Modelling NO2 concentrations at the street level in the GAINS integrated assessment model: projections under current legislation

. NO 2 concentrations at the street level are a major concern for urban air quality in Europe and have been regulated under the EU Thematic Strategy on Air Pollution. Despite the legal requirements, limit values are exceeded at many monitoring stations with little or no improve-ment in recent years. In order to assess the effects of future emission control regulations on roadside NO 2 concentrations, a downscaling module has been implemented in the GAINS integrated assessment model. The module follows a hybrid approach based on atmospheric dispersion calculations and observations from the AirBase European air quality database that are used to estimate site-speciﬁc parameters. Pollutant concentrations at every monitoring site with sufﬁcient data coverage are disaggregated into contributions from regional background, urban increment, and local roadside increment. of each with a model of the appropriate scale: 28 × 28 km grid based on the EMEP Model for the regional background, × 7 urban increment based on the CHIMERE Chemistry Transport Model, and a chemical box for the roadside increment.


Introduction
NO 2 is a traffic-related air pollutant that is of major concern for public health, both through direct effects and as a precursor to tropospheric ozone formation (WHO, 2006). For the European Union (EU), legally binding limit values on NO 2 concentrations have been defined in the Air Quality Directive (EU, 2008) and its Daughter Directives. Member States are facing considerable difficulties to meet these limit values, with 22 out of 27 Member States still recording exceedances as of 2010 (EEA, 2012). In spite of the past reductions in emissions of total nitrogen oxides (NO x = NO + NO 2 ), NO 2 concentrations at roadside stations have shown only small decreases or even increases during recent years, which has been The GAINS Integrated Assessment Model 1 (Amann et al., 2011) is employed in the current revision of the EU Thematic Strategy on Air Pollution (TSAP) as a policy analysis tool to provide an outlook on the likely development of air quality under different emission control strategies, and assess their costs and benefits. The GAINS model brings together information on the sources and impacts of air pollutant and greenhouse gas emissions and their interactions. GAINS addresses air pollution impacts on human health and ecosystems, in addition to the mitigation of greenhouse gas emissions. As an input to the revision of the TSAP, the impact of different emission scenarios on compliance with air quality limit values is assessed. GAINS air quality impact calculations are currently done on a resolution of roughly 7 × 7 km for human health. This resolution is sufficient to describe urban background conditions; compliance with air quality limit values, however, is determined at single monitoring stations which may be located in street canyons at traffic hot spots and therefore require a significantly finer model resolution.
Modelling pollution levels inside street canyons with heavy traffic is challenging, owing mostly to the physical description of the air flow and the quantification of emissions, whereas the chemical processes are relatively well understood. Detailed models have shown impressive results in describing pollution levels within street canyons (for an overview see e.g. Vardoulakis et al., 2003), attempting to explicitly model the air flow through computational fluid dynamics (Sabatino et al., 2007;Murena et al., 2009) or introducing a parametrisation of turbulence (Berkowicz, 2000). Owing to the complexity of input information required, models used in these studies are often limited to single cities or even single street canyons (Denby, 2011). The focus of the GAINS model, however, is to provide estimates for the evolution of air quality and compliance with limit values in the EU as a whole. This results in the need to model concentrations in the immediate vicinity of hundreds of traffic stations. It is infeasible to gather sufficient data to employ detailed street canyon models for all these stations. We therefore base our modelling scheme on a combination of bottom-up modelling and past observations, employing the chemical description used in parametrised models such as OSPM (Berkowicz, 2000) but estimating physical input parameters from past observations.
In this article, we describe the downscaling scheme that has been implemented in the GAINS model to assess compliance with limit values at individual monitoring stations reporting to the European air quality database (AirBase) 2 . Starting from monitoring data, we disaggregate measured annual mean concentrations at roadside monitoring sites into the appropriate measured background and a local traffic increment. We then attempt to explain both contributions to the extent possible with the available models and relate them to the relevant emission sources; unexplained residuals are kept constant. The same scheme is applied for stations classified as background or industrial stations, however without calculation of a roadside increment.
With this modelling scheme, GAINS is able to estimate NO 2 concentrations for more than 1950 monitoring stations, more than 400 of which are traffic stations, including a large fraction of the currently exceeding stations. Scenario calculations until 2030 have been provided as an input to the revision of the TSAP (Amann et al., 2013). As an application of the modelling scheme, we here discuss the projected evolution of compliance with NO 2 limit values under current legislation.
The remainder of this article is organised as follows. Section 2 provides a detailed description of the modelling scheme. Section 3 presents a validation of modelled concentrations against past observations. Necessary simplifications contained in the modelling scheme lead to station-specific uncertainties and limitations of the model, which are discussed in Sect. 4. In Sect. 5, the evolution of NO 2 concentrations under current legislation is assessed. Section 6 presents a summary and draws conclusions.

