Annual-scale wind power prediction method based on cluster division and attention mechanism
Through the annual-scale wind power power prediction method of cluster division and attention mechanism, the dual graph convolution neural network and cross-modal mutual distillation network are used to extract the spatiotemporal characteristics of the wind farm, combined with LSTM training, the problem of low prediction accuracy on the annual scale of the traditional method is solved, achieving higher prediction accuracy and reliability.
Patent Information
- Application Number
- CN202510594798.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2025-04-15
- Filing Date
- 2025-05-09
- Publication Date
- 2025-08-08
AI Technical Summary
Traditional wind power power prediction methods are difficult to fully explore the spatiotemporal characteristics of the cluster on the annual scale of provincial clusters, resulting in low prediction accuracy.
Using a method based on cluster division and attention mechanism, the spatial and temporal feature vectors of wind farms are extracted through dual graph convolution neural networks, feature fusion is performed by combining cross-modal mutual distillation networks, and training is performed through long-term and short-term memory neural networks to generate an annual-scale wind power prediction model.
It improves the accuracy and reliability of annual-scale wind power power prediction, can deeply utilize spatiotemporal information, maximize the utilization of characteristic differences inside and outside the cluster, and adaptively focus on critical moments and regions.
Smart Images
Figure CN120454050A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of new energy power prediction and relates to an annual-scale wind power prediction method based on cluster division and attention mechanism. Background Art
[0002] As a clean, renewable energy source, wind power has become a vital component of global energy supply. However, wind farm power output is subject to significant volatility and uncertainty, influenced by factors such as weather and wind speed. Accurate wind power forecasting is crucial for power system scheduling, maintaining stability, and improving the economic benefits of wind power.
[0003] Annual wind power forecasting is particularly complex, involving long-term trend analysis and the impact of seasonal fluctuations in wind energy. This is particularly challenging, as it requires not only considering climate change, cyclical changes (such as seasonal winds), and atmospheric patterns, but also processing large amounts of time series data. Traditional wind power forecasting methods, when applied to annual wind power forecasts for provincial clusters, often exhibit limitations, making it difficult to fully exploit the spatiotemporal characteristics of the clusters and resulting in low power forecast accuracy. Summary of the Invention
[0004] The purpose of the present invention is to solve the problem that the traditional wind power prediction method in the above-mentioned background technology usually shows limitations when applied to the annual-scale wind power prediction of provincial clusters, it is difficult to fully explore the spatiotemporal characteristics of the cluster, and the power prediction accuracy is low. An annual-scale wind power prediction method based on cluster division and attention mechanism is provided.
[0005] The purpose of the present invention is achieved by the following technical measures:
[0006] A method for annual wind power forecasting based on clustering and attention mechanism includes the following steps:
[0007] S1: Preprocess the collected data from each wind farm to further improve the data quality of each wind farm through outlier removal, missing value filling, filtering and denoising, and normalization methods;
[0008] S2: The data processed in step S1 is divided into two parts: geographic information of the wind farm and meteorological and power information of the wind farm. The spatial graph and feature graph are constructed respectively and input into the dual graph convolutional neural network (DGCN). Alternating convolution is performed to extract the spatiotemporal feature vectors of each wind farm.
[0009] S3: Perform cross-modal inter-distillation based on the spatiotemporal feature vectors of each wind farm extracted in step S2 to divide the clusters;
[0010] S4: The clustering results and the extracted spatiotemporal feature vectors of each wind farm are input into the enhanced attention mechanism to conduct comparative learning between and within clusters, construct positive and negative samples for each cluster, and enhance the importance of effective features of wind farm clusters and the time series relationship at the annual scale;
[0011] S5: The weights of the enhanced attention mechanism, the clustering results after division, and the extracted spatiotemporal feature vectors of each wind farm are input into the long short-term memory neural network (LSTM) for forward and backward training. The training is stopped after the validation set loss does not decrease after 10 iterations. The final annual-scale wind power prediction model is obtained, and the numerical weather forecast data of each wind farm in the next year is input into the final annual-scale wind power prediction model for prediction to obtain the final annual-scale wind power.
[0012] Furthermore, the specific steps of S1 include:
[0013] S1.1: Use the quartile method to remove abnormal data values from each wind farm and fill in missing values;
[0014] S1.2: The padded data is input into the Kalman filter model for noise reduction;
[0015] S1.3: The noise-reduced data shall be normalized and the processed data shall be checked to ensure data integrity.
