Ocean weather prediction is crucial for various applications, such as global climate prediction, marine environmental protection, and offshore production. However, current data-based marine weather prediction methods have limitations when predicting multiple variables in a particular area, failing to meet the efficiency and accuracy requirements of practical applications. In the realm of ocean weather variations, the presence of highly interconnected spatial and temporal continuations, coupled with the mutual influence of individual variables, underscores the utmost importance of effectively capturing dynamic correlations encompassing space, time, and variables to accurately predict ocean weather. To address this, we developed a novel approach called StHCFormer, which is a multivariate spatiotemporal hybrid convolutional attention network. The first key component of StHCFormer is the spatiotemporal hybrid convolutional attention (StHCA) module, which leverages a hybrid convolutional attention mechanism to explore both global spatial representations and local features. Additionally, the module incorporates temporal attention to capture the temporal dependence of weather records and effectively captures the dynamic correlations among multiple variables through channel deflation and weighted residuals. To ensure balanced variable losses, we introduced the concept of homoscedasticity uncertainty loss to dynamically adjust the multitask weights. This guarantees a global optimal solution and leads to more accurate multivariate ocean weather prediction. Finally, we conducted a comprehensive evaluation and comparison of the StHCFormer model with other state-of-the-art algorithms using the ERA5 dataset in the Philippine Sea. The results demonstrated that StHCFormer outperforms existing methods in marine multivariate field weather prediction.
1
1 January 2026 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Wind speed prediction is an essential spatiotemporal forecasting task in wind energy systems, yet it remains challenging due to the nonlinear and dynamic characteristics of atmospheric processes. The evolution of wind is governed by physical laws, which can be effectively described using partial differential equations (PDEs). To improve forecasting reliability and accuracy, this paper proposes a novel network model, termed DynWindNet, which integrates equation-based dynamics with data-driven dynamics within a unified framework. Specifically, an interactive dual-branch architecture is designed, where a Physics–Data Coupling Module (PDCM) enables adaptive information exchange between the two dynamics via attention-based gating mechanisms. In addition, a frequency-aware enhancement module (FAEM) is introduced to refine the representations of the data-driven branch by selectively emphasizing informative frequency components. Experimental results on the ERA5 dataset demonstrate that DynWindNet consistently outperforms representative baseline methods across atmospheric pressure levels. Overall, the proposed framework provides an effective approach for integrating physics-guided evolution modeling with deep spatiotemporal representation learning in wind field forecasting.