Power-off protection auxiliary decision-making method and system based on multi-agent collaborative adaptive graph structure retrieval enhancement generation technology

Abstract

The invention discloses a power-off protection auxiliary decision-making method and system based on a multi-agent collaborative adaptive graph structure retrieval enhancement generation technology, and belongs to the technical field of intelligent power grids. In order to solve the illusion problem existing in the application of a large model in the professional field of relay protection, the problem of index errors caused by time ambiguity in knowledge modeling, and the problems of low reasoning efficiency and insufficient accuracy of a traditional retrieval enhancement generation method, the method comprises the following steps: constructing a power-off protection knowledge graph; performing tense semantic analysis on an input query by utilizing a large language model, and decomposing the query into sub-queries when the query contains a time constraint; retrieving the tense sub-graph, and determining effective nodes by using a time sequence self-adaptive three-layer interactive retrieval module; retrieving an effective source text from a tense source text power-off protection database; and inputting the effective graph information, the effective text information and query into a large language model to generate a power-off protection auxiliary decision answer. The method is used for power-off protection of the power grid.

Description

Technical Field

[0001] This invention relates to a power outage protection auxiliary decision-making method and system, belonging to the field of smart grid technology. Background Technology

[0002] With the rapid development of ultra-high voltage power grids and the continuous expansion of power system scale, their operating characteristics and fault modes are becoming increasingly complex and diverse. How to achieve rapid and accurate diagnosis and handling of power system faults has become a key issue in ensuring the safe and stable operation of the power grid.

[0003] Traditional relay protection diagnosis and decision-making methods mainly rely on expert experience and expertise, which have limited knowledge coverage and are difficult to expand and integrate. In recent years, artificial intelligence and deep learning methods have been gradually applied to power system fault diagnosis and auxiliary decision-making, achieving some success. However, general large-scale models still face the following problems when applied in specialized fields such as relay protection:

[0004] 1. Large model illusion leads to irrelevant answers: Large models are mostly trained on general corpora, lacking in-depth coverage of professional knowledge such as relay protection. In the absence of reliable external knowledge support, large models are prone to "illusions," generating answers that do not conform to facts or professional standards, affecting their reliability and practicality in critical power system scenarios.

[0005] 2. Temporal ambiguity in knowledge modeling leads to indexing errors: In real-world scenarios, knowledge evolves over time. The same entity may have different attributes (such as fault type, fault node, etc.) at different points in time. Existing GraphRAG methods often ignore temporal features when constructing knowledge graphs or generating indexable embeddings, resulting in facts from specific times being merged under the same entity node. This lack of clear temporal separation leads to errors in knowledge indexing and retrieval.

[0006] 3. Low inference efficiency and poor accuracy: Existing methods typically rely on a single agent to extract information using a fixed pattern, lacking the ability to adaptively model multi-level information in graph data. When processing graph structures, a uniform information granularity is often used, capturing only local text or simple topological relationships, failing to effectively integrate multi-dimensional information such as node attributes, edge structures, and global features. This static aggregation mechanism limits the model's ability to express complex graph structures, leading to the loss of semantic information and reducing inference efficiency and accuracy.

[0007] Therefore, there is an urgent need for a power outage protection auxiliary decision-making method that can integrate time-series knowledge, support multi-level adaptive retrieval, and have self-correction capabilities, in order to improve the accuracy, interpretability, and real-time performance of fault diagnosis and decision-making. Summary of the Invention

[0008] The purpose of this invention is to address the problems of illusion in the application of large models in the field of relay protection, indexing errors caused by temporal ambiguity in knowledge modeling, and low reasoning efficiency and insufficient accuracy of traditional retrieval-enhanced generation methods. This invention provides a power outage protection auxiliary decision-making method and system based on multi-agent collaborative adaptive graph structure retrieval-enhanced generation technology.

[0009] The present invention discloses a power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology, which includes:

[0010] Step 1: Based on the temporal source text power outage protection database, extract entities and relationships between entities in the power outage protection field, and label the entities and relationships with time attributes to construct a power outage protection knowledge graph;

[0011] Step 2: Receive the query and use the large language model to perform temporal semantic parsing on the query. When the query contains time constraints, decompose the query into sub-queries corresponding to the time attributes in the power outage protection knowledge graph.

[0012] Step 3: For each subquery, based on its corresponding time attribute, retrieve the corresponding temporal subgraph from the power outage protection knowledge graph; use the temporal adaptive three-layer interactive retrieval module to extract the descriptive knowledge and entity associations related to the subquery from the temporal subgraph, and determine the entity containing the most relevant descriptive knowledge as the valid node;

[0013] Step 4: Based on the effective nodes and their association structure in the power outage protection knowledge graph, retrieve the associated effective source texts from the temporal source text power outage protection database;

[0014] Step 5: The descriptive knowledge and entity associations extracted based on the temporal subgraph are used as valid graph information, and the retrieved valid source text is used as valid text information. These are then input into the large language model along with the query to generate a power outage protection auxiliary decision answer.

