Non-technical loss identification and early warning method and system based on power grid topology inference

Abstract

The application relates to the technical field of power grid loss analysis, and provides a non-technical loss identification and early warning method and system based on power grid topology inference. The application infers a power distribution network physical topology graph in real time through multi-source fusion data, compares the power distribution network physical topology graph with a static file topology graph, determines a topology abnormal node, and identifies electricity stealing behavior. Meanwhile, fine line loss allocation is performed according to the real-time inferred power distribution network physical topology and line parameters, a physical section with an abnormally high loss rate is identified, line loss abnormalities are accurately locked to specific line sections, and accurate targets are provided for on-site inspection. Meanwhile, the real-time inferred power distribution network physical topology graph is used as a basic structure, and an electricity consumption behavior graph is constructed according to the electricity consumption behavior characteristics of users, and an electricity consumption behavior abnormal group is identified. The application combines three detection methods of topology comparison, fine line loss allocation and electricity consumption behavior graph analysis, and realizes multi-dimensional and high-precision identification of non-technical loss.

Description

Technical Field

[0001] This application relates to the field of power grid loss analysis technology, specifically to a non-technical loss identification and early warning method and system based on power grid topology inference. Background Technology

[0002] Non-Technical Loss (NTL) mainly refers to the loss of electrical energy in the power system caused by reasons other than the power equipment itself, such as electricity theft, metering device failure, and wiring errors. Among these, electricity theft is the main cause.

[0003] Traditional electricity theft detection methods primarily rely on statistical analysis of user electricity consumption. They construct models of normal user electricity usage behavior to identify users exhibiting abnormal patterns. Common methods include electricity consumption prediction based on time series analysis, user grouping based on clustering algorithms, and anomaly detection based on outlier detection. These methods compare a user's actual electricity consumption with predicted values ​​or the average of similar users, calculating the degree of deviation as an indicator of suspected electricity theft.

[0004] The transformer substation line loss rate (the ratio of the difference between the total electricity consumption of the substation and the sum of the electricity consumption of each sub-meter to the total electricity consumption) is an important indicator reflecting electricity theft. An abnormally high line loss rate usually indicates potential electricity theft in that substation. Traditional methods based on transformer substation line loss analysis monitor daily and monthly trends in the line loss rate and set thresholds for early warning. Line loss allocation methods attempt to distribute the total loss to each user according to certain rules (such as electricity consumption ratio) to identify abnormal contributors. However, traditional methods lack topological information support and cannot accurately locate the electricity theft sections.

[0005] In recent years, deep learning technology has been widely used in electricity theft detection. Typical methods include electricity waveform feature extraction based on convolutional neural networks (CNNs), electricity time series modeling based on long short-term memory networks (LSTMs), and reconstruction error detection based on autoencoders. These methods train normal users' behavioral patterns and mark users with large deviations as suspects. However, existing methods of this type mainly focus on the temporal characteristics of individual users and are difficult to detect group anomalies.

[0006] In the field of anti-electricity theft, existing research has focused on constructing anti-electricity theft knowledge graphs and identifying abnormal patterns through methods such as subgraph matching. Knowledge graphs model entities (such as users, transformers, and meters) and their relationships (such as power supply connections and geographical locations) as graph structures, supporting graph queries, graph reasoning, and abnormal pattern mining. When constructing knowledge graphs, these schemes typically use topological relationships derived from marketing archive systems (such as user-transformer relationships and metering point locations). However, the topology in the knowledge graph is a static archive topology, not a real-time physical topology. Electricity theft itself can lead to inconsistencies between the physical topology and the archive topology (unauthorized wiring), in which case reasoning based on the static archive topology may fail. Summary of the Invention

[0007] To address one of the aforementioned technical deficiencies, this application provides a method and system for identifying and warning of non-technical losses based on power grid topology inference. By combining three technical means—multi-source fusion data, comparison of the distribution network physical topology map with the static archive topology map, refined line loss allocation, and electricity consumption behavior graph analysis—it achieves multi-dimensional and high-precision identification of non-technical losses (electricity theft).

[0008] The first aspect of this application provides a non-technical loss identification and early warning method based on power grid topology inference, including: The physical topology of the distribution network is inferred in real time based on multi-source fusion data, which is obtained by associating and fusing real-time measurement data, marketing archive data and historical alarm data of each node in the power grid. The real-time inferred physical topology map of the distribution network is compared with the static archive topology map to identify topology anomaly nodes and generate first-level alarm information containing topology anomaly nodes. Based on the real-time inferred physical topology of the distribution network, the loss of each branch is calculated according to the real-time measurement data. Branches with losses exceeding the preset threshold are identified as abnormal line loss sections, and secondary alarm information containing abnormal line loss sections is generated. Based on the real-time inferred physical topology of the distribution network, an electricity consumption behavior map is constructed according to the characteristics of users' electricity consumption behavior. Based on the electricity consumption behavior map, abnormal electricity consumption groups are identified, and three-level alarm information containing abnormal electricity consumption groups is generated. The first-level alarm information, second-level alarm information, and third-level alarm information are integrated and correlated to generate audit information.

[0009] In this embodiment of the application, the step of inferring the physical topology of the distribution network in real time based on multi-source fusion data includes: The current multi-source fusion data is input into a trained graph neural network model. The physical connection relationship of the current distribution network is inferred through the graph neural network model, and the physical topology map of the distribution network is output. The physical topology map of the distribution network includes a set of nodes and a set of edges representing the physical connection relationship of the nodes.

[0010] In this embodiment of the application, the step of comparing the real-time inferred distribution network physical topology map with the static archived topology map to determine topology anomaly nodes and generating first-level alarm information containing topology anomaly nodes includes: Extract a static archive topology graph from the marketing archive data, the archive topology graph containing a set of nodes and a set of edges for archive records; The real-time inferred physical topology map of the distribution network is compared with the archived topology map to identify edges whose edge sets are inconsistent with those of the archived topology map. These inconsistent edges are identified as topology anomaly edges. A first-level alarm message is generated for each topology anomaly edge, and the first-level alarm message includes the topology anomaly node associated with the topology anomaly edge.

[0011] In this embodiment of the application, the distribution network physical topology diagram based on real-time inference, which calculates the loss of each branch based on real-time measurement data, includes: A breadth-first search is performed on the physical topology of the distribution network to obtain the hierarchy of each node; Starting from the deepest node, the injected current of each node is calculated based on the real-time measurement data of each node, and then accumulated upstream to obtain the upstream voltage. Starting from the root node, calculate the voltage of each downstream node based on the upstream voltage and line impedance, and calculate the active power loss of each branch.

