A detection method and device for short-circuit faults in a distribution network

Abstract

The present invention discloses a method and device for detecting short-circuit faults in a distribution network, which relates to the field of power operation and maintenance technology. The detection method includes: obtaining the topological structure of the target distribution network and the voltage and current data of each node during the fault, constructing graph structure data and adding fault labels; constructing a fault detection model including an LSTM network and a multi-layer GCN network, and introducing an adaptive frequency aggregation mechanism in each layer of the GCN network; training the fault detection model using graph structure data with fault labels, and jointly optimizing the fault detection model using a physical information constraint loss function and a classification loss function; and performing fault detection on the distribution network to be detected using the trained fault detection model. The present invention can accurately identify the type and location of faults through the joint modeling of temporal and spatial features, effectively shorten the fault response time, and improve the safety and stability of power grid operation.

Description

Technical Field

[0001] The present invention relates to the technical field of power operation and maintenance, and in particular to a method and device for detecting short-circuit faults in a distribution network. Background Art

[0002] As a core component of the power system, the distribution network undertakes the critical task of transmitting electricity from substations to end users. Due to its wide coverage, diverse equipment types, and complex operating environment, its reliability and stability directly impact the safety and economic efficiency of power supply. However, during distribution network operation, various short-circuit faults (such as single-phase ground faults, two-phase interphase shorts, and three-phase interphase shorts) can cause voltage sags, abnormal currents, and even power outages, severely impacting power quality and user experience. Therefore, accurately and rapidly detecting and locating distribution network faults is crucial for ensuring power system safety and reducing operation and maintenance costs.

[0003] Currently, distribution network fault detection technologies primarily include methods based on physical models, signal processing, and machine learning. Physical model-based fault detection methods include impedance ranging, traveling wave methods, and fault component methods. Impedance ranging and fault component methods require accurate line impedance and system equivalent parameters. However, these parameters vary widely in real power grids, making them difficult to accurately obtain. Methods like the traveling wave method require high-precision synchronous sampling equipment, resulting in high computational complexity and difficulty in achieving real-time detection. Signal processing-based methods include Fourier transform (FFT), wavelet transform (WT), and empirical mode decomposition (EMD). Limitations include the sensitivity of wavelet transform and empirical mode decomposition to noise, which can easily lead to misjudgments. Complex signal processing algorithms require high computational resources, making them unsuitable for real-time applications. Their applicability is limited, and their versatility is poor. Among machine learning-based methods, decision trees, support vector machines (SVMs), and random forests have been introduced to the field of distribution network fault detection. These methods build classification models based on historical fault data for fault identification. However, their reliance on feature engineering and inadequate processing of time series features make them difficult to adapt to dynamic environments. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method and device for detecting short-circuit faults in distribution networks, so as to solve the technical problems that traditional distribution network fault detection technology has many deficiencies in the face of complex grid structures, time series feature extraction, and noise interference, and to improve the accuracy and real-time performance of fault detection.

[0005] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0006] In a first aspect, the present invention provides a method for detecting a short-circuit fault in a distribution network, comprising:

[0007] Obtain the topology of the target distribution network and the voltage and current data of each node during the fault period, construct graph structure data and add fault labels;

[0008] Build a fault detection model that includes an LSTM network and a multi-layer GCN network, and introduce an adaptive frequency aggregation mechanism in each GCN layer;

[0009] The fault detection model is trained using graph structure data with fault labels, and the fault detection model is jointly optimized using a physical information constraint loss function and a classification loss function;

[0010] The trained fault detection model is used to perform fault detection on the distribution network to be tested.

