Power field multi-step intelligent reasoning question-answering system and method based on question decomposition and self-adaptive iteration hybrid retrieval

Abstract

The invention discloses an electric power field multi-step intelligent reasoning question answering system and method based on problem decomposition and self-adaptive iteration mixed retrieval, and aims to solve the problems that existing electric power question answering is complex, problem reasoning is not coherent, multi-step knowledge association is weak, retrieval noise is large, and electric power safety regulations are prone to being violated. The method comprises the following steps: S1, constructing a'question-thinking chain-answer 'fine tuning data set in the power field, and performing fine tuning on a Qwen3-32B model; s2, disassembling the complex problem of the user into an ordered sub-problem sequence; s3, dynamically judging a retrieval demand based on an electric power professional vocabulary library, and generating an intermediate answer by adopting sparse and dense fused mixed retrieval; and S4, verifying the candidate answer through the electric power safety regulation hard logic constraint library, and triggering negative feedback regeneration if the candidate answer is illegal until a safety compliance answer is output. According to the method, the accuracy, traceability and safety compliance of complex fault diagnosis and operation guidance questions and answers are improved through'adaptive iterative hybrid retrieval + safety regulation hard constraint 'closed loop.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and intelligent technology in the power industry, specifically to a multi-step intelligent reasoning question-answering system and method in the power field based on a hybrid retrieval of question decomposition and adaptive iteration, applicable to scenarios such as power dispatching, equipment operation and maintenance, fault diagnosis, emergency response, work guidance, and training. Background Technology

[0002] In the power industry, the safety and reliability of power grid operation and dispatch are directly related to social production and people's livelihood. With the large-scale integration of renewable energy, the development of inter-regional interconnected power grids, and the impact of complex electricity load characteristics, the operation of the power system faces unprecedented challenges. Industry personnel need to frequently consult operating procedures, emergency plans, and equipment maintenance documents in their daily work, which not only consumes a lot of time but is also prone to misjudgment due to the complexity of information.

[0003] In recent years, Large Language Models (LLMs) have made breakthrough progress in the field of natural language understanding and generation, but there are still significant shortcomings in their application in the power sector: (1) Insufficient power expertise and reasoning chains. When faced with complex problems involving multiple constraints, multiple stages, and strong causal dependencies, incoherent reasoning or conclusions lacking basis are prone to occur. (2) Serious illusion problem. Large language models are prone to generating seemingly reasonable but actually erroneous content. In power safety production scenarios, incorrect operational suggestions may lead to serious consequences. (3) Insufficient timeliness. Power industry technical standards, operating procedures, equipment parameters, and other information are updated frequently, while the knowledge of large language models ends at the training time point, making it impossible to obtain the latest industry dynamics and standard changes. (4) Poor interpretability. The answers generated by general models lack clear basis and source, making it difficult to trace and verify, which contradicts the power industry's requirements for decision-making transparency and auditability.

[0004] Existing Retrieval Augmentation (RAG) technology, by introducing external authoritative knowledge through a "retrieve first, generate later" approach, can alleviate the problems of insufficient knowledge and timeliness to some extent. However, it still has shortcomings in complex power-related problem scenarios: First, complex problems lack structured decomposition and multi-step dependent reasoning support, resulting in insufficient multi-hop knowledge association; second, the retrieval strategy lacks adaptive control for power-related technical terms, which can easily lead to recall bias or noise introduction; and third, it lacks a hard constraint verification mechanism for power safety regulations, meaning that generated content may still violate safety regulations. Therefore, there is an urgent need for a multi-step reasoning question-answering system and method for the power field, which can achieve a closed loop of problem decomposition, adaptive hybrid retrieval, and safety compliance verification to improve the accuracy, reliability, and security of question-answering results. Summary of the Invention

[0005] This invention aims to provide a multi-step intelligent reasoning question-answering system and method for the power industry based on problem decomposition and adaptive hybrid retrieval. It addresses the problems of unclear reasoning chains, weak multi-step knowledge connections, high retrieval noise, and easy violation of power safety regulations in existing power question-answering systems under complex problem scenarios. The system achieves multi-step, explainable, traceable, and highly reliable intelligent reasoning question-answering in the power industry, providing reliable support for power dispatching, equipment operation and maintenance, fault handling, and work guidance. The multi-step intelligent reasoning question-answering system for the power industry based on problem decomposition and adaptive hybrid retrieval includes a CoT fine-tuning module, a sub-problem decomposition module, an adaptive iterative retrieval module, and a security verification module. These modules work collaboratively to complete the multi-step reasoning question-answering task, as detailed below: 1. CoT Fine-tuning Module: Used to construct the power domain thinking chain (CoT) fine-tuning dataset and perform full-parameter supervised fine-tuning of the preset large language model, endowing the model with power domain logical reasoning and interpretable derivation capabilities; its construction process involves the collection of original power domain text data, automatic generation of "question-thinking chain-answer" supervised samples based on model API, sample quality screening, and fine-tuning training.

