Hourly ozone prediction for a 24-h horizon using neural networks

This study is an attempt to verify the presence of non-linear dynamics in the ozone time series by testing a ‘‘dynamic’’ model, evaluated versus a ‘‘static’’ one, in the context of predicting hourly ozone concentrations, one-day ahead. The ‘‘dynamic’’ model uses a recursive structure involving a cascade of 24 multilayer perceptrons (MLP) arranged so that each MLP feeds the next one. The ‘‘static’’ model is a classical single MLP with 24 outputs. For both models, the inputs consist of ozone and of exogenous variables: past 24-h values of meteorological parameters and of NO2; concerning the ozone inputs, the ‘‘static’’ model uses only the 24 past measurements, while the ‘‘dynamic’’ one uses, also, the previously forecast ozone concentrations, as soon as they are predicted by the model. The outputs are, for both configurations, ozone concentrations for a 24-h horizon. The performance of the two models was evaluated for an urban and a rural site, in the greater Paris. Globally, the results indicate a rather good applicability of these models for a short-term prediction of ozone. We notice that the results of the recursive model were comparable with those obtained via the ‘‘static’’ one; thus, we can conclude that there is no evidence of non-linear dynamics in the ozone time series under study.