Electric power design agent intelligent processing system based on AI large model

Abstract

The invention relates to the technical field of electric power design, and discloses an electric power design agent intelligent processing system based on an AI large model, which comprises an interface interaction module, a professional model module, a dialogue management module and a professional algorithm module. The interface interaction module orderly calls idle modules according to the instruction sequence; the professional model module periodically retrieves multi-source data, constructs a professional data set through dynamic denoising, secondary classification processing and verification, generates a fine tuning data set through various knowledge distillation modes, and optimizes an industry large model by adopting a low-rank adapter; the dialogue management module is connected with input and output and instruction generation; and the professional algorithm module dynamically calls five types of professional models and adapts to computing power resources. According to the system, a traditional tool data island is broken through, the problems of general LLM professional shortages and'illusion 'are solved, automatic closed loop from natural language interaction to professional design tasks is achieved, manual dependence and error risks are reduced, and power design efficiency, quality and system stability are improved.

Description

Technical Field

[0001] This application relates to the technical field of power design, and in particular to an intelligent processing system for power design intelligent agents based on an AI large model. Background Technology

[0002] In the field of power engineering design, graphic drawing is a core component throughout the entire design process, directly impacting the safety, economy, and feasibility of the project. Power engineering design requires comprehensive consideration of multiple factors, including power grid topology, equipment parameters, and regulatory requirements. It heavily relies on the designer's professional knowledge and practical experience, and involves a large amount of tedious work such as load calculations, short-circuit current analysis, equipment selection and matching, consulting numerous design specifications, and accurately drawing various types of diagrams, including main electrical wiring diagrams, plan and section diagrams, and secondary circuit diagrams.

[0003] Traditional computer-aided design (CAD) systems focus on graphic drawing and editing, which simplifies drawing operations to some extent but fails to address the computational analysis and data collaboration needs in professional design. Furthermore, specialized analysis software such as short-circuit current calculation, load forecasting, relay protection setting, and power flow calculation are often independently developed and fragmented, resulting in incompatible data formats and scattered storage, creating data silos. Designers must frequently switch between CAD software, various professional calculation tools, and standard document libraries, manually entering, converting, and verifying data. This not only increases repetitive work and prolongs the design cycle but also increases the risk of data deviations due to human error, ultimately affecting the quality of engineering design and the safety of subsequent construction.

[0004] In recent years, Large Language Models (LLMs) have demonstrated strong technical advantages in natural language understanding and generative interaction, offering new possibilities for intelligent human-computer interaction. However, the training data for general-purpose LLMs lacks deep integration with professional knowledge in the power industry, making it difficult to accurately understand the technical connotations, logical connections, and engineering application contexts of professional terms such as "short-circuit current," "relay protection setting," "electrical main wiring," and "grounding resistance calculation." Furthermore, general-purpose LLMs lack the ability to collaboratively drive backend professional design software and calculation tools, often resulting in "illusion" responses. They exhibit significant shortcomings in the accuracy of technical parameters, compliance with standards, and engineering feasibility, and their reliability needs improvement. Summary of the Invention

[0005] To improve the reliability of large-scale models in power design applications, this application provides an intelligent processing system for power design agents based on AI large-scale models, employing the following technical solution:

[0006] An intelligent processing system for power design agents based on AI large-scale models includes the following modules:

[0007] The interface interaction module is used to obtain interaction commands, call the professional model module, dialogue management module and professional algorithm module according to the interaction commands, obtain new interaction commands in real time to obtain command sequence, and call the idle modules in the professional model module, dialogue management module and professional algorithm module in sequence according to the order of the interaction commands in the command sequence.

[0008] The professional model module is used to periodically retrieve multi-source data from a specified industry to obtain professional knowledge content, construct a professional dataset based on the professional knowledge content, verify the professional dataset, and after verification, perform knowledge distillation on the professional dataset to obtain a fine-tuned dataset. Based on the fine-tuned dataset, the model parameters in the preset industry-wide language model are updated using a preset model fine-tuning algorithm. Based on the real-time fine-tuned industry-wide language model, semantic understanding and intent recognition are performed in response to the input language content, and the parsing results are generated and returned.

[0009] The dialogue management module acquires input information and returns task results based on the constructed dialogue engine, generates language content based on the input information, inputs the language content into the professional model module, obtains the parsing results returned by the professional model module, identifies the professional content in the parsing results, and generates calling instructions based on the professional content.

[0010] The specialized algorithm module is used to dynamically invoke electrical calculation models, hardware selection models, specification review models, load simulation models, and drawing modeling models in response to call commands; generate output files corresponding to the specialized content, and send the output files to the dialogue management module.

[0011] By adopting the above technical solutions, the interface interaction module sequentially calls idle professional models, dialogue management, and professional algorithm modules according to the instruction sequence. This effectively breaks down the "data silos" that separate CAD and professional analysis software in traditional power design, avoiding repetitive work such as designers frequently switching tools and manually transferring data, and significantly shortening the design cycle. The professional model module relies on multi-source industry data to construct and verify a fine-tuned dataset through knowledge distillation optimization. This allows for real-time fine-tuning of large industry models, overcoming the shortcomings of general LLMs that lack power expertise and are prone to "illusions." It enables LLMs to accurately understand the engineering context of professional terms such as "short-circuit current" and "relay protection setting," improving the reliability of semantic understanding and intent recognition. The dialogue management module connects input / output and instruction generation. Combined with the professional algorithm module, it dynamically calls dedicated models such as electrical calculations, hardware selection, specification review, and drawing modeling, and generates output files. This achieves a closed loop from natural language interaction to automated execution of professional design tasks. This reduces over-reliance on designer experience, minimizes the risk of data deviation caused by human error, and ensures that engineering design meets the requirements of safety, economy, and feasibility, comprehensively improving the efficiency and quality of power engineering design.

[0012] Optionally, the industry-wide language model has multiple candidate libraries, including models such as Qwen, DeepSeek, and Llama. The models in the candidate libraries are run sequentially for a preset test duration. During the test duration, output files are randomly obtained, and corresponding input feedback information is obtained from the output files. Feedback values ​​are set for the output files based on the input feedback information. The models in the candidate libraries are sorted based on the feedback values, and the model at a preset position is selected as the base model of the industry-wide language model.

