Power distribution system fault positioning method and device

Abstract

The invention is suitable for the technical field of power distribution networks, and particularly relates to a power distribution system fault positioning method and device. According to the method, an attribute undirected graph is constructed according to collected real-time voltage and / or real-time current phases of measurement points of each phase in N measured buses in a power distribution system, so that a vertex attribute matrix and a vertex adjacency matrix are obtained, vertex characteristics are updated according to the vertex attribute matrix and the vertex adjacency matrix, and the vertex characteristics are updated according to the updated vertex characteristics. Obtaining a vertex feature vector of each vertex, performing down-sampling clustering on the attribute undirected graph according to the vertex feature vectors to obtain hyper-vertex sub-graphs, performing feature extraction on the hyper-vertex sub-graphs by using read-out operation to obtain graph feature vectors, and performing classification on the graph feature vectors to obtain the attribute undirected graph. And determining the fault probability that any line section formed by the measurement points is a fault line section to obtain a fault position. Through a mode of forming graph data, accurate analysis of the graph data is realized, so that classified calculation of fault probability is carried out to obtain a fault position of the power distribution system.

Description

Technical Field

[0001] The present application is applicable to the technical field of power distribution network, and in particular, relates to a method and device for locating faults in a power distribution system. Background Art

[0002] Failures in distribution systems are inevitable due to various natural and man-made reasons. If the faults are not cleared in time, irreversible damage to power equipment may occur, or even power outages may occur. Therefore, it is important to accurately locate the faulty line segment so that fault clearing operations can be performed in a timely manner. Traditionally, fault location in distribution systems is mainly accomplished by using a set of coordinated protection devices. Assuming a single power source and a radial distribution system, the location of the fault is determined by monitoring these coordinated protection devices. However, with the popularity of distributed generation (DG), distribution systems are increasingly equipped with multiple power sources that result in bidirectional power flows, making traditional fault location methods ineffective.

[0003] Situational awareness of modern distribution systems has been enhanced by deploying advanced metering infrastructure, micro phasor measurement units (μPMUs), and communication networks within distribution systems. This has enabled the application of artificial intelligence (AI) techniques, including artificial neural networks, support vector machines, and genetic algorithms, which are widely used for fault location. Typically, the aforementioned AI-based fault location methods involve manually extracted fault features. However, manual feature extraction is time-consuming, and these features may not be effective in locating fault lines with weak fault features. In addition, the above AI-based fault location methods are limited to structured data that conforms to the data model and follows a consistent order. Recent advances in deep learning (DL), especially in the fields of convolutional neural networks (CNN) and recurrent neural networks (RNN), have attracted the attention of academia and industry. Due to the advantages of deep learning in self-learning feature extraction and managing unstructured data with spatial and temporal correlations (such as images, videos, and audio), recent research focuses on using DL-based methods to improve the efficiency of fault location in distribution systems.

[0004] Despite the outstanding performance of DL in various fields, its application in fault location in distribution systems is still hindered by many obstacles. First, the spatial information in the measurements is crucial to handle the changes in the distribution system topology. Existing DL-based fault location methods are limited to data defined in the Euclidean domain and cannot handle the spatial information in the measurements. Second, despite the high accuracy of μPMUs, measurement errors are inevitable due to various factors such as potential / current transformers, time synchronization, and communication. In the presence of measurement errors, existing DL-based fault location methods may produce inaccurate results. Finally, in order to improve the accuracy of DL-based fault location methods, synthetic fault data are usually used to provide sufficient training data. However, the inherent modeling mismatch leads to inconsistent distributions between synthetic fault data and field fault data. When the fault location model learned from synthetic fault data is applied to field fault data, the cross-domain generalization error may lead to inaccurate results.

[0005] Therefore, how to optimize fault data to improve the accuracy of fault location becomes an urgent problem to be solved. Summary of the invention

[0006] In view of this, an embodiment of the present application provides a method and device for locating faults in a power distribution system to solve the problem of how to optimize fault data to improve the accuracy of fault locating.

[0007] In a first aspect, an embodiment of the present application provides a method for locating a fault in a power distribution system, the method comprising: Collect the real-time voltage and / or real-time current phase of the measurement point of each phase in the N measured buses in the power distribution system, and construct an attribute undirected graph according to the real-time voltage and / or real-time current phase of the measurement point, wherein the vertices of the attribute undirected graph are the measurement points, the attributes of the attribute undirected graph are the attribute vectors of the corresponding measurement points, the edges of the attribute undirected graph are the line segments connecting the two vertices, and N is an integer greater than zero; According to the attribute undirected graph, a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information are obtained, and vertex features are updated according to the vertex attribute matrix and the vertex adjacency matrix to obtain a vertex feature vector of each vertex, wherein the initial vertex feature of the vertex is the attribute vector in the vertex attribute matrix; According to the vertex feature vector, downsampling and clustering are performed on the vertices in the attribute undirected graph to obtain a super vertex subgraph, and feature extraction is performed on the super vertex subgraph using a readout operation to obtain a graph feature vector; Classification is performed on the graph feature vector to determine the fault probability that any line section formed by the measurement points is a fault line section, and the fault location of the power distribution system is determined based on the fault probability.

[0008] In a second aspect, an embodiment of the present application provides a power distribution system fault location device, the power distribution system fault location device comprising: An undirected graph construction module, used for collecting the real-time voltage and / or real-time current phase of the measurement point of each phase in the N measured buses in the power distribution system, and constructing an attribute undirected graph according to the real-time voltage and / or real-time current phase of the measurement point, wherein the vertices of the attribute undirected graph are the measurement points, the attributes of the attribute undirected graph are the attribute vectors of the corresponding measurement points, the edges of the attribute undirected graph are the line segments connecting the two vertices, and N is an integer greater than zero; A feature vector updating module is used to obtain a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information according to the attribute undirected graph, and update the vertex features according to the vertex attribute matrix and the vertex adjacency matrix to obtain a vertex feature vector of each vertex, wherein the initial vertex feature of the vertex is the attribute vector in the vertex attribute matrix; A feature vector extraction module is used to downsample and cluster the vertices in the attributed undirected graph according to the vertex feature vector to obtain a super-vertex subgraph, and use a read operation to extract features from the super-vertex subgraph to obtain a graph feature vector; a fault location module is used to classify the graph feature vector, determine the fault probability that any line section formed by the measurement points is a fault line section, and determine the fault location of the distribution system according to the fault probability.

[0009] Compared with the prior art, the embodiments of the present application have the following beneficial effects: the present application collects the real-time voltage and / or real-time current phase of the measuring point of each phase in the N measured buses in the distribution system, constructs an attributed undirected graph based on the real-time voltage and / or real-time current phase of the measuring point, obtains a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information based on the attributed undirected graph, updates the vertex features based on the vertex attribute matrix and the vertex adjacency matrix, obtains the vertex feature vector of each vertex, performs downsampling clustering on the vertices in the attributed undirected graph based on the vertex feature vector, obtains a super-vertex subgraph, extracts features from the super-vertex subgraph using a readout operation, obtains a graph feature vector, performs classification on the graph feature vector, determines the fault probability that any line section formed by the measuring points is a fault line section, and determines the fault location of the distribution system based on the fault probability. By forming graph data, accurate analysis of the graph data is achieved, thereby performing classification calculation of the fault probability to obtain the fault location of the distribution system. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0011] Figure 1 It is a flow chart of a method for locating a fault in a power distribution system provided in Embodiment 1 of the present application; Figure 2 It is a structural schematic diagram of a three-phase power distribution system including four busbars under test provided in Example 1 of the present application; Figure 3 It is a schematic diagram of a framework based on a graph attention network provided in Example 1 of the present application; Figure 4 This is a schematic diagram of a framework based on graph attention and transfer learning network provided in Example 2 of the present application; Figure 5 This is a topological diagram of a 123 bus test feeder provided in Example 3 of the present application; Figure 6 This is a sample schematic diagram of a load level of an experiment provided in Example 3 of the present application; Figure 7 This is a schematic diagram of the results of a fault simulation of an experiment provided in Example 3 of the present application; Figure 8 This is a schematic diagram of the results of a fault simulation of Experiment 2 provided in Example 3 of the present application; Fig. 9 This is a sample schematic diagram of a load level of Experiment 2 provided in Example 3 of the present application; Fig.10 This is a schematic diagram comparing the overall positioning accuracy of different methods under different numbers of training samples provided in Example 3 of the present application; Fig.11 This is a comparative schematic diagram of the overall positioning accuracy without using enhanced data for training provided in Example 3 of the present application; Fig.12 This is a schematic diagram of comparing the overall positioning accuracy under training with different numbers of enhanced training samples provided in Example 3 of the present application; Fig.13 This is a comparative schematic diagram of overall positioning accuracy under different trade-off parameter training provided in Example 3 of the present application; Fig.14 It is a structural schematic diagram of a power distribution system fault locating device provided in Example 4 of the present application. DETAILED DESCRIPTION

[0012] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0013] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0014] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0015] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0016] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0017] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0018] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0019] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0020] It should be understood that the size of the serial numbers of the steps in the following embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0021] In order to illustrate the technical solution of the present application, a specific embodiment is provided below for illustration.