[0015] Preferably, the specific method for constructing the power outage protection knowledge graph in step 1 includes:

[0016] Step 1-1: Divide the temporal source text knowledge base into multiple time periods, and divide the text within each time period into text blocks of fixed size;

[0017] Temporal source text knowledge base Divided into: ,in, express A subset of text within a time period;

[0018] The text blocks are divided as follows: ;in, Indicates the time period The inner division of the first Each text processing unit serves as the basic unit for information extraction.

[0019] Steps 1-2: Use a large language model to identify entities, their types, and descriptions in each text block, and identify the relationships between these entities, their types, and descriptions; define these descriptions as descriptive knowledge.

[0020] The entity is: The expression is:

[0021] ;

[0022] in, Representing the time in a knowledge graph The next One entity, Represents a set of entities. express Time attribute, Indicates time Time Entity The description of knowledge; Representing entities The time corresponding to knowledge at a specific moment. Represents the collection of time attributes of an entity;

[0023] Steps 1-3: Label the identified entities and relationships with time attributes to distinguish the knowledge of the same entity at different points in time, thereby constructing the power outage protection knowledge graph. .

[0024] Preferably, in step 2, when the query does not contain time constraints, the query remains unchanged, and a retrieval is performed based on all the knowledge in the power outage protection knowledge graph.

[0025] Preferably, the specific method for retrieving the corresponding temporal subgraph from the power outage protection knowledge graph for each subquery in step 3 includes:

[0026] Using a pre-trained embedding model, the knowledge contained in the nodes of the temporal subgraph is encoded to generate a node embedding library;

[0027] Specifically:

[0028] Construct a subgraph retrieval tool for power outage protection For a timestamp of Subtemporal query From the knowledge graph of power outage protection Retrieval temporal subgraph The nodes and relationships in the power outage protection subgraph only include time points. The knowledge contained therein;

[0029] Encode all knowledge in the nodes using a pre-trained embedding model to generate Node embedding library :

[0030]

[0031] in, Represents the embedding vector of a node. Represents an embedded model. This indicates an entity with a tense marker. express A real vector space of dimension 1.

[0032] Preferably, the specific method for extracting descriptive knowledge and entity relationships related to the subquery from the temporal subgraph and determining valid nodes in step 3 includes:

[0033] Candidate nodes are retrieved using a node retrieval tool based on the similarity between the node embedding library and the query embedding.

[0034] The knowledge retrieval tool extracts the most relevant descriptive knowledge from candidate nodes as effective knowledge.

[0035] Nodes containing the aforementioned valid knowledge are identified as valid nodes.

[0036] Preferably, the retrieval process of the node retrieval device and the knowledge retrieval device is collaboratively controlled by the adaptive graph information extraction module; the adaptive graph information extraction module includes a planning agent, a thinking agent, and an execution agent, wherein:

[0037] The planning agent is used to establish reasoning paths;

[0038] The thinking agent is used to identify the type of graph information required for the current reasoning step;

[0039] The executing agent is used to dynamically perform retrieval operations based on the reasoning path and information type.

[0040] Preferably, the specific method for retrieving the associated valid source text from the temporal source text power failure protection database in step 4 includes:

[0041] Based on the connectivity of the effective nodes and their adjacent nodes in the knowledge graph, the candidate source text containing the effective nodes is scored, where the more 1-hop adjacent nodes a text block contains, the higher the score.

[0042] Based on the scores, the top scorers will be selected. Each source text is considered as the valid source text. ,in, This indicates the maximum number of source texts returned by the retrieval.

[0043] Preferably, the specific method for generating the power outage protection auxiliary decision answer in step 5 includes:

[0044] The valid graph information, the valid text information, and the query are input into the large language model to generate the answer corresponding to the query:

[0045] ;

[0046] in, Indicates time The above is the answer generated for the subquery. This indicates that after temporal decomposition, the time corresponding to... Subqueries; Represents the generation function of a large language model;

[0047] The effective graph information includes several relevant knowledge extracted from the time-adaptive three-layer interactive retrieval module. and the effective node Related relational databases The valid text information includes several valid source texts retrieved from the temporal source text power failure protection database. .

[0048] Preferably, step 5 further includes:

[0049] When the query is decomposed into multiple subqueries, the answer corresponding to each subquery is input into the large language model along with the original query for comprehensive processing to generate the final power outage protection auxiliary decision answer:

[0050] ;

[0051] in, This indicates the original query. The generated final power outage protection auxiliary decision answer.