[0013] By adopting the above technical solution, and by providing multiple alternative model libraries such as Qwen, DeepSeek, and Llama, each model is run sequentially within a preset test duration, and output files are randomly selected to collect input feedback information and set feedback values. The base model is then selected based on the feedback values, thus breaking the limitation of single model selection.

[0014] Optionally, the professional model module obtains multi-source data from a preset industry knowledge base. The multi-source data is a collection of data of multiple data types. The multi-source data is extracted and cleaned to form an initial sample set. The initial sample set is denoised to obtain sample data. The sample data is sorted according to data type and the sampling values ​​corresponding to the data type are set. Professional knowledge content is constructed based on the sample data.

[0015] Calculate the total amount of data of multiple data types in the multi-source data, and calculate the ratio of the amount of sample data of the same data type to the total amount of data as the data ratio; if the data ratio is less than the preset first ratio, adjust the denoising coefficient of the type according to the positive correlation of the data ratio, otherwise, adjust the denoising coefficient of the type according to the negative correlation of the data ratio; where the larger the denoising coefficient, the more noise data is removed, and the smaller the denoising coefficient, the less noise data is removed.

[0016] By adopting the above technical solution, multi-source data is extracted from the industry knowledge base and processed through extraction, cleaning, and denoising to form sample data. The data is then sorted by data type and sample values ​​are set, ensuring the structure and relevance of the power industry knowledge base data. At the same time, the denoising coefficient is dynamically adjusted based on the proportion of data types. When the data proportion is lower than the first ratio, the denoising coefficient is reduced to retain scarce data, and when the proportion meets the standard, the denoising coefficient is increased to eliminate redundant noise. This not only avoids the loss of key professional data (such as niche design scenario specifications and special equipment parameters) but also ensures the purity of massive mainstream data, effectively improving the diversity, balance, and accuracy of the knowledge base data.

[0017] Optionally, the professional dataset, built based on professional knowledge content, also includes:

[0018] The professional knowledge content is processed a second time based on the data type, which includes role instruction type, instruction response pair type, task instruction sequence type and code task type;

[0019] For data of role-based instruction types, define the model's professional identity, behavioral guidelines, and boundary constraints;

[0020] For data of the same type as the instructions, the core knowledge points are accurately extracted from technical documents and transformed into a question-and-answer format;

[0021] For task instruction sequence type data, decompose complex engineering design tasks into multi-step dialogue processes, select model thinking paths and tool call logic;

[0022] For data of code task type, the professional calculation principle data is converted into executable code format and corresponding content explanation is added;

[0023] After secondary processing, a professional dataset is obtained. The professional dataset is then validated, including format checks, content checks, and sampling audits.

[0024] By adopting the above technical solutions, professional knowledge content is subjected to targeted secondary processing according to four data types: role instructions, instruction response pairs, task instruction sequences, and code tasks. The professional identity and constraints of the model are clarified, core knowledge points are extracted and transformed into questions and answers, complex design tasks are decomposed and tool calling logic is sorted out, and professional calculation principles are transformed into executable code with explanations. This comprehensively covers the knowledge transfer, task execution, and tool collaboration needs of power design scenarios, making the dataset structure more in line with the industry's large model fine-tuning requirements. At the same time, through a triple verification mechanism of format checking, content verification, and sampling review, problems such as data deviation and format errors are effectively eliminated.

[0025] Optionally, the knowledge distillation process includes the following:

[0026] Load the professional dataset, read the samples in the professional dataset and perform distillation operation, construct system prompt words and messages based on the read samples, and call the industry's large language model;

[0027] If the call is successful, extract and clean up the model's answer; otherwise, record the error and set the answer to empty.

[0028] High-quality data points are constructed based on the extracted responses and saved to a temporary dataset;

[0029] If there are still samples that have been read in the professional dataset, then read the next sample and perform the distillation operation; otherwise, save the temporary dataset to the fine-tuning dataset.

[0030] By adopting the above technical solution, loading professional datasets sample by sample, constructing targeted system prompt words and message call industry-wide language models, and using the professional capabilities of the models to distill and optimize the samples, while using a rigorous processing logic of "successfully extracting and cleaning answers, and recording errors and leaving blanks for failures", invalid data is effectively filtered out and the reliability of the distilled data is ensured.

[0031] Optionally, knowledge distillation also includes the following:

[0032] Build first, then distill: After building the professional dataset, knowledge distillation is performed.

[0033] Alternatively, distillation can be performed while building: knowledge distillation can be performed immediately by calling the industry's large language model as each data sample in the professional dataset is built.

[0034] Alternatively, iterative optimization through distillation: update the professional dataset according to a preset fixed period, calculate the success rate of calling the industry's large language model during the knowledge distillation process, calculate the distillation feedback value based on the success rate and the preset reference rate, and adjust the fixed period based on the negative correlation of the previous knowledge distillation before updating the professional dataset.

[0035] By adopting the above technical solutions, three flexible collaborative methods—build-before-distillation, build-while-distillation, and distillation iterative optimization—are adopted to adapt to different data construction scenarios in power design. Build-before-distillation facilitates batch data processing and improves distillation efficiency, while build-while-distillation can optimize the quality of individual samples in real time and reduce subsequent rework. Distillation iterative optimization can calculate distillation feedback values ​​based on the model call success rate and negatively adjust the dataset update cycle to achieve dynamic closed-loop optimization of data construction and distillation.

[0036] Optionally, select a supported pre-trained fine-tuning model library, select a fine-tuning algorithm, and select a low-rank adapter to update the model parameters.

[0037] By adopting the above technical solutions, it is possible to flexibly select adaptation algorithms from the pre-trained fine-tuning model library and use low-rank adapter (LoRA) to update model parameters. This not only breaks through the limitations of a single fine-tuning algorithm and can specifically match the professional needs of power design scenarios and the characteristics of the base model, but also significantly reduces the computational cost and memory usage of model fine-tuning through low-rank adaptation technology, avoiding the waste of resources caused by full parameter fine-tuning, and achieving efficient training under limited computing power.