[0012] In this embodiment of the application, the distribution network physical topology diagram based on real-time inference, which calculates the loss of each branch based on real-time measurement data, includes: calculating the actual total loss of the transformer area based on real-time measurement data; The actual total loss is proportionally allocated to each branch to obtain the loss of each branch; where the allocation coefficient of each branch is the product of the square of the branch current and the resistance.

[0013] In this embodiment of the application, the power distribution network physical topology map based on real-time inference, which constructs an electricity consumption behavior map based on the electricity consumption behavior characteristics of users, includes: For each user node, extract the electricity consumption behavior feature vector based on the user's historical electricity consumption time-series data; Using the physical topology of the distribution network as the basic graph structure, the electricity consumption behavior feature vector of each user node is added as a node attribute to form an attribute graph, which is then used as the electricity consumption behavior map.

[0014] In this embodiment of the application, the step of determining abnormal electricity consumption groups based on electricity consumption behavior maps includes: Based on the electricity consumption behavior map, the nodes are divided into multiple communities using graph neural networks or graph clustering algorithms. The nodes in each community are connected on the distribution network physical topology map and have similar electricity consumption behaviors. Identify groups with unusual electricity consumption behaviors in multiple communities.

[0015] In this embodiment of the application, the identification of abnormal electricity consumption groups within multiple communities includes: For each community, calculate the average reconstruction error of the nodes within the community. If the average reconstruction error of a community is significantly higher than the average reconstruction error of other communities, the community is identified as a group with abnormal electricity consumption behavior. Alternatively, for each node, calculate the local outlier value of electricity consumption behavior within the node's topological neighbors. If the local outlier value of a node and its neighbors is higher than the normal value, determine that the node and its neighbors as a group with abnormal electricity consumption behavior. Alternatively, compare the electricity consumption statistics of different communities in the same transformer area. If the electricity consumption statistics of a certain community are significantly higher, then that community is identified as a group with abnormal electricity consumption behavior.

[0016] In this embodiment of the application, the process of fusing and associating first-level alarm information, second-level alarm information, and third-level alarm information to generate audit information includes: Alarm messages of Level 1, Level 2, and Level 3 are prioritized; Level 1 alarm messages have the highest priority, and Level 3 alarm messages have the lowest priority. Calculate the overall suspicion level for all users or regions, associate alarm information of different priorities in the same region based on the overall suspicion level, and generate audit information based on the overall suspicion level.

[0017] In this embodiment of the application, the non-technical loss identification and early warning method based on power grid topology inference further includes: overlaying the real-time inferred distribution network physical topology map and alarm information onto a geographic information system map and displaying it on a visualization interface.

[0018] A second aspect of this application provides a non-technical loss identification and early warning system based on power grid topology inference, comprising: The data source layer is used to acquire real-time measurement data, marketing archive data, and historical alarm data, and to correlate and merge them to obtain multi-source fused data; The data processing layer is used for: Real-time inference of the physical topology of the distribution network based on multi-source fusion data; The real-time inferred physical topology map of the distribution network is compared with the static archive topology map to identify topology anomaly nodes and generate first-level alarm information containing topology anomaly nodes. Based on the real-time inferred physical topology of the distribution network, the loss of each branch is calculated according to the real-time measurement data. Branches with losses exceeding the preset threshold are identified as abnormal line loss sections, and secondary alarm information containing abnormal line loss sections is generated. Based on the real-time inferred physical topology of the distribution network, an electricity consumption behavior map is constructed according to the characteristics of users' electricity consumption behavior. Based on the electricity consumption behavior map, abnormal electricity consumption groups are identified, and three-level alarm information containing abnormal electricity consumption groups is generated. The early warning and display layer is used to integrate and correlate first-level, second-level, and third-level alarm information to generate investigation information.

[0019] The aforementioned technical solution uses multi-source fusion data to infer the physical topology of the distribution network in real time. By comparing this physical topology with a static archived topology, it identifies anomalous nodes and thus recognizes "unauthorized wiring" type electricity theft. Simultaneously, it performs refined line loss allocation based on the real-time inferred distribution network physical topology and line parameters, identifying physical sections with abnormally high loss rates and pinpointing the anomalies to specific line segments, providing precise targets for on-site inspections. Furthermore, using the real-time inferred distribution network physical topology as a foundation, it constructs an electricity consumption behavior map based on user electricity behavior characteristics to identify groups with abnormal electricity consumption behavior. This solution combines topology comparison, refined line loss allocation, and electricity consumption behavior map analysis to achieve multi-dimensional and high-precision identification of non-technical losses (electricity theft).

[0020] Other features and advantages of the technical solution of this application will be described in detail in the following detailed implementation section. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a non-technical loss identification and early warning method based on power grid topology inference provided in this application embodiment; Figure 2 This is a block diagram of the Spatiotemporal Graph Convolutional Network (ST-GCN) provided in an embodiment of the present invention; Figure 3 This is a block diagram of a non-technical loss identification and early warning system based on power grid topology inference provided in this application embodiment; Figure 4 This is a schematic diagram of the overall architecture and data flow of a non-technical loss identification and early warning system based on power grid topology inference provided in a specific embodiment of this application. Detailed Implementation

[0022] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0023] This application provides a non-technical loss identification and early warning method and system based on power grid topology inference. It infers the physical topology of the distribution network in real time through multi-source fusion data. The physical topology is compared with a static archived topology. When inconsistencies are found (e.g., nodes not connected in the archived topology appear connected in the physical topology), these are used as high-confirmation evidence of electricity theft, triggering an alarm and directly identifying "unauthorized wiring" type electricity theft. Simultaneously, using real-time inferred (high-precision) distribution network physical topology and line parameters (or parameter estimation), the total loss of the transformer area is distributed hierarchically to each branch segment and terminal node, identifying physical segments with abnormally high loss rates. This achieves refined line loss allocation based on the real physical topology, precisely pinpointing line loss anomalies to specific line segments, providing accurate targets for on-site inspections. Furthermore, using the real-time inferred distribution network physical topology as the basic structure, user electricity consumption characteristics (such as load curves, nighttime electricity consumption, and fluctuation patterns) are used as node attributes to construct an electricity consumption behavior graph. Algorithms such as graph clustering and graph anomaly detection are applied to identify collaborative electricity consumption anomalies among topologically adjacent users. This application combines three detection methods: topology comparison, refined line loss allocation, and electricity consumption behavior graph analysis. It also performs multi-level fusion and priority ranking of the three detection results to achieve multi-dimensional and high-precision identification of non-technical losses (electricity theft) to guide the efficient allocation of inspection resources.

