Neural architectures for global solar irradiation and air temperature prediction

Abstract

This paper presents a study on neural architectures for the prediction of global solar irradiation and air temperature time series, a useful task for thermal energy management systems. In this contribution, the highly cyclic nature of the variables is carefully considered in the normalization step and the neural architecture design. The standard neural approach is confronted to the absolute daily and the absolute tri-hourly architectures for the prediction of the next 24 hours. For generalization purpose, models are assessed and compared on data from two sites in France. Results show that the absolute models outperform the reference model and some naive models. A complexity analysis also outlines the interest of the proposition. © 2012 Springer-Verlag.