A power design knowledge base construction method fusing multi-modal data and RAG technology

Abstract

The application relates to a power design knowledge base construction method fusing multi-modal data and RAG technology, and belongs to the technical field of power software development. The method comprises the following steps: collecting and information extracting of multi-source heterogeneous original data; constructing a multi-dimensional knowledge element structure containing parameters, specifications and case relationships by classifying, processing and cross-modal associating the data; based on a vector, graph and relational database hybrid storage architecture, respectively realizing efficient semantic vector retrieval, knowledge graph relationship management and business data synchronization; dynamically optimizing the retrieval results, and outputting compliance conclusions and bases by a double-driven reasoning mechanism; and enhancing the retrieval to generate services to ensure that the output content conforms to the specifications. The method realizes systematic management and intelligent application of power design knowledge.

Description

Technical Field

[0001] This invention belongs to the field of power software development technology, specifically relating to a method for constructing a power design knowledge base that integrates multimodal data and RAG technology. Background Technology

[0002] In the field of power engineering design, design work relies on massive amounts of diverse data, including various design specifications, engineering case studies, equipment parameter tables, electrical drawings, and expert experience records. This data collectively constitutes the core knowledge resources supporting the quality and efficiency of power design. As the power industry moves towards intelligent and digital transformation, business scenarios such as substation design, transmission line design, and secondary circuit design place higher demands on the efficiency, accuracy, and relevance of knowledge retrieval. Therefore, building a power design knowledge base capable of efficiently integrating, storing, and applying knowledge has become a key requirement for the industry's development.

[0003] Currently, the field of power design faces several technical bottlenecks in knowledge management and application: Firstly, the integration of multimodal data is challenging. Power design data exhibits significant multimodal characteristics, with text-based specification documents, tabular equipment parameters, graphical electrical drawings, and image-based field records stored in different formats, and there is a lack of effective correlation mechanisms between these different modalities.

[0004] Secondly, the automation and specialization of data processing are insufficient. Existing technologies for processing power design data largely rely on manual intervention. For example, the identification of symbols in electrical drawings requires manual labeling of equipment types, and the extraction and normalization of professional terms in text specifications require manual processing. This is not only time-consuming and labor-intensive, but also prone to data deviation due to human error. At the same time, general data processing tools are difficult to adapt to the professional characteristics of the power field. For example, general natural language processing models cannot accurately identify power-specific terms such as "grounding resistance requirements" and "insulation level standards," and general image recognition tools cannot distinguish the differences in electrical symbols for different voltage levels. As a result, the processed data cannot meet the professional requirements of the knowledge base.

[0005] Third, the knowledge storage architecture suffers from poor adaptability. Traditional power design knowledge storage often employs a single database format, such as using relational databases to store tabular data and file servers to store drawings and text. This fails to meet the storage needs of different data modalities—relational databases struggle to store the semantic information of drawings and the vector features of text, while file servers cannot achieve rapid data retrieval and related calls. Furthermore, the existing storage architecture lacks knowledge update and version control mechanisms. When new design specifications are released or engineering cases are updated, it is difficult to quickly integrate the new knowledge into the existing knowledge base, causing the knowledge base content to lag behind industry development.

[0006] Fourth, the knowledge application service capability is weak. Current power design knowledge applications are mostly limited to simple keyword searches, unable to achieve cross-modal joint queries. For example, when searching for a design scheme for a certain type of equipment, it is impossible to simultaneously obtain the corresponding standard clauses and similar cases. In terms of knowledge reasoning and generation, there is a lack of intelligent reasoning capabilities that combine power industry rules. Compliance verification of design schemes requires manual comparison with standards, which is inefficient and prone to missing key clauses. At the same time, existing generative technologies do not fully integrate power industry knowledge, and the output design suggestions often have problems such as inconsistencies with industry standards and unprofessional parameter descriptions, making it difficult to directly support design decisions. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides a method for constructing a power design knowledge base that integrates multimodal data and RAG technology. The objective of this invention can be achieved through the following technical solutions:

[0008] A method for constructing a power design knowledge base that integrates multimodal data and RAG technology includes:

[0009] S1: Recognize the high-frequency file format of power design, extract core information from the file based on the preset intelligent mapping network architecture of heterogeneous power data fields, and complete the collection of multi-source heterogeneous raw data;

[0010] S2: Perform full-process multimodal data processing, including classification, specialized processing, and cross-modal association, on the multi-source heterogeneous raw data to form structured knowledge information; based on the design parameters-specification clauses-case feature association rules, construct cross-modal mapping relationships between multi-source modal data to generate a multidimensional associated knowledge element structure;

[0011] S3: The structured knowledge information is stored; the vector database generates semantic vectors through a bidirectional encoder representation model and combines product quantization indexing technology to achieve efficient retrieval of semantic vectors from text and drawings; the graph database constructs a multidimensional relational knowledge graph based on the knowledge triplet of design scheme-design specification-typical case, and completes the implicit relationships between entities through preset association rules; the relational database divides data tables based on power design business scenarios and establishes a real-time data synchronization mechanism with the graph database.

[0012] S4: The hybrid search engine integrates semantic vector search and power keyword search. Based on a dynamic weighting algorithm, it adjusts the weight ratio of the search according to the query type and outputs search results ranked by relevance. The knowledge reasoning service constructs a dual-drive reasoning mechanism of rule engine and graph neural network to collaboratively output reasoning conclusions and evidence. The search enhancement generation service integrates a large language model, based on customized dynamic prompt templates, and outputs compliant generated results through compliance verification plugins.

[0013] As a preferred embodiment of the present invention, the method for recognizing the high-frequency file format of power design is as follows:

[0014] Based on a two-stage architecture of intelligent pre-classification and accurate parsing, the system scans file format identification information using a dynamic feature fingerprint database and combines this with a format classification model to predict file type; based on the prediction result, a differentiated parsing process is triggered.

[0015] By invoking the graphic semantic enhancement algorithm and the primitive recognition algorithm, the graphic information of the drawing file is separated, and the association logic between graphic elements is automatically constructed through the topology relationship reconstruction engine. Based on the intelligent semantic tree parsing system, the title level, clause number, and table reference relationship are identified through the fusion technology of natural language processing and knowledge graph, and a structured knowledge network of the document content is constructed.

[0016] Based on the built-in adaptive table understanding module of the power industry-specific parser, the semantic relationships of cells are dynamically analyzed to extract and model the relationships between row headers, column headers and corresponding cell data.

[0017] Specifically, the intelligent mapping network architecture for heterogeneous fields of power data includes:

[0018] The field feature learning model based on dynamic attention mechanism extracts features from core fields in the power design field to construct a standard power data field feature library. Through reinforcement learning dynamic programming algorithm, it intelligently matches the original fields extracted from the file format with the standard fields in the standard power data field feature library, and optimizes the field correspondence through simulated game process. Based on multimodal cross-validation technology, it pairs the content of the equipment specification field in the table with the equipment model description in the text, and performs consistency checks by combining multimodal information such as drawing annotations and voice annotations.

