A method and system for identifying power grid branch parameters based on adaptive multi-hop graph attention network

Abstract

The present invention discloses a method and system for identifying power branch parameters based on an adaptive multi-hop graph attention network in the field of power branch parameter identification technology, aiming to solve the optimization problem of existing power branch parameter identification technology. The method includes: building a simulated power grid, using the simulated power grid to simulate and record the measurement data of each branch to obtain a data set to be identified; preprocessing the data set to be identified to obtain a processed data set; constructing an adjacency matrix and a multi-hop matrix based on the topological structure of the simulated power grid; inputting the processed data set, the adjacency matrix, and the multi-hop matrix into a trained power branch parameter identification model; performing feature aggregation based on the processed data set, the adjacency matrix, and the multi-hop matrix through an adaptive multi-hop graph attention network module to obtain a power branch feature set; using the SE attention module to weight the power branch feature set and then inputting it into a multi-layer fully connected layer for regression to obtain predicted power branch parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid parameter identification, and in particular to a power grid branch parameter identification method and system based on an adaptive multi-hop graph attention network. Background Art

[0002] Accurate grid branch parameters are crucial for grid operation and maintenance, serving as the foundation for various power grid analysis and calculations, including fault diagnosis, state estimation, and line loss calculation. However, in the actual operation of a power grid, influenced by factors such as operating conditions and climatic conditions, the parameter values ​​recorded in the grid database often deviate from the actual line parameter values, potentially impacting the reliability of grid dispatch decisions. Therefore, developing high-precision grid branch parameter identification methods is crucial to ensuring the safe and stable operation of the power grid.

[0003] Existing methods for identifying branch parameters in power grids are primarily based on measurement data from supervisory control and data acquisition (SCADA) systems and synchronized vector measurement units (PMUs). These methods employ least squares or other optimization algorithms to solve the problem, thereby identifying branch parameters. However, these methods are often susceptible to the accuracy of the measurement data. When noise is present in the measurement data, the identification results obtained by these methods can exhibit significant deviations. With the development of artificial intelligence (AI) technology, many researchers have begun using deep learning methods for parameter identification, particularly those based on graph neural networks (GNNs), such as graph convolutional networks (GCNs) and graph attention networks (GATs). Graph neural networks utilize topological information and comprehensively consider measurement data from multiple adjacent branches, effectively smoothing out noise in the measurement data and improving parameter identification accuracy when the measurement data quality is poor.

[0004] However, the existing parameter identification methods based on graph neural networks still have shortcomings. The data quality of different branches is often different, and the distribution of noise is usually uneven. The data quality of some branches is good, and aggregating the information of other branches may introduce noise, resulting in the dilution of their own effective information. For some branches, the branch measurement data quality is poor within a certain spatial range. At this time, it is necessary to introduce information from farther branches to better suppress noise, but effective information will be lost in the process of propagation between different graph nodes. This also means that the node in the center of the noise area cannot fully utilize the effective information of farther branches, and the noise suppression ability is insufficient. In addition, in the existing parameter identification methods based on graph neural networks, the branch features extracted by the graph neural network are directly used for parameter regression without considering the dependency of each feature channel, which causes the model to be unable to focus on features that are more important to the task, affecting the identification accuracy. Summary of the Invention

[0005] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a power grid branch parameter identification method and system based on an adaptive multi-hop graph attention network, which can adapt to different noise conditions of different branches and reduce the error of power grid branch parameter identification when there is noise in the measurement data. At the same time, the dependency between each feature channel is considered so that the model can focus on the feature channels that are more important to the task, thereby further improving the identification accuracy.

[0006] In order to solve the above technical problems, the present invention is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present invention provides a method for identifying power grid branch parameters based on an adaptive multi-hop graph attention network, comprising:

[0008] Building a simulated power grid, using the simulated power grid to simulate and record measurement data of each branch to obtain a data set to be identified;

[0009] Preprocessing the data set to be identified to obtain a processed data set;

[0010] Constructing an adjacency matrix and a multi-hop matrix according to the topological structure of the simulated power grid;

[0011] The processed data set, adjacency matrix and multi-hop matrix are input into the trained power grid branch parameter identification model: based on the processed data set, adjacency matrix and multi-hop matrix, feature aggregation is performed through an adaptive multi-hop graph attention network module to obtain a power grid branch feature set; the power grid branch feature set is weighted using the SE attention module to obtain a power grid branch weighted feature set; the power grid branch weighted feature set is input into a multi-layer fully connected layer for regression to obtain predicted power grid branch parameters.

[0012] Optionally, the step of simulating the simulated power grid and recording the measurement data of each branch to obtain a data set to be identified includes:

[0013] The load node active power, load node reactive power, generator output and branch reactance parameters to be identified in the simulated power grid are randomly changed. A power flow calculation is performed every time a change occurs, and the measured data of each branch is recorded to obtain the data set to be identified.

[0014] Optionally, the preprocessing includes performing Z-Score normalization on the dataset to be identified to obtain a processed dataset;

[0015] The calculation formula for the Z-Score standardization is:

[0016]

[0017] in, represents the characteristics of the processed dataset, Represents the characteristics of the dataset to be identified, Represents the average value set of each feature in the data set to be identified, Represents the set of standard deviations of each feature in the dataset to be identified.

[0018] Optionally, constructing an adjacency matrix and a multi-hop matrix according to the topological structure of the simulated power grid includes:

[0019] According to the topological structure of the simulated power grid, a graph is constructed with branches as nodes and the adjacent relationships between branches as edges. ;in, Representation diagram, Represents a collection of nodes, represents a set of edges;

[0020] According to the diagram Constructing the adjacency matrix based on the adjacent relationship between the nodes ; Among them, the adjacency matrix The diagonal elements of are 0, , Indicates the nodes and There are edges between nodes. Indicates the nodes and There are no edges between the nodes. represents a real number, N is the number of graph nodes;

[0021] According to the adjacency matrix , calculate the intermediate matrix ;

[0022] The intermediate matrix Set the elements greater than or equal to 1 in the matrix to 1 and set the diagonal elements to 0 to obtain the multi-hop matrix .

