Accurate projection of wind speed is essential for assessing wind energy resources and strategically planning future energy development. However, current Global Climate Models (GCMs) display discrepancies between simulated and observed near-surface wind speed (NSWS), which limits the precise evaluation of wind energy potential. Inspired by the principles of transfer learning, this paper employs a neural network model, referred to as WindNet, to project changes in NSWS across China. WindNet learns the characteristics of NSWS through self-supervised pretraining and is subsequently fine-tuned to function as an NSWS reconstruction model. The results show that the WindNet model effectively captures the spatiotemporal features of NSWS compared to GCMs outputs, particularly in high-altitude regions of China, thereby enhancing spatial resolution and prediction accuracy. The root-mean-square error (RMSE) was reduced to 0.5 m