristics of the wind power of the target wind farm cluster. .

[0145] S520. The spatial topological features and the temporal evolution features of wind power are fused to generate the joint spatiotemporal features of wind power of the source wind farm group.

[0146] Specifically, embodiments of the present invention can align and stitch together spatial topological features and temporal evolution features of wind power to form a unified spatiotemporal feature representation. This fusion process ensures compatibility between spatial and temporal dimensions through feature alignment operations, and then organically combines the two types of features through methods such as concatenation or weighted fusion. The fused high-dimensional spatiotemporal features can comprehensively reflect the spatial structural correlation and temporal power variation characteristics of wind farm clusters, and are fed as input into the subsequent fully connected prediction layer to achieve accurate prediction of future wind power.

[0147] This invention utilizes graph convolutional networks to extract spatial features based on dynamic spatial topology relationships, combines temporal convolutional networks to capture the temporal evolution features of historical power sequences, and constructs joint spatiotemporal features of wind power by fusing the two types of features. This comprehensively characterizes the spatiotemporal dependence of wind farm clusters, significantly improving the accuracy and generalization ability of wind power prediction.

[0148] Optionally, in the above Figure 5 Based on one or more corresponding embodiments, in another optional embodiment provided by the present invention, step S520 may specifically include:

[0149] Spatial topological features and wind power temporal evolution features are aligned and stitched together in the feature dimension to obtain fused features; the fused features are then reduced in dimension and integrated to obtain joint spatiotemporal features of wind power.

[0150] Specifically, due to differences in dimensionality or structure between spatial topological features and wind power temporal evolution features, feature alignment operations are required. For example, fully connected layers or linear transformations can be used to... and Mapping to the same feature dimension (e.g., both becoming d-dimensional) ensures shape compatibility. After alignment, for each wind farm node, its spatial and temporal features are concatenated along the feature dimension to obtain the joint spatiotemporal features of wind power. To achieve the fusion of spatiotemporal features:

[0151] .

[0152] Finally, the combined spatiotemporal characteristics of wind power are input into the fully connected prediction layer, i.e., the wind power prediction model, which outputs the wind power prediction results for each wind farm in future time periods. :

[0153] ,

[0154] in, These are the weight parameters for the fully connected prediction layer; This is the bias term for the fully connected prediction layer.

[0155] This invention improves the expressive power of joint spatiotemporal features of wind power and the accuracy of prediction models by aligning and splicing spatial topological features and wind power temporal evolution features in the feature dimension, combined with dimensionality reduction and integration operations, thereby effectively fusing multi-source spatiotemporal information.

[0156] Optionally, in the above Figure 4 Based on one or more corresponding embodiments, in another optional embodiment provided by the present invention, step S160 may specifically include:

[0157] The process involves obtaining actual wind power data accumulated during the operation of the target wind farm cluster, along with matching wind power prediction data; calculating the prediction error between the actual and predicted wind power data; using the prediction error as a feedback signal to construct a loss function; employing the backpropagation algorithm to simultaneously calculate the gradient of the graph attention network and the target wind farm prediction model using the loss function; and using the gradient descent algorithm to collaboratively update the graph attention network and the target wind farm prediction model based on the calculated gradient.

[0158] Specifically, in the model running phase, for each prediction time point t, the wind power prediction result of any wind farm can be output from the target wind farm prediction model. Simultaneously, actual wind power data of the same wind farm observed at that time point are obtained. The prediction error is obtained by calculating the difference between the two. Mean Squared Error (MSE) is typically used as the error metric. The calculation method is as follows:

[0159] ;

[0160] in, This represents the number of wind farms in the wind farm cluster.

[0161] Based on the aforementioned prediction error, a loss function is defined. This loss function quantifies the model's prediction performance and serves as the optimization objective. This loss function is then backpropagated to each layer of the model to calculate its impact on the graph attention network parameters. And the parameters of the target wind farm prediction model gradient:

[0162] ;

[0163] ;

[0164] This process is based on the chain rule, calculating the sensitivity of parameters to loss layer by layer to ensure that gradient information is accurately transmitted to the graph attention network and the wind power prediction model.

[0165] Furthermore, embodiments of the present invention can utilize the calculated gradient information based on a preset learning rate. and Gradient descent is used to update the parameters of the graph attention network and the wind power prediction model:

[0166] ;

[0167] ;

[0168] Where k is the preset number of iterations.

[0169] This invention uses prediction error as a feedback signal to construct a loss function and utilizes the backpropagation algorithm to perform collaborative gradient calculation and update on the graph attention network and the target wind farm prediction model. This enables end-to-end adaptive optimization of the model, enhances the ability to express dynamic spatial topological relationships and spatiotemporal features, and thus significantly improves the accuracy and robustness of cross-wind farm wind power prediction.