[0023] Optionally, the feature aggregation through the adaptive multi-hop graph attention network module includes:

[0024] The processed dataset is input into the multi-layer adaptive multi-hop graph attention network layer. The grid branch feature set output by the adaptive multi-hop graph attention network layer is used as the first The input of the adaptive multi-hop graph attention network layer is used to traverse all the adaptive multi-hop graph attention network layers in turn to obtain the power grid branch feature set; among them, when When, The input of the layer-adaptive multi-hop graph attention network layer is the processed dataset;

[0025] The said The data processing flow of the layer-adaptive multi-hop graph attention network layer includes:

[0026] The said The grid branch feature set output by the adaptive multi-hop graph attention network layer is input to the The first branch of the adaptive multi-hop graph attention network layer uses a fully connected layer and ReLU activation function to extract features and obtain the first branch feature set;

[0027] The said The grid branch feature set and adjacency matrix output by the adaptive multi-hop graph attention network layer are input to the first The second branch of the layer-adaptive multi-hop graph attention network layer: add self-loops to the adjacency matrix to obtain a self-loop adjacency matrix; obtain the neighbor nodes of each node according to the self-loop adjacency matrix, calculate the attention score of each node and its neighbor nodes through the graph attention mechanism, and normalize the attention score to obtain the aggregation weight, aggregate the neighbor node features according to the aggregation weight as the output of a single attention head, splice or average the outputs of multiple single attention heads to obtain the second branch feature set; wherein, if the first The last layer is an adaptive multi-hop graph attention network layer, which obtains the second branch feature set by averaging the outputs of multiple single attention heads; otherwise, the second branch feature set is obtained by concatenating the outputs of multiple single attention heads;

[0028] The said The grid branch feature set and multi-hop matrix output by the adaptive multi-hop graph attention network layer are input to the The third branch of the adaptive multi-hop graph attention network layer: add self-loops to the multi-hop matrix to obtain a self-loop multi-hop matrix; obtain the multi-hop neighbor nodes of each node according to the self-loop multi-hop matrix, calculate the attention score of each node and its multi-hop neighbor nodes through the graph attention mechanism, and normalize the attention score to obtain the multi-hop aggregation weight, aggregate the multi-hop neighbor node features according to the multi-hop aggregation weight as the output of a single attention head, splice or average the outputs of multiple single attention heads to obtain the third branch feature set; wherein, if the first The layer is the last layer of the adaptive multi-hop graph attention network layer, which obtains the third branch feature set by averaging the outputs of multiple single attention heads; otherwise, the third branch feature set is obtained by concatenating the outputs of multiple single attention heads;

[0029] The first branch feature set, the second branch feature set and the third branch feature set are input into the Adaptive multi-hop information gating fusion module of the adaptive multi-hop graph attention network layer: After splicing the first branch feature set, the second branch feature set and the third branch feature set, input them into the first fully connected layer, and input them into the second fully connected layer after the ReLU activation function, and normalize them using the softmax function to obtain the normalized fusion weight; use the normalized fusion weight to perform weighted fusion on the first branch feature set, the second branch feature set and the third branch feature set to obtain the first branch feature set. The set of network branch features output by the layer-adaptive multi-hop graph attention network layer.

[0030] Optionally, the first branch feature set is obtained by the following formula:

[0031]

[0032] in, Indicates the The first branch feature set of the layer-adaptive multi-hop graph attention network layer, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, Indicates the The first of the power grid branch feature sets output by the layer-adaptive multi-hop graph attention network layer The characteristics of the nodes, represents the ReLU activation function, Indicates the The fully connected layer of the first branch of the adaptive multi-hop graph attention network layer;

[0033] Adding self-loops to the adjacency matrix is ​​achieved by the following formula:

[0034]

[0035] in, represents the adjacency matrix, represents the self-loop adjacency matrix, represents the identity matrix;

[0036] The graph attention mechanism is used to calculate the attention score of each node and its neighboring nodes using the following formula:

[0037]

[0038] in, Indicates the second branch Node and The attention score of each node, represents the LeakyReLU activation function, represents the weight parameter of the graph attention mechanism in the second branch, represents transpose, represents the weight matrix in the second branch, Indicates the The first of the power grid branch feature sets output by the adaptive multi-hop graph attention network layer The characteristics of the nodes, Represents a splicing operation;

[0039] The normalization of the attention score is achieved by the following formula:

[0040]

[0041] in, Indicates the second branch Node and The aggregate weight of the nodes, represents the softmax activation function, Indicates based on The obtained The set of neighbor nodes of a node, Indicates the second branch Node and The attention score of each node, represents the natural exponential function;

[0042] The second branch feature set is obtained by concatenating the outputs of multiple single attention heads using the following formula:

[0043] ,

[0044] in, represents the second branch feature set, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, represents the number of attention heads in the second branch, represents the attention head number, represents the ELU activation function;

[0045] The second branch feature set is obtained by averaging the outputs of multiple single attention heads using the following formula:

[0046] ;

[0047] The said The grid branch feature set output by the layer-adaptive multi-hop graph attention network layer is obtained by the following formula:

[0048]

[0049] in, Indicates the third branch feature set The characteristics of the nodes, represents the first fully connected layer, represents the second fully connected layer, Indicates the The fusion weight of the first branch feature of each node, Indicates the The fusion weight of the second branch feature of each node, Indicates the The fusion weight of the third branch feature of each node, Indicates the The grid branch feature set output by the layer-adaptive multi-hop graph attention network layer, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, is the splicing function.

[0050] Optionally, the weighting of the power grid branch feature set by using the SE attention module includes:

[0051] The power grid branch feature set is input into the graph convolution layer, and after the ReLU activation function, the attention score of each node in each channel is obtained;

[0052] The softmax function is used to normalize the attention scores of each node in each channel, and the normalized attention scores are used for weighted pooling to obtain compressed information representing each channel;

[0053] Use two fully connected layers to stimulate the compressed information of each channel and obtain the weight of each channel;

[0054] The power grid branch feature set is multiplied by the channel weights using the weights of the respective channels to obtain a power grid branch weighted feature set.

