A cloud-edge collaborative system and method for real-time identification and fault location of power distribution network topology

Through the collaborative work of the perception layer, edge layer and cloud platform layer of the cloud-edge collaborative architecture, real-time identification of distribution network topology and accurate fault location are achieved, solving the problems of high communication bandwidth pressure, insufficient real-time fault location and weak adaptive capability when topology changes in existing technologies, and possessing adaptive capability to topology changes.

CN122490381BActive Publication Date: 2026-11-17BEIJING SMARTCHIP MICROELECTRONICS TECHNOLOGY CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202610976840.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-02
Publication Date
2026-11-17
Estimated Expiration
2046-07-02

AI Technical Summary

Technical Problem

Existing fault location methods for power distribution networks suffer from problems such as high communication bandwidth pressure, insufficient real-time fault location, and weak system adaptability during topology changes.

Method used

A cloud-edge collaborative architecture is adopted. Data is collected through the perception layer, lightweight feature extraction and local anomaly detection are performed at the edge layer, and lightweight compression and dynamic topology inference of the graph neural network model are performed at the cloud platform layer. Fault location is achieved by combining graph search algorithm and a triggered incremental update mechanism is adopted to respond to topology changes.

Benefits of technology

It enables real-time identification of distribution network topology and accurate fault location, reduces communication bandwidth pressure and cloud computing burden, and has the ability to adapt to topology changes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122490381B_ABST
    Figure CN122490381B_ABST
Patent Text Reader

Abstract

This invention relates to the field of distribution network fault detection, providing a cloud-edge collaborative system and method for real-time distribution network topology identification and fault location. It includes a sensing layer, an edge layer, and a cloud platform layer. The sensing layer collects raw electrical quantity data; the edge layer performs lightweight feature extraction and local anomaly detection on the raw electrical quantity data uploaded by the sensing layer; the cloud platform layer deploys a graph neural network model, performs real-time global topology inference based on the extracted lightweight feature data to obtain a dynamic distribution network topology structure, and uses a graph search algorithm to locate the fault location based on the distribution network topology structure. This invention employs a collaborative approach of lightweight feature extraction by an edge-side lightweight model, dynamic topology inference by a graph neural network model on the cloud platform, and triggered incremental updates of model parameters to achieve real-time identification of distribution network topology and accurate fault location. It also enables the system to adapt to topology changes while significantly reducing communication bandwidth pressure and cloud computing burden.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of distribution network fault detection, specifically to a cloud-edge collaborative distribution network topology real-time identification and fault location system and method. Background Technology

[0002] Fault location and topology identification in power distribution networks are key technologies for ensuring the safe and stable operation of power systems. Currently, the existing technical approaches in this field mainly include the following categories: One approach is a centralized fault location method. The data acquisition terminal uploads all raw electrical quantity data to the cloud, relying on static topology files for centralized calculation. This method results in enormous pressure on communication bandwidth, an excessive burden on cloud computing, failure to meet real-time fault response delays, and an inability to adapt to dynamic topology changes, leading to a significant decrease in location accuracy.

[0003] Second, there are static fault location methods based on graph theory. These methods achieve automated location based on graph theory analysis, but all analysis is completed on a single node. They cannot utilize the edge to achieve local preprocessing and real-time early warning, and all raw data must be transmitted to the cloud, which does not alleviate the communication pressure. After topology changes, all data needs to be re-identified, resulting in high computational overhead. Fault location and topology inference are independent of each other, and real-time topology can not be used to improve location accuracy.

[0004] Thirdly, there are existing cloud-edge collaborative fault detection methods. These methods deploy time-series analysis models at the edge for rapid detection and use spectral analysis for fault location. However, the location accuracy is limited, and the time-series analysis models at the edge have weak dynamic adaptability to changes in the distribution network topology.

[0005] In summary, existing technologies lack technical solutions that can simultaneously address a series of issues such as high communication bandwidth pressure, insufficient real-time and accuracy of fault location, and weak system adaptability during topology changes. Summary of the Invention

[0006] To address the problems of existing technologies, this invention provides a cloud-edge collaborative distribution network topology real-time identification and fault location system and method, which can realize real-time topology identification and accurate fault location in dynamically changing distribution networks, and enable the system to adapt to topology changes, while significantly reducing communication bandwidth pressure and cloud computing burden.

[0007] The first aspect of the present invention provides a cloud-edge collaborative distribution network topology real-time identification and fault location system, comprising: a sensing layer, an edge layer and a cloud platform layer; The sensing layer includes terminal equipment deployed in the power distribution network for collecting raw electrical quantity data; The edge layer includes an intelligent fusion terminal deployed in the distribution substation and a lightweight edge model, which is used to perform lightweight feature extraction and local anomaly detection on the raw electrical quantity data uploaded by the perception layer, and upload the extracted lightweight feature data and the anomaly data obtained from the local anomaly detection to the cloud platform layer. The cloud platform layer is equipped with a graph neural network model and a fault location engine. The graph neural network model is used to perform real-time global topology inference based on lightweight feature data uploaded from all edge layers to obtain a dynamic distribution network topology. The fault location engine is used to locate the fault location based on the abnormal data uploaded from the edge layers and the dynamic distribution network topology using a graph search algorithm. The edge-side lightweight model is deployed to the edge layer after the cloud platform layer performs lightweight compression based on the graph neural network model. The update parameters of the edge-side lightweight model in the edge layer are obtained by the cloud platform layer after updating the graph neural network model under the condition of event triggering and then sent to the edge layer.

[0008] In this embodiment of the invention, the edge-side lightweight model performs dimensionality reduction feature extraction on the time series data obtained based on the original electrical quantity data to obtain lightweight feature data.

[0009] In this embodiment of the invention, the edge-side lightweight model performs dimensionality reduction feature extraction on time series data obtained based on the original electrical quantity data, including: using principal component analysis to project the time series data of each time window onto the first k principal components to obtain a k-dimensional feature vector; or, using a lightweight autoencoder network to compress the time series data into a k-dimensional feature vector; where k takes the value of 3 to 10.

[0010] In this embodiment of the invention, the intelligent fusion terminal has a built-in anomaly detection module. The anomaly detection module determines in real time whether there are voltage sags, overcurrents, or harmonic distortion anomalies based on the original electrical quantity data or lightweight feature data. When an anomaly such as voltage sag, overcurrent, or harmonic distortion is detected, the original electrical quantity data or corresponding lightweight feature data within the time window before and after the anomaly occurs are uploaded to the cloud platform layer as the anomaly data for local anomaly detection.

[0011] In this embodiment of the invention, the graph neural network model of the cloud platform layer is a spatiotemporal graph convolutional network, and each edge layer serves as a node of the spatiotemporal graph convolutional network; The spatiotemporal graph convolutional network takes lightweight feature data of all edge layer nodes within the most recent time window as input, outputs a connection probability matrix between nodes, and obtains a radial power distribution network topology diagram after post-processing.

[0012] In this embodiment of the invention, the fault location engine uses a graph search algorithm based on a dynamic distribution network topology to locate the fault location, specifically including: When abnormal data is received from the edge layer, a list of all abnormal nodes that have experienced voltage drops is collected. Extract the current distribution network topology tree; Calculate the nearest common ancestor node of the list of abnormal nodes in the current distribution network topology tree; Search downstream from the nearest common ancestor node and determine the branch where the fault is located based on the abnormal distribution of voltage drops in the child nodes.

