Atmospheric Environment 42 (2008) 4870–4883 Performance characteristics of MM5–SMOKE–CMAQ for a summer photochemical episode in southeast England, United Kingdom Y. Yu a,b, , R.S. Sokhi a , N. Kitwiroon a , D.R. Middleton c , B. Fisher d a Centre for Atmospheric and Instrumentation Research (CAIR), University of Hertfordshire, Hatfield, Hertfordshire AL10 9AB, UK b Laboratory for Climate, Environment and Disasters of Western China, Cold and Arid Regions Environmental and Engineering Research Institute, Chinese Academy of Sciences, Lanzhou, Gansu 730000, PR China c Met Office, FitzRoy Road, Exeter EX1 3PB, UK d Environment Agency, Reading RG1 8DQ, UK Received 15 July 2007; received in revised form 28 November 2007; accepted 19 February 2008 Abstract In this study a modelling system consisting of Mesoscale Model (MM5), Sparse Matrix Operator Kernel Emissions (SMOKE) and Community Multiscale Air Quality (CMAQ) model has been applied to a summer photochemical period in southeast England, UK. Ozone (O 3 ), nitrogen dioxide (NO 2 ) and particulate matter (PM 2.5 ) concentrations modelled with different horizontal grid resolutions (9 and 3 km) were evaluated against available ground-level observations from the UK Automatic Urban and Rural Network (AURN) and London Air Quality Network (LAQN) for the period of 24–28 June 2001 with a focus on O 3 predictions. This effort, which represents the first comprehensive performance evaluation of the modelling system over a UK domain, reveals that CMAQ’s ability to reproduce surface O 3 observations varies with O 3 concentrations. It underpredicts O 3 mixing ratios on high-O 3 days and overpredicts the maximum and minimum hourly O 3 values for most low-O 3 days. Model sensitivity analysis with doubled anthropogenic NO x or volatile organic compounds (VOC) emissions and analysis of the daylight- averaged levels of OX (sum of O 3 and NO 2 ) as a function of NO x revealed that the undereprediction of peak O 3 concentrations on high-O 3 days is caused by the underprediction of regional contribution and to a lesser extent local production, which might be related to the underestimation of European emissions in EMEP inventory and the lacked reactivity of the modelled atmosphere. CMAQ systematically underpredicts hourly NO 2 mixing ratios but captures the temporal variations. The normalized mean bias for hourly NO 2 , although much larger than that for O 3 , falls well within the generally accepted range of 20% to 50%. CMAQ with both resolutions (9 and 3 km) significantly underpredicts PM 2.5 mass concentrations and fails to reproduce its temporal variations. While model performance for O 3 and PM 2.5 are not very sensitive to model grid resolutions, a better agreement between modelled and measured hourly NO 2 mixing ratios was achieved with higher resolution. Further investigation into the uncertainties in meteorological input, uncertainties in emissions, as well as representation of physical and chemical processes (e.g. chemical mechanism) in the model is needed to identify the causes for the discrepancies between observations and predictions. r 2008 Elsevier Ltd. All rights reserved. Keywords: Model evaluation; CMAQ; PM 2.5 ; Ozone; UK ARTICLE IN PRESS www.elsevier.com/locate/atmosenv 1352-2310/$ - see front matter r 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.atmosenv.2008.02.051 Corresponding author at: Laboratory for Climate, Environment and Disasters of Western China, Cold and Arid Regions Environmental and Engineering Institute, Chinese Academy of Sciences, Lanzhou, Gansu 730000, PR China. Tel./fax: +86 931 4967168. E-mail addresses: [email protected], [email protected] (Y. Yu).
Transcript
ARTICLE IN PRESS
1352-2310/$ - se
doi:10.1016/j.at
�CorrespondEnvironmental
E-mail addr
Atmospheric Environment 42 (2008) 4870–4883
www.elsevier.com/locate/atmosenv
Performance characteristics of MM5–SMOKE–CMAQfor a summer photochemical episode in southeast England,
United Kingdom
Y. Yua,b,�, R.S. Sokhia, N. Kitwiroona, D.R. Middletonc, B. Fisherd
aCentre for Atmospheric and Instrumentation Research (CAIR), University of Hertfordshire, Hatfield, Hertfordshire AL10 9AB, UKbLaboratory for Climate, Environment and Disasters of Western China, Cold and Arid Regions Environmental and Engineering Research
Institute, Chinese Academy of Sciences, Lanzhou, Gansu 730000, PR ChinacMet Office, FitzRoy Road, Exeter EX1 3PB, UKdEnvironment Agency, Reading RG1 8DQ, UK
Received 15 July 2007; received in revised form 28 November 2007; accepted 19 February 2008
Abstract
In this study a modelling system consisting of Mesoscale Model (MM5), Sparse Matrix Operator Kernel Emissions (SMOKE)
and Community Multiscale Air Quality (CMAQ) model has been applied to a summer photochemical period in southeast
England, UK. Ozone (O3), nitrogen dioxide (NO2) and particulate matter (PM2.5) concentrations modelled with different
horizontal grid resolutions (9 and 3km) were evaluated against available ground-level observations from the UK Automatic
Urban and Rural Network (AURN) and London Air Quality Network (LAQN) for the period of 24–28 June 2001 with a focus
on O3 predictions. This effort, which represents the first comprehensive performance evaluation of the modelling system over a
UK domain, reveals that CMAQ’s ability to reproduce surface O3 observations varies with O3 concentrations. It underpredicts O3
mixing ratios on high-O3 days and overpredicts the maximum and minimum hourly O3 values for most low-O3 days. Model
sensitivity analysis with doubled anthropogenic NOx or volatile organic compounds (VOC) emissions and analysis of the daylight-
averaged levels of OX (sum of O3 and NO2) as a function of NOx revealed that the undereprediction of peak O3 concentrations on
high-O3 days is caused by the underprediction of regional contribution and to a lesser extent local production, which might be
related to the underestimation of European emissions in EMEP inventory and the lacked reactivity of the modelled atmosphere.