[0055] Optionally, the attention score of each node in each channel is obtained by the following formula:

[0056]

[0057] in, Indicates that each node is The attention score of each channel, represents the ReLU activation function, represents the graph convolutional layer, The first channels, represents the adjacency matrix, represents the self-loop adjacency matrix, express The degree matrix of Indicates the The weight parameters of the graph convolution layer corresponding to the channels;

[0058] The compression information of each channel is obtained by the following formula:

[0059]

[0060] in, Indicates the compression information of each channel, Indicates the compression information of the first channel, Indicates the compression information of the second channel, Indicates the The compressed information of each channel, Indicates the The compressed information of each channel, represents the total number of channels in the SE attention module, Indicates that each node is The normalized attention score of each channel, Indicates the The node in The normalized attention score of each channel, Indicates the first The node channels, Indicates the The node in The attention score of each channel, Indicates the The node in The attention score of each channel, Indicates the The node in The attention score of each channel, Indicates the The node in The normalized attention score of each channel, Indicates the The node in The normalized attention score of each channel, Indicates the The node in The normalized attention score of each channel;

[0061] The weight of each channel is obtained by the following formula:

[0062]

[0063] in, represents the weight of each channel, represents the Sigmoid activation function, represents the first fully connected layer in the SE attention module, Represents the second fully connected layer in the SE attention module;

[0064] The power grid branch weighted feature set is obtained by the following formula:

[0065]

[0066] in, represents the weighted feature set of the power grid branch, represents the weight of the first channel, represents the weight of the second channel, Indicates the The weight of each channel, Represents the first channel of the power grid branch feature set, The second channel representing the power grid branch feature set, The first channels.

[0067] Optionally, the training process of the power grid branch parameter identification model includes:

[0068] Obtain grid branch parameters and grid branch measurement data, construct training sets and validation sets using grid branch measurement data as features and branch parameters as labels;

[0069] Building a power branch parameter identification model, performing forward propagation on the power branch parameter identification model using the training set to obtain predicted power branch parameters; calculating the training loss using the MSE loss function based on the predicted power branch parameters and the labels of the training set, performing backpropagation based on the training loss, and updating the model parameters using the Adam optimizer;

[0070] The power grid branch parameter identification model is verified using the validation set, and the model weight with the best validation set index is selected to obtain the trained power grid branch parameter identification model;

[0071] The expression of the MSE loss function is:

[0072]

[0073] in, represents the training loss, represents the number of samples in the training set, Indicates the The labels of samples, Indicates the The predicted grid branch parameters of samples.

[0074] In a second aspect, the present invention provides a power grid branch parameter identification system based on an adaptive multi-hop graph attention network, which is applicable to any power grid branch parameter identification method based on an adaptive multi-hop graph attention network described in the first aspect, comprising:

[0075] The data acquisition module is used to: build a simulated power grid, use the simulated power grid to simulate and record the measurement data of each branch to obtain a data set to be identified;

[0076] A preprocessing module is used to preprocess the data set to be identified to obtain a processed data set;

[0077] A matrix construction module, configured to construct an adjacency matrix and a multi-hop matrix according to the topological structure of the simulated power grid;

[0078] An adaptive multi-hop graph attention network module is used to: perform feature aggregation based on the processed data set, the adjacency matrix, and the multi-hop matrix through the adaptive multi-hop graph attention network module to obtain a power grid branch feature set;

[0079] The SE attention module is configured to weight the power grid branch feature set to obtain a power grid branch weighted feature set;

[0080] The regression module is used to: input the power grid branch weighted feature set into a multi-layer fully connected layer for regression to obtain predicted power grid branch parameters.

[0081] Compared with the existing technology, the beneficial effects achieved by the present invention are:

[0082] 1. Adopting an adaptive multi-hop graph attention network module, it aggregates information from nodes with different hop counts, or branches, through different branches. The first branch of the adaptive multi-hop graph attention network layer is used to prevent certain nodes with higher data quality from having their effective information diluted due to the introduction of noise from other nodes. The second branch provides a certain degree of noise suppression capability. Compared with the third branch, the nodes that serve as information sources in the information aggregation process of the second branch are closer to the node itself, and their information is more relevant to the task objectives, making the aggregated node features more effective. The third branch is used to address regional noise issues. By directly connecting to two-hop neighbors, the number of network layers required to aggregate information from remote nodes is reduced, thereby reducing information loss. This allows nodes in the center of the noisy area to effectively utilize the effective information of more distant nodes, achieving the purpose of noise suppression. The adaptive multi-hop information gating fusion module dynamically fuses the node features output by different branches for each node, making the output node features adaptable to different node noise conditions, thereby improving the robustness and accuracy of parameter identification.

[0083] 2. The SE attention module based on self-attention weighted pooling takes into account the dependencies between feature channels and can dynamically adjust the weights of different channels, so that the model can focus on the information of important channels. In addition, unlike the global average pooling used in the original SE attention module, this invention uses weighted pooling based on self-attention and uses a graph convolution layer to calculate the attention score. It comprehensively considers node features and graph structure information, focuses on nodes that can better represent the global information of the channel, reduces the interference of nodes containing noise information on the pooling process, and thus obtains more effective channel weights, thereby improving the accuracy of model parameter identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 A flow chart of a method for identifying power grid branch parameters based on an adaptive multi-hop graph attention network according to an embodiment of the present invention;

[0085] Figure 2 A structural diagram of a power grid branch parameter identification model provided according to an embodiment of the present invention;

[0086] Figure 3 A flowchart of a power grid branch parameter identification model training according to an embodiment of the present invention;

[0087] Figure 4A diagram showing the structure of an adaptive multi-hop graph attention layer according to an embodiment of the present invention;

[0088] Figure 5 This is a structural diagram of the SE attention module provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0089] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present application and the specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. Unless there is a conflict, the embodiments of the present application and the technical features in the embodiments can be combined with each other.

