Artificial Neural Network for Forecasting One Day Ahead of Global Solar Irradiance

Abstract

Due to global warming, the world is seeking to use more renewable energy. In this study, we focus on solar energy, which has been receiving increased amounts of attention in the last few decades. The integration of solar energy into electricity networks requires reliable forecast information of solar resources enabling it to quantify the available energy and allowing it to optimally manage the transition between intermittent and conventional energies. Throughout our research, we investigated different forecasting techniques in order to find which one is appropriate for forecasting the daily global solar irradiance for the region of Rabat. The first-tested approach is linear modeling based on classical ARIMA-GARCH and exponential smoothing models. The second approach proposes non-linear modeling based on Artificial Neural Networks (ANNs) models. Numerous research has demonstrated the ability of ANNs to predict time series of weather data. In this study, we will examine a particular structure of ANNs, Multilayer Perceptron (MLP), which has been used the most among ANN structures in renewable energy and time series forecasting broadly. We used some statistical feature parameters to find the optimal structure of MLP in the univariate case and the multivariate case. The results showed that the MLP with exogenous variables performed better than the other models.