Presentation Information
[Poster 1-02]Toward generalizable deep-learning post-processing for seasonal and subseasonal wind forecasts
*Alessandro Damiani1 (1. Japan Agency for Marine-Earth Science and Technology (JAMSTEC) (Japan))
Reliable atmospheric forcing is essential for marine forecasting, offshore decision support, and operational ocean digital-twin systems. However, seasonal and subseasonal wind forecasts often contain systematic biases, imperfect representations of spatial variability, and limited probabilistic reliability, restricting their direct use in downstream ocean applications.
Here, we assess the generalizability of a previously developed deep-learning post-processing framework across different forecast systems and end-use requirements. Built on a single three-dimensional U-Net backbone, the framework employs two formulations. Regression-based ensemble correction produces continuous, spatially coherent wind fields for downstream physical applications, whereas ordinal probabilistic prediction provides direct information for risk assessment.
We evaluate both formulations across multiple subseasonal prediction systems and compare their performance with conventional statistical methods, including quantile mapping, and machine-learning baselines. Preliminary results show that the framework generally improves forecast skill and usability, although the gains vary across post-processing objectives, forecast leads, and prediction systems. This variability highlights the value of a flexible framework that can be configured for different forecast products and end uses.
Overall, the results demonstrate the potential of deep-learning post-processing to provide more reliable, application-ready atmospheric forcing for marine prediction and ocean digital-twin applications.
Here, we assess the generalizability of a previously developed deep-learning post-processing framework across different forecast systems and end-use requirements. Built on a single three-dimensional U-Net backbone, the framework employs two formulations. Regression-based ensemble correction produces continuous, spatially coherent wind fields for downstream physical applications, whereas ordinal probabilistic prediction provides direct information for risk assessment.
We evaluate both formulations across multiple subseasonal prediction systems and compare their performance with conventional statistical methods, including quantile mapping, and machine-learning baselines. Preliminary results show that the framework generally improves forecast skill and usability, although the gains vary across post-processing objectives, forecast leads, and prediction systems. This variability highlights the value of a flexible framework that can be configured for different forecast products and end uses.
Overall, the results demonstrate the potential of deep-learning post-processing to provide more reliable, application-ready atmospheric forcing for marine prediction and ocean digital-twin applications.
Comment
To browse or post comments, you must log in.Log in