Air pollution exposure monitoring and estimation Part II.‡ Model evaluation and population exposure† Sam-Erik Walker,* Leiv H. Slørdal, Cristina Guerreiro, Frederick Gram and Knut E. Grønskei Norwegian Institute for Air Research, PO Box 100, N-2027 Kjeller, Norway. E-mail: samerik@nilu.no; Fax:+47 63 89 80 50 Received 7th April 1999, Accepted 25th June 1999 The air pollution dispersion model EPISODE has been developed at the Norwegian Institute for Air Research (NILU) over the past several years in order to meet the needs of modern air quality management work in urban areas.The model has recently been used as a basis for exposure calculations of NOx and NO2 in order to assess the eVects of diVerent traYc diversion measures on health and well being for the residents in the Va° lerenga–Ekeberg– Gamlebyen area in Oslo.Here we describe some results from the most recent evaluations of the model for NOx and NO2 at station Nordahl Brunsgate in Oslo for the period 1 October 1996–19 November 1996. In addition examples of population exposure calculations for Oslo performed during the winter period of 1995–96, are also presented. using an explicit diVerence scheme.6 Vertical advection and 1.Introduction diVusion are solved simultaneously using a combined scheme As a basis for modern air-quality management work in urban for the vertical exchange of mass depending on divergence of areas, there is a need for urban scale dispersion models capable the horizontal wind and on the growth of Gaussian scheme of estimating time-varying (hourly) concentrations in arbitrary sz values.2 receptor points within the urban area.Such models should be The Lagrangian part of the model consists of diVerent able to specify the time-varying pollutant gradients even close Gaussian subgrid models for each of the three main emission to sources. The air pollution dispersion model EPISODE has categories: area, line and point sources.The purpose of the been developed at NILU over the past several years in order subgrid models is to calculate concentrations with finer resoto meet these needs.1 Earlier versions of the model have been lution near the diVerent individual sources. The subgrid area applied in several places in Norway and elsewhere.2,3 The source model calculates ground level concentrations in diVerent model has recently been used as a basis for exposure calcu- grid cells from area emissions in the grid cell itself using an lations of NOx and NO2 in order to assess the eVects of integrated Gaussian scheme.7 The calculation procedures for diVerent traYc diversion measures on health and well being pollution contributions from line sources and area sources to for the residents in the Va° lerenga–Ekeberg–Gamlebyen area the urban scale grid system are based on horizontal fluxes of in Oslo.4 pollution to the down wind grid cells.This calculation takes In this article we first give a short description of the into account the diVerent emission heights for the two main EPISODE model and its development over recent years. Then area source types in Oslo, domestic heating and small street results of model evaluation by comparing hourly calculated traYc.For traYc the emission height is set equal to 1 m above concentrations of the two compounds NOx and NO2 with the ground. For domestic heating the emission height is set hourly observed values at the centrally placed station Nordahl equal to the building height in the centre of Oslo which is Brunsgate in Oslo (see Fig. 1) are given. This was the only station where measurements of NOx and NO2 were taken during the period of investigation. Finally a description of a population exposure model together with results of population load in Oslo for diVerent model simulations are presented. 2. Description of the EPISODE dispersion model The EPISODE dispersion model is a combined 3D Eulerian– Lagrangian type model for urban and local-to-regional scale applications.1 The Eulerian part of the model is based on the numerical solution of the atmospheric (mass) conservation equation of the pollutant species, using a three-dimensional Eulerian grid.The equation is solved using a time-splitting approach where alternating sequences of advection and diVusion operators are called every other time step.The horizontal advection is solved using Botts positive definite and monotone 4th degree scheme.5 Horizontal diVusion is solved Fig. 1 The Oslo area with the air quality measurement