A power distribution station engineering file optimization suggestion generation method, system and electronic device

Abstract

The present application relates to a power distribution station engineering file optimization suggestion generation method, system and electronic equipment, and belongs to the technical field of natural language processing. The present application solves the problems of low review efficiency and inaccurate optimization suggestions caused by the diversity of power distribution station engineering files, the complexity of review rules and the dependence on manual experience. The present application includes: based on the rule materials of the electric power industry, a prototype knowledge system containing ontology and rule base is constructed using a large language model; based on expert causes, an expert cause library is obtained using an embedding model; the power distribution station engineering project data is parsed, and combined with the prototype knowledge system, structured semantic entity pairs are extracted using a multi-modal large model; based on the prototype knowledge system and the expert cause library, the semantic entity pairs are reasoned and verified through a multi-agent collaborative review mechanism to generate a review conclusion; the review conclusion is polished using a multi-modal large model to generate optimization suggestions. The present application realizes the improvement of engineering file review efficiency and the accuracy of optimization suggestions.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method, system, and electronic device for generating optimization suggestions for power distribution station engineering documents. Background Technology

[0002] With the increasing complexity and scale of power engineering projects, and the continuous updating of relevant technical standards, the amount of engineering documents involved—including engineering design specifications, load calculation sheets, equipment selection lists, wiring diagrams, project applications, approval documents, and site photos—is exploding. This involves a wide range of heterogeneous data formats. The ability to quickly and accurately identify, analyze, and extract key information closely related to current review needs from this massive and highly specialized power engineering documentation, and to use this information to assist reviewers or automated artificial intelligence algorithms in conducting efficient and accurate analysis and reasoning to generate professional and evidence-based optimization suggestions, has become a major challenge that the power industry urgently needs to address.

[0003] Currently, the review and optimization suggestions for substation engineering documents still mainly rely on the traditional manual review model, supplemented by a large amount of human experience-based judgment. Although these traditional models have played an important role in ensuring project quality, their inherent limitations are becoming increasingly apparent when dealing with highly specialized documents such as substation engineering documents, which are heavily reliant on domain knowledge and have diverse formats.

[0004] These limitations lead to problems such as low review efficiency, insufficient accuracy of results, and high labor costs. Specifically, due to the low efficiency of manual review, it is difficult to cope with the review needs of massive amounts of complex data: when a single project of existing power distribution station engineering data may contain dozens or even hundreds of documents, with a total of hundreds or even thousands of pages of multimodal heterogeneous data, manual review cannot quickly and effectively extract key content from massive amounts of information and perform cross-modal consistency checks, resulting in severely low review efficiency and failing to meet the needs of rapid business response and batch processing. Due to the lack of standardized knowledge bases and explicit expression of expert experience, manual review often relies on the tacit knowledge and intuitive judgment accumulated by reviewers over a long period of time. This tacit knowledge is difficult to effectively inherit and quantify, resulting in uneven experience levels among different reviewers, which can easily lead to inconsistent review standards, omission of key issues, or even misjudgments. Due to the difficulty of multimodal data parsing and cross-modal consistency checks, it is difficult for manual review to accurately and effectively extract key information from these complex data and perform cross-modal semantic consistency checks, which can easily lead to project delays, repeated modifications, and seriously restrict the improvement of business quality and efficiency. The lack of an automated and intelligent decision-making mechanism makes it difficult to generate accurate optimization suggestions, increasing the risk of project rework. Furthermore, it makes it difficult to achieve closed-loop feedback and continuous optimization of the review system, thus limiting the progress of power distribution station project review towards a higher standard of automation and intelligence. Summary of the Invention

[0005] Based on the above analysis, the embodiments of the present invention aim to provide a method, system, and electronic device for generating optimization suggestions for power distribution station engineering documents, in order to solve the problems of low review efficiency and inaccurate optimization suggestions caused by the diversity of power distribution station engineering documents, the complexity of review rules, and reliance on human experience.

[0006] On one hand, embodiments of the present invention provide a method for generating optimization suggestions for substation engineering documents, including the following steps:

[0007] Based on the rules and materials of the power industry, a prototype knowledge system including ontology and rule base is constructed using a large language model; based on expert case examples, an expert case example library is obtained using an embedding model.

[0008] The data of the power distribution station project are analyzed, and structured semantic entity pairs are extracted using a multimodal large model in conjunction with the prototype knowledge system.

[0009] Based on the prototype knowledge system and expert case database, a multi-agent collaborative review mechanism is used to reason and verify semantic entity pairs and generate review conclusions. The multi-agent includes: rule agent, case agent and review agent. All agents are built on the same multimodal big model and are constructed through prompt word templates containing different role-playing and instruction tasks.

[0010] The review conclusions are professionally refined using a multimodal large model to generate optimization suggestions.

[0011] Based on further improvements to the above method, a multi-agent collaborative review mechanism is used to perform reasoning verification on semantic entity pairs and generate review conclusions, including:

[0012] The rule-based agent and the case-based agent each reason about semantic entity pairs and output their preliminary opinions to the review agent.

[0013] The review agent integrates the two preliminary opinions and performs conflict detection to generate a discussion draft. The coordination agent for rules and cases analyzes, votes on, and modifies the discussion draft, and returns the modified preliminary opinions to the review agent to generate a new discussion draft. The coordination agent for rules and cases analyzes, votes on, and modifies the new discussion draft again until all votes pass or the maximum number of iterations is reached. After integrating the preliminary opinions of the final version, the review agent forms the review conclusion.

[0014] Based on further improvements to the above method, the rule-based agent outputs preliminary opinions through the following steps:

[0015] Based on the semantic similarity between semantic entity pairs and the ontologies and rule bases in the prototype knowledge system, the most similar ontologies and rules are recalled.

[0016] Based on semantic entity pairs, similar ontologies, and rules, prompt word templates are constructed to guide the rule-based intelligent agent in logical verification and output the preliminary opinions of the rule-based intelligent agent.

[0017] Based on further improvements to the above method, the case agent outputs preliminary opinions through the following steps:

[0018] Using query question prompt word templates, query questions are generated from semantic entity pairs; based on the embedding vector of the query questions, multiple most similar expert case reasons are retrieved from the expert case reason database;

[0019] Based on the query question, multiple most similar expert case reasons, and semantic entity pairs, a prompt word template is constructed to guide the case agent to perform case reasoning and output the case agent's preliminary opinion.

[0020] Based on further improvements to the above methods, a prototype knowledge system including an ontology and a rule base is constructed using a large language model, including:

[0021] Collect regulatory materials from the power industry and preprocess them;

[0022] Knowledge elements are extracted from pre-processed rule materials using a large language model pre-trained in the power industry.

[0023] Based on the extracted knowledge elements, an ontology and a rule base are constructed to obtain a prototype knowledge system. The ontology includes concepts, attributes, and hierarchical relationships in the field of power engineering, and the rules in the rule base are logical rules expressed in IF-THEN form.

[0024] Based on further improvements to the above methods, the prototype knowledge system is also integrated and updated according to the expert case examples, including:

[0025] Target entities and logical relationships are extracted from expert case descriptions using a large language model;

[0026] The prompt word template guides the large language model to analyze whether the target entity and logical relationship have made corrections or additions to the prototype knowledge system. If so, the prototype knowledge system is integrated and updated to obtain the updated prototype knowledge system.

[0027] Further improvements to the above methods involve integrating and updating the prototype knowledge system, including:

[0028] If the target entity does not exist in the prototype knowledge system, then the concepts and attributes of the target entity are obtained and added to the ontology of the prototype knowledge system.

[0029] If the logical relation does not exist in the prototype knowledge system, then the logical relation is converted into a logical rule expressed in IF-THEN form and added to the rule base of the prototype knowledge system;

[0030] If a logical relationship exists in the prototype knowledge system but is inconsistent with existing rules, the updated rule is output, and after being resolved by experts or a preset priority mechanism, it is added to the rule base of the prototype knowledge system.

