Method for Locating Fault Sections in Distribution Networks and Analyzing Interpretability by Integrating Spatiotemporal Graph Information

Abstract

The present invention discloses a method for locating and interpretable analysis of distribution network fault sections by integrating spatio-temporal graph information, belonging to the technical field of distribution network fault diagnosis and artificial intelligence deep learning. The method includes the following steps: constructing a spatio-temporal graph information fault characterization structure that integrates time continuity and spatial integrity to depict the mapping relationship between spatio-temporal features and fault sections; extracting fault features from two dimensions of time feature capture and spatial feature perception, and establishing a distribution network fault location model that integrates spatio-temporal graph information; designing an interpretable analysis verification module to perform post hoc interpretable analysis on the model's decision basis and its internal working mechanism, and realizing the verification of fault location results. The present invention can achieve high-precision fault location tasks under complex operating scenarios such as new energy output fluctuations and topological changes, and analyze and depict the principle of the model itself and fault prediction, providing technical support for the application of deep learning-based fault location methods in actual systems.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault diagnosis and artificial intelligence deep learning, and particularly relates to a method for locating fault sections and interpretable analysis in a distribution network by integrating spatio-temporal graph information. Background Art

[0002] In complex operation scenarios such as new energy output fluctuations and frequent topological reconstructions, the safe and reliable operation of the distribution system faces new challenges. Therefore, quickly and accurately locating faults can provide effective guidance for maintenance personnel, significantly shorten the fault handling time, and is of great significance for improving the operation reliability of the system.

[0003] Traditional distribution network fault location methods mainly include impedance method, traveling wave method, and artificial intelligence method, etc. The impedance method aims to calculate the distance from the measurement point to the fault point through the ratio of the loop impedance to the unit impedance of the line, and it is one of the most easily implemented methods in on-site operations. However, when applying the impedance method, it is necessary to first perform an equivalent of the network topology. Once the topological structure changes, it is necessary to re-perform the equivalent calculation, which is difficult to adapt to in the case of frequent changes in the distribution network topology. The traveling wave method uses the propagation time of the fault transient traveling wave signal in the line and the traveling wave signal to achieve fault location, and is basically not affected by the transition resistance and the neutral point grounding method. However, the fault traveling wave is easily affected by the complex topological structure of the distribution network, and signal attenuation or distortion occurs during the transmission and reflection processes, resulting in a decrease in the fault location accuracy.

[0004] Currently, the artificial intelligence solutions applied to distribution network fault location mainly include artificial neural networks, convolutional neural networks, long short-term memory neural networks, and other composite models, etc. Although the improved solutions of the fault location models based on deep learning have achieved good application effects, there are still certain limitations. That is, the above solutions usually directly splice multiple time-series signals from different measurement units to form a longer two-dimensional signal, or construct it into a two-dimensional image and send it into the deep learning network, without considering the spatial information. As a result, the spatial characteristics of the interaction between devices are ignored, making the generalization performance of the existing solutions poor in the case of frequent changes in the network topology. At the same time, this type of method fails to perform interpretable analysis on the "black box problem" commonly existing in artificial intelligence models, and the lack of interpretability has become one of the main obstacles restricting the application of deep learning in safety-sensitive tasks such as distribution network fault diagnosis. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for locating fault sections and interpretable analysis in a distribution network by integrating spatio-temporal graph information, which solves the problems of certain limitations and lack of interpretability.

[0006] To achieve the above purpose, the present invention provides a method for locating fault sections and interpretable analysis in a distribution network by integrating spatio-temporal graph information, including the following steps:

[0007] S1. Construct a spatio-temporal graph information fault characterization structure, fuse fault feature information from the perspectives of time continuity and space integrity, and enhance the recognition ability of the classification model;

[0008] S2. Establish a distribution network fault location model that integrates spatio-temporal graph information;

[0009] S3. Design an interpretable analysis and verification module to perform interpretable analysis on the decision-making basis and internal working mechanism.

