An Artificial Intelligence-Based Power Data Management System and Method

Abstract

This invention discloses an artificial intelligence-based power data management system and method, relating to the field of power data acquisition and analysis technology. The invention compresses waveform data into standardized fault feature vectors at the edge and uploads them. At the master station, data sets for the same event are organized according to the event timestamp field and event cache aggregation conditions, reducing the impact of strict alignment of full waveforms across points on the positioning process and ensuring that the master station's inference input maintains a unified format that is both field-based and aggregatable. Switch status codes and topology version identifiers are written into the graph structure data, and a single topology version is locked within one inference cycle, ensuring that the graph inference process remains consistent with the current topology state. Simultaneously, edges corresponding to disconnected topology connections are removed from the adjacency relationships, ensuring that the candidate set of fault sections is consistent with the conduction connectivity relationship.

Description

Technical Field

[0001] This invention relates to the field of power data acquisition and analysis technology, and in particular to a power data management system and method based on artificial intelligence. Background Technology

[0002] Single-phase grounding faults are prone to occur during the operation of medium-voltage distribution networks. When a fault occurs, distribution automation terminals, feeder terminals, and monitoring devices can collect electrical quantities such as zero-sequence current and zero-sequence voltage, and generate waveform data containing high-frequency components during the transient phase of the fault. Various system implementations exist in engineering for the collection, aggregation, and analysis of this type of data.

[0003] In existing technologies, one implementation uses a centralized platform to uniformly analyze waveform data from multiple monitoring points: each monitoring point uploads transient waveforms and event information to the main station or cloud platform, where the central side performs similarity comparisons, extracts the arrival time of traveling waves, and correlates network topology to identify suspected faulty feeders and sections. This type of approach typically requires time alignment of data recorded by multiple terminals and relies on communication links to transmit a large amount of waveform data back to the central side. Another implementation involves on-site detection at the edge terminal: the terminal triggers a grounding alarm based on the amplitude exceeding limits of zero-sequence voltage and zero-sequence current, phase relationship, or a small number of statistical quantities, and uploads the alarm and some waveform segments to the platform for display and processing. In cases of high-resistance grounding, resonant grounding, and grounding accompanied by intermittent arcing, the amplitude and transient characteristics of the fault signal may exhibit time-varying and intermittent characteristics, resulting in uncertainty in the alarm triggering time and transient characteristic extraction. Furthermore, multiple transient disturbances may cause repeated alarms during the same fault process. Another implementation uses programmable logic controllers or script frameworks to configure rules for filtering, suppressing, and simply linking terminal alarms. Rule scripts typically rely on threshold criteria and logical combinations, with parameter configuration depending on operational experience. In complex multi-branch distribution networks and changing operating modes, the adaptability of rule parameters and criteria is significantly affected by on-site conditions.

[0004] In multi-terminal centralized analysis scenarios, the alignment accuracy of multi-source waveform data is affected by terminal time synchronization errors, link delays, and data arrival jitter. Waveform correlation analysis and arrival time comparison are prone to deviations under data misalignment conditions, thus affecting the fault segment judgment and location results. To meet alignment accuracy requirements, engineering typically requires configuring higher-precision time synchronization methods or introducing time synchronization capabilities into the communication network. However, in scenarios with a large number of existing terminals and significant differences in communication conditions, the deployment and maintenance costs of time synchronization are high.

[0005] In edge threshold discrimination and rule script scenarios, the transient characteristics of single-phase grounding faults are affected by fault resistance, fault initial angle, network capacitance distribution and arc extinguishing process. Under some operating conditions, the amplitude of electrical quantities is insufficient to trigger conventional over-limit criteria, or the critical transient moment has been crossed before the criteria are met, causing uncertainty in fault event identification and classification. When alarm suppression adopts delay or filtering methods, the correspondence between alarm triggering time and transient information may also be shifted.

[0006] With the increasing number of sensing points and the growth in data volume in the power distribution network, the power data management system needs to process multi-source asynchronous data, waveform data and alarm event data simultaneously, and meet the timeliness and consistency requirements of on-site handling in terms of data alignment, event correlation, alarm management and fault analysis. Summary of the Invention

[0007] This application provides an artificial intelligence-based power data management system and method to solve the problems of strong dependence on cloud waveform synchronization, easy storming of threshold alarms, and difficulty in locating loops.

[0008] In a first aspect, embodiments of the present invention provide an artificial intelligence-based power data management system, comprising: Multiple edge feature extraction devices are deployed at multiple monitoring points in the power distribution network; and a master station collaborative analysis device is communicatively connected to the multiple edge feature extraction devices. Any of the edge feature extraction devices is configured to acquire local electrical quantity waveform data, perform preprocessing on the electrical quantity waveform data, input it into the first artificial intelligence model, generate a standardized fault feature vector, and send it to the main station collaborative judgment device; The master station collaborative analysis device is configured to receive standardized fault feature vectors from the multiple monitoring points, obtain distribution network topology data, construct graph structure data from the topology data, write the standardized fault feature vectors into the corresponding node features, input them into the second artificial intelligence model, and output fault section location information and fault nature classification information.

[0009] In some embodiments, the first artificial intelligence model is a convolutional neural network model, and the number of model parameters does not exceed a preset parameter limit; the standardized fault feature vector includes at least: a fault nature confidence field, a fault direction field, an arc state field, and an event timestamp field.

[0010] In some embodiments, the standardized fault feature vector further includes a deep feature code output by the convolutional neural network model, and the dimension of the deep feature code is smaller than the number of sample points of the electrical quantity waveform data after preprocessing.

[0011] In some embodiments, the second artificial intelligence model is a graph neural network model; The nodes of the graph structure data correspond to the monitoring points, and the edges correspond to the topological connections defined by switching equipment or line segments in the power distribution network. The graph neural network model outputs a set of fault probabilities for each edge, and the main station collaborative analysis device determines the fault segment location information based on the set of fault probabilities. When constructing the graph structure data, the master station collaborative analysis device writes the fault nature confidence field, fault direction field, arc state field, event timestamp field and depth feature code from the standardized fault feature vector into the node features of the corresponding node, and writes the node validity mask field into the node that has not received the standardized fault feature vector. Each edge of the graph structure data is written with an edge feature containing a switch state encoding field, which is used to characterize the on or off state of the corresponding topology connection; the master station collaborative analysis device performs dynamic updates on the graph structure data based on the topology update timestamp field and generates a topology version identifier, locking a single topology version identifier within one second artificial intelligence model inference cycle; During the reasoning process of the second artificial intelligence model, message transmission is blocked for edges that are represented as disconnected by the switch state encoding field, and message transmission of missing nodes is suppressed based on the node validity mask field. The main station collaborative judgment device performs connectivity constraint aggregation on edges in the conducting state based on the fault probability set, determines the connected subgraph corresponding to the fault segment, and outputs the fault segment location information.

[0012] In some embodiments, any of the edge feature extraction devices is further configured to: When the event triggering conditions are met, multiple standardized fault feature vectors are generated within a preset time window and a vector sequence is formed. The vector sequence and the corresponding sequence index are sent to the main station collaborative analysis device. The main station collaborative analysis device is also configured to perform time-series pattern analysis on the vector sequence, output time-series classification results related to the arc state, and write the fault nature classification information. The preset time window is determined by the window length and step size. The edge feature extraction device adaptively determines the window length and step size used for this vector sequence generation from the window length candidate set and the step size candidate set based on the fault nature confidence field and arc state field in the standardized fault feature vector corresponding to the trigger point. The sequence index includes at least a monitoring point identifier field, a sequence identifier field, a trigger point timestamp field, a window length encoding field, a step size encoding field, and a sequence number field within the sequence; The main station collaborative analysis device uses a temporal convolutional network model to perform temporal pattern analysis on the vector sequence, outputs a temporal classification result including an arc time series category field and a temporal analysis confidence field, and statistically obtains the arcing cycle field, arcing duty cycle field, and number of intermittent fields based on the arc state field and event timestamp field of each vector in the vector sequence. The master station collaborative analysis device writes the arc timing category field, timing analysis confidence field, arcing cycle field, arcing duty cycle field, and number of intermittent fields into the fault nature classification information, and writes the alarm storm identifier field into the fault nature classification information when the number of intermittent fields and the arcing duty cycle field meet the alarm storm criterion.

