In order to address the inherent limitations of traditional analysis methods—such as insufficient integration of professional knowledge and weak interpretability of causal reasoning—when dealing with the complex characteristics of multi-factor nonlinear interactions in power systems, a large model for power production safety accident analysis was proposed that integrates LLM, RAG, and KG. A framework with four core modules was built: knowledge retrieval, knowledge reasoning, answer generation, and performance evaluation. RAG technology was used to accurately retrieve relevant knowledge from professional texts, and KG was leveraged for structured reasoning on accident entities and relationships to make up for retrieval blind spots. Finally, LLM was employed to generate professional and interpretable answers for accident causal analysis. The study comprehensively evaluated the system through subjective expert scoring and objective metrics like ROUGE and BLEU, and results show that in the scenario of power production safety accident analysis, the knowledge enhancement technology of RAG and KG provides universal performance improvement for basic models with a certain parameter scale—it helps models accurately capture professional correlations such as equipment fault transmission chains and enhances the quality of accident cause mining and result evolution reasoning. Large models including DeepSeek-R1 and Qwen2.5-72B significantly improved in the accuracy of parsing professional terms and organizing multi-factor correlations under this mode, among which DeepSeek-R1 achieved a comprehensive score of 4.05, better meeting the precision requirements of the field; meanwhile, there is a model capability threshold for the enhancement effect: after enhancement, Qwen2.5-72B can efficiently parse complex logics like cross-regional power grid fault linkage, balances performance and deployment costs, and is suitable for enterprises' practical needs, while smaller models such as Qwen2.5-14B, due to limited basic reasoning capabilities, fail to process professional information effectively after introducing external knowledge, leading to performance degradation and inability to meet professional requirements.