WaveUformer: a bias correction model for GWSM4C Wave Forecasting

Artificial intelligence (AI) models are being progressively applied to the field of wave forecasting. However, in operational forecast scenarios, these data-driven models exhibit error characteristics different from those of numerical models due to factors such as uncertainties in the driving wind fields. Traditional correction methods have limited capability to correct these data-driven biases, particularly for medium- to long-range forecasts and extreme sea states. To address this issue, this study proposes a deep learning-based post-processing correction model, WaveUformer, specifically designed to correct the forecast results of the AI wave model Global Wave Surrogate Model for Climate simulation (GWSM4C). The model synergistically processes driving wind field data and forecast wave field data, and integrates an adaptive correction mechanism based on forecast lead time with an efficient spatiotemporal attention network to effectively capture the dynamic evolution patterns of errors. Evaluation based on the full-year test data of 2023 shows that WaveUformer reduces the annual mean root mean square error of 24-240-hour significant wave height forecasts from 0.57 m to 0.39 m, achieving an overall relative improvement of 31%. In the case analysis of Typhoon, the model successfully corrected the underestimation bias of extreme conditions and accurately reproduced the spatial structure of high-wave areas. The results demonstrate that WaveUformer can reduce the forecast errors of AI models, improving their forecast accuracy and reliability.