A power dispatching decision method and system based on multi-agent cooperation
By employing a multi-agent collaborative power dispatch decision-making method, this approach addresses the shortcomings of existing technologies, such as insufficient capture of deep topological relationships, inadequate multi-hop reasoning capabilities, difficulties in fusing static knowledge with dynamic data, and insufficient compliance verification. This approach enhances the accuracy and security of power dispatch decisions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-02
AI Technical Summary
Existing power dispatch decision-making technologies suffer from several problems in high-risk, highly specialized power dispatch scenarios: inability to capture deep topological relationships and causal logic between entities, insufficient multi-hop reasoning capabilities, difficulty in integrating static knowledge and dynamic data, lack of self-reflection and compliance verification, and insufficient ability to integrate and generate multi-source information.
A power dispatch decision-making method based on multi-agent collaboration is adopted. Through the collaborative work of task orchestration agent, data reconnaissance agent, graph mining agent, content generation agent and compliance verification agent, user query requests are obtained, intent is identified, real-time status information is obtained, logical chain information is retrieved, information fusion and text rendering are performed, and compliance review and optimization are carried out to finally generate power dispatch decision text.
It ensures the accuracy and security of power dispatching decisions, enhances multi-hop reasoning capabilities by preserving the topological structure and causal paths between entities, achieves seamless integration of static knowledge and dynamic data, ensures that the generated text complies with power dispatching procedures, and improves the professionalism and reliability of decision-making.
Smart Images
Figure CN122132579A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power dispatching technology, and in particular to a power dispatching decision-making method and system based on multi-agent collaboration. Background Technology
[0002] As the control center of the power grid, the power dispatching system shoulders crucial functions such as real-time monitoring, analysis and decision-making, and command issuance. The real-time performance and accuracy of its decisions directly affect the coordination efficiency of power generation, transmission, and distribution, and are fundamental to ensuring the safe and stable operation of the power grid. Delays or deviations in dispatching decisions can range from affecting power quality and economic benefits to potentially triggering localized faults or even widespread power outages, severely impacting social production and daily life.
[0003] Currently, dialogue generation technologies for assisting scheduling decisions mainly include three stages: template matching technology, end-to-end generation technology, and retrieval-enhanced generation (RAG) technology. The simplified implementation process of RAG technology includes: First, a dialogue template library for various scheduling scenarios is pre-built. User queries are mapped to corresponding templates through keyword matching, and a fixed-format response is output after filling in relevant data. Second, the organized scheduling data is used as training samples to train a general or fine-tuned generation model. After receiving user input, the model directly outputs dialogue text through semantic understanding and generation logic. Then, a retrieval engine is built, converting unstructured text, knowledge graphs, and other data into vector form to construct a retrieval library. When a user initiates a query, structured knowledge with clear logical relationships (e.g., device A connects to protection B, protection B triggers action C, and action C must follow procedure D) is compressed and projected into a continuous numerical space. Relevant information is retrieved from the retrieval library through vector similarity calculation. The retrieved results are then input into the generation model along with the user's question, and a response is generated after information fusion. Finally, the processed results are fed back to the scheduler in the form of a natural language dialogue, ensuring the readability of the output text and completing the technical implementation loop of a scheduling interaction.
[0004] However, the existing RAG technology is essentially a flat retrieval based on semantic similarity. The originally clear topological relationships and causal logic between entities are compressed into a continuous numerical space, which makes the retrieval results unable to represent the deep topological relationships and causal logic between entities (such as "equipment-protection-action-procedure"), thus reducing the accuracy of power dispatching decisions. Summary of the Invention
[0005] Therefore, it is necessary to provide a power dispatching decision-making method and system based on multi-agent collaboration to address the aforementioned technical problems.
[0006] This invention provides a power dispatching decision-making method based on multi-agent cooperation, comprising: Obtain user scheduling query requests for power system fault scenarios; The scheduling query request is input into the scheduling decision model, which includes: a task orchestration agent, a data reconnaissance agent, a graph mining agent, a content generation agent, and a compliance verification agent; the task orchestration agent includes: an intent recognition module, an information fusion module, and a final verification module. The intent recognition module parses the scheduling query request to identify the user's query intent and obtain the execution plan. Based on the execution plan, the data reconnaissance agent obtains the real-time status information of the target equipment or line from the power system monitoring data source, and the graph mining agent retrieves the logical chain information associated with the fault type in the pre-built power knowledge graph. The information fusion module semantically aligns real-time status information, logical chain information, and a subset of rules extracted from the compliance rule base to obtain a fusion matrix; the fusion matrix is then rendered using a pre-configured domain text template in the content generation agent to obtain scheduling suggestion text. The compliance verification agent reviews the dispatch suggestion text for rule compliance. If the review fails, feedback information including specific violations and correction guidelines is generated. Based on the feedback information, the content generation agent iteratively optimizes the dispatch suggestion text to obtain the optimized dispatch suggestion text. The final verification module reviews the optimized dispatch suggestion text to obtain the power dispatch decision text.
[0007] Optionally, the scheduling query request is parsed by the intent recognition module to identify the user's query intent and obtain an execution plan, specifically including: The dispatch query request is input into the intent recognition function, and the intent is classified by the big language model in the power dispatching field to obtain the scenario category to which the dispatch query request belongs. The scenario category includes one or more of the following: fault handling, instruction consultation and operation and maintenance. Based on the identified scenario category, the scheduling query request is decomposed into a set of subtasks that can be executed sequentially or in parallel using a task decomposition function; Based on the set of subtasks, a target agent is assigned to each subtask using the agent scheduling function, and corresponding subtask instructions are generated. An execution plan is then formed based on the following formula: P = { P 1, P 2, ..., P n}; P = O ( Q ); O ( Q ) = S ( T ( I ( Q ))); P t+1 = O ( Q , F t ); in, Q To schedule query requests, I ( Q ) is the intent recognition function. T ( I ) is the task decomposition function. S ( T ) is the agent scheduling function. P In order to execute the plan, P i For subtask instructions n The total number of subtasks. i =1, 2, ..., n , The set of outputs from each agent. This represents the number of iterations.