[0090] It should be noted that the term "and / or" in this document simply describes an association between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0091] Example 1:

[0092] The embodiment of the present invention discloses a method for identifying parameters of power grid branches based on an adaptive multi-hop graph attention network. Figure 1 As shown, the specific steps include:

[0093] S1, build a simulated power grid, use the simulated power grid to simulate and record the measurement data of each branch to obtain the data set to be identified;

[0094] S2, preprocessing the data set to be identified to obtain the processed data set;

[0095] S3, constructing the adjacency matrix and multi-hop matrix according to the topological structure of the simulated power grid;

[0096] S4, input the processed data set, adjacency matrix and multi-hop matrix into the trained power branch parameter identification model: based on the processed data set, adjacency matrix and multi-hop matrix, feature aggregation is performed through the adaptive multi-hop graph attention network module to obtain the power branch feature set; the power branch feature set is weighted using the SE attention module to obtain the power branch weighted feature set; the power branch weighted feature set is input into the multi-layer fully connected layer for regression to obtain the predicted power branch parameters.

[0097] Specifically, in step S1, a simulated power grid is used to simulate and record the measurement data of each branch to obtain a data set to be identified, including:

[0098] The load node active power, load node reactive power, generator output and branch reactance parameters to be identified in the random change simulation power grid are calculated once each time a change occurs, and the measured data of each branch are recorded to obtain the data set to be identified; in this embodiment, the data set to be identified includes the node voltages at both ends of the branch and , Active power at both ends of the branch and and reactive power at both ends of the branch and , the parameter to be identified is the branch reactance X.

[0099] In step S2, preprocessing includes performing Z-Score normalization on the dataset to be identified to obtain a processed dataset;

[0100] The calculation formula for Z-Score standardization is:

[0101]

[0102] in, represents the characteristics of the processed dataset, Represents the characteristics of the dataset to be identified, Represents the average value set of each feature in the data set to be identified, Represents the set of standard deviations of each feature in the dataset to be identified.

[0103] In step S3, since the branch parameters to be identified are the characteristics of the edges of the power grid, the original power grid needs to be converted into a graph with branches as nodes and the adjacent relationships between branches as edges. The adjacency matrix and multi-hop matrix are constructed according to the topological structure of the simulated power grid, including:

[0104] According to the topological structure of the simulated power grid, a graph is constructed with branches as nodes and the adjacent relationships between branches as edges. ;in, Representation diagram, Represents a set of nodes, each node corresponds to a branch in the actual power grid, is an edge set, which represents the adjacency relationship between branches;

[0105] According to the figure Constructing the adjacency matrix based on the adjacent relationship between the nodes , the elements in the adjacency matrix The value of depends on whether the branches corresponding to the graph nodes are connected to the same grid node; , Indicates the nodes and There are edges between nodes. Indicates the nodes and There are no edges between the nodes. represents a real number, and the diagonal elements of the adjacency matrix A are set to 0, and N is the number of graph nodes;

[0106] In addition, it is necessary to calculate the multi-hop matrix including one- and two-hop neighbor nodes for the input of the subsequent multi-hop graph attention network; according to the adjacency matrix , calculate the intermediate matrix ;

[0107] The intermediate matrix Set the elements greater than or equal to 1 in the matrix to 1 and set the diagonal elements to 0 to obtain the multi-hop matrix .

[0108] In step S4, refer to Figure 2 As shown in the figure, the power grid branch parameter identification model mainly includes three parts. The first part is the adaptive multi-hop graph attention network module, which contains multiple layers of adaptive multi-hop graph attention network layers. Each layer contains three branches. The first branch only focuses on its own information, inputs the node features into the fully connected layer and the ReLU activation function as the output of the branch, and explicitly isolates the information of other nodes, thereby avoiding the dilution of the effective information of some nodes with higher data quality due to the introduction of noise from other nodes. The second branch uses the graph attention layer to aggregate its own information and one-hop neighbor information, and comprehensively considers the local spatial information to achieve a certain noise suppression capability. The graph attention mechanism can reduce the aggregation weight of neighboring nodes with higher noise. In addition, compared with the third branch, the nodes as the information source in the information aggregation process of the second branch are closer to their own nodes, and their information is more relevant to the task objectives. , the aggregated node features are more effective. The third branch uses the graph attention layer to aggregate the node's own information, one-hop neighbor information and two-hop neighbor information. By directly connecting the two-hop neighbors, it provides a larger information aggregation range. At the same time, it also reduces the loss of remote node information in the traditional multi-layer graph attention network during the propagation process. Even when the quality of the surrounding node data is poor, it can provide effective information, thereby suppressing noise. Then, the adaptive multi-hop information gating fusion module is used to dynamically fuse the node features extracted by the three branches. The adaptive multi-hop information gating fusion module calculates the fusion weights of the three branch output features for each node, and weightedly adds the output features of the three branches as the output features of the node in this layer, so that the output features of each node can adapt to the different noise conditions of different nodes.

[0109] The second part is the SE attention module (SE: Squeeze-and-Excitation Networks) based on self-attention weighted pooling. It is used to perform channel-wise weighting on the node features output by the first part. The module consists of three stages: Squeeze, Excitation, and Scale. In the Squeeze stage, self-attention-based weighted pooling replaces the global average pooling used in traditional SE attention, and a graph convolutional layer is used to calculate the attention score. This comprehensively considers node features and graph structure information, focusing on nodes that better represent the global channel information and reducing the interference of nodes containing noisy information on the pooling process. In the Excitation stage, the weights of each channel are calculated based on the vector output by the Squeeze stage. In the Scale stage, the obtained channel weights are used to weight the node features by channel, allowing the model to focus more on important feature channels.

[0110] The third part is N fully connected layers, which are used to output the parameters to be identified of N branches. The node features after channel weighting in the second part are used as the input of the fully connected layer. Each fully connected layer contains 3 layers, with the number of neurons being 256, 128, and 1 respectively. The first two hidden layers use the ReLU activation function, and the last layer outputs the branch reactance parameters predicted by the corresponding branch of the fully connected layer.

