Power grid flood disaster intelligent early warning and disposal system and method based on digital twinning and multi-agent collaboration
Patent Information
- Application Number
- CN202610157022.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-04
- Publication Date
- 2026-06-23
Smart Images
Figure CN122264507A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system disaster prevention, mitigation, and emergency management technology. Specifically, it relates to an intelligent early warning and response system and method for power grid flood disasters based on digital twins and multi-agent collaboration. It is particularly suitable for complex terrain areas such as mountainous and karst regions, aiming to solve the problems of insufficient accuracy in power grid flood disaster early warning, difficulty in quantifying equipment risk assessment, and static and rigid emergency response plans lacking forward-looking verification in existing technologies. It achieves a closed-loop intelligent response throughout the entire process, from disaster perception and accurate risk diagnosis to intelligent decision-making and virtual simulation of solutions. Background Technology
[0002] With global climate change, extreme heavy rainfall has led to frequent floods, posing a serious threat to power infrastructure. In areas with complex terrain, flash floods and urban flooding are characterized by their suddenness, destructive power, and difficulty in prediction. They can easily cause accidents such as the collapse of transmission towers, flooding of substations, and damage to power distribution facilities, resulting in widespread power outages and significant economic and social losses.
[0003] Currently, existing technologies for power grid response to floods suffer from the following shortcomings: First, at the disaster early warning level, traditional methods are mostly based on site rainfall threshold alarms or macro-hydrological model outputs. The correlation between early warning information and the physical location of power grid equipment is weak, failing to provide precise disaster-causing parameters such as water depth and flow velocity at the equipment level, leading to ambiguous risk positioning. Second, at the equipment risk assessment level, current methods largely rely on historical statistics or expert experience to set fixed risk levels, failing to establish a coupled analysis model between dynamic disaster intensity and the equipment's inherent vulnerability, resulting in insufficiently refined and quantifiable assessment results. Finally, at the emergency decision-making level, response plans typically rely on static contingency plans, lacking the ability to dynamically match and intelligently generate plans based on real-time disaster evolution and similar historical cases, making it difficult to quickly formulate optimal and operable disaster relief and recovery strategies.
[0004] While existing research has attempted to apply neural networks for fault prediction or utilize case-based reasoning to assist decision-making, most studies focus on single aspects and have yet to form a closed-loop intelligent system that seamlessly integrates "high-precision disaster simulation," "equipment-level fault probability prediction," and "intelligent response solution generation." Furthermore, general artificial intelligence models, lacking the constraints of specialized knowledge in the power sector, are prone to producing outputs that do not align with engineering realities—a phenomenon known as "machine illusion."
[0005] Existing technical solutions are mostly open-loop decision-making systems, i.e., a "data input - model calculation - result output" model, lacking the ability to simulate and iteratively optimize decision-making solutions in a virtual environment. Decision-makers find it difficult to predict the effects and risks of different strategies before taking action, resulting in a high degree of uncertainty in the final decision.
[0006] Therefore, it is necessary to propose an intelligent system that can deeply integrate physical mechanisms and data-driven approaches, support multi-agent collaboration and digital twin simulation, and possess full-process closed-loop decision-making capabilities, in order to improve the scientific nature, accuracy, and foresight of power grid flood disaster emergency management. Summary of the Invention
[0007] To overcome the problems of insufficient accuracy in early warning of power grid flood disasters, difficulty in quantifying equipment risk assessment, static and rigid emergency response plans, and lack of forward-looking verification in existing technologies, this invention aims to provide an intelligent early warning and response system and method for power grid flood disasters based on digital twins and multi-agent collaboration. This system constructs a hierarchical multi-agent collaboration framework with digital twins as carriers and management agents as the central hub, and deeply integrates domain knowledge graphs, physical mechanism models, and data-driven models to achieve fully automated and intelligent closed-loop decision support from dynamic disaster perception and precise equipment diagnosis to intelligent plan generation and virtual simulation verification.
