Power grid fault handling method and system based on large model and multiple agents

Abstract

The invention discloses a power grid fault handling method and system based on a large model and multiple agents. Acquiring and collecting power grid multi-source data to form power grid fault information; the method comprises the following steps: performing semantic analysis on power grid fault information, generating a fault disposal strategy through a fine-tuning trained power field large model, disassembling the fault disposal strategy into an executable subtask sequence, and determining an input / output and execution sequence of each subtask; constructing a plurality of agents, and cooperatively controlling and executing a plurality of subtasks of a fault handling strategy; and after the execution of the fault processing strategy is completed, setting a multi-dimensional evaluation index, and carrying out quantitative evaluation on the fault processing effect. According to the large model and multi-agent cooperative power grid fault disposal technical scheme provided by the invention, the intelligent level and the emergency response efficiency of power grid fault disposal are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system automation and artificial intelligence, and relates to a power grid fault handling method and system based on a large model and multiple intelligent agents. Background Art

[0002] In modern power systems, with the continuous expansion of power grid scale and the increasing complexity of its structure, the types and impacts of power grid failures have become more diverse and complex.

[0003] With the development of artificial intelligence technology, intelligent fault handling methods have been introduced. For example, patent application publication number CN113193549A employs network analysis mechanism computing technology to propose a control transaction perception and coordinated fault handling system and method. However, it does not employ artificial intelligence methods, resulting in certain deficiencies in flexibility and in-depth fault mining. Patent application publication number CN118282032A implements a power grid fault handling method based on safety analysis. This method only utilizes safety analysis mechanism computing methods during the handling process, without incorporating artificial intelligence methods to conduct in-depth analysis of power grid fault information. Furthermore, the involved mechanism computing modules, such as static safety calculation and load transfer path calculation, operate independently, lacking a well-organized coordinated strategy. Patent application publication number CN115940124A utilizes a deep convolutional neural network to assist in recommending fault measures and electronic emergency plans based on power grid fault conditions. This method focuses on matching and recommending textual handling measures in fault situations, and does not include generating real-time handling strategies based on the current real-time operation of the power grid.

[0004] To address the above problems, there is an urgent need for a power grid fault handling method and system based on large models and multi-agents. Summary of the Invention

[0005] In order to solve the deficiencies in the prior art, the present invention provides a power grid fault handling method and system based on a large model and multiple intelligent agents.

[0006] The present invention adopts the following technical solutions.

[0007] In a first aspect, the present invention provides a method for handling power grid faults based on a large model and multiple agents, which adopts the following technical solutions: Acquire and aggregate multi-source data of the power grid to form power grid fault information; the multi-source data includes fault time text information, real-time operation data, power grid topology information and environmental variables; Perform semantic analysis on the power grid fault information, generate a fault handling strategy by fine-tuning the trained power domain model, and decompose it into a sequence of executable subtasks, clarifying the input, output, and execution order of each subtask; Construct multiple intelligent agents to collaboratively control and execute multiple subtasks of the fault handling strategy; the intelligent agents include a risk analysis agent, a unit adjustment agent, a load transfer agent, a coordination agent, a safety verification agent, and a control execution agent; After the fault handling strategy is executed, multi-dimensional evaluation indicators are set to perform a quantitative evaluation of the fault handling strategy.

[0008] Furthermore, semantic parsing of the power grid fault information includes: Construct a power grid knowledge graph, where the entity set includes power grid equipment, sites, and topological nodes, and the relationship set includes equipment affiliation, topological connection, and protection association. Transform power grid fault information into structured fault features, including fault device identification, fault type, and fault device status; determine fault causes and associated devices through entity relationship analysis based on the power grid knowledge graph; Utilizing a large model in the electric power sector, a fault handling strategy is generated based on the structured fault characteristics and the matching results of historical fault cases.

[0009] Furthermore, the coordination mechanism of the multiple agents includes: The risk analysis agent is used to generate a risk analysis report based on the real-time operation status of the power grid and send it to the coordination agent; the unit adjustment agent is used to respond to the risk analysis agent, generate a unit adjustment plan based on the multiple optimization objectives of the power grid, and send it to the coordination agent; the load transfer agent is parallel to the unit adjustment agent, and is used to search for load transfer paths based on the power grid topology, generate a transfer plan, and send it to the coordination agent; the coordination agent is used to receive the risk analysis report, unit adjustment plan, and load transfer plan, summarize them, determine the task execution order, and generate a fault handling plan; The safety verification agent is used to perform risk verification on the fault handling plan. If the handling plan passes the risk verification, it is sent to the control execution agent; if it fails, the verification result is fed back to the coordination agent; the control execution agent is used to issue control instructions to the power grid equipment.

