Power distribution network fault positioning method and system fusing space-time physical modeling and graph attention mechanism

Abstract

The invention discloses a power distribution network dynamic fault positioning method and system fusing space-time physical modeling and a graph attention mechanism, and the method comprises the steps: collecting topological structure data, equipment operation parameters and fault signal feature data of a power distribution network, and carrying out the preprocessing and enhancement; secondly, a dynamic topological graph structure reflecting the physical connection relation and the real-time operation state of the power distribution network is constructed; then, designing a fault positioning model fusing a hierarchical global graph attention mechanism and space-time physical modeling; training the model by using the diversified fault scene data; comprehensively evaluating the performance of the model by adopting positioning precision, response speed and robustness indexes; and finally, integrating the optimized model to the intelligent operation and maintenance platform of the power distribution network. According to the method, the fault positioning accuracy and real-time performance in the dynamic topology environment are remarkably improved, and reliable technical support is provided for efficient operation and maintenance and safe operation of the intelligent power distribution network.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of power system protection and control and artificial intelligence, and in particular relates to a method and system for dynamic fault location in distribution networks that integrates spatiotemporal physical modeling and graph attention mechanism. Background Technology

[0002] As a crucial link in the energy transformation, the power distribution network requires consideration of numerous factors in its current upgrade and transformation. These factors include ensuring electricity demand under the current new power system, planning for future electricity demand development, and simultaneously considering economic costs, electricity safety, and fault detection. Fault location in the power distribution network is a vital topic, not only necessary for daily operation and maintenance management and improving efficiency, but also the starting point for building smart distribution networks and the cornerstone of all modern power distribution technology research and application.

[0003] However, traditional fault location methods for distribution networks face numerous challenges during this transformation process, urgently requiring technological innovation and model upgrades to address the following prominent issues: Current research typically assesses risk from a single perspective—distribution network equipment, network structure, or environment—and has not yet developed a fault location method that comprehensively considers multiple factors of the distribution network. This makes it difficult to comprehensively and systematically guide distribution network operation and maintenance management and planning. Furthermore, research methods lack in-depth theoretical research and information pattern mining; models are relatively simple, with constants and boundary constraints relying on experience; they focus on their own fault models, single-level, single-section health states, and post-fault results. The health state of equipment and the distribution network has not yet been organically integrated with maintenance work based on fault cause analysis.

[0004] Graph neural networks (Graph Neural Networks) are a novel type of neural network model based on spatial structural information. Physical structural relationships play a crucial constraining role in the model's learning process. This enables them to possess stronger feature extraction capabilities and faster training speeds when facing increasingly complex node information and dynamically changing topologies. Among these, graph attention mechanisms demonstrate superior performance in addressing these challenges. Graph attention mechanisms, through feature aggregation and information propagation of nodes and edges, can effectively capture the topological relationships and dynamic characteristics of distribution networks. Simultaneously, they adjust network parameters according to changes in the distribution network topology, thereby achieving high accuracy and response speed in fault location under different fault types and complex environments, better realizing the goal of rapid and accurate fault location in distribution networks. Existing graph attention networks primarily achieve feature aggregation through neighbor-level attention mechanisms, but their information propagation range is limited to neighboring nodes, making it difficult to effectively capture long-distance topological dependencies across regions, and they are prone to oversmoothing problems as the number of layers increases.

[0005] In the process of implementing this invention, the inventors of this application discovered the following challenges in current methods for locating faults in power distribution networks: Equipment operation status management is diverse and complex. With the increasing number and types of distribution network equipment, as well as the diversification of equipment operating environments and service lives, the health of the distribution network is affected by multiple factors, leading to a significant increase in the probability of fault occurrence, and the causes are becoming increasingly diversified. At the same time, the dynamic changes in the distribution network topology and the weakness of fault signals pose significant challenges to fault location. Traditional fault location methods, when faced with massive amounts of distribution network operation data, cannot effectively utilize the correlation between data, topological structure, and dynamic evolution characteristics, resulting in one-sidedness and lag in distribution network fault location and analysis, making it difficult to meet the requirements of efficient and safe operation.

[0006] Adapting to dynamic changes in distribution network topology is challenging. The topology of a distribution network constantly changes during operation; load fluctuations, distributed power source integration, equipment maintenance, and fault repair all necessitate adjustments to the distribution network topology. Traditional fault location methods rely on static topology models and massive data-driven training to improve performance. However, these methods cannot identify distribution network topology information, making it difficult to balance sensitivity and adaptability to multiple scenarios. Their ability to update and evolve in the face of rapidly changing new scenarios is insufficient, thus hindering their ability to quickly adapt and maintain efficient operation in situations with frequent changes in distribution network topology.

[0007] Real-time requirements and computational costs are high. In practical applications, power distribution network fault location technology needs to achieve real-time monitoring and location, requiring models to quickly respond to the location of faulty nodes. However, complex deep learning models often require a large amount of computational resources during training and inference, making it difficult to run efficiently on low-power devices. The existing models exhibit a significant contradiction between real-time performance and resource consumption in practical applications, hindering the widespread adoption of the system. Summary of the Invention

[0008] In view of this, the present invention provides a method and system for dynamic fault location in distribution networks that integrates spatiotemporal physical modeling and graph attention mechanism, in order to solve or at least partially solve the existing difficulties in fault location in distribution networks.

[0009] Based on the above-mentioned technical problems, the present invention adopts the following technical solution: A dynamic fault location method for distribution networks that integrates spatiotemporal physical modeling and graph attention mechanisms includes: Step S1: Collect distribution network data, including various operating parameters and electrical characteristic data when a fault occurs in the distribution network, as well as distribution network topology data; Step S2: Preprocess and enhance the collected power distribution network data; Step S3: Based on the preprocessed distribution network data, construct a dynamic topology diagram structure that reflects the physical connection relationship and real-time operating status of the distribution network; Step S4: Construct a fault location model that integrates hierarchical global graph attention and spatiotemporal physical modeling. Utilize the dynamic topology graph structure to capture the spatiotemporal propagation patterns and topological dependencies of fault signals through feature encoding, information aggregation, and topology awareness, thereby achieving fault location. Step S5: Train and optimize the fault location model using training data containing the dynamic topology graph structure and its corresponding labels, and adjust the fault location model parameters by using a joint loss function and optimization algorithm; Step S6: Use a test set to validate the optimized fault location model and evaluate its location accuracy, robustness, and generalization ability. Step S7: Deploy the optimized fault location model into the actual fault location system, obtain feedback from actual power distribution network fault scenarios, and continuously optimize the fault location model based on application feedback.

