Typhoon weather-based power grid material fault prediction and active defense system and method based on digital twinning and multi-agent large model

Abstract

The application discloses a power grid material fault prediction and active defense system and method based on digital twinning and a multi-agent large model, and belongs to the technical field of power system disaster prevention. The system comprises a knowledge graph construction module, a multi-agent collaborative prediction module and a digital twinning simulation pre-play module. By constructing a "disaster-equipment" knowledge graph integrating multi-source data, a structured knowledge base is provided for the system; by using three-level agents of risk perception, probability prediction and strategy generation to work collaboratively, closed-loop decision-making from typhoon warning analysis to executable defense strategy generation is realized; finally, the scheduling scheme is dynamically simulated, deduced and optimized and verified in a digital twinning environment. The application solves the problems of difficult data fusion, disconnection between prediction and decision-making and lack of verification mechanism in traditional methods, and significantly improves the fault prediction capability, emergency response speed and active defense reliability of the power grid in typhoon weather.

Description

Technical Field

[0001] This invention relates to the field of power system disaster prevention, mitigation and intelligent operation and maintenance technology, and in particular to a system and method for predicting and actively defending against faults in key power grid materials during typhoon weather by integrating digital twins, multi-agent large models and knowledge graphs. Background Technology

[0002] As a critical national infrastructure, the safe and stable operation of the power grid is of paramount importance. In recent years, extreme weather events such as typhoons and ice storms have occurred frequently, posing a serious threat to key power grid materials such as transformers and power poles, and easily leading to large-scale power outages. Achieving accurate prediction and rapid proactive defense against power grid material failures under extreme weather conditions is a major challenge for the safe operation of the power system.

[0003] Traditional power grid disaster early warning and emergency decision-making rely heavily on statistical analysis models and human experience, which has significant drawbacks: First, traditional models struggle to integrate heterogeneous data from multiple sources, such as meteorology, geography, equipment records, and operational status, resulting in complex feature engineering and poor generalization capabilities. Second, the prediction models are disconnected from emergency response decisions, failing to form a closed loop from risk perception to dispatch execution. Third, existing methods lack a deep understanding and reasoning ability regarding the complex correlation between disaster chain evolution and equipment failures, leading to poor interpretability of prediction results and difficulty in supporting high-confidence operation and maintenance decisions.

[0004] In recent years, digital twin and large model technologies have shown great potential. However, general large models are insufficient in terms of professional knowledge and understanding of complex mechanisms in the field of power grid disaster prevention, and the reasoning process is prone to "machine illusion". On the other hand, existing digital twin systems mostly focus on condition monitoring and visualization, and lack deep integration with predictive decision-making models, thus failing to form an intelligent closed loop of "prediction-decision-simulation-optimization".

[0005] Therefore, there is an urgent need for an integrated intelligent system that can deeply integrate domain knowledge, achieve accurate quantitative prediction, and simulate and verify decision-making schemes, so as to enhance the resilience and proactive defense capabilities of the power grid in the face of extreme weather. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a power grid material failure prediction and active defense system and method based on digital twin and multi-agent large model, so as to realize accurate prediction of failure probability of key power grid materials, risk assessment, generation of defense strategy and simulation of scheduling scheme under typhoon conditions.

[0007] To achieve the above objectives, this invention provides a power grid material fault prediction and proactive defense system based on digital twins and multi-agent large models, comprising: The knowledge graph construction module is responsible for building and storing the "disaster-equipment" knowledge graph, which deeply integrates typhoon disaster-causing factors (strong winds, heavy rain, storm surge), geographical environmental parameters, inherent attributes of power grid equipment, historical failure modes, and emergency response strategies. A multi-agent collaborative prediction module is connected to the knowledge graph construction module and is used to perform hierarchical collaborative reasoning based on the knowledge graph to generate fault probability prediction and defense strategies. The digital twin simulation and pre-play module is connected to the multi-agent collaborative prediction module and is used to receive the generated defense and scheduling strategies and to perform simulation and optimization of material scheduling paths in a virtual power grid environment. The knowledge graph construction module includes: The knowledge graph structure design unit is responsible for defining the ontology structure of the "disaster-equipment" knowledge graph, covering core concepts such as typhoon disasters (e.g., wind level, path, rainfall), geographical environment, power grid materials, failure modes, and response plans, as well as their interrelationships. The entity recognition and relation extraction unit, based on natural language processing technology, automatically extracts entities and relations from multi-source texts such as historical disaster reports and technical documents to construct and enrich the knowledge graph. The knowledge storage and update unit is responsible for storing, indexing, and dynamically updating the graph data, supporting the rapid import and integration of new knowledge.

