This paper introduces an advanced interactive disaster training scenario generation system powered by Large Language Models (LLMs) to address the critical issues in scenario-based disaster training such as the need for expert knowledge, formalization of the scenario and time consumption. By adopting Retrieval-Augmented Generation (RAG) and the chatbot-based interface as a Human-in-the-Loop (HITL) enables the generation of customised scenarios for different types of disasters and regional characteristics while ensuring the quality. The feasibility of the proposed method is tested by analysing the generated scenarios and interviewing users. We have confirmed that it is possible to generate scenarios tailored to specific disaster and regional characteristics by referring to actual training scenarios and geographical information, using RAG methods to provide domain knowledge specific to disaster response, which vanilla LLMs do not have. Future enhancements will incorporate dynamic real-time data and detailed user feedback to further optimize system performance and applicability in disaster response training.