Wind plays a crucial part during adverse events, such as storms and wildfires, and is a widely leveraged source of renewable energy. Predicting long-term daily local wind speed is critical for effective monitoring and mitigation of climate change, as well as to locate suitable locations for wind farms. Long-term simulations of wind dynamics (until year 2100) are given by various general circulation models (GCMs). However, GCM simulations are at a grid with coarse spatial resolution (>100 km), which renders spatial downscaling to a smaller scale an important prerequisite for climate-impacts studies. In this work, we propose a novel deep learning approach, named Bayesian AIG-Transformer, that consists of an attention-based input grouping (AIG), transformer, and uncertainty quantification. We use the proposed approach for the spatial downscaling of daily average wind speed (AWND), formulated as a multivariate time series forecasting problem, over four locations within New Jersey and Pennsylvania. To calibrate and evaluate our deep learning approach, we use large-scale observations extracted from NOAA’s NCEP/NCAR reanalysis dataset (2.5° × 2.5° resolution), which provides a proxy for GCM data when evaluating the model. Results show that our approach is suitable for the downscaling task, outperforming related machine learning methods.