[0006] 2. Sub-problem decomposition module: This module is used to perform structured parsing of complex problems input by users, extract problem intent, key entities and constraints, and use the fine-tuned domain reasoning model to decompose the complex problem into an ordered sequence of sub-problems with logical dependencies.

[0007] 3. Adaptive Iterative Retrieval Module: This module performs context-aware and adaptive iterative retrieval on the sub-question sequence. Its adaptive approach involves dynamically determining whether a retrieval is triggered by identifying term hits using a self-built power industry terminology database. For sub-questions that trigger a retrieval, a "link context-driven query reconstruction + hybrid retrieval" approach is used to obtain the retrieval context. The hybrid retrieval combines dense semantic retrieval with sparse keyword retrieval to improve recall accuracy and reduce noise in power industry scenarios.

[0008] 4. Security Verification Module: This module summarizes the intermediate answers from multiple steps to generate candidate final answers and performs compliance verification based on the hard logic constraint library of power safety regulations. When the verification fails, it triggers an interception or negative feedback regeneration mechanism to output a final safe answer that meets the safety regulations, thereby reducing the possibility of generated content touching the red line of safety regulations in high-risk scenarios.

[0009] The multi-step intelligent reasoning question-answering method for the power field, based on question decomposition and adaptive hybrid retrieval, is applied to the above system, including the following steps: S1. Construct a fine-tuning dataset of the thinking chain in the power field. Based on the original text data in the power field, automatically generate supervised samples containing "question-thinking chain-answer". Perform full-parameter supervised fine-tuning on the preset large language model to obtain a domain reasoning model with logical reasoning ability in the power field. S2. Obtain the user's complex question and use the domain reasoning model to decompose the complex question into an ordered sequence of sub-questions with logical dependencies; S3. Perform context-aware and adaptive iterative retrieval on the sub-problem sequence: use a self-built power professional terminology database to determine whether a retrieval is triggered, use "link context-driven query reconstruction + hybrid retrieval" to obtain the retrieval context for the sub-problems that trigger the retrieval, and generate corresponding intermediate answers based on the sub-problems and the retrieval context, iterating until all sub-problems are completed; S4. Summarize all intermediate answers to form candidate final answers, and perform compliance verification based on the power safety regulations hard logic constraint library. If the verification fails, trigger an interception or negative feedback regeneration mechanism to output the final result that meets the safety regulations constraints.

[0010] Furthermore, the specific method of step S1 includes the following sub-steps: S1-1, Fine-tuning the construction of the original dataset. Collect relevant data from the power sector. This data should include at least power dispatching regulations, operation and maintenance manuals, equipment documents, relevant books, and standards and specifications. The initial form of this data is raw text data, which may not necessarily contain explicit question-and-answer pairs. This is denoted as the original corpus set.

[0011] in For the first One piece of raw text data, This represents the original corpus size. The original corpus is then processed for deduplication, anonymization, and terminology normalization to obtain a standardized corpus set used for automatically constructing question-and-answer samples. .

[0012] S1-2, Automatic Generation and Quality Screening of CoT Question-Answer Samples. Based on the standardized corpus obtained in step S1-1. The Qwen3-max API is called to automatically construct question-answer samples from the original text, generating supervised data containing "question-thought chain-answer", enabling the fine-tuned data to have a learnable reasoning process.

[0013] For any original text Define the sample generation function as:

[0014] in The training problem is derived from the original text through extraction or rewriting. The reasoning process corresponding to the question. This yields the final answer, resulting in the CoT fine-tuning dataset:

[0015] in, This represents the number of samples generated.

[0016] To ensure the usability and consistency of automatically generated samples, a sample quality scoring function is introduced to comprehensively evaluate factors such as "whether the thought process revolves around the question, whether the answer is consistent with the thought process, and whether there are obvious jumps or unfounded conclusions."

[0017] in To call the Qwen3-max API as a scoring tool, and simultaneously apply filtering constraints:

[0018] in The preset threshold is set to 0.8. At that time, the corresponding samples are removed or marked as samples to be revised, thus obtaining a high-quality fine-tuning dataset:

[0019] S1-3, Full-parameter supervised fine-tuning training, based on the high-quality fine-tuning dataset obtained in step S1-2. Full-parameter supervised fine-tuning training was performed on Qwen2.5-72B-instruct to obtain... This enables the model to acquire logical reasoning capabilities and interpretable reasoning output capabilities in the field of electricity.

[0020] The training objective sequence is defined as a concatenation of thought chains and answers:

[0021] The parameters are optimized using the cross-entropy loss function:

[0022] Furthermore, the specific method of step S2 includes the following sub-steps: S2-1, Structured parsing of complex problems: Receiving complex problem text input from the user. The structured analysis is performed to extract the problem intent, key entities, and constraints, providing a basis for subsequent sub-problem generation and dependency determination. The analysis result is defined as follows:

[0023] in, This represents a set of key entities (such as equipment type, station line name, voltage level, etc.). It represents a set of constraints (such as time range, operating condition boundaries, operational restrictions, etc.). Indicates the type of problem intent (such as fault diagnosis, cause analysis, handling steps, etc.).

