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

Abstract

Short-term predictions of regional wind waves are crucial for coastal and ocean engineering. In this study, we introduce a novel Swin-Transformer-based model, named ST-RWP (Swin Transformer for Regional Wave Prediction), designed to leverage the spatiotemporal relationships of wind velocities and significant wave heights. The model considers inductive bias to capture both local and global dependencies via Convolution and Swin Transformer layers, enabling accurate short-term wave field predictions on unseen data. A rolled-out prediction scheme is employed to extend the forecast horizon efficiently. Trained on the reanalysis dataset offered by European Center for Medium-Range Weather Forecasts, ST-RWP demonstrates excellent performance in predicting wave fields with lead times of 6 and 12 h. However, the model's accuracy degrades when the lead time exceeds 24 h, primarily due to the limited spatial information available at boundary nodes and the low autocorrelation value for such large time span. The dataset exhibits strong spatial and temporal correlations, which are key to the model's success. Our findings indicate that ST-RWP offers an efficient tool for real-time wave field nowcasting, representing a significant advancement in the application of Transformer-based deep neural networks to wave prediction.