Open Access

Claudia Weitnauer

• In the last decades, the critical increase of the emissions of PM10 (Particulate Matter with a medium diameter less than 10 µm), especially in urban areas, has become a problem for the environment and human health. High concentration episodes of fine particles increase the risk of cardiovascular or respiratory tract diseases, as several studies confirm. Local concentrations of PM10 are influenced by meteorological parameters on different scales, e.g. local meteorological conditions and large-scale circulation dynamics. With climate changing rapidly, these connections need to be better understood in order to be able to estimate climate-change-related consequences for air quality management purposes. To detect critical periods of high PM10 concentrations, one focus in recent studies is set on the improvement of accurate short-term deterministic, statistical prediction models or reliable approaches for long-term air quality prediction. The general relationship between local particleIn the last decades, the critical increase of the emissions of PM10 (Particulate Matter with a medium diameter less than 10 µm), especially in urban areas, has become a problem for the environment and human health. High concentration episodes of fine particles increase the risk of cardiovascular or respiratory tract diseases, as several studies confirm. Local concentrations of PM10 are influenced by meteorological parameters on different scales, e.g. local meteorological conditions and large-scale circulation dynamics. With climate changing rapidly, these connections need to be better understood in order to be able to estimate climate-change-related consequences for air quality management purposes. To detect critical periods of high PM10 concentrations, one focus in recent studies is set on the improvement of accurate short-term deterministic, statistical prediction models or reliable approaches for long-term air quality prediction. The general relationship between local particle concentrations and large-scale circulation dynamics, as for example reflected by weather or circulation types, has been proven in several studies. Thus far, only a few systematic attempts have been made to optimize objective classifications concerning their relationship to local PM10 concentrations. Against this background, the aim of this study is to evaluate various approaches for the optimization of circulation type classifications with respect to their relevance for Bavarian cities (Germany) to detect those classifications, that are best suited for the use in statistical downscaling models. The used data set of daily mean PM10 concentrations has been provided by the Bavarian Environmental Agency (Landesamt für Umwelt) from their air quality monitoring network. For the analysed period 1980-2011, measurements from 16 urban stations with over 90% data availability, spread over the target region, are available. (Initial characteristics of this data set concerning basic quality aspects, long-term trends and correspondence between locations are presented in this thesis. To characterize the correspondence between the PM10 measurements of the different stations by spatial patterns, a regionalization by a s-mode Principal Component Analysis is realized on the data set. Objective circulation type classifications are optimized with respect to their synoptic skill for the target variable PM10 in a stepwise procedure. Seasonal patterns emerging from this optimization are varying weighted combinations of three large-scale atmospheric variables in several pressure levels. Their performance is evaluated by using a range of skill scores for varying calibration and validation periods. Results from these analyses have revealed highest synoptic skill for classifications based on a mid-size domain (7.5°-27.5° E and 40°-60° N), 18 number of classes, seasonal classifications, a weighted combination of three large-scale atmospheric input variables and the conditioning by the target variable. Depending on the season and the considered PM10 station, the combinations of classified atmospheric parameters vary. Air temperature (1000 hPa level), relative (850 hPa level) and specific humidity (1000 hPa level), mean sea level pressure, geopotential height at the 500 hPa level as well as zonal wind (500 hPa level) are detected to be the most relevant parameters throughout the seasons. Best performing classifications, in terms of maximum skill from evaluation, have been ascertained for each station and each season. Monthly occurrence frequencies of circulation types (predictors), resulting from the previous optimization of classifications, are related to daily and monthly PM10 indices (predictands) by using different statistical downscaling techniques. The comprehensive set of downscaling tools comprise variants of Synoptic Downscaling and regression-based methods (Multiple Linear Regression, Generalized Linear Models, Random Forests). PMdaily, PMmean and PM50 are used as PM10 indices. The generated transfer models are evaluated via cross-validation using different subintervals of the 1980-2011 period as calibration and validation periods, respectively. Highest model skill in cross-validation is detected for PMmean in winter, using the Random Forest approach, at all stations. Model results for PMdaily have revealed positive but less pronounced skill for variants of the synoptic downscaling. Hence, the synoptic downscaling models are used for estimating PMmean as well. Most suitable downscaling procedures, in terms of model skill determined from cross-validation, are finally applied to CMIP5 climate model data (ECHAM6, EC-Earth) to derive estimates of possible future climate-change-related variations in PM10 concentrations. Projections are used for two time periods (2021-2050, 2071-2100) and two different scenarios (RCP 4.5, RCP 8.5). Possible model biases evoking from the climate models and the downscaling approaches are assessed by numerical and statistical ensembles. A bias correction is applied on the modelled PM10 time series in the observational period as well as on the estimated future particle levels. On the one hand, climate-change-induced variations of particle levels have yielded out a decrease in winter at nearly all stations in Bavaria. An increase of particle levels is estimated for summer months, independent from considered scenario and time step, on the other hand. Detected changes in winter and summer correspond with variations in the frequencies of occurrence of PM10-relevant large-scale atmospheric dynamics. An increase of zonal, cyclonic weather and circulation types for example is estimated in winter and a decrease in summer until the end of this century. In the transitional seasons, variations of estimated PM10 levels are less pronounced, more variable and remain insignificant in most of the cases. Finally, the model skill of the circulation-type-based downscaling models in cross-validation and estimated future climate-change-induced variations of particle levels are compared to results from statistical downscaling models based on local meteorological parameters as predictors. In a two-step process, the local transfer models have been developed in the framework of the research project "Particulate matter and climate change in Bavaria", funded by the German Research Foundation. First of all, local PM10-relevant meteorological variables are downscaled from large-scale atmospheric fields and are applied as predictors in statistical downscaling models for estimating PM10 concentrations afterwards. Circulation-type-based downscaling models and local transfer models use objective circulation type classifications, that are optimized with respect to their synoptic skill for the target variable PM10. All downscaling approaches are evaluated via the aforementioned cross-validation procedure. Detected changes in future PM10 concentrations have revealed similar tendencies of decreasing levels in winter and increasing ones in summer for all downscaling approaches. Nevertheless, estimations from circulation-type-based models have shown partly pronounced differences concerning climate models, numerical and statistical ensembles and downscaling approaches compared to results from local transfer models.…