[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an intelligent early warning and response system for power grid flood disasters based on digital twins and multi-agent collaboration, characterized in that it includes: a data access and preprocessing module, a knowledge graph in the field of power grid flood disasters, a digital twin environment uniformly managed by digital twin agents, and digital twin agents (DTA), monitoring and early warning agents (MWA), fault diagnosis agents (FDA), and case handling agents (CHA) that collaborate in the digital twin environment through a message middleware. The data access and preprocessing module is used to standardize the access and processing of multi-source data, which includes at least: grid meteorological forecast data, hydrological monitoring data, digital elevation model data, soil data, geographic information and attribute parameters of power grid equipment, and structured historical case data; A knowledge graph for the field of power grid flood disasters is stored in the form of a graph database. Its node entities include at least: power grid equipment, geographical units, disaster events, failure modes, and response measures; its edge relationships describe at least the association between: equipment location, event cause, and measures targeted. The Digital Twin Agent (DTA), serving as the system's management hub and virtual space scheduler, is configured to: manage and maintain the digital twin of the power grid and geographical environment, achieving data synchronization and mapping between physical entities and virtual models; provide model library management and computing resource scheduling services; receive and execute simulation tasks initiated by other agents, create parallel simulation branches in the digital twin, simulate the execution process of the disposal plan and perform "hypothesis analysis," and provide a "Solution Simulation Evaluation Report" containing quantitative evaluation results. The core scheduling and deduction functions of the digital twin intelligent agent are implemented through the following modules: ① Virtual-to-Real Mapping and Data Synchronization Module: Establishes and maintains a one-to-one correspondence between physical entities and virtual models in digital twins, and periodically synchronizes key status data through preset interfaces.
[0009] ② Model Library Management and Scheduling Module: Manages integrated simulation models of various types (hydrological, structural, electrical, etc.), and adopts a priority-based (… The dynamic resource scheduling strategy responds to the computational requests of the agent. Here, Urgency represents the task urgency, and Complexity represents the computational complexity. and These are the weighting coefficients.
[0010] ③ Simulation and deduction service engine: Receives structured disposal plans from the case disposal agent as "deduction scripts", clones the current state in the digital twin to create parallel deduction branches, schedules relevant models to execute virtual operations in sequence, monitors key indicators, and finally generates a "Solution Deduction Evaluation Report" containing quantitative results such as recovery time, resource consumption, and risk changes.
[0011] The Monitoring and Early Warning Intelligent Agent (MWA) is configured to: drive a disaster simulation chain consisting of a rainfall prediction module, a runoff calculation module, and a flood evolution module, coupled sequentially, in a digital twin environment scheduled by DTA, to generate flood inundation depth field, flow velocity field, and duration field for the target area in the future; and perform spatial overlay analysis of the above field data with the coordinates of power grid equipment to extract the disaster intensity characteristics of each equipment location, and output a preliminary risk list including equipment identification and its corresponding water depth, flow velocity, and expected soaking duration; The rainfall prediction module employs a Long Short-Term Memory (LSTM) network model, and its internal computation process follows the LSTM gating mechanism. Forgotten Gate: Input Gate: Candidate cell status: Cell status update: Output gate: Hidden output: In the above formula This represents the forget gate weight matrix. It is the hidden state of the previous time step. This represents the input at the current time step. Represents the forgetting gate bias. It is the sigmoid activation function; This represents the input gate weight matrix. This represents the input gate bias.
[0012] Represents the cell state weight matrix. This indicates a cell state bias. This indicates the cell state at the previous time step. This represents the output gate weight matrix. This is used to bias the output gate.
[0013] The runoff calculation module uses the Soil Conservation Service Curve Number (SCS-CN) method, and its calculation formula is as follows: In the formula, S represents the maximum potential retention capacity, CN is the curve value, and P is the cumulative rainfall. This represents the effective rainfall.
[0014] The flood evolution module employs a two-dimensional shallow water equation solved using the finite volume method, and its governing equations are: In the formula, h is the water depth and t is the time; These are the velocity components in the x and y directions, respectively; , Represents spatial coordinates; R represents source and sink terms. It is the acceleration due to gravity. The density of water; Indicates the elevation of the bed bottom; , It represents the components of the bed shear stress in the x and y directions.
[0015] The fault diagnosis agent (FDA) is configured to: receive a preliminary risk list, and for each device in the list, call the corresponding fault probability prediction model optimized by an optimized algorithm according to its type; the model is a BP neural network model with initial weights and thresholds optimized by the Grey Wolf optimization algorithm; its input is a feature vector that integrates disaster intensity features and the device's own state parameters, and its output is the fault probability value of the device; and it summarizes all prediction results to generate a list of critical equipment fault probabilities sorted by probability value.