[0008] Optionally, a data reconnaissance agent obtains real-time status information of the target equipment or line from power system monitoring data sources, specifically including: Retrieve the real-time status retrieval instruction from the execution plan. The real-time status retrieval instruction includes: target device identifier and query intent. Based on the following formula, a structured query statement corresponding to the query intent is generated according to the target device identifier and the predefined SCADA data table structure: ; The real-time operating status data of target equipment or lines is obtained from power system monitoring data sources using structured query statements, and then converted into natural language descriptions. The real-time status information is obtained based on the following formula: ; in, For real-time status information, For structured query statements, For natural language conversion functions, For data query functions, For semantic transformation functions, For query intent, This is the SCADA data table structure.
[0009] Optionally, a graph mining agent retrieves logical chain information associated with fault types from a pre-built power knowledge graph, specifically including: Retrieve the knowledge association retrieval instructions from the execution plan. The knowledge association retrieval instructions include: anchor entities and query content. Starting with the anchor entity, execute within the power knowledge graph K By traversing the subgraph, a set of graph paths relating the associated anchor entities to the deep topological relationships is obtained; Based on the following formula, the graph path set is reordered using a cross-encoder of a large language model to obtain the relevance score of each graph path, and the graph paths with a relevance score greater than a set threshold are selected as the multiple logical chains most relevant to the scheduling query request: ; The logical chain information is obtained by summarizing all logical chains based on the following formula: ; ; in, The relevance score between scheduling query requests and graph paths is used. Schedule query requests, Graph path, For power knowledge graph, Linear transformation weight matrix, Bias term, Sigmoid activation function For the output function of the embedding layer, For anchor point entities, This is logical chain information.
[0010] Optionally, the information fusion module semantically aligns real-time status information, logical chain information, and a subset of rules extracted from the compliance rule base to obtain a fusion matrix, specifically including: Real-time status information from data reconnaissance agents is used as real-time data, and logical chain information from graph mining agents is used as graph knowledge chain. Extract a subset of rules from the compliance rule base based on the current task scenario, and use the extracted subset of rules as compliance constraints; Based on the following formula, real-time data, graph knowledge chains, and compliance constraints are semantically aligned to obtain the fusion matrix: ; in, For the fusion matrix, For real-time data, For graph knowledge chain, For compliance constraints, This is a vertical splicing operation.
[0011] Optionally, the fusion matrix is rendered using a pre-configured domain text template in the content generation agent to obtain scheduling suggestion text, specifically including: Retrieve the initial text generation instructions from the execution plan; Select a domain text template based on the scheduling query request. The domain text template includes at least one of the following categories: fault handling, instruction consultation, and operation and maintenance. Based on the following formula, the semantic information in the fusion matrix is filled and rendered according to the template logical structure to obtain the scheduling suggestion text: ; in, Functions for constructing the fusion matrix, For template adaptation functions, For text generation functions, To optimize the correction function.
[0012] Optionally, a compliance verification agent reviews the scheduling suggestion text for rule compliance. If the review fails, feedback information including specific violations and correction guidelines is generated. Based on the feedback information, a content generation agent iteratively optimizes the scheduling suggestion text, specifically including: Retrieve compliance verification instructions from the execution plan; Based on the following formula, the dispatch suggestion text is matched with the preset power dispatch procedure rule base according to the compliance verification instruction to obtain the rule compliance review result: ; in, For the results of the rule compliance review, For rule matching functions, For compliance judgment functions, To revise the suggested function; Based on the following formula, the scheduling suggestion text is iteratively optimized using a content generation agent based on feedback information to obtain the optimized scheduling suggestion text: ; in, The optimized scheduling suggestion text, To optimize the number of times, For the first Feedback on the compliance verification.
[0013] Optionally, the optimized scheduling suggestion text is reviewed by the final verification module to obtain the power dispatching decision text, which specifically includes: The optimized dispatch suggestion text was reviewed from three aspects: logical coherence, standardization of power dispatch technical terminology, and adaptability to dispatch query requests. After verification, the optimized dispatch suggestion text was used as the final power dispatch decision text.
[0014] This invention provides a power dispatching decision system based on multi-agent cooperation, comprising: The query request module is used to obtain user scheduling query requests for power system fault scenarios; The model building module is used to input scheduling query requests into the scheduling decision model. The scheduling decision model includes: a task orchestration agent, a data reconnaissance agent, a graph mining agent, a content generation agent, and a compliance verification agent. The task orchestration agent includes: an intent recognition module, an information fusion module, and a final verification module. The intent recognition module is used to parse scheduling query requests to identify the user's query intent and obtain an execution plan. The information acquisition module is used to obtain real-time status information of target equipment or lines from power system monitoring data sources through a data reconnaissance agent based on the execution plan, and to retrieve logical chain information associated with fault types in a pre-built power knowledge graph through a graph mining agent. The initial generation module is used to semantically align real-time status information, logical chain information, and a subset of rules extracted from the compliance rule base through the information fusion module to obtain a fusion matrix; the fusion matrix is then rendered using a pre-configured domain text template in the content generation agent to obtain scheduling suggestion text. The compliance review module is used to review the dispatch suggestion text for rule compliance through a compliance verification agent. If the review fails, feedback information including specific violations and correction guidelines is generated, and the dispatch suggestion text is iteratively optimized by a content generation agent based on the feedback information. The optimized dispatch suggestion text is then reviewed by the final verification module to obtain the power dispatch decision text.
