A multi-wind farm multi-feature wind speed prediction method based on graph embedding and GIN-GRU

Abstract

The application relates to the technical field of wind speed prediction, and discloses a wind speed prediction method for multiple wind farms and multiple features based on graph embedding and GIN-GRU. First, on the spatial level, the overall wind speed features in a wind farm at a moment are extracted through graph embedding and Mahalanobis distance to build a network graph, then the network graph is input into GIN, multiple graphs are input in time steps one by one, and a spatial convolution layer is used to aggregate the neighbor information of nodes, finally, the output of GIN is input into GRU to aggregate the information at the current moment and the historical moment, the representation of each node is fused with the adjacent information at the current moment and the historical moment, and the output of GRU is a time sequence prediction result which is collected from the mutual relationship among wind turbines. The application utilizes the combined advantages of graph embedding and GIN-GRU, can more comprehensively understand and predict the wind speed change under the multiple wind farm environment, and provides a more accurate and efficient tool for the management of wind farms and the dispatching of power grids.

Description

Technical Field

[0001] This invention relates to the field of wind speed prediction technology, specifically to a wind speed prediction method based on graph embedding and GIN-GRU for multiple wind farms and multiple features. Background Technology

[0002] In the renewable energy sector, especially wind power generation, accurate wind speed forecasting is crucial for stable grid operation and energy management. Traditional wind speed forecasting methods mainly rely on physical models or statistical methods, but these methods face challenges such as limited data processing capabilities and insufficient forecast accuracy. With the development of machine learning and deep learning technologies, data-driven wind speed forecasting methods have gradually emerged, providing new solutions.

[0003] Known technologies, such as machine learning methods like Support Vector Machines (SVM), Random Forests, and Extreme Learning Machines (ELM), as well as deep learning methods like Backpropagation (BP), Recurrent Neural Networks (RNN), and Long Short-Term Memory Networks (LSTM), have been used for wind speed prediction and have achieved certain results. However, these methods still have limitations when dealing with nonlinear and non-stationary wind speed data. To overcome these limitations, a novel wind speed prediction method is proposed, which combines graph embedding and graph neural networks (such as GIN-GRU, i.e., graph isomorphic network and gated recurrent unit). This method utilizes graph embedding to capture the spatial relationships between wind farms, while GIN-GRU can effectively process time-series data, improving the accuracy of wind speed prediction. Through this multi-wind-farm, multi-feature wind speed prediction method, wind speed can be better understood and predicted, thereby optimizing wind farm operation and grid energy allocation. Summary of the Invention

[0004] The purpose of this invention is to address the limitations of existing wind speed prediction methods and to propose a wind speed prediction method based on graph embedding and GIN-GRU for multiple wind farm features.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] A wind speed prediction method based on graph embedding and GIN-GRU for multiple wind farm features includes the following two components:

[0007] S1. Construct a network graph G′=(V,E′); S2. Based on the network graph G′=(V,E′), build a spatiotemporal prediction model based on graph GIN and GRU;

[0008] S1 includes the following steps:

[0009] S1.1.1 Extraction of primary and secondary features based on random forest;

[0010] S1.1.2, Dimensionality reduction of secondary features based on PCA;

[0011] S1.1.3. The main features and the reduced secondary features are concatenated to form the dataset for building the network graph G = (V, E);

[0012] S1.2.1 Construct the original network graph G = (V, E) using the dataset described in S1.1.3;

[0013] S1.2.2 Graph embedding is performed using a deep learning method based on GraphSAGE;

[0014] S1.2.3. Then, based on the node embedding vector representation of Mahalanobis distance, a new network graph G′=(V,E′) is established;

[0015] S2 includes the following steps:

[0016] S2.1, Constructing a framework for GIN based on the network graph G′=(V,E′);

[0017] S2.2, Substitute the output of GIN into the recurrent neural network GRU as the input of GRU;

[0018] S2.3 The output of the recurrent neural network GRU is the time series prediction result of the relationship between wind turbines.

[0019] Based on the above technical solution, the present invention can be further improved as follows.

[0020] Preferably, step S1.1.1, the extraction of primary and secondary features based on random forest, includes the following steps:

[0021] The Gini importance index is used to extract the main features of the random forest. The calculation of Gini importance is first performed by constructing a random forest from each decision tree trained on the training set sampled by Bootstrap. Then, during the construction of each tree, a subset of features obtained by cross-validation is randomly selected from all features to find the best branch features to optimize the model performance according to the principle of minimizing the Gini index. Finally, unbiased estimation is performed from the unsampled data, and the importance of each feature is calculated using the Gini index. The features are sorted in descending order to select the main and secondary features.

[0022] Preferably, S1.1.2, PCA-based secondary feature dimensionality reduction, includes the following steps:

[0023] By calculating the eigenvalues ​​and eigenvectors based on the covariance matrix, and then reducing the dimensionality of the corresponding matrices of the secondary features, we can obtain the meteorological elements among the secondary features that can significantly affect wind speed changes.

[0024] Preferably, step S1.1.3, concatenating the main features and the dimensionality-reduced secondary features to construct the dataset for the network graph G = (V, E), includes the following steps:

[0025] 1) Calculate the Gini index. The dataset has j features (X1, X2, X3, ..., X...). j Given C categories and I decision trees, the Gini index of node m is:

[0026]

[0027] Where: G m Let m represent the Gini index of node m. This represents the probability estimate that sample m at node c belongs to class c.

[0028] 2) Calculate the change in the Gini index, characteristic quantity X. j Importance score of node m The change in the Gini index before and after branching at node m is used to represent the value:

[0029]

[0030] Where: G l and G r These are the Gini indices of the two new nodes l and r after node m branches;

[0031] 3) Calculate the importance of the i-th tree and define the feature value X. j If the set of nodes appearing in the i-th tree is M, then the feature quantity X j The importance of the i-th tree is represented as

[0032]

[0033] In the formula: Representing the characteristic quantity X j The importance of the i-th tree;

[0034] 4) Calculate the final importance, feature X j Its importance in RF can be expressed as

[0035]

[0036] In the formula: Representing the characteristic quantity X j Its importance in RF;

[0037] 5) Based on the importance of the features, we select the top k features according to the importance score gradient to obtain the feature matrix K after random forest filtering. The matrix N of the remaining jk features is reduced by PCA to retain the principal components.