[0016] The core position update formula of the gray wolf optimization algorithm is as follows: In the formula , , This represents the distance vector between the current individual gray wolf and the three leader wolves α, β, and δ. , , This represents a vector of random coefficients, used to enhance the randomness of the search. , , Let α, β, and δ be the position vectors of the three leader wolves; It is the current position vector of the individual gray wolf. , , This represents the vector of the middle position after adjustment based on the positions of the three leader wolves; , , This is the convergence coefficient vector, which controls the search step size and direction; This represents the updated position vector of the individual gray wolf.
[0017] The Case Handling Agent (CHA) is configured to: construct a feature vector for the current disaster scenario based on the current disaster scenario characteristics and the list of failure probabilities of the key equipment; retrieve similar cases from the historical case database using a weighted comprehensive similarity algorithm, whereby the knowledge graph provides semantic constraints and extensions for case retrieval; and integrate the handling measures of the retrieved cases to generate a preliminary handling plan. The CHA is further configured to: initiate a virtual simulation request to the Digital Twin Agent (DTA), submitting the preliminary handling plan as a simulation script to the DTA's simulation simulation service engine; receive the "Scheme Simulation Evaluation Report" returned by the DTA, analyzing the bottlenecks and risks in the simulation results; adjust and optimize the preliminary plan based on the analysis report, and through optional multiple rounds of "generation-simulation-optimization" iterations, finally generate a structured, virtually verified, tiered emergency response and recovery plan.
[0018] The formula for calculating the weighted comprehensive similarity is: In the formula This represents the overall similarity between the current scenario and the j-th historical case; It is the feature vector of the current scene. It is the feature vector of the j-th historical case; , , These are weighting coefficients, corresponding to the weights of feature similarity, geographical proximity, and seasonal matching, respectively. L is the geographical distance between the current scene and the location of historical cases; L is the geographical distance normalization parameter. This indicates the season in which the current scenario and historical cases occurred; This indicates that the value is determined by the degree of seasonal difference; if the seasons are the same, the value is 1, otherwise it is 0.
[0019] The measures taken in the retrieved cases are integrated and optimized based on the rules in the knowledge graph and real-time power grid constraints to generate a structured, tiered emergency response and recovery plan. The collaboration among the Digital Twin Agent (DTA), Monitoring and Early Warning Agent (MWA), Fault Diagnosis Agent (FDA), and Case Handling Agent (CHA) is specifically as follows: the DTA is responsible for overall coordination and resource allocation after the system is triggered; after the MWA completes its analysis, it publishes a message containing the preliminary risk list through the message middleware; the FDA subscribes to the message, is triggered to start and completes fault probability prediction, and then publishes a new message; the CHA subscribes to relevant message streams, starts the case retrieval and solution generation process in parallel, and can initiate virtual simulation requests to the DTA, iteratively optimizing the solution based on simulation feedback, thereby realizing asynchronous, decoupled pipeline collaboration with the DTA as the central hub, and supporting feedback optimization.
[0020] The system also includes an application interaction layer for receiving and integrating the outputs of various intelligent agents to generate a unified decision-making product. Specifically, based on a digital twin, it renders the inundation field into a flood risk base map using a geographic information system, and overlays the equipment location and probability information from the key equipment failure probability list onto the base map in a layered color scheme to form a panoramic view of power grid flood risk; simultaneously, it automatically compiles the key equipment failure probability list, the structured disposal plan, and related virtual simulation process summaries into a comprehensive decision-making report.
[0021] The system ultimately outputs a power grid flood risk map, a list of critical equipment failure probabilities, a customized response plan report, and a virtual simulation and evaluation appendix for the plan.
[0022] Secondly, the present invention provides a method for intelligent early warning and handling of power grid flood disasters based on the system described in the first aspect, characterized by comprising the following steps: S1: In response to disaster early warning, the digital twin agent (DTA) synchronizes multi-source data, updates and manages the digital twin; S2: The monitoring and early warning intelligent agent (MWA) drives the disaster simulation chain in the digital twin environment, generates a flood field and superimposes the coordinates of power grid equipment, and outputs a preliminary risk list; S3: The Fault Diagnosis Agent (FDA) receives the preliminary risk list, calls the BP neural network model optimized by the Grey Wolf Optimization Algorithm, predicts the failure probability of the equipment in the list, and outputs a list of failure probabilities of key equipment. S4: The case handling agent (CHA) retrieves historical cases based on the current scenario and the fault probability list to generate a preliminary handling plan, and requests the DTA to perform a virtual simulation; S5: The simulation and deduction service engine of DTA performs the scheme deduction in the digital twin and feeds back the evaluation report to CHA; S6: CHA iteratively optimizes the plan based on the simulation and evaluation report, and outputs the final disposal plan that has been virtually verified; S7: The application interaction layer comprehensively renders the risk map, fault list and handling plan to form a complete decision-making product.
