A method and system for power grid fault location based on feature decoupling graph networks

Abstract

A feature-decoupled graph network method and system for locating power grid line faults includes: First, using power fault analysis software, a time-domain simulation is performed to generate a large dataset consisting of bus features and fault locations. A line graph is then used to construct the network topology of power grid nodes and transmission lines. Next, the graph structure data is multiplied by a weight matrix to transform it into a new hidden space, and attention scores between nodes are calculated in the first-order neighborhood. Simple factor graph classification labels are set, and an additional head is introduced in the unwrapping layer as a discriminator to prevent factor graph degradation. Further feature aggregation is achieved by taking the weighted sum of neighboring nodes, merging factor graphs, and recombinating the decoupled features onto a single graph. Finally, features at the corresponding fault locations are extracted from the aggregated graph features and mapped to a one-dimensional fault location. This invention does not rely on the prior knowledge required for users to build mathematical and physical models, has higher versatility, and is simple to operate.

Description

Technical Field

[0001] This invention belongs to the field of power grid transmission technology, specifically relating to a power grid fault location method and system based on feature decoupling graph networks. Background Technology

[0002] Power transmission systems frequently face problems such as regional power outages and equipment damage due to line faults. This not only affects the reliability and stability of the power grid system but can also threaten social order and public safety. Therefore, it is necessary to detect, classify, and locate power grid faults as early as possible to avoid prolonged power outages and cascading failures. This not only ensures the operational quality of the power grid system but also reduces revenue losses caused by power outages and saves labor costs associated with manual fault location.

[0003] Most power grid faults originate from line faults in transmission lines. Line faults occurring during power system operation are categorized into short-circuit faults and open-circuit faults, with short-circuit faults being more common. Short-circuit faults can be further divided into symmetrical and asymmetrical faults, with five subtypes: single-line-to-ground, line-to-line, double-line-to-ground, three-phase-to-ground, and three-phase short circuit. For power grid line fault problems, fault diagnosis involves three steps: fault detection, fault classification, and fault location.

[0004] Current power grid fault location methods can be divided into two categories: methods based on mathematical and physical models and methods based on measurement data and statistical analysis. Methods based on mathematical and physical models mainly employ traveling wave and impedance-based approaches, which require the establishment of mathematical models based on prior knowledge. These methods have poor versatility and are computationally complex and cumbersome. Statistical analysis methods based on measurement data mainly employ traditional machine learning, deep learning, and reinforcement learning methods. These methods require sufficient historical data to train and adjust the computational model, thus placing very high demands on the quality of the dataset.

[0005] Chinese invention patent CN119064723A discloses a method for locating single-phase open-circuit faults in power distribution networks. By setting a comparison quantity, the comparison quantity is directly compared with the monitored quantity to obtain the offset. The offset is then used to determine the condition of the monitoring point. The determination is made directly based on the monitored voltage and the number of existing monitoring points, without the need for additional measurements from external electronic equipment, resulting in fast location speed.

[0006] Chinese invention patent CN118962342A discloses a method for locating faults in medium-voltage distribution network cable lines. Addressing the problem of overlapping initial and reflected traveling waves in short-circuit scenarios, which can hinder fault location, this method uses a reference high-frequency signal to pinpoint the fault location. It also proposes a large data sample library that can adapt to distribution network cable lines with different parameters, improving the accuracy of cable fault location and reducing the workload of cable fault investigation.

[0007] Chinese invention patent CN118965204A discloses a data analysis-based equipment fault location method and related components. By acquiring historical equipment operation data, and selecting either a fault model or preset fault rules based on the number of fault tags corresponding to the feature data, fault detection is automated, improving efficiency.

[0008] Chinese invention patent CN118884129A discloses an artificial intelligence-based method for fault location in distribution networks and a storage medium. By using data collected from power sensors in a multi-level distribution network, electrical parameters are determined to identify the characteristic values ​​and deviations of each level. The fault type is then identified by combining the network topology and fault identification model. Finally, fault location is achieved using fault location methods based on impedance and traveling waves.

[0009] With the increasing application of new and complex architectures such as next-generation loads and equipment in existing power distribution network systems, traditional protection schemes are no longer sufficient to meet the new challenges. Furthermore, with the increase in the amount of available measurement data, traditional fault location methods suffer from three main problems:

[0010] (1) Traditional mathematical physics model-based methods require the establishment of mathematical models based on prior knowledge, which has poor versatility and is computationally complex and cumbersome.

