A dynamic feedback and incremental learning method and system for a large power generation model

Abstract

This invention belongs to the interdisciplinary field of artificial intelligence and energy and power technology, and relates to a dynamic feedback and incremental learning method and system for large-scale power generation models. The method includes: acquiring user feedback and generating structured feedback records; extracting structured knowledge after quality filtering, and generating high-priority knowledge items through three-dimensional quality assessment and priority ranking; detecting and resolving conflicts with the existing knowledge base; employing at least two of LoRA fine-tuning, knowledge distillation, parameter protection, and sample replay for incremental learning; and, after multi-dimensional effect evaluation, deploying the decision-making model and / or updating the knowledge base, thus avoiding the performance degradation risk that may be introduced by unconditionally updating the model after each incremental learning iteration. Through a closed-loop process of feedback collection, quality filtering, knowledge extraction and ranking, conflict resolution, incremental learning, evaluation, and decision-making, the method collaboratively achieves high efficiency, stability, and reliability in the continuous updating of large-scale power generation models.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and energy and power technology, specifically involving a dynamic feedback and incremental learning method and system for large-scale power generation models. Background Technology

[0002] The application of large language models in the power generation field is accelerating, supporting key tasks such as fault diagnosis, operation optimization suggestions, and technical document generation. However, the rapid technological iteration, continuous changes in operating conditions, and the emergence of new fault modes in the power generation field require models to have the ability to continuously learn new knowledge to maintain their effectiveness and timeliness.

[0003] In current technologies, the continuous updating of large language models faces four main challenges. First, model updates rely on full retraining, which takes weeks, incurring high computational costs and easily leading to the model forgetting old knowledge—a problem known as "catastrophic forgetting." Second, user feedback is not effectively utilized. Valuable learning resources such as user ratings, corrections, and supplements to the model's answers are currently used only for performance evaluation or simple preference alignment, without being systematically converted into learning signals for model parameter updates. Third, there is a lack of effective detection and resolution mechanisms for conflicts between new and old knowledge. The model may provide contradictory answers to the same question at different times, severely impacting user trust, especially in the power generation field where knowledge accuracy is paramount. Fourth, existing incremental learning methods offer limited protection for large models. Single strategies such as parameter regularization, sample replay, and LoRA fine-tuning are insufficient to simultaneously meet the dual requirements of efficient learning of new knowledge and full retention of old knowledge.

[0004] In summary, existing technologies suffer from problems such as long update cycles, high costs, insufficient feedback utilization, lack of knowledge conflict handling, and limited incremental learning protection effects. There is an urgent need for a technical solution that can achieve efficient, continuous, and low-cost updates to large-scale power generation models. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic feedback and incremental learning method and system for large-scale power generation models, in order to solve the technical problems in the existing technology of large language models during continuous updating, such as long full retraining cycle and high computational cost, failure to systematically convert user feedback information into learning signals, lack of automated detection and resolution mechanism for conflicts between new and old knowledge, and difficulty in balancing efficient learning of new knowledge and full retention of old knowledge by a single incremental learning strategy.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a dynamic feedback and incremental learning method for large-scale power generation models, characterized in that it includes: The system acquires user feedback information on the output of the large-scale power generation model, classifies, labels, and preprocesses the feedback information to generate structured feedback records, and performs quality filtering on the structured feedback records to remove invalid, low-reliability, and malicious feedback to obtain valid feedback data. Structured knowledge is extracted from valid feedback data, and a three-dimensional quality assessment is performed on the structured knowledge to obtain a comprehensive score. The priority ranking of structured knowledge is determined based on the comprehensive score, and high-priority knowledge items are generated. Conflict detection is performed between high-priority knowledge items and knowledge in the pre-defined existing knowledge base. When a conflict is detected, a pre-defined conflict resolution strategy is used to process it, resulting in a knowledge item after conflict resolution. At least two collaborative methods based on LoRA fine-tuning, knowledge distillation regularization, important parameter protection and sample replay are used to incrementally learn the power generation model by using the knowledge entries after conflict resolution as training signals to obtain the updated power generation model. A multi-dimensional effect evaluation is conducted on the updated power generation model, and updated strategy decisions are generated based on the evaluation results. Based on the updated strategy decision, execute model deployment operations and / or update the existing knowledge base to complete a dynamic feedback and incremental learning cycle.

[0007] Optionally, the feedback information includes rating feedback, correction feedback, supplementary feedback, follow-up questioning feedback, and adoption feedback; the classification labeling includes three methods: active collection, passive collection, and automatic labeling. Active collection invites users to evaluate the model's answers through interface prompts, passive collection infers implicit feedback by analyzing users' subsequent behavior, and automatic labeling automatically marks cases where users explicitly point out errors as negative samples.

[0008] Optionally, the three-dimensional quality assessment includes: reliability assessment, which calculates a reliability score by combining the credibility score of the knowledge source and the quality of historical feedback; consistency assessment, which calculates a consistency score by measuring the semantic similarity between new knowledge and existing knowledge in the knowledge base; and novelty assessment, which calculates a novelty score by measuring the degree of difference between new knowledge and existing knowledge. The formula for calculating the comprehensive score is as follows: Overall score = α × reliability + β × consistency + γ × novelty α=0.3, β=0.3, γ=0.4 Structured knowledge with a comprehensive score higher than a preset threshold is entered into the priority learning queue and identified as a high-priority knowledge item.

[0009] Optionally, the conflict detection includes the following three methods: logical consistency detection, which checks whether the new knowledge conflicts with the attribute constraints of the knowledge graph in the knowledge base; context verification, which uses the model's contextual understanding ability to determine whether the new knowledge is reasonable in a specific context; and majority voting, which performs credibility-weighted voting on multiple knowledge sources for the same fact. The preset conflict resolution strategies include: a time priority strategy, which prioritizes new knowledge when its timestamp is updated and the time difference exceeds a preset number of days; an authority priority strategy, which prioritizes knowledge from the authority source when it conflicts with knowledge from ordinary sources; and an uncertainty expression strategy, which presents and explains the conflicting knowledge in the model's response when the correctness cannot be determined.

[0010] Optionally, the parameters for LoRA fine-tuning are set to rank r=8, scaling factor α=16, and learning rate lr=2×10⁻⁶. -4 Training rounds 3-5, batch_size=8; The knowledge distillation regularization uses the large power generation model before incremental learning as the teacher model, with distillation temperature T=2.0 and distillation loss weight λ1=0.5; The critical parameter protection applies a gradient decay factor γ=0.1 to parameters whose importance score is higher than a preset threshold; The sample replay is a representative sample selected from the old knowledge test set, and the number of replay samples accounts for 20% of the total number of training samples.

