Graph neural network wind power generation power prediction method based on tree structure guidance

Abstract

The invention discloses a graph neural network wind power generation power prediction method based on tree structure guidance. In order to solve the problem that an existing deep learning model neglects interaction among input variables in a wind power generation prediction task and lacks interpretability, a tree structure is constructed by utilizing a hierarchical relationship of the variables in a wind power generation process, and a tree structure model binarization matrix and an adjacent matrix are combined; obtaining a graph adjacency matrix based on tree structure guidance; inputting the input variable and the tree structure guided graph adjacency matrix into a graph neural network model to predict the wind power generation power; through a key sub-graph extraction module in the graph neural network model, variables which play a key role in predicting the wind power generation power and the connection relation between the variables can be obtained. And quantizing the contribution degree of each process variable to wind power generation power prediction through anti-fact explanation. According to the method, the problems of insufficient reliability and poor interpretability of a prediction model in wind power generation power prediction are effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of wind power generation, and in particular to a wind power generation power prediction method based on a tree structure guided graph neural network. Background Art

[0002] As wind power continues to grow in renewable energy, accurate wind power forecasting is crucial for grid dispatching, energy management, and equipment maintenance. Traditional wind power forecasting methods rely primarily on statistical models (such as autoregressive models and support vector regression) or machine learning methods (such as random forests and ensemble learning). While these methods can fit the changing trends of wind power data to a certain extent, they often struggle to effectively capture the complex linear relationships between multiple variables, such as wind speed, direction, temperature, and air pressure, limiting forecast accuracy.

[0003] In recent years, deep learning-based prediction methods (such as recurrent neural networks and convolutional neural networks) have demonstrated promising performance in wind power forecasting due to their powerful feature extraction capabilities. However, existing deep learning-based prediction methods still face the following key challenges: 1) Inadequate modeling of inter-variable interactions: Traditional deep learning methods typically treat input variables as independent or simply stacked features, failing to explicitly model the physical connections and dynamic couplings between variables in the wind power generation process, limiting the model's predictive performance. 2) Poor model interpretability: Wind power generation requires prediction models to be both highly accurate and interpretable, enabling operators to understand the underlying decision-making. Existing deep learning models are mostly black-box models, lacking quantitative analysis of the contributions of key variables, making them difficult to support subsequent decision-making. To address these issues, previous studies have attempted to introduce graph neural networks to model inter-variable relationships. However, existing methods typically rely on data-driven graph structure learning, ignoring the hierarchical relationships between variables in wind power systems. This results in illogical graph structures or excessive redundant connections. Furthermore, existing methods lack effective mechanisms for extracting and interpreting the substructure of key variables, making it difficult to achieve reliable and transparent predictions in complex scenarios.

[0004] Therefore, there is an urgent need for a wind power prediction method that can combine domain knowledge, explicitly model variable interactions, and have interpretability to improve prediction accuracy and industrial applicability. Summary of the Invention

[0005] To address the technical issues presented in the prior art, the present invention provides a tree-structured graph neural network-based wind power prediction method. This method first constructs a tree structure using the hierarchical relationships of variables in the wind power generation process to explicitly model the coupling relationships between the variables. The tree-structured model's binarized matrix is ​​merged with the adjacency matrix to obtain a tree-structured graph adjacency matrix. The input variables and the tree-structured graph adjacency matrix are then input into a graph neural network model to predict wind power. A key subgraph extraction module within the graph neural network model is used to identify variables critical to predicting wind power and the connections between them. A counterfactual output value is set, and the input test data is perturbed and then input into the graph neural network model so that the output approaches the set value. This allows the model to be interpreted from a counterfactual interpretation perspective. By generating counterfactual samples, the impact of each process variable on the wind power prediction result is quantified, improving the model's interpretability and revealing the model's decision-making mechanism. Furthermore, the variable contribution values ​​obtained from the counterfactual interpretation help identify key variables, support system operation and maintenance, and optimize control. This enhances model transparency and improves the intelligent monitoring and control capabilities of the wind power system. It has important development significance for realizing intelligent operation and maintenance of wind power generation.