[0011] (2) It is not possible to flexibly merge measurement data from different buses. Instead, the faulty line is identified by checking one by one and then the faulty data is processed.

[0012] (3) Traditional machine learning methods struggle to model the topology of distributed networks, not to mention the possibility of topology changes. This limitation leads to insufficient accuracy and precision in predictions when dealing with complex power grid systems.

[0013] Therefore, the introduction of feature-based decoupling graph neural networks is crucial. It can effectively process graph data, model the nodes and lines of the power grid, and utilize the topological relationships between nodes for information transmission and learning, thereby better reflecting the actual operating status of the power grid. Summary of the Invention

[0014] This invention aims to overcome the technical problems of existing technologies, such as difficulty in modeling power networks, poor versatility, computational complexity, and insufficient feature mining, and proposes a power grid fault location method and system based on feature decoupled graph networks.

[0015] This invention applies a feature-decoupled graph convolutional network to the power grid fault location problem, fully utilizing the data and spatial features of the bus to achieve highly accurate power grid fault location prediction. Furthermore, the system effectively processes graph data through the graph convolutional network to model the nodes and lines of the power grid, and utilizes feature decoupling technology to obtain hidden features and structures in the power grid data. This improves the accuracy and precision of power grid fault location, thereby effectively enhancing power grid security, promoting the development of smart grids, and providing new ideas and directions for research and practice in the field of power system fault location.

[0016] The first aspect of this invention relates to a power grid fault location method based on feature decoupling graph networks, comprising the following steps:

[0017] Step S01: Based on the power grid topology and sensor data, perform time-domain simulation using power fault analysis software to generate a large dataset consisting of bus characteristics and fault locations;

[0018] Step S02: Construct the network topology of the power grid nodes and transmission lines using a line graph method, so that the data features on the edges are transferred to the graph nodes, and multiple graphs are organized into a large graph through batch processing;

[0019] Step S03: Multiply the graph structure data by the weight matrix to transform it into a new hidden space. Calculate the strength of the relationship between nodes by the attention score of the edges in the neighborhood according to the graph attention network mechanism. After calculating the attention score of each edge of each factor graph, a single factor graph can be represented, thus completing the decoupling and representation of the factor graph.

[0020] Step S04: Simply set the factor graph classification labels and introduce an additional head in the unwound layer as a discriminator to avoid the degradation of the factor graph;

[0021] Step S05: Further feature aggregation is achieved by taking the weighted sum of its neighboring nodes, i.e., the graph convolution step. After the aggregation step is completed, different factor graphs represent different types of feature subsets in the power grid structure. The factor graphs are merged to recombine the decoupled features onto a single graph.

[0022] Step S06: Extract the features at the corresponding fault locations from the aggregated image, map them to the one-dimensional fault locations, and use an activation function to complete the final transformation.

[0023] Specifically, step S01 includes:

[0024] Simulations were performed using the PSASP power fault analysis software. A specific power system topology was constructed, and network parameters such as the number of grid nodes, network generation method, fault line number, and fault type were modified. Based on the time-domain simulation results generated by the software, power grid fault location data samples with fault location labels were obtained. The fault location labels are integers between 0 and 100, and the data features include bus voltage, current, resistance, active power, bus amplitude, bus frequency, V / I angle, and voltage phase angle.

[0025] Specifically, step S02 includes:

[0026] Since the data characteristics of the power grid are generated on transmission lines (such as current, voltage, and power), a line graph approach is used to construct the network structure to facilitate decoupling and other processing of these characteristics. For an undirected graph G = (V, E), the line graph L(G) = (V', E') represents the adjacency relationships between edges in G, where V = {v1, v2, ..., v...} m} is a graph node v i The set, E = {e1, e2, ..., e n} is an edge e in the graph i The set, The method for constructing L(G) is as follows: for any edge e in G i ∈E, all of which are treated as a vertex v of L(G). i '∈V'; and for each group G, the same vertex v is shared. i For any two edges ∈ V, connect their corresponding vertices in L(G) to form an edge, from which vertex v i Convert to e in L(G) i '∈E'. Specifically, the connection combinations of power grid nodes are numbered. If the input graph L(G) during training has graph nodes v i '=e i =(v i ,v k ), v j '=e j =(v i ,v j If L(G) has an edge e, then L(G) has an edge e. i '=(v i ',v j ')=v i Converting to a line graph allows data features on the edges to be transferred to the graph nodes, which is more conducive to graph-level feature decoupling. Finally, batch processing of the images simplifies the embedding of graph structures.