[0052] The present invention discloses a power outage protection auxiliary decision-making system based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology, which includes:

[0053] The knowledge graph construction module is used to build a power outage protection knowledge graph based on temporal source text;

[0054] The temporal query parsing module is used to perform temporal parsing and decomposition on the input query;

[0055] The time-adaptive three-layer interactive retrieval module is used to adaptively retrieve relevant graph information from the power outage protection knowledge graph based on the parsed query.

[0056] The source text retrieval module is used to supplement relevant source text based on the retrieval results;

[0057] The answer generation module is used to integrate the graph information with the source text to generate an answer that assists in decision-making.

[0058] Advantages of the present invention: The power outage protection auxiliary decision-making method and system based on multi-agent collaborative adaptive graph structure retrieval enhancement generation technology described in this invention constructs a temporal knowledge graph, realizes multi-agent collaborative adaptive retrieval, and integrates multi-source information for enhanced generation, forming a complete, accurate and reliable intelligent decision-making scheme. It effectively solves the problems of illusion, temporal confusion and rigid reasoning in professional fields of large models, and significantly improves the accuracy, interpretability and real-time performance of fault diagnosis and decision-making in power outage protection scenarios.

[0059] Specifically, the present invention has the following three outstanding advantages:

[0060] 1. Significantly enhanced temporal awareness, eliminating retrieval errors caused by time confusion: By constructing a knowledge graph with time attributes and performing temporal semantic parsing and decomposition on queries, the system can accurately distinguish entity states and knowledge at different points in time, achieving reliable answers to time-sensitive queries. This mechanism fundamentally avoids knowledge retrieval errors caused by confusion of time information such as years and months, and is particularly suitable for business scenarios with strong time correlations, such as power grid fault history analysis and protection setting change tracing.

[0061] 2. A multi-layer adaptive retrieval mechanism significantly improves information extraction accuracy and reasoning efficiency: Employing a time-series adaptive three-layer interactive retrieval module, the system combines node retrieval with knowledge retrieval and introduces a multi-agent collaborative strategy involving planning, thinking, and execution. The system can dynamically adjust the retrieval granularity and reasoning path based on query complexity. This design effectively filters redundant and noisy information, significantly improving the recall and accuracy of relevant knowledge, while avoiding the "over-reasoning" or "under-reasoning" problems caused by fixed retrieval patterns in traditional methods. It achieves a balance between efficiency and accuracy on complex graph structures.

[0062] 3. Multi-source information fusion and self-reflection mechanisms ensure answer reliability and system robustness: By fusing structured descriptive knowledge extracted from the knowledge graph, entity relationships, and semantic fragments retrieved from the original text, the system provides rich and mutually corroborating context for the large language model, significantly suppressing model illusion. The further introduced multi-perspective self-reflection module performs consistency checks and error corrections on the reasoning process, enabling the system to continuously optimize and greatly improving the reliability, logical consistency, and overall system robustness of the output answer. Attached Figure Description

[0063] Figure 1 This is an overall flowchart of a power outage protection auxiliary decision-making system based on multi-agent collaborative adaptive graph structure retrieval enhancement proposed in this invention;

[0064] Figure 2 This is a result evaluation graph of a power outage protection auxiliary decision-making method based on multi-agent collaborative adaptive graph structure retrieval enhancement proposed in this invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0067] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0068] Example 1:

[0069] The following is combined with Figure 1 and Figure 2 This embodiment describes a power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology, which includes:

[0070] Step 1: Based on the temporal source text power outage protection database, extract entities and relationships between entities in the power outage protection field, and label the entities and relationships with time attributes to construct a power outage protection knowledge graph;

[0071] Step 2: Receive the query and use the large language model to perform temporal semantic parsing on the query. When the query contains time constraints, decompose the query into sub-queries corresponding to the time attributes in the power outage protection knowledge graph.

[0072] Step 3: For each subquery, based on its corresponding time attribute, retrieve the corresponding temporal subgraph from the power outage protection knowledge graph; use the temporal adaptive three-layer interactive retrieval module to extract the descriptive knowledge and entity associations related to the subquery from the temporal subgraph, and determine the entity containing the most relevant descriptive knowledge as the valid node;

[0073] Step 4: Based on the effective nodes and their association structure in the power outage protection knowledge graph, retrieve the associated effective source texts from the temporal source text power outage protection database.

[0074] Step 5: The descriptive knowledge and entity associations extracted based on the temporal subgraph are used as valid graph information, and the retrieved valid source text is used as valid text information. These are then input into the large language model along with the query to generate a power outage protection auxiliary decision answer.

[0075] Furthermore, the specific method for constructing the power outage protection knowledge graph in step 1 includes:

[0076] Step 1-1: Divide the temporal source text knowledge base into multiple time periods, and divide the text within each time period into text blocks of fixed size;

[0077] Temporal source text knowledge base Divided into: ,in, express A subset of text within a time period;

[0078] The text blocks are divided as follows: ;in, Indicates the time period The inner division of the first Each text processing unit serves as the basic unit for information extraction.

