Distribution network station area topology construction and fault reasoning method based on graph neural network

Abstract

The application discloses a distribution network station area topology construction and fault reasoning method based on a graph neural network, and relates to the technical field of power distribution network operation and maintenance and intelligent analysis, which comprises the following steps: collecting multi-source measurement data of each node of a distribution network station area and abstracting the data into a graph structure; inputting the graph structure into a multi-task graph neural network model, constructing a station area topology through a link prediction task, and reasoning the hierarchical relationship of each node through a node hierarchical classification task; detecting a topology change event in real time, triggering an incremental graph updating mechanism to update a local topology; identifying a fault type and locating a fault section based on the constructed topology; establishing a joint loss function to perform end-to-end joint optimization on the topology construction and fault reasoning; and outputting a diagnosis report. The application integrates the topology construction and fault reasoning into a unified framework to realize joint optimization, adapts to dynamic changes of the topology through an incremental updating mechanism, and still maintains high precision and high robustness under the condition of sparse measurement data.

Description

Technical Field

[0001] This invention relates to the field of distribution network operation and maintenance and intelligent analysis technology, specifically to a method for distribution network topology construction and fault reasoning based on graph neural networks. Background Technology

[0002] Distribution network areas are a crucial user-facing link in the power system. Their topology (i.e., the physical connections and hierarchical relationships between transformers, lines, switches, smart meters, and other equipment) forms the foundation for distribution network operation management, state estimation, fault diagnosis, and optimized control. With the large-scale integration of distributed power sources (such as distributed photovoltaics) and the widespread deployment of electric vehicle charging infrastructure, the source-grid-load uncertainty of distribution networks is increasing, electrical coupling characteristics are becoming more complex, and topological forms are becoming more diversified. Compared to transmission networks, distribution networks are characterized by shorter transmission distances, more line branches, more flexible operating modes, and incomplete deployment of measurement equipment, making topology construction and fault diagnosis more challenging.

[0003] Currently, the construction of distribution network topology and fault reasoning mainly rely on the following technical approaches:

[0004] (I) Traditional Topology Identification Methods. Existing methods for topology identification in distribution networks mainly include state estimation, correlation analysis, and data learning. These methods mostly extract data features and analyze node connection relationships through traditional search or sorting methods. However, traditional methods involve a large amount of computation when evaluating all possible topologies, making it difficult to meet real-time requirements and adapt to the complex electrical coupling characteristics under high-proportion distributed power source access.

[0005] (II) Fault Diagnosis Methods Based on Traditional Deep Learning. In recent years, deep learning models, such as Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM), have been widely used in power grid analysis. However, most studies are based solely on data analysis of existing signals, neglecting network topology analysis, which has certain limitations. Furthermore, most deep learning algorithms lack the ability to generalize to topology changes, often requiring retraining when the distribution network topology changes, and their accuracy is low when measurements are insufficient and the labeling rate is low.

[0006] (III) Preliminary Applications of Graph Neural Networks in Power Systems. Graph Neural Networks (GNNs) can effectively capture the relationships between nodes and global information of the graph, providing a new approach to address the shortcomings of traditional deep neural networks in processing graph-structured data. In recent years, researchers have begun to apply GNNs to distribution network topology identification or fault diagnosis. For example, some schemes use graph convolution to aggregate node and structural information of the distribution network for topology identification; others utilize graph attention networks (GAT) to construct fault location models for distribution networks, fully mining fault characteristics by combining the distribution network topology. In addition, active distribution network fault location methods based on spatiotemporal graph convolutional networks have also been proposed, which locate faults by combining the spatiotemporal correlation of distribution network fault information.

[0007] A comprehensive analysis of existing technologies reveals the following shortcomings:

[0008] (1) The separation of topology construction and fault reasoning. Most existing technologies solve topology identification and fault diagnosis as two independent problems. Topology identification focuses on restoring network connectivity, while fault diagnosis focuses on locating faults in a known topology. There is a lack of organic integration and collaborative reasoning mechanism between the two.

[0009] (2) Insufficient adaptability to dynamic topology changes. Distribution networks operate flexibly and their topology changes frequently. Operations such as the integration of distributed energy resources, network reconfiguration, and fault isolation lead to frequent topology changes. Existing AI-based fault location methods lack generalization ability to topology changes, and often require retraining after topology changes.

