Power system fault diagnosis method and system based on deep learning

Abstract

The application is suitable for the field of power system operation and maintenance technology, and provides a power system fault diagnosis method and system based on deep learning, which comprises the following steps: obtaining and storing all relay protection device action logic data; constructing a protection logic knowledge graph containing action logic correlation by a logic digitization engine; constructing a virtual power grid fault simulation model by using a graph neural network; recording action process data by simulating a fault scene; analyzing logical interaction characteristics, identifying implicit risks, and generating optimization suggestions. The application solves the problems of heterogeneous data integration difficulty, complex fault simulation deficiency, and implicit risk identification difficulty, and improves fault diagnosis accuracy and practicality.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power system operation and maintenance, and particularly relates to a power system fault diagnosis method and system based on deep learning. BACKGROUND

[0002] With the rapid development of power systems towards high voltage, large capacity and cross-regional direction, as the core component to ensure the safe operation of the power grid, the application scale and type of the relay protection device continue to expand, covering line protection devices, transformer protection devices and bus protection devices of different manufacturers, different models and different ages. In recent years, deep learning technology has gradually become an important technical direction in the field of power system fault diagnosis due to its strong feature learning ability and complex pattern recognition advantage, aiming to break through the dependence on artificial experience of traditional diagnosis methods and improve the efficiency, accuracy and automation level of fault identification.

[0003] However, the action logic data of the relay protection device in the power system has problems such as heterogeneous format and complex association, which makes it difficult for the deep learning model to obtain complete system-level logic information, restricting the further improvement of its diagnosis capability. SUMMARY

[0004] The purpose of the present application is to provide a power system fault diagnosis method based on deep learning, aiming to solve the technical problems existing in the prior art identified in the background.

[0005] The present application is implemented as follows: a power system fault diagnosis method based on deep learning, the method comprising:

[0006] Obtain the action logic data of all relay protection devices in the power system, analyze the action logic association relationship of all relay protection devices, and store the action logic data to a preset database;

[0007] Translate and model the action logic data through a logic digitization engine, convert the logic description into a standardized computable logic unit, and construct a protection logic knowledge graph containing the action logic association relationship of all relay protection devices through entity recognition and relationship mapping;

[0008] Based on the protection logic knowledge graph, use a graph neural network to construct a virtual power grid fault simulation model for the topology structure of the power system and the logic association relationship of the relay protection device, to simulate the parallel interactive operation state of the protection device in the power system;

[0009] Simulate a number of fault scenarios through the virtual power grid fault simulation model, and record the action process data of all relay protection devices under each fault scenario;

[0010] The logic interaction characteristics of the relay protection device in the action process data are analyzed, hidden risks in the power system are identified, and adaptive logic optimization suggestions are generated in combination with a protection logic knowledge graph.

[0011] As a further scheme of the present application, the action logic data includes configuration files, setting value sheets and binary code form logic description data corresponding to line protection devices, transformer protection devices and bus protection devices of different manufacturers, different models and different ages.

[0012] As a further scheme of the present application, the action logic data of all relay protection devices in the power system is obtained, specifically including:

[0013] Action logic data of all relay protection devices is collected from various data sources of the power system;

[0014] The collected action logic data is format-analyzed and semantically extracted, the logic rules, parameter settings and dependency relationships in each relay protection device are identified, and an action logic correlation graph is constructed;

[0015] The analyzed action logic data and its action logic correlation are stored in a preset database and indexed.

[0016] As a further scheme of the present application, the protection logic knowledge graph containing the action logic correlation of all relay protection devices is constructed, specifically including:

[0017] The action logic data is loaded by a logic digitalization engine, and the action logic data is uniformly translated to convert heterogeneous logic descriptions into standardized computable logic units;

[0018] Based on the translated computable logic units, entity recognition and relationship mapping are performed, the relay protection device, logic criterion and action rule are identified as entities, and the time sequence, condition and dependency relationship between entities are defined as edges to construct a preliminary graph structure;

[0019] The preliminary graph structure is integrated into a protection logic knowledge graph by graph database technology storage and management, and the protection logic knowledge graph contains the action logic correlation of all relay protection devices.

[0020] As a further scheme of the present application, the preliminary graph structure is integrated into a protection logic knowledge graph by graph database technology storage and management, specifically:

[0021] The entity nodes and relationship edges in the preliminary graph structure are topologically sorted and attribute fused to establish a unified node identification system and relationship weight matrix;

[0022] The structured data is persistently stored using a bulk loading interface of a graph database to form a complete protection logic knowledge graph.

[0023] As a further scheme of the present application, the construction of the virtual power grid fault simulation model specifically comprises:

[0024] Based on the protection logic knowledge graph, a feature representation of the logical correlation between the power system topology and the relay protection device is extracted using a graph neural network, and the graph neural network models the interaction dependency and time sequence constraint between the relay protection devices through node embedding and edge weight learning;

[0025] The graph neural network model is optimized in combination with a reinforcement learning algorithm to construct a virtual power grid fault simulation model, and the virtual power grid fault simulation model simulates the parallel interactive operation state of the protection devices in the power system, including initializing a simulation environment, setting a parallel computing framework and integrating real-time interaction logic.