[0022] See also Figure 1 , is a flow chart of a method for locating a power distribution system fault provided in the first embodiment of the present application. The method for locating a power distribution system fault can be applied to a computer device in a network environment. The computer device executes a corresponding program to implement the method. The method for locating a power distribution system fault includes the following steps: Step S101, collecting the real-time voltage and / or real-time current phase of the measuring point of each phase in N measured buses in the power distribution system, and constructing an attribute undirected graph according to the real-time voltage and / or real-time current phase of the measuring point.

[0023] In this embodiment, the vertices of the attribute undirected graph are measurement points, the attributes of the attribute undirected graph are attribute vectors corresponding to the measurement points, the edges of the attribute undirected graph are line segments connecting two vertices, and N is an integer greater than zero.

[0024] For example, a distribution system consisting of an Nb bus and an Nl line segment has a single fault, which can be divided into a three-phase (TP) short circuit, a line to ground (LG), a double line to ground (DLG) or a line to line (LL) fault. At this time, the faulty line segment is located by measuring the pre-fault and post-fault phasor measurements collected by the busbar through the corresponding detection technology. The measurement location where the detection technology is installed is the measurement point, and the corresponding measurement data may include voltage, current, current phase, etc.

[0025] In this embodiment, the measurement can be performed using a micro-synchronous phasor measurement unit (μPMU), which can obtain high-resolution measurements of voltage and current phasors at a maximum sampling rate of 120 Hz, which helps to accurately locate faults in the distribution system. In practice, only a few buses in the distribution system are equipped with μPMUs, and the real-time status of the distribution system under study cannot be fully observed.

[0026] In order to take into account both spatial information and measurement data, an attributed undirected graph G = (U, E, M) is defined for the distribution system, where U is the set of vertices, each vertex u∈U represents a phase of the bus, E is the set of edges, and M is the vertex attribute matrix, each element of which represents the attribute data of the corresponding vertex. Let e uv =(u,v)∈E represents the line segment connecting vertex u and vertex v. The first-order neighborhood of vertex u is defined as N(u)={v∈u|(u,v)∈E}. If the adjacency matrix A is a symmetric matrix with an entry, then a uv ={(1,if e uv ∈E; 0, otherwise), M∈R |U|×2 is a matrix whose row vector m u Represents the attribute vector of vertex u.

[0027] Optionally, constructing an attribute undirected graph according to the real-time voltage and / or real-time current phase of the measuring point includes: taking the measuring point with the real-time voltage and the real-time current phase as the first-class vertex, and determining the attribute vector of the first-class vertex as the real-time voltage and the real-time current phase of the corresponding measuring point; Taking the measurement points with real-time voltage but without real-time current phase as the second-type vertices, and determining the attribute vectors of the second-type vertices as the preset vectors; Get the edges connecting the first-category vertices and the second-category vertices, and construct an attribute undirected graph based on the first-category vertices, the second-category vertices, the attribute vector of each vertex, and the edges.

[0028] Among them, the measurement points with missing data cannot be used as vertices. The measurement points that can measure the voltage and current phase are the first type of vertices, and the vertices whose phases cannot be measured are the second type of vertices. The attribute vectors of the first type of vertices can be represented by their real-time voltage and real-time current, while the attribute vectors of the second type of vertices can be represented by a preset vector.

[0029] like Figure 2As shown, it is a structural schematic diagram of a three-phase power distribution system including four measured buses provided in the first embodiment of the present application. The vertex attribute vector of the measured bus is the measurement vector (ΔV, ΔΦ) of each phase, and the vertex of the unmeasurable node is taken as the intermediate vertex. Among them, the b and c phases of the measured bus 3 and the b phase of the bus 4 are missing, and the current phase of the measured bus 2 is considered to be an unmeasurable node. Therefore, it belongs to the intermediate vertex. Figure 2 In the above example, U = {1a, 1b, 1c, 2a, 2b, 2c, 3a, 4a, 4c} represents the vertex set, and u∈{1a, 1b, 1c, 3a, 4a, 4c} corresponds to the attribute vector m. u =(ΔV u , ΔΦ u ), v∈{2a,2b,2c} corresponds to the attribute vector m v =(1,1), u is a vertex of the first type, and v is a vertex of the second type.

[0030] Step S102, obtaining a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information according to the attribute undirected graph, and updating the vertex features according to the vertex attribute matrix and the vertex adjacency matrix to obtain a vertex feature vector of each vertex.

[0031] The initial vertex feature of the vertex is the attribute vector in the vertex attribute matrix.

[0032] In this embodiment, the processing of step S102, step S103 and step S104 can be placed in a model for execution, which can specifically be a model of a graph attention network (Graph Attention Networks, GATs) architecture. Of course, models of other network architectures can also be used to meet the requirements of feature extraction and classification of graph data, etc., which is not limited here.

[0033] From the attribute undirected graph, we can extract the vertex attribute matrix representing the attribute vector and the vertex adjacency matrix representing the control information. By updating the features of the vertex with the features of the adjacent vertices, we can capture the vertex-level features, thereby obtaining a more accurate vertex feature vector based on the adjacent vertices.

[0034] For example, GAT based on spatial graph convolution (SG-Conv), such as Figure 3As shown, a framework schematic diagram based on a graph attention network provided in Example 1 of the present application is provided, and the input is spatial information (i.e., the adjacency matrix A) and measurement data (i.e., the vertex attribute matrix M). SG-Conv captures vertex-level features by updating the features of a vertex with the features of adjacent vertices, where the initial features of a vertex are its attribute vectors. After each SG-Conv operation, the vertices are downsampled using a pooling operator. Then, the vertex-level features are embedded using a readout operator to generate graph-level features as input to a multi-layer perceptron (MLP) classifier. The output of the MLP is processed using a softmax activation function to provide a probability distribution indicating the likelihood that any line segment i∈E is a faulty line segment.

[0035] Optionally, the vertex features are updated according to the vertex attribute matrix and the vertex adjacency matrix to obtain a vertex feature vector for each vertex, including: Get the pre-trained parameterized weight matrix and parameterized weight vector; For any target vertex, determine the adjacent vertices of the corresponding target vertex according to the target vertex adjacency matrix, and determine the vertex feature vectors of the target vertex and the adjacent vertices according to the vertex attribute matrix; According to the parameterized weight matrix, the vertex feature vectors of the target vertex and the adjacent vertex are concatenated to obtain a concatenation result, and the transpose of the parameterized weight vector is multiplied by the concatenation result to obtain a multiplication result; Use a preset activation function to activate the multiplication result to obtain a single activation result, traverse all vertices adjacent to the adjacent vertex, and determine the sum of all single activation results as a comprehensive activation result; The ratio of a single activation result to the comprehensive activation result is used as the significance coefficient of the target vertex's feature to the adjacent vertices. The significance coefficient represents the significance of a vertex's feature to its adjacent vertices. Obtain the significance coefficients of the features of all vertices adjacent to the target vertex and the corresponding vertex feature vectors to the target vertex, and obtain the updated vertex feature vector of the target vertex based on all the significance coefficients and vertex feature vectors combined with the parameterized weight matrix.

[0036] This embodiment defines a significance coefficient that can be learned from each measurement end-to-end. This coefficient quantifies the importance of each measurement for feature extraction while taking into account the measurement accuracy. For weighted distribution system fault location that is inversely proportional to the measurement error, defining a significance coefficient can better avoid the error caused by considering only equal weights for adjacent vertices.

