A knowledge graph expansion method and system for large-scale model scheduling multi-agent power grid fault response strategy

Abstract

A method and system for expanding a knowledge graph for multi-agent power grid fault response strategies for large-scale scheduling, comprising: modeling task dispatch as a mixed integer programming problem and solving it using a relaxation-correction algorithm; extracting and normalizing core knowledge tasks and outputting structured data; summarizing the resulting text fragments to prompt a large language model to retain key entities, relations, and domain-specific terms; using LLM-based named entity recognition technology, combined with prompts and dictionary / ontology filtering in the power scheduling field, to identify relevant entities in the text and normalize them into a standard form in the knowledge graph; detecting logical contradictions between newly extracted triples and existing relations in the knowledge graph, and using LLM-based debate prompts to classify and resolve conflicts; and aggregating multiple verification signals to calculate the global confidence, clarity, and relevance scores for each triple.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, and specifically relates to a method for expanding a knowledge graph of a large-model dispatching multi-agent power grid fault response strategy, and also relates to a knowledge graph expansion system for a large-model dispatching multi-agent power grid fault response strategy. Background Art

[0002] In recent years, against the backdrop of frequent natural disasters caused by climate change, with the increasing scale of power grids, the access of large-scale intermittent renewable energy, and the rapid advancement of new power system construction, the uncertainty of power grid operation has increased, and the types and number of dispatch objects have grown exponentially. The current power grid dispatch plan based on physical models has problems such as slow optimization decision-making calculation speed, long time consumption, and insufficient adaptability to multiple uncertain scenarios. In particular, in the intraday stage, it still often relies on manual control by dispatchers.

[0003] The new power system, based on ultra-high voltage AC / DC hybrid grids, operates in increasingly complex and variable modes, exhibiting a high degree of uncertainty. Dispatching operations at all levels face the risk of unknown failures at any given moment during real-time operation, posing a significant threat to grid security, supply assurance, and the integration of renewable energy. Grid fault response plans are a crucial mechanism for ensuring proactive and safe grid operation. However, the current technical support for fault response plan preparation and management, which relies heavily on the accumulated experience of control personnel, is unable to meet the requirements for safe production and stable operation of the new power system. They also lack proactive and effective control decision-making and comprehensive risk prevention and control measures in scenarios such as severe source and load fluctuations, equipment maintenance, and grid failures.

[0004] The increasing time-variability and complexity of power grid operation modes have made traditional dispatch decision-making mechanisms unsustainable, becoming a pressing issue for complex power grid operations both domestically and internationally. To ensure the safe and economical operation of power systems, a classic two-step dispatch decision-making mechanism has gradually emerged. Operational experts manually select typical operating modes, analyze grid safety and stability through offline analysis programs, and manually formulate operating rules, ultimately creating an "operation manual." Dispatchers then use these operating rules as a safety boundary for dispatching and control, aiming to achieve the optimal operating point within that safety boundary.

[0005] Maintaining domain knowledge graphs is even more difficult in the field of power dispatching systems, which have complex and specialized terminology. Traditional knowledge graph construction methods include manual annotation management, which, while reliable, is difficult to maintain on large-scale data. Automated construction methods based on traditional natural language processing (NLP) often have difficulty handling domain-specific terminology and context-dependent relationships in scientific and technological literature. In addition, extracting and integrating knowledge into existing knowledge bases requires robust mechanisms for pattern alignment, consistency, and conflict resolution. In high-risk applications, the cost of inaccuracies in these systems can be very high. Traditional solutions that rely on expert rule bases to identify faults and rely on dispatcher experience to manually generate fault response strategies are difficult to cope with the challenges brought about by factors such as the large-scale grid integration of new energy sources and complex grid environments in new power systems.

[0006] With recent advances in large language models enabling significant progress in contextual understanding and reasoning, the research community is increasingly exploring multi-agent systems, where several specialized agents work collaboratively to solve complex tasks. These systems leverage the strengths of individual agents, each optimized for a specific subtask, and support cross-agent validation and iterative improvement of outputs. This multi-agent framework shows promise in areas ranging from decision making to structured data extraction, providing robustness through redundancy and collaboration. Summary of the Invention

[0007] In order to address the deficiencies in the prior art, the present invention provides a method and system for expanding the knowledge graph of large-model scheduling multi-agent power grid fault response strategies.

