An Artificial Intelligence Method for Identifying Potential Faults in Power Distribution Network Lines

Abstract

This invention provides an artificial intelligence-based method for identifying potential faults in distribution network lines, comprising: S1. Hierarchical deployment of measurement nodes for distribution network lines and synchronous acquisition of multi-source heterogeneous data; S2. Multi-scale preprocessing and fractional-order transient feature enhancement of multi-source heterogeneous data; S3. Construction of a three-layer coupled dynamic heterogeneous graph structure constrained by Riemannian manifolds; S4. Construction of a topology-aware multimodal graph neural network architecture; S5. Phased training and adaptive updating of the TM-GNN model through federated meta-learning; S6. Real-time inference and identification of potential faults in distribution network lines based on the TM-GNN inference model. This invention achieves early identification, precise location, risk quantification, and trend prediction of potential faults in distribution network lines, constructing a closed-loop process from data acquisition, intelligent analysis, early warning output to operation and maintenance feedback iteration, comprehensively improving the reliability of distribution network operation and the level of intelligent operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of intelligent distribution network condition monitoring and fault early warning technology, and in particular to an artificial intelligence method for identifying potential faults in distribution network lines. Background Technology

[0002] As a core component of the power system directly facing end users, the reliability of the 10kV distribution network directly determines the power supply quality and user experience. Early hidden dangers such as line aging, tree obstructions, and poor contact are the core causes of distribution network line faults and power outages. Therefore, early identification, accurate location, and proactive warning of potential line faults are the core requirements and key challenges of distribution network operation and maintenance.

[0003] Current technologies for detecting hidden dangers in power distribution network lines have multiple core shortcomings, as detailed below: First, traditional operation and maintenance models have fundamental limitations. Conventional methods relying on manual inspections and periodic drone inspections suffer from long inspection cycles, limited coverage, and significant susceptibility to environmental and human factors. They cannot achieve continuous real-time monitoring of line status, are unable to capture the dynamic and subtle characteristics of early-stage potential problems, and cannot provide early warnings of faults. Furthermore, conventional single-point online monitoring devices can only perform isolated measurements of single physical quantities such as partial discharge and temperature, failing to reflect the overall operating status of the line at the distribution network system level. Their sensitivity to early-stage potential problems is severely insufficient, easily leading to missed or false diagnoses.

[0004] Second, existing AI-based identification solutions have shortcomings in core technologies. Current power distribution network hazard identification solutions that incorporate machine learning and deep learning often employ conventional models such as convolutional neural networks and long short-term memory networks. These models can only process time-series or image data individually, generally ignoring the inherent topological connections and spatial structural characteristics of the power distribution network. They fail to effectively integrate data from multiple measurement points, losing spatial propagation correlation information of hazard features, ultimately leading to insufficient hazard identification and location accuracy and a high false alarm rate.

[0005] Third, existing graph neural networks have significant shortcomings in the application of power distribution networks. The few existing power distribution network analysis schemes that incorporate graph neural networks only construct static graph structures based on simple physical topologies, failing to characterize the dynamic electrical connections and nonlinear characteristics of hazard propagation in the power distribution network's operating state. Furthermore, the convolutional, attention, and time-series processing modules of these models are mostly unidirectional and sequential designs, lacking bidirectional interaction and parameter co-optimization mechanisms, thus hindering the deep refinement of hazard features. Simultaneously, existing schemes do not embed the core physical laws of the power system as hard constraints throughout the model process, making them prone to learning false features that contradict physical laws. This results in severely insufficient model generalization ability and extremely poor adaptability to different lines and regions. Summary of the Invention

[0006] This invention provides an artificial intelligence-based method for identifying potential faults in distribution network lines. It addresses some of the problems existing in current distribution network fault identification technologies, enabling early identification, accurate location, risk quantification, and trend prediction of potential faults in distribution network lines. It constructs a closed-loop process from data collection, intelligent analysis, early warning output to operation and maintenance feedback iteration, thereby comprehensively improving the reliability of distribution network operation and the level of intelligent operation and maintenance.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: An artificial intelligence-based method for identifying potential faults in power distribution network lines includes: S1. Complete the hierarchical deployment of measurement nodes for the distribution network lines, and synchronously collect multi-source heterogeneous raw measurement datasets with unified timestamps through each node. At the same time, obtain the initial physical topology wiring diagram of the target line from the power grid production management system. S2. Using the multi-source heterogeneous original measurement dataset and the initial physical topology wiring diagram as input, complete the data standardization preprocessing and output the standardized node feature matrix, the real-time risk matrix of all line nodes and the time series similarity results of electrical quantities between nodes; S3. Using the standardized node feature matrix, the real-time risk matrix of all line nodes, the time-series similarity results of electrical quantities between nodes, and the initial physical topology wiring diagram as input, construct a three-layer coupled real-time dynamic heterogeneous graph and synchronously output the edge feature matrix corresponding to the real-time dynamic heterogeneous graph. S4. Build a topology-aware multimodal graph neural network model, using real-time dynamic heterogeneous graphs, standardized node feature matrices, and edge feature matrices as inputs, to complete feature extraction and hazard identification multi-task outputs; S5. Using the topology-aware multimodal graph neural network model as the object, perform phased training and optimization, and output a converged deployable inference model; S6. Based on the deployable inference model, input the real-time operation data of the distribution network lines, perform end-to-end inference, and output a standardized identification result set of real-time fault hazards of the lines.

[0008] The artificial intelligence method for identifying potential faults in power distribution network lines also includes the following in this manual: S7. Based on the standardized identification result set of real-time line fault hazards, complete the operation and maintenance early warning push and on-site handling feedback data collection, and after associating and labeling the feedback data, send it back to the model training and optimization stage of S5 to complete the model adaptive update and the whole process closed loop.

[0009] In this specification, in the data standardization preprocessing of S2, after the multi-scale feature standardization extraction is completed, the extracted millisecond-level transient features are subjected to targeted enhancement processing by an adaptive fractional calculus operator; the order of the fractional calculus operator is adaptively optimized based on the physical topological adjacency matrix of the real-time dynamic heterogeneous graph generated in the previous time segment, and the enhanced features are used to generate the standardized node feature matrix of the current time segment.

[0010] In this specification, in step S3, after completing the construction of the three-layer coupled real-time dynamic heterogeneous graph, the real-time dynamic heterogeneous graph in Euclidean space is mapped to the Riemannian manifold space to construct a Riemannian manifold metric tensor and a geodesic distance matrix adapted to the nonlinear characteristics of the distribution network topology. The Riemannian manifold metric tensor and the geodesic distance matrix are synchronously input into the topology-aware multimodal graph neural network model as core constraint parameters for model feature extraction.

[0011] In this specification, the topology-aware multimodal graph neural network model of S4 is equipped with a topology-adaptive graph convolution module with Riemannian manifold constraints and a multi-head graph attention module with enhanced physical constraints. The feature output of the graph convolution module is used as the input of the attention module, and the output of the attention module back-optimizes the Riemannian manifold metric tensor of the graph convolution module. The two modules achieve bidirectional interactive collaborative optimization.

[0012] In this specification, the topology-aware multimodal graph neural network model of S4 is equipped with a temporal graph convolutional network unit and a hazard feature decoupling module with Tsallis non-extensive entropy constraints. The spatiotemporal fusion features output by the temporal graph convolutional network unit are used as the input of the hazard feature decoupling module. The decoupling result of the hazard feature decoupling module back-optimizes the weight parameters of the temporal graph convolutional network unit, and the two units achieve bidirectional interactive collaborative optimization.

[0013] In this specification, during the phased training optimization of S5, large-scale unsupervised pre-training is first performed. The unsupervised pre-training adopts a method of joint optimization of three loss functions: graph reconstruction loss, state prediction loss, and contrast loss. By minimizing the weighted sum of the three losses through backpropagation, the model learns the basic spatiotemporal representation of the distribution network and the node topology association characteristics.

[0014] In this specification, during the phased training optimization of S5, supervised multi-task fine-tuning is performed after unsupervised pre-training is completed. The supervised multi-task fine-tuning adopts a method of joint optimization of five loss functions: node classification loss, edge-level risk quantification loss, root cause localization loss, physical constraint loss, and entropy constraint loss. Among them, the physical constraint loss is used to force the model output to conform to the core physical laws of power balance and Kirchhoff's laws in the power distribution network, so as to avoid the model learning false correlation features.

[0015] In this specification, during the end-to-end inference process of S6, the generated fault hazard identification result of the current time section will back-optimize the parameters of the standardized preprocessing of the next time section data and the construction weight of the real-time dynamic heterogeneous graph, so as to realize the bidirectional interaction and real-time optimization between the inference process and the preprocessing and graph structure construction stages.

[0016] In this specification, the adaptive update of the model in S7 adopts a federated meta-learning-driven incremental update mechanism, and builds a three-level update architecture of local edge nodes, regional edge clusters, and cloud analysis center. Local edge nodes perform lightweight incremental learning, regional edge clusters perform model parameter aggregation in their respective regions, and cloud analysis center performs global model optimization and knowledge distillation, thereby achieving cross-regional adaptation of the model and learning of small sample vulnerabilities while protecting data privacy.