[0024] Figure 1 This is a flowchart of a non-technical loss identification and early warning method based on power grid topology inference provided in an embodiment of this application. Figure 1 As shown, the non-technical loss identification and early warning method based on power grid topology inference provided in this embodiment includes the following steps: S110, Real-time inference of the physical topology of the distribution network based on multi-source fusion data, wherein the multi-source fusion data is obtained by associating and fusing real-time measurement data, marketing archive data and historical alarm data of each node in the power grid; S120 compares the real-time inferred physical topology map of the distribution network with the static archive topology map to identify topology anomaly nodes and generate a first-level alarm message containing information on topology anomaly nodes. S130, based on the real-time inferred physical topology of the distribution network, calculates the loss of each branch according to the real-time measurement data, identifies the branch with loss exceeding the preset threshold as the abnormal line loss section, and generates a secondary alarm information containing the abnormal line loss section information. S140, based on the real-time inferred physical topology of the distribution network, constructs an electricity consumption behavior map according to the electricity consumption behavior characteristics of users, identifies abnormal electricity consumption groups according to the electricity consumption behavior map, and generates a three-level alarm information containing information on abnormal electricity consumption groups; S150 integrates and correlates Level 1, Level 2, and Level 3 alarm information to generate audit information.

[0025] In step S110 above, a unified data fusion interface is used to associate and fuse multi-source information such as real-time measurement data (e.g., voltage, current, power), marketing archive data (e.g., household transformer relationship, metering point information, user profile), and historical alarm data (e.g., cover opening records, on-site inspection results) to obtain multi-source fused data.

[0026] The current multi-source fused data is input into a trained graph neural network model. The model infers the physical connections of the current distribution network and outputs a physical topology map of the distribution network. (Distribution network physical topology map) , containing a set of nodes (Such as electricity meters, branch boxes, transformers, etc.) and the set of edges representing the physical connection relationships of nodes. Edge set Each edge carries estimated line impedance parameters (resistance R and reactance X), which can be obtained from marketing archive data or identified through real-time measurement data.

[0027] In a specific embodiment, the graph neural network model employs a spatiotemporal graph convolutional network (ST-GCN). For example... Figure 2 As shown, the Spatiotemporal Graph Convolutional Network (ST-GCN) includes a spatiotemporal graph encoder and a dual-branch decoder. The spatiotemporal graph encoder includes temporal convolutional layers and spatial graph convolutional layers, used to extract spatiotemporal correlation features from multi-source fused data and output node spatiotemporal embedding feature vectors. The dual-branch decoder includes a voltage reconstruction decoder and an edge prediction decoder. The voltage reconstruction decoder calculates the predicted voltage based on the node spatiotemporal embedding feature vectors, and the edge prediction decoder predicts the node's edge input probability based on the node spatiotemporal embedding feature vectors.

[0028] The spatial graph convolutional layer in the spatiotemporal graph encoder contains a learnable adjacency matrix. The initial value of the learnable adjacency matrix is ​​set as a matrix based on the correlation coefficients of node measurement data, and it is updated during training via gradient descent. The spatial graph convolutional layer measures the potential connection strength between nodes using the learnable adjacency matrix. The current features of a node are linearly transformed by the trainable weight matrix, then weighted and summed according to the normalized weights of the learnable adjacency matrix, and finally passed through an activation function to obtain the updated features.

[0029] The voltage reconstruction decoder includes a differentiable push-back simulation layer. The differentiable push-back simulation layer simulates the voltage drop calculation based on linearized power flow by using the spatiotemporal embedding feature vector of the node and the known node injected power data to obtain the predicted voltage.

[0030] The voltage reconstruction decoder includes a differentiable push-back simulation layer. This layer simulates voltage drop calculations based on linearized power flow without requiring precise line parameters, using the node spatiotemporal embedding feature vectors and known node injected power data. .

[0031] The edge prediction decoder is implemented by a multilayer perceptron. Its input is the concatenation of the embedded feature vectors of any two nodes, and the output is mapped to a probability value between 0 and 1 by the sigmoid function. The probability values ​​of all node pairs formed by two nodes constitute the node edge probability.

[0032] The loss function of the Spatiotemporal Graph Convolutional Network (ST-GCN) model includes: voltage reconstruction loss. Physical consistency loss Edge prediction regularization loss The voltage reconstruction loss is the error between the predicted voltage and the actual voltage measurement. The physical consistency loss is determined based on the node edge probability under the constraint of the node current law. The edge prediction regularization loss introduces a regularization term to calculate the L1 norm of the probabilistic adjacency matrix, preventing the spatiotemporal graph convolutional network from outputting an overly dense or sparse connection structure. Total loss Represented as: ;in and These are hyperparameters used to balance the weights of various losses.

[0033] Voltage rebuilding loss : Calculation model predicts voltage Compared with the actual measured voltage input The mean square error (MSE) between the voltage and the signal, as the core of the self-supervised signal, requires the model to learn the correct topology and correlations in order to accurately reconstruct the voltage. Voltage reconstruction loss. Represented as: .

[0034] Physical consistency loss Including KCL constraint loss based on Kirchhoff's current law and radial constraint loss The formula for calculating KCL constraint loss is: ; in, The probability adjacency matrix represents the probability of nodes entering an edge. With nodes There is a probability that an edge exists between them. This indicates that the summation is performed for all non-roots ( )node.

[0035] Radial constraint loss Loops are avoided by penalizing non-zero elements in the powers of the predicted adjacency matrix. The total physical consistency loss is... ,in For weight hyperparameters.

[0036] The aforementioned Spatiotemporal Graph Convolutional Network (ST-GCN) is an unsupervised learning graph neural network model. In the dual-branch decoder, the voltage reconstruction decoder calculates the predicted voltage based on the spatiotemporal embedding feature vector of the node, and the edge prediction decoder predicts the node edge entry probability based on the node spatiotemporal embedding feature vector. The spatiotemporal graph convolutional network model uses voltage reconstruction loss as the main driving signal, and simultaneously integrates physical consistency loss based on the node edge entry probability under the constraint of the node current law for training, so as to realize the direct and efficient inference of the distribution network topology from multi-source fused data.