[0010] Preferably, in step S1, the time continuity is the electrical change quantity of the fault information mined from the continuous time series of the voltage and current collected by the measurement points;

[0011] The space integrity is the switch change quantity of the topological node link relationship during the operation of the distribution network.

[0012] Preferably, in step S1, the expression of the spatio-temporal graph information fault characterization structure constructed is:

[0013]

[0014] where p j is the measurement point numbered j in the distribution line; respectively represent the instantaneous values of the three-phase voltage and three-phase current collected by the measurement point at p j moment.

[0015] Preferably, in step S2, the fault location model includes a time series information capture module and a space feature perception module;

[0016] To capture the fault features contained in the time series, the time series information capture module selects the gated convolutional unit as the main structure of the time series information capture module, and realizes feature enhancement by merging the features on the continuous time segments. The model structure consists of a gated convolutional unit composed of two traditional one-dimensional convolutional layers, and its mathematical expression is:

[0017]

[0018] where: H l and H l+1 are the input and output of the Gated-CNN module in the l-th layer respectively; and are convolutional kernels of the same dimension; and are the biases of the model; is the convolutional operation; σ1(·) is the gating function, taking the Sigmoid activation function;

[0019] To better represent the spatial dependence relationship and propagation mode of fault points in the distribution network, the spatial feature perception module uses the topological structure of the distribution network based on the graph convolutional neural network to capture the fault impact by aggregating node features, and its mathematical expression is:

[0020]

[0021] Where: X l ∈R N×C is the input matrix of the l-th hidden layer, N is the number of measurement points, and C is the number of sampling points for each measurement point; σ2(·) is the ReLU activation function; represents the sum of the adjacency matrix A and the identity matrix; represents the degree matrix corresponding to the adjacency matrix A; W ∈ R C×Q , represents the weight matrix of the convolutional layer, and Q is the number of hidden layer nodes. Finally, the output of the l-th hidden layer f(A, X l ) can be obtained.

[0022] Preferably, usually, the distribution network reconstruction (such as the change in node connectivity caused by switch operation) occurs in a local area. Due to its natural property of aggregating local neighborhood information, GCN can adjust the information transmission path in a timely manner without a large-scale adjustment of the entire model, so it has a strong adaptability to the new topological structure.

[0023] Preferably, in step S3, the established interpretable analysis and verification module evaluates the importance of graph nodes by identifying the feature nodes concerned by the model during the decision-making process by combining graph data and gradient information. The specific steps are as follows:

[0024] S31. Class activation loss update

[0025] Assume that the number of measurement points in the topology is n, the number of sampling channels for each measurement point is m, the sampling length is l, and the fault location tasks for c sections of the distribution network need to be completed. Then the spatio-temporal graph information fault characterization structure is X = {X1, X2,..., X n}, X ∈ R n×m×l is used as the input in the interpretability stage, and its label is Y = {Y1, Y2,..., Y n}. Among them, X i represents the time series signal collected at the i-th measurement point, and Y i represents the fault section corresponding to this signal.

[0026] Construct a new STGCN network with the same structure as the trained STGCN network and retain the trained weight parameters. Define a new class activation loss function L CAM as shown in Equation (4):

[0027]

[0028] Among them, Z i ={Z1, Z2, Z3, …, Z n}, is the final classification output after the X i input model; the maximum value of the final classification output obtained through the argmax(·) function. Using the defined class activation loss function L CAM as the loss function of STGCN for model training, so that the prediction result Z i of each measurement point is as close as possible to Y i .