[0019] Specifically, the process of classifying the multi-source heterogeneous raw data includes:

[0020] Based on a cross-modal recognition model, the original data is initially screened for intelligent modality, automatically classifying drawing, text, table, and image modal data into corresponding data groups, and optimizing the classification accuracy through a dynamic weighting algorithm. Combined with the power design business scenario, a secondary semantic clustering is performed on the initially classified data using domain knowledge graphs and self-attention mechanisms. A semi-supervised annotation system based on generative adversarial networks is constructed to automatically generate scene labels containing multi-dimensional information such as business scenarios, data features, and application requirements, and a dynamic label update mechanism is established.

[0021] Specifically, the method for specialized processing of drawing data in S2 is as follows:

[0022] Based on dynamic graphic recognition algorithms, three-dimensional spatial patterns are extracted from multi-dimensional graphic elements such as electrical symbols, equipment layout locations, and line routes in drawings. The extracted multi-dimensional graphic elements are compared with a dynamic standard graphic library in the field of power design based on knowledge graphs, and combined with equipment operation scenarios, to achieve deep semantic analysis of symbols.

[0023] By extracting information from parameter annotations, a parameter compliance prediction network is constructed. Based on the parameter requirements in power design specifications, equipment operation data under simulated operating conditions are used to predict potential risks caused by parameter annotations and dynamically mark parameter annotations with risks. The processed graphic information and parameter information are organized into structured drawing data with relationships through graph database technology, and knowledge graph nodes containing design intent, technical specification basis, and potential optimization suggestions are generated.

[0024] Specifically, the generation process of the multidimensional related knowledge element structure is as follows:

[0025] A knowledge element intelligence core framework is constructed based on a dynamic semantic network. Through cross-modal semantic fusion technology, key information such as equipment model, engineering stage, and design indicators is extracted to establish a dynamic correlation graph between parameters. Based on natural language processing and intelligent retrieval technologies, corresponding design specification clauses are matched and embedded. Through augmented reality and virtual reality technologies, engineering case summaries are provided, and links to 3D visualized case demonstrations are generated. Based on blockchain technology, the original files are hashed and stored, and the file type and storage location are marked. Through machine learning algorithms, in-depth data analysis is performed to generate parameter recommendations, specification application suggestions, and decision support information.

[0026] Specifically, the method for generating semantic vectors from the vector database is as follows:

[0027] Based on the dynamic evolutionary bidirectional encoder representation model, iterative pre-training is performed in the power design field using reinforcement learning mechanism to achieve adaptive learning of the bidirectional encoder representation model in the power field. Based on the attention mechanism and cross-modal fusion technology, the bidirectional encoder representation model is used to extract deep semantic understanding and multi-dimensional features of the input content to generate the semantic vector. Based on the generated semantic vector, a multi-level semantic clustering network is constructed using the product quantization indexing technology. The multi-level semantic clustering network divides semantically similar vectors and realizes parallel retrieval of similar semantic data based on the quantum state superposition principle.

[0028] Specifically, the process of completing the implicit relationships between entities in the graph database includes:

[0029] Based on the power design association rule base, the graph neural network is used to perform rule matching on existing entities in the knowledge graph; semantic similarity between entities is calculated based on node embedding vectors to check the relationship between entities; based on the reinforcement learning model, combined with the current technological development trends and policy changes in the power industry, the operation and maintenance requirements corresponding to equipment models are extracted from the structured knowledge information to construct the association between high-voltage circuit breakers and operation and maintenance requirements; based on the completion of implicit relationships, the knowledge graph is updated through version control and change auditing mechanisms, and blockchain technology is used to store the update records to achieve traceability of entity relationships.

[0030] Specifically, the method for partitioning data tables and establishing a real-time data synchronization mechanism in the relational database is as follows:

[0031] Customized tables are built based on power design business scenarios. Standardized data table templates are established according to the core business scenarios of substation design, transmission line design, and distribution system design, and the filling specifications of basic and professional fields for each scenario are clarified. Data correlation is enhanced through a semantic tagging system, and semantic retrieval and linkage analysis of cross-scenario data are completed based on the common features of semantic tags.

[0032] Based on the graph database, a dual synchronization mode of event-driven and intelligent prediction is constructed. Based on the dual synchronization mode, when the content of the relational database field is updated, potential data changes are predicted through historical data, and the entity structure of the graph database is optimized in advance. Based on version traceability chain technology, data change logs are recorded, and attribute contradictions between the relational database and the graph database are identified and resolved through conflict resolution algorithms.

[0033] Specifically, the dynamic weighting process of the hybrid search engine includes:

[0034] By using natural language processing and multimodal information analysis, we can perform deep scenario identification of user query needs, construct a three-dimensional demand model that includes query intent, application scenario, and data type, and complete the query type determination. Based on knowledge graph reasoning technology, we can associate standard names, clause terms and power professional knowledge networks to increase the retrieval weight of power keywords, and mine implicit associations through graph path analysis to complete intelligent navigation across standards.

[0035] Based on the style transfer model, the semantic vector retrieval weight is enhanced and the design concepts input by users are analyzed to generate multiple versions of semantically enhanced query statements and expand the matching range of similar solutions. Based on the fusion strategy of dynamically adjusting keyword retrieval and semantic vector retrieval, the retrieval results are mapped to the three-dimensional demand model for multi-dimensional scoring. Combined with generative summarization technology, intelligent previews are generated for the results, and value-added services such as association specification interpretation and similar solution comparison are provided.

[0036] Specifically, the knowledge reasoning service, through the process of constructing a dual-driven reasoning mechanism of rule engine and graph neural network, includes:

[0037] The rule engine compares the user's reasoning requirements with the mandatory clauses of the sub-rule library by calling the sub-rule library corresponding to the design scenario, and outputs a preliminary compliance judgment. The graph neural network performs reasoning based on the entity association relationship of substations in the knowledge graph, and combines it with a spatiotemporal dynamic data enhancement model to access dynamic data of meteorological and geological changes in real time, and analyzes the similarity between the scheme and compliance cases and the fit of the regulatory clauses. Based on multimodal evidence chain aggregation technology, the logical judgment of the rule engine is fused with the semantic reasoning results of the graph neural network, and the reasoning logic basis is generated through a visualized decision tree, outputting the reasoning conclusion and basis with confidence score.

[0038] Specifically, the prompt template and verification of the enhanced search generation service include:

[0039] Driven by a dual engine of dynamic scene perception and intelligent semantic adaptation, the system generates customized dynamic prompt templates. By analyzing user input features and combining them with the knowledge graph of the power design stage, it calls the architecture generation mode to output a structured solution framework that includes a deduction diagram of the overall project layout and a recommendation matrix for the selection of major equipment.

[0040] The compliance verification is based on a reinforcement learning feedback mechanism to build an adversarial verification environment that includes a positive specification library and a negative error case library. It uses a multimodal comparison model to perform semantic similarity calculation and logical contradiction detection on the generated results and design specifications. Through risk heat map visualization technology, a risk heat map is generated to display the risk distribution and impact range.

