HIWindCast: A Deep Learning Model for Operational High-Resolution Tropical Cyclone Wind Forecasting

Accurate and timely forecasting of tropical cyclone (TC) winds remains a critical challenge. We introduce HIWindCast, a novel deep learning model that integrates advanced AI weather prediction (AIWP) models with high-resolution satellite observations to overcome key limitations in current methods. Traditional numerical models are costly, and AIWP models trained on coarse (~25 km) reanalysis data often underestimate TC intensity. HIWindCast employs a specialized algorithm to simultaneously downscale and bias-correct AIWP forecasts, using high-resolution (~6 km) satellite-derived wind observations for training. A principal advantage of HIWindCast is its broad applicability: it can be seamlessly applied to any AIWP model trained on ERA5 data. During Typhoon Ragasa (2025), HIWindCast reduces maximum wind speed error by up to 57% and radius of maximum wind error by up to 70% across three leading AIWP models. This advancement provides a crucial tool for operational forecasting, delivering timely and reliable predictions essential for mitigating TC impacts.