发电技术 ›› 2025, Vol. 46 ›› Issue (3): 454-466.DOI: 10.12096/j.2096-4528.pgt.25084

• AI在新型电力系统中的应用 • 上一篇    

基于大语言模型的图检索增强生成技术在核电领域的应用与展望

徐浩然1, 张瑾昀1, 马歆1, 雷文强2, 曹杰铭2   

  1. 1.中国核动力研究设计院,四川省 成都市 610213
    2.四川大学计算机学院,四川省 成都市 610065
  • 收稿日期:2025-02-17 修回日期:2025-04-30 出版日期:2025-06-30 发布日期:2025-06-16
  • 通讯作者: 雷文强
  • 作者简介:徐浩然(1990),男,硕士,馆员,主要从事计算机应用工作,549445686@qq.com
    张瑾昀(1996),男,硕士,馆员,研究方向为数字化研发、知识图谱;
    马歆(1995),女,硕士,馆员,从事知识工程、计算机应用工作;
    雷文强(1992),男,博士,教授,研究方向为自然语言处理、信息检索等,本文通信作者,wenqianglei@scu.edu.cn
    曹杰铭(2000),男,硕士研究生,研究方向为大语言模型。
  • 基金资助:
    国家自然科学基金面上项目(62272330)

Applications and Prospects of Graph Retrieval-Augmented Generation Technology Based on Large Language Models in the Nuclear Power Field

Haoran XU1, Jinyun ZHANG1, Xin MA1, Wenqiang LEI2, Jieming CAO2   

  1. 1.Nuclear Power Institute of China, Chengdu 610213, Sichuan Province, China
    2.College of Computer Science, Sichuan University, Chengdu 610065, Sichuan Province, China
  • Received:2025-02-17 Revised:2025-04-30 Published:2025-06-30 Online:2025-06-16
  • Contact: Wenqiang LEI
  • Supported by:
    General Projects of National Natural Science Foundation of China(62272330)

摘要:

目的 为应对新型核电系统在保障能源供应、推动清洁能源转型及实现“双碳”目标过程中面临的结构多样化、负荷不确定性、数据复杂性等挑战,并解决大语言模型应用于核电领域时存在的知识局限性、幻觉及推理成本高等问题,研究了大语言模型与知识图谱结合,特别是图检索增强生成(graph retrieval-augmented generation,GRAG)技术的应用潜力,以构建更智能、高效、可靠的核电系统信息处理能力。 方法 分析了大语言模型与知识图谱在核电领域的研究现状、各自优缺点及结合的互补性;重点介绍了GRAG技术相较于传统检索增强生成(retrieval-augmented generation,RAG)技术的优势,并探讨了其在核电风险评估、智能问答辅助、知识管理与决策支持、故障诊断与预测等场景的具体应用;梳理了引入并微调大语言模型、构建领域知识图谱、实现GRAG增强的技术路径。最后,从混杂数据下的知识图谱构建、大语言模型认知推理与决策、人机交互可控性等方面对未来研究进行了展望。 结论 结合知识图谱的GRAG技术能有效缓解大语言模型在专业领域的知识局限性和幻觉问题,增强生成内容的可解释性与可靠性。研究结果可为未来优化核电领域知识图谱的构建、提升大语言模型在复杂推理任务中的能力、开发与核电领域专家高效互动的人工智能体提供参考。

关键词: 双碳, 核电, 智能电网, 人工智能(AI), 深度学习, 大语言模型, 知识图谱, 检索增强生成

Abstract:

Objectives To address the challenges faced by new types of nuclear power systems in ensuring energy supply, promoting clean energy transition, and achieving “dual carbon” goals, such as structural diversification, load uncertainty, and data complexity, and to solve problems associated with the application of large language models in the nuclear power domain, such as knowledge limitations, hallucinations, and high reasoning costs, the study explores the application potential of combining large language models with knowledge graphs, especially graph retrieval-augmented generation (GRAG) technology. This combination aims to build more intelligent, efficient, and reliable information processing capabilities for nuclear power systems. Methods The research status of large language models and knowledge graphs in the nuclear power domain, their respective advantages and disadvantages, and the complementarity of their integration are analyzed. The advantages of GRAG technology over traditional retrieval-augmented generation (RAG) are highlighted, along with its specific applications in scenarios such as nuclear power risk assessment, intelligent question-answering assistance, knowledge management and decision support, and fault diagnosis and prediction. Furthermore, the technical pathway for introducing and fine-tuning large language models, constructing domain-specific knowledge graphs, and implementing GRAG enhancement is outlined. Finally, an outlook on future research is provided, covering areas such as knowledge graph construction under heterogeneous data, cognitive reasoning and decision-making of large language models, and the controllability of human-computer interaction. Conclusions GRAG technology combined with knowledge graphs can effectively alleviate the knowledge limitations and hallucination problems of large language models in specialized domains, enhancing the interpretability and reliability of the generated content. The research findings can provide references for the future optimization of knowledge graph construction in the nuclear power domain, enhancing the capabilities of large language models in complex reasoning tasks, and developing artificial intelligence agents for efficient interaction with experts in the nuclear power field.

Key words: dual carbon, nuclear power, smart grid, artificial intelligence (AI), deep learning, large language model, knowledge graph, retrieval-augmented generation

中图分类号: