Modelling pollutants sub-grid scale variability due tosurface emission heterogeneity at urban areas

Myrto Valari and Laurent Menut

Institut Pierre-Simon Laplace, Laboratoire de Meteorologie Dynamique,Ecole Polytechnique, 91128 Palaiseau, France.

Abstract.We propose a sub-grid scale model, that takes into account surface emission heterogeneityand models pollutants variability at scales smaller than model resolution. Grid-averageemission flux is split into a sub-grid mosaic of different types of emitting surfaces (traffic,residential, etc...). The model is implemented into a mesoscale chemistry transport modeland results for a 3-months simulation over Paris (June 1st to August 31) are compared withsurface measurements at traffic and background stations. The model is able to differentiatecorrectly pollutants concentrations over sub-grid scale micro-environments (close to streetsor over residential areas) improving correlation with measurements up to 25% for [PM10]and up to 72% for [NO2]. Decrease in model bias reaches 30% for [NO] and 11% for [O3]compared to grid-average concentrations. Different sensitivity of ozone production to itsprecursors emission rates is observed between sub-grid simulations. Modelled [O3] sub-gridscale variability is found to vary by a factor of 3 depending on the photochemical regime,with the NO − x-sensitive photochemistry favouring variability creation compared to theV OC-sensitive regime.

1. Introduction

Recent studies confirm the evidence of atmosphericpollution adverse health effects (WHO [2004]; Bellet al. [2005]). Health risk evaluation in health impactassessment studies (HIA) may whether consider ho-mogeneous exposure to air pollution, regardless pol-lutants spatial variability (Blanchard et al. [2008];Medina [2006]), or keep record of individuals ex-posure during their daily activities (Penard-Morandet al. [2005]). Each one of these approaches uses dif-ferent types of air quality data. The former, mostcommonly, uses measured concentrations at sites sell-ected to represent background pollution (backgroundsites), while the latter, uses high resolution simula-tions (few tenths of meters) of emissions dispersionwithin city-blocks (street canyons), where exposureactually takes place (Vardoulakis et al. [2003, 2005]).The evaluation of an average citizens exposure hasthe evident inconvenience that in reality, variabilityin emission sources spatial distribution leads to sharpvariations in pollutants concentrations inside cities(Menut [2003]). Disepersion models, on the otherhand, are limited to simulations around localizedemission sources; modelled concentrations at urbancenter scale are very sensitive to boundary conditionsand models domain need to be large enough aroundall the included sources (Menut [2003]; Brucher et al.

[2000]).Mesoscale chemistry transport models (CTM), atleast to the knowledge of the authors, have not beenused in (HIA) evaluation. Even though they havebeen proved to provide reliable concentrations (vanLoon et al. [2007]; Vautard et al. [2007]) over domainsof tenths of kilometres around urban centers, theirrelatively low resolution (a few kilometres) preventsthem from being directly associated with human ex-posure. In the present study, we propose a methodthat takes into account emission heterogeneity atscale smaller than CTM resolution (sub-grid scale)and attches the associated variability on mesoscaleCTM grid-average concentrations. This additionalinformation, is of significant importance when humanexposure comes into question.Within a common CTM, emissions are dilluted in-stantly in the entire grid-cell, an area of some squaresof kilometres. At the moment of their release, allinformation on emissions spatial distribution at sub-grid scale is lost. Surface emission heterogeneity how-ever, can affect modelled concentrations throughoutthe atmospheric boundary layer (Galmarini et al.[2007]; Auger and Legras [2007]). In reality, fastchemical reactions occur close to sources and emit-ted species are consumed locally rather than be-ing directly dispersed in the entire grid-cell volume.By modelling pollutants concentrations at sub-grid

1

2 Myrto Valari and Laurent Menut

scales we obtain a description of the concentrationfield close to emission sources.The principal advantage of the proposed sub-gridmodel is the simplicity of its implementation andits low computational cost. Sub-grid variability inthe context of this study represents the statisticalprobability distribution of pollutants concentrationover a sub-grid mosaic of different types of emittingsurfaces. In this sense, the proposed model com-bines the deterministic calculation of grid-average’standard’ CTM concentration with a statistical de-scription of emission sources spatial distribition overthe grid cell area. This hybrid approach betweendeterministic modelling and statistical downscaling,is the main difference with recent studies address-ing the same issue of transferring emission sub-gridvariability to mesoscale model output concentrations(Galmarini et al. [2007]), where concentration vari-ance equation is added and explicitly solved by theReynolds-Average Navier-Stokes model.The case study and model setup are presented insection 1. The principal assumptions and method-ological aspects of the implementation are discussedin section 2. Section 3 is an overall presentationof model validation based on the comparison of 3-months simulation results with surface measurements.In section 4 we zoom in measurement sites repre-senting different types of air pollution (traffic andbackground) and we compare the temporal evolutionof measured concentrations with sub-grid concentra-tions. In section 5 we focus on the sub-grid modelresponse with respect to the non-linear aspects ofNOx-V OC-O3 chemistry. The principal foundingsof the study are synthesized in section 4.

