REMFLOW: RAG-Enhanced Multi-Factor Rainfall Flooding Warning in Sponge Airports via Large Language Model

The integration of rainfall flooding data, particularly with the development and application of the Storm Water Management Model (SWMM), has significantly enhanced the predictive capabilities of flood warning systems. However, existing models often lack relevant environmental context, leading to the incorporation of external knowledge bases. Previous machine learning approaches have primarily focused on structured knowledge, neglecting unstructured data and high-dimensional semantic information. To address these issues, we propose REMFLOW, a Retrieval-Augmented Generation (RAG) framework designed to improve the analysis and application of multi-factor rainfall flooding data. We employ an embedding model to encode time-series flood data, which is then stored in a knowledge base. Simultaneously, a Large Language Model (LLM) is used to encode multi-factor rainfall flooding data and retrieve relevant knowledge. Subsequently, an adaptive context fusion mechanism is applied to integrate the extracted knowledge and generate the final outputs. Experimental results from various rainfall flooding prediction tasks demonstrate that the REMFLOW framework significantly outperforms baseline models in the precise predictive analysis of multi-factor rainfall flooding data, highlighting its effectiveness.