Plume-in-grid modeling for particulate matter

Transcription

1 Atmospheric Environment 40 (2006) Plume-in-grid modeling for particulate matter Prakash Karamchandani a,, Krish Vijayaraghavan a, Shu-Yun Chen a, Christian Seigneur a, Eric S. Edgerton b a Atmospheric and Environmental Research, Inc., 2682 Bishop Drive, Suite 120, San Ramon, CA 94583, USA b Atmospheric Research and Analysis, Inc., 410 Midenhall Way, Cary, NC 27513, USA Received 10 May 2006; received in revised form 21 June 2006; accepted 25 June 2006 Abstract Three-dimensional grid models are now being commonly used to simulate particulate matter (PM) concentrations, especially fine PM (PM 2.5 ), which includes a significant fraction of secondary species formed in the atmosphere. These models usually do not address the subgrid-scale effects associated with emissions from large elevated point sources. This can lead to errors in the calculation of PM 2.5 concentrations downwind of these sources because of unrealistic representations of the dispersion and transformation processes that govern PM 2.5 formation. Here, we describe the development, application and evaluation of a plume-in-grid (PiG) model for PM. The model, based on an existing PiG model for ozone, is extended to include in the plume component state-of-the-science treatments of aerosol chemistry and dynamics as well as aqueous chemistry that are consistent with the treatments used in the host grid model. Application of this model to several SO 2 and NO x emitting power plants in the southeastern United States shows that the PiG treatment leads to significant differences in sulfate and total inorganic nitrate concentrations. Comparisons of model results, with and without PiG treatment, against measurements characterizing specific plume events, show that the PiG treatment captures the plume events more often and generally better than the standard grid-based approach. r 2006 Elsevier Ltd. All rights reserved. Keywords: PM 2.5 ; Subgrid-scale; Plume chemistry; CMAQ; SCICHEM; SEARCH 1. Introduction Three-dimensional (3-D) grid models are frequently used to predict the impacts of emission controls on concentrations of pollutants such as ozone and fine particulate matter (PM 2.5 ). However, such models are limited in their ability to correctly represent the near-source dispersion, transport, and chemistry of emissions from elevated sources. These Corresponding author. Tel.: ; Fax: address: pkk@aer.com (P. Karamchandani). limitations are due to the inability of grid models, with typical horizontal resolutions of a few kilometers to tens of kilometers, to resolve stack plumes with initial dimensions of tens of meters. A common approach to address these limitations is to use Plume-in-Grid (PiG) modeling, in which a subgridscale representation of stack plumes is embedded in the 3-D model. This approach has traditionally been limited to ozone modeling studies (Seigneur et al., 1983; Kumar and Russell, 1996; Karamchandani et al., 2002; Vijayaraghavan et al., 2006). A detailed discussion of the errors associated with ignoring the subgrid-scale features of point source plumes on the /$ - see front matter r 2006 Elsevier Ltd. All rights reserved. doi: /j.atmosenv

2 P. Karamchandani et al. / Atmospheric Environment 40 (2006) formation of gaseous pollutants has been provided by Karamchandani et al. (2002). The errors associated with the gridded representation of point source plumes are just as important for PM 2.5 as they are for ozone, since the chemistry leading to formation of PM 2.5 constituents, such as sulfate and nitrate, is significantly slower in the plume of a large NO x point source than it is in the background air (Karamchandani et al., 1998; Karamchandani and Seigneur, 1999). This is particularly true in the initial stages of plume dispersion, when the plume dimensions are small relative to the grid resolution. Thus, using a purely grid-based approach can potentially lead to an incorrect estimation of the impacts of large SO x and NO x point sources on ambient PM 2.5 concentrations. Karamchandani et al. (2002) previously developed a PiG model for ozone based on the US EPA Community Multiscale Air Quality model (CMAQ). This model is referred to as CMAQ with Advanced Plume Treatment (CMAQ-APT), and includes a state-of-the-science reactive puff model, SCICHEM, to resolve stack plumes within CMAQ. Here, we present the improvement of SCICHEM and CMAQ-APT to include a state-of-the-science treatment for PM. This treatment, referred to as the Model of Aerosol Dynamics, Reaction, Ionization, and Dissolution (MADRID), was incorporated in CMAQ by Zhang et al. (2004), and is extended in this work to SCICHEM and CMAQ-APT. The resulting PiG model, CMAQ-MADRID-APT, is applied to the southeastern United States for 2 months in 2002 with stacks from 14 power plants selected for PiG treatment. For comparison purposes, the grid model without PiG treatment, CMAQ-MADRID, is applied for the same domain and time periods. 2. The PiG model CMAQ-MADRID-APT consists of a reactive plume model, SCICHEM (Karamchandani et al., 2000), embedded into the host 3-D grid-based model, CMAQ-MADRID (Zhang et al., 2004). Brief descriptions of CMAQ-MADRID and SCICHEM are provided in the following sub-sections. SCICHEM was embedded into the host grid model following the established protocols for incorporating new science modules into CMAQ ( CMAS-CodingGuidelines.pdf). The general approach for the coupling of the plume model with the host model is described in detail by Karamchandani et al. (2002) The host grid model, CMAQ-MADRID The host model is based on the October 2004 release (Version 4.4) of the US EPA s CMAQ model and includes a condensed version of the MADRID PM 2.5 treatment of Zhang et al. (2004). CMAQ was developed by EPA to address multiscale multipollutant air pollution problems (Byun and Schere, 2006). CMAQ treats the emissions, transport, dispersion, chemical transformations, gas-particle conversion and removal processes that govern the behavior of chemical pollutants in the atmosphere. Emissions include those from area sources (e.g., industrial, residential, agricultural, mobile and biogenic emissions) and point sources (e.g., power plants, smelters, and refineries). The plume rise of point source emissions is treated in a pre-processor to CMAQ. Transport processes include advection, large-scale convection and, in the presence of cumulus clouds, subgrid-scale convection. Dispersion includes both horizontal and vertical dispersion. Chemical transformations modeled include reactions in the gas phase and reactions in the aqueous phase (i.e., in cloud droplets). The formation of secondary aerosols and the gas-particle partitioning of semi-volatile chemical species are simulated. Dry deposition is simulated for gases and particles. Wet deposition is simulated for precipitating clouds that are resolved by the grid system as well as for clouds that are treated at the subgridscale. In CMAQ-MADRID, the EPA treatment of aerosols is replaced by a condensed version of the MADRID treatment of Zhang et al. (2004). The primary differences between the EPA and MADRID treatments are in their representation of the particle size distribution and the formation of secondary organic aerosols (SOA). The former uses a modal representation of the size distribution, while the latter uses a sectional representation. Here, MADRID was applied with two size sections. The condensed version of MADRID used here includes 8 explicit SOA precursors (Pun et al., 2004) instead of the 14 precursors used in the original MADRID formulation. This condensed version was found to produce slightly higher SOA concentrations than the original formulation, but the differences between the two versions were negligible compared to current uncertainties in our knowledge