[0026] By comparing the historical fault data with the simulation output, the parameters of the virtual power grid fault simulation model are adjusted to verify the accuracy and stability of the virtual power grid fault simulation model.

[0027] As a further scheme of the present application, the action process data includes: the logical judgment time sequence of the protection device, the decision result and the interaction influence data.

[0028] The fault scenarios include: cross-line fault, developing fault, protection refusal fault and protection misoperation fault.

[0029] As a further scheme of the present application, the simulation generates a plurality of fault scenarios, and records the action process data of all relay protection devices under each fault scenario, specifically comprising:

[0030] A plurality of fault scenarios are simulated by the virtual power grid fault simulation model, including cross-line fault, developing fault, protection refusal fault and protection misoperation fault.

[0031] Under each fault scenario, all relay protection devices are operated, and the action process data of each relay protection device is recorded in real time.

[0032] The recorded action process data is stored in an analysis database and is preprocessed and formatted.

[0033] As a further scheme of the present application, the identification of the implicit risk in the power system specifically comprises:

[0034] Extract the local timing mode of the action of the relay protection device, learn the decision path dependence relationship across time steps through the long short-term memory network, establish a protection action behavior benchmark model based on historical data under normal conditions, and compare the real-time action process data with the benchmark model by using a density-based clustering algorithm, and mark the interaction behavior as abnormal interaction behavior when the action timing difference exceeds ±20ms and the criterion output is inconsistent with the expectation;

[0035] Based on the analysis result of the logical interaction feature, combining the protection logic knowledge graph, the risk source and the influence range are located through pattern matching and graph traversal technology, and the implicit risk in the power system is identified, including the overstep trip logic vulnerability and the logic conflict;

[0036] The logical association relationship between the protection devices is traversed, the shortest action path between the protection devices is calculated, if the action timing of the upstream relay protection device in the protection logic knowledge graph is earlier than that of the direct downstream relay protection device, and the time difference exceeds the cooperation time limit specified by the fixed value, it is determined that the overstep trip logic vulnerability exists;

[0037] The logical expression outputs of the protection devices of different manufacturers under the same fault condition are compared, when the same input condition produces exclusive judgment results, it is marked as a logic conflict;

[0038] According to the identified implicit risk, an adaptive logic optimization suggestion is generated, and the optimization effect is verified through simulation.

[0039] Another object of the present application is to provide a deep learning-based power system fault diagnosis system, which comprises:

[0040] A data acquisition and storage module is used to acquire the action logic data of all relay protection devices in the power system, analyze the action logic association relationship of all relay protection devices, and store the action logic data to a preset database;

[0041] A knowledge graph building module is used to uniformly translate and model the action logic data through a logic digitization engine, convert the logic description into a standardized computable logic unit, and build a protection logic knowledge graph containing the action logic association relationship of all relay protection devices through entity recognition and relationship mapping;

[0042] A simulation model building module is used to build a virtual power grid fault simulation model based on the protection logic knowledge graph, using a graph neural network to simulate the parallel interaction operation state of the protection devices in the power system;

[0043] A fault scenario simulation module is used to simulate a plurality of fault scenarios through the virtual power grid fault simulation model, and record the action process data of all relay protection devices under each fault scenario.

[0044] An invisible risk identification module is configured to analyze logical interaction features of the relay protection device in the action process data, identify invisible risks in the power system, and generate adaptive logic optimization suggestions in combination with the protection logic knowledge graph.

[0045] The present application has the following advantages:

[0046] The present application effectively solves the heterogeneous problem of action logic data of relay protection devices of different manufacturers, different models and different years, realizes unified integration and system-level associated modeling of data, provides complete and accurate input basis for deep learning models, and breaks through the data island limitation in traditional methods; the virtual power grid fault simulation model constructed based on the graph neural network and the reinforcement learning can truly simulate the parallel interaction operation state of the protection device in the power system, comprehensively cover complex scenes such as cross-line fault, developing fault, protection refusal fault and protection misoperation fault, and generate action process data to provide reliable support for invisible risk identification, and make up for the defects of single simulation scene and large deviation from actual operation in traditional simulation scene; the long short-term memory network learns the time sequence dependence relationship, the density-based clustering algorithm marks abnormal interaction behavior, and the risk source is located and the influence range is analyzed in combination with the protection logic knowledge graph, which can not only accurately identify invisible risks such as overstep tripping logic vulnerability and logic conflict, but also generate adaptive logic optimization suggestions based on the risk type, avoid the subjectivity and trial and error risk of traditional manual suggestions, and greatly improve the comprehensiveness, accuracy and practicality of power system fault diagnosis. At the same time, the present application has good scalability and continuous iteration capability, and only needs to add entities and associated relationships in the knowledge graph when new equipment is put into operation, without the need to reconfigure the whole model, which can adapt to the updating of power system equipment and the adjustment of operation parameters, long-term guarantee the safe and stable operation of power grid, and reduce the cost of manual operation and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 A flowchart of the power system fault diagnosis method based on deep learning provided by the embodiment of the present application is shown in the figure;

