Advancing Tropical Cyclone Intensity Forecasts Over the South China Sea With a Meteorology‐Aware and Attention‐Enhanced Deep Learning Model

Abstract Accurate tropical cyclone (TC) intensity (TCI) prediction is critical for effective disaster preparedness. Although machine learning‐based weather prediction models demonstrate strong performance in simulating TC tracks, their TCI predictions show systematic biases. To overcome this limitation, this study proposes TC‐FANet, a deep learning framework built upon a temporal convolutional network (TCN) backbone, incorporating a feature‐aware self‐attention (FSA) mechanism that dynamically reweights atmospheric features according to TCI changes. TC‐FANet achieves an overall mean MAE of 5.90 m s −1 across all lead times (24–120 hr), representing a 46.02% reduction compared with all evaluated baseline categories, and delivers a 40.23% improvement over deep learning baselines during rapid intensification (RI) phases. SHAP analyses revealed that the FSA module consistently elevates RI‐critical controls (tropical cyclone fullness) while down‐weighting redundant or persistence‐like signals, revealing nonlinear, interaction‐driven determinants of TCI. These findings highlight the potential of DL architectures to provide interpretable insights for operational TCI forecasting.