[0079] Steps 1-2: Use a large language model to identify entities, their types, and descriptions in each text block, and identify the relationships between these entities, their types, and descriptions; define these descriptions as descriptive knowledge.

[0080] The entity is: The expression is:

[0081] ;

[0082] in, Representing the time in a knowledge graph The next One entity, Represents a set of entities. express Time attribute, Indicates time Time Entity The description of knowledge; Representing entities The time corresponding to knowledge at a specific moment. Represents the collection of time attributes of an entity;

[0083] Steps 1-3: Label the identified entities and relationships with time attributes to distinguish the knowledge of the same entity at different points in time, thereby constructing the power outage protection knowledge graph. .

[0084] Furthermore, in step 2, when the query does not contain time constraints, the query remains unchanged, and a retrieval is performed based on all the knowledge in the power outage protection knowledge graph.

[0085] Furthermore, the specific method for retrieving the corresponding temporal subgraph from the power outage protection knowledge graph for each subquery in step 3 includes:

[0086] Using a pre-trained embedding model, the knowledge contained in the nodes of the temporal subgraph is encoded to generate a node embedding library;

[0087] Specifically:

[0088] Construct a subgraph retrieval tool for power outage protection For a timestamp of Subtemporal query From the knowledge graph of power outage protection Retrieval temporal subgraph The nodes and relationships in the power outage protection subgraph only include time points. The knowledge contained therein;

[0089] Encode all knowledge in the nodes using a pre-trained embedding model to generate Node embedding library :

[0090]

[0091] in, Represents the embedding vector of a node. Represents an embedded model. This indicates an entity with a tense marker. express A real vector space of dimension 1.

[0092] Furthermore, the specific method for extracting descriptive knowledge and entity relationships related to the subquery from the temporal subgraph and determining valid nodes in step 3 includes:

[0093] Candidate nodes are retrieved using a node retrieval tool based on the similarity between the node embedding library and the query embedding.

[0094] The knowledge retrieval tool extracts the most relevant descriptive knowledge from candidate nodes as effective knowledge.

[0095] Nodes containing the aforementioned valid knowledge are identified as valid nodes.

[0096] Furthermore, the retrieval process of the node retrieval device and the knowledge retrieval device is collaboratively controlled by the adaptive graph information extraction module; the adaptive graph information extraction module includes a planning agent, a thinking agent, and an execution agent, wherein:

[0097] The planning agent is used to establish reasoning paths;

[0098] The thinking agent is used to identify the type of graph information required for the current reasoning step;

[0099] The executing agent is used to dynamically perform retrieval operations based on the reasoning path and information type.

[0100] Furthermore, the specific method for retrieving relevant valid source text from the temporal source text power failure protection database in step 4 includes:

[0101] Based on the connectivity of the effective nodes and their adjacent nodes in the knowledge graph, the candidate source text containing the effective nodes is scored, where the more 1-hop adjacent nodes a text block contains, the higher the score.

[0102] Based on the scores, the top scorers will be selected. Each source text is considered as the valid source text. ,in, This indicates the maximum number of source texts returned by the retrieval.

[0103] Furthermore, the specific method for generating the power outage protection auxiliary decision answer in step 5 includes:

[0104] The valid graph information, the valid text information, and the query are input into the large language model to generate the answer corresponding to the query:

[0105] ;

[0106] in, Indicates time The above is the answer generated for the subquery. This indicates that after temporal decomposition, the time corresponding to... Subqueries; Represents the generation function of a large language model;

[0107] The effective graph information includes several relevant knowledge extracted from the time-adaptive three-layer interactive retrieval module. and the effective node Related relational databases The valid text information includes several valid source texts retrieved from the temporal source text power failure protection database. .

[0108] Furthermore, step 5 also includes:

[0109] When the query is decomposed into multiple subqueries, the answer corresponding to each subquery is input into the large language model along with the original query for comprehensive processing to generate the final power outage protection auxiliary decision answer:

[0110] ;

[0111] in, This indicates the original query. The generated final power outage protection auxiliary decision answer.

[0112] Example 2:

[0113] The following is combined Figure 1 and Figure 2 This embodiment describes a power outage protection auxiliary decision-making system based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology, which includes:

[0114] The knowledge graph construction module is used to build a power outage protection knowledge graph based on temporal source text;

[0115] The temporal query parsing module is used to perform temporal parsing and decomposition on the input query;

[0116] The time-adaptive three-layer interactive retrieval module is used to adaptively retrieve relevant graph information from the power outage protection knowledge graph based on the parsed query.

[0117] The source text retrieval module is used to supplement relevant source text based on the retrieval results;

[0118] The answer generation module is used to integrate the graph information with the source text to generate an answer that assists in decision-making.

