A Study on Knowledge-Based Chain-of-Thought Reasoning Methods for Power Information Operation and Maintenance

The operation and maintenance (O&M) of power information systems have become increasingly complex due to the integration of diverse technologies. This paper explores the application of Knowledge-Based Chain-of-Thought (CoT) reasoning methods to improve O&M processes. While In-Context Learning (ICL) and CoT prompting techniques enhance reasoning performance in large language models (LLMs), hallucinations—where models generate outputs that contradict commonsense or facts—remain a challenge, often due to insufficient background knowledge. Pretraining or fine-tuning with knowledge graphs can mitigate this, but such methods are impractical for closed-source models like ChatGPT or GPT-4. To address this, we focus on improving inference by integrating external knowledge during reasoning. Preliminary experiments show that contextual entity knowledge and prior knowledge from pretraining data are key factors. Based on these findings, we propose a knowledge-augmented prompting method that incorporates a knowledge-driven CoT reasoning framework, enhancing the generation of structured reasoning paths. Our evaluation of fidelity and factual consistency demonstrates that this approach significantly improves the reliability and scalability of traditional O&M methods.