[0015] The power dispatching decision-making method and system based on multi-agent cooperation provided in this invention have the following advantages compared with the prior art: This invention retrieves fault-related "equipment-protection-action-procedure" logical chain information from a power knowledge graph. By using graph relationship retrieval instead of flat semantic retrieval, it can completely preserve the inherent topological structure and causal path between entities. This avoids the relationship loss and logic dilution caused by compressing structured knowledge into a vector space in traditional dialogue generation technology. It ensures that the recalled information is not a fragmented piece of information but a coherent reasoning chain with embedded domain logic, thereby improving the accuracy of power dispatching decisions. Attached Figure Description
[0016] Figure 1 This is a diagram illustrating the overall system architecture of a power dispatching decision-making method based on multi-agent collaboration, as provided in one embodiment. Figure 2 This is a flowchart illustrating the knowledge graph construction and indexing process of a power dispatching decision-making method based on multi-agent collaboration, as provided in one embodiment. Figure 3 This is a time-series swimlane diagram of a power dispatching decision-making method based on multi-agent cooperation provided in one embodiment. Detailed Implementation
[0017] 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 embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0018] While existing dialogue generation technologies for assisting dispatching decisions have improved interaction efficiency, they still face the following core bottlenecks in high-risk, highly specialized power dispatching scenarios: 1. Unstructured retrieval leads to broken logical chains: Traditional RAG is based on vector similarity for flat retrieval, which cannot capture the deep topological relationships and causal logic between entities (such as "device-protection-action-procedure"), and is prone to missing key constraints.
[0019] 2. Lack of planning and reasoning ability for complex tasks: A single model architecture is difficult to handle multi-hop reasoning tasks such as "first querying the real-time status, then combining fault waveform analysis, and finally giving suggestions based on the procedure".
[0020] 3. Difficulty in integrating static knowledge with dynamic data: Existing technologies struggle to seamlessly integrate SCADA real-time cross-sectional data with static principle knowledge in knowledge graphs within the same semantic space.
[0021] 4. Lack of self-reflection and compliance verification: The general model lacks strict boundary constraints, which can easily generate operation instructions that violate the "Power Dispatch Regulations" (such as disconnecting the switch under load), posing a safety hazard.
[0022] 5. Insufficient ability to integrate and generate multi-source information: Existing technologies have not designed a dedicated content generation mechanism for power dispatching scenarios, making it difficult to efficiently integrate real-time data, graph logic, and compliance rules into structured, professional, and natural dialogue text.
[0023] This invention provides a power dispatching decision-making method based on multi-agent cooperation, the method comprising: Obtain user scheduling query requests for power system fault scenarios.
[0024] The scheduling query request is input into the scheduling decision model, which includes: a task orchestration agent, a data reconnaissance agent, a graph mining agent, a content generation agent, and a compliance verification agent. The task orchestration agent includes: an intent recognition module, an information fusion module, and a final verification module.
[0025] The intent recognition module parses the scheduling query request to identify the user's query intent and obtain an execution plan. The execution plan includes: real-time status acquisition instructions, knowledge association retrieval instructions, multi-source information fusion instructions, preliminary text generation instructions, and compliance verification instructions.
[0026] Based on the execution plan, the data reconnaissance agent obtains real-time status information of target equipment or lines from power system monitoring data sources, and the graph mining agent retrieves logical chain information associated with fault types in a pre-built power knowledge graph.
[0027] The information fusion module semantically aligns real-time status information, logical chain information, and a subset of rules extracted from the compliance rule base to obtain a fusion matrix. The fusion matrix is then rendered using pre-configured domain text templates in the content generation agent to obtain scheduling suggestion text.
[0028] A compliance verification agent reviews the dispatch suggestion text for rule compliance. If the review fails, feedback information including specific violations and correction guidelines is generated. Based on the feedback, a content-generating agent iteratively optimizes the dispatch suggestion text to obtain the optimized dispatch suggestion text. The final verification module reviews the optimized dispatch suggestion text to obtain the power dispatch decision text.
[0029] Specifically, the intent recognition module parses the scheduling query request to identify the user's query intent and obtain an execution plan, which includes: The dispatch query request is input into the intent recognition function, and the intent is classified by the big language model in the power dispatching field to obtain the scenario category to which the dispatch query request belongs. The scenario category includes one or more of the following: fault handling, instruction consultation and operation and maintenance.
[0030] Based on the identified scenario category, the scheduling query request is broken down into a set of subtasks that can be executed sequentially or in parallel using a task decomposition function.
[0031] Based on the set of subtasks, a target agent is assigned to each subtask using the agent scheduling function, and corresponding subtask instructions are generated. An execution plan is then formed based on the following formula: P = { P 1, P 2, ...,P n}; P = O ( Q ); O ( Q ) = S ( T ( I ( Q ))); P t+1 = O ( Q , F t ); in, Q To schedule query requests, I ( Q ) is the intent recognition function. T ( I ) is the task decomposition function. S ( T ) is the agent scheduling function. P In order to execute the plan, P i For subtask instructions n The total number of subtasks. i =1, 2, ..., n , For the feedback set of each agent, This represents the number of iterations.
[0032] Specifically, the real-time status information of target equipment or lines is obtained from power system monitoring data sources through a data reconnaissance intelligent agent, including: Retrieve the real-time status retrieval instruction from the execution plan. The real-time status retrieval instruction includes: target device identifier and query intent.