[0006] The present invention specifically comprises the following steps: Step 1. Collect variable data during the real-time operation of wind power generation, divide the collected variable data into a training set and a test set, and normalize the data of the training set and the test set; As a preferred embodiment, step 1 is specifically as follows: collecting the real-time operation of wind power generation D dimensional variable data x and 1-dimensional power generation y , divide these data into training set and test set; perform zero-mean normalization on the collected training set and test set.

[0007] Preferably, the variable data include temperature, humidity, dew point, wind speed and wind direction at different heights during the power generation process of the wind power system.

[0008] Step 2. Cluster and stratify the collected variable data, and build a tree structure model based on the hierarchical relationship of the variables. The tree structure model includes V nodes and E The root node of the tree structure represents the entire wind power generation system, the parent nodes in the tree structure are multiple groups of classified variable data, and the leaf nodes of the tree structure are the individual variables in each group of variable data. The connections between the root node and the parent node, and the connections between the parent node and the leaf nodes in the tree structure are represented by a binary matrix; the structural relationship between the variables is calculated using the K-nearest neighbor algorithm to obtain the adjacency matrix between the variables. The binary matrix of the tree structure model is merged with the adjacency matrix to obtain the graph adjacency matrix A guided by the tree structure; Step 3. Align the variable samples x in the training set and test set with the dimensions of the graph adjacency matrix A to obtain the input variable z.

[0009] Furthermore, an aggregation matrix is ​​introduced to align the dimensions of the input variable sample x with the graph adjacency matrix A.

[0010] Step 4. Establish a graph neural network model guided by a tree structure. The graph neural network model includes a graph neural network encoding module, a key subgraph extraction module, a counterfactual explanation module, and a prediction module. The graph neural network encoding module consists of two graph neural networks and a fully connected network. The first layer of the graph neural network is used to extract hidden information from the aggregated input variable z and the tree-structured graph adjacency matrix A. The second layer of the neural network is used to extract deeper hidden information. The final fully connected layer outputs the extracted hidden information to obtain the mean and variance of the obfuscated variable u and the mean and variance of the latent representation variable h. The key subgraph extraction module is used to calculate the probability of edges in the key subgraph. The key subgraph extraction module consists of two fully connected networks, where the first layer of the fully connected network is used to extract the hidden information of u, h and A, and the second layer of the fully connected network is used to output the auxiliary variable B; The counterfactual explanation module consists of two fully connected networks, where the first fully connected network is used to extract the hidden information in z and u, and the second fully connected network is used to output the counterfactual potential representation variable h cf The function of the counterfactual explanation module is to generate a counterfactual latent representation h different from the original latent representation variable h according to the confusion variable u and the input variable z. cf , which is used by the model to subsequently calculate the contribution value of the counterfactual explanation.

[0011] The prediction module includes a graph neural network, a long short-term memory neural network and a fully connected layer, where the graph neural network is used to extract u, h and key subgraph A IB The long short-term memory neural network is used to perform time series modeling on the hidden information, and the last fully connected layer is used to output the final hidden information as the predicted value of wind power generation.

[0012] Step 5. Input the dimensionally aligned data in the training set into the graph neural network model constructed in step 4 and use the loss function to train the model parameters to obtain a trained graph neural network model.

[0013] As a preference, the following loss function is used to train the model: ; in is the regularization parameter, is the distribution of u output by the graph neural network encoder, is the distribution of h output by the graph neural network encoder. The distribution generated by the counterfactual explanation module, is the prior distribution of u, which is the standard normal distribution. represents the expectation, KL(·||·) represents the KL divergence between the two distributions, are the parameters of the graph neural network encoder, Parameters of the counterfactual explanation, key subgraph extraction, and prediction modules.

[0014] Step 6. Input the dimensionally aligned test data and the tree-based adjacency matrix into the trained graph neural network model. Perform counterfactual interpretation on the graph neural network model and use this to obtain variable contribution values. Use this counterfactual interpretation to verify the model's performance.

[0015] Furthermore, the specific steps of the counterfactual interpretation process described in step 6 are as follows: Step 6.1. After the graph neural network model is trained, the test data is input into the graph neural network model. The graph neural network model uses the key subgraph extraction module to obtain the key subgraphs and key variables that are most relevant to the power prediction value to achieve model explanation of feature attribution.

