Power grid fault analysis method and system
Patent Information
- Application Number
- CN202611079562.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-08-18
AI Technical Summary
[0003]数据融合能力差：难以同时处理高频时序数值（如电压、电流）与非结构化文本日志（如告警信息、操作记录），导致大量语义信息丢失
[0048] 1. This application utilizes a time-series large model to map operation and maintenance data to a high-dimensional semantic space, combines static topology with real-time high-dimensional embedding vectors to construct a dynamic spatiotemporal graph, fully explores the deep correlation of cross-modal data, provides a unified data structure for fault propagation analysis, and improves the robustness of the model in complex scenarios.
Smart Images
Figure CN122594749A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of power system automation and artificial intelligence technology, specifically a power grid fault analysis method and system. Background Technology
[0002] With the large-scale integration of renewable energy and the increasing complexity of power grid structures, power grid operation data exhibits characteristics of massive volume, multiple sources, heterogeneity, and high dimensionality. Traditional power grid fault analysis methods mainly suffer from the following shortcomings:
[0003] Poor data fusion capability: It is difficult to process high-frequency time-series values (such as voltage and current) and unstructured text logs (such as alarm information and operation records) at the same time, resulting in the loss of a large amount of semantic information.
[0004] Separation of static topology and dynamic state: Traditional power grid topology diagrams only describe the connection relationship of equipment and cannot reflect the dynamic changes of equipment operating status in real time. They are even more difficult to capture the spatiotemporal propagation process of faults along the topology.
[0005] Cause inference is difficult: After a failure occurs, it is difficult to automatically distinguish the temporal sequence and spatial propagation path. Root cause localization relies on human experience, which is inefficient and prone to errors.
[0006] Single evidence utilization: Existing root cause localization methods are usually based only on thresholds or simple logic, failing to integrate multi-dimensional evidence such as device status deviation, topological criticality, semantic alarm level, and similarity of historical cases.
[0007] Therefore, there is an urgent need for an intelligent method that can uniformly process multimodal data, dynamically characterize the spatiotemporal propagation of faults, and integrate multi-source evidence for quantitative root cause localization. Summary of the Invention
[0008] To overcome the shortcomings of existing technologies, this invention provides a power grid fault analysis method and system. This method can achieve automated, high-precision, and interpretable root cause localization of power grid faults, shorten power outage time, and improve the reliability of power grid operation.
[0009] To achieve the above objectives, the technical solution adopted by the present invention is: a power grid fault analysis method, comprising:
[0010] Obtain the static topology diagram of the power grid;
[0011] Acquire operation and maintenance data of various devices in the power grid;
[0012] Input the operation and maintenance data into the pre-trained time series large model and output a high-dimensional embedding vector for each device node at time t.
[0013] Based on the static topology graph, a dynamic spatiotemporal graph is constructed by inputting high-dimensional embedding vectors into the corresponding device nodes.
[0014] The dynamic spatiotemporal graph is input into a preset spatiotemporal graph neural network model to obtain a weighted directed graph of fault propagation. The spatiotemporal graph neural network is used to infer the order of faults and the propagation influence weights along the topological edges.
[0015] Based on the weighted directed graph of fault propagation and the first preset threshold of the propagation influence weight, a fault propagation chain graph is obtained;
[0016] Based on the fault propagation chain graph, the root cause confidence level is calculated for nodes with an in-degree of 0 on the propagation chain.
[0017] Through the above technical solution, the power grid fault analysis method disclosed in this application integrates a time-series large model and a graph neural network. It uses the time-series large model to map operation and maintenance data to a high-dimensional semantic space, combines static topology with real-time high-dimensional embedding vectors to construct a dynamic spatiotemporal graph, fully explores the deep correlation of cross-modal data, provides a unified data structure for fault propagation analysis, and improves the robustness of the model in complex scenarios.
[0018] Furthermore, the step of inputting the dynamic spatiotemporal graph into a preset spatiotemporal graph neural network model to obtain a weighted directed graph of fault propagation includes:
[0019] Real-time monitoring of abnormal scores of high-dimensional embedding vectors of each device node; if an abnormal score is detected to exceed a second preset threshold, in the dynamic spatiotemporal graph, starting from the device node, all device nodes within a preset number of hops are extracted to form a local subgraph.
