A Dual-Stage Retrieval-Augmented Knowledge Generation Approach for Intelligent Power Operation and Maintenance

Large Language Model (LLM) is an important component of natural language processing. With the rapid development of artificial intelligence application, LLMs have achieved excellent results through vertical domain integration with different fields. However, in the domain of power operation and maintenance, due to the vast amount of specialized knowledge, frequent content updating etc, there are numerous challenges for the LLMs to generate specialized knowledge of power operation and maintenance. Currently, researchers have attempted to use general LLM to build power operation specialist, but it has limited interaction with users, which cannot address the collaboration issues. At the same time, the range of power operation questions is extremely broad, which poses challenges to the judgment and processing capabilities of LLM. To address these problems, we propose a Dual-Stage Retrieval-Augmented Knowledge Generation Approach for Intelligent Power Operation and Maintenance called PoD-RAG. This method introduces dual-stage retrieval-augmented generation technology, ensuring the reliability and usability of the generated content for power operation and maintenance. In the first stage, PoD-RAG exploits vertical question-answer pairs for supervised fine-tuning, assuring the professionalism and compliance of the LLM. In the second stage, the method provides answers through RAG, then uses a self-reflection mode to allow the LLM to autonomously judge the generated content, balancing the security and professional needs of power operation and maintenance. To validate the effectiveness of the method, we conducted experiments using professional data and knowledge base from power operation and maintenance. The results indicate that PoD-RAG is highly stable in generating professional responses in the power domain, providing support for power infrastructure construction.