[0119] In this invention, such as Figure 1The diagram shows the overall flowchart of a power outage protection auxiliary decision-making system based on multi-agent collaborative adaptive graph structure retrieval enhancement proposed in this invention. Using a single-time question-and-answer example, the diagram illustrates the entire process of the proposed method: constructing a knowledge graph from temporal source text, receiving and parsing queries, extracting relevant knowledge and effective nodes through a three-layer interactive retrieval module with temporal adaptive retrieval, supplementing the retrieval of effective source text, and finally generating the auxiliary decision-making answer using a large language model. First, a large number of documents / images / source texts in the power outage protection dataset are converted into a knowledge graph (TG) and a searchable text index. For queries with time attributes (such as 2016), the time dimension is extracted and temporal decomposition retrieval is performed. Effective nodes and documents (Top-n, Top-t, Top-k) are retrieved in the graph through three types of relationships: nodes, time, and knowledge. Then, "effective graph information," "effective knowledge," and "effective source text" are extracted. Finally, the answer is synthesized by a large model. The system includes adaptive information extraction and self-reflection modules for online error correction and iterative improvement. This design combines complex graph structure learning with simple classification, achieving both high efficiency and high accuracy.

[0120] The method specifically includes the following steps:

[0121] Step 1: Abstract the information from the temporal source text power outage protection database into a power outage protection knowledge graph. ,in For entities in a text dataset, Define the relationships between entities; and obtain the entities based on this graph. expression;

[0122] Step 2: Construct a power outage protection temporal query engine and use the powerful temporal semantic awareness of the large language model to analyze, reason, and decompose the temporal constraints in the query.

[0123] Step 3: Design a time-adaptive three-layer interactive retrieval module to extract knowledge and relationships from the power outage protection knowledge graph and select the most relevant knowledge and relationships for the query.

[0124] Step 4: Extract the most relevant and valid source text from the temporal source text power failure protection database to supplement the completeness of knowledge;

[0125] Step 5: Enhance the large model to generate answers based on the relevant knowledge, relationships, and most relevant valid source texts from the power outage protection knowledge graph.

[0126] In step one, a knowledge graph for power outage protection is constructed. The specific process is as follows:

[0127] Step A1: Accumulate a power outage protection dataset by collecting texts such as books, documents, and power outage technical reports, and define it as a temporal source text knowledge base. Divide it into multiple time periods This facilitates temporal evolution modeling. Then, the text within each time period is divided into text blocks of fixed size. , as a processing unit.

[0128] Step A2: Extract graph information using a large language model. First, identify all entities in the text block, including their names, types, and descriptions. Then, identify the relationships between related entities, clarifying the types and descriptions of relationships between source and target entities. Since the semantic information of a knowledge graph mainly exists in the descriptions of entities and relationships, these descriptions are defined as knowledge. Simultaneously, a time attribute is introduced into entities and relationships, and the corresponding timestamps are associated with the knowledge to distinguish knowledge of the same entity at different points in time. During the graph update process, knowledge with the latest timestamp is continuously added to entities and relationships, ultimately constructing a power outage protection knowledge graph. .

[0129] Step A3: Based on the aforementioned power outage protection knowledge graph Define entity Its expression is:

[0130]

[0131] in express Time attribute, Indicates time Time Entity The knowledge above.

[0132] The key to this method's ability to analyze text data across different time periods lies in constructing a graph based on time-stamped knowledge rather than simple entity relationships. Text entities are abstracted as graph nodes, and relationships between entities are abstracted as edges. Weights are assigned to edges based on contextual associations, thus constructing a weighted graph. The final time... Time Entity and entity The knowledge provided here serves as a standard for subsequent knowledge retrieval.

[0133] In step two, a power outage protection temporal query device is constructed, given the original query. First, a large language model is used to identify and parse its temporal constraints. If the query does not contain time constraints, the original query remains unchanged, which means that the problem requires traversing knowledge across all time periods.

[0134] If the query includes multiple time constraints, the reasoning power of the large language model can be used to decompose the time into multiple independent subqueries. Each subquery corresponds to a single point in time.

[0135] In this implementation, the original queries are categorized based on whether they have time constraints, and subsequent queries are performed for different time constraints. This ensures that all original queries can complete the retrieval and generation within their time requirements, which helps improve the model's accuracy and final performance.

[0136] The specific process of the time-adaptive three-layer interactive retrieval module in step three is as follows:

[0137] Step B1: Construct a power outage protection subgraph retrieval tool ( For a timestamp of Subtemporal query From the knowledge graph of power outage protection Retrieval temporal subgraph The nodes and relationships in the power outage protection subgraph only include time points. The knowledge contained within. A pre-trained embedding model is used to generate node embeddings; specifically, we encode all the knowledge within the nodes to generate... Node embedding library Its expression is:

[0138]

[0139] in, Indicates the dimension of the output vector.

