Intelligent troubleshooting method and system for faults of power distribution and utilization ring network

Abstract

The invention relates to the technical field of the intelligent power grid industry, in particular to a power distribution and utilization ring network fault intelligent troubleshooting method and system. According to the method, a looped network situation primitive is generated through multi-source heterogeneous data fusion. The primitives are input into synchronous compression transformation to extract transient fault initial features, and the transient fault initial features are mapped into fault propagation path features through a graph convolutional network considering topological structure constraints; then, inputting the features into a space-time correlation prediction model, integrating defect type judgment and trend pre-judgment features, and generating a fault diagnosis evidence set; and starting an investigation mechanism based on the evidence set to determine an interval to be investigated, and performing closed-loop verification by constructing a fusion judgment matrix based on multi-modal physical sensing information to realize physical space locking of a fault defect point. According to the method, deep learning diagnosis and multi-mode positioning are integrated, and the detection precision and the positioning efficiency of looped network defects and the identification capability of transient hidden faults are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of smart grid industry technology, specifically to a method and system for intelligent fault diagnosis in distribution ring networks. Background Technology

[0002] With rapid socio-economic development and continuous growth in electricity demand, building stable and efficient intelligent power distribution systems has become a core objective of the modern power industry. Intelligent power distribution systems, through the deep integration of advanced information technology, communication technology, and power automation technology, aim to improve the operational efficiency, power supply reliability, and self-recovery capabilities of the power distribution network. However, in actual operation, intelligent power distribution systems still face severe challenges related to faults. Factors such as natural environmental influences, equipment aging, and external damage can all lead to various faults in power distribution lines, seriously threatening power supply quality and user experience.

[0003] Currently, common techniques for fault detection and location in intelligent power distribution systems include real-time monitoring of operating parameters such as current, voltage, and temperature. When monitoring data shows abnormalities, the system can preliminarily determine that a fault may exist in the intelligent power distribution system. Some existing technologies receive fault information reported by users, calculate regional fault values ​​using preset models, and then compare them with fault thresholds to narrow down the fault area. Simultaneously, some methods also attempt to incorporate image information to classify and identify specific areas to assist in indicating risk locations. These technologies aim to achieve preliminary detection of cable, transmission line, or network faults in intelligent power distribution systems.

[0004] However, fault detection and location in current intelligent power distribution systems face severe challenges. Existing fault detection methods are inaccurate and inefficient, relying primarily on single or limited operating parameters, making it difficult to effectively identify complex fault types. For cable, transmission line, or network faults in intelligent power distribution systems, especially in complex ring network structures, traditional methods struggle to accurately pinpoint the fault location. This leads to vague fault location, often only providing a broad "risk area," severely impacting fault repair efficiency and power supply reliability. Existing technologies lack sufficient sensitivity to transient, concealed, or developing faults in intelligent power distribution systems, hindering effective early warning and timely handling. These core deficiencies necessitate an intelligent solution that can effectively improve the accuracy and efficiency of fault diagnosis in intelligent power distribution systems.

[0005] To address this, a method and system for intelligent fault diagnosis of power distribution ring networks is proposed, which solves the shortcomings of existing technologies in fault diagnosis accuracy, as well as the limitations of data sources and analysis dimensions. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for intelligent fault diagnosis in power distribution ring networks. It achieves intelligent diagnosis and dynamic definition of faults through multi-source heterogeneous data fusion and deep learning models, and completes the precise location of fault points by combining multi-modal physical sensing information, thereby realizing accurate fault diagnosis and precise location of intelligent power distribution ring networks.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for intelligent fault diagnosis of power distribution ring networks includes: Acquire instantaneous electrical data, environmental disturbance signals, switchgear operation information, and historical operating condition data of the intelligent power distribution ring network; based on a multi-source heterogeneous data fusion strategy of timestamp alignment and spatial topology association, map the data to a unified spatiotemporal dimension to generate ring network status primitives; The signal time-frequency analysis inputs the ring network situation primitives into the synchronous compression transform to extract the instantaneous fault initiation features; the fault initiation features are then input into a graph convolutional network that considers the ring network topology constraints to map them into fault propagation path features. The initial fault characteristics and fault propagation path characteristics are input into the spatiotemporal correlation prediction model, and fault type discrimination characteristics and trend prediction characteristics are integrated to generate an intelligent fault diagnosis evidence set. Establish a troubleshooting initiation mechanism based on intelligent diagnostic evidence, using the intelligent fault diagnosis evidence set as the trigger mechanism for the troubleshooting process, and determine the fault-to-investigate interval based on the fault propagation path characteristics in the evidence set. A multimodal fusion decision matrix based on physical sensing information is constructed, which includes geographic information matching, acoustic source localization, thermal imaging detection, and fiber optic vibration analysis. Closed-loop verification is performed on the area to be investigated to determine the physical spatial location of the fault point.

[0008] Preferably, the process of generating the ring network status primitive includes: acquiring instantaneous electrical data of the intelligent distribution ring network, the instantaneous data including instantaneous values ​​of three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence current; acquiring environmental disturbance signals, the environmental disturbance signals including temperature signals, humidity signals, wind speed signals, and rainfall signals around the line; acquiring operating information of intelligent distribution switchgear, the operating information including the opening and closing status of circuit breakers, the opening and closing status of isolating switches, the number of recloser operations, and the opening and closing times; acquiring historical operating condition data of the intelligent distribution system, the historical operating condition data including ring network operating voltage records and ring network operating data over a past period. The system records current flow and fault alarm events; it timestamps and calibrates the instantaneous electrical data, environmental disturbance signals, switching equipment operation information, and historical operating condition data of the intelligent distribution ring network to achieve time synchronization between various intelligent distribution data sources; based on the actual physical connection relationship of the intelligent distribution ring network, it establishes an intelligent distribution ring network topology model, and uses the topology model to accurately map and logically associate the spatial physical locations of each intelligent distribution data source; it integrates the various types of time-synchronized intelligent distribution data with the spatial physical location mapping and logical association results, and uses a multi-feature-level fusion method to generate intelligent distribution ring network status primitives in a unified spatiotemporal coordinate system.

[0009] Preferably, the process of inputting the ring network situation primitive into the synchronous compression transform signal time-frequency analysis to extract the instantaneous fault initiation features includes: obtaining and separating the three-phase voltage instantaneous data and three-phase current instantaneous data of different monitoring points on each feeder and tie line from the ring network situation primitive; the ring network situation primitive, as a real-time data structure, includes references and simplified storage of associated original high-frequency data during generation; inputting the separated original high-frequency instantaneous data into the synchronous compression transform algorithm; performing time-frequency analysis on the voltage instantaneous data and current instantaneous data through the synchronous compression transform algorithm to obtain the energy distribution at different times and frequencies; identifying and extracting voltage mutation points, current mutation points, harmonic component anomalies, and transient oscillation components from the time-frequency analysis results based on a preset fault feature threshold; the fault feature threshold is determined based on statistical analysis of historical operating condition data of the smart distribution ring network and electromagnetic transient simulation results; and using the voltage mutation points, current mutation points, harmonic component anomalies, and transient oscillation components as instantaneous fault initiation features in the form of structured feature vectors to characterize the initial occurrence state of faults in the smart distribution ring network.

[0010] Preferably, the process of inputting the initial fault features into a graph convolutional network considering the ring network topology constraints and mapping them to fault propagation path features includes: constructing a topology graph of the smart distribution ring network, where the nodes of the graph represent each feeder node and tie node of the smart distribution ring network, and the edges of the graph represent the connection relationships between feeders and ties in the smart distribution ring network; using the initial fault features as the initial input features for the corresponding nodes in the topology graph; and performing multi-layer feature aggregation and transformation on the topology graph through the graph convolutional network, so that the feature learning of each node integrates its own initial fault features to... The system collects and transforms the initial fault characteristics of adjacent nodes. During the feature aggregation and transformation process, the propagation weights of the graph convolutional network are adjusted using the topological constraints of the smart distribution ring network. The magnitude of the propagation weights is inversely proportional to the electrical impedance of the connection, giving priority to feature associations along the actual electrical connection path to reflect the propagation trend of the fault in the smart distribution ring network. Through the output layer of the graph convolutional network, the fault propagation direction and fault impact range are extracted from the learned node features and mapped to the fault propagation path features of the smart distribution ring network to detect cable and transmission line faults in the smart distribution ring network.