[0013] In some embodiments, the event triggering condition includes: when the fault nature confidence field is greater than a preset confidence threshold and the fault direction field meets a preset direction criterion, the edge feature extraction device initiates the generation and transmission of the vector sequence of the preset time window.

[0014] In some embodiments, any of the edge feature extraction devices includes a data preprocessing unit configured to perform at least two of the following before the electrical quantity waveform data is input into the first artificial intelligence model: noise filtering, amplitude normalization, resampling, and fixed-length truncation.

[0015] Secondly, embodiments of the present invention provide an artificial intelligence-based power data management method, including: Local electrical quantity waveform data are acquired from multiple monitoring points and preprocessed. The preprocessed electrical quantity waveform data is input into the first artificial intelligence model to generate a standardized fault feature vector. The standardized fault feature vector is sent to the main station collaborative analysis device; The main station collaborative analysis device receives standardized fault feature vectors from multiple monitoring points and obtains distribution network topology data. The topology graph data is constructed into graph structure data and the standardized fault feature vector is written into the corresponding node features; The graph structure data is input into the second artificial intelligence model, which outputs fault segment location information and fault nature classification information.

[0016] In some embodiments, an event triggering judgment is performed before the data is sent to the main station collaborative analysis device; When the confidence field of the fault nature in the standardized fault feature vector is greater than the preset confidence threshold, a vector sequence is generated within a preset time window and sent to the main station collaborative judgment device. The main station collaborative judgment device updates the fault nature classification information based on the vector sequence.

[0017] In some embodiments, when the master station collaborative analysis device receives the standardized fault feature vector, it writes the vectors from different monitoring points into the event cache area according to the event timestamp field. When the event buffer meets the preset aggregation conditions, the graph structure data is constructed and the reasoning of the second artificial intelligence model is executed; the preset aggregation conditions include at least one of: reaching a preset monitoring point number threshold or exceeding a preset waiting time threshold.

[0018] Through the above technical solution, the present invention can achieve at least the following beneficial effects: To address the issues of timing errors, communication delays, and out-of-order arrival of transient data from multiple terminals, the edge side compresses waveform data into standardized fault feature vectors and sends them up. The main station organizes the data set of the same event according to the event timestamp field and event cache aggregation conditions, reducing the impact of strict alignment of full waveforms across points on the positioning process and ensuring that the main station's inference input maintains a unified format that is both field-based and aggregable.

[0019] To address the issue of inconsistencies between location results and actual switch states caused by frequent changes in distribution network topology, a switch state code field and a topology version identifier are written into the graph structure data. A single topology version is locked within one inference cycle to ensure that the graph inference process is consistent with the current topology state. At the same time, edges corresponding to disconnected topology connections are removed from the adjacency relationships to ensure that the candidate set of fault sections is consistent with the conduction connectivity relationship.

[0020] To address the issue of incomplete input to the main station due to monitoring points not uploading, uploading delays, or missing uploading data, a node validity mask field and a missing data marker field are introduced into the node features. This allows missing nodes to be explicitly marked in the graph inference input, and together with the event aggregation rules, constrains the inference trigger time and the boundary of the input set, thereby improving the tractability of missing data scenarios.

[0021] To address the issue of intermittent electric arcs causing multiple transient disturbances in the same fault process, leading to repetitive alarms and alarm storms, a vector sequence with a sequence index structure is generated after triggering at the edge. The main station uses a temporal convolutional network to perform temporal pattern analysis on the vector sequence, outputting the arc time series category and corresponding confidence level. Based on the arc state field and event timestamp field, the arcing period, arcing duty cycle, and number of intermittent arcs are statistically analyzed, aggregating intermittent arcing behavior into a temporal feature description of the same event.

[0022] To address the issue of inconsistent boundaries in the representation of faulty sections, the master station performs connectivity constraint aggregation after obtaining the fault probability set of each edge. It calculates connected components on the set of connected edges and generates a set of candidate sections. Combining the correlation degree and the aggregated fault probability value, it determines the faulty connected subgraph and then maps it to the switchgear or line segment identifier to form the section output, so that the section representation has a field structure with consistent topological boundaries. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0024] Figure 1 This is a framework diagram of the power data management system in the embodiment.

[0025] Figure 2 This is a flowchart of the power data management method in the embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0028] Example 1: like Figure 1 As shown, this embodiment proposes an artificial intelligence-based power data management system, including: Multiple edge feature extraction devices are deployed at multiple monitoring points in the power distribution network; and a master station collaborative analysis device is communicatively connected to the multiple edge feature extraction devices. Any of the edge feature extraction devices is configured to acquire local electrical quantity waveform data, perform preprocessing on the electrical quantity waveform data, input it into the first artificial intelligence model, generate a standardized fault feature vector, and send it to the main station collaborative judgment device; The master station collaborative analysis device is configured to receive standardized fault feature vectors from the multiple monitoring points, obtain distribution network topology data, construct graph structure data from the topology data, write the standardized fault feature vectors into the corresponding node features, input the second artificial intelligence model, and output fault section location information and fault nature classification information. In one embodiment, the electrical quantity waveform data includes at least one of zero-sequence current waveform and zero-sequence voltage waveform, and each waveform data segment contains a sampling sequence and corresponding sampling time information.

[0029] The standardized fault feature vector is defined as a structured data record, including a vector header field and a vector payload field. The vector header field includes at least a monitoring point identifier field, an event identifier field, and an event timestamp field. The vector payload field includes at least a classification-related field and a feature encoding field output by the first artificial intelligence model. In this embodiment, the monitoring point identifier field is a unique identifier field within the distribution network, and its value can be derived from the equipment serial number of the edge feature extraction device or the monitoring point number issued by the master station collaborative analysis device. The event identifier field is a field that identifies a single triggered associated process within the same monitoring point. The event identifier field can be generated by combining the monitoring point identifier field and the trigger point timestamp field. In one implementation, the event identifier field is generated by the monitoring point identifier field and the trigger point timestamp field according to a preset string concatenation rule, and an incrementing count value is added when the same trigger point timestamp field is repeated to distinguish different events. The event timestamp field has a recording precision of milliseconds or seconds. The event timestamp field can be generated based on the time source of the edge feature extraction device. The time source includes at least one of satellite time synchronization, network time synchronization and local clock. When the time source is switched, the edge feature extraction device writes the time source type flag and the switching time flag into the local event status record associated with the event identifier field, so that the main station collaborative analysis device can verify the consistency of the timestamp source within the same event when aggregating in the event buffer.

[0030] The power distribution network topology data is defined as a data set describing the relationship between monitoring points and topology connections, including a set of node identifiers, a set of edge identifiers, and a field showing the association between nodes and edges; wherein each node identifier corresponds one-to-one with a monitoring point, and each edge identifier corresponds one-to-one with a topology connection.

[0031] The graph structure data is defined as a graph data object that maps the distribution network topology data to a set of nodes and a set of edges, and the received standardized fault feature vector is written into the node feature field of the corresponding node according to the monitoring point identifier field; when there is a monitoring point that has not sent the standardized fault feature vector, the corresponding node feature field is written into the missing marker field.