[0170] Although the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous.

[0171] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0172] Corresponding to the above method embodiments, this invention also provides a cross-wind farm wind power prediction device, including: a source wind farm group information data acquisition unit, a dynamic spatial topology relationship construction unit, a wind power joint spatiotemporal feature construction unit, a source wind farm prediction model acquisition unit, a target wind farm group historical data acquisition unit, a model adjustment unit, and a model adaptive update unit.

[0173] The source wind farm cluster information data acquisition unit is used to obtain the geographical coordinate information, meteorological observation data and first historical wind power data of the source wind farm cluster.

[0174] The dynamic spatial topology construction unit is used to construct the dynamic spatial topology of the source wind farm group using geographic coordinate information, meteorological observation data, and first historical wind power data.

[0175] The wind power joint spatiotemporal feature construction unit is used to construct the wind power joint spatiotemporal features of the source wind farm group by utilizing dynamic spatial topological relationships and first historical wind power data.

[0176] The source wind farm prediction model acquisition unit is used to train the model using the spatiotemporal characteristics of wind power and the first historical wind power data to obtain the source wind farm prediction model.

[0177] The target wind farm cluster historical data acquisition unit is used to obtain the second historical wind power data of the target wind farm cluster.

[0178] The model adjustment unit is used to adjust the source wind farm prediction model using the second historical wind power data to obtain the target wind farm prediction model. The target wind farm prediction model is used to predict the wind power of the target wind farm group and output wind power prediction data.

[0179] The model adaptive update unit is used to adaptively update the model parameters of the target wind farm prediction model by using the actual wind power data continuously accumulated during the operation of the target wind farm, as well as the matching wind power prediction data, so as to adapt to the changes in the operating status of the target wind farm group.

[0180] Optionally, the dynamic spatial topology relationship construction unit may include: a static spatial distribution feature acquisition subunit, a dynamic correlation feature acquisition subunit, and a dynamic spatial topology relationship generation subunit.

[0181] The static spatial distribution characteristics acquisition sub-unit is used to obtain the static spatial distribution characteristics of the source wind farm group using geographic coordinate information.

[0182] The dynamic correlation feature acquisition sub-unit is used to obtain the dynamic correlation features of the source wind farm group using meteorological observation data and first historical wind power data.

[0183] The dynamic spatial topology generation sub-unit is used to fuse static spatial distribution characteristics with dynamic correlation characteristics to generate the dynamic spatial topology of the source wind farm group.

[0184] Optionally, the static spatial distribution feature acquisition sub-unit can be used to obtain the Euclidean distance between any two wind farms in the source wind farm group; based on the Euclidean distance, spatial similarity is calculated through a normalization function; and based on the spatial similarity between all wind farm pairs, the static spatial distribution feature of the source wind farm group is constructed.

[0185] Optionally, the dynamic spatial topology generation sub-unit can be used to perform a weighted summation of static spatial distribution features and dynamic correlation features using weighted fusion parameters; the result of the weighted summation is then processed by an activation function to generate the dynamic spatial topology of the source wind farm group.

[0186] Optionally, the dynamic correlation feature acquisition sub-unit can be used to align and fuse meteorological observation data and first historical wind power data to construct a node feature vector for each wind farm in the source wind farm group; map the node feature vector to a high-dimensional feature space to obtain a high-dimensional feature representation; based on the high-dimensional feature representation, use a graph attention network to calculate the attention coefficient between any two wind farms; normalize the attention coefficient to obtain the dynamic correlation features of the source wind farm group.

[0187] Optionally, the model adjustment unit can be used to freeze the parameters of the graph convolutional network in the source wind farm prediction model; and to fine-tune the temporal convolutional network, feature fusion parameters, and prediction layer parameters in the source wind farm prediction model using the second historical wind power data, so as to obtain the target wind farm prediction model.

[0188] Optionally, the wind power joint spatiotemporal feature construction unit can be used to extract the spatial topological features of the source wind farm group based on dynamic spatial topological relationships using graph convolutional networks; process the first historical wind power data using temporal convolutional networks to extract the temporal evolution features of the wind power of the source wind farm group; and fuse the spatial topological features and the temporal evolution features of the wind power to generate the joint spatiotemporal features of the wind power of the source wind farm group.

[0189] Optionally, the joint spatiotemporal feature construction unit for wind power can be used to align and splice spatial topological features with wind power temporal evolution features in the feature dimension to obtain fused features; and to reduce and integrate the fused features to obtain joint spatiotemporal features of wind power.

[0190] Optionally, the model adaptive update unit can be used to obtain the actual wind power data continuously accumulated during the operation of the target wind farm group, as well as the matching wind power prediction data; calculate the prediction error between the actual wind power data and the wind power prediction data; use the prediction error as a feedback signal to construct a loss function; use the backpropagation algorithm to calculate the gradient of the graph attention network and the target wind farm prediction model simultaneously using the loss function; and use the gradient descent algorithm to collaboratively update the graph attention network and the target wind farm prediction model based on the calculated gradient.

