Power distribution network line parameter identification method based on graph convolutional neural network
Through the method based on graph convolution neural network, Chebishev graph convolution neural network and TPE Bayesian optimization, the problem of insufficient identification accuracy of distribution network line parameters in traditional methods is solved, and high-precision identification is achieved in complex power grid environments, improving the security and regulation capabilities of distribution networks.
Patent Information
- Application Number
- CN202510447496.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-10
- Publication Date
- 2025-07-18
AI Technical Summary
When traditional distribution network line parameter identification methods deal with complex power grid topology and large amounts of uncertain data, the identification accuracy is insufficient and are easily affected by incomplete data or noise, making it difficult to achieve efficient and accurate parameter identification.
The method based on graph convolutional neural network is adopted to standardize the feature data of the distribution network nodes and model the graph structure, and use the Chebischev graph convolutional neural network to perform neighbor feature aggregation and higher-order polynomial feature propagation. Combining the cross entropy loss function and the TPE Bayesian optimization method, the model hyperparameters are optimized to achieve efficient identification of line parameters.
It realizes high-precision line parameter identification in complex power grid environments, has good noise resistance and generalization capabilities, and provides solid technical support for the safe operation and intelligent regulation of the distribution network.
Smart Images
Figure CN120336758A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of distribution networks, and particularly to a method for identifying distribution network line parameters based on a graph convolutional neural network. Background Art
[0002] With the continuous development of the power system, the identification of distribution network line parameters has become an important task in power system research and engineering applications. Accurately identifying distribution network line parameters is of great significance for improving power grid dispatching capabilities, optimizing power transmission efficiency, reducing energy losses, and ensuring the stability and security of the power system. Traditional methods for identifying distribution network line parameters mainly rely on physical models, mathematical models, and traditional optimization algorithms. However, due to the influence of various factors on line parameters in the power system, traditional methods face many limitations, especially when dealing with complex power grid topologies and large amounts of uncertain data, the identification accuracy is often affected.
[0003] Traditional line parameter identification methods include optimization methods based on power flow calculation, least squares method, weighted least squares method, etc. These methods estimate the admittance parameters of the line by fitting the data of the power grid operation state and using tools such as voltage sensitivity matrix and state estimation model. However, these methods are often difficult to achieve efficient and accurate parameter identification in the face of large-scale, complex topology, and dynamic changes of the power grid. In addition, traditional methods are also easily affected by incomplete data or noise, resulting in large deviations in parameter estimation results. Therefore, the scalability and accuracy of traditional methods are still significantly insufficient when dealing with complex distribution network problems.
[0004] In recent years, with the development of artificial intelligence technology, distribution network line parameter identification methods based on machine learning and deep learning have gradually emerged. Common artificial intelligence methods include support vector machine (SVM), random forest, deep neural network (DNN), etc. These methods can automatically learn the potential laws in the data and improve the prediction accuracy through optimization algorithms. Different from traditional methods, artificial intelligence methods do not rely on explicit physical models, but obtain models with strong adaptability and generalization ability through a large amount of data training. For example, support vector machine classifies or regresses by finding the optimal hyperplane and is suitable for dealing with nonlinear problems; deep neural network can automatically extract features and perform complex pattern recognition through a multi-layer network structure. Although these methods have achieved good results in some studies, there are still problems such as strong data dependence, large computational resources required for the training process, and difficulty in effectively combining power grid topology information. Summary of the Invention
[0005] In order to solve the above technical problems, the present invention provides a method for identifying distribution network line parameters based on a graph convolutional neural network with simple algorithm, high efficiency and accuracy.