[0020] The local subgraph is input into a preset spatiotemporal graph neural network model, and spatiotemporal graph neural network inference is performed to obtain a weighted directed graph of fault propagation.
[0021] Furthermore, the spatiotemporal graph neural network inference includes:
[0022] The spatiotemporal graph neural network model traces back the event windows before and after the fault occurs, obtains the evolution sequence of the historical state of each device node, confirms the order in which the faults occurred, and obtains the directed graph of fault propagation.
[0023] Furthermore, the spatiotemporal graph neural network inference also includes:
[0024] Based on the fault propagation directed graph, the propagation influence weights of the fault propagation along the topological edges are calculated to obtain the fault propagation weighted directed graph.
[0025] Furthermore, the calculation of root cause confidence for nodes with an in-degree of 0 on the propagation chain based on the fault propagation chain graph includes:
[0026] In the fault propagation chain graph, the fault propagation chain is obtained by traversing from the node with an in-degree of 0 along the directed edge to the node with an out-degree of 0.
[0027] Calculate the state deviation score, topological criticality score, semantic alarm score, and prior matching score of the source node in the fault propagation chain, respectively.
[0028] The root cause confidence of the source node is obtained by weighted summation of the state deviation score, topological criticality score, semantic alarm score, and prior matching score.
[0029] Furthermore, after acquiring the operation and maintenance data of each device in the power grid, the method further includes converting the operation and maintenance data into basic units with the same structure. The operation and maintenance data includes at least time-series indicators and device logs. The conversion of the operation and maintenance data into basic units with the same structure includes:
[0030] The time-series indicators are segmented and vectorized to obtain the numerical token of the first dimension;
[0031] The device logs are semantically encoded to obtain a text token of the first dimension;
[0032] The numerical tokens and text tokens are concatenated into a unified token sequence in chronological order.
[0033] Furthermore, before converting the operation and maintenance data into basic units with the same structure, step data preprocessing is also included. The data preprocessing includes at least cleaning, deduplication, and timestamp alignment. The cleaning step includes at least removing data with format errors, missing fields, or values exceeding the range.
[0034] The deduplication step includes removing duplicate data caused by retransmission based on packet identifiers;
[0035] The timestamp alignment step involves aligning data from different sources to the same timeline based on the original timestamps of the data itself, using a watermarking mechanism.
[0036] Furthermore, after calculating the root cause confidence level for nodes with an in-degree of 0 on the propagation chain based on the fault propagation chain graph, the system also includes outputting a diagnostic report containing the fault propagation chain graph and the root cause confidence level.
[0037] A power grid fault analysis system, comprising:
[0038] The topology acquisition module is used to acquire the static topology diagram of the power grid;
[0039] The data access and processing module is used to acquire operation and maintenance data of various devices in the power grid;
[0040] The Time Series Large Model module is used to input operation and maintenance data into a pre-trained time series large model and output a high-dimensional embedding vector for each device node at time t.
[0041] The dynamic spatiotemporal graph construction module is used to construct a dynamic spatiotemporal graph by inputting high-dimensional embedding vectors into the corresponding device nodes based on the static topology graph.
[0042] The spatiotemporal graph neural network inference module is used to input a dynamic spatiotemporal graph into a preset spatiotemporal graph neural network model to obtain a weighted directed graph of fault propagation. The spatiotemporal graph neural network is used to infer the sequence of faults and the propagation influence weights along the topological edges.
[0043] The fault propagation chain graph construction module is used to obtain the fault propagation chain graph based on the weighted directed graph of fault propagation and the first preset threshold of the propagation influence weight.
[0044] The scoring module is used to calculate the root cause confidence level for nodes with an in-degree of 0 on the fault propagation chain graph.
[0045] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described power grid fault analysis method.
[0046] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described power grid fault analysis method.