[0011] Optionally, constructing labeled graph structure data includes:

[0012] Construct an adjacency matrix based on the topological structure of the target distribution network , the adjacency matrix The row and column numbers correspond to the nodes, and the adjacency matrix of OK Listed as :

[0013]

[0014] A node characteristic matrix is constructed according to the voltage and current data of each node during the fault period, including:

[0015] Preprocess the acquired voltage and current data, and construct a node feature matrix based on the preprocessed voltage and current data , node feature matrix The row and column numbers correspond to the sampling time and node, and the node feature matrix of OK Listed as :

[0016]

[0017] Where, For nodes At the sampling time The pre-processed voltage and current data;

[0018] According to the adjacency matrix and the node feature matrix Building graph-structured data .

[0019] Optionally, the preprocessing includes data cleaning and normalization.

[0020] Optionally, the training of the fault detection model using graph structure data with fault labels and the joint optimization of the fault detection model using a physical information constraint loss function and a classification loss function include:

[0021] Initialize the model parameters of the fault detection model;

[0022] The graph structure data with fault labels is input into the fault detection model as input data. The node feature matrix in the graph structure data is input into the LSTM network to extract time series features. The time series features are input into the multi-layer GCN network. The adjacency matrix in the graph structure data is used to realize inter-node feature propagation to extract time series features with spatial features, and then mapped to obtain the fault classification probability of each node.

[0023] The total loss is obtained by calculating the physical information constraint loss function and the classification loss function based on the fault classification probability and fault label of each node;

[0024] Based on the total loss, the Adam optimizer is used to update the model parameters, and the gradient clipping technique is combined to limit the gradient norm;

[0025] Repeat the forward propagation and back propagation process until the total loss reaches a preset threshold or the number of training times reaches a set value, completing the training of the fault detection model.

[0026] Optionally, introducing an adaptive frequency aggregation mechanism in each layer of the GCN network includes:

[0027] Adaptive frequency aggregation is used to decompose the low-frequency and high-frequency components in the GCN network, and the ratio of low-frequency and high-frequency components is dynamically adjusted through learnable weights to enhance the adaptability of the GCN network to different node features. The expression of adaptive frequency aggregation is:

[0028]

[0029] Where, Output features for the GCN network, Input features and characteristics for GCN network The difference, are learnable weights, The features obtained by adaptive frequency aggregation are used as the input features of the next layer of GCN network.

[0030] Optionally, the physical information constraint loss function includes:

[0031] Kirchhoff's current law losses :

[0032]

[0033] Where, is the total number of nodes, For nodes The set of neighbor nodes of For nodes Flow Node The current value, For nodes Flow Node The current value;

[0034] Kirchhoff's voltage law losses :

[0035]

[0036] Where, For nodes voltage value.

[0037] Optionally, the classification loss function is:

[0038]

[0039] Where, is the classification loss, is the total number of categories, For category The weight of For nodes The fault label, For nodes Predicted as category probability;

[0040]

[0041] Where, is the total number of nodes, For category The total number of nodes, is the smoothing factor.

[0042] In a second aspect, the present invention provides a device for detecting a short-circuit fault in a distribution network, comprising:

[0043] A data processing module is configured to obtain the topology of the target distribution network and the voltage and current data of each node during the fault period, construct graph structure data and add fault labels;

[0044] The model building module is configured to build a fault detection model that includes an LSTM network and a multi-layer GCN network, and introduce an adaptive frequency aggregation mechanism in each layer of the GCN network;

[0045] a model training module configured to train the fault detection model using graph structure data with fault labels, and to jointly optimize the fault detection model using a physical information constraint loss function and a classification loss function;

[0046] The model deployment module is configured to perform fault detection on the distribution network to be detected using the trained fault detection model.

[0047] In a third aspect, the present invention provides an electronic device, including a processor and a storage medium;

[0048] The storage medium is used to store instructions;

[0049] The processor is configured to operate according to the instructions to execute the steps of the above method.

