Reliable solar irradiance prediction using ensemble learning-based models: A comparative study

•Accurate prediction of solar irradiance is vital to design photovoltaic systems.•Ensemble models reduce the variance of prediction error while obtaining a low bias.•Four ensemble learning models have been compared for solar irradiance prediction.•Variables contribution assessment was investigated using three ensemble models.•Results show the superior performance of the ensemble learning models. Accurately predicting solar irradiance is important in designing and efficiently managing photovoltaic systems. This paper aims to provide a reliable short-term prediction of solar irradiance based on various meteorological factors using ensemble learning-based models that take into account the time-dependent nature of the solar irradiance data. The use of ensemble learning models is motivated by their desirable characteristics in combining several weak regressors to achieve an improved prediction quality relative to conventional single learners. Furthermore, they reduce the overall prediction error and have the ability to combine different models. In this paper, we first investigate the prediction performance of the well-known ensemble methods, Boosted Trees, Bagged Trees, Random Forest, and Generalized Random Forest in short-term prediction of solar irradiance. The performance of these ensemble methods has been compared to two commonly known prediction methods namely Gaussian process regression, and Support Vector Regression. Typical Meteorological Year data are used to verify the prediction performance of the considered models. Results showed that ensemble methods offer superior prediction performance compared to the individual regressors. Furthermore, the results showed that the ensemble models have a consistent and reliable prediction when applied to data from different locations. Lastly, variables contribution assessment showed that the lagged solar irradiance variables contribute significantly to the ensemble models, which help in designing more parsimonious models.