A method, device and system for analyzing the action of a relay protection
By introducing a "semantic + topology" dual-drive mechanism and an engineering-based human-machine feedback loop into relay protection action analysis, the problems of insufficient numerical calculation accuracy and interpretability of large language models in relay protection analysis are solved, achieving efficient and accurate causal inference and meeting the requirements of interpretability and traceability in relay protection analysis.
Patent Information
- Application Number
- CN202611138981.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-29
- Publication Date
- 2026-11-06
AI Technical Summary
In existing technologies, when relay protection action reasoning relies on large language models, there are problems such as insufficient numerical calculation accuracy, lack of interpretability of the reasoning process, and susceptibility of historical case retrieval to semantic interference. Maintenance personnel cannot intuitively perceive the causal relationship between changes in sampled data and relay protection action logic, which fails to meet the requirements of relay protection analysis.
A dual-drive mechanism of "semantics + topology" is adopted. By obtaining the natural language description of the fault and the main logical execution path, bidirectional retrieval is performed in the pre-built historical fault case library and standard action sequence library. Chain causal inference is performed using a large language model, and high-precision numerical calculation is performed by combining external scientific computing tools. An engineered human-machine feedback closed-loop mechanism is designed to achieve interpretability and traceability.
It significantly improves the transparency, accuracy, and efficiency of relay protection action analysis, meets the stringent requirements of relay protection analysis for interpretability and traceability, eliminates interfering cases where "the textual descriptions are similar but the action sequences are completely different", and ensures the reliability and credibility of the analysis results.
Smart Images
Figure CN123049146A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of relay protection technology, specifically relating to a relay protection action analysis method, device, and system. Background Technology
[0002] Relay protection devices are the first line of defense for ensuring the safe and stable operation of power systems. When a system fault occurs, the protection device, based on the collected electrical quantity information (such as voltage and current waveforms), and according to internally preset logic criteria (such as impedance circle characteristics and differential ratio braking characteristics), quickly identifies the fault and outputs a trip command. Accurate analysis of the correctness of the protection action is of great significance for fault diagnosis, setting optimization, and device development.
[0003] In recent years, with breakthroughs in the field of natural language processing using Large Language Models (LLMs), some research has begun to explore their application in power system fault diagnosis. However, general-purpose LLMs face a significant "technical gap" when dealing with highly specialized fields like relay protection, which require extremely high numerical accuracy.
[0004] Numerical Hallucination: Large models are essentially probabilistic text generation models with weak arithmetic capabilities. When performing tasks such as impedance trajectory calculation, harmonic analysis, and phasor comparison, they are prone to numerical biases, leading to unreliable analysis results.
[0005] The reasoning process is black-boxed: Existing knowledge-based retrieval augmented generation (RAG) technology is essentially still a black-box intelligent diagnostic approach. Its core reasoning process relies entirely on the parameterized generation mechanism of a large model, lacking an explicit display of the internal logic and actions of the protection system. Maintenance personnel cannot intuitively perceive the causal relationship between changes in sampled data and the relay protection's operational logic, failing to meet the stringent requirements of process transparency and causal traceability in relay protection analysis. Summary of the Invention
[0006] The purpose of this invention is to provide a method, device, and system for analyzing relay protection actions, in order to solve the technical problem that the reasoning process of relay protection actions in the prior art relies solely on textual descriptions, and maintenance personnel cannot intuitively perceive the causal relationship between changes in sampled data and the logic of relay protection actions, thus failing to meet the requirements of relay protection analysis.
[0007] To address the aforementioned technical problems, the first aspect of this invention provides a method for analyzing relay protection actions, the method comprising:
[0008] S1. Obtain the natural language description and main logical execution path of this fault;
[0009] S2. Based on the natural language description, query similar fault cases in the historical fault case database;
[0010] Based on the main logical execution path, match similar standard action sequences in the standard action sequence library;
[0011] The historical fault case library stores natural language descriptions of historical fault cases; the standard action sequence library stores standard action sequences corresponding to all standard protection logic.
[0012] The fault case and the standard action sequence with the highest comprehensive similarity are selected from the similar fault cases and the similar standard action sequences; the comprehensive similarity is obtained by weighting the semantic similarity corresponding to the fault case and the topological similarity corresponding to the standard action sequence.
[0013] S3. Inject the fault cases with the highest comprehensive similarity and the standard action sequence into the thought chain as prompt words of the large language model, and generate the protection action evaluation conclusion through chain causal inference by the large language model.
[0014] In one possible implementation, S2 further includes: calculating the protection criterion setting value using a pre-set relay protection quantitative calculation tool; the prompt word also includes the protection criterion setting value.