[0050] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The present invention provides a method and device for detecting short-circuit faults in distribution networks. The method comprises: 1) constructing a fault detection model comprising an LSTM network and a multi-layer GCN network, and introducing an adaptive frequency aggregation mechanism into each GCN layer. This combines time series modeling and adaptive frequency aggregation, effectively improving the accuracy and reliability of fault detection. 2) Training the fault detection model using graph-structured data with fault labels and jointly optimizing it using a physical information constraint loss function and a classification loss function. This introduces PINN physical constraints during training, ensuring that the model adheres to the fundamental principles of power systems while performing data-driven learning, thereby improving generalization and physical interpretability. Compared to traditional fault detection methods based on rules or single machine learning models, the present invention simultaneously considers the temporal evolution trends and spatial topological information of the distribution network. Using AFA and PINN, the method enhances sensitivity to abnormal faults and physical consistency, resulting in a model with strong robustness even with small amounts of data or unseen fault patterns. This makes the model particularly suitable for fault detection and location in actual distribution network environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 1 is a flow chart of a method for detecting a short-circuit fault in a distribution network provided by an embodiment of the present invention;

[0054] Figure 2 This is a simulation wiring diagram of a three-microgrid structure distribution network provided by an embodiment of the present invention;

[0055] Figure 3This is a graph showing the loss and accuracy during training provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0057] Example 1:

[0058] like Figure 1 As shown, an embodiment of the present invention provides a method for detecting a short-circuit fault in a distribution network, comprising the following steps:

[0059] Step S1: Obtain the topology of the target distribution network and the voltage and current data of each node during the fault period, construct graph structure data and add fault labels.

[0060] Specifically in this embodiment, the target distribution network selects a three-microgrid structure distribution network, and uses the simulation software MATLAB / SIMULINK to align and simulate. The specific simulation wiring diagram is as follows: Figure 2 As shown, node 0 is the main bus node, nodes 1, 2, and 3 are public load line nodes 3, 4, and 5, node 4 is bus 1 node, node 5 is load bus 1 node, node 6 is A load line node, nodes 7, 8, and 9 are photovoltaic A1, A2, and A3 line nodes, node 10 is battery line 1 node, node 11 is bus 2 node, node 12 is load bus 2 node, node 13 is B load line node, nodes 14, 15, and 16 are photovoltaic B1, B2, and B3 line nodes, node 17 is bus 3 node, node 18 is load bus 3 node, node 19 is C load line node, nodes 20, 21, 22, 23, and 24 are photovoltaic automation line 1, 2, 3, 4, and 5 nodes, nodes 25 and 26 are wind turbine grid-connected line 1 and 2 nodes, node 27 is battery line 2 node, and node 28 is a load line node.

[0061] In the simulation software, faulty nodes were manually set, and two random faults were applied to each of the 29 nodes in the simulation model. Ten fault types were set: AG, BG, CG, ABG, ACG, BCG, AB, AC, BC, and ABC, representing a ground fault on each of the three phases A, B, and C; a ground fault on each of the two phases AB, AC, and BC; a short circuit fault between the two phases AB, AC, and BC; and a short circuit fault between the three phases ABC. The simulation time was set to 1 second, and a fault occurred between 0.5 and 0.6 seconds. During the fault period, three-phase voltage and current data were extracted from each node at a sampling frequency of 500 Hz. A total of 2900 data sets were obtained, with a 27:2 ratio of faulty node data to non-faulty node data. The simulation was repeated ten times.

[0062] Specifically in this embodiment, constructing labeled graph structure data includes:

[0063] Construct an adjacency matrix based on the topological structure of the target distribution network , the adjacency matrix The row and column numbers correspond to the nodes, and the adjacency matrix of OK Listed as :

[0064]

[0065] The node characteristic matrix is constructed based on the voltage and current data of each node during the fault period, including:

[0066] Preprocess the acquired voltage and current data, and construct the node feature matrix based on the preprocessed voltage and current data , node feature matrix The row and column numbers correspond to the sampling time and node, and the node feature matrix of OK Listed as :

[0067]

[0068] Where, For nodes At the sampling time The pre-processed voltage and current data;

[0069] According to the adjacency matrix and node feature matrix Building graph-structured data .

