Transformer transformer low-voltage line loss anomaly diagnosis method and device based on graph convolutional neural network

Abstract

The application discloses a kind of based on graph convolutional neural network's low voltage line loss abnormal diagnosis method and device of table area, it is related to distribution network line loss intelligent diagnosis field.The method abstracts table area node as graph node, connection relation is abstracted as edge, with operating data and data quality information as node feature, input graph convolutional neural network aggregation neighborhood feature, output table area line loss state and node abnormal probability;When detecting abnormality, the abnormal score of node is calculated by fusing power imbalance sensitivity, spatial correlation similarity and data quality abnormality degree, tracing along power supply path, outputting suspected abnormal node and the like information.The device includes corresponding module for implementing the above method.The present application can distinguish normal reverse sending electricity from metering file abnormality, reduce data misjudgment, refine diagnosis to branch and user level, and improve line loss management efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnosis of distribution network line losses, and particularly to a method for diagnosing abnormal low-voltage line losses in transformer substations based on graph convolutional neural networks, and also to a device for diagnosing abnormal low-voltage line losses in transformer substations based on graph convolutional neural networks. Background Technology

[0002] Low-voltage distribution transformer areas are located at the end of the power supply and consumption chain in the power system, and their line loss levels directly affect the operating efficiency, metering accuracy, and management level of the distribution network. In actual operation, transformer area line losses are affected by a combination of factors, including line impedance, power supply radius, conductor cross-section, load distribution, three-phase imbalance, metering device status, data acquisition quality, accuracy of household-transformer relationships, user electricity consumption behavior, and distributed renewable energy backfeeding. When abnormally high or negative line losses occur in a transformer area, failure to promptly identify the cause of the anomaly will result in a waste of power resources and affect the efficiency of line loss management and the level of lean management of the distribution network.

[0003] Currently, low-voltage transformer area line loss diagnosis mainly relies on methods such as manual inspection, threshold judgment, power balance analysis, and machine learning analysis. Manual inspection requires staff to check transformer area files, on-site wiring, meter operating status, data collection records, and user electricity consumption data one by one, which is time-consuming, inefficient, and highly dependent on personnel experience. Threshold judgment and power balance analysis rely solely on indicators such as power input, power output, and line loss rate to identify anomalies, making it difficult to accurately distinguish between metering anomalies, data collection anomalies, file anomalies, line loss anomalies, and differences between normal load fluctuations. Although traditional machine learning methods can build identification models using historical data, they typically treat data from each monitoring point as tabular features, focusing primarily on changes in electrical quantities of individual users or the transformer area's main meter. In fact, the nodes in a low-voltage distribution area are not independent of each other. There are clear upstream and downstream power supply relationships between the low-voltage side of the distribution transformer, low-voltage outgoing lines, branch boxes, meter boxes, and user electricity meters. An abnormality in a user's meter, a wiring error in a meter box, a failure to collect data in a branch box, or an abnormality in the reverse active power at a grid connection point will all affect the power balance results of the upstream branch box, low-voltage outgoing lines, and the total meter side of the distribution area along the power flow path. This makes it difficult for existing machine learning analysis to accurately determine the source of the anomaly, often resulting in an excessively large range of suspected anomalies, making it difficult to accurately guide on-site troubleshooting.

[0004] With the large-scale integration of distributed photovoltaic and other new energy sources, the operational status of distribution transformer areas has become increasingly complex. Distributed new energy grid connection points may generate reverse active power, causing changes in the power flow direction within the transformer area. During certain time periods, an increase in reverse active power on the user side may lead to a decrease in the line loss rate or even negative losses in the transformer area. This phenomenon may originate from normal reverse power feeding, or it may be related to factors such as incorrect metering direction, incorrect grid connection point file assignment, asynchronous data collection times, or abnormal main and sub-meter multipliers. Existing line loss diagnosis methods are prone to misjudging normal reverse power feeding as abnormal negative losses, and may also overlook genuine anomalies caused by metering or file issues, still requiring manual investigation.

[0005] Furthermore, the low-voltage distribution area operation data itself suffers from quality issues such as insufficient data collection completeness, communication interruptions, abnormal meter operating status, and time synchronization problems. These data issues directly affect the line loss calculation results and resemble actual abnormal power consumption, often leading to system misjudgments. These issues can only be addressed by maintenance personnel checking each problem individually, making it difficult to distinguish data quality problems from actual power consumption anomalies. Summary of the Invention

[0006] This invention proposes a method and device for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks. Its purpose is to address the problems of existing low-voltage transformer substation line loss diagnosis methods that fail to fully utilize the physical topology of the substations, ignore upstream and downstream power supply relationships between nodes leading to inaccurate anomaly location, misjudging normal backfeeding as abnormal negative losses or missing real anomalies during distributed renewable energy backfeeding, and data quality fluctuations causing system misjudgments that are difficult to distinguish from actual power consumption anomalies. This invention aims to improve the rationality and accuracy of line loss anomaly diagnosis, reduce the scope of on-site investigation, and enhance the efficiency of line loss management.

[0007] The technical solution of this invention is as follows:

[0008] A method for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks includes:

[0009] Obtain the physical topology data of the low-voltage distribution transformer area, abstract the low-voltage side of the distribution transformer, low-voltage outgoing lines, branch boxes, user meter boxes, user energy meters, and distributed new energy grid connection points as nodes in the graph structure, abstract the actual line connection relationship between nodes as edges in the graph structure, construct a topology graph structure representing the power transmission relationship of the transformer area, and construct the corresponding adjacency matrix.

[0010] The running data and data quality information of each node within a preset time window are collected. After preprocessing, the running data and data quality information are combined to construct a node dynamic feature vector. A node feature matrix is ​​formed based on the dynamic feature vectors of all nodes.

[0011] The adjacency matrix and the node feature matrix are input into a graph convolutional neural network. The graph convolutional layer propagates and aggregates the node's own features and the features of its neighboring nodes to obtain the substation-level implicit representation and the node-level implicit representation. The substation line loss status is output based on the substation-level implicit representation.

[0012] When the line loss status of the transformer area is abnormally high or abnormally negative, the abnormal probability of each node is output according to the implicit representation of the node level. The abnormality score of the node is calculated by combining the influence of the node power change on the line loss index of the upper branch or transformer area, the spatial correlation similarity between the node and the neighboring nodes, and the abnormality of the node data quality.

[0013] Based on the topological connection relationship of the transformer area and the direction of power flow, the system searches step by step along the power supply path. Combining the node anomaly score and the branch power imbalance, the cumulative anomaly contribution of the path is calculated, a priority verification path is generated, and the diagnostic results are output. The diagnostic results include the transformer area line loss status, anomaly type tendency, a list of suspected anomaly nodes, and the priority verification path.

[0014] As a further improvement to the method for diagnosing abnormal low-voltage line losses in transformer substations based on graph convolutional neural networks, the operating data includes three-phase current, three-phase voltage, power factor, forward active power, reverse active power, net active power, and distributed renewable energy output data.

[0015] The data quality information includes data collection completeness rate, meter operating status, communication status, and time synchronization status.

[0016] The dynamic feature vector also includes node type identifier, topology level, branch number, phase affiliation, identifier of whether it has reverse power transmission capability, and identifier of whether there is a suspected anomaly in the topology file.

[0017] As a further improvement to the method for diagnosing abnormal low-voltage line loss in transformer areas based on graph convolutional neural networks, self-connections are added to the adjacency matrix and normalization is performed; when line parameters are available, a weighted adjacency matrix is ​​constructed using edge weights, and the edge weights are determined based on line length, conductor cross-section, or historical load correlation.

[0018] As a further improvement to the aforementioned method for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks, the formula for calculating the node anomaly score is as follows:

[0019]

[0020] in, Represents a node In the time window The probability of node anomalies, Represents a node In the time window Sensitivity to power imbalance Represents a node Its neighboring nodes in the time window The average spatial association similarity is determined based on the node-level implicit representations of a node and its neighboring nodes. Represents a node In the time window The degree of data quality anomalies, , , , This represents the weighting coefficient.

