uEMEP provides a consistent methodology for redistributing/downscaling gridded CTM emissions and concentrations, in this case from the EMEP MSC-W model, to higher-resolution subgrids within the CTM grids. Proxy data, that represent the spatial distribution of the emissions, are used to redistribute emissions to subgrids. Concentrations are then recalculated at the subgrid level, using a Gaussian model, whilst removing the local contribution of the CTM at these subgrids but keeping the non-local contributions. This procedure consistently avoids double counting of emissions.

Throughout this paper, we refer to the downscaling grids in uEMEP as “subgrids”. These may be any size but current applications use a range of between 25 and 250 m. When referring to the EMEP MSC-W model, we use the term “grid”. This may also vary dependent on the application but is usually in the range of 2 to 15 km. The term “local” is also used. Local, in regard to EMEP, means the local EMEP grids, so terms such as “local fraction” refer to a particular grid and the other EMEP grids in the “local region”, for example, within a range of 5×5 EMEP grids. When referring to “local” in uEMEP, we also refer to subgrid contributions from within this local EMEP region. This is visualised in Fig. 1. “Non-local”, in regard to uEMEP, refers to any contribution that is not downscaled or calculated with uEMEP, usually contributions from outside the local EMEP region but these can also be other source sectors not accounted for by uEMEP. We will refer to concentrations provided by the EMEP MSC-W model simply as EMEP concentrations.

uEMEP can be run using two downscaling methods, both of which make use of Gaussian dispersion modelling to describe the redistribution of concentrations at high resolution. Both methods make use of the recently developed “local fraction” functionality in the EMEP model (Sect. 2.3; Wind et al., 2020) to avoid double counting of emissions and to allow near-seamless integration of the two models. The two downscaling methods are as follows:

Emission redistribution. Redistribution of the EMEP gridded emissions using emission proxy data to subgrids and calculation of concentrations through Gaussian modelling, with removal of the locally emitted EMEP concentrations.

Typical source sectors downscaled using uEMEP include traffic, residential combustion, shipping and industry. The sectors addressed will depend on the availability of high-resolution data for distributing them. uEMEP is only applied to the primary emissions of PM10, PM2.5 and NOx and does not involve any complex chemistry or secondary formation of particles. The concentrations of NO2 and O3 are calculated with uEMEP using a simplified chemistry scheme (Sect. 3.4 and 3.5).

The choice of downscaling method will depend on the quality of the high-resolution proxy or emission data available, whether the calculation will be made on hourly or annual data and on the EMEP model resolution. The first downscaling method, emission redistribution, will be applied when only approximate proxy data for emissions are available, which will be the case in many large-scale applications. Examples of such proxy data may be population density, road network data or land use data. The second downscaling method, independent emissions, is suitable for when high-quality emission data are available that are consistent between uEMEP and EMEP. When the gridded emission data are entirely consistent with the proxy data, i.e. the proxy data are given as emissions and are aggregated to the CTM grid emissions, the two methods are equivalent.

The total subgrid concentrations CSG(i,j) at subgrid indexes (i,j) are calculated by adding the local Gaussian concentration CSG,local(i,j) and the non-local part of the EMEP grid concentration CSG,nonlocal(i,j) and is written simply as 1CSGi,j=CSG,locali,j+CSG,nonlocali,j, where we use the subscript notation “SG” to denote any subgrid value and in further equations the subscript “G” to indicate any EMEP grid value. CSG,local(i,j) is determined directly from the subgrid dispersion calculation: 2CSG,locali,j=∑i′=1nx∑j′=1nyESG,locali′,j′Ui,j,i′,j′⋅Iri,j,i′,j′,h(i′,j′),zi,j. Here, ESG,local is the emission attributed to each subgrid and U is the wind speed, which in uEMEP is dependent on both the source and receptor subgrid values (Sect. 3.1). nx and ny represent the extent of the subgrid calculation window. The function I(r,h,z) is the dispersion intensity function (Sect. 3.1) that determines the dispersion of the emission source ESG,local with the horizontal spatial vector r(i,j,i′,j′) between the receptor grid points (i,j) and the source grid points (i′,j′) at height z(i,j). The source height h(i′,j′) is also specified. The contribution from every proxy emission subgrid (i′,j′), within the defined subgrid calculation window (nx,ny) is calculated and summed at the receptor subgrid (i,j) centred within subgrid calculation window; see Fig. 1.

