Intercomparison of the Community Multiscale Air Quality Model and

Environ. Sci. Technol. 2005, 39, 5742-5753

Intercomparison of the Community Multiscale Air Quality Model and CALGRID Using Process Analysis

or CBIV chemical mechanisms in the CMAQ cases. These results demonstrate the need for specialized process field measurements to confirm whether we are modeling ozone with valid processes.

S U S A N M . O ’ N E I L L * ,† A N D BRIAN K. LAMB‡ Pacific Wildland Fire Sciences Laboratory, USDA Forest Service, 400 N 34th Street, Suite 201, Seattle, Washington 98103, and Washington State University, Pullman, Washington 99164-2910

1. Introduction

This study was designed to examine the similarities and differences between two advanced photochemical air quality modeling systems: EPA Models-3/CMAQ and CALGRID/CALMET. Both modeling systems were applied to an ozone episode that occurred along the I-5 urban corridor in western Washington and Oregon during July 11-14, 1996. Both models employed the same modeling domain and used the same detailed gridded emission inventory. The CMAQ model was run using both the CB-IV and RADM2 chemical mechanisms, while CALGRID was used with the SAPRC-97 chemical mechanism. Output from the Mesoscale Meteorological Model (MM5) employed with observational nudging was used in both models. The two modeling systems, representing three chemical mechanisms and two sets of meteorological inputs, were evaluated in terms of statistical performance measures for both 1- and 8-h average observed ozone concentrations. The results showed that the different versions of the systems were more similar than different, and all versions performed well in the Portland region and downwind of Seattle but performed poorly in the more rural region north of Seattle. Improving the meteorological input into the CALGRID/CALMET system with planetary boundary layer (PBL) parameters from the Models-3/CMAQ meteorology preprocessor (MCIP) improved the performance of the CALGRID/CALMET system. The 8-h ensemble case was often the best performer of all the cases indicating that the models perform better over longer analysis periods. The 1-h ensemble case, derived from all runs, was not necessarily an improvement over the five individual cases, but the standard deviation about the mean provided a measure of overall modeling uncertainty. Process analysis was applied to examine the contribution of the individual processes to the species conservation equation. The process analysis results indicated that the two modeling systems arrive at similar solutions by very different means. Transport rates are faster and exhibit greater fluctuations in the CMAQ cases than in the CALGRID cases, which lead to different placement of the urban ozone plumes. The CALGRID cases, which rely on the SAPRC97 chemical mechanism, exhibited a greater diurnal production/loss cycle of ozone concentrations per hour compared to either the RADM2 * Corresponding author phone: (206)732-7851; fax: (206)732-7801; e-mail: [email protected]. † USDA Forest Service. ‡ Washington State University. 5742

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 39, NO. 15, 2005

The Pacific Northwest is home to several large urban areas surrounded by forests, mountains, and agriculture. As population and industry have grown, periods of air quality degradation have arisen. One such episode occurred in July of 1996 when ozone concentrations approached and exceeded the National Ambient Air Quality Standards (NAAQS) at monitoring sites downwind of Seattle, WA and Portland, OR. These elevated levels occurred under light wind, sunny conditions. The predominant wind direction was from the northwest, and pollutants from the three major population areas of Vancouver, BC, Seattle, WA, and Portland, OR, along with the vehicle emissions from the Interstate-5 corridor, were transported to the southeast toward the Cascade Mountains. Thus high ozone levels occurred in relatively rural areas and near class I wilderness areas. Russell and Dennis (1), in their critical review of photochemical models, indicated that the greatest uncertainties in models are due to modeling inputs and the modeling process itself and that the models themselves have surpassed our ability to apply them well. Eulerian grid models are open systems; therefore, they cannot be subject to verification because only a closed system can be proved “true”. However, they can be subject to confirmation where confirmation is a matter of degree. Thus models are useful for corroborating a hypothesis, elucidating discrepancies in other models, and for exploring “what if” questions (2). This paper addresses these issues associated with photochemical grid modeling in particular by (1) evaluation of the CMAQ modeling system with a previously modeled Pacific Northwest domain and available observational data and (2) a detailed model intercomparison using process analysis results. This study continues the work of Barna et al. (3), Barna and Lamb (4), and Barna et al. (5) by applying the Models-3/CMAQ Eulerian grid modeling system to the Cascadia region of British Columbia, Washington, and Oregon. Process analysis (PA) is a novel approach developed by Jeffries and Tonnesen (6) that tracks the contribution of a process to the species conservation equation. CMAQ incorporates two types of process analysis: integrated process rate (IPR) and integrated reaction rate (IRR). IPR tracks each process’ contribution to the species conservation equation, where the major processes are horizontal advection, vertical advection/diffusion, chemical production/loss, and deposition. IRR tracks each chemical reaction’s contribution to the chemical production/destruction term in the species conservation equation. Jeffries and Tonnesen (6) developed process analysis to investigate ozone chemistry in the SAPRC90 and CBVI chemical mechanisms in a Lagrangian box model. Using VOC/NOX mixing ratios ranging from 5 to 38, they found that while the two mechanisms exhibited similar ozone maxima and reactivity, the SAPRC90 mechanism was more reactive and produced more ozone, especially at lower VOC/NOX mixing ratios. Jang et al. (7) and Jang et al. (8) applied IPR and IRR type PA, respectively, to investigate the effects of grid resolution on ozone formation and transport in the U.S. East Coast using the regional acid deposition model (RADM). Jiang et al. (9) coupled a backtrajectory analysis with process analysis results in the MM5/ 10.1021/es048403c CCC: $30.25