[0016] Furthermore, the specific steps of S2 include:
[0017] S2.1: Construct a spatial graph based on the geographic information of the longitude and latitude coordinates and location distances of each wind farm processed in S1.3, with the longitude and latitude coordinates as nodes of the graph and the location distances as edges of the graph;
[0018] S2.2: Construct a feature graph based on the meteorological information and power data of each wind farm, with each meteorological information and power data as the nodes of the graph, and the similarity and information entropy between each feature as the edges of the graph;
[0019] S2.3: Input the spatial graph and feature graph into DGCN to construct the adjacency matrix and node feature matrix;
[0020] S2.4: Alternate convolution is performed on the spatial graph and feature graph. DGCN transfers and aggregates the feature information of nodes through the adjacency matrix of each graph to extract the spatiotemporal feature vector of each wind farm.
[0021] Furthermore, the specific steps of S3 include:
[0022] S3.1: Divide the spatiotemporal eigenvectors extracted by DGCN in S2.4 into annual time series modes and spatial eigenmodes containing wind farm information;
[0023] S3.2: A teacher network and a student network are established based on the annual time series mode and the spatial feature mode, respectively. The student network is trained based on the annual time series data, while the teacher network is trained based on the spatial feature data. Training is stopped after the student network validation set loss does not decrease after three iterations, resulting in a trained student network. Training is stopped after the teacher network validation set loss does not decrease after five iterations, resulting in a trained teacher network.
[0024] S3.3: Concatenate and cluster the features of the two modalities of the trained teacher network and the trained student network using cross-modal mutual distillation to divide the clusters.
[0025] Furthermore, the specific steps of S4 include:
[0026] S4.1: Based on the clustering results of S3.3 and the spatiotemporal feature vectors of each wind farm extracted by DGCN in S2.4, perform intra-cluster self-attention mechanism training and inter-cluster attention mechanism training. Stop training after the validation set loss does not decrease after 5 iterations. Preliminary determine the weights of the self-attention mechanism and the inter-cluster attention mechanism.
[0027] S4.2: Construct cluster positive and negative samples based on cluster division results and feature similarity;
[0028] S4.3: Construct a contrastive loss model for clustered positive and negative samples to enhance cluster learning for positive and negative samples, strengthen the weight of the attention mechanism, and highlight samples and features that have a strong impact on prediction power as well as time series relationships at the annual scale.
[0029] Furthermore, the specific steps of S5 include:
[0030] S5.1: Build an LSTM and determine the hidden layer size, learning rate, loss function, and optimizer parameters.
[0031] S5.2: Input the enhanced attention mechanism weights in S4.3, the clustering results after division, and the extracted spatiotemporal feature vectors of each wind farm into the LSTM;
[0032] S5.3: Perform forward training and backward training separately to enable the LSTM to better understand the changing trends of the spatiotemporal characteristics of wind farms, especially the long-term trend of wind power on an annual scale. Stop training when the validation set loss does not decrease after 10 iterations, and obtain the final annual wind power prediction model.
[0033] S5.4: Input the numerical weather forecast data of each wind farm for the next year into the final annual-scale wind power prediction model to perform prediction and obtain the final annual-scale wind power.
[0034] The beneficial effects achieved by the present invention are: a method for annual-scale wind power prediction based on clustering and attention mechanism, which performs alternating convolution operations on the spatial map and feature map of the wind farm through a dual graph convolutional neural network, deeply exploring the complex spatial relationships and feature information between wind farms; combining cross-modal mutual distillation networks for feature fusion and network, comprehensively considering the geographical location, meteorological characteristics and power data of the wind farm, and rationally dividing the wind farm into multiple clusters; through comparative learning within and between clusters, positive and negative samples are generated, and the training of the enhanced attention mechanism is further strengthened. The model can adaptively focus on the most critical moments and areas for prediction, maximizing the use of feature differences within and outside the cluster. Finally, the model inputs the data into the LSTM for bidirectional training, fully exploring long-term time series features, realizing the deep utilization of spatiotemporal information and optimizing the wind power prediction effect, and improving the accuracy and reliability of predictions at the annual scale. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 It is a structural block diagram of the overall process of the present invention. DETAILED DESCRIPTION
[0036] The following is a detailed description with reference to the accompanying drawings in the embodiments of the present invention. The purpose of the present invention is to provide an annual-scale wind power prediction method based on cluster division and attention mechanism.
[0037] like Figure 1 As shown in FIG, a method for annual wind power prediction based on clustering and attention mechanism includes the following steps:
[0038] S1: Preprocess the collected data from each wind farm to further improve the data quality of each wind farm through outlier removal, missing value filling, filtering and denoising, and normalization methods. The specific steps are as follows:
[0039] S1.1: Use the quartile method to remove abnormal data values from each wind farm and fill in missing values;
[0040] S1.2: The padded data is input into the Kalman filter model for noise reduction;
[0041] S1.3: The noise-reduced data shall be normalized and the processed data shall be checked to ensure data integrity.
