Spatial-temporal Offshore Current Field Forecasting Using Residual-learning Based Purely CNN Methodology with Attention Mechanism

Abstract

Spatial-temporal current forecasting is indispensable for ocean engineering and marine science exploration, for instance aiding in the conservation and protection of marine ecosystems, planning shipping-routes and determining the length and fuel consumption of sea-going voyages, obtaining deeper insights into the distribution of heat flux within the ocean, which is vital for better understanding climate changes, and so on. Most present related-studies primarily focused on single location or grid-cell-based forecasting, such methodologies are site-specific and neglect the importance of spatial-temporal fidelity. Furtherly, the Recurrent Neural Networks-based methods previously employed exhibit low efficiency in terms of model convergence concerning practical engineering purposes, and numerical weather models are time-consuming and computational expensive. A newly improved Unet-based model using residual-learning with attention strategy is proposed for 2D sea surface current (SSC) velocity predictions with a more efficient perspective. Several machine-learning methodologies were adopted for a better performance comparison. The final predictions demonstrated its superiorities that the proposed neural-learning method outperforms the other established approaches with spatial-resolved mean RMSE less than 0.009 m/s and 0.006 m/s. As a promising surrogate for SSC predictions, the proposed methodology has strong potential in operation marine monitoring and engineering constructions.