[0011] Optionally, the multidimensional performance evaluation includes four indicators: mastery of new knowledge, retention of old knowledge, comprehensive quality score, and knowledge consistency index. The rules for the update strategy decision are as follows: when the mastery of new knowledge is ≥95%, the retention of old knowledge is ≥97%, the overall quality score is ≥0.80, and the knowledge consistency index is ≥0.90, an incremental update decision is output; when the mastery of new knowledge is ≥95% but the retention of old knowledge is <97%, a distillation recovery decision is output; when the mastery of new knowledge is <95%, the overall quality score is <0.80, or the knowledge consistency index is <0.90, a full retraining decision is output.

[0012] Optionally, the knowledge base includes a factual knowledge base, an experiential knowledge base, and a strategy knowledge base; The fact knowledge base stores deterministic knowledge, with an update delay of ≤1 hour; the experience knowledge base stores expert experience and best practices, with an update delay of ≤24 hours; the strategy knowledge base stores decision-making strategies and parameter recommendations, with an update delay of ≤1 hour; each update of the knowledge base retains a version snapshot, and the version snapshot retention period is ≥180 days.

[0013] In a second aspect, the present invention provides a dynamic feedback and incremental learning system for large-scale power generation models, characterized in that it comprises: The data acquisition and processing unit is used to acquire user feedback information on the output of the large-scale power generation model, classify, label, and preprocess the feedback information to generate structured feedback records; and perform quality filtering on the structured feedback records to remove invalid, low-reliability, and malicious feedback to obtain valid feedback data. The knowledge extraction and evaluation unit is used to extract structured knowledge from effective feedback data, perform three-dimensional quality evaluation on the structured knowledge to obtain a comprehensive score, determine the priority ranking of structured knowledge based on the comprehensive score, and generate high-priority knowledge items. The conflict detection and handling unit is used to detect conflicts between high-priority knowledge items and knowledge in the preset existing knowledge base. When a conflict is detected, a preset conflict resolution strategy is used to process it and obtain the knowledge item after conflict resolution. The incremental learning unit is used to perform incremental learning on the power generation model by employing at least two collaborative methods based on LoRA fine-tuning, knowledge distillation regularization, important parameter protection and sample replay, using the knowledge entries after conflict resolution as training signals to obtain the updated power generation model. The evaluation unit is used to conduct multi-dimensional performance evaluation of the updated power generation model and generate updated strategy decisions based on the evaluation results. The update unit is used to make decisions based on the update strategy, execute model deployment operations and / or update the existing knowledge base, and complete a dynamic feedback and incremental learning cycle.

[0014] In a third aspect, the present invention provides an electronic device, characterized in that it includes a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement the dynamic feedback and incremental learning method for large-scale power generation models as described in any one of the preceding claims.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the dynamic feedback and incremental learning method for a large-scale power generation model as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention generates structured feedback records by classifying, labeling, and preprocessing user feedback information, and then performs quality filtering on these records to remove invalid, low-reliability, and malicious feedback. This ensures the quality of feedback data entering subsequent processes from the source, avoids interference from low-quality feedback on the model learning process, and provides a reliable data foundation for subsequent knowledge extraction.

[0017] This invention extracts structured knowledge from effective feedback data, performs a three-dimensional quality assessment on the extracted structured knowledge to obtain a comprehensive score, and then determines the priority ranking of knowledge based on the comprehensive score to generate high-priority knowledge items. It can automatically transform unstructured user feedback into structured learnable knowledge and achieve efficient allocation of learning resources based on knowledge quality, ensuring that high-value knowledge enters the subsequent learning process first and avoiding the occupation of limited learning resources by low-value knowledge.

[0018] This invention detects conflicts between high-priority knowledge items and existing knowledge in a pre-defined knowledge base, and processes the detected conflicts using a pre-defined conflict resolution strategy to obtain conflict-resolved knowledge items. This allows for the pre-identification and resolution of contradictions between new and old knowledge before it enters incremental learning, ensuring consistency between the knowledge base and the model output, and preventing the model from outputting conflicting information due to learning contradictory knowledge.

[0019] This invention employs at least two collaborative methods based on LoRA fine-tuning, knowledge distillation regularization, protection of important parameters, and sample replay to incrementally learn a large-scale power generation model using conflict-resolved knowledge entries as training signals. This approach protects old knowledge from multiple dimensions while learning new knowledge, effectively mitigating the catastrophic forgetting problem during incremental learning. As a result, the model can balance the learning efficiency of new knowledge with the retention of old knowledge even under conditions of large parameter scale.

[0020] This invention evaluates the updated power generation model in multiple dimensions, generates update strategy decisions based on the evaluation results, and then executes model deployment operations and / or updates the existing knowledge base according to the update strategy decisions. It can adaptively determine subsequent operations based on the actual effect of incremental learning, avoiding the performance degradation risk that may be introduced by unconditionally updating the model after each incremental learning, and ensuring the stability and reliability of model services in the production environment.

[0021] In summary, this invention achieves high efficiency, stability, and reliability in the continuous updating of the large-scale power generation model through a complete closed-loop process of feedback collection, quality filtering, knowledge extraction and prioritization, conflict detection and resolution, incremental learning, effect evaluation, and decision execution. This is achieved through the synergistic effect of multiple stages, including effective utilization of feedback information, knowledge quality control, automatic conflict resolution, protection against forgetting of old knowledge, and adaptive decision-making for update strategies. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a diagram illustrating the overall architecture of the dynamic feedback and incremental learning system according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the incremental learning engine training process and parameter configuration according to an embodiment of the present invention; Figure 4 This is a comparison chart of the incremental learning effects in multiple scenarios according to embodiments of the present invention; Figure 5 This is a system structure block diagram according to an embodiment of the present invention; Figure 6 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0024] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0025] See Figure 1 This invention discloses a dynamic feedback and incremental learning method for large-scale power generation models, comprising: The system acquires user feedback information on the output of the large-scale power generation model, classifies, labels, and preprocesses the feedback information to generate structured feedback records, and performs quality filtering on the structured feedback records to remove invalid, low-reliability, and malicious feedback to obtain valid feedback data. Structured knowledge is extracted from valid feedback data, and a three-dimensional quality assessment is performed on the structured knowledge to obtain a comprehensive score. The priority ranking of structured knowledge is determined based on the comprehensive score, and high-priority knowledge items are generated. Conflict detection is performed between high-priority knowledge items and knowledge in the pre-defined existing knowledge base. When a conflict is detected, a pre-defined conflict resolution strategy is used to process it, resulting in a knowledge item after conflict resolution. At least two collaborative methods based on LoRA fine-tuning, knowledge distillation regularization, important parameter protection and sample replay are used to incrementally learn the power generation model by using the knowledge entries after conflict resolution as training signals to obtain the updated power generation model. A multi-dimensional effect evaluation is conducted on the updated power generation model, and updated strategy decisions are generated based on the evaluation results. Based on the updated strategy decision, execute model deployment operations and / or update the existing knowledge base to complete a dynamic feedback and incremental learning cycle.

