A unified deep learning framework for subseasonal wind forecast post-processing
- 1Japan Agency for Marine-Earth Science and Technology (JAMSTEC)
Abstract
Skillful subseasonal-to-seasonal (S2S) wind forecasts are critical for optimizing marine operations, renewable energy management, and climate risk assessment, but their skill at these lead times is often modest. Here, we develop a unified deep learning–based post-processing framework to improve near-surface SINTEX-F2 wind forecasts over the North Pacific, built on a 3D U-Net backbone. The framework is configured for bias correction, statistical downscaling, and probabilistic prediction at lead times of up to 12 weeks. Compared to statistical and machine-learning baselines, the U-Net removed systematic bias, consistently reduced RMSE, improved correlation, and provided a more realistic representation of forecast variability and wind-speed distributions. Downscaling enhanced spatial coherence at higher resolution without degrading deterministic skill. The framework also improved probabilistic performance, reduced the Continuous Ranked Probability Score (CRPS), and increased the Ranked Probability Skill Score (RPSS) across all tested seasons and lead times. An ordinal classification approach provided the highest categorical skill, whereas regression-based ensemble correction recovered most of the probabilistic gain and retained physically coherent wind fields for downstream applications. In an additional benchmark across five forecast systems, the U-Net reduced week-3–4 North Pacific RMSE and CRPS by 13% and 15%, respectively. These results show that a unified deep learning framework can improve the skill and usability of S2S wind forecasts.
Related
- Subseasonal forecast improvements from sea ice concentration data assimilation in the Antarctic
- The role of the oceans for subseasonal prediction: insights from eddy-permitting and eddy-rich coupled forecast systems
- Seasonal hydropower forecast framework using a combination of deep learning and physics-based methods
- Advancing subseasonal forecasting with machine learning.
- Uncertainty analysis of a machine learning-aided ensemble forecast for coastal flood maps
BibTeX
@article{damiani2026unified,
title = {{A unified deep learning framework for subseasonal wind forecast post-processing}},
author = {A. Damiani and Y. Miyazawa and M. Nonaka and J. V. Ratnam and K. R. Patil},
journal = {Scientific Reports},
year = {2026},
month = jul,
doi = {10.1038/s41598-026-63299-0},
url = {https://doi.org/10.1038/s41598-026-63299-0},
note = {CC BY},
}
RIS
TY - JOUR
TI - A unified deep learning framework for subseasonal wind forecast post-processing
AU - A. Damiani
AU - Y. Miyazawa
AU - M. Nonaka
AU - J. V. Ratnam
AU - K. R. Patil
PY - 2026
DA - 2026/07/01
JO - Scientific Reports
DO - 10.1038/s41598-026-63299-0
UR - https://doi.org/10.1038/s41598-026-63299-0
AB - Skillful subseasonal-to-seasonal (S2S) wind forecasts are critical for optimizing marine operations, renewable energy management, and climate risk assessment, but their skill at these lead times is often modest. Here, we develop a unified deep learning–based post-processing framework to improve near-surface SINTEX-F2 wind forecasts over the North Pacific, built on a 3D U-Net backbone. The framework is configured for bias correction, statistical downscaling, and probabilistic prediction at lead times of up to 12 weeks. Compared to statistical and machine-learning baselines, the U-Net removed systematic bias, consistently reduced RMSE, improved correlation, and provided a more realistic representation of forecast variability and wind-speed distributions. Downscaling enhanced spatial coherence at higher resolution without degrading deterministic skill. The framework also improved probabilistic performance, reduced the Continuous Ranked Probability Score (CRPS), and increased the Ranked Probability Skill Score (RPSS) across all tested seasons and lead times. An ordinal classification approach provided the highest categorical skill, whereas regression-based ensemble correction recovered most of the probabilistic gain and retained physically coherent wind fields for downstream applications. In an additional benchmark across five forecast systems, the U-Net reduced week-3–4 North Pacific RMSE and CRPS by 13% and 15%, respectively. These results show that a unified deep learning framework can improve the skill and usability of S2S wind forecasts.
SN - 2045-2322
ER -