[0010] Furthermore, the execution steps of the risk analysis agent include: Collect grid operating parameters, equipment rated parameters, and safe operating thresholds of operating parameters in real time, calculate the mean of the grid operating parameters within a sliding window, and preliminarily determine that an over-limit situation exists if the difference between the grid operating parameter and the corresponding mean exceeds the threshold; Based on the preliminary judgment, the time series prediction method is used to obtain the predicted value of the operating parameter. If the predicted value exceeds the threshold, it is determined that there is an over-limit risk. When the over-limit risk is confirmed, the current load values ​​of all nodes affected by the over-limit risk and requiring load transfer are weighted and summed to obtain the load to be transferred. A risk analysis report is generated, including the over-limit equipment and its corresponding load to be transferred.

[0011] Furthermore, the execution steps of the unit adjustment agent include: An optimization model is constructed with the goal of minimizing power balance and economic cost of the power grid. The objective function includes economic cost term, load balance term, and frequency stability term. The economic cost item is the total power generation cost of all generator sets, the load balancing item is the supply and demand deviation penalty of all load nodes; the frequency stability item is the frequency deviation penalty of all frequency monitoring points; The constraints of the objective function include power balance constraints and upper and lower limit constraints of the generator set output; solving the optimization model obtains the unit adjustment plan.

[0012] Furthermore, the execution steps of the load transfer agent include: For the power grid topology, the shortest path is searched as the load transfer path based on the edge weight, where the influencing factors of the edge weight include line impedance and transmission capacity; Calculate the transfer scheme score for the searched load transfer path, where the transfer scheme score is the inverse of the transfer cost; the numerator is 1, and the denominator is the weighted sum of voltage drop, line loss, and transfer capacity; The load transfer path corresponding to the maximum transfer plan score is taken as the optimal transfer plan.

[0013] Furthermore, the multi-dimensional evaluation indicators include fault handling efficiency, power grid restoration efficiency, resource utilization efficiency, and economic loss reduction; wherein, the fault handling time is calculated by the time difference between the time when the fault occurs and the time when the fault is confirmed and eliminated; the ratio of the fault handling time to the historical handling time of the same type of fault is used as the fault handling efficiency; The grid restoration efficiency is calculated by taking the ratio of the total load restored to the total load before the fault as the numerator and the weighted average of the power restoration time in each area of ​​the grid as the denominator. The total resource input is calculated by counting the number of devices used and the duration of use during the fault handling process. The amount of restored power is used as the actual resource output. The ratio of the actual resource output to the total resource input is used as the resource utilization efficiency. The difference between the estimated economic loss without taking any action and the actual economic loss under the fault handling strategy is used to represent the reduction in economic loss.

[0014] In a second aspect of the present application, a method for handling power grid faults based on a large model and multiple agents is provided, which adopts the technical solution described in the first aspect of the present application, and the system includes: Fault perception and information aggregation module; used to obtain and aggregate multi-source data of the power grid to form structured power grid fault information; the multi-source data includes fault time text information, real-time operation data, power grid topology information and environmental variables; Large model reasoning and task decomposition module; used to perform semantic analysis on the power grid fault information, generate fault handling strategies through the power field large model, and decompose them into executable subtask sequences, clarifying the input, output and execution order of each subtask; An agent collaborative execution module; used to construct multiple agents to collaboratively control and execute multiple subtasks of the fault handling strategy; the agents include a risk analysis agent, a unit adjustment agent, a load transfer agent, a coordination agent, a safety verification agent, and a control execution agent; The treatment effect evaluation and feedback optimization module is used to set multi-dimensional evaluation indicators after the fault treatment strategy is executed, and to conduct a quantitative evaluation of the fault treatment strategy.