[0013] In this embodiment of the invention, the specific implementation method of the cloud platform layer to obtain a lightweight edge model by lightweight compression of the graph neural network model includes: the cloud platform layer trains the graph neural network model, and compresses the trained graph neural network model into a lightweight model through knowledge distillation, quantization and pruning, and the lightweight model is deployed to the edge layer as a lightweight edge model.

[0014] In this embodiment of the invention, the cloud platform layer retrains the graph neural network model under the condition of event triggering, performs lightweight compression on the retrained graph neural network model to obtain an updated edge-side lightweight model, and sends the parameters of the updated edge-side lightweight model to the edge layer through a standardized interface. After receiving the updated parameters of the edge-side lightweight model, the edge layer replaces the parameters of the current edge-side lightweight model with the parameters of the updated edge-side lightweight model.

[0015] In this embodiment of the invention, the cloud platform layer retrains the graph neural network model under the condition of event triggering, including: when the cloud platform layer detects that the statistical distribution of the lightweight feature data uploaded by the edge layer has drifted to exceed a preset threshold, or when the cloud platform layer determines, based on the on-site verification feedback data, that the detection accuracy of the current edge-side lightweight model of the edge layer on a specific anomaly type has dropped to exceed a preset threshold, the graph neural network model is retrained.

[0016] In this embodiment of the invention, the cloud platform layer adopts a triggered incremental update mechanism to dynamically update the distribution network topology. Specifically, when the cloud platform layer detects a switch change event, it determines the area of ​​influence centered on the switch, extracts lightweight feature data of all nodes within the area of ​​influence, runs a local graph neural network model to re-infer the sub-topology to obtain the updated local sub-graph, and merges the updated local sub-graph with the original global topology to form a new distribution network topology diagram.

[0017] In this embodiment of the invention, the cloud platform layer is further used to send the located fault location to the operation and maintenance terminal, use the actual fault location fed back by the operation and maintenance terminal as a labeled sample to retrain the graph neural network model, and iteratively update the edge-side lightweight model corresponding to the graph neural network model.

[0018] In this embodiment of the invention, the fault location engine of the cloud platform layer is also used to output a fault location result including the fault location and confidence level, wherein the confidence level is calculated based on at least one of the following factors: the number of voltage sag nodes, the consistency of abnormal amplitude, and the clarity of topology branches.

[0019] A second aspect of this invention provides a method for real-time identification and fault location of distribution network topology using cloud-edge collaboration, comprising: Raw electrical quantity data is collected through a sensing layer, which includes terminal equipment deployed in the power distribution network; Lightweight feature extraction and local anomaly detection are performed on the raw electrical quantity data uploaded from the perception layer by using a lightweight edge-side model and intelligent fusion terminal deployed at the edge layer. The extracted lightweight feature data and the anomaly data obtained from local anomaly detection are then uploaded to the cloud platform layer. By deploying a graph neural network model at the cloud platform layer, global topology inference is performed in real time based on lightweight feature data uploaded from all edge layers, resulting in a dynamic distribution network topology. The fault location engine deployed at the cloud platform layer locates the fault location based on the abnormal data uploaded from the edge layer and the dynamic distribution network topology using a graph search algorithm. The edge-side lightweight model is deployed to the edge layer after the cloud platform layer performs lightweight compression based on the graph neural network model. The update parameters of the edge-side lightweight model in the edge layer are obtained by the cloud platform layer after updating the graph neural network model under the condition of event triggering and then sent to the edge layer.

[0020] In this embodiment of the invention, the lightweight feature extraction and local anomaly detection of the raw electrical quantity data uploaded by the perception layer through the edge-side lightweight model and intelligent fusion terminal deployed at the edge layer includes: Dimensionality reduction feature extraction is performed on time series data obtained from raw electrical quantity data to obtain lightweight feature data; The built-in anomaly detection module of the intelligent fusion terminal can determine in real time whether there are voltage sags, overcurrents or harmonic distortion anomalies based on raw electrical quantity data or lightweight characteristic data.

[0021] In this embodiment of the invention, the graph neural network model deployed through the cloud platform layer performs real-time global topology inference based on the lightweight feature data uploaded by all edge layers. This includes: using a spatiotemporal graph convolutional network, treating each edge layer as a node of the spatiotemporal graph convolutional network, using the lightweight feature data of all edge layer nodes within the most recent time window as the input of the spatiotemporal graph convolutional network, outputting the connection probability matrix between nodes through the spatiotemporal graph convolutional network, and obtaining a radial distribution network topology diagram after post-processing.

[0022] In this embodiment of the invention, the fault location engine deployed through the cloud platform layer locates the fault location based on the abnormal data uploaded from the edge layer and a graph search algorithm based on the dynamic distribution network topology, including: When abnormal data is received from the edge layer, a list of all abnormal nodes that have experienced voltage drops is collected. Extract the current distribution network topology tree; Calculate the nearest common ancestor node of the list of abnormal nodes in the current distribution network topology tree; Search downstream from the nearest common ancestor node and determine the branch where the fault is located based on the abnormal distribution of voltage drops in the child nodes.

[0023] In this embodiment of the invention, the cloud-edge collaborative distribution network topology real-time identification and fault location method further includes: dynamically updating the distribution network topology structure through a triggered incremental update mechanism at the cloud platform layer. Specifically, when the cloud platform layer detects a switch change event, it determines the influence area centered on the switch, extracts lightweight feature data of all nodes within the influence area, runs a local graph neural network model to re-infer the sub-topology to obtain an updated local subgraph, and merges the updated local subgraph with the original global topology graph to form a new distribution network topology structure graph.

[0024] This invention is based on a three-layer collaborative architecture of perception layer, edge layer and cloud platform. It adopts a collaborative approach of using a lightweight edge-side model of edge layer to extract lightweight features, graph neural network model of cloud platform for dynamic topology inference, and trigger-based incremental update of model parameters to achieve real-time identification of distribution network topology and accurate fault location. It also enables the system to adapt to topology changes, while significantly reducing communication bandwidth pressure and cloud computing burden.

[0025] Other features and advantages of the technical solution of the present invention will be described in detail in the following detailed embodiments section. Attached Figure Description

[0026] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is an architecture diagram of the cloud-edge collaborative distribution network topology real-time identification and fault location system provided in this embodiment of the invention; Figure 2 This is a flowchart of the cloud-edge collaborative distribution network topology real-time identification and fault location method provided in the embodiments of the present invention. Detailed Implementation

[0027] To make the technical solutions and advantages of the embodiments of the present invention clearer, the exemplary embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0028] As described in the background section, existing methods for fault location and topology identification in power distribution networks suffer from problems such as high communication bandwidth pressure, insufficient real-time fault location, and weak system adaptability during topology changes.