[0037] In step S120 above, a static archive topology map is extracted from the marketing archive data. The archive's topology graph contains a set of nodes. (Such as electricity meters, branch boxes, transformers, etc.) and the set of records. Among them, the edge set of archival records Based on household transformer relationships and line records, a real-time inferred physical topology map of the distribution network is constructed. With archive topology map By comparing the edges, the set of edges in the physical topology graph of the distribution network can be identified. The set of edges of the archive topology graph Inconsistent edges are identified as topologically abnormal edges (or topologically missing edges). A first-level alarm message is generated for each topologically abnormal edge, which includes information about the topologically abnormal node associated with the edge.

[0038] In step S130 above, based on the real-time inferred distribution network physical topology map, the loss of each branch is calculated according to the real-time measurement data. Specifically, this includes: performing a breadth-first search on the distribution network physical topology map to obtain the level (depth) of each node; starting from the deepest node, calculating the injected current (or complex power) of each node according to the real-time measurement data of each node, and accumulating it upstream to obtain the upstream voltage; starting from the root node, calculating the voltage of each downstream node according to the upstream voltage and line impedance, and calculating the active power loss of each branch; wherein, the line impedance can be obtained from the edge set information of the distribution network physical topology map.

[0039] Alternatively, the actual total loss of the transformer area can be calculated based on real-time measurement data, and the actual total loss can be proportionally allocated to each branch to obtain the loss of each branch; where the allocation coefficient of each branch is the product of the square of the branch current and the resistance.

[0040] In step S140 above, based on the real-time inferred distribution network physical topology map, an electricity consumption behavior map is constructed according to the user's electricity consumption behavior characteristics. Specifically, this includes: for each user node, extracting an electricity consumption behavior feature vector based on the user's historical electricity consumption time-series data; and using the distribution network physical topology map... Based on the basic graph structure, the electricity consumption behavior feature vector of each user node is appended as a node attribute to form an attribute graph. ,in, For a set of nodes, The set of edges that characterizes the physical connections between nodes. The node feature matrix is ​​used to represent the attribute graph as a power consumption behavior map.

[0041] Identifying abnormal electricity consumption groups based on electricity consumption behavior maps specifically includes: dividing nodes into multiple communities using graph neural networks or graph clustering algorithms (such as graph autoencoders, GraphSAGE combined with clustering, etc.) based on electricity consumption behavior maps, where nodes in each community are connected on the distribution network physical topology map and have similar electricity consumption behaviors; and identifying abnormal electricity consumption groups within multiple communities.

[0042] To identify groups with abnormal electricity consumption behavior within multiple communities, the following three methods can be used: First, a method based on reconstruction error: train a graph autoencoder to learn the feature representations of normal users. For each community, calculate the average reconstruction error of nodes within that community. If the average reconstruction error of a community is significantly higher than that of other communities, that community is identified as a group with abnormal electricity consumption behavior. Second, a method based on local outlier factor: for each node, calculate the local outlier factor (LOF) value of electricity consumption behavior within the node's topological neighbors. If the LOF values ​​of a node and its neighbors are generally high (higher than normal), that node and its neighbors are identified as a group with abnormal electricity consumption behavior. Third, a method based on community comparison: compare the electricity consumption statistics (such as the average percentage of nighttime electricity consumption) of different communities in the same transformer area. If the electricity consumption statistics of a community are significantly higher, that community is identified as a group with abnormal electricity consumption behavior.

[0043] In step S150 above, the first-level alarm information, second-level alarm information, and third-level alarm information are prioritized; among them, the first-level alarm information has the highest priority, and the third-level alarm information has the lowest priority. A comprehensive suspicion level is calculated for all users or regions (e.g., sorted in descending order). Based on the comprehensive suspicion level, alarm information of different priorities in the same region (such as the same server area or the same branch) is associated, and investigation information is generated based on the comprehensive suspicion level.

[0044] In an optional embodiment, the non-technical loss identification and early warning method based on power grid topology inference further includes: overlaying the real-time inferred distribution network physical topology map and alarm information onto a Geographic Information System (GIS) map and displaying it on a visualization interface. Through the visualization interface, the inferred physical topology, anomaly markers (topology mismatch points, abnormal line loss sections, users with abnormal behavior), and suspected electricity theft paths are overlaid on the GIS map, providing inspectors with intuitive on-site navigation.

[0045] Figure 3 This is an architecture diagram of a non-technical loss identification and early warning system based on power grid topology inference provided in an embodiment of this application. Figure 3As shown, the non-technical loss identification and early warning system includes: a data source layer, a data processing layer, and an early warning and display layer. The data source layer acquires real-time measurement data, marketing archive data, and historical alarm data, and correlates and merges them to obtain multi-source fused data. The data processing layer infers the physical topology of the distribution network in real time based on the multi-source fused data; compares the real-time inferred physical topology with the static archive topology to identify topology anomaly nodes and generate first-level alarm information containing these nodes; based on the real-time inferred physical topology, it calculates the loss of each branch according to real-time measurement data, identifies branches with losses exceeding a preset threshold as abnormal line loss sections, and generates second-level alarm information containing these sections; based on the real-time inferred physical topology, it constructs an electricity consumption behavior map based on user electricity consumption behavior characteristics, identifies abnormal electricity consumption behavior groups based on the map, and generates third-level alarm information containing these groups. The early warning and display layer merges and correlates the first-level, second-level, and third-level alarm information to generate audit information.

[0046] Specifically, the data processing layer infers the physical topology of the distribution network in real time based on multi-source fusion data, including: inputting the current multi-source fusion data into a trained graph neural network model, inferring the physical connection relationship of the current distribution network through the graph neural network model, and outputting the physical topology of the distribution network. The physical topology of the distribution network includes a set of nodes and a set of edges representing the physical connection relationship of the nodes.

[0047] Specifically, the data processing layer compares the real-time inferred distribution network physical topology map with the static archive topology map to identify topology anomaly nodes and generate a Level 1 alarm message containing the topology anomaly nodes. This includes: extracting the static archive topology map from the marketing archive data, which contains a set of nodes and a set of edges for archive records; comparing the real-time inferred distribution network physical topology map with the archive topology map to find edges where the edge set of the distribution network physical topology map does not match the edge set of the archive topology map, identifying the inconsistent edges as topology anomaly edges, and generating a Level 1 alarm message for each topology anomaly edge, which includes the topology anomaly node associated with the topology anomaly edge.