[0042] S2: The data processed in step S1 is divided into two parts: geographic information of the wind farm and meteorological and power information of the wind farm. The spatial graph and feature graph are constructed respectively and input into the dual graph convolutional neural network DGCN for alternating convolution to extract the spatiotemporal feature vectors of each wind farm. The specific steps are as follows:
[0043] S2.1: Construct a spatial graph based on the geographic information of the longitude and latitude coordinates and location distances of each wind farm processed in S1.3, with the longitude and latitude coordinates as nodes of the graph and the location distances as edges of the graph;
[0044] S2.2: Construct a feature graph based on the meteorological information and power data of each wind farm, with each meteorological information and power data as the nodes of the graph, and the similarity and information entropy between each feature as the edges of the graph;
[0045] S2.3: Input the spatial graph and feature graph into DGCN to construct the adjacency matrix and node feature matrix;
[0046] S2.4: Alternate convolution is performed on the spatial graph and feature graph. DGCN transfers and aggregates the feature information of nodes through the adjacency matrix of each graph to extract the spatiotemporal feature vector of each wind farm.
[0047] S3: Perform cross-modal inter-distillation based on the spatiotemporal feature vectors of each wind farm extracted in step S2 to divide the clusters. The specific steps are as follows:
[0048] S3.1: Divide the spatiotemporal eigenvectors extracted by DGCN in S2.4 into annual time series modes and spatial eigenmodes containing wind farm information;
[0049] S3.2: A teacher network and a student network are established based on the annual time series mode and the spatial feature mode, respectively. The student network is trained based on the annual time series data, while the teacher network is trained based on the spatial feature data. Training is stopped after the student network validation set loss does not decrease after three iterations, resulting in a trained student network. Training is stopped after the teacher network validation set loss does not decrease after five iterations, resulting in a trained teacher network.
[0050] S3.3: Concatenate and cluster the features of the two modalities of the trained teacher network and the trained student network using cross-modal mutual distillation to divide the clusters.
[0051] S4: The clustering results and the extracted spatiotemporal feature vectors of each wind farm are input into the enhanced attention mechanism to conduct comparative learning between and within clusters, construct positive and negative samples for each cluster, and enhance the importance of effective features of wind farm clusters and the time series relationship at the annual scale. The specific steps are as follows:
[0052] S4.1: Based on the clustering results of S3.3 and the spatiotemporal feature vectors of each wind farm extracted by DGCN in S2.4, perform intra-cluster self-attention mechanism training and inter-cluster attention mechanism training. Stop training after the validation set loss does not decrease after 5 iterations. Preliminary determine the weights of the self-attention mechanism and the inter-cluster attention mechanism.
[0053] S4.2: Construct cluster positive and negative samples based on cluster division results and feature similarity;
[0054] S4.3: Construct a contrastive loss model for clustered positive and negative samples to enhance cluster learning for positive and negative samples, strengthen the weight of the attention mechanism, and highlight samples and features that have a strong impact on prediction power as well as time series relationships at the annual scale.
[0055] S5: The weights of the enhanced attention mechanism, the clustering results after division, and the extracted spatiotemporal feature vectors of each wind farm are input into the long short-term memory neural network (LSTM) for forward and backward training. The training is stopped after the validation set loss does not decrease after 10 iterations. The final annual-scale wind power prediction model is obtained. The numerical weather forecast data of each wind farm in the next year is input into the final annual-scale wind power prediction model for prediction to obtain the final annual-scale wind power. The specific steps are as follows:
[0056] S5.1: Build an LSTM and determine the hidden layer size, learning rate, loss function, and optimizer parameters.
[0057] S5.2: Input the enhanced attention mechanism weights in S4.3, the clustering results after division, and the extracted spatiotemporal feature vectors of each wind farm into the LSTM;
[0058] S5.3: Perform forward training and backward training separately to enable the LSTM to better understand the changing trends of the spatiotemporal characteristics of wind farms, especially the long-term trend of wind power on an annual scale. Stop training when the validation set loss does not decrease after 10 iterations, and obtain the final annual wind power prediction model.
[0059] S5.4: Input the numerical weather forecast data of each wind farm for the next year into the final annual-scale wind power prediction model to perform prediction and obtain the final annual-scale wind power.
