Accurate and high-resolution air temperature prediction is important in many different applications. Hourly air temperature forecasting in mountainous areas is necessary and important because mountainous areas are becoming increasingly important areas of human activities. At present, scientists successfully employ numerical weather prediction (NWP) models, such as the WRF model, to achieve reliable forecasts. However, air temperature forecasting and modeling over complex geographical zones are difficult tasks. The WRF model is a mesoscale model and does not adequately account for the influence of terrain on the air temperature. It is important to downscale larger-scale models to a much finer scale. In this paper, a statistical temperature downscaling method based on geographically weighted regression (GWR) and diurnal temperature cycle (DTC) models is proposed. A statistical downscaling scheme is designed to forecast the hourly air temperature, at a 30-m spatial resolution, up to 24 h in advance. Compared to WRF simulations, RMSE of the combined downscaling model decreased 1.01 ℃ at the automatic weather station level and 0.80 ℃ at the spatial level, and MAE decreased by 0.81 ℃ and 0.69 ℃, respectively, at these two levels. The results reveal that the combined downscaling model performs very well in correcting and downscaling the air temperature in WRF simulations in mountainous areas.