[0048] Specifically, the data processing layer, based on the real-time inferred physical topology of the distribution network, calculates the losses of each branch according to real-time measurement data. This includes: performing a breadth-first search on the physical topology of the distribution network to obtain the hierarchy of each node; starting from the deepest node, calculating the injected current of each node based on real-time measurement data, and accumulating it upstream to obtain the upstream voltage; starting from the root node, calculating the voltage of each downstream node based on the upstream voltage and line impedance, and calculating the active power loss of each branch. Alternatively, it calculates the actual total loss of the distribution area based on real-time measurement data, and proportionally allocates the actual total loss to each branch to obtain the loss of each branch; where the allocation coefficient for each branch is the product of the square of the branch current and the resistance.

[0049] Specifically, the data processing layer constructs an electricity consumption behavior graph based on the real-time inferred distribution network physical topology map and the electricity consumption behavior characteristics of users. It then identifies groups with abnormal electricity consumption behavior based on this graph. This process includes: for each user node, extracting electricity consumption behavior feature vectors based on the user's historical electricity consumption time-series data; using the distribution network physical topology map as the base graph structure, attaching the electricity consumption behavior feature vectors of each user node as node attributes to form an attribute graph, which serves as the electricity consumption behavior graph. Based on the electricity consumption behavior graph, graph neural networks or graph clustering algorithms are used to divide the nodes into multiple communities. Nodes within each community are connected on the distribution network physical topology map and exhibit similar electricity consumption behaviors; and abnormal electricity consumption behavior groups within multiple communities are identified.

[0050] This involves identifying groups with abnormal electricity consumption behavior within multiple communities. Three methods can be used to identify such groups: First, a reconstruction error-based method: For each community, calculate the average reconstruction error of nodes within that community. If the average reconstruction error of a community is significantly higher than that of other communities, that community is identified as having abnormal electricity consumption behavior. Second, a local outlier (LOF)-based method: For each node, calculate the local outlier (LOF) value of electricity consumption behavior within its topological neighbors. If the LOF values ​​of a node and its neighbors are generally high (higher than normal), that node and its neighbors are identified as having abnormal electricity consumption behavior. Third, a community comparison method: Compare the electricity consumption statistics of different communities within the same transformer area. If the electricity consumption statistics of a community are significantly higher than normal, that community is identified as having abnormal electricity consumption behavior.

[0051] Figure 4 This is a schematic diagram illustrating the overall architecture and data flow of a non-technical loss identification and early warning system based on power grid topology inference, provided as a specific example of this application. (See diagram for details.) Figure 4As shown, in one specific embodiment, the system consists of three parts: a data source layer, a data processing layer, and an early warning and display layer. The data source layer receives real-time measurement data (from smart meters), marketing archive data (from the marketing business system), and historical alarm and audit data (from maintenance records). The core processing modules of the data processing layer include: a physical topology inference module, an archive topology extraction module, a physical-archive topology comparison module, a refined line loss allocation module, and a user electricity consumption behavior map construction module. The data processing layer ultimately generates three levels of alarm information. The multi-level early warning fusion engine of the early warning and display layer merges and prioritizes the three levels of alarm information, displays it through a GIS visualization platform, pushes it to the on-site audit terminal, and receives on-site verification feedback, forming a closed-loop optimization.

[0052] The multi-source data fusion module at the data source layer collects data from different business systems. The data collection interfaces include: real-time measurement data interface, marketing record data interface, and historical alarm and audit data interface. The real-time measurement data interface obtains data from the electricity information collection system or IoT management platform. The data format is JSON or CSV, and includes: user identifier, timestamp, voltage (A / B / C phases), current (A / B / C phases), active power, reactive power, power factor, etc. The collection frequency is 15 minutes / time or higher. The marketing record data interface obtains data from the marketing business system. The data format is a relational database table or API interface, and includes: basic user information, user-transformer relationship (area, transformer), metering point information (meter ID, installation location, wiring method), line records (line number, length, model, impedance parameters), etc. The historical alarm and audit data interface obtains historical electricity theft alarm records (time, user, alarm type), on-site audit results (verification, cause, handling method), cover opening records, magnetic field interference alarms, etc., from the operation and maintenance management system. Data preprocessing includes time alignment, outlier handling, unified entity identification, and data standardization. Time alignment aligns data from different sources to a unified timestamp (e.g., 15-minute granularity), with missing values ​​filled using linear interpolation or forward imputation. Outlier handling removes data that significantly exceeds the physical measurement range (e.g., voltage > 1.2 times rated value, negative current). Unified entity identification establishes a unified system for identifying users, meters, transformer areas, and lines, linking marketing records with measurement data. Data standardization normalizes numerical data (e.g., voltage per unit value) to facilitate subsequent algorithmic processing.

[0053] The physical-archive topology comparison module automatically compares the real-time inferred physical topology with the static archive topology of the marketing system, detecting topology inconsistencies caused by unauthorized wiring, alterations, etc., and issuing a level one alarm. Based on real-time measurement data, the module uses graph neural network models (such as the Spatiotemporal Graph Convolutional Network (ST-GCN) model) to infer the physical connections of the current distribution network and generate a distribution network physical topology map. ,in For a set of nodes (meters, branch boxes, transformers). This represents the inferred set of edges (physical connections). A static archive topology graph is extracted from the marketing archive database. , It is a set of edges for the archive records (constructed based on household-transformer relationships and line archives). This is a set of nodes (meters, branch boxes, transformers). The goal of the comparison is to identify... and Inconsistent edges are mainly divided into missing edges and redundant edges. Missing edges do not exist in the file but exist in the physical topology (i.e., newly added connections made without authorization). Redundant edges exist in the file but do not exist in the physical topology (i.e., connections lost due to line cuts or modifications). The comparison steps are as follows: (1) Node matching: Ensure that the two graphs use the same node identifiers (user ID, meter ID, device ID), and establish a node mapping table; (2) Edge set difference calculation: Calculate the newly added edge set ; Calculate the missing edge set ; (3) Anomaly detection: If If the edge is not empty, it indicates suspected unauthorized or haphazard connections; each newly added edge will be recorded as a topological anomaly. If the edge is not empty, there may be a broken line or an error in the record, which will be recorded as a missing edge in the topology. A level 1 alarm message is generated for each abnormal edge in the topology. The confidence level of the alarm message can be calculated based on factors such as the confidence level of the topology inference model and the number of abnormal edges.

