Machine Learning Site Adaptation for Optimizing Heliosat-4 Global Horizontal Irradiance using Adjacent Satellite Cells

Accurate estimation of global horizontal irradiance (GHI) from satellite-based models is critical for solar energy applications, especially in regions with sparse ground measurements. Traditional site adaptation (SA) methods adjust satellite-derived irradiance using local measurements, but are often limited by information from a single satellite cell. In this study, we propose a novel SA framework that incorporates modeled GHI values from adjacent satellite cells as additional predictors in a machine learning (ML) setting. This approach captures spatial variability in cloud cover and improves estimation accuracy while using fewer input features. We evaluate the performance of two ML models—Multilayer Perceptron (MLP) and Random Forest (RF)—using data from five sites in northwestern Argentina. The proposed method achieves up to a 4.5% reduction in root mean square error (RMSE) compared to conventional SA, while decreasing model complexity by reducing the number of input variables from 14 to 5. Results demonstrate consistent improvements across all sites, particularly in those with moderate initial bias. These findings indicate that spatially-aware models leveraging adjacent-cell data offer an effective and computationally efficient strategy for operational GHI estimation.