[0140] Step B2: Build a node retriever With knowledge retrieval The combined interactive approach, firstly, for subtemporal queries The same embedding strategy is applied to ensure consistency in text information processing. Subsequently, to identify candidate nodes relevant to the query, a nearest neighbor search method is used, where cosine similarity is used to measure the similarity between the query embedding and the node embedding. The most relevant nodes were selected as candidate nodes. The search is defined as follows:

[0141]

[0142] Based on this similarity, before retrieval Each element provides a set of elements considered most relevant to the query. Simultaneously, an adaptive graph information extraction module is used to hierarchically define paths using three types of collaborative agents, clarify the required graph information, and perform flexible graph information extraction.

[0143] Next, valid knowledge will be extracted from the candidate nodes. Specifically, for each piece of knowledge in the candidate nodes, an embedding will be generated separately using the same embedding strategy.

[0144]

[0145] Then, cosine similarity is used to measure the distance between the query and the knowledge embedding, and the top [query / knowledge embedding] is selected. The most similar knowledge blocks are considered as valid knowledge. .

[0146]

[0147] Finally, include knowledge The node is considered a valid node. .

[0148] The complete process of the adaptive graph information extraction module involves the coordinated operation of three agents. First, the planning agent establishes a structured reasoning path by merging query and contextual information, ensuring that the reasoning process follows a coherent logical order rather than arbitrary retrieval. Next, the thinking agent refines the scope of information extraction by identifying the specific graph-related knowledge required for the current reasoning step, avoiding unnecessary retrieval from the entire graph. Finally, the execution agent dynamically adjusts the retrieval strategy according to the reasoning steps, ensuring that the extracted knowledge maintains structural integrity and interdependence, and iterates within this framework, enabling the large language model to more effectively utilize complex graph structures.

[0149] Step B3: A multi-perspective self-reflection module is introduced. After the adaptive graph information extraction module generates the answer, the self-reflection module evaluates the logical consistency between the reasoning path and the final response, identifying potential errors or biases. If discrepancies are detected, the self-reflection module summarizes key reasoning points and records error patterns. Furthermore, the self-reflection module combines reverse reasoning and multi-perspective analysis to improve the adaptive graph information extraction module's understanding of queries and contextual information.

[0150] In this implementation scheme, the time-adaptive three-layer interactive retrieval module is the core innovation of this invention. Unlike traditional GraphRAG methods that use a single-channel graph structure retrieval approach, i.e., relying on node embedding or relational path similarity calculation to achieve relevant knowledge retrieval, this invention uses a multi-level interactive retrieval mechanism to refine the granularity of the retrieval content, helping to reduce interference from redundant node information. Simultaneously, the adaptive graph information extraction module uses three types of collaborative agents to hierarchically define paths, clarify the required graph information, and perform flexible graph information extraction. Even in non-temporal scenarios, It also improves the model's ability to extract relevant evidence from complex knowledge graphs. Furthermore, the self-reflection module can identify retrieval errors and improve the understanding of contextual information, avoiding the illusion problem caused by reasoning bias.

[0151] The method in step four is as follows:

[0152] Candidate source texts are scored based on the connectivity of valid nodes and their neighbors. A text block receives a higher score if it contains more 1-hop neighbors. Finally, the top t source texts with the highest scores are selected as the final valid source texts. .

[0153] In this implementation scheme, the reliability of the generated content can be increased by re-evaluating the source text and supplementing it into the subsequent generation process, ensuring that the large language model can access the complete semantic context.

[0154] The method in step five is specifically as follows:

[0155] First, query answers are generated. Valid graph information, valid text information, and the query input are then fed into a large model for enhancement. Valid graph information includes top-k relevant knowledge extracted from a three-layer interactive knowledge retrieval system. and with valid nodes Related relational databases This enables the large language model to understand the topological information of the retrieved subgraph; at the same time, the effective text information base includes top-t relevant source texts extracted by the effective source text retriever. This ensures that the large language model has access to the complete semantic context. The generator enhances the knowledge between these two perspectives to produce accurate answers to queries. .

[0156]

[0157] Ultimately, the original query extends along all sub-temporal queries. and their corresponding answers It is fed into a large language model to generate a comprehensive final answer. .

[0158]

[0159] In this implementation, for complex queries with multiple time constraints, the system hierarchically aggregates the subquery results from different time points. All sub-temporal queries and their corresponding intermediate answers are sequentially input into the large language model, and the final comprehensive answer is generated through multi-perspective semantic fusion and consistency constraints. This design effectively integrates knowledge across time dimensions, improving logical consistency and information completeness in temporal question answering tasks.