[0008] The present invention adopts the following technical solutions.

[0009] The first aspect of the present invention relates to a method for expanding a knowledge graph of a large-scale model scheduling multi-agent power grid fault response strategy. At the resource allocation level, task dispatch is modeled as a mixed integer programming problem and solved using a relaxation-correction algorithm. The core task is to extract and normalize knowledge from various scheduling text documents such as scheduling procedures and scheduling logs, and output JSON structured data containing standardized text content and metadata. The normalized text is parsed into coherent fragments and irrelevant content is filtered out. The normalized document is divided into fragments, each of which is assigned a relevance score. The significance of the fragment in the power dispatch field relative to the current knowledge graph is evaluated based on domain-specific instructions. The obtained text fragments are summarized to prompt the large language model to retain key entities, relations, and domain-specific terms. LLM-based named entity recognition technology is used, combined with prompts and dictionary / ontology filtering in the power dispatch field, to identify relevant entities in the text and normalize them into a standard form in the knowledge graph. Logical contradictions between newly extracted triples and existing relations in the knowledge graph are detected, and conflicts are classified and resolved using LLM-based debate prompts. Multiple verification signals are aggregated to calculate the global confidence, clarity, and relevance scores for each triple.

[0010] Model the task dispatch as a mixed integer programming problem with the objective function:

[0011]

[0012] For intelligent agents Control exploration intensity,

[0013] Counts global tasks,

[0014] For intelligent agents capacity,

[0015] is a priority indicator.

[0016] Task priority decision is based on the fusion of semantic understanding of language model and reinforcement learning exploration strategy through basic utility function Quantification tasks In the current system state Potential value ;

[0017] Introducing the multi-armed bandit strategy, superimposing the exploration term on the utility value:

[0018]

[0019] in, Control the intensity of exploration, Counts global tasks, Record the historical execution times of a specific task.

[0020] Priority indicators:

[0021]

[0022] For safety constraints,

[0023] To ensure continuous supply,

[0024] and is the weight,

[0025] Comprehensive multi-dimensional considerations to form a dynamically adjusted priority queue .

[0026] Each snippet is assigned a relevance score, which is calculated as:

[0027]

[0028] It is a document fragment after standard processing. It is the original knowledge graph. is the inference model,

[0029] if , the reading module will discard the fragment, and the retained fragment will be passed to the summary module;

[0030] in is the domain-calibrated threshold.

[0031] The original entities are mapped to normalized entities by minimizing the distance function in the joint embedding space, and the new entities are labeled and added to the candidate vertex set.

[0032] Compute global confidence, clarity, and relevance scores for each triple;

[0033] The scores are comprehensively evaluated through weighted average or logical function, and based on the set threshold, it is decided whether the triples should be finally integrated into the knowledge graph.

[0034] The second aspect of the present invention relates to a large-scale model scheduling multi-agent power grid fault response strategy knowledge graph expansion system using the method of the first aspect of the present invention;

[0035] At the resource allocation level, the central scheduling module models task dispatching as a mixed integer programming problem and solves it using a relaxation-correction algorithm.

[0036] The entity extraction module extracts and normalizes knowledge from various scheduling text documents such as scheduling procedures and scheduling logs, and outputs JSON structured data containing standardized text content and metadata.

[0037] The reading module parses the normalized text into coherent segments and filters out irrelevant content. It then segments the normalized document into segments, assigning each segment a relevance score. Based on domain-specific instructions, the module evaluates the segment's significance in the power dispatch domain relative to the current knowledge graph.

[0038] The summarization module summarizes the obtained text fragments, prompting the large language model to retain key entities, relations and domain-specific terms;

[0039] The entity extraction module uses LLM-based named entity recognition technology, combined with prompts and dictionary / ontology filtering in the power dispatching field, to identify relevant entities in the text and normalize them into a standard form in the knowledge graph;

[0040] The schema alignment module matches the newly extracted entities and relations with the existing schema of the knowledge graph;

[0041] The conflict resolution module detects logical contradictions between newly extracted triples and existing relations in the knowledge graph and classifies and resolves conflicts using LLM-based debate prompts;

[0042] The evaluation module aggregates multiple verification signals and computes global confidence, clarity, and relevance scores for each triplet.

[0043] A third aspect of the present invention relates to a terminal, comprising a processor and a storage medium;

[0044] The storage medium is used to store instructions;

[0045] The processor is configured to operate according to the instructions to execute the steps of the method according to the first aspect of the present invention.