[0011] Preferably, the process of generating the intelligent fault diagnosis evidence set includes: jointly feeding the instantaneous fault initiation features and fault propagation path features into a pre-trained spatiotemporal correlation prediction model; the spatiotemporal correlation prediction model is based on a deep learning architecture, including a recurrent neural network layer for processing time series data and a graph attention network layer for processing spatial topological relationships; extracting the temporal dynamic features of fault occurrence and development through the recurrent neural network layer; evaluating the spatial correlation strength of the fault in the intelligent distribution ring network topology through the graph attention network layer in combination with the fault propagation path features; fusing the temporal dynamic features and spatial correlation features to form a comprehensive representation of the current ring network fault state; based on the comprehensive representation, determining the current fault type through the model's output layer, the fault type discrimination features including short-circuit fault, ground fault, open-circuit fault, and insulation degradation fault; based on the comprehensive representation, predicting the fault development trend in the future through the model's output layer, the trend prediction features including the fault range expansion trend, the fault severity increase trend, and the fault location drift trend; and combining the fault type discrimination features and the trend prediction features to form an intelligent fault diagnosis evidence set, indicating the fault status of the intelligent distribution ring network.

[0012] Preferably, the process of determining the fault investigation interval based on the fault propagation path characteristics in the evidence set includes: issuing a fault alarm signal when the fault type discrimination characteristics in the intelligent fault diagnosis evidence set reach a preset alarm threshold; using the fault alarm signal as an explicit instruction to initiate the intelligent power distribution fault investigation process; extracting the fault propagation path characteristics contained in the intelligent fault diagnosis evidence set, wherein the fault propagation path characteristics indicate the direction and range of fault propagation along the intelligent power distribution ring network topology; and calculating and defining a physical area containing all potential fault points based on the fault propagation path characteristics, combined with the geographical information and equipment distribution of the intelligent power distribution ring network, using a real-time updated topology path analysis algorithm, wherein the area is defined as the fault investigation interval for accurate fault location determination, guiding subsequent physical sensor information collection and verification.

[0013] Preferably, the process of determining the physical space locking of the fault point includes: acquiring geographic information matching data within the investigation interval, the geographic information matching data being derived from the spatial location information of the intelligent power distribution line equipment; acquiring acoustic source positioning data, the acoustic source positioning data being derived from the abnormal sound acquisition and algorithm processing of the intelligent power distribution line; acquiring thermal imaging detection data, the thermal imaging detection data being derived from the temperature scan of the intelligent power distribution line and its surrounding environment; acquiring fiber optic vibration analysis data, the fiber optic vibration analysis data being derived from the vibration mode monitoring and analysis of the intelligent power distribution line; performing feature extraction and standardization processing on the acquired data to transform heterogeneous data into unified fault tendency evidence; inputting the fault tendency evidence into a multimodal fusion decision matrix, the decision matrix performing weighted fusion decision based on Bayesian inference; determining the precise location with the highest fault probability within the investigation interval based on the output of the fusion decision matrix, the precise location being used as the physical space locking result of the fault point; comparing and verifying the physical space locking result with the intelligent fault diagnosis evidence set to complete the closed-loop verification of the intelligent power distribution ring network fault.

[0014] A power distribution ring network fault diagnosis system includes: The data processing and analysis module acquires instantaneous electrical data, environmental disturbance signals, switchgear operating information, and historical operating condition data of the intelligent power distribution ring network. Based on a multi-source heterogeneous data fusion strategy of timestamp alignment and spatial topology association, it maps the data to a unified spatiotemporal dimension to generate ring network status primitives. The data processing and analysis module is used to input the ring network status primitives into a synchronous compression transformation algorithm for signal time-frequency analysis to extract instantaneous fault initiation features. It is also used to input the fault initiation features into a graph convolutional network model considering the ring network topology constraints to map them into fault propagation path features. The intelligent diagnosis module inputs the instantaneous fault initial characteristics and fault propagation path characteristics into the spatiotemporal correlation prediction model, integrates fault type discrimination characteristics and trend prediction characteristics, and generates an intelligent fault diagnosis evidence set. The troubleshooting and location module establishes a troubleshooting initiation mechanism based on the intelligent fault diagnosis evidence. It uses the intelligent fault diagnosis evidence set as the trigger mechanism for the troubleshooting process and determines the fault's investigation interval based on the fault propagation path characteristics in the evidence set. The troubleshooting and location module is used to construct a multimodal fusion decision matrix based on physical sensor information, including geographic information matching, acoustic source localization, thermal imaging detection, and fiber optic vibration analysis. It also performs closed-loop verification on the investigation interval to determine the physical spatial location of the fault point.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention generates ring network situation primitives through a multi-source heterogeneous data fusion strategy, extracts instantaneous fault initiation features by combining synchronous compression transformation, and maps fault propagation path features using a graph convolutional network that considers the constraints of the ring network topology. This substantially solves the shortcomings of traditional fault diagnosis methods, such as incomplete extraction of complex and variable fault features, insensitivity to transient and early faults, and neglect of the influence of the ring network topology. Thus, in the scenario of smart distribution ring networks, it significantly improves the comprehensiveness, timeliness, and accuracy of fault judgment.

[0016] 2. This invention collaboratively inputs the instantaneous fault initiation characteristics and fault propagation path characteristics into a spatiotemporal correlation prediction model, enabling deep integration of fault type discrimination features and trend prediction features to generate an intelligent fault diagnosis evidence set, and dynamically determine the fault's investigation range based on this set. This collaborative design overcomes the limitations of existing technologies in fault early warning and range narrowing, achieving proactive diagnosis and efficient troubleshooting guidance for faults in intelligent distribution ring networks.

[0017] 3. Within the defined investigation area, this invention constructs a fusion decision matrix based on multimodal physical sensing information such as geographic information matching, acoustic source localization, thermal imaging detection, and fiber optic vibration analysis. Closed-loop verification is then performed to ultimately determine the physical spatial location of the fault point. This collaborative fusion and verification of multimodal evidence completely overcomes the problems of low positioning accuracy and susceptibility to single-sensor limitations in existing technologies. It provides a highly robust and accurate final location method for faults in smart power distribution ring networks, effectively shortening fault repair time and improving power supply reliability. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an intelligent fault diagnosis method for distribution network ring networks according to the present invention. Figure 2 This is a schematic diagram of the fault feature extraction process according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent fault diagnosis system for power distribution ring networks according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figures 1 to 3 This invention provides an intelligent fault diagnosis method and system for power distribution ring networks, the technical solution of which is as follows: Example 1:

[0021] Reference Figure 1 This is a flowchart illustrating an intelligent fault diagnosis method for distribution network ring networks according to the present invention. This embodiment provides a specific application of the intelligent fault diagnosis method for distribution network ring networks, and the specific steps include: Acquire instantaneous electrical data, environmental disturbance signals, switchgear operation information, and historical operating condition data of the intelligent power distribution ring network; based on a multi-source heterogeneous data fusion strategy of timestamp alignment and spatial topology association, map the data to a unified spatiotemporal dimension to generate ring network status primitives; The signal time-frequency analysis inputs the ring network situation primitives into the synchronous compression transform to extract the instantaneous fault initiation features; the fault initiation features are then input into a graph convolutional network that considers the ring network topology constraints to map them into fault propagation path features. The initial fault characteristics and fault propagation path characteristics are input into the spatiotemporal correlation prediction model, and fault type discrimination characteristics and trend prediction characteristics are integrated to generate an intelligent fault diagnosis evidence set. Establish a troubleshooting initiation mechanism based on intelligent diagnostic evidence, using the intelligent fault diagnosis evidence set as the trigger mechanism for the troubleshooting process, and determine the fault-to-investigate interval based on the fault propagation path characteristics in the evidence set. A multimodal fusion decision matrix based on physical sensing information is constructed, which includes geographic information matching, acoustic source localization, thermal imaging detection, and fiber optic vibration analysis. Closed-loop verification is performed on the area to be investigated to determine the physical spatial location of the fault point.

