Small-scale solar radiation forecasting using ARMA and nonlinear autoregressive neural network models

Abstract

This paper aims to introduce an approach for multi-hour forecasting (915 h ahead) of hourly global horizontal solar radiation time series and forecasting of a small-scale solar radiation database (30- and 1-s scales) for a period of 1 day (47,000 s ahead) using commonly and available measured meteorological solar radiation. Three methods are considered in this study. First, autoregressive–moving-average (ARMA) model is used to predict future values of the global solar radiation time series. However, because of the non-stationarity of solar radiation time series, a phase of detrending is needed to stationarize the irradiation data; a 6-degree polynomial model is found to be the most stationary one. Secondly, due to the nonlinearity presented in solar radiation time series, a nonlinear autoregressive (NAR) neural network model is used for prediction purposes. Taking into account the advantages of both models, the goodness of ARMA for linear problems and NAR for nonlinear problems, a hybrid method combining ARMA and NAR is introduced to produce better results. The validation process for the site of Ghardaia in Algaria shows that the hybrid model gives a normalized root mean square error (NRMSE) equals to 0.2034 compared to a NRMSE equal to 0.2634 for NAR model and 0.3241 for ARMA model.