[0010] Furthermore, step S1 specifically includes the following steps: Step S1.1: Collect distribution network topology data, including unique node ID, node type, parent node ID, feeder, voltage level, latitude and longitude, altitude, supplementary line parameters, and switch status; Step S1.2: Collect fault event data, including fault type, start and end time, fault node or section, and fault event of the feeder to which it belongs. Fault events include switch action records and electrical measurement data before and after the fault. Step S1.3: Collect equipment health status data, including specific status parameters of transformer equipment, switchgear, cable equipment, and capacitor equipment; Step S1.4: Store all collected data in a unified structured format and ensure time series alignment.

[0011] Furthermore, step S2 specifically includes the following steps: Step S2.1: Fill in the electrical measurement data and equipment health status data using a linear interpolation method; Step S2.2: Perform outlier cleaning on all collected data based on statistical principles; Step S2.3: Standardize or normalize the electrical measurement data and equipment health status data; Step S2.4: Enhance the acquired electrical measurement data by injecting noise, random shielding features, or time window perturbation.

[0012] Furthermore, step S3 specifically includes the following steps: Step S3.1: Based on the collected distribution network topology data, construct the basic elements of the graph including nodes and edges, construct the static electrical properties of nodes and edges, and generate the adjacency matrix of the graph structure at the initial time. Step S3.2: Weight and dynamically update the static electrical attribute correlation and the adjacency matrix at the initial time to form a time-varying weighted adjacency matrix: Step S3.3: Encode the features of nodes and edges to construct the distribution network node feature matrix. Sum of edge feature matrices ; Step S3.4: Generate fault labels for fault events. Based on the fault event record table, map the fault location at the corresponding time to the edge label to obtain the graph-label pair (G(t), y(t)), where G(t) represents the graph data at that time and y(t) represents the label data corresponding to the graph data at that time.

[0013] Furthermore, step S4 specifically includes the following steps: Step S4.1: Map the feature matrices of the distribution network nodes and the feature matrices of the edges as inputs to node feature vectors and edge feature vectors: Step S4.2: Fuse node features and adjacent edge information through a topology-gated attention mechanism; the topology-gated attention mechanism extracts global topology information by aggregating the mean of adjacent edge features, and generates dynamic gating coefficients by combining node features with a learnable weight matrix and activation function to weighted aggregate the features of nodes and edges; Step S4.3: Based on the time-varying weighted adjacency matrix, the multi-head graph attention layer adaptively aggregates neighborhood information. The attention scores between the node and its neighbors are calculated and normalized by Softmax to obtain the attention weights. The node features are then updated by weighted summation of the neighbor features. Multiple independent attention heads are used for parallel computation. Step S4.4: Adaptive sampling and feature compression of the graph structure are performed through a dynamic pooling mechanism; the dynamic pooling mechanism fuses the context information of node and edge features through global gating, and calculates the local importance of nodes based on multi-head attention through local gating, and calculates the final feature weights by combining the two through the Hadamard product; Step S4.5: Using a fully connected mapping and a Softmax classifier, the processed node representation vector is mapped to the corresponding fault location probability distribution.

[0014] Furthermore, step S5 specifically includes: Step S5.1: Divide the generated graph-label pairs into training set, validation set and test set according to a predetermined ratio; Step S5.2: Define a joint loss function, which consists of a cross-entropy loss used to measure the difference between the predicted probability and the true distribution, and an attention consistency constraint term used to constrain the differences in the attention matrices of adjacent layers to maintain the consistency of topological association. Step S5.3: Use a gradient descent-based optimization algorithm to update the model parameters with the joint loss function as the objective, and combine the learning rate scheduling strategy and early stopping strategy to control the training process; Step S5.4: When calculating the loss, denormalize the model predictions and the true values, and convert both the predictions and the true values ​​back to the original scale; Step S5.5: During training, save the model parameters, optimizer state, and related configurations when the validation metrics are optimal.

[0015] Furthermore, step S6 specifically includes: Step S6.1: Under the same dataset and parameter configuration, conduct comparative experiments with at least one existing model such as FCNN, GraphSAGE, and GCN, and quantify the performance comparison with accuracy as the main indicator; Step S6.2: By removing or replacing at least one key component of the model, such as the edge feature encoder, topological attention enhancement, dynamic graph pooling layer, multi-head attention mechanism, and physical data augmentation strategy, construct several ablation groups and compare and analyze the impact of each component on the model localization accuracy. Step S6.3: Inject random noise, construct non-uniform interference, simulate partial data loss, or introduce at least one type of interference into the test data, and test the change in positioning accuracy of the model under different interference scenarios. Step S6.4: Test the model using untrained new distribution network topology data or different power load conditions data to evaluate its positioning accuracy and verify its generalization ability.

[0016] Furthermore, step S7 specifically includes: Step S7.1: Deploy the trained and validated model to the server, and design a real-time data input interface for communication with power monitoring equipment to receive and preprocess real-time operating data; Step S7.2: Based on the fault location feedback data generated during actual system operation and the newly collected sample data, periodically or trigger-based retraining and hyperparameter adjustment are performed on the deployed model.

[0017] On the other hand, the present invention provides a dynamic fault location system for distribution networks that integrates spatiotemporal physical modeling and graph attention mechanisms, comprising: Data acquisition module: It is used to collect data of the distribution network, including various operating parameters and electrical characteristics data of the distribution network when a fault occurs, as well as the distribution network topology data; Data preprocessing and enhancement module: It is used to preprocess and enhance the collected power distribution network data; Dynamic graph construction module: It is used to construct a dynamic topology graph structure that reflects the physical connection relationship and real-time operating status of the distribution network based on preprocessed distribution network data; Fault localization model construction module: It is used to construct a deep learning model that integrates hierarchical global graph attention and spatiotemporal physical modeling. By utilizing the constructed dynamic topology graph structure, it captures the spatiotemporal propagation law and topological dependency relationship of fault signals through feature encoding, information aggregation and topology awareness, so as to realize fault localization. Model training module: It is used to train and optimize the deep learning model using training data containing the dynamic topology graph structure and its corresponding labels, and to adjust the parameters of the deep learning model through a joint loss function and optimization algorithm; Evaluation module: It is used to validate the optimized deep learning model using a test set, and to evaluate the localization accuracy, robustness and generalization ability of the optimized deep learning model. Model optimization and control module: It is used to deploy the optimized deep learning model to the actual fault location system, obtain feedback in the actual power distribution network fault scenario, and continuously optimize the model based on the application feedback.