[0021] As a further improvement to the graph convolutional neural network-based method for diagnosing low-voltage line loss anomalies in transformer substations, the power imbalance sensitivity in the node anomaly scoring is calculated using the following formula:

[0022]

[0023] in, Indicates time window Includes nodes Time branch The amount of power imbalance, Indicates within the time window Remove nodes in the following calculation Or for nodes The branch power imbalance is estimated normally after replacement, and the power imbalance sensitivity is... Used for quantization nodes The degree of contribution to the imbalance of branch power.

[0024] As a further improvement to the aforementioned method for diagnosing abnormal low-voltage line losses in transformer substations based on graph convolutional neural networks, nodes The average spatial association similarity with its neighboring nodes is calculated by the following formula:

[0025]

[0026] in, Represents a node The set of neighboring nodes;

[0027] For nodes With nodes In the time window The spatial similarity is calculated using the following formula:

[0028]

[0029] in, and Representing nodes respectively and nodes Hidden features in the last layer of graph convolution.

[0030] As a further improvement to the graph convolutional neural network-based method for diagnosing low-voltage line loss anomalies in transformer substations, the degree of data quality anomaly in the node anomaly score is calculated using the following formula:

[0031]

[0032] in, Represents a node In the time window The completeness of the data collection Represents a node In the time window The degree of data quality anomaly identification.

[0033] As a further improvement to the graph convolutional neural network-based method for diagnosing low-voltage line loss anomalies in transformer substations, the node anomaly probability in the node anomaly score is calculated using the following formula:

[0034]

[0035] in, and These represent the weights and biases of the node-level output layer, respectively. For nodes The node-level implicit representation is the implicit feature output by the last layer of graph convolution.

[0036] As a further improvement to the graph convolutional neural network-based method for diagnosing low-voltage line loss anomalies in transformer substations, the step-by-step search along the power supply path includes:

[0037] For scenarios with abnormally high losses, the search proceeds step by step along the forward power supply path from the low-voltage side of the distribution transformer to the user side.

[0038] For abnormal negative loss scenarios, in addition to searching downstream from the low-voltage side of the distribution transformer along the forward power supply path, we also start from the distributed new energy grid connection point or user node with reverse active power and trace the reverse impact path upstream. Based on whether the changes in reverse active power, net active power and negative loss change trends are consistent, as well as the metering direction, topology affiliation and collection status of the grid connection point, we determine whether it is normal reverse power supply or abnormal.

[0039] This invention also discloses a low-voltage line loss anomaly diagnosis device for transformer substations based on graph convolutional neural networks, comprising:

[0040] The topology construction module is used to obtain the physical topology data of the low-voltage distribution transformer area, and abstract the low-voltage side of the distribution transformer, low-voltage outgoing lines, branch boxes, user meter boxes, user energy meters and distributed new energy grid connection points as nodes in the graph structure, and abstract the actual line connection relationship between nodes as edges in the graph structure to construct a topology graph structure representing the power transmission relationship of the transformer area, and construct the corresponding adjacency matrix.

[0041] The data acquisition and preprocessing module is used to collect the running data and data quality information of each node within a preset time window. After preprocessing, the running data and data quality information are combined to construct a node dynamic feature vector, and a node feature matrix is ​​formed based on the dynamic feature vectors of all nodes.

[0042] A graph convolutional neural network is used to take the adjacency matrix and the node feature matrix as input, and to propagate and aggregate the node's own features and the features of its neighboring nodes through graph convolutional layers to obtain the substation-level hidden representation and the node-level hidden representation. The substation-level hidden representation is then used to output the substation line loss status.

[0043] The anomaly scoring module is configured to: when the line loss status of the transformer area is abnormally high loss or abnormally negative loss, output the anomaly probability of each node according to the node-level implicit representation, and calculate the node anomaly score by combining the influence of node power changes on the line loss index of the upper branch or transformer area, the spatial correlation similarity between the node and neighboring nodes, and the degree of node data quality anomaly.

[0044] The tracing and output module is used to search along the power supply path step by step according to the topological connection relationship of the transformer area and the direction of power flow. It calculates the cumulative abnormal contribution of the path by combining the node anomaly score and the branch power imbalance, generates the priority verification path, and outputs the diagnostic results including the transformer area line loss status, anomaly type tendency, suspected abnormal node list and priority verification path.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. This invention introduces the physical topology of low-voltage distribution transformer areas into line loss diagnosis. It abstracts each power supply and consumption link as a graph structure node and the actual line connection relationship as an edge. Through a graph convolutional neural network, it propagates and aggregates the node's own features and neighborhood features along the topology structure. This allows the model to simultaneously utilize the node's own operating status and the operating information of upstream and downstream adjacent nodes to learn the correlations such as load changes, power accumulation, branch power supply, and reverse power flow. This captures the causal chain of anomalies propagating along the power flow path, and the anomaly location results are consistent with the actual power supply path of the transformer area, improving the rationality and accuracy of diagnosing abnormal high losses and abnormal negative losses.

[0047] 2. In the construction of the topology graph, this invention introduces a weighted adjacency matrix, which determines the edge weight based on the line length, conductor cross-section or historical load correlation, quantifies the degree of influence of different lines on power flow transmission, and balances the influence of the difference in the number of neighbors of different nodes on the feature aggregation result through self-connection and normalization processing. This not only preserves the characteristics of the terminal nodes, but also improves the accuracy of feature aggregation and enhances the sensitivity to long lines and high impedance branches.

[0048] 3. To address the reliability issue of collected data, this invention incorporates data quality information such as data completeness rate, meter operating status, communication status, and time synchronization status as node features, inputting them along with the operating data into a graph convolutional neural network. This allows the model to synchronously perceive the data reliability of each node during power balance analysis. By using neighborhood aggregation, it automatically reduces the impact of low-quality data on the implicit representations of upstream and downstream nodes, thereby reducing misjudgments caused by data collection failures, missing frozen power, abnormal meter operating status, or time asynchrony. This distinguishes data quality issues from actual power consumption anomalies, improving the reliability of diagnostic results.

[0049] 4. In the scenario of distributed renewable energy access, this invention treats the grid connection point of distributed renewable energy as an independent graph node to participate in line loss analysis. When abnormal negative loss is detected, in addition to searching along the forward power supply path, it also traces the reverse impact path upstream from the grid connection point or user node with reverse active power. Combining the changes in reverse active power, net active power and negative loss change trends, as well as the metering direction, topology affiliation and collection status, it distinguishes between normal reverse power transmission and real anomalies such as abnormal metering direction, abnormal grid connection point affiliation or asynchronous collection time, thereby reducing the probability of false positives and false negatives.

[0050] 5. After detecting abnormal line loss in a transformer area, this invention calculates a node anomaly score by integrating the node anomaly probability, power imbalance sensitivity, spatial correlation similarity, and data quality anomaly degree output by the fusion model. It then accumulates the anomaly contribution level by level along the power supply path, outputting suspected abnormal nodes, suspected abnormal branches, and priority verification paths. This allows the diagnostic results to be further refined from transformer area-level status identification to the branch and user levels. Simultaneously, it quantifies the node anomaly score and the impact path, improving the interpretability of the results. This provides data support for metering device verification, household transformer relationship verification, branch line verification, distributed new energy grid connection point verification, and user abnormal power consumption investigation.

[0051] 6. The method and apparatus described in this invention can be trained and operated based on existing data from existing electricity information collection systems and distribution automation systems without the need for additional dedicated sensing devices. Furthermore, the results of manual verification, metering anomalies, collection anomalies, and line loss mitigation can be written back to the sample library for periodic fine-tuning or incremental training of the model. This allows the model to gradually adapt to new operating states as the load structure of the transformer area changes, the proportion of new energy access changes, and the topology is updated. This makes it suitable for large-scale deployment, reduces the dependence of maintenance personnel on experience, and improves the efficiency of line loss mitigation. Attached Figure Description

[0052] Figure 1 A flowchart of a method for diagnosing abnormal low-voltage line loss in transformer substations based on graph convolutional neural networks provided in an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram illustrating an application scenario of the graph convolutional neural network-based method for diagnosing abnormal low-voltage line losses in transformer substations, as provided in an embodiment of the present invention.

