Power distribution network fault identification drive accurate isolation method based on edge computing architecture

Abstract

The invention discloses a power distribution network fault identification driving accurate isolation method based on an edge computing architecture, and relates to the technical field of power system automation, and the method comprises the following steps: S001, collecting voltage and current signals at each edge computing node of a power distribution network, analyzing the signal fluctuation amplitude and propagation time delay of different nodes based on the same fault event, and obtaining a fault event; and calculating a fault identification judgment difference degree between the nodes. According to the method, the identification credibility is quantified and abnormal node judgment is corrected through fault identification judgment difference analysis and topological constraint modeling; self-adaptive identification logic optimization is realized in combination with historical and real-time data; control priority and pre-simulation analysis are introduced, so that the feasibility and safety of the isolation action are improved; and finally, a closed-loop mechanism integrating recognition, simulation and feedback is constructed, and the fault response accuracy, coordination and operation stability of the power distribution network under a distributed architecture are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, and more specifically to a method for precise isolation driven by fault identification in distribution networks based on an edge computing architecture. Background Technology

[0002] "Fault Identification-Driven Precise Isolation in Distribution Networks Based on Edge Computing Architecture" refers to introducing edge computing technology into the distribution network system, deploying fault identification and isolation decision-making capabilities in edge computing nodes (such as distribution terminals, intelligent ring main units, or pole-mounted switch control units) close to the field. This enables local real-time monitoring of the power grid's operating status, rapid fault identification, and precise control response. When faults such as short circuits, grounding, or equipment anomalies occur, edge nodes can instantly determine the fault type and location based on locally collected multi-source data such as current, voltage, and switch status using embedded intelligent algorithms. They then automatically execute isolation operations according to preset strategies, such as precisely cutting off the faulty segment, locally disconnecting switches, or linking adjacent equipment to reconstruct the power supply path. This method not only avoids the response lag problems caused by communication delays or data bottlenecks in traditional centralized control but also significantly improves the speed, accuracy, and local autonomy of fault handling, effectively ensuring the operational stability and power supply continuity of the distribution network.

[0003] The existing technology has the following shortcomings: In existing distributed power grid architectures, edge computing nodes are typically deployed at key switching units, feeder terminals, or branch nodes to independently perform fault detection and isolation decisions. However, due to the lack of a unified global fault perception model and decision coordination mechanism among edge nodes, when a fault occurs in a certain area, different nodes may arrive at conflicting identification results based on their own local perception data (such as voltage dips, current surges, or zero-sequence component changes). For example, a node upstream of the fault point may determine that the line it is connected to is the fault source and prepare to perform isolation actions, while a downstream node, due to perception lag or insufficient detection sensitivity, may mistakenly believe that the line is operating normally and fail to perform the corresponding response, or conversely identify a non-faulty area as an abnormal area. Such judgment errors propagate among multiple nodes, easily causing interference with isolation control commands and disordered execution sequences. This may not only lead to the accidental disconnection of power supply to non-faulty areas, but also the risk of faulty areas not being isolated in a timely and effective manner. Especially in ring network or distributed power source access scenarios, such conflicting fault response behaviors will cause serious consequences such as power supply topology instability, current feedback, and voltage imbalance, severely affecting the reliability and security of the system power supply.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a precise isolation method for distribution network fault identification driven by an edge computing architecture. This method quantifies the reliability of fault identification and corrects abnormal node judgments through fault identification judgment difference analysis and topology constraint modeling; it combines historical and real-time data to achieve adaptive identification logic optimization; it introduces control priority and pre-simulation analysis to improve the feasibility and safety of isolation actions; and finally, it constructs a closed-loop mechanism integrating identification, simulation, and feedback, significantly improving the accuracy, coordination, and operational stability of fault response in distribution networks under a distributed architecture, thereby solving the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for accurate isolation driven by fault identification in power distribution networks based on edge computing architecture, comprising the following steps: S001: Collect voltage and current signals at each edge computing node of the distribution network, analyze the signal fluctuation amplitude and propagation delay of different nodes based on the same fault event, and calculate the difference in fault identification and judgment between nodes. S002, based on the difference in fault identification and judgment between nodes, combined with the distribution network topology, a fault identification consistency assessment model is constructed according to the difference in fault identification and judgment. The probability analysis of the fault type and location identified by each node is performed, and the identification reliability level index is output. S003, for edge computing nodes whose trust level index is lower than the set threshold, the fault identification logic is reconstructed based on cross-validation of historical operating conditions and real-time data, and the fault identification results of the edge computing nodes are adaptively corrected. S004, Based on the adaptively corrected fault identification results, calculate the control priority of each edge computing node in the fault isolation operation; S005, based on the control priority of each edge computing node, perform isolation operation pre-simulation analysis to evaluate the operation path and risk under topology and power flow conditions; S006, based on the dynamic pre-simulation analysis results of isolation operation, monitor the isolation execution status, extract response deviations and use them to correct the identification model and control strategy, and construct a closed-loop control mechanism of identification-simulation-feedback.

[0007] Preferably, step S001 includes: Three-phase voltage and three-phase current signals are collected at each edge computing node. 20,000 sample points are recorded per second using a sampling frequency of 20kHz. DC bias removal and normalization are performed on each group of signals. A mother wavelet function with abrupt change enhancement characteristics is used to perform continuous wavelet transform, calculate wavelet modulus and extract local maximum modulus points at each scale, construct modulus extreme value paths, and identify the abrupt change point time when each node first senses fault disturbance; Based on the sensing time, voltage and current characteristic parameters and network topology connections uploaded by edge nodes, a judgment difference matrix is ​​constructed and the fault identification judgment difference between nodes is quantified.

[0008] Preferably, step S002 includes: After obtaining the fault identification and judgment difference matrix between edge computing nodes, a directed graph structure containing node connection direction, feeder length, conductor type and impedance parameters is constructed by combining the physical topology connection relationship of the distribution network. Based on the fault type, fault location and perception time identified by each edge computing node, the similarity of type, proximity of location and perception time order are compared with the identification results of adjacent nodes to generate a compatibility score vector. The compatibility score vector is input into the multi-factor Bayesian joint probability model. By combining the historical identification accuracy of nodes with prior knowledge of fault propagation, the probability value of each node identification result being supported as a real fault judgment is inferred. Based on the joint probability value obtained from reasoning, the difference in node judgment, the topology depth, the degree of response time lag, and the historical recognition accuracy, a recognition credibility level scoring function is constructed, and the recognition credibility level index is output.

[0009] Preferably, step S003 includes: For edge computing nodes whose confidence level index is below the set threshold, their historical operating condition data is called up and the three-phase voltage signal and three-phase current signal in the current fault event are matched with the waveform using a dynamic time warping algorithm to filter similar historical samples. Extract multidimensional feature parameters of the current waveform and historical samples and compare them using the K-nearest neighbor algorithm to identify the deviation between the judgment result and the historical label. Adjust the feature weight coefficients and recognition rule base in the recognition logic based on historical sample label information, and reconstruct the recognition logic; The corrected recognition logic is reapplied to the original sampled data, the corrected result is output and the consistency of the correction is evaluated. If the consistency is improved, the corrected logic is marked as a learnable template.

