A novel methodology for interpreting air quality measurements from urban streets using CFD modelling

lable at ScienceDirect

Atmospheric Environment 45 (2011) 5230e5239

Contents lists avai

Atmospheric Environment

journal homepage: www.elsevier .com/locate/atmosenv

A novel methodology for interpreting air quality measurements from urbanstreets using CFD modelling

Efisio Solazzo a,*, Sotiris Vardoulakis a,b, Xiaoming Cai a

aDivision of Environmental Health & Risk Management, School of Geography, Earth & Environmental Sciences, University of Birmingham, Birmingham B15 2TT, UKbCentre for Radiation, Chemical & Environmental Hazards, Health Protection Agency, Chilton, Didcot, Oxon OX11 0RQ, UK

a r t i c l e i n f o

Article history:Received 1 December 2010Received in revised form4 May 2011Accepted 5 May 2011

Keywords:Urban street canyonsCFDPassive diffusion tubesRoadside monitoring

* Corresponding author. Current address: EuropeanCentre, Institute for Environment and Sustainability,

E-mail address: [email protected] (E.

1352-2310/$ e see front matter � 2011 Elsevier Ltd.doi:10.1016/j.atmosenv.2011.05.022

a b s t r a c t

In this study, a novel computational fluid dynamics (CFD) based methodology has been developed tointerpret long-term averaged measurements of pollutant concentrations collected at roadside locations.The methodology is applied to the analysis of pollutant dispersion in Stratford Road (SR), a busy streetcanyon in Birmingham (UK), where a one-year sampling campaign was carried out between August 2005and July 2006. Firstly, a number of dispersion scenarios are defined by combining sets of synoptic windvelocity and direction. Assuming neutral atmospheric stability, CFD simulations are conducted for all thescenarios, by applying the standard k-3 turbulence model, with the aim of creating a database of nor-malised pollutant concentrations at specific locations within the street. Modelled concentration for allwind scenarios were compared with hourly observed NOx data. In order to compare with long-termaveraged measurements, a weighted average of the CFD-calculated concentration fields was derived,with the weighting coefficients being proportional to the frequency of each scenario observed during theexamined period (either monthly or annually). In summary the methodology consists of (i) identifyingthe main dispersion scenarios for the street based on wind speed and directions data, (ii) creatinga database of CFD-calculated concentration fields for the identified dispersion scenarios, and (iii)combining the CFD results based on the frequency of occurrence of each dispersion scenario during theexamined period. The methodology has been applied to calculate monthly and annually averagedbenzene concentration at several locations within the street canyon so that a direct comparison withobservations could be made. The results of this study indicate that, within the simplifying assumption ofnon-buoyant flow, CFD modelling can aid understanding of long-term air quality measurements, andhelp assessing the representativeness of monitoring locations for population exposure studies.

� 2011 Elsevier Ltd. All rights reserved.

1. Introduction

The increasing urbanisation and use of road transport havecreated concern about traffic-related air pollution and associatedhealth effects in urban environments (Baik et al., 2007; Vardoulakiset al., 2007; Blocken et al., 2008). In urban streets in particular,where motor vehicle density is higher and pollutant dispersionreduced due to the presence of buildings, air pollution levels oftenexceed national air quality limit values. In the UK, over 200 localauthorities have already declared Air Quality Management Areas(AQMA) mainly due to exceedences of the annual NO2 and the 24-hPM10 limit values in city centres affected by heavy and/or congested

Commission Joint ResearchIspra, Italy.Solazzo).

All rights reserved.

road traffic (Harrison et al., 2008). Developing air quality actionplans to reduce the damaging health effects of pollutants in urbanenvironment is now a major environmental challenge to localgovernment (Woodfield et al., 2006). In addition to monitoringtools, local air quality assessments require appropriate modellingtools that can help interpret measurements and test future emis-sion scenarios (Vardoulakis et al., 2007). A large variety of modelsare available to this scope ranging from parameterised dispersionmodels, to chemistry-transport models (see, e.g. Vardoulakis et al.,2003; Holmes and Morawska, 2006).

Numerical computational fluid dynamics (CFD) modelling canplay an important role in characterising the mechanical processesgoverning air pollutant dispersion within urban areas. Reynolds-Averaged NaviereStocks Equations (RANS), Large-Eddy Simula-tions (LES), and Direct number simulations (DNS) models have allproved to be useful in assisting with the interpretation of three-dimensional flows and dispersion patterns in complex geometries

Fig. 1. Plan view of SR. The section of the street studied is highlighted with light blue.The position of the continuous monitoring station is also indicated. (For interpretationof the references to colour in this figure legend, the reader is referred to the webversion of this article.)

E. Solazzo et al. / Atmospheric Environment 45 (2011) 5230e5239 5231

(Baik et al., 2007; Buccolieri et al., 2010; Blocken et al., 2008; Caiet al., 2008; Coceal et al., 2007) as well as in interpreting airquality measurements frommonitoring stations (e.g. Murena et al.,2008).

RANS-based CFD models remain the most practical and widelyapplied tools in industry and academia, mainly due to thesubstantial computational resources required to run LES and DNSmodels over large domains (Lien and Yee, 2004). The standard k-3model is the most documented and validated RANS turbulencemodel for urban dispersion applications (Vardoulakis et al., 2011a).Furthermore, it is available in most commercial CFD softwarepackages and it is widely used by industry.

Intensive monitoring campaigns such as the MUST experiment(Buccolieri and Di Sabatino, 2011), the Nantes ’99 experiment(Louka et al., 2002) and others, have employed the k-3 CFDmodel tomimic complex atmospheric flow patterns in urban areas. In thiscontext, CFD can provide high resolution spatial analysis of the flowfeatures. It should be noted that routinely used atmosphericdispersion models (e.g. Gaussian), although capable of predictingpollutants concentrations over long and short time periods, cannotprovide information about localised pollution effects or complexflow fields.