[0017] In summary, this invention has at least the following beneficial effects: It significantly improves the sensitivity of identifying early-stage hidden dangers in distribution network lines. By using a fractional-order calculus transient feature enhancement operator, it effectively amplifies the weak transient signals of early-stage hidden dangers, solving the problems of insensitivity and easy omission of early-stage potential hidden danger features in existing technologies, and achieving early warning of line hidden dangers. It achieves deep fusion of multi-source heterogeneous data and topological information of the distribution network. Through a three-layer coupled dynamic heterogeneous graph structure constrained by Riemannian manifolds, it fully characterizes the physical connection characteristics, dynamic electrical correlations, and hidden danger propagation laws of the distribution network, solving the problems of lost spatial structure information and insufficient fusion of multi-source data in existing technologies, and improving the comprehensiveness and accuracy of hidden danger feature extraction. It greatly improves the physical consistency of the model and the credibility of the identification results. By embedding the core physical laws of the power system into the entire process of model encoding, processing, and training, it constructs an attention mechanism and loss function for physical constraints, avoiding the model learning false correlation features and effectively reducing the false alarm rate of hidden danger identification. This system achieves deep collaborative optimization across all modules of the model. By employing a bidirectional interaction mechanism between the graph convolution module and the attention module, and between the temporal processing module and the feature decoupling module, it replaces the unidirectional serial design of existing technologies. This enables targeted purification and multi-dimensional collaborative optimization of hazard features, enhancing the model's feature representation capabilities. It significantly improves the model's cross-scenario adaptability and few-shot learning ability. Through a federated meta-learning-driven incremental adaptive update mechanism, it achieves cross-regional knowledge transfer and rapid learning of few-shot hazards while protecting data privacy. This addresses the shortcomings of existing centralized training models, such as insufficient generalization ability and poor identification of rare hazards. Furthermore, it constructs a closed-loop system for the entire process of distribution network hazard identification and operation and maintenance. Through deep integration of operation and maintenance feedback data and model updates, it achieves continuous iterative optimization of the model, while providing precise decision support for distribution network operation and maintenance, thus improving the intelligence level and operational reliability of distribution network operation and maintenance. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the artificial intelligence identification method for potential faults in power distribution lines involved in this invention.

[0019] Figure 2 This is a schematic diagram of the main business process for identifying potential faults in power distribution lines involved in this invention.

[0020] Figure 3 This is a schematic diagram illustrating the staged training and adaptive update of the TM-GNN model involved in this invention.

[0021] Figure 4 This is a schematic diagram of the core process of hazard feature extraction and multi-task identification involved in this invention. Detailed Implementation

[0022] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] refer to Figure 1 and Figure 2 This embodiment provides an artificial intelligence-based method for identifying potential faults in distribution network lines. First, based on the functional importance of 10kV distribution network line nodes, a hierarchical multi-measurement point data acquisition system is constructed to achieve synchronous acquisition of multi-source heterogeneous data covering electrical quantities, state quantities, image data, and environmental data, providing a standardized basic data source for full-process analysis. Second, through fractional-order calculus transient feature enhancement and adaptive filtering preprocessing based on topological constraints, early weak fault characteristics are amplified and data is standardized. Simultaneously, bidirectional parameter optimization is achieved with the subsequent dynamic graph structure construction module, replacing the unidirectional data transmission mode of existing technologies. Subsequently, a three-layer coupled dynamic heterogeneous graph structure with Riemannian manifold constraints is constructed to simultaneously characterize the physical connection characteristics, real-time electrical correlations, and fault propagation patterns of the distribution network. This maps the graph structure of Euclidean space to Riemannian manifold space, adapting to the nonlinear manifold characteristics of the distribution network topology and solving the core pain point of topological information loss in existing technologies.

[0024] Based on this, the present invention constructs a topology-aware multimodal graph neural network (TM-GNN) architecture. Through the bidirectional interaction between the topology-adaptive graph convolution module constrained by Riemannian manifold and the multi-head graph attention module enhanced by physical constraints, it achieves deep extraction and targeted purification of hazard features. Through the bidirectional collaboration between the hazard feature decoupling module constrained by Tsallis non-extensive entropy and the temporal graph convolutional network, it achieves accurate separation and spatiotemporal joint modeling of multiple types of hazard features, solving the problems of hazard feature coupling and insufficient classification accuracy in existing technologies.

[0025] The model training adopts a phased progressive learning framework. First, it learns the basic spatiotemporal representation of the distribution network through large-scale unsupervised pre-training. Then, it achieves accurate adaptation to the hazard identification task through multi-task joint supervised fine-tuning. At the same time, it introduces a federated meta-learning-driven incremental adaptive update mechanism to achieve cross-regional knowledge transfer and rapid learning of small-sample hazards while protecting data privacy. Finally, it outputs the results of hazard identification, location, risk quantification and trend prediction through end-to-end real-time inference, and forms a closed loop based on the feedback data from the operation and maintenance site to achieve continuous iterative optimization of the model.

[0026] S1. Distribution network line hierarchical measurement node layout and multi-source heterogeneous data synchronous acquisition; taking a 10kV distribution network line as an example, complete the standardized hierarchical layout of measurement nodes for the entire 10kV distribution network line, synchronously acquire multi-source heterogeneous raw data covering all dimensions, and obtain the initial physical topology wiring diagram of the target line.

[0027] 1. Based on the functional importance, equipment type, and data acquisition requirements of 10kV distribution network line nodes, all measurement nodes along the entire line are divided into three fixed levels, named Level-1 critical nodes, Level-2 important nodes, and Level-3 ordinary nodes. Specific deployment rules and examples of data acquisition content are as follows: Level-1 critical nodes are deployed in the switching stations and ring main units of the 10kV distribution network lines, performing full data acquisition; examples of data acquisition content include voltage U, current I, active power P, reactive power Q, 2nd to 50th harmonics, three-phase imbalance, temperature at core equipment locations, ambient humidity, partial discharge, infrared thermal image data of equipment and lines, and other relevant data. The data acquisition includes: visible video data, real-time meteorological data, and lightning location data; the sampling frequency for voltage, current, active power, reactive power, harmonics, and three-phase imbalance is fixed at 1kHz; Level-2 critical nodes are deployed in the branch boxes and pole-mounted switches of the 10kV distribution network, performing core data acquisition; examples of acquired data include voltage U, current I, temperature at core equipment locations, and partial discharge pulse data; the sampling frequency for voltage and current is fixed at 100Hz; Level-3 ordinary nodes are deployed in the user access points and line segmentation points of the 10kV distribution network, performing basic status acquisition; examples of acquired data include line voltage monitoring data and fault indicator action data.

[0028] 2. All measurement nodes have built-in GPS / BeiDou clock synchronization modules. Each piece of raw data is accompanied by a unique synchronization timestamp, and the timestamp accuracy is perfectly matched with the sampling accuracy of the corresponding data.

[0029] 3. Obtain the initial physical topology wiring diagram of the target 10kV distribution network line from the power grid production management system. The initial physical topology wiring diagram fully records the actual electrical connection relationship between all measurement nodes and equipment, line impedance parameters, and line length parameters of the entire line.

[0030] Processing results: 1. A multi-source heterogeneous raw measurement dataset of 10kV distribution network lines with synchronization timestamps, named 1. t represents the corresponding time segment; 2. Initial physical topology wiring diagram of the target 10kV distribution network line, named 3. A complete record of basic attributes of all measurement nodes along the entire line, including node number, node level, deployment location, equipment type, equipment commissioning years, and line electrical parameters.

[0031] S2. Multi-scale preprocessing and fractional-order transient feature enhancement of multi-source heterogeneous data; The original measurement dataset output from S1 is standardized and preprocessed to address issues such as spatiotemporal misalignment, inconsistent feature dimensions, missing data, and outlier interference; a fractional-order calculus transient feature enhancement operator is introduced to amplify weak transient signals of early-stage potential problems; bidirectional interaction with the dynamic graph structure construction in S3 is achieved, with the feature output of this step serving as the core input for the S3 graph structure construction, and the topology parameters output from S3 inversely optimizing the filtering and enhancement operator parameters of this step. This step consists of five sequential sub-steps, with bidirectional data transfer between sub-steps. The specific implementation details are as follows: S2.1. Spatiotemporal alignment calibration of multi-source data; 1. The input data source is the multi-source heterogeneous raw measurement dataset with synchronization timestamps output from S1. 2. Using the timestamp generated by the GPS / BeiDou clock synchronization module in S1 as the sole reference, perform spatiotemporal alignment on raw data of different levels, sampling frequencies, and types; for millisecond-level sampled data, align with 1 millisecond as the time unit; for second-level sampled data, align with 1 second as the time unit; for minute-level sampled data, align with 1 minute as the time unit. 3. Perform spatial dimension matching on the same type of data from different nodes at the same time segment to ensure that, at the same moment, the measurement data of all nodes along the entire line can correspond to the initial physical topology wiring diagram. Based on the corresponding node positions, generate a spatiotemporally aligned dataset, named .