[0032] The fault segment location information is defined as an output record that includes at least a segment identifier field and a segment confidence field; the fault nature classification information is defined as an output record that includes at least a fault category field and a fault category confidence field.

[0033] In this embodiment, the first artificial intelligence model is a convolutional neural network model, and the number of model parameters does not exceed the preset parameter limit; the standardized fault feature vector includes at least: a fault nature confidence field, a fault direction field, an arc state field, and an event timestamp field; In this embodiment, the first artificial intelligence model is obtained through at least one of offline training and offline training followed by retraining; the training data source categories include at least one of historical waveform samples, simulated waveform samples, and samples associated with on-site handling reports. The labeling methods for fault category identifiers include at least one of manual report labeling, labeling based on maintenance handling results, and semi-automatic review labeling; the labeling methods for the fault direction field and the arc status field include at least one of manual review labeling and rule-based extraction labeling.

[0034] The training objective category of the first artificial intelligence model includes at least one of the classification loss category and the multi-task loss category, wherein the classification loss category corresponds to the classification training objective of the fault category identifier, and the multi-task loss category corresponds to the joint training objective of the fault category identifier, the fault direction field and the arc state field.

[0035] When the first artificial intelligence model is deployed on the edge feature extraction device side, it is stored in the form of a combination of model parameter file and model structure description file, and is consistent with the model parameter quantity constraint corresponding to the preset parameter upper limit. When the first artificial intelligence model is updated, the edge feature extraction device writes the model version mark field and attaches the model version mark field to the generated standardized fault feature vector, so that the main station collaborative judgment device can form a distinguished record of different model version outputs.

[0036] The acquisition method and training data source categories of the second artificial intelligence model and the temporal convolutional network model are consistent with those of the first artificial intelligence model. The training samples of the second artificial intelligence model use graph structure data as input sample units, and the training samples of the temporal convolutional network model use vector sequences as input sample units. The training target categories all include classification loss categories, and the classification labels correspond to the fault section location information and the arc timing category field, respectively.

[0037] In one implementation, the fault nature confidence field is defined as a category confidence list, which contains one or more elements, each containing a fault category identifier and a corresponding confidence value. The fault category identifier is taken from a preset fault category set, which includes at least two of the following: single-phase grounding, high-resistance grounding, and intermittent arcing grounding. The fault direction field is defined as a discrete value field, whose values ​​belong to a direction value set. The direction value set includes at least three values: along the reference topology direction, inverse reference topology direction, and unknown. The reference topology direction is defined as the topology direction from the monitoring point node to the power source node in the distribution network topology map data.

[0038] The arc state field is defined as a discrete value field, and the values ​​of the discrete value field belong to the arc state value set; the arc state value set includes at least three values: arcing, arc extinguishing, and uncertain. The event timestamp field is defined as a time stamp representing the moment when the edge feature extraction device generates the standardized fault feature vector, and together with the monitoring point identifier field, it determines the event cache writing order.

[0039] In this embodiment, the standardized fault feature vector further includes a deep feature code output by the convolutional neural network model, and the dimension of the deep feature code is smaller than the number of sample points of the electrical quantity waveform data after preprocessing. In one implementation, the deep feature encoding is defined as a fixed-length numerical vector, which is generated by the intermediate layer output or the reciprocal layer output of the first artificial intelligence model; the deep feature encoding and the fault nature confidence field together constitute part of the vector payload field of the standardized fault feature vector.

[0040] The deep feature encoding can be encoded as a fixed-point array or a floating-point array, and bound and stored with the event identifier field; when using fixed-point array encoding, the standardized fault feature vector includes a quantization scale field to characterize the mapping relationship between the fixed-point array and the floating-point representation.

[0041] In this embodiment, the second artificial intelligence model is a graph neural network model; The nodes of the graph structure data correspond to the monitoring points, and the edges correspond to the topological connections defined by switching equipment or line segments in the power distribution network. The graph neural network model outputs a set of fault probabilities for each edge, and the main station collaborative analysis device determines the fault segment location information based on the set of fault probabilities. When constructing the graph structure data, the master station collaborative analysis device writes the fault nature confidence field, fault direction field, arc state field, event timestamp field and depth feature code from the standardized fault feature vector into the node features of the corresponding node, and writes the node validity mask field into the node that has not received the standardized fault feature vector. Each edge of the graph structure data is written with an edge feature containing a switch state encoding field, which is used to characterize the on or off state of the corresponding topology connection; the master station collaborative analysis device performs dynamic updates on the graph structure data based on the topology update timestamp field and generates a topology version identifier, locking a single topology version identifier within one second artificial intelligence model inference cycle; During the reasoning process of the second artificial intelligence model, message transmission is blocked for edges that are represented as disconnected by the switch state encoding field, and message transmission of missing nodes is suppressed based on the node validity mask field; the main station collaborative judgment device performs connectivity constraint aggregation on edges in the conducting state based on the fault probability set, determines the connected subgraph corresponding to the fault segment, and outputs the fault segment location information. In one implementation, the topology connection corresponding to the edge is defined by a line segment or switch interval, and each edge includes an edge endpoint field and an edge status field; the edge endpoint field includes a start node identifier and an end node identifier, and the edge status field includes at least one status marker, either closed or open. The output of the graph neural network model is defined as a fault probability set, which contains one or more edge probability elements, each containing an edge identifier and a fault probability value. The master station collaborative analysis device generates a candidate segment set based on the fault probability set. The candidate segment set contains one or more candidate segment elements, each containing at least a segment identifier field, a segment confidence field, and a corresponding edge identifier list field; the segment confidence field is obtained by aggregating the edge probability elements in the corresponding edge identifier list field. The candidate segment set is sorted according to the segment confidence field and then written into the fault segment location information.

[0042] In one possible implementation, the master station collaborative analysis device constructs graph structure data based on the distribution network topology data. This graph structure data includes a set of nodes and a set of edges. Nodes correspond to monitoring points, and edges correspond to topological connections defined by switching equipment or line segments in the distribution network. The master station collaborative analysis device assigns a unique node identifier to each node and a unique edge identifier to each edge. A one-to-one mapping relationship is established between these edge identifiers and the switching equipment identifiers or line segment identifiers in the topology graph data.

[0043] 1) Node feature composition The master station collaborative analysis device writes standardized fault feature vectors from monitoring points into the node feature fields of the corresponding nodes. The node feature fields include: fault nature confidence field, fault direction field, arc state field, event timestamp field, and deep feature encoding field.

[0044] The main station collaborative analysis device writes missing marker fields and node validity mask fields to nodes that have not sent standardized fault feature vectors within the same event aggregation period. The node validity mask field takes a first state to indicate that the node feature is valid, and a second state to indicate that the node feature is missing. The main station collaborative analysis device writes a time alignment difference field into the node feature field to characterize the deviation of the node's event timestamp field from the aggregated reference time in the main station collaborative analysis device's event buffer.

[0045] In this embodiment, the node validity mask field is an enumerated field, with the first state and the second state corresponding to valid and missing, respectively. In one implementation, the first state is written as 1, and the second state is written as 0. The missing flag field is a missing identifier field consistent with the node validity mask field. The missing flag field is written into the node feature field and maintains consistency constraints with the node validity mask field: when the node validity mask field is 0, the corresponding fault nature confidence field, fault direction field, arc state field, event timestamp field, and deep feature encoding field are written into the missing flag data; when the node validity mask field is 1, the above fields are written into the corresponding field values ​​in the received standardized fault feature vector. The arc state field is an enumerated field. In one implementation, arcing is written as 1, arc extinguishing as 0, and uncertainty as 2. The fault direction field is an enumerated field. In one implementation, along the reference topology direction is written as 1, against the reference topology direction as 0, and unknown as 2.