[0006] The technical solution of the present invention to solve the above technical problems is: a method for identifying line parameters of a distribution network based on a graph convolutional neural network, comprising the following steps:
[0007] Step 1: Perform feature construction and standardization on the power feature data of each node in the collected distribution network;
[0008] Step 2: Adopt a distribution system graph structure modeling method to construct the real structure of the distribution network into a graph model, and generate an adjacency matrix with consistent physical structure;
[0009] Step 3: Use the Chebyshev graph convolutional neural network to aggregate neighbor features, perform long-distance feature propagation through high-order polynomials, so that the nodes contain the topological and physical information of their multi-layer neighbor nodes, then map the node features to edge features, and then convert the edge features into line parameters through a linear layer;
[0010] Step 4: Use the cross-entropy loss function to train the model, and at the same time use the TPE Bayesian optimization method to optimize the hyperparameters of the model;
[0011] Step 5: Identify the line parameters of the distribution network through the optimized model.
[0012] For the above method for identifying line parameters of a distribution network based on a graph convolutional neural network, the specific process of the first step is as follows:
[0013] (1-1) Perform feature construction on the node active power P, reactive power Q, and voltage amplitude U of each node in the collected distribution network, and construct the node features into node features X, X ∈ R N×F , where N is the number of nodes, is the feature dimension of each node, and construct the edge features to be obtained into an edge feature matrix Y, Y ∈ R E×2 , where E is the number of edges, and each edge has two target features, namely conductance G and susceptance B;
[0014] (1-2) Perform mean standardization on all node features:
[0015]
[0016] where X′ is the standardized node feature, μ is the mean, X i is the i-th node feature, σ is the standard deviation,
[0017] For the above method for identifying line parameters of a distribution network based on a graph convolutional neural network, the specific process of the second step is as follows:
[0018] (2-1) Abstract the substations, distribution transformers, ring main units, and busbars in the distribution network as nodes in the graph, and the node feature X exists on the nodes; map each distribution line to an edge in the graph, then the edge features to be obtained are concentrated on the edges; then construct the adjacency matrix from the connection relationships between the nodes. Assume there are N nodes in the distribution network, and the adjacency matrix A is an N×N matrix. The relationship between nodes i and j is defined as follows:
[0019]
[0020] The adjacency matrix A is a symmetric matrix, that is, A ij = A ji , convert the topological structure of the distribution system into the graph structure of the graph neural network, and construct a physically consistent adjacency matrix;
[0021] (2-2) When modeling the graph structure of the distribution system, self-loops are cancelled, that is, the diagonal elements of the constructed adjacency matrix are all 0, to ensure that only the connection relationships between busbars with real physical meanings are retained in the graph structure.
[0022] For the above method for identifying line parameters of a distribution network based on a graph convolutional neural network, in step (3), the specific process of aggregating neighbor features using a Chebyshev graph convolutional neural network is as follows:
[0023] The graph convolutional network mines the topological structure relationship between branches through the adjacency matrix A, so as to capture the spatial information of the distribution network, and then mines the features of each node through the input node features of active power P, reactive power Q, and voltage amplitude U after conversion; the graph convolutional network uses the Laplacian matrix to process graph data, and the Laplacian matrix L lap is defined as:
[0024] L lap = D deg - A(3)
[0025] where D deg is the degree matrix;
[0026] Normalize L lap to obtain as shown in the following formula:
[0027]
[0028] where λ max is the largest eigenvalue of the graph Laplacian, and I is the identity matrix.
[0029] For the above method for identifying line parameters of a distribution network based on a graph convolutional neural network, in step (3), the specific process of propagating features over long distances through high-order polynomials is as follows:
[0030] Adopt the ChebConv of graph convolution based on Chebyshev polynomials. The idea is to approximately replace the Laplacian operator in the standard GCN with Chebyshev polynomials. The Chebyshev polynomials are expressed as:
[0031] T k (r) = cos(k * arccos(r)) (5)
[0032] Among them, T k (r) is the k-th order Chebyshev polynomial when the input variable is r, k is the order of the Chebyshev polynomial, k ≥ 2, and r is the input variable. In order to calculate the high-order Chebyshev polynomial, the following recursive formula of the Chebyshev polynomial is adopted:
[0033]
[0034] In the ChebNet of the graph convolutional neural network based on Chebyshev polynomials, the update of node features adopts the expansion of Chebyshev polynomials to perform multi-order filtering on the normalized Laplacian matrix . The update process of node features in the l-th layer is as follows:
[0035]
[0036] Among them, h (l) is the node feature of the l-th layer, is the activation function, is the coefficient of the k-th order polynomial, K is the total order of the Chebyshev polynomial, is the Chebyshev polynomial of the Laplacian matrix.