CMAQ systematically underpredicts hourly NO2 mixing ratios but captures the temporal variations. The normalized mean bias
for hourly NO2, although much larger than that for O3, falls well within the generally accepted range of�20% to�50%. CMAQ
with both resolutions (9 and 3km) significantly underpredicts PM2.5 mass concentrations and fails to reproduce its temporal
variations. While model performance for O3 and PM2.5 are not very sensitive to model grid resolutions, a better agreement between
modelled and measured hourly NO2 mixing ratios was achieved with higher resolution. Further investigation into the uncertainties
in meteorological input, uncertainties in emissions, as well as representation of physical and chemical processes (e.g. chemical
mechanism) in the model is needed to identify the causes for the discrepancies between observations and predictions.
r 2008 Elsevier Ltd. All rights reserved.
Keywords: Model evaluation; CMAQ; PM2.5; Ozone; UK
e front matter r 2008 Elsevier Ltd. All rights reserved.
mosenv.2008.02.051
ing author at: Laboratory for Climate, Environment and Disasters of Western China, Cold and Arid Regions
and Engineering Institute, Chinese Academy of Sciences, Lanzhou, Gansu 730000, PR China. Tel./fax: +86 931 4967168.
ARTICLE IN PRESSY. Yu et al. / Atmospheric Environment 42 (2008) 4870–4883 4871
1. Introduction
Despite the substantial efforts in reducing pollu-tant emissions in the last decades, especially fromtransportation and industry, pollutant concentra-tions in many European cities are still likely toexceed ambient air quality standards and guidelines.Elevated concentrations of tropospheric ozone (O3),nitrogen dioxide (NO2) and fine particulate matterare known to have negative effects on human healthand the environment (e.g. Englert, 2004; Koop andTole, 2006), and to a larger extent have importantinfluence on the global atmospheric chemistry andclimate change (Jenkin and Clemitshaw, 2000). Inorder to minimize the environmental and healthimpacts of pollutants such as O3, organizations suchas the United Nations Economic Commission forEurope (UNECE) and the Commissions for EuropeCommunities (EC) have proposed and agreed toprotocols designed to reduce concentrations of thesepollutants in Europe (e.g. UNECE, 1999; CEC,1999, 2002). In the United Kingdom, air pollutantsare also subject to standard specified by theNational Air Quality Strategy and European AirQuality Framework Directives.
Air quality models are now widely used toestimate the spatial distribution and evolution oftropospheric pollutant concentrations, resultingfrom both local emissions and long-range transport.They are also valuable tools for the exploration ofemission control strategies to mitigate elevatedconcentrations of pollutants such as O3, NO2 andparticulate matter (e.g. PM10 and PM2.5). Duringthe last two decades different air quality models,ranging from simple statistical models to fully three-dimensional (3-D) comprehensive Eulerian models,have been developed in Europe and elsewhere. TheCommunity Multiscale Air Quality (CMAQ) modeldeveloped by the US Environment ProtectionAgency (US EPA), has been increasingly used inNorth America (e.g. Hogrefe et al., 2004; Eder andYu, 2006; Smyth et al., 2006; Byun et al., 2007) andAsia (e.g. Zhang et al., 2006a; Chen et al., 2007) forboth scientific studies and regulatory assessment.However its application and validations for Eur-opean domains are very limited (e.g. Sokhi et al.,2006; Vautard et al., 2007; Jimenez et al., 2007). Infact, apart from the reported works of Sokhi et al.(2006) and Cocks et al. (2003) there are nopublished studies on the use of CMAQ for studyingair pollution episodes for the UK. Dispersionmodels with simplified treatment for meteorology
and/or chemistry, such as Trajectory Model withAtmospheric Chemical Kinetics (TRACK) (Leeet al., 2000), Ozone Source–receptor Model(OSRM; Hayman et al., 2002) and ADMS (CERC,1998), have been adopted as policy tools in the UK.These approaches are not adequate for cases wherecomplex multi-pollutants and multiscale interac-tions and coupling between atmospheric chemistryand dynamics are involved.
The purpose of the present study is to applyCMAQ to the UK domain and evaluate its abilityto simulate ambient concentrations of O3, NO2 andPM2.5 in the southeast of England during a summer(June 2001) air pollution episode. The modelledconcentrations have been compared with observa-tions from several ground-based monitoring sta-tions. The paper focuses on O3 by undertakingfurther sensitivity studies and analysis of OX (thesum of O3 and NO2) to help understand the O3
behaviour during the period and possible sources ofmodel discrepancies.
The rest of the paper is structured as follows. Themodelling system and its configuration are brieflyintroduced in Section 2, along with details of modelinput preparation. Evaluation results are presentedin Section 3 focusing on O3, NO2 and PM2.5, withdiscussions given in Section 4. Conclusions arepresented in Section 5.
