Bayesian Variable Selection in Neural Networks for Short-Term Meteorological Prediction

Abstract

This work examines the influence of Bayesian variable selection on neural architectures for global solar irradiation and air temperature time series prediction. These models, 3 neural architectures with differing input and output processing strategies [2], predict all time slots in the 24 hours ahead period, with inputs solely taken from local measurements of the 24 last hours. Qualitative and computational points of view are considered for the comparison of Bayesian and non-Bayesian learning, with a specific care for salient variable sets analysis. For generalization purpose, models are assessed and compared on data from two contrasted sites in France. The input space appeared to be reduced by at least 34%, and up to 73%, with a small prediction quality loss (1.3% on average), and a good repeatability of selected salient variables across sites.