Predictability analysis of SO2 time series by linear and non-linear forecasting approaches

Sulphur dioxide daily concentration time series recorded at four stations in Ravenna (Italy) during the period 1981-1990 have been analyzed. The predictability of these series has been evaluated using two possible forecasting approaches: the global autoregressive approximation and the local autoregressive approximation. The comparison of the predictive capability of the two techniques allows us to discriminate between low- dimensional chaos from random dynamics. In the global approach, the data are considered as a realization of a stochastic process with a large number of degrees of freedom, while in the local one as a deterministic process generated by few non-linear dynamical equations. As a by-product our predictability analysis gives the best autoregressive approximation to forecast future values of the signal. In our case, the global approach is superior over the local one, all the possible implications with the dynamics of the process are discussed. A preliminary time series analysis reveals the presence of trends and seasonal components which have to be removed before applying the two forecasting approaches above. Finally scaling laws in the frequency domain are investigated, the power spectra of the filtered time series follow a f(-?) law, f being the frequency. Our findings are typical fingerprints of a broad class of fractal stochastic systems.