Nowcasting of typhoon tracks based on LLM and RAG

Abstract Accurate typhoon track nowcasting is vital for navigation and coastal disaster prevention. This research integrates a Large Language Model (LLM) with Retrieval-Augmented Generation (RAG) technology for typhoon path prediction. Leveraging LLMs as the predictive foundation, the approach tailors forecasts to individual typhoon characteristics. The methodology involves collecting satellite imagery, standardizing data, and employing optical flow methods to track typhoons and derive path coordinates. These coordinates are preprocessed and embedded into the LLM. RAG enhances the LLM’s predictive performance, enabling effective forecasting. Increasing typhoon-specific embedded data further improves accuracy. Using the FY-4 dataset, the method achieved an average absolute error of 10.78 km in 12-hour predictions. The study demonstrates that LLM-RAG integration excels in nowcasting.