[0010] Preferably, step S004 includes: Based on the adaptively corrected fault identification results, the identification confidence level value, fault perception time, topology location index and historical isolation execution performance index of each edge computing node are extracted. The extracted indicators are normalized and a fixed-weighted model is used to calculate the global decision weight value. A control priority list is generated from high to low based on the global decision weight value of each node, and a local response waiting time window is set; When a high-priority node completes the isolation operation and broadcasts the execution status, the low-priority node cancels the action; otherwise, the next-level node takes over the task and completes the fault isolation.

[0011] Preferably, step S005 includes: Based on control priority, extract the isolation operation information and identification results of the current edge computing node; By calling the distribution network topology diagram and power flow boundary conditions, an isolation operation impact path diagram is constructed; Perform power flow recalculation and analysis of distribution network isolation operations to assess voltage fluctuations, current changes, and topology breakage risks. The feasibility and risk level of isolation operations are determined based on the power flow simulation results, and an isolation action simulation analysis report is generated for reference by edge nodes and control centers.

[0012] Preferably, step S006 includes: After the isolation operation is completed at the edge computing node, the isolation process status monitoring is started, and the voltage, current and power flow direction change data before and after the operation are collected and transmitted to the simulation verification platform. The dynamic pre-simulation analysis results completed before the isolation operation are called up and compared with the actual monitoring data point by point to extract multiple deviation indicators; Based on various deviation indicators, the scoring weights of the fault identification consistency evaluation model and the feature extraction rules of the fault identification logic are jointly modified. The revised model and logic are distributed to each edge computing node and the simulation predictions, actual responses, deviation results and corrections are archived to form a closed-loop control mechanism of identification, simulation and feedback.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention calculates the difference in fault identification judgments between edge nodes, constructs a consistency evaluation model based on the topology, and quantifies the reliability of the identification results, effectively identifying judgment deviations and potentially misidentified nodes. Simultaneously, cross-validation of historical data and real-time information enables adaptive correction of the identification logic, ensuring that the judgment of each node more closely matches actual operating conditions. Furthermore, the introduction of a control priority mechanism and dynamic pre-simulation analysis helps to comprehensively assess the feasibility and risks of isolation before action execution, avoiding misoperation during power outages. Finally, through monitoring and feedback correction of actual response deviations, the identification model and control strategy are continuously optimized, achieving a self-evolving closed-loop control integrating identification, simulation, and feedback. This scheme overcomes the conflicts and miscontrol problems caused by isolated judgments of traditional edge nodes, effectively improving the stability, accuracy, and system-level coordination of fault response, ensuring the safe and reliable operation of the power distribution system in a distributed architecture. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of the method for precise isolation driven by fault identification in power distribution networks based on edge computing architecture, as described in this invention. Detailed Implementation

[0016] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0017] This invention provides, for example Figure 1 The edge computing-based fault identification-driven precise isolation method for power distribution networks, as shown, includes the following steps: S001, collect voltage and current signals at each edge computing node of the distribution network, and perform fluctuation amplitude analysis and propagation delay analysis on the voltage and current signals of different edge computing nodes under the same fault event, and the difference in fault identification judgment between computing nodes. To quantify the differences in fault identification results among multiple edge computing nodes in a distribution network under the same fault event, and to support subsequent consistency modeling and result reliability assessment, the following specific steps are implemented: In each edge computing node of the distribution network, a three-phase voltage sensor and a three-phase current transformer are installed and connected to the monitored feeder circuits. All edge computing nodes are uniformly configured with a high-precision time synchronization unit with microsecond-level time synchronization capability, and a dual-redundant time synchronization method using BeiDou or GPS is adopted to achieve millisecond-level clock consistency between different nodes. In each edge computing node, a high-frequency analog sampling module is used to collect three-phase voltage and three-phase current signals. The sampling frequency is set to 20kHz, that is, 20,000 sample points are collected per second to fully capture the electrical waveform changes within 0.5 seconds before and after the fault occurs, specifically including the steady-state waveform before the fault trigger, the jump waveform at the moment of fault triggering, and the continuous disturbance waveform during the fault maintenance period. Each set of sampled data is accompanied by a local sampling timestamp and stored in the node's internal cache. The sampling process covers every instantaneous voltage and current value of phases A, B, and C.

[0018] In the local computing unit of each edge computing node, mathematical feature extraction and disturbance initiation point identification are performed sequentially on the acquired three-phase voltage and three-phase current signals. For the current signal, the root mean square value mutation rate within a specified time window, the maximum descent slope of the current waveform within 25 milliseconds, and the maximum amplitude difference after the current mutations of phases A, B, and C are calculated. For the voltage signal, the drop rate of the three-phase voltage amplitude, the rise time interval of the negative sequence voltage component, and the duration of the jump from the stable value to the fault peak of the zero sequence voltage are extracted. Simultaneously, continuous wavelet transform combined with the maximum modulus extreme value pursuit algorithm is applied to identify the first significant mutation point in each set of signal curves. The time of this mutation point is defined as the first perception time of the fault by the node. All the extracted feature values ​​and perception times are uniformly structured and packaged into feature vector data packets, and uploaded to the edge collaborative analysis node in real time through a secure data channel.

[0019] The purpose of sequentially extracting mathematical features and identifying disturbance initiation points from the acquired three-phase voltage and current signals is to accurately characterize the dynamic response features of electrical fault events within the sensing range of each edge node, and to determine the first occurrence time of the fault disturbance with high precision, thereby enabling a quantitative comparison of response behaviors between nodes. Specifically, mathematical feature extraction refers to performing numerical calculations and statistical processing on the signal waveform in the time and frequency domains, including calculating the amplitude drop rate and phase angle change amplitude of the three-phase voltage, the peak change amount, rising or falling slope, the rate of change of zero-sequence and negative-sequence components, and the asymmetry factor of the waveform envelope, etc., to characterize the intensity, morphology, and phase disturbance characteristics of the fault signal; disturbance initiation point identification, through wavelet transform or abrupt change point detection algorithms, extracts the occurrence time of the first electrical disturbance from high-frequency fluctuations, i.e., the time stamp of the edge node sensing the fault propagation. The combination of these two elements not only constitutes a complete information set for the local perception of fault events, but also provides high-resolution, structured data for subsequent fault identification consistency analysis, node credibility assessment, and isolation action decision-making, thereby significantly improving the accuracy of response to fault behavior and the reliability of judgment.