[0048] Figure 2 A flowchart of obtaining action logic data of all relay protection devices in the power system provided by the embodiment of the present application is shown in the figure;

[0049] Figure 3 A flowchart of constructing a protection logic knowledge graph containing action logic associated relationships of all relay protection devices provided by the embodiment of the present application is shown in the figure;

[0050] Figure 4 A flowchart of constructing a virtual power grid fault simulation model provided by the embodiment of the present application is shown in the figure;

[0051] Figure 5 A flowchart for simulating generation of several fault scenarios is provided for the embodiments of the present application.

[0052] Figure 6 A flowchart for identifying implicit risks in a power system is provided for the embodiments of the present application.

[0053] Figure 7 A structural block diagram of a deep learning-based power system fault diagnosis system is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0054] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0055] Figure 1 A flowchart of a deep learning-based power system fault diagnosis method is provided for the embodiments of the present application, as shown in Figure 1 The method comprises:

[0056] S100, acquiring action logic data of all relay protection devices in a power system, analyzing the action logic correlation of all relay protection devices, and storing the action logic data into a preset database;

[0057] In the data acquisition stage, all action logic data sources of the relay protection devices in the power system are comprehensively covered, and these data sources include real-time operation data sources. Because the action logic of the relay protection devices in the power system has dual attributes of dynamic response and static reference, for example, the action logic of the line protection device not only contains the dynamic response rule that the current exceeding the setting value in real-time operation triggers the protection, but also contains the static logic of “protection action priority sorting” fixed at the factory, and the lack of any type of data will cause the whole logic to be incomplete.

[0058] S200, uniformly translating and modeling the action logic data through a logic digitization engine, converting the logic description into a standardized computable logic unit, and constructing a protection logic knowledge graph containing the action logic correlation of all relay protection devices through entity recognition and relationship mapping;

[0059] In power systems, the logic description forms of protection devices from different manufacturers and different eras are quite different: some new devices store logic in the standardized IEC61850 protocol format, some old devices rely on binary codes or paper documents to record logic, and some devices express logic through graphical methods such as ladder diagrams and function block diagrams. If these heterogeneous data are not uniformly processed, the deep learning model will not be able to effectively learn the logic rules due to the confusion of input formats. Therefore, the logic digitization engine needs to design special parsing rules for different formats of data: for graphical ladder diagrams, graphical element recognition technology is used to convert components such as contacts, coils, and timers into structured logic expressions of conditional judgments, execution actions, and timing parameters; for binary codes, through decompilation and logic flow restoration technology, the conditional branches and execution sequences in the code are extracted, and the trigger threshold and subsequent actions of each logic judgment are determined; for paper documents described in natural language, domain dictionary-based natural language processing technology is used to extract core elements.

[0060] After completing the unified translation, a preliminary graph structure needs to be constructed through entity recognition and relationship mapping. The core requirement of power system fault diagnosis is to understand the coordinated response logic of multiple devices when a fault occurs, rather than isolated analysis of a single device. Therefore, entity recognition needs to break through the limitation of only identifying protection devices and include logic criteria and action rules into the entity category, forming a three-in-one entity system of protection devices, logic criteria, and action rules.

[0061] The preliminary graph structure needs to be integrated into a protection logic knowledge graph through graph database technology. The core reason for choosing a graph database instead of a traditional relational database is that the association of power system protection logic has multi-dimensional, strong coupling, and dynamic characteristics: a relational database needs to query through complex associations of multiple tables to obtain the logic chain between devices, while a graph database can directly locate associated entities through node traversal. In the integration process, topological sorting can determine the order of logic execution to avoid timing confusion during deep learning model training caused by logic loops; attribute fusion integrates multiple source attributes of the same entity into a single node to avoid the need for the model to retrieve information from multiple data sources due to scattered attributes; a unified node identification system is used to achieve data uniqueness and traceability; and the use of batch loading interfaces is to adapt to the large number of protection devices in power systems, such as a provincial power grid containing thousands of devices. Through batch loading, tens of thousands of entity and relationship data can be quickly written into the graph database, avoiding the inefficiency caused by storing one by one.

[0062] S300, based on the protection logic knowledge graph, a virtual power grid fault simulation model is constructed by using a graph neural network to associate the topology of the power system and the logic of the relay protection device, to simulate the parallel interactive running state of the protection devices in the power system;

[0063] In power systems, the actions of protection devices are not isolated events, but are influenced by the topology and logical association of the power grid. Traditional methods are difficult to capture both types of association simultaneously, while the node embedding and edge weight learning of graph neural networks can solve this problem. Node embedding technology converts each entity in the knowledge graph into a low-dimensional vector, which not only contains the properties of the entity itself, but also integrates the association information with its surrounding entities. Edge weight learning assigns reasonable weights to different types of relationships (time sequence, condition, and dependency) in the knowledge graph through iterative optimization.