[0041] The beneficial effects of this invention are as follows:

[0042] Through a two-stage architecture of intelligent pre-classification and precise parsing, the power industry-specific parser can automatically identify high-frequency design file formats such as drawings, text, and tables, eliminating the need for manual format conversion. The differentiated parsing link can specifically handle data of different modalities, avoiding the fragmented extraction problem of general parsing tools, and significantly improving data collection efficiency compared to traditional manual collection. The intelligent mapping network architecture for heterogeneous fields in power data constructs a standard field feature library through a dynamic attention mechanism, optimizes field matching relationships by combining reinforcement learning dynamic programming algorithms, and then performs multimodal cross-validation, effectively solving the problems of field mismatch and information contradiction.

[0043] For drawing data, a 3D spatial model is used to extract graphic elements. Combined with dynamic standard graphic library comparison and equipment operation scenario analysis, deep semantic parsing of electrical symbols is achieved. The parameter compliance prediction network, based on design specifications and simulated operating condition data, predicts parameter annotation risks in advance, avoiding subsequent design rework and reducing engineering hazards caused by parameter errors. Through the association rules of design parameters, specification clauses, and case features, a multimodal data mapping relationship is constructed. The generated multidimensional associated knowledge element structure integrates equipment parameters, specification clauses, AR / VR 3D cases, and blockchain-stored information, eliminating the need for manual cross-platform retrieval and significantly improving knowledge acquisition efficiency. Based on cross-modal recognition models and secondary semantic clustering of domain knowledge graphs, data can be automatically classified according to business scenarios such as substation design and transmission line design. Multidimensional scene labels are generated through a semi-supervised annotation system of generative adversarial networks, enabling rapid location of scene-specific data during subsequent calls and reducing data management costs.

[0044] The vector database uses a dynamic evolutionary bidirectional encoder representation model to generate semantic vectors. Combined with a multi-level semantic clustering network and the quantum state superposition principle, it achieves parallel retrieval, significantly improving retrieval speed compared to traditional single-database systems, making it particularly suitable for rapid access to massive amounts of design data. Product quantization indexing technology further compresses vector storage space, reducing hardware resource consumption. The graph database completes implicit entity relationships through a power design association rule base and reinforcement learning model, avoiding the tedious manual completion of relationships. The relational database's scenario-specific tables and event-driven + intelligent prediction synchronization mode enable real-time data linkage with the graph database. The graph database's version control and change auditing mechanisms, along with blockchain hash notarization of the knowledge element structure, completely record the generation, update, and change trajectory of knowledge data. Subsequent verification allows for precise tracing of data sources and reasons for changes, meeting the power design industry's compliance requirements for data traceability and auditability, and avoiding design disputes caused by unclear data sources.

[0045] The hybrid search engine's three-dimensional demand model can deeply identify user query intent, and the dynamic weighting algorithm adjusts search weights accordingly. Knowledge graph navigation and generative summary services also provide interpretation of related specifications and comparison of similar solutions, reducing users' manual filtering time. The knowledge reasoning service's rule engine + graph neural network dual-drive mechanism ensures mandatory clause verification through sub-rule bases while enabling flexible reasoning by combining spatiotemporal dynamic data. Multimodal evidence chain aggregation and visualized decision trees clearly display the reasoning logic, helping designers understand the reasoning basis and reducing decision-making risks. Attached Figure Description

[0046] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0047] Figure 1This is a flowchart illustrating a method for constructing a power design knowledge base that integrates multimodal data and RAG technology according to the present invention.

[0048] Figure 2 A flowchart for specialized processing of drawing data in this invention;

[0049] Figure 3 This is a flowchart illustrating the dynamic weighting process of the hybrid search engine of the present invention. Detailed Implementation

[0050] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0051] Please see Figure 1 A method for constructing a power design knowledge base that integrates multimodal data and RAG technology includes:

[0052] S1: Recognize the high-frequency file format of power design, extract core information from the file based on the preset intelligent mapping network architecture of heterogeneous power data fields, and complete the collection of multi-source heterogeneous raw data;

[0053] S2: Perform full-process multimodal data processing, including classification, specialized processing, and cross-modal association, on the multi-source heterogeneous raw data to form structured knowledge information; based on the design parameters-specification clauses-case feature association rules, construct cross-modal mapping relationships between multi-source modal data to generate a multidimensional associated knowledge element structure;

[0054] S3: The structured knowledge information is stored; the vector database generates semantic vectors through a bidirectional encoder representation model and combines product quantization indexing technology to achieve efficient retrieval of semantic vectors from text and drawings; the graph database constructs a multidimensional relational knowledge graph based on the knowledge triplet of design scheme-design specification-typical case, and completes the implicit relationships between entities through preset association rules; the relational database divides data tables based on power design business scenarios and establishes a real-time data synchronization mechanism with the graph database.

[0055] S4: The hybrid search engine integrates semantic vector search and power keyword search. Based on a dynamic weighting algorithm, it adjusts the weight ratio of the search according to the query type and outputs search results ranked by relevance. The knowledge reasoning service constructs a dual-drive reasoning mechanism of rule engine and graph neural network to collaboratively output reasoning conclusions and evidence. The search enhancement generation service integrates a large language model, based on customized dynamic prompt templates, and outputs compliant generated results through compliance verification plugins.

[0056] Specifically, the method for recognizing the high-frequency file format of power design is as follows:

[0057] Based on a two-stage architecture of intelligent pre-classification and accurate parsing, the system scans file format identification information using a dynamic feature fingerprint database and combines this with a format classification model to predict file type; based on the prediction result, a differentiated parsing process is triggered.

[0058] By invoking the graphic semantic enhancement algorithm and the primitive recognition algorithm, the graphic information of the drawing file is separated, and the association logic between graphic elements is automatically constructed through the topology relationship reconstruction engine. Based on the intelligent semantic tree parsing system, the title level, clause number, and table reference relationship are identified through the fusion technology of natural language processing and knowledge graph, and a structured knowledge network of the document content is constructed.

[0059] Based on the built-in adaptive table understanding module of the power industry-specific parser, the semantic relationships of cells are dynamically analyzed to extract and model the relationships between row headers, column headers and corresponding cell data.

[0060] This embodiment uses the identification process of three types of high-frequency documents (electrical drawing documents, design specification text documents, and equipment parameter table documents) in the substation design scenario when a provincial power design institute builds a power design knowledge base as an example to explain in detail the workflow of the power industry-specific parser.

[0061] Based on the predicted results of the "drawing file", a dedicated parsing link for graphic files is triggered (taking the 110kV substation main connection DWG drawing as an example):

[0062] Graphic information separation: The parser calls the graphic semantic enhancement algorithm and the primitive recognition algorithm to perform layer parsing on the DWG drawing, and automatically separates the electrical symbol layer (including circuit breaker, disconnector, current transformer, and voltage transformer symbols), the line layer (including 10kV bus, 35kV bus, and main transformer connection line), and the annotation layer (including equipment model annotation "ZN28-12" and line voltage annotation "110kV"), while excluding auxiliary layers in the drawing (such as grid lines and annotation text boxes).

