An improved Transformer-based deep learning approach enhances typhoon track and intensity prediction across China

    Accurate multi-step typhoon track and intensity prediction remains a critical yet hightly challenging task due to complex nonlinear atmospheric dynamics and strong spatiotemporal variability of typhoon evolution. Traditional statistical methods and numerical modeling approaches often struggle to sustain predictive accuracy, particularly for medium- to long-range forecasts. In this study, we propose a Transformer-based deep learning approach to jointly predict typhoon trajectories and maximum wind speed using multi-source meteorological data. The model is trained on historical best-track of typhoon datasets across China and reanalysis data from 1949 to 2023, incorporating key environmental variables to characterize large-scale atmospheric conditions that modulate typhoon development. All input features are subjected to rigorous quality control, normalization, and temporal alignment, and are structured  sequential samples to capture the temporal evolutionary  dynamics of typhoon systems. The Transformer architecture is adopted to effectively model long-range dependencies and complex spatiotemporal interactions inherent in typhoon evolution.

   To quantitatively evaluate the performance of the proposed approach, two baseline models, XGBoost and Long Short-Term Memory (LSTM), are implemented for comparison. Experimental results demonstrate that the proposed model significantly outperforms both baselines across all evaluated forecast lead times. At short lead times, the model achieves RMSE values of 200–244 km for track prediction, corresponding to a 30% reduction relative to LSTM and a 58% reduction compared to XGBoost. At longer lead times (48h), the model maintains robust and competitive performance, with track forecasting RMSE ranging from 273–480 km, achieving up to 26% lower errors than LSTM and 52% reduction relative to XGBoost. For intensity prediction, the model also shows strong performance, with an MAE of 2.37 m/s, RMSE of 3.12 m/s, and R² of 0.68, indicating strong agreement with best-track observational records.

    Overall, our results confirm that the proposed framework effectively captures the complex spatiotemporal dynamics of typhoon systems, providing marked improvements in predictive accuracy, stability, and robustness. This framework holds strong application potential for operational typhoon forecasting and typhoon-induced disaster risk management.