[0010] (3) Inference capability is weak under incomplete measurement data. Problems such as incomplete deployment of distribution network measurement equipment and data transmission loss lead to sparse measurement data. Existing methods show a significant decrease in topology construction accuracy and fault reasoning accuracy under incomplete data scenarios.

[0011] (4) Lack of a unified end-to-end framework. From topology construction to fault reasoning, existing solutions mostly adopt a phased, multi-model architecture, with prominent issues of information loss and error accumulation in the intermediate links, making it difficult to achieve end-to-end optimization from raw measurement data to fault reasoning results. To address this, we propose a distribution network area topology construction and fault reasoning method based on graph neural networks. Summary of the Invention

[0012] The technical problem to be solved by this invention is to overcome the existing defects and provide a distribution network area topology construction and fault reasoning method based on graph neural network. This method can integrate distribution network area topology construction and fault reasoning into a unified graph neural network framework to achieve end-to-end joint optimization, improve the adaptability to dynamic topology changes and scenarios with incomplete measurement data, and improve the accuracy and real-time performance of topology construction and fault reasoning. This method can effectively solve the problems in the background technology.

[0013] To achieve the above objectives, the present invention provides the following technical solution: a method for topology construction and fault reasoning of distribution network areas based on graph neural networks, comprising the following steps:

[0014] S1: Collect multi-source measurement data of each node in the distribution network area, and abstract the distribution network area into a graph structure based on the multi-source measurement data. The nodes of the graph structure correspond to power components, and the edges correspond to electrical connection relationships.

[0015] S2: Input the graph structure into the multi-task graph neural network model, construct the topology of the transformer area through the link prediction task, and infer the hierarchical relationship of each node in the transformer area through the node hierarchical classification task.

[0016] S3: Real-time detection of topology change events. When a topology change is detected, a graph update mechanism is triggered to update the local topology of the changed region.

[0017] S4: Based on the constructed transformer area topology and multi-source measurement data, identify fault types and locate fault sections;

[0018] S5: Establish a joint loss function and perform end-to-end joint optimization of the topology building task and the fault reasoning task through the joint loss function;

[0019] S6: Outputs a diagnostic report containing topology and fault reasoning results.

[0020] Furthermore, the multi-source measurement data in step S1 includes at least node voltage time-series data, node current time-series data, active power time-series data, reactive power time-series data, and switch status data; the nodes include one or more of the following: transformer nodes, switch nodes, line end nodes, and smart meter nodes.

[0021] Furthermore, in step S1, the graph structure includes a set of nodes, a set of edges, and a node feature matrix. The node feature matrix is ​​composed of multi-source measurement data. The edge set is initialized as follows: when there are historical topology records, an initial candidate topology structure is constructed based on the historical topology records; when there are no historical topology records, the edge set is initialized as a fully connected graph.

[0022] Furthermore, the multi-task graph neural network model in step S2 includes a graph encoding module, a topology decoding module, and a node-level inference module. The graph encoding module uses a graph attention network or a graph isomorphic network as the encoder, and iteratively aggregates the neighbor node information of each node through a message passing mechanism to generate a node embedding representation containing local topology information and electrical measurement information. The topology decoding module predicts the probability of a connection relationship between node pairs based on the node embedding representation to reconstruct the transformer area topology. The node-level inference module outputs the probability distribution of each node belonging to each level category, including transformer level, branch line level, and user meter level.

[0023] Furthermore, the link prediction task and the node-level classification task share the parameters of the graph coding module. The node-level classification task serves as an auxiliary task and is jointly trained with the link prediction task to participate in the parameter update of the graph coding module.

[0024] Furthermore, in step S3, real-time detection of topology change events specifically includes: calculating the statistical feature vector within the sliding window for the measurement data sequence of each node, wherein the statistical feature vector includes at least the mean, variance, skewness, and kurtosis; calculating the distance metric between the current window statistical feature vector and the historical baseline window statistical feature vector; when the distance metric exceeds a preset threshold, it is determined that a topology change event has been detected, and the time of change and the set of affected nodes are recorded.

[0025] Furthermore, the graph update mechanism in step S3 is an incremental graph update mechanism, which specifically includes: taking the set of affected nodes corresponding to the detected topology change event as the center, expanding the neighbor range by a preset number of hops to construct a local subgraph, where the preset number of hops is a positive integer from 1 to 3; performing forward propagation calculation of the graph neural network only on the nodes within the local subgraph range to update the node embedding representation and edge connection probability of each node within that range; the node embedding representation and topology connection relationship outside the local subgraph range in the global graph structure remain unchanged.