[0031] Based on further improvements to the above method, an expert case base is obtained using an embedding model, based on expert case causes, including:

[0032] After the expert case preprocessing text is processed, it is sliced ​​into main paragraphs according to the preset segment word count; then, for each main paragraph, the transition paragraphs are extracted according to the preset number of overlapping words before and after it, and the main paragraphs are spliced ​​with the corresponding transition paragraphs to generate the final text of each paragraph.

[0033] Embedding vectors of each text segment are extracted using an embedding model and stored in an expert case database.

[0034] Based on further improvements to the above method, knowledge elements are extracted from the preprocessed rule material, including:

[0035] Using cue word engineering, identify and extract entities in the field of power engineering;

[0036] Extract semantic relationships between entities;

[0037] Extract rules and constraints represented by logical expressions or pseudocode.

[0038] Based on further improvements to the above methods, the data of power distribution station projects are analyzed, including:

[0039] Obtain each page of the document containing information about the power distribution station project;

[0040] Use document parsing tools to extract the text blocks and their layout information from each page of the document;

[0041] The semantic type of each text block is identified based on the layout information, and the text content of each text block is extracted.

[0042] Based on further improvements to the above method, the layout information includes: the coordinates of the text block, font, color, font size, line height, horizontal alignment, whether it is bold and italic; the semantic types include: title, body text, table and legend.

[0043] Based on further improvements to the above methods, and combined with a prototype knowledge system, structured semantic entity pairs are extracted using a multimodal large model, including:

[0044] Construct prompt word templates containing different entity extraction tasks, and use the prototype knowledge system as background knowledge and constraints in the prompt word templates to guide the multimodal large model to perform entity extraction tasks on each text block, and output the obtained entities and their values ​​in JSON key-value pair format as semantic entity pairs.

[0045] Based on further improvements to the above methods, entity extraction tasks include: extracting project metadata, extracting engineering parameters, identifying equipment configuration, identifying basic drawing information, identifying drawing technical requirements, and identifying at least one of standards and specifications.

[0046] Based on further improvements to the above method, entities in the field of power engineering include: equipment name, equipment parameters, equipment number, and process name; semantic relationships between entities include: type, capacity relationship, and usage method; rules and constraints include: IF-THEN rules, constraint conditions, calculation formulas, and recommended value ranges.

[0047] Further improvements to the above method, after generating optimization suggestions, also include: using expert feedback and evaluation of the generated optimization suggestions as new expert case reasons, and updating the expert case reason library and prototype knowledge system.

[0048] Further improvements to the above methods utilize a multimodal large model to professionally refine review conclusions. This involves defining clear roles, tasks, and refinement requirements within the refinement prompt template, refining dynamically introduced review conclusions, and outputting a final review opinion and optimization suggestion report. Refinement requirements include: professional language style, clear structure, emphasis on key points, and objectivity of reserved opinions.

[0049] On the other hand, embodiments of the present invention provide a system for generating optimization suggestions for substation engineering documents, including:

[0050] The prototype knowledge system construction module is used to build a prototype knowledge system containing an ontology and a rule base based on rule materials from the power industry and using a large language model.

[0051] The expert case library construction module is used to generate an expert case library based on expert case causes using an embedding model.

[0052] The project information extraction module is used to analyze the data of power distribution station engineering projects and, in conjunction with the prototype knowledge system, extract structured semantic entity pairs using a multimodal large model.

[0053] The review conclusion generation module is used to generate review conclusions by reasoning and verifying semantic entity pairs based on the prototype knowledge system and expert case database through a multi-agent collaborative review mechanism. The multi-agent includes: rule agent, case agent and review agent. All agents are built based on the same multimodal big model and are constructed through prompt word templates containing different role-playing and instruction tasks.

[0054] The optimization suggestion generation module is used to professionally refine the review conclusions using a multimodal large model and generate optimization suggestions.

[0055] Based on further improvements to the above system, the system also includes:

[0056] The feedback update module is used to update the expert case library and prototype knowledge system by taking the experts' feedback and evaluation of the generated optimization suggestions as new expert case reasons.

[0057] This invention also provides an electronic device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it executes the substation engineering document optimization suggestion generation method described above.

[0058] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0059] 1. By constructing a maintainable prototype knowledge system and expert case database, unified semantic parsing and structured information extraction of multimodal data were achieved, and automated verification was performed using rule knowledge and engineering logic. On this basis, a multi-agent collaborative mechanism was introduced, combined with a multimodal large model, to generate professional, accurate, and operable intelligent review assistance opinions and optimization suggestions. This achieved an intelligent assisted review and optimization suggestion generation system with full-chain capabilities of "identification-reasoning-review-optimization suggestion generation," improving the efficiency and quality of the review process, promoting the intelligent level of power distribution station engineering document review, and providing a replicable intelligent solution for other power engineering review scenarios, thus helping the power industry to fully realize digital transformation.

[0060] 2. It realizes the automated parsing and structured information extraction of multimodal engineering data, and combines the automatic verification of rule knowledge and engineering logic, which greatly reduces the time spent on manual review, improves the standardization level and accuracy of the review process, and effectively avoids errors or omissions caused by human negligence, thereby greatly improving the overall efficiency and quality of the review work.

[0061] 3. By integrating a large power industry model with a multi-agent collaborative mechanism (rule agent, case agent, and review agent), deep reasoning is performed to generate specific and quantifiable suggestions that not only comply with standards but also have high guiding significance. These suggestions cover multiple core dimensions such as technical parameters, equipment selection, and construction technology, providing strong professional support for engineering decisions and ensuring the operability of the optimization plan.

[0062] 4. A maintainable prototype knowledge system and expert case library have been constructed, transforming implicit and scattered expert experience into explicit and systematic knowledge assets. This not only helps to effectively inherit and reuse valuable expert experience and accelerate the training of new personnel, but also enables the system capabilities to evolve continuously through continuous feedback learning and closed-loop optimization, forming a knowledge ecosystem with autonomous learning and continuous optimization capabilities.

[0063] 5. Through a multi-agent collaborative review mechanism, the system efficiently and accurately verifies power distribution station engineering documents and generates high-quality, professional optimization suggestions. This not only improves the reliability of the review results but also simulates the advantages of collaborative human expert teams, overcoming the limitations of a single model. The system embeds identification results, reasoning conclusions, and review suggestions into structured template outputs to support human-machine collaboration. Finally, through feedback optimization, a closed-loop mechanism is formed that integrates feedback adjustment, autonomous repair, and experience accumulation, propelling power engineering review towards a higher standard of automation and intelligence.

[0064] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0065] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0066] Figure 1 This is a flowchart of a method for generating optimization suggestions for power distribution station engineering documents according to Embodiment 1 of the present invention;

[0067] Figure 2 This is a schematic diagram of the multi-agent collaborative review mechanism in Embodiment 1 of the present invention;

[0068] Figure 3 This is a schematic diagram of a power distribution station engineering document optimization suggestion generation system in Embodiment 2 of the present invention. Detailed Implementation

[0069] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the invention, but are not intended to limit the scope of the invention. It should be understood that the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0070] Those skilled in the art should understand that the following specific embodiments or implementation methods are a series of optimized configurations listed to further explain the specific content of the invention. These configurations can be combined or used in conjunction with each other, unless the invention explicitly states that some or a specific embodiment or implementation method cannot be associated with or used in conjunction with other embodiments or implementation methods. Furthermore, the following specific embodiments or implementation methods are merely optimized configurations and are not intended to limit the scope of protection of the invention.