[0038] Optionally, periodic searches include the following:

[0039] Vectorized retrieval is performed based on a preset retrieval engine. After the retrieval, preliminary results are returned. The industry's large language model is used to parse the multiple fragmented paragraphs extracted from the preliminary results to generate comprehensive results. The comprehensive results are then added to multi-source data.

[0040] The portion of the non-contiguous paragraphs in the preliminary results whose width falls within the preset judgment range is considered fragmented paragraphs, and the remaining portion is considered unfragmented paragraphs. A comprehensive proportion is calculated based on the amount of data in the comprehensive results and the amount of data in the unfragmented content, and the judgment range width of the fragmented paragraphs is adjusted according to the positive correlation of the comprehensive proportion value.

[0041] By adopting the above technical solution, vectorized retrieval is carried out based on a preset retrieval engine. The fragmented segments in the preliminary results are deeply analyzed by an industry-wide large language model to generate coherent comprehensive results. This effectively solves the problems of contextual breakage and scattered key information caused by segmentation in traditional retrieval, making the content of multi-source data more complete and semantically more coherent. At the same time, the comprehensive ratio value is calculated based on the comprehensive results and the amount of non-fragmented content, and the judgment range of fragmented segments is positively adjusted. This allows fragmentation recognition to dynamically adapt to the characteristics of different types of data in the power industry, avoiding missed detection of key fragmented information or misjudgment of non-fragmented content, and significantly improving the accuracy and adaptability of retrieval results.

[0042] Optionally, intent recognition includes the following:

[0043] Extract entity semantics from the language content, classify entity semantics according to the content intent, perform contextual analysis on entity semantics within the same intent category to obtain analysis results, and hierarchically divide the analysis results into structured tasks and executable information according to the execution order. Generate parsing results based on structured tasks and executable information.

[0044] By adopting the above technical solution, the professional entity semantics in the language content are extracted first, and then the intent is classified according to the content. At the same time, contextual correlation analysis is carried out under the same intent, which effectively avoids isolated interpretation and semantic deviation of power professional statements. This ensures that the core needs of expressions such as "What parameters are needed for load forecasting?" and "Please verify the short-circuit current of this line" can be accurately captured. Furthermore, by hierarchically obtaining structured tasks and executable information through execution sequence, the ambiguous natural language instructions are transformed into logically clear and parameter-specific standardized parsing results. This perfectly connects the collaboration between the dialogue management module and the professional algorithm module, allowing backend electrical calculation, specification review and other models to be directly and accurately invoked, which greatly reduces the communication cost and misjudgment risk of task execution.

[0045] Optionally, dynamically calling electrical calculation models, hardware selection models, specification review models, load simulation models, and drawing modeling models includes the following:

[0046] The same computing resources are allocated to the electrical calculation model, hardware selection model, specification review model, load simulation model, and drawing model, in that order.

[0047] The computing power consumption of electrical calculation model, hardware selection model, specification review model, load simulation model and drawing model is obtained in real time, and the average computing power consumption of each model is calculated within the preset calculation time to obtain the average computing power.

[0048] The computing power adjustment value is calculated based on the latest computing power usage and average computing power, and the computing power resources of each model are adjusted in a positive correlation with the computing power adjustment value of each model.

[0049] By adopting the above technical solution, computing resources are first evenly allocated to five types of professional models, including electrical calculations and hardware selection, ensuring resource balance in the initial operation of each model and avoiding task lag due to resource imbalance during the start-up phase. Through real-time monitoring of computing power usage and calculation of the average computing power over a preset time period, a computing power adjustment value is derived based on the latest usage, and resource allocation is adjusted accordingly to achieve dynamic adaptation of computing power resources. This allows high-load models (such as complex electrical calculations and large-scale drawing modeling) to receive more computing power support, while low-load models release redundant resources, effectively avoiding the coexistence of computing power waste and resource shortages. Combined with an architecture where each model is encapsulated as an independent API / microservice and a highly available and easily scalable deployment platform, it supports distributed operation and elastic scaling, while ensuring system stability in high-concurrency scenarios. This allows various professional tasks triggered by users (such as selecting the capacity of a 50MW load transformer) to quickly obtain the appropriate computing power, significantly improving model call response speed and execution efficiency, and ensuring the timeliness and accuracy of output results.

[0050] In summary, this application includes at least one of the following beneficial technical effects: Through the synergistic linkage of four core modules—interface interaction, professional models, dialogue management, and professional algorithms—it breaks down the "data silos" that separate CAD and professional analysis software in traditional power design, avoiding repetitive work such as designers frequently switching tools and manually transferring data. Furthermore, through technologies such as multiple alternative model selection, multi-source data processing with dynamic adjustment of denoising coefficients, secondary processing and triple verification of four types of targeted data, flexible collaborative knowledge distillation, and low-rank adaptation with efficient fine-tuning, it overcomes the shortcomings of general-purpose LLMs, which lack power industry expertise and are prone to "illusions," enabling large industry models to accurately understand professional terminology and engineering context. Simultaneously, with the help of precise intent recognition, dynamic computing power adaptation of five types of professional models, and independent API / microservice deployment, it achieves a complete closed loop from natural language interaction to the automated execution of professional tasks such as electrical calculations, equipment selection, specification review, and drawing modeling, reducing over-reliance on designer experience and minimizing the risk of human error. Attached Figure Description

[0051] Figure 1This is a technical architecture description diagram: It shows the overall technical architecture of the system, covering the relationship between the dialogue engine (Dify), the large power industry model, professional AI agents (electrical calculation, drawing modeling, etc.), and the power industry professional knowledge base. It is the core framework diagram for module collaboration.