[0037] In the above method for identifying the line parameters of a distribution network based on a graph convolutional neural network, in step three, the specific process of mapping node features to edge features is as follows:
[0038] Adopt the SumAggregation method to calculate the edge feature e ij , and the specific form is as follows:
[0039]
[0040] Among them: is the final node feature of node i, is the final node feature of node j, that is, the result after propagation through L layers of ChebConv.
[0041] In the above method for identifying the line parameters of a distribution network based on a graph convolutional neural network, in step three, the specific process of converting edge features to line parameters through a linear layer is as follows:
[0042] Introduce a linear layer to learn the mapping relationship from high-dimensional features to circuit parameters; the role of the linear layer is to convert the edge feature e ij into the final required circuit parameters G and B, and the mathematical expression is as follows:
[0043]
[0044] where: W e is the weight matrix of the linear transformation, and its size depends on the dimension of the edge feature and the number of output parameters; b is the bias term used to adjust the mean of the prediction result; is the circuit parameter predicted by the m-th circuit.
[0045] In the above method for identifying circuit parameters of a distribution network based on a graph convolutional neural network, in step three, after converting the edge feature into a circuit parameter, a Chebyshev graph convolutional network is constructed, and the Chebyshev graph convolutional network training attempts to minimize:
[0046]
[0047] where M is the number of all edges, y m is the actual circuit parameter of the m-th circuit, λ is the regularization coefficient, controlling the strength of regularization; is the L2 regularization term, representing the square of the L2 norm of all model parameters, that is, the sum of the squares of the parameters:
[0048]
[0049] where W is all the learnable parameter weights in the model, w p is the p-th parameter weight, and U is the number of all parameters in the model.
[0050] In the above method for identifying circuit parameters of a distribution network based on a graph convolutional neural network, in step four, when applying TPE for hyperparameter optimization, the objective function y = f(x) is the loss function of the model on the validation set:
[0051]
[0052] where x is the hyperparameter of the model, and x ∈ X′, X′ is the d-dimensional hyperparameter space, X′ = {x1, x2,..., x d}, x d is the d-th hyperparameter; x contains four hyperparameters, namely the number of layers of the network, the hidden feature dimension of each layer, the order of the Chebyshev polynomial, and the learning rate;
[0053] The goal of TPE is to find the optimal hyperparameter x by minimizing f(x) * :
[0054]
[0055] Under the TPE framework, TPE conditionally models the distribution of the hyperparameter x:
[0056]
[0057] Among them, P(x|y) is the conditional probability distribution, indicating the probability that the parameter configuration x appears under the condition that the objective function value is y. y is the objective function, and y * is the threshold of the loss function. l(x) represents the conditional probability density function of the hyperparameter x in the excellent region, and l(x) = P(x|y < y * ). g(x) represents the conditional probability density function of the hyperparameter
[0058] x in the poor region, and g(x) = P(x|y ≥ y * ); when dividing l(x) and g(x), let γ = p(y < y * ). γ represents the quantile of the TPE algorithm, and its range is between (0, 1).
[0059] For the above method for identifying the line parameters of a distribution network based on a graph convolutional neural network, in step four, during the hyperparameter optimization process, TPE selects the expected improvement EI as the acquisition function to guide the search process. The expected improvement calculates the expected value of improving the objective function after selecting a certain hyperparameter. The definition of EI is:
[0060]
[0061] EI y (x) is the expectation of the target value, and ∞ is proportional to; according to equation (16), when the value of g(x) / l(x) is maximized, EI reaches its highest value, thereby generating the best hyperparameter.