Overview
The EU air quality legislation specifies two different limit values for NO 2 concentrations: a limit of 40 µg m −3 on annual mean concentrations, and a limit value of 200 µg m −3 on hourly mean concentrations, not to be exceeded more than 18 times per calendar year. Our modelling scheme is designed to calculate annual mean concentrations and thus cannot directly provide estimates on compliance with the limit value on hourly mean concentrations. However, in practice the annual mean limit value presents a significantly more stringent target than the hourly limit value: in 2009, all but five stations complying with the annual mean limit value also complied with the hourly limit value. The same pattern is confirmed throughout previous observational years. Therefore, although we focus only on annual means, conclusions are also possible about compliance with the hourly limit value.
The modelling scheme combines past monitoring data with bottom-up emission modelling. The starting point of all calculations is monitoring data reported to AirBase in 2009. To ensure quality of the data, we consider only stations with more than 80 % temporal coverage of the hourly data. Air-Base comprised more than 2500 NO 2 monitoring stations in the EU in 2009, with a substantial subset of these monitoring data also used to determine compliance with limit values.
A schematic overview of the modelling scheme is shown in Fig. 1 observed background 3 and the roadside increment. This first step of disaggregation is essential, since the two contributions have different sources and hence need to be modelled independently. For a station labelled as background station in AirBase, no disaggregation of the observed concentration is undertaken.
As a second step, we attempt to explain the two contributions with the model and relate them to the relevant emissions. Background concentrations are calculated as a regional scale background and an urban scale increment, as described in detail in Sect. 2.2. The residual from calculated to observed background concentrations constitutes an unexplained part that is kept constant for scenario applications. The roadside increment is calculated with a simple parametrised box model, detailed in Sect. 2.3.
In the third step, the application to future emission scenarios, we modify each component following the change in related emissions, using the parameters determined in step two (see Sect. 2.4).

Modelling background concentrations
By definition, background concentrations at a roadside station are made up from all contributions other than local traffic in the street where the station is located. This encompasses large-scale transboundary transport of pollution, but also an urban background increment caused by low-level emissions from within the rest of the city.
Due to the short chemical lifetime of NO 2 , it is advantageous to treat NO 2 as a member of the reactive nitrogen (NO x ) family. All bottom-up background calculations described in this chapter are done for NO x , and only in a final step is the partitioning into NO 2 and NO calculated. NO x concentrations resulting from emissions in the entire EU are modelled in GAINS based on linear transfer coefficients calculated with the EMEP MSC-W Chemical Transport Model (CTM; Simpson et al., 2012) at a resolution of 0.5 • (lon) × 0.25°(lat) or roughly 28 × 28 km; see Sect. 2.2.1 for further details. To capture sub-28 km structures, we use a full year simulation performed with the CHIMERE CTM (Menut et al., 2013;Schmidt et al., 2001;Vautard et al., 2005;Bessagnet et al., 2008) at a resolution of 0.125°× 0.0625°or roughly 7×7 km. The CHIMERE grid is nested in the EMEP grid, so that 4 × 4 CHIMERE grid cells match exactly one EMEP grid cell.
Both models use the same spatial disaggregation of emissions, which has been generated specifically for this work, as detailed in Appendix A.
The spatial domain of the CHIMERE model, which is also the domain considered in the station modelling scheme as a whole, can be seen in Fig. 2.
The different steps of the background modelling are discussed in detail in the following subsections.

The 28 × 28 km background: EMEP
The EMEP sensitivity simulations use meteorological fields of the years 2006, 2007, 2008, 2009, and 2010, and the emissions expected in the year 2020. In each sensitivity run, NO x emissions from one source region were reduced by 15 %, allowing the change in concentrations at each grid point to be related to the change in emissions. In total, 53 source regions were included: EU-28, Switzerland, Norway, Iceland, Serbia and Montenegro, Belarus, Bosnia-Herzegovina, Turkey, Ukraine, Moldova, FYR Macedonia, Georgia, Armenia, Albania, Azerbaijan, Russia (European part), and 10 sea regions as defined by Campling et al. (2013).
The NO x transfer coefficient ν(i, r) from source region r to receptor grid cell i = i lon i lat is defined as with [NO x ] denoting the annual mean NO x concentration in grid cell i, E(r) the annual total NO x emissions from source region r, and the labels base and red referring to the base case (full emissions) and the reduced emission model run.
From the transfer coefficients, concentrations [NO x ](i) are calculated as The constant δ(i) quantifies the residual NO x concentration emerging from hemispheric background, natural sources and non-linearities in the system. It is calculated by inserting the base case emissions and concentrations in Eq. (2), For most European regions, this residual is less than ±10 % of base case modelled NO x , with maxima of −20 % or around −3 µg m −3 in the Po valley.

Downscaling to 7 × 7 km: CHIMERE
We employ the CHIMERE CTM at a resolution of 0.125 • × 0.0625 • using meteorological fields and emissions for 2009. Significant modifications have been included in the CHIMERE code to ensure optimal simulation of urban conditions: WRF 3-D meteorological variables were substituted by ECMWF-IFS data, to avoid overestimation of wind speed and thus too fast dispersion of emissions within urban areas (Miglietta et al., 2012). Furthermore, to account for the urban canopy influence on meteorology, wind speed and vertical diffusion (K z coefficient) were modified empirically. In the lowest model layer, the wind speed was multiplied by a factor 0.5, in accordance with Solazzo et al. (2009), who found in a computational fluid dynamics study that the wind speed in the urban canopy layer is roughly half of that in the undisturbed boundary layer.
In the CTM output, a robust linear relation exists between the low-level NO x emissions and concentration increments.
Let m = m lon m lat identify a 7 km CHIMERE grid cell and i(m) identify the 28 km EMEP grid cell that contains m.
Higher-than-average NO x low-level emission density e L in sub-grid m(i) leads to a corresponding increase in ground level NO x concentrations above the EMEP grid average, and vice versa for negative deviations. We can find a regression coefficient ξ(i) linking concentration increments to emissions so that Here e L (i(m)) denotes the average low-level emission density in the 28 km grid, calculated by averaging the 7 km emission densities e L (m) over the respective 28 km grid i. e L contains all sources of emissions that are released into the lowest vertical layer of CHIMERE: domestic, road traffic, and non-road traffic emissions (in the SNAP nomenclature 4 , sectors 2, 7, and 8). ξ is calculated for every EMEP grid cell by regressing emission increments e L (m) − e L (i(m)) against concentration increments [NO x ](m)−[NO x ](i(m)) and contains all local meteorological characteristics. A map of ξ for the whole domain is shown in Fig. 2. The impact on concentrations of a single additional ton of pollutant being emitted in a grid cell differs significantly throughout Europe, owing to the different meteorological conditions. For example, frequent stagnant conditions lead to higher values of ξ in the Alps and the Po valley than in northern Europe. Since this resolution-dependent concentration increment is relevant mostly in urban areas, we refer to it also as urban increment, although it is calculated for every EMEP grid cell regardless of its location. EMEP grid cells containing parts of the same urban area are combined in the regression analysis, thus enhancing the statistical significance of the calculation. Each major city is thus assigned a single characteristic value of ξ . Two examples of the data behind the ξ calculation in European cities are shown in Fig. 3: Paris (top panel) and Milan (bottom panel). Although both cities show a compact linear correlation between emissions and their concentration effects, the slopes (ξ values) differ by a factor of two. Similarly high R 2 values are found throughout most of Europe, in particular in urban areas where emission and concentration increments are high.
The modelled 7 × 7 km concentration given by Eq. (4) has a tendency to underestimate observed urban background concentrations in medium-sized and smaller cities. This is at least partly related to the fact that 7 km is still rather large for smaller cities, which do not fill entire grid cells, especially if they are distributed over several grid cells. Thus we undertake a further disaggregation of emissions into urban and suburban parts of a given CHIMERE grid cell m, which is described further in Sect. 2.2.3.

Further disaggregation: inner-urban concentrations
Structures smaller than the 7 km CHIMERE grid are not reflected in the emission inventory nor in the model output. Many emissions, however, scale with population density, which is available at a much finer resolution. Here we have used population density on a resolution of roughly 1 × 1 km (Gallego, 2010) to determine urban polygon shapes for European cities with a population > 100 000, as detailed in Appendix B. These polygons are used to split CHIMERE grid cells partly containing urban areas into an urban part (red in Fig. 4) and a suburban part (yellow in Fig. 4). Emissions from the domestic sector and from all traffic sources except heavy duty trucks are redistributed according to population numbers inside and outside the urban polygon, so that for these sectors, s, the inner urban emission density, e u , from region r in m is then given by with pop u and A u representing population and area inside the urban polygon, pop tot and A tot those of the entire grid cell. NO x concentrations inside the urban polygon are calculated from Eq. (4) using the inner urban emission density e u instead of the 7 km grid average e. Outside the urban polygon, emission densities are lowered from the grid average by the equivalent factor. Whether the urban or the suburban concentration is taken as representative for the background of a given traffic station (or the modelled concentration at a background station) depends entirely on the location of the station inside or outside the urban polygon.
The background calculation scheme detailed in Sects. 2.2.1-2.2.3 is used to calculate NO x concentrations. In a final step, NO 2 concentrations are calculated from NO x by applying the modelled NO 2 /NO x concentration ratio from the 2009 CHIMERE simulation in each grid cell individually.

Modelling the roadside increment
For traffic stations, the local roadside increment is calculated with a simple parametrised box model that considers NO x -O 3 photochemistry and mixing with the background. The chemical core has been described by Palmgren et al. (1996) and Düring et al. (2011) and is also used in the OSPM parametrised street pollution model (Berkowicz, 2000).
The model calculates steady-state NO 2 concentrations in the street canyon from roadside NO x concentrations, the share of NO 2 in local NO x road traffic emissions, concentrations of NO 2 , NO x , and O 3 in the urban background, and the characteristic mixing time of air in the street canyon.
The mixing time of air is a station-specific characteristic, depending on the layout of buildings around the station, typical wind speeds etc. Since these data are not readily available for all traffic stations in AirBase, we estimate the mixing time from past observations. Details on the traffic increment model formulation are given in Sect. 2.3.1. The Monte Carlo-based parameter estimation process is described in Sect. 2.3.2.

The parametrised model
A detailed description of the steady-state model is given by Düring et al. (2011). Only NO oxidation by O 3 and NO 2 photolysis are considered, as all other reactions involving NO x happen on timescales distinctly longer than the typical residence times of air in the street canyon (around 40-100 s). The chemical equations involved are with k and J the reaction constants for NO 2 formation and photolysis, respectively. Under the influence of NO and NO 2 emissions on the one hand and mixing with the background on the other hand, steady-state roadside NO 2 concentrations are given by (Düring et al., 2011) [ The following quantities are used: [

Estimation of site-specific parameters
In Eq. (9), a traffic station is characterised by a single parameter, the characteristic residence time of air in the street canyon (τ ). If all other quantities entering the calculation are known from observations, it is possible to invert Eq. (9) and calculate τ . However, in doing so, all uncertainty is concentrated in this one parameter. Reasonable ranges for τ are known to be between 40-100 s (Düring et al., 2011). For several traffic stations, direct inversion of Eq. (9) gives physically meaningless values far outside this range or even below zero. Parts of the difficulties here are connected to uncertainties in the other input quantities, in particular the measured background concentrations. There are good arguments for different ways of calculating an appropriate background concentration from observations (see Sect. 2 above for details on the methodology applied here), but none can be truly representative for the background air above the roof level close to the traffic station with which the air in the street canyon is mixed: by definition, background stations are typically located somewhat remotely from busy roads. Even within one city, observed urban background concentrations of NO 2 and NO x vary strongly (Cyrys et al., 2012). It is thus reasonable to assume that the observed city average background may not be fully representative of the local background concentrations entering Eq. (9) -in some cases, a representativeness correction is needed. This is designed here as a relative correction in order to make it easily transferable to future scenario years in which the mean value may have changed substantially: The same parameter λ NO x is used in Eqs. (10) and (11) to ensure consistency between NO 2 and NO x background representativeness corrections. σ NO x , σ NO 2 , and σ O 3 are the average relative standard deviations of NO x , NO 2 and O 3 background monitoring stations, averaged over all European cities with more than two background stations operational in 2009.
We also recognise that the NO 2 emission share p is not known exactly for a given station, but do not include it in our analysis, as a relative correction cannot be easily transferred to the future (p is determined by the technology mix which changes over time).
With the above parameters, we reformulate the problem of characterising a traffic station as an estimation problem: given the measured value of NO 2 , and the measured (or modelled, in the case of p) input quantities, what are the most likely values for the parameters τ , λ NO x , and λ O 3 ?
Estimation theory provides several solution pathways for such a problem. Most analytical ones, however, require the model equation to be linear in its parameters, which is not the case with Eq. (9). We solve the estimation problem by minimising a penalty function through Monte Carlo simulation, similar to the approach described by Dilks et al. (1992). For this purpose, we assign distributions to the parameters: λ NO x and λ O 3 are assumed to follow a standard linear distribution (mean = 0, variance = 1), while τ is assumed to be log-normally distributed with median and width approximately as given in the literature:τ = 70 s and σ τ = 30 s. The log-normal distribution is chosen mainly to strictly avoid negative numbers of τ . The parameters are assumed to be uncorrelated.
Mathematically, we seek to estimate parameters from the a priori information supplied in the input distributions. Equation (9) is strongly under-determined if three parameters are inserted: from just one observed variable, annual mean [NO 2 ], it is impossible to extract enough information to estimate three parameters -in this case, the solution would depend strongly on the assumed a priori distributions.
A way forward is to exploit more of the actual information available from AirBase and apply Eq. (9) not to annual mean, but to hourly values. Theoretically, up to 365 × 24 = 8760 measurement points are available for each station in the base year. However, the time series is dominated by short-term fluctuations in turbulent mixing conditions and emissions, which are not represented in the simple model formulation.
To eliminate the noise from short-term meteorology, we consider the annual average weekly pattern instead, in which the time series has been averaged according to the hour of the week (see Fig. 5a). Thus we effectively use 24 × 7 = 168 data points for the estimation process, thereby turning the under-determined equation into an over-determined system of 168 equations with three unknowns.
For the estimation process, we acknowledge that also observed roadside [NO 2 ] is not known perfectly well but subject to some (observational) uncertainty. Effectively, we search for a solution to Eq. (9) which matches observed [NO 2 ] as well as necessary (to stay within the limits of observational accuracy), while altering input parameters as little as possible from their a priori values. For this purpose, we introduce a penalty function that weights residuals in calculated steady-state [NO 2 ] ss against departures of input parameters from their expected values: σ obs represents the acceptable deviation between model and observation, which is empirically set to 1 % of the observed concentration. The solution is found by minimising f . Since the calculation is performed for every hour of the week, we minimise the sum of f calculated for every hour. We perform a Monte Carlo simulation, drawing for each parameter a large number of samples (n = 100 000). The best 1 % of solutions then make up the a posteriori distributions, which can be quite well approximated as Gaussian distributions (see Fig. 5c-e).
Their mean values are taken as the solution, while the width of the distributions may be used in an uncertainty analysis. The aim here is to determine a robust solution; we do not necessarily require the solution to be optimal in the sense of estimation theory.
An example for the parameter estimation is shown in Fig. 5 for the Landshuter Allee traffic station, a monitoring station located at a busy road in Munich. The observed weekly pattern of roadside NO 2 (panel a), showing the characteristic fingerprint of a traffic station, is well reproduced by the model; however, the explained concentration range is large if the whole a priori range of parameters is considered (green area). The estimation process narrows the relatively broad a priori distributions of parameters (black lines in panels c-e) to the most likely values (red lines in c-e), resulting in the a posteriori range of modelled NO 2 concentrations (yellow area in a).
The parameters determined in this step are meant to give a simple representation of the station characteristics. They are calculated for the year 2009 and are assumed to remain unchanged in the past and the future.