[0191] Optionally, the model adjustment unit can be used to set an initial fine-tuning learning rate based on the final learning rate of the source wind farm prediction model in the pre-training stage, wherein the initial fine-tuning learning rate is less than the final learning rate; during the fine-tuning iteration of the temporal convolutional network, feature fusion parameters and prediction layer parameters in the source wind farm prediction model, the initial fine-tuning learning rate is exponentially decayed according to a preset number of cycles.

[0192] This invention provides a cross-wind farm wind power prediction device, which is used to: obtain the geographical coordinate information, meteorological observation data, and first historical wind power data of a source wind farm group; construct the dynamic spatial topology of the source wind farm group using the geographical coordinate information, meteorological observation data, and first historical wind power data; construct the joint spatiotemporal characteristics of wind power of the source wind farm group using the dynamic spatial topology and first historical wind power data; train a model using the joint spatiotemporal characteristics of wind power and first historical wind power data to obtain a source wind farm prediction model; obtain second historical wind power data of a target wind farm group; adjust the source wind farm prediction model using the second historical wind power data to obtain a target wind farm prediction model, wherein the target wind farm prediction model is used to predict the wind power of the target wind farm group and output wind power prediction data; and adaptively update the model parameters of the target wind farm prediction model using the actual wind power data continuously accumulated during the operation of the target wind farm and the matching wind power prediction data to adapt to changes in the operating status of the target wind farm group. This invention constructs a dynamic spatial topology by integrating static geographic information and dynamic meteorological information, and further combines historical wind power data to generate joint spatiotemporal features, thereby accurately modeling the spatiotemporal coupling evolution law of wind farm clusters. It utilizes transfer learning and dynamic updates to achieve high-precision and high-stability cross-wind farm power prediction to meet the actual operation requirements of the power system.

[0193] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0194] The cross-wind farm wind power prediction device includes a processor and a memory. The aforementioned source wind farm group information data acquisition unit, dynamic spatial topology relationship construction unit, wind power joint spatiotemporal feature construction unit, source wind farm prediction model acquisition unit, target wind farm group historical data acquisition unit, model adjustment unit, and model adaptive update unit are all stored as program units in the memory. The processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0195] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured. By adjusting kernel parameters, dynamic spatial topological relationships can be constructed by fusing static geographic information and dynamic meteorological information. Furthermore, historical wind power data is combined to generate joint spatiotemporal features, thereby accurately modeling the spatiotemporal coupling evolution of wind farm clusters. Transfer learning and dynamic updates are used to achieve high-precision, high-stability cross-wind farm power prediction to meet the actual operational needs of the power system.

[0196] This invention provides a computer-readable storage medium storing a program that, when executed by a processor, implements the cross-wind farm wind power prediction method.

[0197] This invention provides a processor for running a program, wherein the program executes the cross-wind farm wind power prediction method during runtime.

[0198] This invention provides an electronic device, which includes at least one processor, at least one memory connected to the processor, and a bus; wherein the processor and the memory communicate with each other via the bus; the processor is used to call program instructions in the memory to execute the aforementioned cross-wind farm wind power prediction method. The electronic device in this document can be a server, PC, PAD, mobile phone, etc.

[0199] The present invention also provides a computer program product, which, when executed on an electronic device, is suitable for executing a program that initializes a method for predicting wind power across wind farms.

[0200] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0201] In a typical configuration, an electronic device includes one or more processors (CPUs), memory, and a bus. The electronic device may also include input / output interfaces, network interfaces, etc.

[0202] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, like read-only memory (ROM) or flash RAM, and memory includes at least one memory chip. Memory is an example of computer-readable media.

[0203] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0204] In the description of this invention, it should be understood that if the terms "upper", "lower", "front", "rear", "left" and "right" are used to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the position or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0205] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0206] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0207] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the present invention.

Claims

1. A method for predicting wind power across wind farms, characterized in that, include: Obtain the geographic coordinates, meteorological observation data, and first historical wind power data of the source wind farm cluster; Using the geographic coordinate information, the meteorological observation data, and the first historical wind power data, a dynamic spatial topology relationship of the source wind farm group is constructed; Using the dynamic spatial topology and the first historical wind power data, the joint spatiotemporal characteristics of wind power of the source wind farm group are constructed; The source wind farm prediction model is obtained by using the combined spatiotemporal characteristics of wind power and the first historical wind power data for model training. Obtain the second historical wind power data of the target wind farm group; The source wind farm prediction model is adjusted using the second historical wind power data to obtain the target wind farm prediction model, wherein the target wind farm prediction model is used to predict the wind power of the target wind farm group and output wind power prediction data. By using the actual wind power data continuously accumulated during the operation of the target wind farm, and the matching wind power prediction data, the model parameters of the prediction model for the target wind farm are adaptively updated to adapt to changes in the operating status of the target wind farm group.