[0026] Furthermore, the incremental graph update mechanism also includes a topology confidence assessment step: a confidence score is assigned to each predicted edge. When the confidence score of an edge is lower than a preset confidence threshold, the edge is marked as pending confirmation and is not included in the final topology output until subsequent measurement data is accumulated and reassessed. When the confidence score of an edge is higher than the preset confidence threshold, the edge is marked as confirmed and included in the final topology output.

[0027] Furthermore, in step S4, fault type identification and fault segment localization are implemented using a temporal graph neural network, which includes a spatial graph convolution module and a time series modeling module. The spatial graph convolution module aggregates the spatial neighbor information of each node based on the topological connection relationship of the graph structure, and the time series modeling module captures the temporal evolution characteristics of the fault signal along the time dimension. The spatial graph convolution module and the time series modeling module are stacked alternately to simultaneously capture the spatial topological features and temporal evolution characteristics of the fault signal. The fault type includes any one or more of the following: single-phase ground fault, two-phase short-circuit fault, two-phase short-circuit ground fault, three-phase short-circuit fault, and open-circuit fault.

[0028] Furthermore, in step S5, the joint loss function is obtained by weighted summation of the topology construction loss, the fault inference loss, and the consistency constraint loss between the topology construction result and the fault inference result. The consistency constraint loss is used to constrain the nodes or edges marked as faults in the fault segment location result to form a connected subgraph in the topological connection relationship, and the magnitude of the change in electrical quantity caused by the fault decreases as the topological distance from the fault segment increases. When the topology construction result and the fault inference result violate the consistency constraint, the consistency constraint loss applies a penalty, forcing the model to correct the topology construction and / or fault inference results in joint training to meet physical consistency.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: The distribution network topology construction and fault reasoning method based on graph neural networks has the following advantages:

[0030] 1. This invention integrates topology construction and fault reasoning into a single model using a multi-task graph neural network framework, achieving end-to-end joint optimization and avoiding information loss and error accumulation in traditional staged methods. It organically integrates and mutually promotes the three tasks: topology construction via link prediction, hierarchical relationship reasoning via node-level classification, and fault identification and localization via a temporal graph neural network.

[0031] 2. This invention, through a dynamic topology change detection mechanism and an incremental graph update mechanism, can quickly and adaptively adjust when the topology changes. Only the local subgraph of the changed region is reconstructed, while the rest of the global graph remains unchanged, eliminating the need for complete retraining and significantly reducing computational overhead and response time.

[0032] 3. This invention utilizes the message passing mechanism of graph neural networks to effectively compensate for the lack of local measurement data by leveraging topological information. In scenarios with sparse measurement data, information from adjacent nodes can be supplemented through graph aggregation, while still maintaining high topology construction accuracy and fault reasoning accuracy.

[0033] 4. This invention uses graph neural networks for topology construction and fault reasoning, and the computational complexity is approximately linearly related to the graph size; combined with the incremental update mechanism, it only processes locally changing areas, which can meet the time requirements for real-time operation and maintenance of distribution network areas.

[0034] 5. This invention introduces topology-fault consistency constraint loss, embedding physical laws such as power flow equations as consistency constraints into the loss function, ensuring that the fault reasoning results conform to the physical laws of the power system (faults propagate along the topology, and electrical quantities decay with topological distance), thereby improving the credibility and interpretability of the reasoning results. Attached Figure Description

[0035] Figure 1 This is a flowchart of the workflow of the present invention. Detailed Implementation

[0036] 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, and 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.

[0037] Please see Figure 1 This invention provides a technical solution: a method for topology construction and fault reasoning of distribution network areas based on graph neural networks, comprising the following steps:

[0038] S1: Collect multi-source measurement data of each node in the distribution network area, and abstract the distribution network area into a graph structure based on the multi-source measurement data. The nodes of the graph structure correspond to power components, and the edges correspond to electrical connection relationships. The multi-source measurement data includes at least node voltage time series data, node current time series data, active power time series data, reactive power time series data, and switch status data. The nodes include one or more of the following: transformer nodes, switch nodes, line end nodes, and smart meter nodes.

[0039] Furthermore, in step S1, the graph structure includes a node set, an edge set, and a node feature matrix. The node feature matrix is ​​composed of multi-source measurement data. The edge set is initialized as follows: when historical topology records exist, an initial candidate topology structure is constructed based on the historical topology records; when no historical topology records exist, the edge set is initialized as a fully connected graph.