[0023] The beneficial effects of this invention are as follows: It has achieved a closed-loop intelligent decision-making process that integrates perception, diagnosis, decision-making, and simulation: through the hierarchical collaboration of four intelligent agents with DTA as the central hub, the four links of disaster simulation, risk assessment, scheme generation, and effect simulation are linked into a feedback-enabled and optimizable organic whole, which greatly improves the systematicness, scientific nature, and efficiency of emergency response.
[0024] A "physical-virtual" parallel interactive decision-making paradigm was constructed: by introducing digital twins and management intelligent agents, a virtual simulation environment that interacts synchronously with the physical power grid was created, enabling emergency response plans to undergo multiple rounds of "hypothesis analysis" and iterative optimization before execution, which significantly reduces the risks and uncertainties of actual decision-making.
[0025] It deeply integrates physical mechanisms and data-driven approaches: the physical model chain of "GreyWolf Optimizer-Back Propagation Neural Network (GWO-BPNN) + two-dimensional shallow water equations" ensures the interpretability and accuracy of disaster simulation, while the intelligent prediction model of "GWO+BPNN" achieves accurate fitting of the nonlinear vulnerability relationship of equipment. The combination of the two ensures the scientific nature of prediction from macro disaster situation to micro risk.
[0026] Enhancing reliability through knowledge graphs throughout: Domain knowledge graphs provide parameters for the model, constrain the prediction range, and provide handling rules, effectively guiding each agent to operate within a professional track, significantly suppressing "machine illusions," and improving the engineering credibility of the output results.
[0027] The decision-making solutions combine historical wisdom, real-time adaptability, and forward-looking predictability: by drawing on historical experience through multi-dimensional weighted case similarity matching, and by using DTA virtual simulation to proactively evaluate and optimize the solutions, the generated solutions are not only reasonable and flexible, but also have higher feasibility and success rate. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the overall system architecture provided in Embodiment 1 of the present invention. Figure 2 This is a flowchart illustrating the internal logic and data processing of the Monitoring and Early Warning Intelligent Agent (MWA) in Embodiment 2 of the present invention. Figure 3 This is a flowchart of the GWO-BPNN model construction and optimization process used by the fault diagnosis agent (FDA) in Embodiment 3 of the present invention. Figure 4 This is a schematic diagram illustrating the principle of case retrieval and solution generation by the case handling agent (CHA) in Embodiment 4 of the present invention. Figure 5 This is a sequence diagram of the complete workflow of the system from disaster early warning to response decision-making in Embodiment 5 of the present invention. Figure 6 This is a schematic diagram of the workflow of the internal simulation and deduction service engine of the digital twin intelligent agent (DTA) in Embodiment 6 of the present invention. Detailed Implementation
[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.
[0030] Example 1 See Figure 1 This embodiment reveals the overall architecture of the system of the present invention. The system adopts a four-layer decoupled design of "physical perception, virtual mapping, intelligent collaboration, and interactive application", mainly including: physical entity layer 101, digital twin layer 102, intelligent agent collaboration layer 103, and application interaction layer 104.
[0031] The physical entity layer 101 corresponds to real-world power grid equipment, meteorological and hydrological stations, and geographical environments, acquiring data through sensors and monitoring systems. The digital twin layer 102, built under the management of the digital twin agent (DTA) 1021, is a high-fidelity virtual power grid environment, integrating a 3D virtual scene, a multi-dimensional model library, real-time data mirroring, and a domain knowledge graph 1022. The agent collaboration layer 103 is the core, containing the digital twin agent (DTA) 1030 as the management center, and three functional agents: the Monitoring and Early Warning Agent (MWA) 1031, the Fault Diagnosis Agent (FDA) 1032, and the Case Handling Agent (CHA) 1033. DTA 1030 is responsible for maintaining the twin layer 102 and providing simulation services for other agents. MWA 1031, FDA 1032, and CHA 1033 communicate asynchronously and collaborate under the coordination of DTA through message middleware 1034, forming a dynamic decision pipeline that supports feedback optimization. Each agent can actively query the knowledge graph 1022 when executing tasks. The application interaction layer 104 is responsible for integrating, visualizing, and encapsulating the various results generated by the agents, ultimately presenting them to the user in the form of risk maps, ranking lists, solution reports, and simulation process replays, supporting scheduling and emergency command decisions.
