Method for power system dynamic topology and line parameter identification based on graph embedding learning

Abstract

This invention belongs to the field of power system parameter identification technology and discloses a method for identifying dynamic topology and line parameters of power systems based on graph embedding learning. The method involves: extracting initial line parameters using a π-type equivalent circuit model; constructing a node feature matrix by combining node voltage, power, and line parameters of adjacent branches, and normalizing the node feature matrix; constructing a joint identification model to reconstruct the topology graph and correct node features; using a regression model for line parameter prediction and evaluation; and using RMSE to evaluate the identification accuracy of the joint identification model. This invention designs a graph embedding learning framework based on first-order and second-order similarity. The joint identification model constructed based on this framework can effectively capture the local and global characteristics of the power grid structure, integrate power grid topology information and electrical parameters, and significantly improve the accuracy of parameter identification under conditions of sparse monitoring data.

Description

Technical Field

[0001] This invention belongs to the field of power system parameter identification technology, specifically involving a method for identifying dynamic topology and line parameters of power systems based on graph embedding learning. Background Technology

[0002] In recent years, distributed generation (DG) integration into the power grid has brought significant advantages; however, it has also presented enormous challenges to maintaining grid stability, reliability, and efficiency. One key challenge is the accurate identification of line parameters. Accurate line parameter identification not only improves the accuracy of grid state estimation but also effectively assists in the safe grid connection of DG, optimizes power distribution and dispatch strategies, and reduces system operational risks. Accurate parameter identification is closely related to the grid's topology information; fully utilizing topology information is crucial for improving parameter identification accuracy. With the increasing penetration rate of DG, the grid topology is exhibiting increasingly dynamic characteristics, making traditional parameter identification methods based on static, fixed topologies inadequate for adapting to this change.

[0003] Parameter identification methods can be broadly classified into three categories: (1) physical model-based methods; (2) numerical optimization or statistical inference-based methods; and (3) artificial intelligence-based methods. Physical model-based methods generally perform power flow calculations or matrix analysis on the power system, which cannot cope with the complex and dynamic changes that actually exist in the power grid. For the second type of method, the parameter identification task is usually regarded as an optimization problem, and the system state and parameters are inferred by using intelligent optimization algorithms or constructing optimization objective functions.

[0004] In recent years, there has been some analysis and research on power system parameter identification based on artificial intelligence technologies and algorithms. However, these methods mainly rely on static topology structures, making it difficult to fully capture real-time topology changes in the power grid caused by distributed generation and other dynamic factors. Graph embedding has made significant progress in the past decade, becoming an important means of learning low-dimensional representations of nodes, edges, or the entire graph while preserving structural and relational information. However, how to use it to cope with the sparsity of node features in power systems remains an unsolved problem.

[0005] Therefore, accurately estimating these parameters is crucial for enabling real-time monitoring, fault detection, and system control of smart grids. Meanwhile, data collection plays a vital role in parameter identification. Various methods exist for capturing grid measurements, typically relying on phasor measurement units (PMUs), Supervisory and Data Acquisition Systems (SCADA) and smart meters (SMs). While PMUs provide high-precision data, including voltage phase angles, their high infrastructure and maintenance costs make large-scale deployment in the grid difficult. This results in incomplete or biased topology information of the grid in scenarios where actual monitoring data is sparse, thus affecting the accuracy of parameter identification using traditional methods.

[0006] Therefore, how to solve the problem of line parameter identification under the condition of sparse monitoring data, so as to improve the accuracy of parameter identification, is the technical problem that this invention aims to solve. Summary of the Invention

[0007] The purpose of this invention is to provide a method for identifying dynamic topology and line parameters of power systems based on graph embedding learning, so as to solve the problems mentioned in the background art.

[0008] The objective of this invention is achieved as follows: a method for identifying dynamic topology and line parameters of power systems based on graph embedding learning, characterized by the following steps:

[0009] Step S1: Extract initial line parameters using a π-type equivalent circuit model;

[0010] Step S2: Construct a node feature matrix by combining the node voltage, power, and line parameters of adjacent branches, and then normalize the node feature matrix.

[0011] Step S3: Construct a joint identification model and reconstruct the topology graph to correct node features;

[0012] Step S4: Use a regression model to predict and evaluate the line parameters, and use RMSE to evaluate the identification accuracy of the joint identification model.

[0013] Preferably, in step S1, the initial line parameters are extracted using a π-type equivalent circuit model, specifically as follows:

[0014] Step S1-1: Using the π-type equivalent circuit model, obtain the power at both ends of the line, specifically:

[0015] Node The power at that point is denoted as active power. and reactive power The current amplitude was calculated. :

[0016] ;

[0017] ;

[0018] ;

[0019] in, For the electricity meter at the node Active power monitored at the location For the electricity meter at the node Reactive power monitored at the location For the electricity meter at the node Voltage monitored at the location;

[0020] Node The power at that point is denoted as active power. and reactive power The current amplitude was calculated. :

[0021] ;

[0022] ;

[0023] ;

[0024] in, For the electricity meter at the node Active power monitored at the location For the electricity meter at the node Reactive power monitored at the location For the electricity meter at the node Voltage monitored at the location;

[0025] According to the π-type equivalent circuit model, the sum of the power at the nodes at both ends of the line is equal to the power consumed by the impedance and admittance. The expression for the sum of the power at the nodes at both ends of the line is:

[0026] ;

[0027] in, , Represent Dot and Complex apparent power at a point; This represents the equivalent impedance of the line; The susceptance at both ends of the line; For the nodes that flow through The current vector; , These are the nodes at both ends of the line. and nodes The voltage; The imaginary unit;

[0028] Step S1-2: Calculate the resistance and reactance parameters to obtain the line conductance parameters. .

[0029] Preferably, in step S1-2, the resistance and reactance parameters are calculated to obtain the line conductance parameters. Specifically:

[0030] Decomposing power into active power and reactive power, we obtain the following two sets of equations:

[0031] ;

[0032] ;

[0033] in, The susceptance at both ends of the line; For resistance; For reactance; For nodes active power, For nodes active power, For nodes reactive power, For nodes reactive power, For the electricity meter at the node Voltage monitored at the location For the electricity meter at the node Voltage monitored at the location;

[0034] The susceptance obtained from solving the system of equations ,resistance and reactance The conductance is obtained by combining the expressions for the parameters with the impedance-admittance relationship. :

[0035] .