[0027] Specifically, step S03 includes:

[0028] The transformed graph structure undergoes feature decoupling. The decoupling method used is graph-level decoupling, meaning that multiple subgraphs are generated after decoupling. The input graph is decomposed into several factor graphs during the feature decoupling stage; the specific number needs to be designed and input manually based on the actual problem. Node feature sets are used. The set of edges e = {e1, e2, ..., e} n The input graph structure is represented. Where h... i This represents the data characteristics of the F dimension on the transmission line, i.e. Multiply the input feature structure H by a linear transformation matrix. Transform to a new hidden space H′, and obtain

[0029] This invention uses a mechanism similar to graph attention networks to calculate attention scores for edges:

[0030] s ij =ψ(Wh i ,Wh j (1)

[0031] in It is a shared attention mechanism responsible for calculating the attention mechanism score s of the edges of the factor graph e after taking the features of nodes i and j as input. ij It was implemented using a single-layer perceptron. The attention mechanism score is s. ij This demonstrates the importance of node j's characteristics to node i. When the score is 0, it is considered that other nodes do not contribute to that node.

[0032] This invention employs a masking attention mechanism to incorporate graph structures into the convolution process between nodes. Specifically, only those conforming to the parameters are computed. The contribution of node j to node i under certain conditions, where Let i be a neighborhood of node i. Let this neighborhood be of order one, which is the set of all points directly connected to node i.

[0033] This invention generates a new factor graph by directly generating the coefficients of the edges. A threshold is set, and edges connecting nodes i and j with scores below this threshold are considered as having weak relationships. Then, the transformed features are used to generate the following factor coefficients:

[0034]

[0035] Normalizing the attention score to the interval [0,1] (where e is the natural constant) yields the coefficient E used to represent the edge between nodes i and j in the factor graph e. ije After calculating the attention score for each edge of each factor graph, E can be used.e This invention represents a single factor graph. Thus, it achieves the decoupling and representation of factor graphs.

[0036] Step 4 specifically includes:

[0037] This optimization problem is transformed into a graph classification problem by simply labeling the factor graphs to find an approximate solution. That is, the differences in labels guide the generation of different topological structures in the factor graphs. Therefore, the head is used as the discriminator, defined as follows:

[0038]

[0039] in It is a three-layer graph autoencoder that transforms the features H' and the generated factor graph E. e The features are used as input to generate new node features, which are then mapped onto simple labels through a fully connected layer. The readout function is used to average and aggregate the features of all nodes in the graph, thereby generating a graph-level feature representation.

[0040] The loss function used to train the discriminator is defined as follows:

[0041]

[0042] Where Ne is the number of factor graphs and N is the number of training samples, which is obtained by multiplying the number of factor graphs by the number of input graphs. It is the distribution of sample i. This represents the probability that the generated factor graph has label c. 1 e＝c This is an indicator function that takes a value of 1 when the predicted label is correct and 0 otherwise. The formula calculates the sum of the logarithms of the probability distributions of correct predictions for all factor maps for a single sample, then calculates the negative average of the logarithms of all samples, thus obtaining the loss value generated during one training iteration.

[0043] The loss of the entire framework is defined as These represent the decoupling layer loss and the task loss, respectively, and cross-entropy loss is used as the method for estimating the loss between the target and the actual loss.

[0044] Step 5 specifically includes:

[0045] During the aggregation process, new node features are generated by taking the weighted sum of their neighboring nodes, which is essentially a standard graph convolution process. The general formula for graph convolution can be written as:

[0046]

[0047] in This is a method of normalizing the adjacency matrix A, which ensures that the contribution of each node's features is relatively balanced during graph convolution operations. Similarly, this invention achieves this by normalizing the features of factor graph nodes by the square root of their degrees, i.e., E. ije / c ij E ije c is the coefficient of the edge from node i to node j in the factor graph e. ij This is the normalization term calculated based on the degrees of nodes i and j. The convolution aggregation process is described as follows:

[0048]

[0049] in This represents the new feature obtained by aggregating node i from factor graph e at layer l+1. W represents all the neighboring nodes of node i in the input graph. (I) It is the linear transformation matrix in the characteristic decoupling step.