[0160] To illustrate the advantages of the proposed method, the multi-agent collaborative adaptive exploration graph structure retrieval enhancement generation method described in this invention is experimentally compared with LLM, GraphRAG, and Vanilla RAG. For all RAG-based methods, stella-en-1.5B-v5 is used as the embedding model for dense retrieval. The retrieval strategy for the source text is parameterized by the product of the block size and the top-t. In the main experiments, the block size is set to 1000, and the top-t is set to 5. For GraphRAG and the multi-agent collaborative adaptive exploration graph structure retrieval enhancement generation method, the total number of retrieved context tokens is limited to 1600 tokens. Specifically, for the multi-agent collaborative adaptive exploration graph structure retrieval enhancement generation method, the number of candidate nodes (n) is set to 30, and the top-k most relevant knowledge units are selected based on the graph retrieval score, where k=15.

[0161] For the power outage protection dataset, a time-evolutionary long text dataset was constructed to ensure that most of the data is invisible to LLM (Local Power Management System). The basic corpus consisted of publicly available reports on Jiangsu power generation enterprises' technical supervision during peak summer demand, books and papers related to circuit fault detection and location methods, and distribution network protection and restoration strategies from 2016 to 2024. Using the MinerU tool, the source text was converted from PDF to Markdown format, retaining only the text content. Subsequently, an automated time-constrained problem-answer pair construction process was designed, including single-time-constrained, dual-time-constrained, multi-time-constrained, and unconstrained problems, to test the method.

[0162] For evaluation metrics, a language model-based metric is used to assess the quality of the model's responses, and Qwen2.5-72B is used to evaluate their consistency with the true answers. Each sample is evaluated independently by the model three times, and a majority voting strategy is applied: if the model considers the answer correct in at least two of the three evaluations, the answer is considered correct.

[0163] Reference Figure 1 The figure illustrates the entire process of question answering using this method, with the final large model score measured by the accuracy of the answers. The results are shown in Table 1.

[0164] Table 1. Model evaluation results on the power outage protection dataset.

[0165] like Figure 2 The graph shown is a result evaluation chart of the method of this invention, illustrating a performance comparison of LLM, Vanilla RAG, Graph RAG, and the method of this invention (this method) under four different conditions: single-time question answering, dual-time question answering, multi-time question answering, and no-time question answering. Figure 2 The overall performance results were compared and analyzed from multiple dimensions. Among all LLM backbones, our method significantly outperformed all baselines. In particular, our method showed a more significant improvement under complex temporal constraints. In the table above, our method's score was 13.71 points higher than GraphRAG on average, 9.74 points higher than Vanilla RAG on average, and 63.66 points higher than LLM on average. This is because complex temporal constraints often introduce more temporal redundancy, reducing the accuracy of node retrieval. The experimental results robustly demonstrate the effectiveness of our method in reducing temporal retrieval errors. Furthermore, for question answering without specific time constraints, our method also outperformed Vanilla RAG and LLM, further validating the ability of its three-layer interactive retrieval framework to improve the performance of non-basic LLMs, which was significantly worse. This indicates that the necessary knowledge to answer these questions is not inherently contained in LLMs. The accuracy of the response depends primarily on the ability to effectively retrieve relevant text.

[0166] These comparative results clearly demonstrate that the multi-agent collaborative adaptive exploration graph structure retrieval enhancement generation method proposed in this invention has stronger feature extraction capabilities and higher accuracy compared to existing technologies, verifying the advanced nature and effectiveness of the proposed solution.

[0167] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology, characterized in that, It includes: Step 1: Based on the temporal source text power outage protection database, extract entities and relationships between entities in the power outage protection field, and label the entities and relationships with time attributes to construct a power outage protection knowledge graph; Step 2: Receive the query and use the large language model to perform temporal semantic parsing on the query. When the query contains time constraints, decompose the query into sub-queries corresponding to the time attributes in the power outage protection knowledge graph. Step 3: For each subquery, based on its corresponding time attribute, retrieve the corresponding temporal subgraph from the power outage protection knowledge graph; use the temporal adaptive three-layer interactive retrieval module to extract the descriptive knowledge and entity associations related to the subquery from the temporal subgraph, and determine the entity containing the most relevant descriptive knowledge as the valid node; Step 4: Based on the effective nodes and their association structure in the power outage protection knowledge graph, retrieve the associated effective source texts from the temporal source text power outage protection database; Step 5: The descriptive knowledge and entity associations extracted based on the temporal subgraph are used as valid graph information, and the retrieved valid source text is used as valid text information. These are then input into the large language model along with the query to generate a power outage protection auxiliary decision answer.