[0008] The multi-agent cooperative prediction module includes: The risk perception agent is responsible for receiving typhoon warning information, conducting preliminary screening and association recall of risk areas and key equipment based on the knowledge graph, and deciding whether to transfer the prediction task to the probability prediction agent based on the preset confidence threshold. The probabilistic predictive agent is responsible for making refined probabilistic predictions of specific power grid equipment fault types. This agent integrates the correlation reasoning ability of large models with the power system risk assessment quantification algorithm to form a two-layer architecture of fault hypothesis generation and quantitative risk assessment, and outputs quantified fault probability and comprehensive risk value. The strategy generation agent is responsible for predicting the output of the probability prediction agent, accessing the local emergency response case library, and generating defense strategy text containing specific material allocation plans through retrieval enhancement generation technology.

[0009] The risk perception agent, as the system's front-line perception unit, is responsible for preliminary and rapid analysis and risk localization of typhoon warnings. It includes: Early warning information analysis unit: responsible for receiving and analyzing standardized early warning information from meteorological departments, and extracting key disaster-causing factors, such as the typhoon's center location, movement path, maximum wind speed, radius of influence, and rainfall intensity.

[0010] Knowledge Graph Fast Retrieval Unit: Based on the parsed disaster-causing factors, it matches and traverses the "Disaster-Equipment" knowledge graph to quickly recall geographical areas, equipment types, and known vulnerabilities that have a high historical correlation with the weather pattern.

[0011] Regional Risk Confidence Assessment Unit: This unit performs a preliminary analysis of the search results and calculates the overall risk confidence score for the affected areas. This confidence score integrates the degree of match between weather intensity and equipment vulnerability. Its output is a quantified confidence score, Cr.

[0012] Task referral decision unit: A confidence threshold τ is preset. When Cr < τ, the risk is considered low and only monitoring is required; when Cr ≥ τ or the region belongs to the preset extremely high risk set Xs, a task referral instruction is automatically generated, and the specific equipment list and risk clues are transferred to the probabilistic predictive agent for in-depth analysis.

[0013] The probabilistic predictive agent is the core computation and inference engine of the system, responsible for refining the predictions of specific devices. Failure probability prediction. It employs a two-layer architecture, specifically including: The fault hypothesis generation layer performs deep correlation reasoning based on risk cues received from the risk-aware agent. Specifically, it includes: Evidence fusion unit: integrates real-time weather data, the ledger attributes of the target equipment (such as wind resistance level, years of operation) and its current operating status.

[0014] Hypothesis reasoning unit: Based on the fused evidence, it queries and infers all possible fault modes in the knowledge graph, forming a set of fault hypotheses to be calculated, H={h1,h2,...,hm}. For example, it generates the hypothesis "Transformer A experiences bushing rupture at a wind speed of 52m / s".

[0015] The quantitative risk assessment layer provides precise mathematical calculations for each failure hypothesis. Specifically, it includes: The unit calculates the probability of service disruptions: it incorporates a multi-state weather model based on IEEE standards (such as normal, severe, etc.). (Catastrophe), calculating the probability of equipment downtime under different operating conditions based on weather condition parameters. Consequence Severity Assessment Unit: Using an exponential utility function, the severity of the failure, overload, and voltage limit exceedance can be calculated respectively.

[0016] Comprehensive risk calculation unit: Based on the probability of accident occurrence and the severity of consequences, calculate the underload risk index, overload risk index and voltage over-limit risk index, and obtain the comprehensive risk assessment value by weighted summation.

[0017] Probability Integration and Output Unit: Combines the prior probability provided by the knowledge graph with the conditional probability obtained by quantitative calculation, and outputs the final predicted probability of the fault hypothesis through Bayesian method or average weighting, and simultaneously outputs its comprehensive risk value and the mapped standard risk level.

[0018] Policy-generating agent: Responsible for transforming predictions into actionable plans. It includes: The strategy retrieval unit receives a list of high-probability faults output by the probability prediction agent. It vectorizes the fault information using pre-trained models such as BERT and employs efficient similarity retrieval tools like FAISS to quickly retrieve the top K historical response case text blocks most similar to the current situation from a locally built emergency response case library.

[0019] Hint Engineering and Context Building Unit: Combines retrieved case text, current fault prediction details, available resource list, and other information according to a preset hint template to build a hint rich in contextual information.

[0020] Strategy generation and optimization unit: Input the constructed prompts into the base model to generate preliminary and professional defense strategies and resource allocation scheme texts.

[0021] Strategy formatting and output unit: The generated text strategy is structured, key actions, required materials, starting point, destination and other elements are extracted and encapsulated into a standardized instruction format, and then output to the digital twin simulation pre-play module.

[0022] The digital twin simulation and pre-visualization module includes: The 3D virtual scene construction unit, based on the geographic information system and equipment model library, constructs a high-precision restored 3D virtual scene of the power grid; the real-time data-driven unit synchronously maps real-time meteorological data, equipment status and forecast results to the virtual scene. The route planning and simulation unit, based on the scheduling requirements output by the strategy-generating agent and combined with the real-time impact range of the typhoon (such as the wind circle and rain belt), dynamically plans multiple alternative material transportation routes and conducts parallel simulation to evaluate the travel time and safety risks of each route. The scheme optimization and output unit compares and analyzes the simulation results, selects and outputs the comprehensive optimal material scheduling path and its estimated parameters.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By combining knowledge graph-based associative reasoning with quantitative risk assessment models, probabilistic and quantifiable prediction results can be generated for specific equipment and specific failure modes, and the reasoning basis can be provided, which significantly improves the accuracy, granularity and interpretability of predictions and supports high-confidence decision-making. 2. It has achieved full automation of the entire process from receiving early warnings to outputting structured scheduling instructions, freeing maintenance personnel from tedious information screening and solution design, and reducing decision response time from hours to minutes, greatly improving the efficiency of emergency response.

[0024] 3. Through digital twin simulation and rehearsal, resource allocation routes can be verified and optimized in virtual space in advance, mitigating execution risks caused by weather and road conditions, and outputting the optimal solution that has been "stress-tested." Combined with security verification of the strategy, this provides dual assurance for the feasibility and compliance of the solution in the real world. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall system framework in Embodiment 1 of the present invention; Figure 2 This is a block diagram of the knowledge graph construction module in this invention; Figure 3 This is a block diagram of the multi-agent collaborative prediction module in this invention; Figure 4 This is a block diagram of the digital twin simulation pre-visualization module in this invention; Figure 5 This is a flowchart illustrating the overall process of the fault prediction and active defense system in this invention. Figure 6 This is a diagram illustrating the knowledge graph construction process in this invention. Figure 7 This is a diagram illustrating the multi-agent collaborative prediction process of the present invention. Detailed Implementation The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, so that the objectives, technical solutions, and advantages of the present invention will be clearer. It should be noted that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] like Figure 1 As shown, this embodiment provides a power grid material fault prediction and proactive defense system based on digital twins and a multi-agent large model. The system includes a knowledge graph construction module, a multi-agent collaborative prediction module, and a digital twin simulation module. These three modules are connected sequentially to form a closed loop of data and decision flow.

[0027] The knowledge graph construction module is responsible for building and storing the "disaster-equipment" knowledge graph, which deeply integrates typhoon disaster-causing factors (strong winds, heavy rain, storm surge), geographical environmental parameters, inherent attributes of power grid equipment, historical failure cases, and emergency response strategies. A multi-agent collaborative prediction module, connected to the knowledge graph construction module, is used for hierarchical collaborative reasoning based on the knowledge graph to generate fault probability prediction and defense strategies; it consists of three agents with clearly defined functions and working collaboratively. - Risk-aware intelligent agent: As the system's front-line sentinel, it receives and analyzes typhoon warning information in real time, and through rapid querying of the knowledge graph, it initially screens out high-risk areas and key equipment, and determines whether to initiate the deep prediction process based on the quantified risk confidence level.

[0028] - Probabilistic Predictive Agent: Upon receiving a high-risk task, it adopts a two-layer architecture of "fault hypothesis generation" and "quantitative risk assessment". The upper layer generates specific fault hypotheses based on knowledge graphs and real-time data reasoning; the lower layer integrates power system risk assessment quantification algorithms (such as multi-state weather models and utility functions) to calculate the probability of occurrence of each hypothesis and the comprehensive risk value, and outputs accurate quantitative prediction results.

[0029] - Strategy Generation Agent: Based on prediction results, it employs Retrieval Augmentation (RAG) technology to recall similar historical cases from the local emergency response case database. Combined with the current context, it generates specialized and actionable defense strategy texts using a large model, and structures these texts into standardized dispatch instructions containing resources, routes, and actions. The digital twin simulation pre-play module is connected to the multi-agent collaborative prediction module and receives standardized scheduling instructions. This module constructs a high-precision three-dimensional virtual power grid scene and injects real-time meteorological data and prediction results. Under the premise of considering the dynamic impact of typhoons, it performs multi-path parallel simulation and deduction of material scheduling schemes, evaluates the timeliness and safety risks of each path, and outputs the comprehensive optimal scheduling execution scheme through comparison and optimization, forming a decision verification closed loop.

[0030] This embodiment utilizes a knowledge graph construction module to form a structured domain knowledge base, providing precise semantic constraints for intelligent agents. A multi-agent collaborative prediction module simulates expert decision-making, realizing an intelligent pipeline from risk perception and accurate prediction to strategy generation. A digital twin simulation module verifies and optimizes scheduling schemes in a virtual space, forming a decision-making closed loop. These modules work collaboratively to form an integrated proactive defense system encompassing perception, prediction, decision-making, and simulation, significantly improving the intelligence and reliability of the power grid's emergency response to typhoon disasters.

[0031] A method for predicting and actively defending power grid material faults based on the above system includes the following steps: (1) constructing and dynamically updating a "disaster-equipment" knowledge graph; (2) based on the knowledge graph, achieving end-to-end intelligent decision-making from early warning analysis to defense strategy generation through the collaboration of three-level intelligent agents: risk perception, probability prediction, and strategy generation; (3) performing simulation and multi-scheme comparison and optimization of the generated defense strategy in a digital twin environment, and outputting the verified executable scheme.

[0032] Figure 2 The figure shows a block diagram of the knowledge graph construction module in this invention. The knowledge graph construction module includes: The knowledge graph structure design unit is responsible for defining the ontology structure of the "disaster-equipment" knowledge graph, covering core concepts such as typhoon disasters (e.g., wind level, path, rainfall), geographical environment, power grid materials, failure modes, and response plans, as well as their interrelationships. The entity recognition and relation extraction unit, based on natural language processing technology, automatically extracts entities and relations from multi-source texts such as historical disaster reports and technical documents to construct and enrich the knowledge graph; The knowledge storage and update unit is responsible for storing, indexing, and dynamically updating the graph data, and supports the rapid import and integration of new knowledge.

[0033] like Figure 3 As shown, the multi-agent cooperative prediction module includes: The risk perception agent is responsible for receiving typhoon warning information, performing preliminary screening and recall of risk areas and key equipment based on the knowledge graph, and determining whether to transfer the prediction task to the probabilistic prediction agent based on a preset confidence threshold; specifically including: --Early Warning Information Analysis Unit: Responsible for receiving and analyzing standardized early warning information from meteorological departments, and extracting key disaster-causing factors, such as the typhoon's center location, movement path, maximum wind speed, radius of influence, and rainfall intensity.

[0034] --Knowledge Graph Fast Retrieval Unit: Based on the parsed disaster-causing factors, it matches and traverses the "Disaster-Equipment" knowledge graph to quickly recall geographical areas, equipment types, and known vulnerabilities that have a high historical correlation with the weather pattern.

