ABSTRACT Developing a spatial downscaling method using machine learning techniques has emerged as a significant endeavor in meteorological forecasting. In this study, we propose a novel Residual-Pyramid-Attention Super Resolution (RPASR) model that effectively addresses the checkerboard artifacts issue prevalent in existing approaches, offering a unique perspective on spatial downscaling. Moreover, we propose a unique scaling technique during the data training to enhance the prediction performance. To validate and apply the proposed method, we utilize both low and high-resolution meteorological variables obtained from a numerical weather prediction model and observations in the Zhangbei region, China. The results demonstrate the exceptional performance of the RPASR model in spatial downscaling, exhibiting superior accuracy in meteorological prediction. Notably, the model outperforms interpolation techniques, significantly reducing deviations by 50% for variables such as 2 m temperature (T2), 2 m humidity (rh2), and 10 m wind speed (wspd10). In practical applications, our proposed model surpasses the DeepSD super-resolution model and low-resolution simulations as a whole, particularly for T2, where it enhances the T2 correlation coefficient by 1.2% compared to DeepSD. Additionally, it reduces the T2 root mean square error by 10.6% and the rh2 root mean square error by 8.6% compared to the low-resolution simulations. Overall, the RPASR model demonstrates robust validation and exhibits promising potential for advancing spatial downscaling methods in meteorological forecasting.