When using the emission redistribution method, ESG,local is calculated using the EMEP grid emissions EG(I,J) and the proxy data for emissions, Pemission. Pemission is normalised within the EMEP grid in the following way to determine the subgrid emission ESG,local: 3ESG,locali′,j′=EG(I,J)Pemissioni′,j′∑i′=1nx∑j′=1nyPemissioni′,j′. When using the independent emission method, the local subgrid emissions ESG,local are specified directly.

The term CSG,nonlocal(i,j), given in Eq. (1) and further derived in Eqs. (9) and (10), is the non-local contribution from the EMEP grid at the specific subgrid (i,j). Though this is based on the non-local contribution provided by EMEP at grids (I,J) interpolation due to the moving window (Sect. 2.4) surrounding each receptor subgrid means that non-local contributions are specified at the subgrid level. The gridded non-local contribution, CG,nonlocal(I,J), is derived from the “local fraction” calculation in EMEP. The methodology is described completely in Wind et al. (2020) but the essential elements are reproduced here.

The local fraction methodology corresponds to a tagging method where pollutants from different origins are tagged and stored individually. In this case, the tagging occurs relative to the surrounding grid cells of any individual grid. This means that emissions from any grid cell are tagged and followed through the various model processes out to neighbouring grid cells. Tagged species are assumed to be inert species, primary PM and NOx, for this downscaling application as chemical reactions are not included in the tagging. It is generally not computationally possible, or in this application necessary, to follow all grid cell contributions to all other grids within the EMEP model domain. The local fraction region extent is then limited. In Wind et al. (2020), the local fraction region extent (nlf) was tested up to a 161×161 EMEP grids on low-resolution EMEP runs for Europe. Generally, 21×21 EMEP grids were found to be computationally and memory efficient. In the uEMEP application, the local fraction region needs only be as large as the uEMEP calculation window, i.e. the allowed distance from the receptor subgrid to the emission subgrids. In the forecasting application discussed in Sect. 4, this requires only a 5×5 EMEP grid local fraction region. Sensitivity to the size of this region is discussed in Sect. S5.2. For each EMEP grid (I,J), there will be an associated local fraction grid LF(I,J,Ilf,Jlf) that specifies the fraction of the surrounding grids contributing to the (I,J) grid. Ilf and Jlf are indexed from -nlf/2 to +nlf/2.

With use of the local fraction, the local (CG,local) and non-local (CG,nonlocal) contributions from any particular primary pollutant to an EMEP grid (I,J) are given by the sum of all the contributing local fraction contributions of the local sources (s=1 to nsource) and the non-local contribution, specified by 4CG,localI,J,Ilf,Jlf,s=LFI,J,Ilf,Jlf,sCGI,J5CG,nonlocalI,J=CGI,J-∑I′=-nmw/2+nmw/2∑J′=-nmw/2+nmw/2∑s=1nsourceCG,localI,J,I′,J′,s. Note that in Wind et al. (2020) CG,local and CG are termed LP (local pollutant) and TP (total pollutant), respectively, and the local fraction index is specified here using (Ilf,Jlf) instead of (Δxs⋅Δys). This change is for compatibility with the notation used for the uEMEP application.

When determining the local and non-local EMEP contribution at any uEMEP subgrid receptor, a moving window methodology is applied. The aim of the moving window calculation is to represent as well as possible the local and non-local EMEP contributions at any one subgrid, in effect creating an EMEP grid that is centred on the receptor subgrid. The moving window is centred on the receptor subgrid (i,j) and its size is specified by the number of EMEP grids it covers (nmw,nmw). The moving window region is the same as the uEMEP calculation window in extent, which is also defined by the number of subgrids (nx,ny) (Sect. 2.2). nmw is given by the user but it must not be larger than the area covered by the EMEP local fraction region (nlf), i.e. nmw≤nlf-1. Figure 1 shows an example where nlf=5 and nmw=4.

