A Transformer-Based Sea Surface Wind Error Correction Method

Abstract: In order to reduce the sea surface wind field (SSW) prediction errors in numerical weather prediction (NWP) due to the uncertainties in the initial physical parameters and physical process approximations, this paper proposes an SSW correction model based on a converter architecture that combines adaptive partitioning, overlapping patch segmentation, and spatiotemporal attention mechanisms. The model can effectively capture local changes in wind speed and direction, mitigate boundary effects, and enhance the model's ability to learn spatio-temporal features. By learning the forecast data and error data of the past three days, the model can understand the evolution of the error and predict the future error trend, thus reducing the SSW forecast error and improving the forecast accuracy of the numerical model. The experimental results in the Pacific Northwest region show that the calibration model performs well in the short-term calibration (24-hour forecast) and significantly reduces the SSW root mean square error (RMSE). Although the correction effect of the 48- and 72-hour forecasts is reduced, the model still maintains good stability, adaptability, and reliability.