[0032] See Example 2 Figure 2 This embodiment details the internal workflow of the Monitoring and Early Warning Agent (MWA). When the system is triggered, the MWA begins execution. First, it acquires high spatiotemporal resolution gridded rainfall forecast data from the data layer. Then, it enters the core disaster simulation chain: First, it uses a pre-trained Long Short-Term Memory (LSTM) network model to downscale and correct biases in the original forecast, obtaining a more accurate localized gridded rainfall sequence. Second, it inputs the corrected rainfall data, along with soil type and moisture information, into the SCS-CN runoff generation model. This model calculates the net rainfall (direct runoff) generated by each grid cell based on regional characteristics. Third, using the aforementioned net rainfall as input, it drives a numerical model based on two-dimensional shallow water equations to simulate the flood evolution process on a high-precision digital elevation model. This model calculates the dynamic changes in physical quantities such as water depth and flow velocity at each spatial location within a future time period by solving a set of partial differential equations. Finally, the MWA reads the geographic location layer of the power grid equipment and performs spatial overlay analysis with the simulated flood field, extracting the maximum water depth, maximum flow velocity, and expected immersion duration for each equipment location, forming a detailed "preliminary risk list." The list is published via a message middleware as input for subsequent agents.
[0033] See Example 3 Figure 3This embodiment focuses on the construction and optimization process of the core prediction model in the Fault Diagnosis Agent (FDA)—GWO-BPNN. Initialization is performed first, determining the structure of the BP neural network and the parameters of the gray wolf optimization algorithm. A key step is using the gray wolf optimization algorithm to search for optimal initial weights and thresholds for the BP network. All network parameters are encoded as "gray wolf" position vectors, and a wolf pack is randomly initialized. Using the network's prediction error on the validation set as the fitness evaluation criterion, the algorithm iteratively guides the wolf pack's movement and search based on the position of the alpha wolf (α, β, δ). After reaching the maximum number of iterations, the algorithm outputs the gray wolf with the optimal position. Decoding this optimal position yields a set of initial network parameters superior to the randomly initialized ones. The BP neural network is initialized using these optimized parameters and formally trained on the training set, ultimately resulting in a well-trained GWO-BPNN fault prediction model with stronger generalization capabilities. In online applications, the FDA inputs the device feature vector into this model to obtain high-precision fault probability values.
[0034] Example 4 See Figure 4 This embodiment illustrates the principle of the Case Handling Agent (CHA) generating intelligent response plans. The process begins with feature extraction. CHA fuses the macroscopic features of the current disaster with critical equipment failure information output by the FDA to construct a feature vector V_current for the current scenario. The core step is similar case retrieval: CHA traverses the historical case database, calculating the similarity between the feature vector V_case_j of each historical case and V_current using a weighted comprehensive similarity formula. The K most similar historical cases are retrieved as a reference set. Next comes preliminary plan generation. CHA analyzes the measures taken and their effects in these K cases, while simultaneously querying relevant rules from the knowledge graph to form a preliminary response plan. The crucial optimization loop begins with a virtual simulation request. CHA sends the preliminary plan to the Digital Twin Agent (DTA) to request a virtual simulation. DTA executes the simulation and provides feedback. DTA's simulation service engine creates parallel branches within the digital twin, strictly executing the plan's timing sequence, simulating the entire process from power outage operations and resource allocation to power restoration, and monitoring key indicators. After the simulation, DTA provides CHA with a quantitative "Solution Simulation Evaluation Report." Solution analysis and optimization follow. CHA analyzes the simulation report, identifying bottlenecks in the solution, such as resource conflicts or excessively long recovery times. If the solution fails to meet expectations, adjustments and optimizations are made. Iteration and output: This "generation-simulation-optimization" process can be iterative. Until the solution meets all constraints and optimization objectives, CHA outputs the final structured emergency response and recovery plan, along with a simulation evaluation summary.