[0029] S32. Generation of spatial weight map

[0030] Input the spatio-temporal graph information fault characterization structure X i into a new STGCN network with L CAM as the loss function to calculate the weight contribution of the s-th measurement point to the final classification result as shown in Equation (5):

[0031]

[0032] Among them, d l+1 is the output dimension of the graph convolutional layer; y e refers to the probability that the classification output is the e-th section;

[0033] The physical meaning of is the importance of the i-th feature point of the s-th measurement point to the classification result. It can be calculated that when a fault occurs in the e-th section, the weight vector of each measurement point is as shown in Equation (6):

[0034]

[0035] According to the weight vector of each measurement point and the topological adjacency matrix A, judge the influence of the node measurement information on the positioning result during the model classification through the spatial weight size of each node. If topological reconstruction occurs, the connection relationship of the spatial weight map changes, and it can be retrained in the STGCN module and the classification result confusion matrix can be analyzed to verify the importance of the node to the fault of a specific section.

[0036] Therefore, the present invention adopts a method for power distribution network fault section location and interpretability analysis that integrates spatio-temporal graph information with the above structure, and has the following beneficial effects:

[0037] The present invention has good generalization ability under complex operation scenarios such as new energy output fluctuations, noise interference, and topological reconstruction, and can provide an effective solution for solving the problem of fault location in dynamically operating distribution networks. At the same time, the interpretable analysis module designed in the present invention can quantify the contribution degree of each measurement point information in the output result, and verify the rationality of fault section location by reconstructing the spatial weight map, providing a new research idea for the interpretability research in the field of distribution network fault diagnosis.

[0038] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a topological diagram of a 10kV distribution network model for simulation verification;

[0040] Figure 2 It is a schematic flow diagram of a fault section location scheme;

[0041] Figure 3 It is a process diagram of model training;

[0042] Figure 4 It is a spatial weight map when line L4B2 fails;

[0043] Figure 5 It is a spatial weight map when line L4B2 fails after the adjacency matrix is modified. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0045] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left", "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0046] Embodiment

[0047] The present invention provides a method for locating and interpretability analysis of distribution network fault sections by fusing spatio-temporal graph information, including the following steps:

[0048] S1. Construct a spatio-temporal graph information fault characterization structure;

[0049] The constructed spatio-temporal graph information fault characterization structure integrates fault feature information from the perspectives of time continuity and space integrity, enhancing the recognition ability of the classification model. Specifically as follows:

[0050] Time continuity mainly refers to mining the electrical change quantity of fault information based on the continuous time series of voltage and current collected by measurement points.

[0051] Space integrity mainly refers to the switch change quantity of the topological node link relationship during the operation of the distribution network.

[0052] The constructed spatio-temporal graph information fault characterization structure can be expressed by Equation (1):

[0053]

[0054] Among them, p j is the measurement point numbered j in the distribution line; respectively represent the instantaneous values of the three-phase voltage and three-phase current collected by the measurement point at time p j

[0055] S2. Establish a distribution network fault location model integrating spatio-temporal graph information;

[0056] The established fault location model mainly includes a time series information capture module and a space feature perception module. The specific module descriptions are as follows:

[0057] Time series information capture module:

[0058] To capture the fault features contained in the time series, a gated convolutional unit is selected as the main structure of the time series information capture module, and feature enhancement is achieved by merging the features on continuous time segments. The model structure consists of a gated convolutional unit composed of two traditional one-dimensional convolutional layers, and its mathematical expression is as Equation (2):

[0059]

[0060] Among them: H l and H l+1 are the input and output of the Gated-CNN module at the l-th layer respectively; and are convolutional kernels of the same dimension; and are the biases of the model; is the convolutional operation; σ1(·) is the gating function, taking the Sigmoid activation function.

[0061] Space feature perception module ​

[0062] To better represent the spatial dependence relationship and propagation mode of fault points in the distribution network, based on the graph convolutional neural network, the topological structure of the distribution network is used to capture the fault impact by aggregating node features. The mathematical expression of the model is as shown in (3):

[0063]

[0064] Where: X l ∈R N×C is the input matrix of the l-th hidden layer, N is the number of measurement points, and C is the number of sampling points for each measurement point; σ2(·) is the ReLU activation function; represents the sum of the adjacency matrix A and the identity matrix, that is, the node can obtain the comprehensive information of itself and surrounding nodes; represents the degree matrix corresponding to the adjacency matrix A; W ∈ R C×Q , representing the weight matrix of the convolutional layer, and Q is the number of hidden layer nodes. Finally, the output of the l-th hidden layer f(A, X l ) can be obtained.