 2005 American Chemical Society Published on Web 06/29/2005

CALMET/CALGRID system to assess the nature of ozone formation and transport downwind of Seattle. In this study, IPR type PA is applied to show how two different modeling systems with three different chemical mechanisms arrive at their solutions. Several questions can be posed whose answers are formulated in this paper. First, how do the CMAQ and CALGRID modeling systems perform when compared to observations? Barna et al. (3) have addressed this with regard to the MM5/CALMET/CALGRID modeling system, but this is the first application of CMAQ to the Pacific Northwest. How do the two modeling systems compare to each other? What are the differences between the two model formulations? Thus an ensemble approach is applied with six different cases examined, and process analysis is invoked to help understand how the models arrive at their solutions. This study is unique because it is one of a few photochemical model intercomparisons utilizing both traditional statistical measures and process analysis. Furthermore, it is the first study presenting CMAQ results for the Pacific Northwest. This study is the first phase of a two-phase project designed to study ozone and aerosol formation in the Pacific Northwest as well as investigating the applicability of CMAQ to the region. O’Neill et al. (10) and Chen et al. (11) present results for the second phase of the research, which addresses aerosol formation as well. First, this paper briefly describes the domain, observational network used to verify the model results, and the system inputs: meteorology, emission inventory, and initial and boundary conditions. Sections 3 and 4 are discussions of model results for both traditional (statistical and time series) analysis and process analysis. Finally, concluding remarks are provided regarding the model intercomparison.

2. Domain Description, Observational Network, and Model Inputs The model domain, shown in Figure 1, extends from British Columbia to south of Portland Oregon and from the Pacific Ocean to just east of the Cascade Mountains. Key features of the domain are the rugged terrain and complex coastline. The major cities of Vancouver, BC, Seattle, WA, and Portland, OR lie along the interstate highway 5 (I-5) corridor between the coastal range and Cascade Mountains. The Columbia River Gorge basin passes south-centrally through the domain from east to west. Although the gorge is a subgrid scale feature, it nevertheless influences wind patterns in the Portland region. A network of 12 pollutant (ozone) observation stations is distributed along the I-5 corridor and in the Cascade Mountains as shown in Figure 1. Unfortunately, ozone precursor observations are not available. Three stations are located between Seattle, WA and Vancouver, BC: Custer, Darrington, and Getchel. Custer and Getchel are located directly downwind of I-5, while Darrington is further east within the Cascade Mountains. Five stations are located within and downwind of the Seattle/Puget Sound region: Lake Sammamish, Enumclaw, Paradise, Pack Forest, and Packwood. Lake Sammamish is near the urban core of Seattle. Enumclaw is located in the foothills of the Cascade Mountains southeast of Seattle. Pack Forest lies southeast of Seattle and I-5, while Packwood is in a rural area further to the east and in the lee-side of Mount Rainier. Paradise is on the side of Mount Rainier at approximately 1700 m above sea level. Four stations are located in the Portland region: Sauvie Island, Mountain View, Milwaukie High School, and Carus. All these stations are distributed in a north-south line along the I-5 corridor passing through Portland, OR. MM5 version 2 was applied to simulate the wind field for the period of July 11-15, 1996. A three-nested domain with

FIGURE 1. Map of the modeling domain showing terrain height and locations of 12 surface ozone monitors where CA ) Carus, MH ) Milwaukie High School, MV ) Mountain View, SI ) Sauvie Island, PW ) Packwood, PF ) Pack Forest, PA ) Paradise, EW ) Enumclaw, LS ) Lake Sammamish, GE ) Getchel, DA ) Darrington, and CU ) Custer. 45 km, 15 km, and 5 km grid cell spacing was selected where the 45 km domain extended across the Pacific Ocean to capture the synoptic scale flow features. Four-dimensional data assimilation (FDDA) analysis nudging was employed in the 45 km domain solution, and FDDA observational nudging (“obs-nudging”) was applied in the 5 km domain. Barna and Lamb (4) completed an in-depth study of the meteorological wind field applied in this work. They investigated three meteorological model configurations to determine the optimal wind field to use for this modeling period. The first wind field was created by MM5 and did not employ observational nudging, and the second wind field combined the first MM5 solution with CALMET’s objective analysis feature to adjust the MM5 predictions toward the observed winds. This was a time-consuming iterative approach. The third case was an MM5 simulation incorporating observational nudging. In addition to evaluating predicted meteorology against a network of 45 surface stations clustered mostly along the I-5 corridor (north-south west of the Cascade Mountains), performance of the meteorological model was also evaluated in terms of system success in predicting ozone concentrations. Barna and Lamb (4) found that the MM5 observational nudged case used here performed well in the Seattle area, where the meteorological model results compare well with the surface observations in terms of wind speed, wind direction, and temperature. Few data were available to evaluate MM5 predictions in the mountainous regions. MM5 performance in the Portland region was not quite as good but much improved in the MM5 observational nudging case versus the other two cases. In both MM5 runs (observational nudged and non-nudged), MM5 predicts easterly flow out of the Columbia River Gorge into Portland, while the observations show a westerly and northwesterly flow. The observational nudging case however has lower wind speeds as compared to the nonobservational nudged case. This has large implications for air quality VOL. 39, NO. 15, 2005 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