[0015] Based on the above technical solution, this application includes at least one of the following beneficial effects: 1. This invention can rapidly complete key operations such as unit adjustment and load transfer, significantly shortening fault resolution time and minimizing the impact of faults on the power grid. Furthermore, relying on a mechanism for evaluating and optimizing fault resolution results and feedback, it enables the system to dynamically evolve. This holistic approach transcends the limitations of traditional single-point optimization methods, improving grid fault resolution efficiency throughout the entire process from analysis and decision-making to task execution and iterative optimization, providing a systematic solution for the safe and stable operation of the power grid.

[0016] 2. Leveraging the powerful logical reasoning capabilities of large models, in-depth analysis and processing of power grid fault information can be achieved, accurately locating the root cause of the fault and the scope of impact. Simultaneously, the collaborative execution advantages of multiple intelligent agents can be leveraged to efficiently complete fault handling tasks, thereby significantly improving the safety and stability of power grid operations, shortening fault handling time, and reducing losses caused by faults.

[0017] 3. By training and optimizing the large language model enhanced by the knowledge graph, it is equipped with the ability to deeply understand the professional terms in the power field, reason about cross-domain knowledge associations, and logically deduce complex fault scenarios; on this basis, by constructing a collaborative architecture of multi-agent systems such as risk analysis agents, unit adjustment agents, and load transfer agents, combined with power grid topology data and real-time monitoring information, it can achieve rapid fault location, accurate analysis and dynamic handling strategy generation, effectively improving the intelligence level of power grid fault handling and emergency response efficiency, and improving the power grid fault handling level. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 1. A flow chart of a method for handling power grid faults; Figure 2 Schematic diagram of the collaborative process of intelligent agents; Figure 3 This is the system structure diagram of the power grid fault handling system. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the present invention clearer and more accurate, the technical solutions of the present invention are described in detail below through multiple specific embodiments. The embodiments used in the present invention are only used to explain the present invention and are not intended to limit the content of the present invention.

[0020] As an embodiment of the present application, refer to Figure 1 ,This embodiment provides a specific implementation of a power grid fault handling method based on a large model and multiple intelligent graphs.

[0021] S1: Fault perception and information aggregation: Acquire multi-source data and aggregate them to form structured fault-related information.

[0022] The integrated intelligent alarm system obtains fault event text information. An example format is "xx:xx:xx, xxkV equipment tripped due to protection, causing a power outage at xx station. The pre-accident load was xxMW." This fault event text provides a description of the time of occurrence, equipment, voltage level, protection action, and the scope of the power outage.

[0023] Obtain real-time operating data of equipment and systems from the SCADA system, including electrical quantities (voltage, current, system frequency), power quantities (active power, reactive power), status quantities (circuit breaker opening and closing status, line flow), and rated parameters (rated voltage, rated current).

[0024] Obtain grid structure and equipment configuration information, including node connection relationships, line parameters (line length, impedance), substation main wiring method, and transformer parameters (transformer ratio, operating gear).

[0025] Environmental variables related to power grid operation are obtained from the meteorological information system, including wind speed, wind direction, rainfall, temperature, and humidity.

[0026] As a step in this embodiment, data from different information sources is formatted and synchronized based on a unified data interface standardization protocol. Fault information, real-time operation data, topology data, and meteorological parameters are aggregated into a fault sensing and information aggregation module for centralized management.

[0027] S2: Large model reasoning and task decomposition; perform semantic understanding and logical analysis of power grid fault information, and decompose the fault handling process into subtasks.

[0028] 2.1: Data input and semantic understanding; The fault information collected by S1 (fault information, real-time operation data, topology data, meteorological parameters) is input into the large model reasoning and task decomposition steps.

[0029] 2.2: The large model reasoning and task decomposition steps include the large model training optimization module, the knowledge graph embedding module and the reasoning and task decomposition module.

[0030] 2.2.1: Large model training and optimization: A large language model adapted to the power sector is trained using a combination of supervised learning and reinforcement learning.

[0031] In the supervised learning stage, the cross-entropy loss function is used to optimize the model parameters so that the large language model can accurately classify and identify power grid fault types.

[0032] During the reinforcement learning phase, a policy gradient approach is employed, with a reward function designed to evaluate decision-making outcomes and update policy parameters, enabling the large language model to gradually learn the optimal fault handling strategy. During the reinforcement learning process, through continuous interaction with the power grid operating environment, the model's decision strategy is dynamically adjusted based on the actual fault handling results. This allows the large model to gradually learn the optimal decision-making plan for power grid fault handling tasks, thereby improving its ability to respond to complex fault scenarios.