station Nordahl Brunsgate and the meteorology station Valle Hovin. †Presented at AIRMON ’99, Geilo, Norway, February 10–14, 1999. ‡For Part I, see ref. 15. J. Environ. Monit., 1999, 1, 321–326 321estimated to be around 20 m above the ground. The building ent (stability), and wind speed and direction at the centrally placed station Valle Hovin in Oslo (see Fig. 1). The wind and height is also used to calculate the initial vertical dispersion (mixing) of pollution due to building eVects from both dom- stability data at this station are used as input to the diagnostic wind field model MATHEW.9 The hourly wind field data estic heating and traYc.The concentration contributions from most of the traYc sources (roads) are calculated using the produced by this model are then input to the EPISODE dispersion model. subgrid scale line source model in EPISODE. This subgrid model is based on using an integrated steady-state Gaussian Horizontal and vertical turbulence (sv and sw) are calculated for each grid cell using NILU’s meteorological pre-processor plume model from US EPA (HIWAY-2).8 Each line source has a user defined influence zone within which the subgrid line MEPDIM.10 This pre-processor is based on atmospheric similarity theory, using temperature- and wind-profiles together source model is capable of calculating concentrations in individual receptor points.In Oslo, all active line sources (around with surface roughness in order to calculate the above mentioned turbulence parameters.11,12 In addition it also calculates 400) have a common influence distance of 500 m. The concentration contribution in a receptor point is defined by integrated several other meteorological parameters such as mixing height, friction velocities and Monin–Obukhov lengths etc.also used Gaussian dispersion along the road. The EPISODE model also includes a subgrid scale point source model, but the point for dispersion calculations. sources in Oslo are so few, and they contribute very little to the total concentration levels of NOx and NO2, so this part 5. Model evaluation of the model will not be described here.The model also contains a module for calculating photo- The model evaluation was performed by comparing the modelcalculated concentration at the station Nordahl Brunsgate in chemical equilibrium between NO, NO2 and O3 in all grid cells and receptor points on an hourly basis. The calculation Oslo with a corresponding measured (observed) concentration at the same station each hour during the period of investi- of NO2 is based on this balance and on dispersion calculations of NOx and Ox (NO2+O3).gation, October 1–November 19, 1996, a total of 1198 h. The comparison was performed for the two pollutant species NOx For Oslo, a 22 km×18 km grid was defined with horizontal resolution of 1000 m in each direction.Vertically, 3 layers were and NO2. Model evaluation was not performed for PM2.5 or PM10 and could not be performed for O3 since it was not used with resolution 20, 30 and 150 m from the ground upwards. Background concentrations of ozone in Oslo were measured locally. Time series plots of model calculated (predicted) and meas- estimated on an hourly basis, using measured values of ozone at the two regional stations Jeløya and Prestebakke in Norway ured (observed) NOx and NO2 concentrations are shown in Figs. 2 and 3 for a representative part of the investigation (maximum at the two stations for each hour of measurements). Ideally, stations both north of and south of Oslo should have been used in order to give the best estimate for regional ozone concentration for Oslo during the investigation period. However, during the autumn, ozone gradients during northerly wind-transport episodes are generally small in these regions.Therefore no great systematic errors are introduced by using only stations south of Oslo. 3. Emission database The emission database for Oslo contains hourly emission data from the three diVerent source categories, domestic heating (area sources), road traYc (area and line sources), and industry (point sources) for the diVerent types of compounds such as NOx, NO2, and particulate matter PM2.5, PM10, etc.For domestic heating the database consists of a Fig. 2 Observed and model calculated hourly concentrations of NOx 22 km×18 km emission field (grid) with horizontal resolution at station Nordahl Brunsgate in Oslo for a part of the evaluation of 1000 m in either direction.These