Combining background and roadside increment models
On the European average, NO 2 at urban traffic stations is made up of roughly equal contributions from urban background and roadside increment in the base year. In individual countries, this allocation varies between 40 % and 70 %. The roadside increment itself is composed of roughly 50 % primary emitted NO 2 and 50 % oxidised NO emissions. This apportionment into "primary" and "secondary" NO 2 varies between different countries, depending on the country-specific value of p determined by the fleet mix (particularly the share of diesel vehicles).
Once the different contributions from regional background, urban increment, and traffic increment have been established from observations for the base year 2009, and the station-specific parameters have been determined from Monte Carlo-based estimation, the model is ready for scenario analysis. Essentially, this means replacing the emis- The NO 2 /NO x partitioning ratio used to derive urban background NO 2 concentrations from urban background NO x is kept constant for the future, in accordance with the assumption that urban background O 3 levels remain unchanged. A discussion on the possible limitations of this approach is given in Sect. 4.
The formulation of the traffic increment model implies that knowledge of the exact amount of traffic emissions generated in the individual street canyon is not necessary. Instead, roadside NO x concentrations are needed. Since NO x is chemically inert at timescales of seconds to minutes, we calculate roadside NO x concentrations in any scenario year y by scaling the observed increment in the base year y 0 with the trend in road traffic emissions,

Emission calculation
Emissions are calculated in the GAINS model bottom-up from a given projection of activity (fuel consumption), accounting for fleet turnover and penetration of vehicles with different stages of emission controls. We calculate emissions for countries (using national emission factors), and for average urban conditions in each country. Vehicle fleets are known to differ in composition in urban areas from the national average (Carslaw et al., 2011b): the share of heavy trucks is much lower in urban driving than nationally, yet there are more buses and motorised twowheelers. To account for these general differences we determined for each vehicle category its share of distance driven in urban areas, based on model estimates from COPERT  or national inventories where available (Carslaw et al., 2011a;Knoerr et al., 2012). Also unit emissions differ in urban areas from the country average: in particular for heavy duty vehicles it is known that NO x exhaust after-treatment has not functioned well at urban speeds, and hence unit emissions are several times higher than, e.g. for highway driving. We account for these differences by emissions factors derived explicitly for typical urban driving conditions for each country, with vehicle category and Atmos. Chem. Phys., 14, 813-829, 2014 www.atmos-chem-phys.net/14/813/2014/  Table 1). In the GAINS model, p shares from the Handbook Emission Factors for road transport (HBEFA v3.1, http://www.hbefa.net) are used, which are based on chassis dynamometer tests over various driving cycles (Hausberger et al., 2009). There are no indications that this share differs between driving conditions. Therefore the same shares are used for both national and urban driving.