[0046] A fourth aspect of the present invention relates to a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the method described in the first aspect of the present invention.

[0047] The beneficial effect of the present invention is that, compared with the prior art, the present invention provides a large-scale model scheduling multi-agent power grid fault response strategy knowledge graph expansion method and system, with the comprehensive goals of ensuring safety, supply, and consumption, based on the multi-objective balanced plan generation safety constraint rules, and by expanding the knowledge graph of fault prediction and fault response strategy, completes the intelligent generation of active safety response strategy plans for power grid events. This method is expected to achieve dynamic adjustment of active safety response strategies for power grid events based on pre-matching migration and deduction of power grid fault events, online verification and generation of plan deviations, and post-plan rolling update and re-verification technology, match power grid fault prevention and control strategies through power grid fault prediction and early warning results, recommend fault handling and prevention steps, assist control personnel in online fault handling, and achieve iterative optimization of active prevention and control strategies for power grid events based on feedback from handling results.

[0048] The beneficial effects of the present invention are also as follows:

[0049] 1. Multi-agent Architecture: This system uses multiple specialized agents to collaborate on tasks such as entity extraction, pattern alignment, and conflict resolution. A cross-validation mechanism is employed. For example, the relationship extraction agent collaborates with the pattern alignment agent to verify entity legitimacy, resolving logical contradictions through a large language model debate mechanism. This improves the reliability of extracted knowledge. For example, the relationship extraction agent verifies candidate entities based on pattern consistency output, while the conflict resolution agent resolves contradictions through a large language model-based debate mechanism.

[0050] 2. Domain-adaptive prompt strategy: This allows the system to maintain accuracy while processing specific contexts and can adapt to the terminology and logic of different fields, especially for various dispatching logs, dispatching procedures and other dispatching domain text data in the power dispatching field.

[0051] 3. Central control scheduling mechanism: The central scheduling module is the core scheduling module of the multi-agent framework, which uses a two-layer optimization mechanism to achieve dynamic task scheduling and resource coordination. Its task priority decision is based on the fusion of the semantic understanding of the language model and the reinforcement learning exploration strategy through the basic utility function. Quantification tasks In the current system state To balance the game between known high-reward tasks and new task exploration, multi-agents introduce a multi-armed bandit strategy, superimposing the exploration term on the utility value:

[0052]

[0053] in Control the intensity of exploration, Counts global tasks, Record the historical execution times and priority indicators of specific tasks:

[0054]

[0055] Combines semantic value and security constraints , continuous supply Multi-dimensional consideration of specific rules in the field of power dispatching, etc., to form a dynamically adjusted priority queue .

[0056] 4. Modular design: ensures scalability and supports dynamic updates when new entities or relationships appear, allowing new scheduling domain knowledge to be expanded into the existing knowledge graph at any time. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a module diagram of a knowledge graph expansion method for large-scale model scheduling of multi-agent power grid fault response strategies according to the present invention. DETAILED DESCRIPTION

[0058] To make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in the present invention are only part of the embodiments of the present invention, not all of the embodiments. Based on the spirit of the present invention, all other embodiments not described in the present invention that are obtained by ordinary technicians in this field based on the embodiments described in the present invention without making creative work should fall within the scope of protection of the present invention.

[0059] In a first aspect of the present invention, a method for expanding a knowledge graph of a large-scale model-based multi-agent power grid fault response strategy is provided, the method comprising:

[0060] At the resource allocation level, task dispatch is modeled as a mixed integer programming problem and solved using a relaxation-correction algorithm.

[0061] The core task is to extract and normalize knowledge from various scheduling text documents such as scheduling procedures and scheduling logs, and output JSON structured data containing standardized text content and metadata;

[0062] Parse the normalized text into coherent segments and filter out irrelevant content. Segment the normalized document into segments, assign a relevance score to each segment, and evaluate its significance in the power dispatch field relative to the current knowledge graph based on domain-specific instructions.

[0063] Summarize the resulting text snippets, prompting large language models to retain key entities, relations, and domain-specific terms;

[0064] Adopting LLM-based named entity recognition technology, combined with prompts and dictionary / ontology filtering in the field of power dispatching, to identify relevant entities in the text and normalize them into a standard form in the knowledge graph;

[0065] Match the newly extracted entities and relations with the existing schema of the knowledge graph;

[0066] Detect logical contradictions between newly extracted triples and existing relations in the knowledge graph, and classify and resolve conflicts using LLM-based debate prompts;

[0067] Aggregate multiple verification signals and compute global confidence, clarity, and relevance scores for each triplet.