[0035] Example 5 See Figure 5This embodiment depicts the complete collaborative workflow and timing of the system from startup to decision-making. The entire process is triggered by a rainstorm warning event (time T0). The Digital Twin Agent (DTA), acting as the central controller, is first activated (process A). It synchronizes the latest data and updates the digital twin, then issues a simulation task to the Monitoring and Warning Agent (MWA). The MWA starts, executes its internal disaster simulation chain within the digital twin, and publishes a "Preliminary Risk List" message upon completion (time T1). The Fault Diagnosis Agent (FDA) subscribes to this message and immediately starts upon receiving it (process B), performing batch fault probability calculations and publishing a "Critical Equipment Fault Probability List" message upon completion (time T2). The Case Handling Agent (CHA) can work in parallel. It can begin preliminary retrieval after acquiring disaster characteristics early on, and quickly generates a preliminary plan upon receiving the detailed list from the FDA (process C1). Subsequently, the CHA initiates a "Virtual Simulation Request" to the DTA (time T3). After receiving the request, DTA invokes the simulation engine to execute the simulation (process C2), and upon completion, sends the "Simulation Evaluation Report" back to CHA (time T4). CHA optimizes the solution based on the report (process C3) and finally releases the "Structured Disposal Plan" message (time T5). The application layer listens for all key messages, and after the core products are complete, it integrates and renders the results (process D), finally displaying comprehensive decision information uniformly on the user interface (time T6). The entire process (T0 to T6) achieves asynchronous and efficient collaboration and feedback optimization among multiple agents with DTA as the central hub, enabling the completion of the entire intelligent decision-making process from risk perception to virtual verification in a relatively short time.
[0036] Example 6 See Figure 6This embodiment details the workflow of the simulation and deduction service engine in the Digital Twin Agent (DTA). When the DTA receives a deduction request from the Case Handling Agent (CHA), the engine starts. Step S601: Solution Analysis and Scenario Cloning. The engine analyzes the structured handling solution submitted by the CHA, extracting key operation instructions and timing. Subsequently, it clones a complete snapshot of the current state in the digital twin, creating an independent "parallel deduction branch" environment to ensure that the deduction does not affect the twin's main state. Step S602: Timing Scheduling and Model Execution. The engine schedules relevant models in the digital twin model library step by step according to the timeline of the structured handling solution from the Case Handling Agent. For example, at the "Power Outage Operation" moment, the power grid flow calculation model is called to simulate the network state after the switch is disconnected; at the "Repair Team Departure" moment, the traffic simulation model is called to estimate the arrival time; during the "Flood Evolution" period, the hydrodynamic model continues to drive. Step S603: Process Monitoring and Data Acquisition. During the deduction execution process, the engine monitors predefined key performance indicators in real time. Step S604: Conflict Detection and Logical Processing. The engine detects logical conflicts or rule violations during the simulation, such as duplicate resource allocation or incorrect operation sequence, and automatically processes or marks them according to preset rules. Step S605: Generating a Simulation Evaluation Report. After the simulation, the engine summarizes all monitoring data and compares and analyzes the simulation results with the contingency plan objectives. The report quantitatively outputs the estimated total recovery time, the time consumed at each stage, peak resource usage, remaining risk level, etc., and points out the bottlenecks and potential risks in the solution. Step S606: Result Feedback. The engine returns the "Solution Simulation Evaluation Report" to the requesting CHA, providing accurate data support for solution optimization. Through this engine, DTA transforms the digital twin from a static display model into a dynamic decision-making experimental field, which is the core technical support for the system to achieve "parallel decision-making".
[0037] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A smart early warning and response system for power grid flood disasters based on digital twins and multi-agent collaboration, characterized in that, include: The data access and preprocessing module is used to access and standardize multi-source heterogeneous data from weather forecasts, hydrological monitoring, geographic information and power grid equipment status in real time. A knowledge graph for the field of power grid flood disasters is stored in the form of a graph database. It stores power grid equipment entities, geographical unit entities, disaster event entities, fault mode entities, and response measure entities, and stores edge relationships that represent the causal relationships, spatial relationships, and logical constraints between the entities. The digital twin environment is a virtual simulation environment constructed based on the aforementioned multi-source heterogeneous data and synchronized with the physical power grid in terms of space and state. The digital twin intelligent agent communicates with the digital twin environment and the knowledge graph of the power grid flood disaster domain to maintain bidirectional data synchronization between the digital twin environment and the physical entity; manages and schedules integrated hydrological, structural mechanics and power grid electrical simulation models; receives external schemes and initiates independent parallel inference branches in the digital twin environment, executes virtual operations and outputs quantitative evaluation reports; The monitoring and early warning intelligent agent, communicating with the digital twin intelligent agent and message middleware, is used to: based on the digital twin environment, sequentially drive the coupled physical model chain to generate a gridded flood field for the target area; perform spatial overlay analysis of the flood field and the geographical coordinates of the power grid equipment to form and publish a preliminary equipment-level risk list. A fault diagnosis agent is used to subscribe to and receive the preliminary risk list; For each device in the list, an optimized neural network fault probability prediction model is invoked to output a dynamic fault probability value and generate a list of fault probabilities for critical equipment. The case handling intelligent agent is communicatively connected to the DTA and the knowledge graph, and is used to: retrieve historical cases in the knowledge graph and generate preliminary handling plans based on the list of failure probabilities of key equipment and the characteristics of the current disaster scenario; initiate a virtual simulation request to the DTA, and iteratively optimize the plan based on the simulation evaluation report fed back by the DTA, and output the final handling plan that has been virtually verified. The application interaction layer is used to integrate and visualize the preliminary risk list, the key equipment failure probability list, and the final disposal plan, and generate a comprehensive decision report.