[0038] 6) Decentralize the data by centering the dimensionality-reduced matrix N, that is, subtract the mean of each feature so that the mean of each feature is 0;

[0039] 7) Calculate the covariance matrix. Perform matrix multiplication on the centered data to obtain a covariance matrix C, which represents the correlation between each feature.

[0040]

[0041] In the formula: C represents the correlation between each feature;

[0042] 8) Calculate eigenvalues ​​and eigenvectors. Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. If vector v is an eigenvector of matrix C, then it satisfies the eigenvalue calculation formula, where λ is the eigenvalue corresponding to eigenvector v, and the eigenvector represents the projection direction of each feature in the new feature space. Calculate the principal component influence factors based on the calculated eigenvalues. Define the cumulative contribution rate ≥ 85% as the selected b (b < m) principal component factors. The eigenvalue calculation formula is as follows:

[0043] Cv=λ (6);

[0044] 9) Select principal components, sort them in descending order of eigenvalues, and select the first p eigenvectors as the principal component matrix P, where p is the dimension of the new feature space, that is, the dimension after dimensionality reduction.

[0045] 10) Dimensionality reduction: Multiply matrix X and transformation matrix P to obtain the dimension-reduced matrix Y.

[0046] Y = PN (7)

[0047] The K and Y matrices are concatenated to obtain a new matrix W1. This eliminates the adverse effects of redundant factors on prediction accuracy while retaining the contribution of the main features to wind speed prediction. Furthermore, to facilitate the study of relationships between multiple wind turbines, the wind speed characteristics of the target wind turbine are used as labels for random forest selection. The features of the remaining n-1 wind turbines are processed using the methods described in 1)-10) to obtain matrices W2, W3...W n The matrix W obtained by concatenating the matrices is used as the dataset for building the network graph.

[0048] Preferably, step S1.2.1, constructing the original network graph G = (V, E) using the dataset described in S1.1.3, includes the following steps:

[0049] The matrix W, after feature processing using S1.1.3, is used to construct a network graph based on data correlation as input for graph embedding, resulting in the original network graph G = (V, E), where V = {v1, v2, v3, ... v...}. n} represents a set of nodes, whose node feature matrix is ​​matrix W. T The dimension is N x D, where N is the number of nodes and D is the dimension of the features of each node. It is a set of edges, representing the connections between nodes. For a node v in the network... i and v j If there is a relationship between them, then there exists an edge represented by e. ij =1 means that e ij A value of 0 indicates that the nodes are not connected. The correlation between the N nodes in the original network graph is expressed using the Pearson correlation coefficient.

[0050]

[0051] In the formula: δ is the correlation coefficient, x i For node v i The i-th dimension features For node v i The average value of features across all dimensions, y i For node v j The i-th dimension features For node v j The average value of features across all dimensions.

[0052] Preferably, step S1.2.2, performing graph embedding using a deep learning method based on GraphSAGE, includes the following steps:

[0053] Sampling: For each target node v in the node set V, randomly sample S from its neighboring node set. k Each node forms a subgraph;

[0054] Aggregation: For each target node v in the node set V, use an aggregation function AGGREGATE. k , and the feature vectors of its neighboring nodes Compression and transformation are performed to obtain a new feature vector. Then, by using an aggregation function that calculates the average of the feature values ​​of neighboring nodes, the unordered neighboring features are transformed into a fixed-dimensional vector, which can then be concatenated with the features of the target node. The aggregation formula is as follows:

[0055]

[0056] In the formula: Let σ represent the embedding vector of node ν at layer k, and let Q represent the activation function. k This represents a learnable weight matrix, CONCAT represents the concatenation operation, and AGGREGATE represents the AGGREGATE operation. k Let N(v) represent the aggregation function of the k-th layer, and let N(v) represent the set of neighbors of node v.

[0057] Update: For each target node v in the node set V, its own feature vector and aggregated feature vectors The vectors are concatenated or added together, then passed through a fully connected layer and an activation function to obtain the embedding vector of the target node. This allows for the fusion and nonlinear transformation of features of each target node v and its neighboring nodes. The updated formula is as follows:

[0058]

[0059] In the formula: COMBINE represents the concatenation or addition operation, Q k+1 This represents the learnable weight matrix.

[0060] Preferably, step S1.2.3, which involves using the node embedding vector representation based on Mahalanobis distance to establish a new network graph G′=(V,E′), includes the following steps:

[0061] The Mahalanobis distance method is used to optimize node embedding vectors. This method is suitable for high-dimensional data with a large number of meteorological factors and can adapt to situations where the data is not independent and identically distributed across dimensions. The calculation formula is as follows:

[0062]

[0063] in, This is the node embedding vector obtained above, μ is the mean vector of the sample distribution, and ∑ is the covariance matrix of the sample distribution. -1 It is the inverse of the covariance matrix, and then d is calculated based on each node. mahal The magnitude and correlation coefficient σ are used to determine the connection relationship of the optimized edges in the original network graph G=(V,E). Make modifications.

[0064]

[0065] This leads to the creation of a new network graph G′=(V,E′).