[0054] The refined line loss allocation module utilizes real-time physical topology and line parameters (or parameter estimation) to distribute the total loss of the distribution area step-by-step to each line segment and node according to the physical path, identifying abnormal line loss sections as secondary alarms. The refined line loss allocation module is input with the real-time inferred physical topology map, node measurement data, and total meter readings for the distribution area. Distribution network physical topology map. It includes nodes (transformers, branch boxes, meters) and edges (line segments), each edge carrying estimated line impedance parameters (obtainable from archives or identified through measurement data). Node measurement data includes the active power of each node (meter). reactive power Voltage amplitude The total power consumption of the transformer substation is the active power of the total meter at the transformer substation's output point. .

[0055] The refined line loss allocation module employs the forward-backward substitution method from distribution network power flow calculation, calculating the active power loss of each branch level by level from the end node to the root node (transformer). The algorithm steps are as follows: (1) Topology layering: Taking the transformer as the root node, the physical topology diagram of the distribution network is... Perform a breadth-first search to obtain the level (depth) of each node. (2) Forward calculation: Starting from the deepest node, calculate the injection current (or complex power) of each node and accumulate it upstream; (3) Back-substitution calculation: Starting from the root node, calculate the voltage of each downstream node based on the upstream voltage and line impedance, and calculate the active power loss of each branch: ;in, The magnitude of the current flowing through this branch. Branch resistance; (4) Accumulate total loss: Add up the losses of all branches to obtain the theoretically calculated total loss. Then, calculate the actual total loss. ; This represents the sum of active power from all meters. Calculate the deviation between theoretical and actual losses: ;like If the loss exceeds the threshold (e.g., 10%), it indicates that there is non-technical loss.

[0056] To further pinpoint abnormal sections, the actual total loss is proportionally allocated to each branch, with the allocation coefficient for each branch being the product of the square of the branch current and its resistance. Then, the abnormal loss of each branch is calculated: ; Branches with abnormal losses exceeding the threshold are marked as abnormal line loss sections, and their abnormality severity is calculated. For each abnormal line loss section, a level-two alarm is generated.

[0057] The user electricity consumption behavior graph construction module combines real-time physical topology with user electricity consumption behavior characteristics to construct an electricity consumption behavior graph. It uses graph mining algorithms to identify abnormal patterns in group electricity consumption, serving as a level-three alarm. During node feature construction, for each user node, behavioral feature vectors are extracted based on its historical electricity consumption time-series data. Optional features include: electricity consumption statistics (daily average electricity consumption, variance, peak-to-valley difference, nighttime electricity consumption percentage, etc.); electricity consumption pattern features (typical pattern coefficients obtained from load curve decomposition, such as latent vectors extracted through an autoencoder); and abnormal event features (historical number of times the cover was opened, number of times voltage exceeded limits, etc.). The electricity consumption behavior graph construction process uses the distribution network physical topology map as a reference. Based on the basic graph structure, the behavioral feature vector of each user node is appended as a node attribute to form an attribute graph. ,in The node feature matrix is ​​used. Graph neural networks or graph clustering algorithms (such as graph autoencoders and GraphSAGE combined clustering) are used to divide the nodes into multiple clusters. The nodes in each cluster are topologically connected and have similar electricity consumption behaviors. Then, the following methods are used to identify abnormal patterns in the clusters: (1) Reconstruction error-based method: Train graph autoencoders to learn the feature representation of normal users. For each cluster, calculate the average reconstruction error of the nodes in the cluster. If the average reconstruction error of a certain cluster is significantly higher than that of other clusters, there may be a group anomaly; (2) Local outlier factor-based method: For each node, calculate the local outlier factor (LOF) of electricity consumption behavior within its topological neighbors. If the LOF value of a certain node and its neighbors is generally high, it is marked as an abnormal group; (3) Community comparison method: Compare the electricity consumption statistics (such as the average proportion of nighttime electricity consumption) of different communities in the same transformer area. If the indicator of a certain community deviates significantly, it is marked. For the discovered abnormal groups, a three-level alarm is generated.

[0058] The multi-level early warning fusion engine integrates Level 1, Level 2, and Level 3 alarms, prioritizes them, and associates them with relevant evidence to provide decision support for investigators. The priority order is as follows: Level 1 alarms (topology mismatch) have the highest confidence and strongest certainty, and are therefore the highest priority; Level 2 alarms (abnormal line loss sections) are next; and Level 3 alarms (abnormal behavior patterns) serve as auxiliary clues and have the lowest priority. Then, alarms of different levels within the same area (such as the same server area or the same branch) are correlated and aggregated to form a comprehensive suspicion report. For example, if a certain area simultaneously has both Level 1 alarms (unauthorized wiring) and Level 2 alarms (abnormal line loss), the overall suspicion level is extremely high.

[0059] The formula for scoring the overall suspicion level is as follows: ;in, These represent the intensity (e.g., confidence level, degree of abnormality) of level 1, 2, and 3 alarms, respectively. As weight, and Calculate the overall suspicion level for all users or regions, and sort them in descending order.

[0060] An inspection work order is automatically generated based on the overall suspicion level. The information in the inspection work order includes: target area (transformer area, branch, specific user); alarm list (details of alarms at all levels); on-site verification suggestions (such as focusing on checking a certain section of line or a few meters); GIS navigation information (map coordinates, path), etc.

[0061] The GIS visualization platform overlays the inferred physical topology and alarm information onto a GIS map, visually displaying the distribution of suspected electricity theft. The data layers in the GIS visualization include: Basic layers: satellite map, street map, power equipment distribution map (transformers, poles).

[0062] Physical topology layer: Displays the inferred physical connections in real time in the form of lines, and can distinguish between different voltage levels and different phases; Archive Topology Layer: Displays the connection relationships of archive records in a dashed overlay format for easy comparison; Alarm layers: Level 1 alarm: Red flashing icon marks topology anomalies (such as unauthorized wiring); Level 2 alarm: Yellow highlight shows abnormal line loss sections; Level 3 alarm: Purple area circles groups with abnormal behavior.

[0063] When inspectors click the alarm icon, a details box pops up displaying alarm information, related users, alarm confidence level, etc. The alarm confidence level can be calculated based on factors such as the confidence level of the topology inference model (as shown in the neural network model) and the number of abnormal edges.

[0064] Compared with existing technical solutions, this application has the following significant technical advantages: 1. This method employs multi-source fusion data to infer the physical topology of the distribution network in real time. The inferred physical topology is then compared with the static archived topology recorded by the marketing system. By calculating the edge set differences between the two graphs, inconsistencies between physical connections and archived records caused by unauthorized wiring, meter bypassing, or line alterations are directly identified and used as highly definitive evidence of electricity theft. This solution does not rely on statistical analysis of user electricity consumption; instead, it utilizes an automatic comparison mechanism between physical and archived topologies to achieve direct and accurate identification of unauthorized wiring.