To our knowledge, however, CFD has not been used before tosimulate long-term averaged air quality measurements, e.g. frompassive sampling, in local air quality management applications.This is because of the large computational resources and longsimulation running time required which make, at present, routineCFD modelling use impractical. A way of overcoming this drawbackcould be by transferring CFD results from one scenario to anotherby applying simple mathematical transformations without theneed of re-running the model on each occasion. For example, oncea CFD simulation has been run over an extended domain for a givenset of input conditions (meteorology and emissions), an appro-priate methodology could be developed to allow the use of theresults for another set of input conditions.

The main aim of this work is therefore to develop such a meth-odology, with the following specific objectives: (a) to identify theresponse of themodel to changes in emissions under different windconditions; (b) to run the CFD simulations for this set of scenarios;(c) to create a database of normalised concentrations or pollution“footprints” at specific roadside locations; (d) to develop a statis-tical methodology for combining CFD simulations according to thespecific meteorological conditions and emission factors of thestreet canyon under study. This methodology can be thought asa version of the emulator function (Fisher et al., 2010), who definedan emulator as a statistical representation of a simulator (in thiscase the results of the CFD model). In fact, with a minor modifica-tion, the methodology presented here can be used in model diag-nostic analysis, for quantifying the variation in concentration levelsunder different emission and wind scenarios.

The methodology developed in this study has been applied tostudy the dispersion of air pollutants in Stratford Road (SR), a busystreet canyon within an AQMA in Birmingham (UK), using moni-toring data collected between August 2005 and July 2006. Passivesampling data of BTEX (benzene, toluene, ethylbenzene andxylenes) and NOx data from a continuous monitoring station havebeen used to evaluate the methodology.

2. Methods

2.1. Monitoring site and sampling campaign

An extensive air quality monitoring campaignwas carried out inSR between August 2005 and July 2006. Passive samplers wereplaced along a straight section of SR (150 m length) between two

major junctions (Vardoulakis et al., 2009, 2011b) (Fig. 1). Theaverage building heights H are 12 and 10 m on the east and westside, respectively. The total width W of the street is 22 m onaverage, with large pavements (5.5e6.0 m) on both sides and thestreet axis clockwise bearing from the north is 153� (Fig. 2).

BTEX passive samplers were placed on lampposts and buildingfacades, as shown in Fig. 2. Location SR(A) corresponds to triplicateBTEX samplers co-located with an automatic NOx chem-iluminescence analyser. The limit of detection of the diffusion tubeswas 0.2 ppb and accuracy 14.67% (Gradko Environmental, http://www.gradko.co.uk). The NOx analyser recorded hourly mean NOxconcentrations within �5% of uncertainty bounds. An evaluation ofthe passive samplers used in this field campaign has been pre-sented elsewhere (Vardoulakis et al., 2009).

Urban background measurements were collected in CentenarySquare (BC) (around 5 kmnorth-west of SR), using passive samplersand continuous gas analysers. Hourly averaged wind and temper-ature data from the synoptic weather stations situated in Coleshill(around 11 km north-east of SR) and hourly traffic data (density,average speed and fleet composition) from the Birmingham CityCouncil were obtained for the same period.

With the aim of studying the street canyon effect on the spatialdistribution of air pollutants, vertical profiles of pollutant concen-trations were measured in a 2D section of SR, in the vicinity of theautomatic air quality monitoring station, halfway between the twojunctions. Such section can be considered as a low-rise asymmetricstreet canyon, with an aspect ratio of H/W z 0.5.

Fig. 2. Stratford Road geometry and receptor positions. The red squares indicate thepositions of the BTEX samplers. The automatic NOx monitor is indicated with the greenbox on the east pavement. (For interpretation of the references to colour in this figurelegend, the reader is referred to the web version of this article.)

E. Solazzo et al. / Atmospheric Environment 45 (2011) 5230e52395232

2.2. CFD modelling of roadside concentrations

2.2.1. Meteorological dataThe hourly meteorological data collected during the one-year

campaign have been grouped into a selected number of disper-sion scenarios, based on combinations of wind direction and speed(Table 1). Eight wind direction sectors were defined, at equalintervals of 45�. For each wind direction sector, four wind speeds ulwere selected, ranging from 0 to 4 m s�1, with intervals of 1.0 m s�1

(Table 1). A total of thirty-two combinations of wind direction andspeed have been simulated using a CFD model.

The number of scenarios selected represents the optimumamount of model runs for obtaining annual average benzeneconcentration that fall within 50% of the observed values (as dis-cussed in detail in section 2.2.4). In Fig. 3 the annually averagedwind distribution is presented for the wind ranges of Table 1,showing high frequency of south and south-west components(WD5eWD7) (15e20%). The recorded wind speed was mostfrequently (w75%) within the ranges of u2 and u3.

2.2.2. Inlet boundary condition of wind flowIn this study, the standard k-3 turbulence model provided by the

commercial code FLUENT (www.FLUENT.com), was used. Themodel was extensively evaluated against wind tunnel data forwind, concentration and turbulence, with and without additionalturbulence generated by vehicular traffic (Solazzo et al., 2008a and2008b). The evaluation was based on the recommendations of theCOST-732 Action for quality assurance and improvement of micro-scale meteorological models (http://www.mi.uni-hamburg.de/Home.484.0.html).

To prepare the input for the CFD analysis of SR, simulations wereconducted to generate an inflow profile mimicking those devel-oping over real urban canopies. A 3D domain was created (a cross-section is shown in Fig. 4) with two simplified buildings creating an

Table 1Wind direction and speed ranges adopted for analysis.