Since we need to account for all source contributions from EMEP within the moving window and since the subgrids are not centred in the middle of the EMEP grids, the local contribution from the EMEP grids for any particular source sector s can be written as 6CG,locali,j,s=∑I′=-nmw/2+nmw/2∑J′=-nmw/2+nmw/2CG,localI,J,I′,J′,s⋅wi,j,I′,J′,s. Here, the weighting variable w(i,j,I′,J′,s) refers to the weighting of the EMEP grid relative to the receptor subgrid and the index I,J refers to the EMEP grid which contains the uEMEP subgrid (i,j). For EMEP grids entirely within the moving window, this weighting will be unity, but for EMEP grids only partially within the moving window this weighting will be less than unity as part of that EMEP grid will also contribute to the non-local concentrations.

There are two methods implemented in uEMEP for specifying these weights. The simplest and most often used is area weighting where only the area fraction of the EMEP grid that is within the moving window for that particular receptor subgrid is included in the local contribution. This is illustrated in Fig. 1 and is usually sufficient for the calculation, especially when the number of EMEP grids covered by the moving window is larger than 3×3. Mathematically, the area weighting, wa, can be written as 7wai,j,I′,J′=a(i,j)∩A(I′,J′)A(I′,J′), where A(I′,J′) is the area and position of each EMEP grid, a(i,j) is the area and position of the moving window centred at the receptor subgrid point (i,j) and a(ij)∩A(I′J′) is the overlapping area of these two regions. For the case where nmw=1, this area weighting is equivalent to a bilinear interpolation of the surround EMEP grids. Area weighting is not dependent on the source.

When the moving window only covers a limited number of EMEP grids and when high-resolution emission data are used that are compatible with the EMEP grid emissions, this weighting can also be based on the high-resolution emission data. This better represents the moving window concept because it reflects the effect of moving the EMEP grid to be centred on the receptor subgrid in a more realistic way. In this case, the emission weighting term (we) on the edge of the moving window will be determined by the fraction of the total subgrid emissions within the moving window and within the EMEP grid, instead of the area. This can be written as 8wei,j,I′,J′,s=∑ei,j,I′,J′,s:∈a(i,j)∩A(I′,J′)∑ei,j,I′,J′,s:∈A(I′,J′), where the numerator is the sum of the emissions within the intersection of a(i,j) and A(I′,J′) and the denominator is the sum of the emissions within A(I′,J′). The resulting total concentration, using this method, may be higher or lower than the original EMEP concentrations because it reflects the impact of moving the EMEP grid in space. This is easiest to visualise if the moving window is the same size as the EMEP grid. If the moving window was centred on an area with concentrated emissions, that are in reality spread over two EMEP grids, then when using the emission weighting the new EMEP local contribution would be higher, the non-local lower and the total would be different; see Fig. 2. The opposite is also true if the moving window were placed over a region with low emissions, the local contribution would be lower and the non-local higher. Due to this, it is not possible simply to subtract the local EMEP contribution from the total to get the non-local EMEP contribution, as detailed in Eq. (5).

To address this, the non-local EMEP contribution is also calculated using the moving window with Eq. (9). The first term is the non-local contribution for a particular source and is calculated with the area weighting distribution since non-local contributions, those outside the local fraction region, do not have any associated emission or local fraction for weighting. An additional correction term, second term in Eq. (9), accounts for the non-local contributions from local contributions on the EMEP edge grids, those parts of the EMEP grids that are outside the moving window and not included as a local contribution in Eq. (6). In those cases, the local EMEP contribution outside the moving window must be converted to a non-local contribution and subtracted from the calculated non-local value (first term in Eq. 9). 9CG,nonlocali,j,s=∑I′=I-nmw/2I+nmw/2∑J′=J-nmw/2J+nmw/2CG,nonlocalI′,J′,s⋅wai,j,I′,J′-∑I′=-nmw/2I′≠0+nmw/2∑J′=-nmw/2J′≠0+nmw/2CG,localI,J,I′→I,J′→J,s⋅wi,j,I′,J′,s+CG,localI,J,I→I′,J→J′,s⋅wi,j,I,J,s In Eq. (9), the weighting term w represents either the emission (we) or the area (wa) weighting, depending on the choice of weighting method.