2. Case-study and model setup

The case-study is based on a 3-months (June 1st toAugust 31st 2006) simulation of gas and particulatematter chemistry and transport over Paris (France),with CHIMERE (www.lmd.polytechnique.fr/chimere)model. Model domain (160×130km2) is centeredaround Paris city, with a 3km horizontal resolutionand 10 vertical layers (995-500hPa).Input meteorological data (off-line meteorology) arecalculated with the MM5 model (Grell et al. [1994])at a 5km horizontal resolution with 61 vertical lay-ers (995-200hPa). Model configuration considersa simple ice scheme (Dudhia [1989b]), the plane-tary boundary layer (PBL) Medium Range ForecastModel (MRF) scheme Hong and Pan [1996] and the

Land surface model 5-layer thermal scheme (Dudhia[1989a]). A meteorological diagnostic processing isapplied within CHIMERE model providing the finalturbulence-related parameters (Schmidt et al. [2001];Vautard et al. [2001]) reducing CTM vertical resolu-tion to the aforementioned 10 vertical layers.Initial and boundary conditions are taken from acoarser CHIMERE domain simulation over France(15km horizontal resolution) following a one-waynesting method. Chemical boundary conditions forthe coarse resolution simulation are driven by GO-CART model monthly climatologies (Ginoux et al.[2001]) for aerosol species and by the LMDz-INCAglobal chemical weather forecast system for gas-phasespecies (Van Leer [1979]; Tiedtke [1989]; Hourdin andArmengaud [1999]; Hauglustaine et al. [2004]).Emission input for Paris area are taken from the 1kmresolution inventory issued by AIRPARIF (http://www.airparif.asso.fr)under the ESMERALDA project (http://www.esmeralda-web.fr) providing hourly mean emission fluxes for gasand particulate matter species considered in modelchemical mechanism.At European level, sources of pollutants are definedby considering the type of emission activity, fuels andother criteria such as type of process, abatement tech-niques, operating conditions etc... (http://www.citepa.org/emissions/methodologie).AIRPARIF anthropogenic emission inventory pro-vides mean annual emission flux rates for pollutants 1

released from eleven different source types (table 1).Source categories are characterized by different tem-poral profiles and chemical compositions. This a pos-teriori information is applied on raw emission data toprovide refined emission input for CTM simulations.Total NMV OC mass, for example, is split into 375volatile organic compounds according to the chemicalcomposition of the corresponding source type. At asecond level, different NMV OCs are aggregated to alimited number of model species2 taking into accounttheir reactivity respectively to model chemistry.A common CTM procedure, considers that emis-sion is homogeneous over the grid-cell surface andwipes out all trace of the differentiation betweensource types at the moment of the aggregation ofraw data to model grid-average emission flux rates.

1NOx,V OC,CO2,CO,SO2,NH3,CH4,black and organiccarbon,particulate matter classified to three size bins

2α-pinene (C10H6), ethylene (C2H4), ethane (C2H6),propylene (C3H6), α-butane (C4H10), isoprene (C5H8),formaldehyde (C2HO), acetaldehyde (C2H4O), methyl ethylketone (C4H8O), o-xylene (C8H10)

Modelling pollutants sub-grid scale variability due to surface emission heterogeneity at urban areas 3

# Source type1 Combustion in energy and transformation industries2 Non-industrial combustion plants3 Combustion in manufacturing industry4 Production processes5 Extraction and distribution of fossil fuels and geothermal energy6 and other product use7 Road transport8 Other mobile sources and machinerys9 Waste treatment and disposal10 Agriculture11 Other sources and sinks

Table 1. Definition of the different emission sources,relating emission type and fuels, following the SNAP (Selected Nomenclature for Air Pollution) referencenomenclature (http://www.citepa.org/emissions/methodologie).

When different types of emitting sources coexist inthe same model grid-cell, emission sub-grid variabil-ity acts as a source of variability around modelledgrid-average concentration. Traffic transport emis-sion, for example, is characterized by low V OC

NOxra-

tios (mean daily V OCNOx Traffic

≈ 0.3), whereas moreV OC than NOx is released over residential build-ings (mean daily V OC

NOx Residential≈ 5). Note that the

difference in these ratios is much more due to vari-ation in NOx than V OC emission; similar V OCmass is emitted in both traffic or residential micro-environments( NOx Traffic

NOx Residential≈ 10, V OCResidential

V OCTraffic≈ 1).

Variability in its precursors emission ratios may in-duse O3 formation under different local photochemi-cal regimes.

3. Methodology

3.1. Preparation of the sub-grid scaleemission input

For the present study we preserved the classificationof emitted pollutants over different source types. In-stead of a single, mean emission flux rate over the en-tire grid-cell we considered a sub-grid mosaic of fourtypes of emitting surfaces: road traffic emission, res-idential/tertiary emission, emission due to outdoorsleisure activity (e.g. gardening, river navigation etc...), all other source. Each sub-grid emission repre-sents the contribution of a certain source type to thestandard grid-average input flux. For instance, mapsof figure 1 show NO emission flux due only to traf-

fic (left) or only to polluting activities in residentialhouseholds (right).