[0037] x u i ∈R J It is represented as the feature vector of vertex u entering the i-th SG-Conv operator. For i=0, xu 0 =x u T , where T represents the transposition operation. Similarly, this application defines x u (i+1) ∈R J is the updated feature vector of vertex u after passing through the i-th SG-Conv operator. In addition, there is a learnable parameterized weight matrix W in an SG-Conv operator i and parameterized weight vector Among them, W i ∈R (J ' ×J) As a linear transformation, it is shared among the vertices of the i-th SG-Conv operator. In addition, the cardinality J′ of the output feature vector may be different from the cardinality J of the input feature vector, which is a hyperparameter. Then, in the i-th SG-Conv operator, the normalized saliency coefficient is defined as The significance of the feature of vertex v to its adjacent vertex u can be expressed as: Where R is the leaky rectified linear unit (ReLU) activation function, is the parameterized weight vector of the i-th SG-Conv operator, || is the concatenation operator, W i is the parameterized weight matrix of the SG-Conv operator. Among them, the vertices with larger measurement errors have smaller significance coefficient values, and their features contribute less to the adjacent points. Optionally, based on all the significance coefficients and vertex feature vectors, combined with the parameterized weight matrix, an updated vertex feature vector of the target vertex is obtained, including: Multiply the significance coefficient of the feature of each vertex adjacent to the target vertex, the vertex feature vector of the corresponding vertex, and the parameterized weight matrix to obtain the product result of each vertex adjacent to the target vertex; The sum of all product results is multiplied by the nonlinear parameter in the exponential linear unit to obtain the updated vertex feature vector of the target vertex.

[0038] Among them, the updated vertex feature vector is calculated by the following formula, and the vertex features are updated by applying the exponential linear unit (ELU) nonlinearity σ based on the significance coefficient, as follows: In the formula, x u i+1The updated vertex feature vector can be input into the i+1th SG-Conv operator. The above method can accurately use the features of the adjacent vertices to update the features of the current vertex, thereby achieving feature re-expression.

[0039] Step S103, based on the vertex feature vectors, downsample and cluster the vertices in the attribute undirected graph to obtain a super vertex subgraph, and use a readout operation to extract features from the super vertex subgraph to obtain a graph feature vector.

[0040] Among them, Figure 3 In the method, based on the vertex feature vector, a pooling operation is performed after each SG-Conv operator to realize the downsampling clustering of the attribute undirected graph, thereby obtaining the clustering result. Each clustering result is represented by a vertex subgraph, and the corresponding vertex is defined as a super vertex, thereby obtaining a super vertex subgraph. Then, the reading of the super vertex subgraph is realized by the readout operation, thereby obtaining the graph feature vector.

[0041] Optionally, according to the vertex feature vector, down-sampling and clustering are performed on the vertices in the attribute undirected graph to obtain a super vertex subgraph, and a readout operation is used to extract features from the super vertex subgraph to obtain a graph feature vector, including: For any vertex, obtain the significance coefficient of the vertex to all vertices adjacent to the vertex, determine the sum of all significance coefficients as the importance score of the vertex, and traverse to obtain the importance scores of all vertices; According to the importance scores of all vertices and the vertex feature vectors of the corresponding vertices, combined with a preset pooling ratio, a super vertex is determined from the vertices in the attribute undirected graph, and the adjacency matrix and the vertex feature matrix of the super vertex are determined as a super vertex subgraph; Use the readout operation to extract features from the super-vertex subgraph and obtain the graph feature vector.

[0042] The pooling operator downsamples the vertices of the graph by clustering them into subgraphs, each of which is represented by a supervertex. Then, a single supervertex is generated during the readout operation to represent the entire graph. The pooling operator after the i-th SG-Conv operator is defined as the i-th pooling operator. The pooling operator constructs a supervertex set U i+1 , where |U i+1 |=κ i |U i |, where the pooling ratio κ i ∈(0,1) is a hyperparameter such that the set U i+1 The maximum number of fault discrimination features is retained. Based on the above significance coefficient The physical interpretation of is the coefficient The larger the value, the more important the vertex v is to the feature of vertex u. Therefore, this application defines an important score To express the relative importance of the features of vertex v to its neighbors, the determination formula is as follows: Among them, the larger value The vertices of |U are crucial to represent the graph because they contribute more to feature embedding and / or have high connectivity. Therefore, the required |U i+1 |Supervertices can be selected as follows: U i+1 =U idx i Among them, i is the set of importance scores of all vertices in the SG-Conv operator, is a function that returns the index of the top value in SG-Conv and is an indexing operator. The adjacency and vertex feature matrices after pooling can be determined as: A i+1 =A idxidx’ i Among them, A idxidx’ i Represents the adjacency matrix A i exists in the idxth th row and idx th Elements of the column, X idx i is the vertex feature matrix indexed row by row, and × is the element-wise matrix multiplication. Finally, a single super vertex with a graph-level feature vector xr is generated. After the readout operation, the MLP uses the graph-level feature vector xr to perform the classification task, giving the probability that any line section is a fault line section.

[0043] Optionally, obtain pre-trained parameterized weight matrices and parameterized weight vectors, including: An attribute undirected graph formed by test results obtained by performing a test experiment on the power distribution system and a probability that the label of the corresponding attribute undirected graph is a line section k are obtained as field data samples; The first network based on the graph attention mechanism is trained using field data samples to obtain a pre-trained first network, and a pre-trained parameterized weight matrix and a parameterized weight vector are extracted from the pre-trained first network.

[0044] Optionally, the training process of the first network is as follows: For the attribute undirected graph in the field data sample, a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information are obtained. According to the vertex attribute matrix and the vertex adjacency matrix, combined with the parameterized weight matrix and the parameterized weight vector of the current iteration, the vertex features are updated to obtain the vertex feature vector of each vertex; According to the vertex feature vector, the vertices in the attribute undirected graph are downsampled and clustered to obtain a super vertex subgraph, and the super vertex subgraph is extracted using a readout operation to obtain a graph feature vector; Classification is performed on the graph feature vector to determine the fault probability that any line section formed by the measurement points is a fault line section. According to the fault probability and label, combined with minimizing the cross entropy loss function, the parameterized weight matrix, parameterized weight vector and classification parameters of the current iteration are trained until the iteration conditions are met to obtain the pre-trained first network.

[0045] Since the GAT-based fault localization method proposed in this application is multi-label classification, GAT is trained by minimizing the cross entropy loss function as follows: In the formula, N is the total number of samples, G is j is the input graph of the jth sample, P(y j = k) is the label y of the jth sample j is the probability of event k, that is, line segment k is the fault line segment, θ f and θ m are the parameters of the feature extractor and MLP, respectively. For GAT, use the input graph G j The predicted probability of a failure event occurring in segment k.

[0046] Step S104, classifying the graph feature vector, determining the fault probability that any line section formed by the measurement points is a fault line section, and determining the fault location of the power distribution system according to the fault probability.

[0047] Among them, the graph feature vector is classified based on the multi-layer perceptron (MLP) classifier. MLP uses the graph-level feature vector xr to perform the classification task and gives the probability that any line section is a fault line section, thereby performing fault location based on the probability.

[0048] The present application collects the real-time voltage and / or real-time current phase of each phase of the measurement point of the N measured buses in the distribution system, and constructs an attributed undirected graph based on the real-time voltage and / or real-time current phase of the measurement point. Based on the attributed undirected graph, a vertex attribute matrix representing the attribute vector of the measurement and a vertex adjacency matrix representing the spatial information are obtained. Based on the vertex attribute matrix and the vertex adjacency matrix, vertex features are updated to obtain a vertex feature vector of each vertex. Based on the vertex feature vector, vertices in the attributed undirected graph are downsampled and clustered to obtain a super-vertex subgraph, and feature extraction is performed on the super-vertex subgraph using a readout operation to obtain a graph feature vector. The graph feature vector is classified to determine the fault probability that any line section formed by the measurement points is a fault line section, and the fault location of the distribution system is determined based on the fault probability. By forming graph data, accurate analysis of graph data is achieved, thereby performing classification calculations of fault probabilities to obtain the fault location of the distribution system.

[0049] See also Figure 4 , is a schematic diagram of a framework based on graph attention and transfer learning network provided in Example 2 of the present application. In order to train a GAT with acceptable reliability, a large amount of fault data needs to be collected under various fault conditions. However, since power system failures are uncommon, field fault data are usually insufficient. Synthetic fault data can be used as a suitable substitute for field fault data. However, the distribution difference between synthetic data and field data will reduce the performance of fault location.

