Wave forecast in the Atlantic Ocean using a double-stage ConvLSTM network

Accurate forecasting of ocean waves is of great importance to the safety of marine transportation. Despite wave forecasts having been improved, the current level of prediction skill is still far from satisfactory. Here, the authors propose a new physically informed deep learning model, named Double-stage ConvLSTM (D-ConvLSTM), to improve wave forecasts in the Atlantic Ocean. The waves in the next three consecutive days are predicted by feeding the deep learning model with the observed wave conditions in the preceding two days and the simultaneous ECMWF Reanalysis v5 (ERA5) wind forcing during the forecast period. The prediction skill of the d-ConvLSTM model was compared with that of two other forecasting methods—namely, the wave persistence forecast and the original ConvLSTM model. The results showed an increasing prediction error with the forecast lead time when the forecasts were evaluated using ERA5 reanalysis data. The d-ConvLSTM model outperformed the other two models in terms of wave prediction accuracy, with a root-mean-square error of lower than 0.4 m and an anomaly correlation coefficient skill of ∼0.80 at lead times of up to three days. In addition, a similar prediction was generated when the wind forcing was replaced by the IFS forecasted wind, suggesting that the d-ConvLSTM model is comparable to the Wave Model of European Centre for Medium-Range Weather Forecasts (ECMWF-WAM), but more economical and time-saving.