[0016] Step 6.2. Explain the graph neural network model using counterfactual explanations. Set a counterfactual output and optimize the input variable z so that the model's output approximates the set counterfactual output. The greater the difference between the optimized variable and the original variable, the greater the variable's contribution to the prediction. Quantify the contribution of each variable based on the difference between the counterfactual explanation and the original sample.

[0017] Step 7. Align the variable data in any wind power generation process using the same method as step 3, input it into the model verified in step 6, and output the predicted wind power value.

[0018] The present invention uses a tree-structured graph neural network to explicitly model variable interactions, and combines information bottlenecks with counterfactual explanation mechanisms to effectively solve the problems of insufficient reliability and poor interpretability of prediction models in wind power generation forecasting.

[0019] Through a tree-structured graph neural network, the hierarchical structure of process variables in a wind power generation system is explicitly embedded into the graph adjacency matrix, effectively modeling the hierarchical dependencies and physical coupling relationships between variables, enhancing the realism and expressiveness of the graph representation. Furthermore, by leveraging graph information bottleneck techniques, key subtrees and subgraph structures highly relevant to power prediction are automatically mined, compressing information redundancy and effectively focusing on predictive elements, significantly improving the model's prediction accuracy and generalization. Furthermore, by visualizing key subgraphs, the inherent coupling mechanisms between variables are revealed, and a counterfactual explanation mechanism is introduced to quantify the marginal contribution of each variable to the prediction results, making the model output not only usable but also interpretable and traceable. This results in a complete technical closed loop: "hierarchical variable integration - key feature extraction - target prediction - mechanism explanation and control recommendations." This addresses the common problems of existing deep learning models in industrial forecasting tasks, such as insufficient modeling of variable interactions and poor interpretability. It also provides an intelligent modeling approach for wind power prediction that combines high accuracy, transparency, and reliability. This approach offers a practical and widely applicable solution for intelligent monitoring, optimized operation, and maintenance of wind power generation systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Schematic diagram of the wind power generation variable tree structure in the embodiment; Figure 2 This is a structural diagram of a graph neural network model guided by a tree structure in an embodiment; Figure 3 It is the prediction result diagram in the embodiment; Figure 4 The key subgraph, key subtree and key variable graph in the embodiment; Figure 5 Graph showing the normalized contribution values ​​based on counterfactual explanations in the embodiment. DETAILED DESCRIPTION

[0021] The present invention will be further described below with reference to the accompanying drawings and examples.

[0022] The embodiments of the present invention and their implementation process are as follows: Step 1. In this example, using real wind power generation data as an example, meteorological variables and wind power data are collected as variable data. Meteorological instruments are used to collect temperature, humidity, dew point, and wind energy at different altitudes. Power data is also collected from wind turbines. The data is sampled every one hour, and a total of 5,000 sample points are collected. The first 3,000 sample points are used as the training set, and the last 2,000 sample points are used as the test set. Both training and test samples are zero-mean normalized. A detailed description of the variable data in this example is shown in Table 1. The variable data x collected in this example is 8-dimensional.

[0023] Table 1 Variable numbers and variable descriptions

[0024] Step 2. Cluster and stratify the variable data, and construct the following hierarchical relationships between variables: Figure 1 The tree structure model shown.

[0025] The tree structure model of this embodiment T It includes 8 nodes and 11 edges, where the root node of the tree structure represents the entire wind power generation system , the parent node in the tree structure is the multiple groups of variable data after classification The leaf nodes of the tree structure are the variables in each group of variable data. .

[0026] The connection between the root node and the parent node, and the connection between the parent node and the leaf node in the tree structure are represented by a binary matrix, that is, [A par ,A par-leaf ]; Use K nearest neighbor algorithm to calculate the structural relationship between variables and obtain the adjacency matrix between variables [A leaf ] Merge the binary matrix of the tree structure model with the adjacency matrix to obtain the graph adjacency matrix A guided by the tree structure; is a sparse asymmetric graph adjacency matrix: .

[0027] In the above formula, Represents the connection from the root node to the parent node in the tree structure (wind power generation system to variable data), Represents the connection from parent node to child node in the tree structure (classified variable group to variables in each variable group), Represents the mutual relationship between child nodes (variables) in the tree structure. Obtained through the K-nearest neighbor algorithm.