[0047] By employing the above technical solutions, the beneficial effects of the present invention are as follows:
[0048] 1. This application utilizes a time-series large model to map operation and maintenance data to a high-dimensional semantic space, combines static topology with real-time high-dimensional embedding vectors to construct a dynamic spatiotemporal graph, fully explores the deep correlation of cross-modal data, provides a unified data structure for fault propagation analysis, and improves the robustness of the model in complex scenarios.
[0049] 2. This application breaks through the limitation of traditional diagnosis relying on a single chain of evidence by using a multi-source evidence fusion root cause credibility scoring mechanism, providing maintenance personnel with a clear fault evolution path and intuitive root cause probability, and shortening the average fault location time.
[0050] To make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0052] Figure 1 This is a flowchart of the power grid fault analysis method in an embodiment of the present invention. Detailed Implementation
[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0054] It should be noted that in the description of this invention, the terms "first," "second," etc., are used only for descriptive purposes and to distinguish similar objects; there is no order between them, nor should they be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means two or more.
[0055] Example: This example discloses a power grid fault analysis method, including the following steps:
[0056] Step S10: Obtain the static topology diagram of the power grid;
[0057] Step S20: Obtain operation and maintenance data of each device in the power grid;
[0058] Step S30: Input the operation and maintenance data into the pre-trained time series large model and output the high-dimensional embedding vector of each device node at time t.
[0059] Step S40: Based on the static topology graph, input the high-dimensional embedding vector into the corresponding device node to construct a dynamic spatiotemporal graph;
[0060] Step S50: Input the dynamic spatiotemporal graph into a preset spatiotemporal graph neural network model, perform spatiotemporal graph neural network inference, and obtain a weighted directed graph of fault propagation. The spatiotemporal graph neural network inference includes at least inferring the order of faults and the propagation influence weights along the topological edges.
[0061] Step S60: Remove edges in the weighted directed graph of fault propagation where the propagation influence weight is lower than the first preset threshold to obtain the fault propagation chain graph.
[0062] Step S70: Based on the fault propagation chain graph, calculate the root cause confidence level for nodes with an in-degree of 0 on the propagation chain.
[0063] Step S10, obtaining the static topology diagram of the power grid, includes:
[0064] Obtain the static topology of the power grid from the power grid asset management system or topology parsing module, and represent it as an undirected or directed graph G = (V, E). Where the node set V = {v1, v2, ..., v...} n} represents equipment nodes in the power grid, including but not limited to inverters, circuit breakers, disconnect switches, transformers, busbars, transmission lines, loads, etc.; edge sets This represents the physical connection or logical dependency between devices.
[0065] Step S20, obtaining the operation and maintenance data of each device in the power grid includes:
[0066] S20.1, Data Acquisition: Data is acquired in real-time from SCADA systems, synchronous phasor measurement units, and log servers using a lightweight data agent deployed at substations or power plants, employing standard industrial IoT protocols such as MQTT and OPC-UA. Specifically:
[0067] SCADA systems can provide telemetry data such as voltage, current, active / reactive power, and frequency, as well as remote signaling data such as circuit breaker / disconnector position. The sampling frequency is usually 1~10Hz.
[0068] The synchronous phasor measurement unit can provide high-precision, high-frequency synchronous phasor data such as voltage / current amplitude, phase angle, and frequency change rate, with a sampling frequency typically of 50~100Hz;
[0069] Log servers can provide text data such as device operation logs, protection action events, alarm information, and operation records, and are event-driven.
[0070] S20.2, Data preprocessing: The multi-source heterogeneous data collected in step S20.1 is sent to a Kafka message queue cluster and stored by topic partitioning. The Apache Flink stream processing framework is used for real-time processing such as cleaning, deduplication, and timestamp alignment, including:
[0071] The cleaning process includes removing abnormal data points that are formatted incorrectly, have missing fields, or are clearly outside the reasonable range, such as negative voltage values.
[0072] The deduplication step includes using unique identifiers in data packets, such as device ID + timestamp + sequence number, to remove duplicate records caused by network retransmission;
[0073] The timestamp alignment process involves using an event time processing mechanism to align data from different sources to the same timeline based on the original timestamps carried by the data itself, using a watermarking mechanism. It should be noted that due to potential microsecond-level deviations in the time bases of different data sources or variations in network transmission latency, time sequence discrepancies between data sources can occur.