[0071] Example 1

[0072] A specific embodiment of the present invention discloses a method for generating optimization suggestions for substation engineering documents, such as... Figure 1 As shown, it includes the following steps:

[0073] S1. Based on the rule materials of the power industry, construct a prototype knowledge system including ontology and rule base using a large language model; based on expert case reasons, obtain an expert case reason library using an embedding model;

[0074] S2. Analyze the data of the power distribution station project and extract structured semantic entity pairs using a multimodal large model in conjunction with the prototype knowledge system.

[0075] S3. Based on the prototype knowledge system and expert case database, the semantic entity pairs are reasoned and verified through a multi-agent collaborative review mechanism to generate review conclusions.

[0076] S4. Use a large language model to professionally refine the review conclusions and generate optimization suggestions.

[0077] In implementation, this embodiment aims to construct an intelligent assisted review and optimization suggestion generation system with a full-chain capability encompassing "identification-reasoning-review-optimization suggestion generation." By building a maintainable prototype knowledge system, it achieves unified semantic parsing and structured information extraction from multimodal data, and uses rule-based knowledge and engineering logic for automated verification. Based on this, a multi-agent collaborative mechanism is introduced, combined with a large-scale power industry model, to generate professional, accurate, and actionable intelligent review assistance opinions and optimization suggestions, and a feedback mechanism is implemented to continuously optimize system capabilities.

[0078] The following provides a detailed explanation of each step.

[0079] The review of power distribution station engineering documents relies heavily on standardized industry knowledge and expert experience. To achieve the systematic application and maintainability of this knowledge, the core knowledge base required is first constructed according to step S1, including a maintainable prototype knowledge system based on rule-based materials and a vectorized expert case library based on historical expert case studies.

[0080] The construction process of the prototype knowledge system is not merely a simple compilation of rule materials, but a process of in-depth refinement, structuring, formalization, and maintainability of the rule engine through advanced natural language processing technology and knowledge engineering methods. The construction process includes steps S11-S13, and to ensure the real-time performance, accuracy, and completeness of the knowledge system, step S14 implements the integration of expert case data and a dynamic maintenance mechanism for the knowledge system.

[0081] S11. Collect regulatory materials from the power industry and preprocess them.

[0082] It should be noted that the collected power industry regulations and materials include: technical standards, design guidelines, construction specifications, and recommended parameters for equipment selection issued by power grid companies.

[0083] Furthermore, the collected rule materials are preprocessed, including:

[0084] Optical Character Recognition (OCR) technology is used to process scanned documents and PDF images to ensure all materials are machine-readable. Text content is extracted using document parsing tools, and formatting and text cleaning are performed to remove irrelevant headers, footers, watermarks, and other noise. Layout analysis is used to initially identify structural information such as chapters, headings, and paragraphs, resulting in a structured Markdown version of the material.

[0085] S12. Using a large language model pre-trained in the power industry, knowledge elements are extracted from the pre-processed rule materials.

[0086] It should be noted that the preprocessed set of rule materials is as follows: ,in This represents a document. It uses a large pre-trained language model from the power industry as the core tool, and designs sophisticated prompt word engineering. Guiding large language models In-depth analysis and knowledge extraction are performed on the preprocessed rule materials. The extraction process aims to make the implicit business rules and judgment logic scattered in the text explicit and structured. At the same time, the formalized knowledge system facilitates the extraction of key information in the later text, and the abstract generalization avoids excessive adherence to the details and overfitting during model reasoning.

[0087] Specifically, knowledge elements are extracted from the preprocessed rule materials, including:

[0088] Using cue word engineering, identify and extract entities in the field of power engineering, including: equipment name, equipment parameters, equipment number, and process name;

[0089] For example, equipment names include: transformer, circuit breaker; equipment parameters include: voltage level, capacity.

[0090] Extract semantic relationships between entities, including: type, capacity, and usage.

[0091] Extract rules and constraints represented by logical expressions or pseudocode, including: IF-THEN rules, constraints, calculation formulas, and recommended value ranges.

[0092] It should be noted that rules and constraints are guided by prompts to the large language model to summarize and generalize the logical structure and element requirements of the rule materials, covering all aspects of power design and clearly defining the relationships and judgment conditions between various review elements. For example, the rationality judgment of the power supply scheme needs to comprehensively consider the matching degree between transformer capacity and user load, whether the access point meets the power supply radius limit, whether there are safety hazards in the line laying path, etc., and strictly follow industry standards.

[0093] Extracted knowledge elements , can be represented as:

[0094] ,

[0095] in, Represents a set of entities Entities in;

[0096] Representing a set of relations medium entity The relationship between them;

[0097] Representing a set of rules The first in One rule.

[0098] Prompt words for:

[0099] "You are an expert in analyzing power engineering specifications. Please perform the following tasks from the input document:\n1. Identify all power engineering entities.\n2. Extract semantic relationships between entities.\n3. Extract conditions—action rules, constraints, or calculation formulas."

[0100] For example, when the input document is:

[0101] The extracted knowledge element is: "For distribution transformers with a capacity exceeding 800kVA, circuit breakers and relay protection systems should be installed; the combination of load switches and fuses is prohibited. Oil-immersed transformers with ONAN cooling are used."

[0102] "{\n"entities":["Distribution Transformer","800kVA","Circuit Breaker","Relay Protection System","Load Switch + Fuse","Oil-Immersed Transformer","ONAN"],\n"relations":[\n["Distribution Transformer","Capacity","Exceeding 800kVA"],\n["Distribution Transformer","Should Be Configured","Circuit Breaker"],\n["Distribution Transformer","Should Be Configured","Relay Protection System"],\n["Distribution Transformer","Prohibited from Use","Load Switch + Fuse"],\n["Distribution Transformer","Model","Oil-Immersed Transformer"],\n["Oil-Immersed Transformer","Cooling Method","ONAN"]\n],\n"rules":[\n"IF Transformer.Capacity>800kVA THEN Must.Configuration=Circuit Breaker + Relay Protection AND Prohibited.Use=Load Switch + Fuse"\n]\n}.

[0103] S13. Based on the extracted knowledge elements, construct an ontology and a rule base to obtain a prototype knowledge system; wherein, the ontology includes concepts, attributes and hierarchical relationships in the field of power engineering, and the rules in the rule base are logical rules represented in IF-THEN form.

[0104] It should be noted that, based on the extracted knowledge elements, a multi-layered and multi-granular prototype knowledge system is constructed. This system can use a combination of ontology and rule base to represent entity relationships and logical rules separately.

[0105] The ontological structure is represented as ,in, For a set of concepts, A set of relationships between concepts. It is a set of attributes. For example, "Equipment" is the parent class of "Transformer" and "Circuit Breaker". "Transformer" has attributes such as "Capacity", "Voltage Rating", and "Cooling Method".

[0106] This embodiment uses prompt words. Indicating that the large language model is in Based on this, it assists in the identification of ontology concepts, the construction of hierarchical structures, and the definition of attributes, represented as:

[0107] .

[0108] Prompt words for:

[0109] "You are an electrical engineering ontology engineer. Based on the extracted entities, relationships, and rules, construct a structured electrical engineering ontology. Tasks: 1. Define the hierarchical structure (is-a relationship) of the top-level concepts and their subclasses. 2. Define the attributes that each concept should possess."

[0110] For example, based on the example output in step S12, the obtained ontology is: "{\n"concepts":[\n{"name":"device","parent":null},\n{"name":"transformer","parent":"device"},\n{"name":"oil-immersed transformer","parent":"transformer"},\n{"name":"protection device","parent":"device"},\n{"name":"circuit breaker","parent":"protection device"},\n{"name":"relay protection system","parent":"protection device"},\n{"name":"load switch + fuse","parent":"protection device"}\n],\n"attributes":{\n"transformer":["capacity","voltage level","model"],\n"oil-immersed transformer":["cooling method"],\n"protection device":["applicable transformer capacity"]\n}\n}".