[0033] Based on the following formula, a structured query statement corresponding to the query intent is generated according to the target device identifier and the predefined SCADA data table structure: ; The real-time operating status data of target equipment or lines is obtained from power system monitoring data sources using structured query statements, and then converted into natural language descriptions. The real-time status information is obtained based on the following formula: ; in, For real-time status information, For structured query statements, For natural language conversion functions, For data query functions, For semantic transformation functions, For query intent, This is the SCADA data table structure.
[0034] Specifically, the process involves using a graph mining agent to retrieve logical chain information associated with fault types from a pre-constructed power knowledge graph. Retrieve knowledge association retrieval instructions from the execution plan. Knowledge association retrieval instructions include: anchor entities and query content.
[0035] Starting with the anchor entity, execute within the power knowledge graph K By traversing the subgraph, a set of graph paths relating the anchor point entities to the deep topological relationships is obtained.
[0036] Based on the following formula, the graph path set is reordered using a cross-encoder of a large language model to obtain the relevance score of each graph path, and the graph paths with a relevance score greater than a set threshold are selected as the multiple logical chains most relevant to the scheduling query request: ; The logical chain information is obtained by summarizing all logical chains based on the following formula.
[0037] ; ; in, The relevance score between scheduling query requests and graph paths is used. Schedule query requests, Graph path, For power knowledge graph, Linear transformation weight matrix, Bias term, Sigmoid activation function For the output function of the embedding layer, For anchor point entities, This is logical chain information.
[0038] Specifically, the information fusion module semantically aligns real-time status information, logical chain information, and a subset of rules extracted from the compliance rule base to obtain a fusion matrix, which includes: Real-time status information from the data reconnaissance agent is used as real-time data, and logical chain information from the graph mining agent is used as the graph knowledge chain. A subset of rules is extracted from the compliance rule base based on the current task scenario, and this extracted subset of rules is used as compliance constraints.
[0039] Based on the following formula, real-time data, graph knowledge chains, and compliance constraints are semantically aligned to obtain the fusion matrix: ; in, For the fusion matrix, For real-time data, For graph knowledge chain, For compliance constraints, This is a vertical splicing operation.
[0040] Specifically, the fusion matrix is rendered using pre-configured domain text templates in the content generation agent to obtain scheduling suggestion text, which includes: Retrieve the preliminary text generation instructions from the execution plan. Select a domain text template based on the scheduling query request. The domain text template includes at least one of the following categories: fault handling, instruction consultation, and operation and maintenance.
[0041] Based on the following formula, the semantic information in the fusion matrix is filled and rendered according to the template logical structure to obtain the scheduling suggestion text: ; in, Functions for constructing the fusion matrix, For template adaptation functions, For text generation functions, To optimize the correction function.
[0042] Specifically, a compliance verification agent reviews the scheduling suggestion text for rule compliance. If the review fails, feedback information including specific violations and correction guidelines is generated. Based on the feedback information, a content generation agent iteratively optimizes the scheduling suggestion text, including: Retrieve compliance verification instructions from the execution plan. Based on the following formula, match the dispatch suggestion text with the preset power dispatch procedure rule base according to the compliance verification instructions to obtain the rule compliance review result:
[0043] ; in, For the results of the rule compliance review, For rule matching functions, For compliance judgment functions, To revise the suggested function.
[0044] Based on the following formula, the scheduling suggestion text is iteratively optimized using a content generation agent based on feedback information to obtain the optimized scheduling suggestion text: ; in, The optimized scheduling suggestion text, To optimize the number of times, For the first Feedback on the compliance verification.
[0045] The optimized scheduling suggestion text is reviewed by the final verification module to obtain the power dispatching decision text, which specifically includes: The optimized dispatch suggestion text was reviewed from three aspects: logical coherence, standardization of power dispatch technical terminology, and adaptability to dispatch query requests. After verification, the optimized dispatch suggestion text was used as the final power dispatch decision text.
[0046] A specific embodiment of the present invention is provided: like Figure 1 As shown, the scheduling decision model includes five core intelligent agents: task orchestration agent, graph mining agent, data reconnaissance agent, compliance verification agent, and content generation agent.
[0047] 1. Knowledge graph construction based on pattern constraints.
[0048] The semantic understanding capabilities of LLM are used to replace the traditional sequence labeling model. To ensure the accuracy of power knowledge, a standardized Chinese graph schema is predefined and strict constraints are implemented during the construction phase.
[0049] 1.1 Definition of Graph Patterns.
[0050] Entity type set: {"Substation", "Transformer", "Busbar", "Circuit Breaker", "Disconnect Switch", "Protection Device", "Automatic Device", "Dispatching Procedure", "Operation Ticket"}.
[0051] The set of relationship types is: {"Electrical connection", "Belongs to", "Protection range includes", "Action trigger", "Lock-in condition", "Procedure constraint", "Adjacent association"}.
[0052] 1.2 Knowledge Extraction Prompt Template.
[0053] Triple extraction is performed using a standardized template consisting of "role setting + constraints + thought chain demonstration": [Character Setting] You are a senior expert in building knowledge graphs for power dispatching. Based on the defined [Chinese Schema], please extract entity relation triples from the input text.
[0054] [Constraints] Creating new entity types or relationship types outside of the schema definition is strictly prohibited.
[0055] [Mind Chain Demonstration] Input: "Main transformer differential protection tripped, tripping all switches on both sides."
[0056] Thinking process: Identify that the "differential protection" (protection device) belongs to the "main transformer" (transformer); the "differential protection" action causes the "switch" (circuit breaker) to trip.
[0057] Output JSON: [{"Header Entity":"Differential Protection","Type":"Protection Device","Relationship":"Action Trigger","Tail Entity":"Switch","Type":"Circuit Breaker"}).