These local and non-local calculations are carried out for each emission source individually so the non-local contribution is also dependent on source and the non-local component for any particular source will also contain the local contributions from the other sources. This makes creating a final non-local contribution complicated. To solve this, all the source specific CG,local+CG,nonlocal contributions are averaged and the sum of the CG,local source contributions are subtracted to obtain the final CG,nonlocal. The final non-local contribution at each subgrid CSG,nonlocal (Eq. 1) is equivalent to the EMEP non-local CG,nonlocal contribution and is calculated by 10CSG,nonlocali,j=1nsource∑s=1nsource(CG,locali,j,s+CG,nonlocali,j,s)-∑s=1nsourceCG,locali,j,s. In the case of area weighting, where the sum of local and non-local contribution is the same as the original EMEP total concentration, then the first term in the summation is equivalent to the original EMEP concentration without summation. The method is illustrated in one dimension in Fig. 2.

The calculation based on emission weighting is computationally more expensive than the area weighting and is only used when necessary and appropriate, e.g. when nmw=1 and when subgrid and grid emissions are consistent with each other.

A standard Gaussian narrow plume dispersion model formulation, e.g. Seinfeld and Pandis (1998), is used in the subgrid dispersion calculations with multiple reflections from the surface (z=0) and boundary layer height (z=H). Generically, the Gaussian plume calculation can be written as 11Cx,y,z=QUI(x,y,z), where for the sake of clarity we have dropped references to subgrid indexes as given in Sect. 2 and use coordinates instead of indices. Here, C is the concentration, Q is the emission source strength and I is the plume intensity given by 12I(x,y,z)=12πσyσzexp⁡-y22σy2∑i=1i=6exp⁡-z-hi22σz2. Here, hi represents the plume emission height and five additional virtual plume emission heights after single and double reflections from the surface and boundary layer top (H) given by 13hi=hemis,-hemis,2H-hemis,2H+hemis,-2H+hemis,-2H-hemis. For the well-mixed plume case, when σz is of the order of H, we define a threshold beyond which the plume concentration is constant throughout the boundary layer. This is specified to occur when σz>0.9H, leading to an intensity given by 14I(x,y,z)=12πσyHexp⁡-y22σy2. The Gaussian dispersion parameters (σz and σy) used in Eq. (12) may be determined empirically (Smith, 1973; Martin, 1976; Turner, 1994; Liu et al., 2015) or through a range of methods based on theoretical and semi-empirical considerations (Seinfeld and Pandis, 1998). Venkatram (1996) also discusses the relationship between empirically and theoretically based dispersion parameters. Standard Gaussian plume models do not take into account variable vertical profiles of wind speed or diffusivity. Some non-Gaussian descriptions are available based on the application of power laws to these profiles and the vertical integration of the diffusion equation (Chaudhry and Meroney, 1973; van Ulden, 1978; Venkatram et al., 2013) but this then creates the problem of defining power laws that “fit” varying wind and dispersion profiles over the entire boundary layer. Instead of this, we use the centre of mass of the plume (zcm) to define the height at which the advective wind speed and eddy diffusivity (Kz) are defined and allow this to vary dependent on the plume travel distance, giving a similar effect to the plume dispersion as the non-Gaussian vertically integrated derivation. A similar methodology is employed by the OPS model (Sauter et al., 2018). We then use a combination of eddy diffusivity (Kz) vertical profiles, Lagrangian timescales and centre of mass plume placement, along with initial values σz0 and σy0, to determine σz and σy values. The aim of this combination is to provide realistic plume dispersion over short distances but to asymptotically approach the same Kz values used in the EMEP model dispersion scheme over longer distances. In addition, the methodology is implementable at all emission heights and takes into account both surface roughness and atmospheric boundary layer height.

