Characteristic decoupling graph network-based power grid fault positioning method and system

Abstract

The invention discloses a feature decoupling graph network method and system for positioning a power grid line fault, and the method comprises the steps: firstly, carrying out the time domain simulation through power fault analysis software, generating a large data set composed of bus features and fault positions, and constructing a topological structure of a network of power grid nodes and transmission lines in a line graph manner; then, the graph structure data multiplied by the weight matrix is converted into a new hidden space, and an attention score between nodes is calculated in a first-order neighborhood; a factor graph classification label is simply set, and an additional head is introduced into an unwrapping layer to serve as a discriminator to avoid degradation of the factor graph; further feature aggregation is realized by taking the weighted sum of neighbor nodes, factor graphs are combined, and the decoupled features are recombined on one graph; and finally, extracting features on corresponding fault positions of the aggregated image features, and mapping the features to one-dimensional fault positions. The method does not depend on priori knowledge required by a user to establish a mathematical physical model, and is higher in universality and simple to operate.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid transmission, and specifically relates to a power grid fault location method and system based on a feature decoupling graph network. Background Art

[0002] The power transmission system often faces problems such as regional power outages and equipment damage caused by line faults. This not only affects the reliability and stability of the power grid system, but even poses a threat to social order and public safety. Therefore, it is necessary to detect, classify, and locate the faults occurring in the power grid as early as possible to avoid long-term power outages and cascading failures. This can not only ensure the operation quality of the power grid system, but also reduce the revenue losses caused by power failures and save the labor costs for manually locating faults.

[0003] Most power grid fault problems originate from line faults in transmission lines. The types of line faults occurring during the operation of the power system are divided into short-circuit faults and open-circuit faults, among which short-circuit faults are more likely to occur. Short-circuit faults can be divided into symmetrical faults and asymmetrical faults, with a total of five subtypes: single line to ground, line to line, double line to ground, three-phase to ground, and three-phase short circuit. For the power grid line fault problem, fault diagnosis is divided into three steps: fault detection, fault classification, and fault location.

[0004] Current power grid fault location methods can be divided into methods based on mathematical and physical models and statistical analysis methods based on measurement data. Methods based on mathematical and physical models mainly adopt methods based on traveling waves and impedance, which require establishing mathematical models according to prior knowledge, with poor generality and complex and cumbersome calculations; statistical analysis methods based on measurement data mainly adopt traditional machine learning methods, deep learning, and reinforcement learning methods, which require sufficient historical data to train and adjust the calculation model, so the quality requirements for the data set are very high.

[0005] Chinese invention patent CN119064723A discloses a method for locating single-phase open-circuit faults in a distribution network. By setting a comparison quantity, the comparison quantity is directly compared with the monitored quantity to obtain an offset quantity, and the offset quantity is judged to obtain the situation of the monitoring point. The situation is directly judged according to the monitored voltage and the number of existing monitoring points, without the need to externally connect other electronic devices for additional measurements, and the positioning speed is fast.

[0006] Chinese invention patent CN118962342A discloses a method for locating faults in medium-voltage distribution network cable lines. For the problem that the initial traveling wave and the reflected traveling wave are prone to overlap in the short line scenario, resulting in inability to locate, the fault point can be located based on the reference high-frequency signal; a large data sample library is proposed, which can adapt to distribution network cable lines with different parameters, improve the accuracy of cable fault location, and reduce the workload of cable fault troubleshooting.

[0007] Chinese Invention Patent CN118965204A discloses a device fault location method and related components based on data analysis. After obtaining the historical data of device operation, one of the fault models or preset fault rules is selected for fault detection according to the number of fault labels corresponding to the feature data, realizing automated fault detection and improving the fault detection efficiency.

