Solar Radiation Prediction Using Radial Basis Function Models

Abstract

Accurate weather information is essential for developing an efficient solar power generation system. In this study one year hourly meteorological data for Kaitaia, New Zealand has been obtained from The National Climate Database of New Zealand to predict solar radiation. Twelve models with different combinations of input variables were formed. Three artificial neural networks (ANN), Multilayer Perceptron (MLP), Nonlinear Autoregressive Network with Exogenous Inputs (NARX), and Radial Basis Function (RBF) using Levenberg-Marquardt (LM) back propagation learning algorithm were trained and tested for the twelve models. The performance of each approach was assessed by calculating mean square error (MSE) and regression values. The results shows that models with a higher number of input variables irrespective of the number of neurons and delays provide better accuracy and improved results for regression values. In addition, the RBF network outperforms the NARX and MLP approaches. Furthermore, the 24-hour and 4-day ahead predicted solar radiation values of the RBF, NARX and MLP approaches are presented and, the results shows that the RBF network performs better than NARX and MLP approaches.