[0074] S20.3, converting the multi-source heterogeneous data processed in step S20.2 into basic units with the same structure, including:
[0075] High-frequency time-series metric tokenization treats a continuous time series as a segment. Each segment is mapped to a numerical vector (i.e., a numerical token) in the first dimension using a trainable linear projection layer or a one-dimensional convolutional layer. The continuous time series includes the most recent 10 voltage waveform sampling points within 100ms. For example, the input segment dimension is (batch, patch_size), and the output numerical token dimension is (batch, d_model), where d_model = 256 or 512.
[0076] Device log tokenization involves segmenting each log text into words and converting them into first-dimensional text tokens using a pre-trained text encoder or the embedding layer and positional encoding built into the Transformer.
[0077] Sequence concatenation combines numerical tokens and text tokens into a unified token sequence [T1, T2, ... T] in chronological order. L ], where L is the sequence length. This sequence preserves the temporal relationship and semantic information of numerical values and text.
[0078] Step S30: Input the operation and maintenance data into the pre-trained time series large model, and output the high-dimensional embedding vector of each device node at time t, including:
[0079] The pre-trained temporal large model receives the token sequence generated in step S20 and uses the network parameters obtained from pre-training to output a high-dimensional embedding vector for each device node v at the current timestamp t. , where d is the embedding dimension, ranging from 128 to 768. The high-dimensional embedding vector integrates the current operating status of the device, semantic anomaly features in the device logs, and implicit cross-modal correlations. The current operating status of the device includes normal voltage, high current, and closed switch; the semantic anomaly features in the device logs include excessively high temperature and communication interruption; the implicit cross-modal correlation can be: although there are no alarm logs at present, the numerical waveform is similar to the waveform before a certain historical fault.
[0080] The pre-trained temporal model can be built based on the PyTorch framework and the Transformer architecture. The pre-trained temporal model employs an encoder structure, with input being a token sequence in a unified format. The pre-training task of the pre-trained temporal model includes at least the following:
[0081] Temporal masking prediction randomly masks some tokens in a token sequence, requiring the model to predict the masked tokens based on the context, so that the model can learn the statistical patterns of the time series.
[0082] Cross-modal alignment forces the model to make numerical tokens and text tokens at the same time close in the latent space, thereby establishing a semantic association between numerical waveforms and text descriptions.
[0083] Step S40: Based on the static topology graph, the high-dimensional embedding vector is input into the corresponding device node to construct a dynamic spatiotemporal graph, including:
[0084] Based on the static topology graph G = (V, E) obtained in step S10, the high-dimensional embedding vectors of all nodes generated in step S30 at time t are... Organized into a state matrix The state matrix is then input into node v in the static topology graph, so that each node carries its time-varying state vector. Thus, the static graph evolves into a dynamic spatiotemporal graph G. t = (V, E, X) t );
[0085] Where V and E are fixed topologies, X t It changes over time t. This data structure provides the foundation for subsequent capture of spatial propagation of faults, which can be represented as the evolution of the correlation between the state vectors of adjacent nodes over time.
[0086] Step S50: Input the dynamic spatiotemporal graph into a preset spatiotemporal graph neural network model, perform spatiotemporal graph neural network inference, and obtain a weighted directed graph of fault propagation, including:
[0087] The high-dimensional embedding vector of each device node is monitored in real time, and its anomaly score is calculated. The anomaly score includes values based on Mahalanobis distance or KL divergence with historical normal distributions. When the anomaly score of a device node exceeds a second preset threshold θ... anomaly At that time, the spatiotemporal graph neural network inference is triggered. The second preset threshold θ anomaly A value of 3.0 to 5.0 is acceptable.
[0088] In some feasible embodiments, when the anomaly score of a certain device node exceeds a second preset threshold θ anomaly In the dynamic spatiotemporal graph obtained in step S40, starting from the device node, all device nodes within a preset number of hops are extracted to form a local subgraph. The local subgraph is then input into a preset spatiotemporal graph neural network model, and spatiotemporal graph neural network inference is performed to obtain a weighted directed graph of fault propagation. The preset number of hops can be 2 to 5.