[0068] This method proposes a new agent framework that uses the LLM large language model through a multi-agent collaboration. Each agent focuses on a different task in the knowledge graph expansion task pipeline.

[0069] At the resource allocation level, CCA models task dispatch as a mixed integer programming problem with the objective function:

[0070]

[0071] It is required to minimize the total resource consumption with priority weights, and the constraints ensure that the agent The resource load does not exceed its capacity This model supports differentiated resource allocation for heterogeneous computing units (CPU / GPU / TPU) and achieves rapid solutions through a relaxation-correction algorithm. In terms of system design, CCA can adaptively discover high-value task types using time-varying exploration items, utilizing a weighted architecture to achieve multi-objective trade-offs. The modular design of the resource model is compatible with horizontal scalability and supports asynchronous task processing for thousands of agents. A built-in fault-tolerance mechanism automatically increases the priority of stalled tasks through functions, and a timeout retry strategy ensures task reliability. This hybrid architecture, combining semantic reasoning with operational optimization, enables CCA to maintain a balance between scheduling efficiency and robustness in dynamic environments.

[0072] The extraction module is able to complete the core task of efficiently extracting and normalizing knowledge from various scheduling text documents such as scheduling procedures and scheduling logs. Through semantic analysis and format standardization, this module uniformly converts heterogeneous raw text into a structured text representation, while extracting key metadata (including date, time, location, device name, device ownership, and other unique identifiers). Ultimately, the extraction module outputs JSON structured data containing standardized text content and metadata, where the text content is stored as a single string or an array that retains the original hierarchical structure, so that subsequent reading modules can perform content screening and segmentation based on domain relevance scores. This process significantly improves the adaptability of unstructured text to the input of knowledge graph construction through automated error correction and semantic enhancement.

[0073] The reading module parses the normalized text into coherent segments (such as logs, procedures, relevant standards, etc.) and filters out irrelevant content. For normalized documents, the reading module will convert Split into Each snippet is assigned a relevance score, which is calculated as:

[0074]

[0075] According to the instructions of a specific field, the significance of the fragment in the field of power dispatching relative to the current knowledge graph is evaluated. (in is a domain-calibrated threshold), the reading module discards the fragment, and the retained fragments are passed to the summarization module.

[0076] Summarize the text fragments obtained by the reading module and define:

[0077] $$u_j=\text{LLM}{summ}(s_j,P{summ})$$

[0078] Where P{summ} is used to prompt the large language model to retain key entities, relations, and domain-specific terms. This summarization ensures that the entity extraction agent and the relationship extraction agent receive high-signal and low-noise text input.

[0079] The entity extraction module uses LLM-based named entity recognition (NER) technology, combined with hints from the power dispatch domain and dictionary / ontology filtering, to identify relevant entities in the text and normalize them into a standard form within the knowledge graph. Original entities are mapped to normalized entities by minimizing the distance function in the joint embedding space. New entities are then labeled and added to the candidate vertex set.

[0080] The schema alignment module is responsible for matching newly extracted entities and relations with existing schemas in the knowledge graph. Unmatched entities and relations are mapped to known types through domain-specific classification using the LLM. If no suitable match can be found, they are marked as candidate additions for subsequent review.

[0081] The conflict resolution module detects logical contradictions between newly extracted triples and existing relations in the knowledge graph. Using LLM-based debate prompts, it classifies and resolves conflicts. Based on the system's confidence, it decides whether to discard the conflicting triples or submit them for manual review and resolution.

[0082] The evaluation module aggregates multiple verification signals to calculate the global confidence, clarity, and relevance scores for each triple. These scores are comprehensively evaluated through weighted averaging or logistic functions. Based on a set threshold, the module decides whether to integrate the triple into the knowledge graph, ensuring the high quality of the integrated knowledge.

[0083] The quality of the newly added triples is evaluated by average confidence, average clarity, and average relevance. Average confidence reflects the average confidence score of all new triples; average clarity measures the clarity and directness of each relationship; and average relevance reflects the importance of the relationship in the field.