[0015] In one possible implementation, the evaluation conclusion of the protection action includes: the fault type and its criterion verification, the deviation between the main logical execution path and the standard action sequence, the causal chain of the current relay protection action, and whether the action is correct or not.
[0016] In one possible implementation, the protection action evaluation conclusion further includes a confidence score, which is obtained by weighted summation of a data quality index, an action margin index, and a logical consistency index; the data quality index is used to represent the quality of fault data in the warning words; the action margin index is used to represent the reliability of the relay protection action; and the logical consistency index is used to represent the logical consistency of the actions of each protection element.
[0017] In one possible implementation, the process of querying similar failure cases in a historical failure case database based on the natural language description includes: converting the natural language description into a semantic query vector, and retrieving failure cases in the historical failure case database that correspond to the K semantic vectors with the highest semantic similarity to the semantic query vector.
[0018] The historical failure case library stores historical failure cases in the form of semantic vectors.
[0019] In one possible implementation, the process of matching similar standard action sequences in the standard action sequence library based on the logical execution path backbone includes: converting the logical execution path backbone into a graph embedding vector, and matching the M standard action sequences with the highest topological similarity to the graph embedding vector in the standard action sequence library.
[0020] The standard action sequence library stores the standard action sequences using graph embedding vectors as indexes.
[0021] In one possible implementation, the data quality metrics Motion margin index Logical consistency index They are respectively:
[0022]
[0023]
[0024]
[0025] in, For the integrity rate of the recording channel; The continuity rate of the sampled data; This is the timescale synchronization deviation factor; , and All are preset weights. ; The number of independent protection components consistent with the deduction conclusion; The total number of protection components participating in the simulation; As an action threshold; This represents the actual action value.
[0026] In one possible implementation, the method further includes: S3, after receiving the user's evaluation conclusion on the protection action, storing the knowledge of the corrected protection action evaluation conclusion in the historical fault case library and the standard action sequence library.
[0027] To address the aforementioned technical problems, a second aspect of the present invention provides a relay protection action analysis device, comprising a processor, the processor being configured to execute a computer program to implement the steps of the method in any possible implementation of the first aspect of the present invention.
[0028] To solve the above-mentioned technical problems, a third aspect of the present invention provides a relay protection action analysis system, comprising:
[0029] Data access module: used to obtain the natural language description and main logical execution path of this fault;
[0030] The intelligent agent scheduling module is used to query similar fault cases in the historical fault case library based on the natural language description; match similar standard action sequences in the standard action sequence library based on the logical execution path backbone; and select the fault case and standard action sequence with the highest comprehensive similarity from the similar fault cases and the similar standard action sequences.
[0031] The comprehensive similarity is obtained by weighting the semantic similarity corresponding to the fault case and the topological similarity corresponding to the standard action sequence; the historical fault case library stores the natural language description of historical fault cases; the standard action sequence library stores the standard action sequences corresponding to all standard protection logics;
[0032] Causal inference engine: It is used to inject the fault cases with the highest comprehensive similarity and standard action sequences into the thought chain as prompt words of the large language model, and generate protection action evaluation conclusions through chain causal inference by the large language model.
[0033] The beneficial effects of this invention are as follows: This invention simultaneously acquires the natural language description of the fault and the logical execution path backbone of the protection device, performs bidirectional searches in a pre-built historical fault case library and a standard action sequence library, and selects the most matching case and standard sequence using a weighted comprehensive score of semantic similarity and topological similarity. Subsequently, the case and sequence are injected as structured prompts into a large language model, and the protection action evaluation conclusion is output using chain-like causal inference. This method innovatively proposes a "semantic + topological" dual-drive mechanism, performing dual matching of historical cases based on "semantic similarity + action timing topology," effectively eliminating interfering cases where "the textual descriptions are similar but the action timings are completely different," and eliminating the illusion of contextual interference. This significantly improves the transparency, accuracy, and efficiency of the analysis, meeting the stringent requirements of interpretability and traceability in relay protection action analysis. This invention solves the technical problem of existing technologies where the relay protection action reasoning process relies solely on textual descriptions, preventing maintenance personnel from intuitively understanding the causal relationship between changes in sampled data and the relay protection action logic, thus failing to meet the requirements of relay protection analysis. Attached Figure Description
[0034] Figure 1 This is an overall flowchart of the implementation method for relay protection action analysis of the present invention;
[0035] Figure 2 This is a flowchart of the dual-drive retrieval process in the implementation of the relay protection action analysis method of the present invention;