[0040] S2: Input the graph structure into the multi-task graph neural network model. Construct the transformer area topology through the link prediction task, and infer the hierarchical relationship of each node in the transformer area through the node hierarchical classification task. The multi-task graph neural network model includes a graph encoding module, a topology decoding module, and a node hierarchical inference module. The graph encoding module uses a graph attention network or graph isomorphic network as the encoder. Iteratively aggregates the neighbor node information of each node through a message passing mechanism to generate a node embedding representation containing local topology information and electrical measurement information. The topology decoding module predicts the probability of the existence of connection relationships between node pairs based on the node embedding representation to reconstruct the transformer area topology. The node hierarchical inference module outputs the probability distribution of each node belonging to each hierarchical category. The hierarchical categories include transformer level, branch line level, and user meter level. The link prediction task and the node hierarchical classification task share the parameters of the graph encoding module. The node hierarchical classification task is used as an auxiliary task to jointly train with the link prediction task and participate in the parameter update of the graph encoding module.

[0041] S3: Real-time detection of topology change events. When a topology change is detected, a graph update mechanism is triggered to update the local topology of the changed region. Real-time detection of topology change events specifically includes: calculating the statistical feature vector within a sliding window for the measurement data sequence of each node. The statistical feature vector includes at least the mean, variance, skewness, and kurtosis; calculating the distance metric between the current window's statistical feature vector and the historical baseline window's statistical feature vector. When the distance metric exceeds a preset threshold, a topology change event is detected, and the time of the change and the set of affected nodes are recorded. The graph update mechanism is an incremental graph update mechanism, specifically including: expanding outwards by a preset number of hops from the set of affected nodes corresponding to the detected topology change event as the center. The graph update mechanism constructs a local subgraph with a preset hop count of 1 to 3 positive integers. Forward propagation of the graph neural network is performed only on nodes within the local subgraph, updating the node embeddings and edge connection probabilities of each node within that subgraph. Node embeddings and topological connections outside the local subgraph remain unchanged in the global graph structure. The incremental graph update mechanism also includes a topological confidence evaluation step: a confidence score is assigned to each predicted edge. When the confidence score of an edge is lower than a preset confidence threshold, the edge is marked as pending confirmation and temporarily excluded from the final topological output until re-evaluation after subsequent measurement data accumulation. When the confidence score of an edge is higher than the preset confidence threshold, the edge is marked as confirmed and included in the final topological output.

[0042] S4: Based on the constructed transformer area topology and multi-source measurement data, fault type identification and fault segment location are performed. Fault type identification and fault segment location are implemented using a time-series graph neural network, which includes a spatial graph convolution module and a time series modeling module. The spatial graph convolution module aggregates the spatial neighbor information of each node based on the topological connection relationship of the graph structure, and the time series modeling module captures the temporal evolution characteristics of the fault signal along the time dimension. The spatial graph convolution module and the time series modeling module are stacked alternately to simultaneously capture the spatial topological features and temporal evolution characteristics of the fault signal. The fault types include any one or more of the following: single-phase ground fault, two-phase short-circuit fault, two-phase short-circuit ground fault, three-phase short-circuit fault, and open-circuit fault.

[0043] S5: Establish a joint loss function to perform end-to-end joint optimization of the topology construction task and the fault reasoning task. The joint loss function is obtained by weighted summation of the topology construction loss, the fault reasoning loss, and the consistency constraint loss between the topology construction result and the fault reasoning result. The consistency constraint loss is used to constrain the nodes or edges marked as faults in the fault segment location result to form a connected subgraph in the topological connection relationship, and the magnitude of the change in electrical quantity caused by the fault decreases as the topological distance from the fault segment increases. When the topology construction result and the fault reasoning result violate the consistency constraint, the consistency constraint loss applies a penalty, forcing the model to correct the topology construction and / or fault reasoning results in joint training to meet physical consistency.

[0044] S6: Outputs a diagnostic report containing topology and fault reasoning results.

