A Swin-Transformer-based deep-learning model for rolled-out predictions of regional wind waves.

This article presents the development and evaluation of ST-RWP (Swin Transformer for Regional Wave Prediction), a novel deep-learning model combining convolutional neural networks (CNNs) and Swin Transformer layers to forecast short-term regional wind wave fields. Trained on the ERA5 reanalysis dataset over the North Atlantic Ocean, ST-RWP effectively captures spatiotemporal dependencies in wind velocities and significant wave heights, achieving accurate predictions up to 12 hours ahead using a rolled-out prediction scheme. The model demonstrates superior performance compared to existing Vision Transformer and CNN-LSTM approaches, particularly due to its inductive bias that leverages strong local and global correlations in the data. Limitations include reduced accuracy beyond 24-hour lead times and higher errors near domain boundaries, attributed to diminished temporal correlations and limited spatial information, respectively. This study highlights the potential of Transformer-based architectures for efficient real-time wave forecasting and provides insights for future improvements in long-term prediction and boundary condition integration.