nominal emission fields period containing the highest concentrations. are then scaled using daily and weekly factors in order to produce actual emission field values on an hourly basis. The scaling procedure also includes a temperature dependent factor. Road traYc is the most important source of NOx and NO2 pollution in Oslo.The emission database for road traYc consists of about 2400 line sources (roads) with individual characteristics for each road such as: start and end position (co-ordinates), the total width and slope (elevation), the speed limit, types of vehicles, daily and weekly variations in traYc intensity, and the amount of heavy duty traYc. No seasonal adjustments to the traYc emissions have been included.The individual road data is input to a separate emission module for calculating emissions of NOx and NO2 on each individual road. The emission database currently does not contain any volatile organic compounds (VOCs). 4. Meteorological database Fig. 3 Observed and model calculated hourly concentrations of NO2 The meteorological database for Oslo consists of hourly meas- at station Nordahl Brunsgate in Oslo for a part of the evaluation period containing the highest concentrations.ured values of air temperature, vertical air temperature gradi- 322 J. Environ. Monit., 1999, 1, 321–326period (November 9–November 19, 1996). As can be seen observed and model-calculated values for NOx and NO2 was found to be reasonable high, 0.73 for NOx and 0.68 for NO2.from the figures, the observed and predicted concentrations corresponded fairly well for most of the period. Sensitivity This is also reflected in the index of agreement, being 0.85 and 0.78 respectively for NOx and NO2. analyses have indicated that the diVerences between observed and predicted concentrations, to a large extent can be explained The correspondence is also quite good when comparing mean values, 89 mg m-3 (observed) and 83 mg m-3 (predicted) by the non-representativeness of using only the observed thermal stability DT at the station Valle Hovin (which lies for NOx, and 39 mg m-3 (observed) and 36 mg m-3 (predicted) for NO2.For NOx the maximum concentration observed outside of the city centre), as a measure of the stability conditions at Nordahl Brunsgate.The build-up and weakening (1249 mg m-3) is somewhat higher than the maximum concentration calculated (863 mg m-3), while for NO2 the two num- of pollution concentrations in the central parts of Oslo is a result of a more complex process where mechanical turbulence, bers show a very good correspondence, 121 and 123 mg m-3 respectively.For both components the bias calculated by locally generated heat from buildings, and local turbulence from road traYc, influences the local wind- and turbulence- taking the diVerence in predicted and observed mean values divided by the observed mean value are also quite small being fields and the vertical exchange of pollution. In addition to the time series plots, a set of model evaluation 0.064 and 0.068 respectively for NOx and NO2.The correspondence is also quite good when considering the parameters have also been calculated for the whole period of investigation (October 1–November 19, 1996) (1198 h). These systematic (RMSEs) and unsystematic (RMSEu) portions of the total root mean square error RMSE. For a good model parameters are shown in Table 1.A description of these parameters with their definitions is given in the Appendix the unsystematic portion of the RMSE is much larger than the systematic, while a large systematic RMSE indicates a (Section 8). As can be seen from Table 1 the correlation between poor model.13 As can be seen from Table 1 the unsystematic Fig. 4 (a) The population distribution within the model domain.(b) The population load (hourly NO2 concentrations with 100 mg m-3 as threshold value) with all emissions included. (c) The population load with a 10% reduction in emissions from road traYc. (d) The population load without emissions from area-distributed stationary sources (i.e. mostly domestic heating). J. Environ. Monit., 1999, 1, 321–326 323Table 1 Model evaluation parameters comparing hourly observed and model predicted NOx and NO2 concentrations at Nordahl Brunsgate station in Oslo for the period October 1–November 19, 1996 (a total of 1198 h) Parameter Unit Observed NOx Predicted NOx Observed NO2 Predicted NO2 Mean mg m-3 88.8 83.2 38.5 35.9 Max mg m-3 1248.6 863.3 121.4 122.8 s mg