[0111] The adaptive multi-hop graph attention network module includes multiple layers of adaptive multi-hop graph attention network layers, refer to Figure 4 As shown in the figure, each adaptive multi-hop graph attention network layer includes an adaptive multi-hop information gating fusion module and three parallel branches. The adaptive multi-hop information gating fusion module is used to fuse the features of the three parallel branches and use them as the output of the adaptive multi-hop graph attention layer. Feature aggregation is performed through the adaptive multi-hop graph attention network module, including:

[0112] The processed dataset is input into the first branch of the adaptive multi-hop graph attention network layer. The first branch inputs the node features into the fully connected layer with 108 output channels and applies the ReLU activation function to the output to obtain the first branch feature set. This branch only focuses on the node information itself.

[0113] The processed data set and the adjacency matrix are input into the second branch of the adaptive multi-hop graph attention network layer: self-loops are added to the adjacency matrix to obtain the self-loop adjacency matrix; the neighbor nodes of each node are obtained according to the self-loop adjacency matrix, and the attention scores of each node and its neighbor nodes are calculated through the graph attention mechanism, and the attention scores are normalized to obtain the aggregation weights. The neighbor node features are aggregated according to the aggregation weights as the output of a single attention head, and the outputs of multiple single attention heads are spliced ​​or averaged to obtain the second branch feature set; among them, if the first The layer is the last layer of the adaptive multi-hop graph attention network layer, and the second branch feature set is obtained by averaging the outputs of multiple single attention heads; otherwise, the second branch feature set is obtained by concatenating the outputs of multiple single attention heads; in this embodiment, the graph attention layer contains 3 attention heads, each single attention head has 36 output channels, and the total number of output channels is 108. If this branch is located in the last layer of the adaptive multi-hop graph attention network layer, the output feature is the average of the outputs of multiple attention heads, that is, the number of output channels of each of the three attention heads is 108, and the final number of output channels is also 108. This branch focuses on the node's own information and one-hop neighbor information;

[0114] The processed data set and the multi-hop matrix are input into the third branch of the adaptive multi-hop graph attention network layer: self-loops are added to the multi-hop matrix to obtain a self-loop multi-hop matrix; the multi-hop neighbor nodes of each node are obtained according to the self-loop multi-hop matrix, and the attention scores of each node and its multi-hop neighbor nodes are calculated through the graph attention mechanism, and the attention scores are normalized to obtain the multi-hop aggregation weights. The multi-hop neighbor node features are aggregated according to the multi-hop aggregation weights as the output of a single attention head, and the outputs of multiple single attention heads are spliced ​​or averaged to obtain the third branch feature set; this branch focuses on the node's own information, one-hop neighbor information and two-hop neighbor information. The operation of the third branch is similar to that of the second branch, except that the scope of directly aggregating spatial information is expanded; among them, if the first branch The layer is the last layer of adaptive multi-hop graph attention network layer, which obtains the third branch feature set by averaging the outputs of multiple single attention heads; otherwise, the third branch feature set is obtained by concatenating the outputs of multiple single attention heads;

[0115] The first branch feature set, the second branch feature set and the third branch feature set are input into the adaptive multi-hop information gating fusion module: after the first branch feature set, the second branch feature set and the third branch feature set are spliced, they are input into the first fully connected layer, the output channel of the first fully connected layer is 3*108, and after the ReLU activation function, they are input into the second fully connected layer, the output channel is 3, and the attention scores of the three branches are obtained. They are normalized using the softmax function to obtain the fusion weights of each branch feature; the fusion weights are used to perform weighted fusion on the first branch feature set, the second branch feature set and the third branch feature set. In this process, the fusion weights are calculated for each node and weighted fusion is performed to adapt to the different noise conditions of different nodes to obtain the power grid branch feature set.

[0116] The first branch feature set is obtained by the following formula:

[0117]

[0118] in, Indicates the The first branch feature set of the layer-adaptive multi-hop graph attention network layer, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, Indicates the The first of the power grid branch feature sets output by the adaptive multi-hop graph attention network layer The characteristics of the nodes, represents the ReLU activation function, Indicates the The fully connected layer of the first branch of the adaptive multi-hop graph attention network layer;

[0119] Adding self-loops to the adjacency matrix is ​​achieved through the following formula:

[0120]

[0121] in, represents the adjacency matrix, represents the self-loop adjacency matrix, represents the identity matrix;

[0122] The attention score of each node and its neighboring nodes is calculated through the graph attention mechanism using the following formula:

[0123]

[0124] in, Indicates the second branch Node and The attention score of each node, represents the LeakyReLU activation function, represents the weight parameter of the graph attention mechanism in the second branch, represents transpose, represents the weight matrix in the second branch, Indicates the The first of the power grid branch feature sets output by the adaptive multi-hop graph attention network layer The characteristics of the nodes, Represents a splicing operation;

[0125] Normalizing the attention score is achieved through the following formula:

[0126]

[0127] in, Indicates the second branch Node and The aggregate weight of the nodes, represents the softmax activation function, Indicates based on The obtained The set of neighbor nodes of a node, Indicates the second branch Node and The attention score of each node, represents the natural exponential function;

[0128] The output of multiple single attention heads is concatenated to obtain the second branch feature set through the following formula: ,

[0129] in, represents the second branch feature set, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, represents the number of attention heads in the second branch, represents the attention head number, represents the ELU activation function;

[0130] The second branch feature set is obtained by averaging the outputs of multiple single attention heads using the following formula:

[0131] ;

[0132] No. The grid branch feature set output by the layer-adaptive multi-hop graph attention network layer is obtained by the following formula:

[0133]

[0134] in, Indicates the third branch feature set The characteristics of the nodes, represents the first fully connected layer, represents the second fully connected layer, Indicates the The fusion weight of the first branch feature of each node, Indicates the The fusion weight of the second branch feature of each node, Indicates the The fusion weight of the third branch feature of each node, Indicates the The grid branch feature set output by the layer-adaptive multi-hop graph attention network layer, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, is the splicing function.