[0063] Association Logic Construction: Through the topology relationship reconstruction engine, the spatial location and connection relationship of the separated graphic elements are analyzed—identifying the line connection path of "110kV incoming line terminal - disconnecting switch - circuit breaker - current transformer - high voltage side of main transformer", automatically constructing the association logic of "equipment symbol - line route - voltage level", and generating structured graphic data: "disconnecting switch (symbol number S1) → circuit breaker (symbol number QF1) → connecting line (voltage 110kV) → main transformer (symbol number T1)".

[0064] Based on the predicted result of "text file", a dedicated parsing link for text files is triggered (taking the PDF file of "110kV-750kV Substation Design Specification" as an example):

[0065] Structure Recognition: The parser calls the intelligent semantic tree parsing system, combining natural language processing technology and power design knowledge graph to perform chapter structure analysis on the PDF file—automatically identifying the first-level heading "Chapter 4 Site Selection and General Layout", the second-level heading "4.2 General Layout", and the third-level heading "4.2.3 Equipment Layout Spacing Requirements", and extracting the clause numbers under each heading (such as "4.2.3-1" and "4.2.3-2").

[0066] Knowledge Network Construction: Through knowledge graph fusion technology, the content of clauses and their citation relationships are linked. The requirement of "minimum distance between main transformer and building" in clause "4.2.3-1" is identified, and the table "Table 4-3 Limits for distance between equipment and building" referenced by the clause is located. A structured knowledge network of "chapter title - clause number - clause content - referenced table" is automatically constructed, transforming the scattered text content into knowledge nodes with clear hierarchy and explicit association.

[0067] Based on the prediction results of the "table file", a dedicated parsing link for table-type files is triggered (taking the XLSX table of substation main equipment parameters as an example):

[0068] Cell semantic association analysis: The parser calls the built-in adaptive table understanding module to perform semantic analysis on the header and content of the XLSX table—identifying the row titles as "Equipment Type", "Rated Capacity", "Rated Voltage" and "Short Circuit Breaking Current", and the column titles as "Main Transformer 1", "Main Transformer 2", "110kV Circuit Breaker 1" and "110kV Circuit Breaker 2", and determining the cell correspondence between "Main Transformer 1 - Rated Capacity" and "110kV Circuit Breaker 1 - Short Circuit Breaking Current" through semantic matching.

[0069] Relationship Modeling and Data Extraction: Automatically extract data from each cell and establish a relationship model of "Equipment Name - Parameter Item - Parameter Value" - such as "Main Transformer 1 - Rated Capacity - 50MVA", "Main Transformer 1 - Rated Voltage - 110kV / 10kV", "110kV Circuit Breaker 1 - Short Circuit Breaking Current - 31.5kA". At the same time, cells with missing data (such as "Main Transformer 2 - Short Circuit Breaking Current") are marked with "To be supplemented" to ensure the structure and completeness of parameter data.

[0070] Through the above implementation process, the power industry-specific parser can automatically and professionally identify and parse high-frequency power design documents, and obtain structured graphic, text, and tabular data without manual intervention, providing high-quality raw data support for subsequent multimodal data processing and knowledge base construction.

[0071] Specifically, the intelligent mapping network architecture for heterogeneous fields of power data includes:

[0072] The field feature learning model based on dynamic attention mechanism extracts features from core fields in the power design field to construct a standard power data field feature library. Through reinforcement learning dynamic programming algorithm, it intelligently matches the original fields extracted from the file format with the standard fields in the standard power data field feature library, and optimizes the field correspondence through simulated game process. Based on multimodal cross-validation technology, it pairs the content of the equipment specification field in the table with the equipment model description in the text, and performs consistency checks by combining multimodal information such as drawing annotations and voice annotations.

[0073] Specifically, the process of classifying the multi-source heterogeneous raw data includes:

[0074] Based on a cross-modal recognition model, the original data is initially screened for intelligent modality, automatically classifying drawing, text, table, and image modal data into corresponding data groups, and optimizing the classification accuracy through a dynamic weighting algorithm. Combined with the power design business scenario, a secondary semantic clustering is performed on the initially classified data using domain knowledge graphs and self-attention mechanisms. A semi-supervised annotation system based on generative adversarial networks is constructed to automatically generate scene labels containing multi-dimensional information such as business scenarios, data features, and application requirements, and a dynamic label update mechanism is established.

[0075] In this embodiment, taking the classification of design data for a newly constructed 500kV substation project as an example, the specific implementation process is as follows:

[0076] The raw data collected by the engineering team included various types of documents such as CAD format substation electrical main wiring diagrams, PDF version of "DL / T5352-2018 High Voltage Power Distribution Equipment Design Code", Excel format main transformer equipment parameter tables, and on-site geological survey photos. The cross-modal recognition model first activated the ResNet-BERT hybrid architecture to extract document features: the image feature extraction module identified vector graphic features of the CAD drawings and pixel texture features of the photos, while the text feature extraction module parsed the character sequence information of the PDF and Excel files.

[0077] After initial classification, the system found two CAD files misclassified as image files (due to interference from embedded preview images). The dynamic weighting algorithm then adjusted the classification parameters, increasing the weight of graphic topological features from 0.5 to 0.7 and decreasing the weight of pixel features from 0.3 to 0.1. After reclassification, all files were accurately assigned to the drawing group (12 CAD files), the text group (8 PDF specifications), the table group (5 Excel parameter tables), and the image group (20 on-site photos), improving the classification accuracy from 92% in the initial screening to 99%.

[0078] Based on a knowledge graph in the power design field (containing 23 core entity relationships such as "substation design - equipment selection - standards and specifications"), the system further segments the initial screening data according to business scenarios:

[0079] The drawing group uses a self-attention mechanism to focus on key features such as "wiring method" and "equipment layout", and clusters 12 CAD drawings into three subcategories: "main transformer system design drawing", "GIS power distribution device drawing" and "grounding grid layout drawing". Among them, the grounding grid layout drawing is automatically associated with the "grounding resistance requirement" entity in the knowledge graph.

[0080] Based on the hierarchical relationship of clause numbers (such as "3.2.1 Conductor Selection"), and combined with the self-attention mechanism to enhance the recognition of professional terms such as "short-circuit current" and "insulation coordination", the PDF specification documents are clustered into "Equipment Selection Specification", "Electrical Safety Specification" and "Construction Acceptance Specification".

[0081] The table group uses semantic analysis of the table header to associate the parameter table with the corresponding equipment entity. For example, the "Main Transformer Parameter Table" matches the "Oil-immersed Transformer" node in the knowledge graph, and the "GIS Equipment Parameter Table" is associated with the "Gas Insulated Switchgear" entity.

[0082] Generative adversarial networks use 10 manually labeled typical datasets as seed samples to automatically generate multi-dimensional labels:

[0083] Generate tags for "Grounding Grid Layout Diagram": Business Scenario = Substation Grounding System Design, Data Characteristics = Horizontal Grounding Electrode Topology + Vertical Grounding Electrode Spacing, Application Requirements = Grounding Resistance Verification;

[0084] For Section 3.4 of DL / T 5352-2018, the following tags are generated: Business Scenario = Power Distribution Equipment Selection, Data Characteristics = Short-Circuit Current Withstand Value + Dynamic and Thermal Stability Verification, Application Requirements = Equipment Parameter Compliance Review.