[0022] Furthermore, the process of generating the ring network status primitive includes: acquiring instantaneous electrical data of the intelligent distribution ring network, the instantaneous data including instantaneous values ​​of three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence current; acquiring environmental disturbance signals, the environmental disturbance signals including temperature, humidity, wind speed, and rainfall signals around the line; acquiring operating information of intelligent distribution switchgear, the operating information including the opening and closing status of circuit breakers, the opening and closing status of isolating switches, the number of recloser operations, and the opening and closing times; acquiring historical operating condition data of the intelligent distribution system, the historical operating condition data including ring network operating voltage records and ring network data over a past period. The system records operating current and fault alarm events; it timestamps and calibrates the instantaneous electrical data, environmental disturbance signals, switchgear operating information, and historical operating condition data of the intelligent distribution ring network to achieve time synchronization between various intelligent distribution data sources; it establishes an intelligent distribution ring network topology model based on the actual physical connection relationship of the intelligent distribution ring network, and uses the topology model to accurately map and logically associate the spatial physical locations of each intelligent distribution data source; it integrates the various types of time-synchronized intelligent distribution data with the spatial physical location mapping and logical association results, and uses a multi-feature-level fusion method to generate intelligent distribution ring network status primitives in a unified spatiotemporal coordinate system.

[0023] Specifically, in a typical smart distribution ring network application scenario, smart sensors deployed on feeders and tie lines collect instantaneous values ​​of three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence current every 10 milliseconds. Smart sensors or meteorological monitoring equipment deployed along the lines provide temperature, humidity, wind speed, and rainfall signals every minute. The ring network automation terminal reports the opening and closing status of circuit breakers, isolating switches, recloser operations, and opening and closing times in real time, updating every second. Historical operating data is imported from management at hourly granularity, including voltage and current fluctuation records for the past 30 days and details of all fault alarm events.

[0024] After completing the acquisition and aggregation of multi-source data, and receiving all the aforementioned data, the timestamps of high-frequency instantaneous data are first calibrated at the millisecond level. For example, synchronization is performed using a network time protocol to ensure that data recorded by different sensors are precisely aligned on the timeline. For low-frequency environmental data and switch status data, second-level timestamp calibration is performed. After calibration, a topology diagram containing all feeders, tie lines, nodes, and switchgear is established based on pre-stored intelligent distribution ring network geographic information data. Each node and edge in the diagram is associated with its precise geographic coordinates and physical attributes. Various types of data are mapped to the corresponding nodes or edges in the topology diagram according to their acquisition location.

[0025] Finally, a multi-feature-level fusion method is used to integrate and generate intelligent distribution ring network status primitives under a unified spatiotemporal coordinate system. The fusion process employs feature vector concatenation. Instantaneous electrical data features, environmental disturbance signal features, switchgear operating information features, and historical operating condition features are concatenated into a unified ring network status primitive feature vector, serving as a fusion representation of multi-source heterogeneous data in a unified spatiotemporal dimension. The specific steps are as follows: Feature extraction and downsampling: For electrical instantaneous data acquired over 10 milliseconds, the effective value and statistical mean are calculated within each 1-second time window, serving as high-frequency features. Environmental signals and switch status quantities are directly used as low-frequency features. Mean normalization: Min-maximum normalization is employed to unify the dimensions of all features.

[0026] ; in, These are the eigenvalues ​​to be normalized. Represents the original feature values. and These are the minimum and maximum observed values ​​of this feature type in historical operating condition data. This processing effectively reduces the dynamic range and floating-point operation complexity in subsequent calculations.

[0027] Feature vector concatenation and fusion: Instantaneous electrical quantity features, environmental quantities, switch state quantities, and historical quantities, after being synchronized, spatially mapped, and normalized, are concatenated and fused using a multi-dimensional feature vector concatenation and fusion method to form a ring network situational primitive under a unified spatiotemporal dimension. The ring network situational primitive is based on a quadruple... It is stored and represented in the form of , where For timestamps, and For spatial coordinates, and This is the fused high-dimensional feature vector. The feature vector includes at least: the normalized effective value of the three-phase voltage, the effective value of the zero-sequence current, the line segment temperature value, and the circuit breaker status indicator (0 or 1).

[0028] This invention comprehensively gathers multi-source heterogeneous data from intelligent distribution ring networks and effectively overcomes the traditional data silo problem through timestamp calibration and spatial topology correlation. The generated intelligent distribution ring network situation primitives, due to their clear data structure and the integration of downsampling and normalization optimization, improve the computational efficiency of subsequent fault diagnosis and analysis models, thereby enhancing the comprehensiveness of data perception and the accuracy of fault location and diagnosis.

[0029] Furthermore, the process of inputting the ring network situation primitive into the synchronous compression transform signal time-frequency analysis to extract the instantaneous fault initial characteristics includes: obtaining and separating the three-phase voltage instantaneous data and three-phase current instantaneous data of different monitoring points on each feeder and tie line from the ring network situation primitive; the ring network situation primitive, as a real-time data structure, includes references and simplified storage of associated original high-frequency data during generation; inputting the separated original high-frequency instantaneous data into the synchronous compression transform algorithm; performing time-frequency analysis on the voltage instantaneous data and current instantaneous data through the synchronous compression transform algorithm to obtain the energy distribution at different times and frequencies; identifying and extracting voltage mutation points, current mutation points, harmonic component anomalies, and transient oscillation components from the time-frequency analysis results based on preset fault feature thresholds; the fault feature thresholds are determined based on statistical analysis of historical operating condition data of the smart distribution ring network and electromagnetic transient simulation results; and using the voltage mutation points, current mutation points, harmonic component anomalies, and transient oscillation components as instantaneous fault initial characteristics in the form of structured feature vectors to characterize the initial occurrence state of faults in the smart distribution ring network. Reference Figure 2 A flowchart illustrating the process of fault feature extraction.

[0030] Specifically, from the generated intelligent distribution ring network situational awareness unit, for each feeder (e.g., feeder A, feeder B) and tie line (e.g., tie line L1) within the ring network, multiple monitoring points (e.g., monitoring points M1, M2) are extracted in time series using a data referencing mechanism. This ring network situational awareness unit, as an integrated real-time data structure, contains a unique index or pointer to the associated original high-frequency electrical instantaneous data storage location. This unit uses a simplified storage method to record the spatiotemporal location information and basic statistics of the data, while ensuring non-destructive access to the original high-frequency waveform data through an indexing mechanism. During time-frequency analysis, the original high-frequency waveform data at that location and time is accurately retrieved based on the spatiotemporal coordinates and embedded references of the ring network situational awareness unit. For example, for the current data sequence of monitoring point M1 from 500 milliseconds before the fault to 500 milliseconds after the fault, the retrieved high-frequency electrical instantaneous data is input into a synchronous compression transform for time-frequency analysis. The synchronous compression transform algorithm performs continuous wavelet transform on the input data, using Morlet wavelet basis functions as the fundamental transform kernel functions and implementing it based on the generalized S-transform framework. The initial time window length is set to 20ms (one power frequency cycle). By calculating the instantaneous frequency of the wavelet coefficients, the wavelet coefficients are redistributed and precisely compressed on the time-frequency plane, resulting in a time-frequency representation with a final frequency resolution of 1Hz and extremely high time-frequency aggregation, clearly showing the precise distribution of signal energy at different times and frequencies. For example, during normal operation, the current signal time-frequency diagram shows energy concentration at the 50Hz fundamental frequency, while during a fault, high-frequency transient components or specific harmonic energy enhancement may occur. To ensure numerical stability under low signal-to-noise ratio or non-stationary signals, a lower bound threshold is set for the amplitude of the wavelet coefficients before calculating the instantaneous frequency; wavelet coefficients below this threshold are not included in the instantaneous frequency calculation and redistribution to avoid outliers caused by the denominator approaching zero, ensuring the robustness of the time-frequency aggregation effect. Subsequently, these time-frequency analysis results are analyzed to identify whether there are any transient abrupt changes. For example, under a specific operating condition, if the instantaneous voltage value drops by more than 20% of its rated value within 10 milliseconds, or the instantaneous current value rises by more than 300% of its steady-state value within 20 milliseconds, it can be identified as a sudden change point. The "rated value" in the voltage sudden change threshold is taken from the system's preset standard voltage value (e.g., 10kV), and the criterion for judging a drop of more than 20% is this rated value. The "steady-state value" in the current sudden change threshold is defined as the root mean square (RMS) value of the current data within 100 milliseconds before the fault occurs. At the same time, by analyzing the continuous presence or abnormal enhancement of non-fundamental frequency energy on the time-frequency graph (e.g., the energy of the 3rd and 5th harmonic components exceeding the normal threshold), and the presence of broadband energy pulses with a bandwidth between 500 Hz and 5 kHz and a duration of less than 100 milliseconds, abnormal harmonic components and transient oscillation components are identified.The "normal threshold" for the abnormal harmonic components is set as a percentage of the harmonic current relative to the fundamental current, referring to the State Grid standard (e.g., 2.5% for the 3rd harmonic and 1.5% for the 5th harmonic). The frequency range of the transient oscillation components (500 Hz to 5 kHz) is selected based on statistical analysis of transient signals from common faults in distribution networks. These identified voltage abrupt changes, current abrupt changes, abnormal harmonic components, and transient oscillation components are established as instantaneous fault initiation features in the form of structured feature vectors to characterize the instantaneous onset of ring network faults.