[0033] 2.2.2: Knowledge graph embedding and reasoning; Construct a power grid knowledge graph, where entities are equipment, sites, and topological nodes, and relationships are connection relationships and operating status; use the TransE model to map power grid entities and relationships into a low-dimensional vector space to implement knowledge representation learning. As an example of this embodiment, a provincial power grid company constructs a knowledge graph containing the following elements: Entity types: 500kV substation (entity ID: S1), 220kV main transformer (ID: T1), relay protection device (ID: R1); relationship types: affiliation (transformer → substation), connection (device → line), protection object (device → device).

[0034] Construct a set of positive sample triples , generate a negative sample set by randomly replacing the head / tail entity , for example, generates an error triple In this application, positive samples represent real grid topology relationships (such as device affiliation), while negative samples refer to illegal relationships that are constructed (such as random replacement of devices or substations).

[0035] For triples , and are head / tail entities, The TransE model is based on the assumption that , embedding the vector through the model standard loss function.

[0036] 2.2.3: Generate fault handling process; As a technical step in this embodiment, natural language processing technology is first used to semantically parse unstructured fault information and convert it into structured fault feature data. Entity relationship analysis based on a knowledge graph is then used to determine the fault's root cause, impact range, and associated power grid equipment. For example, if the input is "500kV substation A's #1 main transformer heavy gas protection tripped," the output is {Device: Main Transformer A1, Fault Type: Heavy Gas Tripped, Status: Tripped}."

[0037] Through graph embedding and semantic reasoning, combined with a historical fault case library and handling experience, we can generate a fault handling process that covers fault event grouping and assessment, risk analysis, and contingency plan matching or strategy generation. The historical fault case library includes fault description, initial operation mode, post-fault operation mode, phased fault handling strategy, fault recovery, and information notification. As a technical step of this embodiment, based on the historical fault case library, the large language model trained in step 2.2.1 is used to generate the optimal handling strategy using the fault feature data as the state.

[0038] The processing flow generated by the large model is analyzed and the Hierarchical Task Network (HTN) planning algorithm is used to decompose the processing flow generated by the large model into subtasks. Specifically: Decomposing functions by tasks , according to the original task and system status , generating a set of executable subtasks. The input and output requirements of each subtask, as well as the execution order between subtasks, are clearly defined through the constraints of the power system's internal logic.

[0039] in, is the original task, that is, the optimal disposal strategy obtained; The current system status, that is, the current power system status, such as the current grid operating parameters, equipment opening and closing status, etc.

[0040] Decompose the task into multiple executable subtasks , and through the constraints Make it clear The input and output requirements of each subtask, as well as the execution order.

[0041] As an example of this embodiment, the original task For the task "Restore power to main transformer #1 at 500 kV substation A," the system state is "C = {Current load: 400 MW, Standby main transformer A2 capacity: 500 MW, Adjacent substation available load transfer: 200 MW}." Table 1 shows an example of the subtask decomposition results generated by the task decomposition module based on the original task and system state.

[0042] Table 1 Task decomposition example

[0043] For example, for risk analysis tasks, the input is real-time monitoring data and equipment parameters, and the output is a list of devices that exceed the limit and the amount of load to be transferred; The task of formulating the unit adjustment plan takes the results of the equipment exceeding the limit as input and outputs the power adjustment strategy; The load transfer task outputs the optimal transfer path and alternative solutions based on the load to be transferred and network topology data; The safety verification task receives the disposal plan and real-time status of the power grid, and outputs safety assessment results and risk warnings; Coordinate tasks to integrate the solutions of various agents and output optimized comprehensive disposal strategies and execution priorities; The control execution task takes the final disposal plan as input and outputs the execution results and actual effect data.

[0044] This application uses large-scale model reasoning and task decomposition modules to perform semantic understanding and logical analysis of power grid fault information, enabling multi-dimensional fault feature extraction and accurate fault type identification. Based on the reasoning results, the complex fault handling process is decomposed into multiple interrelated subtasks, such as fault location refinement, isolation plan formulation, and power restoration priority sorting. At the same time, combining the power grid topology with real-time operating data, it dynamically optimizes the task execution sequence and resource allocation strategy, providing a clear execution framework for subsequent multi-agent collaborative operations.

