Transformer-based full-track simulation of tropical cyclones

Abstract

Full-track simulation of tropical cyclones (TCs) is the most widely adopted approach to estimate wind speeds for structure design and TC hazard assessment. This study presents two Transformer-based models, TCformer and eTCformer, for TC full-track simulation. TCformer is designed to directly simulate the entire TC track, aligning with the statistical approach, while eTCformer incorporates large-scale environmental factors such as wind velocity components and vertical wind shear into the Transformer framework, corresponding to the statistical-dynamical approach. The primary advantages of the two DL models lie in their efficiency and flexibility, particularly their ability to capture complex nonlinear interactions among arbitrary variables. Comparative analysis demonstrates that Transformer-based models outperform traditional methods across multiple metrics, exhibiting enhanced robustness and better preservation of spatial correlation patterns, especially in intensity simulations. Furthermore, the accuracy of virtual TC landfall simulations and their evaluated impacts on key coastal cities also underscores the Transformer-based models potential to advance typhoon risk analysis. Overall, the proposed models demonstrate potential as robust tools for reliable TC sample modeling, as well as enhanced understanding of behavioral patterns and improved assessment of both climate change impacts and disaster risks.