[0026] This invention achieves continuous updates to a large-scale power generation model by establishing a complete closed-loop process, from user feedback collection, quality filtering, knowledge extraction and prioritization, to conflict detection and resolution, incremental learning, effect evaluation, and decision execution. Specifically, the invention first classifies, labels, and filters the collected user feedback, effectively eliminating invalid and malicious feedback and providing a reliable data foundation for subsequent processes. Based on this, it performs three-dimensional quality evaluation and prioritization on the extracted structured knowledge, automatically transforming unstructured feedback into learnable knowledge and achieving efficient allocation of learning resources. Furthermore, after conflict detection and resolution with the existing knowledge base, the invention employs at least two collaborative incremental learning methods from LoRA fine-tuning, knowledge distillation regularization, important parameter protection, and sample replay. This allows the model to effectively protect old knowledge from multiple dimensions while learning new knowledge, significantly mitigating the catastrophic forgetting problem in large-scale model incremental learning. Finally, the invention ensures the stability and reliability of the model service in the production environment by conducting multi-dimensional effect evaluations of the updated model and adaptively deciding whether to deploy the model or update the knowledge base based on the evaluation results. The various steps of this invention work together synergistically to achieve high efficiency, knowledge consistency, and deployment stability in the continuous updating process of the large-scale power generation model.

[0027] In some embodiments, the feedback information includes rating feedback, correction feedback, supplementary feedback, follow-up feedback, and adoption feedback; the classification and labeling includes three methods: active collection, passive collection, and automatic labeling. Active collection invites users to evaluate the model's answers through interface prompts; passive collection infers implicit feedback by analyzing users' subsequent behavior; and automatic labeling automatically marks cases where users explicitly point out errors as negative samples. By setting five feedback types—rating, correction, supplementary, follow-up, and adoption—a full-dimensional feedback collection framework covering users' evaluation, error pointing, supplementation, follow-up, and adoption of model outputs is constructed, ensuring that all kinds of potential learning signals can be effectively captured. Simultaneously, through the cooperation of active collection, passive collection, and automatic labeling, it is possible to actively obtain users' explicit evaluations, mine their implicit feedback intentions by analyzing their subsequent behavior, and automatically label errors explicitly pointed out by users as negative samples. This achieves efficient and comprehensive collection of feedback information without extensive manual intervention, providing a rich and high-quality data source for subsequent knowledge extraction and incremental model learning.

[0028] In some embodiments, the three-dimensional quality assessment includes: reliability assessment, which calculates a reliability score by combining the credibility score of the knowledge source and the quality of historical feedback; consistency assessment, which calculates a consistency score by measuring the semantic similarity between the new knowledge and the existing knowledge in the knowledge base; and novelty assessment, which calculates a novelty score by measuring the degree of difference between the new knowledge and existing knowledge. The formula for calculating the comprehensive score is as follows: Overall score = α × reliability + β × consistency + γ × novelty α=0.3, β=0.3, γ=0.4 Structured knowledge with a comprehensive score higher than a preset threshold is entered into a priority learning queue and identified as high-priority knowledge items. The extracted structured knowledge is quantitatively evaluated across three dimensions: reliability, consistency, and novelty. A comprehensive score is calculated with weights of α=0.3, β=0.3, and γ=0.4. High-priority knowledge items are then selected based on a preset threshold. The benefits are as follows: reliability assessment ensures the accuracy of knowledge from the perspective of source credibility; consistency assessment maintains the internal coordination between new knowledge and the existing knowledge system through semantic similarity calculation; and novelty assessment effectively identifies knowledge with added information through quantitative differentiation. The highest weight (0.4) is assigned to novelty in the weighting configuration, ensuring the system remains sensitive to incremental information while maintaining knowledge reliability and consistency. Threshold selection narrows the processing objects of subsequent conflict detection and incremental learning to high-value knowledge that has been verified by quality, avoiding the ineffective use of system resources by low-quality or repetitive knowledge, ensuring efficient allocation of learning resources, and providing a quality prerequisite for the accuracy of subsequent conflict detection and the effectiveness of incremental learning.

[0029] In some embodiments, the conflict detection includes the following three methods: logical consistency detection, which checks whether the new knowledge conflicts with the attribute constraints of the knowledge graph in the knowledge base; context verification, which uses the model's contextual understanding ability to determine whether the new knowledge is reasonable in a specific context; and majority voting, which performs credibility-weighted voting on multiple knowledge sources for the same fact. The pre-defined conflict resolution strategies include: a time-priority strategy, which prioritizes new knowledge when its timestamp is updated and the time difference exceeds a preset number of days; an authority-priority strategy, which prioritizes knowledge from the authority source when it conflicts with knowledge from ordinary sources; and an uncertainty expression strategy, which presents and explains both conflicting knowledge in the model's response when correctness cannot be determined. By combining logical consistency detection, contextual verification, and majority voting, new knowledge is comprehensively detected from three dimensions: knowledge graph rule constraints, semantic rationality judgment, and multi-source credibility weighting, significantly improving the coverage and accuracy of conflict detection. Furthermore, for detected conflicts, three hierarchical resolution strategies—time-priority, authority-priority, and uncertainty expression—are employed to differentiate between different types and severity of conflicts. In particular, the uncertainty expression strategy transparently displays the knowledge conflict situation to the user when correctness cannot be clearly determined, preventing the model from blindly choosing one side and outputting incorrect information. The aforementioned detection and resolution mechanisms work together to automate the handling of conflicts between old and new knowledge, ensure internal consistency during the knowledge base update process, effectively prevent the model from generating conflicting outputs due to learning contradictory knowledge, and enhance the credibility and user trust of the Power Generation model in practical applications.

[0030] In some embodiments, the parameters for LoRA fine-tuning are set to rank r=8, scaling factor α=16, and learning rate lr=2×10⁻⁶. -4 Training rounds 3-5, batch_size=8; The knowledge distillation regularization uses the large power generation model before incremental learning as the teacher model, with distillation temperature T=2.0 and distillation loss weight λ1=0.5; The critical parameter protection applies a gradient decay factor γ=0.1 to parameters whose importance score is higher than a preset threshold; The sample replay is a representative sample selected from the old knowledge test set, and the number of replay samples accounts for 20% of the total number of training samples.

[0031] The above four strategies form a complementary and enhanced protection network from four dimensions: parameter structure, output distribution, gradient update, and training data. This enables the model to effectively absorb new knowledge while retaining existing knowledge to the maximum extent, achieving a good balance between the efficiency of learning new knowledge and the degree of retention of old knowledge.

[0032] In some embodiments, the multidimensional performance evaluation includes four indicators: new knowledge mastery, old knowledge retention, comprehensive quality score, and knowledge consistency index. The rules for the update strategy decision are as follows: when the mastery of new knowledge is ≥95%, the retention of old knowledge is ≥97%, the overall quality score is ≥0.80, and the knowledge consistency index is ≥0.90, an incremental update decision is output; when the mastery of new knowledge is ≥95% but the retention of old knowledge is <97%, a distillation recovery decision is output; when the mastery of new knowledge is <95%, the overall quality score is <0.80, or the knowledge consistency index is <0.90, a full retraining decision is output.

