This paper presents a comparison between different prediction models for solar radiation application. The present study assessed the performance of multi-layer perceptron (MLP) as well as boosted decision tree, and used a new combinition of these models with linear regression for the prediction of daily global solar irradiation (DGSR). The performance of the studied models was validated using a real dataset measured at the Applied Research Unit for Renewable Energies (URAER) situated in the south of Algeria. Different input combinations have been analysed in order to select the relevant input parameters for DGSR prediction. The results acheived show that the MLP model perfoms better than the others models in terms of statistical indicators: normalised root mean square error (0.033) and R² (97.7%).

This paper proposes a new approach for daily horizontal global solar radiation prediction using only air temperature as input. The proposed approach is based on support vector machine (SVM) by fusing the outcomes of two SVM models. The first one is used to estimate the horizontal global solar radiation whereas, the second SVM-model is used to estimate the error produced by the first model. Finally, weighted sum is applied to predict the global solar radiation. The proposed approach (Corrected-SVM) has been evaluated for Ghardaïa site (Algeria) which is characterized by a semi-arid climate conditions, using measurement of three years (2013-2015). It was found that the proposed Corrected-SVM model improves the performance of the conventional SVM-model. The obtained statistical parameters (Mean Absolute Bias Error (MABE), Root Mean Square Error (RMSE), Relative Square Error, and Correlation Coefficient(r)) of C-SVM and conventional SVM model are: 1.623 MJ/m2, 2.286 MJ/m2, 11.35, 94.20% and 1.713 MJ/m2, 2.575 MJ/m2, 12.61, and 93.06%, respectively.

Accurate estimation of renewable energy sources plays an important role in their integration into the grid. An unexpected atmospheric change can produce a range of problems related to various solar plant components affecting the electricity generation system. Global solar radiation (GSR) assessment has been increased in the past decade due to its important use in photovoltaic application. In this paper, we propose the use of machine learning-based models for daily global and direct solar radiation forecasting in a semi-arid climate, using a combination set of meteorological parameters on a horizontal surface in the Ghardaïa region. The models are presented and implemented on 3-year measured meteorological data at Applied Research Unit for Renewable Energies (URAER) at Ghardaïa city between 2014 and 2016. The results show that both MLP and RBF models perform well for three-step-ahead forecasting with a slight improvement in MLP models in terms of statistical metrics.