[0036] Preferably, in step S2, the node voltage, power, and line parameters of adjacent branches are used to construct a node feature matrix, and the node feature matrix is ​​normalized, specifically as follows:

[0037] Node voltage on the node side Active power reactive power The line parameters of the adjacent branches are combined to form the node feature vector:

[0038] ;

[0039] in, , , Representing the nodes respectively Average resistance, reactance, and susceptance parameters of connected branches; , Represents the transpose of a matrix;

[0040] By summarizing all node features, a node feature matrix is ​​constructed:

[0041] ;

[0042] in, This represents the total number of nodes in the power grid;

[0043] The node feature matrix is ​​normalized to eliminate the dimensional differences between different feature components. The 1st normalized node feature is then calculated. Each component is expressed as: ,in, It is the first node feature before reduction One portion, and These refer to the first of all node features. The minimum and maximum values ​​of each characteristic component;

[0044] Normalize each component of all node features according to the above formula, and recombine them according to the row and column dimensions of the original node feature matrix to obtain the normalized node feature matrix.

[0045] Preferably, the joint identification model includes an encoder module, a decoder module, and a feature correction and optimization module, specifically:

[0046] The encoder module has a multi-layer nonlinear mapping structure, specifically including:

[0047] Input layer: Receives the adjacency matrix of the original power grid topology and the normalized node feature matrix;

[0048] Multi-layer nonlinear feature compression layer: The input data is progressively reduced in dimensionality and compressed in a layer-by-layer nonlinear transformation through a fully connected nonlinear network layer;

[0049] Output layer: Outputs the dimensionality-reduced low-dimensional embedding vector matrix;

[0050] The decoder module specifically includes:

[0051] Feature reconstruction layer: The structure is symmetrical to the multi-layer nonlinear feature compression layer of the encoder. It receives the low-dimensional embedding vector matrix output by the encoder output layer, and performs the inverse process of encoding through a fully connected network with an increasing number of nodes and nonlinear mapping to complete the decompression of high-dimensional features and the restoration of feature correlation between nodes.

[0052] Output layer: The output dimension is the same as the input dimension of the encoder input layer, and it is used to output the probability distribution of the reconstructed adjacency matrix after recovery;

[0053] The feature correction and optimization module specifically includes:

[0054] Neighborhood sampling unit: Based on the reconstructed adjacency matrix output by the decoder, the set of neighboring nodes of the target node is filtered according to the reconstructed topological connection weights to determine the local receptive field of the node;

[0055] Feature aggregation unit: aggregates the electrical features of all nodes in the neighborhood to generate an intermediate feature representation of the target node;

[0056] State Update Unit: Through a learnable linear transformation, the neighborhood aggregation features and the original node features are weighted and fused to output the corrected node feature vector.

[0057] Preferably, in step S3, constructing a joint identification model and reconstructing the topology graph to correct node features specifically involves:

[0058] Step S3-1: Based on the power grid topology, construct unweighted and weighted networks, and calculate the first-order and second-order similarity of the unweighted and weighted networks, specifically as follows:

[0059] Define an unweighted network ,in, It is a vertex set. It is an edge set, for Each edge in Assign binary weights , used to represent Nodes and The connection relationships between nodes

[0060] Calculate Nodes and Similarity from a node to all nodes.

[0061] Define a weighted network ,in, It is a vertex set. It is an edge set. It is the set of weights of all edges in a weighted network, used to characterize the degree of electrical characteristic correlation between nodes;

[0062] Binary weights based on unweighted networks Construct a binary adjacency matrix The binary adjacency matrix No. Line number Column elements ;

[0063] The node feature vectors in the normalized node feature matrix are used to calculate the node similarity using the cosine similarity formula. With nodes Weights between , to indicate Nodes and First-order similarity between nodes:

[0064] ;

[0065] First-order similarity is calculated using cosine similarity:

[0066] ;

[0067] Second-order similarity is calculated using cosine similarity:

[0068] ;

[0069] in, For nodes With nodes Feature vector and The angle between them For nodes With nodes The angle between the first-order similarity distribution vectors, , They are nodes ,node Normalized feature vectors For feature vectors The One portion, For feature vectors The One portion, is the dimension of the node feature vector. For nodes With the first in the power grid First-order similarity of nodes For nodes With the first in the power grid First-order similarity of nodes This represents the total number of nodes in the power grid.

[0070] Step S3-2: Convert the binary adjacency matrix The normalized node feature matrix is ​​input into the encoder simultaneously, and a low-dimensional embedding vector matrix is ​​constructed using the low-dimensional embedding vector output by the encoder. Specifically:

[0071] The encoding stage extracts the implicit structural relationships between nodes through layer-by-layer nonlinear transformation, as shown in the formula:

[0072] ;

[0073] ;

[0074] in, This is the initial input; Representing the The weight matrix of the layer is used to weight the first layer. The input features of the layer undergo a linear transformation; Representing the The input features of the layer; Representing the The bias vector of the layer is used to provide the bias vector of the first layer. Add a bias term to the linear transformation result of the layer; For activation functions;

[0075] In the nonlinear transformation of each layer, the distance consistency of local features between nodes is constrained by first-order similarity, and the distribution consistency of global structure between nodes is constrained by second-order similarity, so that the reduced features retain both the local connectivity characteristics and the global topology characteristics of the power grid.

[0076] After multiple mappings, a low-dimensional embedding vector is obtained for each node. The low-dimensional embedding vectors of all nodes are then stacked row by row in node number order to obtain a low-dimensional embedding vector matrix. ;

[0077] Step S3-3: The decoding module receives the embedded vector matrix. The reconstruction process is optimized by using a hybrid loss function to decode and reconstruct the low-dimensional embedding vector matrix into a new adjacency matrix. ;

[0078] Step S3-4: Based on the reconstructed adjacency matrix The newly constructed topology graph corrects the initial node feature data through sampling and aggregation methods.

[0079] Preferably, in step S3-3, the decoding module receives the embedding vector matrix. The reconstruction process is optimized by using a hybrid loss function to decode and reconstruct the low-dimensional embedding vector matrix into a new adjacency matrix. Specifically:

[0080] In the joint identification model, the decoder receives a low-dimensional embedding vector matrix containing the structural and node feature information of the power grid after encoding and fusion. ;

[0081] The reconstruction process is supervised and optimized using a hybrid loss function, the formula of which is: ,

[0082] in, A hybrid loss function to guide the decoder in reconstructing the adjacency matrix;

[0083] This is the weight for the first-order similarity loss, used to balance the proportion of the first-order similarity loss in the total loss;

[0084] It is a first-order similarity loss, expressed as: ,in, express Nodes and The first-order similarity value between nodes Corresponding embedding vector matrix The Middle row nodes Structural characterization Corresponding embedding vector matrix The Middle row nodes Structural characterization;

[0085] This is the weight for the second-order similarity loss, used to balance the proportion of the second-order similarity loss in the total loss;

[0086] It is a second-order similarity loss, expressed as: ,in, Represents the original adjacency matrix. The reconstructed adjacency matrix output by the decoder is the decoder's reconstruction of the original adjacency matrix. Prediction results after topology reconstruction Represents the weight matrix;

[0087] This is the weight of the regularization term, used to balance the proportion of regularization loss in the total loss;

[0088] yes Regularization, used to prevent overfitting of the model, is expressed as: ,in, Indicates the first The weight matrix of the layer, Indicates the first The bias matrix of the layer, This represents the total number of layers in the model.