[0024] S2-2: Sub-problem candidate generation, calling the domain reasoning model obtained in step S1. This approach breaks down complex problems into several candidate subproblems to cover the key reasoning steps required by the original problem. The sequence of candidate subproblems is defined as follows:

[0025] in, The number of candidate subproblems. Indicates the first Each candidate sub-question is expressed as a single, searchable, and answerable objective question.

[0026] To ensure that the decomposition results are "sufficient and not redundant," a coverage constraint is introduced, and a coverage function is defined:

[0027] in, This indicates the degree to which the candidate sub-problem covers the key information and reasoning objective of the original problem. The preset coverage threshold is set to 0.8. When the coverage is insufficient, sub-problems are added or the granularity of sub-problems is refined.

[0028] S2-3: Logical Dependency Identification and Dependency Graph Construction. This involves determining the sequential dependencies between candidate subproblems and constructing a dependency graph of subproblems with logical dependencies. Define the dependency determination function:

[0029] in Indicates an answer Need to Using intermediate conclusions or evidence as premises, construct a dependency graph accordingly:

[0030] in, Represents the set of sub-problem nodes. Represents the set of dependent edges, edges Indicates "solve first" , then solve To ensure that dependencies are executable, acyclic constraints are introduced. When a circular dependency is detected, conflicting subproblems are rewritten, merged, or split to eliminate circular dependencies and preserve necessary reasoning paths.

[0031] S2-4: Generation of an ordered sequence of subproblems under dependency constraints, in the dependency graph Performing a topological sort on the above yields an ordered sequence of subproblems that satisfy logical dependencies:

[0032] in, The final output is an ordered sequence of subproblems, satisfying the following order constraints:

[0033] in, This indicates that the corresponding subproblem is in the sequence The position index within.

[0034] Furthermore, the specific method of step S3 includes the following sub-steps: S3-1. Power Industry Terminology Database and Search Trigger Determination: A pre-constructed power industry terminology database is used to determine whether a search is required for the current sub-question. The terminology database is represented as follows:

[0035] in For the first The electrical engineering terminology should include at least the following categories: equipment names, voltage levels, protection and relay terms, operating mode terms, and keywords for regulations and clauses.

[0036] For the first Sub-problem text Perform term hit statistics to obtain the hit count. .

[0037]

[0038] in, This represents the word segmentation function. This indicates an indicator function (1 for a hit, 0 for otherwise). When This means that a search is required; otherwise, the search can be skipped, and the model can directly generate an intermediate answer using the existing context, thereby reducing the noise and overhead of invalid searches.

[0039] S3-2, First Child Problem The retrieval and answer generation process involves triggering a decision in step S3-1 before the retrieval. Then Perform a mixed search as a retrieval query to obtain the retrieval context. ;like ,make .

[0040] When a retrieval is triggered, the context retrieval process is represented as follows:

[0041] in, This represents the hybrid retrieval process, which is elaborated in sections S3-4. Subsequently, the model is based on... Generate the first intermediate answer .

[0042]

[0043] when This indicates that an intermediate answer is generated when there is no search evidence.

[0044] S3-3. Starting with the second sub-question, before retrieval, reconstruct the current sub-question based on the preceding link information to make it more suitable for retrieval. Define the preceding link information as:

[0045] when When a retrieval is needed, the model... Generate search-friendly refactored queries :

[0046] when When no search is needed, you can directly set .

[0047] S3-4, Hybrid retrieval execution and context generation: When a retrieval is needed, the reconstructed context is used. Perform a mixed search to obtain the search context. .

[0048] Dense retrieval uses vector similarity:

[0049] in To embed the vector mapping function, Qwen3-Embedding-8B is used as the vectorization model. For cosine similarity, For documents in the database.

[0050] Sparse search scores are obtained using BM25 keyword matching. The BM25 scoring formula has been optimized as follows:

[0051] in Keywords Inverse document frequency, for In the document word frequency in , To adjust the parameters, For document length, This represents the average length of knowledge base documents. The weighting coefficients for electrical engineering terminology are set as follows: core terms are weighted at 1.5-2.0, and common terms are weighted at 1.0.

[0052] Calculate the overall score for hybrid retrieval:

[0053] in This is the weighting coefficient, with a value range of [0.4, 0.6]. According to... Select the candidate document with the highest score documents ( () as the final search result.

[0054] S3-5. Based on the preceding link information, the current sub-question, and its retrieval context, the domain inference model generates the first... An intermediate answer :

[0055] The results of the current round are incorporated into the link information for use in the reconstruction and retrieval of subsequent sub-problems:

[0056] right By sequentially executing steps S3-1 to S3-5, complete preceding link information is obtained:

[0057] Will The output is used in subsequent steps for result aggregation and security verification.

