Machine Learning for Site Adaptation of Satellite-DerivedSolar Irradiance in Northwestern Argentina

Accurate estimation of global horizontal irradiance (GHI) is essential for solar energy resource assessment, particularly in regions with limited ground-based measurements. This study evaluates the performance of three machine learning models—Simple Linear Regression (SLR), Extreme Gradient Boosting (XGB), and Multilayer Perceptron (MLP)—for the site adaptation of satellite-derived GHI data in five locations in Northwestern Argentina. Two satellite products, CAMS and LSA-SAF, were used as input data. Themodels were assessed using standard error metrics (MBE, MAE, RMSE), and their residual patterns were analyzed. Results show that LSA-SAF data led tolower errors compared to CAMS, especially in highaltitude sites. While complex models like MLP and XGB marginally improved accuracy in some cases, SLR offered comparable results with higher robustness. The analysis also identified systematic biases and discretization effects in tree-based models. These findings suggest that, under current data conditions,simpler models may offer reliable performance. Enhancing input data quality and incorporating additional meteorological features may yield greater improvements than increasing model complexity