2. Model set-up and input preparation
2.1. Modelling period and CMAQ configuration
The surface O3 distribution in the UK ischaracterized by a marked gradient from south tonorth with the highest concentrations in the southand east of the British Isles (UK PORG, 1997). Thedegree of severity of summertime photochemicalepisodes largely depends on daytime air tempera-ture, and high summertime air pollution events arealmost always associated with anticyclonic condi-tions and temperatures in excess of 28–30 1C (UKPORG, 1997). Long-range transport of O3 and itsprecursors from the European continent may con-tribute significantly to the elevated O3 concentra-tions during these photochemical episodes (Derwentet al., 2003). Although the general synoptic causesof episodes are known, there is an importantscientific and policy need to be able to explain thebehaviour of pollutants, spatially and temporally,under such meteorological conditions. One suchevent occurred during 24–26 June 2001 when warm
ARTICLE IN PRESS
Harwell
Surface stationsWind profilerRadiosounde
Air quality stations
Rural
Urban Centre
Urban BackgroundSuburban
Met Stations
Herstmonceaux
Wattisham
Lullington Heath
23 11
4
5Rochester
Fig. 1. (a) Four nested CMAQ modelling domains; (b) the 3 km-
grid CMAQ domain marked with locations of UK Hourly
Weather Observation sites, UK Automatic Urban and Rural
Network (AURN) sites and London Air Quality Network
(LAQN) observation sites used for model evaluation. 1—London
Y. Yu et al. / Atmospheric Environment 42 (2008) 4870–48834872
weather (with maximum temperatures reaching30 1C on 26 June 2001) prevailed over much of thesouth and east of the UK. The UK AutomaticUrban and Rural Network (AURN) recorded apeak O3 concentration of 198 mgm�3 (�99 ppb) atLullington Heath near the south coast of Englandon 26 June and a peak NO2 concentration of161 mgm�3 (�84 ppb) at an urban site on the sameday. The UK Air Quality Expert Group (AQEG,2004) has identified this episode as an example of aNO2 episode related to a summertime photochemi-cal episode and it was therefore selected forevaluating the Mesoscale Model (MM5)–SparseMatrix Operator Kernel Emissions (SMOKE)–CMAQ modelling system in the present study. Themodelling time period began at 12 UTC 22 June andended at 12 UTC 28 June 2001. The first twosimulation days were used as a ‘spin-up’ period asrecommended by Berge et al. (2001) and Jimenez etal. (2007) and the analyses focused on the following4 days after which the episode dissipated as a lowpressure system brought relatively cooler and morechangeable weather with occasional outbreaks ofrain or showers.
In this study, the US EPA’s CMAQ version 4.4was used with a modified version of the Carbon-Bond Mechanism version IV (CB-IV) chemicalmechanism. Fig. 1a shows the CMAQ modellingdomain that consists of four nested domains withresolutions of 81, 27, 9 and 3 km. The coarsestdomain covers most of the Europe and the finest3 km-grid domain covers the southeast of England.Vertically there are 26 s-levels extending from thesurface to an altitude of about 14 km. Vertical layerswere unevenly distributed with fifteen layers in thelowest kilometre and a surface layer of approxi-mately 14m above ground level (AGL). Fig. 1bshows the enlarged 3 km-grid domain and thelocations of measurement sites referred to in thispaper.
2.2. Model input preparation
2.2.1. Meteorology
The Fifth-generation Pennsylvania State Univer-sity–National Center for Atmospheric Research(NCAR) MM5, version 3 (Dudhia et al., 2004)was used to generate meteorological fields forCMAQ. The MM5 was configured to have fournested domains, covering and aligning with theCMAQ domains shown in Fig. 1a with each of theMM5 domain being at least five grid cells larger
than the corresponding CMAQ domains. TheEuropean Centre for Medium-Range WeatherForecasts (ECMWF) 11� 11 reanalysis data avail-able at every 6 h were used to provide initial andboundary conditions for the coarsest MM5 domain.The physical options used in MM5 include theMedium Range Forecast (MRF) PBL scheme, theDudhia simple ice microphysics scheme, the cloudradiation scheme and the five-layer soil model. TheAnthes–Kuo cumulus parameterization scheme wasused for the coarsest model domain, the Grellcumulus parameterization scheme was used for the27- and 9-km grid domains and no cumulus schemewas used for the 3-km resolution domain.
ARTICLE IN PRESSY. Yu et al. / Atmospheric Environment 42 (2008) 4870–4883 4873
2.2.2. Emissions
Annual anthropogenic emissions for six pollu-tants, i.e. NOx, non-methane volatile organiccompounds (NMVOCs), sulphur dioxide (SO2),carbon monoxide (CO), ammonia (NH3) and fineand coarse particulate matter (i.e. PM2.5 and PMcoarse), were taken from the European Monitoringand Evaluation Programme (EMEP) for year 2002(http://www.emep.int) and used for all the CMAQdomains except model grid cells covering the UK(including Northern Ireland, Scotland, England andWales), where the 1-km spatial resolution NationalAtmospheric Emissions Inventory (NAEI) data(http://www.naei.org.uk/) were used. Point sourceemissions were extracted from the European Pollu-tant Emission Register (EPER, http://www.eper.cec.eu.int/) and NAEI database, for non-UK andUK point sources, respectively. The SMOKE model(Carolina Environmental Program, 2003) was usedto process these annual emissions to a temporallyresolved, spatially distributed and speciated model-ready emissions data for CMAQ. NMVOC emis-sions were split into model species represented in theCB-IV chemical mechanism. Different speciationprofiles were derived for different activity sectorsbased on the detailed UK volatile organic com-pounds (VOC) speciation given in Dore et al.(2004). It is assumed that the speciation profile forthe UK could be applied across Europe withoutfurther adjustment. This assumption was consideredto be reasonable as vehicle exhaust emissions, fuelevaporative emissions and solvents are likely tohave similar profiles across north west Europe,although uncertainty in the spatial distribution ofindividual VOC emissions may be large for otherparts of Europe. Temporal profiles were developed,taking into account monthly, weekday/weekend andhourly variations, for each country, activity sectorsand pollutant using information provided by theInstitute for Energy Economics and Rational Use ofEnergy, University of Stuttgart (IER, privatecommunication) and information available in Jen-kin et al. (2000).