[0089] The spatiotemporal graph neural network comprises two core components: a temporal attention mechanism and a spatial attention mechanism, wherein:
[0090] The time attention mechanism includes capturing a time window before and after the fault trigger time t0. ,in Usually, 5 to 10 seconds are taken;
[0091] For each node v, its historical state vector sequence within that time window is... Input is fed into a multi-head temporal attention layer;
[0092] The temporal attention layer calculates the importance weight of each time step in the sequence and extracts the chronological order of events. For example, the model can determine that the time step of the voltage drop at node A is earlier than the time step of the overload log at node B, thus inferring that A is the cause and B is the result, resulting in the directed edge A→B.
[0093] Spatial attention mechanisms include those based on Graph Attention Networks (GAT), where for each directed edge... (in (And there is a connection relationship), calculate the propagation influence weight from node i to node j. .
[0094] The calculation method for the weight of the transmission effect is as follows:
[0095] ;
[0096] in, Let W be the node state vector, W be the learnable weight matrix, and a be the attention vector. This represents vector concatenation. Let N be the set of neighbors of node i;
[0097] The propagation influence weight reflects the strength of the influence of the abnormal state of node i on the state of node j. If If the edge is significantly higher than other edges, it indicates a high probability that the fault will propagate from edge i to edge j.
[0098] Step S60: Remove edges in the weighted directed graph of fault propagation where the propagation influence weight is lower than a first preset threshold, resulting in a fault propagation chain graph including:
[0099] After the spatiotemporal graph neural network model completes inference on a dynamic spatiotemporal graph or a weighted directed graph of fault propagation, it obtains the propagation influence weights of all directed edges. Set the first preset threshold. The A value of 0.3 to 0.5 can be used to remove all weakly correlated edges whose propagation influence weights are lower than the first preset threshold from the dynamic spatiotemporal graph or the weighted directed graph of fault propagation. The retained strongly correlated edges constitute the fault propagation chain graph.
[0100] Step S70, based on the fault propagation chain graph, calculate the root cause confidence level for nodes with an in-degree of 0 on the propagation chain, including:
[0101] In the fault propagation chain graph, starting from a node with an in-degree of 0 (i.e., a node to which no other node points), traversing along directed edges to a node with an out-degree of 0 yields several directed paths. Each directed path constitutes a fault propagation chain, in the form of: .
[0102] The terminal node is typically a device that trips due to protective action or ultimately fails. This propagation chain diagram visually illustrates the entire process of a fault, from its source and path to its final destination.
[0103] For the source node of the obtained fault propagation chain, i.e., the node with an in-degree of 0, the system comprehensively calculates the state deviation score. Topological keyness score Semantic Alarm Score Matching score with prior knowledge The scores across four dimensions are used to obtain the root cause credibility through a weighted summation, where:
[0104] State deviation score Calculate the high-dimensional embedding vector of the source node. The degree of deviation from its historical normal distribution. Specifically, Mahalanobis distance or log-likelihood based on a Gaussian mixture model can be used. The greater the deviation, the higher the score, with the range normalized to [0, 1], where:
[0105] .
[0106] Topological keyness score Based on the generated fault propagation chain graph, the source node should have an in-degree of 0. For nodes with an in-degree of 0, If multiple nodes have an in-degree of 0, normalization can be performed based on their out-degree or influence range. For example, if there are two nodes H and Z with an in-degree of 0, where node H has an out-degree of 10 and node Z has an out-degree of 2, after normalization, node H's score is 1.0, and node Z's score is 2 / 10 = 0.2. Graph centrality metrics, such as eigenvector centrality, can also be used for fine-tuning. Eigenvector centrality considers not only the number of neighbors a node has but also the importance of those neighbors. If a node connects to many important nodes, it becomes more important. The higher.
[0107] Semantic Alert Score The system invokes large language models, such as LLaMA, ChatGLM, or GPT series, to analyze the latest log text from the source node. The large language model is prompted to evaluate the severity level of the logs, outputting categories such as fatal, severe, warning, and informational, mapped to a score of 0.9–0.1. Simultaneously, the model can extract key entities from the logs, such as overcurrent, overvoltage, and high temperature, as supplementary evidence.