[0050] The feature factor graphs generated by decoupling are merged according to the following formula, so that the decoupled features are recombined on a single graph:

[0051]

[0052] Where || denotes a connection operation. This yields the feature-enhanced output network.

[0053] Step 6 specifically includes:

[0054] The self-attention mechanism maps all node features to a one-dimensional fault location.

[0055] x=λ*Sigmoid(SektAttention(H”)),

[0056]

[0057] This completes the task of locating power grid faults.

[0058] A second aspect of the invention relates to a system for implementing the fault location method based on feature decoupling graph networks of the invention, comprising:

[0059] The power grid line fault dataset construction module is used to generate a large dataset consisting of bus features and fault locations;

[0060] The line graph conversion and graph structure construction module uses line graphs to construct the power network topology structure of power grid nodes and transmission lines, transferring data features on transmission lines to graph nodes, and organizing multiple graphs into a large graph through batch processing.

[0061] The feature decoupling module multiplies the graph structure data by a weight matrix to transform it into a new hidden space. Within the neighborhood, it calculates the attention scores of edges according to the graph attention network mechanism to obtain the strength of the relationship between nodes. After calculating the attention score of each edge of each factor graph, it can represent a single factor graph, thus completing the decoupling and representation of the factor graph.

[0062] The classification label setting module is used to simply set the classification labels of the factor graph and introduce an additional head in the unwrap layer as a discriminator to avoid the degradation of the factor graph.

[0063] The feature aggregation and merging module further aggregates features by taking the weighted sum of its neighboring nodes, i.e., the graph convolution step. After the aggregation step is completed, different factor graphs represent different types of feature subsets in the power grid structure. Merging factor graphs recombines the decoupled features onto a single graph.

[0064] The power grid fault location regression task module is used to extract features at the corresponding fault locations from the aggregated graph features, map them to one-dimensional fault locations, and use an activation function to complete the final transformation.

[0065] A third aspect of the present invention relates to a fault location apparatus based on a feature-decoupled graph network, characterized in that it includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the fault location method based on the feature-decoupled graph network of the present invention.

[0066] A fourth aspect of the present invention relates to a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the fault location method based on feature decoupling graph networks of the present invention.

[0067] Compared with existing technologies, the innovation of this invention lies in the fact that it does not rely on the prior knowledge required by users to build mathematical and physical models. Instead, it fully utilizes the data characteristics and topological space characteristics of the power grid bus to model the power grid table data as a graph structure. By combining two algorithmic architectures—decoupled representation learning and graph convolutional neural networks—it achieves both topology-based modeling capabilities and robust representation learning capabilities. Ultimately, it achieves more universal, simpler to operate, and more realistic power grid data characteristics, enabling the discovery of hidden patterns and structures in power network graph data, thereby improving positioning accuracy and robustness. Attached Figure Description

[0068] Figure 1 The diagram shows the system structure of the present invention, wherein: 1-power grid line fault dataset construction module, 2-line graph conversion and graph structure construction module, 3-feature decoupling module, 4-feature aggregation and merging module, and 5-power grid fault location regression task module.

[0069] Figure 2 This is a schematic diagram of the feature decoupling steps of the present invention.

[0070] Figure 3 This is a schematic diagram of the device of the present invention. Detailed Implementation

[0071] 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.

[0072] Example 1

[0073] This embodiment relates to a power grid fault location method based on feature decoupling graph networks, including the following steps:

[0074] Step S01: Based on the power grid topology and sensor data, perform time-domain simulation using power fault analysis software to generate a large dataset consisting of bus characteristics and fault locations;

[0075] Step S02: Construct the network topology of the power grid nodes and transmission lines using a line graph method, so that the data features on the edges are transferred to the graph nodes, and multiple graphs are organized into a large graph through batch processing;

[0076] Step S03: Multiply the graph structure data by the weight matrix to transform it into a new hidden space. Calculate the strength of the relationship between nodes by the attention score of the edges in the neighborhood according to the graph attention network mechanism. After calculating the attention score of each edge of each factor graph, a single factor graph can be represented, thus completing the decoupling and representation of the factor graph.

[0077] Step S04: Simply set the factor graph classification labels and introduce an additional head in the unwound layer as a discriminator to avoid the degradation of the factor graph;

[0078] Step S05: Further feature aggregation is achieved by taking the weighted sum of its neighboring nodes, i.e., the graph convolution step. After the aggregation step is completed, different factor graphs represent different types of feature subsets in the power grid structure. The factor graphs are merged to recombine the decoupled features onto a single graph.