Biogenic emissions of isoprene and monoterpeneswere calculated based on the following formula(Guenther et al., 1995; Sanderson, 2002):
Fi ¼ �iDgi
where Fi is the emission flux (mgm�2 h�1), ei is an eco-system (i) dependent emission factor (mgCg�1 h�1)and D is the foliar density. Values of ei and D aretaken from Sanderson (2002); gi is the environmental
correction factor accounting for the dependence ontemperature and radiation (Guenther et al., 1995).The spatial distribution of ecosystems was establishedby firstly aggregating the 100m resolution Coordina-tion of Information on the Environment (CORINE)Land Cover data for Europe (CLC2000, http://dataservice.eea.europa.eu/dataservice/) to the CMAQmodel grids. Then the 44 CORINE land use classeswere aggregated into four ecosystems (i.e. grass,broadleaf forest, needle leaf forest and shrub) andthe fraction of the area of each grid cell covered byeach ecosystem class and the associated emissions ofisoprene and monoterpenes was calculated using thehourly temperature and solar radiation values fromMM5. Nitrogen oxide (NOx) released from soil andformed by lightning are not included in the presentstudy. The possible effect on model results will bediscussed later.
2.2.3. Initial and boundary conditions
The initial and boundary conditions for thecoarsest CMAQ domain were generated based onmonthly mean data from the UK Met Office global3-D Lagrangian tropospheric chemistry model(STOCHEM). This model outputs concentrationsof 26 species with a horizontal resolution of 51latitude� 51 longitude and nine vertical layersextending from surface up to 150 hPa (Collinset al., 1997). The initial and boundary conditionsfor the inner three domains are provided by thecoarser domain.
3. Model evaluation and simulation results
3.1. Meteorological predictions
The MM5 modelled near-surface temperature,wind speed and wind direction were compared tohourly weather observations from 29 land surfacestations archived at the Met Office Integrated DataArchive System (MIDAS; UK MeteorologicalOffice, 2006). These quantities were selected becausethey reflect the nature of the local thermodynamiccirculation and govern contaminant distributions inair quality models. Several standard statisticalmeasures were employed for the evaluation. Theseinclude the mean observed and modelled values, themean bias (MB), the normalized mean bias (NMB),the mean error (ME), the normalized mean error(NME), the root-mean-square error (RMSE) andthe index of agreement (IA). While calculation of
ARTICLE IN PRESSY. Yu et al. / Atmospheric Environment 42 (2008) 4870–48834874
these statistics is straightforward for wind speed andtemperature, it poses a problem for circular data,i.e. wind direction. To get around this problem, a‘modified’ wind direction, following Lee and Fer-nando (2004), was used, where a 3601 is either addedto or subtracted from the predicted wind directionto minimize the absolute difference between theobserved and predicted wind direction. Table 1summarizes the performance statistics for MM5calculated values based on near-surface data fromthe 29 stations displayed in Fig. 1b, along withdefinitions of statistical measures. These valuesreflect averages over space (all monitoring stationswithin the 3 km-grid CMAQ domain) and time (allhours in the simulation period). The table showssmall MEs (and biases) for the 2-m temperature and10-m wind speed and direction, and low RMSE. Ingeneral the model captured the observed near-surface temperatures and winds quite well. OverallIA values of 0.97 for 2-m temperature and 0.75 and0.93 for 10-m wind speed and direction, wereachieved. The IA, which is a measure of howwell the solutions represent the spatial variability(Willmott et al., 1985), indicates a good overallagreement between observations and model predic-tions.
A qualitative comparison of the modelled me-teorological fields with observations is shown inFig. 2. The figure shows the vertical profiles of windspeed, wind direction and temperature at Herst-monceaux (see Fig. 1b for location) on 26 June 2001when the highest temperature was experienced.Herstmonceaux is the only radiosounde stationwithin the 3-km grid MM5 domain. The model isable to reproduce the major features of the observed
Table 1
Performance statistics of modelled temperature and wind speed and di
Variables Temp. (K) WS (m s�1) W
Mean obs. O 18.2 3.4 1
Mean sim. M 18.9 3.0 1
Total N 4599 4414 43
MB 0.7 �0.3
NMB (%) 3.7 �8.8
ME 1.4 1.2
NME (%) 7.6 36.6
RMSE 1.7 1.5
Index of agreement 0.97 0.75
am ¼ modelled, o ¼ observed.
wind and temperature fields and the modelledprofiles show a generally good agreement withmeasurements.
