Energy-Conserving Deep Learning Model for Accurate Ocean Wave Forecasting

Abstract Traditional wave forecasting models, despite being based on energy conservation equations, are computationally expensive. Existing deep learning models, while efficient, often suffer from energy dissipation in long-term forecasts. This paper proposes OceanCastNet (OCN), a novel energy-balanced deep learning wave forecasting model. OCN maintains energy balance by incorporating wind fields at current, previous, and future time steps, along with wave fields at current and previous time steps. Experiments demonstrate that OCN achieves short-term forecast accuracy comparable to traditional models, outperforming the widely used WaveWatch III model. The importance of energy constraints for improving long-term forecast performance is confirmed through a simple meteorological model, OCN-wind, highlighting the generalizability of the energy conservation principle in deep learning atmospheric models. This finding provides new ideas for future research on deep learning geophysical fluid models.