[0064] For the core architecture of the virtual power grid fault simulation model, the topology structure and logical association in the protection logic knowledge graph need to be directly mapped to the basic framework of the simulation scene during the initialization of the simulation environment, and the initial state of each device needs to be determined. The parallel computing framework is set up to simulate the parallel interaction of multiple protection devices in the real power system. The parallel framework can process the action process data generated by multiple devices in real time through distributed computing resources, avoiding simulation delays caused by data congestion; integrating real-time interaction logic ensures that the dynamic influence between devices is transmitted in real time.

[0065] The verification step adjusts the parameters by comparing historical fault data with simulation output, which is a necessary step to ensure the reliability of the model. Historical fault data is the real benchmark accumulated in the actual operation of the power system, containing key information such as fault type, protection device action time sequence, and decision results. By comparing the output of the simulation model under the same fault conditions with the historical data, if the action time sequence of the simulation under a certain type of fault is found to be significantly different, the model parameters need to be adjusted.

[0066] S400, simulate and generate a number of fault scenarios through the virtual power grid fault simulation model, and record the action process data of all relay protection devices under each fault scenario;

[0067] Cross-line fault, developing fault, protection refusal fault, and protection misoperation fault are key scenarios that can easily cause widespread power outages and are difficult to cover by traditional diagnosis in power systems: cross-line fault involves the coordinated response of protection devices on multiple lines, which can expose inconsistencies in different line protection criterion interpretations; developing fault evolves dynamically over time, requiring protection devices to adjust their action strategies at different stages, which can lead to misjudgment due to improper timing coordination; protection refusal / misoperation fault is a direct consequence of protection logic mismatch and can trigger a chain reaction.

[0068] When designing these scenarios, the topological association and logical dependency in the protection logic knowledge graph should be considered: when generating a cross-line fault scenario, the intersection nodes of lines with different voltage levels are selected as fault points according to the connection relationship between lines and buses in the graph, to simulate the complex situation of fault current flowing into multiple lines at the same time; when generating a developing fault scenario, the dynamic change rate of fault characteristics is set according to the logical criterion triggering sequence of the protection device in the graph, to ensure that the scenario can reproduce the dynamic criterion environment faced by the protection device in the real fault.

[0069] In the real-time recording of action process data stage, the key is to capture the full-link dynamic information of the fault response of the protection device. Traditional fault diagnosis only stores the final state of the action / inaction of the protection device, and cannot trace the key process such as why the action is taken, when the judgment starts, and how it interacts with other devices, while the implicit risk is hidden in the process data. Therefore, the recording should focus on three core dimensions: logical judgment timing, which needs to be recorded to the millisecond level to record the full-process time nodes of the protection device from sensing the fault signal to outputting the action instruction, including: fault signal detection → current criterion judgment → voltage criterion judgment → action instruction output, and the time consumption of each step.

[0070] S500, analyze the logical interaction features of the relay protection device in the action process data, identify the implicit risks in the power system, and generate adaptive logic optimization suggestions in combination with the protection logic knowledge graph.

[0071] In the power system, the action timing of the protection device has strong regularity. Under normal working conditions, the time consumption of the same type of device from fault signal detection to criterion judgment to action output is relatively fixed, and the interaction timing between different devices also has stable coordination relationship. Traditional methods mostly use fixed threshold to judge timing anomalies, which is difficult to adapt to timing fluctuations under different working conditions, while the advantage of LSTM is that it can learn the decision path dependency relationship across time steps.

[0072] When using the density-based clustering algorithm (DBSCAN) for anomaly labeling, the interaction behavior deviating from the baseline model is identified by the difference in data density. The advantage of DBSCAN is that it does not need to define the type of anomaly in advance, and can automatically find the sparse area (anomaly point) in the data set, which is especially suitable for implicit anomalies that are difficult to predict in the power system. In specific operation, the real-time action process data needs to be converted into a feature vector (including action timing value, criterion output state, and interaction signal transmission time), and the normal feature vector output by the baseline model is input into the DBSCAN algorithm: the algorithm calculates the Euclidean distance between data points, and determines that the area with high density is a normal cluster and the isolated point with very low density is an anomaly; at the same time, combined with the two strong anomaly signals of action timing difference exceeding ±20ms and criterion output inconsistent with the expectation, a double anomaly labeling mechanism is formed.

[0073] The action logic data includes configuration files, setting sheets, and logical description data in binary code form for line protection devices, transformer protection devices, and bus protection devices from different manufacturers, models, and eras.

[0074] like Figure 2 As shown, the acquisition of the operation logic data of all relay protection devices in the power system specifically includes:

[0075] S110 collects the operation logic data of all relay protection devices from various data sources in the power system;

[0076] S120 performs format parsing and semantic extraction on the collected action logic data, identifies the logic rules, parameter settings and dependencies in each relay protection device, and constructs an action logic relationship map;

[0077] S130, store the parsed action logic data and its action logic relationships in a preset database and create an index.