[0028] Step 3. Align the variable samples x in the training set and test set with the dimensions of the graph adjacency matrix A to obtain the input variable z.

[0029] In this embodiment, dimension alignment is performed by introducing an aggregation matrix, specifically: Introducing an aggregation matrix Propagate and aggregate information in leaf nodes to parent nodes and root nodes.

[0030] Aggregation Matrix , ; in, i yes V The index ofd yes D The index of For the d process variables, for D The identity matrix of dimension , Indicates the The set of leaf nodes under the node, Indicates the The number of leaf nodes under a node.

[0031] Aggregated nodes , where z is the node feature guided by the tree structure, and x is a sample point collected. D The variable x of dimension is expanded to V The dimensional variables z, z are input into the graph neural network together with the tree structure guided graph adjacency matrix A to ensure the consistency of the input dimensions.

[0032] Step 4. Create Figure 2 The graph neural network model shown is based on tree structure guidance; the graph neural network model includes a graph neural network encoding module, a key subgraph extraction module, a counterfactual explanation module and a prediction module.

[0033] The encoding module of the graph neural network consists of two graph neural networks and a fully connected network. The first layer of the graph neural network is used to extract hidden information from the aggregated node variable z and the tree structure guided adjacency matrix A. The second layer of the neural network is used to extract deeper hidden information. The final fully connected layer outputs the extracted hidden information to obtain the mean and variance of the confusion variable u and the mean and variance of the latent representation variable h. The graph neural network encoding module is represented as follows: ; Where GNN is the graph neural network encoder, z and A are the inputs of the graph neural network encoder, as well as are the mean and variance of u and h respectively, u is a confounding variable used for subsequent counterfactual explanation, and h is a potential representation variable; u and h are calculated using the resampling method.

[0034] The key subgraph extraction module consists of two fully connected networks, where the first layer of the fully connected network is used to extract the hidden information of u, h and A, and the second layer of the fully connected network is used to output the auxiliary variable B. The key subgraph extraction module is used to calculate the probability of the edges in the key subgraph. : , ; Among them, FC represents the key subgraph extraction module, is the input of the key subgraph extraction module, B is the auxiliary variable, Compute the key subgraph using continuous probability relaxation of discrete variables; key subgraph ,in is the slack variable and t is the scaling variable.

[0035] The counterfactual explanation module consists of two fully connected networks, where the first fully connected network is used to extract the hidden information in z and u, and the second fully connected network is used to output the counterfactual potential representation variable h cf The function of the counterfactual explanation module is to generate a counterfactual latent representation h different from the original latent representation variable h according to the confusion variable u and the input variable z. cf , which is used by the model to subsequently calculate the contribution value of the counterfactual explanation. The counterfactual explanation module is represented as follows: ,in is the counterfactual latent representation variable, For the counterfactual explanation module.

[0036] The prediction module includes a graph neural network, a long short-term memory neural network and a fully connected layer, where the graph neural network is used to extract u, h and key subgraph A IB The long short-term memory neural network is used to perform time series modeling on the hidden information, and the last fully connected layer is used to output the final hidden information as the predicted value of wind power generation.

[0037] The prediction module is represented as follows: , where Pred represents the prediction module, is the predicted value of power generation.

[0038] Step 5. Input the dimensionally aligned data in the training set into the graph neural network model constructed in step 4 to train the model parameters and obtain a trained graph neural network model. In this embodiment, the following loss function is used to train the model: ; in is the regularization parameter, is the distribution of u output by the graph neural network encoder, is the distribution of h output by the graph neural network encoder. The distribution generated by the counterfactual explanation module, is the prior distribution of u, which is the standard normal distribution. represents the expectation, KL(·||·) represents the KL divergence between the two distributions, are the parameters of the graph neural network encoder, Parameters of the counterfactual explanation, key subgraph extraction, and prediction modules.

[0039] Step 6. Input the dimensionally aligned test data and the tree-guided graph adjacency matrix into the trained graph neural network model. Perform counterfactual interpretations on the graph neural network model and use these interpretations to obtain variable contributions. Use these counterfactual interpretations to verify the model's performance.

