Sunshine duration is an important atmospheric indicator used in many agricultural, architectural, and solar energy applications (photovoltaics, thermal systems, and passive building design). Hence, it should be estimated accurately for areas with low-quality data or unavailable precise measurements. This paper aimed to obtain a sunshine duration measurement database in Algeria’s south region and also to study the applicability of computational models to predict them. This work develops ensemble learning models for assessing daily sunshine duration with meteorological datasets that include daily mean relative humidity, daily mean air temperature, daily maximum air temperature, daily minimum air temperature, and daily temperature range as input. The study proposes a unique hybrid model, combining grey wolf and stochastic fractal search (GWO-SFS) optimization algorithms with the random forest regressor ensemble. A pre-feature selection process improved the newly suggested model. Various commonly adopted algorithms in relevant studies have been considered as references for evaluating the new hybrid algorithm. The accuracy of models was examined as a function of some frequently used statistical pointers, as well as the Wilcoxon rank-sum test. Besides, the models were evaluated according to the several input combinations. The numerical experiments show that the proposed optimization ensemble with feature preprocessing outperforms stand-alone models in terms of prediction accuracy and robustness, where relative root mean square errors are reduced by over 20% for all considered locations. In addition, all correlation coefficients are higher than 0.999. Moreover, the proposed model, with RMSEs lower than 0.4884 hours, shows significantly superior performances compared to previously proposed models in the literature.