3 7282 ARTICLE IN PRESS P. Karamchandani et al. / Atmospheric Environment 40 (2006) of SOA precursor emissions and SOA formation (Pun et al., 2004) The embedded reactive plume model The reactive plume component of CMAQ- MADRID-APT is the Second-order Closure Integrated puff model (SCIPUFF) with CHEMistry (SCICHEM). Plume transport and dispersion are simulated with SCIPUFF, a model that uses a second-order closure approach to solve the turbulent diffusion equations (Sykes et al., 1988, 1993; Sykes and Henn, 1995). The plume is represented by a myriad of 3-D puffs that are advected and dispersed according to the local micrometeorological characteristics. Each puff has a Gaussian representation of the concentrations of emitted inert species. The overall plume, however, can have any spatial distribution of these concentrations, since it consists of a multitude of puffs that are independently affected by the transport and dispersion characteristics of the atmosphere. SCIPUFF can simulate the effect of wind shear since individual puffs will evolve according to their respective locations in an inhomogeneous velocity field. As a puff grows larger, it may encompass a volume that cannot be considered homogenous in terms of the meteorological variables. A puffsplitting algorithm accounts for such conditions by dividing puffs that have become too large into a number of smaller puffs. Conversely, puffs may overlap significantly, thereby leading to an excessive computational burden. A puff-merging algorithm allows individual puffs that are affected by the same (or very similar) micro-scale meteorology to combine into a single puff. In the PiG version of SCICHEM, puffs are dumped to the host model grid (i.e., the puff contents are transferred to the 3-D grid system and the puff is destroyed) when the horizontal size of the puff is comparable to the horizontal grid resolution. The formulation of nonlinear chemical kinetics within the puff framework is described by Karamchandani et al. (2000), who also provide details on the SCICHEM model formulation and the evaluation of the model with helicopter plume measurements from the 1995 Southern Oxidants Study (SOS) in Nashville/Middle Tennessee. Chemical species concentrations in the puffs are treated as perturbations to the grid concentrations on which the puffs are overlaid. The chemical reactions within the puffs are simulated using a general framework that allows any chemical kinetic mechanism to be treated. In this study, a modified version of the Carbon-Bond Mechanism (CBM-IV) is used in SCICHEM for consistency with the host model, CMAQ-MADRID. This mechanism includes the reactions of the SOA precursors and condensable organic products that are not explicitly treated in CBM-IV. The MADRID aerosol modules were incorporated into SCICHEM. These include modules for the thermodynamics of inorganic species (ISOROP- PIA; Nenes et al., 1999), the condensed SOA representation of Pun et al. (2004), and aerosol dynamics. In addition, the aqueous-phase chemistry and equilibrium modules of CMAQ-MADRID were also incorporated into SCICHEM to treat the scavenging of soluble gases and aerosols and the oxidation of SO 2 to sulfate in clouds. Other changes to SCICHEM include the addition of heterogeneous chemistry on the surface of aqueous particles and cloud droplets according to the treatment of Zhang et al. (2004). 3. Model application The base model (CMAQ-MADRID) and the PiG version of the model (CMAQ-MADRID-APT) were applied to the southeastern US for 2 months, January and July The top panel of Fig. 1 shows the modeling domain, which consists of 168 by 177 grid cells in the horizontal, with a grid resolution of 12 km. The vertical grid structure consists of 19 layers from the surface to about 15 km, with finer resolution near the surface. This grid is identical to the inner grid used by the Visibility Improvement State and Tribal Association of the Southeast (VISTAS), the Regional Planning Organization (RPO) responsible for managing regional haze for the southeastern United States The modeling datasets (meteorology, emissions, initial and boundary conditions, photolysis rates) for the study described here were provided by VISTAS ( For the CMAQ-MADRID-APT simulations, elevated point source emissions from 14 coal-fired power plants (CFPPs) in Alabama, Florida, Georgia, and Mississippi were selected for explicit treatment with the PiG approach. The bottom panel of Fig. 1 shows the locations of the CFPPs in a sub-domain of the modeling domain. For the 14 sources selected here, the computational time for the