[0111] Furthermore, the rules and constraints within the knowledge elements are formalized into executable logical rules. This embodiment selects a simple rule representation method based on logical implication to ensure the readability and ease of implementation of the rules. Each rule consists of an antecedent and a consequence, forming an IF-THEN form. When the antecedent is satisfied, the consequence is derived or triggered. A general rule can be represented as:

[0112] ,

[0113] in, These represent the preconditions for the rules, such as comparisons of entity attributes and checks on the existence of relationships. Represents the conclusion or triggered action of a rule. The premises and conclusions may contain complex logical operators (AND, OR, NOT) and numerical comparisons.

[0114] Logically, this can be expressed as:

[0115] ,

[0116] Similarly, constructing structured cue words Guiding large language models Based on the extracted rules and constraints, formalized logical rules are output in the IF-THEN form, which includes premises and conclusions.

[0117] For example, the formalized output of the rules extracted from the knowledge elements above is: "IF(Transformer.Capacity>800kVA)THEN(Must.Main Protection=="Circuit Breaker + Relay Protection")AND(Prohibited.Use=="Load Switch + Fuse")".

[0118] A rule base consists of a series of such formalized logical rules: In the subsequent automated verification phase, these rules will be matched and executed by the inference engine to arrive at the review conclusion.

[0119] Ultimately, the prototype knowledge system It is the main body and rule base The combination of: .

[0120] It should be noted that the goal of the prototype knowledge system is to provide a clear and callable knowledge base for subsequent data analysis and automated verification, ensuring that all auditing activities are based on unified and authoritative industry standards.

[0121] S14. Integrate and update the prototype knowledge system.

[0122] First, the collected expert case examples This includes: expert judgment cases for specific engineering problems, typical defect reports, review meeting minutes, as well as verified excellent engineering design schemes, construction logs, and acceptance reports.

[0123] It should be noted that expert opinions are usually presented as one or two sentences of reasons and suggestions for modification for each major category of issues. This format makes the knowledge-sharing process simple and easy to handle. Different opinions are also relatively independent of each other and can be used separately without considering the relationship between them.

[0124] Furthermore, utilizing large language models Combine with prompt word template Perform in-depth analysis to extract the target entities and logical relationships contained within. , is represented as:

[0125] .

[0126] Prompt template The task is as follows: "You are an expert in extracting knowledge from electrical engineering cases. Please conduct an in-depth analysis of the expert's case, extracting the reusable professional knowledge hidden within it. Your tasks are: 1. Identify the target entity: Find the core equipment, parameters, or systems directly related to the opinion, and clarify their specific attributes. Format: Target entity: [Concept].[Attribute] 2. Infer logical relationships: Analyze the causal relationships, conditional judgments, or empirical rules implicit in the opinion. Format: Logical relationship: [Premise] → [Conclusion / Recommendation] Please analyze based on the following expert case:" Please strictly follow the above format for output and do not include any other content.

[0127] For example, when the expert's case input is: "The main transformer capacity of this station is 1000kVA, and it should be replaced with a circuit breaker scheme.", the following will be extracted: "Target entity: protection equipment.type\n Logical relationship: transformer capacity > 800kVA → configure circuit breaker scheme".

[0128] Furthermore, by using prompt word templates to guide the large language model to analyze whether the target entities and logical relationships have made corrections or additions to the prototype knowledge system, the prototype knowledge system is then integrated and updated to obtain the updated prototype knowledge system.

[0129] It should be noted that the prompt word template for:

[0130] "You are a power engineering knowledge base maintenance expert. The existing knowledge system includes: {{" The target entities and logical relationships extracted from the expert case are as follows: {{ Please analyze the expert's case and determine whether it modifies or supplements the existing knowledge system. If so, please make minimal modifications and output the updated rules. If there are no conflicts, please output the original text.

[0131] Furthermore, the prototype knowledge system is integrated and updated, including: the addition of concepts and attributes, the addition and modification of rules, and conflict detection and resolution.

[0132] Specifically, if the target entity does not exist in the prototype knowledge system, it means that the expert's case introduces new concepts and attributes that are not included in the existing ontology, and the concepts and attributes in the target entity are obtained and added to the ontology of the prototype knowledge system.

[0133] If the logical relation does not exist in the prototype knowledge system, it means that the expert has discovered a new rule, and the logical relation is converted into a logical rule expressed in IF-THEN form and added to the rule base of the prototype knowledge system.

[0134] If a logical relationship exists in the prototype knowledge system but is inconsistent with existing rules, it indicates that there is a conflict between the expert experience in the expert case and the old knowledge. The updated rule is output, and after being resolved by expert confirmation or a preset priority mechanism, it is added to the rule base of the prototype knowledge system.

[0135] The above process can be represented as follows:

[0136] .

[0137] It should be noted that if new rules or expert cases are added in the future, the rules will be supplemented according to the new rules and the prototype knowledge system will be integrated and updated according to the new expert cases, in accordance with steps S11-S14. This will enable the prototype knowledge system to continuously learn, evolve and improve itself, and always keep pace with the latest industry standards and expert experience, thereby providing a solid and dynamic knowledge foundation for the intelligent auxiliary review and optimization suggestions of power distribution station engineering documents.

[0138] S15. Based on expert case causes, obtain an expert case cause library using an embedding model.

[0139] Specifically, the text after the expert case preprocessing is sliced ​​into main segments according to the preset segment word count; then, transition segments are extracted from each main segment according to the preset number of overlapping words before and after it; the main segments are then spliced ​​with the corresponding transition segments to generate the final text of each segment.

[0140] Embedding vectors of each text segment are extracted using an embedding model and stored in an expert case database.

[0141] For example, the number of segmented words is set to 1000, and the number of overlapping words is set to 100; the embedding model adopts the GTE-large-zh model.

[0142] In step S2, based on the prototype knowledge system and the data of the power distribution station project, a multimodal large model is used. Using lightweight document parsing tools as the parsing engine, a recognition system with multimodal integrated perception is built to achieve efficient and accurate parsing and structured information extraction of these multimodal materials.

[0143] It should be noted that the project documents for substation engineering are generated in various formats during the technical and economic review process, exhibiting significant multi-source, nested, and heterogeneous characteristics. The content spans various data formats, including engineering design specifications, load calculation sheets, equipment selection lists, wiring diagrams, project applications, approval documents, and site photos. A complete project document includes documents in various heterogeneous formats such as Word documents, Excel spreadsheets, PDF reports, CAD drawings, scanned copies, and equipment photographs.

[0144] Analysis of data from power distribution station projects, including:

[0145] Obtain each page of the document containing information about the power distribution station project;

[0146] Use document parsing tools to extract each text block and its layout information from each page of the document; the layout information includes: the coordinates of the text block, font, color, font size, line height, horizontal alignment, whether it is bold and italic;

[0147] The semantic type of each text block is identified based on the layout information, and the text content of each text block is extracted; the semantic types include: title, body text, table and legend.

[0148] Specifically, the MinerU tool is used to extract text and layout information from each page of the document, obtaining the text content and layout information of each text block to form the text sequence of each page of the document: ,in, Indicates the first A text sequence of a page document. and They represent the first Page 1 of the document The text content and layout information of each text block.

[0149] It should be noted that the MinerU tool can identify the type of each text block, including but not limited to: title, text, table, and figure, with corresponding semantic types of title, body text, table, and figure legend.

[0150] For text blocks of heading and body types, the text content can be directly parsed. However, for text blocks of table and legend types, other tools are needed to extract the text content, including:

[0151] For table-type text blocks, tools such as MarkItDown are used to extract the table content from the table area, including rows, columns, cells, and header information, and convert it into Markdown format text content;

[0152] For text blocks of the illustration type, a large language model is used to generate summary text content.

[0153] Furthermore, after obtaining the text content and semantic type of each text block, as relatively independent semantic context fragments, the multimodal large model is... As the core parsing engine, it enables deep semantic understanding and structured information extraction of power distribution station engineering data.