Claims
1. Annual scale wind power forecasting method based on clustering and attention mechanism, characterized by Follow these steps: S1: Preprocess the collected data from each wind farm to further improve the data quality of each wind farm through outlier removal, missing value filling, filtering and denoising, and normalization methods; S2: The data processed in step S1 is divided into two parts: geographic information of the wind farm and meteorological and power information of the wind farm. The spatial graph and feature graph are constructed respectively and input into the dual graph convolutional neural network (DGCN). Alternating convolution is performed to extract the spatiotemporal feature vectors of each wind farm. S3: Perform cross-modal inter-distillation based on the spatiotemporal feature vectors of each wind farm extracted in step S2 to divide the clusters; S4: The clustering results and the extracted spatiotemporal feature vectors of each wind farm are input into the enhanced attention mechanism to conduct comparative learning between and within clusters, construct positive and negative samples for each cluster, and enhance the importance of effective features of wind farm clusters and the time series relationship at the annual scale; S5: The weights of the enhanced attention mechanism, the clustering results after division, and the extracted spatiotemporal feature vectors of each wind farm are input into the long short-term memory neural network (LSTM) for forward and backward training. The training is stopped after the validation set loss does not decrease after 10 iterations. The final annual-scale wind power prediction model is obtained, and the numerical weather forecast data of each wind farm in the next year is input into the final annual-scale wind power prediction model for prediction to obtain the final annual-scale wind power.
2. The annual wind power forecasting method based on clustering and attention mechanism according to claim 1 is characterized in that: In the step S1, the specific steps include: S1.1: Use the quartile method to remove abnormal data values from each wind farm and fill in missing values; S1.2: The padded data is input into the Kalman filter model for noise reduction; S1.3: The noise-reduced data shall be normalized and the processed data shall be checked to ensure data integrity.
3. The annual wind power forecasting method based on clustering and attention mechanism according to claim 2 is characterized in that: In the step S2, the specific steps include: S2.1: Construct a spatial graph based on the geographic information of the longitude and latitude coordinates and location distances of each wind farm processed in S1.3, with the longitude and latitude coordinates as nodes of the graph and the location distances as edges of the graph; S2.2: Construct a feature graph based on the meteorological information and power data of each wind farm, with each meteorological information and power data as the nodes of the graph, and the similarity and information entropy between each feature as the edges of the graph; S2.3: Input the spatial graph and feature graph into DGCN to construct the adjacency matrix and node feature matrix; S2.4: Alternate convolution is performed on the spatial graph and feature graph. DGCN transfers and aggregates the feature information of nodes through the adjacency matrix of each graph to extract the spatiotemporal feature vector of each wind farm.
4. The annual wind power forecasting method based on clustering and attention mechanism according to claim 3 is characterized in that: In the step S3, the specific steps include: S3.1: Divide the spatiotemporal eigenvectors extracted by DGCN in S2.4 into annual time series modes and spatial eigenmodes containing wind farm information; S3.2: A teacher network and a student network are established based on the annual time series mode and the spatial feature mode, respectively. The student network is trained based on the annual time series data, while the teacher network is trained based on the spatial feature data. Training is stopped after the student network validation set loss does not decrease after three iterations, resulting in a trained student network. Training is stopped after the teacher network validation set loss does not decrease after five iterations, resulting in a trained teacher network. S3.3: Concatenate and cluster the features of the two modalities of the trained teacher network and the trained student network using cross-modal mutual distillation to divide the clusters.
5. The annual wind power forecasting method based on clustering and attention mechanism according to claim 4 is characterized in that: In the step S4, the specific steps include: S4.1: Based on the clustering results of S3.3 and the spatiotemporal feature vectors of each wind farm extracted by DGCN in S2.4, perform intra-cluster self-attention mechanism training and inter-cluster attention mechanism training. Stop training after the validation set loss does not decrease after 5 iterations. Preliminary determine the weights of the self-attention mechanism and the inter-cluster attention mechanism. S4.2: Construct cluster positive and negative samples based on cluster division results and feature similarity; S4.3: Construct a contrastive loss model for clustered positive and negative samples to enhance cluster learning for positive and negative samples, strengthen the weight of the attention mechanism, and highlight samples and features that have a strong impact on prediction power as well as time series relationships at the annual scale.
6. The annual scale wind power forecasting method based on clustering and attention mechanism according to claim 5 is characterized in that: In the step S5, the specific steps include: S5.1: Build an LSTM and determine the hidden layer size, learning rate, loss function, and optimizer parameters. S5.2: Input the enhanced attention mechanism weights in S4.3, the clustering results after division, and the extracted spatiotemporal feature vectors of each wind farm into the LSTM; S5.3: Perform forward training and backward training separately to enable the LSTM to better understand the changing trends of the spatiotemporal characteristics of wind farms, especially the long-term trend of wind power on an annual scale. Stop training when the validation set loss does not decrease after 10 iterations, and obtain the final annual wind power prediction model. S5.4: Input the numerical weather forecast data of each wind farm for the next year into the final annual-scale wind power prediction model to perform prediction and obtain the final annual-scale wind power.