4 P. Karamchandani et al. / Atmospheric Environment 40 (2006) smaller region encompassing these four states and the immediate surrounding areas, these sources represent approximately 33% and 7% of the total SO x and NO x emissions, respectively. Primary emissions of PM sulfate from the 14 CFPPs are approximately 2% of the SO 2 emissions. For convenience, we will refer to the CMAQ- MADRID-APT simulation as the APT simulation in the discussion that follows. To determine the effect of using a PiG treatment, we conducted the following additional simulations: 1. A simulation with CMAQ-MADRID in which the 14 CFPPs referred to above were treated like the remaining point sources (i.e., without PiG treatment). We will refer to this simulation as the base simulation. 2. A simulation with CMAQ-MADRID in which the emissions from the 14 CFPPs were excluded from the simulation. We will refer to this simulation as the background simulation. We first present the model performance evaluation results for the base and APT versions of CMAQ-MADRID. Then, we examine the differences between the two versions in their predictions of sulfate, nitrate and PM 2.5 concentrations as well as their calculations of the contributions of the 14 CFPPs to these concentrations Operational evaluation of model performance Fig. 1. Modeling domain (top) and sub-domain (bottom) showing the locations of CFPPs selected for explicit PiG treatment (hollow circles) and SEARCH monitoring sites (solid triangles). PiG simulation increased by about 19% compared to a conventional grid simulation. The stack heights for the 14 sources range from about 60 m to about 300 m; however, the actual release heights of the species emitted from the sources are determined from plume rise calculations and vary temporally due to variations in prevailing conditions (e.g., wind speeds, temperatures, etc.). These sources represent approximately 7% and 2% of the total SO x and NO x emissions over the entire modeling domain, respectively. However, over the Measurements from many monitoring networks were used to evaluate and compare the model performance of CMAQ-MADRID and CMAQ- MADRID-APT. These data sources include: (1) EPA s Air Quality System (AQS) database; (2) the Interagency Monitoring of Protected Visual Environments (IMPROVE) network; (3) the Clean Air Status and Trends Network (CASTNET); and (4) the Southeastern Aerosol Research and Characterization study (SEARCH) network. While the first three networks have a nation-wide focus and provide routine measurements, SEARCH is a highly instrumented, eight-station regional network in the states of Alabama, Florida, Georgia and Mississippi (Hansen et al., 2003). The eight stations (see Fig. 1) are arranged in four urban rural pairs in each of the four states. SEARCH provides integrated filter-based measurements as well as yearround continuous measurements (1 60 min averages) of PM 2.5 and gas components. Additional