[0154] It should be noted that the multimodal large model can process information from multiple modalities such as text, images, and tables, and integrates the prototype knowledge system built in the previous stage. Enhancements can be made. A general multimodal large model can be used, such as the Qwen-VL large model, or the Guangming Power large model with vertical fine-tuning specific to the power industry.

[0155] Specifically, by combining the prototype knowledge system and utilizing a multimodal large model, structured semantic entity pairs are extracted, including:

[0156] Construct prompt word templates containing different entity extraction tasks, and use the prototype knowledge system as background knowledge and constraints in the prompt word templates to guide the multimodal large model to perform entity extraction tasks on each text block, and output the obtained entities and their values ​​in JSON key-value pair format as semantic entity pairs.

[0157] It should be noted that the entity extraction task includes: extracting project metadata, extracting engineering parameters, identifying equipment configuration, identifying basic information in drawings, identifying technical requirements in drawings, and identifying at least one of the standards and specifications.

[0158] Among them, project metadata is the overall background information of the project, such as the type and scale of the project;

[0159] The basic information of a drawing is its length, width, and height;

[0160] Technical requirements for drawings are the specific parameter values ​​identified in the drawings, while standards and specifications are industry standards mentioned and referenced in the explanatory text, including explanatory notes.

[0161] Based on the semantic type of each text block, the corresponding prompt word template is executed. For example, a text block with the semantic type of "legend" only needs to execute the prompt word template corresponding to the entity extraction tasks of recognizing basic drawing information and recognizing drawing technical requirements.

[0162] For example, the prompt template for extracting engineering parameters is as follows:

[0163] Based on the existing knowledge system: {{ Extract all information related to [engineering parameters] from the following table, including [parameter name] and [value / unit], and output it in JSON format, such as {"parameter name":"voltage level","value":"10kV"}.

[0164] The prompt word template for identifying device configuration is as follows:

[0165] This text describes the configuration of [transformers] in a substation. Based on your existing knowledge: {{ Identify its [model], [capacity], [quantity] and other information, and output it in JSON format, such as {"model":"SCB11","capacity":"800kVA","quantity":"1"}.

[0166] The prompt word template for entity extraction tasks not only includes the instructions for the extraction task, but also incorporates a prototype knowledge system and examples as context. For each text block, the corresponding prompt word template is executed according to its semantic type, thereby achieving named entity recognition and semantic alignment for each text block. The expression is as follows:

[0167] ;

[0168] in, Indicates a prompt word, This represents the extracted semantic entity tags, such as "transformer model", "voltage level", "power supply path", etc. Let it be the value it takes in the original text.

[0169] When a document mentions "using an 800 kVA transformer, model SCB11," the multimodal large model, guided by prompts, automatically parses it into more standardized key elements: {"type":"transformer","value":"SCB11-800kVA"}, effectively filtering out irrelevant content and redundant information. This processing method greatly improves the extraction of semantic information from the document, and the model has the ability to comprehensively understand across pages and paragraphs, automatically integrating interdependent information from preceding and following paragraphs into a unified semantic result.

[0170] For fuzzy descriptions, such as "a transformer with a larger capacity," the model can infer the possible specific numerical range based on context and prototype knowledge (e.g., based on typical load conditions) and perform standardization. Simultaneously, the model possesses context completion capabilities, automatically integrating interdependent information from preceding and following sections into a unified semantic result, resolving issues of scattered and disjointed information. For example, if a document describes "a new load of 1000kW" in the preceding section and "a transformer to be selected" in the following section, the model can correlate the two and infer the transformer capacity requirement. This high-order reasoning approach based on a large language model differs from earlier strategies using fixed keywords or regular expressions, demonstrating greater diversity and flexibility.

[0171] Based on the unified semantic parsing and extraction of multimodal large models, the core task of power distribution station engineering review is to determine whether the engineering data is reasonable, complete, and in compliance with current industry norms and technical standards.

[0172] Therefore, the core of step S3 is to utilize structured semantic entity tags. Through knowledge prototype system Automated logic verification is performed by constructing a multi-agent collaborative review mechanism to simulate the deep understanding and reasoning of power design schemes by human experts, achieving a highly professional judgment and decision-making process, and ultimately generating high-quality review conclusions. This process not only verifies the data but also simulates the deep understanding and reasoning of power design schemes by human experts through scheme design, achieving a highly professional judgment and decision-making process.

[0173] It should be noted that the multi-agent system includes: rule-making agent, case-based agent, and review agent; all agents are based on the same large multimodal model. It is constructed using prompt templates that include different role-playing and instruction tasks.

[0174] Among them, the rule agent is used based on a pre-built prototype knowledge system. Ontology in and rule base This system performs automated and logical compliance verification on input semantic entity pairs. The agent aims to identify explicit, quantifiable violations or inconsistencies.

[0175] The Case Agent is used to provide experience, risk warnings, and multi-dimensional solutions by retrieving historical expert cases that are most similar to the current project data.

[0176] The case-based intelligent agent is designed based on engineering practice. While rule materials often stem from key review elements summarized from past experience, they don't always provide appropriate and timely guidance for scenario details, actual tasks, or the latest developments. Furthermore, power engineering reviews are often quite complex, requiring experts to rely on explicit rule materials and implicit personal experience, employing complex, multi-step reasoning that may not be documented on paper before arriving at a final opinion. Simply reasoning based on rule materials is insufficient for understanding, analyzing, and effectively addressing large and complex real-world project scenarios. Therefore, the case-based intelligent agent aims to overcome the limitations of pure rule verification, making it particularly suitable for handling complex, ambiguous-boundary, or experience-based engineering problems.

[0177] The rule agent and the case agent perform rule verification and case reasoning respectively. On this basis, a review agent with higher authority is introduced as the core coordinator to control the overall process and form the final conclusion.

[0178] In other words, the review agent is responsible for integrating the preliminary opinions from the rule-making agent. Preliminary opinions from case-specific intelligent agents This process identifies conflicting and complementary viewpoints, and through an iterative "analysis-voting-revision" mechanism, ultimately arrives at a comprehensive, rigorous, and consistent review conclusion. .

[0179] Specifically, a multi-agent collaborative review mechanism is used to perform reasoning verification on semantic entity pairs and generate review conclusions, including:

[0180] ① The rule-based agent and the case-based agent each reason about the semantic entity pairs and output their preliminary opinions to the review agent.

[0181] It should be noted that the rule-based intelligent agent outputs its initial opinions through the following steps:

[0182] Based on the semantic similarity between semantic entity pairs and the ontologies and rule bases in the prototype knowledge system, the most similar ontologies are recalled. and rules ;

[0183] Based on semantic entity pairs, similar ontologies, and rules, a prompt word template is constructed to guide the rule-based agent in logical verification, and the preliminary opinion of the rule-based agent is output. .

[0184] Specifically, for each extracted semantic entity pair The rule-based agent recalls the most similar ontology by calculating its semantic similarity (cosine similarity) with relevant paragraphs in the ontology and rule base. and logical rules Then, based on the prompt words, input the most similar ontology, logical rules, and semantic entity pairs, and let the rule-based intelligent agent perform reasoning.

[0185] The prompt template for the rule-based intelligent agent is: "You are a power compliance review expert, specializing in referencing the entity relationships and rule base retrieved through retrieval to perform logical verification and compliance checks on the extracted engineering entities.\nThe rules for retrieval and recall include: {{ }};\nThe retrieved ontology includes: {{ }};\nThe input engineering semantic entity pairs are: {{ Task: Match relevant rules, determine if there is a violation, and output the violation item, the basis for the rule, and preliminary suggestions. Finally, compile a preliminary compliance report.