3.2. Implementation of the sub-grid scalecalculation

In a common CTM calculation, the variation of pollu-tants grid-average concentration within a model timestep interval integrates production and loss terms re-lated to modelled physical and chemical processesunder consideration (i.e. primary emission, chemi-cal production or consumption, horizontal transport,vertical mixing, scavering by rain, dry deposition onthe underlying surface). All model variables are gridand time averaged. In the proposed sub-grid model,sub-grid emission flux Ei is released over the sub-gridarea Ai considering that the corresponding sourcelies over the Ai

A area fraction (E =∑n

i=1 Ei,A =∑ni=1 Ai). In this sense, we assume for instance that

traffic emission Etraffic is released over the Atraffic

Asub-grid fraction of the grid-cell area which is coveredwith roads (figure 2).A schematical representation of sub-grid model im-plementation is shown in figure 3. Grid-cell area is di-vided in the following 4 sub-grid surfaces, each corre-sponding to a single emission source type: i)roads fortraffic, ii)buildings for residential emission, iii)parks,gardens, rivers etc... for outdoor leisure activityrelated emission, iv)the rest of grid-cell area is at-tributed to the rest of emission sources (left part offigure 3). Note that the division in three emissiontypes, illustrated in figure 3, is only an indicative ex-

4 Myrto Valari and Laurent Menut

Figure 1. NO emission fluxes over Paris due to week day morning traffic (left) and residential activities(right) at 1×1km2 horizontal resolution on 02/06/2006 at 07:00. Red triangles show the location of mea-surement stations of the local air quality network (AIRPARIF) assigned to measure atmospheric pollutantsconcentrations close to traffic sources and black circles give the location of background stations.

Figure 2. The fraction of model grid-cells occupied by streets and avenues in Paris city at 1×1km2 resolution.

ample and that in fact, an unconditional number ofsub-grid emissions can be defined.Independent simulations over each sub-grid space cal-culate pollutants concentration variation during atime-step interval. Sub-grid model provides differentconcentration values (sub-grid concentrations) rep-resenting the deviation range around the ’standard’grid-average concentration due to the proximity tolocal emission sources (middle part of figure 3). Inthe following we will use the terms traffic and resi-dential sub-grid concentrations. At the end of each

time-step, sub-grid concentrations define the concen-tration sub-grid scale variability and they are aver-aged to the ’standard’ grid-average concentration forthe continuation of the simulation (right part of fig-ure 3).It should be noted that we have made the assumptionthat air parcels are neither transported nor mixed inbetween sub-grid areas and only emission is consid-ered at sub-grid scale. Mass transport inwards andoutwards model grid-cells at each time-step alwaysconsiders grid-average terms for both meteorological

Modelling pollutants sub-grid scale variability due to surface emission heterogeneity at urban areas 5

variables (wind speed for transport, eddy coefficientfor mixing etc...) and species concentrations.In this context, sub-grid space should be seen as theprobability that during a model time-step intervalan air parcel is found over a certain type of emittingsurface, rather than it has been physically advectedthere by the wind. The absence of sub-grid trans-port ensures that each sub-grid simulation remainscompletely unaffected by any other emission source.Sub-grid emission flux rates can not enter directlyin the calculation; less mass is emitted over sub-gridspace since sub-grid emission Ei is always smallerthan grid-average emission (Ei < E). This artificialmass discontinuity would lead to erroneous advectionand mixing of air parcels at grid-average scale. Forthis reason sub-grid emission flux is normalized bythe area fraction so that sub-grid emission no longerrepresents the part of the total emission due to a cer-tain source but rather what the total emission wouldbe if the whole grid-cell surface was occupied by thecorresponding surface type.

4. Overall comparison of model resultswith surface measurements

Sub-grid model implementation is applied on thecentral part of the domain representing downtownParis (12 model grid-cells over 12×9km2), includingthe largest and more complex part of anthropogenicemission sources (figure 4).We use the two available types of measurements sites,namely, background and traffic stations, for modelvalidation (see also figure 4 for locations). We com-pare both ’standard’ model grid-average concentra-tion and traffic and residential sub-grid concentra-tions with measurements. The comparison includesNO2, PM10, NO and O3. It should be noted thatonly background measurements exist for ozone.Sub-grid concentrations other than traffic and resi-dential could not be compared with measurementssince no measurement site could be directly asso-ciated with the corresponding micro-environment.Given that traffic and residential related emissionrepresent the largest part of emission burden overthe studied area, we assume that pollutants variabil-ity can be well represented by the difference betweentraffic and residential sub-grid concentrations. Thefollowing analysis is hence, based on this considera-tion and ignores the rest of modelled sub-grid con-centrations.

4.1. Sub-grid model validation

Model results for a 3-months simulation period (June1st to August 31st 2006) are compared with mea-surements according to common validation criteria(Honore et al. [2008]). The interest of the comparisonis to test whether sub-grid model, by running inde-pendent simulations over different emission sources,is able to provide sub-grid concentrations closer tothe measurements at the corresponding site type,than grid-average concentration. In this way we canvalidate our considerations on emission heterogeneityprojection via land use fractions on model sub-gridspace.The surface type attributed to residential emissioncovers by far the largest part of the studied area andso we consider that residential sub-grid concentra-tion represents background pollution and it is com-pared with measured concentration at backgroundsites. Measured concentrations at traffic stations arecompared with the traffic sub-grid concentrations.For the calculation of model bias and root meansquare error (RMSE) of [NO2], [NO] and [O3] weuse the daily measured and observed maximal val-ues. For [PM10] the mean daily values are comparedinstead (table 2).

Criterion FormulaBIAS 1

N

∑Ni=1(Mi −Oi)

RMSE√

1N

∑Ni=1(Mi −Oi)2

CORR∑N

i=1(Oi−Omean)(Mi−Mmean)√∑N

i=1(Oi−Omean)2

√∑N

i=1(Mi−Mmean)2

Table 2. Mathematical formulas of the criteria ap-plied to compare model results with measurements:model error (BIAS), root mean square error (RMSE)and correlation coefficient, where ’M’ stands formodel value and ’O’ for observations.

