Short-Term Solar Irradiance Forecasting Using Deep Learning Techniques: A Comprehensive Case Study

Abstract

Reliable estimation of solar irradiance is required for many solar energy applications such as photovoltaics, water heating, cooking, solar microgrids, etc. Deep Learning techniques have shown outstanding behaviour for analysing complex datasets efficiently with high accuracy. Multi-Layer Perceptron (MLP), Long-Short Term Memory (LSTM), and Gated Recurrent Unit (RGU) techniques are found to be the most competitive techniques in the literature for solar irradiance forecasting. Therefore, in this study, a comparative analysis of those models is carried out using eleven years of NASA satellite data for training and testing. The grid search technique is used to optimize the networks architectures to ensure the best performance of the models for forecasting daily global solar irradiance. The results show that all models have similar accuracy with a mean square error close to 0.017 kWh/m2/day. However, the speed of training varies between 17 and 208 seconds for each model where GRU has shown higher speed than LSTM despite of containing more layers due to their computational complexity. The MLP is found to be the most efficient model due to using a low number of parameters 49,281 as compared to 1,025,793 for GRU. The study is of importance for reliable solar irradiance forecasting for any location worldwide.