[0035] --Regional Risk Confidence Assessment Unit: Performs preliminary analysis of the search results and calculates the comprehensive risk confidence score for the affected areas. This confidence score integrates the degree of match between weather intensity and equipment vulnerability. Its output is a quantified confidence score Cr.

[0036] --Task referral decision unit: A confidence threshold τ is preset. When Cr < τ, the risk is considered low and only monitoring is required; when Cr ≥ τ or the region belongs to the preset extremely high risk set Xs, a task referral instruction is automatically generated, and the specific equipment list and risk clues are transferred to the probabilistic predictive agent for in-depth analysis.

[0037] The probabilistic predictive agent is the core computing and inference engine of the system, responsible for performing refined fault probability predictions for specific devices. Internally, it employs a two-layer architecture, specifically including: (i) Fault hypothesis generation layer, which performs deep correlation reasoning based on risk cues received from the risk-aware agent. This includes: Evidence fusion unit: integrates real-time weather data, the ledger attributes of the target equipment (such as wind resistance level, years of operation) and its current operating status.

[0038] Hypothesis reasoning unit: Based on the fused evidence, it queries and infers all possible fault modes in the knowledge graph, forming a set of fault hypotheses to be calculated, H={h1,h2,...,hm}. For example, it generates the hypothesis "Transformer A experiences bushing rupture at a wind speed of 52m / s".

[0039] (II) Quantitative Risk Assessment Layer: This layer provides precise mathematical calculations for each failure hypothesis. Specifically, it includes: The unit calculates the probability of service disruptions: it incorporates a multi-state weather model based on IEEE standards (such as normal, severe, etc.). (Catastrophe), calculating the probability of equipment downtime under different operating conditions based on weather condition parameters. Consequence Severity Assessment Unit: Using an exponential utility function, the severity of the failure, overload, and voltage limit exceedance can be calculated respectively.

[0040] Comprehensive risk calculation unit: Based on the probability of accident occurrence and the severity of consequences, calculate the underload risk index, overload risk index and voltage over-limit risk index, and obtain the comprehensive risk assessment value by weighted summation.

[0041] Probability Integration and Output Unit: Combines the prior probability provided by the knowledge graph with the conditional probability obtained by quantitative calculation, and outputs the final predicted probability of the fault hypothesis through Bayesian method or average weighting, and simultaneously outputs its comprehensive risk value and the mapped standard risk level.

[0042] Policy-generating agent: Responsible for transforming predictions into actionable plans. Specifically includes: --Policy Retrieval Unit: Receives a list of high-probability faults output by the probability prediction agent. The fault information is vectorized using pre-trained models such as BERT, and efficient similarity retrieval tools such as FAISS are used to quickly retrieve the top K historical response case text blocks most similar to the current situation from a locally built emergency response case library.

[0043] --Prompt Engineering and Context Building Unit: Combines retrieved case text, current fault prediction details, available resource list, and other information according to a preset prompt template to build a prompt rich in contextual information.

[0044] --Strategy Generation and Optimization Unit: Inputs the constructed prompts into the base model to generate preliminary, specialized defense strategies and resource allocation scheme texts.

[0045] --Strategy Formatting and Output Unit: This unit structures the generated text strategy, extracts key actions, required resources, departure point, destination, and other elements, and encapsulates them into a standardized instruction format. It then outputs this information to the digital twin simulation pre-play module and the visualization and decision support module.

[0046] The strategy generation agent employs "retrieval-enhanced generation" technology, combining the general generation capabilities of large models with emergency response knowledge specific to the power grid domain. This ensures the professionalism, accuracy, and operability of the generated strategies. It addresses the issue of insufficient knowledge in vertical domains within general large models, ensuring that the generated strategies are no longer vague suggestions but specific, feasible, and historically supported precise solutions. Through the precise division of labor and collaboration among these three agents, this invention achieves end-to-end intelligent decision-making from "weather warnings" to "executable and verifiable dispatch instructions," significantly enhancing the power grid's proactive defense capabilities against typhoons.