[0018] Furthermore, it also includes: Results visualization module: It is used to display the fault location results on the distribution network operation platform in a multi-level and multi-dimensional way. This module can intuitively present the location of fault nodes or branches in the form of highlights, thermal distribution or color markings in the topology map, and supports hierarchical zoom and regional aggregation to help quickly locate the fault range.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. The model uses a dynamic graph structure to reflect the topology changes of the distribution network in real time and simultaneously captures the evolution of electrical quantities such as voltage and current over time. For example, during a fault, the model can not only identify sudden changes in electrical quantities near the fault point, but also determine the scope of the fault's impact by combining topology changes, effectively distinguishing between real faults and electrical quantity fluctuations caused by network reconfiguration, thus improving the modeling capability for complex fault scenarios. 2. By incorporating key information such as line impedance, switch status, and fault indicator signals into the model through edge feature encoding, the limitations of traditional methods that rely solely on node features are avoided, effectively improving the accuracy of feature fusion and fault location. 3. Data augmentation extends the model's adaptability to abnormal operating conditions through topology perturbation, electrical noise injection, and time-series data interpolation. The dynamic pooling mechanism divides feature aggregation regions based on real-time topology. When local data is missing or communication is abnormal, only the positioning accuracy of the corresponding region is affected, while other regions can still maintain high accuracy based on complete information, effectively avoiding global performance degradation caused by local anomalies.

[0020] 4. Verification experiments show that the method of the present invention is significantly superior to existing methods in terms of accuracy, generalization ability and anti-interference performance, and has good application prospects. Attached Figure Description

[0021] Figure 1 This is a flowchart of a dynamic fault location method for distribution networks that integrates spatiotemporal physical modeling and graph attention mechanism according to the present invention; Figure 2 The following is a specific embodiment of the dynamic fault location method for power distribution networks that integrates spatiotemporal physical modeling and graph attention mechanism according to the present invention: the model training convergence process is shown in Figures (a)-(c), and the model loss convergence process is shown in Figures (d)-(e). Figure 3 The graph shows the performance evaluation results of the dynamic fault location method for power distribution networks that integrates spatiotemporal physical modeling and graph attention mechanism in this invention on a test dataset. Detailed Implementation

[0022] Through extensive research and practice, the inventors of this application have discovered the following challenges in the current field of distribution network fault location: 1) Equipment operation status management is diverse and complex. With the increasing number and types of distribution network equipment, as well as the diversification of equipment operating environments and service lives, the health of the distribution network is affected by multiple factors, leading to a significant increase in the probability of fault occurrence, and the causes are becoming increasingly diverse. At the same time, the dynamic changes in the distribution network topology and the weakness of fault signals pose significant challenges to fault location. Traditional fault location methods cannot effectively utilize the correlation between data, topological structure, and dynamic evolution characteristics when faced with massive amounts of distribution network operation data, resulting in one-sidedness and lag in distribution network fault location and analysis, making it difficult to meet the requirements of efficient and safe operation.

[0023] 2) Difficulty in adapting to dynamic changes in topology. The topology of a distribution network is constantly changing during operation. Factors such as load fluctuations, distributed power source integration, equipment maintenance, and fault repair all cause adjustments to the distribution network topology. Traditional fault location methods rely on static topology models and massive data-driven training to improve performance. However, they cannot identify distribution network topology information, making it difficult to balance sensitivity and adaptability to multiple scenarios. Their ability to update and evolve in the face of rapidly changing new scenarios is insufficient, thus making it impossible to quickly adapt and maintain efficient operation when the distribution network topology changes frequently.

[0024] 3) High real-time requirements and computational costs. In practical applications, power distribution network fault location technology needs to achieve real-time monitoring and location, requiring models to quickly respond to the location of faulty nodes. However, complex deep learning models often require a large amount of computational resources during training and inference, making it difficult to run efficiently on low-power devices. The existing models exhibit a significant contradiction between real-time performance and resource consumption in practical applications, hindering the widespread adoption of the system.

[0025] Based on the above considerations, this invention proposes a dynamic fault location system method for power distribution networks that integrates spatiotemporal physical modeling and graph attention mechanism. The aim is to overcome the problems of low fault location accuracy, poor robustness, and insufficient generalization ability in the existing technology, and to achieve multi-source heterogeneous data fusion, dynamic graph structure modeling, and attention mechanism optimization, so as to maintain high accuracy and stability under complex working conditions.

[0026] To achieve the above objectives, the main concept of this invention is as follows: This paper first proposes a dynamic fault location method for distribution networks that integrates spatiotemporal physical modeling and graph attention mechanisms. By constructing a deep learning model that can simultaneously capture the spatial topological features of the distribution network and the evolution characteristics of fault events, high-precision identification of fault locations is achieved. Addressing the problem that traditional methods rely on manual rules and struggle to effectively utilize complex topological structures, this invention introduces dynamic graph modeling. Buses and load nodes are treated as graph nodes, and lines and switches as graph edges, with added electrical attributes. This dynamically constructs a graph structure reflecting the topological state and operational characteristics, thereby enhancing the representation of complex electrical information in the distribution network. To address the challenge of strong spatial dependence in fault signal propagation, this invention designs a feature-encoded, hierarchical global graph attention network that organically combines local neighborhood, regional aggregation, and global topological information. Under a dynamic pooling mechanism, it achieves adaptive selection of key nodes and edges, thus accurately capturing the transmission patterns of fault signals. During training, a joint loss function is constructed using cross-entropy and attention consistency constraints, and the Adam optimizer is used to dynamically adjust model parameters and hyperparameters to ensure training convergence and model generalization performance. To ensure robustness under different operating environments, this invention verifies, through comparative experiments, ablation experiments, and anti-interference experiments, that the constructed model maintains high accuracy and stability under various scenarios such as noise interference, topology disturbances, and data loss. Ultimately, this method can be deployed in real-time in actual distribution network operation, performing fault location based on real-time collected operating parameters and topology status. Combined with a feedback mechanism, the model is continuously optimized, achieving intelligent operation and maintenance and efficient fault diagnosis of the distribution network. This effectively overcomes the problems of strong dependence, poor adaptability, and insufficient generalization ability of traditional fault location methods.