[0089] By minimizing this loss function, the weights and bias parameters of the encoder and decoder are iteratively updated. Training stops when the loss function converges to a preset threshold or reaches the maximum number of iterations, and the decoder outputs the reconstructed adjacency matrix.

[0090] Preferably, in steps S3-4, the reconstructed adjacency matrix is ​​used as the basis for... The newly constructed topology graph refines the initial node feature data through sampling and aggregation methods, specifically:

[0091] The topology graph uses power grid nodes as vertices to reconstruct the adjacency matrix. The element value in the table is the connection weight of the edge, used to determine the node. With nodes Are there any topological connections between them?

[0092] Based on the reconstructed adjacency matrix Update features using the new topology graph:

[0093] The local receptive domain of a node is determined by sampling from its neighbors.

[0094] ;

[0095] in, Represents a node The set of neighbors; The reconstructed adjacency matrix The element values ​​in the table represent the nodes in the reconstructed topology. With nodes Connection weights between them;

[0096] The features of neighboring nodes are weighted and aggregated to obtain the nodes. Intermediate feature representation:

[0097] ;

[0098] in, For nodes In the The eigenvectors of the next iteration; Represents a node The set of neighbors; Represents nodes in the reconstructed topology With nodes Connection weights between them; It is a learnable linear transformation weight matrix; It is a non-linear activation function;

[0099] The aggregated features are then fused with the original features of the nodes to achieve dynamic correction of the node features:

[0100] ;

[0101] Among them, nodes initial input features The nodes obtained after normalization eigenvectors, This is to aggregate and update the node features. For balance coefficient, For nodes The corrected features.

[0102] Preferably, in step S4, a regression model is used to predict and evaluate the line parameters, specifically as follows:

[0103] The node topology is abstracted into a line graph topology. Based on the reconstructed topology, the corrected node features are reorganized into line feature vectors, specifically:

[0104] The node topology is abstracted into a line graph topology, ignoring the original node topology, treating lines as new nodes, and extending node features to line features.

[0105] Line susceptance The line parameters of branch conductance are used as new nodes, and the node characteristics are extended to line characteristics.

[0106] Using the predictive variables at both ends of the line, based on The line susceptance is calculated using a regression model. ;

[0107] Combined with line susceptance ,based on Branch conductance is calculated using a regression model. ;

[0108] RMSE is used to evaluate the quality of model fit, specifically:

[0109] ;

[0110] in, For the sample size, It is the first Each sample corresponds to the actual value of the electrical parameters of the transmission line. It is the first The predicted values ​​of electrical parameters of the transmission line corresponding to each sample.

[0111] Compared with the prior art, the present invention has the following improvements and advantages:

[0112] 1. The line parameter identification method based on graph embedding learning proposed in this invention can learn the distributed representation of power grid nodes in a low-dimensional vector space. Based on the electrical attributes of the nodes and the topology information of the network, it integrates the power grid topology information and electrical parameters through the collected power and voltage data, effectively inferring and completing missing information, and significantly improving the accuracy of parameter identification in the case of sparse monitoring data.

[0113] 2. This invention designs a graph embedding learning framework based on first-order and second-order similarity. First-order similarity is used to represent the local direct connection characteristics of nodes, and second-order similarity is used to represent the global neighborhood structural similarity of nodes. This allows the dimensionality-reduced features to retain both the local connection characteristics and global topological characteristics of the power grid, thereby constructing a topology identification model and further improving the accuracy of parameter estimation.

[0114] 3. This invention corrects node characteristics by reconstructing the topology graph, which can adapt to real-time changes in the network topology caused by noise or distributed generation (DG) in the power grid, and outputs high-precision and high-reliability line parameters verified by root mean square error (RMSE). This significantly improves the accuracy of parameter identification in dynamic power grids with sparse monitoring data, and enhances the stability and reliability of power grid operation. Attached Figure Description

[0115] Figure 1 This is a technical architecture diagram of the present invention;

[0116] Figure 2a This is a schematic diagram of the original adjacency matrix before dynamic identification in the IEEE 118-node system.

[0117] Figure 2b This is a schematic diagram of the reconstructed corrected adjacency matrix in the IEEE 118-node system.

[0118] Figure 3 This is a schematic diagram comparing the RMSE of line conductance under noise-free conditions.

[0119] Figure 4a This is a schematic diagram comparing RMSE under 0.2% Gaussian noise.

[0120] Figure 4b This is a schematic diagram showing the comparison of RMSE under 0.5% Gaussian noise;

[0121] Figure 4c This is a schematic diagram showing the RMSE comparison under 1% Gaussian noise;

[0122] Figure 4d This is a schematic diagram comparing RMSE under missing line features;

[0123] Figure 5a Schematic diagram of the IEEE 118 node feeder for connecting to DG;

[0124] Figure 5b This is a diagram showing the comparison of RMSE after integrating DG. Detailed Implementation

[0125] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0126] like Figure 1As shown, the present invention discloses a method for identifying dynamic topology and line parameters of a power system based on graph embedding learning. The dynamic topology is achieved through topology reconstruction. The identification method improves the accuracy and robustness of power system line parameter identification through graph embedding learning and topology reconstruction, and includes the following steps:

[0127] Step S1: Extract initial line parameters using a π-type equivalent circuit model;

[0128] Step S2: Construct a node feature matrix by combining the node voltage, power, and line parameters of adjacent branches, and then normalize the node feature matrix.

[0129] Step S3: Construct a joint identification model and reconstruct the topology graph to correct node features;

[0130] Step S4: Use a regression model to predict and evaluate the line parameters, and use RMSE to evaluate the identification accuracy of the joint identification model.

[0131] Step S1-1: Using the π-type equivalent circuit model, obtain the power at both ends of the line, specifically:

[0132] For the transmission line to be identified, the original monitoring nodes at both ends are defined as nodes. and nodes ,node For the line side The inner equivalent node of the equivalent circuit, node For the line side The inner equivalent node of the equivalent circuit;

[0133] From node To the node Electricity meter measurement data The power remains unchanged, but due to the impact on ground admittance, the node... The power at that point is denoted as active power. and reactive power The nodes were calculated. Current amplitude :

[0134] ;

[0135] ;

[0136] ;

[0137] Similarly, the nodes are derived. Current amplitude at :

[0138] ;

[0139] ;

[0140] ;

[0141] in, For the electricity meter at the node Active power monitored at the location For the electricity meter at the node Reactive power monitored at the location For the electricity meter at the node Voltage monitored at the location For the electricity meter at the node Active power monitored at the location For the electricity meter at the node Reactive power monitored at the location For the electricity meter at the node Voltage monitored at the location; and The nodes that flow through are respectively and The current vector, and The nodes that flow through are respectively and The current vector, , and , satisfy The circuit constraint relationship, both are used to characterize The current distribution relationship between the node-to-ground admittance branch and the line series branch in the equivalent circuit provides circuit boundary conditions for the subsequent solution of line impedance parameters, and can also be used to verify the current conservation relationship at both ends of the line.