[0029] To address the aforementioned issues, this invention provides a cloud-edge collaborative system and method for real-time topology identification and fault location in distribution networks. The system comprises a perception layer, an edge layer, and a cloud platform layer. The perception layer includes terminal devices deployed in the distribution network for collecting raw electrical quantity data. The edge layer includes intelligent fusion terminals deployed in distribution substations and a lightweight edge model for performing lightweight feature extraction and local anomaly detection on the raw electrical quantity data uploaded by the perception layer, and uploading the extracted lightweight feature data and the anomaly data detected locally to the cloud platform layer. The cloud platform layer deploys a graph neural network model and a fault location engine. The graph neural network model performs real-time global topology inference based on the lightweight feature data uploaded by all edge layers to obtain a dynamic distribution network topology. The fault location engine uses a graph search algorithm based on the anomaly data uploaded by the edge layers and the dynamic distribution network topology to locate the fault location, and pushes the located fault location to the operation and maintenance terminal. In this invention, the edge-side lightweight model is deployed to the edge layer after being lightweighted and compressed from the graph neural network model by the cloud platform layer. The update parameters of the edge-side lightweight model are obtained by the cloud platform layer after updating the graph neural network model under event-triggered conditions and then distributed to the edge layer, thus maintaining the synergy between the edge-side lightweight model in the edge layer and the graph neural network model in the cloud platform layer. Based on a three-layer collaborative architecture of the perception layer, edge layer, and cloud platform, this invention employs a collaborative approach of extracting lightweight features from the edge-side lightweight model in the edge layer, dynamic topology inference from the graph neural network model in the cloud platform, and trigger-based incremental updates of model parameters. This enables real-time identification of distribution network topology and accurate fault location, and gives the system the ability to adapt to topology changes, while significantly reducing communication bandwidth pressure and cloud computing burden.

[0030] The lightweight model on the edge side of the edge layer performs dimensionality reduction feature extraction on the time series data obtained from the original electrical quantity data, resulting in lightweight feature data. Specifically, principal component analysis can be used to project the time series data of each time window onto the first k principal components, obtaining a k-dimensional (low-dimensional) feature vector; or, a lightweight autoencoder network can be used to compress the time series data into a k-dimensional (low-dimensional) feature vector, where k takes the value of 3 to 10.

[0031] The intelligent fusion terminal has a built-in anomaly detection module. This module determines in real time whether there are voltage sags, overcurrents, or harmonic distortion anomalies based on raw electrical quantity data or lightweight feature data. When a voltage sag, overcurrent, or harmonic distortion anomaly is detected, an event upload mechanism is triggered. The raw electrical quantity data or corresponding lightweight feature data within the time window before and after the anomaly occurs are uploaded to the cloud platform layer as the local anomaly detection data.

[0032] The graph neural network model of the cloud platform layer is a spatiotemporal graph convolutional network (ST-GCN), with each edge layer serving as a node in the spatiotemporal graph convolutional network. The spatiotemporal graph convolutional network takes lightweight feature data of all edge layer nodes within the most recent time window as input, outputs a connection probability matrix between nodes, and after post-processing, obtains a radial distribution network topology diagram, which is then stored in a real-time topology database.

[0033] The fault location engine uses a graph search algorithm based on the dynamic distribution network topology to locate the fault location. Specifically, it includes: when receiving abnormal data uploaded from the edge layer, collecting a list of all abnormal nodes that have experienced voltage drops; extracting the current distribution network topology tree from the real-time topology graph database; calculating the nearest common ancestor (LCA) node of the abnormal node list in the current distribution network topology tree; and searching downstream from the LCA node to determine the branch where the fault is located based on the abnormal distribution of voltage drops in the child nodes.

[0034] The specific implementation of lightweight compression of graph neural network models by the cloud platform layer to obtain lightweight models on the edge side includes: the cloud platform layer trains the graph neural network model, and compresses the trained graph neural network model into a lightweight model through knowledge distillation, quantization and pruning. This lightweight model is then deployed to the edge layer as a lightweight model on the edge side.

[0035] The cloud platform layer retrains the graph neural network model under event-triggered conditions. It then performs lightweight compression on the retrained model to obtain an updated lightweight edge-side model. The parameters of this updated model, along with its version number, are sent to the edge layer via a standardized interface. Upon receiving the updated parameters, the edge layer replaces the parameters of its current lightweight edge-side model with those of the updated model. Specifically, the cloud platform layer retrains the graph neural network model when it detects a shift in the statistical distribution of the lightweight feature data uploaded by the edge layer exceeding a preset threshold, or when the cloud platform layer determines, based on on-site verification feedback data, that the detection accuracy of the current lightweight edge-side model for a specific anomaly type has decreased to exceed a preset threshold.

[0036] The cloud platform layer employs a triggered incremental update mechanism to dynamically update the distribution network topology in response to changes in the distribution network topology. When the cloud platform layer detects a switch change event (a change in the distribution network topology), it does not perform a full global topology re-inference. Instead, it determines the affected area centered on the switch, extracts lightweight feature data of all nodes within the affected area, runs a local graph neural network model to re-infer the sub-topology to obtain an updated local subgraph, and merges the updated local subgraph with the original global topology to form a new distribution network topology diagram.

[0037] After locating the fault location, the cloud platform layer sends the fault location to the operation and maintenance terminal. The actual fault location reported by the operation and maintenance terminal is stored in the database as a labeled sample. The labeled sample is used to retrain the graph neural network model and iteratively update the corresponding edge-side lightweight model.

[0038] The fault location engine at the cloud platform layer outputs fault location results including fault location and confidence level. The confidence level is calculated based on at least one of the following factors: number of voltage sag nodes, consistency of abnormal amplitude, and clarity of topology branches. The confidence level is marked with different colors on the operation and maintenance terminal APP to assist the operation and maintenance terminal and operation and maintenance personnel in decision-making.

[0039] Figure 1 This is an architecture diagram of the cloud-edge collaborative distribution network topology real-time identification and fault location system provided in an embodiment of the present invention. Figure 1 As shown, the cloud-edge collaborative distribution network topology real-time identification and fault location system provided in this embodiment adopts a three-layer collaborative architecture of perception layer (smart meter / terminal), edge layer (smart fusion terminal of distribution area) and cloud platform layer (cloud / master station).

[0040] The perception layer consists of terminal devices such as smart meters deployed on the user side and feeder side. These terminal devices collect raw electrical quantity data such as voltage, current, and power at a fixed frequency (such as per second or per cycle) and upload them to the edge computing gateway of their respective transformer area through local communication networks (such as RS485, power line carrier, and low-power wireless).

[0041] Terminal devices (such as smart meters) are used to collect electrical quantity data at high frequencies. The sampling frequency is at least one cycle (50 points per second for a 50Hz system) or higher (e.g., 1024 points / cycle for power quality analysis). The collected electrical quantity data includes: three-phase voltage amplitude V, phase angle θ, active power P, reactive power Q, current I, and event logs (such as voltage sags, overcurrent, etc.). Communication protocols such as Modbus, DL / T645, and IEC 61850 are used between the terminal devices and the edge computing gateway. The terminal devices upload the collected electrical quantity data to the edge computing gateway via RS485 or a wireless module.

[0042] The edge layer consists of intelligent fusion terminals (i.e., edge computing nodes) deployed in distribution transformer substations or switching stations. These intelligent fusion terminals are responsible for aggregating data from all terminals within their jurisdiction, performing local storage, preprocessing, feature extraction, and anomaly detection, and uploading only the refined high-value features or fault events to the cloud. Simultaneously, the edge layer receives lightweight edge-side models from the cloud and dynamically updates the local models.

[0043] The intelligent fusion terminal has the following built-in modules: data preprocessing and caching module, lightweight feature extraction model (such as PCA / natural coding / decision tree), anomaly detection module (i.e., local anomaly detection and event triggering module), feature compression and uploading module, and lightweight model receiving and updating module.

[0044] The data preprocessing and caching module receives raw measurement data reported by the sensing terminals under its jurisdiction, performs time alignment (eliminating clock differences between terminals), outlier removal (such as data exceeding the physical measurement range), and missing value imputation (filling short-term missing values ​​using linear interpolation or historical averages). The cleaned data is then cached in a local database (such as the time-series database InfluxDB) in time-series format. Data preprocessing can employ methods such as Kalman filtering and sliding window filtering.