[0062] The beneficial effects of the present invention are as follows: The present invention uses the distribution system diagram structure for modeling, maps the real distribution network structure completely into the graph model of the graph neural network, enabling the graph neural network to directly mine features on the real graph structure; at the same time, through the Chebyshev graph convolutional network, features are mined from more neighbor nodes, and then the mined features are mapped into high-dimensional edge features through node features, and then the edge features are converted into the required line parameters through a linear layer; in addition, the introduction of TPE Bayesian optimization enables the model to handle various measurement data situations, ensuring high-precision line parameter identification results and providing a solid technical support for the safe operation and intelligent regulation of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] Figure 1 It is the flow chart of the present invention.
[0064] Figure 2 It is the topological structure diagram of Embodiment 1.
[0065] Figure 3 It is a schematic diagram of the results of adopting the present invention under different measurement errors in Embodiment 1.
[0066] Figure 4 It is a schematic diagram of the predicted value and relative error results under a 2% measurement error in Embodiment 1.
[0067] Figure 5 It is the topological structure diagram of Embodiment 2.
[0068] Figure 6 It is a comparison diagram of the identification results of adopting the present invention and other models under different measurement error scenarios in Embodiment 2. Specific implementation manners
[0069] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.
[0070] As Figure 1 shown, a method for identifying line parameters of a distribution network based on a graph convolutional neural network includes the following steps:
[0071] Step 1: Perform feature construction and standardization on the power feature data of each node in the collected distribution network.
[0072] The specific process of Step 1 is as follows:
[0073] (1-1) Perform feature construction on the node active power P, reactive power Q, and voltage amplitude U of each node in the collected distribution network, and construct the node features as node feature X, X ∈ R N×F , N is the number of nodes, is the feature dimension of each node, and construct the edge features to be obtained as edge feature matrix Y, Y ∈ R E×2 , E is the number of edges, and each edge has two target features, namely conductance G and susceptance B;
[0074] (1-2) Perform mean standardization on all node features:
[0075]
[0076] where X′ is the standardized node feature, μ is the mean, X i is the i-th node feature, σ is the standard deviation,
[0077] Step 2: Adopt a distribution system graph structure modeling method to construct the real structure of the distribution network as a graph model, and generate an adjacency matrix with consistent physical structure.
[0078] The specific process of the second step is as follows:
[0079] (2-1) Abstract the substations, distribution transformers, ring main units, and busbars in the distribution network as nodes in the graph, and the node feature X exists on the nodes; map each distribution line to an edge in the graph, then the edge features to be calculated are concentrated on the edges; then construct the adjacency matrix from the connection relationships between the nodes. Assume there are N nodes in the distribution network, and the adjacency matrix A is an N×N matrix. For the relationship between node i and node j, the following definitions are as follows:
[0080]
[0081] The adjacency matrix A is a symmetric matrix (for an undirected graph), that is, A ij = A ji , and convert the topological structure of the distribution system into the graph structure of the graph neural network, constructing a physically consistent adjacency matrix;
[0082] (2-2) To make the modeling more in line with physical reality, self-loops are removed during the graph structure modeling of the distribution system, that is, the diagonal elements of the constructed adjacency matrix are all 0, to ensure that only the connection relationships between busbars with real physical meanings are retained in the graph structure.
[0083] Step three: Use the Chebyshev graph convolutional neural network for neighbor feature aggregation, perform long-distance feature propagation through high-order polynomials, so that the nodes contain the topological and physical information of their multi-layer neighbor nodes, then map the node features to edge features, and then convert the edge features into line parameters through a linear layer.