5743

modeling since winds out of the Gorge advect clean air to the Portland urban area diluting ozone and ozone precursors. CALMET serves as the meteorological preprocessor for CALGRID, while MCIP is the meteorological preprocessor for CMAQ. Both models rediagnose the PBL parameters and, thus, introduce differences into the meteorology input to the two Eulerian grid modeling systems. Two approaches were taken with the CALMET meteorology. First CALMET was run in a pass through mode using only the MM5 obsnudged solution as input. The second approach involved replacing the CALMET PBL parameters with those from MCIP to reduce differences between the two modeling systems. The following PBL parameters were substituted from MCIP into CALMET: friction velocity, convective velocity scale, planetary boundary layer height, air temperature, MoninObukhov length, roughness length, and terrain elevation. The horizontal and vertical structures employed in both the CMAQ and CALGRID modeling systems were very similar. Both used 71 columns, 130 rows, and 17 layers. Of the 17 layers, 16 were within the maximum PBL height of approximately 4 km. The 17th layer in the CMAQ simulations extended to the top of the MM5 domain, while the CALGRID 17th layer extended to 4.7 km. Vertical cross-sections of the 4-day period shows that mixing and ozone generation were contained within the lowest 10 layers (∼1000 m) in both modeling systems. CALGRID also had a 20 m first layer, while the first layer in MCIP/CMAQ was 40 m deep. Also, the CALGRID vertical layers were specified using the average MCIP cell face values to minimize vertical interpolation of the mm5 wind field in CALGRID and to make the two modeling systems as similar as possible. CMAQ (12) represents the latest state-of-the-science in air quality models. Developed by EPA, it represents a concerted effort on the part of the scientific community to develop a modeling system using a one-atmosphere approach (13). CMAQ solves the species conservation equation (14) and offers a choice of chemical mechanisms and a choice of chemical solvers. It includes cloud processes, aerosol processes, plume-in-grid modeling, and advanced diagnostics such as process analysis (15). In this study, both the RADM2 and CBIV chemical mechanisms were applied using CMAQ v4.1. Process analysis was invoked in both cases. The MM5/CALMET/CALGRID (v1.6, level 970627) modeling system is an Eulerian grid modeling system that also solves the species conservation equation (16). It assumes clear-sky photolysis, a reasonable assumption for this photochemical event (4), and includes the SAPRC97 chemical mechanism (17). CALGRID applies operator splitting to individually solve the components of the species conservation equation. Thus the code is modular, and it was a straightforward task to implement IPR type process analysis. This was accomplished by saving the concentration field before the vertical advection/diffusion, horizontal advection/diffusion, and chemistry subroutines and then taking the difference between the saved and new concentration fields after each process. This is similar to the logic applied within CMAQ. While both modeling systems are similar in their mathematical formulation solving the species conservation equation, they are fundamentally different in how mass consistency is achieved. Byun (18) discusses how mass conservation is the most important physical constraint for an air quality simulation and offers formulations for achieving it in (19), while Lee et al. (20) illustrate the effects it can have within an air quality model. CALGRID’s meteorological prepreprocessor, CALMET, recalculates the vertical wind component to satisfy the continuity equation (21). CMAQ uses the methods described in refs 12 and 18 to incorporate the mass consistency adjustment within the air quality module itself and treats mass consistency as another process in the species 5744

9


conservation equation. This fundamental difference has implications for the transport processes in the two modeling systems. The Washington State Department of Ecology compiled an extensive emission inventory for the SAPRC97 chemical mechanism. Mobile, area source, and point source emissions were then respeciated to convert the SAPRC97 lumped chemical species to CBIV and RADM2 lumped chemical species for input into the CMAQ modeling system. The respeciation from SAPRC97 to RADM2 was a straightforward process. The respeciation from SAPRC97 to CBIV was less straightforward and required allocation of species based on carbon bond characteristics. The GLOBEIS biogenic emission model (22) was applied directly to obtain speciated biogenic emissions for each of the three chemical mechanisms. GLOBEIS requires input of wind speed, temperature, short wave and long wave radiation, and water vapor mixing ratio from MM5 to estimate emissions of 18 terpene species, isoprene, and an “other” VOC category. The 18 terpene species were summed and allocated to the TERPB category in RADM2 and CBIV and the OLE3 category in SAPRC97. The other VOC category was speciated for each chemical mechanism as specified by ref 23. Boundary conditions, based on the work of Jiang (24), were set to clean Pacific air conditions on the western boundary and clean-forested conditions were applied on the north, south, and east boundaries. The clean-forested conditions were used as model initial conditions.

3. Results To address the first question posed, “how do the CMAQ and CALGRID models perform compared to the observations and to each other?”, ozone formation and transport are examined across the domain, time series are analyzed at the 12 monitoring stations, and statistical measures are calculated by grouping the monitoring sites regionally. Six cases were obtained as follows: Case 1. CMAQ/RADM2 Case 2. CMAQ/CBIV Case 3. CALGRID/C Case 4. CALGRID/CM

Case 5. 1E Case 6. 8E

CMAQ with the RADM2 chemical mechanism CMAQ with the CBIV chemical mechanism CALGRID with the SAPRC97 chemical mechanism, CALMET only meteorology CALGRID with the SAPRC97 chemical mechanism and CALMET meteorology improved with MCIP PBL parameters 1-h ensemble average of the cases 1-5 8-h ensemble average of cases 1-5

Cases 1-4 involve two modeling systems, three meteorological simulations, and three chemical mechanisms. Case 5 is an ensemble average of the first five cases and was performed to determine if an ensemble approach improved the model performance. Finally, because it has been indicated that air quality models perform better over a longer time period (25) and because of the new 8-h ozone National Ambient Air Quality Standard (NAAQS), performance of an 8-h ensemble average is investigated. Ozone concentration contour maps are shown in Figure 2 for Sunday July 14, 1996 at 3 p.m. PDT. Figure 2a depicts ozone levels for case 4 (CALGRID/CM), while Figure 2b depicts ozone levels for case 2 (CMAQ/CBIV). These contours are indicative of the behavior of ozone formation and transport on all 4 days. Sunday is of special interest because that was the day the highest ozone levels were achieved. High NOX (NO2 + NO) sources in the urban areas caused

