A Method for Generating Safety Measures of Hydropower Plant Work Tickets Integrating Knowledge Graph and Large Language Models

Abstract

To address the low efficiency and error-prone nature of manually formulating safety measures in hydropower plant work tickets, as well as the difficulty of existing methods in balancing accuracy and compliance, this paper proposes a safety measure generation method integrating a knowledge graph (KG) and large language models (LLMs). A work ticket KG is constructed to capture historical associations between work contents and safety measures, while a safety regulation vector database is established as a rigid compliance constraint, forming a dual-modal safety knowledge base. Named entity recognition and a weighted similarity retrieval algorithm are employed to obtain relevant subgraphs from the KG. Based on the retrieved context, a locally deployed LLM generates candidate safety measures. Furthermore, a compliance review mechanism, combining opposite-action veto and KG co-occurrence statistics, is introduced to evaluate both compliance and necessity. Experiments on 1,600 test work tickets show that the proposed method achieves a precision of 91.5%, a recall of 87.8%, and a compliance score of 86.7, outperforming all baseline configurations and improving the reliability and automation of work-ticket processing.