[0186] The case agent outputs preliminary opinions through the following steps:

[0187] Using query prompt word templates, query questions are generated from semantic entity pairs;

[0188] Retrieve multiple most similar expert causes from the expert cause database based on the embedding vector of the query question;

[0189] Based on the query question, multiple most similar expert case reasons, and semantic entity pairs, a prompt word template is constructed to guide the case agent to perform case reasoning and output the case agent's preliminary opinion.

[0190] Specifically, for each extracted semantic entity pair The case agent extracts core elements (such as project type, scale, equipment configuration parameters, etc.) from semantic entity pairs and generates natural language query questions using query question prompt word templates. The query is a simple question asking about the details of the project.

[0191] The query prompt word template is as follows:

[0192] "Based on semantic entities {{ The system generates a concise and clear natural language question for the core elements of each project, which is then used for semantic retrieval in the expert case database.

[0193] For example, the query generated for the semantic entity pair {"type":"transformer","value":"SCB11-800kVA"} is: "Is a transformer of model SCB11 with a capacity of 800kVA in compliance with the specifications? What operational risks might exist?"

[0194] After obtaining the embedding vector of the query question using an embedding model, a semantic search based on cosine similarity is performed in a vectorized expert case database to recall the Top-K most similar historical expert cases. .

[0195] It's important to note that the recalled historical expert case studies are not considered absolute authoritative guidance for the current task. Instead, they are provided to the case agent as contextual learning references to help it conduct in-depth analysis and mimic the structure of expert case studies, outlining the root causes of problems, expert judgment logic, solutions, and final outcomes. The case agent, following the decision-making processes in these experiences, effectively reasons about current engineering data to infer potential risks, optimization opportunities, or applicable solutions for the current project. .

[0196] Finally, the prompt word template for the constructed case agent is as follows:

[0197] "You are an experienced expert in electrical engineering, specializing in referencing and retrieving the most similar historical cases, drawing on their analytical logic, and performing logical verification and compliance checks on the extracted engineering entities. The query is:" The most similar case retrieved is: {{ }};\nThe input engineering semantic entity pairs are: {{ Task: Imitate the expert thinking in the case study, analyze the current project materials, determine whether there are any violations, and output the violations, the legal basis, and preliminary suggestions. Finally, compile a preliminary compliance report.

[0198] If the recalled expert case involves a device that malfunctioned under similar conditions, the case agent can issue a warning about the current device even without sufficient theoretical support. This reasoning model, which relies more on rich experience than conventional rules for opinion generation, greatly expands the practicality and diversity of optimization.

[0199] ② The review agent integrates the two preliminary opinions and performs conflict detection to generate a discussion draft. The rule agent and case agent are then coordinated to analyze, vote on, and revise the discussion draft. The revised preliminary opinions are then returned to the review agent to generate a new discussion draft. The rule agent and case agent are then coordinated to analyze, vote on, and revise the new discussion draft again until all votes are passed or the maximum number of iterations is reached. After integrating the preliminary opinions of the final version, the review agent forms the review conclusion.

[0200] It should be noted that, as Figure 2 As shown, after receiving the initial comments from the rule agent and the case agent, the review agent's first task is to integrate them into a draft discussion document and identify any conflicts. For example, the rule agent might point out a violation of a parameter, while the case agent provides an empirical precedent that is acceptable under specific conditions. The review agent needs to identify this conflict and send it to the rule agent and the case agent for analysis.

[0201] Prompt template for reviewing intelligent agents for:

[0202] "You are the overall person in charge of the substation project review, responsible for coordinating the opinions of the rule-based intelligent agent and the case-based intelligent agent to ensure that the final conclusion is comprehensive, rigorous, and conflict-free. Your tasks are as follows: 1. Opinion integration: Receive and read the preliminary compliance report generated by the rule-based intelligent agent {{ }} and the preliminary compliance report generated by the case intelligence agent {{ 2. Conflict Detection: Carefully compare the two reports to identify direct contradictions (e.g., the rule considers it compliant, the case considers it non-compliant) or potential conflicts (e.g., rule recommendation A, case recommendation B; or the case supplements a significant risk not covered by the rule). 3. Generate Discussion Draft: Output the integrated preliminary draft and detailed conflict explanation. Ensure the output format is strictly JSON.

[0203] Finally, the discussion draft output by the review agent is expressed as follows:

[0204] ,

[0205] in, and These represent the review agent's... Consolidation of opinions and conflict statements in the subsequently generated draft discussion; and Representing the rule-based agent and the case-based agent respectively. Preliminary opinions generated later.

[0206] Furthermore, the review agent coordinates the rule agent and the case agent to iteratively analyze, vote, and revise the draft discussion until all votes pass or the maximum number of iterations is reached, including:

[0207] After receiving the draft discussion, the rule agent and the case agent adjust the prompt word template and then perform voting reasoning. That is, they reason based on the previous original input and the current draft discussion, vote on whether to pass or not, and if not, they provide modification suggestions and reasons, and generate a revised preliminary opinion.

[0208] At this point, for the rule-based agent, the adjusted prompt word template is:

[0209] "You are an electricity compliance review expert, specializing in referencing retrieved entity relationships and rule bases to perform logical verification and compliance checks on extracted engineering entities. You have submitted a preliminary report:\n\n" Now, the review agent has submitted the following draft comments and conflict analysis: Draft comments: Conflict Analysis: The rules for retrieval and recall include: {{ }};\nThe retrieved ontology includes: {{ }};\nThe input engineering semantic entity pairs are: {{ Task: Review the draft comments and conflict analysis to determine if you approve the current draft comments. If not, clearly state the required modifications and specific reasons (refer to the recalled rule base and ontology).

[0210] For the case agent, the adjusted prompt word template is as follows:

[0211] "You are an experienced expert in power engineering, specializing in referencing the most similar historical cases retrieved through case studies, drawing on their analytical logic to perform logical verification and compliance checks on the extracted engineering entities. You have submitted a preliminary report:\n\n" Now, the review agent has submitted the following draft comments and conflict analysis: Draft comments: Conflict Analysis: The query statement is: The most similar case retrieved is: {{ }};\nThe input engineering semantic entity pairs are: {{ Task: Review the draft comments and conflict analysis to determine if you approve the current draft comments. If not, clearly state the required changes and the specific reasons (refer to the most similar case in the recall).

[0212] Preliminary opinions after voting on rule-based intelligent agent reasoning Preliminary opinions after voting on case-based intelligent agent reasoning The two initial opinions are sent together to the review agent. Similar to the previous steps, the review agent integrates the opinions from both sources and identifies any conflicts or complementarities between them. If the review agent identifies that either agent's vote is unsuccessful, or if conflicts still exist, a new draft discussion is generated and sent to both the rule agent and the case agent for further analysis, voting, and revision, until both agents pass the vote or the maximum number of iterations is reached. If both the rule agent and the case agent pass the vote and there are no conflicts, or the maximum number of iterations is reached, the review agent outputs the final draft discussion as the review conclusion and exits the iteration process.

[0213] For example, the maximum number of iterations is set to 5.

[0214] Considering that the rule agent and case agent in step S3 may not always achieve complete consistency, but the maximum number of iterations has been reached, step S4 involves professionally refining the review conclusions based on the multimodal large model using the polishing prompt template. This involves defining clear roles, tasks, and polishing requirements in the polishing prompt template, refining the dynamically introduced review conclusions, and outputting a final review opinion and optimization suggestion report. Polishing requirements include: professional language style, clear structure, emphasis on key points, and objectivity of reserved opinions. This ensures overall consistency in opinions, while areas with significant conflicts are reserved with explanatory notes, resulting in optimization suggestions.

