Downscaling and Predicting Downward Shortwave Radiation: A Case Study Using Ambient Weather and Open-Meteo Data

This study presents a methodology for downscaling and predicting shortwave radiation. To address coarse-resolution limitations in global datasets, this study integrates local high-frequency measurements with lower-frequency global outputs. Specifically, we align 15-minute Open-Meteo observations with 5-minute Ambient Weather observations via interpolation, creating a unified higher-resolution dataset. Two machine learning models, XGBoost and a newly introduced Nearest Neighbor-XGBoost (NN-XGBoost), are then trained to forecast DSWR for one day. By leveraging localized smoothing from nearest neighbors alongside gradient boosting, NN-XGBoost surpasses standard XGBoost in error reduction and R2. These findings can aid applications ranging from microgrid management to precision agriculture, showcasing a scalable and generalizable data science framework for downscaling and predicting DSWR. Key Words: Downscaling, Downward shortwave radiation, forecast.