[0050] In order to ensure that the optimal parameters of the GAT network learned using the comprehensive fault data can be successfully applied to the fault location of the field fault data, transfer learning (TL) is introduced in the fault location of the distribution system. In TL, the domain of the classification problem is defined as D = {X, p}, where X is the feature space and X∈X is the feature, that is, the feature vector x in this paper r ,p is the variable x r The joint probability distribution of and y. The subscripts S and T denote the domains of synthetic and field fault data samples, respectively. In the problem, both the classification based on synthetic data and field fault data use spatial information and measurement data as input. Therefore, the feature space X of the synthetic and field fault data domains is the same (i.e., X S =X T However, due to modeling inaccuracies, there will be a difference between the joint distribution of the two domains (i.e., p S ≠p T ), which leads to cross-domain generation errors between the two domains and hinders the direct application of parameters learned based on synthetic fault data to classification problems using field fault data.

[0051] Based on causal learning, the goal is to minimize the joint distribution difference between the two domains by performing separate priors and covariate shifts under appropriate conditions. For the GAT-based power system fault localization problem, the cause y (i.e., a specific type of fault occurring at a specific location) is associated with the effect x r (i.e. spatial information and measurement data), where the marginal probability distribution p(y) depends on the noise e y , and the conditional probability distribution p(x r |y) relies on imitating causality and noise The mechanism φ(y).

[0052] For fault location in distribution systems, the noise that affects the probability of a certain type of fault occurring at a specific location y Depends on several factors, including weather conditions, infrastructure vulnerability, animal movement, etc. r Noise is the combined effect of several factors, such as measurement error, load variation, and DG status. Then, the noise e y and can be viewed as two independent terms. Considering measurement errors, load changes, DG states and other disturbance factors as noise, φ(y) can be viewed as a deterministic mechanism completely governed by the physical laws of power flow. as additive noise.

[0053] The variation of p(y) across domains S and T is caused by inaccurately modeling the probability of failure using inconsistent information, i.e. 2) If Then p(x r |y) The variation across the two domains is mainly caused by inaccurately modeling measurement errors, load variations, and DG states, i.e., ex r S =ex r T . Due to e y and independence, and is an additive noise, this application can be solved by minimizing p(y S ) and p(y T ) and p(x r S |y S ) and p(x r T |y T ) to minimize the joint distribution p(x) between the two do-mains r ,y)=p(x r |y)p(y) difference, which can be achieved by changing ey and To be completed individually.

[0054] Then, we can minimize the difference in the joint distribution between S and T by following the steps below: 1) By minimizing d(p(y S ), p(y T ))make and Align so that p(y S )≈p(y T ), 2) Given p(y S )≈p(y T ), d(p(x r S |yS), p(x r T |yT)) By passing ex r S With ex r T Align so that p(x r S )≈p(x r T )minimum.

[0055] Where d(·) is the maximum mean difference (MMD) mean. In the first step, d(p(y S ), p(y T )) makes p(y S )≈p(y T ) is indeed a priori offset. Since, in the problem of this application, labels from both domains S and T are available, we can reweight the loss function To implement this a priori shift, look like this: Where α(k)=P(k) S / P(k) T , P(k) S and P(k) T Pay attention to the event probability of segment k being faulty in S and T respectively. According to the reweighted loss function, when solving the optimal parameters of the GAT network, segment samples that are more likely to fail in the S domain will be given greater weights.

[0056] Using class-based weights, we can enforce p(x r S )≈p(x r T )(also called covariate shift) to minimize d(p(x rS |y S ), p(x r T |y T )). Because, in the problem of this application, the feature is to use a parameter θ f The feature extractor extracts the feature, and this application can be solved by minimizing d(p(x r S ), p(x r T )) f To force p(x r S )≈p(x r T ). Based on the empirical expression of MMD, a parameter θ is expressed f The difference d(p(x r S ), p(x r T ))’s loss function L d (θ f ) is defined as follows: in, is a Gaussian kernel function with a standard deviation of δ, which is usually determined by the following calculation: Where MSD is the median of all samples x in domain S r The square of the distance between them.

[0057] In order to learn fault discrimination and domain invariant features, a trade-off term γ∈(0,1) is introduced into the overall loss function L of the GAT-based distribution fault location problem, which can be expressed as: Among them, the optimal parameters In order to improve the estimation of the empirical difference loss, a mini-batch gradient descent algorithm with a learning rate of ρ is used to solve (θ f * ,θ m * ).

[0058] like Figure 4 As shown, the synthetic samples and field samples are separated into N^ S <N S and N^ T <N T Both synthetic and live samples consider the shared parameters (θ f ,θ m ). In the forward pass, mini-batches from synthetic samples are used to evaluate the loss function The weight is α(k). Features generated based on small batches of synthetic samples and live samples Used to evaluate the difference loss function L d (θ f ). Then the loss function and L d (θ f ) is combined into the overall loss function L c (θ f ,θ m ). In order to reduce the variance of the gradient, the average gradient over the mini-batch is used in the back propagation To update the parameters (θ f ,θ m ).

[0059] Optionally, obtain pre-trained parameterized weight matrices and parameterized weight vectors, including: The obtained attribute undirected graph formed by the test results obtained by performing the test experiment on the power distribution system and the probability that the label of the corresponding attribute undirected graph is the line section k are used as the field data sample; The obtained attribute undirected graph formed by synthetic data obtained by simulating the power distribution system and the probability that the label of the corresponding attribute undirected graph is line section k are used as synthetic data samples; The second network integrating the graph attention mechanism and the transfer learning is trained by using the field data samples and the synthetic data samples to obtain a pre-trained second network, and a pre-trained parameterized weight matrix and a parameterized weight vector are extracted from the pre-trained second network.

[0060] Optionally, the training process of the second network is as follows: For the attribute undirected graph in the field data sample, a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information are obtained. According to the vertex attribute matrix and the vertex adjacency matrix, combined with the parameterized weight matrix and the parameterized weight vector of the current iteration, the vertex features are updated to obtain the vertex feature vector of each vertex; According to the vertex feature vector, the vertices in the attribute undirected graph are downsampled and clustered to obtain a super vertex subgraph, and the super vertex subgraph is extracted using a readout operation to obtain a graph feature vector; Classification is performed on the graph feature vector to determine the fault probability that any line section formed by the measurement points is a fault line section. According to the fault probability and label, combined with the minimization of the cross entropy loss function, the parameterized weight matrix, parameterized weight vector and classification parameters of the current iteration are trained until the iteration conditions are met to obtain a pre-trained graph attention network.

[0061] Both of the above two training processes can obtain a network that implements graph data analysis. The second embodiment of the present application specifically adopts a transfer learning method to obtain a trained network. The following are specific experimental results based on a network combined with transfer learning.

[0062] For example, the performance of the above distribution system fault location method in an IEEE 123 bus test feeder is evaluated, and the fault location performance is specifically evaluated by the location accuracy rate (LAR).

[0063] like Figure 5 As shown, it is a topological diagram of a 123-bus test feeder provided in Example 3 of the present application, which consists of 114 overhead line segments and is configured with various phase configurations. Training and test data sets are generated, and faults are simulated on different line segments on four different types of lines (i.e., TP, LG, DLG, and LL faults). For simulation purposes, the present application considers distributed loads, and the system load level varies according to the probability mass function (PMF). In addition, when switches 18-135 and 54-94 are operated, the system topology also changes. In addition, the present application considers DGs connected to buses 33, 39, and 83. In order to simulate the transfer learning process, the present application designs two experimental settings for domains S and T, respectively, as shown below: 1) Experimental setup in S domain: The experimental setup in S domain aims to obtain synthetic fault data. This application considers all four types of faults, five different fault impedance values ​​(0.0001, 0.001, 0.01 and 0.1pu), eight different DG state combinations and four different switch combinations. For each fault type, the samples generated by this application are different phase combinations considering TP, DLG and LL faults at the same time. Figure 6 As shown, it is a sample schematic diagram of a load level of an experiment provided in Example 3 of the present application, from Figure 6 40 load level samples are randomly selected from the PMF shown. Then, for a given combination of fault impedance value, DG state, switch state and load level, there are a total of 582 synthetic fault samples in this application. Therefore, a total of 3724800 synthetic fault samples are generated for domain S.

[0064] 2) Experimental setup in T domain: This application designs this experimental setup to generate field failure data. Figure 7 As shown, it is a schematic diagram of the results of a fault simulation of an experiment provided in Example 3 of the present application. This application does not consider all possible fault locations, but only Figure 7 Some fault locations are randomly selected from the PMF shown in Figure 8 to simulate 850 faults. For each selected fault location, the application Figure 8 A fault type is randomly selected with the probability of occurrence shown. Figure 8 As shown, it is a schematic diagram of the results of the fault simulation of Experiment 2 provided in Example 3 of the present application. The present application uses a normal distribution to sample the fault phase, fault impedance value, DG state and switch state of each fault. Fig. 9 As shown, it is a sample schematic diagram of a load level of Experiment 2 provided in Example 3 of the present application, from Fig. 9 The corresponding load levels are randomly sampled in the PMF shown. Therefore, a total of 850 field fault samples are generated for domain T.