[0085] When the engineering team imported the newly released "GB 50169-2024 Code for Construction and Acceptance of Grounding Devices", the dynamic label update mechanism automatically triggered label system iteration by monitoring the version information in the code number: the code basis field of the original "grounding design" related labels was expanded from "DL / T 5352-2018" to multiple version associations, and a new feature label "updated soil resistivity testing method" was added. The system also supplemented the labels of all grounding-related drawings and parameter tables with the application requirement prompt "must comply with Section 4.2 of GB 50169-2024" through knowledge graph association.

[0086] Through this classification process, the data processing work that originally required 3 engineers and 2 days to complete can now be completed automatically by the system in just 1.5 hours, laying a high-quality data foundation for subsequent knowledge extraction and storage.

[0087] Specifically, the method for specialized processing of drawing data in S2 is as follows:

[0088] Based on dynamic graphic recognition algorithms, three-dimensional spatial patterns are extracted from multi-dimensional graphic elements such as electrical symbols, equipment layout locations, and line routes in drawings. The extracted multi-dimensional graphic elements are compared with a dynamic standard graphic library in the field of power design based on knowledge graphs, and combined with equipment operation scenarios, to achieve deep semantic analysis of symbols.

[0089] By extracting information from parameter annotations, a parameter compliance prediction network is constructed. Based on the parameter requirements in power design specifications, equipment operation data under simulated operating conditions are used to predict potential risks caused by parameter annotations and dynamically mark parameter annotations with risks. The processed graphic information and parameter information are organized into structured drawing data with relationships through graph database technology, and knowledge graph nodes containing design intent, technical specification basis, and potential optimization suggestions are generated.

[0090] Specifically, the generation process of the multidimensional related knowledge element structure is as follows:

[0091] A knowledge element intelligence core framework is constructed based on a dynamic semantic network. Through cross-modal semantic fusion technology, key information such as equipment model, engineering stage, and design indicators is extracted to establish a dynamic correlation graph between parameters. Based on natural language processing and intelligent retrieval technologies, corresponding design specification clauses are matched and embedded. Through augmented reality and virtual reality technologies, engineering case summaries are provided, and links to 3D visualized case demonstrations are generated. Based on blockchain technology, the original files are hashed and stored, and the file type and storage location are marked. Through machine learning algorithms, in-depth data analysis is performed to generate parameter recommendations, specification application suggestions, and decision support information.

[0092] In this embodiment, taking the selection and design of the main transformer of a 220kV smart substation as an example, the specific generation process of the multi-dimensional related knowledge element structure is as follows:

[0093] The core framework of the dynamic semantic network first performs cross-modal semantic fusion on multi-source data: extracting the "SFZ11-240000 / 220" main transformer model and "three-phase dual-winding" wiring method from the CAD format main transformer system design drawing; parsing 28 core parameters such as rated capacity 240MVA, short-circuit impedance 14%, and cooling method "forced oil circulation air cooling" from the Excel parameter table; and obtaining the project progress information of the "equipment selection stage" from the construction schedule table of the project management system. The correlation between "capacity-impedance-cooling method" is strengthened through a self-attention mechanism to construct a dynamic correlation graph, where the rated capacity node is automatically associated with historical operating data of "load rate 75%", and the short-circuit impedance node is associated with the system parameter constraint of "110kV side short-circuit current 25kA".

[0094] The natural language processing module performs deep analysis of the text library, accurately matching the main transformer parameters with the relevant standard clauses: based on Article 4.2.3 of "DL / T 574-2010 Guidelines for the Operation and Maintenance of Transformer Tap Changers", the maintenance requirement of "insulation test should be performed on the non-excitation tap changer every 5 years" is embedded in the tap changer parameter node; through intelligent retrieval technology, Article 6.3 of "GB1094.1-2013 Power Transformers Part 1: General Rules" is located, and the insulation requirement of "temperature rise limit of 105K" is associated with the winding material parameters; based on the characteristics of the cooling method, the mandatory clause of "forced oil circulation air cooling system should have dual power supply switching function" in "DL / T 572-2010 Operating Procedures for Power Transformers" is automatically matched, forming a chain of standard basis.

[0095] The system utilizes augmented reality and virtual reality engines to generate two types of visualization resources: Three similar 220kV substation main transformer installation cases are selected from the case library, generating labeled AR demonstration links. Clicking these links allows users to view 3D annotations of the "safe distance between the 240MVA main transformer and the firewall"; VR technology recreates the entire process of "on-site installation of SFZ11 series main transformers" in a provincial power grid, highlighting 3D operation guidance for key procedures such as "precision control of transformer placement." Each case link is associated with a project overview summary, including key indicators such as "commissioning time: 2020" and "average annual load rate: 68%", providing an intuitive presentation of design references.

[0096] A consortium blockchain architecture is used for hash-based notarization of core data: original CAD drawings generate unique hash values, stored in a location labeled "Power Design Cloud Platform / 220kV Substation Project / Electrical Primary / Version 3"; parameter table files are associated with a blockchain timestamp "2025-09-15 14:30:22", recording the designer's digital signature; matching specification documents form a version chain, clearly defining the revision relationship of "DL / T 574-2010" replacing "DL / T 574-1995", with a special note on the revision point of "adding content on dry-type on-load tap changers". All notarized information is made immutable through smart contracts, supporting traceability and auditing throughout the design process.

[0097] The decision-making model is trained based on historical data from 50 similar projects, and outputs multi-dimensional auxiliary information: a capacity recommendation curve is generated through regression analysis, suggesting that "240MVA (recommendation rate 92%) should be selected first, and 180MVA (recommendation rate 78%) should be selected as an alternative" under the current load conditions; regarding the short-circuit impedance parameter, combined with the system short-circuit current calculation, it is suggested that "14% impedance value can meet the current limiting requirements, but the breaking capacity of the 110kV side circuit breaker needs to be verified"; the specification application suggestion module specifically reminds that "the selection of tap changers in this project must comply with the newly added requirements of online oil filtration devices in DL / T 574-2010", and the maintenance process reference path in Appendix H is marked.

[0098] The knowledge metastructure generated through this process enables a complete link between equipment parameters and operation and maintenance specifications. When designers query a parameter, they can simultaneously obtain related specifications, 3D case studies, historical data, and decision-making suggestions, reducing the data integration work that originally took 6 hours to 45 minutes.

[0099] Specifically, the method for generating semantic vectors from the vector database is as follows:

[0100] Based on the dynamic evolutionary bidirectional encoder representation model, iterative pre-training is performed in the power design field using reinforcement learning mechanism to achieve adaptive learning of the bidirectional encoder representation model in the power field. Based on the attention mechanism and cross-modal fusion technology, the bidirectional encoder representation model is used to extract deep semantic understanding and multi-dimensional features of the input content to generate the semantic vector. Based on the generated semantic vector, a multi-level semantic clustering network is constructed using the product quantization indexing technology. The multi-level semantic clustering network divides semantically similar vectors and realizes parallel retrieval of similar semantic data based on the quantum state superposition principle.