Grid-average model (’standard’ simulation) overesti-mates [NO2], [PM10] and [NO] at all backgroundstations and underestimates them at all traffic sta-tions (figure 5 left). This model response is log-ical considering that these pollutants are directlyemitted at the traffic network and their concentra-tion decreases with the distance from source loca-tion. Smaller bias is observed with the residentialsub-grid concentration than with the grid-average atbackground stations for [NO2] and [PM10]. For[NO] model bias may increase or decrease with theresidential sub-grid model depending on the station.

6 Myrto Valari and Laurent Menut

Figure 3. The blue arrows-pathway (top) represents a ’Standard’ model procedure, where emission is in-stantly diluted to the volume of a grid-cell in the begining of model time-step (left). Production/loss termsaffect pollutants concentration during the time-step interval (middle) leading to a mean concentration value(right). The red arrows-pathway represents the implemented sub-grid model methodology (bottom), whereemission heterogeneity (left) is preserved inside the time-step interval (middle) leading to discrete sub-gridconcentrations and allowing the definition of sub-grid variability before averaged to the grid-average value.

Bias decrease in the sub-grid model is especially pro-nounced at traffic stations, for all sites and all pri-mary pollutants.Model bias for [O3] at background stations is alsosmaller when residential sub-grid concentrations arecompared with measurements instead of the grid-average values. Grid-average model underestimatesozone at all sites most probably because dilution ofhigh NO emissions at the entire grid-cell volumeresults in enhanced [NO]+[O3] reaction rate. Thefact that more O3 is modelled in the residential sub-grid simulation shows that differentiation in sourceswithin the sub-grid model counteracts dilution effect.Smaller RMS error is observed with the sub-grid traf-fic and residential model for all stations and pollu-tants (figure 5 middle). Note that even for [NO] atbackground stations RMS error is lower when res-

idential sub-grid concentrations are compared withmeasurements instead of the grid-average concentra-tions. Given that RMSE has the tendency to favourlarge descrepancies, we conclude that grid-averageconcentrations remain most of the time closer tomeasurements than sub-grid concentrations (lowerBIAS), but when they miss grid-average model de-screpancies are much larger than sub-grid model er-rors (larger RMS error).At background stations sub-grid residential [NO2] isless correlated with measurements than grid-averageconcentrations but a much better correlation is ob-served at traffic stations (figure 5 right). At bothtraffic and background stations sub-grid [PM10] ismore correlated with measured concentrations thangrid-average concentrations. Sub-grid model resultsfor NO show that sub-grid model is less correlated

Modelling pollutants sub-grid scale variability due to surface emission heterogeneity at urban areas 7

Figure 4. Central part of model domain, representing Paris city centre. Model grid at 3km horizontalresolution (points) and measurement stations at traffic (red triangles) and background (black circles) sites arealso shown.

with measurements at background sites, whereas attraffic sites correlation may improve or worsen de-pending on the station.O3 sub-grid concentrations are more correlated withmeasurements at all background stations but the im-provement is clearly less pronounced than for therest of the studied pollutants. This is probably dueto the fact that O3 is a secondary pollutant, whilethe other discussed species are directly emitted fromtraffic sources. Emission heterogeneity has logicallygreater influence on primary pollutants variability.Synthesizing those remarks we conclude that the sep-aration between traffic and residential micro-environmentswithin the sub-grid model provides a more realisticrepresentation of pollutants concentration. This con-clusion is more pronounced at areas at the proximityto traffic sources than for background pollution. Sub-grid modelled [NO] concentrations enhance overesti-mation of mesured concentrations at background sta-tions and they are less correlated with measurements.

5. Zoom in individual stations

’Standard’ model output concentration is a 9km2

grid-average value. Sub-grid concentrations give thedeviation range around this mean value due to emis-sion heterogeneity inside the cell. By selecting sur-face stations included in the same grid-cell, onecan compare ’real’ variability with modelled sub-gridvariability. Figure 6 compares the difference between

traffic and residential sub-grid concentrations withthe difference in measured [NO2] at the traffic (TR3)and background stations (BG5) (see figure 4 for loca-tions). Even though modelled variability has a ten-dency to underestimate the measured one, a goodcorrelation exists between the two values (R2 ≈ 0.7).Figure 7 compares the temporal evolution of sub-grid modelled [NO2] concentrations with measure-ments at the traffic and background sites (TR3 andBG5), from June 30th to July 30th. Traffic station(TR3) and background station (BG5) are both withinthe same model grid-cell; differences between trafficand residential sub-grid concentrations for [NO2] canreach 100µg/m3. Traffic sub-grid concentration fol-lows much closer the temporal evolution of nitrogendioxide at the traffic station, whereas the residentialsub-grid concentration is more competent to repro-duce [NO2] evolution at the background station.There are no [PM10] measurements at traffic andbackground stations both located in the same modelgrid-cell. Traffic and residential sub-grid concentra-tions are consequently compared with measurementsat traffic and background stations included in thecorresponding grid-cell. Sub-grid model response isstudied respectively to the grid-average concentra-tion.Traffic sub-grid concentration better represents mea-sured [PM10] temporal evolution at the traffic sta-tion (TR1) and residential sub-grid concentrationis closer to measurements at the residential station

8 Myrto Valari and Laurent Menut

BIAS RMSE R2

Figure 5. Comparison between ’Standard’ simulation (blue crosses) and sub-grid model (red crosses) at forNO2, PM10, NO and [O3], based on bias error (left), root mean square error (middle) and correlation withmeasurements (right) at background and traffic stations.