[0065] Domains S and T contain different numbers of fault events that follow different distributions. In addition, they consider different distributions of fault types, fault impedance values, DG states, switch states, and load levels. Therefore, they have different joint distributions.

[0066] A. Data augmentation with measurement errors, as follows: In order to consider the impact of measurement error on fault location, the present application enhances the μPMU data with additive noise; more specifically, the present application adds Gaussian noise with a predefined mean and standard deviation to the μPMU data. Then, the augmented measurement value V aug,mea and φ aug,mea It can be determined as follows: V aug,mea =V true -V RME V true ,φ aug,mea =φ true -φ RME φ true Among them, V true and φ true Represent the true value of the voltage amplitude and phase angle, V RME and φ RME Denote the amplitude and phase angle errors, respectively. This application considers a typical μPMU accuracy model for synthesizing fault data, where independent Gaussian noise is considered as the measurement error source of the μPMU with an average value of zero, V RME and φ RME The standard deviations of are in the range of [0.005, 0.020] and [0.001, 0.025], respectively. For field fault data, this application considers a more practical μPMU accuracy model based on real data in

[30] . More precisely, the μPMU measurement error is given by V RME The mean is in [0.242, 0.382], the standard deviation is in [0.001, 0.008], and the skewness is in [0.100.0.280] due to Gaussian noise. PEThe mean is in [0.301, 0.421], the standard deviation is in [0.005, 0.011], and the skewness is in [0.038.0.108].

[0067] B. The implementation of GAT and TL coupling is as follows: The fault location method of distribution system combined with GAT and TL is implemented in MATLAB. The simulation is given to be run on a PC equipped with Intel Core i7-6700 CPU, 3.4GHz processor and 8GB RAM. Since this application regards one stage of the bus as a vertex, the total number of vertices is 230. For the vertex with μPMU, this application uses the measurement vector (ΔV, ΔΦ) as its attribute. The attribute of the intermediate vertex is (1,1), then GAT takes a 230×230 adjacency matrix a and a 230×2 attribute matrix M as input. This application performs hyperparameter tuning based on 25% of the data in domain s. This application uses three SG-Conv layers and two fully connected layers. The second dimension of each weight matrix W of all three SG-Conv layers is set to 2. The pooling ratio κ of the two pooling layers is 0.5. This application uses the ADAM optimizer with a learning rate of 0.001 to train the model with a mini-batch size of 32 and a weight decay of 0.001. The parameters δ and γ are set to 13.3 and 0.3 respectively. This application sets the training stop epoch to 300.

[0068] C. The penetration levels of μPMUs are as follows: In most actual power distribution systems, the number of μPMUs deployed is insufficient to achieve full observability of the system status. This application studies the impact of different levels of μPMU penetration on the performance of the method proposed in this application. This application considers that 15% to 35% of the total busbars have μPMUs, and adopts a greedy method to place μPMUs to obtain sufficient system observability for fault location. This application uses 80% of the data in domain S as training data and the remaining 20% ​​as test data. The fault location results of the method proposed in this application for μPMUs with different penetration levels are shown in Table 1, where the overall LAR is the fault location result without knowing the fault type. The LARs results show that the method can locate all four types of faults with high accuracy. The overall LARs value shows that the higher the penetration of μPMU, the easier the fault location. When the penetration level exceeds 20%, the overall LAR is greater than 95%. When the penetration level exceeds 30%, the change in the overall LARs is less than 3%. Therefore, even if only a few μPMUs are deployed, the method proposed in this application can accurately locate the fault.

[0069] As shown in Table 1, LARs are usually greater than 90% but less than 100%. On this basis, one-hop LAR is further considered to analyze the fault line profile of the wrong location. Specifically, if the predicted fault line segment is directly connected to the actual fault line segment, the fault is considered to be correctly located under the one-hop LAR metric. The one-hop LAR results of the method proposed in this application are shown in Table 2, indicating that when the μPMU penetration level exceeds 20%, all single-hop LARs exceed 98%. When the μPMU penetration level exceeds 30%, all faults can be located at least within one hop of the actual fault line segment. The reason why the fault location performance is significantly improved when the single-hop LAR is considered is the shorter length of the line segment in the distribution system. In the remaining case studies, this application considers that only 35% of the busbars are measured.

[0070] Table 1 μPMUs penetration level 15％ 20％ 25％ 30％ 35％ LAR under TP failure 0.914 0.958 0965 0.979 0.980 LAR under LG failure 0.952 0.979 0.987 0.991 0.990 LAR under DLG failure 0.919 0.962 0.983 0.987 0.988 LAR under LL fault 0.890 0.913 0.979 0.981 0.981 Comprehensive LAR 0.915 0.949 0.959 0.984 0.985 Table 2 μPMUs penetration level 15％ 20％ 25％ 30％ 35％ One-hop LAR under TP failure 0.974 0.988 0.993 1.000 1.000 One-hop LAR under LG failure 0.982 0.989 1.000 1.000 1.000 One-hop LAR under DLG failure 0.978 0.982 0.996 1.000 1.000 One-hop LAR under LL failure 0.969 0.983 0.994 1.000 1.000 Comprehensive one-hop LAR 0.976 0.986 0.996 1.000 1.000 D. Comparison and discussion with baseline methods are as follows: In order to evaluate the performance of the GAT and TL methods proposed in this application in fault location of distribution systems, this application implements four baseline methods: 1) CNN: A CNN-based fault location method that takes into account the μPMU measurements on the selected bus. The main difference between the CNN-based method and the method proposed in this application is whether spatial information is embedded. 2) Spectral GCN: A fault localization method based on spectral GCN. The basis of this baseline method is the eigendecomposition of the graph Laplacian matrix. Then, the graph filter operator is defined based on the Fourier transform of the graph. 3) DCTLN: The DCTLN-based fault diagnosis method solves the domain difference problem by minimizing the MMD of marginal distributions. 4) DTN: The DTN method extends marginal distribution adaptation to joint distribution adaptation, taking into account the sum of the differences between the marginal distribution and the conditional distribution.

[0071] The CNN and Spectral GCN baseline methods are only trained and tested on the S domain, without considering the impact of cross-domain differences. When implementing the DCTLN and DTN baseline methods, this application uses their domain adaptation modules instead of the domain adaptation method proposed in this application, focusing on domain adaptation.

[0072] First, the method of the present application is compared with the CNN and spectral GCN baseline methods, using 80% of the data from the S domain as training data and the remaining 20% ​​for testing. Since both the training data and the test data are in the S domain, there is no cross-domain generalization error. The LAR results of the method of the present application, the spectral GCN method, and the CNN baseline method are summarized in Table 3. The LARs of the method of the present application and the spectral GCN method are significantly higher than those of the CNN method. In addition, when considering a one-hop LAR, the fault location accuracy of the method of the present application and the spectral GCN method reaches 100%, while the accuracy of the CNN method is only 92.5%. When the number of hops increases, the accuracy of the CNN method improves. However, when considering a five-hop LAR, the LAR of the CNN method is still less than 98%. The results show that the performance of the distribution system fault location method can be significantly improved by embedding topology information.

[0073] Table 3 This method Spectral GCN CNN Comprehensive LAR 0.986 0.985 0.882 Comprehensive one-hop LAR 1.000 1.000 0.925 Comprehensive two-hop LAR 1.000 1.000 0.934 Comprehensive three-hop LAR 1.000 1.000 0.949 Comprehensive four-hop LAR 1.000 1.000 0.956 Comprehensive five-hop LAR 1.000 1.000 0.976 Then, this application examines the sensitivity of these three methods to the size of the training dataset. Fig.10 As shown, it is a schematic diagram comparing the overall positioning accuracy of different methods provided in Example 3 of the present application under different numbers of training samples. It can be seen that the method proposed in the present application is better than the CNN method. This is because the system topology is interfered by various factors, such as switch operation and connection / disconnection of DG, and the CNN method cannot effectively extract spatial information. On the other hand, the method proposed in the present application can effectively extract spatial information even with a very small training sample size. When the training sample size is small, this method is better than the spectral GCN method. This is because although the spectral GCN method can effectively extract spatial information, it cannot handle the fault location problem in topologies that have never been encountered in the training samples.