[0054] Figure 3 This is a schematic diagram of the weighted fusion of node anomaly scores in an embodiment of the present invention;

[0055] Figure 4 This is a structural block diagram of a low-voltage line loss anomaly diagnosis device based on graph convolutional neural network provided in an embodiment of the present invention. Detailed Implementation

[0056] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0057] Example 1

[0058] This embodiment provides a method for diagnosing abnormal low-voltage line losses in distribution transformer areas based on graph convolutional neural networks. This method is applied to low-voltage distribution transformer areas to locate abnormal line losses.

[0059] Specifically, this method abstracts the low-voltage side of the distribution transformer, low-voltage outgoing lines, branch boxes, user meter boxes, user energy meters, and distributed new energy grid connection points as graph nodes, and the actual line connection relationships between nodes as edges. It uses the operational data and data quality information of each node as node features, inputting them into a graph convolutional neural network for topological neighborhood feature aggregation, and outputting the line loss status of the distribution area and the probability of anomalies for each node. When the line loss status of the distribution area is abnormally high or abnormally negative, it further integrates the node's power imbalance sensitivity, spatial correlation similarity, and data quality anomaly degree to calculate the node anomaly score. Tracing the source along the power flow path, it outputs suspected abnormal nodes, branches, and priority verification paths. This fully utilizes the physical topology of the distribution area, effectively distinguishing between normal backfeeding of distributed new energy and metering / archive anomalies, reducing misjudgments caused by data acquisition quality issues, and extending the diagnostic results from the distribution area level to the branch and user levels. It provides clear priority paths for on-site verification, significantly improving the efficiency of line loss management, and solving the problem that existing technologies cannot accurately identify abnormally high or abnormally negative losses and locate abnormal nodes when distributed new energy backfeeds or data quality fluctuates.

[0060] refer to Figure 1 The method includes steps S1, S2, S3, and S4. In its implementation, the method is divided into two phases. The first phase is the offline modeling phase, which includes topology graph construction in step S1, feature construction and sample labeling in step S2, and model training in step S3, training the graph convolutional neural network based on historical operating data of the transformer substations. The second phase is the online diagnosis phase, i.e., step S4, where the latest collected operating data is input into the trained graph convolutional neural network through the same data processing flow, outputting the line loss status of the transformer substations, and performing node anomaly scoring and source tracing analysis when anomalies are detected. Both phases share the same topology graph structure and feature construction method.

[0061] Step S1: Obtain the physical topology data of the low-voltage distribution transformer area, construct a topology graph structure representing the power transmission relationship of the transformer area, and construct the corresponding adjacency matrix.

[0062] refer to Figure 2A low-voltage distribution substation may include branch boxes 21, 22, and 23 extending from the low-voltage side 10 of the distribution transformer; user electricity meters 31 and 32 extending from branch box 21; distributed renewable energy grid connection point 41 and user electricity meter 33 extending from branch box 22; and user electricity meter 34 and distributed renewable energy grid connection point 42 extending from branch box 23. In practical implementation, the physical topology data of the low-voltage distribution substation is first acquired. This topology data can originate from the substation archive system, the distribution network geographic information system, the electricity consumption information collection system, or on-site verification data. The topology data includes the connection relationships between the low-voltage side of the distribution transformer and the low-voltage outgoing lines, the connection relationships between the low-voltage outgoing lines and the branch boxes, the connection relationships between the branch boxes and the user meter boxes, the connection relationships between the user meter boxes and the user electricity meters, and the connection relationships between the distributed renewable energy grid connection point and the substation lines.

[0063] In step S1, after acquiring the data, the equipment number, user number, meter box number, and line branch number of the transformer area are uniformly coded to eliminate naming differences for the same equipment in different systems. For missing or conflicting topology information, verification can be performed by combining transformer area files, data acquisition terminal affiliation, on-site wiring records, and historical line loss management records.

[0064] After obtaining the distribution area topology, the low-voltage side of the distribution transformer, low-voltage outgoing nodes, branch boxes, user meter boxes, user energy meters, and distributed new energy grid connection points are abstracted as nodes in the graph structure, and the physical line connection relationships between nodes are abstracted as edges in the graph structure.

[0065] Specifically, the diagram structure of the low-voltage distribution substation is set as follows:

[0066]

[0067] in, Represents a set of nodes. Denotes the set of edges. This represents the adjacency matrix.

[0068] There are a total of [number] units within the [area name] area. If there are nodes, then the adjacency matrix It can be represented as:

[0069]

[0070] When constructing a graph structure, if there is a direct electrical connection between two nodes, a corresponding edge is created in the graph structure. The adjacency matrix is ​​undirected, i.e. Edge direction information is not involved in the construction of the adjacency matrix; it is only used in subsequent source tracing analysis to determine the forward or reverse search path along the power flow. For scenarios with distributed renewable energy backflow, the reverse power flow does not change the existence relationship of physical topology edges, but is reflected through node characteristics such as reverse active power, net active power, and grid connection point operating status.

[0071] When constructing the graph structure, edge weights can be assigned based on line length, conductor cross-section, line type, connection type, or historical load correlation. If a line branch is long, has a small conductor cross-section, or has a high historical loss contribution, then that branch can be assigned a corresponding weight.

[0072] Specifically, the weighted adjacency matrix can be represented as:

[0073]

[0074] That is, only when the node With nodes Only edges with direct electrical connections are added to the weighted adjacency matrix; edges without connections are marked as 0. Subsequent self-connection and normalization calculations are directly based on the weighted adjacency matrix. conduct.

[0075] Specifically, the edge weights can be expressed as:

[0076]

[0077] in, The impedance of the line is determined by referring to a table based on the conductor type, cross-section, and length, or by taking the ratio of the line length to the conductor cross-section. Get Node With nodes The absolute values ​​of the Pearson correlation coefficients for the net active power series of each historical time window; For nodes With nodes The ratio of losses to losses in historical normal samples; , , The preset weighting coefficients all have values ​​greater than or equal to zero, and Preferred , , .

[0078] If detailed data such as line impedance and conductor parameters are missing, the edge weight is directly set to 1. That is, the weighted adjacency matrix degenerates into an adjacency matrix that only represents connection relationships.

[0079] To facilitate computation in graph convolutional neural networks, in the weighted adjacency matrix By adding self-joins, we can ensure that each node has a degree of at least 1, preserve the characteristics of terminal nodes, and prevent isolated nodes from disappearing. This results in:

[0080]

[0081] in, Represents the identity matrix.

[0082] according to Calculate the degree matrix:

[0083]

[0084] The normalized adjacency matrix is ​​obtained:

[0085]

[0086] The normalized adjacency matrix is ​​used to balance the impact of differences in the number of neighbors of different nodes on the feature aggregation results. When refined data is lacking, the above self-joining and symmetric normalization processes are performed in the same way.

[0087] Step S1 constructs the objective physical connections in the low-voltage distribution area into a graph structure, enabling the subsequent model to use topological relationships for feature propagation, fully reflecting the upstream and downstream power transmission relationships, laying the foundation for accurately identifying abnormal propagation paths. Furthermore, self-connection and normalization processes ensure the numerical stability of the graph convolutional layer, and edge weight options help improve the model's sensitivity to long lines and high-impedance branches.

[0088] Step S2: Collect the running data and data quality information of each node within a preset time window. After preprocessing, combine the running data and data quality information to construct a node dynamic feature vector. Based on the dynamic feature vectors of all nodes, form a node feature matrix. This step specifically includes the following sub-steps.