[0142] According to the π-type equivalent circuit model, the sum of the power at the nodes at both ends of the line is equal to the power consumed by the impedance and admittance. The expression for the sum of the power at the nodes at both ends of the line is:

[0143] ;

[0144] in, , Represent Dot and Complex apparent power at a point; This represents the equivalent impedance of the line; The susceptance at both ends of the line; For the nodes that flow through The current vector; , These are the nodes at both ends of the line. and nodes The voltage; The imaginary unit;

[0145] In step S1-2, the resistance and reactance parameters are calculated to obtain the line conductance parameters. Specifically:

[0146] Based on the derivation in step S1-1 The power balance equation for the equivalent circuit shows that the complex apparent power can be decomposed into active power components and reactive power components, and the line equivalent impedance... It can also be broken down into resistance components. and reactance component The complex equations for the total power at both ends of the line obtained in step S1-1 are decoupled and decomposed according to the real active power and the imaginary reactive power respectively. Substituting these into the formulas for calculating the amplitude of the series branch current derived in step S1-1, the following two sets of equations are obtained:

[0147] ;

[0148] ;

[0149] in, The susceptance at both ends of the line; For resistance; For reactance; For nodes active power, For nodes active power, For nodes reactive power, For nodes reactive power, For the electricity meter at the node Voltage monitored at the location For the electricity meter at the node Voltage monitored at the location;

[0150] The aforementioned steps have yielded the line resistance by solving the active and reactive power balance equations. Reactance With susceptance To fully extract all parameters of the equivalent line, it is necessary to further solve for the conductance, which characterizes the active power loss of the line. This is achieved by simultaneously solving the equivalent transformation relationship between line impedance and admittance and the nodal power balance equations, and then performing the following solution operations. :

[0151] Step S1-2-1: Based on The impedance-admittance equivalent relationship of the equivalent circuit is established, and the constraint equations for the series impedance and parallel admittance of the line are established:

[0152] ;

[0153] in, The equivalent admittance of the series branch of the line is given. The equivalent impedance of the series branch of the line is obtained through the previously solved... , The constraint relationship between the real and imaginary parts of the admittance can be determined;

[0154] Step S1-2-2: Combining the total active power balance equations at both ends of the line, , Substituting the constraint relationships, the conductivity is derived. The solution formula is as follows:

[0155] ;

[0156] Step S1-2-3: Apply the solution obtained above , , Substituting into the above formula, we can complete the conductivity calculation. The numerical solution yields the complete initial equivalent parameters of the line, providing a complete electrical parameter basis for subsequent node feature construction.

[0157] Step S2: Construct a node feature matrix from the node voltage, power, and line parameters of adjacent branches, and normalize the node feature matrix, specifically as follows:

[0158] Based on step S1, the node voltage on the node side is... Active power reactive power The line parameters of the adjacent branches are combined to form the node feature vector:

[0159] ;

[0160] in, , , Representing the nodes respectively The average resistance, reactance, and susceptance parameters of the connected branches, , It is the transpose matrix;

[0161] After summarizing all node features, a node feature matrix is ​​constructed:

[0162] ;

[0163] in, This represents the total number of nodes in the power grid.

[0164] The node feature matrix is ​​normalized to eliminate the dimensional differences between different feature components, resulting in a normalized feature matrix:

[0165] ,in, It is the first of the reduced node features One portion, It is the first node feature before reduction One portion, and These refer to the first of all node features. The minimum and maximum values ​​of each characteristic component;

[0166] After normalizing each component of all node features according to the above formula, the components are recombined according to the row and column dimensions of the original node feature matrix to obtain the complete normalized feature matrix.

[0167] Step S3: Construct a joint identification model and reconstruct the topology graph to correct node features. The structure of the joint identification model is as follows:

[0168] The joint identification model is cascaded in the logical order of topological feature encoding, topological structure reconstruction, and node feature correction, and includes an encoder module, a decoder module, and a feature correction and optimization module. The specific composition of each module is as follows:

[0169] The encoder module is built based on the Structural Deep Network Embedding (SDNE) algorithm. It is a cascaded multi-layer nonlinear mapping structure, specifically consisting of an input layer, multiple nonlinear feature compression layers, and an output layer.

[0170] The input layer serves as the model input entry point, used to receive the adjacency matrix of the original power grid topology and the node feature matrix normalized in step S2;

[0171] The multi-layer nonlinear feature compression layer is a continuously stacked fully connected nonlinear network layer. It progressively reduces the dimensionality and compresses the high-dimensional sparse input data through layer-by-layer nonlinear transformation. In the nonlinear transformation of each layer, the distance consistency of local features between nodes is constrained by first-order similarity to represent the local direct connection characteristics between nodes, and the distribution consistency of global structure between nodes is constrained by second-order similarity to represent the similarity of global neighborhood structure between nodes. This allows the dimensionality-reduced features to retain both the local connection characteristics and global topological characteristics of the power grid. The first-order and second-order similarity are integrated as core structural constraints into the embedding vector learning process of the encoder and the topology reconstruction optimization process of the decoder.

[0172] The output layer is used to output the dimensionality-reduced low-dimensional dense node embedding vector matrix to complete the extraction of key topology information of the power grid.

[0173] The decoder module adopts a multi-layer fully connected network structure symmetrical to the encoder structure, specifically consisting of a feature reconstruction layer and an output layer, as follows:

[0174] The feature reconstruction layer is symmetrical to the multi-layer nonlinear feature compression layer structure of the encoder. It receives the low-dimensional embedding vector output by the encoder output layer and performs the inverse process of encoding through a fully connected network with an increasing number of nodes and nonlinear mapping to complete the decompression of high-dimensional features and the restoration of feature correlation between nodes.

[0175] The output layer, whose output dimension is exactly the same as the input dimension of the encoder input layer, is used to output the probability distribution of the reconstructed adjacency matrix after recovery, so as to complete the reconstruction of the power grid topology.

[0176] The feature correction and optimization module is logically cascaded according to sampling-aggregation-update, consisting of a neighborhood sampling unit, a feature aggregation unit, and a state update unit, as follows:

[0177] The neighborhood sampling unit uses the reconstructed adjacency matrix output by the decoder as a basis, and filters the set of neighboring nodes of the target node according to the reconstructed topological connection weights to determine the local receptive field of the node.

[0178] The feature aggregation unit aggregates the electrical features of all nodes in the neighborhood in a weighted or pooled manner to generate an intermediate feature representation of the target node;

[0179] The state update unit uses a learnable linear transformation to weightedly fuse neighborhood aggregation features with the original node features, and finally outputs the corrected node feature vector.