[0052] Figure 2 It is a flowchart of the data construction process: it presents the full-process construction logic of the professional dataset (SFT dataset), including multi-source data preprocessing, type-based structuring processing, quality verification and other steps;

[0053] Figure 3 It is a knowledge distillation flowchart: It clarifies the steps of knowledge distillation to optimize professional datasets, from sample loading and model calling to data point generation and saving, forming a standardized process;

[0054] Figure 4 This is a basic flowchart of the knowledge base construction based on RAGFlow: it illustrates the retrieval and data processing mechanism of the power industry knowledge base, including key steps such as vectorized retrieval, fragmented paragraph parsing, and comprehensive result generation. Detailed Implementation

[0055] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0056] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0057] This application discloses an intelligent processing system for power design based on a large AI model, belonging to the interdisciplinary field of power engineering design informatization and artificial intelligence. The system takes "natural language-driven power design" as its core objective and solves the problems of isolated tools, inefficient knowledge acquisition, and insufficient intelligent human-computer interaction in traditional power design through the collaborative efforts of four modules: interface interaction, professional models, dialogue management, and professional algorithms.

[0058] This embodiment presents an intelligent processing system for power design agents based on a large AI model, comprising the following modules:

[0059] (a) Interface Interaction Module

[0060] 1. Core Function Positioning

[0061] The interface interaction module serves a dual role as both a "unified interaction entry point" and a "resource scheduling hub":

[0062] As a user interaction portal, it provides an intuitive and efficient natural language interaction interface, encapsulating complex power design processes (such as equipment selection and specification review) in a simple dialogue experience, realizing "natural language-driven design".

[0063] The system acquires user interaction commands in real time and generates command sequences. Based on the command order, it calls idle modules from the professional model module, dialogue management module, and professional algorithm module to avoid resource contention among multiple modules and ensure task execution efficiency.

[0064] The collaborative context management system remembers key information from multiple rounds of interaction, such as user-inputted load parameters and design standard requirements, to ensure the continuity of interaction logic.

[0065] 2. Customized interactive interface integration (adapted to multiple scenarios)

[0066] To achieve the goal of being an "ubiquitous intelligent design assistant," the module supports three customized front-end integration methods, covering the needs of all power design scenarios:

[0067] Embedded Integration with CAD Software: By developing native plugins, the intelligent assistant is directly embedded into mainstream power design environments such as AutoCAD, MicroStation, and Revit. The plugins utilize CAD software API interfaces for deep interaction, reading design information such as equipment models, line routes, and bus parameters from drawings. Simultaneously, the design results generated by the intelligent agent (such as standardized wiring components and parameter optimization suggestions) are directly written into the drawings or used to create new design elements. Designers can verify compliance with specifications in real time during the drawing process (such as whether the conductor cross-sectional area meets dynamic and thermal stability requirements) and generate typical design modules with one click without switching software, significantly improving workflow efficiency.

[0068] General-purpose platform embedded integration: Based on Web Components or micro-frontend technologies, intelligent assistants are encapsulated as reusable components and integrated into existing enterprise work platforms, including enterprise portals, collaborative office tools such as DingTalk / WeChat Work, knowledge management systems, and low-code development platforms. After integration, it supports unified enterprise identity authentication and single sign-on, allowing engineers to quickly access professional support in a familiar office environment. Typical scenarios include design question consultation during mobile work, real-time specification verification during design review meetings, and intelligent retrieval entry points for the enterprise knowledge system. It boasts advantages such as ease of use, rapid deployment, and easy unified management.

[0069] API-based system integration: The core capabilities of the intelligent assistant (such as semantic understanding and agent invocation) are exposed as services through RESTful API interfaces, accompanied by comprehensive API documentation, SDK development toolkits, and debugging examples. It supports asynchronous task processing, batch design task submission, and webhook event subscription (such as task completion notifications), and can be seamlessly integrated into the power industry's automated design pipeline. As an intelligent decision-making node, it injects natural language processing capabilities with large language models into traditional expert systems (such as relay protection setting systems), promoting the intelligent upgrade of traditional design systems.

[0070] (II) Professional Model Module

[0071] The professional model module is the "core of professional knowledge" and "foundation of model capabilities" of the system. Through the entire process of "data acquisition-processing-optimization-model fine-tuning", it provides the system with accurate and reliable power professional knowledge and semantic understanding capabilities. The specific functions are broken down as follows:

[0072] 1. Periodic retrieval and multi-source data processing

[0073] Reference Figure 4 The module leverages the RAGFlow open-source search enhancement and generation engine to perform multi-source data retrieval in the industry, ensuring the timeliness and coverage of the data.

[0074] Search process: Based on RAGFlow's vectorized search technology, relevant data is retrieved from the power industry knowledge base containing design specifications, equipment manuals, engineering cases, etc., and preliminary results are returned;

[0075] Fragmented paragraph processing: Because knowledge base documents need to be segmented into small paragraphs when being added to the database to balance retrieval efficiency and accuracy, this can easily lead to breaks in the context of key information, such as the splitting of standard clauses. The module calls an industry-wide large language model to perform deep analysis and semantic integration on the fragmented paragraphs in the initial results, generating a logically coherent "comprehensive result" and supplementing it to the multi-source dataset;

[0076] Dynamically adapt search strategy: Calculate the "comprehensive proportion value" based on the "comprehensive result data volume / non-fragmented content data volume" and adjust the width of the judgment range of fragmented paragraphs accordingly; if the comprehensive proportion value is high, it indicates that the fragmentation parsing effect is good, and the judgment range is appropriately expanded to include more potential fragmented information; if the comprehensive proportion value is low, it indicates that the parsing effect is insufficient, and the judgment range is narrowed to reduce invalid fragmented data, ensuring the completeness and accuracy of search results.

[0077] 2. Construction of the professional dataset (SFT dataset)

[0078] Reference Figure 2A high-quality SFT dataset is the foundation for model fine-tuning. The module constructs the dataset through a systematic process of "multi-source integration - structured processing - quality verification".

[0079] Data sources: Covering multi-source heterogeneous data in the power industry, including national / industry technical standard documents (such as "GB50217-2018 Power Engineering Cable Design Standard"), design manuals (such as "Power Engineering Electrical Design Manual"), academic papers (research related to power system calculation and equipment selection), equipment manufacturer manuals (equipment parameters such as transformers and circuit breakers), and domain expert experience and knowledge (typical design schemes and solutions to difficult problems).

[0080] Build process:

[0081] 1. Preprocessing stage: Clean the raw data (remove duplicate content), remove noise (filter irrelevant information), and standardize the format (unify document encoding and table structure) to form an initial sample set;

[0082] 2. Structured Processing Stage: Based on predefined templates for four data types, the initial sample set undergoes in-depth processing to ensure data suitability for model fine-tuning requirements.