Validation: explaining historical NO 2 trends
Validating a model designed to provide results for more than 1950 individual monitoring stations across Europe poses a considerable challenge. Here we show a validation of the background modelling in the base year for all background stations on the one hand, and a validation of trends of the station-based model results on the other hand.
As a validation of calculated (explained) NO 2 urban background, Fig. 6   is lower for urban background stations within defined urban polygons. This highlights the added value of the last downscaling step beyond the 7 × 7 km resolution within cities, which becomes even more important for smaller cities. For practical reasons, a lower limit of 100 000 inhabitants was used when calculating urban polygon shapes from population density maps, leading to the underestimation of urban background concentrations in cities falling below this limit (panel b).
The ability of the parametrised traffic station model to reproduce well the observed strong variations in NO 2 concentrations in the base year is demonstrated exemplarily in Fig. 5a, with similarly strong agreement also observed at most other traffic stations.
For years other than the base year, the NO x traffic increment is scaled with the trend in urban NO x traffic emissions. This relation is shown in Figure 7 for Germany: here the trend in observed NO x roadside concentration increments is compared to the trend in NO x road traffic emissions, using either national average or urban driving conditions. The observed values contain all roadside stations with a NO x increment >15 µg m −3 in 2009 and observations available for at least nine out of the ten years 2000-2009. The spread of trends measured at individual stations is indicated as shaded area (standard deviation of the mean). Observed trends in NO x traffic increments are captured perfectly by the calculated NO x emission trends, however, only if urban driving conditions are employed. Similarly good agreement is found in other countries with sufficiently large numbers of traffic stations to allow for a good statistical sample (in particular, France, Austria and the UK). Figure 8 shows the trend of NO 2 and NO x concentrations measured and modelled at all roadside monitoring stations in the EU. Only stations with monitoring data available for at least nine out of the ten years are selected. The traffic station model reproduces both trends almost perfectly well. While NO x concentrations decline by around 25 %, NO 2 concentrations are stagnant. This discrepancy is related to strong increases in primary NO 2 emissions (Fig. 10, panel b) due to increased use of diesel vehicles, offsetting the effect of NO x emission reductions.
Further distinguishing into Member States, Fig. 9 shows for all EU Member States with sufficient data absolute NO 2 concentrations averaged by station category (rural background, urban background, and roadside stations) for the period 2000-2009. Since the model is calibrated to monitoring data in 2009, consistency between observations and model calculations is inherently forced in this year; the other years, however, are independent of observations. To improve consistency in the observational record, only stations with more than five years of data available in the ten year period have been selected, and only countries with at least three such traffic stations are shown in Fig. 9. The same set of stations is used here for both observations and model results. Trends are in general well reproduced for all station types, especially in Austria, Germany, Netherlands, France, Sweden, and some other countries. In some countries (notably Belgium, Italy, Portugal, UK) modelled trends at traffic stations are more optimistic than observations show. A potential reason for this behaviour is the strong increase of the share of primary NO 2 in road traffic NO x emissions (p in Eq. 9) in these countries, which is taken into account in the street canyon module but not for the urban background: in the same countries modelled urban background NO 2 concentrations also decline too fast. This is further discussed in Sect. 4.

Uncertainties and robust solutions
Owing to the scope of the model linking European-scale emission calculations with impacts in individual street canyons, simplifications are taken that induce uncertainty. A systematic source of uncertainty stems from the assumptions on urban background NO 2 . In the absence of reliable predictions regarding the evolution of future urban background O 3 levels, these are assumed to remain constant. Furthermore, it is assumed that urban background stations and the air masses they represent are sufficiently remote from traffic emissions to ensure that the partitioning of NO 2 and NO in urban background NO x is determined by ambient O 3 rather than the initial share of NO 2 in traffic emissions (p). Under these assumptions, urban background NO 2 follows the same trend as urban background NO x . In reality, this is not completely fulfilled (see e.g. Keuken et al., 2009), and urban background NO 2 declined more slowly than NO x , especially in countries with a strong increase in the share of primary NO 2 in NO x emissions. Altogether, this results in backward calculated urban background NO 2 in the year 2000 over-estimating observations by ∼ 2 µg m −3 on the European average. However, the influence of the background trend on modelled roadside concentrations is limited, as the traffic increment calculation does take into account p variations. From GAINS calculations (Fig. 10b) as well as recent observations (Carslaw and Rhys-Tyler, 2013) the increases of p are expected to slow down in the future, further decreasing the relevance of this systematic deviation.
The NO 2 /NO x emission shares used in the model are subject to considerable uncertainty (see Table 1). However, errors in p are partially compensated by the parameter Atmos. Chem. Phys., 14, 813-829 -Errors in observations or meta information. The direct involvement of observations can be both benefi-cial and detrimental to the quality of the analysis, as it induces a strong dependency on the quality of monitoring data and station meta-information supplied by AirBase. Apart from measurement uncertainties, the positioning and classification of stations into traffic or background stations is crucial in our scheme for the disaggregation of observed concentrations into urban background and roadside increment, and the identification of the unexplained component.
-National emission trends. Emission trends are calculated on a national basis using national total or urban fleet compositions and average emission factors under average national or urban driving conditions. The same is true for the share of NO 2 in NO x emissions. However, it is known from detailed analyses that traffic composition and emissions can vary significantly between stations in the same city or region.
From these considerations, it is obvious that results for individual stations have to be used with caution. Rather than actual site-specific predictions, they illustrate the likely evolution of pollution levels if local conditions are well in line with national average trend assumptions, and if no additional local measures are taken. In practice, the influence of local action or inaction may well dominate the actual resulting trend. It is not attempted to model these influences here, but rather to quantify the concentration changes that may be expected from European-wide emission control efforts, and quantify the complementary need for additional local action for achieving compliance.
As a way to deal with the random uncertainties connected with predictions for individual stations, a statistical evaluation of results is optimal. One way to do this is illustrated in Fig. 9, showing country-wide averages of concentrations for different station categories. For compliance estimates, however, the important information is the concentration at each site compared to the limit value. Therefore we analyse stations aggregated into compliance classes according to their modelled concentrations.