[0033] This system quantifies incremental learning effectiveness from different dimensions using four indicators: new knowledge mastery, old knowledge retention, overall quality score, and knowledge consistency index. Based on the achievement of these four indicators, it outputs three decisions: incremental update, distillation recovery, or full retraining, thus constructing a complete hierarchical decision-making system. Specifically, when all four indicators meet the criteria, an incremental update decision is output, directly updating the model and maximizing the efficiency of incremental learning. When only old knowledge retention fails to meet the criteria, a distillation recovery decision is output, specifically addressing the problem of old knowledge forgetting. When new knowledge mastery, overall quality score, or knowledge consistency index fails to meet the criteria, a full retraining decision is output, preventing low-quality updates from entering the production environment. In this scheme, the distillation recovery decision serves as an intermediate processing path between incremental update and full retraining, addressing the problem of insufficient old knowledge retention without triggering costly full retraining, thus avoiding unnecessary computational overhead. This technical solution achieves adaptive evaluation and differentiated processing of incremental learning effectiveness, ensuring optimal matching between update strategy and learning effect, maximizing computational resource utilization efficiency while guaranteeing model update quality.

[0034] In some embodiments, the knowledge base includes a fact knowledge base, an experience knowledge base, and a strategy knowledge base; The fact knowledge base stores deterministic knowledge, with an update delay of ≤1 hour; the experience knowledge base stores expert experience and best practices, with an update delay of ≤24 hours; the strategy knowledge base stores decision-making strategies and parameter recommendations, with an update delay of ≤1 hour; each update of the knowledge base retains a version snapshot, and the version snapshot retention period is ≥180 days.

[0035] The knowledge base is divided into three categories: factual knowledge base, experiential knowledge base, and strategy knowledge base. These categories store deterministic knowledge, expert experience and best practices, and decision-making strategies and parameter recommendations, respectively. Different update delay constraints are imposed (factual knowledge base ≤ 1 hour, strategy knowledge base ≤ 1 hour, experiential knowledge base ≤ 24 hours). This ensures rapid response to frequently changing knowledge (facts and strategies) while also taking into account the need for thorough verification of experiential knowledge, achieving a balance between refined knowledge classification management and update efficiency. Furthermore, by retaining a version snapshot for each update with a retention period of ≥ 180 days, any change to a knowledge item can be traced back to a specific version and time point. This provides a data foundation for backtracking and comparison in conflict resolution and version rollback in the production environment, meeting the power generation industry's requirements for knowledge rigor and auditability.

[0036] Example 1: A dynamic feedback and incremental learning method for large-scale power generation models, see [link to example]. Figure 2 ,include: Step 1: Automatic collection of user feedback Users submit queries through the model service layer, and the model returns the generated results. The system automatically records the complete context of this interaction, including the user's query text, the model's output text, the inference confidence score, and related knowledge items. The system also proactively invites users to evaluate the model's response through interface prompts (active collection), and infers implicit feedback (passive collection) based on subsequent user behavior (such as whether they ask the question again or view information from other sources).

[0037] Feedback collection time window t∈[1, 86400] seconds (default 24 hours); rating feedback threshold θ_rating≥3 points (out of 5) is considered positive feedback; invalid interaction threshold θ_reask≥2 times (users asking the same question ≥2 times is considered a negative signal). Feedback data integrity check passes (field non-empty rate ≥95%), no sensitive information desensitization fails; failure criteria: feedback data missing rate >5% or the filtering module is triggered to refuse to receive feedback when batch robot-generated feedback is detected.

[0038] Step 2: Feedback Quality Filtering The feedback quality filtering module performs multi-level filtering on the original feedback. The first level is format filtering, removing duplicate, empty, and garbled feedback; the second level is credibility scoring, which calculates a credibility score by comprehensively considering the user's historical feedback quality, feedback text length, and information richness; the third level is malicious feedback identification, which detects abnormal behaviors such as score manipulation and spamming based on behavioral characteristics.

[0039] Credibility score threshold θ_cred ≥ 0.6 (range 0~1); minimum text length L_min ≥ 10 characters; information richness score threshold θ_info ≥ 0.4; malicious behavior detection confidence threshold θ_malicious ≥ 0.8. Feedback passes all three levels of filtering and has a credibility score ≥ 0.6; failure criteria: feedback with a credibility score < 0.6 enters the pending review queue, feedback with a credibility score < 0.3 or detected malicious behavior features is directly discarded and logged.

[0040] Step 3: Knowledge Extraction The feedback knowledge extraction module analyzes the filtered feedback. For corrective feedback, the system extracts the correct knowledge provided by the user, identifies elements such as fault type, phenomenon description, and suggested measures, and constructs structured knowledge triples (entity, relation, attribute). For supplementary feedback, the system extracts new information and identifies its knowledge category (factual / experience-based / strategy-based). For adoption feedback, the system analyzes the adopted solution and extracts success patterns.

[0041] Entity recognition accuracy must be ≥85%; relation extraction F1 score must be ≥0.80; knowledge category recognition accuracy must be ≥90%; knowledge triple completeness must be ≥90%. Knowledge triple extraction must be complete and entity linking success rate ≥85%. Failure criteria: If entity linking failure rate >15% or key elements are missing from the knowledge triple, return to step 2 to supplement feedback information or mark it as pending manual review.

[0042] Step 4: Knowledge Quality Assessment A three-dimensional quality assessment is performed on the extracted knowledge entries. Reliability assessment: A reliability score is calculated by combining the credibility ratings of the knowledge source users and the quality of historical feedback. Consistency assessment: The semantic similarity between new knowledge and existing knowledge in the knowledge base is calculated; consistency score = 1 - average semantic distance. Novelty assessment: The degree of difference between new knowledge and existing knowledge is calculated; novelty = 1 / (1 + overlap rate), with greater difference indicating higher novelty. The overall score = α × reliability + β × consistency + γ × novelty, where α = 0.3, β = 0.3, and γ = 0.4.

[0043] The reliability weight is α=0.3; the consistency weight is β=0.3; the novelty weight is γ=0.4; the comprehensive scoring threshold is θ_quality≥0.65 (high scores enter the priority learning queue); the lower limit of the reliability score is θ_reliable≥0.5; the lower limit of the consistency score is θ_consistent≥0.6. Knowledge items with a comprehensive score ≥0.65, a reliability score ≥0.5, and a consistency score ≥0.6 enter the priority learning queue; failure criteria: comprehensive scores between 0.5 and 0.65 enter the pending review queue; knowledge items with a comprehensive score <0.5 or a consistency score <0.4 are judged as conflicting knowledge and transferred to the knowledge conflict detection and resolution module for processing.