[0008] Chinese Invention Patent CN118884129A discloses a distribution network fault location method and storage medium based on artificial intelligence. By determining the electrical parameters based on the data collected by the power sensors in the multi-level distribution network to determine the characteristic values and deviations of each level, combining the power grid topology relationship and the fault identification model to identify the fault type, and then combining the fault location methods based on impedance method and traveling wave for fault location.

[0009] With the continuous application of new-generation loads and devices and other emerging complex architectures in the existing power distribution system network, traditional protection schemes are no longer sufficient to cope with new challenges. In addition, with the increase in the number of available measurement data, traditional fault location methods mainly have three problems:

[0010] (1) Traditional methods based on mathematical and physical models need to establish mathematical models according to prior knowledge, with poor generality and complex and cumbersome calculations;

[0011] (2) It is unable to flexibly merge measurement data from different buses, but determines the fault line by checking one by one and then processes the fault data;

[0012] (3) Traditional machine learning methods are difficult to model the topology of distribution networks, let alone the possibility of topological structure changes. This limitation leads to insufficient accuracy and precision of prediction results when dealing with complex power grid systems.

[0013] Therefore, the introduction of a graph neural network based on feature decoupling is crucial. It can effectively process graph data, model the nodes and lines of the power grid, and use the topological relationship between nodes for information transmission and learning, so as to better reflect the actual operation state of the power grid. Summary of the Invention

[0014] The present invention aims to overcome the technical problems in the prior art, such as difficult modeling of power networks, poor generality, complex calculations, and insufficient feature mining, and proposes a power grid fault location method and system based on a feature decoupled graph network.

[0015] The present invention applies the feature decoupled graph convolutional network to the power grid fault location problem, fully utilizes the data features and spatial features of the bus, and realizes the prediction of the power grid fault location with relatively high accuracy. In addition, after effectively processing the graph data through the graph convolutional network, the system models the nodes and lines of the power grid, and uses the feature decoupling technology to obtain the hidden features and structures in the power grid data, which can improve the accuracy and precision of power grid fault location, thereby effectively improving the safety of the power grid, promoting the development of the smart grid, and providing new ideas and directions for the research and practice in the field of power system fault location.

[0016] The first aspect of the present invention relates to a power grid fault location method based on a feature decoupled graph network, including the following steps:

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

[0018] Step S02: Use the line graph method to construct the network topology of power grid nodes and transmission lines, transfer the data features on the edges to the graph nodes, and organize multiple graphs into a large graph through batch processing;

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

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

[0021] Step S05: Further feature aggregation is achieved by taking the weighted sum of its neighbor nodes, that is, the graph convolution step. After the aggregation step, different factor graphs represent different types of feature subsets in the power grid structure. Merge the factor graphs and recombine the decoupled features on one graph;

[0022] Step S06: Extract the features at the corresponding fault locations of the aggregated graph features, map them to one-dimensional fault locations, and complete the final conversion using the activation function.

[0023] Among them, step S01 specifically includes:

[0024] Perform simulation according to the PSASP power fault analysis software; construct the specific topological structure of the power system, and generate results through time-domain simulation according to the software by modifying network parameters such as the number of grid nodes, network generation method, fault line number, fault type, etc., to obtain power grid fault location data samples with fault location labels. The fault location label is an integer between 0 and 100, and the data features include bus voltage, current, resistance, active power, bus amplitude, bus frequency, V / I angle, voltage phase angle, etc.

[0025] Among them, step S02 specifically includes:

