Comparative performance of AI and numerical models in forecasting typhoon-induced waves

Abstract

Accurately forecasting typhoon-induced waves remains a significant challenge in marine disaster prevention and mitigation. Traditional numerical models, based on the Navier-Stokes equations, face inherent limitations in accurately capturing wave dynamics, particularly under complex conditions like typhoons. These models struggle with wind data inaccuracies and complex ocean topography. In contrast, this study designs a novel AI-driven deep learning model (LSTM-Self Attention-Dense), leveraging four decades of satellite altimeter data to significantly enhance prediction accuracy. Through three deep learning experiments and four numerical simulations, the model's performance is evaluated against traditional methods. The results demonstrate that the deep learning model significantly reduces prediction errors, achieving a 26.63 % reduction in root mean square error (RMSE) and an 87.91 % reduction in bias, particularly in high sea conditions. These findings underscore the clear advantages of AI-driven approaches over traditional numerical models, providing a valuable enhancement for improving the accuracy and reliability of marine forecasting.