[0047] like Figure 4 As shown, the digital twin simulation pre-playing module includes a 3D virtual scene construction unit, a real-time data-driven unit, a path planning and deduction unit, and a scheme optimization and output unit. Its specific workflow and analysis and optimization process are as follows: 3D Virtual Scene Construction Unit: This unit forms the static foundation of the digital twin system. It integrates Geographic Information System (GIS) data, a 3D model library of power grid equipment, and power grid topology data to construct a 3D virtual scene. This scene not only includes geographical features such as terrain, roads, and buildings, but also precisely integrates physical models of key equipment such as transformers, towers, and lines. It also binds equipment ledger attributes (such as coordinates, model, and wind resistance rating) and status interfaces to these models, providing a precise geometric and physical information foundation for subsequent dynamic simulations.

[0048] Real-time Data-Driven Unit: This unit acts as a bridge connecting the virtual world and real-time dynamics, injecting "life" into static scenes. It receives and integrates two types of real-time data streams: dynamic disaster data from weather forecasts and equipment risk prediction results from upstream probabilistic predictive agents (PPAs). Through data spatiotemporal synchronization and mapping technology, this unit drives the real-time calculation of the "weather-disaster" model in the virtual scene, dynamically presenting the disaster effects of wind, rain, and flooding under the influence of typhoons. It also overlays equipment risks as visual layers (such as risk heatmaps marked with different colors) onto the 3D scene, thereby constructing a dynamic risk environment that reflects the current and future short-term disaster situation.

[0049] Path Planning and Derivation Unit: This unit is the core of solution generation and preliminary verification. It receives structured scheduling instructions from the Policy Generating Agent (SGA). In this dynamic risk environment, this unit performs the following key steps: Multi-path dynamic programming: Instead of calculating a single shortest path, it is guided by multiple objectives such as "timeliness and safety" and uses the A* algorithm, which integrates risk-cost graphs, to simultaneously plan multiple alternative material transportation routes. These routes form a set of differentiated solutions in terms of distance, risk exposure, and estimated travel time.

[0050] Solution Optimization and Output Unit: This unit is the final stage of decision support and closed-loop optimization. It comprehensively evaluates and optimizes all simulated alternative paths. Multi-attribute comprehensive evaluation: Establish an evaluation index system covering "timeliness" (total time consumption), "safety" (comprehensive risk value), and "reliability" (success rate of passage). Employ multi-attribute decision-making methods (such as TOPSIS and weighted scoring) to quantitatively score each route. The system can flexibly adjust the weights of each indicator according to the real-time focus of emergency command (e.g., focusing on timeliness in the initial stage of repairs, and focusing on risk during periods sensitive to personnel safety).

[0051] Optimal Solution Generation and Output: Based on the comprehensive evaluation results, the optimal scheduling path is recommended, and its key decision parameters are output, such as: optimal path number, detailed trajectory, estimated total time, major risk areas encountered, and avoidance measures. Simultaneously, a comparative analysis of several top-ranked solutions is provided to assist decision-makers in making the final decision.

[0052] like Figure 5 As shown, this embodiment of the invention presents the overall process of a power grid material fault prediction and proactive defense method based on digital twins and multi-agent large models.

[0053] Step 01: Through the knowledge graph construction module, multi-source data is systematically integrated to form a structured domain knowledge base. The entity and relation extraction unit automatically extracts key information from unstructured text, and the graph structure construction unit organizes this information according to the power grid domain ontology model. A dynamic update mechanism ensures the knowledge base can continuously evolve, providing accurate and timely knowledge support for intelligent agent reasoning. Through this systematic knowledge graph construction process, the automated conversion from multi-source heterogeneous data to structured knowledge is achieved, providing a reliable knowledge foundation for subsequent intelligent reasoning and solving problems such as lagging knowledge updates and data silos in traditional methods. Specifically, this includes: Step 01_1: Data Acquisition and Preprocessing Stage. The system collects data from multiple heterogeneous data sources, including meteorological departments, geographic information systems, power grid SCADA systems, and equipment ledger databases. The data cleaning unit removes noisy data and outliers, while the data standardization unit converts data from different sources into a unified format, laying the foundation for subsequent processing.