[0026] Since the data features of the power grid are all generated on the transmission lines (such as current, voltage, power, etc.), in order to facilitate the decoupling and other processing of the features, a line graph is used to construct the network structure. For an undirected graph G=(V, E), the line graph L(G)=(V', E') represents the adjacency relationship between the edges of G, where V={v 1 ,v 2 ,...,v m} is the set of graph nodes v i , E={e 1 ,e 2 ,...,e n} is the set of graph edges e i , The method of constructing L(G) is as follows: for any edge e i ∈E in G, it is taken as a vertex v i ’∈V'; and for each group of two edges in G that share the same vertex v i ∈V, the corresponding vertices in L(G) are connected as an edge, and thus the vertex v i is converted into e i '∈E'. Specifically, number the connection combinations of the power grid nodes. If the input graph L(G) has graph nodes v i '=e i =(v i ,v k ), v j '=e j =(v i ,v j ), then L(G) has an edge e i '=(v i ',v j ')=v i . Converting to a line graph can transfer the data features on the edges to the graph nodes, which is more conducive to the decoupling of graph-level features. Finally, complete the batch processing of the images to make the embedding of the graph structure simpler.

[0027] Among them, step S03 specifically includes:

[0028] Decouple the features of the converted graph structure. The feature decoupling adopted belongs to graph-level decoupling, that is, multiple subgraphs are generated after decoupling. The input graph is decomposed into several factor graphs in the feature decoupling stage, and the specific number needs to be designed and input manually according to the actual problem. Use the node feature set and the set of edges e = {e 1 , e 2 ,..., e n} to represent the input graph structure. Among them, h i represents the data features of dimension F on the transmission line, that is Multiply the input feature structure H by a linear transformation matrix and transform it to a new hidden space H', obtaining

[0029] The present invention uses a mechanism similar to the graph attention network to calculate the attention scores of edges:

[0030] s ij = ψ(Wh i , Wh j ) (1)

[0031] Among them is a shared attention mechanism, which is responsible for calculating the attention mechanism score s ij of the edges of the factor graph e after taking the features of nodes i and j as inputs, and adopts the form of a single-layer perceptron in implementation. The attention mechanism score s ij indicates the importance of the features of node j to node i. When the score is 0, it is regarded that other nodes do not contribute to this node.

[0032] The present invention uses a masked attention mechanism to add the graph structure to the convolutional process between nodes. Specifically, only calculate the contribution of node j that meets the condition to node i, where is a certain neighborhood of i. Set this neighborhood to the first order, that is, the set of all points directly connected to node i.

[0033] The present invention generates a new factor graph by directly generating the coefficients of the edges. Set a threshold, and regard the two nodes i and j connected by the edges with scores lower than this threshold as two nodes with weak mutual relationships. Then use the transformed features to generate the following factor coefficients:

[0034]

[0035] Normalize the attention score to the interval [0, 1] (where e is the natural constant) to obtain the coefficient E representing the edge between nodes i and j of the factor graph e ije After calculating the attention scores of each edge of each factor graph, E e can be used to represent a single factor graph. Thus, the present invention completes the decoupling and representation of the factor graph.

[0036] Among them, step 4 specifically includes:

[0037] Convert this optimization problem into a graph classification problem to find an approximate solution by simply setting the factor graph labels. That is, different topological structures of the factor graph are generated by guiding through the differences in labels. Therefore, the head is used as a discriminator, and the definition is as follows:

[0038]

[0039] Among them is a three-layer graph autoencoder that takes the transformed feature H' and the generated factor graph E e as inputs to generate new node features, and then maps them 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, thereby generating a graph-level feature representation.

[0040] The definition of the loss function for training the discriminator is shown as follows:

[0041]

[0042] Among them, Ne is the number of factor graphs, N is the number of training samples, which is obtained by multiplying the number of factor graphs by the number of input graphs. is the distribution of sample i, represents the probability that the generated factor graph has label c. 1 e＝c 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 all factor graphs predicted correctly for a single sample, and then statistically calculates the sum of the logarithms of all samples and takes the negative average value, thereby obtaining the loss value generated by one training.