[0045] Lightweight feature extraction models (i.e., edge-side lightweight models) perform dimensionality reduction and feature extraction on time-series data obtained from preprocessing raw electrical quantity data to produce low-dimensional feature vectors. This reduces the amount of data uploaded to the cloud platform layer while retaining key information required for fault diagnosis and topology inference. Lightweight feature extraction models can employ linear dimensionality reduction models based on PCA (Principal Component Analysis), nonlinear dimensionality reduction models based on autoencoders, or anomaly feature extraction models based on decision trees.

[0046] The PCA-based linear dimensionality reduction model is trained using historical normal operation data. The original data (such as three-phase voltage and current) of each time window (e.g., 1 second) are projected onto the first k principal components to obtain a k-dimensional feature vector. The value of k is usually 3 to 10, and the compression ratio can reach more than 100:1.

[0047] The autoencoder-based nonlinear dimensionality reduction model employs a lightweight autoencoder network (such as a 3-layer fully connected network with a hidden layer node count much smaller than the input dimension). After training in the cloud, the data is compressed and distributed to the edge layer, where the autoencoder network is used to compress the original data into low-dimensional feature vectors. This approach better preserves nonlinear features, but requires slightly more computation.

[0048] Anomaly feature extraction models based on decision trees can be deployed using binary classification decision trees (such as CART). The input is statistical features within a short time window (such as RMS values, harmonic content, negative order components, etc.), and the output is the anomaly probability, thereby achieving fault detection. These decision tree-based anomaly feature extraction models can be trained in the cloud and then deployed to the edge layer.

[0049] The lightweight feature extraction model receives a fixed-length sliding window of data (such as sampling points for 10 consecutive power frequency cycles) and outputs a fixed-length feature vector F (dimension d) in JSON or Protobuf format, which includes timestamp, terminal ID, and feature vector values.

[0050] The local anomaly detection and event triggering module is used to determine whether there are any anomalies (such as voltage drops, overcurrent, harmonic distortion, etc.) based on extracted feature vectors or directly on rules. If an anomaly is detected, the event upload mechanism is triggered.

[0051] The anomaly detection module (i.e., the local anomaly detection and event triggering module) employs lightweight threshold judgment or isolated forest anomaly detection algorithms to determine in real time whether voltage sags, overcurrents, or harmonic distortion anomalies exist based on raw electrical quantity data or lightweight feature data. When a voltage sag, overcurrent, or harmonic distortion anomaly is detected, an event upload mechanism is triggered. The raw electrical quantity data or corresponding lightweight feature data (along with the anomaly type label) within the time window before and after the anomaly occurs is uploaded to the cloud platform layer as the local anomaly detection data. For example, rule triggering: if the effective voltage value is lower than 90% of the rated value for more than two cycles, it is determined to be a voltage sag event. Model triggering: the feature vector is input into the anomaly classification model (such as logistic regression or a small neural network) distributed in the cloud, and the anomaly probability is output; if it exceeds a threshold, it is triggered. Upload decision: the content to be uploaded is determined based on the importance level of the triggered event. Normal operating conditions: only compressed feature vectors are uploaded periodically (e.g., every 15 minutes) for long-term trend analysis and model updates in the cloud. Abnormal events: Immediately upload the raw data or compressed features of a time window before and after the occurrence of the abnormality (e.g., 2 seconds before the event and 3 seconds after the event), and attach an event type label.

[0052] The feature compression and upload module is used to compress and encode the data to be uploaded (feature vectors or original fragments) and upload it to the cloud via a secure communication protocol (such as MQTT with TLS). Specifically, lightweight compression algorithms such as Zstandard or Snappy can be used to further reduce the amount of data transmitted.

[0053] The lightweight model receiving and updating module listens for model update commands sent from the cloud, receives new lightweight model files, and replaces the current local model (such as a lightweight feature extraction model, anomaly detection model, etc.), enabling dynamic upgrades of the edge-side model and ensuring a smooth transition between old and new models. During the initial update of the edge-side lightweight model, all parameters and structure are updated. Subsequent updates are made based on actual conditions; if the model structure changes, both parameters and structure are updated.

[0054] The cloud platform layer can be a distribution automation master station or a public / private cloud platform. It aggregates edge feature data from the entire network, runs complex graph neural network models for global topology inference, and establishes a real-time topology map. Based on real-time topology and voltage anomaly information, it uses graph search algorithms to accurately locate faulty sections and pushes the location results to on-site maintenance terminals. The cloud platform layer is also responsible for offline training based on accumulated data, optimizing the graph neural network model, and then distributing the updated graph neural network model, after it has been lightweighted, to the edge layer.

[0055] The internal modules of the cloud platform layer (cloud / main site) include: data receiving and decompression module, full feature database, complex GNN topology inference model, real-time topology graph database, fault location engine, location result distribution module, on-site verification and feedback module, offline data and model update module, model lightweight compression module, and model distribution module.

[0056] The data receiving and decompression module receives data packets uploaded by all intelligent fusion terminals, decompresses and decrypts them, and parses them into structured data, storing it in the full feature database. The full feature database uses a distributed time-series database (such as TimescaleDB) to store feature vectors of all nodes, abnormal event records, historical fault information, on-site verification feedback, and other information. The complex GNN topology inference model infers the distribution network topology in real time based on the node feature data of the entire network. Specifically, a pre-trained graph neural network (such as ST-GCN) model is deployed and runs periodically (e.g., every minute) or in an event-triggered manner. The input is the feature vectors of all nodes within the most recent time window (e.g., 5 minutes), and the output is the inter-node connection probability matrix. After post-processing, a real-time radial topology map is obtained and stored in the real-time topology map database. When a topology change event (such as a switch change) is detected, the topology is re-inferred only for the affected local area to achieve incremental updates of the topology structure.

[0057] The real-time topology graph database is used to store the currently inferred distribution network topology (graph structure), including node connection relationships, line parameter estimates, etc., and uses a graph database or relational database to store the adjacency list.

[0058] Upon receiving a fault event, the fault location engine quickly locates the faulty branch based on the real-time topology graph and reported abnormal node information. Specifically, it employs the Lowest Common Ancestor (LCA) algorithm based on graph theory for location, as follows: When a fault event is triggered, all relevant smart fusion terminals report a list S of nodes experiencing voltage drops (which may include multiple meter nodes); the fault location engine extracts the current topology from the real-time topology graph database, constructing a real-time topology tree with the distribution network as a tree structure rooted at the substation outlet; for all nodes in the node list S, its Lowest Common Ancestor (LCA) node in the topology tree is calculated, and this node is the upstream branch point where the fault occurred; searching downstream from the LCA node, if all child nodes of a branch experience voltage drops, then that branch is the faulty branch; if some child nodes are normal, the fault point is located between the branch point and the first abnormal child node; the fault location information is output along with a confidence level. Algorithms such as tree doubling are used to achieve an O(logN) LCA query, meeting real-time requirements.

[0059] The fault location result distribution module pushes the fault location results to the on-site terminals (such as mobile apps and tablets) of relevant maintenance personnel via mobile communication networks (4G / 5G), including the fault location, fault type, and suggested handling measures. After on-site verification by maintenance personnel, the actual fault information (such as fault cause, actual fault point, and handling result) is fed back through the on-site terminal. This information is sent back to the cloud and stored in the database of the on-site verification and feedback module for subsequent model optimization and accuracy assessment.

[0060] The offline training and model update module utilizes accumulated historical data and review feedback to periodically retrain the GNN topology inference model and the edge-side lightweight model, thereby improving model performance.