[0083] Role instruction type: Define the "Power Design Expert" role attributes in the model, and clarify the behavioral guidelines (such as prioritizing answers based on the latest national standards and industry specifications) and boundary constraints (for uncertain technical information, users must be clearly prompted to check the original documents and should not make subjective assumptions).

[0084] Instruction-Answer Pair Type: Core knowledge points are precisely extracted from technical documents and transformed into standard question-and-answer format, for example: "Instruction: Please explain the core purpose of power system short-circuit current calculation; Answer: Short-circuit current calculation is the process of analyzing the current value at the fault point during a short-circuit fault. Its purpose is to provide a basis for circuit breaker breaking capacity selection, relay protection setting, and conductor dynamic and thermal stability verification, so as to ensure the safe operation of the system."

[0085] Task-Instruction Sequence Type: Complex power design tasks (such as selecting the main transformer capacity of a 110kV substation) are broken down into multi-step dialogue processes, clarifying the model's thinking path and tool call logic. For example: "User: needs to select the main transformer capacity for a 110kV substation; Assistant: needs to first obtain the long-term planned maximum active load, power factor, number of transformers and main wiring configuration, and then call the [Load Calculation Agent] to perform load statistics and capacity verification."

[0086] Code task type: Transform electrical calculation principles into executable code snippets and add professional explanations, such as: "Instruction: Write a Python function to calculate the three-phase short-circuit current (Ohm method); Answer: def calculate_3p_short_circuit_current(v_nominal,impedance_per_unit):'''Parameter v_nominal is the rated voltage (kV), impedance_per_unit is the per-unit impedance; the return value is the three-phase short-circuit current (kA)'''importmath;i_sc=v_nominal / (math.sqrt(3)*impedance_per_unit);returni_sc";

[0087] 3. Quality Verification Phase: Through a triple verification process of "automated format check (verifying whether the data format meets the fine-tuning requirements) + professional terminology accuracy verification (checking the correct use of terms such as "relay protection setting" and "electrical main wiring") + manual sampling review (randomly selecting 10%-15% of the samples for power design experts to evaluate data quality), a high-quality SFT dataset with unified format, accurate content, and strong adaptability is finally formed.

[0088] 3. Knowledge Distillation Process

[0089] Reference Figure 3 Knowledge distillation is a "secondary optimization" of specialized datasets. Its core purpose is to leverage the capabilities of large models to improve the professionalism and reliability of the data. The specific process is as follows:

[0090] Initial sample preparation: Extract original technical documents and specification manuals from the power industry knowledge base to form an initial sample set containing professional instructions (such as "verify the short-circuit current of a 10kV line") but with rough or blank answers.

[0091] Model invocation and response generation: The instructions of the initial sample are combined with carefully designed system prompts, such as "Please generate an accurate and rigorous response based on GB50054-2011 'Low-voltage Power Distribution Design Code' as an electrical design expert". The system calls large language models such as DeepSeek through API, sets low temperature parameters to reduce the randomness of responses, and generates high-quality responses that meet industry standards.

[0092] Data cleaning and filtering: Perform basic cleaning (removing redundant expressions and correcting format errors) and error handling on the model-generated answers. If the model call fails, such as due to API timeout, record the error information and set the sample answer to empty to avoid invalid data from being mixed in.

[0093] Dataset Generation: High-quality cleaned responses are paired with the original instructions to construct standardized "instruction-output" data points, which are then saved to a temporary dataset. All initial samples are iteratively processed, with brief delays introduced to avoid API rate limitations. Finally, the temporary dataset is aggregated and saved in JSON format, forming a high-quality fine-tuning dataset specifically for model fine-tuning. To further ensure quality, the dataset undergoes manual review and verification to ensure the accuracy of key technical parameters and citations.

[0094] Technical Implementation: The knowledge distillation process is developed based on Python, calling large models with comprehensive knowledge coverage and high-quality generation, such as DeepSeek-R1, to ensure the professional adaptability of the distilled data.

[0095] 4. Fine-tuning of the industry's large language model

[0096] The core goal of fine-tuning is to enable the general model to "understand" the professional context and task logic of power design, rather than simply imparting knowledge. This knowledge is provided by the RAG knowledge base, and the specific implementation is as follows:

[0097] Base model selection: From open-source, user-friendly, and highly versatile large models such as Qwen, DeepSeek, and Llama, a small number of samples were tested. The test dimensions included understanding of professional terminology, Chinese logical reasoning, and accuracy of tool calls. The model that is most suitable for power design tasks was selected as the base.

[0098] Fine-tuning solution: The LoRA (low-rank adapter) parameter fine-tuning technology is adopted. By inserting a low-rank adapter into the Transformer layer of the model, only the adapter-related parameters are updated, accounting for about 1%-5% of the total parameters of the model, which greatly reduces the computational cost and memory requirements, while ensuring that the fine-tuning effect is close to the level of full parameter fine-tuning.

[0099] Fine-tuning the framework and process:

[0100] Basic framework: Based on HuggingFace's three core libraries, transformers (model loading and training), datasets (dataset processing), and peft (efficient parameter fine-tuning), a standardized and reproducible training process is constructed.

[0101] Alternative framework: Supports the LlamaFactory open-source integrated platform, which is compatible with a variety of mainstream large models and has a built-in Grado-based WebUI interface. Users can complete model selection (such as selecting Llama-2-7B), hyperparameter configuration (such as learning rate and training epochs), PEFT method settings (such as LoRA and QLoRA), training status monitoring (real-time viewing of training loss), and result evaluation (visualization of test set accuracy) through graphical operations without programming, which greatly reduces the technical threshold for fine-tuning.

