Model selection for local and regional meteorological normalisation of background concentrations of tropospheric ozone

Meteorological normalisation of time series of air quality data aims to extract anthropogenic signals by removing natural fluctuations in the collected data. We showed that the currently used procedures to select normalisation models can cause over-fitting to observed data and undesirable smoothing of anthropogenic signals. A simulation study revealed that the risk of such effects is particularly large when: (i) the observed data are serially correlated, (ii) the normalisation model is selected by leave-one-out cross-validation, and (iii) complex models, such as artificial neural networks, are fitted to data. When the size of the test sets used in the cross-validation was increased, and only moderately complex linear models were fitted to data, the over-fitting was less pronounced. An empirical study of the predictive ability of different normalisation models for tropospheric ozone in Finland confirmed the importance of using appropriate model selection strategies. Moderately complex regional models involving contemporaneous meteorological data from a network of stations were found to be superior to single-site models as well as more complex regional models involving both contemporaneous and time-lagged meteorological data from a network of stations.