[0065] Usually, the reconstruction of the distribution network (such as the change in node connectivity caused by switch operation) occurs in a local area. Due to its natural property of aggregating local neighborhood information, GCN can adjust the information transmission path in a timely manner without a large-scale adjustment of the entire model. Therefore, it has a strong adaptability to the new topological structure.

[0066] S3. Design an interpretable analysis and verification module to conduct interpretable analysis on the decision basis and internal working mechanism.

[0067] The interpretable analysis and verification module, by combining graph data and gradient information, identifies the characteristic nodes that the model pays attention to during the decision-making process, so as to evaluate the importance of graph nodes. The specific steps are as follows:

[0068] S31. Class activation loss update

[0069] Assume that the number of measurement points in the topology is n, the number of sampling channels for each measurement point is m, the sampling length is l, and the fault location tasks for c sections of the distribution network need to be completed. Then the spatio-temporal graph information fault characterization structure is X = {X1, X2, …, X n}, X ∈ R n×m×l as the input in the interpretability stage, and its label is Y = {Y1, Y2, …, Y n}. Among them, X i represents the time-series signal collected at the i-th measurement point, and Y i represents the fault section corresponding to this signal.

[0070] Further, construct a new STGCN network with the same structure as the trained STGCN network and retain the trained weight parameters, and define a new class activation loss function L CAM As shown in Equation (4):

[0071]

[0072] where Z i ={Z1, Z2, Z3, …, Z n}, is the final classification output after X i is input into the model; the maximum value of the final classification output obtained through the argmax(·) function. Using the defined class activation loss function L CAM as the loss function of STGCN for model training, making the prediction result Z i of each measurement point as close as possible to Y i .

[0073] S32. Generation of spatial weight map

[0074] Input the spatio-temporal graph information fault characterization structure X i into the new STGCN network with L CAM as the loss function, and calculate the weight contribution of the s-th measurement point to the final classification result As shown in Equation (5):

[0075]

[0076] where d l+1 is the output dimension of the graph convolutional layer; y e refers to the probability that the classification output is the e-th section; physically means the importance of the i-th feature point of the s-th measurement point to the classification result. It can be calculated that when a fault occurs in the e-th section, the weight vector of each measurement point As shown in Equation (6):

[0077]

[0078] According to the weight vector of each measurement point and the topological adjacency matrix A, judge the influence of the node measurement information on the positioning result during the model classification through the spatial weight size of each node. If topological reconstruction occurs, the connection relationship of the spatial weight map changes, and it can be retrained in the STGCN module and the classification result confusion matrix can be analyzed to verify the importance of the node to the fault of a specific section.

[0079] Build in MATLAB / Simulink as Figure 1The shown 10 kV distribution network simulation model includes a power source, transformers, lines, loads, and distributed power sources. Among them, the main transformer adopts the Dyn11 connection method, with a transformation ratio of 110 kV / 10.5 kV. Its neutral point is grounded through an arc suppression coil, and the compensation degree is 8%; the lengths and types of lines L1B1, L1B2, …, L6B3 have been marked in the figure, and the line parameters are shown in Table 1; the connection of branches is changed by the opening and closing states of the tie switches S1~S6 to achieve topological reconstruction. Under normal circumstances, only switches S1~S3 are in the closed state; the distributed power sources DG1~DG3 are respectively connected to buses 1~3. It is a μPMU device with a sampling frequency of 10 kHz, which can collect high-precision voltage / current waveform data with time-stamp alignment. The data within a 60 ms time window before and after the start-up criterion action is uniformly intercepted, including 1 cycle before the fault and 2 cycles after the fault.