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

[0044] Among them, step 5 specifically includes:

[0045] In the aggregation process, new node features are generated by taking the weighted sum of their neighbor nodes, that is, the ordinary graph convolution process. The general formula of graph convolution can be written as:

[0046]

[0047] Among them is a way to standardize the adjacency matrix A, which can ensure that the contribution of each node feature is relatively balanced when performing graph convolution operations. Similarly, the present invention achieves this purpose by normalizing the factor graph node features by the square root of their degrees, that is, E ije / c ij E ije is the coefficient of the edge from node i to node j in the factor graph e, and c ij is the normalization term calculated according to the degrees of nodes i and j. The convolution aggregation process is expressed as:

[0048]

[0049] Among them represents the new feature aggregated by node i from the factor graph e at the l + 1 layer, represents all the neighbor nodes of node i in the input graph, and W (I) is the linear transformation matrix in the feature decoupling step.

[0050] The feature factor graph generated by decoupling is merged according to the following formula to recombine the decoupled features on one graph:

[0051]

[0052] where || represents the concatenation operation. Thus, the output network with enhanced features is obtained.

[0053] Among them, step 6 specifically includes:

[0054] Mapping all node features to a one-dimensional fault location through the self-attention mechanism.

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

[0056]

[0057] Since then, the power grid fault location task has been completed.

[0058] The second aspect of the present invention relates to a system for implementing the fault location method based on the feature decoupling graph network of the present invention, including:

[0059] A power grid line fault dataset construction module for generating a large dataset composed of bus features and fault locations;

[0060] The line graph conversion and graph structure construction module constructs the power network topology of power grid nodes and transmission lines in the form of a line graph, transfers the data characteristics on the transmission lines to the graph nodes, and organizes multiple graphs into a large graph through batch processing;

[0061] The feature decoupling module multiplies the graph structure data by a weight matrix and converts it to a new hidden space. Calculate the attention scores of the edges according to the graph attention network mechanism within the neighborhood to obtain the strength of the relationship between nodes. After calculating the attention scores of each edge of each factor graph, a single factor graph can be represented, 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 graphs, and introduce an additional head as a discriminator in the disentangling layer to avoid the degradation of the factor graphs;

[0063] The feature aggregation and merging module realizes further feature aggregation by taking the weighted sum of its neighbor nodes, that is, the graph convolution step. After the aggregation step, different factor graphs represent different types of feature subsets in the power grid structure. Merge the factor graphs and recombine the decoupled features on one graph;

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

[0065] The third aspect of the present invention relates to a fault location device based on a feature decoupled graph network, which is characterized in that it includes a memory and one or more processors. An executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the fault location method based on the feature decoupled graph network of the present invention.

[0066] The fourth aspect of the present invention relates to a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the fault location method based on the feature decoupled graph network of the present invention.

[0067] Compared with the prior art, the innovation of the present invention lies in: the invention does not rely on the prior knowledge required by users to establish mathematical and physical models, but makes full use of the data characteristics and topological space characteristics of the power grid bus, models the power grid tabular data as a graph structure, combines two algorithm architectures of decoupled representation learning and graph convolutional neural network, and while having the modeling ability based on the topological structure, ensures a strong representation learning ability. Finally, it achieves the beneficial effects of higher versatility, simpler operation, better fitting the real power grid data characteristics, being able to discover hidden patterns and structures in the power network graph data, thereby improving the positioning accuracy and robustness. Description of the Drawings

[0068] Figure 1 It is a schematic diagram of the system structure of the present invention, where: 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, 5 - power grid fault location regression task module.

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

[0070] Figure 3 It is a schematic diagram of the device of the present invention. Specific implementation manner

[0071] In order to make the content of the present invention easier to be clearly understood, the present invention will be further described in detail below according to specific embodiments in conjunction with the accompanying drawings.