[0020] The continuous wavelet transform combined with the maximum modulus extreme value tracking algorithm is a high-precision time-frequency analysis method for detecting abrupt changes in electrical signals. In this step, it is used to identify the time point at which the first abnormal disturbance appears in the voltage or current signal collected by the edge computing node.

[0021] The specific implementation steps are as follows: The three-phase voltage or current signals acquired by the edge nodes are de-DC biased and normalized to obtain the purified signal sequence. Select a mother wavelet function with abrupt change enhancement characteristics (such as the Mexican Hat wavelet), perform continuous wavelet transform on the signal sequence, and obtain the wavelet coefficient matrix of the signal at different scales; The modulus of the wavelet coefficients at each scale is calculated, and the local maximum modulus points are extracted within each scale to form a set of modulus extrema points; By analyzing the location continuity across scales, we can track the paths of adjacent modulus extreme points at multiple scales, construct modulus extreme paths, and evaluate their continuous scale range and intensity variation trends. The starting time of the extreme modulus path that appears simultaneously at multiple scales and has the highest modulus intensity is selected as the significant abrupt change point when the edge node first senses the fault disturbance. This time point is used to quantify the initial differences in the responses of multiple edge nodes to the same fault event, and is a key basis for subsequent calculation of the difference degree. It has high sensitivity, strong noise resistance and cross-scale feature extraction capabilities, and is suitable for fault identification scenarios in power distribution networks where electrical signals are complex and variable and disturbances are frequent.

[0022] In the edge collaborative analysis nodes, based on the preset network topology data in the distribution network SCADA system, the physical connection location of each edge computing node in the feeder network is obtained, and a directed connection graph is constructed in the form of edge-nodes. For each pair of edge computing nodes with feeder connections, parameters such as sensing time tags, current surge slope, negative sequence voltage occurrence delay, and zero sequence current change rate are extracted from the data packets uploaded by the nodes. The differences in the response of each type of electrical characteristic are calculated for each node pair. For example, in terms of propagation delay, the result of subtracting the fault sensing time of the upstream node from the fault sensing time of the downstream node is calculated to represent the time difference of fault wave propagation; in terms of current characteristics, the absolute difference of the maximum amplitude of the A-phase current surge between the two nodes is calculated; in terms of voltage characteristics, the difference in the formation time of the negative sequence voltage between the two nodes is calculated. All characteristic difference values ​​are weighted and summed according to the preset characteristic importance weights to form a unified judgment difference index, which is used to quantitatively describe the consistency of the node's response to the same fault event.

[0023] For all connected node pairs in the edge computing system, their difference indices are integrated into a matrix structure to construct a fault identification difference matrix. The horizontal and vertical axes of this matrix correspond to the edge computing nodes currently involved in the fault identification analysis, and each cell in the matrix represents the difference indices between that node pair. Based on this difference matrix, the combination of edge computing nodes with the worst response consistency in this fault can be further analyzed and identified, including nodes with severely delayed response times, significantly low feature change amplitudes, or insufficient sensitivity to disturbances. This provides precise input for the next step of constructing the identification reliability level.

[0024] Voltage and current signals are collected at various edge computing nodes in the distribution network. Fluctuation amplitude and propagation delay analyses are performed on the voltage and current signals of different edge computing nodes under the same fault event. The difference in fault identification judgment between computing nodes is established to create a technical mechanism for quantifying the differences in the responses of different nodes to fault events, providing a foundation for achieving consistent multi-node collaborative identification. Because each edge computing node is located in a different physical structure, the voltage or current disturbances it senses exhibit differences in propagation delay, waveform amplitude attenuation, or local feature distortion. Therefore, the identification results of all nodes cannot be simply considered equally reliable. This step uses electrical characteristics such as current surge slope, voltage drop rate, phase angle offset amplitude, and the rate of change of zero-sequence and negative-sequence components, combined with the time stamp of each node's first fault perception, to form a multi-dimensional quantitative index to describe the behavioral differences of each node in fault identification. This difference not only reveals the consistency level of node responses but can also be used to identify nodes with lagging responses or abnormal judgments, providing a quantitative basis for subsequent reliability level assessment, adaptive identification correction, and control priority ranking, thereby improving the overall accuracy, coordination, and execution stability of fault handling in the edge computing environment.

[0025] S002, based on the difference in fault identification judgment between nodes, combined with the spatial topology of the distribution network, constructs a fault identification consistency evaluation model that includes line connection relationships, performs probabilistic reasoning analysis on the fault type and fault location identified by each edge computing node, and outputs the identification credibility level index. To achieve consistent analysis of fault identification results for multiple edge computing nodes under the same fault event, and to avoid erroneous isolation or out-of-order response caused by conflicting judgments between nodes, a fault identification consistency assessment model is constructed based on the difference in fault identification judgments between nodes and in combination with the spatial topology of the distribution network. This is implemented through the following steps: After obtaining the fault identification and judgment difference matrix between edge computing nodes, the difference values ​​between any two edge computing nodes with physical connections are normalized to obtain the relative response deviation score between each node pair under the current fault event. To construct an accurate spatial topology, geographic information data, cable wiring diagrams, and feeder segment wiring logic from the distribution network SCADA system are retrieved to build a directed graph model. Each node represents an edge computing node installed on a feeder terminal or pole-mounted switch control unit, and directed edges represent the physical cable connection direction. Each connection edge is bound to the actual feeder segment length (in meters), conductor type (e.g., YJV-22, LGJ-95), impedance parameters (in ohms / km), and phase spacing information. The geographical coordinates of the node location, feeder segment number, and power supply direction are also recorded to establish a spatial topology structure that accurately reflects the power transmission path.

[0026] The specific fault type (e.g., phase B short circuit, three-phase short circuit, phase C open circuit) and the identified fault location (in units of "45 meters downstream of this node") of each edge computing node under the current fault event are used as the criteria for verification. Based on the aforementioned topology, the direct upstream and downstream adjacent nodes of this node are identified, and the fault identification results of these adjacent nodes are extracted, including their identified fault type, initial sensing time (in milliseconds), and their estimated fault location (e.g., "30 meters upstream of this node"). By comparing the consistency of judgments among these adjacent nodes, specifically including fault type similarity (1 for complete consistency, 0.5 for one phase difference, and 0 for complete inconsistency), location proximity (distance difference less than 10 meters for high confidence), and the relationship between sensing times (positive weight for conforming to the physical propagation order), a judgment compatibility scoring vector is constructed, with nodes as the unit.

[0027] Based on the constructed spatial topology, a multi-factor Bayesian joint probability model is built using the judgment compatibility score vector of each edge computing node as input to analyze the recognition credibility of each node in the entire network. The prior part of this model is constructed based on historical fault data from the distribution network operation, including the spatial propagation patterns of different types of faults, typical paths of voltage and current disturbances, and propagation delay distributions. The conditional probability part is dynamically updated based on the judgment difference value of each node in the current event, the consistency score with the type and location of neighboring nodes, and the degree of matching with the spatial propagation model. By inputting the recognition data of each node into this model, the probability value that the recognition result of each node is supported as a "true fault judgment" under the current network structure and data observation conditions can be inferred.