[0215] It should be noted that the template for polishing prompts is as follows:

[0216] "You are the chief reviewer of the substation project and have received voting feedback from the rule-based and case-based agents. After multiple rounds of revisions, the final review conclusion has been reached. Your task is to professionally polish the final version of the review conclusion to conform to the written expression style of a senior power expert.\n\nRequirements:\nLanguage style: Professional, rigorous, clear, and concise. Use formal engineering terminology and avoid colloquial expressions.\nClear structure: The conclusion should be presented in points, with logical progression. First, state the core issues, then explain the basis (rules and / or cases), and finally provide clear and actionable optimization suggestions.\nHighlight key points: Emphasize key risks and mandatory requirements.\nHandling reservations: If there are reservations between agents in the final conclusion, they should be objectively listed in the form of "Remarks" or "Supplementary Explanation," without concealing the differences.\n\nOutput format: Plain text report, without JSON or other markup.\nPlease polish the final conclusion based on the following content:\n{{ Please output the final review comments and optimization suggestions report after polishing:

[0217] It should be noted that step S4 guides the multimodal large model to imitate the style of expert opinions, transforming the internal review conclusions generated by the review agent into a final review opinion and optimization suggestion report that imitates the writing style of power experts, is professional, clear, and has operability.

[0218] Furthermore, to achieve closed-loop feedback and continuous optimization of the intelligent review and opinion optimization system, after generating optimization suggestions, the following steps are also included: taking the experts' feedback and evaluation of the generated optimization suggestions as new expert case reasons, and updating the expert case reason database and prototype knowledge system.

[0219] Specifically, a core driving mechanism is constructed to realize the feedback model of identification and reasoning generation. By introducing experts to provide feedback and evaluation on the final modification opinions, new expert case reasons are constructed, the expert case reason database is updated, and the prototype knowledge system is integrated and updated, thereby driving the intelligent evolution of the entire review process and promoting the power engineering review to a higher standard of automation and intelligence.

[0220] Compared with existing technologies, this embodiment provides a method for generating optimization suggestions for power distribution station engineering documents. By constructing a maintainable prototype knowledge system and an expert case library, it achieves unified semantic parsing and structured information extraction of multimodal data, and uses rule knowledge and engineering logic for automated verification. On this basis, a multi-agent collaborative mechanism is introduced, combined with a multimodal large model, to generate professional, accurate, and operable intelligent review assistance opinions and optimization suggestions. This realizes an intelligent assisted review and optimization suggestion generation system with full-chain capabilities of "identification-reasoning-review-optimization suggestion generation," improving the efficiency and quality of the review process, promoting the intelligent level of power distribution station engineering document review, and providing a replicable intelligent solution for other power engineering review scenarios, thus helping the power industry to fully realize digital transformation.

[0221] The method in this embodiment realizes the automated parsing and structured information extraction of multimodal engineering data. Combined with the automatic verification of rule knowledge and engineering logic, it significantly reduces the time spent on manual review, improves the standardization level and accuracy of the review process, and effectively avoids errors or omissions caused by human negligence, thereby greatly improving the overall efficiency and quality of the review work.

[0222] The method in this embodiment integrates a large power industry model with a multi-agent collaborative mechanism (rule agent, case agent, and review agent) to perform deep reasoning, generating specific and quantifiable suggestions that not only comply with standards but also have high guiding significance. These suggestions cover multiple core dimensions such as technical parameters, equipment selection, and construction technology, providing strong professional support for engineering decisions and ensuring the operability of the optimized solutions.

[0223] The method in this embodiment constructs a maintainable prototype knowledge system and expert case library, transforming implicit and scattered expert experience into explicit and systematic knowledge assets. This not only helps to effectively inherit and reuse valuable expert experience and accelerate the training of new personnel, but also enables the system capabilities to continuously evolve through continuous feedback learning and closed-loop optimization, forming a knowledge ecosystem with autonomous learning and continuous optimization capabilities.

[0224] The method in this embodiment utilizes a multi-agent collaborative review mechanism to efficiently and accurately verify power distribution station engineering documents and generate high-quality, professional optimization suggestions. This not only improves the reliability of the review results but also simulates the advantages of collaborative human expert teams, overcoming the limitations that may exist with a single model. The identification results, reasoning conclusions, and review suggestions are embedded into a structured template output to support human-machine collaboration. Finally, through feedback optimization, a closed-loop mechanism is formed that enables feedback adjustment, autonomous repair, and experience accumulation, propelling power engineering review towards a higher standard of automation and intelligence.

[0225] Example 2

[0226] Another embodiment of the present invention discloses a system for generating optimization suggestions for substation engineering documents, thereby implementing the method for generating optimization suggestions for substation engineering documents in Embodiment 1. The specific implementation of each module is described in the corresponding description in Embodiment 1. Figure 3 As shown, the system includes:

[0227] The prototype knowledge system construction module 101 is used to construct a prototype knowledge system containing an ontology and a rule base based on rule materials from the power industry and using a large language model.

[0228] The expert case library construction module 102 is used to obtain an expert case library based on expert case reasons using an embedding model.

[0229] The project information extraction module 103 is used to analyze the data of the power distribution station project and extract structured semantic entity pairs by combining the prototype knowledge system and using a multimodal large model.

[0230] The review conclusion generation module 104 is used to generate review conclusions by reasoning and verifying semantic entity pairs based on the prototype knowledge system and expert case database through a multi-agent collaborative review mechanism.

[0231] The optimization suggestion generation module 105 is used to professionally refine the review conclusions using a multimodal large model and generate optimization suggestions.

[0232] Furthermore, the system also includes:

[0233] The feedback update module is used to update the expert case library and prototype knowledge system by taking the experts' feedback and evaluation of the generated optimization suggestions as new expert case reasons.

[0234] Since the substation engineering document optimization suggestion generation system of this embodiment and the aforementioned substation engineering document optimization suggestion generation method have similarities and can be mutually referenced, this description is redundant and will not be repeated here. Because this system embodiment shares the same principle as the above method embodiment, it also possesses the corresponding technical effects of the above method embodiment.

[0235] Example 3

[0236] Another embodiment of the present invention discloses an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0237] It should be noted that when the processor executes the computer program, it implements the steps of the substation engineering file optimization suggestion generation method in the method embodiment.

[0238] The electronic device in this embodiment can be a server, a terminal, or any other electronic device with the necessary computing and / or processing capabilities.

[0239] In one specific embodiment, the electronic device includes a processor, a memory, and a communication interface connected via a system bus. The processor provides the necessary computing, processing, and / or control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs, and serves as a working area for runtime, temporarily storing intermediate and final results. The internal memory provides the operating environment for the operating system and computer programs in the non-volatile storage medium. The communication interface of the electronic device is used for network connection and communication with external devices. When the computer program is executed by the processor, it performs the steps of the substation engineering document optimization suggestion generation method in the method embodiment.

[0240] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0241] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A substation engineering file optimization recommendation generation method, characterized by, The method comprises the following steps: Based on the rule materials of the power industry, a prototype knowledge system containing ontology and rule base is constructed using a large language model; Based on the expert case, an embedded model is used to obtain an expert case base; The power distribution station engineering project data is parsed, and combined with the prototype knowledge system, a structured semantic entity pair is extracted using a multi-modal large model; Based on the prototype knowledge system and the expert case base, the semantic entity pair is reasoned and verified through a multi-agent collaborative review mechanism to generate a review conclusion; The multi-agent includes a rule agent, a case agent and a review agent, all of which are based on the same multi-modal large model and are constructed by using prompt word templates containing different role-playing and instruction tasks; the review conclusion is generated, including: the rule agent reasons the semantic entity pair according to the ontology and rule base in the prototype knowledge system; the case agent reasons the semantic entity pair by searching for the most similar historical expert case as the reference content for context learning, and outputs the respective preliminary opinions to the review agent; the review agent integrates and detects the conflicts of the two preliminary opinions to generate a discussion draft, and coordinates the rule agent and the case agent to adjust the prompt word module according to the discussion draft, and then analyzes, votes and revises, and returns the revised preliminary opinions to the review agent to generate a new discussion draft again, and coordinates the rule agent and the case agent to analyze, vote and revise the new discussion draft again, until all votes are passed or the maximum iteration number is reached, and the review agent integrates the final version of the preliminary opinions to form the review conclusion; The review conclusion is polished using a multi-modal large model to generate an optimization suggestion.

