2. Methodological developments 23
2.1.3. Emission processing
Adapting and downscaling TNO-MACC III emission data for the Berlin-Brandenburg area
As already outlined in Section 1.3, some studies suggest that air quality modelling at high resolution only improves the model results if input data, most importantly emission input data, is available at a similarly high resolution (e.g. Schaap et al., 2015; Tie et al.,
pollutants was tested, and the results are described in detail in Article 1.
The downscaling was done in two steps: first it was done for Berlin only (Article 1 and 3) based on locally available proxy data including population density (Berlin Senate Depart-ment for Urban DevelopDepart-ment and the EnvironDepart-ment, 2011a) and traffic densities (Berlin Senate Department for Urban Development and the Environment, 2011b) for Berlin. As it was tested successfully, the downscaling was refined in a second step to include a larger geographical area around the city of Berlin, using in addition data on the location of roads from OpenStreetMap and, in cooperation with TNO, data on population density from LandScan2010. This had the aim of improving the representation of air pollutants also at the edge and just outside of the city. The second downscaling step further limited the emission (SNAP) categories to be downscaled to those for which the available proxy data were assumed to actually improve the distribution, e.g. road transport emissions, residential heating emissions and product use emissions. The second stage of the emission downscaling is described in detail in Article 2 and its Supplementary Material.
Technically, the downscaling redistributes emissions from the TNO-MACC III inventory within one original, coarse grid cell (ca. 7km x 7km) into 7 x 7 sub grid cells, and thus conserves the original emissions within each coarse grid cell. For this, the proxy data was re-gridded to the TNO-MACC III coarse grid as well as the fine sub grid, and factors were calculated indicating the share of population/traffic in each sub grid cell, compared to the
“parent” coarse grid cell. These factors were then used to distribute the emissions onto the fine sub grid.
In addition to the downscaling, airport emissions for Berlin were corrected: the original version of the TNO-MACC III emission inventory still includes airport emissions from the Tempelhof airport, which was closed to air traffic in 2008. Thus, airport emissions from all three airports in the original inventory were summed and then distributed on the two still operating airports Sch¨onefeld and Tegel. The distribution is based on their respective passenger and freight transport volumes.
Temporal and vertical distribution of emissions
For modelling air quality with a high temporal resolution, better results are achieved when distributing the annual total anthropogenic emissions temporally. As the distribution is not part of the WRF-Chem model, separate pre-processing routines need to be developed.
Like the emission downscaling, the temporal distribution of emissions was refined in two
veloped in order to process the emissions for model simulations contributing to Article 2.
The latter reads in all necessary specifications (temporal and vertical distribution, NOx, VOC and PM splitting) from easily modifiable csv-files.
In the first stage, the temporal distribution of emissions was based on profiles calculated by Builtjes et al. (2002), including distribution factors depending on the month of the year (annual cycle), the day of the week (weekly cycle) and the hour of the day (diurnal cycle).
In the second stage, the diurnal cycle of traffic emissions was replaced by a diurnal cycle calculated based on traffic counts for Berlin, assuming a linear scaling of traffic emissions with traffic counts as also done by Builtjes et al. (2002). Article 2 describes this procedure.
Distributing anthropogenic emissions vertically instead of releasing them into the first model layer does not change the model results considerably when using a coarse model resolution (Mar et al., 2016). However, the results from Article 1 suggest that a vertical distribution might be more important for higher model resolutions of only a few km.
This is why anthropogenic emissions are distributed vertically for the model simulations contributing to Article 2 of this thesis, as described in the article. In brief, emissions from residential heating, energy and non-energy industry as well as airports are distributed into up to seven vertical layers. The vertical distribution is also included in the routines written for the second stage of the emission pre-processing.
In addition to the modification of Berlin airport emissions described above, a further modification to the original data is introduced. Airport emissions are included in the category of non-road transport emissions in the inventory. Emissions for Berlin from the emission category road transport emissions are split into airport emissions and non-airport emissions. This is possible, as in the case of Berlin non-airport emissions are indicated as point sources. Emissions for airports are then distributed vertically considering both emissions of airport ground transportation and the LTO-cycle (landing, take-off), while other non-road transport emissions are only released into the first model layer.
2.2. Model evaluation
Since the WRF-Chem model can be set up with numerous different combinations of pa-rameterizations, domain configurations and settings related to the model dynamics, it is necessary that every new setup is evaluated against measurement data (also see Sec-tion 1.3.2). The main focus of Article 1 of this thesis is an operaSec-tional model evaluaSec-tion
For the diagnostic model evaluation, a Kolmogorov-Zurbenko filter (Zurbenko, 1986) is used in order to decompose modelled and observed time series. Following Galmarini et al.
(2013); Hogrefe et al. (2000); Solazzo et al. (2017b), the time series are decomposed into a long-term (LT, >21 days), synoptic (SY, 2.5-21 days), diurnal (DU, 0.5-2.5 days) and intra-diurnal (ID, < 0.5 day) component, each representing processes characteristic for these time scales. Specifically, the components are defined as follows, with the Kolmogorov-Zurbenko filter kzm,k, the time window m and smoothing parameter k, time series x and time t:
ID(t) = x(t)− kz3,3(x(t)) (2.1)
DU (t) = kz3,3(x(t))− kz13,5(x(t)) (2.2) SY (t) = kz13,5(x(t))− kz103,5(x(t)) (2.3)
LT (t) = kz103,5(x(t)) (2.4)
The mean square error (MSE) of each component is then assessed by breaking it down into bias, variance error (σ) and minimum achievable mean square error (mMSE) as follows (Solazzo and Galmarini, 2016):
M SE = (mod− obs)2+ (σmod− rσobs)2+ mM SE (2.5) Where mod and obs indicate model results and observations, respectively, and r denotes the correlation coefficient. The decomposition and error apportionment is described in detail in Article 2. For the decomposition, the “kza” library of the R programming language was used. An exemplary R-script for spectrally decomposing a NOx time series can be found in the Appendix to this thesis.