[0028] The multi-factor Bayesian joint probability model is a probabilistic inference method based on Bayesian theory, used to quantitatively estimate the credibility of an event in the presence of multiple uncertainties and conditional variables. In this step, the model integrates multiple dimensions of conditional information, including the identification results of multiple edge computing nodes under the same fault event, the degree of difference in fault judgments between nodes, the positional relationship of nodes in the distribution network topology, the temporal order of fault responses, and historical identification accuracy, to calculate the probability value of each node's current identification result as the true fault judgment. Specifically, the model uses historical distribution network fault propagation data as prior knowledge to construct the spatial-temporal dependency relationship of fault response behavior between nodes, and substitutes the currently observed judgment differences, response delays, and positional consistency as conditional variables. Through Bayesian inference, the posterior probability is obtained, i.e., "the probability that the node's identification result is supported by other nodes." This probability value not only reflects the credibility of a single node's judgment in the entire network but also provides a core basis for subsequent identification credibility level scoring and fault control weight allocation, thereby improving the overall accuracy, consistency, and reliability of identification judgments and control decisions.

[0029] Based on the inference results of this probabilistic model, and combining the edge computing node's own judgment difference, topology depth, response time lag, and historical recognition accuracy, a recognition confidence level scoring function is constructed, with the score value set between 0 and 1. This function employs a fixed-weighted linear combination structure, for example, with the joint probability value accounting for 40%, judgment difference for 25%, topology depth (e.g., whether it is a central node) for 15%, response time for 10%, and historical recognition accuracy for 10%. The output of the scoring function is the recognition confidence level index for each edge computing node in the current fault event. In the scoring results, nodes with a confidence level higher than 0.85 are considered to have high confidence and are prioritized for subsequent isolation control strategy formulation; nodes with a confidence level lower than 0.5 will enter the recognition correction process, and their judgment results will not be adopted temporarily.

[0030] The reliability rating function is a scoring mechanism used to comprehensively evaluate the reliability of edge computing nodes' judgments during fault identification. Its core function is to transform multiple factors affecting identification reliability into a unified numerical score to determine whether each node's identification results are sufficiently reliable and can serve as a basis for subsequent isolation and control decisions. In this step, the scoring function is constructed using a multi-factor fusion approach. First, the identification support probability inferred from a Bayesian joint probability model is used as the primary evaluation criterion. This probability reflects whether the node's judgment is supported by the data behavior of surrounding nodes. Second, consistency scores between the node and its neighboring nodes in terms of fault type, fault location, and perception time are introduced to measure whether the node's judgment is consistent with the judgments of its topological neighborhood. Third, the node's location within the distribution network topology is considered, such as whether it is on a main line, its distance from the power source, or whether it is an important branch point; these factors affect the completeness of its information collection and the importance of its judgment. Simultaneously, the node's response time to the current fault is compared with other nodes to determine if there is a significant response lag. Finally, the accuracy of the node's identification results in historical fault events is introduced as a long-term performance criterion. The scoring function integrates all the above factors with predefined weights and outputs a value between 0 and 1, representing the reliability of the node's identification result. A higher score indicates that the node's identification result in the current fault event is more trustworthy and can be prioritized for isolation decisions; nodes with lower scores may have judgment biases and need to enter an adaptive correction process or have their priority in the control system reduced. This function plays a crucial role in the entire identification consistency assessment model, enabling judgment stratification, filtering effective judgment results, and improving decision accuracy.

[0031] Based on the differences in fault identification judgments among nodes and combined with the spatial topology of the distribution network, a fault identification consistency evaluation model incorporating line connection relationships is constructed. This model performs probabilistic reasoning analysis on the fault type and location identified by each edge computing node and outputs an identification reliability index. Its purpose is to evaluate the judgment behavior of multiple edge computing nodes under the same fault event from a global perspective, quantifying the reliability of each node's judgment result and resolving identification conflicts caused by the localization of information sources and inconsistent judgment standards among nodes. In an edge architecture, different nodes may give different fault type and location judgments based on their own collected voltage and current disturbance information. Without a unified evaluation mechanism, this will lead to inconsistent fault identification results, out-of-order isolation actions, and even mutual interference. This step establishes a logical connection between the topology and judgment behavior, incorporating information such as node physical connection relationships, electrical path direction, location proximity, and response time into the consistency model. Combined with the differences in fault identification judgments among nodes, a probabilistic reasoning method is used to evaluate the support of each node's judgment result. By comparing its judgment with that of neighboring nodes and whether it conforms to topological logic and electrical propagation laws, an identification reliability index is output, assigning a clear reliability weight to the identification result of each node. The credibility level index can not only be used to eliminate erroneous judgments with large deviations and improve the stability and robustness of identification, but also provide key references for subsequent isolation control priority ranking, adaptive identification correction and simulation decision-making, thereby building a trust mechanism for multi-node collaborative identification and significantly improving the identification accuracy and decision consistency of the distribution network under complex fault conditions.

[0032] S003, for edge computing nodes whose trust level index is lower than the set threshold, call historical operating condition data and combine it with the currently collected real-time data, reconstruct the fault identification logic through cross-validation, and adaptively correct the fault identification result of the edge computing node. To improve the accuracy of low-reliability node identification results and ensure the consistency and reliability of the judgment results output by each edge computing node during the distribution network fault identification process under the global coordination mechanism, for edge computing nodes with an identification reliability level index below a preset threshold (e.g., 0.5), cross-validation analysis is performed based on historical operating condition data and current real-time acquired data to reconstruct the fault identification logic of the node, thereby achieving adaptive correction of its fault identification results. This process specifically includes the following steps: Edge computing nodes with a trust level index below the threshold are marked as "nodes to be corrected" in the current fault event. Then, their historical operating data, including but not limited to voltage and current timing curves, disturbance duration, node response time, action type, and final fault type and location label, are retrieved based on various typical operating conditions recorded during their historical operation. This historical data includes, but is not limited to, events such as single-phase grounding, phase-to-phase short circuit, three-phase short circuit, switch tripping, cable breakage, and switch failure to operate. The retrieval of historical operating data is based on a similar case search according to the initial fault type identified by the current node. A waveform matching method based on dynamic time warping is used to select several sets of historical samples from the historical database that are most similar to the currently sampled waveform characteristics, which are then used as the basis for subsequent logic calibration.