[0058] Furthermore, the specific method of step S4 includes the following sub-steps: S4-1. Obtain the complete preceding link information from S3. Used to generate candidate final answers. The final answer is represented as:

[0059] Complete and Input together into the fine-tuned model This generates the final candidate answers.

[0060] S4-2. Compliance verification based on the hard logic constraint library of power safety regulations: Constructing a hard logic constraint library for power safety regulations. For candidate answers The constraint library is represented as follows: (This is a validation process.)

[0061] in This represents the total number of rules in the constraint library. For any... Define the verification result as:

[0062] Among them, when the rules are met If not satisfied The overall compliance assessment is as follows:

[0063] when This indicates that the verification has passed. This indicates that there may be illegal or high-risk content.

[0064] S4-3, Negative Feedback Regeneration and Safety Output, When When that happens, extract the set of rules that failed:

[0065] And trigger the negative feedback regeneration mechanism, As a hard constraint hint input model, the candidate answers are corrected to obtain new candidate answers:

[0066] in And set the maximum number of respawns to The process terminates when the validation passes or the maximum number of attempts is reached. The final safe answer that passes the validation is represented as:

[0067] in, When it reaches If the request still fails, the system will execute an interception policy and prompt the user to request manual review or provide security alternatives based on the rule base.

[0068] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0069] This application provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0070] The beneficial effects of this invention are as follows: 1. By constructing a Chain of Thought (CoT) fine-tuning dataset for the power industry and performing full-parameter supervised fine-tuning on the basic large language model, the model acquires logical reasoning ability and interpretable derivation ability in the power industry, enabling it to handle complex problems with multiple constraints, multi-stage conditions and strong causal dependencies more stably, and improving the reasoning accuracy and consistency in complex scenarios.

[0071] 2. Through a multi-step reasoning mechanism of "complex problem decomposition - logical dependency sorting", the original complex problem is decomposed into an ordered sequence of sub-problems with logical dependencies, which significantly enhances the ability to connect knowledge in multiple steps and the integrity of the reasoning chain, and reduces problems such as missing key premises and jumping conclusions caused by a one-time answer.

[0072] 3. By introducing an adaptive retrieval triggering mechanism driven by a power industry terminology database, we can achieve "retrieval when it is appropriate and non-retrieval when it is not appropriate". Combined with query reconstruction driven by link context, the retrieval expression is made more in line with the language style of power regulations, reducing retrieval bias and irrelevant recall noise, and improving retrieval efficiency and answer stability.

[0073] 4. A hybrid retrieval strategy that combines dense semantic retrieval and sparse keyword retrieval is adopted, taking into account both semantic relevance and the ability to accurately match professional terms. In the power industry scenario, it can simultaneously improve recall coverage and accuracy, provide more reliable evidence support for multi-step reasoning, and thus improve the credibility and traceability of the final answer.

[0074] 5. Introducing a hard logic constraint library of power safety regulations into the output stage for compliance verification, and triggering an interception or negative feedback regeneration mechanism when the verification fails, can effectively reduce the risk of the generated content touching the red line of power safety regulations, and improve the safety and compliance of high-risk scenarios such as work instructions and accident handling. Attached Figure Description

[0075] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a schematic diagram of the fine-tuning process of the model in this invention; Figure 3 The flowchart for the adaptive iterative retrieval enhancement of this invention is shown below. Detailed Implementation

[0076] like Figure 1 , Figure 2 , Figure 3As shown, the S1 hardware and software environment deployment is as follows. The hardware environment includes: a multi-core high-performance CPU (≥32 cores), at least 128GB of memory, and a high-speed SSD (≥2TB capacity) to support efficient storage and retrieval of the vector database. The software environment includes: a Linux CentOS 7.9 operating system, a Python 3.9 runtime environment, FAISS 1.7.4 library, PyTorch 2.1.0 deep learning framework, integration of the Qwen3-Embedding-8B embedding model and the Qwen3-32B generative model, and configuration of the Qwen3-max API call component for data generation.

[0077] S2. Fine-tuning the Dataset and Model. High-quality text was selected from authoritative documents such as DL / T 572-2021 "Operating Regulations for Power Transformers," DL / T596-2021 "Operating Regulations for High-Voltage Switchgear," and DL / T 593-2018 "Operating Regulations for Power Systems," as well as internal enterprise operation and maintenance records and historical accident cases, forming a raw corpus of 20,000 text entries. The Qwen3-max API was used to automatically generate a fine-tuning dataset from the raw corpus. Through a quality screening process, approximately 16,000 high-quality samples were obtained for model fine-tuning. Using Qwen3-32B as the base model, fully parameter-supervised fine-tuning training was performed using the aforementioned fine-tuning dataset.