[0080] Table 1 Line Parameters

[0081]

[0082] According to the foregoing simulation settings, different fault types, transition resistances, fault initial phase angles, load levels, fault lines, and relative positions are set to simulate fault scenarios. The simulation sample parameters of specific fault cases are shown in Table 2, and a total of 1620 fault samples are generated. In addition to the above fault samples, normal operation samples are obtained by changing different load levels. Starting from the moment when the system is completely stable under each normal operation condition, with a time window length of 60 ms, the window slides backward 100 times at intervals of 10 measurement points in sequence. The specific normal operation sample parameters are shown in Table 3, and a total of 300 normal operation samples are generated. After randomly shuffling the above 1920 samples, they are randomly divided into a training set and a test set according to a ratio of 4:1, and the F1 score is used to evaluate the classification effect of the fault location model. Its calculation formula is:

[0083]

[0084] Among them: True Positive (TP) represents a sample that is actually positive and is predicted to be positive; False Positive (FP) represents a sample that is actually negative but is predicted to be positive; False Negative (FN) represents a sample that is actually positive but is predicted to be negative. The F1 score can take into account the classification model's Precision (P) and Recall (R), and is used to comprehensively evaluate the model's location accuracy. It can be considered that the larger the value of F1, the better the classification performance of the model.

[0085] Table 2 Line Parameters

[0086]

[0087]

[0088] The present invention is specifically implemented through the following technical solutions: A method for locating fault sections and interpretable analysis of a distribution network by integrating spatio-temporal graph information, comprising the following steps:

[0089] 1. Acquisition and preprocessing of raw data

[0090] For the above voltage or current time series Y = {y1, y2, …, y k}(y k being the sampling value at the k-th point), perform data normalization operation, and scale it to the interval [-1, 1], as shown in Equation (8):

[0091]

[0092] where: y max is the maximum value in the time series Y; y min is the minimum value in the series; is the value after data normalization, then the original time series is transformed into

[0093] At the same time, for the Figure 1 shown distribution network, an adjacency matrix A1 can be constructed according to the spatial relationship of 5 measurement points to describe the connection relationship between each point in the topological graph. When switch S1 changes from closed to open and S2 changes from open to closed, the adjacency matrix changes to A2, as shown in Equation (9):

[0094]

[0095] 2. Model training and performance optimization

[0096] After preprocessing the raw data and generating the adjacency matrix, use the processed data set to offline train a deep learning model with better performance. In the training stage, divide the data set into a training set and a test set according to a certain ratio. Among them, the training set is used to train the network parameters of the model, and the hyperparameters are adjusted according to the classification effect of the model to select the optimal network architecture and model performance. During the training process, cross-entropy is used as the loss function to quantify the difference between the model output result and the true label; the Adam optimizer is used to calculate the minimization of the loss function, optimize the weights and biases of each layer of the model, and check the convergence of the model. In this paper, the initial learning rate is taken as 10 -5 , the attenuation coefficient is taken as 0.5, and the attenuation period is 3. Finally, save the fault location model that has been trained and has good performance for subsequent applications.

[0097] 3. Online application and interpretable analysis

[0098] During online operation, each measurement point collects three-phase voltage and current signals of each line. At the same time, to avoid repeated detection and waste of computing resources, the sudden change of phase current and zero-sequence overcurrent are used as the starting criteria in this paper, as shown in Equations (10) and (11).

[0099]

[0100] In Equation (8), is the effective value of the phase current channel at any time t k , N is the interval period; I n1 is the maximum value of the sudden change of each phase current during normal operation. Equation (9) is set to prevent the insufficient sensitivity of the sudden change of phase current starting element in the case of high-resistance grounding faults, are the instantaneous values of the three-phase currents at any time t k respectively; I n2 is the reference current, which is taken as 1A here; k set1 and k set2 are the starting thresholds, which are taken as 0.2 and 0.1 in this paper respectively.