[0078] like Figure 3 As shown, the construction of a protection logic knowledge graph containing the operational logic relationships of all relay protection devices specifically includes:

[0079] S210, Load action logic data through the logic digitization engine, perform unified translation on the action logic data, and convert heterogeneous logic descriptions into standardized computable logic units;

[0080] S220, based on the translated computable logic unit, performs entity recognition and relation mapping, identifies relay protection devices, logic criteria and action rules as entities, and defines the timing, conditions and dependencies between entities as edges to construct a preliminary graph structure;

[0081] S230, the preliminary graph structure is integrated into a protection logic knowledge graph by storing and managing graph database technology. The protection logic knowledge graph contains the operation logic relationships of all relay protection devices.

[0082] The role of the protection logic knowledge graph is to serve as a unified representation carrier for heterogeneous logic, transforming the natural language descriptions of configuration files, the numerical parameters of setting sheets, and the logic flow graphs of binary code into a unified graph structure, thus solving the problem of integrating data from multiple sources. It also provides a foundation for fault simulation; when constructing a virtual power grid fault simulation model, the protection device nodes and logical connection edges in the graph directly serve as inputs to the graph neural network, supporting the simulation of parallel and interactive operation of protection devices. Furthermore, it supports the location and optimization of latent risks; when identifying latent risks, the graph traversal technology quickly traces the risk source, and optimization suggestions are generated based on the parameters in the graph.

[0083] Because the protection devices in the power system come from different manufacturers and different years, the action logic forms are various and have no unified correlation, the traditional method cannot establish a collaborative analysis framework between the logics, and the traditional fault diagnosis ignores the logical dependence between devices, while the knowledge graph can clearly show these correlations, and technologies such as graph neural networks require node and edge structures as input, which exactly meets this requirement. The effects after establishment include breaking data silos, realizing the interconnection and intercommunication of different protection device action logics, improving simulation and diagnosis accuracy, making simulation models simulate interaction based on real logical associations, accurately locating associated fault chains during risk diagnosis, and reducing model update costs, which only need to add entity relationships when new equipment is put into operation without reconstructing the entire model.

[0084] The preliminary graph structure is integrated into a protection logic knowledge graph by using a graph database technology, and the method comprises the following steps:

[0085] The entity nodes and relationship edges in the preliminary graph structure are topologically sorted and attribute fused to establish a unified node identification system and a relationship weight matrix;

[0086] The structured data is persistently stored using the batch loading interface of the graph database to form a complete protection logic knowledge graph.

[0087] As shown in Figure 4 The method for constructing a virtual power grid fault simulation model comprises the following steps:

[0088] S310, based on the protection logic knowledge graph, a graph neural network is used to extract the feature representation of the topological structure of the power system and the logical correlation between the relay protection devices, the graph neural network models the interactive dependence and time sequence constraints between the relay protection devices through node embedding and edge weight learning;

[0089] S320, a graph neural network model is optimized by combining a reinforcement learning algorithm to construct a virtual power grid fault simulation model, the virtual power grid fault simulation model simulates the parallel interactive running state of the protection devices in the power system, including initializing the simulation environment, setting up a parallel computing framework and integrating real-time interaction logic;

[0090] S330, by comparing historical fault data with simulation output, adjusting the parameters of the virtual power grid fault simulation model, verifying the accuracy and stability of the virtual power grid fault simulation model.

[0091] The action process data includes the logical judgment time sequence, decision results and interactive influence data of the protection devices.

[0092] The action process data refers to the dynamic operation data of all relay protection devices in each fault scenario in the virtual power grid fault simulation model, including the logic judgment time sequence, i.e., the millisecond-level time sequence from perceiving the fault signal to outputting the action instruction, the decision result, i.e., the state output of the protection device in the fault process and the criterion satisfaction condition, and the interaction influence data, i.e., the influence record of the action of a protection device on the associated device.

[0093] The action process data, as the core basis for identifying the implicit risk, finds the action time sequence mismatch between the protection devices through time sequence alignment analysis, compares and marks the abnormal behavior with the protection action behavior benchmark model under normal working conditions, and traces the risk propagation path in combination with the protection logic knowledge graph.

[0094] The traditional method only records the final action result of the protection device, cannot obtain the intermediate logic process of why the action / inaction, the implicit risk is essentially a dynamic coordination problem between protection devices, and must be identified by recording the time sequence and interaction influence in the action process. The simulation verification optimization suggestion also needs to compare the action process data before and after optimization.

[0095] The fault scenarios include cross-line fault, developing fault, protection refusal fault and protection misoperation fault.