2. The method according to claim 1, characterized in that, The step of constructing the dynamic spatial topology of the source wind farm group using the geographic coordinate information, the meteorological observation data, and the first historical wind power data includes: Using the geographic coordinate information, the static spatial distribution characteristics of the source wind farm group are obtained; Using the meteorological observation data and the first historical wind power data, the dynamic correlation characteristics of the source wind farm group are obtained; The static spatial distribution characteristics and the dynamic correlation characteristics are fused to generate the dynamic spatial topology of the source wind farm group.

3. The method according to claim 2, characterized in that, The step of obtaining the static spatial distribution characteristics of the source wind farm group using the geographic coordinate information includes: Obtain the Euclidean distance between any two wind farms in the source wind farm group; Based on the Euclidean distance, spatial similarity is calculated using a normalization function; Based on the spatial similarity between all wind farm pairs, the static spatial distribution characteristics of the source wind farm group are constructed.

4. The method according to claim 2, characterized in that, The step of fusing the static spatial distribution characteristics with the dynamic correlation characteristics to generate the dynamic spatial topology of the source wind farm group includes: The static spatial distribution features and the dynamic correlation features are weighted and summed using weighted fusion parameters. The weighted summation result is processed by an activation function to generate the dynamic spatial topology of the source wind farm group.

5. The method according to claim 2, characterized in that, The step of obtaining the dynamic correlation characteristics of the source wind farm group using the meteorological observation data and the first historical wind power data includes: The meteorological observation data and the first historical wind power data are aligned and fused to construct a node feature vector for each wind farm in the source wind farm group; The node feature vectors are mapped to a high-dimensional feature space to obtain a high-dimensional feature representation; Based on the high-dimensional feature representation, the attention coefficient between any two wind farms is calculated using a graph attention network. The attention coefficients are normalized to obtain the dynamic correlation characteristics of the source wind farm group.

6. The method according to claim 1, characterized in that, The step of adjusting the source wind farm prediction model using the second historical wind power data to obtain the target wind farm prediction model includes: Freeze the parameters of the graph convolutional network in the source wind farm prediction model; Using the second historical wind power data, the temporal convolutional network, feature fusion parameters, and prediction layer parameters in the source wind farm prediction model are fine-tuned to obtain the target wind farm prediction model.

7. The method according to claim 1, characterized in that, The step of constructing the joint spatiotemporal characteristics of wind power in the source wind farm cluster using the dynamic spatial topology and the first historical wind power data includes: Based on the dynamic spatial topology relationship, the spatial topology features of the source wind farm group are extracted using a graph convolutional network; The first historical wind power data is processed using a temporal convolutional network to extract the temporal evolution characteristics of wind power in the source wind farm group; The spatial topological features and the temporal evolution features of wind power are fused to generate the joint spatiotemporal features of wind power of the source wind farm group.

8. The method according to claim 7, characterized in that, The process of fusing the spatial topological features and the temporal evolution features of wind power to generate the joint spatiotemporal features of the source wind farm cluster includes: The spatial topological features and the wind power time-series evolution features are aligned and stitched together along the feature dimension to obtain the fused features; The fusion features are reduced in dimensionality and integrated to obtain the joint spatiotemporal features of wind power.

9. The method according to claim 5, characterized in that, The step of adaptively updating the model parameters of the prediction model for the target wind farm using the actual wind power data continuously accumulated during the operation of the target wind farm and the matching wind power prediction data includes: Obtain the actual wind power data continuously accumulated during the operation of the target wind farm group, and the matching wind power prediction data; Calculate the prediction error between the actual wind power data and the predicted wind power data; The prediction error is used as a feedback signal to construct a loss function; The gradient of the graph attention network and the target wind farm prediction model is calculated simultaneously using the backpropagation algorithm and the loss function. The gradient descent algorithm is used to collaboratively update the graph attention network and the target wind farm prediction model based on the calculated gradient.

10. The method according to claim 6, characterized in that, The step of fine-tuning the temporal convolutional network, feature fusion parameters, and prediction layer parameters in the source wind farm prediction model using the second historical wind power data includes: Based on the final learning rate of the source wind farm prediction model in the pre-training stage, an initial fine-tuning learning rate is set, wherein the initial fine-tuning learning rate is less than the final learning rate; During the fine-tuning iteration of the temporal convolutional network, feature fusion parameters, and prediction layer parameters in the source wind farm prediction model, the initial fine-tuning learning rate is exponentially decayed according to a preset number of cycles.

Images