[0101] If the current signal of a certain measurement point meets the requirements of the above starting criteria, then based on the starting time of the criteria, the voltages and currents collected by each measurement point are collected in the form of Equation (1) and data preprocessing is carried out, and the fault location model saved by offline training is called for judgment, and the fault diagnosis result can be obtained. In the interpretability analysis stage, the interpretability analysis module is used to analyze the classification mechanism of the fault location model. By calculating the attribution value of the input spatio-temporal features, the influence degree of the input information of a single measurement point on the output result is obtained; combined with the spatial weight map, the correlation between the measurement information and the output category is analyzed from a global perspective, and the decision-making basis of the model for different spatial feature distributions is intuitively explained, so as to understand the internal working mechanism of the model.

[0102] The flow chart of the proposed method for fault section location and interpretability analysis of distribution network integrating spatio-temporal graph information is as Figure 2 shown.

[0103] By Figure 3 showing the performance of the model on the training set and the test set during the training process, it can be seen that at the 160th iteration, the model begins to converge and tend to be stable, and the accuracy rate on the test set can reach 92.9%, indicating that the proposed scheme in this paper has extremely strong fault location ability and is not affected by fault types, transition resistances, and fault initial phase angles.

[0104] Taking the fault of line L4B2 as an example, by selecting the data under a certain operating condition and inputting it into the interpretability analysis module, the result can be obtained as Figure 4The spatial weight map shown. It can be seen that relatively large weights are obtained at measurement points P4, P11, and P12, that is, the data characteristics of these three measurement points have made relatively large contributions to judging the fault of this line, and the data contribution at P11 is the largest. Analyzing the above results, it can be known that the positions of the three measurement points are relatively close to the corresponding faulty line, and the fault propagation path is relatively shorter than that of other measurement points. Therefore, the most obvious fault characteristics can be obtained, and thus the decision-making contribution to the faults at the corresponding positions of the model is relatively large. When a fault occurs in other lines under a certain operating condition, the weight values of the measurement points are shown in Table 3.

[0105] Table 3 Weight values of measurement points when faults occur in each line

[0106]

[0107]

[0108] When topology reconstruction occurs (i.e., S2 is disconnected and S5 is closed), the connection relationship between measurement points P 11 and P 12 is cut off, the input adjacency matrix is modified, and the spatio-temporal graph structure data is input into the STGCN module for testing. Except for the situation where 3 additional samples of line L4B2 faults (true label is 9) are misjudged as line L4B3 faults (predicted result is 10), the judgment results of the remaining lines remain unchanged.

[0109] Further analyze the internal principle of the above decision made by the model, and export the spatial weight map corresponding to the samples of line L4B2 faults that are correctly judged by the model as shown in Figure 5 where the connection relationship between measurement points P 11 and P 12 has been cut off. It can be seen that relatively large weights can still be obtained at measurement points P4, P 11 and P 12 , and the weight values at other measurement points are 0. At this time, the range of the faulty section output by the model can be narrowed down to between the three measurement points; at the same time, the weight at P 11 is still the largest among the three, indicating that the occurrence of topology reconstruction does not affect the model to give a correct judgment. Through the above steps, the contribution degree of the information of each measurement point in the output result can be quantified, providing an important basis for improving the reliability of the output result of the fault location model.

