WaveResNet: Deep residual learning with dynamic predictable-zone constraint for deterministic ocean wave forecasting

Reliable short-term wave forecasting is essential for ensuring the safety and efficiency of offshore operations. Traditional forecasting methods based on linear wave theory or numerical solvers are either limited by the linear assumption or high computational costs when the measured wave field is not a Gaussian process. In this study, WaveResNet is presented as a deep learning framework designed to address this challenge. The model is capable of producing high-fidelity wave time series forecasts at the target position, using a single spatial snapshot of surface wave elevation measurements as a data source. Firstly, a spatio-temporally evolving theoretical predictable zone is derived, adapted to the measurement area. This provides a physics-constrained domain for reliable forecasts with data-driven models. Second, WaveResNet is established and it employs deep residual learning to forecast and further refine the result. It is trained using a novel dynamic loss modulation strategy that emphasizes the theoretical predictable zone, guiding the model to focus on physically realistic regions for wave forecasting. The proposed WaveResNet model demonstrates excellent performance across diverse sea states, with promising application possibilities for wave forecasting to facilitate different marine operations.