Wind direction sectors Wind speed ranges

337.5� � WD1 < 22.5�

22.5� � WD2 < 67.5�

67.5� � WD3 < 112.5� 0.0 � u1 < 1.0 m s�1

112.5� � WD4 < 157.5� 1.0 � u2 < 2.0 m s�1

157.5� � WD5 < 202.5� 2.0 � u3 < 3.0 m s�1

202.5� � WD6 < 247.5� 3.0 � u4 � 4.0 m s�1

247.5� � WD7 < 292.5�

292.5� � WD8 < 337.5�

isolated street canyon. The advantage of this approach is that byimposing periodic boundary condition to the side planes of thedomain, at solution convergence, the mean wind and turbulencehave adjusted to an underlying roughness of a large number ofstreet canyons of similar dimensions. Four simulations were run,corresponding to the input wind speed ranges ul (Table 1) and forwind direction perpendicular to the street canyon axis. Eachsimulation was initialised with a uniform wind flow (the meanvalue of each wind speed range was selected to initialise each run),turbulent kinetic energy k, and dissipation profiles 3 of the form:

u ¼ ulk ¼ 0:1u2l3 ¼ C0:75

m k0:5l ðkzÞ�1(1)

with l ¼ 1e4, z the height (in m) above the ground, and k ¼ 0.4 andCm ¼ 0.09 empirical parameters. Turbulent kinetic energy intensityk was set to 10% of ul

2. Wind velocity and turbulence profilesobtained with this procedure in the middle cross-section of thecanyon were exported and used as inlet boundary conditions forthe simulation of pollutant dispersion in SR under different winddirections.

For a simple verification of the modelling results, the verticalprofiles of mean wind velocity u(z) have been interpolated witha logarithmic equation of the type:

uðzÞu*

¼ 1kln�z� dz0

�(2)

where d and z0 (the zero-plane displacement height and theaerodynamic roughness length respectively) have been set equal to0.6H and 0.3H, respectively (Solazzo et al., 2010). The frictionvelocity u* in Eq. (2) is thus the only fitting parameter assumed totake the following values: 0.035, 0.105, 0.20, and 0.30 m s�1 forinput wind velocity in the range of ul. The ratio ul=u* falls within therange of 7e13, which is in perfect agreement with the results of DiSabatino et al. (2008) for the flow over urban residential neigh-bourhoods, such as SR. This simple check ensures that realisticvalues of mean wind profiles are obtained with this procedure.

The preparation of the inlet boundary conditions of wind flowcould be omitted by directly running CFD for SR with the inletprofiles of Eq. (1), as it is usually done in practice. The procedureadopted in this study, however, offers input wind conditions thatare more representative of densely built urban areas.

2.2.3. Computational settingsThe main CFD analysis was carried out by creating the compu-

tational domain of Fig. 5. The full 3D geometry of SRwas reproducedon a circular horizontal disk with a radius of 150m. The disk rotateswith respect to the inflow plane and the rest of the modellingdomain,which are fixed. For each simulation, the rotation anglewasselected according to the input wind direction of Table 1. The inflowplane was at a distance of 5H from the rotating disk (H ¼ 12 m).Velocity profiles from analysis discussed in the previous sectionwere set at the inflow plane. The outflow plane is placed at 15Hdownstream the disk and the top of the domain at 10H. The distancebetween the lateral planes and the buildings depends on the anglethe disk is rotated with respect to the inflow. For perpendicular andparallel wind flow, the distance is at the minimum (2.5H) andmaximum (7H), respectively. Sensitivity analysis ensured that theflow and concentration fields at the section under study (ata distance of 6H from the edges of the disk) were not affected by theflow near the lateral planes (Solazzo, 2009).

The grid of the domainwas selected based on a sensitivity test oneight types of unstructured grids (non-aligned tetrahedralelements), with increasing resolution (from coarse to fine). The aim

Fig. 3. Annually averaged hourly wind direction (a) and velocity (b) for the ranges of Table 1.

E. Solazzo et al. / Atmospheric Environment 45 (2011) 5230e5239 5233

of the tests was to identifying the conditions for which the meanwind velocity and turbulent kinetic energy at the outflow planeremained unvaried for further grid refinements. The grid fulfillingthis condition was generated by stretching the cells size fromaminimum size of 0.04H at street level up to 0.5H above the roof, fora total of 2.5�106 cells in the street canyon. The grid for the volumeof the circular plane was generated by extrusion of the disk mesh,creating a cylinder entirely contained in the computational domain.The regions of the domain outside of the cylinderweremeshedwithtetrahedral cells with averaged size of 0.13H, for a total of 3.5 � 106

cells. The CFD k-3 model required about 11 h on a single CPUworkstation to finalise the run with this mesh structure.

Vehicle emissions were simulated by means of two parallel linesources at street level placed at equal distance of 0.75H from theflanking buildings. The boundary condition for the sources was setas a velocity inlet of a passive tracer. A default turbulent Schmidtnumber of 0.7 was retained (Riddle et al., 2004; Di Sabatino et al.,2007). For each wind scenario (wind direction and speed), a CFDsimulation was run with an arbitrary emission rate, assuming thatthe concentration field depends linearly on the emission rate.Therefore, for the 32wind scenarios, emissionwas set to 0.01 kg s�1

Fig. 4. CFD model setting and boundary conditions for the computational domain ofStratford Road.

and then concentrations were obtained by linearly scaling with theappropriate emission factor, as detailed in Section 2.2.4.

Traffic produced-turbulence (TPT) effects were taken intoaccount by adding an extra turbulent kinetic energy term, s2w0,acting in the bottom half of the canyon:

s2w0 ¼ bV2 ¼ K1ð1� expð � nv=K2ÞÞV2 (3)

where K1 ¼ 0.0012, K2 ¼ 0.031 are empirical parameters, nv is thedensity of vehicles per unit length of the street (m�1), and V is themean vehicle speed (m s�1). Equation (3) was derived by Solazzo(2009) based on CFD analysis of TPT in a street canyon.