[0066] Preferably, step S2.1, constructing the GIN framework based on the network graph G′=(V,E′), includes the following steps:

[0067] First, we utilize GIN aggregation. To ensure that different inputs are mapped to different outputs, we inject during modeling, using summation as the aggregation function. A multilayer perceptron can simulate the combination of functions. The node representations of the GIN are updated using the following method:

[0068]

[0069] In the formula: Let N(v) be the feature representation of node (v) at level (k), and N(v) be the set of neighboring nodes of (v), ∈ (k) It is a trainable parameter learned at layer (k), which allows the model to adaptively adjust the importance of the center node features and the neighbor features;

[0070] Secondly, GIN is used to read out the graph and aggregate the feature vectors of all nodes in the new network graph G′=(V,E′) into a single graph-level feature vector for graph classification or other graph-level tasks. The CONCAT method is used to sum the node features in each iteration as the graph features of that iteration, and then the graph features of each iteration are concatenated to obtain the final graph features.

[0071]

[0072] Here, h G This is the representation of the new network graph G′=(V,E′). G is the feature representation of node (v) at level (k), and G is the set of all nodes in the graph.

[0073] Preferably, step S2.2, using the output of GIN as the input to the recurrent neural network GRU, includes the following steps:

[0074] S2.2.1 Calculate the values ​​of the update gate and the reset gate. The update gate controls how much of the hidden state from the previous time step is retained in the current time step, and the reset gate controls how much of the hidden state from the previous time step is included in the calculation of the current time step. The formula for this step is:

[0075]

[0076] Among them, z t and r t These are the values ​​for updating the gate and resetting the gate, respectively. σ is the sigmoid activation function used to facilitate gradient calculation and backpropagation. W z and W r It is the weight matrix, b z and b r It is the bias vector, b r It is the hidden state from the previous moment, x t It is the input at the current moment;

[0077] S2.2.2 Calculate the candidate hidden state, which is the output of the current input and the reset hidden state from the previous time step after passing through a tanh activation function. The formula for this step is:

[0078]

[0079] Among them, h t-1 It is a candidate hidden state, W h It is the weight matrix, b h It is the bias vector, and ⊙ is the element-wise multiplication operation;

[0080] S2.2.3 Calculate the hidden state at the current time step, which is the weighted average of the hidden state at the previous time step and the candidate hidden states. The weights are determined by the update gate. The formula for this step is:

[0081]

[0082] Among them, h t It is the hidden state at the current moment, z t It updates the value of the gate. It is a candidate hidden state.

[0083] Compared with the prior art, the technical solution of this application has the following beneficial technical effects:

[0084] 1. This invention uses RF-PCA technology to extract primary features and reduce the dimensionality of secondary features to reduce the dimensionality of the wind speed prediction dataset, thereby reducing the computational and storage overhead of the model and improving interpretability.

[0085] 2. This invention uses graph embedding technology to better capture the spatial relationships between wind farms, and GIN-GRU combines the advantages of graph neural networks and recurrent neural networks to effectively process time series data, thereby improving the accuracy of wind speed prediction.

[0086] 3. This invention can transform the complex topology of wind farms into a low-dimensional representation through graph embedding technology, which helps the model learn and understand the complex spatial dependencies between wind farms.

[0087] 4. This invention uses the GIN-GRU model to dynamically learn the changing trends and patterns of wind speed data, which is particularly effective for short-term and ultra-short-term wind speed prediction.

[0088] 5. This invention can adapt to the specific characteristics and conditions of different wind farms, and has strong adaptability and generalization ability for wind speed prediction in multi-wind farm environments.

[0089] 6. Compared with other deep learning models, such as LSTM, GRU has a simpler structure and fewer parameters, thus reducing computational complexity and improving computational efficiency during training and prediction.

[0090] Overall, this method leverages the combined advantages of graph embedding and GIN-GRU to more comprehensively understand and predict wind speed changes in multi-wind farm environments, providing a more accurate and efficient tool for wind farm management and grid scheduling. Attached Figure Description

[0091] Figure 1 This is a flowchart of the feature selection process of the present invention;

[0092] Figure 2 The flowchart is embedded in the diagram of this invention;

[0093] Figure 3 A network graph flowchart is constructed for this invention;

[0094] Figure 4 This is a visualization of the gated recurrent neural network of the present invention;

[0095] Figure 5 This is a flowchart of the isomorphic recurrent neural network of the present invention;

[0096] Figure 6 The Gini importance of the top ten features that contribute the most to the other multiple wind turbine features, with the wind turbine feature number 0 of this invention being a height of 10m.

[0097] Figure 7 This invention provides a comparison chart of the wind speed prediction curve of the isomorphic gated circulation network at a height of 10m for wind turbine No. 0, the wind speed prediction curve of the correlation height network, and the actual wind speed curve.

[0098] Figure 8 This invention provides a comparison chart of the wind speed prediction curve of the isomorphic gated circulation network at a height of 30m for wind turbine No. 0, the wind speed prediction curve of the correlation height network, and the actual wind speed curve.

[0099] Figure 9 This is a comparison chart of the wind speed prediction curve and the actual wind speed curve of the isomorphic gated circulation network at a height of 10m for wind turbine No. 0 in this invention.

[0100] Figure 10 This is a comparison chart of the wind speed prediction curve and the actual wind speed curve of the isomorphic gated circulation network at a height of 30m for wind turbine No. 0 in this invention.

[0101] Figure 11 This is a comparison of the wind speed prediction curve of the wind turbine at a height of 10m (numbered 0 in this invention) using a gated recurrent network with other deep learning models.

[0102] Figure 12 This is a comparison of the wind speed prediction curve of the wind turbine at a height of 30m (numbered 0 in this invention) using a gated recurrent network with other deep learning models. Detailed Implementation

[0103] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0104] In the embodiments, by Figure 1-12 This paper presents a wind speed prediction method based on graph embedding and GIN-GRU for multiple wind farm features, comprising the following two parts:

[0105] S1. Construct a network graph G′=(V,E′); S2. Based on the network graph G′=(V,E′), build a spatiotemporal prediction model based on graph GIN and GRU;

[0106] S1 includes the following steps:

[0107] S1.1.1 Extraction of primary and secondary features based on random forest;

[0108] S1.1.2, Dimensionality reduction of secondary features based on PCA;

[0109] S1.1.3. The main features and the reduced secondary features are concatenated to form the dataset for building the network graph G = (V, E);

[0110] S1.2.1 Construct the original network graph G = (V, E) using the dataset described in S1.1.3;

[0111] S1.2.2 Graph embedding is performed using a deep learning method based on GraphSAGE;

[0112] S1.2.3. Then, based on the node embedding vector representation of Mahalanobis distance, a new network graph G′=(V,E′) is established;

[0113] S2 includes the following steps:

[0114] S2.1, Constructing a framework for GIN based on the network graph G′=(V,E′);

[0115] S2.2, Substitute the output of GIN into the recurrent neural network GRU as the input of GRU;

[0116] S2.3 The output of the recurrent neural network GRU is the time series prediction result of the relationship between wind turbines.