[0033] The waveform matching method based on the dynamic time warping algorithm is a calculation method for measuring the similarity of two time series after nonlinear alignment on the time axis. In this invention, it is used to solve the similarity assessment problem of fault waveforms under conditions such as propagation delay and differences in disturbance response speed on the time scale. The method's function is to accurately select historical samples that are most similar in terms of change trend, disturbance pattern, and response process by performing nonlinear alignment matching between the voltage or current waveforms collected from the current fault event and waveforms labeled with fault type and location in the historical database, when the reliability level of edge computing node identification is low. This provides a reliable reference for the dynamic correction of the identification logic. The specific steps include: First, normalizing the current fault sampling waveform to ensure its amplitude range matches historical samples, while also unifying the sampling frequency and number of sampling points; second, iterating through all waveform samples in the historical database and using a dynamic time warping algorithm to perform matching calculations between the current waveform and each historical sample waveform, allowing nonlinear stretching and compression on the time axis to obtain the shortest "alignment path distance" between the two waveforms at different time scales; third, sorting the alignment path distances of all historical samples and selecting the top few samples with the smallest distances as the candidate sample set with the highest similarity; finally, extracting the corresponding fault type, location, and disturbance characteristics from these samples, cross-comparing them with the current node identification results to determine if there are significant deviations, and using this as the basis for adjusting logical weights and correcting identification rules. This method focuses on trend matching of the waveform morphology itself by tolerating time-series offsets, exhibiting strong robustness in actual distribution network faults, and effectively addressing complex situations such as response delays and signal noise disturbances in sampling at different nodes, thereby improving the ability of historical experience data to assist in the current judgment process.

[0034] The time-series signals of three-phase voltage and three-phase current collected during the current fault event at this node are aligned with the retrieved historical sample data using feature vector alignment. Specifically, this involves extracting key feature parameters from multiple dimensions, such as normalized current drop amplitude, negative-sequence voltage formation rate, maximum change in zero-sequence current amplitude, fault initiation sensing time point, and phase angle disturbance rate, to establish a multi-dimensional feature vector space. Based on this, the K-nearest neighbor algorithm is used for multi-dimensional spatial projection comparison, calculating the Euclidean distance between the current data and each set of historical samples, and selecting the closest sets of samples for fault type and location reference labeling. If the current node identification result deviates significantly from the majority of judgment results of similar historical samples, the possibility of bias or misinterpretation of features in the original judgment is identified, and the logical reconstruction process begins.

[0035] By combining the label results corresponding to historical samples, the identification logic function of the current node is dynamically reconstructed. The reconstruction process includes adjusting the feature weight coefficients in the fault identification model; for example, increasing the weight of the negative sequence voltage change rate on ground fault identification, reducing the weight interference from insufficient sensitivity to instantaneous current drops, and resetting the threshold judgment conditions for multi-feature combinations. Furthermore, based on common misjudgment types in historical data, the logical relationships in the identification rule base are adjusted, such as adding a special case identification rule for "voltage not dropping but zero-sequence current suddenly rising." All logic reconstruction processes are data-driven and retain records of each modification for subsequent analysis and evaluation of the correction effect.

[0036] The corrected identification logic is reapplied to the original sampled data of the current fault event for secondary identification, outputting the corrected fault type and location judgment results. Simultaneously, a consistency comparison analysis is performed on the judgment results before and after correction. If the corrected result shows improved consistency with the judgment results of most neighboring nodes in the spatial topology, and the judgment time is advanced or misjudgment features are eliminated, then the correction is confirmed to be effective, and the corrected logic is marked as a "learnable template" for automatic matching of similar working conditions in the future. If the corrected result still lacks consistency or accuracy, the node is marked as "recognition limited," and manual review or retraining of its identification logic model is recommended.

[0037] For edge computing nodes whose reliability index falls below a set threshold, historical operating condition data is invoked and combined with currently collected real-time data. Through cross-validation, the fault identification logic is reconstructed, adaptively correcting the fault identification results for that edge computing node. Its core function is to improve the accuracy of identification judgments, compensate for identification deviations caused by local perception distortion, feature misjudgment, or rule failure, and ensure the stability and consistency of the entire distribution network edge identification system. In an edge computing architecture, a node's independent fault judgment may be affected by multiple factors, such as abnormal sensor sampling, incomplete feature extraction, or unreasonable rule threshold settings, leading to significant deviations between its identification results and those of neighboring nodes. By introducing historical operating condition data, waveform features and identification tags that have been verified accurately in similar past fault cases can be invoked and cross-compared with the currently collected data to identify whether the current judgment is within an abnormal range. Furthermore, based on the comparison results, the original identification logic of that node is specifically adjusted, including recalibrating the judgment priority of voltage and current characteristics, updating the threshold range, and introducing special case identification conditions, making the identification logic more closely aligned with current operating conditions. The corrected identification results will again participate in the network-wide identification consistency analysis to ensure that the information output by the node conforms to the overall topology judgment trend and timing propagation logic, thereby avoiding control command conflicts or isolation action errors caused by misjudgments by individual nodes. This step not only realizes the self-learning and self-repair of the identification process, but also builds an intelligent fault-tolerant mechanism for the entire distribution network in the face of dynamically changing environments, which is a key supporting link for realizing highly reliable distributed identification and control.

[0038] S004. Based on the adaptively corrected fault identification results, calculate the global decision weight of each edge computing node in the current fault event, and determine the control priority of each edge computing node in the fault isolation operation according to the decision weight allocation principle. To address the potential control conflicts arising from multiple edge computing nodes simultaneously possessing control capabilities under the same fault event, a method is proposed that calculates global decision weights based on adaptively corrected fault identification results. Then, a clear control priority is assigned to each edge computing node according to this decision weight, ensuring that isolation operations are executed logically, safely, and orderly within the distribution network. This process is achieved through the following steps: After all edge computing nodes complete fault identification and adaptive correction through cross-validation based on historical operating conditions, the identification result parameters of each node are extracted. These parameters include the final fault type (e.g., three-phase short circuit, B-phase grounding), the identified fault location (e.g., 35 meters downstream from this node), the time when the node first senses the fault (e.g., detecting a voltage drop within 12 milliseconds after the fault occurs), and the node's physical location in the distribution network topology, such as whether it is located on the main line or a terminal branch. Simultaneously, the execution records of whether the node accurately identified and successfully issued control commands in the past 30 fault events are obtained, including the success rate (e.g., 83.3%), the false alarm rate (e.g., 3.4%), and the response latency (e.g., an average of 16 milliseconds). These parameters will serve as inputs for calculating decision weights.

[0039] All input factors are normalized. A scoring range of 0 to 1 is used, with the initial score directly derived from the recognition reliability level. The advance detection time is converted into a proportional score by ranking all nodes based on their fault initiation detection time; for example, the first node to detect a fault scores 1, and the latest scores 0. The topology location index is calculated by weighting factors such as whether the node is located on the main feeder, connects to multiple downstream branches, or is close to the distribution transformer. For example, a node located in the middle of the main line and connected to three or more branches scores 0.9. Historical isolation execution performance is scored as the net efficiency value (success rate minus false alarm rate). Subsequently, a fixed-weighted model is used to aggregate all indicators: recognition reliability level accounts for 40%, response advance time for 25%, topology importance for 20%, and historical execution efficiency for 15%. Each node ultimately receives a global decision weight value, which represents its relative control capability and command dominance in this fault response.