[0072] Embodiment 1

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

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

[0075] Step S02: Use the line graph method to construct the topological structure of the power grid nodes and transmission lines, transfer the data features on the edges to the graph nodes, and organize multiple graphs into a large graph through batch processing;

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

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

[0078] Step S05: Further feature aggregation is achieved by taking the weighted sum of its neighbor nodes, that is, the graph convolution step. After the aggregation step, different factor graphs represent different types of feature subsets in the power grid structure. Merge the factor graphs and recombine the decoupled features on one graph;

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

[0080] Among them, step S01 specifically includes:

[0081] Perform simulation according to the PSASP power failure analysis software; construct the specific topological structure of the power system, and by modifying network parameters such as the number of grid nodes, network generation method, fault line number, fault type, etc., generate results through time-domain simulation according to the software, and obtain power grid fault location data samples with fault location labels. The fault location label is an integer between 0 and 100, and the data features include bus voltage, current, resistance, active power, bus amplitude, bus frequency, V / I angle, voltage phase angle, etc.

[0082] Among them, step S02 specifically includes:

[0083] Since the data features of the power grid are all generated on the transmission lines (such as current, voltage, power, etc.), in order to facilitate the decoupling and other processing of the features, a line graph is used to construct the network structure. For an undirected graph G=(V, E), the line graph L(G)=(V', E') represents the adjacency relationship between the edges of G, where V={v 1 , v 2 ,..., v m} is the set of graph nodes v i , E={e 1 , e 2 ,..., e n} is the set of graph edges e i . The method of constructing L(G) is as follows: for any edge e i ∈E in G, it is taken as a vertex v i ’∈V'; and for each group of two edges in G that share the same vertex v i ∈V, the corresponding vertices in L(G) are connected as an edge, and thus the vertex v i is converted into e i '∈E'. Specifically, the connection combinations of the 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 ), then L(G) has an edge e i ' = (v i ', v j ') = v i . Converting to a line graph can transfer the data features on the edges to the graph nodes, which is more conducive to the decoupling of graph-level features. Finally, batch processing of the images is completed to make the embedding of the graph structure simpler.

[0084] Among them, step S03 specifically includes:

[0085] Decouple the features of the converted graph structure. The feature decoupling adopted belongs to graph-level decoupling, that is, multiple subgraphs are generated after decoupling. The input graph is decomposed into several factor graphs in the feature decoupling stage, and the specific number needs to be designed and input manually according to the actual problem. Use the node feature set and the set of edges e = {e 1 , e 2 ,..., e n} to represent the input graph structure. Among them, h i represents the data features of dimension F on the transmission line, that is Multiply the input feature structure H by a linear transformation matrix and transform it to a new hidden space H', obtaining

[0086] The present invention uses a mechanism similar to the graph attention network to calculate the attention scores of edges:

[0087] s ij = ψ(Wh i , Wh j ) (1)

[0088] Among them is a shared attention mechanism, which is responsible for calculating the attention mechanism score s ij of the edges of the factor graph e after taking the features of nodes i and j as inputs, and adopts the form of a single-layer perceptron in implementation. The attention mechanism score s ij indicates the importance of the features of node j to node i. When the score is 0, it is regarded that other nodes do not contribute to this node.

[0089] The present invention uses a masked attention mechanism to add the graph structure to the convolution process between nodes. Specifically, only calculate the contribution of node j that meets the condition to node i, where is a certain neighborhood of i. Set this neighborhood to be first-order, that is, the set of all points directly connected to node i.

[0090] The present invention generates a new factor graph by directly generating the coefficients of the edges. Set a threshold, and regard the two nodes i and j connected by the edges with scores lower than this threshold as two nodes with weak relationships. Then use the transformed features to generate the following factor coefficients:

[0091]

[0092] Normalize the attention score to the range [0, 1] (where e is the natural constant) to obtain the coefficient E representing the edge between nodes i and j of the factor graph e. ije After calculating the attention scores of each edge of each factor graph, E can be used e to represent a single factor graph. Thus, the present invention completes the decoupling and representation of the factor graph.