[0101] Specifically, the process of completing the implicit relationships between entities in the graph database includes:

[0102] Based on the power design association rule base, the graph neural network is used to perform rule matching on existing entities in the knowledge graph; semantic similarity between entities is calculated based on node embedding vectors to check the relationship between entities; based on the reinforcement learning model, combined with the current technological development trends and policy changes in the power industry, the operation and maintenance requirements corresponding to equipment models are extracted from the structured knowledge information to construct the association between high-voltage circuit breakers and operation and maintenance requirements; based on the completion of implicit relationships, the knowledge graph is updated through version control and change auditing mechanisms, and blockchain technology is used to store the update records to achieve traceability of entity relationships.

[0103] In this embodiment, the knowledge graph maintenance of a 220kV substation LW36-126 type SF6 high-voltage circuit breaker is taken as an example. The specific implementation process is as follows:

[0104] The power design rule base pre-sets 128 core rules, including "Equipment Model → Insulation Medium Type → Test Standard" and "Operation and Maintenance Cycle → Equipment Status → Testing Items". A graph neural network scans the existing entities in the knowledge graph: a master node for "LW36-126 circuit breaker" exists, associated with parameters such as rated voltage 126kV and rated current 3150A, and connected to the "SF6 gas" node through the "Use Insulation Medium" relationship, but no association link with the test standard is found. The system automatically triggers a rule matching mechanism, locating rule 42 in the rule base, "SF6 circuit breakers must be associated with the gas tightness test specification," identifying the implicit relationship gap between "LW36-126 circuit breaker" and "DL / T 402-2015 Clause 7.3".

[0105] The TransE algorithm was used to vectorize the nodes in the knowledge graph. The calculated cosine similarity of the embedding vectors of the "LW36-126 circuit breaker" node and the "gas tightness test" node was 0.89 (threshold set at 0.75), confirming a strong semantic association between them. The similarity between the "annual leakage rate ≤0.5%" parameter node and the "environmental protection testing requirements" node reached 0.82, but the current knowledge graph only has one-way data associations and lacks a clear definition of operation and maintenance relationships. Through multiple rounds of similarity verification, the system selected three high-confidence candidate associations that need to be completed: circuit breaker model and gas recovery process, insulating medium and quarterly testing cycle, and operating mechanism and mechanical characteristic test.

[0106] The reinforcement learning model uses "dual-carbon policy requirements" and "SF6 replacement technology trends" as external reward signals: when it detects a policy document from the State Grid stating that "after 2023, it will no longer purchase power distribution equipment containing SF6," it immediately increases the association weight of "environmental protection operation and maintenance requirements." Combining the experience of mixed gas retrofitting of GIS equipment at the Kaifeng substation, it extracts new operation and maintenance content such as "3:7 SF6 / N2 mixed gas ratio detection" from structured data. Through 1000 rounds of simulation training, the model autonomously constructs three types of key associations:

[0107] "LW36-126 Circuit Breaker → Quarterly Gas Composition Analysis" (based on Mixed Gas Operation and Maintenance Specifications).

[0108] "SF6 gas → recovery and treatment process" (related to the latest environmental protection policy requirements).

[0109] "Operating mechanism → Intelligent monitoring terminal installation" (in conjunction with the trend of IoT operation and maintenance technology).

[0110] The knowledge graph adopts a baseline version (V2.0) plus dynamic patching update mechanism to generate a V2.1 version update package, which includes 5 newly added entity relationships and 3 policy basis documents; the changes are hashed and stored through a consortium blockchain to generate a unique identifier, including node ID, association type, update timestamp "2025-09-23 09:47:15" and digital signature of the auditor; the audit log automatically records: "Added 'SF6 gas - annual recovery ≥95%' association, in accordance with Article 5.2 of the 'Sulphur Hexafluoride Gas Power Equipment On-site Recycling Specification'".

[0111] Through this process, eight types of implicit relationships that previously required manual sorting were now fully automated, reducing the knowledge graph's response time to new operation and maintenance requirements from three months to 72 hours. When designers query this type of circuit breaker, the system can automatically display the complete relationship chain from equipment parameters to environmental operation and maintenance, supporting compliance verification and policy adaptability analysis.

[0112] Specifically, the method for partitioning data tables and establishing a real-time data synchronization mechanism in the relational database is as follows:

[0113] Customized tables are built based on power design business scenarios. Standardized data table templates are established according to the core business scenarios of substation design, transmission line design, and distribution system design, and the filling specifications of basic and professional fields for each scenario are clarified. Data correlation is enhanced through a semantic tagging system, and semantic retrieval and linkage analysis of cross-scenario data are completed based on the common features of semantic tags.

[0114] Based on the graph database, a dual synchronization mode of event-driven and intelligent prediction is constructed. Based on the dual synchronization mode, when the content of the relational database field is updated, potential data changes are predicted through historical data, and the entity structure of the graph database is optimized in advance. Based on version traceability chain technology, data change logs are recorded, and attribute contradictions between the relational database and the graph database are identified and resolved through conflict resolution algorithms.

[0115] Specifically, the dynamic weighting process of the hybrid search engine includes:

[0116] By using natural language processing and multimodal information analysis, we can perform deep scenario identification of user query needs, construct a three-dimensional demand model that includes query intent, application scenario, and data type, and complete the query type determination. Based on knowledge graph reasoning technology, we can associate standard names, clause terms and power professional knowledge networks to increase the retrieval weight of power keywords, and mine implicit associations through graph path analysis to complete intelligent navigation across standards.

[0117] Based on the style transfer model, the semantic vector retrieval weight is enhanced and the design concepts input by users are analyzed to generate multiple versions of semantically enhanced query statements and expand the matching range of similar solutions. Based on the fusion strategy of dynamically adjusting keyword retrieval and semantic vector retrieval, the retrieval results are mapped to the three-dimensional demand model for multi-dimensional scoring. Combined with generative summarization technology, intelligent previews are generated for the results, and value-added services such as association specification interpretation and similar solution comparison are provided.

[0118] In this embodiment, taking the query "Seismic Design Specifications and Case References for 220kV Substation GIS Equipment" by an engineer from a power design institute as an example, the specific implementation of the dynamic weighting process of the hybrid search engine is as follows:

[0119] After the engineer inputs the query, the system first performs deep semantic parsing using a BERT-CNN hybrid architecture: extracting dual intent labels "specification query" and "case reference", with intent weights of 0.6 and 0.4 respectively; combining the domain knowledge graph, it matches the scenario chain "substation electrical equipment design → GIS equipment → seismic fortification", with a scenario confidence score of 92%; through multimodal feature analysis, it determines that the required data types are text specifications (accounting for 60%) and drawing cases (accounting for 40%).

[0120] The final three-dimensional requirement model is: {Query intent: Standard consultation + case reference, application scenario: 220kV substation GIS design, data type: text + CAD drawings}. The system determines that the query belongs to the mixed retrieval type of "standard matching + engineering case".