[0180] Step S3-1: Based on the power grid topology, construct unweighted and weighted networks, and calculate the first-order and second-order similarity of the unweighted and weighted networks, specifically as follows:

[0181] First, define a weightless network. ,in, It is a vertex set, corresponding to the nodes in the power grid. It is an edge set, corresponding to the lines in the power grid. Each edge in Assign binary weights ,node With nodes There is a physical electrical connection between them. When there is no physical electrical connection , used to represent Nodes and The connection relationships between nodes;

[0182] Based on the binary connectivity of the unweighted network, the following is calculated: Nodes and Similarity from a node to all nodes;

[0183] Define a weighted network ,in, It is a vertex set. It is an edge set. It is the set of weights of all edges in a weighted network. The weight values ​​are calculated from the electrical feature similarity between nodes and are used to characterize the degree of correlation of electrical features between nodes.

[0184] The node feature vectors in the normalized node feature matrix obtained in step S2 are used to calculate the node similarity using the cosine similarity formula. With nodes Weights between , to indicate Nodes and First-order similarity between nodes:

[0185] ;

[0186] First-order similarity is calculated using cosine similarity:

[0187] ;

[0188] in, For nodes With nodes Feature vector and The angle between them , They are nodes ,node The eigenvectors normalized in step S2, For feature vectors The One portion, For feature vectors The One portion, The dimension of the node feature vector;

[0189] Second-order similarity is calculated using cosine similarity:

[0190] ;

[0191] in, For nodes With nodes The angle between the first-order similarity distribution vectors, For nodes With the first in the power grid First-order similarity of nodes For nodes With the first in the power grid First-order similarity of nodes This represents the total number of nodes in the power grid.

[0192] Step S3-2: Simultaneously input the binary adjacency matrix corresponding to the unweighted network constructed in Step S3-1, along with the normalized node feature matrix, into the encoder. Utilize the low-dimensional embedding vector output by the encoder to construct a low-dimensional embedding vector matrix. Specifically:

[0193] The binary adjacency matrix for The two-dimensional matrix is ​​formed by the binary weights of the unweighted network in step S3-1. The transformation yields: the first element in the matrix Line number Column elements When node With nodes When there is a physical electrical connection When there is no physical electrical connection , used to fully characterize the node connection relationships of the original topology of the power grid;

[0194] The encoder model is trained using first- and second-order similarity, and the loss function guides the feature mapping iteration of each layer in reverse.

[0195] The encoding stage extracts the implicit structural relationships between nodes through layer-by-layer nonlinear transformation, as shown in the formula:

[0196] ;

[0197] ;

[0198] in, Representing the The weight matrix of the layer is used to weight the first layer. The input features of the layer are linearly transformed to achieve layer-by-layer compression of the feature dimension. At the same time, the topological relationship between nodes is learned, which is the core learnable parameter of the encoder's nonlinear mapping. Representing the The input features of the layer, where the initial input is The binary adjacency matrix corresponding to the unweighted network of the power grid, the input of the subsequent layer is the output feature of the previous layer, carrying the power grid topology information extracted layer by layer; Representing the The bias vector of the layer is used to provide the bias vector of the first layer. Adding a bias term to the linear transformation results of the layer improves the encoder's nonlinear fitting capability and adapts to the complex nonlinear characteristics of the power grid topology. The ReLU function is chosen as the activation function to introduce a nonlinear mapping into the linear transformation result, enabling the encoder to learn the higher-order nonlinear structural correlations in the power grid topology and avoid the limitations of the expressive power of the linear model.

[0199] The encoder employs a 3-5 cascaded fully connected network to achieve the aforementioned layer-by-layer mapping. The first layer is the input layer, the middle two to three layers are non-linear feature compression layers, and the last layer is the output layer. The mapping process starts from the initial input... Initially, each layer undergoes linear transformation and nonlinear activation using the formula described above. The feature dimension decreases progressively with each layer. After passing through a multi-layer mapping encoder, the output layer generates a corresponding low-dimensional embedding vector for each node. The low-dimensional embedding vectors of all nodes are then stacked row by row in node number order to obtain a low-dimensional embedding vector matrix. ,matrix for dimension( This represents the total number of power grid nodes. (where is the dimension of the embedding vector), each row corresponds to the structural representation of a node in the embedding space; simultaneously, this embedding vector matrix It integrates the local first-order similarity features and the global second-order similarity structure information of the power grid, and serves as the input feature carrier for the subsequent step S3-3.

[0200] Step S3-3: The decoding module receives the embedded vector matrix. The reconstruction process is optimized by using a hybrid loss function to decode and reconstruct the low-dimensional embedding vector matrix into a new adjacency matrix. Specifically:

[0201] The decoder employs a 3-5 layer cascaded fully connected network structure symmetrical to the encoder. The layer-by-layer mapping process is as follows: first, a low-dimensional embedding vector matrix is ​​received through the feature reconstruction layer. The high-dimensional features are decompressed through fully connected layers with progressively increasing numbers of neurons and nonlinear mapping, restoring the feature associations and connections between nodes. Finally, the probability distribution of the reconstructed adjacency matrix is ​​output by the output layer with dimensions completely consistent with the original input adjacency matrix.

[0202] The above reconstruction process is supervised and optimized using a hybrid loss function, the formula of which is:

[0203] ;

[0204] in, To guide the decoder in reconstructing the adjacency matrix, the hybrid loss function is used. , and For hyperparameters, The range of values ​​is ;

[0205] This is the weight for the first-order similarity loss, used to balance the proportion of the first-order similarity loss in the total loss;

[0206] It is a first-order similarity loss. ,

[0207] in, express Nodes and The first-order similarity value between nodes is derived from the cosine similarity result between nodes calculated in step S3-1, and is used to characterize the degree of local connection between nodes. and They represent Nodes and The node's embedding vector output by the encoder, and the two are the low-dimensional embedding vector matrices ultimately output by the encoder. The row vector, Corresponding matrix The Middle row nodes Structural characterization Corresponding matrix The Middle row nodes Structural characterization;

[0208] This is the weight for the second-order similarity loss, used to balance the proportion of the second-order similarity loss in the total loss;

[0209] It is a second-order similarity loss. , Represents the original binary adjacency matrix. The reconstructed adjacency matrix output by the decoder is the decoder's reconstruction of the original binary adjacency matrix. Prediction results after topology reconstruction Represents the weight matrix, weight matrix The rule for the element values ​​is as follows: for the original adjacency matrix There are physically connected edges (i.e.) Set a higher penalty coefficient, take ( In this embodiment, the value is 10; for edges without physical connections (i.e. )Pick This is used to strengthen the accuracy constraints on the reconstruction of real topology connections;

[0210] This is the weight of the regularization term, used to balance the proportion of regularization loss in the total loss;

[0211] yes Regularization terms are used to prevent the model from overfitting.