Following the methodologies outlined in Seinfeld and Pandis (1998), we describe the dispersion parameters σz and σy as a function of pollutant travel time from source (t) using 15aσzt=σz0+2Kzztft15bσyt=σy0+2Kyztft, where σz0 and σy0 are initial dispersion parameters, Kz(z) and Ky(z) are the vertical profiles of vertical and horizontal eddy diffusivity, and ft is a factor dependent on the Lagrangian integral timescale τl given by 16ft=1+τltexp⁡-tτl-1. There are many varying methods for calculating the Lagrangian integral timescale (Seinfeld and Pandis, 1998; Hanna, 1981; Venkatram, 1984). We use the formulation from Hanna (1981): 17τl=0.6max⁡zemis,zτminu∗, where zemis is the emission height, u∗ is the friction velocity, and zτmin⁡=2 m. Time is calculated from the advective velocity: 18t=max⁡xmin⁡,xU(z), where xmin⁡ is half a subgrid, U(z) is the vertical wind speed profile, and x is the down-plume distance.

In order to be compatible with the EMEP model, the same Kz vertical profile parameterisation is used in Eq. (15a) that is used in EMEP (Simpson et al., 2012). This parameterisation is provided in the Supplement (Eqs. S1–S2).

The centre of mass of the plume is calculated using the same Gaussian formulation with reflection as given in Eq. (12) by integrating the plume intensity over the boundary layer height (H) using 19zcm=∫0HzI(z)dz∫0HI(z)dz. This integral can be analytically solved to give 20zcm=σz2π∑i=1i=6exp⁡-hi22σz2-exp⁡-H-hi22σz2+hi2erfhi2σz+erf(H-hi)2σz, and for the well-mixed case where 21σz>0.9H then zcm=0.5H. The vertical wind profile is calculated in a similar way to Gryning et al. (2007) based on decreasing turbulent shear with height. 22Uz=u∗0κlog⁡zz0-ψm+κzzl1-z2H-zH1+bz2LforL≥0Uz=u∗0κlog⁡zz0-ψm+κzzl1-z2H-zHaz-Lϕm+La(p+1)forL<0 The stability functions (ψm and ϕm) are defined in Eqs. (S3)–(S4) in the Supplement, and the assumptions behind the wind profile derivation are given in Eqs. (S5)–(S8). There is no turning of the wind direction with height. Equation (22) is used to derive u∗0, based on modelled 10 m wind speed, boundary layer height H, Monin–Obukhov length (L) and surface roughness length z0. The vertical wind profile is then derived from this. A minimum wind speed of 0.5 m s-1 for all dispersion calculations has been imposed.

The average of the plume centre of mass height at the receptor point and the emission height, zav=0.5(zcm+hemis), is then used to determine the vertical diffusion Kz(zav) as well as the wind speed U(zav) for use in Eqs. (15) and (19). The entire set of equations (Eqs. 15–22) is solved iteratively to obtain the final σz value at the receptor point. This iteration converges swiftly and generally only two iterations are required.

The horizontal eddy diffusivity Ky is not determined in EMEP so an alternative is required. Ky can be classically related to Kz through the relationship 23Ky(z)=σv(z)2σw(z)2Kz(z), based on the concepts used to define K (Seinfeld and Pandis, 1998). Garratt (1994) provides expressions for the vertical profile σv and σw under unstable conditions where the ratio σv/σw is around 1.85 in the surface layer but decreases to 1 in the mixed layer. Under stable conditions, Nieuwstadt (1984) provides local scaling where this ratio is close to 2. For the current application, we choose the ratio σv/σw=2 and apply it over the whole boundary layer.

It is also possible within the modelling setup to use the simpler empirical formulations of σz and σy, as presented in Eq. (24) and shown in Table 1. This is useful for testing and comparison, and necessary when using the rotationally symmetric plume parameterisation, Sect. 3.2. See Seinfeld and Pandis (1998) for a presentation of these.