[0089] Step S2-1: Data Acquisition. Collect operational data from each monitoring point at a preset time granularity. The time granularity can be determined based on the data integrity and line loss analysis requirements of the electricity information acquisition system, such as 15 minutes, 30 minutes, one hour, or one day. The collected operational data includes one or more of the following: three-phase current, three-phase voltage, power factor, forward active power, reverse active power, net active power, and distributed renewable energy output data.

[0090] Step S2-2: Data Preprocessing. After collecting the operational data, preprocessing is performed, including handling abnormal null values, time alignment, duplicate data removal, missing data imputation, correction of obvious errors, and data normalization. For data gaps caused by acquisition failures, communication anomalies, missing frozen power, or time asynchrony, imputation is performed using data from adjacent time points, adjacent nodes within the same branch, or historical data from the same period. Acquisition completeness rate, meter operating status, communication status, and time synchronization status are retained as auxiliary features. Data that significantly exceeds equipment capacity or does not conform to metering logic is marked as abnormal acquisition data, and its impact on data quality is reflected in subsequent node anomaly scoring to avoid directly equating data quality issues with actual power consumption anomalies.

[0091] For operational data with different dimensions, normalization is applied:

[0092]

[0093] in, Indicates the first The node of the first Feature values ​​before normalization of class features Indicates the first The mean of the class features in the training samples, Indicates the first The standard deviation of class features in the training samples.

[0094] Step S2-3: Construction of Node Dynamic Feature Vectors. After data preprocessing, multiple types of runtime data from the same node within the same time window are combined into a node dynamic feature vector. Specifically, the node... In the time window The dynamic feature vector can be represented as:

[0095] ;

[0096] in, , , Represents a node In the time window The three-phase current, , , Indicates three-phase voltage. Indicates the power factor. Indicates positive active power. Indicates the reverse active power. Indicates net active power. This indicates the data collection completeness rate, which is the percentage collected within the time window. internal nodes The percentage of successfully collected data points out of the total number of data points that should have been collected. This indicates the meter's operating status, derived from the metering automation system. Status messages include: normal, undervoltage, undercurrent, low battery, clock skew, and programming logs. This serves as a data quality identifier, indicating whether there are issues such as excessive time synchronization deviation, communication interruption, missing frozen battery power, or abnormal data markings. This indicates the characteristics of three-phase imbalance. Represents a node The node types are used to differentiate the low-voltage side of distribution transformers, low-voltage outgoing lines, branch boxes, user meter boxes, user energy meters, and distributed new energy grid connection points. Represents a node The topology hierarchy is used to represent the upstream and downstream positions of nodes in the power supply path of the transformer area. Represents a node The branch number is used to indicate the low-voltage branch or meter box range to which the node belongs. Represents a node The different belongings, Represents a node Does it have reverse power transmission capability? Represents a node Does the topology file contain any suspected anomalies?

[0097] The aforementioned electrical quantity characteristics, including three-phase current, three-phase voltage, power factor, forward active power, reverse active power, and net active power, are substituted with normalized values. Node types use one-hot encoding, with the encoding dimension consistent with the total number of node types within the distribution area; the topology hierarchy is a numerical hierarchy, preferably with the low-voltage side of the distribution transformer as the first level, increasing downwards level by level; the branch is represented by the one-hot encoding of the branch number or a numerical mapping after unified encoding; phase attribution uses one-hot encoding; whether it has reverse power transmission capability and whether there are suspected anomalies in the topology file are indicated by binary 0 / 1 identifiers; the data collection completeness rate and status identifier are quantities with clearly defined value ranges and are no longer subject to dimension normalization; data quality identifiers use binary 0 / 1 identifiers; and the three-phase imbalance is a dimensionless ratio.

[0098] Net active power It can be represented as:

[0099]

[0100] That is, the difference between the positive active power and the negative active power.

[0101] The characteristics of three-phase imbalance can be expressed as:

[0102]

[0103] in, The average value of the three-phase current can be expressed as:

[0104]

[0105] Three-phase imbalance characteristics Used to describe the degree of three-phase load imbalance at a node. When a node does not have the conditions for three-phase current acquisition, the corresponding feature bit is set to zero and a missing identifier is set, or the three-phase imbalance characteristics of the meter box or branch box to which it belongs are used as auxiliary information.

[0106] Data collection completeness and meter operating status do not directly indicate the amount of electricity consumed by users, but rather reflect the reliability of the data at that node. When a node experiences abnormal electricity consumption changes and a low data collection completeness, abnormal meter operating status, or abnormal time synchronization, the model will treat that node as a suspected data quality anomaly in subsequent anomaly scoring to reduce the risk of directly classifying data collection issues as actual electricity consumption anomalies.

[0107] All node types use a unified feature template, but the emphasis of feature distribution differs among node types. For ordinary user nodes, forward active power, reverse active power, net active power, power factor, and current variation are the main effective features. For distributed renewable energy grid-connected points, reverse active power and net active power are the main effective features. For transformer low-voltage side and branch box nodes, aggregated current and three-phase imbalance are the main effective features. In the unified template, features from different node types are not included in the data source or are missing, thus reflecting node type differences while ensuring the consistency of feature vector dimensions across all nodes.

[0108] Set time window The feature matrices of all nodes are:

[0109]

[0110] in, This indicates that the graph convolutional neural network operates within a time window. The node feature input.

[0111] Through steps S2-3, the original electrical quantities, data quality information, and topology attributes are uniformly encoded into node features, enabling the subsequent model to simultaneously learn the patterns of power changes, data reliability, and the role of nodes in the transformer area. In particular, by using the collected data quality and meter operating status as feature inputs, the model can automatically reduce the impact of low-quality data.

[0112] Step S2-4: Distribution Area Line Loss Rate Calculation and Sample Labeling. Within each time window, calculate the distribution area line loss rate or branch line loss index based on the power supplied to the distribution area, the user-side settlement power, the reverse active power, and the necessary correction power. Specifically, the distribution area line loss rate can be calculated as follows:

[0113]

[0114] in, Indicates time window Electricity supplied to the main meter in the inner substation area This represents the user-side settlement electricity or the aggregated electricity value of lower-level nodes obtained according to the statistical caliber of line loss in the transformer area. This indicates the corrected electricity amount corresponding to factors such as meter replacement, supplementary payments, refunds, reverse active power from distributed renewable energy sources, data collection gap correction, and metering caliber adjustments. When there are differences in the statistical caliber of line losses in different regions or systems... It can be determined according to the actual measurement rules.

[0115] The calculated transformer area line loss rate Compare with the preset management threshold. If the transformer area line loss rate... If the loss rate exceeds the high-loss threshold and persists for a preset time, such as exceeding 10% for three consecutive time windows, the sample is marked as an abnormally high-loss sample. The sample labeling in steps S2-4 is used for historical sample annotation during the training phase; the transformer area line loss status during the online diagnosis phase is determined by the model output. If the transformer area line loss rate... If the line loss rate is below a reasonable lower limit and continues to reach a preset time, such as being 0% for three consecutive time windows or a set negative loss threshold, then the sample is marked as an abnormal negative loss sample. If the transformer area line loss rate... If the sample falls within a reasonable range, meaning it does not fall into either of the above two categories, it is marked as a normal sample.

[0116] Among them, the high loss threshold and negative loss threshold Determine as follows:

[0117]

[0118]

[0119] in, This indicates the upper limit for line loss management in the transformer area, preferably 10%; This indicates the reasonable lower limit of line loss in the transformer area, preferably 0% or a negative loss lower limit set according to the local line loss statistics. and These represent the mean and standard deviation of the line loss rate for the same type of transformer area or for historical normal samples in that transformer area, respectively. and These represent the 95th and 5th percentile values ​​of the line loss rate for historical normal samples, respectively. The value ranges from 2 to 3, with 3 being the preferred value. When the line loss rate of the transformer area is continuous... A time window higher than When the line loss rate of the transformer substation is continuously high, the sample is marked as an abnormally high loss sample; when the line loss rate of the transformer substation is continuously high... A time window below When this happens, the sample is marked as an abnormally damaged sample.