[0212] , This represents the total number of layers in the model. Indicates the first The weight matrix of the layer, Indicates the first The bias matrix of the layer, and both are learnable parameters of the network. They are initialized by uniform distribution of Xavier before model training, and iteratively updated by backpropagation of the hybrid loss function during training, eventually converging to the optimal value that adapts to the characteristics of the power grid topology.

[0213] The hybrid loss function is minimized by using the Adam optimizer with gradient descent. The weights and bias parameters of the encoder and decoder are updated iteratively. Training stops when the loss function converges to a preset threshold or reaches the maximum number of iterations. The decoder outputs the reconstructed adjacency matrix.

[0214] Step S3-4: Based on the reconstructed adjacency matrix The newly constructed topology graph refines the initial node feature data through sampling and aggregation methods, specifically:

[0215] The topology graph uses power grid nodes as vertices to reconstruct the adjacency matrix. The element values ​​in the table are the connection weights of the edges. When the value exceeds a preset threshold, the node is determined. With nodes There are topological connections between them, thus completing the transformation from reconstructing the adjacency matrix to the topological graph;

[0216] Using the reconstructed topology graph, the initial node characteristics can be reversed based on sampling and aggregation mechanisms, making the characteristics more consistent with the actual operating state of the power grid under the new topology. (Reconstructed adjacency matrix) The structural relationships are used to guide feature updates. First, the local receptive domain of a node is determined through neighbor sampling:

[0217] ;

[0218] in, Represents a node The set of neighbors; Represents nodes in the reconstructed topology With nodes The connection weights between nodes are determined; then, the features of neighboring nodes are weighted and aggregated to obtain the node... Intermediate feature representation:

[0219] ;

[0220] in, For nodes In the The eigenvectors of the next iteration; Represents a node Neighbor set The number of nodes in; Represents nodes in the reconstructed topology With nodes Connection weights between them; It is a learnable linear transformation weight matrix; For non-linear activation functions, ReLU, Sigmoid, or Tanh functions can be selected;

[0221] The aggregated features are then fused with the original features of the nodes to achieve dynamic correction of the node features:

[0222] ;

[0223] in, For nodes The initial input features are specifically the nodes obtained after normalization in step S2. eigenvectors; This refers to the aggregated and updated node features; This is the balance coefficient, and its value range is... ; For nodes The corrected features.

[0224] Through the above three steps of sampling, aggregation, and fusion, the adaptive updating of node features in the reconstructed topology is achieved.

[0225] Step S4: Use a regression model to predict and evaluate line parameters, specifically:

[0226] The original power grid's node topology is abstracted into a line graph topology: ignoring the original node topology, each transmission line in the original power grid is taken as a new node in the line graph topology, and the connection relationship between the lines and nodes in the original power grid is taken as the edge connection relationship of the line graph topology; based on the reconstructed topology, the node feature vectors of the nodes at both ends of the line, after being corrected in step S3, are concatenated to form the line feature vector of the corresponding line, thus completing the expansion from node features to line features;

[0227] Using the predictive variables at both ends of the line, based on The line susceptance is calculated using a regression model. ;

[0228] Combined with line susceptance ,based on Branch conductance is calculated using a regression model. ;

[0229] The regression model can be a multiple linear regression, ridge regression, or deep regression model.

[0230] To mitigate the impact of measurement noise and missing data on errors, regularization constraints or weight penalty terms can be introduced into the model to improve its robustness and generalization ability. The model's predictive performance is evaluated using the root mean square error (RMSE) metric, specifically:

[0231] ;

[0232] in: For the sample size, It is the first Each sample corresponds to the actual values ​​of electrical parameters such as resistance, reactance, susceptance, and conductance of the transmission line. It is the first output of the regression model Each sample corresponds to a predicted value of the electrical parameters of the transmission line. The smaller the RMSE value, the closer the line parameter identification result is to the true value.

[0233] To verify the effectiveness of the present invention, an experimental scenario was constructed based on an IEEE 118-bus feeder network. Modified node feature time-series data was used, and the code implementation employed the PyTorch framework. In the decoding section, the goal was to minimize the loss value, and a traversal method was employed to set... , ; Figure 2a This illustrates the adjacency relationships between nodes in the original power grid topology under the IEEE 118-bus feeder. Figure 2b This demonstrates the node adjacency relationships reconstructed after joint identification using parameter-topology reconstruction.

[0234] This invention uses precision, recall, and F1-score as metrics to evaluate the results of topology identification, as shown in the following expressions:

[0235] ;

[0236] in, , and These represent precision, recall, and F1 score, respectively. , and These represent the number of correctly identified edges, the number of incorrectly added edges, and the number of missed edges, respectively.

[0237] The various metrics after topology reconstruction and identification are shown in Table 1:

[0238] Table 1. Indicators for Topology Reconstruction Identification

[0239]

[0240] Combination Figure 2a , Figure 2b As shown in Table 1, after joint identification via parameter-topology reconstruction, the reconstructed adjacency matrix differs from the original adjacency matrix and is closer to the actual topology. The reconstructed adjacency matrix more dynamically reflects changes in network topology. After obtaining the reconstructed node topology adjacency matrix, the node feature data is corrected in the optimization part using sampling and aggregation methods.

[0241] In simple power grids, line parameters can be directly calculated using physical models, such as the equivalent circuit of a line. As the number of nodes in the power grid increases and the grid structure becomes more complex, regression methods are needed to predict branch parameters. For example... Figure 3 The diagram compares the RMSE of using only the regression algorithm versus reconstructing the feature data before using the regression algorithm. Using corrected data effectively reduces the RMSE, especially showing significant improvement in traditional regression algorithms. Further analysis of the adaptability of different algorithms to corrected data reveals that deep learning methods exhibit more significant performance improvements after data correction.

[0242] To simulate random errors caused by signal interference and environmental influences in real-world scenarios, Gaussian noise can be added to node features. Gaussian noise is typically assumed to have a mean of 0 and a variance of 0. The probability density function of the given information follows a normal distribution.

[0243] ;

[0244] in, Let the noise be a random variable; the node features can be represented as follows after adding noise:

[0245] ;

[0246] In the formula: , and All independently obey Distribution; In order to simulate the small error caused by measurement and the large error caused by environmental anomalies, the following types of noise were added to the example data: 1) Gaussian noise of 0.2%, 0.5%, and 1% was added to the node features respectively; 2) Node feature data was randomly selected and set to zero to simulate the scenario of lost meter data in real-world scenarios; The comparison charts of RMSE under 0.2%, 0.5%, and 1% Gaussian noise and missing line features are shown in 4a, 4b, 4c, and 4d respectively.

[0247] The IEEE 118-bus feeder has 118 nodes and 186 branches. With a random seed of 42, the resulting training set contains 159 branches, and the test set contains 27 branches. It can be seen that the corrected node feature data significantly improves the performance of traditional regression methods, such as linear regression, especially when the topology changes. This difference is because traditional algorithms like linear regression are less robust to noise and missing data, while deep learning methods, with their complex model structures, can adapt to more dynamic changes and noise, demonstrating a stronger performance improvement.

