Optimization of a solar irradiation forecasting tool based on artificial intelligence

Abstract

In current electric markets, where many stakeholders can take part, power network operators need accurate predictions of the energy generated by intermittent renewable sources in order to control the whole system. Therefore, the capacity to accurately forecast solar irradiance is key when it comes to the large-scale integration of solar energy generators in the traditional network. One of the challenges, however, consists of providing accurate very short-term predictions (minutes ahead) due to the variability of solar irradiance caused by different meteorological phenomena. This study addresses this need for very short-term forecasts through the development of an irradiance prediction scheme for 10 minutes ahead. The irradiance prediction algorithm is based on a parallel combination of two different layer recurrent networks and has been trained with a two-year historical database of solar irradiance. The accuracy of the proposed tool has been validated through forecasting a whole year using data that is not in the database used in the training step. This tool was then used to forecast the irradiance in two Spanish locations with different weather conditions to analyse whether the accuracy changes. The accuracy between predicted and actual values demonstrates that this tool outperforms similar forecasters.