[0120] For samples with existing on-site verification results, the anomaly type and location can be further marked. The anomaly types include metering device anomaly, meter multiplier error, wiring anomaly, suspected user electricity theft, abnormal reverse active power of distributed renewable energy, abnormal data collection, and abnormal line loss. The anomaly location can be marked to the user, electricity meter, meter box, branch box, low-voltage outgoing line, or low-voltage side of distribution transformer.

[0121] For samples with only area-level line loss anomaly records without precise location annotations, they can be used for anomaly identification training, but are not mandatory for node location training. For samples caused by reasons such as data acquisition failure, abnormal meter operation status, or time asynchrony, the data quality anomaly type can be labeled simultaneously so that the model can learn the difference between data quality information and actual line loss anomalies.

[0122] The line loss rate of the transformer area is obtained through step S2-4. The calculation can quantitatively characterize the operating status of the transformer area and provide labels for subsequent model training, determining the model's learning objectives and enabling the model to learn the mapping relationship between node features and overall transformer area loss under different operating states.

[0123] Specifically, for distribution areas with distributed renewable energy access, the total net active power of user nodes is defined as follows:

[0124]

[0125] That is, the sum of the net active power of all user nodes within the distribution area, of which This represents the set of user nodes in the distribution area.

[0126] When the sum of net active power of user nodes A negative value indicates that the user side is supplying power in the reverse direction to the transformer area. In this case, the line loss rate of the transformer area may decrease or even become negative. (Summarized net active power of user nodes) It can be used as an auxiliary power characteristic for negative loss analysis in areas with distributed renewable energy. Whether it should be included in the line loss rate calculation should be determined according to the metering rules of the area and the business statistics.

[0127] Step S3: Input the adjacency matrix and the node feature matrix into the graph convolutional neural network. Propagate and aggregate the node's own features and the features of its neighboring nodes through the graph convolutional layer to obtain the substation-level hidden representation and the node-level hidden representation. Output the substation line loss status based on the substation-level hidden representation.

[0128] In step S3, the first The propagation process of layer graph convolution can be represented as:

[0129]

[0130] in, , Indicates the first Hidden features of layer nodes. Indicates the first Layer trainable weight matrix, Indicates the bias term. This represents a non-linear activation function.

[0131] Preferably, the graph convolutional neural network includes a first graph convolutional layer, a second graph convolutional layer, a substation-level pooling layer, a substation state classification layer, and a node anomaly classification layer connected in sequence. Let the feature dimension of the node feature matrix be... The first convolutional layer will The first graph's node features are mapped to 64-dimensional latent features of the first node, and the second graph's convolutional layer maps these 64-dimensional latent features of the first node to 32-dimensional latent features of the second node. Both the first and second graph convolutional layers use the ReLU activation function, and a Dropout layer with a dropout rate of 0.2 is set between the first and second graph convolutional layers.

[0132] The substation-level pooling layer performs node-weighted average pooling on the latent features of each node output from the convolutional layer of the second graph to obtain the substation-level latent representation. The pooling weights for each node are determined by the trainable attention layer based on the node's latent features and node type, ensuring that nodes on the low-voltage side of the distribution transformer, low-voltage outgoing line nodes, branch nodes, user meter box nodes, end-user nodes, and distributed renewable energy grid-connected points each obtain pooling weights appropriate to their operating status.

[0133] Optionally, the transformer area status classification layer uses the Softmax function to output the probabilities of three transformer area line loss states: normal, abnormally high loss, and abnormally negative loss, based on the transformer area-level implicit representation; the node anomaly classification layer uses the Sigmoid function to output the anomaly probability of each node based on the node implicit features output by the convolutional layer in the second figure.

[0134] Through multi-layer graph convolution, the implicit features of each node include its own current, power, power factor and data quality information, as well as the operating status information of its upstream and downstream adjacent nodes. This enables the expression of the correlation between the power transmission and aggregation of low-voltage distribution areas along the topology, allowing user-side anomalies to be transmitted to the upper-level nodes along the topology.

[0135] After passing through several graph convolutional layers, the latent features of nodes can be pooled or summarized to form a region-level latent representation:

[0136]

[0137] in, This represents the output of the last layer of graph convolution. This indicates average pooling, max pooling, or weighted pooling operations. This indicates the implicit representation at the district level.

[0138] The output of the transformer area line loss status identification can be represented as:

[0139]

[0140] in, This represents the probability of line loss status for the transformer substation output by the model. The status categories can include normal, high loss, and negative loss. and These represent the output layer weights and biases, respectively.

[0141] The probability of a node being abnormal can be expressed as:

[0142]

[0143] in, Represents a node The probability of node anomalies, and These represent the weights and biases of the node-level output layer, respectively. For nodes The node-level implicit representation is the implicit feature output by the last layer of graph convolution.

[0144] The output at the transformer substation level is used to determine the line loss status of the transformer substation corresponding to the current time window. The output at the node level is used to represent the degree of correlation between each node and the current line loss anomaly. The output at the node level is not directly equivalent to the final cause of the anomaly, but is used together with the subsequent power imbalance sensitivity, spatial correlation similarity and data quality anomaly degree to sort suspected abnormal nodes.

[0145] Furthermore, during model training, the constructed historical samples are divided into training samples, validation samples, and test samples. Training samples are used to adjust model parameters, validation samples are used to select the number of graph convolutional layers, learning rate, node latent feature dimensions, and regularization parameters, and test samples are used to evaluate the model's diagnostic performance on data not used in training. The model takes the transformer substation topology and node features as input, and substation line loss status labels or abnormal location labels as supervision information. Preferably, for historical samples from the same low-voltage distribution substation, samples within the first 70% of the time period are divided into training samples, samples within the middle 15% of the time period are divided into validation samples, and samples within the last 15% of the time period are divided into test samples. To avoid data leakage caused by data from the same abnormal event or adjacent time windows entering different datasets, all time windows corresponding to the same consecutive abnormal event are divided into the same dataset in the training set, validation set, or test set. The same consecutive abnormal event refers to multiple time windows where abnormal states occur consecutively within the same substation, or where the interval between two adjacent abnormal states does not exceed a preset interval. When using historical samples from multiple low-voltage distribution transformer areas for cross-region training and validation, the transformer area number is used as the group identifier to divide overlapping time windows or time intervals within the same transformer area into the same dataset, so as to avoid highly correlated samples from the same transformer area entering the training set and test set at the same time.

[0146] Preferably, the Adam optimizer is used to train the graph convolutional neural network, with an initial learning rate set to 0.001 and a weight decay coefficient set to... The maximum number of training epochs is set to 200. After each training epoch, the validation set loss is calculated using the validation samples. If the validation set loss does not decrease for 20 consecutive epochs, training is terminated early, and the network parameters corresponding to the lowest validation set loss are saved as the network parameters after training is complete. The number of network layers, hidden feature dimensions, learning rate, weight decay coefficient, maximum number of training epochs, and number of epochs to stop early can all be adjusted based on the validation set results.

[0147] If the task is to identify the status of transformer line loss, the output should be a category such as normal, high loss, or negative loss, and a classification loss in the form of cross-entropy can be used. If the task is to locate abnormal nodes, the output should be the abnormal probability or abnormal score of each node, and a node location loss in the form of binary cross-entropy can be used. If both identifying abnormal states and locating abnormal nodes are required, a multi-task loss function can be set.

[0148]

[0149] in, , , Indicates the weighting coefficient. Represents the set of model parameters. Represents the regularization term, For classification loss function, Locate the loss for the node.

[0150] Preferably, , , .

[0151] For example, when both area labels and node labels exist, one can take... , , When only area-level tags exist, they can be retrieved. , , .