[0102] Fine-tuning capability goals: Through fine-tuning, the model should possess four core capabilities:

[0103] 1. Deep understanding of power industry terminology and context: accurately interpret the engineering meaning of terms such as "short-circuit current calculation", "relay protection setting", and "electrical main wiring", such as distinguishing the applicable scenarios of "single busbar segmented wiring" and "double busbar wiring";

[0104] 2. Master the professional problem-solving paradigm: follow the logical reasoning of an electrical engineer, such as when calculating the cross-sectional area of ​​a conductor, first clarify the load current and laying environment, then select the conductor material, and finally check the voltage drop;

[0105] 3. Adapt to industry output formats: Generate structured text that meets engineering requirements, such as technical calculation report excerpts, equipment selection recommendation forms, and specification review comments, rather than conversational responses;

[0106] 4. Optimize agent collaboration: accurately identify user intent and generate standardized calling instructions, such as "call [Electrical Computing Agent], parameters: rated voltage 110kV, short-circuit impedance 0.05pu", to ensure that the backend agent can reliably execute tasks.

[0107] (III) Dialogue Management Module

[0108] The dialogue management module serves as a "bridge" connecting "user interaction" and "professional function execution." Its core function is to realize input / output processing, command conversion, and process collaboration.

[0109] Dialogue Engine Construction: Building a dialogue engine based on the Dify framework, refer to... Figure 1 It supports natural language input parsing, multi-turn dialogue state management, and formatted output of task results, while also being compatible with various output formats such as text, tables, and drawing links.

[0110] Input / output processing: Receive user input information, such as "Please design a 10kV distribution line", and convert it into "language content" that the professional model module can understand, such as specifying the design scope, voltage level, and load requirements; obtain the parsing results returned by the professional model module, such as task decomposition steps and required parameters, identify the professional content, such as "the load simulation model needs to be called", and generate standardized intelligent agent calling instructions;

[0111] Workflow collaboration assurance: The collaborative context management system and workflow engine, on the one hand, remember the user's historical input, such as the load data previously provided by the user, to avoid repeated queries; on the other hand, according to the complexity of the task, such as a single-step specification query or a multi-step substation design, trigger the corresponding execution process to ensure seamless connection from user instructions to intelligent agent invocation.

[0112] (iv) Professional Algorithm Module

[0113] The specialized algorithm module is the system's "specialized task execution layer." By dynamically invoking various power design-specific models and optimizing the allocation of computing resources, it enables the automated and precise execution of design tasks.

[0114] 1. Professional model invocation mechanism

[0115] The module maintains a "functional service registry" that covers five core professional intelligent agents. All intelligent agents are encapsulated as independent callable APIs or microservices, facilitating flexible expansion and maintenance.

[0116] Electrical calculation model: Responsible for core calculation tasks of the power system, including short-circuit current calculation, load forecasting, power flow calculation, voltage drop verification, conductor dynamic and thermal stability calculation, etc.

[0117] Hardware selection model: Based on calculation results and design specifications, it completes the selection of equipment such as transformers, circuit breakers, disconnect switches, conductors, and cables, and outputs the selected models, technical parameters, and selection basis, such as in accordance with "DL / T5210.1-2018 Power Construction Quality Acceptance Code";

[0118] Standardized review model: Automatically checks whether the design scheme (such as electrical main wiring and power distribution device layout) complies with the latest national / industry standards, and outputs review opinions (such as "the cross-sectional area of ​​the conductor does not meet the requirements of Clause 3.2.1 in GB50217-2018") and rectification suggestions;

[0119] Load simulation model: Based on historical load data and regional development plans, it simulates load change trends under different operating conditions, such as seasonal load fluctuations and long-term load growth, providing data support for equipment selection and system planning;

[0120] Drawing and Modeling: Automatically generates professional drawings such as electrical main wiring diagrams, plan and section diagrams, and secondary circuit diagrams based on design parameters, supports export in CAD format, and can also build 3D design models based on BIM technology.

[0121] When a user submits specific design requirements, such as "select the main transformer capacity based on the current 50MW load," the module receives the call command generated by the dialogue management module, automatically triggers the hardware selection model, extracts key parameters (load capacity 50MW, voltage level, number of units required), and calls the built-in calculation logic, such as main transformer capacity = maximum load × 1.2-1.5 times redundancy coefficient. Finally, it returns selection results that meet industry standards, such as "recommend 2 50MVA three-phase double-winding transformers, model S11-50000 / 110."

[0122] 2. Complex Task Collaboration and Workflow Engine

[0123] For complex tasks involving multiple steps, such as "power distribution line design" and "substation overall design," the module integrates a workflow engine, referencing... Figure 1 To achieve multi-agent collaborative execution:

[0124] Task serialization: Breaking down complex tasks into ordered subtasks. For example, "power distribution line design" requires the sequential execution of "load simulation → path planning → electrical calculation → conductor selection → specification review → drawing and modeling".

[0125] Execution status tracking: Real-time monitoring of the execution progress of each subtask, such as "Electrical calculation in progress" and "Specification review completed", and recording the transmission process of key data (such as calculation results and selection parameters);

[0126] Anomaly handling mechanism: If a subtask fails to execute, such as due to missing load data causing computational interruption, the workflow engine automatically pauses the task and sends a message to the dialogue management module, prompting the user to provide additional information. Once the problem is resolved, the task execution resumes, forming an end-to-end automated closed loop.

[0127] 3. Dynamic adaptation of computing resources

[0128] To ensure system stability and response efficiency in high-concurrency scenarios (such as multiple designers submitting design tasks simultaneously), the module adopts a computing power management strategy of "initial balanced allocation + real-time dynamic adjustment":

[0129] Initial allocation: When the system starts up, basic computing resources are allocated equally to the electrical calculation model, hardware selection model, specification review model, load simulation model, and drawing modeling model to avoid task lag caused by insufficient initial resources for any model;

[0130] Real-time monitoring and calculation: Real-time collection of computing power usage of each model, such as CPU / GPU utilization and memory consumption, and calculation of the "average computing power" within a preset time period every 5 minutes (configurable) to reflect the long-term resource requirements of the model;

[0131] Dynamic adjustment: A "computing power adjustment value" is calculated based on "latest computing power usage - average computing power," and the computing power resources of each model are adjusted accordingly. If a model's latest computing power usage is significantly higher than the average, such as during large-scale graphics modeling, its computing power quota is increased. If a model's computing power usage is consistently lower than the average, such as during periods of low workload for specification review, redundant computing power is released to the resource pool for use by other high-load models. This strategy effectively avoids the coexistence of computing power waste and resource shortage, ensuring the system's response speed and stability under multi-task concurrency.