[0061] The model lightweight compression module compresses trained complex models (such as the GNN neural network model shown in the figure) into lightweight models suitable for edge deployment (such as MobileNet-like structures or binary neural networks) through knowledge distillation, quantization, pruning and other techniques, ensuring that its computational and storage space meet the resource constraints of intelligent fusion terminals.

[0062] The model distribution module distributes the updated lightweight model to each intelligent fusion terminal via protocols such as MQTT, along with a version number, triggering model updates on the edge side.

[0063] The cloud platform layer employs a triggered incremental update mechanism to dynamically update the distribution network topology in response to topology changes, avoiding full recalculation and improving system response speed. The triggered incremental update has two methods: Method 1, direct reporting of switching quantities: Intelligent switching devices (such as circuit breakers and load switches) report change signals to the cloud via intelligent fusion terminals when they operate; Method 2, indirect inference from measurement data: When the cloud-based GNN model periodically infers a sudden change in the correlation pattern of nodes in a certain area, it can infer a possible topology change and trigger local re-inference.

[0064] The local re-inference process includes: taking the switchgear that has changed position as the center, determining the set of nodes within a certain range upstream and downstream (e.g., within 3 hops) to determine the affected area; extracting the feature data of all nodes in the area within the most recent time window from the full feature database to achieve local data extraction; local GNN inference: running a lightweight graph neural network (GNN) model (or utilizing the local subgraph capability of the global model) to perform topological inference only on the affected area, updating the local connectivity, and achieving local GNN inference; splicing the locally updated subgraph with the original global topology (topology fusion) to form a new global real-time topology graph; if the topology change causes a significant change in the feature pattern, the cloud can trigger a fine-tuning update of the lightweight model of the affected intelligent fusion terminal (e.g., using transfer learning), and distribute the updated model to achieve edge model linkage.

[0065] The main algorithm flow of this system is as follows: System initialization phase: 1. The cloud pre-trains a GNN topology inference model and various lightweight models (autoencoders, decision trees, etc.) based on historical data (including simulation data); 2. The cloud performs lightweight compression (quantization, pruning) on ​​complex models to generate edge-deployable versions; 3. The cloud distributes the lightweight models to each intelligent fusion terminal through the cloud model distribution interface; 4. The intelligent fusion terminal loads the model and begins local real-time processing.

[0066] Normal operation phase: 5. The perception layer terminal continuously collects data and reports it to the intelligent fusion terminal; 6. The intelligent fusion terminal preprocesses the data and extracts features, and periodically uploads the feature vectors to the cloud through the edge feature upload interface; 7. The cloud periodically runs the GNN model, infers the global topology based on all features, and stores it in the real-time topology database.

[0067] Fault event triggering phase: 8. The local anomaly detection module of the intelligent fusion terminal detects an anomaly (such as a sudden voltage drop) and triggers event upload; 9. The intelligent fusion terminal caches the original data before and after the event and uploads it to the cloud along with the feature vector through the edge feature upload interface; 10. The cloud fault location engine is awakened and performs the following: a. Obtain the current topology from the real-time topology database; b. Collect a list of all abnormal nodes; c. Calculate the lowest common ancestor (LCA) and locate the faulty branch; 11. The cloud pushes the location results to the relevant operation and maintenance terminals through the fault location result distribution interface.

[0068] On-site handling and feedback phase: 12. Maintenance personnel verify on-site and provide feedback on the actual fault information through the verification feedback interface of the maintenance terminal; 13. Feedback data is stored in the cloud database for subsequent model training.

[0069] Model update phase: 14. The cloud periodically (e.g., weekly) retrains the model using newly accumulated data (including verification labels); 15. After compression, the new model is distributed to the intelligent fusion terminal via the cloud model distribution interface to update the local model; 16. After the intelligent fusion terminal completes the update, the system enters the next cycle.

[0070] Topology change handling: 17. When a switch change is detected (direct reporting or indirect inference), the cloud triggers local re-inference: a. Determine the affected area; b. Extract the features of the regional nodes; c. Run the local GNN to update the sub-topology; d. Merge to form a new global topology; 18. If necessary, the cloud fine-tunes the lightweight model of the affected intelligent fusion terminal and distributes it.

[0071] The intelligent fusion terminal at the edge layer communicates with the cloud platform layer using standardized interface protocols, including: an edge feature upload interface, a cloud model distribution interface, a fault location result distribution interface, and an on-site verification and feedback interface. The edge feature upload interface uses the MQTT over TLS protocol and Protobuf format to upload lightweight feature data and anomaly data. The cloud model distribution interface transmits update parameters and version information of the lightweight model on the edge side. The fault location result distribution interface uses HTTP / 2 or WebSocket (via mobile push service) protocols for communication between the cloud and the maintenance terminal. The on-site verification and feedback interface can use the HTTP POST protocol for communication between the maintenance terminal and the cloud.

[0072] The following section uses a specific single-phase ground fault as an example to describe the process of topology identification and fault location in the system.

[0073] Step 1, Fault Occurrence: At 10:23:45 on a certain day, a single-phase ground fault occurred in phase A of the "Transformer No. 3 area" in the power distribution network. The fault point was located between the "Meter No. 5" and "Meter No. 6" on the branch line.

[0074] Step 2, Edge-side Local Anomaly Detection: The smart converged terminal in this area continuously receives voltage and current data from the 50 smart meters under its jurisdiction. A lightweight feature extraction model (such as a self-encoder) on the gateway calculates the compressed features of each meter in real time. At 10:23:45:50, the gateway detects that the voltage reported by meters 5 and 6 has momentarily dropped to 60% of its rated value, along with a sudden change in zero-sequence current. The local anomaly detection module triggers a "suspected single-phase grounding" event. The gateway immediately caches the raw waveform data from 2 seconds before to 3 seconds after the fault (a total of 5 seconds) and compresses and uploads this data, along with the event tag and meter ID, to the cloud. Simultaneously, a local alarm is pushed to the mobile phones of the maintenance personnel in this area (initial warning).

[0075] Step 3, Cloud Aggregation and Global Analysis: The cloud data receiving module receives the uploaded data from the intelligent fusion terminal and stores it in the full feature database. The fault location engine is triggered, first obtaining the current topology of the transformer area from the real-time topology graph database (this topology was just updated by the GNN model 1 minute ago, accurately reflecting the connection relationships at that time). The location engine extracts a list of all meter nodes that reported voltage drops: {"Meter 5", "Meter 6"}. On the real-time topology tree, the Lowest Common Ancestor (LCA) of these two nodes is calculated. The calculated LCA is "Branch Box B". Further analysis: Searching downstream from Branch Box B, it is found that all child nodes in the branch containing "Meter 5" and "Meter 6" experienced voltage drops, while the voltage in other branches (such as "Meter 7") remained normal. Therefore, the faulty branch is located as "the line segment between Branch Box B and Meters 5 and 6". Output result: "Transformer area No. 3, phase A, single-phase grounding fault occurred in the section from Branch Box B to Meters 5 / 6".

[0076] Step 4, Result Distribution and On-site Verification: The location results are pushed to the mobile phones of the maintenance personnel responsible for the area via the APP, displaying the fault location map, fault type, and suggestions. The maintenance personnel, carrying tools, go to the site and inspect the designated area. They find that the overhead line between "Meter 5" and "Meter 6" has been broken by tree branches, confirming a single-phase grounding fault. The maintenance personnel click "Verification Confirmation" on the APP, enter the fault cause "tree obstruction," and upload on-site photos.