[0135] refer to Figure 5 As shown, the SE attention module includes three stages: Squeeze (compression), Excitation (excitation) and Scale (channel weight multiplication). In the Squeeze stage, unlike the original SE attention module that directly uses global average pooling, the present invention adopts weighted pooling based on self-attention, and uses the graph convolution layer to calculate self-attention, comprehensively considering the node feature information and graph structure information, focusing on nodes that can better represent the global information of the channel, reducing the interference of node noise, so that the generated vector can more effectively compress the information of each channel; in the Excitation stage, a fully connected layer is used to predict the importance of each channel. First, a fully connected layer with a smaller number of output channels is used, followed by a ReLU activation function, and then a fully connected layer with the same number of original input channels is used. Finally, a Sigmoid activation function is used to obtain the weights of each channel; in the Scale stage, the predicted channel weights are multiplied by the original node features according to the channel to obtain channel-weighted node features, so that the model can pay more attention to the information of important channels.

[0136] The SE attention module is used to weight the feature set of the power grid branches, including:

[0137] The power grid branch feature set is input into the graph convolution layer, and after the ReLU activation function, the attention score of each node in each channel is obtained;

[0138] The softmax function is used to normalize the attention scores of each node in each channel, and the normalized attention scores are weighted pooled to obtain a vector representing the compressed information of each channel;

[0139] Use two fully connected layers to stimulate the compressed information of each channel and obtain the weight of each channel;

[0140] The weights of each channel are used to multiply the channel weights of the power grid branch feature set to obtain the power grid branch weighted feature set.

[0141] The attention score of each node in each channel is obtained by the following formula:

[0142]

[0143] in, Indicates that each node is The attention score of each channel, represents the ReLU activation function, represents the graph convolutional layer, The first channels, represents the adjacency matrix, represents the self-loop adjacency matrix, express The degree matrix of Indicates the The weight parameters of the graph convolution layer corresponding to the channels;

[0144] The compression information of each channel is obtained by the following formula:

[0145]

[0146] in, Indicates the compression information of each channel, Indicates the compression information of the first channel, Indicates the compression information of the second channel, Indicates the The compressed information of each channel, Indicates the The compressed information of each channel, represents the total number of channels in the SE attention module, Indicates that each node is The normalized attention score of each channel, Indicates the The node in The normalized attention score of each channel, Indicates the first The node channels, Indicates the The node in The attention score of each channel, Indicates the The node in The attention score of each channel, Indicates the The node in The attention score of each channel, Indicates the The node in The normalized attention score of each channel, Indicates the The node in The normalized attention score of each channel, Indicates the The node in The normalized attention score of each channel;

[0147] The weight of each channel is obtained by the following formula:

[0148]

[0149] in, represents the weight of each channel, represents the Sigmoid activation function, represents the first fully connected layer in the SE attention module, Represents the second fully connected layer in the SE attention module;

[0150] The weighted feature set of the power grid branch is obtained by the following formula:

[0151]

[0152] in, represents the weighted feature set of the power grid branch, represents the weight of the first channel, represents the weight of the second channel, Indicates the The weight of each channel, Represents the first channel of the power grid branch feature set, The second channel representing the power grid branch feature set, The first channels.

[0153] refer to Figure 3The figure shows the training process of the power grid branch parameter identification model. In this embodiment, the MATPOWER library is used to build a simulated power grid of the IEEE39 node system. The load node active power, load node reactive power, generator output and branch reactance parameters to be identified of the IEEE39 node system are randomly changed. A total of 19,200 times are performed, and the range of change is between 80% and 120% of the original value. After each change, a power flow calculation is performed and the measurement data are recorded: the voltage at the nodes at both ends of the branch and , Active power at both ends of the branch and , reactive power at both ends of the branch and , and branch parameters: reactance X, with measurement data as the features of the dataset and branch parameters as the labels of the dataset.

[0154] In addition, the characteristics of the dataset 、 、 、 、 、 Different types of noise were added to simulate the data noise commonly found in actual power grid measurements. The dataset was divided into training, validation, and test sets in a ratio of 8:1:1.

[0155] The training process of the power grid branch parameter identification model includes:

[0156] Obtain grid branch parameters and grid branch measurement data, construct training sets and validation sets using grid branch measurement data as features and branch parameters as labels;

[0157] A power branch parameter identification model was built and forward propagated using the training set to obtain predicted power branch parameters. The training loss was calculated using the MSE loss function based on the predicted power branch parameters and the labels of the training set. Backpropagation was performed based on the training loss, and the model parameters were updated using the Adam optimizer. The batch size was set to 64, the initial learning rate was set to 0.0025, and the learning rate was decayed every 10 epochs with a decay rate of 0.8. A total of 200 epochs were trained.

[0158] The power grid branch parameter identification model is verified using the validation set, and the model weight with the best validation set index is selected to obtain the trained power grid branch parameter identification model.

[0159] The expression of the MSE loss function is:

[0160]

[0161] in, represents the training loss, represents the number of samples in the training set, Indicates the The labels of samples, Indicates the The predicted grid branch parameters of samples.

[0162] Example 2:

[0163] Based on the same inventive concept as the first embodiment, the present invention discloses a power grid branch parameter identification system based on an adaptive multi-hop graph attention network, which is applicable to any power grid branch parameter identification method based on an adaptive multi-hop graph attention network in the first embodiment, including:

[0164] The data acquisition module is used to build a simulated power grid, simulate the simulated power grid and record the measurement data of each branch to obtain the data set to be identified;

[0165] A preprocessing module is used to preprocess the data set to be identified to obtain a processed data set;

[0166] The matrix construction module is used to: construct an adjacency matrix and a multi-hop matrix according to the topological structure of the simulated power grid;

[0167] An adaptive multi-hop graph attention network module is used to: perform feature aggregation based on the processed dataset, adjacency matrix, and multi-hop matrix through the adaptive multi-hop graph attention network module to obtain a set of power grid branch features;

[0168] The SE attention module is used to: weight the power grid branch feature set to obtain the power grid branch weighted feature set;

[0169] The regression module is used to input the weighted feature set of the power grid branch into the multi-layer fully connected layer for regression to obtain the predicted power grid branch parameters.

[0170] The specific functional implementation of each of the above modules can be found in the relevant content of the method in Example 1 and will not be elaborated on here.

