WAVELET PACKET-GAUSSIAN PROCESS REGRESSION MULTIVARIATE AND UNIVARIATE MODEL FOR FORECASTING DAILY SOLAR RADIATION
Khaled Ferkous, Farouk Chellali, Abdalah Kouzou, Belgacem Bekkar, Nacer Hacene
Abstract
In recent years, the accuracy of the prediction of daily global solar radiation (GSR) has received great attention by researchers. In this study, hybrid models using the Gaussian process regression algorithm, wavelet, and wavelet packet decomposition (W, WPD) GPR have been proposed to predict daily solar radiation in Ghardaia city (Algeria). In addition to the hybridization methods, the problem is examined initially by entering meteorological data as input and is extended to the univariate case in the process of estimating the optimum model. For this purpose, three years of data series (2013-2015) have been used in model training, while data from 2016 were used to validate the model. The results showed the effectiveness of the hybrid model, especially WPD-GPR; compared to the classic GPR model, this model gave excellent results when using univariable inputs in terms of root mean square error (RMSE = 0.58 MJ/m2 day), relative root mean square error (rRMSE = 2.83%), mean absolute error (MSE = 0.34 MJ/m2 day), and determination coefficient (R2 = 99.54%).