[0077] Step 5, Closed-Loop Optimization: The verification feedback information is transmitted back to the cloud, stored in the full feature database, and the true location and cause of the fault are marked. In the next offline training cycle (e.g., every Sunday morning), the cloud uses the newly added labeled fault data to retrain the GNN topology inference model and edge feature extraction model, improving the model's accuracy in identifying similar faults. After training is complete, the updated lightweight model (e.g., the optimized autoencoder) is distributed to the intelligent fusion terminal of "Transformer Area No. 3" to achieve online model upgrade.

[0078] Figure 2 This is a flowchart of a cloud-edge collaborative distribution network topology real-time identification and fault location method provided in an embodiment of the present invention. Figure 2 As shown in the figure, the cloud-edge collaborative distribution network topology real-time identification and fault location method provided in this embodiment includes the following steps: S210, raw electrical quantity data are collected through the sensing layer, which includes terminal equipment deployed in the power distribution network; S220 uses a lightweight edge-side model and intelligent fusion terminal deployed at the edge layer to perform lightweight feature extraction and local anomaly detection on the raw electrical quantity data uploaded from the perception layer, and uploads the extracted lightweight feature data and the anomaly data obtained from local anomaly detection to the cloud platform layer. S230 uses a graph neural network model deployed at the cloud platform layer to perform real-time global topology inference based on lightweight feature data uploaded from all edge layers, thereby obtaining a dynamic distribution network topology. The S240 uses a fault location engine deployed at the cloud platform layer to locate faults based on abnormal data uploaded from the edge layer and a graph search algorithm that leverages the dynamic distribution network topology.

[0079] The edge-side lightweight model of the edge layer is deployed to the edge layer after the cloud platform layer performs lightweight compression based on the graph neural network model. The update parameters of the edge-side lightweight model of the edge layer are obtained by the cloud platform layer after updating the graph neural network model under the condition of event triggering and then sent to the edge layer.

[0080] In step S220 above, the lightweight edge model performs dimensionality reduction feature extraction on the time series data obtained based on the original electrical quantity data. Specifically, it includes: using principal component analysis to project the time series data of each time window onto the first k principal components to obtain a k-dimensional (low-dimensional) feature vector, where k takes the value of 3 to 10; or, using a lightweight autoencoder network to compress the time series data into a low-dimensional feature vector.

[0081] The intelligent fusion terminal uses a built-in anomaly detection module to determine in real time whether there are voltage sags, overcurrents, or harmonic distortion anomalies based on raw electrical quantity data or lightweight feature data. When a voltage sag, overcurrent, or harmonic distortion anomaly is detected, an event upload mechanism is triggered, uploading the raw electrical quantity data or corresponding lightweight feature data within the time window before and after the anomaly as the local anomaly detection data to the cloud platform layer.

[0082] In step S230 above, the graph neural network model deployed on the cloud platform layer performs real-time global topology inference based on the lightweight feature data uploaded by all edge layers. Specifically, this includes: using a spatiotemporal graph convolutional network, treating each edge layer as a node of the spatiotemporal graph convolutional network, using the lightweight feature data of all edge layer nodes within the most recent time window as the input of the spatiotemporal graph convolutional network, outputting the connection probability matrix between nodes through the spatiotemporal graph convolutional network, and obtaining a radial distribution network topology diagram after post-processing, which is then stored in the real-time topology database.

[0083] In step S240 above, the fault location engine deployed at the cloud platform layer locates the fault location based on the abnormal data uploaded from the edge layer and the dynamic distribution network topology using a graph search algorithm. Specifically, this includes: when abnormal data uploaded from the edge layer is received, collecting a list of all abnormal nodes that have experienced voltage drops; extracting the current distribution network topology tree from the real-time topology graph database; calculating the nearest common ancestor (LCA) node of the abnormal node list in the current distribution network topology tree; and searching downstream from the LCA node to determine the branch where the fault is located based on the abnormal distribution of voltage drops in the child nodes.

[0084] In a specific embodiment, the cloud platform layer performs lightweight compression of the graph neural network model to obtain a lightweight edge model, which includes: using the trained graph neural network model as the teacher model and the edge network structure to be deployed as the student model, using knowledge distillation technology to make the student model imitate the output of the teacher model, and using quantization and pruning techniques to compress the student model to generate a lightweight edge model that meets the edge computing power and storage constraints.

[0085] The cloud platform layer retrains the graph neural network model under event-triggered conditions, performs lightweight compression on the retrained graph neural network model to obtain an updated lightweight edge-side model, and sends the parameters of the updated lightweight edge-side model, along with its version number, to the edge layer through a standardized interface. Upon receiving the updated parameters, the edge layer replaces the parameters of its current lightweight edge-side model with those of the updated model.

[0086] In a specific embodiment, when the cloud platform layer detects that the statistical distribution of the lightweight feature data uploaded by the edge layer has drifted beyond a threshold, or when the cloud platform layer determines, based on the on-site verification feedback data, that the detection accuracy of the current edge-side lightweight model of the edge layer on a specific anomaly type has decreased beyond a preset threshold, the graph neural network model is retrained.

[0087] The cloud-edge collaborative method for real-time identification and fault location of distribution network topology also includes the following steps: dynamically updating the distribution network topology through a triggered incremental update mechanism at the cloud platform layer. Specifically, when the cloud platform layer detects a switch change event (a change in the distribution network topology), it does not perform a full global topology re-inference, but instead determines the affected area centered on the switch, extracts lightweight feature data of all nodes within the affected area, runs a local graph neural network model to re-infer the sub-topology to obtain an updated local sub-graph, and merges the updated local sub-graph with the original global topology graph to form a new distribution network topology graph.

[0088] In a specific embodiment, the cloud platform layer is also used to send the located fault location to the operation and maintenance terminal, and use the real fault location fed back by the operation and maintenance terminal as a labeled sample to retrain the graph neural network model and iteratively update the corresponding edge-side lightweight model.

[0089] The fault location engine at the cloud platform layer outputs fault location results including fault location and confidence level. The confidence level is calculated based on at least one of the following factors: number of nodes experiencing voltage sags, consistency of abnormal amplitudes, and clarity of topology branches. The confidence level is marked with different colors on the operations and maintenance terminal APP to assist operations and maintenance personnel in decision-making.

[0090] The technical details of the distribution network topology identification and fault location method provided in this embodiment can be understood by referring to the specific implementation details of the distribution network topology identification and fault location system described above, and will not be repeated here.

[0091] The proposed distribution network topology identification and fault location method has the same technical effects and advantages as the aforementioned distribution network topology identification and fault location system.

[0092] Advantage 1: Significantly reduces communication bandwidth pressure and cloud computing burden, improving resource utilization. This invention adopts a three-layer cloud-edge collaborative architecture of "perception layer - edge layer - platform layer," and deploys a lightweight feature extraction model and a local anomaly detection module at the edge layer. The intelligent fusion terminal performs real-time dimensionality reduction and feature extraction on the raw high-frequency data, only periodically uploading compressed feature vectors; and only uploading raw waveform segments when an abnormal event is detected. This mechanism optimizes the data transmission volume from "full raw data" to "high-value features + event-driven data," solving the bandwidth and computing pressure problems at the source. This invention offloads the processing task of massive raw measurement data to the intelligent fusion terminal, uploading only compressed high-value features or abnormal events, reducing communication bandwidth requirements by more than 90% and cloud computing load by more than 80%, effectively solving the network congestion and computing bottleneck problems caused by the full data upload in traditional centralized solutions.