[0132] This embodiment further includes the following optional technical solutions:

[0133] (I) Industry-Specific Language Model Base Selection Scheme

[0134] To ensure the adaptability of the base model, the module sets up a multi-model candidate library and selects the optimal model through "real-world scenario testing + feedback evaluation":

[0135] Alternative model libraries include three major series of models: Qwen (Alibaba Cloud, excellent Chinese language capabilities), DeepSeek (DeepSeek, high-quality generation in professional fields), and Llama (Meta, well-developed open-source ecosystem).

[0136] Selection process: Run each model in sequence within a preset test duration (e.g., 72 hours), randomly select the model output design result files (e.g., calculation reports, selection suggestions), and set feedback values ​​for the output files by user input feedback (e.g., "professional accuracy score" and "engineering adaptability score"); rank the candidate models based on the feedback values, and select the model ranked first (or in a preset position, such as the top two) as the base of the industry's large language model;

[0137] Advantages: It breaks through the limitations of single model selection and ensures that the base model performs optimally in terms of understanding power professional terminology, engineering logic reasoning, and intelligent agent collaboration through feedback evaluation from real design scenarios.

[0138] (II) Dynamic Denoising Scheme for Multi-Source Data

[0139] To balance the preservation of scarcity and the removal of redundancy from multi-source data, the module adopts a dynamic denoising strategy based on data proportion:

[0140] Data proportion calculation: After obtaining multi-source data from the preset industry knowledge base, the total data volume of each data type (such as standard clauses, equipment parameters, and engineering cases) is counted, and the ratio of "sample data volume of a certain type / total data volume of that type" is calculated, which is the data ratio.

[0141] Noise reduction coefficient adjustment rules: If the data ratio is less than the preset first ratio, such as 5%, corresponding to scarce data types, such as design specifications under special climatic conditions, then the noise reduction coefficient is reduced according to the positive correlation of the data ratio. The smaller the noise reduction coefficient, the more data is retained, avoiding the loss of key scarce data; if the data ratio is greater than or equal to the first ratio, corresponding to mainstream data types, such as conventional substation design parameters, then the noise reduction coefficient is increased according to the negative correlation of the data ratio. The larger the noise reduction coefficient, the more noise data is removed, ensuring data purity.

[0142] Advantages: It balances data diversity and accuracy, preventing the loss of key information such as niche design scenario specifications and special equipment parameters, while avoiding the impact of redundant noise (such as repeated fragments of equipment manuals) in mainstream data on model training performance.

[0143] (III) Collaborative Solution for Knowledge Distillation and Data Construction

[0144] The module supports three collaborative modes of knowledge distillation and data construction, adapting to different data scales and design scenario requirements:

[0145] Build first, then distill: First, complete the construction of the full professional dataset, including preprocessing, structuring, and quality verification, and then uniformly optimize the dataset through knowledge distillation. This approach is suitable for scenarios with large datasets that require batch processing, such as the initial construction of an industry knowledge base, and can improve distillation efficiency and reduce the cost of repetitive operations.

[0146] Knowledge distillation during data construction: When constructing each data sample, such as generating a command-response pair, the industry's large language model is immediately invoked for knowledge distillation to optimize sample quality in real time. This is suitable for scenarios with small datasets and high accuracy requirements, such as adding datasets with specially designed tasks, and can reduce subsequent data rework.

[0147] Distillation Iterative Optimization: The professional dataset is updated at a preset fixed cycle (e.g., monthly). During each knowledge distillation process, the success rate of calling the industry's large language model and the percentage of samples that successfully generate high-quality answers are calculated. The distillation feedback value is calculated based on "success rate - preset reference rate (e.g., 90%)". If the success rate is lower than the reference rate, it indicates that the distillation effect is not good, and the update cycle of the next round of datasets is shortened negatively, such as from 30 days to 20 days, to accelerate data optimization. If the success rate is higher than the reference rate, the cycle is appropriately extended to balance the optimization effect and resource consumption. This model forms a closed loop of "data construction - distillation - feedback - adjustment" to continuously improve the quality of the dataset.

[0148] By adopting the above technical solutions, the interface interaction module of this application calls idle modules in an orderly manner according to the instruction sequence, which solves the problem of the separation between CAD and professional software such as short-circuit current calculation and load forecasting in traditional power design. It avoids the repetitive work of designers frequently switching tools and manually transferring data, and the design cycle can be shortened by 30%-50%. The professional model module uses multi-source data dynamic denoising, four types of data structured processing, knowledge distillation optimization, and LoRA fine-tuning to enable the industry large model to accurately understand the engineering context of professional terms such as "short-circuit current" and "relay protection setting", which significantly reduces the "illusion" phenomenon of general LLM and provides reliable knowledge support for design tasks. The dialogue management module connects input and output and instruction conversion, and the professional algorithm module dynamically calls five types of professional models and coordinates complex tasks through the workflow engine to form a fully automated process from natural language instructions to design result output. It reduces over-reliance on designers' experience, especially lowering the learning cost for new designers, and avoids the risk of deviation caused by manual data input and parameter calculation, thus improving the pass rate of design results. Through three customized interaction methods—CAD embedding, general platform integration, and API calls—the system can adapt to the needs of all scenarios such as power design, review, mobile office, and automated production lines, truly realizing an "ubiquitous intelligent design assistant" and providing a brand-new human-machine collaboration paradigm for the power design industry.