[0036] Figure 3 This is a schematic diagram of the plug-in architecture of the scientific computing toolset and the tool calling interface in the implementation method of the relay protection action analysis method of the present invention;
[0037] Figure 4This is a block diagram of the system modules in the implementation method of relay protection action analysis of the present invention;
[0038] Figure 5 This is a block diagram of the module composition in the implementation method of the relay protection action analysis method of the present invention;
[0039] Figure 6 This is a flowchart of the human-machine interaction feedback closed-loop mechanism in the implementation of the relay protection action analysis method of the present invention;
[0040] Figure 7 This is a schematic diagram illustrating the correction of human-computer interaction reasoning conclusions in the implementation of the relay protection action analysis method of the present invention;
[0041] Figure 8 This is a schematic diagram of the device architecture in an embodiment of the relay protection action analysis device of the present invention. Detailed Implementation
[0042] This invention simultaneously acquires the natural language description of the fault and the logical execution path of the protection device, and performs bidirectional searches in a pre-built historical fault case library and a standard action sequence library. The most matching case and standard sequence are selected using a weighted comprehensive score of semantic similarity and topological similarity. Subsequently, the case and sequence are injected as structured cue words into a large language model, and a chain-like causal inference is used to output the protection action evaluation conclusion. This method innovatively proposes a "semantic + topological" dual-drive mechanism, performing dual matching of historical cases based on "semantic similarity + action timing topology," effectively eliminating interfering cases where "the textual descriptions are similar but the action timings are completely different," thus eliminating the illusion of contextual interference. This significantly improves the transparency, accuracy, and efficiency of the analysis, meeting the stringent requirements of interpretability and traceability in relay protection action analysis. This invention solves the technical problem of existing technologies where the relay protection action reasoning process relies solely on textual descriptions, preventing maintenance personnel from intuitively understanding the causal relationship between changes in sampled data and the relay protection action logic, thus failing to meet the requirements of relay protection analysis.
[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings.
[0044] Implementation methods for relay protection action analysis:
[0045] This invention aims to address the technical problems in existing technologies where relay protection fault analysis relies on large language models, such as insufficient numerical calculation accuracy, lack of interpretability in the reasoning process, susceptibility to semantic interference in historical case retrieval, and lack of engineering design in the human-machine feedback mechanism. It provides a relay protection action analysis method and system based on intelligent agent tool invocation.
[0046] This invention constructs a "neural symbolic collaboration" intelligent agent architecture: a large language model is responsible for semantic understanding and causal reasoning, while an external scientific computing toolset handles high-precision numerical calculations. These two are decoupled and collaborate to eliminate numerical illusions. Based on this, a dual-drive memory mechanism of temporal graph neural networks and vector retrieval is introduced to perform dual matching of historical cases based on "semantic similarity" and "action temporal topology," eliminating interfering cases. Simultaneously, an engineered human-computer feedback closed-loop mechanism is designed to standardize feedback parsing and resolve knowledge conflicts, ensuring continuous system learning without performance degradation.
[0047] The relay protection action analysis method of this embodiment is as follows: Figure 1 As shown, the method includes the following steps:
[0048] Step 1: Obtain multi-source heterogeneous operational data of the target to be analyzed and perform preprocessing.
[0049] Specifically, the fault waveform data and action event reports output by the relay protection device in COMTRADE format (a common format for power system transient data exchange) are acquired through the data access module. Preprocessing includes: parsing the waveform file, extracting the sampled data from the analog channel (voltage and current sample values) and the digital channel (protection action signals, circuit breaker positions, etc.); performing time-scale correction, filtering and noise reduction, and extracting fault characteristic electrical quantities.
[0050] The file parsing unit of the data access module first parses the COMTRADE configuration file (.cfg), intelligently identifying and interpreting the physical meaning of analog and digital input channels. It then performs preprocessing steps including data standardization and resampling, multi-source data time synchronization and alignment, fault feature extraction, and data window truncation. The key fault characteristic electrical quantities extracted after preprocessing include: current and voltage values before / after the fault, differential current, zero-sequence component, and phase difference, which are comprehensively determined to be a single-phase ground fault in phase B.
[0051] Step 2: Construct an intelligent agent collaborative reasoning architecture and use a large language model for semantic parsing and task planning.
[0052] The agent possesses a long-term memory module, which employs a dual-drive memory retrieval mechanism combining temporal graph neural networks and vector retrieval, such as... Figure 2 As shown, it specifically includes:
[0053] (1) Establishment of semantic vector index: During the agent construction stage (i.e. training stage), the text information such as fault feature description, causal inference conclusion, and action behavior judgment of historical fault cases is transformed into high-dimensional semantic vectors by embedding encoder and stored in vector database to support fast retrieval based on semantic similarity.