[0074] 1) Adaptability to changing operating conditions: This application further analyzes the impact of changing operating conditions on fault location performance by considering the following three scenarios: 1) The spectral GCN method and the proposed method are trained using 80% of the data from the S domain, and the remaining 20% ​​of the data are used to test them. 2) The two methods are trained using data from the S domain without considering the working condition I, where switches 18-135 are closed, switches 54-94 are open, and the DG of the 83 bus is connected to the system. Then, the present application tests them using data from the S domain obtained under the working condition I. 3) Scenario 3 is similar to Scenario 2, except that this application uses data from the S domain to train two methods without considering working condition II, that is, switches 18-135 are in the open state, switches 54-94 are in the closed state, and the DG on the 33 bus is connected to the system.

[0075] Scenario 2 and Scenario 3 are intended to simulate events encountered during actual testing under operating conditions not considered during training. The comparison results are shown in Table 4. It can be seen from Table 4 that when all possible system topologies are considered during training, both the spectral GCN method and the method proposed in this application are able to obtain higher LARs. However, when encountering new topologies, that is, topologies that were not considered during training, the performance of the spectral GCN method and the CNN method will drop significantly. When considering one-hop LAR, the method proposed in this application achieves a LAR greater than 98% in both scenario 1 and scenario 2. However, for both scenario 1 and scenario 2, the one-hop LARs of the spectral GCN and CNN methods are less than 85%. In addition, when using five-hop LAR, the spectral GCN and CNN methods still mislocate more than 8.5% and 16% of faults, respectively. The main reason is that the spectral GCN method is based on the Fourier basis of the graph, which means that any change in the topology will affect the feature basis. On the other hand, the GAT method of this application performs local convolution on each vertex, thereby maintaining the high versatility of the graph structure.

[0076] Table 4 This method Spectral GCN CNN Operation condition 1 0.985 / 1.000 0.984 / 1.000 0.882 / 0.925 Operation Condition 2 0.957 / 0.982 0.714 / 0.825 0.585 / 0.633 Operation Condition 3 0.961 / 0.988 0.735 / 0.911 0.603 / 0.647 2) Robustness to measurement errors: In order to evaluate the robustness of the method proposed in this application to measurement errors, this application trained and tested the CNN method, the spectral GCN method and the method proposed in this application on the augmented domain S data. The LAR results of the three methods are summarized in Table 5. Under measurement errors, the method proposed in this application can correctly locate all faults within a range of up to two hops from the actual fault segment. However, the two-hop LARs of the diaphragm GCN and CNN methods are only 97.1% and 83.4%, respectively. When considering four-hop LAR, the LAR of the spectral GCN method reaches 100%, while the five-hop LAR of the CNN method is still less than 99%.

[0077] To further evaluate the robustness of the GAT method proposed in this application, we train the three methods on S-domain data and then evaluate them on enhanced S-domain data. Then, during the training process, we gradually increase the ratio of augmented domain S data to domain S data, allowing all three methods to better understand the statistical measurement errors. Fig.11As shown, a comparative schematic diagram of the overall positioning accuracy without using enhanced data for training is provided in Example 3 of the present application. When no enhanced data is used for training, all three methods will result in a significant reduction in the overall LARs. The overall LARs of the spectral GCN and the method proposed in the present application are reduced by 20.71% and 20.93%, respectively, while the overall LAR of the CNN method is reduced by 30.27%. In the absence of enhancement, the spectral GCN and the method proposed in the present application are more robust to measurement errors than the CNN method. This is mainly because the spectral GCN and the method proposed in the present application retain the data structure and operate not only on the measurement data itself but also on the edges, which helps to mitigate the impact of measurement errors. All three methods have the potential to improve the overall LARs by using enhanced data during training. Since the method proposed in the present application uses significant coefficients to identify less reliable measurements, the overall LAR of the method proposed in the present application increases faster than the other two methods, enabling the present application to achieve a higher overall LAR. Although the spectral GCN method is more robust to measurement errors than the CNN method, its overall LAR is still lower than the method proposed in the present application. The main reason is that the robustness of the spectral GCN method to measurement errors depends on extensive training under all possible measurement error conditions. On the other hand, the method proposed in this paper addresses measurement errors by reducing the significance of large error measurements relative to their neighbors, allowing this paper to learn the accuracy of only μPMUs, which requires fewer samples.

[0078] Table 5 This method Spectral GCN CNN Comprehensive LAR 0.976 0.912 0.744 Comprehensive one-hop LAR 0.988 0.944 0.785 Comprehensive two-hop LAR 1.000 0.971 0.834 Comprehensive three-hop LAR 1.000 0.982 0.899 Comprehensive four-hop LAR 1.000 1.000 0.966 Comprehensive five-hop LAR 1.000 1.000 0.986 3) Performance under cross-domain generalization error: In order to evaluate the performance of the method proposed in this application under cross-domain generalization error, this application trained the DCTLN and DTN methods, as well as the performance of the method proposed in this application on augmented domain S data and 80% augmented domain T data. Then, this application tested them using the remaining 20% ​​augmented domain T data. By replacing the TL method with the DCTLN and DTN methods, this application evaluated their domain adaptation modules. The distribution differences of fault impedance values, DG states, switch states and load levels, the occurrence probabilities of various types of faults at different locations, and the differences in the measurement error distributions between domain S and T data explain the distribution differences between the two domains. As Fig.12As shown, a schematic diagram of the comparison of the overall positioning accuracy under training with different numbers of enhanced training samples provided in Example 3 of the present application is shown, in which it can be observed that when only the data applied to domain T is used, the data from domain S performs poorly. The LARs of all three methods increase proportionally with the number of domain T samples considered for TL. Because DCTLN only considers the marginal distribution, its overall LAR is the smallest. By combining the conditional distribution, the DTN method improves the overall LAR compared to the DCTLN method. When the marginal distribution and the conditional distribution are simply added together, the DTN method provides a moderate improvement. Since the method proposed in the present application enhances the consideration of minimizing the difference in joint distribution, when all the training data in T is used, the overall LAR is significantly improved by 10.90% and 15.93% compared to the DTN and DCTLN baseline methods, respectively.

[0079] Furthermore, the sensitivity of the overall transfer learning performance of the fault localization method proposed in this application to the parameter γ is examined. The parameter γ is used to weight the importance of learning domain-invariant features in transfer learning. Larger values ​​of γ lead to increased transferability of learned features at the expense of retaining fewer false positive identification features. Fig.13 As shown, it is a comparative schematic diagram of the overall positioning accuracy under different trade-off parameter training provided in Example 3 of the present application, which shows the overall LAR of the method proposed in the present application in the range of {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9}. The overall LAR increases first and then decreases with the change of γ value, showing a bell-shaped curve. It can be concluded that by selecting a suitable γ value, an appropriate trade-off can be achieved between fault localization performance and transferability.

[0080] This application proposes a new method for fault location in distribution systems based on GAT and TL. By embedding measurement data and spatial information in the attribute graph, GAT can locally extract fault discrimination features, significantly improving the adaptability of the method to system topology changes. In addition, a learnable significance coefficient based on the SA mechanism is proposed to determine the importance of the measurement to feature extraction based on the quality of the measurement data, which can improve the robustness of the method proposed in this application to measurement errors. In addition, a causal analysis of fault location based on GAT is performed, and a method for minimizing cross-domain generalization errors using separate priors and covariate shifts is proposed. This method makes it possible to use synthetic fault data in actual fault location. This application intends to extend this work to distribution systems with multiple faults in the future. In addition, this application also wants to study the impact of incorporating atmospheric data on fault location accuracy.

[0081] Corresponding to the distribution system fault location method in the above embodiment, Fig.14The structure block diagram of the power distribution system fault location device provided in the fourth embodiment of the present application is shown. For the convenience of description, only the part related to the embodiment of the present application is shown.

[0082] See also Fig.14 , the power distribution system fault location device comprises: The undirected graph construction module 141 is used to collect the real-time voltage and / or real-time current phase of the measurement point of each phase in the N measured buses in the power distribution system, and construct an attribute undirected graph according to the real-time voltage and / or real-time current phase of the measurement point, wherein the vertices of the attribute undirected graph are the measurement points, the attributes of the attribute undirected graph are the attribute vectors of the corresponding measurement points, the edges of the attribute undirected graph are the line segments connecting the two vertices, and N is an integer greater than zero; The feature vector updating module 142 is used to obtain a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information according to the attribute undirected graph, and update the vertex features according to the vertex attribute matrix and the vertex adjacency matrix to obtain the vertex feature vector of each vertex, wherein the initial vertex feature of the vertex is the attribute vector in the vertex attribute matrix; the feature vector extraction module 143 is used to perform downsampling clustering on the vertices in the attribute undirected graph according to the vertex feature vector to obtain a super vertex subgraph, and perform feature extraction on the super vertex subgraph using a readout operation to obtain a graph feature vector; The fault location module 144 is used to perform classification on the graph feature vector, determine the fault probability that any line section formed by the measurement points is a fault line section, and determine the fault location of the power distribution system according to the fault probability.