m-3 106.4 112.1 15.7 23.9 BIAS 0.064 0.068 RMSE mg m-3 80.8 17.8 RMSEs mg m-3 25.3 2.7 RMSEu mg m-3 76.8 17.6 Correlation 0.73 0.68 Index of agreement 0.85 0.78 portion of RMSE is much larger than the systematic portion person dose basically is a concentration value, and its spatial distribution points to areas with potentially low air quality. both for NOx and NO2.This indicates that most of the deviance between observed and predicted values were of an unsystematic nature.The population load. The population load, L i,j is an exposure measure which combines the information about the person In order for the model to be more completely evaluated, the sensitivity of the above results to changes in emission dose in a grid square with the population living within this square, Pi,j . inventory or meteorological conditions (period of the year) shall be investigated in a future study.L i,j�Di,j×Pi,j (2) The spatial distribution of the population load reveals popu- 6. Population exposure estimates lated regions with concentrations above the threshold values. However, scarcely populated areas with high person dose After having established the ability of the EPISODE dispersion model to recapture measured concentration levels, the model values will not be distinguished from densely populated regions with small person dose values.A time dependent population was applied for predicting population exposure within the city of Oslo. In addition to NO2, these calculations were also distribution, Pi,j n, can easily be accounted for by the following definition of the population load: performed for PM10 and PM2.5.Since this presentation is meant merely as an example of how the model can be applied for exposure estimates, only results from the NO2 calculations L i,j� 1 N . N n=1 Pi,j n Max[(ci,j n-cT), 0] (3) will be referred to in the following. The exposure calculations presented have been simplified by A simplified time dependency of the population distribution applying a stationary population distribution.This means that would be to have one distribution valid for working hours it has been assumed that the inhabitants have stayed at their and another for the remaining hours. Increasing complexity home address during the entire calculation period. The expo- can be built in depending on available information. sure estimates have then been established by combining the calculated hourly NO2 concentration values at ground level The total population load.Both the person dose and the with the population distribution. population load are two-dimensional fields giving average The computed pollution levels have been compared with values for each grid square within the model domain. A the national air quality guidelines (AQG) given by the measure of the total population load within the domain is Norwegian State pollution control authies.At present the found by simply adding the grid square values of the popuguideline values (or threshold values) for NO2 are 100 mg m-3 lation load, i.e.: for hourly averaged values, 75 mg m-3 for daily averaged values and 50 mg m-3 for half-year averaged values.L Tot�. i,j L i,j (4) 6.1 Population exposure module By combining the calculated air pollution concentrations, the General considerations. The above definitions are just meant population distribution and the AQG values, diVerent measas examples of how various types of exposure measures can ures of population exposure can be constructed. Among these be constructed.Other possible choices of definition are for are the person dose and the population load, which are defined example to just consider episodes with a minimum duration as shown below. when calculating the person dose, or to scale the exceedances so that high concentration values are given more weight than The person dose. The person dose, Di,j is a quantity defined smaller ones. The selection of exposure measure to be used in for each grid square, (i, j), and is defined as: a given situation, or for a given pollutant, should generally be chosen in agreement with the medical community. Population Di,j� 1 N .N n=1 Max[(ci,j n-cT), 0] (1) concentrations based on sub grid calculations will be included in future calculations of a population exposure. where ci,j n is the calculated grid square concentration at time n, cT is a given threshold value (for example the AQG value) 6.2 Results and discussion and N is the total number of concentration values within