[0046] In this embodiment, the aggregated reference time is the reference time marker under the same event identifier field within the event buffer. The aggregated reference time can be the earliest event timestamp field under the event identifier field or the median timestamp field under the event identifier field. The time alignment difference field is the difference marker field between the event timestamp field and the aggregated reference time, and the unit of the difference is milliseconds or seconds. When constructing the graph structure data, the main station collaborative analysis device writes the time alignment difference field into the node feature field. When the absolute value of the time alignment difference field exceeds the preset alignment difference threshold, the corresponding node validity mask field retains its original value and an alignment anomaly marker field is written into the node feature field. The alignment anomaly marker field is stored in association with the event identifier field.

[0047] 2) Edge feature construction and switch state on / off encoding The master station collaborative analysis device constructs edge feature fields for each edge based on the topology graph data. These edge feature fields include: a topology connection type field, a switch status code field, and a topology update timestamp field. The topology connection type field distinguishes between topology connections defined by switching equipment and those defined by line segments. The switch status code field characterizes the conduction status of the edge in the current topology version. A first code indicates the edge is in a conducting state, and a second code indicates it is in a disconnected state. When the topology connection type field represents a line segment, the switch status code field is written with either a conducting state code or a conducting state code derived from the combined states of the switches associated at both ends of the line segment. The master station collaborative analysis device retains only edges with a conducting state code in the adjacency relationship of the graph structure data. Edges with a disconnected state code are marked as non-propagating edges and masked from the adjacency relationship. In this embodiment, the switch status code field is an enumerated field, with the first and second codes corresponding to conducting and disconnected states, respectively. In one implementation, the first code is written as 1, and the second code is written as 0. The edge state field and the switch state encoding field use the same enumeration method. When the edge state field represents an open state, the switch state encoding field is written as 0; when the edge state field represents a closed state, the switch state encoding field is written as 1. The topology update timestamp field is a timestamp field generated by the topology graph data source or recorded by the master station collaborative analysis device. The topology version identifier is a field that identifies the topology state after a topology snapshot or a topology update. In one implementation, the topology version identifier is generated by combining the topology update timestamp field and the version number field of the topology graph data source. A second AI model inference cycle is defined by the inference trigger time and the inference end time. The master station collaborative analysis device reads the current topology version identifier at the inference trigger time and keeps it unchanged within the inference cycle. When the topology update timestamp field changes, the changed topology version identifier takes effect in the next inference cycle.

[0048] 3) Topology dynamic update mechanism The main station collaborative analysis device obtains topology update events or periodic topology snapshots from the topology graph data source, generates a topology version identifier, and records the topology update timestamp field. When a change in switch status, topology connection, or addition / deletion of monitoring points is detected, the main station collaborative analysis device updates the switch status encoding field and topology update timestamp field of the corresponding edge, and performs incremental reconstruction on the affected node set and edge set. Within one second AI model inference cycle, the main station collaborative analysis device locks a single topology version identifier, ensuring that the node set, edge set, and adjacency relationships remain consistent within that inference cycle. The graph structure data after the topology version identifier change takes effect in the next inference cycle.

[0049] 4) Propagation method of missing node mask When the second AI model executes message passing, the main station's collaborative analysis device inputs the node validity mask field as part of the node's features into the model. During edge-level message calculation, the model suppresses message components from missing nodes, preventing nodes with a validity mask field in the second state from outputting valid messages to their neighboring nodes. For edges with a switch state encoding field in the open state encoding, the model does not generate message passing along that edge. Thus, the impact of missing nodes is limited to their local neighborhood, and the topological isolation formed by open edges remains consistent throughout the message passing process.

[0050] 5) Output of the fault probability set of each edge and segment connectivity constraint rules The second artificial intelligence model infers from the graph structure data and outputs a fault probability set corresponding one-to-one with each edge identifier. This fault probability set includes the edge identifier and the fault probability value for each edge. When the main station collaborative analysis device determines the fault segment location information based on the fault probability set, it executes connectivity constraint rules: Edges whose switch status encoding field is "off" are removed from the candidate set; A candidate edge set is generated based on the fault probability set, and the connected components of the candidate edge set are calculated on the adjacency relationship. When multiple connected components coexist, the connected component with the highest correlation to high-value nodes in the fault nature confidence field of node features is selected as the fault connected subgraph; when the correlation is the same, the connected component with the largest fault probability value aggregation result is selected as the fault connected subgraph; in this embodiment, the fault probability value aggregation result is the aggregated value obtained by performing an aggregation operator on the fault probability values ​​of each edge within the connected component; the aggregation operator can be one of the summation operator, mean operator, or maximum value operator, and is fixed in the configuration parameters of the master station collaborative judgment device. The correlation is the number of correlations between the connected component and the high-value node set. The master station collaborative judgment device determines the high-value node set based on the fault nature confidence field in the node features; in one implementation, the high-value node set consists of nodes that meet a preset confidence threshold, and the preset confidence threshold is associated with and stored with the fault category identifier. When the main station collaborative analysis device calculates the correlation degree of a connected component, it counts the number of nodes in the endpoint nodes of the connected component that belong to the set of high-value nodes to obtain the correlation degree of the connected component. When multiple connected components have the same correlation degree and the same aggregation result of fault probability values, the main station collaborative analysis device selects the connected component with the larger number of edges corresponding to the maximum edge fault probability value and writes the edge identifier list field corresponding to the selected connected component into the candidate segment element. Map the fault connectivity subgraph back to the switching equipment or line segment in the topology graph data to determine the boundary switch set of the fault segment; the boundary switch set is determined by the switching equipment corresponding to the topology connection that is adjacent to the fault connectivity subgraph and is in an open state or does not belong to the fault connectivity subgraph. The fault location information is expressed as a topological interval defined by the set of boundary switches and is output in association with the fault nature classification information.

[0051] This implementation transforms topology graph data into a graph structure composed of nodes and edges, and incorporates standardized fault feature vectors from monitoring points into node features. This provides the inference objects of the master station collaborative analysis device with a unified carrier of event semantics and topology semantics. In addition to confidence level, direction, arc state, timestamp, and depth encoding, node features introduce node validity masks and time alignment differences to handle information loss and temporal inconsistencies caused by monitoring points not uploading or uploading delays. Edge features record topology connection types, switch state encodings, and topology update times, enabling explicit expression of edge connectivity and disconnection in adjacency relationships. Topology version locking maintains structural consistency within a single inference cycle. After the model inference outputs the fault probability set for each edge, the master station collaborative analysis device performs segment merging based on connected components as constraints, eliminating disconnected edges and combining high-confidence node correlation and probability aggregation to complete the selection of connected subgraphs. This is then reflected back to switches and lines to determine segment boundaries, thus forming segment location results consistent with the topology state.

[0052] In this embodiment, any of the edge feature extraction devices is further configured to: When the event triggering conditions are met, multiple standardized fault feature vectors are generated within a preset time window and a vector sequence is formed. The vector sequence and the corresponding sequence index are sent to the main station collaborative analysis device. The main station collaborative analysis device is also configured to perform time-series pattern analysis on the vector sequence, output time-series classification results related to the arc state, and write the fault nature classification information. The preset time window is determined by the window length and step size. The edge feature extraction device adaptively determines the window length and step size used for this vector sequence generation from the window length candidate set and the step size candidate set based on the fault nature confidence field and arc state field in the standardized fault feature vector corresponding to the trigger point. The sequence index includes at least a monitoring point identifier field, a sequence identifier field, a trigger point timestamp field, a window length encoding field, a step size encoding field, and a sequence number field within the sequence; The main station collaborative analysis device uses a temporal convolutional network model to perform temporal pattern analysis on the vector sequence, outputs a temporal classification result including an arc time series category field and a temporal analysis confidence field, and statistically obtains the arcing cycle field, arcing duty cycle field, and number of intermittent fields based on the arc state field and event timestamp field of each vector in the vector sequence. The main station collaborative analysis device writes the arc timing category field, timing analysis confidence field, arcing cycle field, arcing duty cycle field, and number of intermittent fields into the fault nature classification information, and writes the alarm storm identifier field into the fault nature classification information when the number of intermittent fields and the arcing duty cycle field meet the alarm storm criterion. In one implementation, the preset time window is defined by a window start time and a window end time, and the vector sequence contains multiple sequence elements; each sequence element contains a sequence index field and a corresponding standardized fault feature vector.