5 7284 ARTICLE IN PRESS P. Karamchandani et al. / Atmospheric Environment 40 (2006) information on SEARCH is available at For the study presented here, the SEARCH network is the most useful because of its detailed measurements and its distribution of monitoring locations in the same states as the 14 CFPPs selected for PiG modeling. Hence, in this paper, we will focus on model performance evaluation using the SEARCH measurements. Table 1 shows the model performance statistics for the base and APT simulations for January 2002, using measurements from the SEARCH network. The corresponding statistics for the July 2002 period are shown in Table 2. The performance measures used are those defined by Seigneur et al. (2000) in their guidance document for the performance evaluation of 3-D grid models, and are calculated as follows: Mean normalized bias ¼ 1 X N P i O i ; (1) N O i Fractional bias ¼ 1 N X N i¼1 Mean normalized error ¼ 1 N i¼1 2 P i O i, (2) P i þ O i X N i¼1 P i O i, (3) Fractional error ¼ 1 X N 2 P i O i N P i¼1 i þ O, (4) i where N is the total number of non-missing measurements, P i the Predicted value at site i, O i the Observed value at site i. O i Observations and predictions were paired in both space and time in the calculation of the model performance statistics. The simulated values correspond to the grid cells containing the SEARCH monitoring locations. For the base simulation, this value is simply the grid cell value. For the APT simulation, this value is the grid cell value plus the contributions of all active puffs (i.e., puffs not previously dumped to the host model) that affect the grid cell. The puff contribution is calculated at the center of the grid cell. However, as shown in the next section where we examine model performance for specific plume events, the APT simulation contains the information required to calculate subgrid-scale variability and this spatial variability can have a great impact on model performance for those monitoring locations that are within a few grid cells downwind of the sources selected for PiG treatment. Tables 1 and 2 show that, for the most part, the model performance statistics for both the base and PiG versions of CMAQ-MADRID are comparable. In particular, for January 2002 (Table 1), the performance statistics are almost identical for most species. There are small differences between the two models for sulfate and somewhat larger differences for SO 2. The mean normalized error and bias are lower in the MADRID-APT simulation than in the MADRID simulation for both sulfate and SO 2, but the fractional error and bias are higher. Both versions of the model significantly overpredict nitric acid concentrations in January. Table 2 shows that there are larger differences between the performance statistics of the two versions of the model for sulfate concentrations in Table 1 Model performance statistics for January 2002 using SEARCH measurements Performance measure CMAQ-MADRID SEARCH CMAQ-MADRID-APT 24-h average SO ¼ 4 concentrations Mean observed value (mgm 3 ) 2.2 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average NO 3 concentrations Mean observed value (mgm 3 ) 1.0 Mean simulated value (mgm 3 ) Mean normalized error (%) 76 77

6 P. Karamchandani et al. / Atmospheric Environment 40 (2006) Table 1 (continued ) Performance measure CMAQ-MADRID SEARCH CMAQ-MADRID-APT Fractional error Mean normalized bias (%) 5 7 Fractional bias Coefficient of determination (r 2 ) h average NH + 4 concentrations Mean observed value (mgm 3 ) 1.0 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) 8 7 Fractional bias Coefficient of determination (r 2 ) h average OC a concentrations Mean observed value (mgm 3 ) 4.4 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average EC b concentrations Mean observed value (mgm 3 ) 0.9 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average PM 2.5 concentrations Mean observed value (mgm 3 ) 11.1 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) 6 5 Fractional bias Coefficient of determination (r 2 ) h average SO 2 concentrations Mean observed value (ppb) 5.3 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average HNO 3 concentrations Mean observed value (ppb) 0.40 Mean simulated value (ppb) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) a Organic carbon. b Elemental carbon.

7 7286 ARTICLE IN PRESS P. Karamchandani et al. / Atmospheric Environment 40 (2006) Table 2 Model performance statistics for July 2002 using SEARCH measurements Performance measure CMAQ-MADRID SEARCH CMAQ-MADRID-APT 24-h average SO ¼ 4 concentrations Mean observed value (mgm 3 ) 4.8 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average NO 3 concentrations Mean observed value (mgm 3 ) 0.4 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average NH + 4 concentrations Mean observed value (mgm 3 ) 1.4 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average OC a concentrations Mean observed value (mgm 3 ) 4.7 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average EC b concentrations Mean observed value (mgm 3 ) 0.7 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average PM 2.5 concentrations Mean observed value (mgm 3 ) 16.0 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average SO 2 concentrations Mean observed value (ppb) 4.2 Mean simulated value (mgm 3 ) Mean normalized error (%) Fractional error

8 P. Karamchandani et al. / Atmospheric Environment 40 (2006) Table 2 (continued ) Performance measure CMAQ-MADRID SEARCH CMAQ-MADRID-APT Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) h average HNO 3 concentrations Mean observed value (ppb) 0.50 Mean simulated value (ppb) Mean normalized error (%) Fractional error Mean normalized bias (%) Fractional bias Coefficient of determination (r 2 ) a Organic carbon. b Elemental carbon. July The PiG version of the model performs better than the base version for sulfate for all the performance measures. For SO 2, most of the performance measures are better for APT than the base, except for the fractional bias and the ratio of the mean simulated value to mean observed value. For all other species, the performance measures are almost identical. As in January 2002, both versions significantly overpredict nitric acid concentrations in July. However, particulate nitrate concentrations are underpredicted by about a factor of 4 on average. This result suggests that not enough nitric acid is being partitioned into the particle phase, possibly because there is insufficient ammonia in the model to form ammonium nitrate. As discussed in Section 3.3, an important aspect of the chemistry of NO x -rich plumes is the inhibition of the SO 2 to sulfate and NO x to nitrate conversion rates in the early stages of plume growth. This feature is correctly treated in the PiG simulation, but not in the base simulation. Thus, there is higher initial conversion of SO 2 to sulfate in the base simulation than in the APT simulation. This explains the larger overpredictions of sulfate concentrations in the base simulation as compared to the APT simulation, particularly in July. The differences between the model performances for sulfate for the APT and base simulations are larger in July than in January because less secondary sulfate is formed in the winter due to reduced photochemical activity. The comparison of model performance statistics for the base and APT simulations suggests that APT simulates sulfate concentrations better than the base model in both January and July However, the aggregate model performance statistics for the base and APT simulations are generally similar and it is difficult to distinguish between the two simulations on the basis of these model performance statistics alone. We present next a more detailed evaluation and comparison of the SO 2 predictions from the base and APT simulations for specific plume events identified in the SEARCH measurements Evaluation of model performance for plume events There were several power plant plume events identified at SEARCH monitoring locations during January and July Many of these events could be associated with unique sources based on the measured SO 2 to NO y ratios and back trajectory analysis, as described below. The identified sources were among the 14 CFPPs included in the APT simulation. Thus, it is useful to compare model results with observations for these plume events, because we expect to see larger differences between the base and APT simulations due to the differences in their treatment of plume chemistry and transport. Source CFPPs for each plume event were identified using two methods. First, the observed dso 2 /dno y ratio was compared with hour-by-hour SO 2 /NO x ratios from in-stack continuous emission monitors (CEMs). Linear regression of SO 2 vs. NO y during the plume event was used to determine dso 2 /dno y. Second, air mass trajectory analysis was used to confirm transport from the CFPP to the receptor site. In practice, 24-h back trajectories were calculated for each event using the interactive