[0096] The cross-line fault refers to a fault that simultaneously involves two or more adjacent lines, the fault current path is complex, and multiple line protection devices are triggered to act simultaneously, which easily leads to protection logic coordination confusion;

[0097] The developing fault refers to a fault that gradually evolves over time, the fault characteristics dynamically change, and the protection device needs to adjust the action strategy at different stages, which easily leads to misjudgment due to improper time sequence coordination;

[0098] The protection refusal fault refers to a protection device that should act when a fault occurs but does not act, which will lead to fault expansion and trigger the action of the upper-level protection;

[0099] The protection misoperation fault refers to the error action of a protection device when there is no fault or a fault outside the protection range, which will lead to power failure in the fault-free area.

[0100] The simulation for these fault scenarios is because they are high-risk scenarios of implicit risk, the cross-line fault easily exposes the problem of inconsistent interpretation of protection criteria for different lines, the developing fault easily triggers the action time sequence mismatch of the protection, and the protection refusal / misoperation is a direct consequence of protection logic mismatch and will trigger a chain reaction; at the same time, the traditional method is difficult to cover these complex scenarios, physical tests or static simulations cannot simulate the multi-path current of cross-line faults and the dynamic evolution of developing faults, leading to the long-term neglect of implicit risks; in addition, covering these scenarios can verify the adaptability of the protection logic knowledge graph and the simulation model, ensuring that the system can accurately identify the risk and output optimization suggestions when such faults occur in the actual power grid.

[0101] As Figure 5 shown, the simulation generates several fault scenarios, and records the action process data of all relay protection devices under each fault scenario, specifically including:

[0102] S410, a plurality of fault scenarios are simulated by a virtual power grid fault simulation model, including cross-line fault, developing fault, protection refusal fault and protection misoperation fault;

[0103] S420, under each fault scenario, all relay protection devices are operated, and the action process data of each relay protection device is recorded in real time;

[0104] S430, the recorded action process data is stored in an analysis database, and is preprocessed and formatted.

[0105] As Figure 6 shown, the implicit risk in the power system is identified, specifically including:

[0106] S510, a local timing pattern of relay protection device action is extracted, a long short-term memory network is used to learn the decision path dependency relationship across time steps, a protection action behavior benchmark model is established based on historical data under normal conditions, and a density-based clustering algorithm is used to compare real-time action process data with the benchmark model, and the interaction behavior whose action timing difference exceeds ±20ms and whose criterion output is inconsistent with the expectation is marked as abnormal interaction behavior;

[0107] Specifically:

[0108] 1. The historical action process data under normal conditions is used to train the benchmark model. The benchmark model includes the action timing mean value and standard deviation of each relay protection device under normal state, and the normal mode of criterion output. For the action timing, the mean value and standard deviation of the historical data are calculated; for the criterion output, the normal output mode is counted;

[0109] 2. The feature vector is extracted from the real-time action process data, including the action timing value and the criterion output value . The feature vector is represented as ;

[0110] 3. The real-time action process data is clustered to identify abnormal points. DBSCAN requires two parameters: neighborhood radius and minimum point number minPts. First, the Euclidean distance between the real-time data points and the core points of the benchmark model is calculated. If a data point is within its neighborhood radius If the number of included reference core points is less than the minimum number minPts, the point is marked as abnormal. At the same time, the time difference is directly checked and the criterion output: If the time difference exceeds ±20ms or the criterion output does not match, it is also marked as abnormal.

[0111] 4. Synthesize the DBSCAN results and direct inspection, and mark the data points that meet any of the following conditions as abnormal interaction behaviors:

[0112] Use the time difference formula If , mark it as a time sequence anomaly, where is the time difference.

[0113] Formula application: Use the criterion output check formula, such as , mark it as a criterion anomaly, where is the expected criterion output in the reference model, which is obtained from historical data statistics, is the actual criterion output in the real-time action process data.

[0114] DBSCAN identifies as a noise point.

[0115] S520, based on the logical interaction feature analysis result, combining the protection logic knowledge graph, locating the risk source and the influence range through pattern matching and graph traversal technology, identifying the implicit risks in the power system, including the over-level trip logic vulnerability and the logic conflict;

[0116] Specifically:

[0117] 1. Based on the protection logic knowledge graph and abnormal interaction behaviors, define risk patterns. For over-level trip, the pattern is that the upstream protection action time is earlier than the downstream protection and the time difference exceeds the coordination time limit; for logic conflict, the pattern is that different protection devices output is mutually exclusive under the same fault condition.

[0118] 2. Traverse the protection logic knowledge graph. Start from the protection device node corresponding to the abnormal interaction behavior, and traverse along the logical association edge.

[0119] For over-level trip logic vulnerability:

[0120] In the traversal process, for each pair of adjacent protection devices (upstream relay protection device and downstream relay protection device ), use the time sequence comparison formula , where is the action time of the upstream relay protection device, is the action time of the downstream relay protection device. If , the upstream relay protection device It was marked as a risk source, and the information was obtained from the customized order to meet the time limit.

[0121] Regarding logical conflicts:

[0122] During the traversal, for the same fault conditions Use the logical conflict checking formula:

[0123] ;

[0124] in and It is a relay protection device and relay protection devices Under fault conditions The judgment output is derived from the action process data. If the outputs are mutually exclusive, the relay protection device... and relay protection devices It has been flagged as a source of risk.