the calculation period.The index n can either count hourly values, Population exposure calculations have been carried out for the city of Oslo for the winter period October 1, 1995 to daily values or other averages for which there exists a threshold value.The person dose can be interpreted as the average March 31, 1996. These calculations have been performed with a stationary population, which have been distributed according exceedance of the threshold value. With this definition no distinction is made between long episodes of small exceedances to their home addresses within the 1 km×1 km regular grid system of the model domain. Fig. 4a shows a colour contour and short periods with large exceedances. Note also that the 324 J. Environ. Monit., 1999, 1, 321–326plot of the applied population distribution. The black contour 8.1 Description of model evaluation performance parameters lines in the figure describe the topography of the model Let T denote the number of data, and let Ot and Pt denote domain. The diVerent colours indicate the number of persons the observed and model-calculated (predicted) values at time within each grid square.t, t=1, ...,T . The following model evaluation parameters may In Figs. 4b, c and d the calculated NO2 population load, then be defined: L i,j as defined by eqn. (2), is shown for three diVerent model O9 : Mean value of observations; simulations.The values shown in Fig. 4b were found when P9 : Mean value of predictions; the model was run with all available emission sources included. Omax : Maximum value of observations; The results shown in Fig. 4c were calculated for a scenario Pmax : Maximum value of predictions; with a 10% reduction in the emissions from road traYc.The sO : Standard deviation of observations; model can also be used in order to test the relative importance sP : Standard deviation of predictions; of diVerent types of sources. This may be done by selectively BIAS : Bias or normalised mean diVerence; removing diVerent source categories one at a time. An example RMSE : Root mean square error; of this is shown in Fig. 4d, where all emissions from area- RMSEs : Systematic RMSE; distributed stationary sources (i.e. emissions from domestic RMSEu : Unsystematic RMSE; heating and small industry not treated as individual point a,b : Intercept and slope of regression line; sources) have been neglected. The calculated population load Corr : Correlation coeYcient; shown in these figures was based on hourly NO2 concen- IA : Index of agreement.trations, and a threshold value of 100 mg m-3 was applied. The parameters are defined through the following set of This threshold value is equal to the AQG for hourly NO2 equations: concentrations given by the Norwegian State pollution control Mean values authorities. The maximum grid value in Fig. 4b is 3280 persons mg m-3. O9= 1 T . T t=1 Ot (A1) This maximum is reduced by almost 30% to 2305 persons mg m-3 as a consequence of the 10% reduction in the traYc emissions (Fig. 4c). By totally neglecting the emissions from P9= 1 T . T t=1 Pt (A2) the area-distributed stationary sources, the maximum calculated population load is reduced slightly more to a value of These denote the usual arithmetical average values of the time 2070 persons mg m-3 (Fig. 4d).Summing the individual grid series Ot and Pt . values in Figs. 4b, c and d, the total population load, as Maximum values defined by eqn. (4), can be computed for each experiment. In this way, total population load values of 43779, 30176 and Omax=max Ot for t=1, ...,T (A3) 29466 persons mg m-3 are found from the data shown in Pmax=max Pt for t=1, ..., T (A4) Figs. 4b, c and d, respectively. Assuming that the total population load is a good measure for population exposure, one These denote the usual maximum values of the time series Ot may infer from these results that a 10% reduction in road and Pt . traYc emissions and a 100% reduction in emissions from area- Standard deviations distributed stationary sources both lead to a 30% improvement in population exposure to NO2.It should be noted that these sO=C 1 T-1 . T t=1 (Ot-O9 )2D0.5 (A5) numbers strongly depend on the threshold value applied in eqn. (1). Calculations like those shown in Fig. 4 give valuable inforsP= C 1 T-1 . T t=1 (Pt-P9 )2D0.5 (A6) mation about possible problem areas within the modelling area. This type of information should be utilised not only These denote the usual standard