Forecasting of daily global solar radiation using wavelet transform-coupled Gaussian process regression: Case study in Spain

Abstract

This work presents a successful application of a new hybrid model in forecasting daily global solar radiation for a site in Spain using time series of solar radiation. The hybrid model incorporates the wavelet transform (WT) and Gaussian process regression (GPR). The WT is used to extract meaningful time-frequency information by decomposing the clearness index time series into a set of well-designed subseries. The future behavior of clearness index is forecasted with the trained GPR model using inputs of those subseries. The daily global solar radiation is obtained by multiplying the forecasted clearness index with extraterrestrial solar radiation. The normalized root mean square error (nRMSE), 9.36%, demonstrates the model's excellent capability in forecasting daily global solar radiation. The proposed model outperforms some other well-established models, including autoregressive moving average (ARMA), non-wavelet GPR, wavelet coupled and non-wavelet artificial neural network (ANN) and support vector regression (SVR) models.