[0125] 3. From the identified set of risk source nodes We begin by using BFS to traverse the protected logic knowledge graph. ,in It is a set of nodes, namely relay protection devices. It is a set of edges, representing the logical connection direction between protection devices.

[0126] Scope of influence Defined as Using the BFS algorithm, from each faulty node... Start accessing all reachable relay protection device nodes These nodes represent protection devices and power grid areas affected by risks, among which, To protect the set of nodes in the logical knowledge graph.

[0127] 4. Based on the above steps, output the identification results of cascading tripping logic vulnerabilities and logic conflicts. The risk sources and scope of impact are used to generate adaptive logic optimization suggestions.

[0128] S530: Traverse the logical relationships between protection devices, calculate the shortest action path between protection devices, and if there is an upstream relay protection device whose action sequence is earlier than that of the direct downstream relay protection device in the protection logic knowledge graph, and the time difference exceeds the coordination time limit specified in the setting sheet, it is judged as a cascading trip logic vulnerability.

[0129] Specifically:

[0130] 1. Based on protection logic knowledge graph From the fault node Start, calculate the shortest path to all other protection device nodes:

[0131] Set , for other nodes , set .

[0132] Create a priority queue and add all nodes to the queue.

[0133] Loop until the queue is empty:

[0134] Take the node with the smallest distance from the queue .

[0135] For each relay protection device node (i.e. there is an edge ), calculate the temporary distance value .

[0136] If the temporary distance value , update and record the path.

[0137] Output the shortest distance and path of each relay protection device node .

[0138] 2. After obtaining the shortest path from the fault node to each relay protection device node , use the value to identify the action timing sequence.

[0139] S540, compare the logical expression outputs of different manufacturers' protection devices under the same fault condition, and when the same input condition produces mutually exclusive judgment results, mark it as a logical conflict;

[0140] S550, according to the identified implicit risk, generate adaptive logic optimization suggestions, and verify the optimization effect through simulation.

[0141] Figure 7 The structural block diagram of the power system fault diagnosis system based on deep learning provided by the embodiment of the application is shown in Figure 7 , the system comprises:

[0142] A data acquisition and storage module 100 is configured to acquire the action logic data of all relay protection devices in the power system, analyze the action logic correlation of all relay protection devices, and store the action logic data in a preset database.

[0143] The knowledge graph building module 200 is configured to perform unified translation and modeling on the action logic data by a logical digitalization engine, convert logical description into a standardized computable logic unit, and build a protection logic knowledge graph containing the action logic correlation of all the relaying protection devices through entity recognition and relationship mapping;

[0144] The simulation model building module 300 is configured to build a virtual power grid fault simulation model by using a graph neural network to model the logical correlation between the power system topology and the relaying protection devices based on the protection logic knowledge graph, and simulate the parallel interactive operation state of the protection devices in the power system.

[0145] The fault scenario simulation module 400 is configured to simulate a plurality of fault scenarios by the virtual power grid fault simulation model, and record the action process data of all the relaying protection devices under each fault scenario.

[0146] The hidden risk identification module 500 is configured to analyze the logical interaction features of the relaying protection devices in the action process data, identify the hidden risks in the power system, and generate adaptive logic optimization suggestions in combination with the protection logic knowledge graph.

[0147] The technical features of the above-described embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described, but as long as the combinations of the technical features do not contradict, they should be considered within the scope of the present disclosure.

[0148] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as limiting the scope of the present patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present patent should be subject to the appended claims.