[0084] Graph statistics: Coverage gain and connectivity gain are used to quantify the structural properties of the enhanced knowledge graph. Coverage gain refers to the number of newly added entities in the knowledge graph; connectivity gain is the net increase in the node degree of existing entities.

[0085] Quality Metrics: The reliability and usability of the knowledge graph are assessed using conflict ratio, LLM-based accuracy, and question-answer consistency. The conflict ratio indicates the proportion of newly extracted edges removed by the conflict resolution agent due to inconsistencies; the LLM-based accuracy is the proportion of new triples judged to be potentially correct out of all new triples; and question-answer consistency is the proportion of answers obtained through knowledge graph traversal that are considered reasonable.

[0086] This method proposes a multi-agent approach to expand the knowledge graph of power grid fault response strategies. This approach leverages multi-agent approaches to address the challenge of expanding the knowledge graph from dispatch domain text. By subdividing tasks into specialized agents, precise knowledge integration is achieved. Experiments validate its advantages over single-agent approaches. This architecture utilizes a large power dispatch model fine-tuned from an open-source large model.

[0087] In a second aspect of the present invention, a large-scale model scheduling multi-agent power grid fault response strategy knowledge graph expansion system using a method is characterized by:

[0088] At the resource allocation level, the central scheduling module models task dispatching as a mixed integer programming problem and solves it using a relaxation-correction algorithm.

[0089] The entity extraction module extracts and normalizes knowledge from various scheduling text documents such as scheduling procedures and scheduling logs, and outputs JSON structured data containing standardized text content and metadata.

[0090] The reading module parses the normalized text into coherent segments and filters out irrelevant content. It then segments the normalized document into segments, assigning each segment a relevance score. Based on domain-specific instructions, the module evaluates the segment's significance in the power dispatch domain relative to the current knowledge graph.

[0091] The summarization module summarizes the obtained text fragments, prompting the large language model to retain key entities, relations and domain-specific terms;

[0092] The entity extraction module uses LLM-based named entity recognition technology, combined with prompts and dictionary / ontology filtering in the power dispatching field, to identify relevant entities in the text and normalize them into a standard form in the knowledge graph;

[0093] The schema alignment module matches the newly extracted entities and relations with the existing schema of the knowledge graph;

[0094] The conflict resolution module detects logical contradictions between newly extracted triples and existing relations in the knowledge graph and classifies and resolves conflicts using LLM-based debate prompts;

[0095] The evaluation module aggregates multiple verification signals and computes global confidence, clarity, and relevance scores for each triplet.

[0096] It is understandable that in order to implement the various functions in the method provided in the above-mentioned embodiment of the present application, the system includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should easily appreciate that, in conjunction with the algorithm steps of each example described in the embodiment disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0097] The embodiment of the present application can divide the system into functional modules according to the above method example. For example, each functional module can be divided according to each function, or two or more functions can be integrated into one processing module. The above integrated modules can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the embodiment of the present application is schematic and is only a logical function division. In actual implementation, there may be other division methods.

[0098] The device includes at least one processor, a bus system, and at least one communication interface. The processor is composed of a central processing unit, a field programmable logic gate array, an application-specific integrated circuit, or other hardware. The memory is composed of a read-only memory, a random access memory, or the like. The memory can be independent and connected to the processor via a bus. The memory can also be integrated with the processor. The hard disk can be a mechanical disk or a solid-state drive, etc. The embodiments of the present invention are not limited to this. The above embodiments are generally implemented through software and hardware. When implemented using a software program, it can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions.

[0099] When loading and executing computer program instructions on a computer, corresponding functions are implemented according to the process provided in the embodiments of the present invention. The computer program instructions involved may be assembly instructions, machine instructions, or codes written in a programming language, etc.

