Improvement of Near-Surface Wind Speed Simulations over China Using Super-Resolution Convolutional Neural Network

Abstract

Global climate models (GCMs) still show limited skill in reproducing near-surface wind speed (NSWS) over mainland China, especially in regions with complex terrain. In this study, we applied an encoder–decoder super-resolution convolutional neural network as a statistical post-processing framework to reconstruct daily NSWS from 14 CMIP6 models to a finer 0.25° grid, using CN05.1 as the reference dataset. The model was trained on the historical period and evaluated over 2000–2014, and its performance was compared with the raw CMIP6 ensemble and NEX-GDDP-CMIP6. The downscaled ensemble substantially improved the simulation of NSWS over mainland China. During 2000–2014, it achieved a regional mean bias of −0.07 m·s −1 and a spatial correlation coefficient of 0.93 for annual mean NSWS, outperforming both CMIP6 and NEX-GDDP-CMIP6. It also better reproduced seasonal patterns, interannual variability, and the spatial distribution of recent changes. Applied to future SSP-RCP scenarios, the downscaled projections indicate a weak but statistically significant decline in area-mean NSWS over the 21st century, with larger decreases under stronger forcing. However, the projected China-wide mean during the next two decades remains close to the 1980–2014 baseline, while localized increases persist in parts of eastern and northeastern China. These results show that super-resolution CNN methods can provide useful post-processing for CMIP6 NSWS over China, although future work should assess the added value of static predictors such as topography and land cover and further evaluate extreme wind behavior.