[0079] Step S06: Extract the features at the corresponding fault locations from the aggregated image, map them to the one-dimensional fault locations, and use an activation function to complete the final transformation.

[0080] Specifically, step S01 includes:

[0081] Simulations were performed using the PSASP power fault analysis software. A specific power system topology was constructed, and network parameters such as the number of grid nodes, network generation method, fault line number, and fault type were modified. Based on the time-domain simulation results generated by the software, power grid fault location data samples with fault location labels were obtained. The fault location labels are integers between 0 and 100, and the data features include bus voltage, current, resistance, active power, bus amplitude, bus frequency, V / I angle, and voltage phase angle.

[0082] Specifically, step S02 includes:

[0083] Since the data characteristics of the power grid are generated on transmission lines (such as current, voltage, and power), a line graph approach is used to construct the network structure to facilitate decoupling and other processing of these characteristics. For an undirected graph G = (V, E), the line graph L(G) = (V', E') represents the adjacency relationships between edges in G, where V = {v1, v2, ..., v...} m} is a graph node v i The set, E = {e1, e2, ..., e n} is an edge e in the graph i The set, The method for constructing L(G) is as follows: for any edge e in G i ∈E, all of which are treated as a vertex v of L(G). i '∈V'; and for each group G, the same vertex v is shared. i For any two edges ∈ V, connect their corresponding vertices in L(G) to form an edge, from which vertex v i Convert to e in L(G) i '∈E'. Specifically, the connection combinations of power grid nodes are numbered. If the input graph L(G) during training has graph nodes v i '=e i =(v i ,v k ), v j '=e j =(v i ,v j If L(G) has an edge e, then L(G) has an edge e. i '=(v i ',v j ')=v i Converting to a line graph allows data features on the edges to be transferred to the graph nodes, which is more conducive to graph-level feature decoupling. Finally, batch processing of the images simplifies the embedding of graph structures.

[0084] Specifically, step S03 includes:

[0085] The transformed graph structure undergoes feature decoupling. The decoupling method used is graph-level decoupling, meaning that multiple subgraphs are generated after decoupling. The input graph is decomposed into several factor graphs during the feature decoupling stage; the specific number needs to be designed and input manually based on the actual problem. Node feature sets are used. The set of edges e = {e1, e2, ..., e} n The input graph structure is represented. Where h... i This represents the data characteristics of the F dimension on the transmission line, i.e. Multiply the input feature structure H by a linear transformation matrix. Transform to a new hidden space H′, and obtain

[0086] This invention uses a mechanism similar to graph attention networks to calculate attention scores for edges:

[0087] s ij =ψ(Wh i ,Wh j (1)

[0088] in It is a shared attention mechanism responsible for calculating the attention mechanism score s of the edges of the factor graph e after taking the features of nodes i and j as input. ij It was implemented using a single-layer perceptron. The attention mechanism score is s. ij This demonstrates the importance of node j's characteristics to node i. When the score is 0, it is considered that other nodes do not contribute to that node.

[0089] This invention employs a masking attention mechanism to incorporate graph structures into the convolution process between nodes. Specifically, only those conforming to the parameters are computed. The contribution of node j to node i under certain conditions, where Let i be a neighborhood of node i. Let this neighborhood be of order one, which is the set of all points directly connected to node i.

[0090] This invention generates a new factor graph by directly generating the coefficients of the edges. A threshold is set, and edges connecting nodes i and j with scores below this threshold are considered as having weak relationships. Then, the transformed features are used to generate the following factor coefficients:

[0091]

[0092] Normalizing the attention score to the interval [0,1] (where e is the natural constant) yields the coefficient E used to represent the edge between nodes i and j in the factor graph e. ije After calculating the attention score for each edge of each factor graph, E can be used.e This invention represents a single factor graph. Thus, it achieves the decoupling and representation of factor graphs.

[0093] Step 4 specifically includes:

[0094] This optimization problem is transformed into a graph classification problem by simply labeling the factor graphs to find an approximate solution. That is, the differences in labels guide the generation of different topological structures in the factor graphs. Therefore, the head is used as the discriminator, defined as follows:

[0095]

[0096] in It is a three-layer graph autoencoder that transforms the features H' and the generated factor graph E. e The features are used as input to generate new node features, which are then mapped onto simple labels through a fully connected layer. The readout function is used to average and aggregate the features of all nodes in the graph, thereby generating a graph-level feature representation.