[0093] Among them, step 4 specifically includes:

[0094] Convert this optimization problem into a graph classification problem to find an approximate solution by simply setting the factor graph labels. That is, guide the factor graph to generate different topological structures through the differences in labels. Therefore, the head is used as a discriminator, defined as follows:

[0095]

[0096] Among them is a three-layer graph autoencoder that takes the transformed feature H' and the generated factor graph E e as inputs to generate new node features, and then maps them 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.

[0097] The definition of the loss function for training the discriminator is shown in the following formula:

[0098]

[0099] Among them, 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. is the distribution of sample i, represents the probability that the generated factor graph has label c. 1 e＝c 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 all factor graph predictions for a single sample being correct, and then statistically calculates the negative average of the logarithm sums of all samples to obtain the loss value generated in one training.

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

[0101] Among them, step 5 specifically includes:

[0102] In the aggregation process, new node features are generated by taking the weighted sum of their neighbor nodes, that is, the ordinary graph convolution process. The general formula for graph convolution can be written as:

[0103]

[0104] Among them is a way to standardize the adjacency matrix A, which can ensure that the contribution of each node feature is relatively balanced when performing graph convolution operations. Similarly, the present invention achieves this purpose by normalizing the factor graph node features by the square root of their degrees, that is, E ije / c ij E ije is the coefficient of the edge from node i to node j in the factor graph e, and c ij is the normalization term calculated according to the degrees of nodes i and j. The convolution aggregation process is expressed as:

[0105]

[0106] Among them represents the new feature aggregated by node i from the factor graph e at the l + 1 layer, represents all the neighbor nodes of node i in the input graph, and W (l) is the linear transformation matrix in the feature decoupling step.

[0107] The feature factor graph generated by decoupling is combined according to the following formula to recombine the decoupled features on one graph:

[0108]

[0109] where || represents the concatenation operation. Thus, the output network with enhanced features is obtained.

[0110] Among them, step 6 specifically includes:

[0111] Mapping all node features to a one-dimensional fault location through the self-attention mechanism.

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

[0113]

[0114] Since then, the power grid fault location task has been completed.

[0115] To illustrate the superiority of the model method proposed by the present invention, the present invention compares the performance using different models. The experiments selected the K-nearest neighbor algorithm, random forest, graph convolution network, and fully connected network to form a comparison with the method of the present invention, and statistically calculated the accuracy of fault location under different fault types. The bolded data are the results with the highest accuracy, and the underlined data are the results with the second highest accuracy. It is not difficult to conclude that the fault location accuracy achieved by the present invention is the highest.

[0116] Fault type 1 Fault type 2 Fault type 3 Fault type 4 K-nearest neighbor algorithm 0.594 0.985 0.992 0.585 Random forest 0.794 1 1 0.808 Graph convolutional network 0.884 <![CDATA 0.988 > <![CDATA 0.996 > <![CDATA 0.967 > Fully connected network <![CDATA 0.975 > 1 1 0.843 The present invention 0.988 1 1 0.974

[0117] Example 2

[0118] This example relates to a system for implementing the fault location method based on the feature decoupled graph network of Example 1, including:

[0119] A power grid line fault dataset construction module for generating a large dataset composed of bus features and fault locations;

[0120] A line graph conversion and graph structure construction module that constructs the power network topology of power grid nodes and transmission lines in the form of a line graph, transfers the data features on the transmission lines to the graph nodes, and organizes multiple graphs into a large graph through batch processing;

[0121] A feature decoupling module that multiplies the graph structure data by a weight matrix and converts it to a new hidden space, calculates the attention scores of the edges according to the graph attention network mechanism within the neighborhood to obtain the strength of the relationship between nodes, and can represent a single factor graph after calculating the attention scores of each edge of each factor graph, thus completing the decoupling and representation of the factor graph;

[0122] A classification label setting module for simply setting the classification labels of the factor graphs, and introducing an additional head as a discriminator in the disentangling layer to avoid the degradation of the factor graphs;

[0123] A feature aggregation and merging module realizes further feature aggregation by taking the weighted sum of its neighbor nodes, that is, the graph convolution step. After the aggregation step, different factor graphs represent different types of feature subsets in the power grid structure. Merge the factor graphs and recombine the decoupled features on one graph;

[0124] A power grid fault location regression task module for extracting the features at the corresponding fault locations of the aggregated graph features, mapping them to one-dimensional fault locations, and completing the final conversion using an activation function.

