Evolutionary procedure based model to predict ground–level ozone concentrations

Abstract

This study aims to predict the next day hourly average ozone (O3) concentrations using threshold autoregressive (TAR) models in which the threshold value and the threshold variable are defined by genetic algorithms. The procedure is also able to generate models with statistically significant regression parameters. The performance of TAR models was then compared to the one obtained with autoregressive (AR) and artificial neural network (ANN) models. Different TAR models were generated, corresponding to different threshold variables and values. For the training period, ANN model presented better results than TAR and AR models. However, in the test period, AR and one of the TAR models achieved better predictions of O3 concentrations than the ANN model. The distinction between the applied models became greater when they were evaluated in the prediction of the extreme values, for which the TAR model presented the best performance. The performance with respect to extreme values is a useful implication for the protection of public health as this model can provide more reliable early warnings about high O3 concentration episodes.