[0117] The step S1.1.1, extraction of primary and secondary features based on random forest, includes the following steps:

[0118] The Gini importance index is used to extract the main features of the random forest. The higher the Gini importance, the more critical the feature is to the model's predictive ability. The calculation of Gini importance is as follows: First, the random forest is constructed by training each decision tree on the training set sampled by Bootstrap. Second, during the construction of each tree, a subset of features obtained by cross-validation is randomly selected from all features, and the optimal branch feature is found according to the principle of minimizing the Gini index to optimize the model performance. Finally, unbiased estimation is performed using unsampled data, and the importance of each feature is calculated using the Gini index. The features are sorted in descending order to select the main and secondary features.

[0119] S1.1.2, PCA-based secondary feature dimensionality reduction, includes the following steps:

[0120] The eigenvalues ​​and eigenvectors are calculated based on the covariance matrix. The eigenvalues ​​can reflect the main information contained in the original multiple variables, that is, the key information in the secondary features. Then, the corresponding matrices of the secondary features are dimensionality reduced to obtain the characteristic meteorological elements that can significantly affect wind speed changes.

[0121] S1.1.3, concatenating the main features and the dimensionality-reduced secondary features to construct the dataset for the network graph G = (V, E) includes the following steps:

[0122] 1) Calculate the Gini index. The dataset has j features (X1, X2, X3, ..., X...). j Given C categories and I decision trees, the Gini index of node m is:

[0123]

[0124] Where: G m Let m represent the Gini index of node m. This represents the probability estimate that sample m at node c belongs to class c.

[0125] 2) Calculate the change in the Gini index, characteristic quantity X. j Importance score of node m The change in the Gini index before and after branching at node m is used to represent the value:

[0126]

[0127] Where: G l and G r These are the Gini indices of the two new nodes L and r after node m branches;

[0128] 3) Calculate the importance of the i-th tree and define the feature value X. j If the set of nodes appearing in the i-th tree is M, then the feature quantity Xj The importance of the i-th tree is represented as

[0129]

[0130] In the formula: Representing the characteristic quantity X j The importance of the i-th tree;

[0131] 4) Calculate the final importance, feature X j Its importance in RF can be expressed as

[0132]

[0133] In the formula: Representing the characteristic quantity X j Its importance in RF;

[0134] 5) Based on the importance of the features, we select the top k features according to the importance score gradient to obtain the feature matrix K after random forest filtering. The matrix N of the remaining jk features is reduced by PCA to retain the principal components.

[0135] 6) Decentralize the data by centering the dimensionality-reduced matrix N, that is, subtract the mean of each feature so that the mean of each feature is 0;

[0136] 7) Calculate the covariance matrix. Perform matrix multiplication on the centered data to obtain a covariance matrix C, which represents the correlation between each feature.

[0137]

[0138] In the formula: C represents the correlation between each feature;

[0139] 8) Calculate eigenvalues ​​and eigenvectors. Perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors. If vector v is an eigenvector of matrix C, then it satisfies the eigenvalue calculation formula, where λ is the eigenvalue corresponding to eigenvector v, and the eigenvector represents the projection direction of each feature in the new feature space. Calculate the principal component influence factors based on the calculated eigenvalues. Define the cumulative contribution rate ≥ 85% as the selected b (b < m) principal component factors. The eigenvalue calculation formula is as follows:

[0140] Cv=λ (6);

[0141] 9) Select principal components, sort them in descending order of eigenvalues, and select the first p eigenvectors as the principal component matrix P, where p is the dimension of the new feature space, that is, the dimension after dimensionality reduction.

[0142] 10) Dimensionality reduction: Multiply matrix X and transformation matrix P to obtain the dimension-reduced matrix Y.

[0143] Y = PN (7)

[0144] The K and Y matrices are concatenated to obtain a new matrix W1. This eliminates the adverse effects of redundant factors on prediction accuracy while retaining the contribution of the main features to wind speed prediction. Furthermore, to facilitate the study of relationships between multiple wind turbines, the wind speed characteristics of the target wind turbine are used as labels for random forest selection. The features of the remaining n-1 wind turbines are processed using the methods described in 1)-10) to obtain matrices W2, W3...W n The matrix W obtained by concatenating the matrices is used as the dataset for building the network graph.

[0145] The steps in S1.2.1, which involve constructing the original network graph G = (V, E) using the dataset described in S1.1.3, include the following:

[0146] The matrix W, after feature processing using S1.1.3, is used to construct a network graph based on data correlation as input for graph embedding, resulting in the original network graph G = (V, E), where V = {v1, v2, v3, ... v...}. n} represents a set of nodes, whose node feature matrix is ​​matrix W. T The dimension is N x D, where N is the number of nodes and D is the dimension of the features of each node. It is a set of edges, representing the connections between nodes. For a node v in the network... i and v j If there is a relationship between them, then there exists an edge represented by e. ij =1 means that e ij A value of 0 indicates that the nodes are not connected. The correlation between the N nodes in the original network graph is expressed using the Pearson correlation coefficient.

[0147]

[0148] In the formula: δ is the correlation coefficient, x i For node v i The i-th dimension features For node v i The average value of features across all dimensions, y i For node v j The i-th dimension features For node v j The average value of features across all dimensions.