[0065] 2. A refined line loss allocation method based on real-time physical topology. Utilizing real-time physical topology maps and line impedance parameters, and employing power flow algorithms such as forward and backward substitution, the total loss of the transformer area is allocated level by level along the physical path to each branch and each line segment. The difference between the theoretical and actual losses of each physical segment is calculated, thereby accurately pinpointing line loss anomalies to specific line segments. This physical model-driven loss analysis method achieves precise "line segment-level" location of line loss anomalies.

[0066] 3. By constructing a user electricity consumption behavior graph based on a topology graph, the real-time physical topology is used as the graph structure, and the multi-dimensional electricity consumption characteristics of users are used as node attributes to form an attribute graph. Based on this, algorithms such as graph clustering, graph autoencoders, and graph anomaly detection are applied to analyze the degree of collaborative deviation among nodes within the community and discover group-based abnormal patterns. This method introduces a spatial correlation dimension to achieve intelligent identification of group-based electricity theft patterns.

[0067] 4. A multi-level early warning fusion and priority ranking mechanism is adopted. Alarms are prioritized according to their degree of anomaly certainty, and alarms of different levels in the same area are aggregated and correlated. A weighted comprehensive suspicion score is then calculated using a weighted comprehensive suspicion scoring formula. This multi-level evidence fusion mechanism effectively avoids misjudgments from single methods and significantly improves the reliability and accuracy of detection.

[0068] 5. The GIS-based multi-layer visualization method overlays the inferred physical topology, archive topology, and alarm information at all levels onto the GIS map for visualization, which can improve the efficiency of on-site inspection operations.

[0069] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application 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 this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0070] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and 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.

[0071] 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.

[0072] 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.

[0073] The optional embodiments of this application have been described in detail above with reference to the accompanying drawings. However, the embodiments of this application are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of this application, various simple modifications can be made to the technical solutions of the embodiments of this application, and these simple modifications all fall within the protection scope of the embodiments of this application. 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 the combination does not violate the spirit of the embodiments of this application, it should also be considered as the content disclosed in the embodiments of this application.

Claims

1. A non-technical loss identification and early warning method based on power grid topology inference, characterized in that, include: Based on multi-source fusion data, the physical connection relationship of the current distribution network is inferred in real time through a graph neural network model, and a physical topology map of the distribution network is output. The physical topology map of the distribution network includes a set of nodes and a set of edges representing the physical connection relationship of the nodes. The multi-source fusion data is obtained by associating and fusing real-time measurement data, marketing archive data and historical alarm data of each node in the power grid. The real-time inferred physical topology map of the distribution network is compared with the static archive topology map to identify topology anomaly nodes and generate first-level alarm information containing topology anomaly nodes. Based on the real-time inferred physical topology of the distribution network, the loss of each branch is calculated according to the real-time measurement data. Branches with losses exceeding the preset threshold are identified as abnormal line loss sections, and secondary alarm information containing abnormal line loss sections is generated. Based on the real-time inferred physical topology of the distribution network, an electricity consumption behavior map is constructed according to the characteristics of users' electricity consumption behavior. Based on the electricity consumption behavior map, abnormal electricity consumption groups are identified, and three-level alarm information containing abnormal electricity consumption groups is generated. The first-level alarm information, second-level alarm information and third-level alarm information are integrated and correlated to generate audit information; The graph neural network model is a spatiotemporal graph convolutional network, which includes a spatiotemporal graph encoder and a dual-branch decoder. The spatiotemporal graph encoder includes a temporal convolutional layer and a spatial graph convolutional layer, used to extract spatiotemporal correlation features from the multi-source fused data and output node spatiotemporal embedding feature vectors. The dual-branch decoder includes a voltage reconstruction decoder and an edge prediction decoder. The voltage reconstruction decoder calculates the predicted voltage based on the node spatiotemporal embedding feature vectors, and the edge prediction decoder predicts the node edge probability based on the node spatiotemporal embedding feature vectors. The loss function of the spatiotemporal graph convolutional network includes voltage reconstruction loss and physical consistency loss. The voltage reconstruction loss is the error between the predicted voltage and the actual voltage measurement value. The physical consistency loss is determined based on the node edge probability under the constraint of the node current law.

2. The non-technical loss identification and early warning method based on power grid topology inference according to claim 1, characterized in that, The process of comparing the real-time inferred physical topology map of the distribution network with the static archived topology map to identify topology anomaly nodes and generate first-level alarm information containing these nodes includes: Extract a static archive topology graph from the marketing archive data, the archive topology graph containing a set of nodes and a set of edges for archive records; The real-time inferred physical topology map of the distribution network is compared with the archived topology map to identify edges whose edge sets are inconsistent with those of the archived topology map. These inconsistent edges are identified as topology anomaly edges. A first-level alarm message is generated for each topology anomaly edge, and the first-level alarm message includes the topology anomaly node associated with the topology anomaly edge.

3. The non-technical loss identification and early warning method based on power grid topology inference according to claim 1, characterized in that, The distribution network physical topology diagram based on real-time inference calculates the losses of each branch based on real-time measurement data, including: A breadth-first search is performed on the physical topology of the distribution network to obtain the hierarchy of each node; Starting from the deepest node, the injected current of each node is calculated based on the real-time measurement data of each node, and then accumulated upstream to obtain the upstream voltage. Starting from the root node, calculate the voltage of each downstream node based on the upstream voltage and line impedance, and calculate the active power loss of each branch.

4. The non-technical loss identification and early warning method based on power grid topology inference according to claim 1, characterized in that, The distribution network physical topology diagram based on real-time inference calculates the losses of each branch based on real-time measurement data, including: Calculate the actual total loss of the transformer area based on real-time measurement data; The actual total loss is proportionally allocated to each branch to obtain the loss of each branch; where the allocation coefficient of each branch is the product of the square of the branch current and the resistance.

5. The non-technical loss identification and early warning method based on power grid topology inference according to claim 1, characterized in that, The power distribution network physical topology map based on real-time inference constructs an electricity consumption behavior map based on users' electricity consumption behavior characteristics, including: For each user node, extract the electricity consumption behavior feature vector based on the user's historical electricity consumption time-series data; Using the physical topology of the distribution network as the basic graph structure, the electricity consumption behavior feature vector of each user node is added as a node attribute to form an attribute graph, which is then used as the electricity consumption behavior map.

