Abstract

This report is in DRAFT form. It is the second part of an assessment of regional transport issues in the Northeast and how well the transport and its effects are represented by models applied to the area.

Contents

• Preface and acknowledgments
• 1. Introduction
  • 1.1 Model Evaluation Objectives
  • 1.2 Modeling Approach
• 2. Modeling Performance for Ozone in the Northeast Corridor
  • 2.1 Approach
  • 2.2 Model Performance for the July 1988 Episode
  • 2.3 Model Performance for the July 1991 Episode
  • 2.4 Model Performance for the July 1995 Episode
    • 2.4.1 Model Performance Statistics
    • 2.4.2 Spatial Patterns and Transport in the 1995 Episode
  • 2.5 Summary
• 3. Model Performance for Western Boundary of the Urban Corridor
• 4. Model Performance Aloft in the Northeast
  • 4.1 Comparison with Ozone on Selected Traverses and Spirals
  • 4.2 Comparison of Verticall Integrated Ozone, NOy, RHC, and CO Data
• 5. Conclusions
• 6. References
• List of Figures
• List of Tables

PREFACE AND ACKNOWLEDGMENTS

This report has been prepared as an informational document for use by the OTAG Air Quality Analysis Workgroup. It includes analyses of NARSTO-Northeast data. It is a working draft document and has not been peer reviewed by the NARSTO-Northeast participants. All conclusions are those of the authors and not of NARSTO-Northeast or the sponsors. It is being circulated for review and comment and will be revised in response to comments received. Suggestions for improvements are welcome.

The authors thank the U.S. EPA and the OTAG modeling centers for providing the UAM-V model results and routine observational data, and thank the NARSTO-Northeast sponsors for making the special study data available. This effort was initiated and coordinated by Jeffrey West of General Public Utilities (GPU) and was made possible by funding from GPU, New England Electric System (NEES), and the Electric Power Research Institute (EPRI). This report is partially based on a separate report titled "Assessment of UAM-V Model Performance in the Northeast" (Londergan and Moore 1997) prepared by Earth Tech for a consortium of Northeast utilities.

1.0 INTRODUCTION

The Ozone Transport Assessment Group (OTAG) performed ozone modeling in the eastern United States in order to characterize regional transport and assess the effects of emission reductions (OTAG, 1996a). The modeling was performed for a large region extending from Florida to northern Michigan and from Texas to Maine. The scientific credibility of the OTAG modeling is derived from the formulation of the model, the quality of the input data, and the performance of the model in these specific applications. This report focuses on the performance of the UAM-V photochemical model in the baseline simulations in the Northeast region. OTAG provided an evaluation of model performance for the entire domain and for four large subregions (OTAG, 1996b), including the Northeast (as a single region). This report examines model performance in four subregions within the Northeast Corridor (Baltimore-Washington, Philadelphia, New York, and Boston nonattainment areas) and in the boundary areas immediately upwind of the Northeast Corridor.

1.1 MODEL EVALUATION OBJECTIVES

This analysis addresses several questions that were not addressed in the OTAG model performance report. These questions are:

1. How well do model simulations agree in time and space to observed surface and aloft ozone and precursor concentrations in the Northeast?
2. How do model simulations compare with the observed vertical distributions of ozone and precursors? Are the layers aloft and carryover of ozone resolved by the model?
3. Can concentration differences be partially explained by differences between measured and modeled transport phenomena?

4. If the modeled transport phenomena and the measurements agreed better, what would be the directional implications for simulated ozone concentrations? How would individual transport phenomena affect simulated downwind concentrations?

However, Question 4 was not addressed in this study because answering it required reapplying the UAM-V model with more accurate windfields which was beyond the scope of the study.

Our analysis includes (1) a review of observed and predicted ozone concentrations in the Northeast Corridor; (2) comparisons of observed and predicted ozone at "boundary sites" located along the western boundary of the Northeast Corridor; and (3) comparison of predicted and observed ozone aloft, based on NARSTO-Northeast aircraft measurements for the 1995 episode. The findings from these analyses are described in subsequent sections, with illustrative examples of each type of analysis.