FIGURE 2. Ozone contours for Sunday July 14, 1996 at 3 pm PDT from (a) CALGRID/CM, (b) CMAQ/CBIV, and (c) an ozone difference plot obtained by subtracting Figure 3a from Figure 3b. O3 titration in the urban cores; however, the NOX then mixed with the biogenic and anthropogenic VOCs to create O3 downwind. The prevailing winds from the northwest yielded areas of elevated ozone concentrations to the south and east of the urban areas along the western slopes of the Cascade Mountains. On first glance, the models appear to perform similarly, however, taking the difference in the contours in parts a and b yields part c of Figure 2, which indicates the following: (1) the CMAQ/CBIV gave greater ozone levels across the Cascade Mountains, (2) the CALGRID/CM solution exhibited greater ozone losses within the urban core areas of Seattle and Puget Sound, and (3) the urban ozone plume placement was different between the two modeling systems. Ozone concentration time series at four of the 12 monitoring stations for cases 1-5 are qualitatively compared with observations in Figure 3. The model results exhibited the diurnal variation expected in ozone levels for sites located in or near urban areas (e.g., Milwaukie High School, Figure 3a). Paradise (Figure 3d) and Packwood, located in the vicinity of Mount Rainier, are the exceptions; these sites exhibit weak diurnal patterns and more fluctuation than observed at the other sites. The Paradise monitor is unique in that it is located at 1650 m above mean sea level. Barna et al. (3) note that similar weak diurnal O3 concentration patterns have been observed by researchers at other high-elevation monitors. The two CALGRID simulations typically performed better at achieving the nighttime ozone lows. The CALGRID/CM case tended to achieve the ozone peaks better than the CALGRID/C case except at Lake Sammamish (Figure 3c) where the CALGRID/C case was the optimal performer. All modeled solutions underestimated the peak ozone concentrations on Sunday at Milwaukie High School (Figure 3a) and Enumclaw (Figure 3b). As a quantitative measure of model performance the following statistical measures were applied: normalized gross error, normalized bias, and index of agreement. The normalized gross error is calculated by

E)

1

N

|pi - oi|

i)1

oi

∑ N

(1)

where pi and oi are the predicted and observed ozone

concentrations, respectively, for measurement i. N is the total number of measurements. The normalized bias is calculated by

B)

N

pi - oi

i)1

oi

1

∑ N

(2)

where a normalized gross error of 35% and a bias less than (15% are suggested by EPA as indicative of acceptable ozone modeling performance. The index of agreement is calculated by N

∑(p - o ) i

I)1-

2

i

i)1

(3)

N

∑(|p - oj| + |o - oj|) i

2

i

i)1

where oj denotes the average observed ozone concentration and a value of 1 indicates perfect agreement between predicted and observed values. Statistical performance is presented on a regional basis by grouping the stations together into three regions: the Portland region, the Seattle region, and the North Seattle region. Table 1 gives the statistics for each of the six cases for each of the regions. The data were calculated using the entire 4-day period, excluding observations less than 10 ppb. The six cases all performed well in the Portland area with indices of agreement ranging from 0.82 to 0.91, correlation coefficients from 0.68 to 0.86, normalized bias less than (15%, and normalized gross errors less than 35%. The 8-h ensemble averaged case performed best across most statistical measures. This result echoes findings reported by Hogrefe et al. (25) and indicate that the models capture the longer averaging periods better than the shorter averaging periods. The next best performer was the 1E case for the statistical measures of index of agreement and correlation coefficient and the CALGRID/CM case for the normalized bias and normalized gross error statistics. However, all cases performed similarly, with the CALGRID/C case typically performing worst. Using the boundary layer parameters from MM5/MCIP improved VOL. 39, NO. 15, 2005 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

5745

FIGURE 3. Ozone time series at four of the 12 monitoring stations for the 4-day period of July 11-14, 1996 where the hour indicates elapsed time from 0000 PDT July 11, 1996. the CALGRID simulation, which can be seen in the improved index of agreement, correlation coefficient, and normalized gross error statistics in the CALGRID/CM case. The three CMAQ cases performed very similarly. Model performance at the Seattle downwind stations was not uniform. The six modeled cases typically performed well at Enumclaw, WA; however, all cases failed to capture the peak ozone concentration of 118 ppb measured on Sunday 5746