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1 like Figure 1 As shown, this embodiment provides a dynamic fault location method for distribution networks that integrates spatiotemporal physical modeling and graph attention mechanisms. This method relies on the Simulink-MATLAB platform to build an IEEE 33-node distribution network simulation system, then constructs a distribution network fault dataset, and through data acquisition, preprocessing, dynamic graph construction, model training, and validation, ultimately achieves accurate fault location in the distribution network. The establishment of this method includes: Step S1: Collect distribution network data, including various operating parameters and electrical characteristic data when a fault occurs in the distribution network, as well as distribution network topology data; Step S2: Preprocess and enhance the collected power distribution network data; Step S3: Based on the preprocessed distribution network data, construct a dynamic topology diagram structure that reflects the physical connection relationship and real-time operating status of the distribution network; Step S4: Construct a fault location model that integrates hierarchical global graph attention and spatiotemporal physical modeling. Utilize the dynamic topology graph structure to capture the spatiotemporal propagation patterns and topological dependencies of fault signals through feature encoding, information aggregation, and topology awareness, thereby achieving fault location. Step S5: Train and optimize the fault location model using training data containing the dynamic topology graph structure and its corresponding labels, and adjust the fault location model parameters by using a joint loss function and optimization algorithm; Step S6: Use a test set to validate the optimized fault location model and evaluate its location accuracy, robustness, and generalization ability. Step S7: Deploy the optimized fault location model into the actual fault location system, obtain feedback from actual power distribution network fault scenarios, and continuously optimize the fault location model based on application feedback.

[0029] In one implementation, step S1 specifically includes: Step S1.1: In this embodiment, the fault dataset is generated by Simulink-MATLAB simulation. An IEEE 33 standard distribution network simulation system is built on the simulation platform. Distribution network topology data is acquired, including unique node IDs, node types, parent node IDs, associated feeders, voltage levels, latitude and longitude, and altitude; line parameters (resistance R, reactance X), switch status, etc., are supplemented to ensure the topology data is complete and reflects the power grid structure.

[0030] Step S1.2: In this embodiment, in order to collect fault data, various fault scenarios such as single-phase grounding, two-phase short circuit, and three-phase short circuit are set up at each node of the distribution network system. Then, relevant electrical data of each fault sample are collected, including fault events containing fault type, start and end time, fault node or section, and feeder to which it belongs; action type, action time, switching action of switching equipment; and steady-state measurements before and after the fault, including three-phase voltage amplitude and phase angle, three-phase current amplitude and phase angle, system frequency, and node power. Step S1.3: In this embodiment, it is also necessary to collect equipment health status data. The electrical attributes collected vary depending on the type of equipment. For transformer equipment, oil temperature, winding temperature, and gas content in the oil are collected; for switchgear equipment, contact resistance and operating life are collected; for cable equipment, insulation resistance and partial discharge are collected; for capacitor equipment, capacitance deviation is collected. Step S1.4: In this embodiment, for data recording and storage, all collected data must be stored in a unified structured format (such as CSV, Parquet, or database) to support subsequent batch processing and real-time querying. Time series data should maintain the original sampling frequency to ensure time alignment and facilitate subsequent processing.

[0031] In one implementation, step S2 specifically includes: Step S2.1: In this embodiment, missing value processing is first performed on the collected data. For missing electrical quantities (such as voltage) in the time series... V t Current I t Linear interpolation is used to fill in the gaps to ensure data integrity. The formula is as follows:

[0032] in, t 1 and t 2 represents adjacent observed time points. For cases where the missing interval is short, data from nearby time points are used directly to supplement the missing data. This represents the data after missing values ​​have been filled in. , These represent data points close to the time of the missing value.

[0033] Step S2.2: In this embodiment, after handling missing values ​​in the data samples, outlier cleaning needs to be performed on the collected data. The 3σ principle is used for rapid screening. x satisfy:

[0034] Then, clean the outliers. After processing the outliers, interpolation or model reconstruction can be used to replace the removed points, and the outlier markers are retained for subsequent training weighting.

[0035] Step S2.3: In this embodiment, the collected data is standardized. To ensure that features of different dimensions and magnitudes can be effectively learned by the same neural network, voltage, current, power, temperature, etc., need to be standardized or normalized. For features that are approximately Gaussian distributed and sensitive to anomalies, Z-score standardization is used, as shown in the following formula:

[0036] in, Represents the standardized features. μ The mean, σ The standard deviation is used; for features with upper and lower bounds or that need to be mapped to a fixed interval, Min–Max normalization is used, as shown in the following formula:

[0037] in, Represents the normalized features. It is the raw data. x min and x max These are the minimum and maximum values ​​of the dataset, respectively. It is normalized data; Step S2.4: In this embodiment, the collected data is augmented. To improve the model's generalization ability under small sample sizes, noisy measurements, and topological changes, various temporal and graph structure augmentation techniques are employed while maintaining physical constraints. Under the premise of not exceeding engineering tolerances, Gaussian noise is injected in proportion to the feature standard deviation to simulate measurement errors and environmental fluctuations, thereby enhancing the model's robustness and enabling it to perform better in the face of uncertainties in real-world environments. The Gaussian injection formula is as follows:

[0038] in, The table contains data after Gaussian noise has been injected. ε This is an additive noise term, represented as Gaussian noise characteristics; Simultaneously, some time steps or feature dimensions are randomly masked to simulate measurement loss, thereby enhancing the model's fault tolerance during training. The feature masking formula is as follows:

[0039] in, m i Represented as a parameter (1) p Bernoulli random variable, This is represented as data after being processed with a random mask; In one implementation, step S3 specifically includes: Step S3.1: In this embodiment, the distribution network topology is modeled. Based on the primary system diagram and equipment attributes of the distribution network, the basic elements of the graph, including nodes and edges, are constructed, and the static electrical attributes of the nodes and edges, such as voltage level, line impedance, and switch status, are also constructed. The adjacency matrix of the graph structure at the initial moment is generated. Step S3.2: In this embodiment, the static electrical attribute correlation and the adjacency matrix at the initial time are weighted and dynamically updated to form a time-varying weighted adjacency matrix. :

[0040] in, To normalize the static electrical property correlation, in this embodiment, it is the normalized line reverse impedance. Represented as a node within the window i , j The correlation, These are the weighting parameters. The topology is updated in real time with switch operations to ensure that the graph structure reflects dynamic operating conditions.