[0152] To reduce the impact of imbalanced samples on training results, different training weights can be assigned to normal samples, abnormally high-loss samples, and abnormally negative-loss samples. Furthermore, the distribution of training samples can be improved by expanding adjacent time windows, supplementing with samples from historically similar transformer areas, or manually verifying sample additions.

[0153] After the model training is completed, the network parameters, feature normalization parameters, and node encoding relationships obtained during training are saved for subsequent online diagnosis.

[0154] Historical line loss management data was fully utilized during model training. Multi-task learning was used to enhance the model's ability to distinguish abnormal patterns, which laid the foundation for the subsequent use of spatial correlation similarity in anomaly detection.

[0155] Step S4: When the line loss status of the transformer area is abnormally high or abnormally negative, output the abnormal probability of each node according to the implicit representation of the node level, and calculate the node abnormal score by combining the influence of the node power change on the upper branch or transformer area line loss index, the spatial correlation similarity between the node and neighboring nodes, and the abnormality of the node data quality. Then, according to the transformer area topology connection relationship and power flow direction, search along the power supply path level by level, calculate the cumulative abnormal contribution of the path by combining the node abnormal score and the branch power imbalance, generate the priority verification path and output the diagnostic results.

[0156] In step S4, during online diagnosis, the latest operating data of the transformer area is obtained according to a preset cycle. After preprocessing and feature construction in step S2, the latest node feature matrix and the adjacency relationship of the transformer area are input into the trained graph convolutional neural network to obtain the line loss status of the transformer area and the abnormal probability of each node.

[0157] If the model output is normal, the time window is recorded as normal operation, and monitoring continues for the next cycle. If the model output is abnormally high loss or abnormally negative loss, the anomaly tracing process begins.

[0158] To avoid false alarms caused by short-term data fluctuations, step S4 is preferably set as follows: if the model output shows abnormally high loss or abnormally negative loss, further confirmation is made by combining the diagnostic results of multiple consecutive time windows. Specifically, the abnormal probability threshold... Based on the diagnostic results of the validation set, the false alarm rate on the validation set is determined to be no more than a preset false alarm rate; preferably, The value ranges from 0.70 to 0.90, with 0.80 being more preferred. When the probability of abnormally high loss or abnormally negative loss in the model output is continuous... All time windows exceeded Or the line loss rate of the transformer area is continuous More than one time window or below Furthermore, when the model output is abnormal, the anomaly localization analysis process is triggered. Preferably, The value ranges from 2 to 3, with 3 being the preferred value.

[0159] Specifically, when the line loss status identification result of the transformer area is abnormal negative loss, the characteristics of the distributed renewable energy grid-connected point, user nodes with reverse active power, and their upstream branch nodes are extracted first to analyze the correlation between changes in reverse active power, changes in net active power, and changes in the transformer area line loss rate. If the changes in reverse active power, changes in net active power, and changes in negative loss trend are consistent, that is, the sum of the net active power of user nodes... When the negative power loss increases simultaneously with the negative power loss in the distribution area, and the metering direction, topology affiliation, and data acquisition status of each grid connection point are normal, it is likely to be judged as normal reverse power transmission; if the sum of net active power of user nodes is... If the trend of negative loss is inconsistent with the trend of negative loss, or if there is an error in the grid connection point or an abnormal measurement direction, it is judged as abnormal.

[0160] Once the exception source tracing is triggered, this step specifically includes the following sub-steps.

[0161] Step S4-1: Spatial Correlation Similarity Calculation. After detecting abnormal line loss, further analysis is performed on the contribution of each node to the anomaly. Spatial correlation similarity can be determined by the degree of similarity of implicit features between the target node and its upstream, downstream, and peer-adjacent nodes, the consistency of load change trends, the correlation of power changes, and topological distance. Nodes With nodes In the time window Spatial association similarity can be represented as:

[0162]

[0163] in, and Representing nodes respectively and nodes Hidden features in the last layer of graph convolution.

[0164] node The average spatial association similarity with its neighboring nodes is calculated by the following formula:

[0165]

[0166] in, Represents a node The set of neighboring nodes.

[0167] If the power consumption change of a node deviates significantly from the normal aggregation relationship of its neighboring nodes, then that node may be related to the current anomaly. For nodes with low spatial correlation similarity, if their power consumption change, data quality status, or reverse active power consumption change are all abnormal, then that node will receive a higher weight in the subsequent anomaly scoring. For nodes with high spatial correlation similarity and whose power consumption change conforms to the power consumption change trend of the upstream branch, their suspected anomaly level can be reduced.

[0168] Step S4-2: Calculation of power imbalance sensitivity. Branch node In the time window The electrical imbalance can be expressed as:

[0169]

[0170] Among them, a branch node refers to an intermediate convergence node that supplies power to downstream objects, such as low-voltage outgoing lines, branch boxes, and multi-outgoing line meter boxes. Represents a branch node Upstream input power, Represents a branch node The set of downstream nodes, Indicates downstream node Net active power, This indicates the electricity adjustment item determined according to the metering and settlement standards of the distribution area. This refers to the reference loss obtained based on historical normal samples, line length, or experience loss levels. It is a pre-set a priori value and can be updated periodically.

[0171] Furthermore, when there are distributed renewable energy grid-connected points or user nodes with reverse active power downstream of a branch node, the downstream power aggregation defaults to using net active power in the calculation. In this case, the reverse active power is already reflected through net active power. Set to zero. If the business system uses a separate settlement or correction method for reverse active power, then the downstream power aggregation will use forward active power, and the reverse active power and its settlement correction amount will be incorporated into the correction item. Handling. Because reverse power supply physically increases the amount of electricity supplied, The active power consumption in the forward and reverse directions is calculated by superimposing the values ​​in the forward direction, with the specific sign determined according to the metering standards of the transformer area. Similarly, corrections for other metering services such as meter replacement, supplementary payments, and refunds are also included. The value and sign of this value are uniformly determined by the statistical caliber of the distribution area's line loss. Therefore, regardless of the reverse active power processing or settlement caliber used by the business system, only the corresponding correction item needs to be defined. The composition of the main formula for the power imbalance remains unchanged. The power imbalance is used to calculate the node power imbalance sensitivity and the cumulative anomaly contribution of the path.

[0172] node To its superior branches The sensitivity to electrical imbalance can be expressed as:

[0173]

[0174] in, Indicates the presence of nodes Time branch The amount of power imbalance, This indicates removing nodes during the calculation. Or for nodes Perform a normal estimate of the branch power imbalance after replacement. The larger the value, the more likely it is to be a node. The more pronounced the impact of the imbalance in the branch's electrical charge, the more significant the impact on the root node that has no parent branch. Take zero.

[0175] Normal estimated replacement can be obtained based on the node's historical electricity consumption during the same period, electricity consumption in adjacent time windows, load curves of similar users, or electricity consumption of similar users in the same meter box. This estimated value is only used to calculate the node's sensitivity to branch electricity imbalance and is not used to replace the original metering data.

[0176] User nodes refer to end-metering nodes that no longer supply power to downstream areas. The power imbalance sensitivity of all types of nodes is calculated uniformly according to the above formula. For distributed renewable energy grid-connected points, this sensitivity mainly reflects the impact of sudden changes in reverse active power on the power balance of their respective branches, and can be used to analyze its correlation with changes in the negative losses of the distribution area.

[0177] Step S4-3: Calculation of Data Quality Anomaly Level. The degree of data quality anomaly can be determined based on the data collection completeness rate, meter operating status, and time synchronization status. Specifically, the degree of data quality anomaly is calculated using the following formula:

[0178]

[0179] in, Represents a node In the time window The completeness of the data collection Represents a node In the time window The degree of data quality anomaly identification, This is obtained by weighted summation of meter anomaly flags, communication anomaly flags, and time synchronization flags. For example, it is calculated as 0.5 multiplied by the meter anomaly flag, plus 0.3 multiplied by the communication anomaly flag, plus 0.2 multiplied by the time synchronization flag. When there are no relevant anomaly flags, Take zero.