Application to the future: attainment of NO 2 limit values under current legislation
Applying the modelling scheme introduced in this article, we here assess the evolution of (possible) compliance with NO 2 limit values until 2030, assuming that only currently approved legislation is successfully implemented. This emission scenario has been used as the baseline scenario in the ongoing revision of the EU Thematic Strategy on Air Pollution. Section 5.1 provides a short description of the assumptions taken for the generation of the emission scenario. With road traffic being the most important source of NO x emissions for monitoring stations exceeding the limit value, the scenario description here is limited to road traffic. A discussion of other sectors, such as industrial emissions, can be found in Amann et al. (2013). The projected evolution of the compliance situation under this emission scenario is discussed in Sect. 5.2.

Trend scenario for NO x emissions from road transport
The prospective NO x emissions under trend assumptions are calculated in five-year steps from 2000 to 2030 combining the following elements: -Transport and consequent fuel demand by vehicle category, based on the latest PRIMES scenario for the European Union (Amann et al., 2013),

National trend for transport demand
The PRIMES model provides a consistent projection of supply and demand of energy in European Member States including the future transport demand by mode. We here use the latest PRIMES 2012/13 Reference scenario to project vehicle mileage and fuel demand. Particularly relevant for NO x emissions is a projected strong growth of diesel cars and an assumed decline in the usage of gasoline cars in almost all EU Member States.

Fleet composition and turnover
The fleet composition for the historic years is based on results of the FLEETS project (Ntziachristos et al., 2008), which produced a consistent data set of detailed vehicle and activity information for road transport for each Member State. This data set has been further updated and reviewed in the framework of the LIFE EC4MACS project (see www.ec4macs.eu) and for this study based on input from national experts. Notably the purchase of used cars was modelled in detail in many countries (Mehlhart et al., 2011), resulting in an older fleet in these countries. The fuel consumption calculated bottom-up is calibrated with the fuel demand as projected by PRIMES, providing an internally consistent data set of fleet composition. A detailed description of the methodology is given by Ntziachristos and Kouridis (2008).

Emission factors for future vehicles
Euro 5 and Euro 6 diesel cars are particularly relevant for future NO x emissions. Real-life emission factors for Euro 5 diesel cars and light trucks are based on recent measurements indicating higher NO x emissions than those of Euro 4 and Euro 3 cars, exceeding the limit value defined for type approval several times (Hausberger, 2010). This analysis employs a national average value of about 870 mg NO x / km, i.e. almost five times higher than the nominal limit value. While the performance of Euro 6 cars under real-world driving conditions is uncertain, we assume here (optimistically) that real-world driving emissions will indeed decline in two phases: a first generation of Euro 6.1 vehicle is assumed to emit about 380 mg NO x / km, i.e. decreasing Euro 5 emissions by the ratio of Euro 6 to Euro 5 limit values. It is further assumed that for a second generation of Euro 6.2 vehicles the type approval testing will be complemented by, e.g. onboard portable emission measurement systems or random cycle tests. As a consequence, new vehicle types from 2017 onwards are assumed to emit 120 mg NO x /km under real-world driving. Early demonstration vehicles have shown that selective catalytic reduction (SCR) technology with injection of 5 Emissions available from http://www.ceip.at.  appropriate amounts of urea could deliver large reductions in NO x emissions over Euro 5 vehicles, both for type-approval and real-world cycles (Demuynck et al., 2012;Weiss et al., 2012).

Baseline trends for NO x emissions
NO x emissions from gasoline cars are expected to decrease quickly, respectively amounting in 2020 and 2030 to only 20 % and 10 % of the 2010 levels. Comparable reductions are expected from heavy duty trucks and diesel buses. NO x emissions from diesel cars start declining later and decline more slowly, respectively reaching 65 % and 31 % of 2010 values in 2020 and 2030. NO x emission trends of passenger cars are shown in Fig. 10a. While total NO x emissions from road transport decrease, the fraction of NO x emitted as NO 2 increases further, however, gradually levelling off at an average of 25 % (Fig. 10b).  Figure 11 shows the modelled evolution of NO 2 concentrations in the EU-28. To reflect unavoidable uncertainties in monitoring data, modelling techniques and future meteorological conditions as discussed in Sect. 4, results are expressed as numbers of stations aggregated in different "compliance categories": annual mean NO 2 concentrations below 35 µg m −3 (i.e. at least 5 µg m −3 below the limit value) indicate rather likely attainment of the limit value (shaded in green). For concentrations computed between 35 and 45 µg m −3 (shaded in yellow), attainment is possible but uncertain due to the factors mentioned above. Stations with modelled concentrations higher than 45 µg m −3 (shaded in red) are unlikely to attain the limit value. The ±5 µg m −3 range was chosen as it corresponds to the maximum deviation between modelled and observed NO 2 at the European average of all traffic stations in any of the years 2000-2008.