[0045] The present invention provides the following embodiments:

[0046] Example 1: Typical Distribution Network Area Topology Construction and Fault Reasoning

[0047] This embodiment uses a real 10kV / 380V distribution network area as an example for verification. The area includes one distribution transformer, six branch boxes, 32 smart meters, and two distributed photovoltaic access points, with a total installed capacity of 400kVA and a power supply radius of approximately 800 meters. Multi-source measurement data for 30 days was continuously collected using the area's smart converged terminal and smart meters at a sampling frequency of once every 15 seconds. This data included three-phase voltage, three-phase current, active power, reactive power, and switch status at each node. The first 20 days of data were used as the training set, the middle 5 days as the validation set, and the last 5 days as the test set.

[0048] The graph structure contains 41 nodes. The multi-task graph neural network model uses a 3-layer graph attention network as the encoder, with each hidden layer having a dimension of 128 and 4 attention heads. The topology decoding module uses a bilinear decoder to predict the connection probabilities between nodes. The node-level inference module uses a single-layer fully connected classification head to output three probability distributions: transformer level, branch line level, and user meter level. The model is trained using the Adam optimizer with a learning rate of 0.001 and a total of 300 training epochs.

[0049] Experimental results on the test set show that the overall accuracy of topology construction is 94.7%, and the accuracy of node hierarchical classification is 96.1%. In 250 simulations of 5 different fault types, the model completes fault identification and location within an average of 85 milliseconds, with a fault type identification accuracy of 96.0% and a fault segment location accuracy of 93.0%.

[0050] Ablation experiments show that: removing the node-level inference module reduces the topology construction accuracy to 89.3%; removing the consistency constraint loss results in approximately 7% of physically inconsistent results; and changing to staged independent training reduces the overall F1-score by approximately 5.8 percentage points, verifying the effectiveness of multi-task joint learning and end-to-end optimization.

[0051] Example 2: Adaptive Update in a Dynamic Topology Change Scenario

[0052] This embodiment verifies the adaptive capability of the present invention under dynamic topology changes. Based on a single transformer substation in the embodiment, three consecutive changing events are set: Event 1 is the switching of power supply by the tie switch, with 8 user meters switching to power supply from the adjacent substation; Event 2 is the load switching after fault isolation, with 5 meters switching to the backup feeder; Event 3 is the sequential commissioning and decommissioning of 2 distributed photovoltaic access points due to changes in sunlight.

[0053] Topology change detection uses a 1-second step size and a window size of 60 sampling points (corresponding to 15 minutes). It calculates the mean, variance, skewness, and kurtosis of each node in real time, and determines a change based on the Mahalanobis distance exceeding a preset threshold. In Event 1, the Mahalanobis distance of the affected node exceeded the threshold 8 seconds after the data transfer. After confirmation through three consecutive windows, the model detected the change event within 12 seconds. Event 2 had a detection delay of 9 seconds; Event 3 had a detection delay of 6 seconds.

[0054] Incremental updates construct a local subgraph by expanding outwards by two hops from the affected node. Event 1, a local update involving 14 nodes, takes 320 milliseconds, a reduction of approximately 85% in computation time compared to full graph reconstruction (2.1 seconds). Event 2 takes 210 milliseconds, and Event 3 takes 85 milliseconds. In the confidence assessment, most new edges are immediately confirmed as valid connections, while a small number of edges with lower confidence are confirmed as valid after accumulating data over three time windows. The final updated topology achieves 98.6% consistency with the actual reconstructed topology. In contrast, the traditional GCN model, under the same topology change without retraining, experiences a drop in fault location accuracy from 92% to approximately 71%, while the updated version recovers to 91.5%.

[0055] Example 3: Robustness Verification in Sparse Measurement Data Scenarios

[0056] This embodiment verifies the robustness of the invention under conditions of missing measurement data. Based on the IEEE 33-node distribution network standard test system, measurement missing ratios of 10%, 20%, 30%, and 40% were set respectively. The missing methods included random missing and continuous missing, and Gaussian white noise (signal-to-noise ratio of 20dB) was added to the data.

[0057] Experimental results show that when the random missing value ratio is 20%, the topology construction recall rate is 93.8% and the fault location accuracy rate is 90.1%. When the missing value ratio increases to 30%, the recall rate remains at 90.5% and the fault location accuracy rate is 86.5%. Even under the extreme missing value condition of 40%, the recall rate still reaches 86.7% and the fault location accuracy rate is 81.3%. In the continuous missing value scenario, by aggregating information from adjacent nodes through the message passing mechanism of graph neural networks to complete the missing features, the topology construction recall rate only decreases by 2 to 3 percentage points compared to the random missing value scenario.