[0121] The knowledge graph reasoning module initiates a bidirectional reasoning mechanism. Starting from the "GIS Equipment" node, it finds "DL / T 5595-2020 Technical Specification for Substation Building Structure Design" along the "Compliance with Standards" relationship, and then links it to "GB 50260-2013 Code for Seismic Design of Power Facilities" through the "Reference Standards" relationship, forming a cross-standard association chain. The weight of professional terms such as "seismic fortification intensity," "GIS equipment layout," and "earthquake condition verification" is increased from the basic value of 0.5 to 0.85. At the same time, the implicit association of "site category → foundation treatment → equipment fixing method" is discovered through graph path analysis, and the keyword weight of "embedded part strength" is automatically increased. Based on the hierarchical relationship of standard clauses, a search navigation path of "general principles → basic design requirements → equipment seismic measures → acceptance standards" is generated, supporting cross-standard clause navigation.

[0122] For the design concepts input by engineers, the style transfer model performs semantic enhancement processing, transforming the original query into three versions: professional expression "method for calculating seismic load on gas-insulated switchgear", engineering expression "seismic support design scheme for GIS equipment", and standard expression "basis for determining the seismic grade of 220kV GIS equipment". Since case retrieval accounts for 40% of the search, the system increases the semantic vector retrieval weight from the default 0.3 to 0.6, and uses a pre-trained Sentence-BERT model from the power industry to generate semantic vectors. Through vector cosine similarity calculation (threshold set to 0.75), it matches the Guangzhou 110kV Lingyun Substation GIS seismic retrofit case (similarity 0.89) and the Lieqiao Substation "Moonlight Treasure Box" seismic design scheme (similarity 0.82).

[0123] The system implements differentiated weighted fusion based on query type. Standardized retrieval adopts "keyword-based + semantic-based" (weight ratio 7:3), while case retrieval adopts "semantic-based + keyword-based" (weight ratio 6:4). The system scores the results from four dimensions: standard matching degree (40%), scenario relevance (30%), update timeliness (20%), and case typicality (10%), generating a comprehensive ranking result.

[0124] The results for the standard category display a summary of Clause 5.3.2 of DL / T 5595-2020, which states that "the seismic fortification intensity of GIS equipment should be increased by 1 degree compared with the basic intensity," and display a "correspondence table between fortification intensity and site category" linked to knowledge graph nodes. The results for the case study category generate a 3D layout thumbnail of the GIS equipment at Lingyun Substation, highlighting the technical feature of "adopting elastic support damping devices," and providing a comparative analysis with the geological conditions of this project. The results also automatically generate a "seismic design compliance checklist," which includes the correspondence between standard clauses and case study measures.

[0125] After implementing dynamic weighting, engineers can obtain the complete knowledge chain from design specifications to engineering practices through a single search, significantly reducing the time cost of cross-platform queries.

[0126] Specifically, the knowledge reasoning service, through the process of constructing a dual-driven reasoning mechanism of rule engine and graph neural network, includes:

[0127] The rule engine compares the user's reasoning requirements with the mandatory clauses of the sub-rule library by calling the sub-rule library corresponding to the design scenario, and outputs a preliminary compliance judgment. The graph neural network performs reasoning based on the entity association relationship of substations in the knowledge graph, and combines it with a spatiotemporal dynamic data enhancement model to access dynamic data of meteorological and geological changes in real time, and analyzes the similarity between the scheme and compliance cases and the fit of the regulatory clauses. Based on multimodal evidence chain aggregation technology, the logical judgment of the rule engine is fused with the semantic reasoning results of the graph neural network, and the reasoning logic basis is generated through a visualized decision tree, outputting the reasoning conclusion and basis with confidence score.

[0128] Specifically, the prompt template and verification of the enhanced search generation service include:

[0129] Driven by a dual engine of dynamic scene perception and intelligent semantic adaptation, the system generates customized dynamic prompt templates. By analyzing user input features and combining them with the knowledge graph of the power design stage, it calls the architecture generation mode to output a structured solution framework that includes a deduction diagram of the overall project layout and a recommendation matrix for the selection of major equipment.

[0130] The compliance verification is based on a reinforcement learning feedback mechanism to build an adversarial verification environment that includes a positive specification library and a negative error case library. It uses a multimodal comparison model to perform semantic similarity calculation and logical contradiction detection on the generated results and design specifications. Through risk heat map visualization technology, a risk heat map is generated to display the risk distribution and impact range.

[0131] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for building a power design knowledge base by fusing multi-modal data with RAG techniques, characterized in that, The method comprises the following steps: S1: identifying the high-frequency file format of power design, collecting multi-source heterogeneous raw data based on the preset power data heterogeneous field intelligent mapping network architecture, and extracting the core information in the file; the method for identifying the high-frequency file format of power design is as follows: Based on the intelligent pre-classification-precise analysis two-stage architecture, the format identification information of the file is scanned through the dynamic feature fingerprint library, and the file type is pre-judged by combining the format classification model; the differential analysis link is triggered based on the pre-judgment result: By calling the graphic semantic enhancement algorithm, the graphic information of the drawing file is separated based on the graphic element recognition algorithm, and the association logic between the graphic elements is automatically constructed through the topological relationship reconstruction engine; based on the intelligent semantic tree analysis system, the title level, clause number and table reference relationship are identified through the natural language processing and knowledge graph fusion technology, and the structured knowledge network of the file content is constructed; Based on the adaptive table understanding module built in the power industry special parser, the semantic association of the cells is dynamically analyzed, and the extraction and relationship modeling of the row title, column title and corresponding cell data are completed; S2: full-process multi-modal data processing of classification, special processing and cross-modal association of the multi-source heterogeneous raw data, forming structured knowledge information; Based on the design parameter-specification clause-case feature association rule, the cross-modal mapping relationship between multi-source modal data is constructed, and a multi-dimensional associated knowledge element structure is generated; The process of classifying the multi-source heterogeneous raw data comprises the following steps: Based on the cross-modal recognition model, the original data is intelligently screened, the drawing, text, table and picture modal data are automatically divided into corresponding data groups, and the classification accuracy is optimized through a dynamic weight algorithm; combined with the power design business scenario, the primary classified data is subjected to secondary semantic clustering by using the domain knowledge graph and the self-attention mechanism; a semi-supervised labeling system based on the generative adversarial network is constructed, and scene labels containing multi-dimensional information such as business scenario, data feature and application demand are automatically generated, and a label dynamic updating mechanism is established; S3: storing the structured knowledge information; the vector database generates semantic vectors through a bidirectional encoder representation model, and realizes efficient retrieval of text and drawing semantic vectors by combining product quantization indexing technology; the graph database constructs a multi-dimensional relationship knowledge graph based on the design scheme-design specification- typical case knowledge triplets, and completes the implicit relationship between entities through a preset association rule; the relationship database divides data tables based on the power design business scenario, and establishes a real-time data synchronization mechanism with the graph database; S4: the hybrid retrieval engine fuses semantic vector retrieval and power keyword retrieval, adjusts the weight proportion of retrieval according to the query type based on a dynamic weighting algorithm, and outputs the retrieval results sorted by correlation degree; the knowledge reasoning service cooperatively outputs the reasoning conclusion and the basis by constructing a rule engine and a graph neural network double-driven reasoning mechanism; the retrieval enhancement generation service integrates a large language model, outputs the generation results conforming to the specification based on a customized dynamic prompt template and a compliance verification plug-in.