[0084] The specific process of using the Chebyshev graph convolutional neural network for neighbor feature aggregation is as follows:
[0085] The graph convolutional network mines the topological structure relationships between branches through the adjacency matrix A, thereby effectively capturing the spatial information of the distribution network, and then mines the features of each node through the input node features of active power P, reactive power Q, and voltage amplitude U; the graph convolutional network uses the Laplacian matrix to process the graph data, and the Laplacian matrix L lap is defined as:
[0086] L lap = D deg - A(3)
[0087] where D deg is the degree matrix;
[0088] To avoid numerical instability problems when calculating graph convolution, normalize L lap to obtain as shown in the following formula:
[0089]
[0090] where λ max is the largest eigenvalue of the graph Laplacian, and I is the identity matrix.
[0091] The specific process of feature propagation over long distances through high-order polynomials is as follows:
[0092] To further improve the computational efficiency and better capture the multi-scale structural information of the graph, the graph convolution ChebConv based on Chebyshev polynomials is adopted. The idea is to approximately replace the Laplacian operator in the standard GCN with Chebyshev polynomials; the Chebyshev polynomials are expressed as:
[0093] T k (r) = cos(k * arccos(r)) (5)
[0094] where T k (r) is the k-th order Chebyshev polynomial when the input variable is r, k is the order of the Chebyshev polynomial, k ≥ 2, and r is the input variable; to calculate the high-order Chebyshev polynomials, the following recursive formula of Chebyshev polynomials is adopted:
[0095]
[0096] In the graph convolutional neural network ChebNet based on Chebyshev polynomials, the update of node features adopts the Chebyshev polynomial expansion, and multi-order filtering is performed on the normalized Laplacian matrix so that information can be propagated at multiple levels on the graph. Specifically, the update process of node features in the l-th layer is as follows:
[0097]
[0098] where h (l) is the node feature of the l-th layer, is the activation function, is the coefficient of the k-th order polynomial, K is the total order of the Chebyshev polynomial, is the Chebyshev polynomial of the Laplacian matrix.
[0099] In ChebNet, neighbor feature aggregation is one of the core ideas of the graph convolutional network. It allows each node to gradually obtain neighbor information at greater distances in different convolutional layers. Since the Chebyshev polynomials are used to approximate the graph Laplacian operator, each layer of convolution can not only directly aggregate the features of neighbor nodes, but also perform feature propagation at greater distances through high-order polynomials, providing more comprehensive data support for the identification of distribution line parameters.
[0100] The specific process of mapping node features to edge features is as follows:
[0101] After the multi-layer ChebConv calculation is completed, the final feature representation of each node reflects the high-order feature expression of the node after multi-layer graph convolution. This feature not only encodes the information of the node itself but also implicitly integrates the topological and physical information of its multi-layer neighbor nodes through the neighborhood propagation mechanism of the graph convolution layer. To obtain edge features from the high-order feature expression of the node, it is necessary to reasonably aggregate the final features of the two end nodes of the line to accurately characterize its electrical properties. The edge feature e is calculated using the summation aggregation method SumAggregation ij , and the specific form is as follows:
[0102]
[0103] Where: is the final node feature of node i, is the final node feature of node j, that is, the result after L-layer ChebConv propagation.
[0104] The specific process of converting edge features to line parameters through a linear layer is as follows:
[0105] The aggregated edge feature e ij is still a high-dimensional vector and needs to be mapped to the line parameters G and B with clear physical meanings. To complete this process, a linear layer LinearLayer is introduced to learn the mapping relationship from high-dimensional features to line parameters; the role of the linear layer is to convert the edge feature e ij into the final required line parameters G and B, and the mathematical expression is as follows:
[0106]
[0107] Where: W e is the weight matrix of the linear transformation, and its size depends on the dimension of the edge feature and the number of output parameters; b is the bias term used to adjust the mean of the prediction result; is the line parameter predicted for the m-th line.
[0108] After converting the edge features to line parameters, a Chebyshev graph convolutional network is constructed, and the Chebyshev graph convolutional network training attempts to minimize:
[0109]
[0110] Where M is the number of all edges, y m is the actual line parameter of the m-th line, λ is the regularization coefficient that controls the strength of regularization; is the L2 regularization term, which represents the square of the L2 norm of all model parameters, that is, the sum of the squares of the parameters:
[0111]
[0112] Among them, W are all the learnable parameter weights in the model, and w p is the p-th parameter weight, and U is the number of all parameters in the model.