[0248] To simulate the impact of distributed generation injection power on the characteristics of the original nodes, distributed generation devices (DGs) with higher penetration rates were introduced into nodes (3, 5, 12, 14, 18, 34, 56, 78, 90, 112), while distributed generation devices with lower penetration rates were introduced into node (64, 82), ensuring that the selected nodes simultaneously covered nodes in both the training and test sets. Figure 5a , Figure 5b As shown, if only classic regression algorithms or machine learning / deep learning algorithms are used, the RMSE value will increase to varying degrees, with the classic regression algorithm showing the largest increase in error. In the three scenarios tested above, the corrected feature data can improve the accuracy of various algorithms, comprehensively enhancing the parameter identification accuracy.

[0249] Besides accuracy, timeliness is equally important for line parameter identification. As power system networks become increasingly complex, the demand for fast algorithms capable of real-time monitoring of line parameters is constantly growing, thus supporting timely subsequent calculations. Table 2 summarizes the execution times of several methods for the IEEE 118-bus test case:

[0250] Table 2 Comparison of computational efficiency of several algorithms

[0251]

[0252] Table 2 shows that the classical regression method takes the shortest time during regression, while the deep learning LSTM algorithm takes the longest time during regression. The overall time increases after joint identification via parameter-topology reconstruction. For the IEEE 118-bus network, after encoding, decoding, and optimization operations, the required time is approximately 3 seconds. Therefore, the identification model proposed in this application is lightweight and can be used for dynamic identification of large-scale power system networks.

[0253] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for identifying dynamic topology and line parameters of a power system based on graph embedding learning, characterized in that: The method includes the following steps: Step S1: Extract initial line parameters using a π-type equivalent circuit model; Step S2: Construct a node feature matrix by combining the node voltage, power, and line parameters of adjacent branches, and then normalize the node feature matrix. Step S3: Construct a joint identification model and reconstruct the topology graph to correct node features; Step S4: Use a regression model to predict and evaluate the line parameters, and use RMSE to evaluate the identification accuracy of the joint identification model.

2. The method for identifying dynamic topology and line parameters of a power system based on graph embedding learning according to claim 1, characterized in that: In step S1, the initial line parameters are extracted using a π-type equivalent circuit model, specifically as follows: Step S1-1: Using the π-type equivalent circuit model, obtain the power at both ends of the line, specifically: Node The power at that point is denoted as active power. and reactive power The current amplitude was calculated. : ； ； ； in, For the electricity meter at the node Active power monitored at the location For the electricity meter at the node Reactive power monitored at the location For the electricity meter at the node Voltage monitored at the location; Node The power at that point is denoted as active power. and reactive power The current amplitude was calculated. : ； ； ； in, For the electricity meter at the node Active power monitored at the location For the electricity meter at the node Reactive power monitored at the location For the electricity meter at the node Voltage monitored at the location; According to the π-type equivalent circuit model, the sum of the power at the nodes at both ends of the line is equal to the power consumed by the impedance and admittance. The expression for the sum of the power at the nodes at both ends of the line is: ； in, , Represent Dot and Complex apparent power at a point; This represents the equivalent impedance of the line; The susceptance at both ends of the line; For the nodes that flow through The current vector; , These are the nodes at both ends of the line. and nodes The voltage; The imaginary unit; Step S1-2: Calculate the resistance and reactance parameters to obtain the line conductance parameters. .

3. The method for identifying dynamic topology and line parameters of a power system based on graph embedding learning according to claim 2, characterized in that: In step S1-2, the resistance and reactance parameters are calculated to obtain the line conductance parameters. Specifically: Decomposing power into active power and reactive power, we obtain the following two sets of equations: ； ； in, The susceptance at both ends of the line; For resistance; For reactance; For nodes active power, For nodes active power, For nodes reactive power, For nodes reactive power, For the electricity meter at the node Voltage monitored at the location For the electricity meter at the node Voltage monitored at the location; The susceptance obtained from solving the system of equations ,resistance and reactance The conductance is obtained by combining the expressions for the parameters with the impedance-admittance relationship. : 。 4. The method for identifying dynamic topology and line parameters of a power system based on graph embedding learning according to claim 1, characterized in that: In step S2, the node voltage, power, and line parameters of adjacent branches are used to construct a node feature matrix, and the node feature matrix is ​​then normalized. Specifically: Node voltage on the node side Active power reactive power The line parameters of the adjacent branches are combined to form the node feature vector: ； in, , , Representing the nodes respectively Average resistance, reactance, and susceptance parameters of connected branches; , Represents the transpose of a matrix; By summarizing all node features, a node feature matrix is ​​constructed: ； in, This represents the total number of nodes in the power grid; The node feature matrix is ​​normalized to eliminate the dimensional differences between different feature components. The 1st normalized node feature is then calculated. Each component is expressed as: ,in, It is the first node feature before reduction One portion, and These refer to the first of all node features. The minimum and maximum values ​​of each characteristic component; Normalize each component of all node features according to the above formula, and recombine them according to the row and column dimensions of the original node feature matrix to obtain the normalized node feature matrix.

5. The method for identifying dynamic topology and line parameters of a power system based on graph embedding learning according to claim 1, characterized in that: The joint identification model includes an encoder module, a decoder module, and a feature correction and optimization module, specifically: The encoder module has a multi-layer nonlinear mapping structure, specifically including: Input layer: Receives the adjacency matrix of the original power grid topology and the normalized node feature matrix; Multi-layer nonlinear feature compression layer: The input data is progressively reduced in dimensionality and compressed in a layer-by-layer nonlinear transformation through a fully connected nonlinear network layer; Output layer: Outputs the dimensionality-reduced low-dimensional embedding vector matrix; The decoder module specifically includes: Feature reconstruction layer: The structure is symmetrical to the multi-layer nonlinear feature compression layer of the encoder. It receives the low-dimensional embedding vector matrix output by the encoder output layer, and performs the inverse process of encoding through a fully connected network with an increasing number of nodes and nonlinear mapping to complete the decompression of high-dimensional features and the restoration of feature correlation between nodes. Output layer: The output dimension is the same as the input dimension of the encoder input layer, and it is used to output the probability distribution of the reconstructed adjacency matrix after recovery; The feature correction and optimization module specifically includes: Neighborhood sampling unit: Based on the reconstructed adjacency matrix output by the decoder, the set of neighboring nodes of the target node is filtered according to the reconstructed topological connection weights to determine the local receptive field of the node; Feature aggregation unit: aggregates the electrical features of all nodes in the neighborhood to generate an intermediate feature representation of the target node; State Update Unit: Through a learnable linear transformation, the neighborhood aggregation features and the original node features are weighted and fused to output the corrected node feature vector.