[0032] S2.2. Multi-scale feature standardization extraction; 1. Input data source is the spatiotemporally aligned dataset output from S2.1. 2. Based on the correlation between time scale and hidden danger characteristics, standardized features are extracted at three fixed scales. All features are appended with timestamps corresponding to the time segment. Specific extraction rules are as follows: Millisecond-level transient features: The extraction objects are transient signals of current, voltage, and partial discharge. Examples of extracted features are the maximum value of the transient signal, signal frequency, pulse width, and duration. The time precision is fixed at 1 millisecond. Second-level steady-state features: The extraction objects are steady-state signals of current and voltage. Examples of extracted features are the effective value of current, the effective value of voltage, and the content of the 2nd to 50th harmonics. The time precision is fixed at 1 second. Minute-level trend features: The extraction objects are ambient temperature and humidity, cable temperature, and line load current. Examples of extracted features are the change in temperature and humidity, the slope of cable temperature rise, and the rate of change of line load. The time precision is fixed at 1 minute. 3. All extracted features are subjected to min-max standardization processing, mapping all feature values ​​to the [0,1] interval to generate an initial node feature matrix of uniform dimension, named Where N is the total number of nodes on the entire line; and D is the feature dimension of a single node.

[0033] S2.3. Fractional Calculus Transient Feature Enhancement; This step introduces the Grünwald-Letnikov fractional calculus operator to target and enhance the weak transient features of early-stage hazards, addressing the weakness of conventional methods in this field in responding to early-stage weak hazard signals. It also enables bidirectional interaction with the S3 graph structure construction. Specific implementation details are as follows: 1. The input data source is the initial node feature matrix output from S2.2. The real-time physical topology adjacency matrix output by S3 2. Construct the Grünwald-Letnikov fractional differential operator. The operator's calculation formula is as follows: ; Let be the fractional differential operator at time t; is the order of the fractional operator, with a value range of [0,2]. The input is a transient characteristic time-series signal; The sampling step size is matched with the sampling frequency of the corresponding data; This is the start time of the signal; The coefficients of the generalized binomial are calculated using the following formula: ,in 3. Gamma function; 4. Real-time physical topology adjacency matrix based on S3 output Adaptive optimization of the order of fractional operators The optimized formula is: ; Let be the adaptive fractional order of node i at time t; The preset base order is fixed at 0.8; The preset order adjustment factor is fixed at 0.6. The physical topological adjacency value between node i and node j at time t is the output result from S3; Let be the transient characteristic variance of node j at time t; Let be the total number of neighboring nodes of node i; 4. Using an adaptively optimized fractional-order operator, the transient features of each node are enhanced to generate an enhanced node feature matrix, named... This matrix will serve as the core input for constructing the S3 dynamic graph structure, enabling bidirectional interaction and fusion between this step and S3. Enhanced node feature matrix. The generating expression is: ; The transient characteristic matrix in milliseconds after enhancement by fractional-order operators is given by... It is generated after processing the transient time sequence of each node; It is a steady-state characteristic matrix on the order of seconds. For the minute-level trend feature matrix, the two types of non-transient features are not subjected to fractional-order enhancement processing, but retain their original standardized values ​​to ensure feature dimensionality. .

[0034] S2.4. Outlier Detection and Correction via Adaptive Neighborhood Topology Filtering; This step introduces an adaptive topological potential field filtering algorithm to achieve accurate outlier detection and correction. It also establishes a bidirectional interaction with the dynamic graph structure in S3. The specific implementation details are as follows: 1. The input data source is the enhanced node feature matrix output from S2.3. The real-time physical topology adjacency matrix output by S3 2. Based on topological potential field theory, construct the neighborhood topological potential field function for each node, and calculate the formula as follows: ; Let be the topological potential value of node i at time t; L2 norm operations; 1. The electrical distance between node i and node j is derived from the node basic attribute ledger output by S1; 2. Based on the topological potential field value, a local prediction model is constructed to calculate the feature prediction value of node i. The calculation formula is as follows: ; In the formula 4. Calculate the deviation between the measured and predicted feature values ​​of node i at time t. The calculation formula is as follows: In the formula Let be the characteristic deviation value of node i at time t; 5. If the characteristic deviation value Exceeding the preset deviation threshold based on historical statistical data and electrical physics laws If the data corresponding to that node is an outlier, then the data is determined to be an outlier; for the data determined to be outliers, the feature prediction value is used. Perform replacement and correction to generate a corrected node feature matrix, named 6. Based on the corrected node feature matrix The adjacency matrix weights of the electrical association layer in S3 are updated in reverse, realizing the bidirectional interactive fusion between this step and S3.

[0035] S2.5. Calculation of Node Risk Basic Dataset and Imputation of Missing Data; 1. Input data sources are the measurement node basic attribute ledger output by S1 and the spatiotemporal alignment dataset output by S2.1. Historical fault data from the power grid operation and maintenance management system, environmental data collected by S1, and real-time physical topology adjacency matrix output by S3. 2. For each node across the entire line, a node risk baseline dataset is calculated and generated. Specific calculation rules are illustrated below: Static historical factors: Based on long-term node operating data, the node's historical fault frequency, average fault recovery time, and equipment aging index are calculated. The equipment aging index is calculated from the equipment's years of operation and the cumulative historical load rate, constituting the node risk baseline value, named... 'i' represents the node number; Dynamic real-time factor: Integrates real-time hazard warning information and abnormal measurement data of the node within the previous 24 hours to generate the node's real-time warning level, named as follows: t represents the current time segment; Environmental catalytic factor: Based on real-time meteorological data, seasonal characteristics, and information on special power supply periods, the environmental risk weighting coefficient for the corresponding node is calculated and named as follows: 3. Based on the above three factors, calculate the real-time risk matrix of the node. The calculation formula is fixed as follows: ; Let be the real-time risk value of node i at time t; The weighting coefficient for the static historical factors is fixed at 0.4. The weighting coefficient for the dynamic real-time factor is fixed at 0.35. The weighting coefficient for environmental catalytic factors is fixed at 0.25; the weighting coefficient satisfies... 4. Real-time physical topology adjacency matrix based on S3 output To impute missing data in the feature matrix using topological constraints, the imputation formula is as follows: ; In the formula 5. After completing the missing data imputation, generate the final standardized node feature matrix, named _t_. This matrix will serve as the core data source for constructing the S3 dynamic graph structure and inputting the S4 model. 6. Calculation results of the time-series similarity of electrical quantities between nodes. The calculation process is as follows: The input for calculating the temporal similarity of electrical quantities between nodes is the temporal data of node electrical quantities obtained after S2.1 spatiotemporal alignment calibration and S2.2 multi-scale feature standardization extraction. Specifically, it is the standardized voltage time series of node i and node j within the same time window. , and standardized current timing sequence , Time window length and the time window length of time-series graph convolutional networks To maintain consistency, 12 time segments are fixed. Step 1: Calculate the similarity of electrical quantity time series patterns based on dynamic time warping; for the voltage and current time series of node i and node j, calculate the dynamic time warping distance respectively to solve the problems of phase offset and time scale inconsistency in the electrical quantity time series signals of the distribution network. The specific calculation formula is as follows: ; For the electrical quantity time sequence (voltage sequence or current sequence) corresponding to node i and node j. Let M be the optimal regularization path for two time series, and M be the total length of the regularization path. For the first in the sequence , The Euclidean distance between the elements is calculated using the following formula: ; Calculate the dynamic time warp distance of the voltage sequence respectively Dynamic time warping distance of current sequence Then, the distance values ​​are normalized to obtain the temporal morphological similarity component, with the following formula: ; In the formula, This represents the maximum value of the dynamic time warping distance of the voltage and current sequences between all nodes on the entire line within the corresponding time window; The time series morphological similarity component has a fixed value range of [0,1]. A larger value indicates a higher morphological similarity between the two electrical quantity time series. Step 2: Calculate the nonlinear correlation similarity of electrical quantity time series based on mutual information; for the voltage and current time series of node i and node j, calculate the mutual information respectively to characterize the nonlinear statistical correlation characteristics of the two time series. The specific calculation formula is as follows: ; The joint probability distribution of the electrical quantity time series of node i and node j is obtained by histogram statistics. Let the marginal probability distributions of the electrical quantity time series at nodes i and j be given, respectively; and the mutual information of the voltage series be calculated. Mutual information of current sequences Then, the mutual information values ​​are normalized to obtain the nonlinear correlation similarity component, as shown in the formula: ; In the formula, This represents the maximum value of the mutual information between voltage and current sequences among all nodes on the entire line within the corresponding time window. The nonlinear correlation similarity component has a fixed value range of [0,1]. A larger value indicates a higher degree of nonlinear correlation between the two electrical quantity time series. Step 3: Fusion to obtain the final inter-node electrical quantity time series similarity result; weighted fusion of the time series morphological similarity component and the nonlinear correlation similarity component yields the final inter-node electrical quantity time series similarity result, using the following formula: ; The weighting coefficient for the temporal morphological similarity component is fixed at 0.5. The weighting coefficients for the nonlinear correlation similarity components are fixed at 0.5; the weighting coefficients satisfy... The final result The value range is fixed at [0,1].