[0031] This invention performs high-resolution, high-density time-frequency analysis on smart power distribution data through synchronous compression transformation, effectively capturing the initial characteristics of instantaneous faults that are difficult to identify using traditional methods. Due to its high time-frequency density, this method reduces the search space and computational complexity of subsequent feature recognition algorithms, enhances the ability to analyze rapidly changing, non-steady-state fault signals in ring networks, and provides crucial initial information for accurate fault diagnosis.

[0032] Furthermore, the process of inputting the initial fault features into a graph convolutional network considering the ring network topology constraints and mapping them to fault propagation path features includes: constructing a topology graph of the smart distribution ring network, where the nodes of the graph represent the feeder nodes and tie nodes of the smart distribution ring network, and the edges of the graph represent the connection relationships between feeders and tie lines in the smart distribution ring network; using the initial fault features as the initial input features for the corresponding nodes in the topology graph; and performing multi-layer feature aggregation and transformation on the topology graph through the graph convolutional network, so that the feature learning of each node integrates its own initial fault features to... The system collects and transforms the initial fault characteristics of adjacent nodes. During the feature aggregation and transformation process, the propagation weights of the graph convolutional network are adjusted using the topological constraints of the smart distribution ring network. The magnitude of the propagation weights is inversely proportional to the electrical impedance of the connection, giving priority to feature associations along the actual electrical connection path to reflect the propagation trend of the fault in the smart distribution ring network. Through the output layer of the graph convolutional network, the fault propagation direction and fault impact range are extracted from the learned node features and mapped to the fault propagation path features of the smart distribution ring network to detect cable and transmission line faults in the smart distribution ring network.

[0033] Specifically, firstly, an undirected graph is constructed based on a pre-established intelligent distribution ring network topology model. The nodes of this graph include all feeder terminal units, tie switches, substation outgoing lines, and other key electrical connection points. For example, node V1 represents a monitoring point on feeder A, and node V2 represents the tie switch connected to it. The edges of the graph represent the actual physical connections between nodes via feeders or tie lines; for example, edge E12 connects V1 and V2. The instantaneous fault initial feature vector extracted from the above steps is assigned as the initial input feature vector of the node corresponding to that monitoring point. Subsequently, this topology graph and the initial features of the nodes are input into a graph convolutional network with three layers. In each layer of graph convolution, the feature update of the nodes follows the standard graph convolutional network (GCN) formula: ;

[0034] in, The updated representation of the feature vector of each node in the (l+1)th layer graph. The graph adjacency matrix is ​​normalized and self-joined. The activation function for the rectified linear unit (ReLU) is... For a trainable weight matrix, This represents the node feature matrix of layer l. This process achieves the fusion of deep learning of node fault information and neighboring node information, while significantly reducing computational complexity and optimizing memory usage compared to traditional fully connected neural networks. During the propagation weight adjustment phase of the graph convolutional network, the graph adjacency matrix is ​​adjusted according to the electrical connection rules of the intelligent power distribution ring network. Adjustments were made. The graph convolutional network adopts a three-layer GCN structure with a weight matrix. elements ij Defined as ij =1 / ij ,in ij This is the electrical impedance between nodes i and j, used to enhance characteristic propagation along low-impedance paths. The electrical impedance... ij The parameters are derived from the line parameters calculated during the distribution network design phase and used as fixed topology attributes before model training. First, an original weighted adjacency matrix is ​​calculated, whose element values ​​are inversely linearly proportional to the electrical impedances of connected nodes i and j. Then, using this original weighted matrix as a basis, self-connection processing and degree matrix normalization are performed according to the requirements of standard graph convolutional networks. This normalized matrix is ​​then used as the matrix in the GCN formula for feature aggregation calculation. Specifically, the weights are inversely proportional to the electrical impedance of the connection. For example, if the impedance of connection i and j... ij If it is smaller, then its corresponding adjacency matrix ij The weights on the adjacent nodes are relatively large, and this weighted adjacency matrix is ​​used to replace the weights in the formula. This ensures that feature aggregation prioritizes the electrical propagation characteristics of the fault along the minimum impedance path. The topological constraint that "the propagation weight is inversely proportional to the electrical impedance of the connection" is mainly applied to electrical faults involving current flow and electromagnetic wave propagation, such as short-circuit faults and ground faults. For non-current or long-term developing faults such as open-circuit faults and insulation degradation, the weights of their propagation features mainly depend on the self-learning capability of the spatiotemporal correlation prediction model. In specific implementation, the system performs weighted fusion of the output of the GCN layer through fault type discrimination features: when the discrimination result is a short-circuit or ground fault, the weights are determined by the inverse impedance relationship; when the discrimination result is an open-circuit or insulation degradation fault, weights based on the historical correlation between nodes are used. After feature aggregation and transformation by the multi-layer graph convolutional network, the network's output layer decodes the learned features of each node and extracts: a list of fault propagation directions (e.g., propagation from V1 along E12 to V2), a node influence probability vector, and a fault influence range identifier (e.g., the influence range covers V2 and all associated load areas). This information is combined to form the fault propagation path characteristics (structured output set) of the intelligent distribution ring network.

[0035] This invention utilizes graph convolutional networks (GCNs) combined with the topological constraints of intelligent distribution ring networks to deeply explore the minimum impedance propagation law of faults. This method overcomes the limitations of traditional analysis in its insufficient utilization of topology. Because the GCN model is highly efficient at extracting features from sparse topologies, it reduces computational resources for model training and inference compared to traditional deep learning models, resulting in more accurate mapping of fault propagation paths and improving the ability of intelligent distribution ring networks to detect cable and transmission line faults.

[0036] Further, the process of generating the intelligent fault diagnosis evidence set includes: jointly feeding the instantaneous fault initiation features and fault propagation path features into a pre-trained spatiotemporal correlation prediction model; the spatiotemporal correlation prediction model is based on a deep learning architecture, including a recurrent neural network layer for processing time series data and a graph attention network layer for processing spatial topological relationships; extracting the temporal dynamic features of fault occurrence and development through the recurrent neural network layer; evaluating the spatial correlation strength of the fault in the intelligent distribution ring network topology through the graph attention network layer in combination with the fault propagation path features; fusing the temporal dynamic features and spatial correlation features to form a comprehensive representation of the current ring network fault state; based on the comprehensive representation, determining the current fault type through the model's output layer, the fault type discrimination features including short-circuit fault, grounding fault, open-circuit fault, and insulation degradation fault; based on the comprehensive representation, predicting the fault development trend in the future through the model's output layer, the trend prediction features including the fault range expansion trend, the fault severity increase trend, and the fault location drift trend; and combining the fault type discrimination features and the trend prediction features to form an intelligent fault diagnosis evidence set, indicating the fault status of the intelligent distribution ring network.