2.2.4. Comparison between CFD and scenario averagedconcentrations

CFD model concentrations have been compared with the hourlyaveraged NOx measurements recorded at the continuous moni-toring station (location SR(A)) on the east side of SR, at z/H ¼ 0.25.NOx concentrations have been normalised according to:

Ci ¼�ci � Cb;i

�Ei

sw;i L H (4)

Fig. 5. Computational domain used for the simulation of the wind scenarios of Table 1.

E. Solazzo et al. / Atmospheric Environment 45 (2011) 5230e52395234

where index i represents the i-th scenario, ci is the modelled (orobserved) concentration, L the length of the line sources, and

sw;i ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiatu2i þ bV2

i

q(5)

is a composite velocity scaling (Kastner-Klein et al., 2003). at is thewind induced turbulence scaling factor, set equal to 0.27 (Solazzo,2009), whereas the TPT component is that of Eq. (3), Cb,i is thebackground concentration of NOx. Ei is the vehicle emission rateused to scale the results of CFD simulations, calculated with theEFT2 toolkit (UK national Air Quality Archive, http://www.airquality.co.uk/archive/). Average NOx emission factors of0.47 g veh km�1 and of 10.31 g veh km�1 were calculated for smallvehicles (cars and LGV) and large vehicles (HDV and buses)respectively, using observed data of vehicles fleet composition andaverage traffic speed in SR. An averaged speed of 20 km h�1 and15 km h�1 for small and large vehicles, respectively, was assumed.

To derive the mean observed NOx data for the i-th scenario, wefirstly identified all hours at which wind conditions fall into the i-thscenario and then averaged the NOx data over these hours. Wedefine fT,i as the hourly frequency of a given i-th scenario during theone-year sampling campaign. The mean concentration for eachscenario, Ci, is:

Ci ¼1fT ;i

XfT ;ik¼1

Ck;i (6)

The standard deviation associated with each scenario is sdCi.

Results of Ci, sdCi, and fT,i for all i ¼ 1,.,32 wind scenarios are

reported in Table 2. Comparisons between scaled modelled andmeasured NOx concentration are reported in Fig. 6 for the eightwind directions of Table 1 (discussed in Section 3.1).

Further to the comparison between CFD and hourly NOx, a novelmethodology for comparing CFD with monthly averagedmeasurements of roadside concentrations was developed.Measurements of monthly averaged benzene concentrationsgathered in SR with passive samplers were used for this. It shouldbe noted that the passive samplers provide monthly averagedmeasurements of benzene concentrations, whilst meteorologicaland traffic data used as model inputs are hourly averages. One ofthe main advantages of this methodology is that it is not dependanton the chemical species emitted. Since CFD has been run fora passive tracer with an arbitrary emission rate, it is possible toconvert the concentration profile for each wind scenario into anynon-reactive compound, providing the appropriate emissionfactors. The methodology is thus developed as follows:

- NOx profiles of Fig. 6 have been converted into benzeneprofiles. Since total NOx and benzene can be considered as inertgases for short atmospheric transport distances and molecularviscosity is neglected, their dispersion is solely governed byatmospheric turbulence, therefore the model results obtained

Table 2Mean normalised concentrations and standard deviation of the mean for the eight WD a

WD (�) u1 u2

C sdC fT C sdCWD1 337.5e22.5 9.2 0.9 305 22.0 2.0WD2 22.5e67.5 10.0 1.4 91 19.7 1.5WD3 67.5e112.5 9.0 0.9 83 16.2 1.2WD4 112.5e157.5 10.7 1.4 63 18.2 1.5WD5 157.5e202.5 7.8 0.9 148 8.9 1.1WD6 202.5e247.5 6.0 0.9 118 5.1 0.9WD7 247.5e292.5 7.8 1.4 93 13.0 1.2WD8 292.5e337.5 11.2 1.1 95 24.9 1.4

for NOx can be converted into benzene concentrations.Benzene emission factors of 0.008 g veh km�1 (small vehicles)and of 0.001 g veh km�1 (large vehicles) were calculated usingthe EFT2 toolkit, assuming an averaged speed of 20 km h�1 and15 km h�1 for small and large vehicles, respectively. Based onthe traffic flow recorded during the sampling campaign,benzene emissions ranged between 0.5 and 4.5 g km�1 h�1.Four emission intervals emi1, emi2, emi3, and emi4 weregenerated from this range (Table 3), corresponding to light,intermediate, dense and congested traffic conditions.

- The cumulative hourly frequency of wind scenarios fT,i andemissions femi of each range were calculated for each month.

- A matrix MC of concentration “footprints” obtained from theCFD calculations (at each receptor) was then created.MC has 16rows (4 wind speeds by 4 emission rates) and 8 columns (winddirections). Similarly, Mf is the frequency matrix, havingdimensions 8 � 16, and whose components are the hourlyfrequency of each combination of observed wind direction,wind speed and emission rate. Mf weights the contribution ofthe concentration components in MC with their frequency.Then, for each receptor the monthly averaged concentrationwas calculated as:

Crec ¼X8k

1fT;k

X16j

hMCðk; jÞ$Mfðj; kÞ

i(7)

where rec indicates the sampling location (FL, F1st, F2nd, DL, DW, D1st,SR(A)).

-The benzene monthly averaged (Fig. 6) and annually averaged(Fig. 8) CFD profiles were calculated based on Eq. (7).

Application of Eq. (7) to model NOx was not needed as both themeteorological input and NOx monitoring data had the same timeresolution, but this method can be applied when combining datawith different time resolutions.