[0149] S1.2.2, the graph embedding using a deep learning method based on GraphSAGE, includes the following steps:

[0150] Sampling: For each target node v in the node set V, randomly sample S from its neighboring node set. k Each node forms a subgraph;

[0151] Aggregation: For each target node ν in the node set V, use an aggregation function AGGREGATE. k , and the feature vectors of its neighboring nodes Compression and transformation are performed to obtain a new feature vector. Then, by using an aggregation function that calculates the average of the feature values ​​of neighboring nodes, the unordered neighboring features are transformed into a fixed-dimensional vector, which can then be concatenated with the features of the target node. The aggregation formula is as follows:

[0152]

[0153] In the formula: Let Q represent the embedding vector of node v at layer k, σ represent the activation function, and Q represent the embedding vector of node v at layer k. k This represents a learnable weight matrix, CONCAT represents the concatenation operation, and AGGREGATE represents the AGGREGATE operation. k Let N(v) represent the aggregation function of the k-th layer, and let N(v) represent the set of neighbors of node v.

[0154] Update: For each target node v in the node set V, its own feature vector and aggregated feature vectors The vectors are concatenated or added together, then passed through a fully connected layer and an activation function to obtain the embedding vector of the target node. This allows for the fusion and nonlinear transformation of features of each target node v and its neighboring nodes. The updated formula is as follows:

[0155]

[0156] In the formula: COMBINE represents the concatenation or addition operation, Q k+1 This represents the learnable weight matrix.

[0157] The steps in S1.2.3, which involve using the node embedding vector representation based on Mahalanobis distance to establish a new network graph G′=(V,E′), include the following:

[0158] The Mahalanobis distance method is used to optimize node embedding vectors. This method is suitable for high-dimensional data with a large number of meteorological factors and can adapt to situations where the data is not independent and identically distributed across dimensions. The calculation formula is as follows:

[0159]

[0160] in, This is the node embedding vector obtained above, μ is the mean vector of the sample distribution, and ∑ is the covariance matrix of the sample distribution. -1 It is the inverse of the covariance matrix, and then d is calculated based on each node. mahal The magnitude and correlation coefficient σ are used to determine the connection relationship of the optimized edges in the original network graph G=(V,E). Make modifications.

[0161]

[0162] This leads to the creation of a new network graph G′=(V,E′).

[0163] The framework for constructing GIN based on the network graph G′=(V,E′) in S2.1 includes the following steps:

[0164] First, we utilize GIN aggregation. To ensure that different inputs are mapped to different outputs, we inject during modeling, using summation as the aggregation function. A multilayer perceptron can simulate the combination of functions. The node representations of the GIN are updated using the following method:

[0165]

[0166] In the formula: Let N(v) be the feature representation of node (v) at level (k), and N(v) be the set of neighboring nodes of (v), ∈ (k) It is a trainable parameter learned at layer (k), which allows the model to adaptively adjust the importance of the center node features and the neighbor features;

[0167] Secondly, GIN is used to read out the graph and aggregate the feature vectors of all nodes in the new network graph G′=(V,E′) into a single graph-level feature vector for graph classification or other graph-level tasks. The CONCAT method is used to sum the node features in each iteration as the graph features of that iteration, and then the graph features of each iteration are concatenated to obtain the final graph features.

[0168]

[0169] Here, h G This is the representation of the new network graph G′=(V,E′). G is the feature representation of node (v) at level (k), and G is the set of all nodes in the graph.

[0170] S2.2, using the output of GIN as the input to the recurrent neural network GRU, includes the following steps:

[0171] S2.2.1 Calculate the values ​​of the update gate and the reset gate. The update gate controls how much of the hidden state from the previous time step is retained in the current time step, and the reset gate controls how much of the hidden state from the previous time step is included in the calculation of the current time step. The formula for this step is:

[0172]

[0173] Among them, z t and r t These are the values ​​for updating the gate and resetting the gate, respectively. σ is the sigmoid activation function used to facilitate gradient calculation and backpropagation. W z and W r It is the weight matrix, b z and b r It is the bias vector, b r It is the hidden state from the previous moment, x t It is the input at the current moment;

[0174] S2.2.2 Calculate the candidate hidden state, which is the output of the current input and the reset hidden state from the previous time step after passing through a tanh activation function. The formula for this step is:

[0175]

[0176] Among them, h t-1 It is a candidate hidden state, W h It is the weight matrix, b h It is the bias vector, and ⊙ is the element-wise multiplication operation;

[0177] S2.2.3 Calculate the hidden state at the current time step, which is the weighted average of the hidden state at the previous time step and the candidate hidden states. The weights are determined by the update gate. The formula for this step is:

[0178]

[0179] Among them, h t It is the hidden state at the current moment, z t It updates the value of the gate. It is a candidate hidden state.

[0180] In summary, firstly, at the spatial level, a network graph is constructed by extracting the overall wind speed characteristics within the wind farm at a certain moment through graph embedding and Mahalanobis distance. Then, the network graph is input into GIN, and multiple graphs are input successively at time steps. Spatial convolutional layers are used to aggregate the neighbor information of nodes. Finally, the output of GIN is input into GRU to aggregate the information of the current moment and the historical moment. The representation of each node incorporates the neighbor information of the current moment and the historical moment. The output of GRU is the time series prediction result that combines the relationships between wind turbines.

[0181] Experimental Example

[0182] Datasets and settings

[0183] The dataset consists of 20 sites along the California coast, with Site IDs 0, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 69, 70, 71, 72, 73, 74, 75, and 76, providing 200 days of forecast data at 15-minute intervals. The preprocessed dataset size was set to 5000, and the ratio of the training, validation, and test sets for the model was set to 6:2:2.

[0184] Experimental equipment

[0185] This experiment is implemented in Python based on the Tensorflow framework and accelerated using the Compute Unified Device Architecture (CUDA). The simulation hardware platform settings are as follows: CPU: Intel Core i7-10875H 2.30GHz, RAM: 32GB, GPU: Nvidia GeForce RTX 2070.