In Fig. 3, we show two example sets of σz curves for near-surface (1 m) and elevated (50 m) releases as calculated with the Kz methodology for three separate stabilities. For reference, the dispersion curves from ASME (American Society of Mechanical Engineers; Smith, 1973) are also shown. These often-used dispersion parameters are relevant for 1 h averaging times. The ASME σz curves are given in Pasquill stability classes and the conversion from their dependency on L and z0 is achieved using the conversion methodology described by Golder (1972). Parameters used in the calculation of the three curves are provided in Table 1.

In addition to the parameterisations presented here, uEMEP also includes parameterisations, provided in the Supplement, for plume meandering and change of wind direction (Sect. S3.4.1), traffic-induced initial dispersion (Sect. S3.4.3) and road tunnel internal deposition and emissions (Sect. S3.4.5).

When applying uEMEP to annual mean emissions, a rotationally symmetric Gaussian plume is used. It is possible to derive an approximate analytical solution to the Gaussian plume equation assuming that wind directions are homogeneously distributed in all directions and that there is no strong dependence of wind speed or stability on wind direction. These conditions are usually not met but it is useful to have such a simplified analytical solution.

The starting point is the Gaussian plume model given in Eq. (12). In this case, we do not derive σy,z using the Kzvalue from EMEP but apply the commonly used power law formulation in order to derive an analytical solution: 24σ(y,z)=σ0(y,z)+a(y,z)xb(y,z). Values for the dispersion parameters a and b may be taken from the literature (Seinfeld and Pandis, 1998) but we use the ASME curves (Smith, 1973) under neutral conditions to specify these.

The rotationally symmetric version of this equation can be derived by rewriting the equation in cylindrical coordinates with appropriate approximations (second order), based on the slender plume assumption, and integrating over all angles. The resulting rotationally symmetric intensity Irot(r,z) as a function of r and z is then written 25Irot(rz)=1π2πrεz1+Berfπ1+B2εθ∑i=1i=6exp⁡-z-hi22εz2, where 26aεz=σ0z+azrbz26bεθ=1rσ0y+ayrby26cB=-εθ2bzεz-σ0zrεθ+byrεθ-σ0yεz. The term B can be less than -1, typically when r<2σ0,y, which can lead to imaginary solutions. This is due to the second-order approximation made in converting to cylindrical coordinates. In that case, we write a second-order approximation based on Taylor series expansion around B=-1 as 27Irotr,z=12πrεzεθ1-π2(1+B)6εθ2+π4(1+B)240εθ4∑i=1i=6exp⁡-z-hi22εz2 forB<-1. A similar derivation has been carried out by Green (1980) using different assumptions for the form of Eq. (24).

In Sect. 3.1 and 3.2, the hourly and annual dispersion parameterisations are described. In both cases, initial values for σ0(y,z) are required. Since we treat the sources as small area emitters, we set the initial σ0y to correspond to these areas. A value of σ0y=Δy/2π≈0.8Δy/2 will give a maximum subgrid centre concentration equivalent to the concentration that would be found if the emissions were distributed evenly in the subgrid. We then write the total initial dispersion to be 28σ0y=σinit,y+0.8Δy2. In all applications of uEMEP, Δx=Δy. The other parameter, σinit,y, is a specific initial dispersion width for each individual emission source, for example, 2 m for traffic and 5 m for shipping, heating and industry. This is generally much smaller than the emissions grids.

The initial value for σ0z is also a combination of a specific emission initial dispersion, for example, σinit,z=5 m for residential wood combustion but also uses the displacement technique for the plume where the start of the plume is displaced upwind by Δx/2 allowing the plume to grow vertically over half the subgrid distance. Tunnel exits are given an initial σinit,z=6 m to represent the extended size of the tunnel portals.

