A sliding EOF-based deep learning framework for spatiotemporal forecasting of wind and wave fields

Accurate wind and wave field forecasting is important for offshore operations and coastal disaster prevention. This study develops a spatiotemporal forecasting framework that couples sliding empirical orthogonal function (EOF) decomposition with deep learning. The proposed sliding EOF module performs decomposition only within historical moving windows, thereby avoiding future information leakage and allowing the EOF representation to respond to locally evolving wind–wave variability. A temporal forecasting network (TFnet) predicts the principal components of the dominant modes, while a residual compensation network (RCnet) recovers fine-scale spatial information lost through EOF truncation. Using ERA5 wind and wave fields from 2001 to 2024 in the northwestern Pacific Ocean, the framework was evaluated at 3, 6, 12, and 24 h lead times. Compared with the strongest deep learning baseline among CNN-Transformer, U-Net, ConvLSTM, and ResNet, the proposed method reduced RMSE by 16.8%–39.6% across variables and lead times while maintaining higher correlation coefficients. Case studies further illustrate that the framework can reproduce the dominant evolution of wind and wave fields during the selected cyclone periods, although local high-intensity conditions and longer-lead wind forecasts remain challenging. These results support the effectiveness of sliding EOF-based spatiotemporal decoupling for regional wind and wave field forecasting.