[0040] Based on the global decision weight values ​​of all nodes, a priority list for edge control under the current fault event is generated by sorting them from highest to lowest. The node ranked first in the list is the node with the highest current control authority and has the right to issue isolation commands first. If the weight score difference between two nodes is less than 0.03, the parallel relationship will be broken based on whether the node has prioritized isolation actions in the previous fault event; if they are still the same, the priority will be assigned to the node closer to the fault location based on the physical distance between the two nodes. In addition, it is checked whether the high-priority node has electrical topology overlap or a common power supply relationship with the next-level node. If so, concurrent execution is avoided by delaying the response window time of the next-level node.

[0041] At the execution level, each edge computing node sets a local response waiting time window based on the received priority list. For example, the first-ranked node will respond immediately at 0 milliseconds after detecting a fault; the second-ranked node will respond after a 30-millisecond delay; and the third-ranked node will respond after a 60-millisecond delay. If a high-priority node completes the isolation operation within the set time and broadcasts the execution status via the communication link, other low-priority nodes will automatically cancel the action to avoid duplicate operations. If a high-priority node fails to execute, the next-level node will automatically take over the task, ensuring that fault isolation is uninterrupted and undelayed.

[0042] Based on the adaptively corrected fault identification results, the global decision weight of each edge computing node in the current fault event is calculated. Then, according to the decision weight allocation principle, the control priority of each edge computing node in the fault isolation operation is determined. This aims to resolve the control conflict problem caused by parallel judgments from multiple nodes in the edge computing architecture, achieving orderliness, stability, and coordination in the fault isolation decision-making process. In the scenario of autonomous control at the edge of a distribution network, different nodes may identify the same fault almost simultaneously. However, due to differences in geographical location, sensing speed, and identification accuracy, each node's judgment of the fault type and location has certain subjectivity and limitations. Without setting control priorities, multiple nodes may issue isolation action commands simultaneously, leading to risks of repeated disconnections, out-of-order actions, and even mutual interference. Therefore, this step first integrates various factors such as the identification reliability level, adaptive correction results, fault response timeliness, electrical importance in the topology, and historical isolation performance of each node to construct a multi-dimensional evaluation system, assigning a quantified global decision weight to each node. This weight value objectively reflects the reliability of the node's identification results, execution capability, and control value in the overall network during this fault event. Subsequently, based on preset weight allocation rules, the nodes are sorted by weight to generate a unique control priority list, clearly defining which node has the highest control say in the current scenario. Other nodes respond or wait for confirmation signals in order of priority, avoiding preemption of commands. This priority-driven control strategy not only ensures clear division of responsibilities and execution paths for fault isolation commands throughout the entire distribution network, but also improves the efficiency, accuracy, and consistency of fault handling, effectively enhancing the overall performance and reliability of intelligent autonomous control of the distribution network under edge computing conditions.

[0043] S005, based on the control priority of each edge computing node, performs dynamic pre-simulation analysis of isolation operations, constructs the impact path of isolation operations, evaluates the power supply interruption, voltage fluctuation and power flow anomaly that isolation operations may cause under the existing topology and power flow boundary conditions, and identifies the feasibility and risks of isolation actions in advance; To ensure that edge computing nodes can accurately predict potential power impacts before actually executing isolation operations, and to avoid cascading power outages, voltage instability, or power flow disturbances caused by erroneous actions, a dynamic isolation action pre-simulation analysis method based on node control priorities is proposed. This method combines the current distribution network topology and real-time power flow boundary conditions to perform path analysis and power consequence assessment for the pre-executed isolation operation. Specifically, it is implemented through the following steps: Based on the control priority list generated in the previous stage, the fault identification results and upcoming isolation operation information of the current top-ranked edge computing node are extracted, including the specific switch equipment number to be disconnected, its location, control direction (e.g., downstream disconnection), and the range of line segments covered by the isolation action. Simultaneously, the current distribution network topology diagram is retrieved from the dispatch data center. This diagram includes the physical connections of all nodes, cable parameters, switch status, load distribution, power supply access point locations, and power flow direction. Starting from the current node, the system automatically traces the upstream and downstream equipment and loads affected by the disconnection action along the topology path, constructing a complete isolation operation impact path diagram, including the isolated loads, potential islanded areas, available tie switches, and bypass branches.

[0044] By combining real-time power flow monitoring data, the current operating voltage, current, load power, and bus potential information of each key node in the affected path are loaded as the initial boundary conditions for the simulation. Subsequently, power flow recalculation is performed, and a distributed power flow algorithm for distribution networks (such as a power flow solution method based on the modified Newton method) is used to simulate the redistribution of network power flow after the planned isolation is completed at the current node. During the simulation, three key indicators are observed: first, whether the voltage amplitude variation of each branch within the isolation area exceeds the allowable fluctuation range of ±7%; second, whether there is excessive current flow or backfeeding in each segment of the non-isolated path; and third, whether the network topology forms island areas or breaks the ring network due to isolation.

[0045] Based on the power flow simulation results, the feasibility and potential risks of this isolation operation under current operating conditions are assessed. If the assessment finds that isolating a branch will cause a voltage drop of more than 10% on the low-voltage side of more than two transformers, or cause the cable to carry current exceeding its rated capacity for a long period of time, or determine that there is a possibility of backflow current causing maloperation of the upper-level protection, then the isolation action will be marked as a "high-risk operation," and the node will be prohibited from executing this action. The priority will be downgraded to the next level node to re-execute the simulation process. If the simulation results show that the network structure remains stable, the voltage and current are within the safety boundaries, and there is no islanding, then the action will be marked as "safe and executable."

[0046] All data results, model parameters, risk analysis conclusions, and execution recommendations from the pre-simulation process are compiled into a complete isolation action simulation analysis report, which is then distributed to relevant edge nodes and the control center via the power distribution communication network as decision support material before formal action execution. High-priority nodes decide whether to immediately execute isolation actions based on the simulation report results, while low-priority nodes adjust their response strategies based on the availability of successor tasks.