6. The method for identifying dynamic topology and line parameters of a power system based on graph embedding learning according to claim 5, characterized in that: In step S3, the joint identification model is constructed, and the topology graph is reconstructed to correct node features, specifically as follows: Step S3-1: Based on the power grid topology, construct unweighted and weighted networks, and calculate the first-order and second-order similarity between the unweighted and weighted networks. Specifically: Define an unweighted network ,in, It is a vertex set. It is an edge set, for Each edge in Assign binary weights , used to represent Nodes and The connection relationships between nodes Calculate Nodes and Similarity from a node to all nodes. Define a weighted network ,in, It is a vertex set. It is an edge set. It is the set of weights of all edges in a weighted network, used to characterize the degree of electrical characteristic correlation between nodes; Binary weights based on unweighted networks Construct a binary adjacency matrix The binary adjacency matrix No. Line number Column elements ; The node feature vectors in the normalized node feature matrix are used to calculate the node similarity using the cosine similarity formula. With nodes Weights between , to indicate Nodes and First-order similarity between nodes: ； First-order similarity is calculated using cosine similarity: ； Cosine similarity is used to calculate second-order similarity: ； in, For nodes With nodes Feature vector and The angle between them For nodes With nodes The angle between the first-order similarity distribution vectors, , They are nodes ,node Normalized feature vectors For feature vectors The One portion, For feature vectors The One portion, is the dimension of the node feature vector. For nodes With the first in the power grid First-order similarity of nodes For nodes With the first in the power grid First-order similarity of nodes This represents the total number of nodes in the power grid. Step S3-2: Convert the binary adjacency matrix The normalized node feature matrix is ​​input into the encoder simultaneously, and a low-dimensional embedding vector matrix is ​​constructed using the low-dimensional embedding vector output by the encoder. Specifically: The encoding stage extracts the implicit structural relationships between nodes through layer-by-layer nonlinear transformation, as shown in the formula: ； ； in, This is the initial input; Representing the The weight matrix of the layer is used to weight the first layer. The input features of the layer undergo a linear transformation; Representing the The input features of the layer; Representing the The bias vector of the layer is used to provide the bias vector of the first layer. Add a bias term to the linear transformation result of the layer; For activation functions; In the nonlinear transformation of each layer, the distance consistency of local features between nodes is constrained by first-order similarity, and the distribution consistency of global structure between nodes is constrained by second-order similarity, so that the reduced features retain both the local connectivity characteristics and the global topology characteristics of the power grid. After multiple mappings, a low-dimensional embedding vector is obtained for each node. The low-dimensional embedding vectors of all nodes are then stacked row by row in node number order to obtain a low-dimensional embedding vector matrix. ; Step S3-3: The decoding module receives the embedded vector matrix. The reconstruction process is optimized by using a hybrid loss function to decode and reconstruct the low-dimensional embedding vector matrix into a new adjacency matrix. ; Step S3-4: Based on the reconstructed adjacency matrix The newly constructed topology graph corrects the initial node feature data through sampling and aggregation methods.

7. The method for identifying dynamic topology and line parameters of a power system based on graph embedding learning according to claim 6, characterized in that: In step S3-3, the decoding module receives the embedding vector matrix. The reconstruction process is optimized by using a hybrid loss function to decode and reconstruct the low-dimensional embedding vector matrix into a new adjacency matrix. Specifically: In the joint identification model, the decoder receives a low-dimensional embedding vector matrix containing the structural and node feature information of the power grid after encoding and fusion. ; The reconstruction process is supervised and optimized using a hybrid loss function, the formula of which is: ; in, A hybrid loss function to guide the decoder in reconstructing the adjacency matrix; This is the weight for the first-order similarity loss, used to balance the proportion of the first-order similarity loss in the total loss; It is a first-order similarity loss, expressed as: ,in, express Nodes and The first-order similarity value between nodes Corresponding embedding vector matrix The Middle row nodes Structural characterization Corresponding embedding vector matrix The Middle row nodes Structural characterization; This is the weight for the second-order similarity loss, used to balance the proportion of the second-order similarity loss in the total loss; It is a second-order similarity loss, expressed as: ,in, Represents the original binary adjacency matrix. The reconstructed adjacency matrix output by the decoder is the decoder's reconstruction of the original binary adjacency matrix. Prediction results after topology reconstruction Represents the weight matrix; This is the weight of the regularization term, used to balance the proportion of regularization loss in the total loss; yes Regularization, used to prevent overfitting of the model, is expressed as: ,in, Indicates the first The weight matrix of the layer, Indicates the first The bias matrix of the layer, This represents the total number of layers in the model. By minimizing this loss function, the weights and bias parameters of the encoder and decoder are iteratively updated. Training stops when the loss function converges to a preset threshold or reaches the maximum number of iterations, and the decoder outputs the reconstructed adjacency matrix.

8. The method for identifying dynamic topology and line parameters of a power system based on graph embedding learning according to claim 6, characterized in that: In step S3-4, based on the reconstructed adjacency matrix The newly constructed topology graph refines the initial node feature data through sampling and aggregation methods, specifically: The topology graph uses power grid nodes as vertices to reconstruct the adjacency matrix. The element value in the table is the connection weight of the edge, used to determine the node. With nodes Are there any topological connections between them? Based on the reconstructed adjacency matrix Update features using the new topology graph: The local receptive domain of a node is determined by neighbor sampling: ； in, Represents a node The set of neighbors; The reconstructed adjacency matrix The element values ​​in the table represent the nodes in the reconstructed topology. With nodes Connection weights between them; The features of neighboring nodes are weighted and aggregated to obtain the nodes. Intermediate feature representation: ； in, For nodes In the The eigenvectors of the next iteration; Represents a node The set of neighbors; Represents nodes in the reconstructed topology With nodes Connection weights between them; It is a learnable linear transformation weight matrix; It is a non-linear activation function; The aggregated features are then fused with the original features of the nodes to achieve dynamic correction of the node features: ； Among them, nodes initial input features The nodes obtained after normalization eigenvectors, This is to aggregate and update the node features. For balance coefficient, For nodes The corrected features.

9. The method for identifying dynamic topology and line parameters of a power system based on graph embedding learning according to claim 1, characterized in that: In step S4, a regression model is used to predict and evaluate the line parameters, specifically as follows: The node topology is abstracted into a line graph topology. Based on the reconstructed topology, the corrected node features are reorganized into line feature vectors, specifically: The node topology is abstracted into a line graph topology, ignoring the original node topology, treating lines as new nodes, and extending node features to line features. Line susceptance The line parameters of branch conductance are used as new nodes, and the node characteristics are extended to line characteristics. Using the predictive variables at both ends of the line, based on The line susceptance is calculated using a regression model. ; Combined with line susceptance ,based on Branch conductance is calculated using a regression model. ; RMSE is used to evaluate the quality of model fit, specifically: ； in, For the sample size, It is the first Each sample corresponds to the actual value of the electrical parameters of the transmission line. It is the first The predicted values ​​of electrical parameters of the transmission line corresponding to each sample.

Images