[0045] S3: Build multiple agents through the agent collaborative execution module, and use multiple agents to control and execute multiple subtasks of the fault handling process.

[0046] Reference Figure 2 , Figure 2 This is a schematic diagram of the agent collaboration process. The agents constructed in this application include a risk analysis agent, a unit adjustment agent, a load transfer agent, a coordination agent, a safety verification agent, and a control execution agent. The following describes the processing flow and mechanism of each agent.

[0047] 3.1: Risk Analysis Agent; Data collection: real-time collection of grid operating parameters (line current, voltage, power, etc.), equipment rated parameters, and safe operating thresholds of each parameter; Over-limit detection: Calculate the mean and standard deviation of real-time grid operating parameters within a sliding window to determine whether there is an over-limit situation; When the difference between the real-time operating parameter and the corresponding mean value exceeds the threshold coefficient, it is preliminarily determined that there is an over-limit situation; Trend prediction: Based on the preliminary judgment, the time series prediction model is used to predict the parameter change trend; if the predicted value also exceeds the threshold, it is determined that there is a risk of exceeding the limit.

[0048] Power flow calculation: When the risk of exceeding the limit is confirmed, the load to be transferred is calculated based on the grid topology, load distribution at each node, and current operating status. The calculation formula is as follows: ; in, is the load to be transferred, is the set of nodes that are affected by the limit and need to transfer load. Grid nodes The current load value; For nodes The load transfer coefficient is obtained from the sensitivity calculation module of the control system, taking into account factors such as the importance of the node and the line transmission capacity limitation.

[0049] Output results; generate a risk analysis report (including the out-of-limit equipment and its corresponding load to be transferred) and send it to the coordination agent.

[0050] 3.2: Crew adjustment agent; An optimization model is constructed with the goal of minimizing power grid power and economic costs. Objective function for: ; in, is the number of generator sets, For the The power generation cost function of a generator set is: For the Active power of each generator set; is the number of load nodes, and Node Maximum load and actual load; is the number of frequency monitoring points, For monitoring points frequency, is the rated power; and is the weight coefficient.

[0051] Constraints include power balance constraints ; Upper and lower limit constraints on generator set output ;in, and Respectively The lower and upper output limits of each generator set.

[0052] The optimization model is solved by using interior point method or genetic algorithm to obtain the unit adjustment plan and send it to the coordination agent.

[0053] 3.3: Load transfer agent; The load transfer path is searched based on the shortest path algorithm in graph theory. ;in, is a node set, is the edge set. Representation node and The edge between The weight of is determined by factors such as line impedance and transmission capacity limitation.

[0054] The shortest path search algorithm is used to search for load transfer paths based on edge weights. After finding the shortest path, the load transfer evaluation model is used to calculate the solution score. The load transfer evaluation model is expressed as: ; in, Score the solution. is the voltage drop, is the line loss, To limit the transfer capacity; 、 and are the weights of voltage drop, line loss and transfer capacity limitation respectively.

[0055] The optimal transfer plan is screened out through the plan score (maximum value), and detailed load transfer operation instructions (such as "close the switch at xx station and open the knife switch at xx station") are sent to the coordination agent.

[0056] The weight coefficients of voltage drop, line loss and transfer capacity limitation can be adjusted according to the actual operation requirements of the power grid. Generally speaking, if the power grid has high requirements for voltage stability, Set to 0.4-0.6; if you pay more attention to reducing line loss, The value can be 0.3-0.5; in areas with large load fluctuations, Can be set to 0.2-0.4. .

[0057] 3.4: Coordinating Agents; Information aggregation; accepting risk analysis reports, unit adjustment plans, and load transfer plans; Priority scheduling; determine the order of task execution in the emergency control phase and the mode adjustment phase; if it is the emergency control phase, apply the unit adjustment plan; if it is the mode adjustment phase, apply the load transfer plan.

[0058] Form a prototype of a complete handling plan, form a prototype of a complete fault handling plan, and send the finalized handling plan to the safety verification intelligent agent.

[0059] 3.5: Security verification agent; As a step of this embodiment, the safety verification intelligent agent accepts the treatment plan output by 3.4 and adjusts the power grid state in the simulation environment (such as adjusting the unit output, switching switches, and transferring loads according to the treatment plan); based on the adjusted power grid state, it establishes the power flow equation and performs Newton-Raphson power flow calculation.