[0044] Step 5: Knowledge Conflict Detection Receive knowledge entries from step 4 with a consistency score <0.6 or a comprehensive score <0.5. First, perform a logical consistency check: check whether the new knowledge conflicts with entity attributes and relational constraints in the knowledge graph (e.g., if the difference between the new knowledge claiming "a certain device has a working efficiency of 95%" and the existing knowledge "the device is designed to have a efficiency of 92%" is greater than 3 percentage points, it is determined to be an attribute conflict). Second, perform contextual verification: use the model's contextual understanding ability to determine whether the new knowledge is reasonable in a specific context. Finally, perform majority voting: for multiple records of the same fact, calculate the credibility weight of each source and select the one with the highest weighted score.

[0045] The attribute conflict threshold Δ_attr > 3 percentage points (a difference of more than 3% for the same numerical attribute is considered a conflict); the contextual reasonableness confidence threshold θ_context ≥ 0.7; the minimum number of samples for majority voting N_min ≥ 3 records; the maximum source authority weight W_max = 1.0 (authoritative source), and the minimum W_min = 0.2 (ordinary user). If the logical consistency check passes (no attribute conflict or conflict value difference ≤ 3%), or the contextual verification confidence is ≥ 0.7, or the knowledge with the highest weighted score in the majority voting is consistent with the new knowledge, the knowledge item is marked as conflict-free and proceeds directly to step 6; the failure standard is: if the attribute conflict difference is detected to be > 3% and the contextual verification confidence is < 0.7 and the majority voting result contradicts the new knowledge, it is considered a knowledge conflict, triggering the conflict resolution process.

[0046] Step 6: Knowledge Conflict Resolution For knowledge items identified as conflicting in step 5, one of the following three strategies is adopted: Strategy 1 (Time Priority): When the timestamp of new knowledge is updated and the time difference Δt > 30 days, the new knowledge is prioritized, and the version of the old knowledge is recorded for traceability. Strategy 2 (Authority Priority): When new knowledge from an authoritative source conflicts with old knowledge from a general source, the knowledge from the authoritative source is prioritized. Authoritative sources include industry standard documents, manufacturer technical manuals, and expert confirmation records. Strategy 3 (Uncertainty Expression): When it is impossible to determine which knowledge is correct, both new and old knowledge are presented in the model's response, explaining the conflict, such as "According to old data, it is A, but the latest information indicates B. Please refer to the comprehensive judgment."

[0047] The time priority threshold Δt > 30 days; the authoritative source determination rule is: the source category is "industry standard", "manufacturer's manual", or "expert confirmation"; the uncertainty expression trigger condition is: neither the authoritative priority nor the time priority can be clearly determined. If the conflict resolution strategy is successfully applied, the knowledge entry update status is marked as "resolved" with a description of the resolution strategy; the failure criterion is: when all strategies fail to provide a clear conclusion, the knowledge entry enters the manual review queue and an alert is issued.

[0048] Step 7: Incremental Learning Execution The incremental learning engine receives high-quality knowledge entries after conflict resolution and executes the following training process. First, training sample pairs are constructed: the input is the user query text, and the output is the extracted and verified correct knowledge. Then, LoRA lightweight fine-tuning is applied: LoRA rank r=8, α=16, dropout=0.05, and learning rate lr=2×10⁻⁶. -4 The training epochs are 3-5, and the batch size is 8. Next, knowledge distillation regularization is applied: the old model is used as the teacher model, the distillation temperature T = 2.0, and the distillation loss weight λ1 = 0.5. Simultaneously, important parameter protection is applied: the parameter importance score threshold θ_importance = 0.75, and a gradient decay factor γ = 0.1 is applied to parameters with an importance score ≥ 0.75. Finally, sample replay is performed: the number of replay samples N_replay = 20% of the training samples, prioritizing old samples with high parameter coverage.

[0049] LoRA rank r=8; LoRA scaling factor α=16; dropout=0.05; learning rate lr=2×10 -4Training epochs: 3-5; batch size: 8; distillation temperature: T = 2.0℃; distillation loss weight: λ1 = 0.5; parameter importance threshold: θ_importance = 0.75; gradient decay factor: γ = 0.1; replay sample ratio: 20%. Training loss converges (loss decreases <0.01 for two consecutive epochs) or reaches the maximum number of epochs (5); validation set new knowledge accuracy ≥90%; old knowledge accuracy retention ≥95%; failure criteria: training loss diverges (loss increases >0.1 for two consecutive epochs) or new knowledge accuracy <85% or old knowledge retention <90%, at which point the incremental learning is terminated and an alert is issued. See also... Figure 3 The system employs a four-strategy collaborative incremental learning engine: LoRA fine-tuning (rank r=8, learning rate 1e-4, parameter update <1%), knowledge distillation regularization (distillation temperature 3, KL divergence loss weight 0.3), important parameter protection (based on gradient importance scoring, protecting the top 10% of parameters), and sample replay (experience replay buffer capacity 5000 samples). The training cycle is <30 minutes, with a new knowledge mastery rate >92% and an old knowledge retention rate >95%. Test data comes from a 600MW unit; the original parameter size of the fault diagnosis model is 7B, with 500 new cavitation fault samples added (5-shot scenario). The diagram employs a four-branch parallel structure: the input node (high-quality knowledge entries + old model parameters) branches into four paths: a LoRA fine-tuning branch (labeled with r=8 and learning rate), a knowledge distillation branch (labeled with temperature 3 and KL weight 0.3), a parameter protection branch (labeled with gradient scores of the top 10%), and a sample replay branch (labeled with a buffer of 5000 entries). The four outputs converge at the incremental model output node. The learning performance evaluation module is located on the right (labeled with new knowledge mastery rate of 92% and old knowledge retention rate of 95%). The branches are marked with dashed lines to indicate collaborative relationships. The bottom indicates that the training time is <30 minutes. This invention is the first to explicitly coordinate four strategies (instead of a single LoRA), with knowledge distillation regularization preventing catastrophic forgetting (retention rate of 95%) and important parameter protection to prevent key knowledge from being overwritten.

[0050] Step 8: Learning Outcome Evaluation A comprehensive evaluation of the incrementally learned model is conducted. The mastery of new knowledge is assessed on the new knowledge test set: the new knowledge test set contains N_new ≥ 100 test samples, and the evaluation metric is accuracy (Acc_new). The retention of old knowledge is assessed on the old knowledge test set: the old knowledge test set contains N_old ≥ 200 test samples, and the evaluation metric is accuracy (Acc_old). A comprehensive quality score is calculated: based on user satisfaction S∈[0,1], task completion rate CR∈[0,1], and knowledge consistency index CI∈[0,1], the comprehensive score QS = 0.4 × S + 0.3 × CR + 0.3 × CI is calculated. Knowledge consistency is evaluated: the semantic consistency between the model's output knowledge and the knowledge base is compared, and the consistency index CI = number of matches / total number of outputs.