It should be noted that the number of wind scenarios wasempirically determined by starting with an arbitrary number ofeighteen scenarios (six wind directions and threewind speeds) andincreasing it until the annual mean concentration of benzene waswithin 50% of the observed values, as required by the EU Directive2008/50/EC for annual modelled benzene concentrations.

3. Results and discussion

3.1. Comparing CFD model results against hourly averaged NOx

concentrations

The comparison between scaled CFD and observed NOxconcentrations is reported in Fig. 6. Given the orientation of thestreet axis and the position of the monitoring station on the east

nd four wind velocities analysed. fT for each dispersion scenario.

u3 u4

fT C sdC fT C sdC fT

538 34.7 2.9 299 40.6 3.9 74307 29.7 2.0 189 32.2 2.3 122305 25.7 2.5 111 30.5 2.0 25262 21.3 1.8 147 22.4 1.9 67775 4.8 1.1 691 5.3 1.6 439560 3.4 1.0 403 3.3 1.1 119419 13.6 1.7 356 16.9 1.9 154457 32.6 2.4 398 39.3 2.0 120

Fig. 6. Comparison between scaled CFD (lines) and observed NOx (symbols) concentrations for the eight analysed wind directions (WD).

E. Solazzo et al. / Atmospheric Environment 45 (2011) 5230e5239 5235

side of the street, WD1, WD2, and WD3 can be regarded as leewardcases. For these cases, the comparison (Fig. 6a, b, c) is overallsatisfactory, although certain model overestimations (WD1, u1 andu2, w50% and 16% respectively, calculated as the ratio between thedifference of observed and modelled values and the observedvalue) and underestimations (WD3, u3, w20%) can be observed.However, the comparison is satisfactory forWD1 u3 and u4 (Fig. 6a),WD2 u2 and u3 (Fig. 6b), andWD3 u1, u2 and u4 (Fig. 6c). WD4 can beconsidered as a “quasi-parallel” case, with the wind flow coming

Table 3Ranges of benzene emissionrates for Stratford Road.

Benzene emission ranges

0.5 < emi1 < 1.5 g km�1 h�1

1.5 < emi2 < 2.5 g km�1 h�1

2.5 < emi3 < 3.5 g km�1 h�1

3.5 < emi4 < 4.5 g km�1 h�1

from the south and thus almost alignedwith the street axis. Also forthis case the comparison is overall satisfactory although the modelslightly underestimates the cases u2 (w20%) and u3 (w16%)(Fig. 6d). The case WD8 is for prevailing winds from the Northdirection, and can also be considered “quasi-parallel”. Because ofthe orientation of the street not being exactly aligned with thenorth-south direction, the wind component normal to the streetaxis is larger for WD8 than for WD4, hence the larger modelledconcentrations (Fig. 6h). ForWD8, the comparison is satisfactory foru4, but the model underestimates the concentrations for u1, u2 andu3 (w25%, 30%, 13% respectively).

Cases WD5, WD6 and WD7 correspond to “windward” condi-tions. For wind coming from south-west (WD5 and WD6), CFDsignificantly overestimates the observed concentrations (Fig. 6e, f).However, for wind coming from north-west (WD7) comparison ismuch improved (Fig. 6g). For WD5 and WD6 the observed scaledconcentrations are clustered together towards low values(between 0 and 10). This trend is not well replicated by the CFD

E. Solazzo et al. / Atmospheric Environment 45 (2011) 5230e52395236

profiles, which predict larger values (up to a factor 4). Thissubstantial model overestimation may be explained by the wellknown tendency of the standard k-3 model to overestimatewindward concentrations, due to the large diffusive component inthe transport equation for k (Cheng et al., 2003; Lien and Yee,2004; Hargreaves and Wright, 2006). Plots of u-wind componentfor the case WD6 (windward) and case WD2 (leeward) are shownin Fig. 7a and b, respectively. In the case of WD6, the upwindbuilding is higher than the downwind one, while the oppositehappens for WD2. The roofs are pitched, with a slope of 45�. Theflow separates on the upwind roof and reattaches on the down-wind roof, inducing a vortical motion in the canyon below, as ithas been observed in regular canyons with H/W z 0.5 (e.g.Kastner-Klein et al., 2004). The shape of the roofs and the differ-ential building height perturb only slightly, and locally, thispattern because of the low roof elevation compared to H (tentimes smaller), and the small differential building heightcompared with W (eleven times smaller). Transport of k (Fig. 7a)and mass (Fig. 7b) are also indicated in Fig. 7. This analysisprovides an example of the additional value (flow, turbulence andconcentration fields) the proposed CFD methodology can providewith respect to parameterised atmospheric dispersion models.

3.2. Comparing CFD model results against monthly and annuallyaveraged benzene concentrations

In this Section, the methodology based on Eq. (7) for comparingCFD and monthly averaged benzene concentrations is evaluated.CFD results are compared against passive sampling measurementsin Figs. 8 and 9. It should be noted that benzene backgroundconcentrations have been subtracted from the monitoring data inFig. 8. The statistical indicators R (correlation coefficient, whichindicates the degree of association between the observed and themodelled concentrations), and FB (mean fractional bias, whichindicates the tendency of the model to over-predict (FB < 0) orunder-predict (FB> 0) the observed concentrations) are reported inTable 4 to assist interpretation of the results in Fig. 8 (monthlycomparison) and Fig. 10 (annual comparison).

Measurements in Fig. 8 show a marked monthly variability,which is more pronounced at street level (receptors DL and FL),whilst at higher receptor positions fluctuations are smaller. Thismay be due to localised effects at street level where turbulence ismore intense (TPT and aerodynamic drag effect). As expected, atpedestrian level (1.5e2.5 m) and closer to the street (receptorsDL and FL), observed concentrations are higher than at thereceptor mounted on the building walls (DW, F1st and D1st, F2nd),as a direct consequence of the proximity to the vehicular emis-sion sources.

