Comparing neural networks and transfer function models for ozone forecasting

摘要

Surface ozone concentrations are determined by complex interactions between precursors and are triggered by meteorological conditions. Ozone concentrations are, in fact, strongly linked to meteorological conditions in the boundary layer and to land-sea breezes at coastal sites. The related relationships are typically complex and nonlinear and might be better captured by dynamical models, namely Neural Networks and Transfer Function models. Aim of our work is the identification of proper Transfer Function models and the estimation of their parameters. Here we present an outline of the methodology that was used to develop the air pollution forecast model for a complex coastal valley. We also investigate the potential for using Neural Networks, namely Multi-Layer Perceptron networks, to forecast ozone pollution, as compared to the multivariate parametric air pollution forecast model and multi-linear regression equations (the most commonly used to forecast Ozone concentrations). Transfer Function models and Neural Networks are examined for the Esino valley (Italy) under different climate and ozone regimes, enabling a comparative study of the two approaches.