[0037] Specifically, the instantaneous fault initiation features extracted in the above steps (e.g., current surge amplitude, transient oscillation frequency components) and the mapped fault propagation path features (e.g., fault initiation point, propagation direction, list of affected nodes) are integrated into a joint input data set and fed into a spatiotemporal correlation prediction model pre-trained on historical fault data. This model employs a graph convolutional recurrent network architecture combined with a multi-head self-attention mechanism.

[0038] The recurrent neural network layer (e.g., a two-layer long short-term memory network (LSTM) or gated recurrent unit (GRU) with 64 hidden units) receives the time series of instantaneous fault initiation characteristics, analyzes the evolution patterns of voltage and current waveforms within 0.5 seconds before and after the fault occurs, and extracts temporal dynamic features such as fault duration and spectral drift rate. The graph attention network layer (e.g., a single-layer graph attention network (GAT) with 8 heads) receives fault propagation path characteristics. For example, by calculating the topological distance and electrical coupling strength between the fault point and other nodes in the network, it assesses the spatial influence distribution of the fault in the smart distribution ring network, forming a quantitative assessment of spatial correlation strength and influence range.

[0039] Temporal dynamic features are extracted through a recurrent neural network layer and pooled in the temporal dimension. A graph attention network layer, combined with fault propagation path features, assesses spatial correlation strength and flattens the node dimension. Subsequently, the pooled temporal feature vector and the flattened spatial feature vector are fed into a fusion layer. For example, they are concatenated and passed through a multilayer perceptron network with two hidden layers (128 and 64 neurons respectively) to form a single vector, serving as a comprehensive representation of the current fault state of the smart distribution ring network. The model's output layer includes two sub-networks: a classifier identifies the current fault type, outputting the probability values ​​for short-circuit faults, ground faults, open-circuit faults, and insulation degradation faults; the other, a regression network, predicts the number of newly affected nodes, the percentage increase in voltage drop, or the fault location center point drift distance (in meters) within the next minute. These identified fault type probability values ​​and predicted trend quantification values ​​together constitute the evidence set for smart distribution fault diagnosis.

[0040] The graph attention network layer further dynamically adjusts the attention weights between nodes based on the topology sensitivity of the smart distribution ring network. The topology sensitivity characterizes the instantaneous impact of changes in topology connections on fault propagation characteristics.

[0041] In the graph attention network layer, topology sensitivity is quantified by real-time calculation of the mutual information gain of fault features between two nodes. When a tie-line switch (e.g., L1) in the ring network changes from closed to open, the system calculates the mutual information of the impact of the initial fault features on the propagation path features before and after the topology change. If the change in mutual information exceeds a preset threshold, it indicates high sensitivity to the topology change. The graph attention network will then perform real-time compensatory increases on the weights of the affected node pairs to strengthen the model's learning of the current topology constraints. This scheme, by introducing a dynamic adjustment mechanism for topology sensitivity, effectively improves the accuracy of feature association under ring network topology reconstruction and complex high-impedance faults, and enhances fault type discrimination accuracy compared to a static weight allocation model.

[0042] This invention achieves accurate identification of fault types and effective prediction of development trends in smart distribution ring networks by deeply fusing instantaneous fault and propagation path features through a spatiotemporal correlation prediction model. This method overcomes the limitations of traditional diagnostic methods that rely solely on single features or simple rules, providing more comprehensive and forward-looking evidence for smart distribution fault diagnosis.

[0043] Furthermore, the process of determining the fault investigation interval based on the fault propagation path characteristics in the evidence set includes: issuing a fault alarm signal when the fault type discrimination characteristics in the intelligent fault diagnosis evidence set reach a preset alarm threshold; using the fault alarm signal as an explicit instruction to initiate the intelligent power distribution fault investigation process; extracting the fault propagation path characteristics contained in the intelligent fault diagnosis evidence set, wherein the fault propagation path characteristics indicate the direction and range of fault propagation along the intelligent power distribution ring network topology; and calculating and defining a physical area containing all potential fault points based on the fault propagation path characteristics, combined with the geographical information and equipment distribution of the intelligent power distribution ring network, using a real-time updated topology path analysis algorithm, wherein the area is defined as the fault investigation interval for accurate fault location determination, guiding subsequent physical sensor information collection and verification.

[0044] Specifically, the system receives a set of evidence for intelligent power distribution fault diagnosis. When the probability value of the main fault type discrimination feature in the evidence set (e.g., short-circuit fault probability) exceeds a preset alarm threshold (e.g., 0.8), and the trend prediction feature (e.g., voltage drop deepening value) reaches a trend warning threshold (e.g., 5%), a fault alarm signal is immediately issued through various means such as audible and visual alarms, SMS notifications, and pop-ups on the intelligent power distribution monitoring platform. This alarm threshold differs from the threshold in the feature extraction stage: the feature extraction threshold is used to identify raw signal anomalies from instantaneous data, while the alarm threshold is set for fault type discrimination features that incorporate spatiotemporal correlation information. It is used to filter transient signals caused by non-fault events such as switching operations, ensuring that the troubleshooting process is initiated only when the model determines that the probability of a fault exceeds a preset safety level. This fault alarm signal is recognized as a mandatory instruction to initiate a manual or automated fault troubleshooting process. Simultaneously, fault propagation path features are extracted from the evidence set; for example, it indicates that the fault starts from monitoring point M3 on feeder C, propagates along tie line L2 to feeder D, and may affect five downstream nodes. The built-in ring network geographic information is invoked to obtain the specific geographic coordinates and equipment distribution (e.g., pole and tower locations, underground cable routes) of feeder C, tie line L2, and feeder D.

[0045] Subsequently, based on the fault propagation path characteristics and geographical information, a path analysis algorithm combining real-time switch status updates to the topology (e.g., Dijkstra's algorithm) is used to calculate and generate a minimum rectangular or polygonal geographical region covering the aforementioned propagation path and all potentially affected nodes in real time. The width of this geographical region (e.g., 50 meters) is determined based on the designed corridor width of the ring network and the positioning error range of the physical sensors. This defined area is the fault investigation zone. This investigation zone is used to precisely narrow down the positioning range, and will subsequently guide physical sensing tools such as drones and handheld testing devices to conduct targeted data collection and verification within this area.

[0046] The process of calculating and defining the physical area containing all potential fault points dynamically adjusts the geographical buffer width of the area to be investigated based on the trend prediction characteristics in the evidence set, adapting to the increasing severity and expanding scope of the fault.

[0047] If the trend prediction indicates that the voltage drop will increase by more than 10% within the next minute, or that more than three new affected nodes are added, the buffer width on both sides of the area to be investigated (e.g., the initial 50 meters) will be increased by 50%, forming a dynamic buffer zone 75 meters wide. This dynamic adjustment mechanism is implemented by mapping the trend prediction characteristics to the geographical coordinate offset of the investigation area, ensuring that a larger safe investigation range is delineated in advance before the fault worsens. This solution, by combining trend prediction with dynamic buffering, effectively reduces the risk of investigation errors when facing rapidly deteriorating faults without significantly increasing the daily investigation range, and improves the responsiveness and foresight of automated investigation equipment.

[0048] This invention achieves rapid and dynamic delineation of fault areas through a troubleshooting mechanism based on intelligent power distribution diagnostic evidence. This method overcomes the problems of traditional manual judgment, such as excessively broad scope, low efficiency, and lack of foresight, providing a more precise and flexible target area for troubleshooting and effectively guiding the subsequent deployment and application of physical sensors.