9 7288 ARTICLE IN PRESS P. Karamchandani et al. / Atmospheric Environment 40 (2006) version of the NOAA HYSPLIT model ( Back trajectories were initiated at the receptor site at the time (nearest hour) of peak measured SO 2 concentration. A specific CFPP was confirmed if the dso 2 /dno y ratio matched the in-stack SO 2 /NO x ratio within 10% and the back trajectory came within 151 of the CFPP. The criteria of 10% is consistent with the findings of Edgerton et al. (2006), who showed that SO 2 :NO y ratios calculated from Continuous Emissions Monitoring (CEM) data at the source were usually within 20% and often agreed to better than 10% with field measurements downwind of the source. Two approaches were used to perform the plumeevent analysis. In the first approach, the simulated values for the grid cells containing the SEARCH monitoring sites were compared with observations. This approach is identical to that used for the calculation of the operational model performance statistics described previously. In the second approach, we used the plume information available in the APT simulation to calculate concentrations at discrete receptor locations. These receptor locations included the actual monitoring site location as well as receptors at intervals of 21 along an arc on each side of the monitoring site. The center of the arc is the CFPP identified as impacting the SEARCH monitoring site for a particular plume event and the radius of the arc is the distance between the CFPP and the monitoring site. The extent of the arc is 301 on each side of the monitoring site location. Fig. 2 provides an illustration of the placement of the array of receptors for a monitoring site located east of the contributing source at a downwind distance of about three grid cells. At this distance, about nine receptors lie in the grid cell containing the monitoring site and the remaining receptors lie in the adjacent grid cells. Using this approach allows us to take advantage of the subgrid-scale features in the APT simulation and also to account for possible small discrepancies in wind direction between the actual winds transporting the plume from the source to the monitoring Fig. 2. Illustration of placement of receptors along an arc downwind of a source to calculate subgrid-scale plume concentrations at and around a given monitoring site.

10 P. Karamchandani et al. / Atmospheric Environment 40 (2006) site and the wind fields used in the simulation. Because this approach is most useful when a single source is responsible for the observed plume event, our analysis is limited to those events with identified and unique sources. Fig. 3 compares measured and simulated SO 2 concentrations at the SEARCH monitoring site in North Birmingham, Alabama (BHM) during a plume event on 30 July The Miller power plant was identified as the source responsible for this plume event. The monitoring site is eastsoutheast of the power plant at a distance of about 24 km (i.e., about two grid cells). Simulated concentrations from both the base and APT simulations are shown in Fig. 3. For the base simulation (dashed line), the concentration is the single grid-volume averaged value for the grid cell containing the monitoring site. For the APT simulation, the approach described earlier is used to calculate concentrations at 31 discrete receptors at a spacing of 21 along an arc centered at the Miller power plant with a radius of about 24 km. Of the 31 receptors, 13 lie in the same grid cell containing the monitoring site. Fig. 3 shows the APT results at 8 of these 13 receptors (the concentrations at the remaining 5 receptors were very similar to those at one of the 8 receptors shown in the figure). The line labeled APT1 represents the calculated concentrations at the actual monitoring location, while the lines labeled APT2 through APT8 represent concentrations at receptors at various distances to the north of the monitoring location but within the same grid cell as and at the same downwind distance from the Miller plant as APT1. Fig. 3 shows that there is considerable subgridscale variability in SO 2 concentrations within the grid cell and the simulated peaks at various locations within the grid cell approach the measured peak as we sample the simulated plume to the north of the monitoring location. The best match between the observed and simulated peak SO 2 concentrations occurs at the receptor labeled APT8, about 11 km to the north of APT1. This subgrid-scale variability is not modeled in the base simulation, where the peak simulated SO 2 concentration is about a factor of two lower than the peak measured concentration. Table 3 compares the peak measured SO 2 plume increments at a number of SEARCH monitoring sites during different days in January and July 2002 with the corresponding values from the base and APT simulations. The measured SO 2 increment is the difference between the measured SO 2 concentration and an inferred background value (calculated as the average of the 1-h concentrations immediately before and after the plume event). The simulated base increments are the differences between the concentrations from the base simulation and the background simulation, while the simulated APT increments are the differences between the concentrations from the APT simulation and the background simulation. Fig. 3. Comparison of measured and simulated SO 2 concentrations at the North Birmingham, Alabama SEARCH monitoring site for a plume event on 30 July The dashed line shows the base simulated value for the grid cell containing the monitoring site, while the solid lines labeled APT1 through APT8 represent values from the APT simulation at 8 discrete receptors within the same grid cell as the monitoring site and at the same distance downwind of the identified source of the plume event.