[0051] The new knowledge test set size is N_new≥100; the old knowledge test set size is N_old≥200; the threshold for new knowledge mastery is Acc_new≥95%; the threshold for old knowledge retention is Acc_old≥97%; the threshold for comprehensive quality score is QS≥0.80; and the threshold for knowledge consistency index is CI≥0.90. The criteria for passing (effective incremental learning): When all four indicators (Acc_new≥95%, Acc_old≥97%, QS≥0.80, and CI≥0.90) are met, incremental learning is considered effective, and an "incremental update" decision is output. The criteria for passing (requiring distillation recovery): When Acc_new≥95% but Acc_old<97%, too much old knowledge has been forgotten, and a "distillation recovery" decision is output. The criteria for failing (requiring full retraining): When Acc_new<95%, QS<0.80, or CI<0.90, incremental learning is considered ineffective, and a "full retraining" decision is output with an alarm.

[0052] See Figure 4This invention employs a four-strategy collaborative incremental learning method, compared to the traditional full-scale retraining method (training all parameters from scratch). Evaluation metrics include: new knowledge mastery rate (%), old knowledge retention rate (%), and training time (minutes). Test scenarios include: fault diagnosis learning of new fault types (cavitation fault, 5-shot), operational suggestion adaptation to new operating conditions (deep peak shaving 200MW), and routine knowledge updates (parameter adjustment). Test data comes from 600MW and 350MW units, with 500 fault diagnosis test samples (new cavitation fault), 300 operational suggestion test samples (new operating conditions), and 1000 routine update test samples. The figure uses three sets of bar charts: the left side represents the fault diagnosis scenario (this invention: new knowledge 92% / old knowledge 95% / training time 25min vs. full retraining: new knowledge 95% / old knowledge 65% / training time 180min); the middle side represents the new working condition adaptation scenario (this invention: new knowledge 88% / old knowledge 93% / training time 20min vs. full retraining: new knowledge 90% / old knowledge 72% / training time 150min); and the right side represents the regular update scenario (this invention: new knowledge 95% / old knowledge 97% / training time 15min vs. full retraining: new knowledge 96% / old knowledge 70% / training time 120min). Specific values ​​are labeled above all bars; the test sample size is labeled at the bottom; and an additional table on the right lists the percentage differences for each scenario. The incremental learning effect (old knowledge retention rate 95%+ vs. full retraining 65-72%) was verified in the three scenarios, with a training time reduction of 85% (25min vs. 180min), demonstrating the significant advantages of incremental learning.

[0053] Step 9: Knowledge Base Update Based on the evaluation results of step 8, perform the corresponding knowledge base update operation. When the decision is incremental update: write the new knowledge resolved in step 6 into the fact knowledge base (deterministic knowledge), experience knowledge base (expert experience), or strategy knowledge base (decision suggestions) according to category, while retaining historical version snapshots. When the decision is distillation recovery: execute the knowledge distillation recovery process, distill key knowledge from the old version model into the current model, and then evaluate it again. When the decision is full retraining: start the full retraining process, and update the knowledge base and model parameter versions synchronously upon completion.

[0054] Atomicity requirements for knowledge base update transactions: all transactions must succeed or all transactions must be rolled back; version snapshot retention period ≥ 180 days; fact knowledge base update delay ≤ 1 hour; experience knowledge base update delay ≤ 24 hours; strategy knowledge base update delay ≤ 1 hour. Successful knowledge base update transaction commit, complete version snapshot generation, and complete update log recording; failure criteria: if transaction commit fails or version snapshot generation is incomplete, roll back to the previous version and issue an alert.

[0055] Step 10: Model service layer synchronization The model service layer receives updated model parameters or a new knowledge base version and completes service and traffic switching. After marking the new model version as available, traffic is gradually switched from the old version to the new version using a canary release strategy (initially 10%, gradually increasing to 100% after 24 hours of observation without any anomalies). The model version record in the version management module is updated synchronously, and the update completion event is reported to the scheduling control module.

[0056] Initial traffic ratio for canary release = 10%; canary observation time t_observe ≥ 24 hours; traffic switching step size = increases by 20% every 4 hours; abnormal rollback threshold: automatic rollback is triggered when the error rate increases by >2% or the latency increases by >50ms. New model service error rate ≤ old model error rate + 0.5%, P99 latency ≤ old model P99 latency + 30ms, user satisfaction not significantly decreased (decreased <5%); failure criteria: automatic rollback to the old version and triggering an alarm when the error rate increases by >2%, the P99 latency increases by >50ms, or the user complaint rate increases by >10%.

[0057] Example 2: Novel Cavitation Fault Learning in Fault Diagnosis Models A novel type of feedwater pump cavitation fault occurred in Unit #2 of a thermal power plant. This fault mode was not learned during the initial training of the model. The large-scale power generation model initially gave an incorrect diagnosis of "feedwater pipeline leakage" based on the user's description of the fault phenomenon. The user subsequently corrected this and provided the correct fault diagnosis knowledge. The system automatically learned the diagnostic knowledge for this new fault from this feedback using the method of this invention.

[0058] Table 1: Parameter Configuration

[0059] Table 2: Test Data

[0060] Step S1 Feedback Collection: The user describes the fault phenomenon as "large fluctuations in the inlet pressure of the water pump, abnormal noise, and unstable flow." The model diagnoses it as "water supply pipeline leakage." The user corrects it as "water pump cavitation fault; the inlet pressure and fluid state should be checked." The system automatically records this interaction as corrective feedback, labeling the feedback type as "fault type correction," and the correction content as "pipeline leakage → cavitation."

[0061] Step S2 Knowledge Extraction: The system extracts structured knowledge triples from the correction feedback: Fault Type (Entity) = Feedwater Pump Cavitation; Typical Phenomena (Attributes) = ["Inlet Pressure Fluctuation", "Abnormal Sound", "Unstable Flow"]; Suggested Measures (Entity) = ["Check Inlet Pressure", "Check Fluid State"]. The knowledge novelty assessment score is 0.92 (high novelty), the reliability score is 0.88, the consistency score is 0.85, and the overall score is 0.88 ≥ 0.65, thus entering the priority learning queue.

[0062] Step S3 Incremental Learning: The system uses the LoRA method for incremental fine-tuning, constructing training sample pairs: Input: "Large fluctuations in feedwater pump inlet pressure, abnormal sound, unstable flow," Output: "Feedwater pump cavitation fault, it is recommended to check inlet pressure and fluid conditions." Training parameters are set as follows: LoRA rank r=8, learning rate 2×10⁻⁶. -4 The training process is performed for 4 epochs with a batch size of 8. At the same time, distillation regularization is applied based on the knowledge of distillation temperature T=2.0, and the distillation loss weight λ1=0.5.