[0049] Furthermore, the process of determining the physical space locking of the fault point includes: acquiring geographic information matching data within the investigation interval, the geographic information matching data being derived from the spatial location information of the intelligent power distribution line equipment; acquiring acoustic source positioning data, the acoustic source positioning data being derived from the abnormal sound acquisition and algorithm processing of the intelligent power distribution line; acquiring thermal imaging detection data, the thermal imaging detection data being derived from the temperature scan of the intelligent power distribution line and its surrounding environment; acquiring fiber optic vibration analysis data, the fiber optic vibration analysis data being derived from the vibration mode monitoring and analysis of the intelligent power distribution line; performing feature extraction and standardization processing on the acquired data to transform heterogeneous data into unified fault tendency evidence; inputting the fault tendency evidence into a multimodal fusion decision matrix, the decision matrix performing weighted fusion decision based on Bayesian inference; determining the precise location with the highest fault probability within the investigation interval based on the output of the fusion decision matrix, the precise location being used as the physical space locking result of the fault point; comparing and verifying the physical space locking result with the intelligent fault diagnosis evidence set to complete the closed-loop verification of the intelligent power distribution ring network fault.

[0050] Specifically, within the defined intelligent power distribution investigation area (e.g., a specific 500-meter section of feeder C), the following operations are performed: Using intelligent power distribution geographic information, the cable routing, tower numbers, and underground pipeline locations within this section are precisely matched with a digital map to generate geographic information matching data, such as the precise GPS coordinate sequence for each cable segment. High-sensitivity acoustic sensors (e.g., microphone arrays) are deployed along this section to collect abnormal sounds from the line (e.g., the hissing sound of partial discharge). The two-dimensional or three-dimensional coordinates of the sound source are calculated using an acoustic source localization algorithm (e.g., time-of-flight localization), generating acoustic source localization data, such as the location of the abnormal sound source, point A. A drone equipped with thermal imaging equipment is used to scan along the line to identify areas with abnormal temperatures (e.g., the temperature at a joint is 10 degrees Celsius higher than the ambient temperature), generating thermal imaging detection data, such as the abnormally high temperature area B. For high-risk sub-areas such as buried cables or critical joints in this section, fiber optic sensors, either installed or integrated, are used to monitor line vibration patterns and analyze for abnormal vibration characteristics (e.g., resonance at a specific frequency or shock waves caused by external impacts). This generates fiber optic vibration analysis data, such as the location of an abnormal vibration event, C. Based on this data (e.g., coordinates of locations A, B, and C, sound intensity, temperature difference, and vibration energy), individual features are extracted (e.g., temperature difference is converted into a fault probability score) and standardized using Min-Max methods, resulting in a unified fault propensity evidence score. .

[0051] The Min-Max-based normalization process follows the formula below to map heterogeneous data to the range [0,1]: ;

[0052] in, This represents the standardized fault propensity evidence score. This represents the characteristic value of the current measurement (e.g., temperature difference or sound intensity). and These represent the minimum and maximum values ​​of the feature type observed in historical operating data, respectively. Subsequently, this evidence is fed into a multimodal fusion decision matrix constructed based on Bayesian inference. This matrix optimizes and fuses all multimodal evidence based on the prior distribution of historical faults and the conditional probability model of sensor failure modes, thereby precisely calculating the comprehensive fault occurrence probability of each candidate point within the investigation interval. The fusion decision matrix uses Bayesian posterior probability to fuse the predicted "fault location center point drift trend" feature as one of the key pieces of evidence. When the drift trend prediction shows that the positioning uncertainty range tends to stabilize or decrease (i.e., the positioning accuracy is improving), the fusion matrix assigns a higher reliability weight to that location, thereby determining the precise location with the highest fault probability as the physical spatial locking result of the fault point. The Bayesian inference model adopts a network structure consisting of a single fault event node and multiple sensor evidence nodes. To balance the efficiency of real-time computation, the model uses conditional independence approximation during inference, but simultaneously introduces an "evidence correlation correction factor" to discount and adjust the joint probability contribution of physically highly correlated evidence (such as acoustic and thermal imaging) to avoid double counting. The network's prior probability and conditional probability table (CPT) are both established by performing frequency statistical estimation on historical fault diagnosis records. For example, when multiple sources of evidence, such as acoustic and thermal imaging, consistently point to a specific location, the probability of a fault at that location will be significantly increased by using a Bayesian posterior probability update mechanism.

[0053] The fusion decision matrix is ​​based on Bayesian inference. It dynamically adjusts the prior and conditional probabilities of the evidence according to environmental disturbance signals from the smart power distribution ring network, offsetting the impact of environmental noise on the reliability of sensor information. Specifically, environmental factors are treated as conditional variables in the Bayesian network, and their influence on the likelihood probability of each sensor is pre-defined in the CPT (Conceptual Probability Table) to naturally reflect changes in sensor reliability under environmental changes. When an environmental disturbance signal indicates the presence of a specific environmental factor in the current investigation area (e.g., rain causing humidity >90%), the conditional probability table (CPT) of the corresponding sensor evidence node is directly corrected in real time according to a pre-established lookup table, thereby correcting the reliability index of each sensor data. If the humidity is too high, the reliability probability value in the CPT corresponding to the thermal imaging evidence is correctively reduced according to the humidity influence curve, while the weight of the fiber optic vibration analysis evidence is correspondingly increased to ensure the robustness of the fusion result. This scheme, by introducing environmental context to dynamically adjust weights, significantly improves the anti-interference capability and decision robustness of the multimodal fusion decision matrix under extreme environmental conditions, improving positioning accuracy to the centimeter level.

[0054] Finally, the precise geographical location with the highest fault probability (e.g., exceeding 0.95) is output as the physical spatial location of the smart power distribution fault point. This physical location result is then compared with the fault diagnosis evidence set to verify the consistency between the diagnostic results and the actual location.

[0055] This invention achieves precise physical spatial location of fault points in intelligent power distribution ring networks through multimodal physical sensor information fusion and Bayesian inference. This method overcomes the limitations and false positive rates of single sensing methods, significantly improving the accuracy and reliability of fault location in intelligent power distribution, and supporting the refinement of intelligent power distribution operation and maintenance.

[0056] Example 2: Reference Figure 3 This embodiment provides a specific application of an intelligent fault diagnosis system for power distribution ring networks, specifically an application deployed in a regional power distribution network control center. The specific steps include: The data processing and analysis module acquires instantaneous electrical data, environmental disturbance signals, switchgear operating information, and historical operating condition data of the intelligent power distribution ring network. Based on a multi-source heterogeneous data fusion strategy of timestamp alignment and spatial topology association, it maps the data to a unified spatiotemporal dimension to generate ring network status primitives. The data processing and analysis module is used to input the ring network status primitives into a synchronous compression transformation algorithm for signal time-frequency analysis to extract instantaneous fault initiation features. It is also used to input the fault initiation features into a graph convolutional network model considering the ring network topology constraints to map them into fault propagation path features. The data processing and analysis module comprises a high-performance server cluster and a distributed data storage system. This module continuously acquires instantaneous values ​​of three-phase voltage and current, instantaneous values ​​of zero-sequence voltage and current, temperature, humidity, wind speed, and rainfall signals, the opening and closing status of circuit breaker isolators, the number of recloser operations and opening and closing times, historical operating data, and alarm event records at a granular level, from field-deployed distribution automation units (RTUs), intelligent distribution sensors, meteorological monitoring stations, and historical databases, all at a granular level. An internal time synchronization unit performs timestamp calibration on all data. The geographic information system database stores the topology model of the intelligent distribution ring network, including the precise geographic coordinates and electrical connections of all lines, nodes, and equipment. Utilizing this information, the data processing and analysis module employs a multi-source heterogeneous data fusion strategy based on a rule engine and machine learning model. For example, it uses Kalman filtering to fuse time-series data, mapping the time-aligned and spatially correlated data onto a unified three-dimensional spatiotemporal grid to generate real-time intelligent distribution ring network status primitives. The data processing and analysis module further integrates a synchronous compression transform algorithm library and a graph convolutional network model. Upon receiving the intelligent distribution ring network status primitives, the module immediately invokes a synchronous compression transform algorithm to perform time-frequency analysis on the instantaneous electrical data within the primitives. For example, it identifies and quantifies transient harmonic energy above 50Hz and extracts abrupt voltage and current changes as initial transient fault features. Simultaneously, the module inputs these initial transient fault features into a pre-trained graph convolutional network model. This model performs feature aggregation and transformation based on the intelligent distribution ring network topology. For instance, it aggregates features from adjacent nodes using a weighted average method to generate fault propagation path features that reflect the direction and extent of fault spread along the line.