[0054] Step 01_2: Entity Recognition and Relation Extraction Stage: A BERT-BiLSTM-CRF hybrid model is used for named entity recognition. First, the BERT model is used to obtain the deep semantic representation of the text. Then, BiLSTM is used to capture the contextual information of the sequence. Finally, the CRF layer is used to decode the label sequence to accurately identify the entity boundaries and types in the text. The relation extraction unit is based on an attention mechanism to identify the semantic relationships between entities.

[0055] Step 01_3: Knowledge Graph Construction Phase: Based on the experience of experts in the power grid field, an ontology model is designed, clarifying the conceptual hierarchy and relational constraints. Knowledge fusion technology is used to resolve conflicts between different data sources, constructing a high-quality knowledge graph. A quality assessment unit checks the completeness and consistency of the constructed graph to ensure knowledge quality.

[0056] Step 01_4: Knowledge Storage and Update Stage: The constructed knowledge graph is stored in a graph database, establishing an efficient indexing mechanism. A dynamic update mechanism monitors the arrival of new data and automatically triggers incremental updates to the knowledge graph, maintaining the timeliness of knowledge.

[0057] Step 02: The multi-agent collaborative prediction module simulates human expert decision-making. The risk perception agent first performs rapid screening, effectively filtering low-risk scenarios. When high risk is detected, the probability prediction agent initiates deep analysis, ensuring prediction accuracy through a two-layer architecture of "hypothesis generation-quantitative evaluation." The policy generation agent ultimately transforms the prediction results into actionable plans. Through the precise division of labor and collaboration among the three agents, the prediction task is progressively deepened, ensuring both system response speed and the accuracy of prediction results and the operability of the strategies, effectively simulating the hierarchical decision-making process of human experts.

[0058] Step 02_1: In the risk perception agent processing stage, the early warning information parsing unit receives and parses standardized meteorological early warnings, extracting key disaster-causing factors (such as central pressure, movement speed, and wind circle radius). The knowledge graph retrieval unit quickly matches relevant regions and equipment in the knowledge graph based on these factors. The confidence calculation unit calculates the regional risk confidence level by considering factors such as weather intensity and equipment vulnerability. The decision-making unit determines whether further analysis is needed based on preset thresholds.

[0059] Step 02_2: In the probability prediction agent processing stage, the evidence fusion unit integrates real-time data and equipment attributes. The fault hypothesis generation unit infers possible fault modes based on a knowledge graph. The outage probability calculation unit calculates the base probability using a multi-state weather model. The consequence severity assessment unit quantifies the fault impact through a utility function. The comprehensive risk calculation unit integrates the probability and consequence, outputting a quantified risk value. The probability integration unit finally generates a fault probability prediction.

[0060] In general step 02_3: the strategy generation agent processing stage, the case retrieval unit searches for similar historical cases in the emergency response database. The context construction unit combines case information with the current context. The strategy generation unit generates specialized defense strategies based on a large model. The strategy formatting unit converts the strategies into standardized instructions.

[0061] Step 03: The digital twin simulation pre-run module verifies the feasibility of the solution in virtual space. Through multi-path parallel simulation, the system can identify potential problems before actual execution, optimize the scheduling scheme, and significantly reduce execution risks.

[0062] This method organically integrates knowledge graphs, multi-agent systems, and digital twins to construct a complete "perception-prediction-decision-rehearsal" intelligent decision-making chain, realizing full-process intelligence from data to knowledge, from knowledge to decision, and from decision to verification, which greatly enhances the power grid's proactive defense capabilities against typhoon disasters.

Claims

1. A power grid material fault prediction and proactive defense system based on digital twin and multi-agent large model, characterized in that, include: The knowledge graph module is used to build and store a "disaster-equipment" knowledge graph that integrates typhoon-causing factors, geographical environmental parameters, power grid equipment attributes, historical failure modes, and response strategies. The intelligent agent collaborative decision-making module is connected to the knowledge graph module and is used to output defense strategy instructions for key materials of the power grid based on the knowledge graph through a three-level intelligent agent collaborative workflow generated by risk perception, fault probability prediction and defense strategy. The digital twin simulation verification module is connected to the intelligent agent collaborative decision-making module. It is used to receive the defense strategy instructions, perform dynamic simulation and optimization of the material scheduling scheme in a high-fidelity virtual power grid environment, and output the verified executable scheme. The knowledge graph module, the intelligent agent collaborative decision-making module, and the digital twin simulation verification module are connected in series.