The only chemistry included in uEMEP is the NOx, O3 chemical reactions. We use a similar methodology to Benson (1984, 1992) known as the discrete parcel method but use a weighted timescale over which the reactions take place. The following chemical reactions are involved, with Ox (O3+NO2) and NOx (NO+NO2) concentrations being conserved: 29aNO+O3→NO2+O229bNO2+hν→NO+O29cO2+O+M→O3+M. Equation (29c) occurs on timescales much faster than the two other reactions and is taken to be instantaneous. The differential equation for the concentration of [NO2] as a function of time is written as 30dNO2dt=k1NOO3-JNO2, where the concentrations are expressed in terms of molecules per cm3 and J is the photolysis rate (s-1) for Eq. (29b) taken from the EMEP model (Simpson et al., 2012). The reaction rate k1 for Eq. (29a) is given by 31k1=1.4×10-12exp⁡-1310Tair(cm3s-1), as in the EMEP model and where Tair is in the atmospheric temperature (K).

We rewrite Eq. (30) in terms of the dimensionless ratios: 32afNO2=NO2NOxandfOx=OxNOx32bJ′=Jk1NOx32ct′=tk1NOx. Equation (30) then becomes 33dfNO2dt′=1-fNO2fOx-fNO2-J′fNO2. The solution to this equation is 34fNO2=B21-Aexp⁡Bt′1+Aexp⁡Bt′+C2, where 35aA=B+C-2fNO2,0B-C+2fNO2,035bB=C2-4fOx35cC=1+fOx+J′, and fNO2,0 is the initial NO2 fraction at time t′=0.

This solution is valid for a box model without dilution through dispersion since it does not take into account how changing NOx and Ox concentrations over the plume travel time will affect the reaction rates. Though this could be accounted for when applied to a single source with assumed dilution rates, by adding a time-dependent diluting term to Eq. (30), this is not practically possible for multiple sources of differing dilutions. The concentrations of NO2 at the start of the plume will be correctly calculated but NO2 concentrations further from the plume will be slightly underestimated, since they do not have the higher initial reaction rates. Eventually, the concentrations will reach photostationary equilibrium and here too NO2 will be correctly calculated. This special case for photostationary equilibrium in Eq. (35) occurs when t′→∞ and Eq. (34) becomes 36fNO2=C-B2. The non-linear nature of Eq. (34) also means that it cannot be consistently applied to Gaussian models since the shape of the plume will change due to the non-linearity. Despite this, this formulation is more physically realistic than the photostationary assumption often used in local-scale air quality modelling or other less physical parameterisations based on empirical fits. See Denby (2011) for an overview of the various NO2 chemistry parameterisation methods used with Gaussian modelling. In order to calculate Eq. (34) in the model application, an initial NOx and Ox concentration must be used and a travel time defined. For multiple sources, this travel time will vary so for each calculated subgrid concentration of NOx from each contributing subgrid source (ns sources) a travel time, ts, is calculated based on the distance and wind speed. This is weighted based on the contribution of each source to the total subgrid NOx concentration. This provides a final weighted travel time tw that is applied in Eq. (34). This ensures that the nearest of the contributing subgrids, often with the highest contributing NOx concentrations, are given a higher travel time weight. A minimum distance, and hence time, of half a subgrid is applied when calculating travel times. 37tw=∑s=1nstsNOxs∑s=1nsNOxs Comparisons with EMEP NO2 calculations show that this chemistry scheme matches the results obtained by EMEP over longer time periods.

When annual mean data are used, the hourly mean formulation cannot be applied. Instead, we use an empirically based conversion of NOx to NO2 based on the type of formulation from Romberg (1996) and updated by Bächlin and Bösinger (2008). A total of 3 years of Norwegian NO2 measurements, 82 measurements in all, have been used to determine this relationship (Fig. 4). 38NO2=aNOxNOx+b+cNOx The fitted constants are determined to be a=20, b=30, and c=0.23. The estimated uncertainty in this conversion is around 10 %, based on the normalised root mean square error of the fitted and observed NO2 concentrations.

This empirical relationship will vary from region to region, largely due to differences in O3 concentrations and photolysis rates that are not included as part of the parameterisation. If used over large regions, for example, Europe, then the uncertainty in the NO2 conversion will increase.

We describe below the implementation steps used in the Norwegian forecasting and analysis system. This implementation of uEMEP uses the independent emission and replacement downscaling method (method 2 in Sect. 2). The following steps are carried out: 1.