[0113] Step 4: Train the model using the cross-entropy loss function, and at the same time optimize the hyperparameters of the model using the TPE Bayesian optimization method.
[0114] Considering that the performance of the Chebyshev graph convolutional network depends on multiple key hyperparameters. To optimize the hyperparameters, the TPE (Tree-structured Parzen Estimator) Bayesian optimization method is adopted. This method proposes a modeling method based on conditional probability, avoiding the bottleneck in Gaussian process modeling, and is especially suitable for high-dimensional and complex hyperparameter search spaces. When applying TPE for hyperparameter optimization, the objective function y = f(x) is the loss function of the model on the validation set:
[0115]
[0116] Among them, x is the hyperparameter of the model, and x ∈ X′, where X′ is a d-dimensional hyperparameter space, X′ = {x1, x2,..., x d}, and x d is the d-th hyperparameter; x contains four hyperparameters, namely the number of layers of the network, the hidden feature dimension of each layer, the order of the Chebyshev polynomial, and the learning rate;
[0117] The goal of TPE is to find the optimal hyperparameter x by minimizing f(x) * :
[0118]
[0119] Under the TPE framework, TPE conditionally models the distribution of the hyperparameter x:
[0120]
[0121] Among them, P(x|y) is the conditional probability distribution, indicating the probability that the parameter configuration x appears under the condition that the value of the objective function is y, y is the objective function, and y * is the threshold of the loss function, l(x) represents the conditional probability density function of the hyperparameter x in the excellent region, l(x) = P(x|y < y * ), g(x) represents the conditional probability density function of the hyperparameter x in the poor region, g(x) = P(x|y ≥ y * ); when dividing l(x) and g(x), let γ = p(y < y* ), where γ represents the quantile of the TPE algorithm, and the range is between (0, 1).
[0122] During the hyperparameter optimization process, TPE selects the expected improvement EI as the acquisition function to guide the search process. After calculating the selection of a certain hyperparameter, the expected value of improving the objective function is calculated. The definition of EI is:
[0123]
[0124] EI y (x) is the expectation of the target value, and ∝ is proportional to; according to Equation (16), when the value of g(x) / l(x) is maximized, EI reaches its highest value, thereby generating the best hyperparameters.
[0125] Step Five: Identify the distribution network line parameters through the optimized model.
[0126] Figure 2 It is the topological structure diagram of Embodiment 1. This figure is the IEEE33-node test feeder, which is a 12.66 kV distribution network with 32 distribution lines and 5 tie lines.
[0127] Figure 3 It is the schematic diagram of the results of adopting the present invention in different measurement errors in Embodiment 1. This figure shows that the model proposed by the present invention can maintain extremely high accuracy in identifying the distribution line parameters in different noise ratio environments, demonstrating strong robustness and reliability.
[0128] Figure 4 It is the schematic diagram of the predicted values and relative error results under 2% measurement error in Embodiment 1. Under 2% Gaussian error, the MAPE values of the line parameters G and B after model identification are 0.44% and 0.50% respectively, and the model demonstrates extremely high accuracy and excellent anti-noise ability.
[0129] Figure 5 It is the topological structure diagram of Embodiment 2. This figure is the IEEE123-node feeder, which is a 4.16 kV distribution network with 122 distribution lines.
[0130] Figure 6 It is the comparison diagram of the identification results of adopting the present invention and other models in different measurement error scenarios in Embodiment 2. As can be seen from Figure 6 it, the present invention can maintain high identification accuracy in different noise environments, and compared with other common methods.
[0131] Regarding the identification performance of the model proposed in the present invention in Example 1, it can be found that the model can maintain extremely high accuracy in an extremely poor measurement data environment. Regarding the analysis of the identification performance of the model proposed in the present invention in Example 2 and comparing it with other deep learning models (CNN, TCN, RESNET), the identification results and generalization ability of the proposed model under different measurement error scenarios are verified.