[0171] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0172] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0173] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0174] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0175] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which are all protected by the present invention.

Claims

1. A method for identifying power grid branch parameters based on an adaptive multi-hop graph attention network, characterized in that: include: Building a simulated power grid, using the simulated power grid to simulate and record measurement data of each branch to obtain a data set to be identified; Preprocessing the data set to be identified to obtain a processed data set; Constructing an adjacency matrix and a multi-hop matrix according to the topological structure of the simulated power grid; Inputting the processed data set, the adjacency matrix, and the multi-hop matrix into a trained power branch parameter identification model; performing feature aggregation through an adaptive multi-hop graph attention network module based on the processed data set, the adjacency matrix, and the multi-hop matrix to obtain a power branch feature set; Using the SE attention module to weight the power grid branch feature set to obtain a power grid branch weighted feature set; inputting the power grid branch weighted feature set into a multi-layer fully connected layer for regression to obtain predicted power grid branch parameters; The feature aggregation is performed by the adaptive multi-hop graph attention network module, including: The processed dataset is input into the multi-layer adaptive multi-hop graph attention network layer. The grid branch feature set output by the adaptive multi-hop graph attention network layer is used as the first The input of the adaptive multi-hop graph attention network layer is used to traverse all the adaptive multi-hop graph attention network layers in turn to obtain the power grid branch feature set; among them, when When, The input of the layer-adaptive multi-hop graph attention network layer is the processed dataset; The said The data processing flow of the layer-adaptive multi-hop graph attention network layer includes: The said The grid branch feature set output by the adaptive multi-hop graph attention network layer is input to the The first branch of the adaptive multi-hop graph attention network layer uses a fully connected layer and ReLU activation function to extract features and obtain the first branch feature set; The said The grid branch feature set and adjacency matrix output by the adaptive multi-hop graph attention network layer are input to the first The second branch of the layer-adaptive multi-hop graph attention network layer: add self-loops to the adjacency matrix to obtain a self-loop adjacency matrix; obtain the neighbor nodes of each node according to the self-loop adjacency matrix, calculate the attention score of each node and its neighbor nodes through the graph attention mechanism, and normalize the attention score to obtain the aggregation weight, aggregate the neighbor node features according to the aggregation weight as the output of a single attention head, splice or average the outputs of multiple single attention heads to obtain the second branch feature set; wherein, if the first The last layer is an adaptive multi-hop graph attention network layer, which obtains the second branch feature set by averaging the outputs of multiple single attention heads; otherwise, the second branch feature set is obtained by concatenating the outputs of multiple single attention heads; The said The grid branch feature set and multi-hop matrix output by the adaptive multi-hop graph attention network layer are input to the The third branch of the adaptive multi-hop graph attention network layer: add self-loops to the multi-hop matrix to obtain a self-loop multi-hop matrix; obtain the multi-hop neighbor nodes of each node according to the self-loop multi-hop matrix, calculate the attention score of each node and its multi-hop neighbor nodes through the graph attention mechanism, and normalize the attention score to obtain the multi-hop aggregation weight, aggregate the multi-hop neighbor node features according to the multi-hop aggregation weight as the output of a single attention head, splice or average the outputs of multiple single attention heads to obtain the third branch feature set; wherein, if the first The layer is the last layer of the adaptive multi-hop graph attention network layer, which obtains the third branch feature set by averaging the outputs of multiple single attention heads; otherwise, the third branch feature set is obtained by concatenating the outputs of multiple single attention heads; The first branch feature set, the second branch feature set and the third branch feature set are input into the Adaptive multi-hop information gating fusion module of the adaptive multi-hop graph attention network layer: After splicing the first branch feature set, the second branch feature set and the third branch feature set, input them into the first fully connected layer, and input them into the second fully connected layer after the ReLU activation function, and normalize them using the softmax function to obtain the normalized fusion weight; use the normalized fusion weight to perform weighted fusion on the first branch feature set, the second branch feature set and the third branch feature set to obtain the first branch feature set. The set of network branch features output by the layer-adaptive multi-hop graph attention network layer.

2. The method for identifying power grid branch parameters based on adaptive multi-hop graph attention network according to claim 1 is characterized in that: The method of using the simulated power grid to simulate and record the measurement data of each branch to obtain a data set to be identified includes: The load node active power, load node reactive power, generator output and branch reactance parameters to be identified in the simulated power grid are randomly changed. A power flow calculation is performed every time a change occurs, and the measured data of each branch is recorded to obtain the data set to be identified.

3. The method for identifying power grid branch parameters based on adaptive multi-hop graph attention network according to claim 1, characterized in that: The preprocessing includes performing Z-Score standardization on the dataset to be identified to obtain a processed dataset; The calculation formula for the Z-Score standardization is: ,in, represents the characteristics of the processed dataset, Represents the characteristics of the dataset to be identified, Represents the average value set of each feature in the data set to be identified, Represents the set of standard deviations of each feature in the dataset to be identified.

4. The method for identifying power grid branch parameters based on adaptive multi-hop graph attention network according to claim 1, characterized in that: The constructing of an adjacency matrix and a multi-hop matrix according to the topological structure of the simulated power grid includes: According to the topological structure of the simulated power grid, a graph is constructed with branches as nodes and the adjacent relationships between branches as edges. ;in, Representation diagram, Represents a collection of nodes, represents an edge set; According to the diagram Constructing the adjacency matrix based on the adjacent relationship between the nodes ; Among them, the adjacency matrix The diagonal elements of are 0, , Indicates the nodes and There are edges between nodes. Indicates the nodes and There are no edges between the nodes. represents a real number, N is the number of graph nodes; According to the adjacency matrix , calculate the intermediate matrix ; The intermediate matrix Set the elements greater than or equal to 1 in the matrix to 1 and set the diagonal elements to 0 to obtain the multi-hop matrix .