[0036] Processing results: 1. Standardized node feature matrix 2. Real-time risk matrix for nodes 3. The results of the time series similarity calculation of electrical quantities between nodes are named as follows: .

[0037] S3. Construction of a three-layer coupled dynamic heterogeneous graph structure constrained by Riemannian manifolds: Based on the standardized node feature matrix output from S2, a three-layer coupled dynamic heterogeneous graph structure is constructed that simultaneously represents the physical connection, real-time operation correlation, and hidden danger propagation characteristics of 10kV distribution network lines. The Riemannian manifold mapping method is introduced to map the graph structure in Euclidean space to the Riemannian manifold space, solving the pain point that conventional graph structures in this field cannot characterize the nonlinear manifold characteristics of distribution network topology. Bidirectional interaction with S2 preprocessing and S4 model encoding layers is realized. The graph structure output of this step serves as the core input for S2 preprocessing parameter optimization and S4 model convolution kernel construction. The feature output of S2 and the convolution output of S4 inversely optimize the manifold metric tensor and graph structure weights of this step.

[0038] This step consists of five sequential sub-steps, with bidirectional data transfer between the sub-steps. The specific implementation details are as follows: S3.1. Physical Topology Layer Construction; 1. Input data source is the initial physical topology wiring diagram output from S1. 1. Real-time switch status signals of the power grid data acquisition and monitoring system; 2. The physical topology layer serves as the static skeleton of the dynamic heterogeneous graph. Using each measurement node of the 10kV distribution network line as a vertex and the actual electrical connections between nodes as edges, a physical topology graph is generated and named... 3. Physical topology diagram The adjacency matrix is The rule for the values ​​of elements in the matrix is: if node i and node j have a direct electrical connection, then If there is no direct electrical connection, then 4. Physical topological adjacency matrix All sub-steps that are fed back to S2 in real time are used as the core constraints for adaptive optimization of S2 parameters.

[0039] S3.2. Construction of the electrical association layer; 1. The input data source is the standardized node feature matrix output by S2. The physical topology adjacency matrix output by S3.1 2. The electrical association layer serves as the dynamic operating layer of the dynamic heterogeneous graph. Its vertices are completely identical to those of the physical topology layer. Edge weights are dynamically calculated based on the real-time association strength between nodes, generating the electrical association graph, named... 3. Electrical Connection Diagram The adjacency matrix is The formula for calculating the elements within a matrix is: ; The electrical quantity time-series similarity matrix among all nodes of the line at time t is calculated using the dynamic time warping and mutual information fusion method, and is derived from the output of S2. 4. Electrically related adjacency matrix; Hadamard product operation; The outlier correction step in S2.4 will be fed back in real time to optimize the weight allocation of the local prediction model.

[0040] S3.3. Construction of the propagation impact layer; 1. The input data source is the real-time node risk matrix output by S2. The physical topology adjacency matrix output by S3.1 2. The propagation impact layer, as the risk evolution layer of the dynamic heterogeneous graph, has vertices completely consistent with the physical topology layer. Edge weights are calculated from the potential probability of the hazard spreading along the path, generating a propagation impact graph, named... 3. Propagation Impact Diagram The adjacency matrix is The formula for calculating the elements within a matrix is: In the formula The risk matrix of all nodes on the line at the previous time segment is derived from the output of S2; 4. Propagation influence adjacency matrix The risk matrix calculation steps are fed back to S2.5 in real time, optimizing the weight allocation of dynamic real-time factors.

[0041] S3.4. Riemannian Manifold Space Mapping and Metric Tensor Construction; This step introduces the Riemannian manifold mapping method to map the three-layer graph structure to the Riemannian manifold space, constructing a metric tensor adapted to the topology characteristics of the distribution network, and realizing bidirectional interaction with the encoding layer of the S4 model. The specific implementation content is as follows: 1. The input data source is the output of S3.1. S3.2 output S3.3 output The normalized node feature matrix output by S2 2. Map the three-layer graph structure of Euclidean space to Riemannian manifold space. 3. The dimension of the manifold space perfectly matches the dimension D of the node features; 4. Construct the metric tensor matrix of the Riemannian manifold space, named The formula for calculating the metric tensor is: ; Let i be the Riemannian manifold metric tensor corresponding to node i at time t; This is the metric weight coefficient for the physical topology layer, with a fixed value of 0.4; This is the metric weighting coefficient for the electrical association layer, with a fixed value of 0.3; The weighting coefficient for the propagation influence layer is fixed at 0.2; The metric weight coefficient for node features is fixed at 0.1; Let be the row vector corresponding to node i in the physical topological adjacency matrix at time t; 4. Based on the Riemannian manifold metric tensor, construct the geodesic distance matrix in the manifold space, named... The formula for calculating geodesic distance is: In the formula Let be the geodesic distance between node i and node j in the Riemannian manifold space at time t; 5. Riemannian manifold metric tensor Distance matrix with geodesic lines This will serve as the core input to the graph convolution kernel and attention module in the encoding layer of the S4 model. At the same time, the feature aggregation result output by S4 will back-optimize the weight coefficients of the metric tensor, realizing the bidirectional interactive fusion between this step and S4.

[0042] S3.5. Three-layer graph structure fusion and dynamic heterogeneous graph generation; 1. The input data source is the output of S3.1. S3.2 output S3.3 output The Riemannian manifold metric tensor output by S3.4 Distance matrix with geodesic lines 2. Based on the physical topology layer, the electrical association layer and the propagation influence layer are coupled and superimposed. Combined with the Riemannian manifold metric tensor, a real-time dynamic heterogeneous graph of the current time segment t is generated, named 3. Synchronously generate dynamic heterogeneous graphs The edge feature matrix, named Where M is the total number of edges in the graph; K is the feature dimension of a single edge; examples of edge features include line impedance, line length, electrical distance, Riemannian geodesic distance, real-time electrical correlation weight, and hazard propagation weight. Edge feature matrix Fixed feature dimensions For each of the 6 edge features, the standardized formula for generating the feature vector of a single edge is: ; The line impedance is standardized and mapped to the [0,1] interval using min-max standardization. The original value comes from the node basic attribute ledger of S1. The line length is standardized and mapped to the [0,1] interval using min-max standardization. The original value comes from the node basic attribute ledger of S1. : The electrical distance is standardized and mapped to the [0,1] interval using min-max standardization. The original value comes from the node basic attribute ledger of S1. The Riemannian manifold geodesic distance is standardized and mapped to the [0,1] interval using min-max standardization. The original values ​​are from S3.4. ; Real-time electrical association weights, i.e., S3.2 The value range is [0,1]. The weight of hazard propagation, i.e., S3.3 The value range is [0,1].

[0043] Processing results: 1. Real-time dynamic heterogeneous graph of the current time segment. Includes physical topology layer Electrical connection layer , Dissemination and Influence Layer 2. Riemannian manifold metric tensor Distance matrix with geodesic lines 3. Dynamic Heterogeneous Diagram The corresponding standardized node feature matrix Edge feature matrix .

[0044] S4. Construction of Topology-Aware Multimodal Graph Neural Network Architecture; The network built in this step is named TM-GNN, and the architecture consists of four sequential layers with bidirectional interaction and fusion between layers. A TM-GNN model adapted to the 10kV distribution network line fault hazard identification scenario is built. The model architecture fully matches the Riemannian manifold-constrained dynamic heterogeneous graph structure output in S3. Accurate hazard feature extraction is achieved through bidirectional interaction between the Riemannian manifold-constrained topology adaptive graph convolution module and the physically-constrained enhanced multi-head graph attention module. Accurate separation of multiple hazard features is achieved through the Tsallis non-extensive entropy-constrained hazard feature decoupling module. The outputs of all modules are bidirectionally bound to the loss function of the subsequent training process, ensuring that the model output fully conforms to the physical laws of the distribution network. Reference Figure 3 and Figure 4 The TM-GNN model architecture consists of four sequential layers. The specific construction rules and examples of the bidirectional interaction mechanism are as follows: S4.1. Model Input Layer Setup; 1. The input data source is the real-time dynamic heterogeneous graph output by S3. Node feature matrix Edge feature matrix Riemannian manifold metric tensor Geodesic distance matrix 2. The input layer consistently receives three types of input content: graph structure input: Vertex feature input: Node feature matrix Where N is the total number of nodes on the entire line, and D is the feature dimension of a single node; edge feature input: edge feature matrix , where M is the total number of edges on the entire line, and K is the feature dimension of a single edge.