[0058] 1.3 Multi-granularity index construction.
[0059] Micro-index: Vector index based on triples.
[0060] Macro Index (Community Summary): The Leiden algorithm is used to cluster the map and generate community summaries (such as "This community describes the 500kV side bus and protection configuration of Anyuan Station"), which supports macro situation retrieval.
[0061] 2. Data Scout Agent and Dynamic Fusion.
[0062] The data reconnaissance agent is responsible for solving the problem of integrating static knowledge with dynamic data. It transforms natural language into structured queries through tool learning.
[0063] 2.1 Structured query template generation.
[0064] [Task] Transform natural language into SQL queries for the SCADA system.
[0065] [Table Structure] Table Name: Real-time Operating Status Table; Fields: Device ID, Switch Status (0=Off, 1=On), Active Power, Protection Signal.
[0066] [Input] "Query the status and load of switch 5012".
[0067] [Output] SELECT switch status, active power FROM real-time operating status table WHERE device ID='5012' AND time='latest'.
[0068] 2.2 Semantic feedback.
[0069] After the agent executes the SQL, it converts the result {switch status: 0, active power: 0} into natural language: “The current 5012 switch is in the open state, and the active power is 0” and feeds it back to the task orchestration agent.
[0070] 2.3 Mathematical representation of data transmission stream.
[0071] Input: Query intent (From scheduling query request), SCADA data schema (Table structure, field definitions).
[0072] Processing function: .
[0073] Output: Real-time status information (Real-time data described in natural language).
[0074] Mathematical expression: .
[0075] Data transmission stream: .
[0076] in, This is a structured query statement.
[0077] 3. Graph Miner Agent and Deep Search.
[0078] like Figure 2 As shown, the graph mining agent performs the GraphRAG operation to resolve the logic breakage problem.
[0079] 3.1 Local Search: Using the faulty device as the anchor point, a K-hop subgraph traversal is performed to obtain the complete logical chain of "device-protection-procedure". Through subgraph traversal, the deep topological relationships and causal logic between entities can be fully captured, avoiding the problem of missing key constraints caused by traditional flat search.
[0080] 3.2 Reranking: To further improve the relevance matching accuracy between search results and scheduling query requests, a cross-encoder based on a Large Language Model (LLM) is introduced to re-rank the retrieved graph paths. With scheduling query request Perform a relevance score.
[0081] Compared to traditional BERT-based encoding methods, this invention uses LLM as the core encoding unit. Its advantages lie in its stronger contextual semantic understanding, domain knowledge modeling, and complex logical reasoning capabilities. It can deeply capture the professional relationships between dispatch query requests and graph paths in power dispatch scenarios (such as the adaptability of fault types to handling procedures and the correspondence between equipment actions and protection logic), avoiding the problem of insufficient semantic understanding in professional domains in general pre-trained models. Furthermore, by fine-tuning the LLM with power dispatch domain corpora, its sensitivity to domain terminology and professional logic can be further enhanced, improving the accuracy of relevance scoring.
[0082] The improved relevance scoring formula is as follows: ; Explanation of the parameters and their meanings in the formula: Scheduling query request With the map path The correlation score ranges from [0, 1], with a score closer to 1 indicating a stronger correlation. Dispatch query requests (e.g., "How to handle the tripping of line 5012"); The path obtained from local retrieval (e.g., "reclosing failure → means → permanent fault → constraint → forced transmission is strictly prohibited"). : Indicates that the query request will be scheduled. With the map path Perform text concatenation to form a unified input sequence (concatenation format: "Schedule query request: { Q}; Graph path: { }”, ensuring that LLM can capture the semantic relationship between the two simultaneously; The LLM embedding layer output function transforms the concatenated text sequence into a high-dimensional semantic vector. The LLM used here can be a large model such as the GPT series or LLaMA series. After fine-tuning with corpus data from the power dispatching domain, it can accurately extract semantic features for specialized scenarios.
[0083] : Linear transformation weight matrix, used to perform dimension mapping and feature enhancement on the high-dimensional semantic vector output by LLM; : Bias term, used to adjust the offset of the linear transformation; The Sigmoid activation function maps the linearly transformed result to the interval [0, 1] to obtain the final relevance score.
[0084] 3.3 Reordering execution process.
[0085] First, all graph paths obtained from the local retrieval are relevance-scored using the formula described above. Then, they are sorted from highest to lowest score, and the Top-K highly relevant paths are selected as the final retrieval results. This ensures that the subsequently generated dialogue text unfolds based on the most core and matching logical chain, improving the accuracy of decision-making. Data transmission streaming mathematical expression:
[0086] Input: Arrangement instructions (Anchor point entity) Scheduling query request ), Electricity Knowledge Graph (node ,side Community Index (Leiden clustering results).
[0087] Processing function: .
[0088] Output: Logical chain information (each) (The path sequence is "entity-relationship-entity") Mathematical expression: ,in .
[0089] Data transmission stream: .
[0090] 4. Compliance Critic Agent and Reflection Mechanism.
[0091] The compliance verification agent acts as the "gatekeeper" of the system and has a built-in rule base based on the "Power Dispatch Regulations".
[0092] 4.1 Compliance verification template.
[0093] [Role] Power Dispatch Safety Inspector.
[0094] [Rule Base] It is strictly forbidden to operate the disconnect switch under load.
[0095] After a failed reclosing or a permanent fault, it is strictly forbidden to blindly force power back on.
[0096] [Input] Proposed command: "Close 5012-1 disconnect switch." (Note: Switch 5012 is in the closed position at this time.)
[0097] [Output] REJECT: Rule 1 violated. Switch detected in the closed position. It is strictly forbidden to open or close the disconnect switch under load.