[0053] The sequence index field is defined as an integer marker that increments sequentially within the same preset time window, and together with the event identifier field, determines the element order of the vector sequence. When the master station collaborative analysis device performs time-series pattern analysis on the vector sequence, it converts the vector sequence into a time-series input tensor, which is composed of sequentially arranged vector payload fields corresponding to the sequence index field. The time-series pattern analysis output is defined as an arc time-series classification record, which includes at least an arc segment count field, an arc extinguishing segment count field, and a segment duration list field; the arc time-series classification record and the arc state field are jointly written into the fault nature classification information.

[0054] In one possible implementation, when the event triggering condition is met, any edge feature extraction device generates multiple standardized fault feature vectors within a preset time window based on the trigger point and forms a vector sequence. The vector sequence and the corresponding sequence index are sent to the main station collaborative analysis device. The main station collaborative analysis device performs time-series pattern analysis on the vector sequence, outputs time-series classification results related to the arc state, and writes fault nature classification information.

[0055] a) Adaptive value selection rules for window length and step size The preset time window includes a candidate set of window length and a candidate set of step size. After generating the first standardized fault feature vector at the trigger point, the edge feature extraction device reads the fault nature confidence field and the arc state field from the vector, and determines the window length and step size used for generating the vector sequence according to the adaptive value selection rule. In this embodiment, both the candidate set of window length and the candidate set of step size are discrete sets, and the elements of the sets are coded discrete positions. The window length encoding field and the step size encoding field are respectively encoding fields that identify the elements in the candidate set of window length and the candidate set of step size. The mapping relationship between the window length encoding field, the step size encoding field and the elements of the sets is consistent in the edge feature extraction device and the main station collaborative judgment device. The high confidence interval and the critical interval are the interval marking caliber obtained by dividing the fault nature confidence field into intervals. In one implementation, the high confidence interval corresponds to the maximum confidence value of the fault nature confidence field being greater than or equal to the preset confidence threshold plus 0.05, and the critical interval corresponds to the maximum confidence value of the fault nature confidence field being between the preset confidence threshold and the preset confidence threshold plus 0.05. The values ​​of the candidate set for window length and the candidate set for step size can be determined based on historical sample quantile tuning or validation set tuning, and stored in the configuration parameters in association with the fault category identifier; when the fault category identifier or the line identifier is different, the corresponding candidate set for window length and the candidate set for step size can be written into different encoding mapping tables. When the fault nature confidence field is in the high confidence interval, the window length is taken as the shorter length in the candidate set of window lengths, and the step size is taken as the larger step size in the candidate set of step sizes, so that the vector sequence covers the key segments near the trigger point; When the fault nature confidence field is in the critical interval or the arc state field is characterized as an unstable state, the window length is taken as the longer length in the candidate set of window lengths, and the step size is taken as the smaller step size in the candidate set of step sizes, so that the vector sequence covers multiple arc ignition and arc extinction changes. When the edge feature extraction device repeatedly meets the event triggering conditions within a preset short period of time, the window length is increased by one candidate level based on the window length of the previous sequence, and the step size is decreased by one candidate level based on the step size of the previous sequence, in order to cover the high-frequency triggering caused by intermittent arcs; when the number of repeated triggers falls below the preset number threshold, the window length and step size are restored to the values ​​determined by the fault nature confidence field and the arc state field.

[0056] b) Trigger point definition and sequence index structure The trigger point is defined as the moment when the edge feature extraction device determines that the event triggering conditions are met, corresponding to the event timestamp field of the first standardized fault feature vector that meets the event triggering conditions. The edge feature extraction device assigns a sequence identifier field to each trigger point of the same monitoring point. The sequence index includes: monitoring point identifier field, sequence identifier field, trigger point timestamp field, window length encoding field, step size encoding field, and sequence number field.

[0057] The edge feature extraction device generates multiple standardized fault feature vectors by sliding the window length within the time range covered by the window length. The sequence number field is incremented in the order of generation. The event timestamp field of each vector records the time position of the waveform segment corresponding to the vector, so that the main station collaborative judgment device can complete the time sequence alignment based on the event timestamp field and the sequence number field.

[0058] c) Types, inputs, and outputs of time series pattern analysis models When the master station collaborative analysis device performs time-series pattern analysis on the received vector sequences, it uses a temporal convolutional network model as the time-series analysis model. The input of the temporal convolutional network model is a vector sequence feature matrix sorted by the sequence number field. Each row of the matrix corresponds to the deep feature encoding field of a standardized fault feature vector. The fault nature confidence field, fault direction field, and arc state field of the vector are concatenated to form the time-series input features. The master station collaborative analysis device calculates the sequence coverage field based on the sequence index, which is used to characterize the ratio of the actual number of received vectors to the theoretical number within the window length. The sequence coverage field is used as an additional input to the temporal convolutional network model.

[0059] The temporal convolutional network model outputs arc time series classification results, which include an arc time series category field and a time series analysis confidence field. The arc time series category field is used to characterize one of the time series modes such as continuous arcing, intermittent arcing, arc extinction recovery, and no arc; the time series analysis confidence field is used to characterize the confidence level of the time series category.

[0060] d) Determination indicators and output fields of arcing cycle characteristics The main station collaborative analysis device extracts arc cycle characteristic judgment indicators based on the output of the temporal convolutional network model and in combination with the arc state field of each vector in the vector sequence. The judgment indicators include at least the cycle indicator, the duty cycle indicator, and the number of intermittent times indicator. The periodicity index is obtained by statistically analyzing the time interval between two adjacent "arc start" states in the vector sequence. The arc start state is defined by the moment when the arc state field jumps from the non-arc state to the arc state. The time interval is calculated by the difference between the event timestamp fields of the corresponding two vectors. The duty cycle index is obtained by statistically analyzing the ratio of the number of vectors in the arcing state within the window length to the total number of vectors within the window length. The arcing state is determined by taking the arcing code from the arcing state field. The number of intermittent events is obtained by counting the number of state transitions between the arc state and the non-arc state within the window length. The number of state transitions is calculated by the number of times the arc state fields of adjacent vectors are inconsistent.

[0061] The master station collaborative analysis device writes the periodic index into the arcing period field, the duty cycle index into the arcing duty cycle field, and the number of intermittent events index into the number of intermittent events field. When the number of intermittent events index is greater than the preset number of intermittent events threshold and the duty cycle index is less than the preset duty cycle threshold, the master station collaborative analysis device writes the alarm storm identifier field as a storm status code to characterize the dense triggering caused by intermittent arcs.

[0062] In this embodiment, the preset intermittent frequency threshold and the preset duty cycle threshold are threshold fields stored in association with the alarm storm criterion, and are associated with at least one of the fault category identifier and the line identifier. When associated with the fault category identifier, different fault category identifiers correspond to different preset intermittent frequency thresholds and preset duty cycle thresholds; when associated with the line identifier, different line identifiers correspond to different preset intermittent frequency thresholds and preset duty cycle thresholds. The alarm storm identifier field is an enumerated field. In one implementation, the storm state code is written as 1, and the non-storm state code is written as 0. The master station collaborative analysis device writes the alarm storm identifier field as 1 when the intermittent frequency index and the duty cycle index meet the alarm storm criterion, and writes it as 0 otherwise. As shown in Table 1.