[0041] Step S3.3: In this embodiment, the features of nodes and edges are encoded to construct the feature matrix of distribution network nodes and edges. Node features include three-phase voltage amplitude and phase angle, three-phase current amplitude and phase angle, active and reactive power, zero-sequence component and negative-sequence component, and equipment health indicators; edge features include line impedance, thermal load, and switch status.

[0042] Step S3.4: In this embodiment, fault labels are generated for fault events. Based on the fault event record table, the fault locations within the corresponding time period are mapped to edge labels, thus obtaining graph-label pairs. G (t) ,y (t) ),in G (t) This represents the graph data at that moment. y (t) This represents the label data corresponding to the graph data at that moment, used for training and validation.

[0043] In one implementation, step S4 specifically includes: Step S4.1: In this embodiment, a model input layer is designed to effectively respond to and process the fault dataset. Electrical feature samples of fault events occurring in the distribution network are used as the model's input data, specifically including the three-phase voltage amplitudes of the nodes. V i With phase angle Three-phase current amplitude I i With phase angle Active power P i reactive power Q i Line impedance z ij and switch status s ijMapped to node feature vectors and edge feature vectors:

[0044] Step S4.2: In this embodiment, to better adapt to the special topology of the distribution network, a topology enhancement and feature fusion module is designed. For the topology characteristic that some nodes in the distribution network have a significantly higher failure probability than ordinary nodes due to their special structural locations, a topology gating attention enhancement module is designed: A feature node matrix is ​​constructed as follows... The edge feature matrix is ,in N For the number of nodes, M Let be the number of sides. D For the feature dimension, global topological information is extracted using the mean of adjacent edge features, and then dynamically gated attention aggregation is performed by combining node features. Feature fusion mechanism. Represented as:

[0045] in, Represents a node The neighborhood group, Representing an edge eigenvectors, For the edge feature dimension, global topological information is extracted by normalizing the mean of adjacent edge features. This is then combined with node features for dynamic gating attention aggregation. Finally, the resulting dynamic weights are combined with historical fault data to generate topological sensitivity grading rules. The dynamic weight aggregation process is represented as follows:

[0046] in, Represents the topology gating coefficient of a node. express sigmoid Activation function This represents the learnable weight matrix. Represents a node i eigenvectors, Representing an edge e ij The module effectively addresses the shortcomings of traditional GNNs in modeling power grid connection patterns, and can form a high-attention ring structure to accurately locate abnormal nodes in fault scenarios. Step S4.3: In this embodiment, to deeply explore the correlation between electrical features and graph structure, a multi-head graph attention layer mechanism is designed. For the fault location model, a multi-head dynamic graph attention mechanism is introduced to adaptively model the correlation between nodes and their neighbors. The process involves embedding nodes and edges that have already undergone feature mapping, and then combining the node and edge features to calculate the attention score:

[0047] in, This represents the attention parameter vector, used to measure the importance of different features. This represents a vector concatenation operation. This represents the activation function, used to enhance nonlinear modeling capabilities. This score... Characterized in the first k In the layer, nodes j For nodes i The information contribution potential. Then, to ensure the comparability of attention weights within the neighborhood, the scoring results are normalized using Softmax:

[0048] in, This represents the normalized attention coefficient, i.e., the node. j For nodes i The relative importance, Represents a node i The set of neighboring nodes. Secondly, node characteristics are updated in real time. i The new feature vector is obtained by weighted summation of neighborhood features:

[0049] in, Represents the ReLU nonlinear activation function. This represents a weighted aggregation of neighbor features. In this way, the feature vector of each node can dynamically fuse the spatiotemporal information of its neighbors and edges. Finally, to improve the robustness and feature representation capability of the model, a weighted aggregation of neighbor features is adopted. K Parallel computation of individual attention heads:

[0050] in, m Indicates the index of the attention head. Indicates the first m The weight matrix of each attention head. The concatenation method is suitable for the middle layer to enhance feature dimensionality and diversity, while the averaging method is suitable for the last layer to improve prediction stability.

[0051] Step S4.4: In this embodiment, to improve the model's adaptability to changes in the distribution network topology, and considering both the overall state and local details of the distribution network system, a dynamic pooling layer module is designed. This module employs a batch-aware mechanism to achieve feature fusion and global gating. By fusing node features with contextual information from edge embeddings, local control... Based on multi-head attention calculation of local importance, through The feature weights are dynamically adjusted. Finally, a hierarchical pooling strategy is used to perform multi-level feature aggregation within each class, preserving the fine-grained structural information of the power grid topology. Specifically, global gating fuses the contextual information of node and edge features through linear transformations and nonlinear activation functions, while local gating models local electrical distance correlations through a multi-head attention mechanism to calculate the local importance of node features. The calculation formulas for global and local gating are as follows:

[0052] in, This is represented as the feature vector after global gating. Z The node feature matrix, W g This is the global gating weight matrix. b g For bias terms, express sigmoid Activation function. Local control calculates the local importance of node features through a multi-head attention mechanism, as shown in the following formula:

[0053]

[0054] in, , and They represent the first Header query, key and value matrix, For feature dimension, For the number of heads, W O This is the output weight matrix. h Indicates the first Attention weight matrix of the head This is represented as the feature vector after local control processing. Concat( ) This indicates a splicing operation. Then, local gate control and global gate control are connected via... The final feature weights are obtained by product calculation, as shown in the following formula:

[0055] in, h G Represents graph-level aggregation features. express product.

[0056] Step S4.5: In this embodiment, the final model output uses a fully connected mapping and Softmax The classifier maps the final node representation vector to the corresponding fault location probability and normalizes it, thereby obtaining the final predicted fault location.

[0057] In one implementation, step S5 specifically includes: Step S5.1: In this embodiment, the preprocessed and enhanced data sample set is divided into a training set, a validation set, and a test set in a 7:2:1 ratio, and used as input to the model for training. The input data includes various electrical characteristic data of the fault samples and distribution network topology data.

