The statistical downscaling of global circulation models presents a significant challenge in selecting appropriate input variables from a vast pool of predictors. To address this, we developed ensemble clustering approach based on the Combining Multiple Clusters via Similarity Graph (COMUSA) method, which integrates k-means and self-organizing maps (SOMs) with the mutual information (MI)-random sampling method. This innovative feature extraction technique demonstrated a 21% improvement in the classification efficacy of large-scale climatic variables. When comparing feature extraction methods, the combination of MI-random sampling and ensemble clustering yielded more accurate results than SOM clustering alone. The most efficient artificial neural networks (ANNs)-based downscaling model was employed to project near- and mid-future precipitation and temperature (2025–2035, 2035–2045), revealing varied outcomes under different scenarios (SSP3–7.0 and SSP5–8.5). Under SSP3–7.0, annual mean precipitation is projected to decrease by 2–3%, while under SSP5–8.5, it is expected to decrease by 4–5%. Temperature projections indicate an increase in 21–27% under SSP3–7.0 and 29–35% under SSP5–8.5 scenarios for the annual mean temperature. The projected scenarios indicate descending and ascending patterns of precipitation and temperature, respectively. Integrating COMUSA ensemble clustering with MI-random sampling enhances the prediction accuracy of the ANN downscaling model, contributing to accurate projections of future precipitation and temperature.