[0083] Optionally, the undirected graph construction module 141 includes: A first-class vertex classification unit is used to take measurement points with real-time voltage and real-time current phase as first-class vertices, and determine the attribute vectors of the first-class vertices as the real-time voltage and real-time current phase of the corresponding measurement points; A second-type vertex classification unit is used to take the measurement points with real-time voltage but without real-time current phase as second-type vertices, and determine the attribute vector of the second-type vertex as a preset vector; The attribute undirected graph construction unit is used to obtain the edges connecting the first type of vertices and the second type of vertices, and construct the attribute undirected graph according to the first type of vertices, the second type of vertices, the attribute vector of each vertex and the edges.

[0084] Optionally, the feature vector updating module 142 includes: A parameter acquisition unit, used to obtain a pre-trained parameterized weight matrix and a parameterized weight vector; An initial vector determination unit is used to determine, for any target vertex, the adjacent vertices corresponding to the target vertex according to the target vertex adjacency matrix, and to determine the vertex feature vectors of the target vertex and the adjacent vertices according to the vertex attribute matrix; A vector concatenation unit is used to concatenate vertex feature vectors of a target vertex and adjacent vertices according to a parameterized weight matrix to obtain a concatenation result, and multiply the transpose of the parameterized weight vector by the concatenation result to obtain a multiplication result; An activation unit, used to perform an activation operation on the multiplication result using a preset activation function to obtain a single activation result, traverse all vertices adjacent to the adjacent vertex, and determine that the sum of all the single activation results is a comprehensive activation result; A significance coefficient calculation unit, used for taking the ratio of a single activation result to a comprehensive activation result as the significance coefficient of the feature of the target vertex to the adjacent vertices, wherein the significance coefficient represents the significance of the feature of a vertex to its adjacent vertices; The vector update unit is used to obtain the significance coefficients of the features of all vertices adjacent to the target vertex and the corresponding vertex feature vectors to the target vertex, and obtain the updated vertex feature vector of the target vertex based on all the significance coefficients and vertex feature vectors combined with the parameterized weight matrix.

[0085] Optionally, the vector update unit comprises: A saliency calculation subunit, used for multiplying the saliency coefficient of the feature of each vertex adjacent to the target vertex, the vertex feature vector of the corresponding vertex and the parameterized weight matrix to obtain the product result of each vertex adjacent to the target vertex; The vector updating subunit is used to multiply the sum of all product results by the nonlinear parameter in the exponential linear unit to obtain the updated vertex feature vector of the target vertex.

[0086] Optionally, the feature vector extraction module 143 includes: A score calculation unit is used to obtain, for any vertex, the significance coefficient of the vertex to all vertices adjacent to the vertex, determine the sum of all significance coefficients as the importance score of the vertex, and traverse to obtain the importance scores of all vertices; A super vertex determination unit, for determining a super vertex from the vertices in the attribute undirected graph according to the importance scores of all vertices and the vertex feature vectors of the corresponding vertices, combined with a preset pooling ratio, and determining the adjacency matrix and vertex feature matrix of the super vertex as a super vertex subgraph; The feature extraction unit is used to extract features from the super-vertex subgraph using a readout operation to obtain a graph feature vector.

[0087] Optionally, the parameter acquisition unit includes: A field sample subunit is used to obtain an attribute undirected graph formed by test results obtained by performing a test experiment on the power distribution system and a probability that the label of the corresponding attribute undirected graph is a line section k as a field data sample; The first training subunit is used to train the first network based on the graph attention mechanism using the field data samples to obtain a pre-trained first network, and extract a pre-trained parameterized weight matrix and a parameterized weight vector from the pre-trained first network.

[0088] Optionally, the training process of the first network is as follows: For the attribute undirected graph in the field data sample, a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information are obtained. According to the vertex attribute matrix and the vertex adjacency matrix, combined with the parameterized weight matrix and the parameterized weight vector of the current iteration, the vertex features are updated to obtain the vertex feature vector of each vertex; According to the vertex feature vector, the vertices in the attribute undirected graph are downsampled and clustered to obtain a super vertex subgraph, and the super vertex subgraph is extracted using a readout operation to obtain a graph feature vector; Classification is performed on the graph feature vector to determine the fault probability that any line section formed by the measurement points is a fault line section. According to the fault probability and label, combined with minimizing the cross entropy loss function, the parameterized weight matrix, parameterized weight vector and classification parameters of the current iteration are trained until the iteration conditions are met to obtain the pre-trained first network.

[0089] Optionally, the parameter acquisition unit includes: A first sample subunit is used to use the obtained attribute undirected graph formed by the test results obtained by performing a test experiment on the power distribution system and the probability that the label of the corresponding attribute undirected graph is a line section k as a field data sample; A second sample subunit is used to use the acquired attribute undirected graph formed by synthetic data obtained by simulating the power distribution system and the probability that the label of the corresponding attribute undirected graph is line section k as a synthetic data sample; The second training subunit is used to train the second network integrating the graph attention mechanism and transfer learning using the field data samples and the synthetic data samples to obtain a pre-trained second network, and extract a pre-trained parameterized weight matrix and a parameterized weight vector from the pre-trained second network.

[0090] Optionally, the training process of the second network is as follows: For the attribute undirected graph in the field data sample, a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information are obtained. According to the vertex attribute matrix and the vertex adjacency matrix, combined with the parameterized weight matrix and the parameterized weight vector of the current iteration, the vertex features are updated to obtain the vertex feature vector of each vertex; According to the vertex feature vector, the vertices in the attribute undirected graph are downsampled and clustered to obtain a super vertex subgraph, and the super vertex subgraph is extracted using a readout operation to obtain a graph feature vector; Classification is performed on the graph feature vector to determine the fault probability that any line section formed by the measurement points is a fault line section. According to the fault probability and label, combined with the minimization of the cross entropy loss function, the parameterized weight matrix, parameterized weight vector and classification parameters of the current iteration are trained until the iteration conditions are met to obtain a pre-trained graph attention network.

[0091] It should be noted that the information interaction, execution process and other contents between the above-mentioned modules are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0092] The computer device of the present application includes: at least one processor, a memory, and a computer program stored in the memory and executable on at least one processor. When the processor executes the computer program, the steps in any of the above-mentioned power distribution system fault location method embodiments are implemented.

[0093] The computer device may include, but is not limited to, a processor and a memory.

[0094] The processor may be a CPU, or other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0095] The memory includes a readable storage medium, an internal memory, etc., wherein the internal memory may be the memory of a computer device, and the internal memory provides an environment for the operation of an operating system and computer-readable instructions in the readable storage medium. The readable storage medium may be a hard disk of a computer device, and in other embodiments may also be an external storage device of a computer device, for example, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on a computer device. Further, the memory may also include both an internal storage unit of a computer device and an external storage device. The memory is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as program codes of computer programs, etc. The memory may also be used to temporarily store data that has been output or is to be output.

[0096] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned device can refer to the corresponding process in the above-mentioned method embodiment, which will not be repeated here. If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned method embodiment when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device capable of carrying computer program code, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. For example, USB flash drive, mobile hard disk, magnetic disk or optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0097] The present application implements all or part of the processes in the above-mentioned embodiment method, and may also be completed through a computer program product. When the computer program product runs on a computer device, the computer device can implement the steps in the above-mentioned method embodiment when executing the computer program product.