[0040] The specific steps of counterfactual explanation are as follows: Step 6.1. After the model is trained, the test data is input into the model. The model uses the key subgraph extraction module to obtain the key subgraphs and key variables that are most relevant to the power prediction value to achieve the model explanation of feature attribution.

[0041] Step 6.2. Explain the model using counterfactual explanations. Set a counterfactual output By optimizing the input variable z, the output of the model can be close to The greater the difference between the optimized variable and the original variable, the greater the role of the variable in the prediction. The optimization objectives of the input variables are as follows: ; in and is the regularization parameter, is the do operator in the do-calculus, for The Laplace matrix of z. Doing the do operation on z is actually inputting z and u into the counterfactual explanation module to obtain , so the do operator can be rewritten into a computable form: ; By optimizing z, we can get a sample of counterfactual explanations The difference between the counterfactual explanation of each variable and the original sample can quantify the contribution of each variable.

[0042] like Figure 3 As shown in the figure, it can be seen that the predicted value of the constructed model is basically consistent with the actual wind power generation power, and the model prediction performance in the test phase is R 2 :0.7145, RMSE:0.1324, MAE:0.1039, which shows that the proposed model has good performance in predicting wind power generation, and verifies that the proposed model can be applied to the field of wind power generation prediction.

[0043] like Figure 4 As shown, the dew point 2m from the ground surface ( ) and the wind speed at 100m above the ground ( ) has little impact on wind power generation prediction. The key subgraphs and key variables extracted by the key subgraph extraction module in the graph neural network model retain the variables that have a great impact on power generation prediction and eliminate the variables that have little impact on power generation prediction ( 、 ); and displays the dependencies between variables through a tree structure. This effectively filters out variables that are not relevant to power generation prediction and retains variables related to wind power generation.

[0044] The quantification of the contribution of counterfactual explanations helps to evaluate the contribution of each variable to the prediction results, identify the dominant factors and secondary influencing factors, and locate the key abnormal variables by the contribution degree when the prediction is biased, thus assisting in fault diagnosis. Figure 5 As shown, it can be seen that the gust wind speed 10m away from the ground ( ) contributes the most to the power generation forecast.

[0045] contrast Figure 4 and Figure 5 It can be seen that the results extracted by the counterfactual explanation module are consistent with the results extracted by the key subgraph module, forming an effective cross-verification. In addition, the counterfactual explanation can also assist in identifying the key variables that affect the prediction error, for example: the counterfactual explanation module identifies The highest contribution to the model's prediction is to improve the gust wind speed 10m above the ground ( ) helps improve the accuracy of wind power forecasts, thereby optimizing overall forecasting performance. In practical applications, when forecast results are abnormal, the variable contributions provided by counterfactual explanations can assist operators in diagnosing deviations and making adjustments, thereby improving the accuracy and robustness of wind power forecast models.

[0046] Step 7. Input the variable data of any wind power generation process into the model verified by step 6, and output the predicted wind power generation value.