9


July 14, 1996. The 8E case performed best overall at this site although all six cases had normalized gross errors less than 35%, and the ensemble cases and CALGRID cases had normalized bias’s within a range of (15%. The CALGRID/ CM and CMAQ/CBIV cases captured the peak ozone concentrations most successfully. It is important to note that at Enumclaw, surrounding grid heterogeneity exists where a neighboring grid cell can yield substantially different ozone time series. Combining the five Seattle stations yielded a picture of the overall statistical performance of the models in this region. The correlation coefficients ranged from 0.31 for the CMAQ/ RADM2 case to 0.60 for the CALGRID/C case. The index of agreement values ranged from 0.71 (CMAQ/CBIV) to 0.83 for the CALGRID/C case. Both CMAQ cases had normalized bias and gross error statistics greater than the EPA recommended values of 15% and 35%, respectively. The two CALGRID solutions and both the ensemble cases had acceptable values with the 8E case exhibiting the lowest gross error of 23%. The CALGRID/C case had the best normalized bias of 2%. Statistically the CALGRID/C case tended to perform best; however, that is mostly due to the fact that it captured the nighttime ozone concentrations more successfully than the other cases. Only at the Lake Sammamish site can it be considered more successful than the other cases at simulating the ozone maxima. Three sites are located north of the Seattle area: Getchel, Darrington, and Custer. Getchel and Custer are primarily impacted by vehicle emissions, while Darrington lies east of I-5 in a valley at the foot of the Cascade Mountains. Custer is located near the border with Canada. At all three stations, the models had difficulties simulating the near zero ozone levels at night and often overshot the weak maximums during the day. The grid cells neighboring Custer and Getchel achieved much lower ozone concentrations at night indicating that ozone titration was occurring but was highly localized. At Darrington peak ozone concentrations were achieved when westerly flow funneled pollutants into the valley. At night, the winds became weak and variable, and as a result ozone levels were maintained near background concentrations in the models while in reality ozone titration occurred. The CALGRID cases were more successful at removing ozone at night, while the CMAQ cases were more successful at achieving the ozone maximum values with some timing problems on the first and third days due to weakening west winds in the model. It is hypothesized that stronger west winds throughout the 4-day period would improve the performance of all the modeled cases at Darrington. The 8E case had the best performance of normalized bias (24%s which still exceeds the EPA limits) and an acceptable normalized gross error of 28%. However this case also exhibited the poorest correlation and index of agreement of the six cases. The CALGRID/C case exhibited the best correlation and index of agreement. In summary, all model cases performed optimally in the southern and central regions of the domain, which are the areas where the largest ozone concentrations were observed, and less optimally in the more rural northern portion of the domain. In general, modifying the CALMET meteorology with the MCIP parameters improved the peak concentrations predicted by the CALGRID/CM case. The CMAQ/ CBIV case consistently achieved the maximum ozone concentrations better across the 4-day period with the exception of the Lake Sammamish and Custer sites where the CALGRID/C and CALGRID/CM, respectively, achieved the maximum ozone concentrations. The 8E case, while often performing well statistically, tended to underpredict ozone concentrations and in many cases failed to predict exceedances of the proposed 80 ppb 8-h ozone air quality standard.

TABLE 1. Summary Statistics Calculated for the Portland, Seattle, and North Seattle Regions for Each of the Six Cases for the Period of July 11-14, 1996a Portland Stations

avg

max

min

std dev

RMSD

Y-intercept

slope

correlation coefficient

index of agreement

normalized bias

normalized gross error

8E 1E CMAQ/RADM2 CMAQ/CBIV CALGRID/C CALGRID/CM 1-h observations 8-h observations

51 51 52 54 46 50 52 52

91 104 110 112 88 102 145 130

22 10 6 7 11 17 10 10

16 18 20 21 16 19 28 25

12 14 14 14 19 15

21 21 20 21 22 19

0.58 0.59 0.62 0.64 0.47 0.58

0.86 0.81 0.81 0.78 0.68 0.75

0.91 0.90 0.91 0.91 0.82 0.89

10 14 15 18 5 10

22 27 27 29 32 28

Seattle Stations

avg

max

min

std dev

RMSD

Y-intercept

slope


index of agreement

normalized bias



49 48 50 50 43 44 47 48

83 94 90 99 89 100 118 92

28 16 11 10 17 18 10 13

11 14 14 17 15 15 20 16

11 14 17 17 13 14

26 25 33 28 16 18

0.48 0.49 0.38 0.47 0.57 0.55

0.52 0.51 0.31 0.34 0.60 0.53

0.80 0.80 0.71 0.74 0.83 0.81

9 16 24 23 2 6

23 32 39 38 27 31

North Seattle Stations

avg

max

min

std dev

RMSD

Y-intercept

slope


index of agreement

normalized bias



51 48 52 49 45 45 35 42

67 75 75 75 74 79 67 59

37 26 27 26 22 21 10 24

7 10 10 10 11 12 16 8

12 18 21 20 16 17

43 36 40 39 30 31

0.18 0.34 0.32 0.28 0.42 0.40

0.05 0.32 0.25 0.18 0.36 0.29

0.40 0.62 0.58 0.57 0.66 0.64

24 66 81 73 51 51

28 71 83 78 59 61

a

The range of standard deviation values for case 1E are calculated from the ensemble average of the three CMAQ and two CALGRID.

4. Process Analysis Process analysis is the means by which the relative contribution to the species conservation equation of each individual process is tracked within the modeling system. The CMAQ system has process analysis capabilities implemented within its architecture, while the CALGRID system required special modifications. The following individual processes were tracked for all grid cells within the CMAQ system: x-direction advection, y-direction advection, z-direction advection, horizontal diffusion, vertical diffusion, mass consistency adjustment, chemistry transformation, and dry deposition. The following processes were tracked in the CALGRID system: horizontal advection, vertical advection, dry deposition, and chemistry transformation. For model intercomparison purposes, the following relations were examined: horizontal advection/diffusion processes, vertical processes (including the mass consistency adjustment in CMAQ and dry deposition), and chemical transformation. The process analysis data are presented for each process in two forms. The first form is a depiction of the change in ozone concentration (∆ O3) for a particular hour. The second form depicts the cumulative effect of the particular process. Process analysis data were extracted for three of the six cases: CMAQ/ CBIV, CMAQ/RADM2, and CALGRID/CM. 4.1. Portland Region. Regional flow patterns through the Portland area for the 4-day period of July 11-14, 1996 were characterized by steady N-NNW winds during the days. The wind directions during the nighttime periods exhibited greater fluctuations and direction changes, often with downslope flows from the mountains southeast of Portland. Sauvie Island lies to the north (upwind) of Portland, while Mountain View is located in a suburban neighborhood of Vancouver,