[0045] S4.2. Model Encoding Layer Construction; The core of the encoding layer consists of a Riemannian manifold-constrained topological adaptive graph convolutional module and a physically-constrained enhanced multi-head graph attention module. The two modules achieve bidirectional interactive fusion. The output of the convolutional module serves as the input of the attention module, and the output of the attention module inversely optimizes the Riemannian manifold metric tensor of the convolutional module. The specific construction rules are as follows: S4.2.1. Construction of a topology-adaptive graph convolution module with Riemannian manifold constraints; 1. The core of the module is graph convolution operation in Riemannian manifold space, which solves the problem that conventional graph convolution in this field cannot adapt to the nonlinear manifold characteristics of distribution network topology. The calculation formula for convolution operation is: ; Let be the feature matrix of all nodes in the l-th layer at time t, with dimension . , Let l be the feature dimension of the l-th layer; ReLU is the non-linear activation function; L is the total number of convolutional layers; The degree matrix is ​​constructed based on the geodesic distance of the Riemannian manifold; An adjacency matrix that integrates information from multiple sources; Let be the learnable weight matrix of the l-th layer; Let be the learnable bias matrix of the l-th layer; The calculation formula is: In the formula, This is a diagonal matrix generating function. The i-th element on the diagonal is the sum of the geodesic distances between node i and all nodes on the entire line. Off-diagonal elements are fixed to 0. The distance of the geodesic line generated by the Riemannian manifold in S3.4; 2. Adjacency matrix integrating multi-source information The calculation formula is: ; I is the identity matrix; for The corresponding degree matrix; The fusion weight coefficients of the electrical correlation matrix can be learned and optimized. The fusion weight coefficients of the propagation influence matrix can be learned and optimized; 3. Based on the attention coefficients output by S4.2.2, the Riemannian manifold metric tensor is optimized inversely. The optimized formula is: ; The optimized Riemannian manifold metric tensor for node i; To measure the learning rate of the tensor, a fixed value of 0.01 is used; The final attention coefficients of node i to node j are output from S4.2.2; 4. The module output is the node aggregation feature matrix in the Riemannian manifold space, named , This matrix represents the feature dimension output by the convolution; it will serve as the core input to the S4.2.2 attention module, enabling bidirectional interaction and fusion between the two modules.

[0046] S4.2.2. Construction of a Physically Constrained Enhanced Multi-Head Graph Attention Module; The module is divided into three serial computation units, which simultaneously realize bidirectional interaction with the graph convolution module. The specific construction rules are as follows: 1. Structural Attention Calculation Unit: Used to calculate the initial feature association scores between nodes. The input data source is the node aggregation feature matrix output from S4.2.1. The calculation formula is: ; Let be the original attention score of node i to its neighbor node j in the k-th attention head at time t; The LeakyReLU nonlinear activation function is used. Calculate the parameter vector for the learnable attention score in the k-th attention head; Let be the linear transformation weight matrix of the learnable features in the k-th attention head; The convolution at node i outputs a feature vector; This is a vector concatenation operation; k is the index of the attention head, and K is the total number of attention heads; 2. Physically Constrained Attention Correction Unit: Used as prior knowledge of the physical laws of the power grid to correct the initial attention score. The input data sources are the output of the structural attention calculation unit and the Riemannian manifold geodesic distance matrix output by S3. The calculation formula is: ; Let be the final attention coefficient of node i to node j in the k-th attention head at time t; For the original attention score Initial attention coefficients after Softmax normalization; For the physical constraint function, input The physical relationship characteristics between node i and node j are, for example, electrical distance, line impedance, historical fault correlation, Riemannian manifold geodesic distance, and whether there is a direct physical connection. The output value range is fixed at [0,1]. 3. Normalization and Multi-Head Aggregation Unit: Used to perform weighted aggregation of node neighborhood information. The Softmax normalization formula is: In the formula Let i be the set of all neighboring nodes of node i; the multi-head feature aggregation formula is: In the formula The final attention feature vector output after aggregating neighborhood information for node i; 4. The module output is the attention enhancement feature matrix for all nodes on the entire network, named... , The attention outputs the feature dimension; this matrix is ​​then fed back into the graph convolution module of S4.2.1 to optimize the Riemannian manifold metric tensor, enabling bidirectional interactive fusion of the two modules.

[0047] S4.3. Model Processing Layer Construction; The core of the processing layer consists of a temporal graph convolutional network unit and a Tsallis non-widespread entropy constraint-based hazard feature decoupling module. The two modules achieve bidirectional interactive fusion. The output of the temporal convolution serves as the input to the feature decoupling module, and the output of the feature decoupling module inversely optimizes the weight parameters of the temporal convolution. At the same time, the Tsallis non-widespread entropy constraint is introduced to address the pain point that conventional methods in this field cannot separate coupled hazard features. The specific construction rules are as follows: S4.3.1. Construction of Temporal Graph Convolutional Network Units; 1. Input data source is the attention enhancement feature matrix output from S4.2. The formula for fusing the temporal and spatial topological relationships of node features is as follows: ; The spatiotemporal fusion node feature matrix is ​​output at time t; It is a long short-term memory network unit; 1. The preset time window length is fixed at 12; 2. Based on the feature decoupling results output by S4.3.2, the weight parameters of the LSTM unit are optimized in reverse. The optimization formula is as follows: ; The optimized LSTM weight matrix; The LSTM weight matrix before optimization; The learning rate of the LSTM unit; The Tsallis non-extensive entropy constraint loss is derived from the output of S4.3.2; 3. The unit output is the spatiotemporal fusion node feature matrix. , This matrix represents the spatiotemporal fusion feature dimension; it will serve as the core input of the S4.3.2 feature decoupling module, enabling bidirectional interactive fusion of the two units.

[0048] S4.3.2. Construction of the Tsallis Non-Extended Entropy Constraint Hidden Danger Feature Decoupling Module; 1. The input data source is the spatiotemporal fusion node feature matrix output from S4.3.1. This is used to separate specific feature vectors for three typical fault hazards in 10kV distribution network lines from spatiotemporal fusion features. Examples of feature vectors for the three types of hazards are as follows: Line aging hazard feature vector: [Harmonic distortion rate growth trend, three-phase imbalance anomaly value, partial discharge mode change characteristics, abnormal temperature rise slope]; Tree obstruction hazard feature vector: [Discharge pulse period, weather correlation coefficient, seasonal correlation coefficient, image recognition confidence level]; Poor contact hazard feature vector: [Contact resistance change rate, temperature-current correlation anomaly value, high-frequency oscillation energy, time-domain reflection characteristics]; 2. Construct a feature decoupling mapping network, with the decoupling formula as follows: ; Let be the decoupling feature matrix of the c-th type of hidden danger at time t, where c takes the values ​​1, 2, and 3, corresponding to the three types of hidden dangers: line aging, tree obstruction, and poor contact, respectively. The decoupling mapping matrix for the c-th type of hidden danger can be learned and optimized. The decoupling bias vector for the c-th type of hidden danger can be learned and optimized. 3. Introduce the Tsallis non-extensive entropy constraint to ensure that the decoupling characteristics of different types of hidden dangers are independent of each other and free from coupling interference. The formula for calculating the Tsallis non-extensive entropy is as follows: ; The Tsallis non-extensive entropy value for the decoupling characteristics of the c-th type of hidden danger; This is the non-extensive entropy coefficient, with a fixed value of 1.5; For the decoupling features of the c-th type of hidden danger, the feature probability distribution corresponding to node i is calculated using the following formula: In the formula, Let L2 norm be the decoupling feature vector of the c-th type of hidden danger at node i. This represents the total number of nodes on the entire line; the calculated value is... satisfy This conforms to the mathematical definition of a probability distribution; 4. Based on the Tsallis non-extensive entropy value, an entropy-constrained loss function is constructed, and the calculation formula is as follows: ; For Tsallis non-extensive entropy constraint loss; These are the weighting coefficients for the entropy constraint loss; 5. Simultaneously construct a hazard risk quantification model, with the calculation formula fixed as follows: ; Let be the risk score of the c-th type of hidden danger at node i at time t, with a value range of [0,1]. The weight coefficients for the m-th dimension feature of the c-th type of hidden danger; Let be the standardized feature value of the m-th dimension of the c-th type of hidden danger at node i at time t; The impact coefficient of the duration of the hidden danger; Let be the duration of the hidden danger characteristics at node i at time t; 6. The module output is a decoupling feature matrix for the three types of hidden dangers. Risk assessment of potential hazards at nodes Entropy-constrained loss The decoupled feature matrix is ​​input in reverse to the temporal graph convolutional network unit of S4.3.1 to optimize the temporal feature extraction weights and achieve bidirectional interactive fusion of the two units.

[0049] S4.4. Model Output Layer Construction; The output layer adopts a multi-task learning framework to simultaneously complete four core tasks. The inputs to the four tasks are the spatiotemporal fusion node features and decoupled hazard features output from S4.3. The specific construction rules are as follows: 1. Task 1: Node-level hazard classification task, used to determine whether a single node has a fault hazard. The output node is a binary classification result of normal or hazard. The classification formula is fixed as follows: ; The value of node i at time t is the hazard classification result, which ranges from [0,1]. If the value is greater than the preset classification threshold, it is determined to be a hazard node. A multilayer perceptron for node classification tasks; 1. For averaging calculations; 2. Task 2: Edge-level risk quantification task, used to quantify the propagation probability of hidden dangers along the route edges, outputting the risk value of each edge, the quantification formula is fixed as follows: ; Let be the risk value corresponding to the edge connecting node i and node j at time t, with a value range of [0,1], representing the probability that the hidden danger will propagate along this edge; Use the Sigmoid activation function; , 3. Task 3: Hazard Root Cause Localization Task, used to locate the source node of potential faults, and calculate the importance of nodes through a reverse attention mechanism, with the formula fixed as follows: ; In the formula Assign an importance score to node i at time t. The higher the score, the greater the probability that the node is the root cause of the hidden danger. 4. Task 4: Hidden danger development trend prediction task, used to predict the development trend of hidden dangers within a preset time period. The prediction formula is fixed as follows: ; ; For the future Node feature prediction matrix of duration; It is a prediction unit for a time-series graph neural network; For the future Predicted trend of potential risks at node i over a given time period; This refers to a multilayer perceptron for trend prediction tasks.