deviations of the time series when designing a monitoring program, but also as a sup- Ot and Pt .plement to the monitoring program itself. As demonstrated Bias above, population exposure calculations can be utilised in diVerent forms of cost–benefit analyses, and as a tool for BIAS=(O9-P9 )/O9 (A7) quantifying the health eVects of various air quality This dimensionless parameter is a measure of the bias of P improvements.versus O. Ideally it should be zero, or close to zero. Root mean square error 7. Acknowledgements RMSE= C1 T . T t=1 (Ot-Pt)2D0.5 (A8) The authors acknowledge fruitful discussions with Dr The RMSE is another measure of the size of the error produced J. Clench-Aas and Dr A. Bartonova during the writing of this by the model.article. Parts of the study were funded by the Norwegian State Systematic and unsystematic RMSE Pollution Control Authority. RMSEs=C1 T . T t=1 (Ot-P� t)2D0.5 (A9) 8. Appendix RMSEu=C1 T . T t=1 (P� t-Pt)2D0.5 (A10) US EPA has given guidelines on procedures to be followed in evaluating air quality models, and a list of recommended where model evaluation performance parameters.14 In this study statistical parameters have been selected in accordance with P� t=a+bOt (A11) these recommendations. Selecting the parameters, results in ref. 13 were also taken into consideration. where a and b are the intercept and slope of the regression J. Environ. Monit., 1999, 1, 321–326 325line: 9. References a=P9-bO9 (A12) 1 S. E.Walker, The EPISODE air pollution dispersion model, version 2.2.Users Guide, Norwegian Institute for Air Research (NILU TR 10/97), Kjeller, Norway, 1997. b=C. T t=1 (Ot-O9 )(Pt-P9 )DNC. T t=1 (Ot-O9 )2D (A13) 2 K. E. Grønskei, S. E.Walker and F. Gram, Atmos. Environ., 1993, 27B, 105. 3 S. Larssen, K. E. Grønskei, F. Gram, L. O. Hagen and S. E. Here Walker, Verification of urban scale time dependent dispersion model with subgrid elements in Oslo, Norway, Air Pollution Modelling and RMSE2=RMSEs2+RMSEu2 (A14) Its Application X, ed.S.-E. Gryning and M. M. Millan, Plenum Press, New York, 1994, pp. 91–99. Systematic and unsystematan square errors give 4 J. Clench-Aas, A. Bartonova, R. Klæboe and M. Kolbenstvedt, in 8th International Symposium on Transport and Air Pollution, ed. valuable information on the possibility of model improve- P.J. Sturm, Technical University Graz, Report of the Institute for ment.13 For a good model the unsystematic portion of the Internal Combustion Engines and Thermodynamics, vol. 76, RMSE is much larger than the systematic, while a large Graz, Austria, 1999. systematic RMSE indicates a poor model. 5 A. Bott, Mon. Weather Rev., 1993, 121, 2637.Correlation coeYcient 6 G. D. Smith, Numerical solution of partial diVerential equations: Finite diVerence methods, Oxford University Press, Oxford, 1985. 7 K. E. Grønskei, Dispersion of pollution from area sources, Corr= 1 T . T t=1 (Ot-O9 )(Pt-P9 )/(sOsP) (A15) Norwegian Institute for Air Research (NILU TR 6/82), Lillestrøm, Norway, 1982. 8 W. B. Petersen, Users Guide for Hiway-2: A Highway Air This is the ordinary Pearson product–moment correlation Pollution Model, US Environmental Protection Agency (EPA-600/ coeYcient. 8-80-018), Research Triangle Park, NC, 1980. Index of agreement 9 C. A. Sherman, J. Appl. Meteor., 1978, 17, 312. 10 T. Bøhler, MEPDIM. The NILU meteorological processor for dispersion modelling. Version 1.0. Model description, Norwegian IA=1-C. T t=1 (Pt¾-Ot¾)2DN. T t=1 (|Pt¾|+|Ot¾|)2 (A16) Institute for Air Research (NILU TR 7/96), Kjeller, Norway, 1996. 11 S. E. Gryning, A. A. M. Holtslag, J. S. Irwin and B. Sivertsen, where Atmos. Environ., 1987, 21, 79. 12 A. P. Van Ulden and A. A. M. Holtslag, J. Appl. Meteorol., 1985, Pt¾=Pt-O9 and Ot¾=Ot-O9 (A17) 24, 1196. 13 C. J. Willmott, Am. Meteorol. Soc. Bull., 1982, 63, 1309. The index of agreement IA varies between 0 and 1, with 0 14 EPA, Interim procedures for evaluating air quality models (revised), EPA-450/4-84-023, OYce of Air Quality Planning and indicating the worst agreement and 1 indicating the best Standards, US Environmental Protection Agency, Research agreement. It has been recommended13 as a better parameter Triangle Park, NC, 1984. to describe the ‘agreement’ between the two time series Ot and 15 J. Clench-Aas, A. Bartonova, T. Bøhler, K. E. Grønskei and Pt than the correlation coeYcient. S. Larssen, J. Environ. Monit., 1999, 1, 313. Paper 9/02776I 326 J. Environ. Monit., 1999, 1, 321–326