11 7290 ARTICLE IN PRESS P. Karamchandani et al. / Atmospheric Environment 40 (2006) Table 3 Comparison of measured and simulated peak SO 2 increments during SEARCH plume events Date Monitoring Site Distance from source (km) Peak SO 2 plume increment (ppb) Measured Simulated Base a Base best b APT c APT best d 21 July YRK July YRK July CTR January YRK July BHM January YRK July BHM July PNS July CTR January YRK July BHM July BHM January YRK July PNS July YRK January YRK January YRK a CMAQ-MADRID grid-cell value in the cell containing the monitoring site. b Best CMAQ-MADRID grid-cell value from cells along the arc shown in Fig. 2. c CMAQ-MADRID-APT point value at the monitoring site location. d Best CMAQ-MADRID-APT value at receptors along the arc shown in Fig. 2. Table 3 shows two sets of plume increments for each simulation. For the base simulation, the first value ( Base ) is the peak incremental SO 2 concentration at the grid cell containing the monitoring site, and the second value ( Base best ) is the peak incremental SO 2 concentration that best matches the measured SO 2 increment in any of the grid cells along the 601 arc shown in Fig. 2. For the APT simulation, the first value ( APT ) is the peak incremental SO 2 concentration calculated at the location of the monitoring site, while the second value ( APT best) is the peak incremental SO 2 concentration that best matches the measured SO 2 increment at any of the array of 31 receptors along the arc in Fig. 2. For both simulations, the timing of the simulated peak increment does not always match the timing of the observed peak, possibly due to differences between the actual wind speeds and those used in the simulations. Also, as discussed previously, selecting the best-matching receptor or grid cell along the arc allows us to account for differences in wind directions of up to 301 on either side of the line joining the source and the monitoring site. The order of plume events in Table 3 is based on the measured SO 2 increments, with the highest values appearing first. The plume events listed are those with a unique identified source, based on the signature of the source. More than half of such plume events occur at the Yorkville, Georgia monitoring site (YRK). In July, three Yorkville plume events are associated with the same source, about 24 km to the northeast of Yorkville. As shown in Table 3, for all three events at Yorkville in July, the peak SO 2 plume increment is better captured in the APT simulation than in the base simulation. For the 21 July episode, both simulations underpredict the measured peak substantially. However, the APT best-match peak is almost twice the base best-match peak and is closer to the measured value. On 5 July, the APT best-match peak plume increment is only about 15% lower than the measured increment, but the base best-match peak is almost a factor of 3 lower than the measured peak. On 31 July, the base best-match value is about 40% lower than the measured peak, but the APT best-match value is only about 6% lower than the measured peak.