Results: attainment of air quality limit values
It does not reflect a defined level of likelihood in the mathematical sense. The set of stations considers only those modelled in the base year; as new monitoring stations are put into operation, the absolute numbers may change in the future. The projected decline in NO x emissions should significantly improve future compliance with NO 2 air quality limit values, decreasing the number of stations severely exceeding the limit value (NO 2 > 45 µg m −3 ) by a factor of four by 2020 in the EU-28; by 2030, if Euro 6 vehicle emission standards work as effectively as assumed here, almost all cases of severe non-compliance should be eliminated. However, by 2030 around 30 stations are still projected to remain in the "attainment uncertain" category, for which reliable conclusions about attainment are impossible due to uncertainties in the modelling scheme as discussed in Sect. 4.
To illustrate the strong geographical contrasts, a map of the spatial distribution of stations in different compliance categories is shown in Fig. 12  achieving compliance are expected to remain for many of the major European cities, such as London, Paris, Rome, Madrid, Milan, the Ruhr area or Athens. Since these major cities are home to a significant proportion of the European population, the few stations exceeding the limit value by 2025 or even 2030 have disproportionally larger implications on possible population exposure than their small absolute number in Fig. 11 would suggest. As discussed in Sect. 4, a fundamental assumption behind the projected concentrations is that emission trends at each station follow the national average (for urban conditions if applicable). Local measures, be they beneficial or detrimental, are not modelled, though they may have a major impact on resulting concentrations. In particular, additional local action may be required to ensure attainment of the limit value at the remaining critical stations.

Conclusions
This paper presents an introduction to the downscaling methodology that has been implemented in the GAINS integrated assessment model for estimating compliance with air quality limit values. Employing a chain of models of different scales, we link European-wide emission control policies with impacts in street canyons.
Results are calculated for more than 400 roadside and around 1550 background monitoring stations reporting to AirBase, including 70 % of the stations exceeding the NO 2 limit value in 2009. Station-specific characteristics of roadside stations are reflected in the parameters estimated from past observations. In the absence of more detailed information for individual stations, emission trends are taken from national calculations, enabling us to analyse station ensembles rather than individual stations. On the other hand, the scheme may as well be applied to model individual stations if localised emission data become available.
The model reproduces well the observed trends in the decade [2000][2001][2002][2003][2004][2005][2006][2007][2008][2009]. NO 2 concentrations were relatively stagnant in this decade, especially at urban traffic hot spots, in spite of declining NO x emissions. This is explained by strong increases in the share of primary NO 2 emissions, predominantly due to increasing shares of diesel cars in European fleets.
In spite of the stagnant NO 2 levels in the past, the model projects declining NO 2 concentrations in the future, if currently approved legislation is successfully implemented. The main reason behind the expected change in trend is the assumption that new exhaust cleaning technologies introduced with the Euro 6 emission standard will deliver strong reductions in real-world NO x emissions from diesel cars from 2015/2017 onwards. In general, our projections show that under the assumptions taken, Europe is on the right track towards achieving compliance with NO 2 limit values in the medium term. At some monitoring stations, however, safe attainment of limit values is not even foreseen until 2030, pointing to the need for additional European legislation or complementary local action to reduce NO x emissions. First the large point sources (LPS) from the fine-scale (0.125 • × 0.0625 • ) TNO-MACC emissions data for 2007 were added to surface emissions to get only one type of emissions; emission heights were then ascribed in a second step using standard profiles as documented by Bieser et al. (2011). For the various activity sectors the processing steps were the following: Atmos. Chem. Phys., 14, 813-829, 2014 www.atmos-chem-phys.net/14/813/2014/ -SNAP 2: the country emissions were re-gridded with coefficients based on population density and French bottom-up data, the methodology was extrapolated to the whole of Europe. To give a more realistic representation of increased domestic heating emissions in winter, a temperature proxy (degree days) was used for the temporal modulation of SNAP 2 emissions.
For countries where MACC-TNO emissions are not available EMEP 0.5 • × 0.5 • emissions were used (Iceland, small countries, Asian countries). This inventory is interfaced with the CHIMERE model using chemical speciation and adequate temporal profiles. The procedure is fully documented by Menut et al. (2012).

Appendix B Urban polygon shapes
Shapes of 473 European cities and urban agglomerations with more than 100 000 inhabitants were constructed based on population density distribution at a 0.01 • × 0.01 • resolution which were obtained from the Institute for Environment and Sustainability, Joint Research Centre in Ispra (see Gallego, 2010).
Population numbers for cities and urban agglomerations were obtained from the www.citypopulation.de database. Information about the location of cities was taken from the Global Rural-Urban Mapping Project (GRUMP, http://sedac.ciesin.columbia.edu/data/dataset/ grump-v1-settlement-points).
From the above-mentioned input data, city and agglomeration shapes were constructed using GIS. The algorithm was based on ascending sorting of grid cells according to a weighting function x, with d distance from city center and σ pop population density in the grid cell. The cut-off for x (determining the city shapes) was based on the condition that the differences between the reported numbers of population living in each city and the population numbers calculated inside the corresponding shapes are minimised. Thus the value of the cut-off was set by the minimum of RMSE calculated for these differences.