[0180] when When the value is high, the diagnostic results can mark the node as a suspected data quality abnormality node and prompt priority to check the acquisition link, meter operating status, frozen power integrity and clock synchronization status.

[0181] Step S4-4: Calculate the node anomaly score. For example... Figure 3 As shown, the average spatial correlation similarity Sensitivity to power imbalance The probability of node anomalies output by the model and the degree of data quality anomalies Perform fusion and calculate node anomaly scores:

[0182]

[0183] Node anomaly scoring A higher score indicates that the node is more likely to be related to the current abnormal line loss. Among these, , , , This represents the weighting coefficient, which can be adjusted according to abnormal scenarios and determined based on historical verification samples, expert experience, or validation set results. For abnormally high-loss scenarios, the weights of power imbalance sensitivity and node anomaly probability can be appropriately increased; for abnormally negative-loss scenarios, the weights of reverse active power related characteristics and data quality anomaly degree can be appropriately increased; for transformer areas with poor data acquisition quality, the weight of data quality anomaly degree can be increased to prioritize data verification.

[0184] Understandably, the node anomaly probability, spatial correlation similarity, and data quality anomaly degree can all be calculated for each node within the distribution area; however, the power imbalance sensitivity can only be calculated for nodes with superior branches. The low-voltage side of the distribution transformer, as the root node of the distribution area, does not have superior branches and is not included in the power imbalance sensitivity calculation. For nodes whose components cannot be calculated, that component is set to zero, and the weight coefficients are renormalized over the remaining components before calculating the node anomaly score.

[0185] For example, ,and .

[0186] In general scenarios, When the power line in the transformer area is experiencing abnormally high losses, When the line is under abnormal negative loss, When the data acquisition quality of the transformer substation is poor, .

[0187] Steps S4-5: Source tracing and path verification. Based on the transformer substation topology, search along the power supply path level by level. During the search, compare the abnormal scores, power differences, and trends of adjacent time windows for each branch node and its downstream nodes. If the input-output power difference of a certain branch is significantly abnormal, and there are one or more nodes with high abnormal scores downstream of that branch, then that branch and its downstream nodes are identified as key suspected areas.

[0188] The step-by-step search along the power supply path includes: for abnormally high loss scenarios, searching step-by-step along the forward power supply path from the low-voltage side of the distribution transformer to the user side; for abnormally negative loss scenarios, in addition to searching downstream from the low-voltage side of the distribution transformer along the forward power supply path, it also starts from the distributed new energy grid connection point or user node with reverse active power, traces the reverse impact path upstream, and determines whether it is normal reverse power transmission or abnormal based on whether the changes in reverse active power, net active power, and negative loss are consistent with the trend, as well as the metering direction, topology affiliation, and collection status of the grid connection point.

[0189] Specifically, for a power flow path to be analyzed The cumulative anomaly contribution value of its path can be expressed as:

[0190]

[0191] in, Representing a path In the time window The path cumulative abnormal contribution, Represents nodes in the path Abnormal scores, Represents branch nodes in the path The amount of power imbalance, The weighting coefficient represents the weighting factor for the branch power imbalance.

[0192] Paths with higher cumulative anomaly contribution values ​​are prioritized for verification. The cumulative anomaly contribution value is used to determine priority verification paths and does not mean that all nodes in the path have anomalies. In actual verification, a comprehensive judgment can be made by combining node anomaly scores, branch power imbalance, data quality status, and on-site business records.

[0193] In particular, for abnormally high losses, priority is given to tracing along the forward power supply path from the low-voltage side of the distribution transformer to the user side to find branches or nodes that widen the difference between the upstream power supply and the downstream power sales. The node anomaly score is then used to determine whether it belongs to suspected metering anomalies, data collection anomalies, wiring anomalies, or abnormal user power consumption.

[0194] For abnormal negative losses, in addition to analyzing along the forward power supply path, we also combine the changes in reverse active power, net active power, metering direction status, topology attribution, and data quality information to distinguish between normal reverse power transmission impact and abnormal metering direction, abnormal grid connection point attribution, or asynchronous data collection time, thereby obtaining the ranking results of suspected abnormal equipment, suspected abnormal branches, and suspected abnormal users.

[0195] Steps S4-6: Output diagnostic results. The diagnostic results include transformer area number, diagnostic time, transformer area line loss status, anomaly type tendency, list of suspected anomaly nodes, node anomaly score, changes in related electrical quantities, topology path involved in the anomaly, priority verification path, and suggested verification objects.

[0196] Specifically, for users or equipment suspected of metering anomalies, it is recommended to check the meter's operating status, multiplier parameters, wiring method, and the integrity of the collected data; for scenarios suspected of topology file errors, it is recommended to check the relationship between the user and transformer, the ownership of the meter box, and the connection relationship of the line branches; for scenarios suspected of abnormal reverse active power of distributed renewable energy, it is recommended to check the metering direction of the grid connection point, the record of reverse active power, and the synchronization status of the collection time; for scenarios suspected of abnormal line losses, it is recommended to check factors such as line aging, joint overheating, three-phase imbalance, and excessively long low-voltage power supply radius.

[0197] The method described in this embodiment can be deployed on a transformer substation line loss platform. The platform reads operational data from the substation's main meter, branch monitoring points, user energy meters, and distributed renewable energy grid-connected points according to a preset cycle. It then constructs a substation map structure based on the low-voltage substation topology archive. The model outputs the substation line loss status, a ranking of suspected abnormal nodes, and priority verification paths. When a substation experiences abnormally high or negative losses for multiple consecutive cycles, the platform further outputs the branch boxes, user meter boxes, user energy meters, or distributed renewable energy grid-connected points that contribute significantly to the anomaly, assisting staff in prioritizing on-site verification. Once the on-site verification confirms the cause of the anomaly, the verification results can be written back to the sample database for periodic model fine-tuning or incremental training, allowing the model to gradually adapt to changes in substation load structure, renewable energy access ratios, and topology updates.

[0198] In low-voltage distribution areas with distributed photovoltaic (PV) grid connection, the system can collect user forward active power, reverse active power, PV grid-connected power generation, grid-connected point metering direction status, and data collection time synchronization status, treating the grid-connected point as an independent node in the graph structure for calculation. When an abnormal negative loss occurs in the distribution area, the aforementioned source analysis method is used to determine whether the negative loss originates from normal reverse power transmission, or from suspected metering direction anomalies, file attribution anomalies, or data collection time asynchrony. This result assists maintenance personnel in identifying priority inspection targets.

[0199] Furthermore, the deployment of this embodiment relies on existing data systems and does not require additional dedicated sensing equipment, thus possessing good economic efficiency and feasibility for widespread adoption.

[0200] Example 2

[0201] This embodiment provides a diagnostic device for low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks. For example... Figure 4 As shown, the device includes a topology construction module, a data acquisition and preprocessing module, a graph convolutional neural network, an anomaly scoring module, and a source tracing and output module.

[0202] The topology construction module is used to obtain the physical topology data of the low-voltage distribution transformer area, and abstract the low-voltage side of the distribution transformer, low-voltage outgoing lines, branch boxes, user meter boxes, user energy meters, and distributed new energy grid connection points as nodes in the graph structure. The actual line connection relationship between nodes is abstracted as the edge in the graph structure, and a topology graph structure representing the power transmission relationship of the transformer area is constructed, and the corresponding adjacency matrix is ​​constructed.

[0203] The data acquisition and preprocessing module is used to collect the running data and data quality information of each node within a preset time window. After preprocessing, the running data and data quality information are combined to construct the node dynamic feature vector, and the node feature matrix is ​​formed based on the dynamic feature vectors of all nodes.

[0204] Graph convolutional neural networks are used to take the adjacency matrix and node feature matrix as input, and propagate and aggregate the features of the nodes themselves and their neighboring nodes through graph convolutional layers to obtain the hidden representation at the transformer substation level and the hidden representation at the node level. The line loss status of the transformer substation is output based on the hidden representation at the transformer substation level.