[0110] Therefore, the present invention adopts the above method for locating and interpretable analysis of distribution network fault sections by integrating spatio-temporal graph information. First, the feasibility of spatio-temporal graph information in the distribution network fault location scheme is analyzed, and a spatio-temporal graph information fault characterization structure that integrates time continuity and spatial integrity is constructed; then, the spatio-temporal graph convolutional network is used to fully extract the temporal waveform information and spatial structure features during the distribution network fault, so as to achieve high-precision fault location in scenarios with frequent topological transformations; finally, an interpretable analysis module of the spatio-temporal graph convolutional network is constructed to conduct post-hoc interpretable analysis on the model decision basis and internal working mechanism. Through the above process, a closed-loop location scheme of "data-modeling-decision-verification" based on machine learning is formed, making the location method have the advantages of high location accuracy and strong robustness, and maintaining good generalization ability in complex operation scenarios such as new energy output fluctuations and topological changes, and having strong application value.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for locating distribution network fault sections and interpretable analysis by integrating spatio-temporal graph information, characterized in that, It includes the following steps: S1. Construct a spatio-temporal graph information fault characterization structure, fuse fault feature information from the perspectives of time continuity and space integrity, and enhance the recognition ability of the classification model; In step S1, the expression of the constructed spatio-temporal graph information fault characterization structure is: (1)； Among them, is the measurement point numbered in the distribution line; , respectively represent the instantaneous values of the three-phase voltage and three-phase current collected by the measurement point at time ; S2. Establish a distribution network fault location model that integrates spatio-temporal graph information; S3. Design an interpretable analysis and verification module to conduct interpretable analysis on the decision basis and internal working mechanism; In step S3, the interpretable analysis and verification module combines graph data and gradient information to identify the feature nodes concerned by the model during the decision-making process, and evaluate the importance of graph nodes. The specific steps are as follows: S31. Class activation loss update Assume that the number of measurement points in the topology is , the number of sampling channels for each measurement point is , the sampling length is , and it is necessary to complete the fault location tasks for sections of the distribution network. Then the fault characterization structure of the spatio-temporal graph information is , As the input of the interpretability stage, its label is ; Among them, represents the time-series signal collected by the th measurement point, and represents the fault section corresponding to this signal; Construct a new STGCN network with the same structure as the trained STGCN network and retain the trained weight parameters, and define a new class activation loss function The expression is as follows: (4)； Among them, is the prediction result, , which is the final classification output after inputting into the model; the maximum value of the final classification output obtained through the function; Activate the loss function with the defined class As the loss function of STGCN for model training, the prediction results of each measurement point as close as possible to ; S32. Spatial weight graph generation The spatio-temporal diagram information fault characterization structure is input into a new STGCN network with as the loss function to calculate the weight contribution of the (5)； Among them, is the output dimension of the graph convolutional layer; refers to the probability that the classification output is the th section; The physical meaning of is the importance of the th feature point of the th measurement point to the classification result; When calculating the weight vectors of each measurement point when the th section fails, the expression is as follows: (6)； According to the weight vectors of each measurement point and the topological adjacency matrix , the influence of the node measurement information on the positioning result during the model classification process is judged by the spatial weight magnitude of each node.

2. A method for locating a distribution network fault section and interpretable analysis by integrating spatio-temporal graph information according to claim 1, characterized in that: In step S1, time continuity is the electrical change amount of fault information mined from the continuous time series of voltage and current collected by the measurement points; Spatial integrity is the switch change amount of the topological node link relationship during the operation of the distribution network.

3. A method for locating a distribution network fault section and interpretable analysis by integrating spatio-temporal graph information according to claim 1, characterized in that: In step S2, the fault location model includes a time series information capture module and a spatial feature perception module; The time series information capture module selects a gated convolutional unit as the main structure of the time series information capture module, and its mathematical expression is: (2)； Wherein: and are the input and output of the Gated-CNN module at the th layer respectively; and are convolutional kernels of the same dimension; and are the biases of the model; is the convolutional operation; is the gating function, taking the Sigmoid activation function; The spatial feature perception module is based on a graph convolutional neural network to capture the fault impact by aggregating node features through the distribution network topology structure, and its mathematical expression is: (3)； Wherein: is the input matrix of the th hidden layer, is the number of measurement points, is the number of sampling points for each measurement point; is the ReLU activation function; , representing the sum of the adjacency matrix and the identity matrix; , representing the degree matrix corresponding to the adjacency matrix ; , representing the weight matrix of the convolutional layer, is the number of hidden layer nodes, and finally the output of the th hidden layer can be obtained.

Images