[0098] 4.2 Mathematical representation of data transmission stream.
[0099] Input: Arrangement instructions (Including content to be verified) ), rule base (A collection of rules from the "Power Dispatch Regulations") Processing function: .
[0100] Output: Verification result ( , (Provide natural language correction suggestions).
[0101] Mathematical expression: .
[0102] Data transmission stream: .
[0103] 5. Content Generator Agent and Multi-Source Fusion Text Generation.
[0104] The content-generating intelligent agent is responsible for integrating multi-source heterogeneous information into professional, accurate, and natural dialogue text that conforms to the power dispatching scenario, thus solving the problem of fragmented integration of multi-source information.
[0105] 5.1 Multi-source input fusion model.
[0106] Define the input data set, including real-time data. (Real-time status information from the data reconnaissance agent), graph knowledge chain (Logic chain information from graph mining agents), compliance constraints (A subset of the rule base preloaded by the compliance verification intelligent agent) to construct a fusion matrix (semicolon (This represents a vertical splicing operation), enabling semantic alignment and structured integration of multi-source information.
[0107] 5.2 Generate templates and constraints.
[0108] A standardized template combining "scenario adaptation, logical progression, and standardized terminology" is used to ensure the professionalism and readability of the output text. [Scenario Adaptation] Fault Handling / Command Consultation / Operation and Maintenance (Automatically switching based on the intent recognition results of the task orchestration agent) [Logical progression] Current situation description → Cause analysis → Handling suggestions → Safety tips.
[0109] [Terminology Standards] Strictly adhere to the "Power Dispatch Technical Terminology," such as "disconnect switch" must not be abbreviated as "knife switch" (except in compliance verification scenarios where the original instruction must be used). [Example Input] (Merged Data): .
[0110] [Example Output]: "According to SCADA real-time data, the distance to line 5012 is..." I After the section protection activated, the switch tripped, and the reclosing operation failed, indicating a permanent fault. Based on power dispatching regulations and related knowledge graph logic, the following measures are recommended: 1. Open the isolating switches on both sides of line 5012 to isolate the fault; 2. Strictly prohibit any forced energization of this line to avoid amplifying the fault's impact; 3. Notify maintenance personnel to conduct a full-line fault inspection and locate the fault point. During the operation, the isolating switch operation specifications must be strictly followed, and opening or closing the switch under load is strictly prohibited.
[0111] 5.3 Multi-round optimization mechanism.
[0112] Receive rejection feedback (if any) from the compliance verification agent, extract correction keywords (such as "forced power supply" and "operation of switch under load"), focus on the correction area through an attention mechanism, re-optimize and generate text, and ensure that the output meets compliance requirements.
[0113] 5.4 Mathematical representation of data transmission streams.
[0114] Input: Arrangement instructions (Including generated scenes) ), fusion matrix ( (For compliance constraints).
[0115] Processing function: .
[0116] Output: Generated text (Professional dialogue text appropriate to the context).
[0117] Mathematical expression: ( To optimize the number of times, For the first (Sub-compliance verification feedback).
[0118] Data transmission stream: .
[0119] 6. Task orchestration agent (Orchestrator Agent) collaborative scheduling mechanism.
[0120] As the core controller of a multi-agent system, the task orchestration agent is responsible for task decomposition, agent scheduling, and feedback integration to ensure the closed-loop operation of the system.
[0121] Mathematical expression of data transmission stream: Input: Scheduling query request (Natural language) and the set of outputs from each intelligent agent ( The index represents the iteration number, and the subscript corresponds to the first iteration of each agent. Secondary feedback output).
[0122] Processing function: .
[0123] Output: Execution Plan P = { P 1, P 2, ..., P n}( (This is a subtask instruction, including the target agent identifier).
[0124] Mathematical expression: P = O ( Q ), O ( Q ) = S ( T ( I ( Q ))), P t+1 = O ( Q , F t ).
[0125] Data transmission stream: .
[0126] Another specific embodiment of the present invention is provided: The following example illustrates how to handle a circuit breaker trip. Figure 3 The complete execution flow of this invention is illustrated below: Step S1: Intent analysis and planning.
[0127] The task orchestration agent receives the user command "Line 5012 has tripped, what should I do?", identifies the intent as fault handling, and formulates an execution plan: [Get real-time status] → [Retrieve related knowledge] → [Multi-source information fusion] → [Generate preliminary text] → [Compliance check] → [Optimize final output].
[0128] Step S2: Real-time status acquisition.
[0129] The data reconnaissance agent receives the query instructions from the task orchestration agent, generates a query statement using an SQL template, and calls the SCADA interface.
[0130] Real-time status information (Real-time data) ): Confirmed that switch 5012 has tripped, distance I protection has activated, and the reclosing signal is "unsuccessful".
[0131] Step S3: Knowledge-related retrieval.
[0132] The graph mining agent uses "distance protection" and "reclosing failure" as anchor entities to perform K-hop subgraph retrieval and path reordering in the knowledge graph.
[0133] Logical chain information (Graphical Knowledge Chain) ): Extract path distance protection -> (belongs to) -> line protection -> (procedure) -> line inspection and reclosing failure -> (means) -> permanent fault -> (constraint) -> forced reclosing is strictly prohibited.
[0134] Step S4: Multi-source information fusion.
[0135] Task orchestration agent receiving and and integrate compliance constraints from the compliance rule base. Construct a fusion matrix The fusion matrix is then sent to the content generation agent.
[0136] Step S5: Preliminary text generation.
[0137] The content generation agent loads fault handling scenario templates based on a fusion matrix. The initial recommendation text is generated as follows: "According to real-time SCADA data, the switch on line 5012 tripped after the distance I protection activated, and the reclosing operation failed, indicating a permanent fault. Recommendations: 1. Close the disconnect switches on both sides of line 5012 to isolate the fault; 2. Attempt to energize the line; 3. Notify maintenance personnel to conduct a full-line inspection."