[0060] For each node in the power grid, the active power equation and reactive power equation are established. , its active power and reactive power They are: ; ; in, and For nodes and The voltage amplitude, and is the node admittance matrix element, is the node voltage phase angle difference, is the number of nodes.

[0061] The Newton-Raphson power flow calculation method is used to iteratively solve the balance equations of node active power and reactive power to obtain the power flow distribution results.

[0062] Determine whether there are risks in the power grid under the disposal plan (line overload, voltage limit, section flow limit, etc.). If there are risks in the plan, analyze the risk type, severity and impact range in detail, generate a safety verification report and feed it back to the coordination agent; if the plan passes the verification, feed the results back to the control execution agent for execution.

[0063] S4: After the fault handling is completed, the handling effect evaluation and feedback optimization module conducts a quantitative evaluation of the fault handling effect based on the pre-set evaluation index system.

[0064] The evaluation index system includes four indicators: fault handling efficiency, power grid restoration efficiency, resource utilization efficiency, and economic loss reduction. The calculation method is as follows: Fault handling efficiency: The fault handling time is calculated by the time difference between the fault occurrence and the time when the fault is confirmed and eliminated. The fault handling efficiency is the ratio of the fault handling time to the historical handling time of similar faults. Grid restoration efficiency is calculated based on the power restoration time, restored power load, and total load before the fault in each area, and is expressed as: ; Where, Restoring efficiency to the electric grid; Indicates area The restored power supply load, Indicates the total load before the fault; The weighted average of the power restoration time in each area, Indicates area The power restoration time, is the regional weight.

[0065] Resource utilization efficiency; calculated based on the resources invested and the amount of restored power during the fault handling process, expressed as: ; Where, is resource utilization; Indicates the amount of power restored. Representation device duration of use; It represents the total amount of resource input.

[0066] Reduced economic losses: Estimate the potential economic losses that would have occurred if effective fault resolution measures had not been taken, such as power outage losses for users and equipment damage, and then calculate the actual economic losses. The difference between the two is the reduced economic losses.

[0067] As an optional implementation step of the embodiment, the four types of indicators are normalized, a comprehensive performance evaluation function is constructed, and a quantitative report of the fault handling (including the four types of indicators + comprehensive score) is output. The evaluation results are fed back to the large model reasoning and task decomposition steps and the intelligent agent collaborative execution module for model retraining, strategy optimization and task allocation adjustment.

[0068] like Figure 3 As shown, Figure 3This is a system architecture diagram of a power grid fault handling system. As one embodiment of the present application, a large-scale model and multi-agent-based power grid fault handling system is provided. This system is implemented using the method described in the first aspect of the present invention and includes a fault perception and information aggregation module, a large-scale model reasoning and task decomposition module, an agent collaborative execution module, and a handling effect evaluation and feedback optimization module.

[0069] The fault perception and information aggregation module retrieves real-time equipment operation data, network topology information and meteorological environment parameters through the power grid dispatching automation system and the meteorological information system; uses standardized data interface protocols to achieve synchronous transmission and acquisition of data from each system; presets numerical threshold rules in the built-in data verification module, and automatically identifies and filters abnormal values ​​that exceed the normal fluctuation range through the data verification module, and classifies and archives them in the data warehouse.

[0070] The large-model reasoning and task decomposition module performs semantic understanding and logical analysis of power grid fault information, and based on the reasoning results, decomposes the complex fault handling process into multiple interrelated sub-tasks; the intelligent agent collaborative execution module combines the power grid topology structure and real-time operation data to dynamically optimize the task execution sequence and resource allocation strategy; after the fault handling is completed, the handling effect evaluation and feedback optimization module quantitatively evaluates the fault handling effect based on a pre-set evaluation indicator system.

[0071] The power grid fault handling system based on large models and multiple intelligent agents includes a perception layer, a data layer, an intelligent decision-making layer, an intelligent execution layer, an application layer, and a feedback optimization layer.

[0072] The perception layer consists of existing grid monitoring equipment and edge computing nodes. The monitoring equipment collects real-time information such as the operating parameters of faulty equipment, grid topology, and meteorological and environmental data. The edge computing nodes perform preliminary cleaning and preprocessing on the collected data before uploading key fault information to the data layer.