[0050] Processing results: The initialized topology-aware multimodal graph neural network TM-GNN model contains a complete input layer, encoding layer, processing layer, output layer, and a bidirectional interaction mechanism for all modules.

[0051] The S5 TM-GNN model undergoes phased training and adaptive updating using federated meta-learning. The initial TM-GNN model built in S4 is trained, optimized, and adaptively updated in stages to ensure convergence and excellent hazard identification accuracy, generalization ability, and dynamic adaptability. A federated meta-learning incremental update framework is introduced to address the shortcomings of conventional methods in this field, such as poor cross-regional adaptability and low accuracy in identifying hazard problems with small samples. The loss function in all training processes is bidirectionally bound to the model modules of S4, and the backpropagation of the loss function directly optimizes the model parameters of each module in S4. This step is divided into three sequential stages, with the specific implementation details as follows: S5.1. Large-scale unsupervised pre-training; 1. Input object is the initialized TM-GNN model output by S4; 2. Input data sources are the unlabeled historical dynamic heterogeneous graph dataset output by S3 and the historical node feature matrix dataset output by S2; 3. Pre-training uses three loss functions for joint optimization, and the total pre-training loss function is fixed as follows: ; The total loss function for pre-training; , , The preset loss weight coefficients are fixed at values ​​of 0.4, 0.35, and 0.25, respectively, and their sum is 1. The graph reconstruction loss is calculated based on the graph masking reconstruction task. Some features of nodes and edges are randomly masked, and the masked features are reconstructed through the model. The mean square error between the reconstructed features and the original features is calculated, which is used to enable the model to learn the inherent relationships and topology between distribution network nodes. The state prediction loss is calculated based on the time-series state prediction task. It predicts the state characteristics of nodes such as voltage and current in the short term in the future through historical time-series characteristics, and calculates the mean square error between the predicted value and the actual value. This is used to allow the model to learn the dynamic evolution law and time dependence of electrical physical quantities. To compare the loss, positive and negative sample pairs are constructed based on the contrastive learning task. Positive sample pairs represent enhanced feature views of the same node at different time points, while negative sample pairs represent feature views of different nodes. The loss function is used to shorten the distance between positive samples and widen the distance between negative samples, thereby enhancing the discriminative power of the model features. 4. The pre-training optimizer is fixed to the AdamW optimizer, with preset training epochs, batch size, and learning rate hyperparameters. Backpropagation minimizes the total pre-training loss. Complete the model pre-training and output the pre-trained TM-GNN model.

[0052] S5.2. Supervised Multi-Task Fine-Tuning; 1. The input is the pre-trained TM-GNN model output from S5.1; 2. The input data source is a labeled 10kV distribution network line fault hazard dataset. Examples of dataset sources include historical fault maintenance records of the power grid, manually inspected and labeled data, and simulated fault test data. The dataset contains dynamic heterogeneous graph data, node feature data, node hazard labels, edge risk labels, and hazard root cause node labels; 3. Fine-tuning uses a multi-task joint loss function for optimization. The total fine-tuning loss function is fixed as follows: ; To fine-tune the total loss function; , , , , The preset loss weight coefficients are fixed at values ​​of 0.3, 0.2, 0.2, 0.2, and 0.1, respectively, so that the sum of the five values ​​is 1. For node-level classification loss, the cross-entropy loss function is used to calculate the deviation between the model node hazard classification results and the true labels, thereby optimizing the hazard classification accuracy. To quantify the loss of edge-level risks, the mean squared error loss function is used to calculate the deviation between the model's edge risk prediction value and the actual risk label, thereby optimizing the accuracy of hazard propagation path judgment. To address the root cause localization loss, a ranking loss function is employed to calculate the deviation between the model's root cause node importance ranking results and the actual root cause labels, thereby optimizing the accuracy of hazard root cause localization. The physical constraint loss is used as a soft constraint regularization term to force the model output to conform to the core physical laws of the power grid and avoid learning false correlations. 4. Physical constraint loss (Tsallis non-extensive entropy constraint loss, output from S4.3.2) is used to optimize the independence of hazard feature decoupling; The calculation formula is: ; The actual power flowing from node i to node j is calculated using Kirchhoff's laws and Ohm's law based on the node voltage and current phasor data output by S2. 5. The estimated power from node i to node j obtained from the feature mapping within the TM-GNN model; 6. The fine-tuning optimizer is fixed to the AdamW optimizer, with preset training epochs, batch size, and learning rate hyperparameters, and the total fine-tuning loss is minimized through backpropagation. Simultaneously optimize the model parameters of all modules in S4, complete the model fine-tuning, and output the fine-tuned converged TM-GNN model.

[0053] S5.3. Federated Meta-Learning-Driven Incremental Adaptive Update; This step introduces the federated meta-learning framework to achieve cross-regional adaptive update and few-sample hazard learning capabilities for the model. It also bidirectionally binds with all model modules in S4. Specific implementation details are as follows: 1. The input object is the finely tuned and converged TM-GNN model output from S5.2; 2. When feedback is received, the input data sources are the operation and maintenance closed-loop feedback data output from S7, the newly added operational data collected in real-time by each edge node, and the newly labeled hazard data; 3. A three-level update architecture is built, consisting of local edge nodes, regional edge clusters, and a cloud analysis center. The inner-loop-outer-loop update mechanism of meta-learning is adopted, combined with the parameter aggregation strategy of federated learning. The specific update rules are as follows: Inner-loop update: Local edge computing nodes perform few-sample incremental learning based on local real-time operational data and newly labeled hazard data to optimize the shallow parameters of the model. The inner-loop update formula is: ; Let be the model parameters for the s-th local edge node; The learning rate for the inner loop; The incremental learning loss function for local nodes is consistent with the fine-tuned total loss function in S5.2; For the local dataset of the s-th local edge node; outer loop update: the regional edge cluster periodically aggregates the model parameter update amounts of local edge nodes within its jurisdiction, performs regional model aggregation, and the aggregation formula is: ; The aggregation model parameters for the r-th region edge cluster; The total number of local edge nodes in the r-th region; Let be the number of dataset samples for the s-th local edge node; Global Update: The cloud analytics center periodically receives model parameter updates from each regional edge cluster, performs global model optimization and knowledge distillation, and updates the global base model. The update formula is: ; These are the parameters of the global basic model; The learning rate for the outer loop; The global optimization loss function is defined as follows: R is the total number of regional edge clusters; global optimization loss is defined as follows: The total loss function for global updates in federated meta-learning: ; The total fine-tuning loss for the model across all regions The average value is used to ensure that the global model fits the hazard identification task across the entire region. Knowledge distillation loss for Riemannian manifold alignment: ;in, The distillation loss weighting coefficient is fixed at 0.2. This represents the total number of regional edge clusters. The Riemannian manifold geodesic distance vector for node i output by the global model; Let i be the Riemannian manifold geodesic distance vector output by the model for the r-th region. L2 regularization loss is used to prevent global model overfitting. ;in, This is the regularization coefficient, with a fixed value of 1e-4; These are the parameters of the global base model. 4. During online updates, the physical constraint loss is always retained. With entropy constraint loss As a regularization term, it ensures that the updated model still conforms to the physical laws of the power grid and the requirements for feature decoupling; 5. The cloud analysis center distributes the updated global basic model to the edge clusters of each region and local edge nodes to complete the full-link adaptive update of the model, and finally outputs a TM-GNN inference model that can be deployed in real time and continuously iterated and optimized.

[0054] Processing results: A converged and optimized TM-GNN inference model for identifying potential faults in 10kV distribution network lines is now deployable.

[0055] S6. Real-time inference and identification of fault hazards in distribution network lines based on the TM-GNN inference model; Based on the TM-GNN inference model trained by the output of S5, end-to-end inference is performed on the real-time operation data of 10kV distribution network lines, outputting complete fault hazard identification, location, risk quantification, and trend prediction results; bidirectional feedback is achieved with S2 preprocessing and S3 dynamic graph structure construction, and the hazard identification results in the inference process back-optimize the preprocessing parameters of S2 and the graph structure weights of S3. This step is divided into four serial sub-steps, with bidirectional data transfer between sub-steps. The specific implementation content is as follows: S6.1. Real-time data acquisition and standardized preprocessing; 1. The input data source is the real-time, timestamped raw measurement data acquired by the hierarchical measurement nodes in step S1, and the initial physical topology wiring diagram of the target line. 2. Following the complete rules of step S2, perform spatiotemporal alignment calibration, multi-scale feature extraction, fractional transient feature enhancement, outlier detection and correction using adaptive neighborhood topology filtering, and calculation of the node risk base dataset and missing data imputation on the real-time raw measurement data, outputting a real-time standardized node feature matrix. Real-time physical topology adjacency matrix Real-time node risk matrix 3. Based on the hidden danger reasoning results output by S6.3, the fractional order operator order and topology filter weights of this step are optimized in reverse to achieve bidirectional interactive fusion between this step and the subsequent reasoning process.

