Abstract: This study addresses the over-reliance on expert scoring, procedural complexity, and strong subjectivity inherent in the Analytic Hierarchy Process(AHP)when determining evaluation index weights and assigning membership degrees to qualitative indicators for large-span underground powerhouse site selection. An innovative site selection decision-making framework integrating knowledge graphs and AHP is proposed. The framework leverages extensive historical case documents as its data foundation and constructs a geological risk knowledge graph using a large language model. Through deep relational knowledge mining, it enables intelligent calculation of indicator weights. The framework further employs LLM-based semantic reasoning to compute membership degrees for qualitative indicators, forming a comprehensive site selection decision-making system. Results indicate that in engineering applications such as the Caiziba Pumped Storage Power Station in Fengjie, Chongqing, the evaluation outcomes of this method align with those of traditional AHP,while the efficiency of the decision-making process improves by over 80%. The main innovation of the study lies in integrating LLM and KG into AHP:using knowledge graphs to derive weights instead of relying on subjective scoring, optimizing qualitative indicator membership degrees through LLM-based semantic reasoning, significantly enhancing efficiency, and providing a scientifically robust and practical technical solution for site selection of large-span underground powerhouses.