[0054] (2) Establishment of the temporal action topology index: During the agent construction phase, based on the standard protection logic diagrams of relay protection devices (such as differential protection logic diagrams and distance protection logic diagrams), all possible action execution paths and their key logic nodes (such as protection start-up, differential criterion, zone / internal discrimination, phase selection, and output) are extracted offline to construct a directed acyclic graph library of standard action sequences. The nodes of this directed acyclic graph represent logical elements, and the directed edges represent the order of execution of standard actions and causal dependencies. Each standard directed acyclic graph is encoded into a graph embedding vector through a temporal graph neural network and stored in the graph database as a benchmark index for subsequent action temporal topology similarity retrieval.
[0055] (3) Dual-drive fusion retrieval: When a new fault analysis task is received, the system simultaneously performs semantic vector retrieval and time sequence graph retrieval.
[0056] Semantic vector retrieval: The feature description of the current fault is transformed into a semantic query vector, and the top K historical cases with semantic similarity are retrieved from the vector database.
[0057] For example, for the analysis task of "a B-phase ground fault occurred on a 220kV line, and the differential protection tripped", the current fault characteristics description should include at least the following: the fault type is a B-phase single-phase ground fault, the voltage level is 220kV, the fault element is the line, the protection action signal is the differential protection action output, and the fault phase is B phase.
[0058] Temporal graph retrieval: The main branch of the actual logical execution path extracted from the current fault (represented in the form of a directed acyclic graph) is fed into the temporal graph neural network to obtain the corresponding graph embedding vector. Then, the graph embedding vector is used to retrieve the directed acyclic graph of the standard action sequence that best matches the actual path in terms of topology from the graph database.
[0059] Dual-drive result fusion: The two retrieval results (semantic vector retrieval and time-series graph retrieval) are weighted and fused. The specific fusion method is as follows:
[0060] Let the similarity score returned by semantic vector retrieval be... The topological similarity score returned by the time-series graph retrieval is Then the overall similarity ,in For preset weighting coefficients, .
[0061] The weighting coefficients are dynamically adjusted based on the fault analysis scenario:
[0062] For fault reports that are primarily text-based (e.g., only protection action messages without detailed waveform data), semantic similarity weighting is used. The preferred value range for the topological similarity weight is 0.65 to 0.75. The corresponding value range is 0.25 to 0.35, for example, it can be taken as... .
[0063] For faults with complete waveform data and extractable complete action timeline paths, semantic similarity weights are used. The preferred value range for the topological similarity weight is 0.35 to 0.45. The corresponding value range is 0.55 to 0.65, for example, it can be taken as... .
[0064] As the default configuration Both can be taken as 0.5.
[0065] The fused comprehensive similarity is used to sort and output the most matching historical cases and standard action sequences. By comparing the deviations between the current actual path and the standard action sequence (such as missing nodes, incorrect order, etc.), the correctness of the protection action (correct action, no action, false action) can be accurately determined.
[0066] The dual-drive mechanism effectively eliminates interfering cases where "the textual descriptions are similar but the timing of the actions is completely different," and provides a standard reference system for judging protective actions.
[0067] The agent scheduling module receives the analysis command "Analyze whether the B-phase differential protection action is correct," performs semantic parsing and task decomposition, and determines the types of tools to be invoked and their execution order (i.e., the task planning mentioned above). The agent invokes the long-term memory module for dual-drive retrieval; see [link to relevant documentation]. Figure 2 The semantic vector retrieval path on the left searches for similar cases in the historical fault case library, while the temporal topology retrieval path on the right performs topological matching between the main branch of the current fault's actual logical execution path and a pre-set directed acyclic graph library of standard action sequences. After the dual matching verification passes, the most matching historical case and standard action sequence are injected into the thought chain context as prompt words for the large model. The agent then generates tool invocation instructions sequentially according to the task plan.
[0068] Chain of Thought (CoT) refers to a cue word engineering technique. Specifically, it involves injecting retrieved similar historical cases and standard action sequences as contextual information into the cue words of a large language model in a structured manner. The aim is to guide and stimulate the large language model to reason step by step according to preset logical steps, thereby generating a more reliable and logical analysis process.
[0069] Step 3: In response to the tool call command, schedule a set of scientific computing tools that are independent of the large language model to perform numerical computations.
[0070] Specifically, the scientific computing toolset adopts a plug-in architecture, supporting the dynamic registration of new computing tools; the interaction between the agent and the scientific computing toolset follows a unified function call specification, and the tools call the gateway (see...). Figure 4 The system automatically matches and schedules any computational tool within the pluggable architecture. The scientific computing toolset includes at least harmonic analysis tools, impedance measurement tools, and fault phase selection tools. The agent interacts with the toolset through standardized interfaces, ensuring that the large model does not directly participate in underlying mathematical operations, thereby eliminating numerical illusions.