3.2. O3, NO2 and PM2.5 predictions
Hourly surface concentrations of O3, NO2 andPM2.5 obtained from the UK AURN (http://www.airquality.co.uk/archive/index.php) and Lon-don Air Quality Network (LAQN, http://www.londonair.org.uk/london/asp/default.asp) were usedin the evaluation. The locations of air qualitymonitoring sites used in the evaluation are shownin Fig. 1b. Only monitoring sites reporting mea-surements for at least 75% of the hours in thestudied period were included in the analysis.Monitoring sites are presented for four categoriesas used in the networks, i.e. rural, urban back-ground, suburban and urban centre. Traffic mon-itoring sites are not considered due to their poorrepresentativeness in the model resolution. Themodel evaluation focuses on the 3-km grid domain.The results from a coarser grid domain, i.e. 9 km,are, therefore, only compared against the observa-tions obtained for locations within the 3-km griddomain. Model values used for evaluation wereextracted from the first model level (about 14mAGL). In addition to the statistical measures used inSection 3.1, the correlation coefficient R was alsocalculated to quantify the model performance.Although no single set of evaluation techniques isuniversally recommended, the statistical measuresused here have been widely used in recent regionalair quality model evaluations (Hogrefe et al., 2004;Eder and Yu, 2006).
0 100 200 300Wind direction (deg.) Wind direction (deg.)
measuredmodelled
270 280 290 300 310 320
0 5 10 15
0 100 200 300
4000
2000
Hei
ght (
m)
0
4000
2000
Hei
ght (
m)
0
4000
2000
Fig. 2. Comparison of modelled and measured vertical profiles of
temperature, wind speed and wind direction at radiosounde
station Herstmonceaux on 26 June 2001.
Y. Yu et al. / Atmospheric Environment 42 (2008) 4870–4883 4875
3.2.1. O3
The modelled O3 time series from both 3- and9-km resolution simulations are compared with themeasured values in Fig. 3 at six ‘representative’ siteschosen to cover the different types of monitoringsites. The model is in general able to capture thediurnal O3 variations for most of the days andexhibits overall good agreement with measurements(with R ¼ 0.7, see Table 2). The modelled O3 mixingratios are in close agreement with observationsduring most of the nighttime hours at the rural site.While CMAQ underpredicts the maximum O3
mixing ratios on high-O3 days, for example, 25and 26 June, it tends to overpredict the maximumand minimum O3 mixing ratios for most low-O3
days, especially at urban centre and suburban sites,with most of the overpredictions occurring duringnight and early morning hours, indicating theunderestimaion of O3 titration, which is consistentwith the underprediction of NO2 mixing ratiosshown in Fig. 6 and Table 3. Overall, the 3 km-gridsimulation gives comparable or slightly better
predictions than the 9 km-grid simulation, especiallyat urban and suburban sites.
Fig. 4 presents a scatter plot of modelled versusmeasured hourly O3 mixing ratios for all themodelling hours and sites. The most importantfeature of model errors revealed by this plot is theoverprediction of hourly O3 concentrations in thelower range and underprediction in the higher rangeof O3 concentrations. Table 2 summarizes thecorresponding O3 performance statistics for the3 km-grid simulation. The statistics for 9 km-gridsimulation are similar to those for 3 km-gridsimulation and are not shown for brevity. The MB(�2.0 ppb) and NMB (�5.3%) show an under-prediction of hourly O3 mixing ratios. For differentcategories of sites (not shown for brevity), NMBranges from �7.8% at urban background sites to�2.7% at suburban sites. The ME and NME are11.9 ppb and 32.4%, respectively, with valuesranging from 23.7% at rural sites to 40.1% aturban centre sites. In total, about 83% of allmodelled values are within a factor of two of thecorresponding measured O3 concentrations. Thestatistical measures for daily maximum 8-h averageO3 mixing ratios are also calculated and summar-ized in the last column of Table 2. The dailymaximum 8-h average O3 mixing ratios wereunderpredicted by 14% based on NMB. The valuesof MB and NMB are higher than those for hourlyO3, indicating that the model is less accurate inreproducing the highest hourly O3 values experi-enced during the episode. US EPA has suggestedinformal criteria for regulatory modelling practicesof 75–15% for NMB and 30–35% for NME(Russell and Dennis, 2000). It is seen from Table 2that our evaluation statistics fall well within thesuggested values. The performance of our modellingsystem is also comparable to or better than someother air quality models used in Europe. Forexample, Schmidt et al. (2001), using the Europeanscale Eulerian chemistry transport model CHI-MERE, obtained a R of 0.58–0.81 for dailymaximum O3 concentrations at representative ruralsites in the UK and Dufour et al. (2005), using thenew MOdele de Chimie Atmospherique a GrandeEchelle (MOCAGE) 3-D multiscale chemistry andtransport model, reported a R of 0.44–0.86 forhourly O3 concentrations in two summer episodesthat occurred during the Experience sur Site pourCOntraindre les Modeles de Pollution atmospher-ique et de Transport d’ Emission (ESCOMPTE)field programme. In general, CMAQ performs
Fig. 3. Comparison of measured and modelled O3 mixing ratios at a selection of six ‘representative’ sites, chosen to cover the different
types of monitoring sites within the 3 km-grid domain (see Fig. 1b for locations).
Table 2
Performance statistics for modelled surface ozone concentrations
(ppb)
Statistics Hourly O3 Max. 8-h mean O3
Mean obs. O 37.0 54.4
Mean sim. M 35.0 46.8
Total N 2128 110
MB �2.0 �7.6
NMB (%) �5.3 �14.0
ME 11.9 14.6
NME (%) 32.4 26.9
RMSE 15.4 18.2
R 0.70 0.40
IA 0.79 0.53
% Within factor
of 2 of measured
83.3 94.5
Table 3
Performance statistics of modelled surface NO2 concentrations
(ppb)
Statistics 3 km 9km
Mean obs. O 21.3
Mean sim. M 15.1 14.2
Total N 2762MB �6.2 �7.1
NMB% �28.9 �33.5
ME 9.7 10.5
NME% 45.5 49.2
RMSE 13.5 13.0
R 0.64 0.58
IA 0.75 0.70
% Within factor of 2 of measured 62 59
Y. Yu et al. / Atmospheric Environment 42 (2008) 4870–48834876
ARTICLE IN PRESS
0
20
20 40 60 80 100 120
40
60
80
100
120
0Measured O3 (ppb)
Mod
elle
d O
3 (p
pb)
Fig. 4. Measured versus modelled hourly O3 mixing ratio.