[0138] Step S6: Reflection and Correction.
[0139] The compliance verification agent receives the initially generated text and performs rule matching: The system detected a risk of operating under load when “closing the disconnect switches on both sides of line 5012” (the current switch is not clearly confirmed to be open). “Attempting to energize” violates the rule that “forced energizing is strictly prohibited for permanent faults”, triggering REJECT feedback and providing the following correction suggestions: “1. Before operation, the switch open status must be confirmed, and the fault should be isolated by ‘opening the disconnect switch’; 2. Delete the relevant statements about energizing and clearly prohibit forced energizing.”
[0140] The content-generating agent receives correction suggestions, focuses on the violation areas through an attention mechanism, and re-optimizes the generated text.
[0141] Step S7: Final output.
[0142] The task orchestration agent confirms that the optimized text passes compliance verification and outputs the final instruction to the user: "According to SCADA data, the 5012 line's distance I protection has activated and reclosing has failed, indicating a permanent fault. Recommendations: 1. Isolate the fault by opening both disconnect switches; 2. Do not force power to this line; 3. Notify personnel to conduct a full-line fault inspection. During operation, strictly adhere to the disconnect switch operation specifications, and do not operate the switch under load."
[0143] Based on the same inventive concept, embodiments of the present invention also provide a power dispatching decision-making system based on multi-agent cooperation, the system comprising: The query request module is used to obtain user scheduling query requests for power system fault scenarios.
[0144] The model building module is used to input scheduling query requests into the scheduling decision model, which includes: a task orchestration agent, a data reconnaissance agent, a graph mining agent, a content generation agent, and a compliance verification agent. The task orchestration agent includes: an intent recognition module, an information fusion module, and a final verification module.
[0145] The intent recognition module is used to parse scheduling query requests to identify the user's query intent and obtain an execution plan.
[0146] The information acquisition module is used to obtain real-time status information of target equipment or lines from power system monitoring data sources through a data reconnaissance agent based on the execution plan, and to retrieve logical chain information associated with fault types in a pre-built power knowledge graph through a graph mining agent.
[0147] The initial generation module semantically aligns real-time state information, logical chain information, and a subset of rules extracted from the compliance rule base using the information fusion module to obtain a fusion matrix. The fusion matrix is then rendered using pre-configured domain text templates in the content generation agent to obtain scheduling suggestion text.
[0148] The compliance review module is used to review the dispatch suggestion text for rule compliance through a compliance verification agent. If the review fails, feedback information including specific violations and correction guidelines is generated, and the dispatch suggestion text is iteratively optimized based on the feedback information using a content-generating agent. The optimized dispatch suggestion text is then reviewed by the final verification module to obtain the power dispatch decision text.
[0149] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.
Claims
1. A power dispatching decision-making method based on multi-agent cooperation, characterized in that, include: Obtain user scheduling query requests for power system fault scenarios; The scheduling query request is input into the scheduling decision model, which includes: a task orchestration agent, a data reconnaissance agent, a graph mining agent, a content generation agent, and a compliance verification agent; the task orchestration agent includes: an intent recognition module, an information fusion module, and a final verification module. The scheduling query request is parsed by the intent recognition module to identify the user's query intent and obtain an execution plan; Based on the execution plan, the data reconnaissance agent obtains the real-time status information of the target equipment or line from the power system monitoring data source, and the graph mining agent retrieves the logical chain information associated with the fault type in the pre-built power knowledge graph. The information fusion module semantically aligns real-time status information, logical chain information, and a subset of rules extracted from the compliance rule base to obtain a fusion matrix; the fusion matrix is then rendered using a pre-configured domain text template in the content generation agent to obtain scheduling suggestion text. The compliance verification agent reviews the dispatch suggestion text for rule compliance. If the review fails, feedback information including specific violations and correction guidelines is generated. Based on the feedback information, the content generation agent iteratively optimizes the dispatch suggestion text to obtain the optimized dispatch suggestion text. The final verification module reviews the optimized dispatch suggestion text to obtain the power dispatch decision text.
2. The power dispatching decision-making method based on multi-agent cooperation as described in claim 1, characterized in that, The step of parsing the scheduling query request through the intent recognition module to identify the user's query intent and obtain an execution plan specifically includes: The dispatch query request is input into the intent recognition function, and the intent is classified by the big language model in the power dispatching field to obtain the scenario category to which the dispatch query request belongs. The scenario category includes one or more of the following: fault handling, instruction consultation and operation and maintenance. Based on the identified scenario category, the scheduling query request is decomposed into a set of subtasks that can be executed sequentially or in parallel using a task decomposition function; Based on the set of subtasks, a target agent is assigned to each subtask using the agent scheduling function, and corresponding subtask instructions are generated. An execution plan is then formed based on the following formula: P = { P 1, P 2, ..., P n}； P = O ( Q )； O ( Q ) = S ( T ( I ( Q )))； P t+1 = O （ Q , F t ）； in, Q To schedule query requests, I ( Q ) is the intent recognition function. T ( I ) is the task decomposition function. S ( T ) is the agent scheduling function. P In order to execute the plan, P i For subtask instructions n The total number of subtasks. i =1, 2, ..., n , The set of outputs from each agent. This represents the number of iterations.