6. The non-technical loss identification and early warning method based on power grid topology inference according to claim 5, characterized in that, The process of identifying groups with abnormal electricity consumption behavior based on electricity consumption behavior patterns includes: Based on the electricity consumption behavior map, the nodes are divided into multiple communities using graph neural networks or graph clustering algorithms. The nodes in each community are connected on the distribution network physical topology map and have similar electricity consumption behaviors. Identify groups with unusual electricity consumption behaviors in multiple communities.

7. The non-technical loss identification and early warning method based on power grid topology inference according to claim 6, characterized in that, The identification of abnormal electricity consumption groups within multiple communities includes: For each community, calculate the average reconstruction error of the nodes within the community. If the average reconstruction error of a community is higher than the average reconstruction error of other communities, the community is identified as a group with abnormal electricity consumption behavior. Alternatively, for each node, calculate the local outlier value of electricity consumption behavior within the node's topological neighbors. If the local outlier value of a node and its neighbors is higher than the normal value, determine that the node and its neighbors as a group with abnormal electricity consumption behavior. Alternatively, compare the electricity consumption statistics of different communities in the same transformer area. If the electricity consumption statistics of a certain community are significantly higher, then that community is identified as a group with abnormal electricity consumption behavior.

8. The non-technical loss identification and early warning method based on power grid topology inference according to claim 1, characterized in that, The process of merging and associating Level 1, Level 2, and Level 3 alarm information to generate audit information includes: Alarm messages of Level 1, Level 2, and Level 3 are prioritized; Level 1 alarm messages have the highest priority, and Level 3 alarm messages have the lowest priority. Calculate the overall suspicion level for all users or regions, associate alarm information of different priorities in the same region based on the overall suspicion level, and generate audit information based on the overall suspicion level.

9. The non-technical loss identification and early warning method based on power grid topology inference according to claim 1, characterized in that, The method further includes: The real-time inferred physical topology of the distribution network and alarm information are overlaid on the geographic information system map and displayed on the visualization interface.

10. A non-technical loss identification and early warning system based on power grid topology inference, characterized in that, include: The data source layer is used to acquire real-time measurement data, marketing archive data, and historical alarm data, and to correlate and merge them to obtain multi-source fused data; The data processing layer is used for: Based on multi-source fusion data, the physical connection relationship of the current distribution network is inferred in real time through a graph neural network model, and the physical topology map of the distribution network is output. The physical topology map of the distribution network includes a set of nodes and a set of edges representing the physical connection relationship of the nodes. The real-time inferred physical topology map of the distribution network is compared with the static archive topology map to identify topology anomaly nodes and generate first-level alarm information containing topology anomaly nodes. Based on the real-time inferred physical topology of the distribution network, the loss of each branch is calculated according to the real-time measurement data. Branches with losses exceeding the preset threshold are identified as abnormal line loss sections, and secondary alarm information containing abnormal line loss sections is generated. Based on the real-time inferred physical topology of the distribution network, an electricity consumption behavior map is constructed according to the characteristics of users' electricity consumption behavior. Based on the electricity consumption behavior map, abnormal electricity consumption groups are identified, and three-level alarm information containing abnormal electricity consumption groups is generated. The early warning and display layer is used to integrate and correlate first-level, second-level, and third-level alarm information to generate audit information; The graph neural network model is a spatiotemporal graph convolutional network, which includes a spatiotemporal graph encoder and a dual-branch decoder. The spatiotemporal graph encoder includes a temporal convolutional layer and a spatial graph convolutional layer, used to extract spatiotemporal correlation features from the multi-source fused data and output node spatiotemporal embedding feature vectors. The dual-branch decoder includes a voltage reconstruction decoder and an edge prediction decoder. The voltage reconstruction decoder calculates the predicted voltage based on the node spatiotemporal embedding feature vectors, and the edge prediction decoder predicts the node edge probability based on the node spatiotemporal embedding feature vectors. The loss function of the spatiotemporal graph convolutional network includes voltage reconstruction loss and physical consistency loss. The voltage reconstruction loss is the error between the predicted voltage and the actual voltage measurement value. The physical consistency loss is determined based on the node edge probability under the constraint of the node current law.

11. The non-technical loss identification and early warning system based on power grid topology inference according to claim 10, characterized in that, The process of comparing the real-time inferred physical topology map of the distribution network with the static archived topology map to identify topology anomaly nodes and generate first-level alarm information containing these nodes includes: Extract a static archive topology graph from the marketing archive data, the archive topology graph containing a set of nodes and a set of edges for archive records; The real-time inferred physical topology map of the distribution network is compared with the archived topology map to identify edges whose edge sets are inconsistent with those of the archived topology map. These inconsistent edges are identified as topology anomaly edges. A first-level alarm message is generated for each topology anomaly edge, and the first-level alarm message includes the topology anomaly node associated with the topology anomaly edge.

12. The non-technical loss identification and early warning system based on power grid topology inference according to claim 10, characterized in that, The distribution network physical topology diagram based on real-time inference calculates the losses of each branch based on real-time measurement data, including: A breadth-first search is performed on the physical topology of the distribution network to obtain the hierarchy of each node; Starting from the deepest node, the injected current of each node is calculated based on the real-time measurement data of each node, and then accumulated upstream to obtain the upstream voltage. Starting from the root node, calculate the voltage of each downstream node based on the upstream voltage and line impedance, and calculate the active power loss of each branch.

13. The non-technical loss identification and early warning system based on power grid topology inference according to claim 10, characterized in that, The power distribution network physical topology map based on real-time inference constructs an electricity consumption behavior map based on users' electricity consumption behavior characteristics, including: For each user node, extract the electricity consumption behavior feature vector based on the user's historical electricity consumption time-series data; Using the physical topology of the distribution network as the basic graph structure, the electricity consumption behavior feature vector of each user node is added as a node attribute to form an attribute graph, which is then used as the electricity consumption behavior map.

14. The non-technical loss identification and early warning system based on power grid topology inference according to claim 13, characterized in that, The process of identifying groups with abnormal electricity consumption behavior based on electricity consumption behavior patterns includes: Based on the electricity consumption behavior map, the nodes are divided into multiple communities using graph neural networks or graph clustering algorithms. The nodes in each community are connected on the distribution network physical topology map and have similar electricity consumption behaviors. Identify groups with unusual electricity consumption behaviors in multiple communities.

Images