[0108] Prior matching score The system performs similarity matching between the current fault's feature vector and historical fault cases stored in the operations and maintenance expert knowledge base. The fault feature vector includes node state deviation patterns, log keywords, topology location, etc. Each case in the knowledge base contains a fault feature vector and the ultimately identified root cause device. The matching degree is calculated using cosine similarity or Euclidean distance; a higher matching degree indicates a better match. The higher.
[0109] The root cause confidence R of the source node is calculated as follows:
[0110] ;
[0111] Wherein, the weight coefficients satisfy It can be set through expert experience, such as W dev W topo W sem W prior The values are 0.3, 0.3, 0.2, and 0.2 respectively, or obtained by learning from historical data using algorithms such as logistic regression. The system outputs the R values of all source nodes and sorts them from highest to lowest.
[0112] It also includes step S80, which outputs a diagnostic report containing a fault propagation chain diagram and root cause confidence levels:
[0113] The final diagnostic report is pushed to the workstations or mobile terminals of maintenance personnel. The report includes a fault propagation chain diagram, a root cause confidence list, and a summary of key evidence. Among these:
[0114] The fault propagation chain diagram is displayed in the form of a directed graph, distinguished by different colors, such as red for the source, yellow for intermediate nodes, and black for the end;
[0115] The root cause confidence list includes the root cause confidence of each candidate source node, as well as details of the four scores;
[0116] Key evidence summaries include: Inverter A: voltage sag (deviation score 0.92), log contains critical keywords, and matches historical case #1023 75%.
[0117] This application also discloses a power grid fault analysis system, including:
[0118] The topology acquisition module is used to acquire the static topology diagram of the power grid;
[0119] The data access and processing module is used to acquire operation and maintenance data of various devices in the power grid;
[0120] The Time Series Large Model module is used to input operation and maintenance data into a pre-trained time series large model and output a high-dimensional embedding vector for each device node at time t.
[0121] The dynamic spatiotemporal graph construction module is used to construct a dynamic spatiotemporal graph by inputting high-dimensional embedding vectors into the corresponding device nodes based on the static topology graph.
[0122] The spatiotemporal graph neural network inference module is used to input a dynamic spatiotemporal graph into a preset spatiotemporal graph neural network model to obtain a weighted directed graph of fault propagation. The spatiotemporal graph neural network is used to infer the sequence of faults and the propagation influence weights along the topological edges.
[0123] The fault propagation chain graph construction module is used to obtain the fault propagation chain graph based on the weighted directed graph of fault propagation and the first preset threshold of the propagation influence weight.
[0124] The scoring module is used to calculate the root cause confidence level for nodes with an in-degree of 0 on the fault propagation chain graph.
[0125] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described power grid fault analysis method.
[0126] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described power grid fault analysis method.
[0127] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
Claims
1. A power grid fault analysis method, characterized in that, include: Obtain the static topology diagram of the power grid; Acquire operation and maintenance data of various devices in the power grid; Input the operation and maintenance data into the pre-trained time series large model and output a high-dimensional embedding vector for each device node at time t. Based on the static topology graph, a dynamic spatiotemporal graph is constructed by inputting high-dimensional embedding vectors into the corresponding device nodes. The dynamic spatiotemporal graph is input into a preset spatiotemporal graph neural network model to obtain a weighted directed graph of fault propagation. The spatiotemporal graph neural network is used to infer the order of faults and the propagation influence weights along the topological edges. Based on the weighted directed graph of fault propagation and the first preset threshold of the propagation influence weight, a fault propagation chain graph is obtained; Based on the fault propagation chain graph, the root cause confidence level is calculated for nodes with an in-degree of 0 on the propagation chain.