[0078] Step S2 includes the following sub-steps: S2-1 Fine-tuning Dataset Construction. 20,000 high-quality original data entries were selected from authoritative documents such as DL / T 572-2021 "Operating Procedures for Power Transformers", DL / T 596-2021 "Operating Procedures for High-Voltage Switchgear", and DL / T 593-2018 "Operating Procedures for Power Systems", as well as internal enterprise operation and maintenance records and historical accident cases. The original data entries underwent deduplication, anonymization, and terminology normalization to obtain a standardized data set. C Call the Qwen3-max API to C Automatically generate "problem-thought chain-answer" supervised samples, and use a quality scoring function. Score ( q i , c i , a i 16,000 high-quality samples with a score ≥ 0.8 were selected to form a fine-tuning dataset. Data .

[0079] S2-2 Domain Inference Model Fine-tuning. Based on the Qwen3-32B model, full-parameter supervised fine-tuning was performed using the cross-entropy loss function. During training, the batch size was set to 8, the learning rate to 2e-5, and the training epochs were 3, ultimately resulting in a domain inference model with multi-step inference capabilities in the power industry. M θ .

[0080] S3. Knowledge Base Construction. Collect the aforementioned authoritative regulations, enterprise operation and maintenance records, and historical cases, and divide the text into blocks of 256 tokens each. Use the Qwen3-Embedding-8B model to vectorize the segmented text, generating 4096-dimensional vectors. Construct an HNSW index based on FAISS (decomposed into 16 subspaces, each with 256 dimensions) to complete the knowledge base initialization. Set up a weekly automatic update task to synchronize with the latest standards and specifications published on the official websites of the National Energy Administration and State Grid, ensuring the timeliness of the knowledge base.

[0081] S4. Demonstrate the question-and-answer process. The implementation example uses a 110kV substation main transformer tripping due to high oil temperature.

[0082] Step S4 includes the following sub-steps: S4-1. When a user inputs the question "The oil temperature of the No. 1 main transformer in a 110kV substation is too high, and the protection has tripped. What should be done? Should it be restarted immediately?", the system analyzes the information and identifies key elements: key entity set (110kV substation, No. 1 main transformer), constraint set (currently in the state after the protection trips), and question intent (processing flow, whether to restart immediately). In the question weight vector, the weights of core terms such as "main transformer", "oil temperature too high", and "tripping" are set to 1.8.

[0083] S4-2, Sub-problem candidate generation. The fine-tuned model is invoked to decompose the problem into a set of candidate sub-problems covering key reasoning stages. In this embodiment, the candidate set is converged into two sub-problems: one of which is... q 1. The main process for answering the question "What should be done first and what should be done next after a power outage?"; the second is... q 2. This is used to answer the question "Is an immediate restart allowed and what are the preconditions?". Therefore, the two sub-questions output by the model in this embodiment are: q 1. "What should be done after the oil temperature of the No. 1 main transformer in a 110kV substation is too high and the protection trips?" q 2: "Under the above conditions, is an immediate restart permissible? If not, what prerequisite checks or confirmation conditions must be met?" Dep ( q 1, q 2) Function determination q 2 Dependenciesq The processing flow of step 1 involves performing a topological sort on the dependency graph and outputting an ordered sequence of subproblems. Q =[ q 1, q 2).

[0084] S4-3, Logical Dependency Determination and Ordered Sequence Output. System Determination. q Question 2 falls under the category of "operational decision-making" questions, and its answer boundaries should be determined by... q The handling process and inspection results of item 1 are constrained, therefore it is determined that... q 2 Logical Dependencies q 1. Based on this, output an ordered sequence of subproblems. Q =[ q 1, q [2], used for subsequent iterative retrieval and intermediate answer generation.

[0085] S4-4, Determine whether q1 has been searched. In this embodiment, q The search was triggered because it matched terms such as "110kV, main transformer, oil overheating, protection action, tripping". Intensive retrieval is based on vector semantic similarity calculation. The sparse search was based on the optimized BM25 algorithm to match professional terms such as "110kV main transformer", "oil temperature too high", "protection action" and "tripping". The weighting weight was 0.5. Five candidate document fragments were initially selected, including Article 5.3.2 of the "Power Transformer Operation Regulations", Section 6.2 of the "Power System Fault Handling Guidelines" and three similar historical fault case documents.

[0086] S4-5, Questions q Answer 1. q1. Input the retrieved documents together into Qwen3-32B. The overall prompt template is: "Role: Operation and Maintenance Engineer. Question: What should be done after the oil temperature of the No. 1 main transformer in a 110kV substation is too high and the protection trips? References: [Article 5.3.2 of the "Power Transformer Operation Regulations": After the oil temperature is too high and the protection trips, it is strictly forbidden to restart immediately. The cause of the overheating must be investigated first...; Section 6.2 of the "Guidelines for Handling Power System Faults": After the transformer protection trips, it should be handled according to the process of 'cause investigation - parameter detection - fault repair - verification and confirmation'...; A similar fault handling case in a power plant in 2023: August 2023, A 110kV main transformer tripped due to excessive oil temperature caused by a cooling system malfunction. The handling procedure is as follows: ...; Equipment Operation and Maintenance Manual: The normal operating oil temperature of the main transformer should be ≤85℃. If it exceeds 95℃, the protection system will trip... [Requirements: 1. Give a clear conclusion; 2. Explain the reasons based on reference materials; 3. List the operation procedures step by step; 4. Identify potential risks and provide preventative measures; 5. If there are no relevant reference materials, explain this clearly.] The initial conclusion is: "First, investigate the cause of the overheating; ensure safety and isolation; quickly determine if there is a significant emergency risk; verify the protection action and temperature measurement information; focus on investigating common causes of excessive oil temperature."