2. The intelligent early warning and response system for power grid flood disasters based on digital twins and multi-agent collaboration as described in claim 1, characterized in that, The rainfall prediction module in the monitoring and early warning intelligent agent specifically adopts a long short-term memory network model, and its operation process follows the LSTM gating mechanism; the runoff calculation module adopts the local soil conservation curve numerical method, and its calculation formula is as follows: In the formula, S represents the maximum potential retention capacity, CN is the curve value, and P is the cumulative rainfall. The effective rainfall is calculated when P > 0.2S, as runoff occurs. When P < 0.2S, no runoff occurs, and the effective rainfall is 0.
3. The intelligent early warning and response system for power grid flood disasters based on digital twins and multi-agent collaboration as described in claim 1, characterized in that, In the BP neural network fault probability prediction model optimized by the gray wolf optimization algorithm used by the fault diagnosis agent, the core position update formula of the gray wolf optimization algorithm is as follows: in, , , This represents the distance vector between the current individual gray wolf and the three leader wolves α, β, and δ. , , This represents a vector of random coefficients, used to enhance the randomness of the search. , , Let α, β, and δ be the position vectors of the three leader wolves; It is the current position vector of the individual gray wolf. , , This represents the vector of the middle position after adjustment based on the positions of the three leader wolves; , , This is the convergence coefficient vector, which controls the search step size and direction; This represents the updated position vector of the individual gray wolf.
4. The intelligent early warning and response system for power grid flood disasters based on digital twins and multi-agent collaboration as described in claim 1, characterized in that, The weighted comprehensive similarity algorithm used by the case handling agent is calculated using the following formula: In the formula, This represents the overall similarity between the current scenario and the j-th historical case; It is the feature vector of the current scene. It is the feature vector of the j-th historical case; , , These are weighting coefficients, corresponding to the weights of feature similarity, geographical proximity, and seasonal matching, respectively. L is the geographical distance between the current scene and the location of historical cases; L is the geographical distance normalization parameter. This indicates the season in which the current scenario and historical cases occurred; This indicates that the value is determined by the degree of seasonal difference; if the seasons are the same, the value is 1, otherwise it is 0.
5. The intelligent early warning and response system for power grid flood disasters based on digital twins and multi-agent collaboration as described in claim 1, characterized in that, The model library management and scheduling module in the digital twin intelligent agent adopts a dynamic resource scheduling strategy, and its task priority calculation formula is as follows: In the formula, To assess the urgency of the task, To calculate the computational complexity, and These are the weighting coefficients.
6. The intelligent early warning and response system for power grid flood disasters based on digital twins and multi-agent collaboration as described in claim 1, characterized in that, The message middleware enables asynchronous and decoupled communication between the monitoring and early warning intelligent agent, the fault diagnosis intelligent agent, the case handling intelligent agent, and the application interaction layer. The digital twin intelligent agent acts as the overall coordinator, responding to system trigger events and sequentially scheduling the startup and execution of the monitoring and early warning intelligent agent, the fault diagnosis intelligent agent, and the case handling intelligent agent to form a closed-loop workflow of "early warning-diagnosis-decision-inference".