Figure 6. Hourly comparison between modelled and measured NO2 sub-grid variability from June 1st toAugust 31st, calculated for measurements at the background station ’Paris 1st’ (GB5) and the traffic stationat ’Bonaparte street’ (TR3).

(BG3) (figure 8). The time series of grid-averageconcentration is also shown in figure 8 and mod-elled sub-grid variability is expressed as the differ-ence between traffic and residential sub-grid concen-trations. In a significant number of cases, measuredmaximal or minimal concentration values are withinthe shadowed area representing the range of mod-elled variability. This remark suggests that attach-ing modelled sub-grid variability on the ’standard’grid-average model output concentration overcomesmodel discrepancies and gives a more realistic pic-ture of [PM10] concentration. Another remark is that

residential sub-grid concentration remains very closeto the ’standard’ model output concentration. Thisjustifies our decision to compare background mea-surements to residential sub-grid modelled concentra-tions and on the same time suggests that our initialconsiderations concerning emission split into differ-ent sources and their projection in sub-grid surfacesis realistic.In what concerns O3 concentration, only backgroundmeasurements were available. Figure 9 (top pan-nel) shows a week-simulation of O3 concentrationcompared to measurements at the background sta-

Modelling pollutants sub-grid scale variability due to surface emission heterogeneity at urban areas 9

Figure 7. Measured and modelled [NO2] at the traffic station on Bonaparte street (TR3) in Paris center(top panel) and at the background station ’Paris 1st’ (BG5) (bottom panel). Time series of mesured values(black crosses) and modelled traffic sub-grid concentration (red line) and residential sub-grid concentration(blue line), from June 30th to July 30th. Sub-grid variability is expressed as the difference between trafficand residential sub-grid concentrations (grey shadow)

tion BG5 in the center of Paris. ’Standard’ modelgrid-average simulation has the tendency to under-estimate [O3], especially during days of exceptionallyhigh photochemical activity (July 12th to 15th). Thisunderestimation is quite common at urban areas withhigh sources of nitrogen oxides. Even though NOemission is localized over high sources (e.g. streets)a ’standard’ simulation dilutes emission at the en-tire grid-cell volume and consequently overestimatesO3+NO reaction rate (see also previous section).The impact of this reaction becomes clear by the lowozone concentrations modelled with traffic sub-gridsimulation. On the contrary, more [O3] is produced inthe residential sub-grid simulation than in the ’stan-dard’ grid-average simulation.At days where ozone concentration remains low (≈80µg/m3), sub-grid model is able to predict cor-rectly the measured daily maximal values with theresidential sub-grid simulation (e.g. July 7th, July10th). This remark shows that separation of emis-sion sources in independent sub-grid scale simula-tions counteracts dilution effect. During days of very

high ozone concentration model underestimation islarge. Adding modelled sub-grid variability to thegrid-average [O3] concentration (error bars in fig-ure 9) shows that the difference between traffic andresidential sub-grid concentrations is of the same or-der of magnitute as model discrepancies. This sug-gests that dillution effect is at large extent responsi-ble for ozone model bias.Having shown that [NO] sub-grid concentrationsmay worsen grid-average model results, without af-fecting neither NO2 nor O3, it is interesting to study[NO] and [O3] concentrations coevolution.Traffic and residential sub-grid and grid-average [NO]concentrations are compared with measurements atthe same background station BG5 as [O3] (figure 9middle pannel). Modelled and measured concentra-tions temporal profiles show a clear anticorrelationbetween [NO] and [O3]. Traffic sub-grid simulationmodels the highest [NO] and lowest [O3]. Measuredbackground [NO] concentration is lower than in allsimulations while measured background [O3] concen-trations are the highest. If we compare modelled

10 Myrto Valari and Laurent Menut

Figure 8. Measured and modelled PM10 concentration from July 31st to August 10nth at the traffic station(TR1) at Victor Basch street (top panel) and at the background station (BG3) ’Paris 13th’ (bottom panel).Time series of mesured values (black crosses) and modelled traffic sub-grid concentration (red line), residentialsub-grid concentration (blue line), and ’standard’ grid-average model output concentration (green line). sub-grid variability is expressed as the the range of deviation from the grid-averaged concentration and it iscalculated by the difference between traffic and residential sub-grid concentrations (grey shadow).

and measured concentrations at the traffic stationTR3 (figure 9 bottom pannel), we may argue thatmodelled traffic and residential sub-grid concentra-tions define a realistic sub-grid variability. Neverthe-less correlation with measurements for [NO] is muchworse than for all other studied pollutants.An interesting remark is that contrary to [O3], [NO2]and [PM10], grid-average [NO] concentration de-taches from the residential sub-grid concentration.Both traffic and residential sub-grid [NO] concentra-tions are significantly higher than grid-average con-centration. This indicates strong chemical reactiv-ity within sub-grid simulations that has a non-lineareffect on grid-average modelled concentration. How-ever, this sign of non-linearity does not affect directlysub-grid modelled [O3] (residential and traffic sub-grid [O3] concentrations are always on oposite sidesof grid-average value). This suggests that the non-linear features of NOx-V OC-O3 photochemistry doesnot affect ozone formation locally (close to emissionsources), where ozone concentration is more driven

by the fast O3+NO reaction, but further downwindat longer temporal scales.The same effect is also present in the grid-average,’standard’ simulation; the large [NO] morning peakmodelled on July 10th (figure 9) does not have a di-rect impact on ozone production on the same after-noon (nevertheless, [NO] afernoon peak on the sameday depletes all ozone during night-time). The non-linear resopnse observed for modelled [NO] concen-trations leads to the question whether NOx-V OC-O3

chemistry within traffic and residential sub-grid sim-ulations is driven by different photochemical regimes.

