An analysis of different deep learning neural networks for intra-hour solar irradiation forecasting to compute solar photovoltaic generators' energy production
Garazi Etxegarai, Asier López, Naiara Aginako, Fermín Rodríguez
Abstract
Renewable energies are the alternative that leads to a cleaner generation and a reduction in CO2 emissions. However, their dependency on weather makes them unreliable. Traditional energy operators need a highly accurate estimation of energy to ensure the appropriate control of the network, since energy generation and demand must be balanced. This paper proposes a forecaster to predict solar irradiation, for very short-term, specifically, in the 10 min ahead. This study develops two tools based on artificial neural networks, namely Long-Short Term Memory neural networks and Convolutional Neural Network. The results demonstrate that the Convolutional Neural Network has a higher accuracy. The tool is tested examining the root mean square error, which was of 52.58 W/m2 for the testing step. Compared against the benchmark, it has obtained an improvement of 8.16%. Additionally, for the 82% of the tested days it has given a less than 4% error between the predicted and the actual energy generation. Results indicate that the forecaster is accurate enough to be implemented on a photovoltaic generation plan, improving their integration into the electrical grid, not only for providing power but also ancillary services.