2. The method of claim 1, wherein, The power data heterogeneous field intelligent mapping network architecture in S1 comprises: The field feature learning model based on a dynamic attention mechanism extracts features of core fields in the power design field to construct a standard power data field feature library. An intelligent matching is performed between the original fields extracted from the file format and the standard fields in the standard power data field feature library through a reinforcement learning dynamic programming algorithm. The field corresponding relationship is optimized through a simulated game process. Based on a multi-modal cross-validation technology, the content of the equipment specification field in the table is paired with the equipment model description content in the text, and consistency checking is performed in combination with multi-modal information such as drawing annotation and voice annotation.

3. The method of claim 1, wherein, The method for processing the drawing data in S2 is as follows: Based on a dynamic pattern recognition algorithm, multi-dimensional graphic elements such as electrical symbols, device layout positions and line directions in the drawing are extracted in a three-dimensional space model. The extracted multi-dimensional graphic elements are compared with a dynamic standard graphic library in the power design field constructed based on a knowledge graph, and deep semantic analysis of the symbols is realized in combination with the device operation scene. A parameter compliance prediction network is constructed through information extraction of parameter annotation. Based on the parameter requirements in the power design specification, the device operation data under simulated working conditions are used to predict potential risks caused by parameter annotation, and the parameter annotation with risks is dynamically marked. The processed graphic information and parameter information are arranged into structured drawing data with associated relationships through a graph database technology, and a knowledge graph node containing design intent, technical specification basis and potential optimization suggestions is generated.

4. The method of claim 1, wherein, The generation process of the multi-dimensional associated knowledge element structure in S2 is as follows: A knowledge element intelligent core framework is constructed based on a dynamic semantic network. Through cross-modal semantic fusion technology, device model, engineering stage and design index key information are extracted to establish a dynamic association graph among parameters. Based on natural language processing technology and intelligent retrieval technology, corresponding design specification clauses are matched and embedded. Through augmented reality technology and virtual reality technology, engineering case summaries are provided, and a three-dimensional visual case demonstration link is generated. Based on blockchain technology, the original file is stored in evidence through hash, and the file type and storage location are marked. Through machine learning algorithm, data is deeply analyzed to generate parameter recommendation, specification application suggestion and decision assistance information.

5. The method of claim 1, wherein, The method for generating a semantic vector by the vector database in S3 is as follows: Based on a dynamic evolutionary bidirectional encoder representation model, an iterative pre-training is performed in the power design field through a reinforcement learning mechanism to realize power field adaptive learning of the bidirectional encoder representation model. Based on an attention mechanism and cross-modal fusion technology, the bidirectional encoder representation model extracts deep semantic understanding and multi-dimensional features of input content to generate the semantic vector. Based on the generated semantic vector, a multi-level semantic clustering network is constructed through the product quantization index technology. The multi-level semantic clustering network divides vectors with similar semantics, and realizes parallel retrieval of similar semantic data based on the principle of quantum state superposition.

6. The method of claim 1, wherein, The process for the graph database to complete the implied relationship between entities in S3 includes: Based on the power design association rule base, the existing entities in the knowledge graph are matched by the graph neural network; the semantic similarity between entities is calculated based on node embedding vectors to check the relationship associated with the entity; based on the reinforcement learning model, combined with the current technical development trend of the power industry and the external dynamic factors of policy changes, the operation and maintenance requirements content corresponding to the equipment model is extracted from the structured knowledge information, and the association relationship of high-voltage circuit breaker-operation and maintenance requirements is constructed; based on the completion of the implied relationship, the knowledge graph is updated through the version control and change audit mechanism, and the updated records are notarized by using the blockchain technology, realizing the traceability of entity relationship.

7. The method of claim 1, wherein, The method for dividing data tables and establishing real-time data synchronization mechanism in the relational database in S3 is: Based on the power design business scenario, a dedicated table is constructed, and based on the core business scenarios of substation design, transmission line design and distribution system design, a standardized data table template is established to clearly define the filling specifications of scene basic fields and professional fields; Through the semantic label system, the data association is enhanced, and based on the common features of semantic labels, semantic retrieval and linkage analysis of cross-scene data are completed; Based on the graph database, an event-driven + intelligent prediction dual synchronization mode is constructed; Based on the dual synchronization mode, when the relational database field content is updated, the potential data changes are predicted through historical data, and the entity structure of the graph database is optimized in advance; and based on the version traceability chain technology, data change logs are recorded, and attribute conflicts between the relational database and the graph database are identified and solved through conflict resolution algorithm.

8. The method of claim 1, wherein, The process of dynamic weighting of the mixed retrieval engine in S4 includes: Through natural language processing and multi-modal information analysis, the user query demand is deeply scene-identified to construct a three-dimensional demand model containing query intent, application scenario and data type, and the query type is determined; based on knowledge graph reasoning technology, the standard name, clause term and power professional knowledge network are associated to increase the weight of power keyword retrieval, and the implicit association is mined through graph path analysis to complete intelligent navigation across specifications; Based on the style transfer model, the semantic vector retrieval weight is enhanced and the user input design concept is analyzed to generate multi-version semantic enhanced query sentences, expand the similar scheme matching range; based on the fusion strategy of dynamically adjusting keyword retrieval and semantic vector retrieval, the retrieval results are mapped to the three-dimensional demand model for multi-dimensional scoring, and the intelligent preview is generated for the results combined with generative summary technology, and the associated specification interpretation, similar scheme comparison value-added services are provided.

9. The method of claim 1, wherein, The process of constructing a rule engine and a graph neural network dual-driven reasoning mechanism for the knowledge reasoning service in S4 includes: The rule engine compares the user reasoning requirements with mandatory clauses of the sub-rule library by calling the sub-rule library corresponding to the design scenario, outputs a preliminary compliance judgment; the graph neural network reasons based on the entity association relationship of the substation in the knowledge graph, and combines the time-space dynamic data enhancement model to access meteorological and geological change dynamic data in real time, analyzes the similarity of the scheme and the compliance case and the fitting degree of the specification clause; based on the multi-modal evidence chain aggregation technology, the logical judgment of the rule engine and the semantic reasoning result of the graph neural network are fused, the reasoning logic basis is generated through the visual decision tree, and the reasoning conclusion and basis with confidence score are output.

10. The method of claim 1, wherein, The search of the prompt template of the enhancement generation service in S4 includes: Based on the dynamic scene perception and intelligent semantic adaptation double-engine driving customized dynamic prompt template generation, the user input characteristics are analyzed, the architecture generation mode is called to output the deduction graph containing the overall layout of the project, and the structured scheme framework of the main equipment selection recommendation matrix; The compliance verification is based on the reinforcement learning feedback mechanism to build an adversarial verification environment containing a positive specification library and a negative error case library, and a multi-modal comparison model is used to calculate the semantic similarity and logical contradiction detection between the generated results and the design specification; through the risk heat map visualization technology, a risk heat map is generated to show the risk distribution and impact range.

Images