A new evolutionary design of an ANN (artificial neural network) ensemble model is developed to generate hourly DNI (direct normal irradiance) estimates. The procedure combines a genetic algorithm for selecting the best inputs with an ANN ensemble method. The ensemble model was calibrated and evaluated using three years of Meteosat-9 images and data measured at 28 high-quality ground stations over an extensive area, mainly in Europe. The most valuable inputs for DNI estimation are shown to be the following: all Meteosat-9 channels except ch08 and ch11; relative air mass m, integral Rayleigh optical thickness or, extraterrestrial global irradiance G(0), beam clear-sky index B-cs, and the cosine of zenith angle theta. No additional atmospheric information such as turbidity, aerosol optical depth or water vapor content are required for the model. Ensemble estimates were nearly unbiased (MBE = 1.98%) and overall RMSE (root mean square error) was 24.29% across an independent spatial and temporal dataset This represents an improvement of 35% over other common methods for estimating DNI. The estimates were reasonably reliable in all seasons, and were more accurate in clear-sky conditions. (C) 2015 Elsevier Ltd. All rights reserved.