[0073] The data layer includes a data warehouse and a data governance module. The data warehouse uses distributed storage technology to store and manage massive amounts of fault data. The data governance module standardizes the data, constructs a knowledge graph for power grid faults, and integrates data such as equipment information, historical fault cases, and response plans, providing data support for large-scale model reasoning and intelligent agent decision-making.

[0074] The intelligent decision-making layer comprises a large-model inference module and a task decomposition module. The large-model inference module integrates large-model training optimization and knowledge graph embedding algorithms. It uses a large-scale language model adapted for the power sector. Through training, it develops a deep understanding of grid fault information and logical reasoning capabilities. It can infer the root cause and impact of faults and generate fault handling processes and strategies. The task decomposition module analyzes the handling processes generated by the large model, breaks them down into specific subtasks, and allocates them based on the tasks and agent capabilities.

[0075] The intelligent execution layer includes a risk analysis agent, a unit adjustment agent, a load transfer agent, a coordination agent, a safety verification agent, and a control execution agent. Each agent performs fault handling tasks according to assigned tasks, interacting with each other and collaborating through communication protocols, implementing corresponding theoretical formulas and methods.

[0076] The application layer provides a visual operation interface for power grid dispatchers and operation and maintenance personnel, displaying real-time fault information, handling process progress, intelligent agent execution status and other data, and supports manual intervention functions, allowing operation and maintenance personnel to adjust the handling strategy and issue instructions to the intelligent agent for execution when the system automatically handles anomalies or requires human decision-making.

[0077] The feedback optimization layer collects data and results from the fault handling process, evaluates and analyzes the large model reasoning accuracy, intelligent agent execution efficiency, collaborative effect, etc. based on the evaluation index system, optimizes the large model training data and algorithm based on the evaluation results, and adjusts the intelligent agent function and collaborative mechanism.

[0078] As an embodiment of the present application, it relates to a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method described in the first aspect of the present invention.

[0079] As an embodiment of the present application, it relates to a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of the method described in the first aspect of the present invention are implemented.

[0080] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solutions of the present invention still include modifications or equivalent replacements for the specific embodiments of the present invention.

Claims

1. A power grid fault handling method based on a large model and multi-agent, characterized in that: include: Acquire and aggregate multi-source data of the power grid to form power grid fault information; the multi-source data includes fault time text information, real-time operation data, power grid topology information and environmental variables; Perform semantic analysis on the power grid fault information, generate a fault handling strategy by fine-tuning the trained power domain model, and decompose it into a sequence of executable subtasks, clarifying the input, output, and execution order of each subtask; Construct multiple intelligent agents to collaboratively control and execute multiple subtasks of the fault handling strategy; the intelligent agents include a risk analysis agent, a unit adjustment agent, a load transfer agent, a coordination agent, a safety verification agent, and a control execution agent; After the fault handling strategy is executed, multi-dimensional evaluation indicators are set to perform a quantitative evaluation of the fault handling strategy.

2. A method for handling power grid faults based on a large model and multiple agents according to claim 1, characterized in that: Performing semantic parsing on the power grid fault information includes: Construct a power grid knowledge graph, where the entity set includes power grid equipment, sites, and topological nodes, and the relationship set includes equipment affiliation, topological connection, and protection association. Transform power grid fault information into structured fault features, including fault device identification, fault type, and fault device status; determine fault causes and associated devices through entity relationship analysis based on the power grid knowledge graph; Utilizing a large model in the electric power sector, a fault handling strategy is generated based on the structured fault characteristics and the matching results of historical fault cases.

3. A method for handling power grid faults based on a large model and multiple agents according to claim 1, characterized in that: The coordination mechanism of multiple agents includes: The risk analysis agent is used to generate a risk analysis report based on the real-time operation status of the power grid and send it to the coordination agent; The unit adjustment agent is used to respond to the risk analysis agent, generate a unit adjustment plan based on the multi-optimization objectives of the power grid, and send it to the coordination agent; The load transfer agent is parallel to the unit adjustment agent and is used to search for load transfer paths based on the power grid topology, generate a transfer plan and send it to the coordination agent; The coordination agent is used to receive risk analysis reports, unit adjustment plans, and load transfer plans, summarize them, determine the order of task execution, and generate fault handling plans; The safety verification agent is used to perform risk verification on the fault handling plan. If the handling plan passes the risk verification, it is sent to the control execution agent; if it fails, the verification result is fed back to the coordination agent; the control execution agent is used to issue control instructions to the power grid equipment.