[0063] Table 1: Relationship between relevant fields and storage key-value pairs for alarm storm judgment criteria Preset interval number threshold Numeric fields Fault Category Identifier / Line Identifier Comparison with the number of intervals field Preset duty cycle threshold Numeric fields Fault Category Identifier / Line Identifier Comparison with the arc duty cycle field Alarm Storm Identifier Field Enumeration fields Event Identifier Field 0 or 1 e) Field structure for writing fault nature classification information The main station collaborative analysis device writes the arc timing category field, timing analysis confidence field, arc cycle field, arc duty cycle field, number of intermittent fields, alarm storm identifier field, and sequence coverage field into the fault nature classification information, so that the fault nature classification information simultaneously includes the fault nature classification conclusions output by the graph neural network model and the arc-related timing conclusions obtained by timing pattern analysis.

[0064] This implementation establishes an executable data structure and discriminative output around the generation of vector sequences after triggering and the timing analysis of the main station collaborative judgment device. The edge side organizes the sequence based on the trigger point and sends the sequence identifier, trigger point time, window length and step size encoding, and sequence number together, enabling the main station collaborative judgment device to complete alignment and tracing. The window length and step size are adaptively selected from the candidate set, adjusting the coverage and sampling density according to confidence level and arc stability, and expanding coverage in the case of short-term repeated triggering to adapt to the dense reporting caused by intermittent arcs. The main station collaborative judgment device uses a temporal convolutional network to perform temporal modeling of deep features and key fields, outputting the arc timing category and corresponding confidence level. Simultaneously, it uses the state transition statistical period, duty cycle, and number of intermittent events to form quantitative indicators for the characteristics of the arcing cycle. These results are written into the fault nature classification information in the form of fields. The alarm storm state is indicated by both high number of intermittent events and low duty cycle, thus mapping intermittent arcs and alarm storms to clear sequence organization, model output, and quantitative indicators.

[0065] In this embodiment, the event triggering condition includes: when the fault nature confidence field is greater than a preset confidence threshold and the fault direction field meets a preset direction criterion, the edge feature extraction device starts the vector sequence generation and transmission of the preset time window; In one implementation, the preset confidence threshold is configured as a parameter on the master station collaborative analysis device side or an edge side configuration parameter, and is stored in association with the fault category identifier; when there are multiple elements in the category confidence list, the determination of the fault nature confidence field is based on the fault category identifier corresponding to the maximum confidence value.

[0066] The preset direction criterion is defined as a set constraint rule for the values ​​of the fault direction field, which includes a set of allowed values ​​and a set of prohibited values. After the edge feature extraction device starts the generation and transmission of the vector sequence, it writes the event identifier field into the local event status record; when the same event identifier field exists in the local event status record and is in a preset active state, the corresponding newly generated standardized fault feature vector is written into the same vector sequence.

[0067] In this embodiment, any of the edge feature extraction devices includes a data preprocessing unit, which is configured to perform at least two of the following before the electrical quantity waveform data is input into the first artificial intelligence model: noise filtering, amplitude normalization, resampling, and fixed-length truncation; In one embodiment, the noise filtering includes at least one of bandpass filtering and notch filtering, and the filtered waveform data is stored in association with the original sampling rate. The amplitude normalization is defined as a process of linearly scaling the waveform data according to a preset scale, which is determined by historical steady-state sample statistics or device range parameters.

[0068] The resampling is defined as the process of converting waveform data to a preset target sampling rate, and generating a new sampling sequence index after resampling. The fixed-length truncation is defined as truncating a sample point sequence of a preset length based on a truncation anchor point. The truncation anchor point is taken from the sample point number corresponding to the event trigger time or from the sample point number corresponding to the time when the zero-sequence mutation exceeds a preset mutation threshold.

[0069] Example 2: Based on Example 1, such as Figure 2 As shown, this embodiment provides a power data management method applied to a collaborative analysis scenario between distribution network monitoring points and a master station collaborative analysis device, including: Step S1: Acquire local electrical quantity waveform data at multiple monitoring points and perform preprocessing; Step S2: Input the preprocessed electrical quantity waveform data into the first artificial intelligence model to generate a standardized fault feature vector; Step S3: Send the standardized fault feature vector to the main station collaborative analysis device; Step S4: The main station collaborative analysis device receives standardized fault feature vectors from multiple monitoring points and obtains distribution network topology data; Step S5: Construct the topology graph data into graph structure data and write the standardized fault feature vector into the corresponding node features; Step S6: Input the graph structure data into the second artificial intelligence model and output the fault section location information and fault nature classification information; In one implementation, generating a standardized fault feature vector includes: generating a vector header field and filling it with a monitoring point identifier field, an event identifier field, and an event timestamp field; generating a vector payload field and filling it with at least two of a fault nature confidence field, a fault direction field, an arc state field, and a deep feature code.

[0070] The process of sending the standardized fault feature vector to the master station collaborative analysis device includes encapsulating the standardized fault feature vector into a message record. The message record contains at least a message sequence number field and a verification field. After performing a consistency check on the verification field, the master station collaborative analysis device writes the standardized fault feature vector into an event buffer. In this embodiment, the message sequence number field is an incrementing sequence number field within the same monitoring point according to the sending order, and the verification field is a verification value field obtained by performing a consistency check on the message record payload. The verification field can be generated using a cyclic redundancy check category or a hash check category. When the consistency check fails, the master station collaborative analysis device writes the corresponding message record into an invalid record set and discards the vector payload of that message record. The invalid record set is stored in association with the event identifier field and includes a monitoring point identifier field, a message sequence number field, and a verification failure flag field. When the consistency check passes, the master station collaborative analysis device writes the standardized fault feature vector into the event buffer and writes an arrival timestamp field into the event index table. The arrival timestamp field is a local timestamp indicating when the master station collaborative analysis device receives the message record, used to form a time reference independent of the event timestamp field in the determination of exceeding a preset waiting time threshold.

[0071] The process of constructing graph structure data from topological graph data includes writing edge state fields into edge feature fields; when the edge state field represents a disconnected state, the corresponding edge is marked as an unavailable edge in the graph structure data.

[0072] In this embodiment, an event triggering judgment is performed before the data is sent to the main station collaborative analysis device; When the confidence field of the fault nature in the standardized fault feature vector is greater than the preset confidence threshold, a vector sequence is generated within a preset time window and sent to the main station collaborative judgment device. The main station collaborative judgment device updates the fault nature classification information based on the vector sequence. In one implementation, generating a vector sequence within a preset time window includes writing a window identifier field to the vector sequence, the window identifier field being stored in association with the event identifier field; the window identifier field is defined as a marker indicating the time window to which a vector sequence of the same event belongs.

[0073] The main station collaborative analysis device updates the fault nature classification information based on the vector sequence by writing the arc timing classification record into the extended field of the fault nature classification information, and retaining the vector sequence index range field associated with the event identifier field.

[0074] In this embodiment, when the master station collaborative analysis device receives the standardized fault feature vector, it writes the vectors from different monitoring points into the event cache area according to the event timestamp field; When the event buffer meets the preset aggregation conditions, the graph structure data is constructed and the reasoning of the second artificial intelligence model is executed; the preset aggregation conditions include at least one of: reaching a preset threshold for the number of monitoring points or exceeding a preset waiting time threshold; In one implementation, the event cache includes an event index table and a vector storage area; the index key of the event index table includes an event identifier field and a monitoring point identifier field, and the vector storage area stores the corresponding standardized fault feature vectors according to the index key. Reaching a preset monitoring point quantity threshold is defined as the number of different monitoring point identifier fields under the same event identifier field in the event index table reaching a preset quantity threshold; exceeding a preset waiting time threshold is defined as the difference between the earliest event timestamp field corresponding to the same event identifier field and the current time of the master station collaborative analysis device exceeding a preset time threshold.