[0057] The intelligent diagnosis module inputs the instantaneous fault initial characteristics and fault propagation path characteristics into the spatiotemporal correlation prediction model, integrates fault type discrimination characteristics and trend prediction characteristics, and generates an intelligent fault diagnosis evidence set. The core of the intelligent diagnostic module is a spatiotemporal correlation prediction model deployed on an independent computing unit. Upon receiving instantaneous fault initiation features and fault propagation path features from the data processing and analysis module, this model immediately initiates inference. Based on a deep learning architecture, the model includes two stacked gated recurrent unit (GRU) layers that process time-series data, extracting the evolution of fault initiation features over time to form temporal dynamic features; and two graph attention network (GAT) layers that process spatial topological relationships. The GAT uses an 8-head attention mechanism to combine fault propagation path features to evaluate the spatial correlation strength of the fault in the smart distribution ring network topology, forming spatial correlation features. The GRU substructure within the model processes the evolution of fault initiation features over time, such as analyzing the onset, development, and decay rates of fault signals; the GAT substructure handles the spatial dependence of fault propagation paths, such as evaluating the impact of faults on different feeder loads. The GAT layer dynamically adjusts the attention weights between nodes by introducing a topology sensitivity mechanism. Before each inference iteration, the topology sensitivity mechanism utilizes the instantaneous fault initiation feature data from the most recent 10 time steps, employing a simplified mutual information approximation method or a pre-trained surrogate model based on Gaussian kernel estimation, to calculate the "mutual information gain" of fault features between two adjacent nodes in real time. This topology sensitivity mechanism is a non-learning correction module in the online inference phase. It uses the real-time calculated mutual information gain as a physical constraint bias, directly superimposing it onto the pre-trained attention score calculated by the graph attention network layer. This mechanism allows the pre-trained model to adapt to instantaneous changes in the ring network's switching state in real time without retraining. If the calculated mutual information gain changes by more than a preset threshold (e.g., 5%) relative to the historical average, indicating a significant impact of the current topology connection change on fault propagation, the system will immediately compensate by increasing the attention weights of that connection before Softmax normalization at the GAT layer. The increase is proportional to the change in mutual information gain to ensure that the weight adjustment matches the actual degree of topology impact. The compensated weights are then Softmax normalized. Through the synergistic effect of these substructures, the model integrates spatiotemporal information to determine the specific type of smart distribution fault. For example, it can identify a single-phase ground fault and predict its potential expansion to a specific branch line within the next 5 minutes, generating a set of smart distribution fault diagnosis evidence that includes fault type discrimination features and trend prediction features.

[0058] The prediction of the trend prediction features is achieved through a regression network integrated into the output of the spatiotemporal correlation prediction model. This regression network consists of two fully connected hidden layers (e.g., each with 128 neurons) with the ReLU activation function. It takes the integrated representation of GRU and GAT as input and outputs quantitative prediction values. The predicted "next minute" time window is determined based on statistical analysis of the average duration of transient faults from initial occurrence to stabilization, aiming to cover the most critical development stage in the early stages of a fault. The "fault location drift trend" does not refer to the movement of the physical fault point itself, but rather to the changing trend of the "uncertainty range" or "location error" of the model's fault location result (i.e., the center point with the highest probability) within the investigation interval over a continuous time window. An increasing trend in the predicted "drift distance" indicates a decrease in location accuracy, while a decreasing trend indicates that the location accuracy is stabilizing. This regression network achieves quantitative prediction of these three trend indicators by learning the evolution of location error in historical fault data.

[0059] The troubleshooting and location module establishes a troubleshooting initiation mechanism based on the intelligent fault diagnosis evidence. It uses the intelligent fault diagnosis evidence set as the trigger mechanism for the troubleshooting process and determines the fault's investigation interval based on the fault propagation path characteristics in the evidence set. The troubleshooting and location module is used to construct a multimodal fusion decision matrix based on physical sensor information, including geographic information matching, acoustic source localization, thermal imaging detection, and fiber optic vibration analysis. It also performs closed-loop verification on the investigation interval to determine the physical spatial location of the fault point.

[0060] This module includes a decision reasoning engine and a physical sensing interface. When the confidence level of the fault type discrimination feature in the intelligent power distribution fault diagnosis evidence set generated by the intelligent diagnosis module exceeds 0.8, and the trend prediction feature indicates that the fault is worsening, the investigation and location module immediately issues an audible and visual alarm and an SMS notification to initiate the investigation process. Based on the fault propagation path characteristics in the evidence set, for example, identifying a specific section of a 10kV overhead line where the fault is propagating, and combining this with data from the intelligent power distribution geographic information system, the module dynamically calculates and defines a covered geographic rectangular area using a path search algorithm (e.g., a heuristic search algorithm), for example, an area 300 meters long and 20 meters wide, as the area to be investigated. The investigation and location module further coordinates external physical sensing devices. Within the area to be investigated, a drone carrying a thermal imager, a ground acoustic sensor array, and fiber optic vibration monitoring equipment along the line is automatically activated or assigned to collect data. The module processes, extracts features from, and standardizes these physical sensor data (e.g., coordinates of abnormal temperature points, locations of discharge sound sources, and abnormal vibration zones) in real time, generating unified fault tendency evidence. This evidence is input into a Bayesian network-based multimodal fusion decision matrix. The decision matrix performs weighted fusion inference on all evidence, ultimately outputting the precise geographical coordinates of the highest fault probability within the investigated interval—for example, a coordinate point accurate to the centimeter level—as the physical spatial location of the smart distribution fault point. The module compares this location result with the diagnostic evidence set of the intelligent diagnostic module for final closed-loop verification of the smart distribution fault.

[0061] This system, through modular design, tightly integrates multi-source data fusion, deep learning diagnostics, and multimodal physical sensing positioning. It overcomes the limitations of existing troubleshooting systems, such as fragmented information, inaccurate diagnosis, and vague positioning, providing comprehensive, intelligent, and high-precision intelligent power distribution ring network fault diagnosis capabilities.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent fault diagnosis in a power distribution ring network, characterized in that, include: Acquire instantaneous electrical data, environmental disturbance signals, switchgear operating information, and historical operating condition data of the intelligent power distribution ring network; A multi-source heterogeneous data fusion strategy based on timestamp alignment and spatial topology association is used to map the data to a unified spatiotemporal dimension to generate ring network situational primitives. The signal time-frequency analysis of the ring network situation primitive input to the synchronous compression transform is used to extract the instantaneous fault initial characteristics; The initial fault characteristics are input into a graph convolutional network that considers the constraints of the ring network topology, and mapped to fault propagation path characteristics. The initial fault characteristics and fault propagation path characteristics are input into the spatiotemporal correlation prediction model, and fault type discrimination characteristics and trend prediction characteristics are integrated to generate an intelligent fault diagnosis evidence set. Establish a troubleshooting initiation mechanism based on intelligent diagnostic evidence, using the intelligent fault diagnosis evidence set as the trigger mechanism for the troubleshooting process, and determine the fault-to-investigate interval based on the fault propagation path characteristics in the evidence set. A multimodal fusion decision matrix based on physical sensing information is constructed, which includes geographic information matching, acoustic source localization, thermal imaging detection, and fiber optic vibration analysis. Closed-loop verification is performed on the area to be investigated to determine the physical spatial location of the fault point.