[0149] The above-described embodiments are only preferred embodiments of the present application and are not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A power system fault diagnosis method based on deep learning, characterized in that, The method includes: Obtain the action logic data of all relay protection devices in the power system, parse the action logic relationship of all relay protection devices, and store the action logic data in a preset database; The action logic data is uniformly translated and modeled by the logic digitization engine, the logic description is converted into standardized computable logic units, and a protection logic knowledge graph containing the action logic associations of all relay protection devices is constructed through entity recognition and relationship mapping. Based on the protection logic knowledge graph, a virtual power grid fault simulation model is constructed by using graph neural networks to analyze the logical relationship between the power system topology and relay protection devices, thereby simulating the parallel interactive operation of protection devices in the power system. By using a virtual power grid fault simulation model, several fault scenarios are generated and the action process data of all relay protection devices under each fault scenario are recorded. Analyze the logical interaction characteristics of relay protection devices in the action process data, identify hidden risks in the power system, and generate adaptive logic optimization suggestions by combining protection logic knowledge graphs. Specifically, the construction of a protection logic knowledge graph containing the operational logic relationships of all relay protection devices includes: The action logic data is loaded through a logic digitization engine, and the action logic data is uniformly translated to convert heterogeneous logic descriptions into standardized computable logic units. Based on the translated computable logic units, entity recognition and relationship mapping are performed. Relay protection devices, logic criteria, and action rules are identified as entities, and the timing, conditions, and dependencies between entities are defined as edges to construct a preliminary graph structure. The initial graph structure is integrated into a protection logic knowledge graph by storing and managing graph database technology. The protection logic knowledge graph contains the operational logic relationships of all relay protection devices. The construction of the virtual power grid fault simulation model specifically includes: Based on the protection logic knowledge graph, a graph neural network is used to extract the feature representation of the logical relationship between the power system topology and the relay protection device. The graph neural network models the interaction dependency and timing constraints between relay protection devices through node embedding and edge weight learning. By combining reinforcement learning algorithms to optimize the graph neural network model, a virtual power grid fault simulation model is constructed. The virtual power grid fault simulation model simulates the parallel interactive operation state of protection devices in the power system, including initializing the simulation environment, setting up the parallel computing framework, and integrating real-time interactive logic. By comparing historical fault data with simulation output, the parameters of the virtual power grid fault simulation model were adjusted to verify the accuracy and stability of the virtual power grid fault simulation model. The identification of hidden risks in the power system specifically includes: The local timing pattern of relay protection device operation is extracted, and the decision path dependency relationship across time steps is learned through long short-term memory network. Based on historical data under normal operating conditions, a protection action behavior benchmark model is established, and a density-based clustering algorithm is used to compare the real-time action process data with the benchmark model. Interaction behaviors with action timing difference exceeding ±20ms and inconsistent criterion output with expectations are marked as abnormal interaction behaviors. Based on the results of logical interaction feature analysis, combined with the protection logic knowledge graph, and through pattern matching and graph traversal techniques, the risk sources and scope of impact are located, and hidden risks in the power system are identified. These hidden risks include overstepping trip logic loopholes and logical conflicts. Traverse the logical relationships between protection devices, calculate the shortest action path between protection devices, and if there is an upstream relay protection device whose action sequence is earlier than that of the direct downstream relay protection device in the protection logic knowledge graph, and the time difference exceeds the coordination time limit specified in the setting sheet, it is judged as a cascading trip logic loophole. By comparing the logic expression outputs of protection devices from different manufacturers under the same fault conditions, when the same input conditions produce mutually exclusive judgment results, they are marked as logical conflicts. Based on the identified hidden risks, adaptive logic optimization suggestions are generated, and the optimization effect is verified through simulation.

2. The method according to claim 1, characterized in that, The action logic data includes configuration files, setting sheets, and logical description data in binary code form for line protection devices, transformer protection devices, and bus protection devices from different manufacturers, models, and eras.

3. The method according to claim 2, characterized in that, The acquisition of the operational logic data of all relay protection devices in the power system specifically includes: Collect the operation logic data of all relay protection devices from various data sources in the power system; The collected action logic data is parsed and semantically extracted to identify the logic rules, parameter settings and dependencies in each relay protection device, and an action logic relationship map is constructed. The parsed action logic data and its action logic relationships are stored in a preset database and an index is created.

4. The method according to claim 1, characterized in that, The process of integrating the preliminary graph structure into a protection logic knowledge graph through graph database technology for storage and management is specifically as follows: The entity nodes and relation edges in the preliminary graph structure are topologically sorted and their attributes are fused to establish a unified node identification system and relation weight matrix. The batch loading interface of the graph database is used to persistently store structured data, forming a complete protection logic knowledge graph.

5. The method according to claim 1, characterized in that, The action process data includes: the logic judgment timing, decision results, and interaction impact data of the protection device; The fault scenarios include: cross-line faults, progressive faults, protection failure to operate faults, and protection malfunctions.

6. The method according to claim 1, characterized in that, The simulation generates several fault scenarios and records the operation process data of all relay protection devices under each fault scenario, specifically including: Through a virtual power grid fault simulation model, several fault scenarios are generated, including cross-line faults, progressive faults, protection failure to operate faults, and protection maloperation faults. In each fault scenario, all relay protection devices are operated, and the action process data of each relay protection device is recorded in real time. The recorded action process data is stored in the analysis database and then preprocessed and formatted.

7. A power system fault diagnosis system based on deep learning, characterized in that, The system includes: The data acquisition and storage module is used to acquire the action logic data of all relay protection devices in the power system, parse the action logic relationship of all relay protection devices, and store the action logic data in a preset database. The knowledge graph building module is used to uniformly translate and model the action logic data through a logic digitization engine, convert the logic description into standardized computable logic units, and construct a protection logic knowledge graph containing the action logic associations of all relay protection devices through entity recognition and relationship mapping. The simulation model building module is used to construct a virtual power grid fault simulation model based on the protection logic knowledge graph and the logical relationship between the power system topology and relay protection devices using graph neural networks, so as to simulate the parallel interactive operation of protection devices in the power system. The fault scenario simulation module is used to simulate and generate several fault scenarios through a virtual power grid fault simulation model, and record the action process data of all relay protection devices under each fault scenario. The hidden risk identification module is used to analyze the logical interaction characteristics of relay protection devices in the action process data, identify hidden risks in the power system, and generate adaptive logic optimization suggestions by combining the protection logic knowledge graph.

Images