[0149] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. An AI large model-based power design agent intelligent processing system, characterized in that, The method comprises the following steps: An interface interaction module is used to obtain an interaction instruction, call a professional model module, a dialogue management module and a professional algorithm module according to the interaction instruction, obtain a new interaction instruction in real time to obtain an instruction sequence, and call idle modules in the professional model module, the dialogue management module and the professional algorithm module in sequence according to the order of the interaction instructions in the instruction sequence; The professional model module is used to periodically search for multi-source data in a set industry to obtain professional knowledge content, construct a professional data set according to the professional knowledge content, verify the professional data set, and perform knowledge distillation on the professional data set to obtain a fine-tuning data set after verification is completed; The model parameters in the preset industry large language model are updated according to the fine-tuning data set using a preset model fine-tuning algorithm; Based on the industry large language model fine-tuned in real time, semantic understanding and intent recognition are performed in response to input language content, and an analysis result is generated and returned; The dialogue management module obtains input information and returns a task result based on a constructed dialogue engine, generates language content according to the input information, inputs the language content to the professional model module, obtains the analysis result returned by the professional model module, identifies professional content in the analysis result, and generates a calling instruction according to the professional content; The professional algorithm module is used to dynamically call an electrical calculation model, a hardware selection model, a specification review model, a load simulation model and a drawing modeling model in response to the calling instruction; and an output file corresponding to the professional content is generated and sent to the dialogue management module. 2.The AI large model-based electric power design agent intelligent processing system according to claim 1, characterized in that, The industry large language model has multiple alternative libraries, the models in the alternative libraries include Qwen, DeepSeek and Llama, the models in the alternative libraries are sequentially used to run for a preset test duration, in the test duration, an output file is randomly obtained, key-in feedback information corresponding to the output file is obtained according to the output file, and a feedback value is set for the output file according to the key-in feedback information; Based on the feedback value, the models in the alternative libraries are sorted, and a model at a preset set position is selected as a base model of the industry large language model. 3.The AI large model-based electric power design agent intelligence processing system according to claim 1, characterized in that, The professional model module obtains multi-source data from a preset industry knowledge base, the multi-source data is a collection of data of multiple data types, the multi-source data is extracted and cleaned to form an initial sample set, noise in the initial sample set is removed to obtain sample data, the sample data is sorted according to data types and a sampling value corresponding to the data types is set; and professional knowledge content is constructed according to the sample data; The total amount of data of multiple data types in the multi-source data is calculated, the ratio of the amount of sample data corresponding to the same data type to the total amount of data is calculated as a data ratio, if the data ratio is less than a preset first ratio, the de-noising coefficient corresponding to the data type is adjusted according to the positive correlation of the data ratio, otherwise, the de-noising coefficient corresponding to the data type is adjusted according to the negative correlation of the data ratio; wherein the larger the de-noising coefficient, the more noise data is removed, and the smaller the de-noising coefficient, the less noise data is removed. 4.The AI large model-based electric power design agent intelligent processing system according to claim 3, characterized in that, According to the professional knowledge content, the professional data set is constructed, which further comprises: The professional knowledge content is processed again according to the data types, wherein the data types include role instruction types, instruction answer pair types, task instruction sequence types and code task types; Data on role instruction type define the professional identity, code of conduct and boundary constraints of the model; Data on instruction answer pair type accurately extract core knowledge points from technical documents and convert them into question and answer form; Data on task instruction sequence type decompose complex engineering design tasks into multi-step dialogue processes, select model thinking paths and tool invocation logic; Data on code task type convert professional calculation principle data into executable code format and add corresponding content explanations; After secondary processing, professional datasets are obtained, and the verification content includes format checking, content checking and sampling auditing. 5.The AI big model-based electric power design agent intelligent processing system according to claim 4, characterized in that, The process of knowledge distillation includes the following: Load the professional dataset, read the samples in the professional dataset for distillation operation, construct system prompts and messages according to the read samples, and call the industry large language model; If the call is successful, extract and clean the model answer, otherwise record the error and set the answer to empty; Construct high-quality data points according to the extracted answer content and save them to the temporary dataset; If there are still read samples in the professional dataset, read the next sample and perform distillation operation; Otherwise, save the temporary dataset to the fine-tuning dataset. 6.The AI big model-based electric power design agent intelligence processing system according to claim 5, characterized in that, Knowledge distillation also includes the following: First distill after building: perform knowledge distillation after building the professional dataset; Or, distill while building: immediately call the industry large language model for knowledge distillation when building each data sample of the professional dataset; Or, iterative optimization of distillation: update the professional dataset according to the preset fixed period, calculate the success rate of calling the industry large language model during knowledge distillation, and calculate the distillation feedback value according to the success rate and the preset reference rate. Adjust the fixed period according to the distillation feedback value before updating the professional dataset. 7.The AI big model based power design agent intelligent processing system according to claim 6, characterized in that, Select the supported pre-training fine-tuning model library, select the fine-tuning algorithm, and select the low-rank adapter to update the model parameters. 8.The AI large model-based power design agent intelligent processing system according to claim 3, characterized in that, Periodic retrieval includes the following: Based on the preset search engine, perform vectorization search, return the preliminary results after search, use the industry large language model to analyze the content of the multiple fragmented paragraphs extracted from the preliminary results to generate a comprehensive result, and add the comprehensive result to the multi-source data; The width of the non-continuous paragraph in the preliminary results is located in the preset judgment range width, which is considered as a fragmented paragraph, and the remaining part is considered as a non-fragmented paragraph; calculate the comprehensive proportion value according to the data amount of the comprehensive result and the data amount of the non-fragmented content, and adjust the judgment range width of the fragmented paragraph according to the comprehensive proportion value. 9.The AI-large model based power design agent intelligent processing system according to claim 1, wherein, Intention recognition includes the following: Extract the entity meaning in the language content, classify the entity meaning according to the content, analyze the context association of the entity meaning in the same intention classification, obtain the analysis result, and obtain the structured task and executable information by hierarchical execution sequence, and generate the analysis result according to the structured task and executable information. 10.The AI-large model based power design agent intelligence processing system according to claim 1, wherein, Dynamic calling of electrical calculation model, hardware selection model, specification review model, load simulation model and drawing modeling model includes the following: The same computing power resources are allocated to the electrical calculation model, the hardware selection model, the specification review model, the load simulation model, and the drawing modeling model in sequence; Real-time acquisition of the computing power occupation of the electrical calculation model, the hardware selection model, the specification review model, the load simulation model, and the drawing modeling model, and calculation of the average value of the computing power occupation of each model within a preset calculation time length to obtain a computing power average value; According to the latest computing power occupation and the computing power average value, a computing power adjustment value is calculated, and the computing power resources of each model are positively correlated adjusted according to the computing power adjustment value of each model.

Images