6. Modelled ozone sub-grid scalevariability as a function of the chemicalregime

Up to this point we showed that sub-grid model isable to represent correctly small scale variability atthe proximity to emission sources, especially for pol-lutants directly emitted at the traffic network (PM10

Modelling pollutants sub-grid scale variability due to surface emission heterogeneity at urban areas 11

Figure 9. Measured and modelled [O3] and [NO] concentration from July 7th to July 15th at the backgroundstation ’PARIS01’ (BG5) (top and middle pannels) and [NO] concentrations at the traffic station ’Bonapartestr.’ (TR3) (bottom pannel). Time series of mesured (black crosses), modelled traffic sub-grid concentration(red line), residential sub-grid concentration (blue line), and ’standard’ grid-average model output concen-tration (green line). Sub-grid variability is expressed as the the range of deviation from the grid-averageconcentration and it is calculated by the difference between traffic and residential sub-grid concentrations(grey shadow). The error bar around the ’standard’ simulation equals the same difference on the hour of thesimulation daily maximal and minimal values.

and NO2). Non-linear effects appear when specieschemical reactivity becomes high. In this section wesetup a sensitivity experiment that allows a closerstudy of NOx-V OC-O3 photochemistry within thesub-grid model. The first step is to put in evidencethe non-linear dependence of ozone formation withrespect to NOx emssion. We identify two days, whereozone formation is driven by different photochemicalregimes. We study ozone sub-grid scale variabilitycreation as a function of ozone precursors emissionratios.We set up 36 simulations by varying NOx and V OCemission rates by steps of 0.25 around their value atthe refference simulation. Variation in emission wasapplied on the whole data set independent of sourcetype (e.g. the same coefficient is applied on NOx

emission released by traffic or residential source). By

varying the ratio of ozone precursors emission, we de-fine areas of different photochemical regimes (Sillman[1995, ]; Sillman and H.Dongyang [2002]; Kanayaet al. [2008]).On July 10th, ozone concentration at the referencesimulation reduces with decrease in V OC emissionand increases with NOx emission increase. Thisresponse defines a V OC-sensitive regime (Sillman[1995]). Four days later, on July 14th, modelledozone production is driven by a different regime,where decrease in NOx emission inhibits ozone pro-duction (NOx-sensitive regime). The remark thatozone production over Paris is often on the transi-tion zone between different regimes agrees with pre-vious studies over the same area (Menut et al. [2000];Sillman et al. [2003]; Deguillaume et al. [2008]).Comparing ozone production on each of the studied

12 Myrto Valari and Laurent Menut

days, we observe rapid change in ozone concentrationunder V OC-sensitive regime ([O3] increases from 67to 127 ppb on 10July 10th), whereas the same fluc-tuations in precursors emission rates leads to a muchlower sensitivity under a NOx sensitive regime ([O3]increases from 64 to 72ppb on 10July 14th).The response of O3 sub-grid scale variability to thesame perturbation precursors emission rates is shownin figure 10 (right). Different response is observed de-pending on the chemical regime. On July 10th, mostof the emission ratios map is dominated by VOC-sensitive chemistry. For low NOx emission, ozonevariability is relatively low but it increases rapidlyand almost linearly with NOx (NOx-limited regime).For higher NOx emission (above 50% of the refferenceemission rate) the regime becomes saturated in NOx

and ozone variability change rate slows down.On July 14th, NOx-sensitive regime dominates mostpart of precursors emission ratio plan. The fast linearincrease in ozone sub-grid variability remains presentat relatively high NOx emission rates (up to the refer-ence simulation), where NOx saturation is observed.Variability becomes much less sensitive to further in-crease in NOx emission.On both days, sub-grid ozone variability increasesby an equivelent factor of 3 under the same fluc-tuation in precursors emission ratios. Note that onJuly 10th (V OC-sensitive regime) ozone concentra-tion varied 10 times more than on July 14th (NOx-sensitive regime) over the same precursors ratio plan.Synthesizing those remarks we can argue that [O3]sub-grid variability increases linearly with increase inNOx emission and is practically insensitive in V OCvariations under a NOx-limited regime. Sensitivityin V OC emission is significantly less pronounced forozone sub-grid variability than for ozone concentra-tion itself under a V OC-sensitive regime.Ozone sub-grid variability dependence on the chem-ical regime is explained by the different V OC

NOxratios

characterizing emission sources (see also the section3.1). The main difference between traffic and res-idential emission is in NOx burden. Consequently,NOx sensitive regime favours ozone sub-grid variabil-ity creation. Even if modelled ozone concentration isless sensitive in precursors emission ratios under aNOx regime, modelled sub-grid variability is highlyunstable.