[0056] S6.2. Real-time Dynamic Heterogeneous Graph Construction; 1. Input data sources are the real-time standardized node feature matrix, real-time physical topology adjacency matrix, and real-time node risk matrix output from S6.1; 2. Following the complete rules of step S3, construct the physical topology layer, electrical association layer, and propagation influence layer in sequence, complete the Riemannian manifold space mapping and metric tensor construction, and fuse them to generate the real-time dynamic heterogeneous graph of the current time segment. Synchronously output the corresponding node feature matrix Edge feature matrix Riemannian manifold metric tensor Geodesic distance matrix 3. Based on the probability results of hazard propagation output in subsequent S6.3, update the adjacency matrix weights of the propagation influence layer in this step in reverse, so as to realize the bidirectional interactive fusion between this step and the subsequent reasoning process.

[0057] S6.3. End-to-end inference of the TM-GNN model; 1. Input object is the trained TM-GNN inference model output from S5; 2. Input data source is the real-time dynamic heterogeneous graph output from S6.2. Node feature matrix Edge feature matrix Riemannian manifold metric tensor Geodesic distance matrix 3. Import the input data into the TM-GNN inference model, and sequentially process it through the input layer, encoding layer, processing layer, and output layer to output four inference results simultaneously, as shown in the following examples: binary classification results of hidden dangers for all nodes along the entire line, including normal and hidden danger judgment results; quantitative value of hidden danger propagation risk for all line edges along the entire line, with a value range of [0,1]; location results and importance ranking of root cause nodes of hidden dangers, with higher importance scores indicating a greater probability of being a root cause node; prediction results of the risk development trend of hidden danger nodes within a preset time period in the future; 4. Based on the hidden danger risk quantification model in S4, calculate the risk score of each hidden danger node corresponding to the three types of hidden dangers to determine the type of hidden danger.

[0058] S6.4. Hazard Identification Result Verification and Classification; 1. The input data source is the TM-GNN model inference result output from S6.3; 2. Based on Kirchhoff's laws and power balance law of power grid, the consistency of the inference result is verified, and false inference results that do not conform to physical laws are eliminated; 3. Based on the hazard risk score, the hazards are divided into three levels: general hazards, important hazards, and emergency hazards according to the preset classification rules, and the final standardized hazard identification result set is generated.

[0059] Processing results: A standardized identification result set of real-time fault hazards in 10kV distribution network lines, which fully includes the location of hazard nodes, hazard type, hazard level, root cause node location results, risk quantification value, and development trend prediction results.

[0060] S7. Hazard Identification Result Output and Operation and Maintenance Closed-Loop Feedback Iteration: Standardize the output of hazard identification results to operation and maintenance personnel, and simultaneously receive operation and maintenance handling feedback data, forming a complete closed loop of data collection, analysis and identification, early warning and handling, and feedback optimization. This provides the core data source for the online adaptive update of the model in S5.3, enabling continuous iterative optimization of the model. 1. The input data source is the standardized identification result set of real-time fault hazards of 10kV distribution network lines output by S6; 2. According to the hazard level, push the identification result set to the terminal equipment of the corresponding operation and maintenance personnel. The push content includes hazard location, hazard type, hazard level, root cause node location, risk quantification value, development trend prediction, and handling priority suggestion; 3. Receive on-site handling feedback data from operation and maintenance personnel. The feedback data includes whether the hazard actually exists, the actual type of the hazard, the hazard handling result, and the handling completion time; 4. Associate and annotate the operation and maintenance feedback data with the measurement data, graph structure data, and model inference results of the corresponding time section to generate a newly added hazard annotation dataset, which is pushed to step S5.3 as the input data source for the online adaptive update of the TM-GNN model, completing the entire closed loop. Processing results: 1. Standardized 10kV distribution network line fault hazard early warning and handling instructions for operation and maintenance; 2. Newly added hazard feedback dataset with annotations for model iterative optimization.

[0061] In some embodiments, physical constraint functions The weighted fusion engineering expression based on multi-factor fuzzy membership is: In the formula, M is the total dimension of the physical constraint features. In this embodiment, M=5, which corresponds to the five physical relationship features disclosed in the original document: electrical distance, line impedance, historical fault correlation, Riemannian geodesic distance, and whether there is a direct physical connection. Let be the fuzzy membership function for the m-th physical constraint feature. The function forms for different features are as follows: Regarding the presence of a direct physical connection: the membership function is a step function; if nodes i and j have a direct physical connection, If there is no direct physical connection, ; For electrical distance, line impedance, and Riemannian geodesic distance characteristics: a semi-trapezoidal descent membership function is adopted, with the formula as follows: ; In the formula, x is the measured value of the corresponding feature. , For the preset threshold of the corresponding feature, where the electrical distance is , Line impedance , The distance of the geodesic lines of the Riemannian manifold , pu is a per-unit value; Based on the correlation characteristics of historical faults: a semi-trapezoidal ascending membership function is adopted, the formula of which is: ; In the formula, x represents the historical fault correlation degree, with a value range of [0,1]. , ; In this embodiment, the output value range of the physical constraint function is fixed at [0,1], and the quantitative embedding of physical laws is realized through multi-factor fuzzy membership fusion.

[0062] In some embodiments, the outlier determination deviation threshold Based on the 3σ criterion and the determination of the topological potential, the specific calculation formula is as follows: ; This is the threshold coefficient, with a fixed value of 3, corresponding to the 99.73% confidence interval of the 3σ criterion; The standard deviation of the characteristic deviation values ​​of node i in the previous 72 hours is calculated in real time based on historical statistical data. The topological potential value of node i is calculated using the topological potential function disclosed in S2.4. In this embodiment, the deviation threshold is adaptively adjusted according to the topological importance of the node and the fluctuation characteristics of historical data, which solves the problem of high misjudgment rate under different working conditions with a fixed threshold and ensures the engineering reproducibility of outlier detection.

[0063] In some embodiments, the generalized binomial coefficients To avoid the high computational cost of the gamma function, based on engineering-based recursive calculations in discrete sampling scenarios, the recursive formula is as follows: ; ; In the formula, The memory length of the fractional operator, i.e., the maximum number of terms in the recursive computation, is given in this embodiment. The value is fixed at 100, corresponding to the memory length of 100 sampling steps.

[0064] In some embodiments, the node hazard judgment threshold is determined based on the maximum Yoden index of the ROC curve and a dynamic correction mechanism for physical constraints. The specific implementation process is as follows: Step 1: Based on the labeled hazard dataset, using the model output... To predict the outcome, plot the ROC curve and calculate the true positive rate (TPR) and false positive rate (FPR) at different thresholds; Step 2: Calculate the Youden index, using the formula: Step 3: Select the threshold corresponding to the maximum value of the Youden index as the basic classification threshold. In this embodiment, the data is calculated based on a typical hidden danger dataset of a 10kV distribution network. The fixed value is 0.6, that is, when When node i is identified as a potential hazard node; Step 4: In real-time inference scenarios, the threshold is dynamically adjusted based on the physical constraint loss. The adjustment formula is as follows: In the formula, This represents the physical constraint loss value at the current time segment. This represents the average physical constraint loss over the past 72 hours.

[0065] In some embodiments, the hazard feature decoupling mapping network is a 3-layer fully connected multilayer perceptron structure, corresponding to the mapping matrix in the decoupling formula. With bias vector The network structure is as follows: Input layer: number of neurons and spatiotemporal fusion feature dimension Completely identical, with no activation function; Hidden layer: the number of neurons is fixed at 64, the activation function is LeakyReLU, the negative slope is fixed at 0.2, a Dropout layer is added, and the inactivation rate is fixed at 0.3 to avoid overfitting; Output layer: the number of neurons is fixed at 4, corresponding to the 4-dimensional feature vector of each type of hazard, the activation function is Sigmoid, ensuring that the output feature value is mapped to the [0,1] interval; In this embodiment, the three types of hazards correspond to independent decoupling mapping networks, the network structures are completely identical, and they are trained and optimized separately to ensure that the feature decoupling of different types of hazards is independent of each other.