[0071] As an example, the impedance measurement tool executes the following formula:
[0072]
[0073] in, For the fault phase voltage phasor, For fault phase current phasors, The zero-order compensation coefficient is... For impedance, The zero-sequence current is used; after calculation, a structured result containing the impedance magnitude and impedance angle is returned to the agent.
[0074] In actual execution, the various tools work together to complete numerical calculations: the harmonic analysis tool performs a Fourier transform on the B-phase current after the fault, extracting the amplitude and phase of the fundamental component; the impedance measurement tool calculates the line impedance according to the above formula; and the logic criterion calculation tool reads the protection settings and verifies the action criteria item by item. Each tool returns structured calculation results for the large model to perform subsequent logical reasoning and evaluation. Step 4: Receive the structured calculation results, combine them with the domain knowledge base to perform chain-like causal deduction, and generate protection action evaluation conclusions.
[0075] The large language model receives precise values returned by scientific computing tools and performs a dual "theory-practice" verification logic: it deduces the set of protection components that should theoretically operate based on fault characteristics and set values; it analyzes the waveform file to extract the actual operation signal sequence; it compares each item to determine whether the operation is correct, fails to operate, or is erroneous; and it visualizes the association between intermediate variables and set value thresholds to generate an explicit causal chain.
[0076] The analysis engine (i.e., the large language model) receives the results of the scientific calculations and the knowledge base retrieval, and performs chained reasoning, such as... Figure 3 As shown, the process involves fault feature identification → standard logic matching → criterion verification → logic node state inference → exit action determination → conclusion output. A weighted calculation model is used to output a confidence score, based on data quality indicators, action margin indicators, and logical consistency indicators. The final decision is a correct action with a confidence level of 0.98.
[0077] The chain-like causal deduction specifically includes: based on fault characteristics and protection settings, deducing the set of protection components that should theoretically operate; analyzing the multi-source heterogeneous operating data and extracting the actual output action signal sequence of the protection device; comparing the theoretical action set with the actual action sequence item by item to determine the correctness, failure to operate, or false operation status of the protection action.
[0078] Causal Chain: This refers to a portion of the analytical conclusions ultimately generated by the method of this invention. It is an explicit, traceable logical chain that clearly demonstrates the complete causal deduction path from "fault characteristics" to "protection activation," then to "criteria fulfillment," and finally to "exit action." The causal chain is the core basis output by the system for explaining whether the protection action is correct or not.
[0079] In the chain-like causal reasoning process, the system guides the large language model to perform multi-step logical reasoning through structured prompts. The core design of the prompts is as follows:
[0080] (1) System role definition and reasoning task description, as shown in Table 1 below:
[0081] Table 1
[0082]
[0083] (2) The input data format for logical reasoning is shown in Table 2 below:
[0084] Table 2
[0085]
[0086] (3) Output requirements for logical reasoning: as shown in Table 3 below. Figure 3 As shown:
[0087] Table 3
[0088]
[0089] (4) Complete prompt word examples (applicable to the analysis task of "whether the B-phase differential protection action is correct") are shown in Table 4 below:
[0090] Table 4
[0091]
[0092] The structured prompts mentioned above ensure that the large language model follows a standardized logical reasoning process when performing causal inference, and that the output results are traceable, verifiable, and practical for engineering applications.
[0093] Step 5: Confidence quantification assessment based on multidimensional indicators.
[0094] While generating conclusions, the agent outputs confidence scores through a weighted calculation model. The calculation is based on data quality indicators, action margin indicators, and logical consistency indicators. The meanings and calculation methods of each indicator are as follows:
[0095] (1) Data quality indicators The formula reflects the completeness and reliability of the input data:
[0096]
[0097] in, For the integrity rate of the recording channel, For the continuity of the sampled data, This is the timescale synchronization deviation factor. , , For the preset weight, the preferred value range is 0.2 to 0.5, and it satisfies the following conditions: By default, it can be allocated equally (each taking 1 / 3). The value ranges from 0 to 1, and the closer the value is to 1, the higher the data quality.
[0098] (2) Motion margin index The formula for calculating the reliability of protective actions is as follows:
[0099]
[0100] Right now:
[0101]
[0102]
[0103] The specific calculation formula chosen depends on the direction of the criterion. As an action threshold; This represents the actual action value; The value is normalized to the range of 0 to 1, and the closer the value is to 1, the more reliable the action is.
[0104] (3) Logical consistency index The formula for verifying the degree of mutual corroboration of the derivation conclusions of protection components based on different principles is as follows:
[0105]
[0106] in, To ensure the number of independent protection components that aligns with the deduction conclusion, This represents the total number of protection components participating in the simulation. The value ranges from 0 to 1. The closer the value is to 1, the more consistent the operating logic of each protection element is.