[0097] The loss function used to train the discriminator is defined as follows:

[0098]

[0099] Where Ne is the number of factor graphs and N is the number of training samples, which is obtained by multiplying the number of factor graphs by the number of input graphs. It is the distribution of sample i. This represents the probability that the generated factor graph has label c. 1 e＝c This is an indicator function that takes a value of 1 when the predicted label is correct and 0 otherwise. The formula calculates the sum of the logarithms of the probability distributions of correct predictions for all factor maps for a single sample, then calculates the negative average of the logarithms of all samples, thus obtaining the loss value generated during one training iteration.

[0100] The loss of the entire framework is defined as These represent the decoupling layer loss and the task loss, respectively, and cross-entropy loss is used as the method for estimating the loss between the target and the actual loss.

[0101] Step 5 specifically includes:

[0102] During the aggregation process, new node features are generated by taking the weighted sum of their neighboring nodes, which is essentially a standard graph convolution process. The general formula for graph convolution can be written as:

[0103]

[0104] in This is a method of normalizing the adjacency matrix A, which ensures that the contribution of each node's features is relatively balanced during graph convolution operations. Similarly, this invention achieves this by normalizing the features of factor graph nodes by the square root of their degrees, i.e., E. ije / c ij E ije c is the coefficient of the edge from node i to node j in the factor graph e. ij This is the normalization term calculated based on the degrees of nodes i and j. The convolution aggregation process is described as follows:

[0105]

[0106] in This represents the new feature obtained by aggregating node i from factor graph e at layer l+1. W represents all the neighboring nodes of node i in the input graph. (l) It is the linear transformation matrix in the characteristic decoupling step.

[0107] The feature factor graphs generated by decoupling are merged according to the following formula, so that the decoupled features are recombined on a single graph:

[0108]

[0109] Where || denotes a connection operation. This yields the feature-enhanced output network.

[0110] Step 6 specifically includes:

[0111] The self-attention mechanism maps all node features to a one-dimensional fault location.

[0112] x=λ*Sigmoid(SektAttention(H”)),

[0113]

[0114] This completes the task of locating power grid faults.

[0115] To demonstrate the superiority of the proposed model method, performance was compared using different models. Experiments were conducted to compare the method with the proposed method using the K-nearest neighbors algorithm, random forest, graph convolutional networks, and fully connected networks, and the accuracy of fault location under different fault types was statistically calculated. The bolded data represents the highest accuracy, and the underlined data represents the second highest accuracy. It is readily apparent that the proposed method achieves the highest fault location accuracy.

[0116] Fault type 1 Fault type 2 Fault type 3 Fault type 4 K-Nearest Neighbors Algorithm 0.594 0.985 0.992 0.585 Random Forest 0.794 1 1 0.808 Graph Convolutional Networks 0.884 <![CDATA[ 0.988 ]]> <![CDATA[ 0.996 ]]> <![CDATA[ 0.967 ]]> Fully connected network <![CDATA[ 0.975 ]]> 1 1 0.843 This invention 0.988 1 1 0.974

[0117] Example 2

[0118] This embodiment relates to a system for implementing the fault location method based on feature decoupling graph networks in Embodiment 1, comprising:

[0119] The power grid line fault dataset construction module is used to generate a large dataset consisting of bus features and fault locations;

[0120] The line graph conversion and graph structure construction module uses line graphs to construct the power network topology structure of power grid nodes and transmission lines, transferring data features on transmission lines to graph nodes, and organizing multiple graphs into a large graph through batch processing.

[0121] The feature decoupling module multiplies the graph structure data by a weight matrix to transform it into a new hidden space. Within the neighborhood, it calculates the attention scores of edges according to the graph attention network mechanism to obtain the strength of the relationship between nodes. After calculating the attention score of each edge of each factor graph, it can represent a single factor graph, thus completing the decoupling and representation of the factor graph.

[0122] The classification label setting module is used to simply set the classification labels of the factor graph and introduce an additional head in the unwrap layer as a discriminator to avoid the degradation of the factor graph.