[0125] Example 3

[0126] Such as Figure 3 , this example relates to a fault location device based on a feature decoupled graph network, including a memory and one or more processors. An executable code is stored in the memory, and when the one or more processors execute the executable code, it is used to implement the fault location method based on the feature decoupled graph network of Example 1.

[0127] Example 4

[0128] The fourth aspect of the present invention relates to a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, it implements the fault location method based on the feature decoupled graph network of the present invention.

[0129] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments. The protection scope of the present invention also covers equivalent technical means that can be conceived by those skilled in the art based on the inventive concept of the present invention.

Claims

1. A fault location method based on a feature decoupling graph network comprises the following steps: Step 1: Based on the grid topology and sensor data, a time domain simulation is performed using power fault analysis software to generate a large data set consisting of bus characteristics and fault locations; Step 2: Use line graphs to construct the power network topology structure for the grid nodes and transmission lines, transfer the data features on the transmission lines to the graph nodes, and organize multiple graphs into a large graph through batch processing; Step 3: Multiply the graph structure data by the weight matrix and transform it into the new hidden space. Calculate the attention score of the edge 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, thereby completing the decoupling and representation of the factor graph. Step 4: Simply set the factor graph classification label and introduce an additional head as a discriminator in the disentanglement layer to avoid the degradation of the factor graph; 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, different factor graphs represent different types of feature subsets in the power grid structure. The factor graphs are merged to reassemble the decoupled features into one graph. Step 6: Extract the features at the corresponding fault position of the aggregated graph features, map them to the one-dimensional fault position, and use the activation function to complete the final conversion.

2. The fault location method based on the feature decoupling graph network according to claim 1, characterized in that: Step 1 includes: According to the PSASP power fault analysis software, simulation is performed; the specific topology of the power system is constructed, and the network parameters such as the number of grid nodes, network generation method, fault line number, fault type, etc. are modified, and the time domain simulation results are generated according to the software to obtain the grid fault location data sample with the fault location label. The fault location label is an integer 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 the feature decoupling graph network according to claim 1, characterized in that: Step 2 includes: The network structure is constructed in the form of a line graph; for an undirected graph G = (V, E), the line graph L(G) = (V', E') represents the adjacency relationship between the edges of G, where V = {v1, v2, ..., v m } is the graph node v i The set of E = {e1,e2,...,e n } is the edge e of the graph i A collection of The way to construct L(G) is as follows: for any edge e in G i ∈E, and regard it as a vertex v of L(G) i '∈V'; and for each group of G that shares the same vertex v i ∈V, connect the corresponding vertices in L(G) as edges, so that vertex v i Convert to e in L(G) i '∈E'; Specifically, the connection combinations of the power grid nodes are numbered. If the input graph L(G) has a graph node v i '=e i =(v i ,v k ), v j '=e j =(v i ,v j ), then L(G) has an edge e i '=(v i ',v j ')=v i Converting to a line graph can transfer the data features on the edge to the graph nodes, which is conducive to the decoupling of graph-level features; finally, batch processing of the image is completed to make the embedding of the graph structure easier.