[0087] S4-6, Judgment q 2. Whether to search. q Question 2 does not contain obvious terms, therefore it is determined not to search. Then... q 1. Preliminary conclusions and q Inputting both 2 into Qwen3-32B yields the candidate conclusion "Immediate restart is strictly prohibited". .

[0088] S4-7. Perform security checks on the candidate conclusions to obtain... The procedure complies with power safety regulations, contains no conflicting content, and can be safely executed. The final output is: "The following is the handling procedure after the No. 1 main transformer of the 110kV substation trips due to excessive oil temperature protection: 1. Safety Isolation: Immediately disconnect the power supply to the transformer, ensure proper grounding protection, and set up warning signs; 2. Information Verification: Verify the protection action record and oil temperature monitoring data to confirm that the tripping cause is excessive oil temperature; 3. Cause Investigation: Focus on checking for faults in the cooling system, winding insulation, and oil circuit; 4. Fault Repair: Repair the identified faults to ensure the cooling system operates normally and the winding temperature drops to ≤85℃; 5. Verification and Confirmation: Operation can only be restored after the insulation resistance is found to be qualified and no other potential faults are detected. Immediate restart after tripping is strictly prohibited; four prerequisites must be met: 'normal cooling system, temperature meets standards, insulation is qualified, and fault is eliminated'."

[0089] S5. Through comparative testing, this system demonstrates superior accuracy by generating more complete handling steps based on problem decomposition and ordered reasoning, reducing omissions and conclusion jumps. Regarding retrieval performance, it employs adaptive triggering and link query reconstruction driven by a power terminology database, combined with a hybrid retrieval system integrating dense and BM25 sparse retrieval, improving recall targeting and reducing noise. In terms of safety and reliability, it performs compliance verification based on a hard logic constraint database of power safety regulations, triggering interception or regeneration of high-risk content to avoid outputting suggestions for non-compliant operations.

Claims

1. A multi-step intelligent reasoning question-answering system for the power industry based on a hybrid retrieval method of problem decomposition and adaptive iterative retrieval, characterized in that, This includes a sequentially coordinated MindChain CoT fine-tuning module, a logic decomposition module, an adaptive iterative hybrid retrieval module, and a security verification module, among which: The CoT fine-tuning module is used to construct a fine-tuning dataset of thought chain instructions in the power field and to perform full-parameter supervised fine-tuning of a large language model, resulting in a model with domain-specific multi-step reasoning capabilities. The data in the power sector is high-quality data selected from the available materials, and then the thought chain is constructed by calling the Qwen3-max API. The logic decomposition module is used to process complex user queries. Decomposed into an ordered sequence of subproblems with logical dependencies. ; The adaptive iterative hybrid retrieval module is used to perform context-aware iterative retrieval around the sub-question sequence. It uses the intermediate conclusions and evidence summaries of preceding sub-questions as prior context for subsequent retrievals, and identifies the professional terminology hit rate of the user's questions based on a pre-built power industry lexicon to determine whether to perform a retrieval. The retrieval employs a hybrid retrieval method that combines sparse and dense retrieval to obtain the retrieval context, ultimately generating an answer sequence. ; The security verification module is used to verify the power safety regulations based on the hard logic constraint library. S The candidate outputs are validated for compliance. When a risky entity or a non-compliant operation path is detected, a negative feedback regeneration mechanism is triggered to output a compliant answer.

2. The multi-step intelligent reasoning question-answering system for the power field based on a hybrid retrieval of problem decomposition and adaptive iterative retrieval as described in claim 1, characterized in that, The CoT fine-tuning module is used to endow large language models with multi-step logical reasoning capabilities in the power domain. Its construction method includes the following steps: S1-1. Fine-tune the construction of the original dataset, collect relevant data in the power field, including at least power dispatching procedures, operation and maintenance manuals, equipment documents, relevant books and standard specifications. The initial form of the data is raw text data, which may not contain explicit question-and-answer pairs, denoted as the original corpus set: in For the first One piece of raw text data, To determine the quantity of the original corpus, deduplication, anonymization, and terminology normalization were performed on the original corpus to obtain a standardized corpus set for automatically constructing question-and-answer samples. ; S1-2. Automatic generation and quality screening of CoT question-answer samples, based on the standardized corpus obtained in step S1-1. The Qwen3-max API is called to automatically construct question-answer samples from the original text, generating supervised data containing "question-thought chain-answer", enabling the fine-tuned data to have a learnable reasoning process; For any original text Define the sample generation function as: in The training problem is derived from the original text through extraction or rewriting. The reasoning process corresponding to the question, This yields the final answer, resulting in the CoT fine-tuning dataset: in, The number of samples generated; To ensure the usability and consistency of automatically generated samples, a sample quality scoring function is introduced to comprehensively evaluate whether the thought process revolves around the question, whether the answer is consistent with the thought process, and whether there are obvious jumps or unfounded conclusions. in To call the Qwen3-max API as a scoring tool, and simultaneously apply filtering constraints: in The preset threshold is set to 0.