[0058] Step S5.2: In this embodiment, considering the characteristics of the power grid topology representation task, a dual joint loss function is adopted, consisting of cross-entropy loss and an attention consistency constraint term. Cross-entropy is the standard loss function for classification problems. As the basic loss for classification tasks, it drives the model to learn the classification features of power equipment nodes, effectively measuring the difference between the predicted probability and the true distribution, making its gradient with respect to the prediction error smoother, which is beneficial to model convergence, and is especially suitable for handling multi-class classification problems of fault location in power grid topology identification. Cross-entropy loss The formula is expressed as:

[0059] in, y i One-hot encoding of the real label. p i The probability distribution predicted by the model. NThis represents the total number of topology categories. The attention consistency constraint term effectively constrains the differences in attention matrices between adjacent layers, forcing the model to learn hierarchical topological relationships. This gives the hierarchical attention mechanism a progressive focusing characteristic, effectively preventing deep networks from losing information about lower-level topological relationships. The regularization term also suppresses interference from abnormal nodes in attention allocation. Its formula is expressed as:

[0060] in, Indicates the first l Attention matrix of layer, Indicates the first l+ Attention matrix of layer 1, Represented as the Frobenius norm, Here is the balance coefficient. The final joint loss function is expressed as:

[0061] Such a joint loss function can ensure classification accuracy while also adapting to the hierarchical topology characteristics of the distribution network.

[0062] Step S5.3: In this embodiment, gradient descent and optimization algorithms are used to adjust the model parameters.

[0063] Using the Adam optimizer allows the model to converge quickly by adjusting the model weights by minimizing the loss function. The update rule for the Adam optimizer is as follows:

[0064] in, yes t+ Model parameters at time 1 yes t Time-model parameters, m t and v t These are momentum and acceleration estimates, respectively. It is the learning rate. At the same time, an early stopping strategy is employed, continuously stopping on the validation set. p indivual epoch Stop if there is no improvement.

[0065] Step S5.4: In this embodiment, the prediction results are destandardized. When calculating the loss, the model predictions and the actual values ​​are destandardized to ensure that the loss reflects the actual data differences. The destandardization formula is:

[0066] By destandardizing, both predicted and actual values ​​are converted back to their original scale, allowing for a more accurate measurement of the model's actual performance.

[0067] Step S5.5: In this embodiment, the optimal model parameters are saved to prepare for model deployment. After each epoch, if the validation metric improves, a snapshot of the model parameters is saved, including weights, optimizer status, current epoch, validation metric, and standardized parameters. The three most recent checkpoints are also retained for rollback. Then, the best model is determined on the validation set using the principal metric's accuracy, and the corresponding weight file is saved. Finally, the model used for fault localization is exported, along with the data standardization parameters, class label mapping, and model hyperparameter configuration file, to ensure online reproducibility.

[0068] In one implementation, step S6 specifically includes: Step S6.1: In this embodiment, a comparative experiment is set up to verify the robustness of the model. By constructing a comparative experiment with existing mainstream models (such as FCNN, GraphSAGE, GCN, etc.), the superiority of the model in distribution network fault location is verified. The experiment is conducted under the same dataset and parameter configuration, and the accuracy is used as the main indicator for quantitative comparison to evaluate the comprehensive performance of the model on the training set, validation set, and test set, ensuring that the model performs well in multiple benchmark methods.

[0069] Step S6.2: In this embodiment, ablation experiments are set up to verify the effectiveness of each module of the model. By removing or replacing key components of the model, such as the edge feature encoder, topological attention enhancement module, dynamic graph pooling layer, multi-head attention mechanism, and physical data augmentation strategy, several ablation groups are constructed, and the contribution of each module to the overall performance is tested one by one. By comparing the changes in accuracy and loss function, the independent value and synergistic effect of each component in performance improvement are demonstrated, thereby verifying the rationality and necessity of the model design.

[0070] Step S6.3: In this embodiment, to verify that the model has a certain anti-interference capability, various interference scenarios are simulated, including random noise pollution, non-uniform interference, partial data loss, and topological abrupt changes, to test the robustness of the model under abnormal environments. The results show that the model maintains a high accuracy even under extreme high noise or large-scale missing data scenarios, proving that it has strong anti-interference capability and catastrophic failure protection mechanism.

[0071] Step S6.4: In this embodiment, to verify that the model has a certain generalization ability, a generalization test is performed on the model under a new, unseen distribution network topology and different power consumption conditions. Through topology reconstruction experiments, it is verified whether the model can adaptively adjust the attention coefficient and maintain high fault location accuracy. The results show that the model still possesses strong transferability and generalization ability under different distribution network topologies and power load scenarios, ensuring its practicality.

[0072] In one implementation, step S7 specifically includes: Step S7.1: In this embodiment, select a suitable deployment method, deploy the model to the client server, design the real-time data input interface of the model, and ensure that it can receive real-time data from the power monitoring equipment; Step S7.2: In this embodiment, the model is continuously optimized. The model is continuously trained using an ever-expanding real-world dataset, and its hyperparameters are adjusted to better adapt to applications in different scenarios. This embodiment constructs a dedicated dataset based on a self-built IEEE 33-node distribution network simulation model on the Simulink-MATLAB platform. This dataset is specifically designed for distribution network fault location research. By constructing a standardized distribution network operating environment and injecting various typical fault types for simulation, the data is systematically collected and organized, thus possessing strong representativeness, controllability, and relevance.

[0073] This dataset comprehensively records the operational characteristics of the distribution network before, during, and after a fault, covering multi-dimensional electrical quantity information at the node level, including the amplitude and phase angle of three-phase voltage and current, active power, reactive power, zero-sequence components, and negative-sequence components, as well as key dynamic features such as system frequency and node power fluctuations. At the topology level, the dataset includes unique node IDs, node types, line impedance parameters (R, X), and switch connectivity, accurately reflecting the static structure and topology evolution of the distribution network. In terms of fault scenario construction, the dataset covers various typical fault types such as single-phase grounding, two-phase short circuits, and three-phase short circuits, and details the fault occurrence time, start and end sections, fault node or section location, and associated feeder, while also including circuit breaker and switch operation records. Through simulations of different load levels and operating conditions, the dataset can demonstrate the fault characteristics of the power grid under different operating conditions, enhancing the data's diversity and application value. The specific contents of the dataset are shown in the table below:

[0074] In this embodiment, the proposed dynamic fault location model for distribution networks, which integrates spatiotemporal physical modeling and graph attention mechanism, is used to test the simulation-generated dataset. The experimental results are as follows: Figure 2 As shown.