4. The fault location method based on the feature decoupling graph network according to claim 1, characterized in that: Step 3 includes: The converted graph structure is feature decoupled; the feature decoupling adopted belongs to graph-level decoupling, that is, multiple subgraphs are generated after decoupling; the input graph is decomposed into several factor graphs in the feature decoupling stage, and the specific number is designed and input according to the actual problem; the node feature set is used and the set of edges e={e1,e2,...,e n } represents the input graph structure, where h i Represents the data characteristics of the F dimension on the transmission line, that is, Multiply the input feature structure H by a linear transformation matrix Transform to a new latent space H′ and get The attention scores of the edges are calculated using a mechanism similar to the graph attention network: s ij =ψ(Wh i ,Wh j ) (1) in It is a shared attention mechanism that takes the features of nodes i and j as input and calculates the attention score s of the edge of factor graph e ij , a single-layer perceptron was used in its implementation; the attention mechanism score s ij It indicates the importance of the feature of node j to node i. When the score is 0, it is considered that other nodes do not contribute to the node. The masked attention mechanism is used to add the graph structure to the convolution process between nodes; specifically, only nodes that meet The contribution of node j to node i under the condition is a neighborhood of i; the neighborhood is set to first order, that is, the set of all points directly connected to node i; A new factor graph is generated by directly generating edge coefficients; a threshold is set, and two nodes i and j connected by edges with scores below the threshold are considered to be two nodes with weak mutual relationship. Then, the transformed features are used to generate the following factor coefficients: Normalize the attention score to the interval [0,1], where e is a natural constant, and get the coefficient E used to represent the edge between nodes i and j in the factor graph e ije ; After calculating the attention score of each edge of each factor graph, we can use E e To represent a single factor graph, the decoupling and representation of the factor graph is completed.

5. The fault location method based on the feature decoupling graph network according to claim 1, characterized in that: Step 4 includes: By simply setting the factor graph labels, the optimization problem is converted into a graph classification problem to seek an approximate solution. That is, the difference in labels guides the factor graph to generate different topological structures. Therefore, the head is used as the discriminator, which is defined as follows: in It is a three-layer graph autoencoder that transforms the transformed features H' and the generated factor graph E e As input, new node features are generated, which are 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 graph-level feature representations. The loss function used to train the discriminator is defined as follows: 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. is the distribution of sample i, Represents the probability that the generated factor graph has label c; 1 e＝c is an indicator function. When the predicted label is correct, the value is 1, otherwise it is 0. This formula calculates the sum of the logarithms of the probability distribution of all factor graphs predicting the correctness of a single sample, and then calculates the logarithm of all samples and takes the negative average, thus obtaining the loss value generated by one training. The loss of the whole framework is defined as They represent the decoupling layer loss and task loss respectively, and use cross entropy loss as the loss estimation method between the target and the actual.

6. The fault location method based on the feature decoupling graph network according to claim 1, characterized in that: Step 5 includes: In the aggregation process, new node features are generated by taking the weighted sum of their neighbor nodes, which is the ordinary graph convolution process. The general formula of graph convolution is written as: in It is a way to standardize the adjacency matrix A to ensure that the contribution of each node feature is relatively balanced when performing graph convolution operations; therefore, similarly, by normalizing the factor graph node features by the square root of their degree, that is, E ije / c ij , E ije is the coefficient of the edge from node i to node j in factor graph e, c ij is a normalized term calculated based on the degree of node i and node j. The convolution aggregation process is expressed as: in represents the new feature obtained by node i in the l+1 layer from the factor graph e. Represents all neighbor nodes of node i in the input graph, W (l) is the linear transformation matrix in the feature decoupling step; The characteristic factor graphs generated by decoupling are merged according to the following formula, so that the decoupled features are recombined on one graph: Here, || represents the connection operation. Thus, the output network with enhanced features is obtained.

7. The fault location method based on the feature decoupling graph network according to claim 1, characterized in that: Step 6 includes: All node features are mapped to one-dimensional fault locations through the self-attention mechanism; Complete the power grid fault location task.

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

9. A fault location device based on a feature decoupling graph network, characterized in that: It comprises a memory and one or more processors, wherein 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 the feature decoupling graph network described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the fault location method based on a feature decoupling graph network described in any one of claims 1 to 7 is implemented.

Images