[0075] The construction of the graph structure data and the execution of inference include: reading a set of standardized fault feature vectors that meet the aggregation conditions from the event cache; writing nodes for which no standardized fault feature vectors were read into a missing marker field; and inputting the missing marker field as part of the node feature field into the second artificial intelligence model. After inference is completed, the event cache generates an evidence record, which includes at least an event identifier field, a list of monitoring point identifiers participating in the aggregation, a set of candidate segments, and fault nature classification information.

[0076] Example 3: This embodiment provides an application in a complex multi-branch distribution network scenario where a single-phase ground fault is accompanied by intermittent arcing. The distribution network contains multiple monitoring points, and edge feature extraction devices are installed at each monitoring point. A master station collaborative analysis device is installed on the master station side and is communicatively connected to each edge feature extraction device. During the operation of the distribution network, there are communication delays and arrival jitter, and the event timestamp fields generated by different monitoring points have deviations. The distribution network topology connection is affected by changes in the switching states, and the topology data source outputs topology update events or periodic topology snapshots.

[0077] In this embodiment, the master station collaborative analysis device preloads distribution network topology data and establishes a one-to-one mapping relationship between node identifiers and edge identifiers and monitoring point identifier fields, switch equipment identifiers, or line segment identifiers; the switch status encoding field and the edge status field use the same enumeration caliber. The master station collaborative analysis device pre-configures a preset monitoring point quantity threshold and a preset waiting time threshold, and enables an event index table and a vector storage area in the event cache area; the preset monitoring point quantity threshold and the preset waiting time threshold can be fixed values ​​or configuration parameters associated with line identifiers. For example, the preset monitoring point quantity threshold is 3, and the preset waiting time threshold is 2 seconds. The edge feature extraction device pre-configures a preset confidence threshold, a window length candidate set, and a step size candidate set, and maintains the mapping relationship between the window length encoding field, the step size encoding field, and the set elements consistent with that of the master station collaborative analysis device.

[0078] When a single-phase ground fault occurs in the distribution network, multiple edge feature extraction devices within the fault's influence range acquire local electrical quantity waveform data and perform preprocessing. The preprocessed electrical quantity waveform data is input into the first artificial intelligence model to generate a standardized fault feature vector containing a fault nature confidence field, a fault direction field, an arc state field, an event timestamp field, and a deep feature encoding field. The edge feature extraction devices determine the event triggering conditions; when the fault nature confidence field is greater than a preset confidence threshold and the fault direction field meets a preset direction criterion, the edge feature extraction devices generate a vector sequence based on the trigger point and send it. The vector sequence carries a sequence index, which includes a monitoring point identifier field, a sequence identifier field, a trigger point timestamp field, a window length encoding field, a step size encoding field, and a sequence number field. The edge feature extraction device determines the current window length and step size from the candidate sets of window length and step size based on the fault nature confidence field and arc state field corresponding to the trigger point, and generates multiple standardized fault feature vectors by sliding according to the step size. When the event triggering condition is repeatedly met within a preset short period of time, the edge feature extraction device changes the window length encoding field and step size encoding field according to the predetermined gear adjustment rules and continues to generate vector sequences. When communication is unavailable, the edge feature extraction device writes the generated standardized fault feature vectors and sequence indexes into the local event status record and retains them until communication is restored before sending them. For example, the local retention time is 30 seconds.

[0079] After receiving message records, the master station collaborative analysis device performs consistency checks on the verification fields and writes the standardized fault feature vectors that pass the checks into the event cache. The event index table uses the event identifier field and the monitoring point identifier field as index keys, and the vector storage area stores the corresponding standardized fault feature vectors according to the index keys and writes them into the arrival timestamp field. The master station collaborative analysis device organizes the data set of the same event in the event cache according to the event identifier field and generates an aggregated reference time based on the event timestamp field. The aggregated reference time is taken as the earliest event timestamp field or the median timestamp field under the same event identifier field. The master station collaborative analysis device calculates the time alignment difference field for each written monitoring point and writes it into the node feature field. When the absolute value of the time alignment difference field exceeds the preset alignment difference threshold, the corresponding node feature field is written into the alignment anomaly flag field. The preset alignment difference threshold is a configuration parameter on the master station side. For example, the preset alignment difference threshold is 200 milliseconds.

[0080] When the event buffer meets the preset aggregation conditions, the master station collaborative analysis device reads the current topology version identifier and locks it in the second AI model inference cycle, constructing graph structure data based on the locked topology version identifier. The master station collaborative analysis device writes the fault nature confidence field, fault direction field, arc state field, event timestamp field, and depth feature encoding field of the received nodes into the node feature field of the corresponding node; for nodes that have not submitted standardized fault feature vectors within the same event aggregation cycle, it writes a missing flag field and writes the node validity mask field as missing. The master station collaborative analysis device writes edge feature fields containing topology connection type field, switch state encoding field, and topology update timestamp field for each edge, and removes edges whose switch state encoding field indicates a disconnected state from the adjacency relationship of the graph structure data.

[0081] The main station collaborative analysis device inputs graph structure data into the second artificial intelligence model for reasoning. During message transmission, the second artificial intelligence model suppresses message components for nodes whose node validity mask field is in a missing state, and does not transmit messages for edges that have been removed from the adjacency relationship. The second artificial intelligence model outputs a fault probability set corresponding one-to-one with each edge identifier. The main station collaborative analysis device generates a candidate edge set based on the fault probability set and calculates connected components on the adjacency relationship. It then performs connectivity constraint aggregation on the connected components and generates a candidate segment set. The main station collaborative analysis device calculates the segment confidence field for each candidate segment element in the candidate segment set. The segment confidence field is obtained based on the aggregation result of the fault probability values. The aggregation operator is fixed in the main station collaborative analysis device configuration parameters as one of the summation operator, mean operator, or maximum value operator. For cases where multiple connected components coexist, the main station collaborative analysis device calculates the correlation degree and selects the faulty connected subgraph. The correlation degree is obtained by the number of nodes belonging to the high-value node set among the endpoint nodes of the connected components. The high-value node set consists of nodes whose fault nature confidence field satisfies a preset confidence threshold. The main station collaborative analysis device maps the fault connectivity subgraph back to the switch equipment identifier or line segment identifier in the topology data, determines the boundary switch set of the fault section, and outputs the fault section location information, which includes the section identifier field and the section confidence field.

[0082] During the same event processing, the master station collaborative analysis device performs time-series pattern analysis on vector sequences from one or more monitoring points. The master station collaborative analysis device converts the vector sequences sorted by their sequence number field into time-series input features. These features are formed by concatenating a deep feature encoding field with a fault nature confidence field, a fault direction field, and an arc state field. A sequence coverage field is calculated based on the sequence index and used as additional input. The master station collaborative analysis device uses a time-series convolutional network model to output an arc time-series category field and a time-series analysis confidence field. Based on the arc state field and event timestamp field in the vector sequence, it statistically analyzes the arcing period field, arcing duty cycle field, and number of intermittent fields. The arcing period field is obtained from the difference in event timestamp fields between two adjacent arcing start states. The arcing duty cycle field is obtained from the ratio of the number of arcing state vectors within the window length to the total number of vectors. The number of intermittent fields is obtained from the number of state transitions between arcing and non-arcing states within the window length. The master station collaborative analysis device applies alarm storm criteria to the intermittent number field and the arcing duty cycle field. When the comparison relationship between the preset intermittent number threshold and the preset duty cycle threshold is met, the alarm storm identifier field is written as a storm status code. The preset intermittent number threshold and the preset duty cycle threshold are associated with at least one of the fault category identifier and line identifier, and stored in the configuration parameters. The master station collaborative analysis device writes the arc timing category field, timing analysis confidence field, arcing cycle field, arcing duty cycle field, intermittent number field, alarm storm identifier field, and sequence coverage field into the fault nature classification information, and outputs it in association with the fault section location information.