[0100] This method proposes a multi-agent approach to expand the knowledge graph of power grid fault response strategies. Leveraging multi-agents, this approach addresses the challenge of expanding the knowledge graph from dispatch domain text. By subdividing tasks into specialized agents, it achieves precise knowledge integration. Experiments validate its advantages over single-agent approaches. This architecture utilizes a large power dispatch model fine-tuned from an open-source large model.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for expanding the knowledge graph of multi-agent power grid fault response strategies for large-scale model scheduling, characterized by: The method comprises: At the resource allocation level, task dispatch is modeled as a mixed integer programming problem and solved using a relaxation-correction algorithm. The core task is to extract and normalize knowledge from various scheduling text documents, including scheduling procedures and scheduling logs, and output JSON structured data containing standardized text content and metadata; Parse the normalized text into coherent segments and filter out irrelevant content. Segment the normalized document into segments, assign a relevance score to each segment, and evaluate its significance in the power dispatch field relative to the current knowledge graph based on domain-specific instructions. Summarize the resulting text snippets, prompting large language models to retain key entities, relations, and domain-specific terms; Adopting LLM-based named entity recognition technology, combined with prompts and dictionary / ontology filtering in the field of power dispatching, to identify relevant entities in the text and normalize them into a standard form in the knowledge graph; Detect logical contradictions between newly extracted triples and existing relations in the knowledge graph, and classify and resolve conflicts using LLM-based debate prompts; Aggregate multiple verification signals and compute global confidence, clarity, and relevance scores for each triplet.

2. A method for expanding the knowledge graph of multi-agent power grid fault response strategies for large-scale model scheduling according to claim 1; characterized in that: Model the task dispatch as a mixed integer programming problem with the objective function: For intelligent agents Control exploration intensity, Counts global tasks, For intelligent agents capacity, is a priority indicator.

3. The method for expanding the knowledge graph of multi-agent power grid fault response strategies for large-scale model scheduling according to claim 1 is characterized by: Task priority decision is based on the fusion of semantic understanding of language model and reinforcement learning exploration strategy through basic utility function Quantification tasks In the current system state Potential value ; is the output of the central scheduling module of the large language model, Introducing the multi-armed bandit strategy, superimposing the exploration term on the utility value: in, Control the intensity of exploration, Counts global tasks, Record the historical execution times of a specific task.

4. The method for expanding the knowledge graph of multi-agent power grid fault response strategies for large-scale model scheduling according to claim 2 is characterized by: Priority indicators: For safety constraints, To ensure continuous supply, and is the weight, Comprehensive multi-dimensional considerations to form a dynamically adjusted priority queue .

5. The method for expanding the knowledge graph of multi-agent power grid fault response strategies for large-scale model scheduling according to claim 1 is characterized by: Each snippet is assigned a relevance score, which is calculated as: It is a document fragment after standard processing. It is the original knowledge graph. is the inference model, if , the reading module will discard the fragment, and the retained fragment will be passed to the summary module; in is the domain-calibrated threshold.

6. The method for expanding the knowledge graph of large-scale model scheduling multi-agent power grid fault response strategy according to claim 1 is characterized by: The original entities are mapped to normalized entities by minimizing the distance function in the joint embedding space, and the new entities are labeled and added to the candidate vertex set.

7. The method for expanding the knowledge graph of large-scale model scheduling multi-agent power grid fault response strategy according to claim 1 is characterized by: Compute global confidence, clarity, and relevance scores for each triple; The scores are comprehensively evaluated through weighted average or logical function, and based on the set threshold, it is decided whether the triples should be finally integrated into the knowledge graph.

8. A system for expanding a knowledge graph of a large-scale model-based multi-agent power grid fault response strategy using the method of any one of claims 1 to 7; characterized in that: At the resource allocation level, the central scheduling module models task dispatching as a mixed integer programming problem and solves it using a relaxation-correction algorithm. The entity extraction module extracts and normalizes knowledge from various scheduling text documents, including scheduling procedures and scheduling logs, and outputs JSON structured data containing standardized text content and metadata. The reading module parses the normalized text into coherent segments and filters out irrelevant content. It then segments the normalized document into segments, assigning each segment a relevance score. Based on domain-specific instructions, the module evaluates the segment's significance in the power dispatch domain relative to the current knowledge graph. The summarization module summarizes the obtained text fragments, prompting the large language model to retain key entities, relations and domain-specific terms; The entity extraction module uses LLM-based named entity recognition technology, combined with prompts and dictionary / ontology filtering in the power dispatching field, to identify relevant entities in the text and normalize them into a standard form in the knowledge graph; The schema alignment module matches the newly extracted entities and relations with the existing schema of the knowledge graph; The conflict resolution module detects logical contradictions between newly extracted triples and existing relations in the knowledge graph and classifies and resolves conflicts using LLM-based debate prompts; The evaluation module aggregates multiple verification signals and computes global confidence, clarity, and relevance scores for each triplet.

9. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images