Hybrid deep learning for large-scale spatiotemporal forecasting of significant wave height

• A novel deep learning model was developed to predict large-scale two-dimensional significant wave height fields. • A hybrid architecture was established by combining local feature extraction and global spatiotemporal dependency modeling. • Comprehensive evaluations were conducted under spatial, temporal, extreme weather, and severe sea state scenarios. Large-scale wave prediction is essential for wave energy development, site selection, and operations and maintenance. In this study, a hybrid deep learning model is proposed for the accurate two-dimensional prediction of Significant Wave Height (SWH) over large spatial domains. Multi-scale convolution is employed to extract spatial features, while a convolution-based multi-head attention mechanism is utilized to capture global temporal dependencies. Through end-to-end training, the nonlinear spatiotemporal coupling between wind and wave fields is effectively captured. ECMWF wind and wave reanalysis datasets from 2000 to 2019 over the Northwest Pacific (0-50°N, 100-150°E) were utilized for model validation. Cross-seasonal wave patterns were accurately reproduced, and stable spatial performance was maintained in nearshore, offshore, and complex terrain regions. Furthermore, wave height trends were tracked accurately over time. Reliable predictions were achieved for forecast horizons within 48 h. Under typhoon-driven high-energy sea states, details of wind-wave coupling were represented accurately, and the maximum Root Mean Square Error (RMSE) was limited to 0.384 m. For rough sea states with SWH greater than 4 m, the absolute error of the predicted occurrence probability was found to be below 0.1% in more than 90% of the domain. These results indicate that strong potential for large-scale wave prediction across various scenarios is demonstrated. Consequently, technical support can be provided for wave energy planning, sea state warnings, and routing decisions.