The CFD modelled concentrations show a temporal trend qual-itatively similar to the observations, although quantitatively CFDtends to underestimate the observed benzene concentrations (FBpositive at all positions in Table 4). Generally, the agreement

Fig. 7. Vectors of u-velocity component overlaid to contours of (a) turbulent kinetic energyblack arrow indicates the free flow wind direction). (For interpretation of the references to

between CFD and diffusion tubes at lower receptors on the east sideof the street (Fig. 8b) is improved compared to the correspondingreceptors on the west side (RF,A > RE, Table 4) (Fig. 8a).

On the west side of the street, the model predicts one mainconcentration peak in November, at all receptor heights, withsecondary peaks in April (receptor DL) and May (receptors DW andD1st). On the east side, two main peaks were predicted by themodel, in November and January at receptors FL and F1st, withsecondary peaks in April (FL) and March (F1st). CFD predicts lowerand less fluctuating concentrations (C ¼ 0.009) at the highestposition F2nd, which is in good agreement with the observations.The occurrence of the predicted benzene peaks is determined bythe different monthly frequency of wind direction and speedcombinations. November is characterised by prevailing winds fromthe south, and high frequency of wind speeds below 2 m s�1

(w50%) (Fig. 2d) which explains the high concentrations observedin this month. In December the frequency of wind speeds below2 m s�1 is about 10% lower with respect to November, whilst the u3component is 7% more frequent, leading to lower measuredconcentrations than November. In particular, observed concentra-tions at the receptor DW on the west side and all the values on theeast side are significantly lower in December (Fig. 8a). CFD profilesalso predict lower concentrations in December compared toNovember, indicating the importance of including several windvelocity classes in the modelling methodology based on Eq. (7).

The CFD model underestimates the observed concentrations onboth sides of the street, including position SR(A), and more mark-edly during winter. Although CFD results for receptors DL and FLpredict the highest monthly concentrations during the one-yearcampaign, the magnitude of the monthly peaks is w80% smallerthan the measurements. This could be due to benzene concentra-tions being higher during winter months in urban areas (Hellénet al., 2005) because of more persistently stable atmosphericconditions and higher emission levels. To test this hypothesis,analysis of atmospheric stability over the Birmingham area for theinvestigated period was conducted. Results, in terms of hourlyoccurrence of PasquilleGifford classes (Pasquill and Smith, 1983),are reported in Fig. 9. Overall, neutral to slightly stable atmosphericconditions (D and E classes) are the most frequent, with very stableconditions (F and G) more persistent during cold months(NovembereJanuary). In particular, occurrence of class D decreasesin winter, being replaced by the more stable class E. As expected,the maximum occurrence of stable conditions is reached for thecold months between November and February. Results for thesummer months, with dominant E and D classes, are in agreementwith findings by Tomlinson et al. (2010) for the same region. Thisanalysis indicates that relatively poor agreement betweenmodelled and observed benzene concentrations during coldmonths (Fig. 8) may be due to the lack of atmospheric stabilityparameterisation in the model simulations.

In Fig. 10, the comparison between CFD and measurements isshown for the annual averaged benzene concentrations.

and (b) concentration of NOx (brighter colours corresponds to higher magnitude; thecolour in this figure legend, the reader is referred to the web version of this article.)

Fig. 8. Comparison between time series of measured benzene concentrations at three receptor locations (symbols) and modelled CFD concentration (lines) (a) on the west (b) andeast side of the canyon.

E. Solazzo et al. / Atmospheric Environment 45 (2011) 5230e5239 5237

Measurements gathered at each receptor during the one-yearsampling campaign were averaged and compared against theCFD concentration profiles. The annual averaged hourly meteoro-logical data of Fig. 3 were used for this comparison. Gaps in theobservations have been replaced with the average between thetwo nearest (in time) available measurements. It is importantnoticing that the CFD profiles in Fig. 10 have been calculated withEq. (7), and thus are not just an average of the monthly resultspresented in Fig. 8. Given the prevalence of the wind directioncomponent from the west (Fig. 2a), higher concentrations areexpected on the west side (leeward) of the canyon. This isconfirmed by the averaged observations, which showed higherconcentrations at receptors DL and DW (Fig. 10a) compared withreceptors FL and F1st (Fig. 8b).

CFD profiles compare satisfactorily with the annual averagedbenzene measurements on the east side of the street at receptorsSR(A), FL, and F2nd (Fig. 10b), and on the west side at receptor DL(Fig. 10a). At the higher receptors F1st and DW the model underes-timates the measurements, repeating the trend observed in themonthly comparison (Fig. 8b), although the agreement is satisfac-tory at receptor F2nd. The comparison between observations andCFD at the lower receptors of both sides (SR(A), FL, DL) is signifi-cantly better than at the higher receptors F1st and DW.

3.3. Applicability and limitations of the methodology

The results presented in the previous sections show how theproposed methodology can be applied to the examination of long-

Fig. 9. Frequency of hourly occurrence of the PasquilleGifford stability categories forthe Birmingham Area between July 2005 and August 2006.

term pollutant concentrations in a street canyon and verificationof compliance against regulatory standards. Given the availabilityof CFD outcomes, such as 3D fields of wind, temperature,concentration, etc., under different scenarios, the methodologyproposed by the present study can readily be applied to otherstreet canyon sites as well as other applications including pop-ulation exposure, ventilation of buildings, pedestrian comfort, andurban planning.