[0063] Step S4 Effect Evaluation: After incremental learning is completed, the accuracy reaches 95.2% on the new fault type test set (150 samples) and remains at 97.1% on the old fault type test set (300 samples). The overall quality score is 0.85 and the knowledge consistency index is 0.92. All four indicators meet the standards, so the incremental learning is deemed effective. The model parameters are updated and the knowledge base is synchronized.

[0064] Example 3: Adaptation of the Operation Recommendation Model to Ultra-Low Load Conditions After completing flexibility modifications, a certain thermal power unit can operate stably at 30% load. The original training data of the large-scale power generation model mainly comes from rated load conditions, and the operational recommendations for ultra-low load conditions are inaccurate. By collecting a large amount of user feedback under ultra-low load conditions, the system automatically extracts new operating condition knowledge and performs incremental learning to achieve operating condition adaptation of the operational recommendation model.

[0065] Table 3: Parameter Configuration

[0066] Table 4: Test Data

[0067] Step S1 Feedback Collection: The system collected 187 user feedbacks regarding operational suggestions under ultra-low load conditions, of which 126 were positive (adopting the model's suggestions) and 61 were negative (indicating that the suggestions were not applicable). A typical example of negative feedback: The original model suggested "lowering the feedwater temperature to increase the steam temperature," but multiple users reported that this suggestion was not applicable under ultra-low load conditions and should be changed to "appropriately increasing the feedwater temperature."

[0068] Step S2 Knowledge Extraction: The system extracts new operating condition knowledge from the feedback: Operating condition (entity) = Load rate < 40%; Optimization suggestion (attribute) = Appropriately increase feedwater temperature by 2~5℃; Explanation of cause (relationship) = Improve hydrodynamic stability. The knowledge novelty assessment score is 0.85, the reliability score is 0.92 (based on consensus confirmation from multiple users), and the consistency score is 0.72 (due to superficial conflicts with traditional knowledge). The comprehensive score = 0.3 × 0.92 + 0.3 × 0.72 + 0.4 × 0.85 = 0.825, and it enters the priority learning queue.

[0069] Step S3 Conflict Detection and Resolution: The system detected a surface conflict between the new knowledge "Increase feedwater temperature under ultra-low load" and the old knowledge "Increased feedwater temperature increases thermal stress." The attribute difference is positive but exceeds the scope of application of the context. System analysis confirms that the old knowledge is applicable to normal load rates (≥40%), while the hydrodynamic characteristics change under ultra-low loads, requiring special handling. The system adopts a condition restriction strategy to add an applicable condition label to the new knowledge, marking the knowledge as "Applicable to load rate <40% conditions," and recording the condition restriction in the knowledge base.

[0070] Step S4 Incremental Learning: The system performs incremental fine-tuning for ultra-low load conditions, constructing training samples: Input: "How to adjust feedwater temperature under ultra-low load conditions", Output: "When the load rate is below 40%, it is recommended to appropriately increase the feedwater temperature by 2~5℃ to improve hydrodynamic stability. Note: This suggestion only applies to ultra-low load conditions; under normal load rates, please refer to the conventional feedwater temperature adjustment strategy." Training parameters: LoRA rank r=8, learning rate 2×10⁻⁶. -4 Train for 5 epochs.

[0071] Step S5 Effect Evaluation: After incremental learning, the accuracy reached 96.8% on the ultra-low load test set (120 samples), the retention rate was 97.6% on the regular load test set (280 samples), the overall quality score was 0.89, and the knowledge consistency index was 0.94. All four indicators met the standards, and the incremental learning was deemed effective.

[0072] Example 4 like Figure 5 As shown, based on the same inventive concept as the above embodiments, the present invention also provides a dynamic feedback and incremental learning system for large-scale power generation models, comprising: The data acquisition and processing unit is used to acquire user feedback information on the output of the large-scale power generation model, classify, label, and preprocess the feedback information to generate structured feedback records; and perform quality filtering on the structured feedback records to remove invalid, low-reliability, and malicious feedback to obtain valid feedback data. The knowledge extraction and evaluation unit is used to extract structured knowledge from effective feedback data, perform three-dimensional quality evaluation on the structured knowledge to obtain a comprehensive score, determine the priority ranking of structured knowledge based on the comprehensive score, and generate high-priority knowledge items. The conflict detection and handling unit is used to detect conflicts between high-priority knowledge items and knowledge in the preset existing knowledge base. When a conflict is detected, a preset conflict resolution strategy is used to process it and obtain the knowledge item after conflict resolution. The incremental learning unit is used to perform incremental learning on the power generation model by employing at least two collaborative methods based on LoRA fine-tuning, knowledge distillation regularization, important parameter protection and sample replay, using the knowledge entries after conflict resolution as training signals to obtain the updated power generation model. The evaluation unit is used to conduct multi-dimensional performance evaluation of the updated power generation model and generate updated strategy decisions based on the evaluation results. The update unit is used to make decisions based on the update strategy, execute model deployment operations and / or update the existing knowledge base, and complete a dynamic feedback and incremental learning cycle.

[0073] Example 5 like Figure 6 As shown, the present invention also provides an electronic device 100 for implementing a dynamic feedback and incremental learning method for a large-scale power generation model; The electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on at least one processor 102, and at least one communication bus 104.

[0074] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of dynamic feedback and incremental learning method for power generation large model by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101.

[0075] The memory 101 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0076] At least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 102 may be a microprocessor or any conventional processor. Processor 102 is the control center of electronic device 100, connecting various parts of electronic device 100 via various interfaces and lines.

[0077] The memory 101 in the electronic device 100 stores multiple instructions to implement a dynamic feedback and incremental learning method for a large-scale power generation model, and the processor 102 can execute multiple instructions to achieve the following: The system acquires user feedback information on the output of the large-scale power generation model, classifies, labels, and preprocesses the feedback information to generate structured feedback records, and performs quality filtering on the structured feedback records to remove invalid, low-reliability, and malicious feedback to obtain valid feedback data. Structured knowledge is extracted from valid feedback data, and a three-dimensional quality assessment is performed on the structured knowledge to obtain a comprehensive score. The priority ranking of structured knowledge is determined based on the comprehensive score, and high-priority knowledge items are generated. Conflict detection is performed between high-priority knowledge items and knowledge in the pre-defined existing knowledge base. When a conflict is detected, a pre-defined conflict resolution strategy is used to process it, resulting in a knowledge item after conflict resolution. At least two collaborative methods based on LoRA fine-tuning, knowledge distillation regularization, important parameter protection and sample replay are used to incrementally learn the power generation model by using the knowledge entries after conflict resolution as training signals to obtain the updated power generation model. A multi-dimensional effect evaluation is conducted on the updated power generation model, and updated strategy decisions are generated based on the evaluation results. Based on the updated strategy decision, execute model deployment operations and / or update the existing knowledge base to complete a dynamic feedback and incremental learning cycle.

[0078] Example 6 If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).

