Hybrid wave height forecasting via integrated physics-based simulation and data-driven correction with contrastive feature fusion

Significant wave height forecasting plays a crucial role in marine engineering, navigation safety, coastal protection, and climate research. However, traditional physics-based numerical simulation methods and purely data-driven models each have inherent limitations. The former require significant computational resources for high-resolution, large-scale simulations and demand stringent initial and boundary conditions, while the latter often fail to accurately capture the underlying physical dynamics of wave processes under sparse observational data or extreme sea conditions. To address these challenges, this study proposes an innovative hybrid forecasting framework that integrates physics-based numerical modeling with data-driven approaches, thereby achieving both physical plausibility and high prediction accuracy in wave height forecasting. The framework first employs numerical models such as WAVEWATCH III – which solves the wave action balance equation – to generate preliminary predictions based on fundamental wave dynamics. These predictions are then refined by a data-driven correction module comprising three submodules: the Edge-Enhanced Wind-Wave Fusion (EWF) module, which enhances the input wave height data by incorporating edge information via the Sobel operator and fuses it with the corresponding wind field information to capture critical wind-wave interactions; the Bidirectional Temporal Feature Fusion (BTFF) module, which integrates geographical spatial information and wave activity weights while employing bidirectional temporal propagation to effectively capture both short-term and long-term nonlinear temporal dependencies; and the Wave-Adaptive Filter (WAF) module, which employs adaptive kernel estimation and spatial displacement prediction to extract and enhance multi-scale local wave features. Finally, a contrastive learning-based feature fusion mechanism projects the physics-based and data-driven predictions into a shared latent space, achieving deep integration of complementary information. Experimental results demonstrate that our model has achieved state-of-the-art performance on the ERA5 dataset.