[0075] Figure 2The diagram illustrates the comparison between the model's predicted fault location and the actual fault location under typical fault scenarios. As shown in the figure, the model can accurately identify and locate fault nodes or sections, and effectively capture the dynamic changes in electrical quantities during a fault. The model achieves over 90% accuracy after approximately 10 iterations, while the loss rate decreases rapidly. After 30 iterations, the model's accuracy on the validation set reaches 97.8%, and the loss rate drops to 0.09.

[0076] like Figure 3 As shown, this is a graph of the performance evaluation results of a dynamic fault location method for power distribution networks that integrates spatiotemporal physical modeling and graph attention mechanism on a test dataset. The radar graph in the upper left shows the model's performance on multi-dimensional performance indicators, the confusion matrix graph in the upper right is used for the analysis of fine-grained classification results, and the dual-axis composite graph at the bottom shows the model's performance at different fault locations.

[0077] Overall, the experimental results show that the power distribution network fault location model proposed in this invention, which integrates hierarchical global graph attention mechanism and spatiotemporal physical modeling, can achieve high-precision and robust fault location under complex operating conditions. This fully demonstrates the advantages of introducing dynamic graph modeling and multi-head attention mechanism, and provides strong technical support for intelligent operation and maintenance and fault diagnosis of power distribution networks.

[0078] Example 2 This embodiment provides a dynamic fault location system for distribution networks that integrates spatiotemporal physical modeling and graph attention mechanisms, including: Data acquisition module: It is used to collect data of the distribution network, including various operating parameters and electrical characteristics data of the distribution network when a fault occurs, as well as the distribution network topology data; Data preprocessing and enhancement module: It is used to preprocess and enhance the collected power distribution network data; Dynamic graph construction module: It is used to construct a dynamic topology graph structure that reflects the physical connection relationship and real-time operating status of the distribution network based on preprocessed distribution network data; Fault localization model construction module: It is used to construct a deep learning model that integrates hierarchical global graph attention and spatiotemporal physical modeling. By utilizing the constructed dynamic topology graph structure, it captures the spatiotemporal propagation law and topological dependency relationship of fault signals through feature encoding, information aggregation and topology awareness, so as to realize fault localization. Model training module: It is used to train and optimize the deep learning model using training data containing the dynamic topology graph structure and its corresponding labels, and to adjust the parameters of the deep learning model through a joint loss function and optimization algorithm; Evaluation module: It is used to validate the optimized deep learning model using a test set, and to evaluate the localization accuracy, robustness and generalization ability of the optimized deep learning model. Model optimization and control module: It is used to deploy the optimized deep learning model to the actual fault location system, obtain feedback in the actual power distribution network fault scenario, and continuously optimize the model based on the application feedback; It also includes a results visualization and decision support module: This module is responsible for visualizing the fault location results and providing decision support for distribution network operation and maintenance personnel. During implementation, the system can mark the fault location results on the topology map and output location accuracy, confidence level indicators, etc., enabling users to intuitively understand the power grid fault situation and take timely measures.

[0079] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic concept of the invention, can make various changes and modifications to the above embodiments without departing from the scope of the invention. Therefore, the appended claims should be interpreted as including the preferred embodiments of the invention and all equivalent variations thereof. Clearly, those skilled in the art can make various modifications to the embodiments of the present invention without departing from its spirit and scope; all modifications falling within the scope of the claims and their equivalents should be covered within the protection scope of the present invention.

[0080] All other parts not described in detail are existing technologies.

Claims

1. A method for dynamic fault location in distribution networks that integrates spatiotemporal physical modeling and graph attention mechanisms, characterized in that, include: Step S1: Collect distribution network data, including various operating parameters and electrical characteristic data when a fault occurs in the distribution network, as well as distribution network topology data; Step S2: Preprocess and enhance the collected power distribution network data; Step S3: Based on the preprocessed distribution network data, construct a dynamic topology diagram structure that reflects the physical connection relationship and real-time operating status of the distribution network; Step S4: Construct a fault location model that integrates hierarchical global graph attention and spatiotemporal physical modeling. Utilize the dynamic topology graph structure to capture the spatiotemporal propagation patterns and topological dependencies of fault signals through feature encoding, information aggregation, and topology awareness, thereby achieving fault location. Step S5: Train and optimize the fault location model using training data containing the dynamic topology graph structure and its corresponding labels, and adjust the fault location model parameters by using a joint loss function and optimization algorithm; Step S6: Use a test set to validate the optimized fault location model and evaluate its location accuracy, robustness, and generalization ability. Step S7: Deploy the optimized fault location model into the actual fault location system, obtain feedback from actual power distribution network fault scenarios, and continuously optimize the fault location model based on application feedback.

2. The method for dynamic fault location in distribution networks integrating spatiotemporal physical modeling and graph attention mechanism as described in claim 1, characterized in that, Step S1 specifically includes the following steps: Step S1.1: Collect distribution network topology data, including unique node ID, node type, parent node ID, feeder, voltage level, latitude and longitude, altitude, supplementary line parameters, and switch status; Step S1.2: Collect fault event data, including fault type, start and end time, fault node or section, and fault event of the feeder to which it belongs. Fault events include switch action records and electrical measurement data before and after the fault. Step S1.3: Collect equipment health status data, including specific status parameters of transformer equipment, switchgear, cable equipment, and capacitor equipment; Step S1.4: Store all collected data in a unified structured format and ensure time series alignment.

3. The method for dynamic fault location in distribution networks integrating spatiotemporal physical modeling and graph attention mechanism as described in claim 1, characterized in that, Step S2 specifically includes the following steps: Step S2.1: Fill in the electrical measurement data and equipment health status data using a linear interpolation method; Step S2.2: Perform outlier cleaning on all collected data based on statistical principles; Step S2.3: Standardize or normalize the electrical measurement data and equipment health status data; Step S2.4: Enhance the acquired electrical measurement data by injecting noise, random shielding features, or time window perturbation.

