A novel depthwise separable U-Net for large-scale wave field prediction

Accurate spatial-temporal ocean wave energy forecasting is critical for advancing global carbon neutrality and clean energy sustainability. While deep neural networks alleviate computational burdens of numerical weather models, prior approaches relying on location-specific or grid-cell samples neglect spatial-temporal correlation and nonlinear dynamics in large 2D wave fields. Recurrent Neural Networks (RNNs) further suffer from poor convergence. To address these gaps, this work proposes a spatial-temporal depthwise separable U-Net model integrating attention mechanisms, residual learning blocks, and depthwise separable convolutions. The U-Net architecture captures multi-scale spatial patterns and propagates energy dynamics across 2D fields, while attention modules prioritize regions of nonlinear interactions (storm zones). Residual blocks stabilize temporal modeling, learning long-term dependencies and abrupt shifts (weather changes), and depthwise separable convolutions efficiently fuse spatial-temporal features, reducing redundancy while preserving variability. The model achieves an RMSE of 0.09 m (1-h) and 0.43 m (12-h) for significant wave height predictions, with PCC of 0.97 and 0.83, respectively. Spatial-averaged RMSEs are 0.01 m (1-h) and 0.09 m (12-h). Experiments demonstrated that the novel model can provide great potential and guidance for operational marine monitoring and renewable energy-based marine constructions.