An operational implementation of themethodology requires thefollowing steps:

1. identifying typical dispersion scenarios based on wind direc-tion, wind speed, traffic volume, atmospheric stability, andpollutant species;

2. creating a database of CFD-calculated concentration fields forthe identified dispersion scenarios at specific locations;

3. combining the CFD results based on the frequency of occur-rence of each dispersion scenario during the examined period;and

4. comparing the model output against the regulatory target.

Given the nature of the averaging procedure embedded in Step 3above, the developed methodology might miss peak values forparticular scenarios. This aspect is of direct relevance when inter-preting the results. For example, neglecting atmospheric stability inthe presented study led to underestimate concentration duringwinter months. This is the main shortcoming of the analysis in thispaper, whichmight be overcome by running some initial sensitivitytests aimed at carefully identifying the number and the nature ofthe dispersion scenarios.

Table 4Statistical model performance indicators (R: correlation coefficient; FB: Fractionalbias) for different receptor locations.

Receptor Monthly Annual

R FB R FB

DL 0.56 0.11 n.a.b 0.032b

DW 0.67 0.06D1st 0.89a 0.08a

FL 0.76 0.17 0.90 0.020F1st 0.40 0.14F2nd 0.16 0.07SR(A) 0.81 0.10

a Based on three data.b Based on two data.

Fig. 10. Comparison between yearly averaged monthly measurements of benzene (black crosses) and CFD results (red squares) on the (a) east, and (b) west side of Stratford Road.(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

E. Solazzo et al. / Atmospheric Environment 45 (2011) 5230e52395238

4. Conclusions

The aim of this study was to explore the operational use of CFDfor interpreting long- and short-term measurements of atmo-spheric pollutants in urban areas. At present, the use of CFD in thistype of applications is limited due to the long simulation time andcomputational costs involved. We propose a methodology thatmoves towards overcoming this limitation by developing a data-base of normalised concentrations computed with a CFD model fora limited number of user-defined dispersion scenarios. We identi-fied the wind conditions as the relevant variables to be taken intoaccount. CFD simulations for a total of thirty-two dispersionscenarios were carried out. The following step was to classify thefrequency of meteorological and emission inputs based on theranges of scenarios identified. The result of this operation isa frequency matrix, whose elements are the frequency of occur-rence of each dispersion scenario during the investigated period.Finally, time averaged concentrations of a pollutant at a givenreceptor point were calculated by linearly combining the corre-sponding concentration data (scaled for a non-reactive pollutantwith the appropriate emission factors) with the frequency matrix.

This methodology has the following potentialities: (a) it allowsto calculate concentrations of different non-reactive pollutant byrescaling the concentration dataset with the appropriate emissionfactors; (b) pollutant concentrations for a different time period atthe same locations can be calculated by modifying the frequencymatrix according to the new meteorological data, without addi-tional CFD simulations; (c) the methodology does not require thewind direction sectors to be equally spaced, or the wind speed tovary within uniform ranges (specific wind direction/speed sectorscan have a finer resolution in the frequency matrix); (d) it providesthree-dimensional flow, turbulence, and concentration fields forthe whole modelling domain. This gives the opportunity to closelyinvestigate localised effects due to peculiar street geometries thatcannot be studied with simpler operational (e.g. Gaussian) models;(e) the methodology presented here could be extended for appli-cation in diagnostic analysis, that is quantifying the variation inpollutant concentrations for varying emissions under differentwind scenarios.

This methodology was tested against hourly averaged NOxmeasurements from an automatic roadside monitoring station aswell as against monthly averaged benzene profiles obtained withpassive sampling in an urban street canyon. We focused ourattention on the simplified case of neutral atmospheric stability

conditions and non-reactive pollutants. Under these conditions,significant model underestimations were observed during wintermonths, especially at street level, on both sides of the street canyon.This was mainly associated with the lack of atmospheric stabilityformulations in the CFD modelling methodology, although someunderestimation of benzene emissions during winter cannot beexcluded. It should be also noted that a limited number of windscenarios was used to test this methodology in the present study.Future work needs to be devoted to incorporating the effects ofatmospheric stability and testing this methodology for morerefined wind scenarios in a wider range of urban locations.

Acknowledgements

The passive sampling campaign was sponsored by BirminghamCity Council, UK. The Environmental Protection team of Birming-ham City Council is gratefully acknowledged for technical assis-tance throughout the survey.

References

Baik, J.-J., Kang, Y.-S., Kim, J.-J., 2007. Modeling reactive pollutant dispersion in anurban street canyon. Atmospheric Environment 41, 934e989.

Blocken, B., Stathopoulos, T., Saathoff, P., Wang, X., 2008. Numerical evaluation ofpollutant dispersion in the built environment: comparisons between modelsand experiments. Journal of Wind Engineering and Industrial Aerodynamics 96,1817e1831.

Buccolieri, R., Di Sabatino, S., 2011. MUST experiment simulations using CFD andintegral models. International Journal of Environment and Pollution 44,376e384.

Buccolieri, R., Sandberg, M., Di Sabatino, S., 2010. City breathability and its link topollutant concentration distribution within urban-like geometries. Atmo-spheric Environment 44, 1894e1903.

Cai, X., Barlow, J.F., Belcher, S.E., 2008. Dispersion and transfer of passive pollutantsin and above street canyons e large-eddy simulations. Atmospheric Environ-ment 42, 5885e5895.

Coceal, O., Dobre, A., Thomas, T.G., 2007. Unsteady dynamics and organized struc-tures from DNS over an idealized building canopy. International Journal ofClimatology 27, 1943e1953.

Cheng, Y., Lien, F.S., Yee, E., Sinclair, R., 2003. A comparison of large eddy simulationswith a standard k-3 Reynolds-Averaged Navier-Stokes model for the predictionof fully developed turbulent flow over a matrix of cubes. Journal of WindEngineering and Industrial Aerodynamics 91, 1301e1328.