[0058] In comparative experiments, the traditional CNN method achieved a fault location accuracy of only 68.5% under conditions of 20% missing data and 20dB noise, while the present invention achieved 90.1%, representing a relative improvement of 31.5%. The accuracy of various fault identification methods remained above 89% under the 20% missing data condition. The superior robustness of this invention is attributed to the message passing mechanism of the graph neural network, which disperses the impact of local missing data; multi-task learning, which provides complementary information; and the consistency constraint loss, acting as physical regularization, which suppresses noise interference.

[0059] Example 4: Verification of generalization ability under different topology scales

[0060] This embodiment verifies the generalization ability of the invention on distribution network topologies of different sizes and complexities. Four test systems were used: an IEEE 13-node system (small), an IEEE 33-node system (medium), an IEEE 123-node system (large), and a real 138-node distribution network system. The model trained on a 41-node distribution area in Embodiment 1 was migrated to the above four systems, with only minor adjustments made to the output layer.

[0061] Migration test results: On an IEEE 13-node system, the topology construction accuracy was 96.8%, and the fault location accuracy was 95.5%; on an IEEE 33-node system, the accuracy was 94.7% and 93.0% respectively; on an IEEE 123-node system, the accuracy was 91.2% and 89.4% respectively; and on a real 138-node system, the accuracy was 90.6% and 88.7% respectively. The results show that even in a large-scale 138-node system, the accuracy remains above 88%, demonstrating the good topology generalization ability of this invention. This generalization ability stems from the fact that the message passing mechanism of the graph neural network does not depend on the specific topology structure, the node features do not contain topology-specific information, and multi-task learning enables the model to master a general topology-electrical mapping relationship.

[0062] Example 5: Edge Computing Platform Deployment and Real-Time Verification

[0063] This embodiment verifies the feasibility of deploying the present invention on a smart terminal in a distribution network area. The hardware platform is a smart converged terminal for distribution network areas (quad-core ARM Cortex-A53 1.5GHz, 2GB memory, NPU computing power 1TOPS). To adapt to the resource constraints of edge devices, the model is subjected to weight quantization (32-bit floating-point compression to 8-bit integer, model size compression from 45MB to 12MB) and graph structure sparsity optimization.

[0064] Performance test results on edge devices: Topology construction for a typical 41-node area took an average of 28 milliseconds, fault inference took an average of 42 milliseconds, and the complete inference process took an average of 68 milliseconds; incremental local updates took an average of 35 milliseconds; 99% percentile latency was within 120 milliseconds, far exceeding the 2-second real-time requirement for distribution network fault handling. After 24 hours of continuous operation, the average power consumption was 3.2W, CPU utilization was approximately 32%, and the NPU speedup was approximately 3.7 times.

[0065] The system operated continuously for 90 days in the actual transformer area, correctly detecting multiple topology change events and completing incremental updates. Several actual faults occurred during this period, and all of them were identified and located within 200 milliseconds after the fault occurred, with the location results consistent with the on-site inspection.

[0066] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for topology construction and fault reasoning of distribution network areas based on graph neural networks, characterized in that, Includes the following steps: S1: Collect multi-source measurement data of each node in the distribution network area, and abstract the distribution network area into a graph structure based on the multi-source measurement data. The nodes of the graph structure correspond to power components, and the edges correspond to electrical connection relationships. S2: Input the graph structure into the multi-task graph neural network model, construct the topology of the transformer area through the link prediction task, and infer the hierarchical relationship of each node in the transformer area through the node hierarchical classification task. S3: Real-time detection of topology change events. When a topology change is detected, a graph update mechanism is triggered to update the local topology of the changed region. S4: Based on the constructed transformer area topology and multi-source measurement data, identify fault types and locate fault sections; S5: Establish a joint loss function and perform end-to-end joint optimization of the topology building task and the fault reasoning task through the joint loss function; S6: Outputs a diagnostic report containing topology and fault reasoning results.

2. The method for topology construction and fault reasoning of distribution network areas based on graph neural networks according to claim 1, characterized in that: The multi-source measurement data in step S1 includes at least node voltage time-series data, node current time-series data, active power time-series data, reactive power time-series data, and switch status data; the nodes include one or more of the following: transformer nodes, switch nodes, line end nodes, and smart meter nodes.