2. The power grid material fault prediction and active defense system based on digital twin and multi-agent large model as described in claim 1, characterized in that, The knowledge graph module defines an ontology structure covering typhoon disasters, geographical environment, power grid materials, fault modes and handling solutions, and performs semantic fusion and storage of multi-source heterogeneous data.

3. The power grid material fault prediction and active defense system based on digital twin and multi-agent large model according to claim 1, characterized in that, The intelligent agent collaborative decision-making module includes: The risk-aware intelligent agent is used to analyze typhoon warning information and query the knowledge graph to output a list of risk areas and key equipment. A probabilistic prediction agent, connected to the risk perception agent, is used to receive the list and, using a two-layer architecture of "fault hypothesis generation and quantitative risk assessment", outputs the quantitative fault probability and comprehensive risk value of the power grid equipment. A strategy generation agent, connected to the probability prediction agent, is used to generate structured defense strategy instructions based on the quantified failure probability and comprehensive risk value, combined with historical emergency response cases.

4. The power grid material fault prediction and active defense system based on digital twin and multi-agent large model according to claim 1, characterized in that, The fault hypothesis generation layer infers and generates a set of fault hypotheses to be evaluated based on evidence fused from the knowledge graph and real-time data. The quantitative risk assessment layer calculates the outage probability based on physical mechanisms and the severity of consequences based on power grid flow for each hypothesis in the set of fault hypotheses, and then synthesizes them to obtain a quantitative risk value.

5. The power grid material fault prediction and active defense system based on digital twin and multi-agent large model according to claim 3, characterized in that, The strategy generation agent employs retrieval-enhanced generation technology. It performs similarity matching between the current fault scenario and a historical emergency response case library, and drives a large language model to generate defense strategy text based on the matching results. The text is then structured and encapsulated into the defense strategy instructions.

6. The power grid material fault prediction and active defense system based on digital twin and multi-agent large model according to claim 1, characterized in that, The digital twin simulation verification module injects typhoon path and intensity forecast data in real time, performs parallel simulations of multiple schemes for the material dispatching paths included in the defense strategy instructions in a virtual environment, and compares and optimizes the schemes based on the travel time and safety risk assessment results obtained from the simulation.

7. The power grid material fault prediction and active defense system based on digital twin and multi-agent large model according to claim 3, characterized in that, It also includes a feedback learning module, which feeds back the optimization results and actual execution effects of the digital twin simulation verification module as new knowledge or cases to the knowledge graph module and the case library of the policy generation agent, so as to realize the self-evolution of the system.

8. A method for predicting and proactively preventing power grid material failures based on digital twins and multi-agent large models, characterized in that, Includes the following steps: S1: Construct and maintain a "disaster-equipment" knowledge graph to achieve semantic fusion of multi-source heterogeneous data; S2: Based on the knowledge graph, a three-level collaborative intelligent decision-making process is generated through risk perception, fault probability prediction, and defense strategy to generate structured power grid material defense strategy instructions. S3: In a digital twin environment, combined with real-time typhoon dynamic data, the defense strategy instructions are simulated and optimized using multiple schemes, and a verified executable scheduling scheme is output.

9. The method for predicting and actively defending power grid material faults based on digital twins and multi-agent large models according to claim 8, characterized in that, The fault probability prediction step in step S2 includes: generating fault hypotheses by performing correlation reasoning based on knowledge graphs and real-time data; calculating the outage probability using a multi-state weather model and calculating the severity of the consequences by combining power grid flow analysis, and comprehensively obtaining a quantitative risk assessment result.

10. The method for predicting and actively defending power grid material faults based on digital twins and multi-agent large models according to claim 8, characterized in that, The defense strategy generation step in step S2 includes: retrieving historical handling cases similar to the current risk situation; using the cases as context to drive a large language model to generate specialized strategy text; and parsing the strategy text into structured instructions containing resources, paths, and actions.

Images