WA, also north of Portland. Milwaukie High School lies slightly south of Portland within the urban core. The mountains southeast of Portland serve as a barrier to transport and help produce the highest modeled ozone concentrations at Carus, which lies at the foot of the mountains downwind of Portland. These differences in location are important in understanding the processes that drive ozone behavior at each of the sites. Figure 4(a-h) depicts change in ozone concentration per hour and the running total of ozone concentration due to chemistry for each of the four Portland monitoring stations. Milwaukie High School exhibited the greatest chemical production of the four sites with 10-20 ppb ozone per hour on Thursday and Friday and 20-35 ppb ozone per hour on Saturday and Sunday. This site also exhibited nighttime scavenging of ozone by NO, thus the time series of all the mechanisms achieved almost zero concentrations at night and the process analysis data reflect this by showing the loss of 8-14 ppb ozone per hour. The chemistry process analysis data for Carus and Sauvie Island were similar to Milwaukie High School, with reduced peaks and valleys. Mountain View was quite different from the other three stations chemically. Overall there was a chemical loss of ozone during most of the 4 days. The SAPRC97 mechanism shows a minimal chemical production on Thursday and Friday of 4 ppb ozone per hour and then exhibited ozone production levels of 15 ppb ozone per hour and 28 ppb ozone per hour on Saturday and Sunday, respectively. The CBIV and RADM2 mechanisms showed a loss of ozone for the entire period with the exception of a 2-6 ppb ozone per hour production on Sunday. Such behavior was expected given that this is a large NOX emission cell. In general, the SAPRC97 chemical mechanism exhibited greater peaks and valleys than the other VOL. 39, NO. 15, 2005 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

5747

FIGURE 4. Chemistry process analysis data for the Portland, OR region stations. Figures in the left column show the change in ozone concentration per hour, and figures on the right show the cumulative ozone concentrations due to chemistry over the 4-day period. two mechanisms. This enhanced reactivity is consistent with results by ref 6. Differences in the models become more apparent when examining the horizontal and vertical process analysis data shown in Figure 5(a-h), which displays the running totals of horizontal and vertical advection and diffusion processes for each of the four Portland monitoring sites. The CMAQ/ CBIV and CMAQ/RADM2 cases behaved very similarly but were often quite different from the CALGRID/CM case. Horizontal advection at Sauvie Island, upwind of the Portland urban area, produced a net gain of ozone in the grid cell over the 4-day period, while vertical advection/diffusion was responsible for a net loss. At Sauvie Island, all three models performed similarly. Such was not the case at suburban Mountain View where the CMAQ cases collectively were very different from the CALGRID/CM case, exhibiting greater 5748

9


ozone production and losses. Horizontal and vertical advection/diffusion processes were much weaker for all three cases at the urban Milwaukie High School site. Finally, the two CMAQ cases were collectively different from the CALGRID/CM case at Carus, downwind of the Portland metro area and lying near the foot of the Cascade Mountains. Differences become more pronounced between the two models as the flow progresses across the emission-rich urban areas and encounters complex terrain. 4.2. Seattle Region. Regional flow patterns through the Seattle area for the 4-day period of July 11-14, 1996 were similar to those in the Portland area with the prevailing winds from the N-NNW during the daytime periods, while the nighttime periods exhibited greater fluctuations and direction changes, often with down-slope flows from the Cascade Mountains lying to the east, and easterly flow through the

FIGURE 5. Horizontal (left column) and vertical (right column) advection/diffusion process analysis data for the Portland, OR region stations. passes. The Cascade Mountains serve as a barrier to transport and the highest modeled ozone concentrations occur for the region near Enumclaw, which lies at the foot of the mountains downwind of the Seattle urban area. Lake Sammamish is in an urban area next to Seattle, while Pack Forest is further south of Enumclaw and also situated at the base of the mountains downwind of the Seattle urban area. Paradise and Packwood are located in the in rural complex terrain near and downwind of Mount Rainier, respectively. Figure 6(a-h) depicts change in ozone concentration per hour and the running total of ozone concentration due to chemistry for each of the five Seattle region monitoring stations. Lake Sammamish exhibited both chemical production and loss of ozone with the net effect that there was net depletion of ozone due to chemistry in the surface layer cell over the 4-day period. The CALGRID/CM case, which relies on the SAPRC97 chemical mechanism, exhibited the greatest diurnal production/loss cycle of ozone concentration per

hour with production at 18 ppb/h on Sunday and losses at 21 ppb/h Thursday night. In comparison, the CMAQ/RADM2 and CMAQ/CBIV cases produced 3-10 ppb ozone per hour, respectively, and reduced ozone concentrations by 18 ppb/h. Downwind of Seattle at Pack Forest and Enumclaw, the chemical process rate results were very similar. At both locations, the urban plume produced 10-20 ppb ozone per hour during the daytime periods, and chemistry was a net addition to the grid cell. All three chemical mechanisms produced ozone at different rates on Thursday and Friday but behaved similarly on Saturday and Sunday. Chemistry was a minor contributor to ozone concentrations at the Paradise and Packwood sites, with the CMAQ/CBIV and CMAQ/RADM2 cases producing slightly more ozone than the CALGRID/CM case. This is possibly due to increased advection/diffusion of pollutants from the urban corridor, as indicated in the horizontal and vertical process analysis data in Figure 7. VOL. 39, NO. 15, 2005 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

5749

FIGURE 6. Chemistry process analysis data for the Seattle, WA region stations. Figures in the left column show the change in ozone concentration per hour, and figures on the right show the cumulative ozone concentrations due to chemistry over the 4-day period. The running totals of the horizontal and vertical process analysis data are displayed in Figure 7(a-h) for each of the five Seattle region stations. Large differences exist between the two CMAQ solutions and the CALGRID case in these data for the downwind urban sites (Enumclaw and Pack 5750

