Abstract: With the advancement of electricity market-oriented reform, electricity market policies are characterized by strong professionalism and frequent iteration. Targeting the issues of hallucination and knowledge lag in general Large Language Models (LLMs), as well as the inability of traditional retrieval-augmented generation (RAG) to comprehend the dynamic evolution mechanism of policies, this paper designed an intelligent question-and-answer system for electricity markets (ChatElet system) that integrates retrieval-augmented generation with the dynamic evolution mechanism of policies. A dual-structure expert knowledge base consisting of "policy ontology + dynamic knowledge graph" was constructed, incorporating rules for policy correlation system establishment, version evolution modeling, timeliness and applicability management, conflict detection, and consistency maintenance. A multi-path retrieval mechanism was devised, combined with semantic re-ranking and a hallucination verification engine, to realize trustworthy generation of output answers. Based on an electricity policy test set containing 500 single-turn Q&A entries and 452 consecutive follow-up rounds, comparative experiments were conducted between the proposed system and baseline systems. Experimental results demonstrate that the ChatElet system achieves scores of 0.628, 0.437 and 0.527 in unigram overlap rate (ROUGE-1), bigram overlap rate (ROUGE-2) and longest common subsequence matching rate (ROUGE-L), respectively. In terms of semantic matching, it attains 0.859 for precision (BERTScore-P), 0.890 for recall (BERTScore-R) and 0.875 for F1-score (BERTScore-F1), respectively, outperforming original LLMs and various RAG-based systems. Meanwhile, the system delivers optimal performance on multiple foundation models, exhibiting cross-model generality. Ablation experiments verify the necessity of collaborative operation of all modules. The system can effectively address problems such as retrieval deviation and content lag, and provide accurate policy interpretation for electricity market participants.

摘要： 随着电力市场化改革的发展，电力市场政策呈现出专业性强、迭代频繁的特点。针对通用大型语言模型(LLM)存在“幻觉”、知识滞后和传统检索增强生成(RAG)难以理解政策动态演化机制的问题，设计一种检索增强生成融合政策动态演化机制的电力市场智能问答系统(ChatElet系统)。构建“政策本体+动态知识图谱”双结构专家知识库，融入政策关联体系构建、版本演化建模、时效与适用性管理、冲突检测与一致性维护规则。设计多路径检索机制，结合语义重排序与幻觉校验引擎，实现输出回答的可信生成。基于含500条单轮问答、452轮连续追问的电力政策测试集，将该系统与基线系统进行对比实验，实验结果表明，ChatElet系统的单字重叠率(ROUGE-1)、双字重叠率(ROUGE-2)、最长公共子序列匹配率(ROUGE-L)分别达到0.628、0.437、0.527，语义匹配精确率(BERTScore-P)、语义匹配召回率(BERTScore-R)、语义匹配F1值(BERTScore-F1)分别达到0.859、0.890、0.875，优于原生模型和各类RAG，同时，该系统在多款基座模型上均实现最优性能，具备跨模型通用性，消融实验验证了各模块协同工作的必要性。该系统能够有效解决检索偏差、内容滞后等问题，为电力市场主体提供精准政策解读。