[0123] The feature aggregation and merging module further aggregates features by taking the weighted sum of its neighboring nodes, i.e., the graph convolution step. After the aggregation step is completed, different factor graphs represent different types of feature subsets in the power grid structure. Merging factor graphs recombines the decoupled features onto a single graph.

[0124] The power grid fault location regression task module is used to extract features at the corresponding fault locations from the aggregated graph features, map them to one-dimensional fault locations, and use an activation function to complete the final transformation.

[0125] Example 3

[0126] like Figure 3 This embodiment relates to a fault location device based on a feature decoupling graph network, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the fault location method based on a feature decoupling graph network of Embodiment 1.

[0127] Example 4

[0128] A fourth aspect of the present invention relates to a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the fault location method based on feature decoupling graph networks of the present invention.

[0129] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A fault localization method based on feature decoupling graph networks, comprising the following steps: Step 1: Based on the power grid topology and sensor data, perform time-domain simulation using power fault analysis software to generate a large dataset consisting of bus characteristics and fault locations. Step 2: Construct the power network topology using a line graph approach to connect the power grid nodes and transmission lines, transferring data features from the transmission lines to the graph nodes, and then organizing multiple graphs into a large graph through batch processing. Step 3: Multiply the graph structure data by the weight matrix to transform it into a new hidden space. Calculate the attention scores of the edges in the neighborhood according to the graph attention network mechanism to obtain the strength of the relationship between nodes. After calculating the attention score of each edge of each factor graph, a single factor graph can be represented, thus completing the decoupling and representation of the factor graph. Step 4: Simply set the factor graph classification labels and introduce an additional head in the unwrap layer as a discriminator to avoid factor graph degradation; Step 5: Further feature aggregation is achieved by taking the weighted sum of its neighboring nodes, i.e., the graph convolution step. After the aggregation step is completed, different factor graphs represent different types of feature subsets in the power grid structure. The factor graphs are merged to recombine the decoupled features onto a single graph. Step 6: Extract the features corresponding to the fault locations in the aggregated image, map them to the one-dimensional fault locations, and use an activation function to complete the final transformation.

2. The fault location method based on feature decoupling graph networks as described in claim 1, characterized in that: Step 1 includes: Simulation was performed using PSASP power fault analysis software; a specific topology of the power system was constructed, and by modifying the number of power grid nodes, network generation method, fault line number, and fault type network parameters, time-domain simulation results were generated based on the software to obtain power grid fault location data samples with fault location labels; the fault location labels are integers between 0 and 100, and the data features include bus voltage, current, resistance, active power, bus amplitude, bus frequency, V / I angle, and voltage phase angle.

3. The fault location method based on feature decoupling graph networks as described in claim 1, characterized in that: Step 2 includes: The network structure is constructed using a line graph approach; for an undirected graph G = (V, E), the line graph L(G) = (V', E') represents the adjacency relationships between edges in G, where V = { , , ... , } is a graph node v i The set, E = { , , ... , } is an edge of the graph The set, The method for constructing L(G) is as follows: for any edge in G... They are all treated as a vertex of L(G). V For each group G, the same vertices are shared. Connect the two edges of L(G) with their corresponding vertices in L(G) to form an edge, from which the vertex... Transformation into L(G) Specifically, the connection combinations of power grid nodes are numbered. If the input graph L(G) during training has graph nodes... ( , ), v j e j ( v j If L(G) has an edge, then L(G) has an edge. v j Converting to a line graph allows data features on the edges to be transferred to graph nodes, which is beneficial for graph-level feature decoupling. Finally, batch processing of the image is completed, making the embedding of graph structures simpler.