[0047] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings, but the above is only a preferred embodiment of the present invention. It should be noted that the technical features involved in the above embodiments and their preferred modes can be reasonably combined and matched into a variety of equivalent schemes by those skilled in the art without departing from or changing the design ideas and technical effects of the present invention; therefore, the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which still do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A wind power generation prediction method based on a tree structure guided graph neural network, characterized by: The specific steps include: Step 1. Collect variable data during the real-time operation of wind power generation, divide the collected variable data into a training set and a test set, and normalize the data in the training set and the test set; Step 2. Cluster and stratify the collected variable data, and build a tree structure model based on the hierarchical relationship of the variables; the tree structure model includes V nodes and E The root node of the tree structure represents the entire wind power generation system, the parent nodes in the tree structure are multiple groups of classified variable data, and the leaf nodes of the tree structure are each variable in each group of variable data; the connection between the root node and the parent node, and the connection between the parent node and the leaf node in the tree structure are represented by a binary matrix; the structural relationship between the variables is calculated using the K nearest neighbor algorithm to obtain the adjacency matrix between the variables; Merge the binary matrix of the tree structure model with the adjacency matrix to obtain the graph adjacency matrix A guided by the tree structure; Step 3. Align the variable samples x in the training set and test set with the dimensions of the graph adjacency matrix A to obtain the input variable z; Step 4. Establish a graph neural network model guided by a tree structure. The graph neural network model includes a graph neural network encoding module, a key subgraph extraction module, a counterfactual explanation module, and a prediction module. The graph neural network encoding module consists of two graph neural networks and a fully connected network. The first layer of the graph neural network is used to extract hidden information from the aggregated input variable z and the tree-structured guided adjacency matrix A. The second layer of the neural network is used to extract deeper hidden information. The final fully connected layer outputs the extracted hidden information to obtain the mean and variance of the obfuscated variable u and the mean and variance of the latent representation variable h. The key subgraph extraction module is used to calculate the probability of edges in the key subgraph. The key subgraph extraction module consists of two fully connected networks, where the first layer of the fully connected network is used to extract the hidden information of u, h and A, and the second layer of the fully connected network is used to output the auxiliary variable B; The counterfactual explanation module consists of two fully connected networks, where the first fully connected network is used to extract the hidden information in z and u, and the second fully connected network is used to output the counterfactual potential representation variable h cf The function of the counterfactual explanation module is to generate a counterfactual latent representation h different from the original latent representation variable h according to the confusion variable u and the input variable z. cf , which is used by the model to subsequently calculate the contribution value of the counterfactual explanation; The prediction module includes a graph neural network, a long short-term memory neural network and a fully connected layer, where the graph neural network is used to extract u, h and key subgraph A IB The long short-term memory neural network is used to perform time series modeling on the hidden information in the image. The last fully connected layer is used to output the final hidden information as the predicted value of wind power generation. Step 5. Input the dimensionally aligned data in the training set into the graph neural network model constructed in step 4 and use the loss function to train the model parameters to obtain a trained graph neural network model. Step 6. Input the dimensionally aligned test data and the tree-based adjacency matrix into the trained graph neural network model, perform counterfactual interpretation on the graph neural network model, and use the counterfactual interpretation to obtain variable contribution values. The performance of the model is verified by referring to the variable contribution values ​​obtained from the counterfactual interpretation. Step 7. Align the variable data in any wind power generation process using the same method as step 3, input it into the model verified in step 6, and output the predicted wind power value.

2. The tree-structure guided graph neural network wind power prediction method according to claim 1, characterized in that: The specific steps of the counterfactual interpretation process described in step 6 are as follows: Step 6.

1. After the GNN model is trained, the test data is input into the GNN model. The GNN model uses the key subgraph extraction module to obtain the key subgraphs and key variables most relevant to the power prediction value to implement the model explanation of feature attribution. Step 6.

2. Explain the graph neural network model using counterfactual explanations. Set a counterfactual output and optimize the input variable z so that the model output can be close to the set counterfactual output; the greater the difference between the optimized variable and the original variable, the greater the role of the variable in the prediction; The contribution of each variable is quantified by the difference between the counterfactual explanation of each variable and the original sample.

3. The tree-structure guided graph neural network wind power prediction method according to claim 1, characterized in that: In step 3, an aggregation matrix is ​​introduced to align the dimensions of the input variable sample x with the graph adjacency matrix A.

4. The tree-structure guided graph neural network wind power prediction method according to claim 1, characterized in that: The step 1 is specifically as follows: collecting the real-time operation of wind power generation D dimensional variable data x and 1-dimensional power generation y , divide these data into training set and test set; perform zero-mean normalization on the collected training set and test set.

5. The tree-structure guided graph neural network wind power prediction method according to claim 1, characterized in that: The variable data include temperature, humidity, dew point, wind speed and wind direction at different heights during the power generation process of the wind power system.

6. The tree-structure guided graph neural network wind power prediction method according to claim 1, characterized in that: The following loss function is used to train the model: ； in is the regularization parameter, is the distribution of u output by the graph neural network encoder, is the distribution of h output by the graph neural network encoder; The distribution generated by the counterfactual explanation module, is the prior distribution of u, which is the standard normal distribution; represents the expectation, KL(·||·) represents the KL divergence between the two distributions, are the parameters of the graph neural network encoder, Parameters of the counterfactual explanation, key subgraph extraction, and prediction modules.

Images