Spatial extrapolation of a satellite-based machine learning model for global horizontal irradiance estimation

Abstract

The availability of accurate solar irradiance estimates in areas without ground measurements is crucial for the design and planning of solar energy systems. This study analyzes the spatial extrapolation capability of an empirical model based on machine learning (ANN) and satellite imagery, trained exclusively with data from the Luján station in Argentina. Its performance is evaluated by applying it to the stations of Paraná, Concepción del Uruguay, and General Villegas, comparing the estimates with ground-based measurements and with two physical reference models (CAMS and NREL). The results show that the model maintains very good performance even at distant stations, with only a slight decrease in accuracy that remains within excellent parameters. In all cases, the model outperforms CAMS and NREL estimates by up to 16% in terms of normalized root mean square error (RMSEₙ). When analyzing errors by sky condition, the ANN model performs best under cloudy and partly cloudy conditions. On clear-sky days, it performs worse than physical models, which is expected given that it does not incorporate solar geometry variables or a clear-sky model as inputs. These findings highlight the robustness of the approach based solely on satellite imagery and its potential for application in regions with limited measurement infrastructure.