Modelled values were extracted from the 3 km-grid simulation;
1:2, 1:1 and 2:1 reference lines are provided.
Y. Yu et al. / Atmospheric Environment 42 (2008) 4870–4883 4877
better at rural sites than at urban sites, where O3
prediction is more sensitive to the representation ofmixing near sources, errors in meteorologicalparameters (e.g. boundary layer height) and titra-tion by local emissions.
From the above analyses it is seen that undermoderate photochemical activity, CMAQ can re-produce the observed O3 concentrations, but themodel tends to underpredict peak O3 mixing ratios(455 ppbv) during this typical summer episode. Theunderprediction of peak O3 concentrations on high-O3 days indicates that the O3-production chemistrymay not be sufficiently reactive. A further examina-tion of the surface O3 time series shown in Fig. 3reveals that urban background (suburban) siteslocated upwind of the London metropolitan area(see Fig. 5a for the wind field), for exampleSevenoaks (Croydon), experienced a nighttimeoverprediction (peak O3 underprediction) that arenot seen for sites located downwind, indicating thatlong-range transport of O3 and its precursors fromthe European continent may contribute to theelevated O3 concentration. This is also indicatedby the more significant underprediction of NO2 forurban background/suburban sites located upwindof the London metropolitan area (see Fig. 6). Someearly studies (e.g. AQEG, 2004; Derwent et al.,2003) have also suggested that the contributionfrom European emissions could be an importantfactor in high-O3 episodes that occur in summer in
the UK. Surface O3 distribution from the 3 km-gridsimulation at 15 UTC is shown in Fig. 5a. Themodelled spatial distributions of O3 were similar tothat measured for 25 and 26 June, but the model didnot reproduce the O3 mixing ratios that exceeded55 ppb at sites affected by urban plume on 25 Juneand at most of the observational sites on 26 June. Itcan be inferred from the wind fields that bothLondon emissions and European emissions mayhave contributed to the spatial distribution of O3 on25 June while the transport into the UK of alreadypolluted air from the European boundary layer maybe more important on 26 June. To investigate thepossible causes of the underestimations, sensitivitystudies with doubled anthropogenic NOx or VOCemissions were carried out. However, it should benoted that there are several other sources ofuncertainties in the model, including inaccuratemeteorological predictions (e.g. cloud cover, PBLheight) and less well represented physical/chemicalprocesses (e.g. entrainment of regionally polluted airfrom aloft, enhanced VOC reactivity), which alsoinfluence the prediction of photooxidants concen-tration and should be addressed in future studies.For example, a recent study by Lee et al. (2006)suggests that entrainment of regionally polluted airfrom aloft may contribute to the rapid increase ofO3 in the morning on high-O3 days. In addition,chemical processes leading to O3 production underepisodic high temperature conditions may be sub-stantially different from that occurring at normalconditions. It was found that the total VOCreactivity to OH could be doubled under high-temperature conditions such as those that occurredduring the August 2003 heatwave (Lee et al., 2006).It is seen from Fig. 5b that doubling the anthro-pogenic NOx emission leads to less O3 over most ofthe domain on 25 June and over northeast part ofthe domain on 26 June but it increases O3
concentration over the southwest domain due tothe transport of European-derived O3 from thesouth and east model boundaries on 26 June. On 25June, doubling anthropogenic VOC emission resultsin increasing O3 concentrations in urban plume,where high NOx are present and oxidation isVOC limited, while on 26 June it increases theO3 transported from the east model boundary.The above analyses indicate that both NOx andVOC emissions may be underestimated in theEMEP inventory, which leads to the underpredictionof O3 and its precursors transported into thestudied domain under episodic conditions. Although
ARTICLE IN PRESS
Fig. 5. (a) Spatial distribution of ozone for two consecutive days at 15 UTC for base case simulation; (b) difference in ozone between cases
with doubled anthropogenic NOx emission and base emission and (c) difference in ozone between cases with doubled anthropogenic VOC
emission and base emission. The coloured points (same colour scale as model results) indicate the measured values.
Y. Yu et al. / Atmospheric Environment 42 (2008) 4870–48834878
doubling of NOx or VOC emissions reduced the lowbias in the modelled O3 concentration compared tothe observations, the model still could not repro-duce the high O3 observed during the episode,indicating the lack of reactivity of the modelledatmosphere.
3.2.2. NO2
Fig. 6 compares the measured NO2 time series withthe modelled results at the same six sites as for O3.Both the diurnal variations and magnitudes of NO2
mixing ratios are well captured by the model at therural site. The model performs better at LondonTeddington (located near the London metropolitanarea) than at Sevenoaks (located upwind of theLondon metropolitan area), suggesting the under-prediction of O3 and its precursors transported fromthe European continent as discussed in Section 3.2.1.The NO2 mixing ratios were appreciably under-predicted at suburban and urban centre sites most ofthe time, indicating the underestimation of NOx
emissions in these areas. The appreciable differenceof predicted NO2 mixing ratios with differenthorizontal grid resolutions indicates that the non-linearity of chemical reactions and heterogeneity
associated with precursor emissions have a significantimpact on model predictions.