[0107] Final confidence level for:
[0108]
[0109] in, For the preset weight, the preferred value range is 0.2 to 0.5, and it satisfies the following conditions: , The default values are 0.2, 0.4, and 0.4 respectively. The confidence level C ranges from 0 to 1, serving as an optional grading method. "High confidence level" "Medium confidence level" This is considered a "low confidence level". The above threshold levels can be adjusted based on actual application scenarios and engineering experience.
[0110] Step 6: Human-computer interaction feedback optimization mechanism.
[0111] The system provides a human-computer interactive reasoning conclusion correction function. After the visualization module displays the causal inference conclusion, users can correct the inference conclusion through interface selection or natural language input. Upon receiving the user's correction command, the system stores the corrected reasoning conclusion as incremental knowledge, which can be used to assist in the deduction of similar fault cases in the future, achieving continuous learning and optimization based on user feedback. Figure 7 As shown.
[0112] Specifically, it includes:
[0113] Feedback parsing and standardization: Parse the user's input correction instructions into a standardized structure, which includes the fault scenario, key features, original reasoning conclusion, correction conclusion, and user confidence level;
[0114] Validity and consistency checks: Check whether the revised conclusions are fundamentally contradictory to the numerical calculation results of the scientific computing toolset, and whether they conflict with existing domain knowledge;
[0115] Conflict resolution and knowledge updating: When a newly revised conclusion conflicts with existing knowledge, the confidence level is compared to overwrite or merge the existing knowledge. If the conflict cannot be resolved automatically, the conflict is sent to expert arbitration.
[0116] Incremental learning and parameter fine-tuning: After adopting the corrected conclusions, the semantic vector index and temporal graph index of the long-term memory module are updated, and the parameters of the confidence assessment model are fine-tuned.
[0117] In the chain-like causal reasoning process, the large language model is guided to perform multi-step logical reasoning through structured prompts. The prompts include: system role definition, input data format specifications, and output step requirements. The output steps include criterion verification, standard path deviation analysis, causal chain construction, and conclusion calculation. The output results of each step are presented in the form of structured tables or explicit causal chains.
[0118] The visualization module generates a logic flow diagram and a structured analysis report. (See attached image.) Figure 5 The system adopts a bypass deployment architecture to access the relay protection fault information master station system, see [link / reference]. Figure 6 No commands are sent to the protection device. Users can correct their reasoning conclusions through interface selection or natural language input. The system stores the corrected reasoning conclusions as incremental knowledge for auxiliary deduction of subsequent similar cases. See [link to relevant documentation]. Figure 7 .
[0119] To verify the effectiveness of this invention, 100 actual relay protection operation cases (including 80 correct operations and 20 false / failed operations) from a power grid in 2023-2024 were selected for testing. The results were compared with traditional manual analysis and existing general large model methods, as shown in Table 5 below:
[0120] Table 5. Comparison of the Invention Method with Traditional Methods and Existing AI Methods
[0121]
[0122] Note: Existing general-purpose large models cannot directly parse COMTRADE format waveform files, nor can they perform numerical calculations such as Fourier transforms and impedance calculations, therefore they cannot independently complete fault analysis. "—" in the table indicates that the indicator is not applicable.
[0123] As shown in the table above, this invention, through the "neural symbolic collaboration" architecture, enables large language models to perform deterministic calculations without directly processing waveform analysis and numerical calculations. Instead, external scientific computing tools complete the calculations, thus completely solving the fundamental problem of "unreadable data and inaccurate numerical calculations" in relay protection fault analysis of general large models. This achieves efficient, accurate, and interpretable intelligent causal inference.
[0124] Implementation method of relay protection action analysis device:
[0125] The architecture diagram of the relay protection action analysis system in this embodiment is shown below. Figure 8As shown, the system includes a memory, a processor, a system bus, and a computer program stored in the memory. The processor and memory communicate and exchange data via the system bus. The processor executes the computer program to implement the steps of the relay protection action analysis method described in the embodiments of the present invention. The specific relay protection action analysis method has been described in sufficient detail in the above embodiments and will not be repeated here. The processor can be a microprocessor (MCU) or other processing system; the memory can be any type of memory that stores information using electrical energy, such as non-volatile storage media (including computer programs, databases), or other types of memory.
[0126] Implementation method of relay protection action analysis system:
[0127] A relay protection action analysis system is provided. The system includes multiple different modules, which cooperate with each other to implement the relay protection action analysis method as described in the implementation method of the relay protection action analysis method. The specific relay protection action analysis method has been described in sufficient detail in the above implementation method of the relay protection action analysis method and will not be repeated here.