3. The method for topology construction and fault reasoning of distribution network areas based on graph neural networks according to claim 1, characterized in that: In step S1, the graph structure includes a set of nodes, a set of edges, and a node feature matrix. The node feature matrix is ​​composed of multi-source measurement data. The edge set is initialized as follows: when there are historical topology records, an initial candidate topology structure is constructed based on the historical topology records; when there are no historical topology records, the edge set is initialized as a fully connected graph.

4. The method for topology construction and fault reasoning of distribution network areas based on graph neural networks according to claim 1, characterized in that: In step S2, the multi-task graph neural network model includes a graph encoding module, a topology decoding module, and a node-level inference module. The graph encoding module uses a graph attention network or a graph isomorphic network as the encoder, and iteratively aggregates the neighbor node information of each node through a message passing mechanism to generate a node embedding representation containing local topology information and electrical measurement information. The topology decoding module predicts the probability of a connection relationship between node pairs based on the node embedding representation to reconstruct the transformer area topology. The node-level inference module outputs the probability distribution of each node belonging to each level category, including transformer level, branch line level, and user meter level.

5. The method for topology construction and fault reasoning of distribution network areas based on graph neural networks according to claim 4, characterized in that: The link prediction task and the node-level classification task share the parameters of the graph coding module. The node-level classification task serves as an auxiliary task and is jointly trained with the link prediction task to participate in the parameter update of the graph coding module.

6. The method for topology construction and fault reasoning of distribution network areas based on graph neural networks according to claim 1, characterized in that: Step S3 involves real-time detection of topology change events, specifically including: calculating the statistical feature vector within a sliding window for the measurement data sequence of each node, wherein the statistical feature vector includes at least the mean, variance, skewness, and kurtosis; calculating the distance metric between the current window statistical feature vector and the historical baseline window statistical feature vector; determining that a topology change event has been detected when the distance metric exceeds a preset threshold, and recording the time of change and the set of affected nodes.

7. The method for topology construction and fault reasoning of distribution network areas based on graph neural networks according to claim 1, characterized in that: The graph update mechanism in step S3 is an incremental graph update mechanism, which specifically includes: taking the set of affected nodes corresponding to the detected topology change event as the center, expanding the neighbor range by a preset number of hops to construct a local subgraph, where the preset number of hops is a positive integer from 1 to 3; performing forward propagation calculation of the graph neural network only on the nodes within the local subgraph range to update the node embedding representation and edge connection probability of each node within that range; the node embedding representation and topology connection relationship outside the local subgraph range in the global graph structure remain unchanged.

8. The method for topology construction and fault reasoning of distribution network areas based on graph neural networks according to claim 7, characterized in that: The incremental graph update mechanism also includes a topology confidence assessment step: a confidence score is added to each predicted edge. When the confidence score of an edge is lower than the preset confidence threshold, the edge is marked as pending confirmation and is not included in the final topology output until subsequent measurement data is accumulated and reassessed. When the confidence score of an edge is higher than the preset confidence threshold, the edge is marked as confirmed and included in the final topology output.

9. The method for topology construction and fault reasoning of distribution network areas based on graph neural networks according to claim 1, characterized in that: In step S4, fault type identification and fault segment localization are implemented using a temporal graph neural network, which includes a spatial graph convolution module and a time series modeling module. The spatial graph convolution module aggregates the spatial neighbor information of each node based on the topological connection relationship of the graph structure, and the time series modeling module captures the temporal evolution characteristics of the fault signal along the time dimension. The spatial graph convolution module and the time series modeling module are stacked alternately to simultaneously capture the spatial topological features and temporal evolution characteristics of the fault signal. The fault types include any one or more of the following: single-phase ground fault, two-phase short-circuit fault, two-phase short-circuit to ground fault, three-phase short-circuit fault, and open-circuit fault.

10. The method for topology construction and fault reasoning of distribution network areas based on graph neural networks according to claim 1, characterized in that: In step S5, the joint loss function is obtained by weighted summation of topology construction loss, fault inference loss, and consistency constraint loss between topology construction results and fault inference results. The consistency constraint loss is used to constrain the nodes or edges marked as faults in the fault segment location results to form a connected subgraph in the topological connection relationship, and the magnitude of the change in electrical quantity caused by the fault decreases as the topological distance from the fault segment increases. When the topology construction results and fault inference results violate the consistency constraint, the consistency constraint loss applies a penalty, forcing the model to correct the topology construction and / or fault inference results in joint training to meet physical consistency.

Images