7. Conclusions

A sub-grid scale model was implemented in a mesoscalechemistry transport model accounting for the vari-ability of surface emission at scales smaller thanmodel resolution. This method allowed to attach pol-lutants variability to the mean output CTM concen-tration and better define pollution within an urbancontext. The proposed implementation is based onthe division of grid-cells in sub-grid surfaces repre-senting areas releasing different types of emission.Emission heterogeneity or turbulence-related sub-grid scale effects on CTM mean output concentra-tions have been previously identified and quanti-fied using explicit methods (e.g. large eddy simula-tions)Krol et al. [2000]; Galmarini et al. [2007]; Ebelet al. [2007]; Vinuesa and Port-Agel [2008]. The ad-vantage of the proposed model is that the effect ofsmaller scales on model grid-average output is cal-culated in a statistical manner without explicitly re-solving the finer scale. We are based on the con-sideration that a statistical distribution of emissionsources over model sub-grid space can generate con-centration variability at sub-grid scale. This hybridapproach, combining deterministic modelling withstatistical description of sub-grid scale heterogene-ity, makes the implementation of the model relativelysimple and at low computational cost.Model results for nitrogen oxides, PM10 and ozonewere compared with measurements at different typesof stations, showing that sub-grid model is able todifferentiate correctly pollutants concentrations atthe proximity to emission sources. Given that ur-ban emissions are principally released over the trafficnetwork or over residential buildings we focused thestudy on the variability induced by the spatial distri-bution of these two emission types. It was shown thatmodelled variability was well correlated to the mea-sured one, suggesting that our initial considerationsconcerning the projection of emission heterogeneitywithin model sub-grid scale were correct. Modelshowed, however, a tendency to underestimate mea-sured variability, indicating that other sources of vari-ability have been neglected.Modelled sub-grid concentrations had a non-linearimpact on grid-average concentration for higly re-active species as NO. We focused on NOx-V OC-O3 photochemistry within the sub-grid model andshowed that ozone sub-grid scale variability may varyby a factor of 3 depending on the photochemicalregime. Ozone concentration wa shown to vary more

Modelling pollutants sub-grid scale variability due to surface emission heterogeneity at urban areas 13

July 10th 2006

O3 (ppbv) O3 variability (ppbv)

July 14th 2006

O3 (ppbv) O3 variability (ppbv)

Figure 10. O3 photochemical production (left) and O3 sub-grid scale variability creation (right) at downtownParis as a function of V OC

NOxratio on July 10th (top) and July 14th (bottom). Red points define emission

coefficients used at each simulation for variation in anthropogenic NOx and V OC emission (Point (1,1)corresponds to the reference simulation.

rapidly with emission perturbations under the NOx-sensitive regime than under a V OC-sensitive one, butthe NOx-sensitive regime had a clear tendency tofavour sub-grid variability creation. These foundingssuggest that photochemical regime indicators may bealso useful for ozone small scale variabiality predic-tion.

References

Auger, L., and B. Legras (2007), Chemical segrega-tion by heterogeneous emissions, Atmospheric En-vironment, 41 (11), 2303–2318.

Bell, L., M., F. Domonici, and M. Samet (2005), Ameta-analysis of time-series studies of ozone andmortality with comparison to the national morbid-ity, mortality and air pollution study, Epidemiol-ogy, 16 (4), 436–445.

Blanchard, M., et al. (2008), Assessment of urbanair pollution im-pact on health concepts and methods., Tech. Rep.http://www.invs.sante.fr/publications/2006/psas/psas.pdf,Institut de veille sanitaire.

Brucher, W., C. Kessler, M. Kerschgens, and A. Ebel(2000), Simulation of traffic-induced air pollutionon regional to local scales, Atmospheric Environ-

14 Myrto Valari and Laurent Menut

ment, 34, 4675–4681.Deguillaume, L., M. Beekmann, and C. Derog-

nat (2008), Uncertainty evaluation of ozone pro-duction and its sensitivity to emission changesover the ile-defrance region during summer pe-riods, Journal of Geophysical Research, 113,doi:10.1029/2007JD009081.

Dudhia, J. (1989a), A multi-layer soil temperaturemodel for mm5, Preprints, The 6th PSU/NCARMesocale Model MM5 Users Workshop, Boulder,CO.

Dudhia, J. (1989b), Numerical study of convectionobserved during the winter monsoon experiment10 using a mesoscale two-dimensional model, At-mospheric Science, 46, 3077–3107.

Ebel, A., M. Memmesheimer, and J. Jacobs (2007),Chemical perturbations in the planetary bound-ary layer and their relevance for chemistry trans-port modelling, Boundary Layer Meteorology, 125,265–278.

Galmarini, S., J.-F. Vinuesa, and A. Martilli (2007),Modelling the impact of sub-grid scale emis-sion variability on upper-air concentration, Atmo-spheric Chemistry and Physics Discussions, 7 (4),12,289–12,326.

Ginoux, P., M. Chin, I. Tegen, J. M. Prospero,B. Holben, O. Dubovik, and S.-J. Lin (2001),Sources and distributions of dust aerosols simu-lated with the GOCART model, Journal of Geo-physical Research, 106, 20,255–20,273.

Grell, G. A., J. Dudhia, and D. R. Stauffer (1994), Adescription of the fifth-generation penn state/ncarmesoscale model (mm5), NCAR Tech. Note TN-398+STR, p. 122.

Hauglustaine, D. A., F. Hourdin, L. Jourdain,M.-A. Filiberti, S. Walters, J.-F. Lamarque,and E. A. Holland (2004), Interactive chemistryin the Laboratoire de Meteorologie Dynamiquegeneral circulation model: Description andbackground tropospheric chemistry evaluation,Journal of Geophysical Research, 109 (D04314),doi:10.1029/2003JD003,957.

Hong, S., and H. Pan (1996), Nonlocal boundarylayer vertical diffusion in a medium-range forecastmodel, Monthly Weather Review, 124, 2322–2339.

Honore, C., et al. (2008), Predictability of europeanair quality: The assessment of three years of op-erational forecasts and analyses by the prev’airsystem, Journal of Geophysical Research, Atmo-spheres, in press.