[0093] Advantage 2: Significantly improves the dynamic adaptability and accuracy of fault location, overcoming the dependence on static topology files. This invention tightly couples real-time topology inference with fault location. A complex GNN model continuously runs in the cloud, updating the topology database in real time based on network-wide features. The fault location engine uses the latest real-time topology and the Least Common Ancestor (LCA) graph search algorithm to accurately locate branches of voltage sag nodes. This method ensures that the location process remains consistent with the actual connection relationship of the power grid, fundamentally overcoming the limitations of static files. This invention can perform fault location based on real-time inferred distribution network topology. When the power grid changes due to switching operations, reconfiguration, etc., the location results still maintain high accuracy, avoiding the location failure problem caused by the reliance on static files in traditional methods, improving the location accuracy by more than 30%.

[0094] Advantage 3: Achieves rapid adaptive updates under topology changes, avoiding full recalculation. Employing a triggered incremental update mechanism, after detecting topology changes through direct reporting of switch data or indirect inference from measurement data, the system automatically identifies the affected area, extracts node features from that area, runs a local GNN model for sub-topology updates, and merges it with the original global topology. Simultaneously, the cloud can trigger fine-tuning updates to the lightweight model of the affected intelligent fusion terminals, achieving rapid adaptation. When topology changes such as switch relocation occur in the distribution network, this system only needs to re-infer and fine-tune the model for the affected local area, reducing update time from minutes to seconds and computational overhead by more than 70%, significantly improving the system's response speed to dynamic topologies.

[0095] Advantage 4: Continuous improvement of edge-side intelligence, avoiding model solidification. This invention constructs an edge-cloud model collaborative update mechanism. The cloud periodically retrains complex models based on accumulated historical data and review feedback, and generates lightweight versions through compression techniques such as knowledge distillation, quantization, and pruning. These lightweight versions are then distributed to intelligent fusion terminals via standard interfaces to replace or update local models, achieving online upgrades at the edge. The lightweight edge-side models (feature extraction, anomaly detection) in this system can be dynamically updated as the cloud model is optimized. The level of edge intelligence continuously improves with data accumulation and algorithm evolution, avoiding the problem of edge models freezing after deployment and failing to adapt to new operating conditions in traditional solutions.

[0096] Advantage 5: A closed-loop optimization mechanism for continuous system learning is established, enabling data-driven iteration. Maintenance personnel report the actual fault location and cause via a terminal app; this data is stored in a cloud database as labeled samples. The cloud-based offline training module periodically retrains the model using these new samples, and the optimized model is then redeployed to the edge, achieving continuous iterative improvement in system performance. This system can feed back the on-site verification results of each fault handling process to the cloud for model retraining, forming a complete closed loop of "fault occurrence-location-verification-optimization." This allows the system's accuracy to continuously improve over time, with long-term fault location accuracy increasing by more than 15%.

[0097] Advantage Six: Real-time local early warning and rapid response to faults, shortening handling time. This invention deploys a lightweight real-time anomaly detection model on the intelligent fusion terminal to perform online analysis of the data stream. Once anomalies such as voltage drops or overcurrents are detected, a local alarm is immediately triggered, and relevant data is cached and uploaded. This mechanism transforms fault response from "centralized analysis in the cloud followed by notification" to a dual guarantee of "edge-based early warning and precise cloud-based location," significantly reducing response latency. This system can achieve millisecond-level local anomaly detection and real-time early warning at the edge, detecting faults several seconds to tens of seconds earlier than traditional centralized solutions, gaining valuable time for rapid isolation and handling, and improving overall fault handling efficiency by 40%.

[0098] Advantage 7: Ensuring system scalability and engineering deployment capabilities. This invention defines a complete set of standardized interface protocols, including an edge feature upload interface (MQTT+Protobuf), a cloud model distribution interface, a fault location result distribution interface, and an on-site verification and feedback interface (REST API). These interfaces clarify the data format, communication protocol, and interaction process, decoupling the various modules of the system and supporting plug-and-play and flexible expansion of heterogeneous devices.

[0099] Advantage 8: Provides confidence information for fault location results, assisting in operation and maintenance decisions. This invention integrates a confidence calculation module into the fault location engine, comprehensively considering factors such as the number of abnormal nodes, consistency of abnormal amplitude, and clarity of topology branches, generating a confidence score for each location result, which is then marked with different colors on the operation and maintenance terminal APP. The fault location results output by the system include confidence indicators (such as those calculated based on the number of nodes experiencing voltage drops, the degree of abnormality, and topology clarity). Operation and maintenance personnel can rationally allocate on-site verification resources according to the confidence level. High-confidence results can be handled directly, while low-confidence results require focused review, improving overall operation and maintenance efficiency by 25%.

[0100] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0101] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0102] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0104] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. Furthermore, it should be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. As long as such combination does not violate the spirit of the embodiments of the present invention, it should also be considered as the content disclosed in the embodiments of the present invention.

Claims

1. A cloud-edge collaborative power distribution network topology real-time identification and fault location system, characterized in that, include: The perception layer, edge layer, and cloud platform layer; The sensing layer includes terminal equipment deployed in the power distribution network for collecting raw electrical quantity data; The edge layer includes an intelligent fusion terminal deployed in the distribution substation and a lightweight edge model, which is used to perform lightweight feature extraction and local anomaly detection on the raw electrical quantity data uploaded by the perception layer, and upload the extracted lightweight feature data and the anomaly data obtained from the local anomaly detection to the cloud platform layer. The cloud platform layer is deployed with a graph neural network model and a fault location engine. The graph neural network model is used to perform real-time global topology inference based on the lightweight feature data uploaded by all edge layers to obtain a dynamic distribution network topology. The graph neural network model is a spatiotemporal graph convolutional network, with each edge layer serving as a node in the spatiotemporal graph convolutional network. The spatiotemporal graph convolutional network takes the lightweight feature data of all edge layer nodes within the most recent time window as input and outputs a connection probability matrix between nodes. After processing, a radial distribution network topology diagram is obtained. The fault location engine is used to locate the fault location based on the abnormal data uploaded by the edge layer and the dynamic distribution network topology using a graph search algorithm. The edge-side lightweight model is deployed to the edge layer after the cloud platform layer performs lightweight compression based on the graph neural network model. The update parameters of the edge-side lightweight model are obtained by the cloud platform layer after updating the graph neural network model under the condition of event triggering and then sent to the edge layer.

2. The cloud-edge collaborative power distribution network topology real-time identification and fault location system of claim 1, wherein, The edge-side lightweight model performs dimensionality reduction feature extraction on time-series data obtained from the original electrical quantity data to obtain lightweight feature data.

3. The cloud-edge collaborative power distribution network topology real-time identification and fault location system of claim 2, wherein, The edge-side lightweight model performs dimensionality reduction and feature extraction on time-series data obtained from the original electrical quantity data, including: Principal component analysis is used to project the time series data of each time window onto the first k principal components to obtain a k-dimensional feature vector. Alternatively, a lightweight autoencoder network can be used to compress the time series data into a k-dimensional feature vector, where k ranges from 3 to 10.

4. The cloud-edge collaborative power distribution network topology real-time identification and fault location system of claim 1, wherein, The intelligent fusion terminal has a built-in anomaly detection module, which determines in real time whether there are voltage sags, overcurrents or harmonic distortion anomalies based on raw electrical quantity data or lightweight feature data. When an anomaly such as voltage sag, overcurrent, or harmonic distortion is detected, the original electrical quantity data or corresponding lightweight feature data within the time window before and after the anomaly occurs are uploaded to the cloud platform layer as the anomaly data for local anomaly detection.