[0070] Among them, preprocessing includes data cleaning and normalization.

[0071] Data cleaning, such as removing noise and outliers and filling in missing values, can provide the model with high-quality and reliable input data, thereby improving training efficiency and prediction accuracy.

[0072] The voltage and current data collected from each node are extremely unbalanced in terms of fault and non-fault conditions. At the same time, there are differences in the dimensions of current and voltage, and there is a large gap in amplitude. The difference in dimensions and amplitude will have a significant impact on model training, resulting in a decrease in model training results. Therefore, the data needs to be normalized.

[0073] Step S2: Build a fault detection model that includes an LSTM network and a multi-layer GCN network, and introduce an adaptive frequency aggregation mechanism in each layer of the GCN network.

[0074] The LSTM (Long Short-Term Memory) network is a recurrent neural network with a special structure. Compared with traditional BP neural networks, the training parameters of LSTM network neurons have added long-term and short-term memory data, which can better capture information about time series changes.

[0075] GCNs (Graph Convolution Networks) are specialized convolutional neural network structures used for graph data processing. Similar to how convolutional neural networks process image data, GCNs extract features from input graph data through convolution operations, which are used for tasks such as node classification, graph classification, and link prediction. GCNs primarily consist of an input layer and a graph convolution layer. The graph convolution layer is a core component that updates the representation of the current node by aggregating features from neighboring nodes. The core concept is neighborhood aggregation of node features, meaning that each node's features are determined not only by its own features but also by the features of its neighboring nodes.

[0076] Through the joint learning of GCN and LSTM, LSTM-GCN can simultaneously consider spatial (node topology) and temporal (time-series variation) information, providing more accurate fault classification and location capabilities. LSTM processes time series features, capturing temporal variations in current and voltage, modeling long-term dependencies and avoiding the misclassification problems caused by traditional methods based solely on single-time-step data. GCN processes grid topology information and models the adjacency relationships of the distribution network, enabling inter-node feature propagation. This ensures that fault detection relies not only on individual node data but also incorporates information from neighboring nodes, improving detection accuracy. The two neural network modules are combined sequentially. The LSTM network takes as input the time series features of the nodes and outputs hidden layer features, which capture the temporal dependencies of node features. The GCN network takes as input the output features of the LSTM network. Each layer of the GCN performs neighbor feature aggregation (message passing), allowing each node to integrate information from neighboring nodes and enhance its perception of the global network state. After multiple layers of graph convolution, the output is the fault classification probability for each node.

[0077] Step S3: The fault detection model is trained using graph structure data with fault labels, and the fault detection model is jointly optimized using the physical information constraint loss function and the classification loss function.

[0078] Specifically in this embodiment, the training process is as follows:

[0079] Step S3.1, initialize the model parameters of the fault detection model; the model parameters include the network parameters of the LSTM network, the network parameters of the multi-layer GCN network, and the weight parameters of the adaptive frequency aggregation mechanism.

[0080] Step S3.2: Input the graph structure data with fault labels as input data into the fault detection model, input the node feature matrix in the graph structure data into the LSTM network to extract the time series features, input the time series features into the multi-layer GCN network, and use the adjacency matrix in the graph structure data to realize feature propagation between nodes to extract the time series features with spatial features, and perform mapping to obtain the fault classification probability of each node.

[0081] Step S3.3: Calculate the physical information constraint loss function and the classification loss function based on the fault classification probability and fault label of each node to obtain the total loss.

[0082] The physical information constraint loss function includes:

[0083] Kirchhoff's current law losses :

[0084]

[0085] Where, is the total number of nodes, For nodes The set of neighbor nodes of For nodes Flow Node The current value, For nodes Flow Node The current value;

[0086] Kirchhoff's voltage law losses :

[0087]

[0088] Where, For nodes voltage value.