3. The power dispatching decision-making method based on multi-agent cooperation as described in claim 1, characterized in that, The process of obtaining real-time status information of target equipment or lines from power system monitoring data sources through a data reconnaissance intelligent agent specifically includes: Obtain the real-time status acquisition instruction from the execution plan, wherein the real-time status acquisition instruction includes: target device identifier and query intent; Based on the following formula, a structured query statement corresponding to the query intent is generated according to the target device identifier and the predefined SCADA data table structure: ； The real-time operating status data of target equipment or lines is obtained from power system monitoring data sources using structured query statements, and then converted into natural language descriptions. The real-time status information is obtained based on the following formula: ； in, For real-time status information, For structured query statements, For natural language conversion functions, For data query functions, For semantic transformation functions, For query intent, This is the SCADA data table structure.
4. The power dispatching decision-making method based on multi-agent cooperation as described in claim 1, characterized in that, The step of retrieving logical chain information associated with fault types from a pre-constructed power knowledge graph using a graph mining agent specifically includes: Obtain the knowledge association retrieval instruction from the execution plan, wherein the knowledge association retrieval instruction includes: anchor entity and query content; Starting with the anchor entity, execute within the power knowledge graph K By traversing the subgraph, a set of graph paths relating the associated anchor entities to the deep topological relationships is obtained; Based on the following formula, the graph path set is reordered using a cross-encoder of a large language model to obtain the relevance score of each graph path, and the graph paths with relevance scores greater than a set threshold are identified as the multiple logical chains most relevant to the scheduling query request: ； The logical chain information is obtained by summarizing all logical chains based on the following formula: ； ； in, The relevance score between scheduling query requests and graph paths is used. To schedule query requests, For the map path, For power knowledge graph, Linear transformation weight matrix, For bias terms, It is the Sigmoid activation function. For the output function of the embedding layer, For anchor point entities, This is logical chain information.
5. A power dispatching decision-making method based on multi-agent cooperation as described in claim 3 or 4, characterized in that, The step involves semantically aligning real-time status information, logical chain information, and a subset of rules extracted from the compliance rule base using an information fusion module to obtain a fusion matrix. Specifically, this includes: Real-time status information from data reconnaissance agents is used as real-time data, and logical chain information from graph mining agents is used as graph knowledge chain. Extract a subset of rules from the compliance rule base based on the current task scenario, and use the extracted subset of rules as compliance constraints; Based on the following formula, real-time data, graph knowledge chains, and compliance constraints are semantically aligned to obtain the fusion matrix: ； in, For the fusion matrix, For real-time data, For graph knowledge chain, For compliance constraints, This is a vertical splicing operation.
6. The power dispatching decision-making method based on multi-agent cooperation as described in claim 1, characterized in that, The step of rendering the fusion matrix using a pre-configured domain text template in the content generation agent to obtain scheduling suggestion text specifically includes: Retrieve the initial text generation instructions from the execution plan; Select a domain text template based on the scheduling query request. The domain text template includes at least one of the following: fault handling, instruction consultation, and operation and maintenance. Based on the following formula, the semantic information in the fusion matrix is filled and rendered according to the template logical structure to obtain the scheduling suggestion text: ； in, Functions for constructing the fusion matrix, For template adaptation functions, For text generation functions, To optimize the correction function.
7. The power dispatching decision-making method based on multi-agent cooperation as described in claim 1, characterized in that, The compliance verification agent reviews the scheduling suggestion text for rule compliance. If the review fails, feedback information including specific violations and correction guidelines is generated. Based on the feedback information, the content generation agent iteratively optimizes the scheduling suggestion text, specifically including: Retrieve compliance verification instructions from the execution plan; Based on the following formula, the dispatch suggestion text is matched with the preset power dispatch procedure rule base according to the compliance verification instruction to obtain the rule compliance review result: ； in, For the results of the rule compliance review, For rule matching functions, For compliance judgment functions, To revise the suggested function; Based on the following formula, the scheduling suggestion text is iteratively optimized using a content generation agent based on feedback information to obtain the optimized scheduling suggestion text: ； in, The optimized scheduling suggestion text, To optimize the number of times, For the first Feedback on the compliance verification.
8. The power dispatching decision-making method based on multi-agent cooperation as described in claim 7, characterized in that, The final verification module reviews the optimized scheduling suggestion text to obtain the power dispatching decision text, which specifically includes: The optimized dispatch suggestion text was reviewed from three aspects: logical coherence, standardization of power dispatch technical terminology, and adaptability to dispatch query requests. After verification, the optimized dispatch suggestion text was used as the final power dispatch decision text.
9. A power dispatching decision-making system based on multi-agent cooperation, characterized in that, include: The query request module is used to obtain user scheduling query requests for power system fault scenarios; The model building module is used to input scheduling query requests into the scheduling decision model, which includes: a task orchestration agent, a data reconnaissance agent, a graph mining agent, a content generation agent, and a compliance verification agent; the task orchestration agent includes: an intent recognition module, an information fusion module, and a final verification module. The intent recognition module is used to parse the scheduling query request to identify the user's query intent and obtain an execution plan; The information acquisition module is used to acquire real-time status information of target equipment or lines from power system monitoring data sources through a data reconnaissance agent based on the execution plan, and to retrieve logical chain information associated with fault types in a pre-built power knowledge graph through a graph mining agent. The initial generation module is used to semantically align real-time status information, logical chain information, and a subset of rules extracted from the compliance rule base through the information fusion module to obtain a fusion matrix; the fusion matrix is then rendered using a pre-configured domain text template in the content generation agent to obtain scheduling suggestion text. The compliance review module is used to review the dispatch suggestion text for rule compliance through a compliance verification agent. If the review fails, feedback information including specific violations and correction guidelines is generated, and the dispatch suggestion text is iteratively optimized by a content generation agent based on the feedback information. The optimized dispatch suggestion text is then reviewed by the final verification module to obtain the power dispatch decision text.