[0186] Error evaluation criteria

[0187] When predicting local wind speed for multiple wind turbines within a wind farm using a network diagram as input, this simulation uses root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R) as evaluation metrics to comprehensively describe the prediction effect. At time k, the prediction error metrics for local wind speed for multiple wind turbines distributed as a point cloud within the wind farm are: mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). Their formulas are as follows:

[0188]

[0189] Where n is the number of samples; y i For the i-th observation; Let be the i-th predicted value. The smaller the MSE, the better the model fits and the smaller the prediction error.

[0190]

[0191] Where n is the number of samples; y i For the i-th observation; Let be the i-th predicted value. The smaller the RMSE, the better the model fits and the smaller the prediction error.

[0192]

[0193] Where n is the number of samples; y i For the i-th observation; Let be the i-th predicted value. The smaller the MAE, the better the model fit and the smaller the prediction error.

[0194]

[0195] Where n is the number of samples; y i For the i-th observation; Let be the i-th predicted value. The smaller the MAPE, the higher the accuracy of the model and the smaller the prediction error.

[0196] Experimental Results and Analysis

[0197] RF-PCA Feature Screening Analysis

[0198] Since the original data contains 46 meteorological elements, previous experiments have shown that extracting characteristic meteorological elements using a single method is inaccurate and somewhat one-sided. Therefore, we first used random forest to select meteorological elements that reflect the main information contained in the original variables, and then used PCA to reduce the dimensionality of the remaining secondary meteorological elements. Finally, we fused the main features and the dimensionality-reduced data to obtain the characteristic meteorological elements that can truly affect wind changes. Figure 6 For wind turbine number 0, the top ten features with the highest contribution to other wind turbine features are determined by their height (10m). Subsequent secondary features are subjected to PCA dimensionality reduction, and then the primary and secondary features are concatenated to form the dataset for building the graph network. (Standard, 20 sites).

[0199] Evaluation and analysis of network graph construction using graph embedding.

[0200] To verify the effectiveness of graph embedding in building network graphs, we compared it with the method of building network graphs solely using Mahalanobis distance, namely the Graph Isomorphism Gated Recurrent Graph (ED-GIN-GRU) model. We chose the same constant for the correlation coefficient σ, and fed the built network graphs into the ED-GIN-GRU model for training, performing 40 days of 10m and 30m wind speed predictions. Model error was evaluated using four metrics: MSE, RMSE, MAE, and MAPE.

[0201] Table 1: Wind speed at a height of 10m

[0202] ED-GIN-GRU GE-GIN-GRU MSE 0.3641 0.2270 RMSE 0.6352 0.4764 MAE 0.4265 0.3586 MAPE 0.2215 0.1251

[0203] Table 2: Wind speed at 30m height

[0204] ED-GIN-GRU GE-GIN-GRU MSE 0.5641 0.3077 RMSE 0.7252 0.5548 MAE 0.5632 0.4194 MAPE 0.3453 0.1128

[0205] Figure 7 and Figure 8 The figures show the predicted and actual trends of wind turbine number 0 under the 10m and 30m graph isomorphic gated cyclic network diagrams and the correlation-based network diagram model, respectively. As can be seen from the figures, both the graph isomorphic gated cyclic network diagram (GE-GIN-GRU) and the correlation-based network diagram (ED-GIN-GRU) can track the changing trends of wind speed signals well overall. ED-GIN-GRU, however, performs poorly in tracking wind speed changes, especially during wind speed fluctuations, resulting in significant prediction errors. ED-GIN-GRU, on the other hand, performs well in this regard, indicating that the wind turbine sites were trained and screened during the graph embedding network diagram construction process. This allows for early prediction of wind speed fluctuations using information from other sites, effectively simulating the unique wind speed trends of different turbines.

[0206] Evaluation and Analysis of GIN-GRU Neural Network

[0207] To further validate the effectiveness of the GIN-GRU neural network, we compared the proposed GE-GIN-GRU model with three popular wind speed prediction models: Convolutional Long Short-Term Memory (CNN-LSTM), Graph Attention Long Short-Term Memory (GAT-GRU), and Graph Attention Gated Recurrent (GAT-LSTM). Wind speed predictions at various altitudes were performed over 40 days on a test set of 20 stations. Model error was evaluated using four metrics: MSE, RMSE, MAE, and MAPE.

[0208] Table 3: Wind speed at a height of 10m

[0209] CNN-LSTM GAT-GRU GAT-LSTM GE-GIN-GRU MSE 7.7851 2.1808 2.3968 0.8457 RMSE 2.7902 1.4768 1.5482 0.9196 MAE 2.6851 1.3567 1.4009 0.7612 MAPE 0.2283 0.1152 0.1213 0.0636

[0210] Table 4: Wind speed at 30m height

[0211] CNN-LSTM GAT-GRU GAT-LSTM GE-GIN-GRU MSE 4.9849 1.3163 1.4214 0.6400 RMSE 2.2327 1.1473 1.1922 0.8000 MAE 2.0489 0.9414 0.9726 0.6076 MAPE 0.1629 0.0759 0.0787 0.0503

[0212] like Figure 9 and Figure 10 Comparison of wind speed predictions for 10m and 30m GIN-GRU respectively.

[0213] The experimental results used the RMSE (Real-Time Sequence) of predicted and actual values ​​as a scoring function to evaluate the model's predictive performance. Tables 1, 2, and 3 show the wind speeds at different altitudes for each model at the four stations. Despite changes in altitude, GIN-GRU maintained high accuracy in wind speed prediction, exhibiting the smallest error compared to other models, demonstrating the effectiveness of the proposed method.