[0089] The classification loss function is:

[0090]

[0091] Where, is the classification loss, is the total number of categories, For category The weight of For nodes The fault label, For nodes Predicted as category probability;

[0092]

[0093] Where, is the total number of nodes, For category The total number of nodes, is the smoothing factor. It is used to prevent the weight from being too large or too small. It is generally set to 0.1-0.5. The larger the value, the more significant the impact of class imbalance.

[0094] Total loss for:

[0095]

[0096] Where, are Kirchhoff's current law loss and Kirchhoff's voltage law loss, for The weight of .

[0097] The present invention introduces PINN physical constraints during the training process, constructs a loss function based on Kirchhoff's current law (KCL) and Kirchhoff's voltage law (KVL), and combines it with a classification loss function to ensure that the model complies with the basic principles of the power system while performing data-driven learning, thereby improving generalization ability and physical interpretability.

[0098] Step S3.4: Update the model parameters using the Adam optimizer based on the total loss, and limit the gradient norm in combination with the gradient clipping technique.

[0099] Step S3.5: Repeat the forward propagation and backpropagation processes until the total loss reaches a preset threshold or the number of training times reaches a set value, completing the training of the fault detection model.

[0100] Traditional GCN networks mostly extract common information from nodes and retain the commonalities of the nodes, which can be regarded as low-pass filters. However, when performing fault detection, low-pass filtering will cause the differences between nodes to be ignored. The difference characteristics between faulty nodes and non-faulty nodes should be maintained so that the node embedding of faulty nodes and non-faulty nodes has a high degree of distinction. Therefore, adaptive frequency aggregation is used to decompose the low-frequency and high-frequency components in the GCN network, and their ratio is dynamically adjusted through learnable weights to enhance the model's adaptability to different types of node features. The expression of adaptive frequency aggregation is:

[0101]

[0102] Where, Output features (low-frequency information) of the GCN network, which are obtained by convolution of the GCN network, represent smooth information propagated from neighboring nodes and are suitable for global pattern recognition (such as group failure). Input features and characteristics for GCN network The difference (high-frequency information) retains local details (such as abnormal points and edge node features). are learnable weights, The features obtained by adaptive frequency aggregation are used as the input features of the next layer of GCN network.

[0103] Step S4: perform fault detection on the distribution network to be detected using the trained fault detection model.

[0104] During deployment, test data is fed into the trained model, which then outputs fault classification results for each node. Faulty nodes and their types are counted by time step, and information about the number of faulty nodes and fault types at each time step is output to assist operations personnel in quickly locating faults. This method supports real-time data processing, inputting operational data and outputting fault detection results in JSON format.

[0105] Combined with the data obtained from the simulation above, training verification is performed. The machine learning framework is Pytorch. The learning rate used in the training process is 0.01, the training epoch is set to 200, the physical information constraint loss weight is set to 0.2, and the data is trained. The loss and accuracy during training are as follows Figure 3 As shown in Table 1, the proposed improved GCN-LSTM model can achieve an accuracy of 96.8% on the 127 laboratory power grid simulation dataset. Compared with other deep learning models, the model proposed in this paper shows extremely high superiority.

[0106] Table 1. Comparison of GCN-LSTM model with other models

[0107]

[0108] At the same time, in order to study the impact of noise on model analysis, noise with a signal-to-noise ratio of 5% and 10% was added to the simulation data, and training and classification were performed again. The accuracy of the GCN-LSTM model under noise is shown in Table 2.

[0109] Table 2. Detection accuracy of GCN-LSTM model under noise

[0110]

[0111] As shown in Table 2, this model has a certain ability to resist noise interference. Under 10% noise interference, the model accuracy only drops by 4%, and the impact of noise interference with a 5% signal-to-noise ratio on the accuracy is even smaller, demonstrating the strong robustness of the model.