2. The intelligent fault diagnosis method for distribution ring networks according to claim 1, characterized in that, The process of generating the ring network status primitive includes: acquiring instantaneous electrical data of the intelligent distribution ring network, including instantaneous values ​​of three-phase voltage, three-phase current, zero-sequence voltage, and zero-sequence current; acquiring environmental disturbance signals, including temperature, humidity, wind speed, and rainfall signals around the line; acquiring operating information of intelligent distribution switchgear, including the opening and closing status of circuit breakers, the opening and closing status of isolating switches, the number of recloser operations, and the opening and closing times; and acquiring historical operating condition data of the intelligent distribution system, including ring network operating voltage records and ring network operating data over a past period. The system records flow and fault alarm events; it timestamps and calibrates the instantaneous electrical data, environmental disturbance signals, switchgear operating information, and historical operating condition data of the intelligent distribution ring network to achieve time synchronization between various intelligent distribution data sources; it establishes an intelligent distribution ring network topology model based on the actual physical connection relationship of the intelligent distribution ring network, and uses the topology model to accurately map and logically associate the spatial physical locations of each intelligent distribution data source; it integrates the various types of time-synchronized intelligent distribution data with the spatial physical location mapping and logical association results, and generates intelligent distribution ring network status primitives in a unified spatiotemporal coordinate system through a multi-feature-level fusion method.

3. The intelligent fault diagnosis method for distribution ring networks according to claim 1, characterized in that, The process of inputting the ring network situation primitive into the synchronous compression transform signal time-frequency analysis to extract the initial characteristics of instantaneous faults includes: obtaining and separating the instantaneous three-phase voltage and three-phase current data of different monitoring points on each feeder and tie line from the ring network situation primitive; the ring network situation primitive, as a real-time data structure, includes references and simplified storage of associated original high-frequency data during generation; inputting the separated original high-frequency instantaneous data into the synchronous compression transform algorithm; performing time-frequency analysis on the instantaneous voltage and current data through the synchronous compression transform algorithm to obtain the energy distribution at different times and frequencies; identifying and extracting voltage mutation points, current mutation points, harmonic component anomalies, and transient oscillation components from the time-frequency analysis results based on preset fault feature thresholds; the fault feature thresholds are determined based on statistical analysis of historical operating data of the smart distribution ring network and electromagnetic transient simulation results; and using the voltage mutation points, current mutation points, harmonic component anomalies, and transient oscillation components as instantaneous fault initial characteristics in the form of structured feature vectors to characterize the initial occurrence state of faults in the smart distribution ring network.

4. The intelligent fault diagnosis method for distribution ring networks according to claim 1, characterized in that, The process of inputting initial fault features into a graph convolutional network considering the topological constraints of the ring network and mapping them to fault propagation path features includes: constructing a topological graph of the smart distribution ring network, where the nodes of the graph represent the feeder nodes and tie nodes of the smart distribution ring network, and the edges of the graph represent the connection relationships between feeders and tie lines in the smart distribution ring network; using the initial fault features as the initial input features of the corresponding nodes in the topological graph; performing multi-layer feature aggregation and transformation on the topological graph through the graph convolutional network, so that the feature learning of each node integrates its own initial fault features and the initial fault features of adjacent nodes; during the feature aggregation and transformation process, using the topological constraints of the smart distribution ring network, adjusting the propagation weights of the graph convolutional network, wherein the propagation weights are inversely proportional to the electrical impedance of the connection, giving priority to feature associations along the actual electrical connection path, reflecting the propagation trend of the fault in the smart distribution ring network; and extracting the fault propagation direction and fault impact range from the learned node features through the output layer of the graph convolutional network, mapping them to fault propagation path features of the smart distribution ring network, and detecting cable and transmission line faults in the smart distribution ring network.

5. The intelligent fault diagnosis method for distribution ring networks according to claim 1, characterized in that, The process of generating an intelligent fault diagnosis evidence set includes: jointly feeding the instantaneous fault initiation features and fault propagation path features into a pre-trained spatiotemporal correlation prediction model; the spatiotemporal correlation prediction model is based on a deep learning architecture, including a recurrent neural network layer for processing time-series data and a graph attention network layer for processing spatial topological relationships; extracting the temporal dynamic features of fault occurrence and development through the recurrent neural network layer; evaluating the spatial correlation strength of the fault in the intelligent distribution ring network topology through the graph attention network layer, combined with the fault propagation path features; fusing the temporal dynamic features and spatial correlation features to form a comprehensive representation of the current ring network fault state; based on the comprehensive representation, determining the current fault type through the model's output layer, the fault type discrimination features including short-circuit fault, ground fault, open-circuit fault, and insulation degradation fault; based on the comprehensive representation, predicting the fault development trend over a future period through the model's output layer, the trend prediction features including fault range expansion trend, fault severity increase trend, and fault location drift trend; and combining the fault type discrimination features and trend prediction features to form an intelligent fault diagnosis evidence set, indicating the fault status of the intelligent distribution ring network.

6. The intelligent fault diagnosis method for distribution ring networks according to claim 1, characterized in that, The process of determining the fault investigation interval based on the fault propagation path characteristics in the evidence set includes: issuing a fault alarm signal when the fault type discrimination characteristics in the intelligent fault diagnosis evidence set reach a preset alarm threshold; using the fault alarm signal as an explicit instruction to initiate the intelligent power distribution fault investigation process; extracting the fault propagation path characteristics contained in the intelligent fault diagnosis evidence set, wherein the fault propagation path characteristics indicate the direction and range of fault propagation along the intelligent power distribution ring network topology; and calculating and defining a physical area containing all potential fault points based on the fault propagation path characteristics, combined with the geographical information and equipment distribution of the intelligent power distribution ring network, using a real-time updated topology path analysis algorithm, wherein the area is defined as the fault investigation interval for accurate fault location determination, guiding subsequent physical sensor information collection and verification.

7. The intelligent fault diagnosis method for distribution ring networks according to claim 1, characterized in that, The process of determining the physical space location of the fault point includes: acquiring geographic information matching data within the investigation area, the geographic information matching data being derived from the spatial location information of the intelligent power distribution line equipment; acquiring acoustic source location data, the acoustic source location data being derived from the abnormal sound acquisition and algorithm processing of the intelligent power distribution line; acquiring thermal imaging detection data, the thermal imaging detection data being derived from the temperature scan of the intelligent power distribution line and its surrounding environment; acquiring fiber optic vibration analysis data, the fiber optic vibration analysis data being derived from the vibration mode monitoring and analysis of the intelligent power distribution line; performing feature extraction and standardization processing on the acquired data to transform heterogeneous data into unified fault tendency evidence; inputting the fault tendency evidence into a multimodal fusion decision matrix, the decision matrix performing weighted fusion decision based on Bayesian inference; determining the precise location with the highest fault probability within the investigation area based on the output of the fusion decision matrix, the precise location being used as the physical space location of the fault point; and comparing and verifying the physical space location result with the intelligent fault diagnosis evidence set to complete the closed-loop verification of the intelligent power distribution ring network fault.

8. A power distribution ring network fault diagnosis system, characterized in that, include: The data processing and analysis module acquires instantaneous electrical data, environmental disturbance signals, switchgear operating information, and historical operating condition data of the intelligent power distribution ring network. Based on a multi-source heterogeneous data fusion strategy of timestamp alignment and spatial topology association, it maps the data to a unified spatiotemporal dimension to generate ring network status primitives. The data processing and analysis module is used to input the ring network status primitives into a synchronous compression transformation algorithm for signal time-frequency analysis to extract instantaneous fault initiation features. It is also used to input the fault initiation features into a graph convolutional network model considering the ring network topology constraints to map them into fault propagation path features. The intelligent diagnosis module inputs the instantaneous fault initial characteristics and fault propagation path characteristics into the spatiotemporal correlation prediction model, integrates fault type discrimination characteristics and trend prediction characteristics, and generates an intelligent fault diagnosis evidence set. The troubleshooting and location module establishes a troubleshooting initiation mechanism based on the intelligent fault diagnosis evidence. It uses the intelligent fault diagnosis evidence set as the trigger mechanism for the troubleshooting process and determines the fault's investigation interval based on the fault propagation path characteristics in the evidence set. The troubleshooting and location module is used to construct a multimodal fusion decision matrix based on physical sensor information, including geographic information matching, acoustic source localization, thermal imaging detection, and fiber optic vibration analysis. It also performs closed-loop verification on the investigation interval to determine the physical spatial location of the fault point.

Images