8. At that time, the corresponding samples are removed or marked as samples to be revised, thus obtaining a high-quality fine-tuning dataset: S1-3, Full-parameter supervised fine-tuning training, based on the high-quality fine-tuning dataset obtained in step S1-2. Full-parameter supervised fine-tuning training was performed on Qwen2.5-72B-instruct to obtain... This enables the model to acquire logical reasoning capabilities and interpretable reasoning output capabilities in the field of electricity. The training objective sequence is defined as a concatenation of thought chains and answers: The parameters are optimized using the cross-entropy loss function: 。 3. The multi-step intelligent reasoning question-answering system for the power field based on a hybrid retrieval of problem decomposition and adaptive iterative retrieval as described in claim 1 or 2, wherein the logical decomposition module comprises: S2-1, Structured parsing of complex problems: Receiving complex problem text input from the user. The structured analysis is performed to extract the problem intent, key entities, and constraints, providing a basis for subsequent sub-problem generation and dependency determination. The analysis result is defined as follows: in, Represents a set of key entities. Represents a set of constraints. Indicates the type of question intent; S2-2: Sub-problem candidate generation, calling the domain reasoning model obtained in step S1. The complex problem is broken down into several candidate sub-problems to cover the key reasoning steps required by the original problem. The sequence of candidate sub-problems is defined as follows: in, The number of candidate subproblems. Indicates the first Each candidate sub-question is expressed as a searchable and answerable single-target question. To ensure that the decomposition results are "sufficient and not redundant," a coverage constraint is introduced, and a coverage function is defined: in, This indicates the degree to which the candidate sub-problem covers the key information and reasoning objective of the original problem. The preset coverage threshold is set to 0.

8. When the coverage is insufficient, sub-problems are added or the granularity of sub-problems is refined. S2-3: Logical Dependency Identification and Dependency Graph Construction. This involves determining the sequential dependencies between candidate subproblems, constructing a dependency graph of subproblems with logical dependencies, and defining the dependency determination function. in Indicates an answer Need to Using intermediate conclusions or evidence as premises, construct a dependency graph accordingly: in, Represents the set of sub-problem nodes. Represents the set of dependent edges, edges Indicates "first solve" , then solve To ensure that dependencies are executable, acyclic constraints are introduced. When a circular dependency is detected, conflicting subproblems are rewritten, merged, or split to eliminate circular dependencies and preserve necessary reasoning paths. S2-4: Generation of an ordered sequence of subproblems under dependency constraints, in the dependency graph Performing a topological sort on the above yields an ordered sequence of subproblems that satisfy logical dependencies: in, The final output is an ordered sequence of subproblems, satisfying the following order constraints: in, This indicates that the corresponding subproblem is in the sequence The position index within.

4. The multi-step intelligent reasoning question-answering system for the power field based on problem decomposition and adaptive iterative hybrid retrieval according to claim 1 or 2, wherein the adaptive iterative hybrid retrieval module comprises: S3-1, Power Industry Terminology Database and Search Trigger Decision: A power industry terminology database is pre-built to determine whether the current sub-question needs to be searched; The terminology database is represented as follows: in For the first The electrical engineering terminology should include at least the equipment name, voltage level, protection and relay terminology, operation mode terminology, and keywords of the regulations and clauses. For the first Sub-problem text Perform term hit statistics to obtain the hit count. : in, This represents the word segmentation function. This indicates an indicator function; it returns 1 if the condition is met, and 0 otherwise. This means that a search is required; otherwise, the search can be skipped and the model can directly generate an intermediate answer using the existing context, thereby reducing the noise and overhead of invalid searches. S3-2, First Child Problem The retrieval and answer generation process involves triggering a decision in step S3-1 before the retrieval. Then Perform a mixed search as a retrieval query to obtain the retrieval context. ;like ,make ; When a retrieval is triggered, the context retrieval process is represented as follows: in, The hybrid retrieval process is described in detail in sections S3-4. Subsequently, the model is based on... Generate the first intermediate answer , when This indicates that an intermediate answer was generated without any search evidence. S3-3. Starting from the second sub-question, before retrieval, the current sub-question is reconstructed based on the preceding link information to make it more suitable for retrieval. The preceding link information is defined as follows: when When a retrieval is needed, the model... Generate search-friendly refactored queries : when When no search is needed, you can directly set , S3-4, Hybrid retrieval execution and context generation: When a retrieval is needed, the reconstructed context is used. Perform a mixed search to obtain the search context. , Dense retrieval uses vector similarity: in To embed the vector mapping function, Qwen3-Embedding-8B is used as the vectorization model. For cosine similarity, For documents in the database, Sparse search scores are obtained using BM25 keyword matching. The BM25 scoring formula has been optimized as follows: in Keywords Inverse document frequency, for In the document word frequency in , To adjust the parameters, For document length, This represents the average length of knowledge base documents. The weighting coefficients for electrical engineering terminology are set as follows: core terms are weighted at 1.5-2.0, and common terms are weighted at 1.