[0128] A relay protection action analysis system includes:
[0129] Data access module: used to execute step 1 in the above implementation method of relay protection action analysis, to obtain external multi-source heterogeneous operation data; and to obtain the natural language description and logical execution path backbone of this fault;
[0130] The intelligent agent scheduling module is used to execute step 2 in the above-described relay protection action analysis method implementation, and is responsible for task planning and intent recognition; it is used to query similar fault cases in the historical fault case library based on the natural language description; it is used to match similar standard action sequences in the standard action sequence library based on the logical execution path backbone; it is also used to select the fault case and standard action sequence with the highest comprehensive similarity from the similar fault cases and the similar standard action sequences; the comprehensive similarity is obtained by weighting the semantic similarity corresponding to the fault case and the topological similarity corresponding to the standard action sequence; the historical fault case library stores the natural language descriptions of historical fault cases; the standard action sequence library stores the standard action sequences corresponding to all standard protection logics;
[0131] Tool Invocation Gateway: Used to execute step 3 in the above implementation method of relay protection action analysis, responsible for parsing invocation instructions and scheduling scientific computing toolsets;
[0132] Causal reasoning engine: used to execute step 4 in the above relay protection action analysis method implementation method, responsible for integrating calculation results for logical reasoning; used to inject the fault cases with the highest comprehensive similarity and standard action sequences into the thought chain as prompt words of the large language model, and generate protection action evaluation conclusions through chain causal reasoning by the large language model.
[0133] Visualization module: used to execute step 5 in the above implementation method of relay protection action analysis, responsible for generating logic flow diagram;
[0134] Feedback Learning Module: This module is used to execute step 6 in the above implementation method for relay protection action analysis. It is responsible for parsing, verifying, resolving conflicts, and incremental learning of human-machine interaction feedback.
[0135] The agent scheduling module has a built-in long-term memory unit, which includes: a vector database: storing semantic embedding vectors of historical cases; a graph database: storing directed acyclic graphs of action sequences of historical cases and graph embedding vectors; and a temporal graph neural network encoder: used to encode the logical execution path of new faults into graph embedding vectors to perform temporal topological similarity retrieval.
[0136] The feedback learning module includes: a feedback parsing unit, used to parse user correction commands into a standardized structure; a verification engine, used to perform validity and consistency checks; a conflict resolution unit, used to compare confidence levels and perform overwrite, merge, or arbitration operations; and an incremental learning unit, used to update the memory index and fine-tune the confidence model.
[0137] It also includes human-computer interaction feedback optimization steps: after the analysis conclusions are displayed in the visualization presentation module, the correction instructions input by the user are received, and the corrected reasoning conclusions are stored as incremental knowledge in the long-term memory module for auxiliary deduction of subsequent similar fault cases.
[0138] This invention has the following characteristics:
[0139] This invention constructs a "neural symbolic collaboration" intelligent agent architecture. A large language model is responsible for semantic understanding and causal reasoning, while an external scientific computing toolset handles high-precision numerical calculations. These two components are decoupled and collaborate to eliminate numerical illusions. The intelligent agent possesses a long-term memory module and employs a dual-drive memory retrieval mechanism combining temporal graph neural networks and vector retrieval. A connection is established between the historical fault case database and the standard action sequence graph database through protection type labeling and reference identifiers. For new faults, semantic similarity retrieval and temporal topology retrieval are performed simultaneously. The results of these two retrievals are weighted and fused to comprehensively evaluate the correctness of protection actions. A causal inference engine combines scientific computing results with a domain knowledge base to generate explicit causal chains and uses a multi-dimensional confidence quantification evaluation model to score the credibility of the analysis conclusions. The system uses a bypass deployment method to connect to the relay protection fault information master station system and designs an engineered human-computer interaction feedback closed-loop mechanism to achieve continuous learning and knowledge updates. This invention achieves the unification of high-precision numerical calculations and fully transparent logical inference in relay protection action analysis.
[0140] Beneficial effects:
[0141] (1) By using the “neural symbolic collaboration” architecture, the logical reasoning of large models and scientific computing tools are decoupled, thus avoiding “numerical illusion”.
[0142] (2) Innovatively, a dual-drive mechanism of "temporal graph neural network + vector retrieval" is proposed to perform dual matching of "semantic similarity + action temporal topology" on historical cases, effectively eliminating interference cases that are "similar in text description but completely different in action temporal sequence" and eliminating the illusion of contextual interference.
[0143] (3) Engineering feedback closed loop: A complete closed loop mechanism of feedback parsing standardization, validity / consistency verification, conflict resolution and incremental learning was designed to ensure that human-computer feedback can be reliably transformed into knowledge accumulation and avoid knowledge pollution and performance degradation.