Claims
1. A method for identifying distribution network line parameters based on a graph convolutional neural network, characterized in that, It includes the following steps: Step 1: Perform feature construction and standardization on the power feature data of each node in the collected distribution network; Step 2: Adopt a distribution system diagram structure modeling method to construct the real structure of the distribution network into a graph model, and generate an adjacency matrix with a consistent physical structure; Step 3: Use a Chebyshev graph convolutional neural network to aggregate neighbor features, perform long-distance feature propagation through high-order polynomials, so that each node contains the topological and physical information of its multi-layer neighbor nodes, then map the node features to edge features, and then convert the edge features into line parameters through a linear layer; Step 4: Train the model using a cross-entropy loss function, and at the same time use the TPE Bayesian optimization method to optimize the hyperparameters of the model; Step 5: Identify the line parameters of the distribution network through the optimized model.
2. The method for identifying distribution network line parameters based on a graph convolutional neural network according to claim 1, wherein The specific process of Step 1 is as follows: (1-1)Construct features for the active power P, reactive power Q, and voltage magnitude U of each node in the collected distribution network, and construct the node features as node feature X, X ∈ R N×F , where N is the number of nodes, is the feature dimension of each node, and construct the edge features to be solved as the edge feature matrix Y, Y ∈ R E×2 , where E is the number of edges, and each edge has two target features, namely conductance G and susceptance B; (1-2) Perform mean standardization on all node features: where X′ is the node feature after standardization, μ is the mean, X i is the i-th node feature, and σ is the standard deviation, 3. The method for identifying distribution network line parameters based on a graph convolutional neural network according to claim 2, characterized in that, The specific process of Step 2 is as follows: (2-1) Abstract the substations, distribution transformers, ring main units, and busbars in the distribution network into nodes in the graph, and the node features X exist on the nodes; map each distribution line to an edge in the graph, then the edge features to be solved are concentrated on the edges; Then construct an adjacency matrix from the connection relationships between nodes. Suppose there are N nodes in the distribution network, and the adjacency matrix A is an N×N matrix. The relationship between node i and node j is defined as follows: The adjacency matrix A is a symmetric matrix, i.e., A ij = A ji , when converting the topological structure of the power distribution system into the graph structure of the graph neural network, a physically consistent adjacency matrix is constructed; (2-2) Cancel self-loops during the distribution system diagram structure modeling, that is, the diagonal elements of the constructed adjacency matrix are all 0, to ensure that only the connection relationships between busbars with real physical meanings are retained in the graph structure.
4. The method for identifying distribution network line parameters based on a graph convolutional neural network according to claim 3, wherein In Step 3, the specific process of aggregating neighbor features using a Chebyshev graph convolutional neural network is as follows: The graph convolutional network mines the topological structure relationship between branches through the adjacency matrix A, so as to capture the spatial information of the distribution network. Then, by inputting the converted node features of active power P, reactive power Q, and voltage amplitude U, the features of each node are mined. The graph convolutional network uses the Laplacian matrix to process graph data, and the Laplacian matrix L lap is defined as: L lap = D deg - A(3) where D deg is the degree matrix; For L lap perform normalization to obtain as shown in the following formula: where λ max is the largest eigenvalue of the graph Laplacian, and I is the identity matrix.
