Large Language Model-Empowered Adversarial Fusion for Typhoon Track Prediction

Abstract

Accurate prediction of typhoon tracks is essential for effective disaster prevention strategies. Given that typhoon tracks can be conceptualized as a special class of time series, promising results have been achieved via the learning transferability of large language models (LLMs) in time series forecasting. However, limitations primarily lie in the incomplete utilization of the temporal and channel features of typhoon tracks. To address this, we propose LAF, an LLM-empowered framework for typhoon track prediction with adversarial fusion. We begin by transferring knowledge from pretrained LLMs to typhoon tracks to capture temporal features. To exploit purer channel features, we design a channel feature extraction strategy to limit the noise introduced. Then, we implement adversarial fusion between temporal and channel features to effectively narrow their discrepancy, facilitating a more comprehensive fusion for typhoon track prediction. Evaluations on Northwest Pacific typhoon track data demonstrate the effectiveness of the proposed model.