Di Sabatino, S., Buccolieri, R., Pulvirenti, B., Britter, R., 2007. Simulations of pollutantdispersion within idealised urban-type geometries with CFD and integralmodels. Atmospheric Environment 41, 8316e8329.

Di Sabatino, S., Solazzo, E., Paradisi, P., Britter, R., 2008. A simple model for spatially-averaged wind profiles within and above an urban canopy. Boundary-LayerMeteorology 127, 131e151.

E. Solazzo et al. / Atmospheric Environment 45 (2011) 5230e5239 5239

Fisher, B., Chemel, C., Hu, R.-M., Sokhi, R., 2010. In: Steyn, D.G., Rao, S.T. (Eds.),Emulating Complex Calculations for Regulatory Decisions. Air PollutionModelling and Its Application XX. Springer, pp. 195e199.

Hargreaves, D., Wright, N., 2006. On the use of k-3 model in commercial CFDsoftware to model the neutral atmospheric boundary layer. Journal of WindEngineering and Industrial Aerodynamics 95, 355e369.

Harrison, R., Stedman, J., Derwent, D., 2008. New directions: why are PM10concentrations in Europe not falling? Atmospheric Environment 42,603e606.

Hellén, H., Kukkonen, J., Kauhaniemi, M., Hakola, H., Laurila, T., Pietarila, H., 2005.Evaluation of atmospheric benzene concentrations in the Helsinki MetropolitanArea in 2000e2003 using diffusive sampling and atmospheric dispersionmodelling. Atmospheric Environment 39, 4003e4014.

Holmes, N.S., Morawska, L., 2006. A review of dispersion modelling and its appli-cation to the dispersion of particles: an overview of different dispersion modelsavailable. Atmospheric Environment 40, 5902e5928.

Kastner-Klein, P., Berkowicz, R., Britter, R., 2004. The influence of street architectureon flow and dispersion in street canyons. Meteorology and Atmospheric Physics87, 121e131.

Kastner-Klein, P., Fedorovich, E., Ketzel, M., Berkowicz, R., Britter, R., 2003. Themodelling of turbulence from traffic in urban dispersion models e part II:evaluation against laboratory and full-scale concentration measurements instreet canyons. Environmental Fluid Mechanics 3, 145e172.

Lien, F.-S., Yee, E., 2004. Numerical modelling of the turbulent flow developingwithin and over a 3-D building array, part I: a high resolution Reynolds-averaged NaviereStokes approach. Boundary Layer Meteorology 112, 427e466.

Louka, P., Vachon, G., Sini, J.-F., Mestayer, P.G., Rosant, J.-M., 2002. Thermal effects onthe airflow in a street canyon e Nantes ’99 experimental results and modelsimulations. Water, Air and Soil Pollution 14, 351e364.

Murena, F., Garofalo, N., Favale, G., 2008. Monitoring CO concentration at leewardand windward sides in a deep street canyon. Atmospheric Environment 42,8204e8210.

Pasquill, F., Smith, F.B., 1983. Atmospheric Diffusion, third ed. Ellis Horwood Limited,Chichester.

Riddle, A., Carruthers, D., Sharpe, A., Mchugh, C., Stocker, J., 2004. Comparisonbetween FLUENT and ADMS for atmospheric dispersion modelling. Atmo-spheric Environment 38, 1029e1038.

Solazzo, E., Cai, X., Vardoulakis, S., 2008a. Modelling wind flow and vehicle inducedturbulence in urban streets. Atmospheric Environment 42, 4918e4931.

Solazzo, E., Cai, X., Vardoulakis, S., 2008b. Improved parameterisation for the windflow and turbulence modelling within an urban street canyon. EnvironmentalModelling and Software 24, 381e388.

Solazzo 2009. Modelling dispersion of air pollutants in urban street canyons. PhDThesis, University of Birmingham (UK).

Solazzo, E., Di Sabatino, S., Aquilina, N., Dudek, A., Britter, R., 2010. Couplingmesoscale modelling with a simple urban model. The Lisbon case study.Boundary-Layer Meteorology 137, 441e457.

Tomlinson, C.J., Chapman, L., Thornes, J.E., Baker, C.J., 2010. Derivation of Birming-ham’s summer surface urban heat island from MODIS satellite images. Inter-national Journal of Climatology. doi:10.1002/joc.2261.

Vardoulakis, S., Fisher, B.E.A., Pericleous, K., Gonzalez-Flesca, N., 2003. Modelling airquality in street canyons: a review. Atmospheric Environment 37, 155e182.

Vardoulakis, S., Lumbreras, J., Solazzo, E., 2009. Comparative evaluation of NO2, NOxand O3 passive diffusion tubes for exposure studies. Atmospheric Environment43, 2509e2517.

Vardoulakis, S., Dimitrova, R., Richards, K., Hamlyn, D., Camilleri, G., Weeks, M.,Sini, J.-F., Britter, R., Borrego, C., Schatzmann, M., Moussiopoulos, N., 2011a.Numerical model inter-comparison for wind flow and turbulence around singleblock buildings. Environmental Modelling & Assessment 16, 169e181.

Vardoulakis, S., Solazzo, E., Lumbreras, J., 2011b. Intra-urban and street scale vari-ability of BTEX, NO2 and O3 in Birmingham, UK: implications for exposureassessment. Atmospheric Environment (in review).

Vardoulakis, S., Valiantis, M., Milner, J., ApSimon, H., 2007. Operational air pollutionmodelling in the UK e street canyon applications and challenges. AtmosphericEnvironment 41, 4622e4637.

Woodfield, N.K., Longhurst, J.W.S., Beattie, C.I., Chatterton, T., Laxen, D.P.H., 2006.Regional collaborative urban air quality management: case studies across GreatBritain. Environmental Modelling & Software 21, 595e599.