[0144] (4) Confidence quantification assessment: Introduce a multi-dimensional confidence assessment mechanism to enable users to intuitively understand the credibility of the analysis results.
[0145] (5) Significant practical benefits: Compared with traditional manual analysis, it significantly shortens the fault analysis time and improves the diagnostic accuracy.
[0146] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still make modifications to the technical solutions described in the foregoing embodiments without creative effort, or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for analyzing the operation of relay protection, characterized in that, The method includes: S1. Obtain the natural language description and main logical execution path of this fault; S2. Based on the natural language description, query similar fault cases in the historical fault case database; Based on the main logical execution path, match similar standard action sequences in the standard action sequence library; The historical fault case library stores natural language descriptions of historical fault cases; the standard action sequence library stores standard action sequences corresponding to all standard protection logic. The fault case and the standard action sequence with the highest comprehensive similarity are selected from the similar fault cases and the similar standard action sequences; the comprehensive similarity is obtained by weighting the semantic similarity corresponding to the fault case and the topological similarity corresponding to the standard action sequence. S3. Inject the fault cases with the highest comprehensive similarity and the standard action sequence into the thought chain as prompt words of the large language model, and generate the protection action evaluation conclusion through chain causal inference by the large language model.
2. The relay protection action analysis method according to claim 1, characterized in that, S2 also includes: calculating the protection criterion setting value using a pre-set relay protection quantitative calculation tool; the prompt word also includes the protection criterion setting value.
3. The relay protection action analysis method according to claim 1, characterized in that, The evaluation conclusions of the protection action include: the fault type and its criterion verification, the deviation between the main logical execution path and the standard action sequence, the causal chain of this relay protection action, and whether the action is correct.
4. The relay protection action analysis method according to claim 2, characterized in that, The evaluation conclusion of the protection action also includes a confidence score, which is obtained by weighted summation of data quality index, action margin index and logical consistency index; the data quality index is used to represent the quality of fault data in the prompt words; the action margin index is used to represent the reliability of relay protection action; the logical consistency index is used to represent the logical consistency of the action of each protection element.
5. The relay protection action analysis method according to claim 1, characterized in that, The process of querying similar failure cases in the historical failure case database based on the natural language description includes: converting the natural language description into a semantic query vector, and retrieving the failure cases corresponding to the K semantic vectors with the highest semantic similarity to the semantic query vector in the historical failure case database. The historical failure case library stores historical failure cases in the form of semantic vectors.
6. The relay protection action analysis method according to claim 1, characterized in that, The process of matching similar standard action sequences in the standard action sequence library based on the logical execution path backbone includes: converting the logical execution path backbone into a graph embedding vector, and matching the M standard action sequences with the highest topological similarity to the graph embedding vector in the standard action sequence library; The standard action sequence library stores the standard action sequences using graph embedding vectors as indexes.
7. The relay protection action analysis method according to claim 4, characterized in that, The data quality indicators Motion margin index Logical consistency index They are respectively: in, For the integrity rate of the recording channel; The continuity rate of the sampled data; This is the timescale synchronization deviation factor; , and All are preset weights. ; The number of independent protection components consistent with the deduction conclusion; The total number of protection components participating in the simulation; As an action threshold; This represents the actual action value.
8. The relay protection action analysis method according to claim 1, 5, or 6, characterized in that, The method further includes: S3, after receiving the user's evaluation conclusion on the protection action, storing the knowledge of the corrected protection action evaluation conclusion in the historical fault case library and the standard action sequence library.
9. A relay protection action analysis device, comprising a processor, characterized in that, The processor is used to execute a computer program to implement the steps of the relay protection action analysis method as described in any one of claims 1 to 8.
10. A relay protection action analysis system, characterized in that, include: Data access module: used to obtain the natural language description and main logical execution path of this fault; The intelligent agent scheduling module is used to query similar fault cases in the historical fault case library based on the natural language description; match similar standard action sequences in the standard action sequence library based on the logical execution path backbone; and select the fault case and standard action sequence with the highest comprehensive similarity from the similar fault cases and the similar standard action sequences. The comprehensive similarity is obtained by weighting the semantic similarity corresponding to the fault case and the topological similarity corresponding to the standard action sequence. The historical failure case database stores natural language descriptions of historical failure cases; The standard action sequence library stores the standard action sequences corresponding to all standard protection logic; Causal inference engine: It is used to inject the fault cases with the highest comprehensive similarity and standard action sequences into the thought chain as prompt words of the large language model, and generate protection action evaluation conclusions through chain causal inference by the large language model.