[0205] The anomaly scoring module is configured to: when the line loss status of the transformer area is abnormally high loss or abnormally negative loss, output the anomaly probability of each node according to the node-level implicit representation, and calculate the node anomaly score by combining the influence of the node's power change on the line loss index of the upper branch or transformer area, the spatial correlation similarity between the node and its neighboring nodes, and the degree of abnormality of the node's data quality.

[0206] The tracing and output module is used to search along the power supply path step by step according to the topological connection relationship of the transformer area and the direction of power flow. It calculates the cumulative abnormal contribution of the path by combining the node anomaly score and the branch power imbalance, generates the priority verification path, and outputs the diagnostic results including the transformer area line loss status, anomaly type tendency, suspected abnormal node list and priority verification path.

[0207] The specific implementation methods and data flow methods of each module in this device are the same as those of the corresponding steps in Embodiment 1, and will not be repeated here. The above module division is only a logical functional division, and in actual implementation, it can be merged, split, or replaced according to hardware deployment, edge computing capabilities, and backend server architecture.

[0208] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0209] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and any variations or modifications may be made to the implementation of the present invention without departing from the stated principles.

[0210] It should be noted that, as will be apparent to those skilled in the art, the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics thereof. The scope of the present invention is defined by the claims rather than the foregoing description.

Claims

1. A method for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks, characterized in that, include: Obtain the physical topology data of the low-voltage distribution transformer area, abstract the low-voltage side of the distribution transformer, low-voltage outgoing lines, branch boxes, user meter boxes, user energy meters, and distributed new energy grid connection points as nodes in the graph structure, abstract the actual line connection relationship between nodes as edges in the graph structure, construct a topology graph structure representing the power transmission relationship of the transformer area, and construct the corresponding adjacency matrix. The running data and data quality information of each node within a preset time window are collected. After preprocessing, the running data and data quality information are combined to construct a node dynamic feature vector. A node feature matrix is ​​formed based on the dynamic feature vectors of all nodes. The adjacency matrix and the node feature matrix are input into a graph convolutional neural network. The graph convolutional layer propagates and aggregates the node's own features and the features of its neighboring nodes to obtain the substation-level implicit representation and the node-level implicit representation. The substation line loss status is output based on the substation-level implicit representation. When the line loss status of the transformer area is abnormally high or abnormally negative, the abnormal probability of each node is output according to the implicit representation of the node level. The abnormality score of the node is calculated by combining the influence of the node power change on the line loss index of the upper branch or transformer area, the spatial correlation similarity between the node and the neighboring nodes, and the abnormality of the node data quality. Based on the topological connection relationship of the transformer area and the direction of power flow, the system searches step by step along the power supply path. Combining the node anomaly score and the branch power imbalance, the cumulative anomaly contribution of the path is calculated, a priority verification path is generated, and the diagnostic results are output. The diagnostic results include the transformer area line loss status, anomaly type tendency, a list of suspected anomaly nodes, and the priority verification path.

2. The method for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks according to claim 1, characterized in that, The operational data includes three-phase current, three-phase voltage, power factor, forward active power, reverse active power, net active power, and distributed renewable energy output data. The data quality information includes data collection completeness rate, meter operating status, communication status, and time synchronization status. The dynamic feature vector also includes node type identifier, topology level, branch number, phase affiliation, identifier of whether it has reverse power transmission capability, and identifier of whether there is a suspected anomaly in the topology file.

3. The method for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks according to claim 1, characterized in that, The adjacency matrix is ​​self-connected and normalized; when line parameters are available, a weighted adjacency matrix is ​​constructed using edge weights, which are determined based on line length, conductor cross-section, or historical load correlation.

4. The method for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks according to claim 1, characterized in that, The formula for calculating the node anomaly score is as follows: ； in, Represents a node In the time window The probability of node anomalies, Represents a node In the time window Sensitivity to power imbalance Represents a node Its neighboring nodes in the time window The average spatial association similarity is determined based on the node-level implicit representations of a node and its neighboring nodes. Represents a node In the time window The degree of data quality anomalies, , , , This represents the weighting coefficient.

5. The method for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks according to claim 4, characterized in that, The power imbalance sensitivity in the node anomaly score is calculated using the following formula: ； in, Indicates time window Includes nodes Time branch The amount of power imbalance, Indicates within the time window Remove nodes in the following calculation Or for nodes The branch power imbalance is estimated after normal replacement, and the power imbalance sensitivity is... Used for quantization nodes The degree of contribution to the imbalance of branch power.

6. The method for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks according to claim 4, characterized in that, node The average spatial association similarity with its neighboring nodes is calculated by the following formula: ； in, Represents a node The set of neighboring nodes; For nodes With nodes In the time window The spatial similarity is calculated using the following formula: ； in, and Representing nodes respectively and nodes Hidden features in the last layer of graph convolution.

7. The method for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks according to claim 4, characterized in that, The degree of data quality anomaly in the node anomaly score is calculated using the following formula: ； in, Represents a node In the time window The completeness of the data collection Represents a node In the time window The degree of data quality anomaly identification.

8. The method for diagnosing abnormal low-voltage line loss in transformer substations based on graph convolutional neural networks according to claim 4, characterized in that, The node anomaly probability in the node anomaly score is calculated using the following formula: ； in, and These represent the weights and biases of the node-level output layer, respectively. For nodes The node-level implicit representation is the implicit feature output of the last layer of graph convolution.

9. The method for diagnosing low-voltage line loss anomalies in transformer substations based on graph convolutional neural networks according to claim 1, characterized in that, The step-by-step search along the power supply path includes: For scenarios with abnormally high losses, the search proceeds step by step along the forward power supply path from the low-voltage side of the distribution transformer to the user side. For abnormal negative loss scenarios, in addition to searching downstream from the low-voltage side of the distribution transformer along the forward power supply path, we also start from the distributed new energy grid connection point or user node with reverse active power and trace the reverse impact path upstream. Based on whether the changes in reverse active power, net active power and negative loss change trends are consistent, as well as the metering direction, topology affiliation and collection status of the grid connection point, we determine whether it is normal reverse power supply or abnormal.

10. A low-voltage line loss anomaly diagnosis device for transformer substations based on graph convolutional neural networks, characterized in that, include: The topology construction module is used to obtain the physical topology data of the low-voltage distribution transformer area, and abstract the low-voltage side of the distribution transformer, low-voltage outgoing lines, branch boxes, user meter boxes, user energy meters and distributed new energy grid connection points as nodes in the graph structure, and abstract the actual line connection relationship between nodes as edges in the graph structure to construct a topology graph structure representing the power transmission relationship of the transformer area, and construct the corresponding adjacency matrix. The data acquisition and preprocessing module is used to collect the running data and data quality information of each node within a preset time window. After preprocessing, the running data and data quality information are combined to construct a node dynamic feature vector, and a node feature matrix is ​​formed based on the dynamic feature vectors of all nodes. A graph convolutional neural network is used to take the adjacency matrix and the node feature matrix as input, and to propagate and aggregate the node's own features and the features of its neighboring nodes through graph convolutional layers to obtain the substation-level hidden representation and the node-level hidden representation. The substation-level hidden representation is then used to output the substation line loss status. The anomaly scoring module is configured to: when the line loss status of the transformer area is abnormally high loss or abnormally negative loss, output the anomaly probability of each node according to the node-level implicit representation, and calculate the node anomaly score by combining the influence of node power changes on the line loss index of the upper branch or transformer area, the spatial correlation similarity between the node and neighboring nodes, and the degree of node data quality anomaly. The tracing and output module is used to search along the power supply path step by step according to the topological connection relationship of the transformer area and the direction of power flow. It calculates the cumulative abnormal contribution of the path by combining the node anomaly score and the branch power imbalance, generates the priority verification path, and outputs the diagnostic results including the transformer area line loss status, anomaly type tendency, suspected abnormal node list and priority verification path.

Images