5. The cloud-edge collaborative power distribution network topology real-time identification and fault location system of claim 1, wherein, The fault location engine uses a graph search algorithm based on the dynamic distribution network topology to locate fault locations, specifically including: When abnormal data is received from the edge layer, a list of all abnormal nodes that have experienced voltage drops is collected. Extract the current distribution network topology tree; Calculate the nearest common ancestor node of the list of abnormal nodes in the current distribution network topology tree; Search downstream from the nearest common ancestor node and determine the branch where the fault is located based on the abnormal distribution of voltage drops in the child nodes.

6. The cloud-edge collaborative power distribution network topology real-time identification and fault location system of claim 1, wherein, The specific implementation methods for the cloud platform layer to perform lightweight compression of the graph neural network model to obtain a lightweight model on the edge side include: The cloud platform layer trains the graph neural network model and compresses the trained graph neural network model into a lightweight model through knowledge distillation, quantization, and pruning. This lightweight model is then deployed to the edge layer as a lightweight model for the edge side.

7. The cloud-edge collaborative power distribution network topology real-time identification and fault location system of claim 1, wherein, The cloud platform layer retrains the graph neural network model under the condition of event triggering, performs lightweight compression on the retrained graph neural network model to obtain an updated edge-side lightweight model, and sends the parameters of the updated edge-side lightweight model to the edge layer through a standardized interface. After receiving the updated parameters of the edge-side lightweight model, the edge layer replaces the parameters of the current edge-side lightweight model with the parameters of the updated edge-side lightweight model.

8. The cloud-edge collaborative distribution network topology real-time identification and fault location system according to claim 7, characterized in that, The cloud platform layer retrains the graph neural network model under event-triggered conditions, including: When the cloud platform layer detects that the statistical distribution of the lightweight feature data uploaded by the edge layer has drifted to a level exceeding a preset threshold, or when the cloud platform layer determines, based on on-site verification feedback data, that the detection accuracy of the current edge-side lightweight model in the edge layer for a specific anomaly type has dropped to a level below a preset threshold, the graph neural network model is retrained.

9. The cloud-edge collaborative distribution network topology real-time identification and fault location system according to claim 1, characterized in that, The cloud platform layer employs a triggered incremental update mechanism to dynamically update the distribution network topology, specifically including: When the cloud platform layer detects a switch change event, it determines the area of ​​influence centered on the switch, extracts lightweight feature data of all nodes within the area of ​​influence, runs a local graph neural network model to re-infer the sub-topology to obtain an updated local sub-graph, and merges the updated local sub-graph with the original global topology to form a new distribution network topology diagram.

10. The cloud-edge collaborative distribution network topology real-time identification and fault location system according to claim 1, characterized in that, The cloud platform layer is also used to send the located fault location to the operation and maintenance terminal, use the actual fault location returned by the operation and maintenance terminal as a labeled sample to retrain the graph neural network model, and iteratively update the edge-side lightweight model corresponding to the graph neural network model.

11. The cloud-edge collaborative distribution network topology real-time identification and fault location system according to claim 1, characterized in that, The fault location engine of the cloud platform layer is also used to output fault location results including fault location and confidence level, wherein the confidence level is calculated based on at least one of the following factors: number of voltage sag nodes, consistency of abnormal amplitude, and clarity of topology branches.

12. A method for real-time identification and fault location of distribution network topology using cloud-edge collaboration, characterized in that, include: Raw electrical quantity data is collected through a sensing layer, which includes terminal equipment deployed in the power distribution network; Lightweight feature extraction and local anomaly detection are performed on the raw electrical quantity data uploaded from the perception layer by using a lightweight edge-side model and intelligent fusion terminal deployed at the edge layer. The extracted lightweight feature data and the anomaly data obtained from local anomaly detection are then uploaded to the cloud platform layer. A graph neural network model deployed on the cloud platform layer performs real-time global topology inference based on lightweight feature data uploaded from all edge layers, resulting in a dynamic distribution network topology. The graph neural network model is a spatiotemporal graph convolutional network, with each edge layer serving as a node. The spatiotemporal graph convolutional network takes lightweight feature data from all edge layer nodes within the most recent time window as input and outputs a connection probability matrix between nodes. After processing, a radial distribution network topology diagram is obtained. The fault location engine deployed at the cloud platform layer locates the fault location based on the abnormal data uploaded from the edge layer and the dynamic distribution network topology using a graph search algorithm. The edge-side lightweight model is deployed to the edge layer after the cloud platform layer performs lightweight compression based on the graph neural network model. The update parameters of the edge-side lightweight model are obtained by the cloud platform layer after updating the graph neural network model under the condition of event triggering and then sent to the edge layer.

13. The method for real-time identification and fault location of distribution network topology based on cloud-edge collaboration according to claim 12, characterized in that, The lightweight edge-side model and intelligent fusion terminal deployed at the edge layer perform lightweight feature extraction and local anomaly detection on the raw electrical quantity data uploaded from the perception layer, including: Dimensionality reduction feature extraction is performed on time series data obtained from raw electrical quantity data to obtain lightweight feature data; The built-in anomaly detection module of the intelligent fusion terminal can determine in real time whether there are voltage sags, overcurrents or harmonic distortion anomalies based on raw electrical quantity data or lightweight characteristic data.

14. The method for real-time identification and fault location of distribution network topology based on cloud-edge collaboration according to claim 12, characterized in that, The graph neural network model deployed through the cloud platform layer performs real-time global topology inference based on lightweight feature data uploaded from all edge layers, including: A spatiotemporal graph convolutional network is adopted, with each edge layer as a node of the spatiotemporal graph convolutional network. The lightweight feature data of all edge layer nodes within the most recent time window are used as the input of the spatiotemporal graph convolutional network. The spatiotemporal graph convolutional network outputs the connection probability matrix between nodes, and after processing, a radial distribution network topology diagram is obtained.

15. The method for real-time identification and fault location of distribution network topology based on cloud-edge collaboration according to claim 12, characterized in that, The fault location engine deployed through the cloud platform layer locates the fault location based on the abnormal data uploaded from the edge layer and a graph search algorithm based on the dynamic distribution network topology, including: When abnormal data is received from the edge layer, a list of all abnormal nodes that have experienced voltage drops is collected. Extract the current distribution network topology tree; Calculate the nearest common ancestor node of the list of abnormal nodes in the current distribution network topology tree; Search downstream from the nearest common ancestor node and determine the branch where the fault is located based on the abnormal distribution of voltage drops in the child nodes.

16. The method for real-time identification and fault location of distribution network topology based on cloud-edge collaboration according to claim 12, characterized in that, The method further includes: The distribution network topology is dynamically updated through a trigger-based incremental update mechanism at the cloud platform layer. Specifically, when the cloud platform layer detects a switch change event, it determines the area of ​​influence centered on the switch, extracts lightweight feature data of all nodes within the area of ​​influence, runs a local graph neural network model to re-infer the sub-topology to obtain the updated local sub-graph, and merges the updated local sub-graph with the original global topology to form a new distribution network topology diagram.

Citation Information

Patent Citations

  • Method and device for identifying low-current single-phase earth fault of power distribution network based on cloud-side cooperation

    CN120744634A

  • Fault positioning method for power distribution network containing distributed power supply based on cloud edge collaboration

    CN121602372A

  • Power distribution network state prediction and fault location method and device, and terminal equipment

    CN121679241A