4. The method for dynamic fault location in distribution networks integrating spatiotemporal physical modeling and graph attention mechanism as described in claim 1, characterized in that, Step S3 specifically includes the following steps: Step S3.1: Based on the collected distribution network topology data, construct the basic elements of the graph including nodes and edges, construct the static electrical properties of nodes and edges, and generate the adjacency matrix of the graph structure at the initial time. Step S3.2: Weight and dynamically update the static electrical attribute correlation and the adjacency matrix at the initial time to form a time-varying weighted adjacency matrix: Step S3.3: Encode the features of nodes and edges to construct the distribution network node feature matrix. Sum of edge feature matrices ; Step S3.4: Generate fault labels for fault events. Based on the fault event record table, map the fault location at the corresponding time to the edge label to obtain a graph-label pair. G (t) ,y (t) ),in G (t) This represents the graph data at that moment. y (t) This indicates the label data corresponding to the graph data at that moment.

5. The method for dynamic fault location in a distribution network that integrates spatiotemporal physical modeling and graph attention mechanism as described in claim 1, characterized in that, Step S4 specifically includes the following steps: Step S4.1: Map the feature matrices of the distribution network nodes and the feature matrices of the edges as inputs to node feature vectors and edge feature vectors: Step S4.2: Fuse node features and adjacent edge information through a topology-gated attention mechanism; the topology-gated attention mechanism extracts global topology information by aggregating the mean of adjacent edge features, and generates dynamic gating coefficients by combining node features with a learnable weight matrix and activation function to weighted aggregate the features of nodes and edges; Step S4.3: Based on the time-varying weighted adjacency matrix, the multi-head graph attention layer adaptively aggregates neighborhood information. The attention scores between the node and its neighbors are calculated and normalized by Softmax to obtain the attention weights. The node features are then updated by weighted summation of the neighbor features. Multiple independent attention heads are used for parallel computation. Step S4.4: Adaptive sampling and feature compression of the graph structure are performed through a dynamic pooling mechanism; the dynamic pooling mechanism fuses the context information of node and edge features through global gating, and calculates the local importance of nodes based on multi-head attention through local gating, and calculates the final feature weights by combining the two through the Hadamard product; Step S4.5: Use fully connected mapping and Softmax The classifier maps the processed node representation vectors to the corresponding fault location probability distributions.

6. The method for dynamic fault location in a distribution network that integrates spatiotemporal physical modeling and graph attention mechanism as described in claim 1, characterized in that, Step S5 specifically includes: Step S5.1: Divide the generated graph-label pairs into training set, validation set and test set according to a predetermined ratio; Step S5.2: Define a joint loss function, which consists of a cross-entropy loss used to measure the difference between the predicted probability and the true distribution, and an attention consistency constraint term used to constrain the differences in the attention matrices of adjacent layers to maintain the consistency of topological association. Step S5.3: Use a gradient descent-based optimization algorithm to update the model parameters with the joint loss function as the objective, and combine the learning rate scheduling strategy and early stopping strategy to control the training process; Step S5.4: When calculating the loss, denormalize the model predictions and the true values, and convert both the predictions and the true values ​​back to the original scale; Step S5.5: During training, save the model parameters, optimizer state, and related configurations when the validation metrics are optimal.

7. The method for dynamic fault location in distribution networks integrating spatiotemporal physical modeling and graph attention mechanism as described in claim 1, characterized in that, Step S6 specifically includes: Step S6.1: Under the same dataset and parameter configuration, conduct comparative experiments with at least one existing model such as FCNN, GraphSAGE, and GCN, and quantify the performance comparison with accuracy as the main indicator; Step S6.2: By removing or replacing at least one key component of the model, such as the edge feature encoder, topological attention enhancement, dynamic graph pooling layer, multi-head attention mechanism, and physical data augmentation strategy, construct several ablation groups and compare and analyze the impact of each component on the model localization accuracy. Step S6.3: Inject random noise, construct non-uniform interference, simulate partial data loss, or introduce at least one type of interference into the test data, and test the change in positioning accuracy of the model under different interference scenarios. Step S6.4: Test the model using untrained new distribution network topology data or different power load conditions data to evaluate its positioning accuracy and verify its generalization ability.

8. The method for dynamic fault location in distribution networks integrating spatiotemporal physical modeling and graph attention mechanism as described in claim 1, characterized in that, Step S7 specifically includes: Step S7.1: Deploy the trained and validated model to the server, and design a real-time data input interface for communication with power monitoring equipment to receive and preprocess real-time operating data; Step S7.2: Based on the fault location feedback data generated during actual system operation and the newly collected sample data, periodically or trigger-based retraining and hyperparameter adjustment are performed on the deployed model.

9. A dynamic fault location system for distribution networks that integrates spatiotemporal physical modeling and graph attention mechanisms, characterized in that, include: Data acquisition module: It is used to collect data of the distribution network, including various operating parameters and electrical characteristics data of the distribution network when a fault occurs, as well as the distribution network topology data; Data preprocessing and enhancement module: It is used to preprocess and enhance the collected power distribution network data; Dynamic graph construction module: It is used to construct a dynamic topology graph structure that reflects the physical connection relationship and real-time operating status of the distribution network based on preprocessed distribution network data; Fault localization model construction module: It is used to construct a deep learning model that integrates hierarchical global graph attention and spatiotemporal physical modeling. By utilizing the constructed dynamic topology graph structure, it captures the spatiotemporal propagation law and topological dependency relationship of fault signals through feature encoding, information aggregation and topology awareness, so as to realize fault localization. Model training module: It is used to train and optimize the deep learning model using training data containing the dynamic topology graph structure and its corresponding labels, and to adjust the parameters of the deep learning model through a joint loss function and optimization algorithm; Evaluation module: It is used to validate the optimized deep learning model using a test set, and to evaluate the localization accuracy, robustness and generalization ability of the optimized deep learning model. Model optimization and control module: It is used to deploy the optimized deep learning model to the actual fault location system, obtain feedback in the actual power distribution network fault scenario, and continuously optimize the model based on the application feedback; The power distribution network dynamic fault location system that integrates spatiotemporal physical modeling and graph attention mechanism is used to execute the steps in the power distribution network dynamic fault location method that integrates spatiotemporal physical modeling and graph attention mechanism as described in any one of claims 1-8.

10. A dynamic fault location system for distribution networks integrating spatiotemporal physical modeling and graph attention mechanism as described in claim 9, characterized in that, Also includes: Results visualization module: It is used to display the fault location results on the distribution network operation platform in a multi-level and multi-dimensional way. This module can intuitively present the location of fault nodes or branches in the form of highlights, thermal distribution or color markings in the topology map, and supports hierarchical zoom and regional aggregation to help quickly locate the fault range.

Images