[0112] In summary, the embodiments of the present invention have the following features:

[0113] (1) The spatiotemporal feature extraction capability of the LSTM-GCN combination. The LSTM layer can effectively capture the time series characteristics of the voltage and current of the distribution network nodes and identify the dynamic change pattern of fault occurrence. By processing data with multiple time steps, the LSTM captures the time dependency and periodicity characteristics, improving the detection capability of sudden faults and gradually evolving faults. The GCN layer extracts the spatial relationship characteristics between nodes based on the topological structure of the distribution network. GCN propagates feature information through the adjacency matrix, making full use of the physical connection characteristics of the power grid, and can accurately locate the fault propagation path and impact range. The combination of the two can realize the spatiotemporal joint detection of distribution network faults, significantly improving the accuracy and robustness of fault identification.

[0114] (2) The adaptive frequency aggregation (AFA) mechanism improves feature extraction quality. After each GCN convolution layer, the AFA mechanism automatically distinguishes and adjusts high-frequency information (abnormal signals, local features) and low-frequency information (smooth features, global patterns). Through learnable parameters, the model adaptively determines the weights of high- and low-frequency features, avoiding the loss of feature information caused by over-smoothing in traditional models. This enhances the model's sensitivity to fault signal details, especially when detecting complex or weak fault signals, and can better retain key features, helping to improve the model's fault location accuracy.

[0115] (3) Introducing physical information constraints to improve the physical consistency of the model. During the model training process, physical constraints based on Kirchhoff's current law (KCL) and Kirchhoff's voltage law (KVL) of the power system were added. These physical laws ensure the conservation relationship between node current and voltage, so that the model conforms to the basic physical laws of the distribution network when making predictions. This makes the model not only data-driven, but also ensures the physical interpretability and consistency of the prediction results, reduces non-physical results caused by overfitting, and significantly improves the reliability of the model. By reducing the dependence on large-scale labeled data and introducing physical constraints, the model can maintain good performance even with a small amount of data by following the physical laws of the power system. This method is particularly important in practical application scenarios where data is scarce. It can reduce data acquisition costs and improve the generalization ability and practicality of the model.

[0116] Example 2:

[0117] An embodiment of the present invention provides a device for detecting a short-circuit fault in a distribution network, comprising:

[0118] A data processing module is configured to obtain the topology of the target distribution network and the voltage and current data of each node during the fault period, construct graph structure data and add fault labels;

[0119] The model building module is configured to build a fault detection model that includes an LSTM network and a multi-layer GCN network, and introduce an adaptive frequency aggregation mechanism in each layer of the GCN network;

[0120] a model training module configured to train a fault detection model using graph-structured data with fault labels and to jointly optimize the fault detection model using a physical information constraint loss function and a classification loss function;

[0121] The model deployment module is configured to perform fault detection on the distribution network to be detected using the trained fault detection model.

[0122] Example 3:

[0123] Based on the detection method provided in Example 1, an embodiment of the present invention provides an electronic device, including a processor and a storage medium;

[0124] The storage medium is used to store instructions;

[0125] The processor is configured to operate according to the instructions to execute the steps of the above method.

[0126] Example 4:

[0127] Based on the detection method provided in Example 1, an embodiment of the present invention provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the above method are implemented.

[0128] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0130] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0131] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0132] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting a short-circuit fault in a distribution network, characterized in that: include: Obtain the topology of the target distribution network and the voltage and current data of each node during the fault period, construct graph structure data and add fault labels; Build a fault detection model that includes an LSTM network and a multi-layer GCN network, and introduce an adaptive frequency aggregation mechanism in each GCN layer; The fault detection model is trained using graph structure data with fault labels, and the fault detection model is jointly optimized using a physical information constraint loss function and a classification loss function; Fault detection is performed on the distribution network to be detected using the trained fault detection model; The physical information constraint loss function includes: Kirchhoff's current law losses : ； Where, is the total number of nodes, For nodes The set of neighbor nodes of For nodes Flow Node The current value, For nodes Flow Node The current value; Kirchhoff's voltage law losses : ； Where, For nodes voltage value.

