Reliable near-surface wind speed (NSWS) data underpin a wide range of research in hydrology, climatology, wind energy, and environmental change. However, estimates of long-term changes in NSWS remain highly uncertain, largely because of incomplete temporal coverage, uneven station distribution, and inhomogeneities in observational records. Here, we release the SgriWind-CN dataset, a new sub-daily gridded NSWS dataset over China, generated from 6-hourly observations collected at 2,384 meteorological stations during 1980–2023. After rigorous quality control and sub-daily homogenization, we apply a deep learning model, PConvUNet, to reconstruct a sub-daily, gridded wind speed dataset. The resulting product effectively reproduces the spatial distribution of NSWS and performs well in data-sparse areas. Quantitative evaluation shows that PConvUNet substantially outperforms conventional interpolation methods: it best preserves the original statistical distribution, as reflected in the lowest Kullback–Leibler divergence, and reduces the mean absolute error on the test set by at least 65.3 % relative to alternative approaches. The model also exhibits an advantage in capturing extreme wind events. A reverse interpolation validation test further confirms the product’s robustness, yielding a national mean bias of less than 0.05 m·s<sup>−1</sup>. Beyond NSWS, the dataset also includes surface roughness, land cover type, elevation, latitude, and longitude to support diverse user applications and independent evaluation. This study aims to provide a robust, high-resolution data foundation for advancing understanding of wind variability and its broader impacts across Earth system sciences. SgriWind-CN is publicly available at <a href="https://doi.org/10.5281/zenodo.22812484" target="_blank" rel="noopener">https://doi.org/10.5281/zenodo.22812484</a> (Bai et al., 2026).