9


Forest) and the rural sites in complex terrain (Paradise and Packwood). Only at the urban site, Lake Sammamish, are the horizontal and vertical process analysis data similar for all three cases. Vertical transport of ozone into the surface layer cell was a large contributor to ozone concentrations at Lake

FIGURE 7. Horizontal (left column) and vertical (right column) advection/diffusion process analysis data for the Seattle, WA region stations. Sammamish (data not shown) because much of the chemistry was occurring aloft. Process analysis data trends at Enumclaw and Pack Forest behaved similarly with horizontal advection/ diffusion being an important contributor to ozone concen-

trations in the two CMAQ cases and a weak sink in the CALGRID/CM case. The trends were reversed in the vertical process analysis data. Paradise and Packwood, situated in the complex terrain of Mount Rainier, were dependent almost VOL. 39, NO. 15, 2005 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

5751

FIGURE 8. The mass consistency adjustment and chemistry process analysis data from the CMAQ model at Enumclaw, WA over the 4-day modeling period. solely upon horizontal and vertical processes for ozone concentrations. It is in these areas of complex terrain that model differences become most pronounced, and it becomes obvious that large differences exist regarding how the models arrive at their solutions. 4.3. Process Analysis Discussion. The process analysis data yields insights into why the differences seen in Figure 2c occur between the CMAQ and CALGRID modeling systems. These differences can be related to the wind flow and the treatment of the wind field which then relates to the mass consistency adjustment made by the two modeling systems. Clean flow enters the domain from the W-NW and encounters pollutant emissions from the urban I-5 corridor that runs north-south from Vancouver, BC to south of Portland, OR. Flow then progresses over the Cascade Mountains. Figure 2c shows few differences between the two modeling systems until emission sources are encountered. Differences in the horizontal and vertical advection/diffusion cause locally different mixing which results in a somewhat different placement of the urban ozone plumes. Then as the terrain becomes more complex across the Cascade Mountains, the enhanced horizontal and vertical advection/diffusion in CMAQ brings greater ozone concentrations to the rural monitors of Paradise and Packwood than CALGRID predicts. Upwind or within the urbanized areas the horizontal and vertical process analysis data are similar between the two modeling systems (e.g. at Sauvie Island, Milwaukie HS, and Lake Sammamish); however, once downwind of the urban area (Carus, Mountain View, Enumclaw, and Pack Forest) or in rural complex terrain (Packwood and Paradise) then the process rates become quite different between the CMAQ and CALGRID cases. The CMAQ cases consistently exhibit larger additions and subtractions of O3 from a grid cell than the CALGRID case. This is possibly related to how the two arrive at a mass consistent wind field. Thus it is necessary to provide some insight into the differences that result between the two modeling systems due to the mass consistency adjustment. The vertical wind component is rediagnosed by CALMET to provide CALGRID with a mass consistent wind field. Investigation of the vertical wind components between the two modeling systems showed that the CALMET/CALGRID vertical winds were approximately 1/2 to an order of magnitude less than the vertical winds in CMAQ. Then, Figure 8 compares the mass consistency adjustment parameter with ozone formation at the Enumclaw, WA site for the CMAQ/ CBIV case. The mass consistency adjustment is quite large and dwarfs chemical production at Enumclaw; similar 5752

9


behavior occurs at the other sites. It should be noted that the mass consistency adjustment should not be viewed alone but should always be included as part of the transport processes. In this study, the mass consistency adjustment was included as one of the vertical transport processes in Figures 5 and 7 and is shown explicitly in Figure 8 to illustrate the importance of addressing it in air quality modeling. These behaviors also help explain the differences seen in Figure 2. Only when coupling the air quality with the meteorological model or by using a meteorological model formulated on the basis of air density and entropy (18), can the issue of mass consistency be removed from eulerian air quality modeling. Work by Kim et al. (26) with the WRF meteorological model shows promise of achieving this.

Acknowledgments Many thanks to Guangfeng Jiang for his chemical speciation expertise, Alex Guenther for providing the GLOBEIS biogenics model, and Robert Yamartino for providing updates to the CALGRID model. This project was funded by the EPA through the Northwest Regional Modeling Center Models-3/CMAQ Demonstration Project. Computing facilities were provided by the Center for Multiphase Environmental Research (CMER) at Washington State University. S. O’Neill was funded by a 3-year EPA Science-To-Achieve-Results (STAR) fellowship.

Literature Cited (1) Russell, A.; Dennis, R. NARSTO critical review of photochemical models and modeling. Atmos. Environ. 2000, 34, 2283-2324. (2) Oreskes, N.; Shrader-Frechette, K.; Belitz, K. Verification, validation, and confirmation of numerical models in the earth sciences. Science 1994, 263, 641-645. (3) Barna, M. G.; Lamb, B. K.; O’Neill, S. M.; Westberg, H.; FigueroaKaminsky, C.; Otterson, S.; Bowman, C.; DeMay, J. Modeling ozone formation and transport in the cascadia region of the Pacific Northwest. J. Appl. Meteorol. 2000, 39, 349-366. (4) Barna, M. G.; Lamb, B. K. Improving ozone modeling in regions of complex terrain using observational nudging in a prognostic meteorological model. Atmos. Environ. 2000, 34, 4889-4906. (5) Barna, M. G.; Lamb, B. K.; Westberg, H. Modeling the effects of VOC/NOX emissions on ozone synthesis in the Cascadia airshed of the Pacific Northwest. J. Air Waste Manage. Assoc. 2001, 51, 1021-1034. (6) Jeffries, H. E.; Tonnesen, S. A comparison of two photochemical reation mechanisms using mass balance and process analysis. Atmos. Environ. 1994, 28, 2991-3003. (7) Jang, J. C. C.; Jeffries, H. E.; Byun, D.; Pleim, J. E. Sensitivity of ozone to model grid resolution - I. Application of high-