2. The method for detecting a short-circuit fault in a distribution network according to claim 1, wherein: The step of constructing labeled graph structure data includes: Construct an adjacency matrix based on the topological structure of the target distribution network , the adjacency matrix The row and column numbers correspond to the nodes, and the adjacency matrix of OK Listed as : ； A node characteristic matrix is constructed according to the voltage and current data of each node during the fault period, including: Preprocess the acquired voltage and current data, and construct a node feature matrix based on the preprocessed voltage and current data , node feature matrix The row and column numbers correspond to the sampling time and node, and the node feature matrix of OK Listed as : ； Where, For nodes At the sampling time The pre-processed voltage and current data; According to the adjacency matrix and the node feature matrix Building graph-structured data .

3. The method for detecting a short-circuit fault in a distribution network according to claim 2, wherein: The preprocessing includes data cleaning and normalization.

4. The method for detecting a short-circuit fault in a distribution network according to claim 1, wherein: The training of the fault detection model using graph structure data with fault labels and the joint optimization of the fault detection model using a physical information constraint loss function and a classification loss function include: Initialize the model parameters of the fault detection model; The graph structure data with fault labels is input into the fault detection model as input data. The node feature matrix in the graph structure data is input into the LSTM network to extract time series features. The time series features are input into the multi-layer GCN network. The adjacency matrix in the graph structure data is used to realize inter-node feature propagation to extract time series features with spatial features, and then mapped to obtain the fault classification probability of each node. The total loss is obtained by calculating the physical information constraint loss function and the classification loss function based on the fault classification probability and fault label of each node; Based on the total loss, the Adam optimizer is used to update the model parameters, and the gradient clipping technique is combined to limit the gradient norm; Repeat the forward propagation and back propagation process until the total loss reaches a preset threshold or the number of training times reaches a set value, completing the training of the fault detection model.

5. The method for detecting a short-circuit fault in a distribution network according to claim 1, wherein: The introduction of the adaptive frequency aggregation mechanism in each layer of the GCN network includes: Adaptive frequency aggregation is used to decompose the low-frequency and high-frequency components in the GCN network, and the ratio of low-frequency and high-frequency components is dynamically adjusted through learnable weights to enhance the adaptability of the GCN network to different node features. The expression of adaptive frequency aggregation is: ； Where, Output features for the GCN network, Input features and characteristics for GCN network The difference, are learnable weights, The features obtained by adaptive frequency aggregation are used as the input features of the next layer of GCN network.

6. The method for detecting a short-circuit fault in a distribution network according to claim 1, wherein: The classification loss function is: ； Where, is the classification loss, is the total number of categories, For category The weight of For nodes The fault label, For nodes Predicted as category probability; ； Where, is the total number of nodes, For category The total number of nodes, is the smoothing factor.

7. A detection device for short-circuit fault in a distribution network, characterized in that: include: A data processing module is configured to obtain the topology of the target distribution network and the voltage and current data of each node during the fault period, construct graph structure data and add fault labels; The model building module is configured to build a fault detection model that includes an LSTM network and a multi-layer GCN network, and introduce an adaptive frequency aggregation mechanism in each layer of the GCN network; a model training module configured to train the fault detection model using graph structure data with fault labels, and to jointly optimize the fault detection model using a physical information constraint loss function and a classification loss function; A model deployment module is configured to perform fault detection on the distribution network to be detected using the trained fault detection model; The physical information constraint loss function includes: Kirchhoff's current law losses : ； Where, is the total number of nodes, For nodes The set of neighbor nodes of For nodes Flow Node The current value, For nodes Flow Node The current value; Kirchhoff's voltage law losses : ； Where, For nodes voltage value.

8. An electronic device, characterized in that: including processor and storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images