Soft computing approach for solar radiation prediction over Abu Dhabi, UAE: A comparative analysis

Abstract

Integration of solar generation into power networks can negatively affect the performance of next generation smart energy grids. Rapidly changing output power of this kind is unpredictable and thus one of the solutions is to predict it by computational intelligence techniques. The stochastic component of solar radiation is highly non-linear in nature because of many factors including time of the year, weather conditions, and geographical locations. In order to uncover the underlying phenomenon, three soft computing techniques for solar radiation forecasting based on neural networks (NN) - artificial neural network (ANN), the adaptive neuro-fuzzy inference system (ANFIS), and nonlinear autoregressive with exogenous inputs (NARX) - are implemented and a comparative analysis is performed. Meteorological variables, such as, relative humidity (RH), temperature (T), wind speed (WS), and sunshine duration (SSD) are used as inputs to the NN models. The global horizontal irradiance (GHI) is estimated using ten years data of Abu Dhabi, the United Arab Emirates (UAE). Different statistical performance indicators are computed. Simulation results show that NARX performs relatively better in this case and generalizes the data well. All the models have the tendency to exhibit more error in spring seasons. This leads to further investigations to cater for seasonality components.