5. The method for identifying power grid branch parameters based on adaptive multi-hop graph attention network according to claim 1, characterized in that: The first branch feature set is obtained by the following formula: ， in, Indicates the The first branch feature set of the layer-adaptive multi-hop graph attention network layer, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, Indicates the The first of the power grid branch feature sets output by the layer-adaptive multi-hop graph attention network layer The characteristics of the nodes, represents the ReLU activation function, Indicates the The fully connected layer of the first branch of the adaptive multi-hop graph attention network layer; Adding self-loops to the adjacency matrix is ​​achieved by the following formula: ， in, represents the adjacency matrix, represents the self-loop adjacency matrix, represents the identity matrix; The graph attention mechanism is used to calculate the attention score of each node and its neighboring nodes using the following formula: ， in, Indicates the second branch Node and The attention score of each node, represents the LeakyReLU activation function, represents the weight parameter of the graph attention mechanism in the second branch, represents transpose, represents the weight matrix in the second branch, Indicates the The first of the power grid branch feature sets output by the layer-adaptive multi-hop graph attention network layer The characteristics of the nodes, Represents a splicing operation; The normalization of the attention score is achieved by the following formula: ， in, Indicates the second branch Node and The aggregate weight of the nodes, represents the softmax activation function, Indicates based on The obtained The set of neighbor nodes of a node, Indicates the second branch Node and The attention score of each node, represents the natural exponential function; The second branch feature set is obtained by concatenating the outputs of multiple single attention heads using the following formula: ， in, represents the second branch feature set, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, represents the number of attention heads in the second branch, represents the attention head number, represents the ELU activation function; The second branch feature set is obtained by averaging the outputs of multiple single attention heads using the following formula: ； The said The grid branch feature set output by the layer-adaptive multi-hop graph attention network layer is obtained by the following formula: ， in, Indicates the third branch feature set The characteristics of the nodes, represents the first fully connected layer, represents the second fully connected layer, Indicates the The fusion weight of the first branch feature of each node, Indicates the The fusion weight of the second branch feature of each node, Indicates the The fusion weight of the third branch feature of each node, Indicates the The grid branch feature set output by the layer-adaptive multi-hop graph attention network layer, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, express Middle The characteristics of the nodes, is the splicing function.

6. The method for identifying power grid branch parameters based on adaptive multi-hop graph attention network according to claim 1, characterized in that: The using the SE attention module to weight the power grid branch feature set includes: The power grid branch feature set is input into the graph convolution layer, and after the ReLU activation function, the attention score of each node in each channel is obtained; The softmax function is used to normalize the attention scores of each node in each channel, and the normalized attention scores are used for weighted pooling to obtain compressed information representing each channel; Use two fully connected layers to stimulate the compressed information of each channel and obtain the weight of each channel; The power grid branch feature set is multiplied by the channel weights using the weights of the respective channels to obtain a power grid branch weighted feature set.

7. The method for identifying power grid branch parameters based on an adaptive multi-hop graph attention network according to claim 6, characterized in that: The attention score of each node in each channel is obtained by the following formula: ， in, Indicates that each node is The attention score of each channel, represents the ReLU activation function, represents the graph convolutional layer, The first channels, represents the adjacency matrix, represents the self-loop adjacency matrix, express The degree matrix of Indicates the The weight parameters of the graph convolution layer corresponding to the channels; The compression information of each channel is obtained by the following formula: ， in, Indicates the compression information of each channel, Indicates the compression information of the first channel, Indicates the compression information of the second channel, Indicates the The compressed information of each channel, Indicates the The compressed information of each channel, represents the total number of channels in the SE attention module, Indicates that each node is The normalized attention score of each channel, Indicates the The node in The normalized attention score of each channel, Indicates the first The node channels, Indicates the The node in The attention score of each channel, Indicates the The node in The attention score of each channel, Indicates the The node in The attention score of each channel, Indicates the The node in The normalized attention score of each channel, Indicates the The node in The normalized attention score of each channel, Indicates the The node in The normalized attention score of each channel; The weight of each channel is obtained by the following formula: ， in, represents the weight of each channel, represents the Sigmoid activation function, represents the first fully connected layer in the SE attention module, Represents the second fully connected layer in the SE attention module; The power grid branch weighted feature set is obtained by the following formula: ， in, represents the weighted feature set of the power grid branch, represents the weight of the first channel, represents the weight of the second channel, Indicates the The weight of each channel, Represents the first channel of the power grid branch feature set, The second channel representing the power grid branch feature set, The first channels.

8. The method for identifying power grid branch parameters based on an adaptive multi-hop graph attention network according to claim 1, characterized in that: The training process of the power grid branch parameter identification model includes: Obtain grid branch parameters and grid branch measurement data, construct training sets and validation sets using grid branch measurement data as features and branch parameters as labels; Building a power branch parameter identification model, performing forward propagation on the power branch parameter identification model using the training set to obtain predicted power branch parameters; calculating the training loss using the MSE loss function based on the predicted power branch parameters and the labels of the training set, performing backpropagation based on the training loss, and updating the model parameters using the Adam optimizer; The power grid branch parameter identification model is verified using the validation set, and the model weight with the best validation set index is selected to obtain the trained power grid branch parameter identification model; The expression of the MSE loss function is: ， in, represents the training loss, represents the number of samples in the training set, Indicates the The labels of samples, Indicates the The predicted grid branch parameters of samples.

9. A power grid branch parameter identification system based on adaptive multi-hop graph attention network, characterized in that: A method for identifying power grid branch parameters based on an adaptive multi-hop graph attention network for implementing any one of claims 1 to 8, comprising: The data acquisition module is used to: build a simulated power grid, use the simulated power grid to simulate and record the measurement data of each branch to obtain a data set to be identified; A preprocessing module is used to preprocess the data set to be identified to obtain a processed data set; A matrix construction module, configured to construct an adjacency matrix and a multi-hop matrix according to the topological structure of the simulated power grid; An adaptive multi-hop graph attention network module is used to: perform feature aggregation based on the processed data set, the adjacency matrix, and the multi-hop matrix through the adaptive multi-hop graph attention network module to obtain a power grid branch feature set; The SE attention module is configured to weight the power grid branch feature set to obtain a power grid branch weighted feature set; The regression module is used to: input the power grid branch weighted feature set into a multi-layer fully connected layer for regression to obtain predicted power grid branch parameters.

Images