Optimization of long-term air quality modelling applying cluster analysis techniques

Abstract

A methodology to classify atmospheric conditions is provided which can be applied efficiently for studies of long-term air quality. It is shown that statistical air quality indicators like average and mean daily maximum concentrations of air pollutants, percentiles, and, with respect to ozone, risk indicators like AOT40 and SOMO35, can be calculated with sufficient accuracy if they are based on modelled concentration data that are calculated only for a few representative meteorological conditions resulting from the classification. The aim of the project was the development of an efficient classification and analysis tool and the demonstration of its general applicability. For the classification, daily meteorological conditions of only one year, namely the year 2000, have been considered. These have not been derived from analysis data or synoptic observations, but from the results of the large-scale model EURAD. Two different classification techniques have been applied, the classical cluster analyses using the WARD and K-MEANS methods, and Kohonen's Self Organizing Maps (SOM). The statistical air quality indicators that were derived using the classification system have been compared with corresponding reference quantities calculated from the results of a continuous annual simulation with the meso-scale Chemistry-Transport-Model (CTM) KAMM/DRAIS. The reference air quality indicators derived from these detailed simulation have also been compared with corresponding values calculated from measurements. The classification of the meteorological conditions of the whole year gave 19, 20, and 21 classes applying the WARD, K-MEANS, and SOM methods, respectively. AOT40 is defined for the periods from April until September and May until July. Therefore, further classifications have been performed for these two periods. Generally, it can be stated that, with only a few exceptions, the difference between statistical measures derived from cluster solutions and detailed simulations are less than ±10%. The only exceptions are for NO and SOMO35. With respect to NO the differences related to the annual mean and the 90-percentile were larger than ±10% considering all clustering methods. With respect to SOMO35 the difference was larger than 10 % for all statistical quantities when the SOM was applied. Assessing the efficiency of the classification methods and whether the resulting classes represent the whole spectrum of possible meteorological conditions, the K-MEANS and WARD methods best fulfilled these criteria.