(8) (9)

(10)

(11)

(12) (13)

(14)

(15)

resolution regional acid deposition model. Atmos. Environ. 1995, 29, 3085-3100. Jang, J. C. C.; Jeffries, H. E.; Tonnesen, S. Sensitivity of ozone to model grid resolution - II. Detailed process analysis for ozone chemistry. Atmos. Environ. 1995, 29, 3101-3114. Jiang, G.; Lamb, B. K.; Westberg, H. Using back trajectories and process analysis to investigate photochemical ozone production in the Puget Sound region. Atmos Environ. 2003, 37, 14891502. O’Neill, S. M.; Lamb, B. K.; Chen, J.; Claiborn, C.; Finn, D.; Otterson, S.; Figueroa, C.; Bowman, C.; Boyer, M.; Wilson, R.; Arnold, J.; Aalbers, S.; Stocum, J.; Swab, C.; Stoll, M.; DuBois, M.; Anderson, M. Modeling Ozone and Aerosol Formation and Transport in the Pacific Northwest with the Community Multiscale Air Quality (CMAQ) Modeling System. Submitted to Environ. Sci. Technol. 2004. Chen, J.; Lamb, B. K.; O’Neill, S. M. Aerosol and Ozone Sensitivity Analysis with the Community Multi-Scale Air Quality (CMAQ) Model for the Pacific Northwest. Submitted to Environ. Sci. Technol. 2004. Byun, D. W.; Ching, J. K. S. Science algorithms of the EPA models-3 Community Multiscale Air Quality (CMAQ) modeling system. EPA/600/R-99/030, U.S. EPA, 1999. Dennis, R. L. The environmental protection agency’s third generation air quality modeling system: An overall perspective. 10th Joint Conference on the Applications of Air Pollution Meteorology with the A&WMA. Jan 11-16, 1998, Phoenix, AZ, pp 255-258. Byun, D.; Young, J.; Gipson, G.; Godowitch, J.; Binkowski, F.; Roselle, S.; Benjey, B.; Pleim, J.; Ching, J.; Novak, J.; Coats, C.; Odman, T.; Hanna, A.; Alapty, K.; Mathur, R.; McHenry, J.; Shakar, U.; Fine, S.; Ziu, A.; Jang, C. Description of the Models-3 Community Multiscale Air Quality (CMAQ) Modeling System. 10th Joint Conference on the Applications of Air Pollution Meteorology with the A&WMA. Jan 11-16, 1998, Phoenix, AZ, pp 264-268. Ching, J.; Byun, D.; Young, J.; Binkowski, F.; Pleim, J.; Roselle, S.; Godowitch, J.; Benjey, W.; Gipson, G. Science Features in Models-3 Community Multiscale Air Quality System. 10th Joint Conference on the Applications of Air Pollution Meteorology with the A&WMA. Jan 11-16, 1998, Phoenix, AZ, pp 269273.

(16) Yamartino, R. J.; Scire, J. S.; Carmichael, G. R.; Chang, Y. S. The CALGRID mesoscale photochemical grid model. Part I: Model formulation. Atmos. Environ. A 1992, 26, 1493-1512. (17) Carter, W. P. L. A detailed mechanism for the gas-phase atmospheric reactions of organic compounds. Atmos. Environ. A 1990, 24, 481-510. (18) Byun, D. W. Dynamically Consistent Formulations in Meteorological and Air Quality Models for Multiscale Atmospheric Studies. Part II: Mass Conservation Issues. J. Atmos. Sci. 1999, 56, 3808-3820. (19) Byun, D. W. Dynamically Consistent Formulations in Meteorological and Air Quality Models for Multiscale Atmospheric Studies. Part I: Governing Equations in a Generalized Coordinate System. J. Atmos. Sci. 1999, 56, 3789-3807. (20) Lee, S. M.; Yoon, S. C.; Byun, D. W. The effect of mass inconsistency of the meteorological field generated by a common meteorological model on air quality modeling. Atmos. Environ. 2004, 38, 2917-2926. (21) Scire, J. S.; Robe, F. R.; Fernau, M. E.; Yamartino, R. J. A user’s guide for the CALMET meteorological model (version 5); Earth Tech Inc.: Concord, MA, 2000. (22) Guenther, A.; Geron, C.; Pierce, T.; Lamb, B.; Harley, P.; Fall, R. Natural emissions of nonmethane volatile organic compounds, carbon monoxide, and oxides of nitrogen from North America. Atmos. Environ. 2000, 34, 2205-2230. (23) Guenther, A. personal communication, 2001. (24) Jiang, G. Ph.D. Thesis, Department of Chemistry, Washington State University, Pullman, WA, 2001. (25) Hogrefe, C.; Rao, S. T.; Kasibhatla, P.; Hao, W.; Sistla, G.; Mathur, R.; McHenry, J. Evaluating the performance of regional-scale photochemical modeling systems: Part II - ozone predictions. Atmos. Environ. 2001, 35, 4175-4188. (26) Kim, S. B.; Byun, D. W.; Kim, S. T.; Fong, N.; Skaramock, W. Mass consistent on-line and off-line air quality modeling paradigms: WRF linked with CMAQ. NCAR MM5 Workshop, 2004. www.mmm.ucar.edu/mm5/workshop/ws04/Session4/ Kim.Seung-Bum.pdf.

Received for review October 12, 2004. Revised manuscript received May 11, 2005. Accepted May 25, 2005. ES048403C

VOL. 39, NO. 15, 2005 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

5753

Intercomparison of the Community Multiscale Air Quality Model and

Recommend Documents