[0098] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0099] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0100] In the embodiments provided in the present application, it should be understood that the disclosed devices / computer equipment and methods can be implemented in other ways. For example, the device / computer equipment embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0101] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0102] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for locating a fault in a power distribution system, characterized in that: The power distribution system fault location method comprises: Collect the real-time voltage and / or real-time current phase of the measurement point of each phase in the N measured buses in the power distribution system, and construct an attribute undirected graph according to the real-time voltage and / or real-time current phase of the measurement point, wherein the vertices of the attribute undirected graph are the measurement points, the attributes of the attribute undirected graph are the attribute vectors of the corresponding measurement points, the edges of the attribute undirected graph are the line segments connecting the two vertices, and N is an integer greater than zero; According to the attribute undirected graph, a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information are obtained, and vertex features are updated according to the vertex attribute matrix and the vertex adjacency matrix to obtain a vertex feature vector of each vertex, wherein the initial vertex feature of the vertex is the attribute vector in the vertex attribute matrix; According to the vertex feature vector, downsampling and clustering are performed on the vertices in the attribute undirected graph to obtain a super vertex subgraph, and feature extraction is performed on the super vertex subgraph using a readout operation to obtain a graph feature vector; Classification is performed on the graph feature vector to determine the fault probability that any line section formed by the measurement points is a fault line section, and the fault location of the power distribution system is determined based on the fault probability.

2. The method for locating a fault in a power distribution system according to claim 1, characterized in that: The constructing of an attribute undirected graph according to the real-time voltage and / or real-time current phase of the measurement point comprises: Taking the measurement points having the real-time voltage and the real-time current phase as first-type vertices, and determining the attribute vectors of the first-type vertices as the real-time voltage and the real-time current phase of the corresponding measurement points; Taking the measurement points having the real-time voltage but not having the real-time current phase as the second-type vertices, and determining the attribute vectors of the second-type vertices as the preset vectors; Obtain edges connecting the first-category vertices and the second-category vertices, and construct an attributed undirected graph based on the first-category vertices, the second-category vertices, the attribute vector of each vertex, and the edges.

3. The method for locating a fault in a power distribution system according to claim 1, characterized in that: The updating of vertex features according to the vertex attribute matrix and the vertex adjacency matrix to obtain a vertex feature vector of each vertex includes: Get the pre-trained parameterized weight matrix and parameterized weight vector; For any target vertex, determine the adjacent vertices corresponding to the target vertex according to the target vertex adjacency matrix, and determine the vertex feature vectors of the target vertex and the adjacent vertices according to the vertex attribute matrix; According to the parameterized weight matrix, concatenate the vertex feature vectors of the target vertex and the adjacent vertex to obtain a concatenation result, and multiply the transpose of the parameterized weight vector by the concatenation result to obtain a multiplication result; Use a preset activation function to perform an activation operation on the multiplication result to obtain a single activation result, traverse all vertices adjacent to the adjacent vertex, and determine that the sum of all single activation results is a comprehensive activation result; The ratio of the single activation result to the comprehensive activation result is used as the significance coefficient of the feature of the target vertex to the adjacent vertices, wherein the significance coefficient represents the significance of the feature of a vertex to its adjacent vertices; Obtain the significance coefficients of the features of all vertices adjacent to the target vertex and the corresponding vertex feature vectors for the target vertex, and obtain the updated vertex feature vector of the target vertex based on all the significance coefficients and vertex feature vectors in combination with the parameterized weight matrix.

4. The method for locating a fault in a power distribution system according to claim 3, characterized in that: The step of obtaining the updated vertex feature vector of the target vertex based on all the significance coefficients and vertex feature vectors in combination with the parameterized weight matrix includes: The significance coefficient of the feature of each vertex adjacent to the target vertex, the vertex feature vector of the corresponding vertex and the parameterized weight matrix are multiplied to obtain the product result of each vertex adjacent to the target vertex; the sum of all product results is multiplied by the nonlinear parameter in the exponential linear unit to obtain the updated vertex feature vector of the target vertex.

5. The method for locating a fault in a power distribution system according to claim 3, characterized in that: The method of downsampling and clustering the vertices in the attribute undirected graph according to the vertex feature vector to obtain a super vertex subgraph, and extracting features from the super vertex subgraph using a readout operation to obtain a graph feature vector includes: For any vertex, obtain the significance coefficient of the vertex to all vertices adjacent to the vertex, determine the sum of all significance coefficients as the importance score of the vertex, and traverse to obtain the importance scores of all vertices; According to the importance scores of all vertices and the vertex feature vectors of the corresponding vertices, combined with a preset pooling ratio, a super vertex is determined from the vertices in the attribute undirected graph, and the adjacency matrix and vertex feature matrix of the super vertex are determined as a super vertex subgraph; a readout operation is used to extract features from the super vertex subgraph to obtain a graph feature vector.

6. The method for locating a fault in a power distribution system according to claim 3, characterized in that: The obtaining of the pre-trained parameterized weight matrix and parameterized weight vector includes: An attribute undirected graph formed by test results obtained by performing a test on the power distribution system and a probability that the label corresponding to the attribute undirected graph is a line section k are obtained as field data samples; The first network based on the graph attention mechanism is trained using the field data samples to obtain a pre-trained first network, and a pre-trained parameterized weight matrix and a parameterized weight vector are extracted from the pre-trained first network.

7. The method for locating a fault in a power distribution system according to claim 6, characterized in that: The training process of the first network is as follows: For the attribute undirected graph in the field data sample, a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information are obtained, and vertex features are updated according to the vertex attribute matrix and the vertex adjacency matrix in combination with the parameterized weight matrix and the parameterized weight vector of the current iteration to obtain a vertex feature vector of each vertex; According to the vertex feature vector, downsampling and clustering are performed on the vertices in the attribute undirected graph to obtain a super vertex subgraph, and feature extraction is performed on the super vertex subgraph using a readout operation to obtain a graph feature vector; Classification is performed on the graph feature vector to determine the fault probability that any line section formed by the measurement points is a fault line section. According to the fault probability and the label, combined with minimizing the cross entropy loss function, the parameterized weight matrix, parameterized weight vector and classification parameters of the current iteration are trained until the iteration conditions are met to obtain a pre-trained first network.

8. The method for locating a fault in a power distribution system according to claim 3, characterized in that: The obtaining of the pre-trained parameterized weight matrix and parameterized weight vector includes: The obtained attribute undirected graph formed by the test results obtained by performing the test experiment on the power distribution system and the probability that the label corresponding to the attribute undirected graph is the line section k are used as the field data sample; Using the attribute undirected graph formed by the synthetic data obtained by simulating the power distribution system and the probability that the label corresponding to the attribute undirected graph is a line section k as a synthetic data sample; Using the field data samples and the synthetic data samples, a second network integrating graph attention mechanism and transfer learning is trained to obtain a pre-trained second network, and a pre-trained parameterized weight matrix and a parameterized weight vector are extracted from the pre-trained second network.

9. The method for locating a fault in a power distribution system according to claim 8, characterized in that: The training process of the second network is as follows: For the attribute undirected graph in the field data sample, a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information are obtained, and vertex features are updated according to the vertex attribute matrix and the vertex adjacency matrix in combination with the parameterized weight matrix and the parameterized weight vector of the current iteration to obtain a vertex feature vector of each vertex; According to the vertex feature vector, downsampling and clustering are performed on the vertices in the attribute undirected graph to obtain a super vertex subgraph, and feature extraction is performed on the super vertex subgraph using a readout operation to obtain a graph feature vector; Classification is performed on the graph feature vector to determine the fault probability that any line section formed by the measurement points is a fault line section. According to the fault probability and the label, combined with minimizing the cross entropy loss function, the parameterized weight matrix, parameterized weight vector and classification parameters of the current iteration are trained until the iteration conditions are met to obtain a pre-trained graph attention network.

10. A power distribution system fault location device, characterized in that: The power distribution system fault locating device comprises: An undirected graph construction module, used for collecting the real-time voltage and / or real-time current phase of the measurement point of each phase in the N measured buses in the power distribution system, and constructing an attribute undirected graph according to the real-time voltage and / or real-time current phase of the measurement point, wherein the vertices of the attribute undirected graph are the measurement points, the attributes of the attribute undirected graph are the attribute vectors of the corresponding measurement points, the edges of the attribute undirected graph are the line segments connecting the two vertices, and N is an integer greater than zero; A feature vector updating module is used to obtain a vertex attribute matrix representing the measured attribute vector and a vertex adjacency matrix representing the spatial information according to the attribute undirected graph, and update the vertex features according to the vertex attribute matrix and the vertex adjacency matrix to obtain a vertex feature vector of each vertex, wherein the initial vertex feature of the vertex is the attribute vector in the vertex attribute matrix; A feature vector extraction module, used to perform downsampling and clustering on the vertices in the attribute undirected graph according to the vertex feature vector to obtain a super vertex subgraph, and perform feature extraction on the super vertex subgraph using a readout operation to obtain a graph feature vector; The fault location module is used to classify the graph feature vector, determine the fault probability that any line section formed by the measurement points is a fault line section, and determine the fault location of the power distribution system according to the fault probability.

Images