12 P. Karamchandani et al. / Atmospheric Environment 40 (2006) In January 2002, there are 6 plume episodes at Yorkville, 4 of which are associated with a source to the northwest at a distance of about 45 km. On 3 and 4 January, the best-match APT peak values compare slightly better with observations than the best-match base peak values. However, both simulations underpredict the measured peaks considerably. On 15 January, the best-match peak increments from both simulations are comparable and about 14% lower than the measured peak. On 7 January, the best-match base value is more than a factor of 2 lower than the measured value, but the best-match APT peak increment is within 4% of the measured peak. The other two Yorkville events in January are associated with a source that is about 150 km to the southwest of Yorkville. On 9 January, we see that the best-match peak increment of 13.1 ppb in the APT simulation is in good agreement with the measured peak value of 11.1 ppb, but the base simulation underpredicts the peak by about 32%. On the other hand, the measured peak is underpredicted significantly on 10 January in both the base and APT simulations, but the base peak is slightly higher than the APT peak. This is one of only two cases in Table 3 in which the base simulation peak increments are in better agreement with the measured peaks than the APT values. There are plume events on 4 days at the North Birmingham, Alabama site (BHM) in July All of these events are associated with a CFPP at a distance of about 24 km to the west-northwest of BHM. As shown in Table 3, the two models also show plume impacts at BHM for 3 of the plumeevent days (30, 20 and 31 July), but neither model predicts a plume event on 21 July. For all the 3 days with simulated plume impacts, the APT best-match peak increments are in much better agreement with observed values than the corresponding base increments. The results for the plume event on 30 July at BHM were shown earlier in Fig. 3. Two plume events at the Centreville, Alabama monitoring site (CTR) are shown in Table 3. The 27 July event is associated with a source to the southwest of Centreville, at a distance of about 60 km. The peak measured SO 2 plume increment is about 33 ppb, while the base and APT best-match simulated values are considerably lower. However, the base simulation value is more than a factor of 10 lower than the measured value, while the APT peak is more than 4 times the base value and about a factor of 2 lower than the measured peak. The 12 July event is associated with a source to the north of Centreville, at a distance of about 83 km. The bestmatch base peak increment underestimates the measured peak by about a factor of 2, while the APT value, which is only about 4% lower than the observed peak, is in much better agreement. Table 3 shows two plume events in July 2002 at the Pensacola, Florida monitoring site (PNS), both of which are associated with a nearby source (about 15 km from PNS) to the north-northeast. On 18 July, the APT best-match SO 2 peak increment is slightly less than a factor of 2 higher than the measured value, while the base simulation value is in much better agreement with and slightly lower (about 11%) than the measured value. On 19 July, both models underpredict the peak measured increment by about a factor of 2, but the APT peak is slightly higher than the base peak. These results underscore the point that many of the subtle differences between the base and APT simulations are not evident when comparing aggregate model performance statistics based on grid-average values. Exploiting the subgrid-scale information in the APT simulation to compare model results for individual plume events allows us to gain a better understanding of the differences between the two models and shows that incorporating a PiG treatment results in significantly better performance in capturing the peak plume increments during the plume events. In the following section, we discuss the differences between the base and APT simulations in terms of their calculated source contributions to sulfate concentrations Effect of PiG treatment Fig. 4 shows the spatial distribution of the contributions of the 14 CFPPs selected for PiG treatment to monthly average ground-level sulfate concentrations in July The results are shown for both the base and APT simulations. These contributions are calculated by subtracting the background simulation (i.e., without the 14 sources) concentrations from the base and APT values. While both models show the same spatial extent of power plant impacts, there are significant differences between the calculated contributions from the two models, particularly in the regions near the 14 sources. As shown in Fig. 4, the maximum CFPP contribution to ambient sulfate concentrations is 5.1 mgm 3 in the base simulation,

13 7292 ARTICLE IN PRESS P. Karamchandani et al. / Atmospheric Environment 40 (2006) Fig. 4. Simulated contributions (mg m 3 ) of the 14 power plants selected for PiG treatment to monthly average surface concentrations of PM 2.5 sulfate during July 2002 in (a) the base simulation and (b) the APT simulation. Fig. 5. (a) Relative changes in power plant contributions to monthly average surface concentrations of PM 2.5 sulfate during July 2002 when a plume-in-grid approach is used; and (b) differences (APT base) between the APT and base monthly average sulfate concentrations (mg m 3 ). while the maximum contribution in the APT simulation is 3 mgm 3, about 41% lower than the base value. Over most of the region impacted by the power plants, the APT simulation results in a smaller contribution of the CFPPs to sulfate concentrations. This is illustrated in Fig. 5a, which shows the relative changes in the calculated power plant contributions when a PiG treatment is used instead of a purely grid-based approach to simulate the fate of elevated point source emissions, and in Fig. 5b, which presents the actual differences (APT base) between the sulfate concentrations predicted in the two simulations. As shown in Fig. 5a, when CMAQ-MADRID-APT is used instead of CMAQ-MADRID, the calculated contributions of the power plants reduce by 5 40% in the regions near the 14 CFPPs (most of Alabama, northwestern Georgia, south-eastern Mississippi and north-western Florida). Even at larger distances from these sources, the calculated impacts in the APT simulation are smaller than the base simulation values. For example, over eastern Tennessee and most of the Carolinas, the power plant contributions to sulfate concentrations in the APT simulation reduce by 1 5% from the base simulation values. Fig. 5b shows that the predicted sulfate concentrations in the APT simulation are about mg/m 3 lower than the base simulation values in most of northern Alabama and northern Georgia