[0047] Based on the control priorities of each edge computing node, dynamic pre-simulation analysis of isolation operations is performed to construct the impact path of isolation operations and assess the potential power supply interruptions, voltage fluctuations, and power flow anomalies that isolation operations may cause under the existing topology and power flow boundary conditions. Its purpose is to verify the impact of the operation on the distribution network's operating status before the formal execution of the fault isolation action, thereby identifying potential risks, avoiding chain reactions caused by blind operations, and ensuring the safety and stability of the power grid operation. In the edge computing architecture, each node has the ability to independently issue isolation control commands. Without effective verification, actions issued by a high-priority node may trigger voltage drops, load imbalances, ring network breaks, and power flow reversals, leading to erroneous power outages in non-faulty areas or the formation of electrical islands, thus affecting the power supply continuity and power quality of the entire distribution system. Therefore, through this step, before the node with the highest control priority initiates an isolation operation, the electrical path involved in the isolation action is constructed based on the current distribution network topology and real-time power flow data. The redistribution of power flow after this path is broken is simulated, and its comprehensive impact on voltage, current, load, and power supply support capacity is assessed. If the simulation reveals that the voltage deviation is too large after isolation, the cable current exceeds the safety limit, or some areas will be isolated, the operation will be deemed high-risk and will not be executed. Instead, it will be transferred to the next priority node for re-simulation and evaluation until the optimal isolation scheme is found. This step, through a closed-loop control mechanism of "verification before execution," effectively improves the decision-making accuracy and response safety of edge intelligent control in dealing with fault scenarios. It ensures that every isolation action is not only timely and fast, but also controllable and reliable, avoiding global operational disorder caused by local judgment errors.

[0048] S006, based on the dynamic pre-simulation analysis results of isolation operation, performs full-process status monitoring of the actual isolation operation process, extracts the deviation index between the actual system response data and the pre-simulation response results, and uses the deviation index to correct the fault identification consistency assessment model and fault identification logic, forming a closed-loop control mechanism of self-optimization of identification, simulation and feedback. To achieve closed-loop self-optimization of the entire process of fault identification and isolation control in distribution networks, a method is proposed that uses dynamic pre-simulation analysis results of isolation operations to monitor the actual isolation execution status throughout the entire process, and dynamically corrects the fault identification consistency assessment model and fault identification logic using deviation feedback. This method constructs a complete closed loop of "identification-simulation-monitoring-feedback" by comparing the fine deviations between prediction and actual response, significantly improving the accuracy and adaptability of identification and control strategies. Specifically, it is implemented through the following steps: After the edge computing node completes the isolation operation according to control priority, the entire electrical condition monitoring process is immediately initiated. The monitoring time is from 50 milliseconds before the operation is triggered to at least 5 seconds after the operation. The monitoring objects include the three-phase voltage change curves upstream and downstream of the isolation point, the discontinuous current waveforms on both sides of the break point, the disappearance or transfer of load current in the isolation area, the power flow path reconstruction after the disconnection action, the voltage recovery time, whether the current direction is reversed, and whether an electrical islanding area is formed. All monitoring data are recorded by the node's local high-frequency sampling device at 2-millisecond intervals, with a unified timestamp, and transmitted in real time to the distribution master station simulation verification platform for accurate alignment and comparison with the pre-simulation model results.

[0049] Retrieve dynamic pre-simulation analysis data completed before the isolation operation. This data includes simulated predictions of voltage, current, and power flow direction at each node under the preset operation. For example, the simulation predicts that after isolation, the voltage at downstream node B will drop from 228V to 215V, the current will drop from 60A to 32A, and the power flow will switch from branch C to branch D. Align the actual monitoring data with the simulation results point by point, perform precise difference calculations and compare the change trends, and extract several deviation indicators, including voltage deviation percentage, consistency of current change trends, matching of power flow direction, voltage recovery time difference, and branch section load error. Set specific judgment thresholds for each deviation indicator. For example, voltage deviation exceeding ±7%, power flow direction completely opposite to the prediction, and current section reconstruction not executed according to the expected path are marked as high-risk deviation items.

[0050] Based on feedback from deviation indicators, the consistency assessment model and fault identification logic are jointly revised. Regarding the consistency assessment model, if the pre-simulation performance of a node deviates significantly from the actual response, its score weight in the consistency scoring of adjacent nodes is reduced, and its historical deviation factor is increased. Simultaneously, "response-prediction consistency" is introduced as a new evaluation dimension when assessing its judgment reliability level. For the revision of the fault identification logic, the feature extraction and judgment conditions leading to misjudgments are analyzed. For example, if it is found that a node is misidentified as a downstream short circuit due to over-reliance on voltage drop amplitude, the voltage sensitivity threshold in its fault identification rules is adjusted, and the weight of the zero-sequence current slope is increased. At the same time, a comparison record between the original rules and the revised logic is retained for effect comparison in subsequent operating conditions.

[0051] The revised identification consistency model and identification logic are loaded onto all edge computing nodes participating in the identification process, overriding the local real-time identification algorithm parameters and being used as a new version in subsequent fault events. Simultaneously, a feedback data archiving mechanism is established, creating a structured record of each isolation operation's simulation predictions, actual responses, deviation results, corrections, and subsequent action results, which is then integrated into the main station data platform as foundational samples for long-term model optimization and identification strategy evolution. As the operating time progresses, the identification model will continuously integrate new fault response data, autonomously optimizing the identification path and dynamically adjusting thresholds, constructing a truly self-learning, self-adaptive, and self-repairing intelligent edge control closed loop, achieving a highly efficient evolution of distribution network fault handling from rule-driven to data-driven approaches.

[0052] Based on the dynamic pre-simulation analysis results of isolation operations, the entire process of actual isolation operations is monitored. Deviation indicators between the actual response data and the pre-simulation response results are extracted, and these deviation indicators are used to correct the fault identification consistency assessment model and fault identification logic. This establishes a closed-loop control mechanism from prediction to verification to optimization, enabling the fault identification and isolation control process of the distribution network to have self-correction, self-learning, and continuous evolution capabilities. In actual operation, even if edge computing nodes possess certain identification and decision-making capabilities, factors such as simplified simulation models, parameter estimation errors, or sudden operational disturbances inevitably exist, leading to differences between the actual execution effect of isolation operations and the expected simulation. By introducing full-process state monitoring, key operational indicators such as breakpoint voltage and current changes, load recovery status, and power flow reconfiguration paths can be collected in real time after operation execution and compared one-to-one with the pre-generated simulation results, thereby accurately extracting the source and type of deviation. For example, if the actual power flow does not transfer along the predicted path, the voltage recovery time exceeds expectations, or an abnormal increase in current occurs in a branch after isolation, these may indicate model deviations or logical defects in the fault identification or simulation process. Based on this deviation information, the weighting coefficients, judgment rules, or evaluation algorithms in the identification model can be adjusted accordingly. This includes redefining the correspondence between voltage dips and fault types, introducing new feature combination logic, or modifying the inter-node consistency scoring mechanism. This feedback correction can not only be used for subsequent processing of current events but also incorporated as training samples into the historical database, driving long-term adaptive optimization of the model. Ultimately, this step achieves a closed loop between identification, simulation, and execution results, enabling the distribution network to progressively improve judgment accuracy and response efficiency, thus constructing a truly self-optimizing edge autonomous control system.