[0079] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0080] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0081] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0082] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0083] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, 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.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A dynamic feedback and incremental learning method for large-scale power generation models, characterized in that, include: Obtain user feedback information on the output of the large-scale power generation model, and classify, label, and preprocess the feedback information to generate structured feedback records; The structured feedback records are quality filtered to remove invalid, low-reliability, and malicious feedback, thus obtaining valid feedback data. Structured knowledge is extracted from valid feedback data, and a three-dimensional quality assessment is performed on the structured knowledge to obtain a comprehensive score. The priority ranking of structured knowledge is determined based on the comprehensive score, and high-priority knowledge items are generated. Conflict detection is performed between high-priority knowledge items and knowledge in the pre-defined existing knowledge base. When a conflict is detected, a pre-defined conflict resolution strategy is used to process it, resulting in a knowledge item after conflict resolution. At least two collaborative methods based on LoRA fine-tuning, knowledge distillation regularization, important parameter protection and sample replay are used to incrementally learn the power generation model by using the knowledge entries after conflict resolution as training signals to obtain the updated power generation model. A multi-dimensional effect evaluation is conducted on the updated power generation model, and updated strategy decisions are generated based on the evaluation results. Based on the updated strategy decision, execute model deployment operations and / or update the existing knowledge base to complete a dynamic feedback and incremental learning cycle.

2. The dynamic feedback and incremental learning method for a large-scale power generation model according to claim 1, characterized in that, The feedback information includes rating feedback, correction feedback, supplementary feedback, follow-up questioning feedback, and adoption feedback; the classification labeling includes three methods: active collection, passive collection, and automatic labeling. Active collection invites users to evaluate the model's answers through interface prompts, passive collection infers implicit feedback by analyzing users' subsequent behavior, and automatic labeling automatically marks cases where users explicitly point out errors as negative samples.

3. The dynamic feedback and incremental learning method for a large-scale power generation model according to claim 1, characterized in that, The three-dimensional quality assessment includes: reliability assessment, which calculates a reliability score by combining the credibility score of the knowledge source and the quality of historical feedback; consistency assessment, which calculates a consistency score by measuring the semantic similarity between new knowledge and existing knowledge in the knowledge base; and novelty assessment, which calculates a novelty score by measuring the degree of difference between new knowledge and existing knowledge. The formula for calculating the comprehensive score is as follows: Overall score = α × reliability + β × consistency + γ × novelty α=0.3, β=0.3, γ=0.4 Structured knowledge with a comprehensive score higher than a preset threshold is entered into the priority learning queue and identified as a high-priority knowledge item.

4. The dynamic feedback and incremental learning method for a large-scale power generation model according to claim 1, characterized in that, The conflict detection includes the following three methods: logical consistency detection, which checks whether the new knowledge conflicts with the attribute constraints of the knowledge graph in the knowledge base; context verification, which uses the model's contextual understanding ability to determine whether the new knowledge is reasonable in a specific context; Majority voting involves weighting the credibility of multiple sources of knowledge for the same fact. The preset conflict resolution strategy includes: a time priority strategy, in which the new knowledge is prioritized when the timestamp of the new knowledge is updated and the time difference exceeds a preset number of days; The authoritative priority strategy prioritizes knowledge from authoritative sources when they conflict with knowledge from ordinary sources; the uncertainty expression strategy presents and explains the conflicting knowledge in the model's response when the correctness cannot be determined.

5. The dynamic feedback and incremental learning method for a large-scale power generation model according to claim 1, characterized in that, The LoRA fine-tuning parameters are set as follows: rank r = 8, scaling factor α = 16, and learning rate lr = 2 × 10⁻⁶. -4 Training rounds 3-5, batch_size=8; The knowledge distillation regularization uses the large power generation model before incremental learning as the teacher model, with distillation temperature T=2.0 and distillation loss weight λ1=0.5; The critical parameter protection applies a gradient decay factor γ=0.1 to parameters whose importance score is higher than a preset threshold; The sample replay is a representative sample selected from the old knowledge test set, and the number of replay samples accounts for 20% of the total number of training samples.

6. The dynamic feedback and incremental learning method for a large-scale power generation model according to claim 1, characterized in that, The multidimensional performance evaluation includes four indicators: mastery of new knowledge, retention of old knowledge, comprehensive quality score, and knowledge consistency index. The rules for the update strategy decision are as follows: when the mastery of new knowledge is ≥95%, the retention of old knowledge is ≥97%, the overall quality score is ≥0.80, and the knowledge consistency index is ≥0.90, an incremental update decision is output; when the mastery of new knowledge is ≥95% but the retention of old knowledge is <97%, a distillation recovery decision is output; when the mastery of new knowledge is <95%, the overall quality score is <0.80, or the knowledge consistency index is <0.90, a full retraining decision is output.

7. The dynamic feedback and incremental learning method for a large-scale power generation model according to claim 1, characterized in that, The knowledge base includes a factual knowledge base, an experiential knowledge base, and a strategy knowledge base; The fact knowledge base stores deterministic knowledge, with an update delay of ≤1 hour; the experience knowledge base stores expert experience and best practices, with an update delay of ≤24 hours; the strategy knowledge base stores decision-making strategies and parameter recommendations, with an update delay of ≤1 hour; each update of the knowledge base retains a version snapshot, and the version snapshot retention period is ≥180 days.

8. A dynamic feedback and incremental learning system for large-scale power generation models, characterized in that, include: The data acquisition and processing unit is used to acquire user feedback information on the output of the large-scale power generation model, and to classify, label and preprocess the feedback information to generate structured feedback records. The structured feedback records are quality filtered to remove invalid, low-reliability, and malicious feedback, thus obtaining valid feedback data. The knowledge extraction and evaluation unit is used to extract structured knowledge from effective feedback data, perform three-dimensional quality evaluation on the structured knowledge to obtain a comprehensive score, determine the priority ranking of structured knowledge based on the comprehensive score, and generate high-priority knowledge items. The conflict detection and handling unit is used to detect conflicts between high-priority knowledge items and knowledge in the preset existing knowledge base. When a conflict is detected, a preset conflict resolution strategy is used to process it and obtain the knowledge item after conflict resolution. The incremental learning unit is used to perform incremental learning on the power generation model by employing at least two collaborative methods based on LoRA fine-tuning, knowledge distillation regularization, important parameter protection and sample replay, using the knowledge entries after conflict resolution as training signals to obtain the updated power generation model. The evaluation unit is used to conduct multi-dimensional performance evaluation of the updated power generation model and generate updated strategy decisions based on the evaluation results. The update unit is used to make decisions based on the update strategy, execute model deployment operations and / or update the existing knowledge base, and complete a dynamic feedback and incremental learning cycle.

9. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the dynamic feedback and incremental learning method for a large-scale power generation model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the dynamic feedback and incremental learning method for large-scale power generation models as described in any one of claims 1 to 7.

Images