[0083] After inference is completed, the main station collaborative analysis device generates a record associated with the event identifier field. The record includes the event identifier field, a list of monitoring point identifiers participating in the aggregation, a set of candidate segments, fault segment location information, and fault nature classification information, and retains the topology version identifier and topology update timestamp fields. The record is archived in the event cache according to the event identifier field and includes a message sequence number field and an arrival timestamp field for traceability.

[0084] In this embodiment, under the conditions of communication delay and arrival jitter, the event buffer organizes a standardized fault feature vector set according to preset aggregation conditions and locks the topology version identifier to construct the graph structure data. The fault probability set output by the second artificial intelligence model is aggregated by connectivity constraints to form a candidate segment set and generate fault segment location information. Under the condition that the arc state changes with time, the vector sequence outputs the arc time series category field through the temporal convolutional network model and counts the arc cycle field, arc duty cycle field and intermittent number field. The alarm storm criterion completes the writing of the alarm storm identifier field. The fault nature classification information and the fault segment location information are associated and output under the same event identifier field and written into the evidence record.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0086] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. An artificial intelligence-based electric power data management system, characterized by, The application comprises: a plurality of edge feature extraction devices respectively arranged at a plurality of monitoring points of a power distribution network; and a main station cooperative research and judgment device in communication connection with the plurality of edge feature extraction devices; any of the edge feature extraction devices is configured to acquire local electrical quantity waveform data, input a first artificial intelligence model after pre-processing the electrical quantity waveform data, generate a standardized fault feature vector and send it to the main station cooperative research and judgment device; the main station cooperative research and judgment device is configured to receive the standardized fault feature vector from the plurality of monitoring points, acquire power distribution network topology graph data, construct the topology graph data into graph structure data and write the standardized fault feature vector into the corresponding node feature, input a second artificial intelligence model, and output fault section positioning information and fault property classification information.

2. The system of claim 1, wherein, The first artificial intelligence model is a convolutional neural network model, and the model parameter quantity does not exceed a preset parameter upper limit; the standardized fault feature vector at least includes a fault property confidence field, a fault direction field, an arc state field and an event timestamp field.

3. The system of claim 2, wherein, The standardized fault feature vector also includes a deep feature encoding output by the convolutional neural network model, and the dimension of the deep feature encoding is less than the sample point number of the electrical quantity waveform data after pre-processing.

4. The system of claim 1, wherein, The second artificial intelligence model is a graph neural network model; the nodes of the graph structure data correspond to the monitoring points, and the edges correspond to the topology connections defined by the switching devices or line sections in the power distribution network; the graph neural network model outputs a fault probability set of each edge, and the main station cooperative research and judgment device determines the fault section positioning information according to the fault probability set; wherein, when constructing the graph structure data, the main station cooperative research and judgment device writes the fault property confidence field, the fault direction field, the arc state field, the event timestamp field and the deep feature encoding in the standardized fault feature vector into the node feature of the corresponding node, and writes a node validity mask field into the node which does not receive the standardized fault feature vector; a switching state encoding field is written into the edge feature of each edge of the graph structure data, and the switching state encoding field is used to represent the on state or off state of the corresponding topology connection; the main station cooperative research and judgment device performs dynamic update on the graph structure data based on a topology update timestamp field and generates a topology version identifier, and locks a single topology version identifier within one second artificial intelligence model inference period; the second artificial intelligence model shields message passing for the edges represented as off state by the switching state encoding field during inference, and suppresses message passing of the missing nodes according to the node validity mask field; the main station cooperative research and judgment device performs connectivity constraint aggregation on the edges in the on state according to the fault probability set, determines the connectivity subgraph corresponding to the fault section and outputs the fault section positioning information.

5. The system of claim 1, wherein, any of the edge feature extraction devices is further configured to: when the event trigger condition is met, generate a plurality of standardized fault feature vectors within a preset time window and form a vector sequence, and send the vector sequence and the corresponding sequence index to the main station cooperative research and judgment device; The master station cooperative research and judgment device is further configured to perform time sequence pattern analysis on the vector sequence, output a time sequence classification result related to the arc state, and write the time sequence classification result into the fault property classification information. The preset time window is determined by a window length and a step size, and the edge feature extraction device adaptively determines the window length and the step size used for generating the current vector sequence from a window length candidate set and a step size candidate set according to a fault property confidence field and an arc state field in the standardized fault feature vector corresponding to the trigger point. The sequence index at least includes a monitoring point identification field, a sequence identification field, a trigger point timestamp field, a window length encoding field, a step size encoding field, and a sequence internal serial number field. The master station cooperative research and judgment device uses a time sequence convolution network model to perform time sequence pattern analysis on the vector sequence, outputs a time sequence classification result including an arc time sequence category field and a time sequence analysis confidence field, and obtains a burning arc period field, a burning arc duty cycle field, and an intermittent number field based on the arc state field and the event timestamp field of each vector in the vector sequence. The master station cooperative research and judgment device writes the arc time sequence category field, the time sequence analysis confidence field, the burning arc period field, the burning arc duty cycle field, and the intermittent number field into the fault property classification information, and writes an alarm storm identification field into the fault property classification information when the intermittent number field and the burning arc duty cycle field satisfy an alarm storm criterion.

6. The system of claim 5, wherein, The event trigger condition includes: when the fault property confidence field is greater than a preset confidence threshold, and the fault direction field satisfies a preset direction criterion, the edge feature extraction device starts the vector sequence generation and sending of the preset time window.

7. The system of claim 1, wherein, Any of the edge feature extraction devices includes a data preprocessing unit configured to perform at least two of noise filtering, amplitude normalization, resampling, and fixed-length truncation before the electrical quantity waveform data is input into the first artificial intelligence model.

8. An artificial intelligence-based power data management method based on the system of any one of claims 1-7, applied to the collaborative research and judgment scene of the monitoring point of the distribution network and the master station, characterized in that, It includes: Local electrical quantity waveform data is obtained at multiple monitoring points and preprocessed; The preprocessed electrical quantity waveform data is input into a first artificial intelligence model to generate a standardized fault feature vector; The standardized fault feature vector is sent to a master station cooperative research and judgment device; The master station cooperative research and judgment device receives the standardized fault feature vector from multiple monitoring points and obtains power distribution network topology graph data; The topology graph data is constructed into graph structure data, and the standardized fault feature vector is written into the corresponding node feature; The graph structure data is input into a second artificial intelligence model to output fault section positioning information and fault property classification information.

9. The method of claim 8, wherein, Before being sent to the master station cooperative research and judgment device, an event trigger judgment is performed; When the fault property confidence field in the standardized fault feature vector is greater than a preset confidence threshold, a vector sequence is generated within a preset time window and sent to the master station cooperative research and judgment device, and the master station cooperative research and judgment device updates the fault property classification information based on the vector sequence.

10. The method of claim 8, wherein, When the master station collaborative analysis device receives the standardized fault feature vector, it writes the vectors from different monitoring points into the event cache area according to the event timestamp field. When the event buffer meets the preset aggregation conditions, the graph structure data is constructed and the reasoning of the second artificial intelligence model is executed; the preset aggregation conditions include at least one of: reaching a preset monitoring point number threshold or exceeding a preset waiting time threshold.

Images