4. The fault location method based on feature decoupling graph networks as described in claim 1, characterized in that: Step 3 includes: The transformed graph structure undergoes feature decoupling; the feature decoupling used is graph-level decoupling, meaning that multiple subgraphs are generated after decoupling; the input graph is decomposed into several factor graphs during the feature decoupling stage, the specific number of which is designed and input according to the actual problem; the node feature set H = { , , ... , The set of edges e = { , ,... , The input graph structure is represented, where... This represents the data characteristics of the F dimension on the transmission line, i.e. Multiply the input feature structure H by a linear transformation matrix. Transform to a new hidden space H′, and obtain H′ = WH, H′ ; Attention scores are calculated for edges using a mechanism similar to graph attention networks: （1） in : It is a shared attention mechanism responsible for calculating the attention mechanism score s of the edges of the factor graph e after taking the features of nodes i and j as input. ij The implementation uses a single-layer perceptron; the attention mechanism score is s. ij This demonstrates the importance of node j's features to node i; when the score is 0, it is considered that other nodes do not contribute to that node. A masking attention mechanism is used to incorporate graph structures into the convolution process between nodes; specifically, only those conforming to the attention mechanism are computed. The contribution of node j to node i under certain conditions, where Let i be a neighborhood of i; set this neighborhood to first order, which is the set of all points directly connected to node i. A new factor graph is generated by directly generating the coefficients of the edges; a threshold is set, and nodes i and j connected by edges with scores below the threshold are considered to be weakly related; then, the transformed features are used to generate the following factor coefficients: （2） The attention score is normalized to the interval [0,1], where e is a natural constant, to obtain the coefficient E used to represent the edge between nodes i and j in the factor graph e. ije After calculating the attention score for each edge of each factor graph, E can be used. e This is used to represent a single factor graph, thus completing the decoupling and representation of the factor graph.

5. The fault location method based on feature decoupling graph networks as described in claim 4, characterized in that: Step 4 includes: Simply define factor graph labels, and use the differences in labels to guide the generation of different topological structures in the factor graph. Use the head as a discriminator, as defined below: （3） in It is a three-layer graph autoencoder that transforms the features H' and the generated factor graph E. e The new node features are generated as input and then mapped to simple labels through a fully connected layer; the Readout function is used to average and aggregate the features of all nodes in the graph to generate a graph-level feature representation. The loss function used to train the discriminator is defined as follows: （4） Where Ne is the number of factor graphs and N is the number of training samples, which is obtained by multiplying the number of factor graphs by the number of input graphs; It is the distribution of sample i. [c] represents the probability that the generated factor graph has label c; It is an indicator function that takes the value 1 when the predicted label is correct and 0 otherwise. This formula calculates the sum of the logarithms of the probability distributions of the correct predictions of all factor maps for a single sample, and then calculates the negative average of the logarithms of all samples to obtain the loss value generated by one training. The loss of the entire framework is defined as ; , These represent the decoupling layer loss and the task loss, respectively, and cross-entropy loss is used as the method for estimating the loss between the target and the actual loss.

6. The fault location method based on feature decoupling graph networks as described in claim 1, characterized in that: Step 5 includes: During the aggregation process, new node features are generated by taking the weighted sum of their neighboring nodes, which is a typical graph convolution process. The general formula for graph convolution is written as: （5） in This is a way to normalize the adjacency matrix A, ensuring that the contribution of each node's features is relatively balanced during graph convolution operations; similarly, by normalizing the features of factor graph nodes by the square root of their degree, E... ije / c ij E ije c is the coefficient of the edge from node i to node j in the factor graph e. ij The normalized term is calculated based on the degrees of nodes i and j. The convolution aggregation process is described as follows: （6） in This represents the new feature obtained by aggregating node i from factor graph e at layer l+1. This represents all the neighboring nodes of node i in the input graph. It is the linear transformation matrix in the characteristic decoupling step; The feature factor graphs generated by decoupling are merged according to the following formula, so that the decoupled features are recombined on a single graph: （7） Where || represents a connection operation; thus, the feature-enhanced output network is obtained.

7. The fault location method based on feature decoupling graph networks as described in claim 1, characterized in that: Step 6 includes: The self-attention mechanism maps all node features to a one-dimensional fault location. （8） Complete the task of locating power grid faults.

8. A system for implementing the fault location method based on feature decoupling graph networks as described in any one of claims 1-7, characterized in that, include: The power grid line fault dataset construction module is used to generate a large dataset consisting of bus features and fault locations; The line graph conversion and graph structure construction module constructs the generated power grid fault location dataset into an actual transmission network structure and converts it into a line graph format; The feature decoupling module calculates attention scores for the transformed graph structure and performs graph-level feature decoupling. The feature aggregation and merging module is used to perform convolution aggregation on the decoupled factor map features and then merge them into a complete feature map. The power grid fault location regression task module is used to extract features at the corresponding fault locations from the aggregated graph features, map them to one-dimensional fault locations, and use an activation function to complete the final transformation.

9. A fault location device based on feature decoupling graph networks, characterized in that, The system includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement the fault location method based on feature decoupling graph networks as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the fault location method based on feature decoupling graph network as described in any one of claims 1-7.

Images