High-resolution emission data for Norway are calculated for each forecast (Sect. 4.2) and are aggregated into the EMEP MSC-W Scandinavian model grid. Some of these emissions require meteorological data.

The EMEP calculations make use of the CAMS-REG-AP_v1.1 regional anthropogenic emission dataset everywhere in Europe (Kuenen et al., 2014; Granier et al., 2019). Only in the 2.5 km Scandinavian calculation, and only in Norway, are the emissions replaced with the aggregated high-resolution dataset. The alternative emissions used in the calculations for Norway are

road traffic exhaust emissions,

The meteorological forecast data used for the European EMEP model calculations are based on the Integrated Forecasting System (IFS, 2020) from the European Centre for Medium-Range Weather Forecasts (ECMWF, 2020). The Scandinavian EMEP model calculation uses the AROME-MetCoOp model for modelling meteorology over Scandinavia (Müller et al., 2017). This last model calculates meteorology at a resolution of 2.5 km and provides forecasts for 66 h in advance. The EMEP MSC-W Scandinavian domain uses the same gridding and projection as the meteorological forecast model but in a smaller domain.

The European EMEP MSC-W model calculation is based on the same daily forecast provided for the Copernicus Atmosphere Monitoring Service (CAMS, 2020; Tarrason et al., 2018) but is run independently and provides boundary conditions for the Scandinavian implementation of EMEP MSC-W at 2.5 km. The Scandinavian EMEP MSC-W calculation includes the Norwegian emission sources described in Sect. 4.2 and also delivers the necessary local fraction information for use in uEMEP.

uEMEP calculates concentrations for all of Norway on grids with resolutions between 50 and 250 m using 1864 individual tiles as described in Sect. 4.1. The resolution of these tiles is defined by the population density and road density information. Tiles with higher population density use 50 m resolution, whilst tiles with lower population density but some traffic have a resolution of 125 m. Tiles with very low traffic density but with shipping or wood burning emissions have a resolution of 250 m, corresponding to the emission resolution. Separate calculations are carried out at measurement sites, 72, with a subgrid resolution of 25 m. An example of a PM10 forecast is shown in Fig. 7.

Here, we present a visual summary of results for NO2, PM10 and PM2.5 for the year 2017, when there were 72 operational air quality stations. Not all stations measure all components so the total number of available stations for NO2 and PM with coverage of more than 75 % is between 34 and 45. The station positions are shown in Fig. 8.

Figure 9 shows the comparison of modelled and observed NO2 for annual average at each station (scatter plot) and daily mean temporal profile averaged over all stations. Included in the scatter plot are the Scandinavian EMEP MSC-W results at 2.5 km. The spatial correlation is quite high, r2=0.81 for uEMEP with little negative bias (FB of -5.9 %). The temporal variation over the whole year is also well represented when averaged over all stations (r2=0.79).

Figure 10 shows the comparison of modelled and observed PM10 for annual average at each station (scatter plot) and daily mean temporal profile averaged over all stations. Included in the scatter plot are the Scandinavian EMEP results at 2.5 km. The spatial correlation is low, r2=0.29 for uEMEP with little negative bias (FB of -9.2 %). The temporal variation over the whole year is well represented when averaged over all stations (r2=0.61) but the model has a negative bias of 4 µg m3 over much of the summer period. Road dust events in the springtime are well captured by the NORTRIP emission model.

Figure 11 shows the comparison of modelled and observed PM2.5 for annual average at each station (scatter plot) and daily mean temporal profile averaged over all stations. Included in the scatter plot are the Scandinavian EMEP results at 2.5 km. The spatial correlation is good, r2=0.49 for uEMEP with little negative bias (FB =-10.5 %). The temporal variation over the whole year is well represented when averaged over all stations (r2=0.67) but the model has a negative bias of 2 µg m3 over much of the summer period. Residential wood combustion (heating) events in the winter are well captured by the emission model MetVed.

In addition to the validation against monitoring data, a number of verification and sensitivity experiments have been undertaken with the model. These include

comparison of single annual mean calculations with the mean of hourly calculations,