2. The power distribution station engineering file optimization suggestion generation method of claim 1, wherein, The rule agent outputs the preliminary opinion by the following steps: According to the semantic similarity of the semantic entity pair and the ontology and rule base in the prototype knowledge system, the most similar ontology and rules are recalled; According to the semantic entity pair, the similar ontology and rules, a prompt word template is constructed to guide the rule agent to perform logical verification and output the preliminary opinion of the rule agent.

3. The power distribution station engineering file optimization suggestion generation method of claim 1, wherein, The case agent outputs the preliminary opinion by the following steps: Using a query question prompt word template, a query question is generated for the semantic entity pair; according to the embedding vector of the query question, a plurality of most similar expert cases are recalled from the expert case base; According to the query question, the plurality of most similar expert cases and the semantic entity pair, a prompt word template is constructed to guide the case agent to perform case reasoning and output the preliminary opinion of the case agent.

4. The power distribution station engineering file optimization suggestion generation method of claim 1, wherein, The large language model is used to construct a prototype knowledge system containing ontology and rule base, which comprises: Collecting rule materials of the power industry and preprocessing; Using a large language model pre-trained in the power industry, knowledge elements are extracted from the preprocessed rule materials; Based on the extracted knowledge elements, an ontology and a rule base are constructed to obtain a prototype knowledge system; wherein the ontology includes concepts, attributes and hierarchical relationships in the power engineering field, and the rules in the rule base are logical rules represented in the form of IF-THEN.

5. The power distribution station engineering file optimization suggestion generation method of claim 1, wherein, The prototype knowledge system is also fused and updated according to the expert case, including: extracting target entities and logical relationships from the expert case by using a large language model; guiding the large language model to analyze whether the target entities and logical relationships make corrections or supplements to the prototype knowledge system through a prompt word template, and if so, fusing and updating the prototype knowledge system to obtain an updated prototype knowledge system.

6. The power distribution station engineering file optimization suggestion generation method of claim 5, wherein, The fusion and update of the prototype knowledge system include: if the target entity does not exist in the prototype knowledge system, obtaining the concepts and attributes in the target entity and adding them to the ontology of the prototype knowledge system; if the logical relationship does not exist in the prototype knowledge system, converting the logical relationship into a logical rule represented in the form of IF-THEN and adding it to the rule base of the prototype knowledge system; if the logical relationship exists in the prototype knowledge system but is inconsistent with the existing rules, outputting the updated rules, and adding them to the rule base of the prototype knowledge system after being resolved by expert confirmation or a preset priority mechanism.

7. The power distribution station engineering file optimization suggestion generation method of claim 1, wherein, The expert case library is obtained based on the expert case by using an embedding model, including: preprocessing the text of the expert case, slicing each main section according to the preset number of segmentation words, and then extracting transition sections for each main section according to the preset number of overlapping words before and after the main section, and splicing the main section and the corresponding transition section to generate the final section text; extracting the embedding vector of each section of text by using an embedding model and storing it in the expert case library.

8. The power distribution station engineering file optimization suggestion generation method of claim 4, wherein, The knowledge elements are extracted from the preprocessed rule materials, including: identifying and extracting entities in the power engineering field by using a prompt word engineering; extracting the semantic relationships between the entities; extracting rules and constraints represented in logical expressions or pseudo code.

9. The power distribution station engineering file optimization suggestion generation method of claim 1, wherein, The power distribution station engineering project data is parsed, including: obtaining each page of the power distribution station engineering project data; extracting each text block and its layout information in each page of the document by using a document parsing tool; identifying the semantic types of each text block according to the layout information and extracting the text content of each text block.

10. The power distribution station engineering file optimization suggestion generation method of claim 9, wherein, The layout information includes the coordinates, font, color, font size, line height, horizontal alignment, boldness, and italicization of the text block; the semantic types include title, body, table, and legend.

11. The power distribution station engineering file optimization suggestion generation method of claim 1, wherein, The structured semantic entity pairs are extracted by using a multi-modal large model in combination with the prototype knowledge system, including: constructing a prompt word template containing different entity extraction tasks, and using the prototype knowledge system as background knowledge and constraints in the prompt word template to guide the multi-modal large model to perform entity extraction tasks on each text block, and outputting the obtained entities and their values in the format of JSON key-value pairs as semantic entity pairs.

12. The power distribution station engineering file optimization suggestion generation method of claim 11, wherein, The entity extraction tasks include at least one of extracting project meta-information, extracting engineering parameters, identifying device configurations, identifying basic information of drawings, identifying technical requirements of drawings, and identifying standards and specifications.

13. The power distribution station engineering file optimization suggestion generation method of claim 8, wherein, The semantic relations among the entities include: a type, a capacity relation, and a use mode; and the rules and constraints include: an IF-THEN rule, a constraint condition, a calculation formula, and a recommended value range.

14. The power distribution station engineering file optimization suggestion generation method of claim 1, wherein, After the optimization suggestion is generated, the system further comprises: updating the expert case library and the prototype knowledge system by taking feedback evaluation of the generated optimization suggestion by an expert as a new expert case.

15. The power distribution station engineering file optimization suggestion generation method of claim 1, wherein, The professional polishing of the review conclusion by the multi-modal large model is performed on the basis of the defined roles, tasks, and polishing requirements in the polishing prompt word template, the review conclusion dynamically introduced is polished, and a final review opinion and an optimization suggestion report after polishing are output; and the polishing requirements include: professional language style, clear structure, highlighted key points, and objective reservation of opinions.

16. A power distribution station engineering file optimization recommendation generation system characterized by, The system comprises: A prototype knowledge system construction module configured to construct a prototype knowledge system comprising an ontology and a rule library based on rule materials of the power industry and by using a large language model; An expert case library construction module configured to obtain an expert case library based on expert cases and by using an embedding model; A project information extraction module configured to parse project information of a power distribution station and extract structured semantic entity pairs by using a multi-modal large model in combination with the prototype knowledge system; A review conclusion generation module configured to generate a review conclusion by reasoning and verifying semantic entity pairs based on the prototype knowledge system and the expert case library and by using a multi-agent collaborative review mechanism; The multi-agent comprises: a rule agent, a case agent, and a review agent, all of which are based on the same multi-modal large model and are constructed by using prompt word templates comprising different roles and instruction tasks; the generation of the review conclusion comprises: reasoning of the semantic entity pairs by the rule agent based on the ontology and the rule library in the prototype knowledge system; reasoning of the semantic entity pairs by the case agent based on a most similar historical expert case to the current project information as reference content for context learning, and output of respective preliminary opinions to the review agent; opinion integration and conflict detection of the two preliminary opinions by the review agent to generate a discussion draft, analysis, voting, and revision of the rule agent and the case agent based on the discussion draft, return of the revised preliminary opinions to the review agent to generate a new discussion draft again, analysis, voting, and revision of the rule agent and the case agent on the new discussion draft again, and generation of a final version of the preliminary opinions by the review agent after opinion integration to form the review conclusion; An optimization suggestion generation module configured to generate an optimization suggestion by professional polishing of the review conclusion by the multi-modal large model.

17. The power distribution station engineering file optimization suggestion generation system of claim 16, wherein, The system further comprises: A feedback updating module configured to update the expert case library and the prototype knowledge system by taking feedback evaluation of the generated optimization suggestion by an expert as a new expert case.

18. An electronic device, comprising: The electronic device comprises a memory, a processor, and a computer program stored on the memory and capable of running on the processor, and the processor executes the computer program to execute the power distribution station engineering file optimization suggestion generation method according to any one of claims 1-15.

Images