4. A method for handling power grid faults based on a large model and multiple agents according to claim 3, characterized in that: The execution steps of the risk analysis agent include: Collect grid operating parameters, equipment rated parameters, and safe operating thresholds of operating parameters in real time, calculate the mean of the grid operating parameters within a sliding window, and preliminarily determine that an over-limit situation exists if the difference between the grid operating parameter and the corresponding mean exceeds the threshold; Based on the preliminary judgment, the time series prediction method is used to obtain the predicted value of the operating parameter. If the predicted value exceeds the threshold, it is determined that there is a risk of exceeding the limit; When the risk of overload is confirmed, the current load values ​​of all nodes affected by the risk of overload and requiring load transfer are weighted and summed to obtain the load to be transferred. Generate a risk analysis report, including out-of-limit equipment and its corresponding load to be transferred.

5. The method for handling power grid faults based on a large model and multiple agents according to claim 3, characterized in that: The execution steps of the unit adjustment agent include: An optimization model is constructed with the goal of minimizing power balance and economic cost of the power grid. The objective function includes economic cost term, load balance term, and frequency stability term. The economic cost item is the total power generation cost of all generator sets, the load balancing item is the supply and demand deviation penalty of all load nodes; the frequency stability item is the frequency deviation penalty of all frequency monitoring points; The constraints of the objective function include power balance constraints and upper and lower limit constraints of generator set output; Solving the optimization model to obtain a unit adjustment plan.

6. A method for handling power grid faults based on a large model and multiple agents according to claim 3, characterized in that: The execution steps of the load transfer agent include: For the power grid topology, the shortest path is searched as the load transfer path based on the edge weight, where the influencing factors of the edge weight include line impedance and transmission capacity; Calculate the transfer scheme score for the searched load transfer path, where the transfer scheme score is the inverse of the transfer cost; the numerator is 1, and the denominator is the weighted sum of voltage drop, line loss, and transfer capacity; The load transfer path corresponding to the maximum transfer plan score is taken as the optimal transfer plan.

7. The method for handling power grid faults based on a large model and multiple agents according to claim 1, characterized in that: The multi-dimensional evaluation indicators include fault handling efficiency, power grid restoration efficiency, resource utilization efficiency, and economic loss reduction; The fault handling time is calculated by the time difference between the fault occurrence and the fault confirmation and elimination time; the fault handling efficiency is the ratio of the fault handling time to the historical handling time of the same type of fault. The grid restoration efficiency is calculated by taking the ratio of the total load restored to the total load before the fault as the numerator and the weighted average of the power restoration time in each area of ​​the grid as the denominator. The total resource input is calculated by counting the number of devices used and the duration of use during the fault handling process. The amount of restored power is used as the actual resource output. The ratio of the actual resource output to the total resource input is used as the resource utilization efficiency. The difference between the estimated economic loss without taking any action and the actual economic loss under the fault handling strategy is used to represent the reduction in economic loss.

8. A power grid fault handling system based on a large model and multiple agents, which executes the power grid fault handling method according to any one of claims 1 to 7, characterized in that: The system comprises: Fault perception and information aggregation module; used to obtain and aggregate multi-source data of the power grid to form structured power grid fault information; the multi-source data includes fault time text information, real-time operation data, power grid topology information and environmental variables; Large model reasoning and task decomposition module; used to perform semantic analysis on the power grid fault information, generate fault handling strategies through the power field large model, and decompose them into executable subtask sequences, clarifying the input, output and execution order of each subtask; An agent collaborative execution module; used to construct multiple agents to collaboratively control and execute multiple subtasks of the fault handling strategy; the agents include a risk analysis agent, a unit adjustment agent, a load transfer agent, a coordination agent, a safety verification agent, and a control execution agent; The treatment effect evaluation and feedback optimization module is used to set multi-dimensional evaluation indicators after the fault treatment strategy is executed, and to conduct a quantitative evaluation of the fault treatment strategy.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is loaded into a processor, the power grid fault handling method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the power grid fault handling method according to any one of claims 1 to 7 is implemented.

Images