5. The method for identifying distribution network line parameters based on a graph convolutional neural network according to claim 4, wherein In Step 3, the specific process of performing long-distance feature propagation through high-order polynomials is as follows: Adopt a graph convolution ChebConv based on Chebyshev polynomials, the idea of which is to approximately replace the Laplacian operator in the standard GCN with Chebyshev polynomials; the Chebyshev polynomials are expressed as: T k (r) = cos(k * arccos(r))(5) where, T k (r) is the k-th order Chebyshev polynomial when the input variable is r, k is the order of the Chebyshev polynomial, k ≥ 2, and r is the input variable; in order to calculate the high-order Chebyshev polynomial, the following recursive formula of the Chebyshev polynomial is adopted: In the ChebNet, a graph convolutional neural network based on Chebyshev polynomials, the update of node features uses the Chebyshev polynomial expansion to perform multi-order filtering on the normalized Laplacian matrix The node feature update process of the l-th layer is as follows: where h (l) is the node feature of the l-th layer, θ is the activation function, are the coefficients of the k-th order polynomial, and K is the total order of the Chebyshev polynomial, is the Chebyshev polynomial of the Laplacian matrix.
6. The method for identifying distribution network line parameters based on a graph convolutional neural network according to claim 5, wherein In Step 3, the specific process of mapping node features to edge features is as follows: Calculate the edge feature e using the SumAggregation method ij , and the specific form is as follows: Wherein: is the final node feature of node i, is the final node feature of node j, that is, the result after propagation through L layers of ChebConv.
7. The method for identifying distribution network line parameters based on a graph convolutional neural network according to claim 6, wherein In Step 3, the specific process of converting edge features into line parameters through a linear layer is as follows: Introduce a linear layer to learn the mapping relationship from high-dimensional features to circuit parameters; the role of the linear layer is to convert the edge feature e ij into the final required circuit parameters G and B, and the mathematical expression is as follows: Where: W e is the weight matrix of the linear transformation, whose size depends on the dimension of the edge features and the number of output parameters; b is the bias term used to adjust the mean of the prediction result; is the line parameter predicted for the m-th line.
8. The method for identifying distribution network line parameters based on a graph convolutional neural network according to claim 7, wherein In Step 3, after converting edge features into line parameters, construct a Chebyshev graph convolutional network, and the Chebyshev graph convolutional network training attempts to minimize: where M is the number of all edges, and y m is the actual line parameter of the m-th line, and λ is the regularization coefficient that controls the strength of regularization; is the L2 regularization term, representing the square of the L2 norm of all model parameters, that is, the sum of the squares of the parameters: Among them, W is all the learnable parameter weights in the model, and w p is the p-th parameter weight, and U is the number of all parameters in the model.
9. The method for identifying distribution network line parameters based on a graph convolutional neural network according to claim 8, wherein In Step 4, when applying TPE for hyperparameter optimization, the objective function y = f(x) is the loss function of the model on the validation set: where x is a hyperparameter of the model, and x ∈ X′, X′ is a d-dimensional hyperparameter space, X′ = {x1, x2,..., x d}, x d is the d-th hyperparameter; x contains four hyperparameters, namely the number of layers of the network, the dimension of hidden features in each layer, the order of the Chebyshev polynomial, and the learning rate; The goal of TPE is to find the optimal hyperparameter x by minimizing f(x). * : In the TPE framework, TPE conditions the distribution of hyperparameters x: Among them, P(x|y) is the conditional probability distribution, representing the probability of the parameter configuration x occurring under the condition that the objective function value is y, where y is the objective function. * is the threshold of the loss function, l(x) represents the conditional probability density function of the hyperparameter x in the excellent region, and l(x) = P(x|y < y * ), g(x) represents the conditional probability density function of the hyperparameter x in the poor region, and g(x) = P(x|y ≥ y * ); when dividing l(x) and g(x), let γ = p(y < y * ), where γ represents the quantile of the TPE algorithm and ranges between (0, 1).
10. The method for identifying distribution network line parameters based on a graph convolutional neural network according to claim 9, wherein In Step 4, during the hyperparameter optimization process, TPE selects the expected improvement EI as the acquisition function to guide the search process. The expected improvement calculates the expected value of improving the objective function after selecting a certain hyperparameter. The definition of EI is: EI y (x) is the expectation of the target value, and ∝ means proportional to; according to Equation (16), when the value of g(x) / l(x) is maximized, EI reaches its highest value, thereby generating the optimal hyperparameters.