Fig. 7 shows a scatter plot of the modelled versusmeasured hourly NO2 mixing ratios for all hoursand sites. Overall, the model reproduced about 62%of the hourly NO2 mixing ratios within a factor oftwo of the measurement. Table 3 summarizes thehourly NO2 performance statistics for both the3 km- and the 9 km-grid resolutions for all the sites.The statistics for NO2 show much larger bias anderror when compared to the same statistics for O3
due to the generally higher sensitivity of NO2 toerrors in emissions and meteorology, especiallyunder stagnant conditions. Overall, the modelunderpredicted NO2 concentrations with a MB of�6.2 ppb and a NMB of �28.9% for the 3 km-gridsimulation and a MB of �7.1 ppb and a NMB of�33.5% for the 9 km-grid simulation. The ME andNME values over all hours and sites are 9.7 ppb and45.5%, respectively, for the 3 km-grid simulationand 10.5 ppb and 49.2%, respectively, for the 9 km-grid simulation. Underprediction of O3 precursors isexperienced by many currently used photochemicalmodels (Russell and Dennis, 2000). The negativeNMB of NO2 from our study falls well within the
Fig. 6. Comparison of measured and modelled NO2 mixing ratios at the same six sites as shown in Fig. 3.
Y. Yu et al. / Atmospheric Environment 42 (2008) 4870–4883 4879
range of �20% to �50% inferred from otherstudies (Hanna et al., 1996).
3.2.3. Fine particulate matter (PM2.5)
Hourly measurements of PM2.5 are available atfour sites within the evaluation domain (one of themis located at roadside and is not included in thepresent study). Fig. 8a compares the measured andmodelled time series of PM2.5 concentrations atHarwell, Rochester and London Bloomsbury. Boththe 9 km- and 3 km-grid resolution simulationsfailed to reproduce the temporal variations andmagnitude of the measured PM2.5 mass concen-trations. Overall, the model tends to underpredictthe PM2.5 mass concentrations with a MB of�8.7 mgm�3 and a NMB of �45% over the threesites. The ME and NME are 9 mgm�3 and 49.3%,
respectively. These values indicate an overall sig-nificant underprediction of PM2.5 mass concentra-tions. A scatter plot of the modelled versusmeasured PM2.5 mass concentrations for all hoursand sites is shown in Fig. 8b. Only 45% of allmodelled PM2.5 concentrations are within a factorof two of the corresponding measurement. Thesestatistics are consistent with the current perfor-mance expected from most air quality models forparticulate matter (e.g. Seigneur, 2001; Bessagnetet al., 2004; Zhang et al., 2006b; Vautard et al.,2007). The limited availability of PM2.5 massconcentration and composition data makes theresults of the model performance analysis lessconclusive and robust; however, the analysis stillprovides a general indication of the model perfor-mance. Several factors could contribute to the
ARTICLE IN PRESS
0
10
20
30
40
50
60
70
80
90
100
0 20 40 60 80 100Measured NO2 (ppb)
Mod
elle
d N
O2
(ppb
)
Fig. 7. Measured versus modelled hourly NO2 mixing ratio.
Modelled values were extracted from the 3 km-grid simulation;
Fig. 8. (a) Comparison of measured and modelled time series of
PM2.5 mass concentrations at Harwell, Rochester and London
Bloomsbury sites; (b) measured versus modelled hourly PM2.5
mass concentration; 1:2, 1:1 and 2:1 reference lines are provided.
Y. Yu et al. / Atmospheric Environment 42 (2008) 4870–48834880
underprediction of PM2.5 mass concentrations,including uncertainties in emissions of particulatematter precursor gases and primary particulatematter and uncertainties in the model treatment ofchemistry and thermodynamics of aerosols. Forexample, sea salt and the interactions between thefine- and coarse-mode particles are not treated inCMAQ v4.4 (e.g. Zhang et al., 2006c); emissionsinventory (e.g. EMEP) may be deficient, as somebiogenic sources are missing and emissions of re-suspension related to traffic on paved or dirt roadsor related to wind are not included. Further studiesare needed to evaluate the relative importance ofthese parameters on overall model performance forPM2.5 prediction.
4. Discussions
The above results on O3 predictions reveal ageneral underprediction of the hourly and dailymaximum 8-h mean O3 mixing ratios by CMAQ forthe studied area, especially on high-O3 days. Under-prediction of daily maximum O3 mixing ratios formost UK rural sites was also reported by Schmidtet al. (2001) and was attributed to the under-estimation of boundary concentrations. Previousstudies on regional O3 distribution across theBritish Isles using both the EMEP and EdinburghLancaster Model for Ozone (ELMO) models alsoindicated the difficulty for these models to capture
the high O3 levels occurring in the southern England(Metcalfe et al., 2002). Additional sensitivity studywith modified boundary conditions (not shown)indicates that boundary condition has very limitedeffect on surface O3 prediction, while correctemission input is more important for a better modelperformance as shown in Section 3.2.1.