[0053] The aforementioned edge computing-based fault identification-driven precise isolation method for distribution networks enables a highly efficient closed-loop control mechanism from local data acquisition to coordinated response across the entire network, significantly improving the accuracy of fault identification and the consistency of isolation control in complex fault scenarios. This invention calculates the difference in fault identification judgments between edge nodes, constructs a consistency evaluation model based on the topology, and quantifies the reliability of the identification results, effectively identifying judgment deviations and potentially misidentified nodes. Simultaneously, cross-validation of historical data and real-time information achieves adaptive correction of the identification logic, ensuring that the judgment of each node better reflects actual operating conditions. Furthermore, the introduction of a control priority mechanism and dynamic pre-simulation analysis helps to comprehensively assess the feasibility and risks of isolation before action execution, avoiding power outage misoperations. Finally, through monitoring and feedback correction of actual response deviations, the identification model and control strategy are continuously optimized, achieving a self-evolving closed-loop control integrating identification, simulation, and feedback. This scheme overcomes the conflicts and miscontrol problems caused by isolated judgments of traditional edge nodes, effectively improving the stability, accuracy, and system-level coordination of fault response, ensuring the safe and reliable operation of the distribution system under a distributed architecture.

[0054] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for precise isolation driven by fault identification in power distribution networks based on edge computing architecture, characterized in that, Includes the following steps: S001: Collect voltage and current signals at each edge computing node of the distribution network, analyze the signal fluctuation amplitude and propagation delay of different nodes based on the same fault event, and calculate the difference in fault identification and judgment between nodes. S002, based on the difference in fault identification and judgment between nodes, combined with the distribution network topology, a fault identification consistency assessment model is constructed according to the difference in fault identification and judgment. The probability analysis of the fault type and location identified by each node is performed, and the identification reliability level index is output. S003, for edge computing nodes whose trust level index is lower than the set threshold, the fault identification logic is reconstructed based on cross-validation of historical operating conditions and real-time data, and the fault identification results of the edge computing nodes are adaptively corrected. S004, Based on the adaptively corrected fault identification results, calculate the control priority of each edge computing node in the fault isolation operation; S005, based on the control priority of each edge computing node, perform isolation operation pre-simulation analysis to evaluate the operation path and risk under topology and power flow conditions; S006, based on the dynamic pre-simulation analysis results of isolation operation, monitor the isolation execution status, extract response deviations and use them to correct the identification model and control strategy, and construct a closed-loop control mechanism of identification-simulation-feedback.

2. The method for precise isolation driven by fault identification in distribution networks based on edge computing architecture according to claim 1, characterized in that, Step S001 includes: Three-phase voltage and three-phase current signals are collected at each edge computing node. 20,000 sample points are recorded per second using a sampling frequency of 20kHz. DC bias removal and normalization are performed on each group of signals. A mother wavelet function with abrupt change enhancement characteristics is used to perform continuous wavelet transform, calculate wavelet modulus and extract local maximum modulus points at each scale, construct modulus extreme value paths, and identify the abrupt change point time when each node first senses fault disturbance; Based on the sensing time, voltage and current characteristic parameters and network topology connections uploaded by edge nodes, a judgment difference matrix is ​​constructed and the fault identification judgment difference between nodes is quantified.

3. The method for precise isolation driven by fault identification in distribution networks based on edge computing architecture according to claim 1, characterized in that, Step S002 includes: After obtaining the fault identification and judgment difference matrix between edge computing nodes, a directed graph structure containing node connection direction, feeder length, conductor type and impedance parameters is constructed by combining the physical topology connection relationship of the distribution network. Based on the fault type, fault location and perception time identified by each edge computing node, the similarity of type, proximity of location and perception time order are compared with the identification results of adjacent nodes to generate a compatibility score vector. The compatibility score vector is input into the multi-factor Bayesian joint probability model. By combining the historical identification accuracy of nodes with prior knowledge of fault propagation, the probability value of each node identification result being supported as a real fault judgment is inferred. Based on the joint probability value obtained from reasoning, the difference in node judgment, the topology depth, the degree of response time lag, and the historical recognition accuracy, a recognition credibility level scoring function is constructed, and the recognition credibility level index is output.

4. The method for precise isolation driven by fault identification in distribution networks based on edge computing architecture according to claim 1, characterized in that, Step S003 includes: For edge computing nodes whose confidence level index is below the set threshold, their historical operating condition data is called up and the three-phase voltage signal and three-phase current signal in the current fault event are matched with the waveform using a dynamic time warping algorithm to filter similar historical samples. Extract multidimensional feature parameters of the current waveform and historical samples and compare them using the K-nearest neighbor algorithm to identify the deviation between the judgment result and the historical label. Adjust the feature weight coefficients and recognition rule base in the recognition logic based on historical sample label information, and reconstruct the recognition logic; The corrected recognition logic is reapplied to the original sampled data, the corrected result is output and the consistency of the correction is evaluated. If the consistency is improved, the corrected logic is marked as a learnable template.

5. The method for precise isolation driven by fault identification in distribution networks based on edge computing architecture according to claim 1, characterized in that, Step S004 includes: Based on the adaptively corrected fault identification results, the identification confidence level value, fault perception time, topology location index and historical isolation execution performance index of each edge computing node are extracted. The extracted indicators are normalized and a fixed-weighted model is used to calculate the global decision weight value. A control priority list is generated from high to low based on the global decision weight value of each node, and a local response waiting time window is set; When a high-priority node completes the isolation operation and broadcasts the execution status, the low-priority node cancels the action; otherwise, the next-level node takes over the task and completes the fault isolation.

6. The method for precise isolation driven by fault identification in distribution networks based on edge computing architecture according to claim 1, characterized in that, Step S005 includes: Based on control priority, extract the isolation operation information and identification results of the current edge computing node; By calling the distribution network topology diagram and power flow boundary conditions, an isolation operation impact path diagram is constructed; Perform power flow recalculation and analysis of distribution network isolation operations to assess voltage fluctuations, current changes, and topology breakage risks. The feasibility and risk level of isolation operations are determined based on the power flow simulation results, and an isolation action simulation analysis report is generated for reference by edge nodes and control centers.

7. The method for precise isolation driven by fault identification in distribution networks based on edge computing architecture according to claim 1, characterized in that, Step S006 includes: After the isolation operation is completed at the edge computing node, the isolation process status monitoring is started, and the voltage, current and power flow direction change data before and after the operation are collected and transmitted to the simulation verification platform. The dynamic pre-simulation analysis results completed before the isolation operation are called up and compared with the actual monitoring data point by point to extract multiple deviation indicators; Based on various deviation indicators, the scoring weights of the fault identification consistency evaluation model and the feature extraction rules of the fault identification logic are jointly modified. The revised model and logic are distributed to each edge computing node and the simulation predictions, actual responses, deviation results and corrections are archived to form a closed-loop control mechanism of identification, simulation and feedback.

Images