[0066] In some embodiments, during the global optimization phase of federated meta-learning, the niche distillation method and distillation loss formula based on Riemannian manifold alignment are as follows: The knowledge distillation employs Riemannian manifold feature space alignment to achieve topological knowledge transfer between the global model and the regional model. The distillation loss formula is: ; To compensate for the knowledge distillation loss, a global optimization loss function is added. China Joint Optimization; The distillation loss weighting coefficient is fixed at 0.2. This is the geodesic distance vector of node i in the Riemannian manifold space, output by the global model. Let R be the geodesic distance vector of node i in the Riemann manifold space output by the model of the r-th region; R is the total number of regional edge clusters, and N is the total number of nodes. This embodiment solves the problem of poor model adaptability caused by large differences in the topology of distribution networks in different regions by knowledge distillation aligned with Riemann manifolds, and replaces conventional logits distillation or feature distillation.

[0067] In some embodiments, the graph reconstruction loss The mean squared error loss with masking weights is constructed, and the calculation formula is as follows: ; Let be the feature masking matrix. If the d-th dimension of the feature of node i is randomly masked, then ,otherwise The concealment ratio is fixed at 15%; The d-th dimension feature value of node i obtained from model reconstruction; The original normalized node feature values ​​before masking are derived from the normalized node feature matrix output by S2. .

[0068] In some embodiments, the state prediction loss The time-series mean squared error loss is used for construction, and the calculation formula is as follows: ; To predict the duration, three time segments are fixed. The future predicted by the model The state feature matrix of node i at time time i has the same feature dimension as the core electrical quantity feature dimension of the node. For the future of real-world data collection The state feature matrix at time node i.

[0069] In some embodiments, the contrast loss The InfoNCE loss function, adapted to the Riemannian manifold space, is used for construction, and the calculation formula is as follows: ; The feature vector of the Riemannian manifold space is obtained by encoding the original features of node i through the model; The feature vector (positive sample) obtained by model encoding of the feature-enhanced view of node i is obtained by Gaussian noise injection and random masking. The feature vector (negative sample) is obtained by encoding the features of node j through the model. ; Let be the similarity function of the Riemannian manifold space, and the calculation formula is: ,in For Riemannian manifold metric tensors Mahalanobis distance; This is the temperature coefficient, with a fixed value of 0.07.

[0070] In some embodiments, the node classification loss A binary cross-entropy loss function with class balancing weights is used to construct a loss function that is suitable for scenarios with imbalanced samples of potential hazards in power distribution networks. The calculation formula is as follows: ; Let i be the actual hazard label for node i. Positive samples (hazard nodes) take a value of 1, and negative samples (normal nodes) take a value of 0. The probability value for the hidden danger classification of node i is output by the model, and its value range is [0,1]. The positive sample balancing weight is set to "number of negative samples / total number of samples"; The negative sample balancing weight is set to "number of positive samples / total number of samples".

[0071] In some embodiments, the edge-level risk quantification loss The mean squared error loss function with topological weights is used for construction, and the calculation formula is as follows: ; This represents the total number of edges along the entire route. It is the set of all edges of the entire path; The risk quantification value of the edge connecting nodes i and j in the model output, with a value range of [0,1]; The true risk label for the edge is obtained by labeling historical fault propagation paths, and its value range is [0,1]. The topological weight of the edge is "the sum of the degrees of node i and node j / the maximum degree of all nodes on the entire path".

[0072] In some embodiments, the root cause localization loss It is constructed using a pairwise sorting hinge loss function, adapted to the root cause node sorting task: ; It is the set of true root cause nodes; It is the set of non-root cause nodes; Assign an importance score to node i from the model output; To maximize the function, ensure that the loss value is non-negative.

Claims

1. An artificial intelligence-based method for identifying potential faults in power distribution network lines, characterized in that, include: S1. Complete the hierarchical deployment of measurement nodes for the distribution network lines, and synchronously collect multi-source heterogeneous raw measurement datasets with unified timestamps through each node. At the same time, obtain the initial physical topology wiring diagram of the target line from the power grid production management system. S2. Using the multi-source heterogeneous original measurement dataset and the initial physical topology wiring diagram as input, complete the data standardization preprocessing and output the standardized node feature matrix, the real-time risk matrix of all line nodes and the time series similarity results of electrical quantities between nodes; S3. Using the standardized node feature matrix, the real-time risk matrix of all line nodes, the time-series similarity results of electrical quantities between nodes, and the initial physical topology wiring diagram as input, construct a three-layer coupled real-time dynamic heterogeneous graph and synchronously output the edge feature matrix corresponding to the real-time dynamic heterogeneous graph. S4. Build a topology-aware multimodal graph neural network model, using real-time dynamic heterogeneous graphs, standardized node feature matrices, and edge feature matrices as inputs, to complete feature extraction and hazard identification multi-task outputs; S5. Using the topology-aware multimodal graph neural network model as the object, perform phased training and optimization, and output a converged deployable inference model; S6. Based on the deployable inference model, input the real-time operation data of the distribution network lines, perform end-to-end inference, and output a standardized identification result set of real-time fault hazards of the lines.

2. The artificial intelligence identification method for potential faults in power distribution network lines according to claim 1, characterized in that, Also includes: S7. Based on the standardized identification result set of real-time line fault hazards, complete the operation and maintenance early warning push and on-site handling feedback data collection, and after associating and labeling the feedback data, send it back to the model training and optimization stage of S5 to complete the model adaptive update and the whole process closed loop.

3. The artificial intelligence identification method for potential faults in power distribution network lines according to claim 1, characterized in that, In the data standardization preprocessing of S2, after the multi-scale feature standardization extraction is completed, the extracted millisecond-level transient features are subjected to targeted enhancement processing through an adaptive fractional calculus operator. The order of the fractional calculus operator is adaptively optimized based on the physical topological adjacency matrix of the real-time dynamic heterogeneous graph generated in the previous time segment. The enhanced features are used to generate the standardized node feature matrix of the current time segment.

4. The artificial intelligence identification method for potential faults in power distribution lines according to claim 1, characterized in that, In S3, after the construction of the real-time dynamic heterogeneous graph with three-layer coupling is completed, the real-time dynamic heterogeneous graph in Euclidean space is mapped to the Riemannian manifold space to construct a Riemannian manifold metric tensor and a geodesic distance matrix that are adapted to the nonlinear characteristics of the distribution network topology. The Riemannian manifold metric tensor and the geodesic distance matrix are synchronously input into the topology-aware multimodal graph neural network model as the core constraint parameters for model feature extraction.

5. The artificial intelligence identification method for potential faults in power distribution network lines according to claim 1, characterized in that, The topology-aware multimodal graph neural network model of S4 is equipped with a topology-adaptive graph convolution module with Riemannian manifold constraints and a multi-head graph attention module with enhanced physical constraints. The feature output of the graph convolution module is used as the input of the attention module, and the output of the attention module back-optimizes the Riemannian manifold metric tensor of the graph convolution module. The two modules achieve bidirectional interactive collaborative optimization.

6. The artificial intelligence identification method for potential faults in power distribution network lines according to claim 1, characterized in that, The topology-aware multimodal graph neural network model of S4 is equipped with a temporal graph convolutional network unit and a hazard feature decoupling module with Tsallis non-extensive entropy constraints. The spatiotemporal fusion features output by the temporal graph convolutional network unit are used as the input of the hazard feature decoupling module. The decoupling result of the hazard feature decoupling module back-optimizes the weight parameters of the temporal graph convolutional network unit. The two units achieve bidirectional interactive collaborative optimization.

7. The artificial intelligence identification method for potential faults in power distribution lines according to claim 1, characterized in that, In the phased training optimization of S5, large-scale unsupervised pre-training is performed first. The unsupervised pre-training adopts a method of joint optimization of three loss functions: graph reconstruction loss, state prediction loss and contrast loss. By minimizing the weighted sum of the three losses through backpropagation, the model learns the basic spatiotemporal representation of the distribution network and the node topology association characteristics.

8. The artificial intelligence identification method for potential faults in power distribution network lines according to claim 1, characterized in that, In the phased training optimization of S5, supervised multi-task fine-tuning is performed after unsupervised pre-training is completed. The supervised multi-task fine-tuning adopts a method of joint optimization of five loss functions: node classification loss, edge-level risk quantification loss, root cause localization loss, physical constraint loss, and entropy constraint loss. Among them, the physical constraint loss is used to force the model output to conform to the core physical laws of power balance and Kirchhoff's laws in the distribution network, so as to avoid the model learning false correlation features.

9. The artificial intelligence identification method for potential faults in power distribution lines according to claim 1, characterized in that, In the end-to-end inference process of S6, the generated current time section fault hazard identification result is used to reversely optimize the parameters of the next time section data standardization preprocessing and the construction weight of the real-time dynamic heterogeneous graph, thereby realizing bidirectional interaction and real-time optimization between the inference process and the preprocessing and graph structure construction stages.

10. The artificial intelligence identification method for potential faults in power distribution network lines according to claim 2, characterized in that, The S7 model adaptive update adopts a federated meta-learning driven incremental update mechanism and builds a three-level update architecture of local edge nodes, regional edge clusters, and cloud analysis center. Local edge nodes perform lightweight incremental learning, regional edge clusters aggregate model parameters for their respective regions, and cloud-based analysis centers perform global model optimization and knowledge distillation, achieving cross-regional model adaptation and learning of potential risks in small samples while protecting data privacy.

Images