2. The power grid fault analysis method as described in claim 1, characterized in that, The step of inputting the dynamic spatiotemporal graph into a preset spatiotemporal graph neural network model to obtain a weighted directed graph of fault propagation includes: Real-time monitoring of abnormal scores of high-dimensional embedding vectors of each device node; if an abnormal score is detected to exceed a second preset threshold, in the dynamic spatiotemporal graph, starting from the device node, all device nodes within a preset number of hops are extracted to form a local subgraph. The local subgraph is input into a preset spatiotemporal graph neural network model, and spatiotemporal graph neural network inference is performed to obtain a weighted directed graph of fault propagation.
3. The power grid fault analysis method as described in claim 2, characterized in that, The spatiotemporal graph neural network inference includes: The spatiotemporal graph neural network model traces back the event windows before and after the fault occurs, obtains the evolution sequence of the historical state of each device node, confirms the order in which the faults occurred, and obtains the directed graph of fault propagation.
4. The power grid fault analysis method as described in claim 3, characterized in that, The spatiotemporal graph neural network inference also includes: Based on the fault propagation directed graph, the propagation influence weights of the fault propagation along the topological edges are calculated to obtain the fault propagation weighted directed graph.
5. The power grid fault analysis method as described in claim 1, characterized in that, The calculation of root cause confidence for nodes with an in-degree of 0 on the propagation chain, based on the fault propagation chain graph, includes: In the fault propagation chain graph, the fault propagation chain is obtained by traversing from the node with an in-degree of 0 along the directed edge to the node with an out-degree of 0. Calculate the state deviation score, topological criticality score, semantic alarm score, and prior matching score of the source node in the fault propagation chain, respectively. The root cause confidence of the source node is obtained by weighted summation of the state deviation score, topological criticality score, semantic alarm score, and prior matching score.
6. The power grid fault analysis method as described in claim 1, characterized in that, After acquiring the operation and maintenance data of each device in the power grid, the process further includes converting the operation and maintenance data into basic units with the same structure. The operation and maintenance data includes at least time-series indicators and device logs. The conversion of the operation and maintenance data into basic units with the same structure includes: The time-series indicators are segmented and vectorized to obtain the numerical token of the first dimension; The device logs are semantically encoded to obtain a text token of the first dimension; The numerical tokens and text tokens are concatenated into a unified token sequence in chronological order.
7. The power grid fault analysis method as described in claim 6, characterized in that, Before converting the operation and maintenance data into basic units with the same structure, step data preprocessing is also included. The data preprocessing includes at least cleaning, deduplication, and timestamp alignment. The cleaning step includes at least removing data with format errors, missing fields, or values that exceed the range. The deduplication step includes removing duplicate data caused by retransmission based on packet identifiers; The timestamp alignment step involves aligning data from different sources to the same timeline based on the original timestamps of the data itself, using a watermarking mechanism.
8. The power grid fault analysis method as described in claim 5, characterized in that, After calculating the root cause confidence level for nodes with an in-degree of 0 on the propagation chain based on the fault propagation chain graph, the system also outputs a diagnostic report containing the fault propagation chain graph and the root cause confidence level.
9. A power grid fault analysis system, characterized in that, include: The topology acquisition module is used to acquire the static topology diagram of the power grid; The data access and processing module is used to acquire operation and maintenance data of various devices in the power grid; The Time Series Large Model module is used to input operation and maintenance data into a pre-trained time series large model and output a high-dimensional embedding vector for each device node at time t. The dynamic spatiotemporal graph construction module is used to construct a dynamic spatiotemporal graph by inputting high-dimensional embedding vectors into the corresponding device nodes based on the static topology graph. The spatiotemporal graph neural network inference module is used to input a dynamic spatiotemporal graph into a preset spatiotemporal graph neural network model to obtain a weighted directed graph of fault propagation. The spatiotemporal graph neural network is used to infer the sequence of faults and the propagation influence weights along the topological edges. The fault propagation chain graph construction module is used to obtain the fault propagation chain graph based on the weighted directed graph of fault propagation and the first preset threshold of the propagation influence weight. The scoring module is used to calculate the root cause confidence level for nodes with an in-degree of 0 on the fault propagation chain graph.
10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the power grid fault analysis method according to any one of claims 1 to 8.
11. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the power grid fault analysis method according to any one of claims 1 to 8.