Offshore wind energy necessitates accurate short-term wind speed predictions for reliable grid operation and efficient wind farm management. Offshore wind speed sequences exhibit considerable non-stationarity and spatiotemporal variability due to multi-scale turbulence, complex marine boundary layers, and nonlinear interactions, which complicate precise forecasting. We introduce GraphDynNet (Graph Dynamic Network), a deep learning framework that concurrently models spatiotemporal continuous dynamics via interconnected modules. The approach utilises VMD for multi-scale decomposition, GAT to identify dynamic spatial correlations among buoy stations, and GRU for modelling temporal dependencies. We offer a Neural ODE module to address the limitations of traditional discrete models by representing wind speed evolution as a continuous dynamic system through differential equations parameterised by neural networks. Experiments conducted at twelve NOAA buoy stations demonstrate that GraphDynNet attains a R² accuracy of 98.5% and decreases RMSE by over 15% compared to leading models. Internal analysis verifies that Neural ODE is essential in reducing discretisation errors, resulting in a 31.4% decrease in RMSE. These findings establish a solid foundation for predicting applications in offshore wind farms.
Dong et al. (Tue,) studied this question.