A coupled Swin transformer-LSTM network for high-resolution ocean wave forecasting: A reanalysis-driven skill assessment in the Chinese marginal seas

Accurate wave forecasting is essential for maritime safety and provides crucial scientific guidance for coastal operations and planning. Most artificial intelligence-based wave forecast models were conducted at coarse resolutions, e.g., 0.25° or 0.5° spatial resolution, and struggled to maintain high forecasting accuracy for extended periods. Here, we introduce the coupled Swin Transformer-LSTM network (SwinLSTM), a hybrid architecture designed to make a spatiotemporal forecast of significant wave height (SWH) at a 0.1-degree resolution over 72-h lead-time in the Bohai Sea, Yellow Sea, and East China Sea. In this study, both historical and lead-time wind fields are taken from the ERA5 reanalysis; therefore, the reported skill reflects a reanalysis-driven (hindcast-style) evaluation that provides an upper-bound estimate under near-perfect wind forcing. The SwinLSTM architecture effectively captures spatial dependencies, simultaneously extracting both long-term and short-term spatiotemporal dependencies in ocean wave dynamics for efficient two-dimensional spatial forecasting. Through sensitivity experiments, the optimal configuration was determined, with historical wind, SWH, topography, and ERA5 reanalysis future wind (used here as a proxy forcing for lead-time prediction) identified as the optimal input combinations using a 6-h encoding time step. Based on comprehensive model evaluation with this optimal configuration, our results demonstrate that for forecast horizons of 1-, 6-, 12-, 24-, 48-, and 72-h, the spatially averaged root mean square error (RMSE) values are 0.113, 0.121, 0.155, 0.190, 0.221, and 0.232 m, respectively, with corresponding spatial correlation coefficients (CC) of 0.989, 0.987, 0.980, 0.972, 0.963, and 0.960. For forecast lead times longer than 12-h, comparisons show that our model is among the best ones in AI-based wave models, showing high prediction accuracy while maintaining satisfactory stability and robustness across different temporal scales. The wave forecast capability and robustness were validated under conditions of cold air outbreaks and typhoon events, demonstrating the model's ability to capture the spatial distribution and temporal evolution of extreme wave events. These findings demonstrate the potential for high-resolution SWH forecasting with enhanced accuracy and efficiency.