7. A method for intelligent early warning and handling of power grid flood disasters based on the system described in any one of claims 1 to 6, characterized in that, Includes the following steps: S1: In response to a disaster early warning signal, the data access and preprocessing module is activated to access and process multi-source heterogeneous data in real time; the digital twin intelligent agent synchronizes the multi-source heterogeneous data, dynamically updates and maintains the digital twin environment to keep it consistent with the physical power grid status; S2: The digital twin agent schedules the activation of the monitoring and early warning agent; the monitoring and early warning agent, within the digital twin environment, sequentially drives the long short-term memory network rainfall prediction model, the SCS-CN runoff calculation model, and the two-dimensional shallow water equation flood evolution model to generate a gridded flood inundation field for the target area in the future; through spatial overlay analysis, the geographical coordinates of the power grid equipment are matched with the flood inundation field, the disaster intensity characteristics of each equipment location are extracted, and a preliminary risk list at the equipment level is formed and published; S3: The fault diagnosis agent subscribes to and receives the preliminary risk list; For each device in the list, a BP neural network fault probability prediction model with initial weights and thresholds optimized by the Grey Wolf optimization algorithm is invoked to make a prediction. The input feature vector of the model integrates disaster intensity features from the preliminary risk list and device status parameters from a knowledge graph or real-time monitoring. Generate and publish a list of critical equipment failure probabilities sorted by failure probability values; S4: The case handling agent subscribes to and receives the list of failure probabilities of the key equipment, and constructs a feature vector of the current scene in combination with the macroscopic characteristics of the current disaster scene; using a weighted comprehensive similarity algorithm, under the semantic constraints of the knowledge graph of the power grid flood disaster domain, it retrieves similar historical cases in the historical case library; and integrates the handling measures of the retrieved cases to generate a preliminary structured handling plan. S5: The case handling agent initiates a virtual simulation request to the digital twin agent and submits the preliminary structured handling plan; S6: The simulation and deduction service engine of the digital twin intelligent agent creates an independent parallel deduction branch in the digital twin environment, clones the current state snapshot; parses and executes the various operations in the preliminary structured disposal plan in sequence, schedules the corresponding hydrological, traffic, and power grid flow simulation models for collaborative simulation, and monitors the predefined key performance indicators in real time; after the deduction is completed, it generates and feeds back the "Scheme Deduction Evaluation Report" containing quantitative evaluation results; S7: The case handling agent receives the "Solution Demonstration and Evaluation Report", analyzes the bottlenecks and risks in the evaluation results, and iteratively optimizes the preliminary structured handling solution; If the solution fails to achieve the preset optimization goal, return to step S5 and initiate a new round of virtual simulation based on the optimized solution; If the solution meets all the preset objectives, the final solution that has been virtually verified will be output. S8: The application interaction layer listens to and integrates the preliminary risk list, the key equipment failure probability list, and the final virtual-verified handling plan. Based on the digital twin environment, it performs integrated visualization rendering to generate a panoramic view of power grid flood risk and automatically compiles and outputs a comprehensive decision report containing the key equipment failure probability list, structured handling plan, and virtual simulation evaluation summary.
8. The intelligent early warning and response method for power grid flood disasters according to claim 7, characterized in that, In step S2, generating a gridded flood inundation field specifically includes: using a long short-term memory network model to perform spatiotemporal downscaling and bias correction on the input gridded rainfall forecast data to obtain a refined gridded rainfall sequence; inputting the corrected rainfall data, soil type, and previous humidity conditions into the SCS-CN model to calculate the net rainfall for each grid cell; using the net rainfall as the source term, driving a numerical model based on two-dimensional shallow water equations on a high-precision digital elevation model to solve for the dynamic changes in water depth and flow velocity of each grid cell in the future time period.
9. The intelligent early warning and handling method for power grid flood disasters according to claim 7, characterized in that, In step S3, the training process of the fault probability prediction model includes: using the Grey Wolf Optimization Algorithm to search for the optimal initial weight and threshold combination for the BP neural network, wherein the prediction error of the network on the validation set is used as the fitness evaluation criterion; after decoding the optimal parameters obtained by the Grey Wolf Optimization Algorithm, the BP neural network is initialized, and then supervised training is performed using labeled historical disaster sample data to finally obtain the trained GWO-BPNN model.
10. The intelligent early warning and handling method for power grid flood disasters according to claim 7, characterized in that, In step S4, the process of retrieving similar historical cases is guided by the semantics of the knowledge graph of the power grid flood disaster domain. The weighted comprehensive similarity calculation integrates the cosine similarity of feature vectors, the geographical proximity between disaster sites, and the matching degree of the disaster season. The knowledge graph is used to semantically associate and expand disaster types, failure modes, and response measures during retrieval to improve the accuracy and coverage of the retrieval.