ARTICLE IN PRESSY. Yu et al. / Atmospheric Environment 42 (2008) 4870–4883 4881
Previous study by Clapp and Jenkin (2001) showsthat the level of OX (the sum of O3 and NO2) at agiven location is made up of NOx-independentregional contribution (the intercept) and NOx-dependent local contribution (the slope). It is thuspossible to estimate which part was underestimatedby CMAQ by comparing measured and modelledOX versus NOx relationship. Daylight averageanalyses were carried out for June 2001 using datafrom 16 monitoring sites, where O3, NO and NO2
were simultaneously measured within the innermostmodel domain. Following Clapp and Jenkin (2001),the data were separated as ‘episode’ days (withdaylight-averaged OX mixing ratio at Teddington450 ppb) and ‘non-episode’ days, which resulted in24, 25 and 26 being selected as ‘episode’ days. It isseen from Fig. 9 that the level of OX wassignificantly higher on ‘episode’ days (black opendots), as a result of the increased regional contribu-tion (the intercept increased from 41 ppb on ‘non-episode’ days to 73 ppb on ‘episode’ days) and a 3%higher local contribution. Thus the enhanced OXlevels occurring during this episode is a contributionof regional transport in a combination of localprocesses, which is consistent with the analysesshown in Section 3.2.1. CMAQ significantly under-predicted the regional contribution during theepisode. The model also underpredicted the localcontribution indicated by the smaller increase ofOX with NOx than is observed (16.9% versus18.4%). These analyses further demonstrate theimportance of long-range transport of O3 and itsprecursors from the European continent to theelevated O3 concentrations during the episode.
[OX] = 0.152[NOx] + 41.2('non-episode' days)
[OX] = 0.184[NOx] + 72.9('episode days')
[OX] = 0.169[NOx] + 56.2(modelled)
0
20
20 40 60 80 100 120
40
60
80
100
120
140
0NOx mixing ratio (ppb)
OX
mix
ing
ratio
(ppb
)
Fig. 9. Variation of daylight-time-averaged mixing ratios of OX
with the level of NOx. Data are presented for each day of June
2001 at 16 sites. The lines were defined by regression analysis of
observed ‘non-episode’, ‘episode’ and modelled ‘episode’ days.
Black solid dots, black open dots and triangles are for ‘non-
episode’ days, ‘episode’ days and modelled ‘episode’ days,
respectively.
As mentioned in Section 2, NOx emissions fromsoil are not included in the present study, which mayalso lead to underprediciton of O3 mixing ratios.Stohl et al. (1996) studied the importance of NOx
emissions from soil on O3 production in Europe andargued that although on European average, bio-genic NO emissions account for only 4% ofanthropogenic NO emissions, they can be relevantin rural areas. The inclusion of NOx emissions fromsoil and their possible effect on O3 predictions willbe a subject of future work.
5. Conclusions
A performance evaluation of MM5–SMOKE–CMAQ modelling system for southeast England,UK, for a summer photochemical episode has beenpresented. The simulated concentrations of O3, NO2
and PM2.5 were compared with ground-levelobservations from the AURN and LAQN. Theevaluation shows that CMAQ tends to underpredicthourly O3 mixing ratios on high-O3 days andoverpredict the maximum and minimum O3 mixingratios for most low-O3 days. Sensitivity studies andanalysis of ambient OX levels (sum of O3 and NO2)as a function of NOx reveal that the transport of O3
and its precursors from the European continent wassignificantly underpredicted by CMAQ, which islikely resulted from the underestimation of Eur-opean emissions, except other physical/chemicalprocesses, e.g. entrainment of regionally pollutedair from aloft and enhanced VOC reactivity underepisodic conditions, that are not well represented inthe model.
In terms of NO2, the model captured themagnitudes and temporal variations generally well,but produced much larger bias and error than thosefor O3 with significant underpredictions at urbancentre and suburban areas, which may be due to theinaccurate meteorological field (e.g. PBL height)and missing or incorrect emissions. For example,NOx emissions from soil were not accounted for inthe present study.
For PM2.5, CMAQ with both resolutions (9 and3km) significantly underpredicted the mass concen-trations and failed to reproduce the temporalvariations, with only 45% of all modelled PM2.5
mass concentrations falling within a factor of two ofthe corresponding measured values. While CMAQperformance for O3 and PM2.5 was not sensitive tomodel resolutions, higher resolution was found to bebeneficial for properly simulating NO2 mixing ratios.
ARTICLE IN PRESSY. Yu et al. / Atmospheric Environment 42 (2008) 4870–48834882
The overall performance of MM5–SMOKE–CMAQ modelling system is comparable to or betterthan similar model predictions by other models usedfor European applications. Further studies, how-ever, are needed to explore how the model responseto uncertainties linked to different processes oftropospheric chemistry modelling, namely large-scale pollution transport, refinement of emissioninventory, as well as representation of the physicaland chemical processes (e.g. enhanced VOC reac-tivity under episodic conditions). Notably, photo-chemical episode may vary from event to event andthe evaluation of one episode is by no meanscomprehensive. More cases need to be conducted toincrease the confidence in these results.
Acknowledgements
This study was supported by the UK Environ-ment Agency and the AIR4EU project fundedunder FP6, which is a member project of theCluster of European Air Quality Research(CLEAR). This work also forms part of COST728 activities on the evaluation and application ofmesoscale models for air pollution research. We arethankful to Richard Derwent for providing theSTOCHEM simulations. Model emissions wereprocessed using the temporal profiles kindly pro-vided by IER at the University of Stuttgart. TheECMWF data were obtained from BADC.
References
AQEG, 2004. Nitrogen dioxide in the United Kingdom. Air
Quality Expert Group first report. Department of the
Environment, London, UK.
Berge, E., Huang, H-C., Chang, J., Liu, TH., 2001. A study of the
importance of initial conditions for photochemical oxidant
modelling. Journal of Geophysical Research—Atmospheres
106 (D1), 1347–1363.
Bessagnet, B., Hodzic, A., Vautard, R., Beekmann, M., Cheinet,
