Improvement of operational wave forecasts in the Western North Atlantic Ocean with a transformer model

This study investigates the enhancement of daily significant wave height (SWH) forecasts in the Western North Atlantic by integrating the Global Ensemble Forecast System (GEFS) with a Transformer-based model. The model is designed to enhance GEFS 10-day wave forecasts daily, utilising forecasted GEFS wind and wave parameters as input features, while buoy data is used only during the training and validation stages. Using buoy data from four National Data Buoy Center (NDBC) stations (41040, 41044, 41048, and 41049), the model is trained on 8 months of data, validated over 2 months, and tested on a 2-month dataset. The accuracy of the Transformer model is assessed on a day-by-day basis for up to 10 days, with notable improvements in forecast accuracy, particularly at station 41044. For this station, the Transformer model shows an overall improvement of over 40 % compared to GEFS. In contrast, the worst performance is observed at station 41048, where the Transformer model achieved a 22 % improvement over 10 days, with a MAPE of 14.06 % and an SI of 0.221. On the 10th day, the Transformer model outperforms GEFS, though the accuracy generally decreased over the forecast period. The results indicate that GEFS wind speed and GEFS SWH are critical features for improving SWH forecasts. These findings highlight the potential of hybrid forecasting approaches to enhance both short- and long-term operational wave forecasts, even in cases where measured data is limited to relatively short periods. • This study investigates the enhancement of daily significant wave height GEFS forecasts in the Western North Atlantic. • A Transformer model is used day-by-day for up to 10 days forecast, and its accuracy is checked with buoy data. • The model is trained on 8 months of data, validated over 2 months, and tested on a 2-month dataset. • On the 10th day, the Transformer model outperforms GEFS, though the accuracy generally decreases over the forecast period.