Research on RAG-Based Cognitive Large Language Model Training Method for Power Standard Knowledge

Electric Standards Knowledge LLM RAG Knowledge Graph Semantic Reasoning

Authors

  • Sai Zhang State Grid Laboratory of Grid Advanced Computing and Applications, China Electric Power Research Institute Co., Ltd., Beijing, China https://orcid.org/0009-0000-0963-8893
  • Xiaoxuan Fan
    fanxiaoxuan08@163.com
    State Grid Laboratory of Grid Advanced Computing and Applications, China Electric Power Research Institute Co., Ltd., Beijing, China https://orcid.org/0009-0006-2463-8983
  • Bochuan Song State Grid Laboratory of Grid Advanced Computing and Applications, China Electric Power Research Institute Co., Ltd., Beijing, China
  • Xiao Liang State Grid Laboratory of Grid Advanced Computing and Applications, China Electric Power Research Institute Co., Ltd., Beijing, China
  • Qiang Zhang State Grid Laboratory of Grid Advanced Computing and Applications, China Electric Power Research Institute Co., Ltd., Beijing, China
  • Zhihao Wang State Grid Laboratory of Grid Advanced Computing and Applications, China Electric Power Research Institute Co., Ltd., Beijing, China
  • Bo Zhang State Grid Wuxi Power Supply Company of Jiangsu Electric Power Co., Ltd., Wuxi, China
Vol. 6 No. 2 (2025): June
Research Articles

Downloads

Electrical standards encompass complex technical requirements across multiple disciplines, making their management and application a significant challenge that urgently requires efficient solutions. This paper proposes a knowledge graph retrieval-enhanced training method for large language models (LLMs). By leveraging a pre-trained language model (PLM), highly similar subgraphs are retrieved from the electrical standards knowledge graph. These subgraphs are then parsed into triples using entity linking and semantic reasoning. The triples are converted into natural language text by the LLM, which combines them with the input question to perform reasoning and generate accurate answers. The proposed method addresses the complexity of question answering for electrical standards and offers a novel approach for managing and applying these standards in the field of electrical engineering. Experimental results demonstrate that this approach significantly enhances the model's understanding of electrical standards, enabling it to generate more accurate answers.