[0214] Figure 10 and Figure 11The prediction trends of other deep learning models and GE-GIN-GRU at wind speeds predicted at heights of 10m and 30m at station number 0 were observed. While other networks generally followed the measured wind speed signal well, they lacked the ability to capture the graph structure and relationships as effectively as GIN-GRU due to the large amount of graph data involved. This required more spatiotemporal nodes for prediction, resulting in a time lag. Similar to GE-GIN-GRU, their prediction error was relatively stable during smooth wind speed changes, but increased significantly during fluctuating wind speeds. In contrast, GE-GIN-GRU's prediction trend was closer to the measured wind speed data than other networks during smooth wind speed changes, especially during periods of smooth wind speed increases. Furthermore, during periods of fluctuating wind speed, GE-GIN-GRU, utilizing its nonlinear fitting capabilities, approximated the measured wind speed signal more closely than the continuous method with smaller fluctuations.

[0215] Combining data mining and neural network techniques, this paper proposes a wind speed prediction model based on a cyclic graph isomorphic network, enabling wind speed prediction for stations in the study area and improving the timeliness and accuracy of forecasts. The proposed RF-PCA algorithm effectively reduces the number of meteorological features involved in the calculation, improving the model's computational efficiency. The GE-GIN-GRU model effectively integrates multiple neural network algorithms, exhibiting stronger generalization ability and higher prediction accuracy. Therefore, this model maintains excellent prediction quality and stability on wind speed datasets at different altitudes. Furthermore, to extract the relationships between wind turbines, a deep learning-based graph embedding method is proposed to construct the network graph, leveraging the powerful spatial feature extraction capabilities of GIN and the strong temporal feature extraction capabilities of GRU. By fully utilizing the advantages of both models, deep extraction of spatiotemporal relationship features is achieved. In future research, combining the location coordinates of wind turbines with the correspondence between their nodes to construct a new network graph is the main objective.

[0216] 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. Furthermore, 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 a process, method, article, or apparatus. Without further limitations, 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 said element.

[0217] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wind speed prediction method based on graph embedding and GIN-GRU for multiple wind farm features, characterized in that, Includes the following two contents: S1. Constructing a network diagram S2, using a network diagram Based on this, a spatiotemporal prediction model based on graph GIN and GRU is constructed; S1 includes the following steps: S1.1.1 Extraction of primary and secondary features based on random forest; S1.1.2, Dimensionality reduction of secondary features based on PCA; S1.1.

3. Concatenate the primary features and the reduced secondary features to construct the network graph. The dataset; S1.2.1 Construct the original network graph using the dataset described in S1.1.

3. ; S1.2.2 Graph embedding is performed using a deep learning method based on GraphSAGE; S1.2.

3. Then, based on the node embedding vector representation using Mahalanobis distance, a new network graph is established. ; S2 includes the following steps: S2.1, Based on network graph The framework for building GIN; S2.2, Substitute the output of GIN into the recurrent neural network GRU as the input of GRU; S2.3 The output of the recurrent neural network GRU is the time series prediction result of the relationship between wind turbines; The step S1.1.1, extraction of primary and secondary features based on random forest, includes the following steps: The Gini importance index is used to extract the main features of the random forest. The calculation of Gini importance is first performed by constructing a random forest from each decision tree trained on the training set sampled by Bootstrap. Then, during the construction of each tree, a subset of features obtained by cross-validation is randomly selected from all features, and the optimal branch features are found according to the principle of minimizing the Gini index to optimize the model performance. Finally, unbiased estimation is performed from the unsampled data, and the importance of each feature is calculated using the Gini index. The features are sorted in descending order to select the main and secondary features. S1.1.2, PCA-based secondary feature dimensionality reduction, includes the following steps: By calculating the eigenvalues ​​and eigenvectors based on the covariance matrix, and then reducing the dimensionality of the corresponding matrices of the secondary features, we can obtain the meteorological elements among the secondary features that can significantly affect wind speed changes. S2.1, based on network graph Building the GIN framework involves the following steps: First, we utilize GIN aggregation. To ensure that different inputs are mapped to different outputs, we inject during modeling, using summation as the aggregation function. A multilayer perceptron can simulate the combination of functions. The node representations of the GIN are updated using the following method: （13） In the formula: Is node (v) at the ( The feature representation of the layer, It is the set of neighboring nodes of (v). It is in the ( A trainable parameter learned by the layer allows the model to adaptively adjust the importance of features of the center node and features of its neighbors; Secondly, GIN is read out to obtain the new network graph. The feature vectors of all nodes are aggregated into a single graph-level feature vector for graph classification or other graph-level tasks. The CONCAT algorithm is used to sum the node features in each iteration as the graph feature for that iteration, and then concatenates the graph features from each iteration to obtain the final graph feature. (14) here, It is a new network diagram The expression, G is the feature representation of node (v) at level (k), and G is the set of all nodes in the graph; S2.2, using the output of GIN as the input to the recurrent neural network GRU, includes the following steps: S2.2.1 Calculate the values ​​of the update gate and the reset gate. The update gate controls how much of the hidden state from the previous time step is retained in the current time step, and the reset gate controls how much of the hidden state from the previous time step is included in the calculation of the current time step. The formula for this step is: （15） in, and These are updating the door's value and resetting the door's value, respectively. The sigmoid activation function is used to facilitate gradient calculation and backpropagation. and It is a weight matrix. and It is a bias vector. It is the hidden state from the previous moment. It is the input at the current moment; S2.2.2 Calculate the candidate hidden state, which is the output of the current input and the reset hidden state from the previous time step after passing through a tanh activation function. The formula for this step is: （16） in, It is a candidate hidden state. It is a weight matrix. It is a bias vector. It is an element-wise multiplication operation; S2.2.3 Calculate the hidden state at the current time step, which is the weighted average of the hidden state at the previous time step and the candidate hidden states. The weights are determined by the update gate. The formula for this step is: （17） in, It is the hidden state at the current moment. It updates the value of the gate. It is a candidate hidden state.

2. The wind speed prediction method based on graph embedding and GIN-GRU for multiple wind farm features according to claim 1, characterized in that, In step S1.1.3, the primary features and the reduced secondary features are concatenated to form the network graph. The dataset includes the following steps: 1) Calculate the Gini index. The dataset has j features ( , , , …, Given C categories and I decision trees, the Gini index of node m is: （1） In the formula: Let m represent the Gini index of node m. This represents the probability estimate that sample m at node c belongs to class c. 2) Calculate the change in the Gini index, characteristic quantity. Importance score of node m The change in the Gini index before and after branching at node m is used to represent the value: （2） In the formula: and These are the two new nodes after branching from node m. , The Gini index; 3) Calculate the importance of the i-th tree and define the feature values. If the set of nodes appearing in the i-th tree is M, then the feature quantity is... The importance of the i-th tree is represented as: （3） In the formula: Representing characteristic quantity The importance of the i-th tree; 4) Calculate the final importance and feature quantity. Its importance in RF can be expressed as: （4） In the formula: Representing characteristic quantity Its importance in RF; 5) Based on the importance of the features, we select the top k features according to the importance score gradient to obtain the feature matrix K after random forest filtering. The matrix N of the remaining jk features is reduced by PCA to retain the principal components. 6) Decentralize the data by centering the dimensionality-reduced matrix N, that is, subtract the mean of each feature so that the mean of each feature is 0; 7) Calculate the covariance matrix by performing matrix multiplication on the centered data to obtain a covariance matrix. This indicates the correlation between each feature. （5） In the formula: Indicates the correlation between each feature; 8) Calculate the eigenvalues ​​and eigenvectors. Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues ​​and corresponding eigenvectors, such as vectors. It is a matrix If the eigenvectors satisfy the eigenvalue calculation formula, then... It is an eigenvector The corresponding eigenvalues ​​and eigenvectors represent the projection direction of each feature in the new feature space. Principal component influence factors are calculated based on the eigenvalues. The b (b < m) principal component factors are defined as having a cumulative contribution rate ≥ 85%. The formula for calculating the eigenvalues ​​is as follows: （6）； 9) Select principal components, sort them in descending order of eigenvalues, and select the first p eigenvectors as the principal component matrix P, where p is the dimension of the new feature space, that is, the dimension after dimensionality reduction. 10) Dimensionality reduction: Multiply matrix X and transformation matrix P to obtain the dimension-reduced matrix Y. （7） Concatenate the K and Y matrices to obtain a new matrix. This approach aims to eliminate redundant factors that negatively impact prediction accuracy while preserving the contribution of key features to wind speed prediction. Furthermore, to facilitate research on the relationships between multiple wind turbines, the wind speed characteristics of the target wind turbine are used as labels for random forest selection. The features of the remaining n-1 wind turbines are then processed using the methods described in 1)-10) to obtain a matrix. , ... The matrix W obtained by concatenating the matrices is used as the dataset for building the network graph.

3. The wind speed prediction method based on graph embedding and GIN-GRU for multiple wind farm features according to claim 1, characterized in that, In step S1.2.1, the original network graph is constructed using the dataset described in S1.1.

3. Includes the following steps: The matrix W, after feature processing using S1.1.3, is used to construct a network graph based on data correlation as input for graph embedding, resulting in the original network graph G=(V,E), where V={ , , ,… } represents a set of nodes, whose node feature matrix is ​​a matrix. The dimension is N x D, where N is the number of nodes and D is the dimension of the features of each node. It is a set of edges, representing the connections between nodes in a network. and If there is a relationship between them, then there exists an edge using... express, This indicates that the nodes are not connected. The correlation between the N nodes in the original network graph is expressed using the Pearson correlation coefficient. (8) In the formula: The correlation coefficient is... For nodes The i-th dimension features For nodes The average value of features across all dimensions For nodes The i-th dimension features For nodes The average value of features across all dimensions.

4. The wind speed prediction method based on graph embedding and GIN-GRU for multiple wind farm features according to claim 3, characterized in that, S1.2.2, the graph embedding using a deep learning method based on GraphSAGE, includes the following steps: Sampling: For each target node in the node set V Randomly sample from its set of neighboring nodes Each node forms a subgraph; Aggregation: For each target node in the node set V Using an aggregate function , and the feature vectors of its neighboring nodes Compression and transformation are performed to obtain a new feature vector. Then, by using an aggregation function that calculates the average of the feature values ​​of neighboring nodes, the unordered neighboring features are transformed into a fixed-dimensional vector, which can then be concatenated with the features of the target node. The aggregation formula is as follows: （9） In the formula: Represents a node In the Layer embedding vectors, This represents the activation function. This represents the learnable weight matrix. This indicates a splicing operation. Indicates the first Layer aggregation functions, Represents a node The set of neighbors; Update: For each target node in the node set V , its own feature vector and aggregated feature vectors The vectors are concatenated or added together, then passed through a fully connected layer and an activation function to obtain the embedding vector of the target node. This allows for the targeting of each node. The updated formula, which combines features from neighboring nodes with nonlinear transformation, is as follows: （10） In the formula: COMBINE represents a concatenation or addition operation. This represents the learnable weight matrix.

5. The wind speed prediction method based on graph embedding and GIN-GRU for multiple wind farm features according to claim 3, characterized in that, S1.2.3, then based on Mahalanobis distance, node embedding vector representation is used to establish a new network graph. Includes the following steps: The Mahalanobis distance method is used to optimize node embedding vectors. This method is suitable for high-dimensional data with a large number of meteorological factors and can adapt to situations where the data is not independent and identically distributed across dimensions. The calculation formula is as follows: （11） in, It is the node embedding vector obtained above. It is the mean vector of the sample distribution. It is the covariance matrix of the sample distribution. It is the inverse of the covariance matrix, and then calculated based on each node. Size and correlation coefficient To determine the connection relationships of optimized edges in the original network graph In Make modifications. (12) This leads to the creation of a new network diagram. .

Images