2. The power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology according to claim 1, characterized in that, The specific methods for constructing the power outage protection knowledge graph in step 1 include: Step 1-1: Divide the temporal source text knowledge base into multiple time periods, and divide the text within each time period into text blocks of fixed size; Temporal source text knowledge base Divided into: ,in, express A subset of text within a time period; The text blocks are divided as follows: ;in, Indicates the time period The inner division of the first Each text processing unit serves as the basic unit for information extraction. Steps 1-2: Use a large language model to identify entities, their types, and descriptions in each text block, and identify the relationships between these entities, their types, and descriptions; define these descriptions as descriptive knowledge. The entity is: The expression is: ； in, Representing the time in a knowledge graph The next One entity, Represents a set of entities. express Time attribute, Indicates time Time Entity The description of knowledge; Representing entities The time corresponding to knowledge at a specific moment. Represents the collection of time attributes of an entity; Steps 1-3: Label the identified entities and relationships with time attributes to distinguish the knowledge of the same entity at different points in time, thereby constructing the power outage protection knowledge graph. .

3. The power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology according to claim 1, characterized in that, In step 2, when the query does not contain time constraints, the query remains unchanged, and a retrieval is performed based on all the knowledge in the power outage protection knowledge graph.

4. The power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology according to claim 1, characterized in that, The specific method for retrieving the corresponding temporal subgraph from the power outage protection knowledge graph for each subquery in step 3 includes: Using a pre-trained embedding model, the knowledge contained in the nodes of the temporal subgraph is encoded to generate a node embedding library; Specifically: Construct a subgraph retrieval tool for power outage protection For a timestamp of Subtemporal query From the knowledge graph of power outage protection Retrieval temporal subgraph The nodes and relationships in the power outage protection subgraph only include time points. The knowledge contained therein; Encode all knowledge in the nodes using a pre-trained embedding model to generate Node embedding library : in, Represents the embedding vector of a node. Represents an embedded model. This indicates an entity with a tense marker. express A real vector space of dimension 1.

5. The power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology according to claim 1, characterized in that, The specific method for extracting descriptive knowledge and entity relationships related to the subquery from the temporal subgraph and determining valid nodes in step 3 includes: Candidate nodes are retrieved using a node retrieval tool based on the similarity between the node embedding library and the query embedding. The knowledge retrieval tool extracts the most relevant descriptive knowledge from candidate nodes as effective knowledge. Nodes containing the aforementioned valid knowledge are identified as valid nodes.

6. The power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology according to claim 5, characterized in that, The retrieval process of the node retrieval device and the knowledge retrieval device is collaboratively controlled by the adaptive graph information extraction module; the adaptive graph information extraction module includes a planning agent, a thinking agent, and an execution agent, wherein: The planning agent is used to establish reasoning paths; The thinking agent is used to identify the type of graph information required for the current reasoning step; The executing agent is used to dynamically perform retrieval operations based on the reasoning path and information type.

7. The power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology according to claim 1, characterized in that, The specific method for retrieving relevant valid source text from the temporal source text power failure protection database in step 4 includes: Based on the connectivity of the effective nodes and their adjacent nodes in the knowledge graph, the candidate source text containing the effective nodes is scored, where the more 1-hop adjacent nodes a text block contains, the higher the score. Based on the scores, the top scorers will be selected. Each source text is considered as the valid source text. ,in, This indicates the maximum number of source texts returned by the retrieval.

8. The power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology according to claim 1, characterized in that, The specific methods for generating the power outage protection auxiliary decision answer in step 5 include: The valid graph information, the valid text information, and the query are input into the large language model to generate the answer corresponding to the query: ； in, Indicates time The above is the answer generated for the subquery. This indicates that after temporal decomposition, the time corresponding to... Subqueries; Represents the generation function of a large language model; The effective graph information includes several relevant knowledge extracted from the time-adaptive three-layer interactive retrieval module. and the effective node Related relational databases The valid text information includes several valid source texts retrieved from the temporal source text power failure protection database. .

9. The power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology according to claim 8, characterized in that, Step 5 further includes: When the query is decomposed into multiple subqueries, the answer corresponding to each subquery is input into the large language model along with the original query for comprehensive processing to generate the final power outage protection auxiliary decision answer: ； in, This indicates the original query. The generated final power outage protection auxiliary decision answer.

10. A system for implementing the power outage protection auxiliary decision-making method based on multi-agent cooperative adaptive graph structure retrieval enhancement generation technology as described in any one of claims 1-9, characterized in that, It includes: The knowledge graph construction module is used to build a power outage protection knowledge graph based on temporal source text; The temporal query parsing module is used to perform temporal parsing and decomposition on the input query; The time-adaptive three-layer interactive retrieval module is used to adaptively retrieve relevant graph information from the power outage protection knowledge graph based on the parsed query. The source text retrieval module is used to supplement relevant source text based on the retrieval results; The answer generation module is used to integrate the graph information with the source text to generate an answer that assists in decision-making.

Images