Hourdin, F., and A. Armengaud (1999), On the use offinite volume methods for atmospheric advectionof trace species: I. test of various formulations ina general circulation model, Monthly Weather Re-view, 127, 822–837.

Kanaya, Y., M. Fukuda, H. Akimoto, N. Takegawa,Y. Komakazi, Y. Yokouchi, M. Koike, andY. Kondo (2008), Urban photochemistry in cen-tral tokyo:2.rates and regimes of oxidant o3+no2production, Journal of Geophysical Research, 113,doi:10.1029/2007JD008671.

Krol, M. C., M. J. Molemaker, and J. V. Guerau deArellano (2000), Effects of turbulence and het-erogeneous emissions on photochemically activespecies in the convective boundary layer, Journalof Geophysical Research, 105 (D5), 6871–6884.

Medina, S. (2006), Full report on health impactassessment of outdoor air pollution., Tech. Rep.http://www.apheis.net/FinalReportNew/WP5OAPreport.pdf,Department of Environmental Health.

Menut, L. (2003), Adjoint modelling for atmosphericpollution processes sensitivity at regional scaleduring the ESQUIF IOP2, Journal of GeophysicalResearch - Atmospheres, ESQUIF Special Section,108 (D17), 8562, doi:10.1029/2002JD002549.

Menut, L., R. Vautard, M. Beekmann, and C. Honore(2000), Sensitivity of photochemical pollutionusing the adjoint of a simplified chemistry-transport model, Journal of Geophysical Research,105 (D12), 15,379–15,402.

Milford, J. B., D. Gao, S. Sillman, P. Blossey, andA. Russel (1994), Total reactive nitrogen (noy) asan indicator of the sensitivity of ozone to reduc-tions in hydrocarbon and nox emissions, Journalof Geophysical Research, 99 (D2), 3533–3542.

Penard-Morand, C., C. Schillinger, A. Armengaud,G. Debotte, E. Chretien, S. Pellier, I. Annesi-Maesano, and ISAAC-France (2005), Assessmentof schoolchildren’s exposure to traffic-related airpollution in the french six cities study using a dis-persion model., AE, 40 (13), 2274–2287.

Schmidt, H., C. Derognat, R. Vautard, and M. Beek-mann (2001), A comparison of simulated and ob-served ozone mixing ratios for the summer of 1998in western europe, Atm. Env., 35, 6277–6297.

Sillman, S. (), The relation between nox and hydro-carbons in urban and polluted rural environments,Atmospheric Environment, 33, 1821–1845.

Sillman, S. (1995), The use of NOy, H2O2 and HNO3

as indicators for ozone-NOx-hydrocarbon sensitiv-

Modelling pollutants sub-grid scale variability due to surface emission heterogeneity at urban areas 15

ity in urban locations, Journal of Geophysical Re-search, 100, 14,175–14,188.

Sillman, S., and H.Dongyang (2002), Some theoret-ical results concerning o3-nox-voc chemistry andnox-voc indicators, J. of Geophys. Res., 107 (D22).

Sillman, S., R.Vautard, L.Menut, and D.Kley (2003),O3-nox-voc sensitivity indicators in paris: resultsfrom models and esquif measurements, Journal ofGeophysical Research, 108 (D17).

Tiedtke, M. (1989), A comprehensive mass fluxscheme for cumulus parameterization in large-scale models, Monthly Weather Review, 117, 1779–1800.

Van Leer, B. (1979), Towards the ultimate conserva-tive difference scheme. V A second order sequel toGodunov’s method, J. Computational Phys., 32,101–136.

van Loon, R., M .and Vautard, et al. (2007), Eval-uation of long-term ozone simulations from sevenregional air quality models and their ensemble, At-mospheric Environment, 41, 2083–2097.

Vardoulakis, S., B. E. A. Fisher, K. Pericleous, andN. Gonzalez-Flesca (2003), Modelling air qualityin street canyons: a review, Atmospheric Environ-ment, 37 (2), 155–182.

Vardoulakis, S., N. Gonzalez-Flesca, B. Fisher, andK. Pericleous (2005), Spatial variability of air pol-lution in the vicinity of a permanent monitor-ing station in central paris, Atmospheric Environ-ment, 39 (15), 2725–2736.

Vautard, R., M. Beekmann, J. Roux, andD. Gombert (2001), Validation of a hybrid fore-casting system for the ozone concentrations overthe Paris area, Atmospheric Environment, 35,2449–2461.

Vautard, R., et al. (2007), Evaluation and intercom-parison of ozone and pm10 simulations by severalchemistry transport models over four europeancities within the citydelta project, AtmosphericEnvironment, 41, 173–188.

Vinuesa, J., and F. Port-Agel (2008), Dynamic mod-els for the subgrid-scale mixing of reactants inatmospheric turbulent reacting flows., J. Atmos.Sci., 65, 1692–1699.

WHO (2004), Health aspects of air pollution resultsfrom the who project ”systematic review of healthaspects of air pollution in europe”, CopenhagenDenmark,http://www.euro.who.int/document/E83080.pdf.

Myrto Valari,Laboratoire de Meteorologie Dy-namique, Ecole Polytechnique, Palaiseau, France.[myrto.valari@lmd.polytechnique.fr]

This preprint was prepared with AGU’s LATEX macros

v5.01, with the extension package ‘AGU++’ by P. W. Daly,

version 1.6 from 1999/02/24.