14 P. Karamchandani et al. / Atmospheric Environment 40 (2006) and small areas of the tri-state region of Alabama, Mississippi and the Florida panhandle. These results are consistent with our understanding of the chemistry of NO x -rich plumes. Both experimental studies (e.g., Richards et al., 1981; Gillani et al., 1998) and theoretical/modeling studies (e.g., Karamchandani et al., 1998, 2002; Karamchandani and Seigneur, 1999; Godowitch, 2004) have shown that the rates of O 3 formation and SO 2 and NO x oxidation in a plume rich in NO x differ significantly from those in the background atmosphere. Karamchandani et al. (1998) defined three stages for plume chemistry. In the first stage, the high nitric oxide (NO) concentrations in the plume lead to a depletion of oxidant levels until sufficient plume dilution has taken place. Therefore, in the vicinity of power plants and other elevated NO x sources, there is a temporary decrease in O 3 and other oxidant levels, even during daytime conditions, due to the titration of O 3 by the high levels of NO in the near-field of the plume. In this stage, there is virtually no formation of sulfate and nitrate from the gas-phase oxidation of their precursors. Farther downwind, in the second stage, the plume starts to mix into the background environment and becomes diluted; accordingly, radical concentrations within the plume start to increase and there is some formation of sulfate and nitrate, but the rates of SO 2 and NO x oxidation are still lower than the background rates, because of depressed oxidant levels. In the last stage, the plume has become sufficiently dispersed that its chemistry is similar to that of the ambient background, and the rates of sulfate and nitrate formation in the plume are comparable to the background rates. These three stages of plume chemistry are qualitatively apparent in Fig. 5b (as seen in the transition from green to red colored regions). Because most of the domain is ammonia-limited and there is sufficient sulfate to neutralize the available ammonia, the effect of the APT simulation on particulate nitrate concentrations is small. Thus, we describe the differences between the two simulations in terms of the effects on total inorganic nitrate (gas-phase nitric acid+particulate nitrate) concentrations. The results for total inorganic nitrate concentrations are qualitatively similar to those for sulfate concentrations and are not displayed here in the interest of space. The maximum contribution of the 14 power plants to the monthly average inorganic nitrate concentrations for July 2002 is 1.1 mgm 3 in the APT simulation, about half the base simulation value of 2 mgm 3. The calculated contributions in the APT simulation of the 14 power plants to inorganic nitrate concentrations are lower by about 5 28% from the base simulation contributions in the regions near the power plants. The sulfate results for January 2002 are shown in Figs. 6 and 7. Fig. 6 shows the contributions of the 14 power plants to monthly average ground-level sulfate concentrations. The results are qualitatively similar to the July 2002 results shown in Fig. 4. The main difference from the July results is that the predicted sulfate levels are lower, as is expected from our understanding of the processes governing the conversion of SO 2 to sulfate and their seasonal Fig. 6. Simulated contributions (mg m 3 ) of the 14 power plants selected for PiG treatment to monthly average surface concentrations of PM 2.5 sulfate during January 2002 in (a) the base simulation and (b) the APT simulation.

15 7294 ARTICLE IN PRESS P. Karamchandani et al. / Atmospheric Environment 40 (2006) Fig. 7. (a) Relative changes in power plant contributions to monthly average surface concentrations of PM 2.5 sulfate during January 2002 when a plume-in-grid approach is used; and (b) differences (APT base) between the APT and base monthly average sulfate concentrations (mg m 3 ). dependence less secondary sulfate is formed in winter than in summer because of the reduced photochemical activity in winter and the consequent reductions in the levels of oxidants. Both the base and APT simulations show the same spatial extent of power plant impacts, but there are large differences between the calculated contributions from the two models in the regions near the 14 sources. Fig. 6 shows that the maximum power plant contribution to ambient sulfate concentrations is 0.8 mgm 3 in the base simulation, while the corresponding contribution in the APT simulation is 0.5 mgm 3, about 38% lower than the base value. As in July 2002, the APT simulation generally results in a smaller contribution of the power plants to sulfate concentrations over most of the region impacted by the power plants in January However, the differences in January are somewhat smaller than those in July because of the reduced formation of secondary sulfate in January versus July (i.e., primary sulfate is a larger fraction of the total sulfate concentrations in January than in July). Fig. 7a shows the relative changes in the calculated power plant contributions to sulfate concentrations when CMAQ-MADRID-APT is used instead of CMAQ-MADRID, while Fig. 7b presents the actual sulfate concentration differences between the APT and base simulations. From Fig. 7a, we see that, in the APT simulation, the calculated contributions reduce by 2 17% from the base simulation values in most of Alabama, Georgia, and Mississippi, and parts of Tennessee. At larger distances from the 14 sources, in southern Kentucky and the Carolinas, the calculated contributions drop by 1 2% from the base values. Fig. 7b shows that the predicted sulfate concentrations in the APT simulation are about mgm 3 lower than the base simulation values in southwestern and northeastern Alabama, northwestern Georgia, southeastern Tennessee, and along the southern part of the boundary between Mississippi and Alabama, and also in small areas of southern Mississippi. The results for total inorganic nitrate are qualitatively similar to those for sulfate, but the differences between the APT and base simulations are smaller (not shown here). The calculated contributions in the APT simulation of the 14 power plants to inorganic nitrate concentrations are lower by about 1 10% from the base simulation contributions in the regions near the power plants. The maximum contribution of the 14 CFPPs to the monthly average inorganic nitrate concentrations for January 2002 is 0.28 mgm 3 in the APT simulation, only slightly lower than the maximum base simulation contribution of 0.31 mgm 3. To better understand the differences between the base and APT simulation results for sulfate and also to understand the seasonal differences, we conducted a mass-budget analysis to estimate the approximate conversion of the SO 2 emissions from the 14 CFPPs to sulfate (this mass budget is approximate because some fraction of the sulfur compounds is deposited during the simulation and SO 2 is removed from the atmosphere at different rates than sulfate). The analysis is based on the sulfate to total sulfur ratios in the power plant emissions and the domain-wide sulfate to total sulfur