0. Calculate the overall score for hybrid retrieval: in This is the weighting coefficient, with a value range of [0.4, 0.6]. Select the candidate document with the highest score The final search results consist of [number] documents. ; S3-5. Based on the preceding link information, the current sub-question, and its retrieval context, the domain inference model generates the first... An intermediate answer : The results of the current round are incorporated into the link information for use in the reconstruction and retrieval of subsequent sub-problems: right By sequentially executing steps S3-1 to S3-5, complete preceding link information is obtained: Will The output is used in subsequent steps for result aggregation and security verification.

5. The multi-step intelligent reasoning question-answering system for the power field based on a hybrid retrieval of problem decomposition and adaptive iterative retrieval as described in claim 1 or 2, wherein the security verification module comprises: S4-1. Obtain the complete preceding link information from S3. Used to generate candidate final answers, the final answer is represented as: Complete and Input together into the fine-tuned model This generates the final candidate answers. S4-2. Compliance verification based on the hard logic constraint library of power safety regulations: Constructing a hard logic constraint library for power safety regulations. For candidate answers The constraint library is represented as follows: (This is a validation process.) in For any constraint library rule, the total number of rules is [number]. Define the verification result as: Among them, when the rules are met If not satisfied The overall compliance assessment is as follows: when This indicates that the verification has passed. This indicates that there may be illegal or high-risk content. S4-3, Negative Feedback Regeneration and Safety Output, When When that happens, extract the set of rules that failed: And trigger the negative feedback regeneration mechanism, As a hard constraint hint input model, the candidate answers are corrected to obtain new candidate answers: in And set the maximum number of respawns to The process terminates when the validation passes or the maximum number of attempts is reached. The final safe answer that passes the validation is represented as: in, When it reaches If the request still fails, the system will execute an interception policy and prompt the user to request manual review or provide security alternatives based on the rule base.

6. A multi-step intelligent reasoning question-answering method in the power field based on a hybrid retrieval of problem decomposition and adaptive iterative retrieval, characterized in that, Applied to the system according to any one of claims 1-5, the method includes the following steps: S1. Construct a fine-tuning dataset of the thinking chain in the power field. Based on the original text data in the power field, automatically generate supervised samples containing "question-thinking chain-answer". Perform full-parameter supervised fine-tuning on the preset large language model to obtain a domain reasoning model with logical reasoning ability in the power field. S2. Obtain the user's complex question and use the domain reasoning model to decompose the complex question into an ordered sequence of sub-questions with logical dependencies; S3. Perform context-aware and adaptive iterative retrieval on the sub-problem sequence: use a self-built power professional terminology database to determine whether a retrieval is triggered, use "link context-driven query reconstruction + hybrid retrieval" to obtain the retrieval context for the sub-problems that trigger the retrieval, and generate corresponding intermediate answers based on the sub-problems and the retrieval context, iterating until all sub-problems are completed; S4. Summarize all intermediate answers to form candidate final answers, and perform compliance verification based on the power safety regulations hard logic constraint library. If the verification fails, trigger an interception or negative feedback regeneration mechanism to output the final result that meets the safety regulations constraints.

7. The multi-step intelligent reasoning question-answering method in the power field based on problem decomposition and adaptive hybrid retrieval as described in claim 6, characterized in that, In step S1, a fine-tuning dataset is automatically generated based on the original power sector text data by calling the Qwen3-max API. The fine-tuning model is Qwen3-32B.

8. The method according to claim 6 or 7, characterized in that, In step S2, the complex problem is decomposed into a set of candidate sub-problems based on the domain reasoning model, and the logical dependencies between the sub-problems are determined. An ordered sequence of sub-problems is output under the dependency constraints.

9. The method according to claim 6 or 7, characterized in that, In step S3, it is determined whether a retrieval is triggered based on the self-built power professional terminology database; for the sub-question that triggers the retrieval, the current sub-question is reconstructed based on the preceding link information before the retrieval, and a hybrid retrieval is used to obtain the retrieval context, wherein the vectorization encoding is generated by Qwen3-Embedding-8B to generate a 4096-dimensional vector, and the hybrid retrieval is obtained by fusing dense semantic retrieval and BM25 sparse retrieval.

10. The method according to claim 6 or 7, characterized in that, In step S4, the candidate final answer is verified for compliance based on the hard logic constraint library of power safety regulations. If the verification fails, an interception or negative feedback regeneration mechanism is triggered.

Images