1.2 Modeling Approach

The primary objective of OTAG modeling was to characterize the contributions to ozone nonattainment in the eastern United States due to regional transport, and to assess the effects of regional emission reductions on peak ozone in identified "problem areas." Our analysis is focused on model performance within the Northeast Ozone Transport Region (OTR), specifically on predictions for the serious/severe nonattainment areas along the Northeast Corridor from Baltimore-Washington to Maine, and on the transport of ozone and precursors into that region and between different urban subregions within the OTR.

For OTAG, the UAM-V air quality model and the RAMS and SAIMM meteorological models were applied using a 12 km grid spacing in the Northeast. This grid spacing is not adequate to resolve the local variations in emission density or in topography and surface characteristics that are important for estimating the time and location of maximum ozone concentrations near a large urban area. The goal of OTAG modeling was to capture the larger regional-scale features of ozone episodes and to characterize the relative contribution of regional versus local emission sources (OTAG, 1996a). Our performance analysis has therefore focused on issues pertaining to regional-scale performance, recognizing the practical compromises that were necessary to achieve OTAG's mission with the available data and computational resources. In order to assess model performance in the Northeast, however, it is important to distinguish the impacts associated with each of the major metropolitan areas that represent the dominant source regions contributing to ozone in the Northeast.

2. MODEL PERFORMANCE FOR OZONE IN THE NORTHEAST CORRIDOR

Model performance for each of the four OTAG episodes was evaluated using subregions that reflect the locations of major metropolitan areas and the predominant southwest-to-northeast transport commonly associated with ozone episodes in the Northeast. Geographic subdivision provides a clearer picture of model performance by providing comparisons of observed and predicted ozone concentrations associated with the "urban plumes" for the Boston, New York City, Philadelphia, and Baltimore-Washington subregions. Monitoring stations in the Northeast were assigned to five categories (the four urban subregions plus "outside of Corridor" sites), as depicted in Figures 2-1, 2-2, 2-3, 2-4, and 2-5. The set of ozone monitors varied by episode; these figures represent a composite based on the 1988, 1991, and 1995 episodes. The total number of ozone monitors in the Northeast is roughly 190, of which about 125 are located in the corridor. Model evaluation was performed for the stations in the Northeast Corridor and stations in the western boundary region (see Section 3).

For the subregions in the corridor, the monitoring network provides relatively good spatial coverage of areas with high observed ozone concentrations. This was confirmed by inspection of the daily maximum 1-hour ozone concentrations by subregion; the top two or three concentration values generally agreed within 10 ppb. (Exceptions are noted below in the discussion of results for individual episode days.)

Three of the four OTAG episodes were included in the analysis: July 4-15, 1988, July 16-21, 1991, and July 10-18, 1995. The initial model "spin up" days and some of the later days were excluded from our analysis because of low peak concentrations. The July 22­29, 1993 OTAG episode was excluded from the analysis because the peak concentrations in the Northeast were significantly lower than in the other three episodes. Ozone exceedances occurred on only three of the eight episode days in the Northeast in 1993 and the maximum observed values were 137, 142, and 163 ppb on these days. In addition, the model simulated peak concentrations were consistently low in the Northeast during the 1993 episode.

2.1 APPROACH

A number of statistical measures were used to evaluate the model performance. Experience has indicated that no one measure is adequate to characterize performance; multiple measures are needed. The specific measures are:

1. The daily maximum 1-hr observed concentration in a subregion to which three predicted concentrations are compared:
   • The daily maximum 1-hr predicted concentration unpaired in space and time. Comparison of this parameter with observations is not a particularly stressful test of a model because it allows for mismatch in space and time. That is, the comparison is made between the highest concentration predicted at any station within the subregion for any hour of the day with the highest observed 1-hr value on that day in the subregion.
   • The daily maximum 1-hr predicted concentration paired in space. This measure is a more stringent test of the model although it does allow for mismatch in timing of the peak ozone. It facilitates comparison of the highest concentration predicted at any hour at the location of the observed maximum to the observed 1-hr maximum concentration in the subregion.
   • The daily maximum 1-hr predicted concentration paired in space and time. This measure is the most stressful test of model performance for peak concentrations. Comparison is made between the concentration predicted at the hour and location of the observed maximum to the observed 1-hr maximum concentration in the subregion.
2. The predicted and observed 90th percentile concentrations for all hours. The 90th percentile concentration is statistically more robust than the maximum 1-hr concentration which behaves more like an extreme value statistic. The 90th percentile concentrations provide a relatively stable measure of the upper tail of the predicted and observed concentration distributions. This measure was not used in the OTAG model performance evaluation (OTAG, 1996b).
3. The mean bias (ppb) and mean normalized bias (%) in the predicted hourly concentrations for all stations and hours where the observed concentrations exceed 40 ppb. These metrics measure bias at all hours of the day and night when concentrations exceed the clean air background concentration of ozone. They are important when assessing transport because transport occurs at all hours. Model bias in the base case simulations is also important because it can be amplified in the nonlinear chemistry in emission control simulations. For example, if consistent underprediction of ozone occurs in a base case simulation because of underestimation of VOC emissions, it is likely that the model will overestimate the benefits of reducing VOC emissions and underestimate the benefits of reducing NO_x emissions. EPA guidelines for the urban scale photochemical model use a 15 percent mean normalized bias as the cutoff for acceptability (EPA, 1991).
4. The mean error (ppb) and mean normalized error (%) in the predicted hourly concentrations for all stations and hours where the observed concentrations exceed 40 ppb. Unlike the bias statistic where overpredicted and underpredicted values cancel one another, the error statistic is the absolute value of the difference. A modest amount of imprecision or error is expected in regional ozone modeling. EPA guidelines for the urban scale photochemical model use a 35 percent mean normalized error as the cutoff for acceptability (EPA, 1991).

These measures are calculated for each day using all of the available ozone data for the region. The gridded model outputs were spatially interpolated to the location of the stations to derive the model estimates at the stations.

Key features of observed and predicted ozone concentrations in the Northeast for each episode are discussed below. Peak concentrations in urban-scale subregions were compared to assess whether the model predictions reproduce the basic features of individual episode days, including regional transport into the corridor and between subregions within the corridor.

2.2 MODEL PERFORMANCE FOR THE JULY 1988 EPISODE

The UAM-V model was applied to the July 4-15, 1988 episode. The highest observed and predicted impacts in the Northeast occurred in the six-day period from July 6 through July 11. The predicted and observed maximum concentrations, 90th percentile concentrations, and mean concentrations as well as model bias and model error for July 6-12 are summarized in Tables 2-1, 2-2, and 2-3. This episode was particularly notable in the Northeast for its high ozone concentrations and long duration. For six consecutive days, maximum observed 1-hour ozone concentrations in the Northeast exceeded 180 ppb, and levels above 120 ppb were observed along the Northeast Corridor from Virginia to Maine. The episode was particularly extreme in the Philadelphia subregion, which experienced peak observed concentrations comparable to those in the New York City and Baltimore-Washington subregions.

On Wednesday, July 6, the highest observed ozone occurred at isolated locations around Baltimore, Philadelphia, and New York City. On July 7, high concentrations (>160 ppb) were more widespread in Baltimore-Washington and Philadelphia subregions, with an isolated peak of 158 ppb in New York City. On Friday, July 8, concentrations above 160 ppb were observed in a broad area stretching from Maryland to Massachusetts, with a maximum of 187 ppb in the Philadelphia subregion. On July 9, observed maximum concentrations were above 160 ppb only in Baltimore-Washington, while on Sunday, July 10, areas with high observed ozone (>160 ppb) included Philadelphia, Baltimore, and the Connecticut shore. On July 11, ozone concentrations above 180 ppb were observed along the Northeast Corridor from Maryland to Massachusetts. On July 12, the maximum observed concentrations in the Corridor dropped to 90-126 ppb.

The UAM-V model simulations reproduced important features of the observed ozone concentrations in the Northeast for this episode. While the magnitude of maximum predictions unpaired in space were 14 ppb lower than observed on average, the model predicted five consecutive days with peak impacts above 180 ppb. The model's maximum ozone estimates paired in space and paired in space and time were substantially lower than the observed maximum in most cases, which indicates the model does not predict the precise location and timing of the peak well. For example, at the Danbury, CT station where 222 ppb was observed at 1600 on July 11 the model predicted 121 ppb at the hour of the observed maximum and 142 ppb two hours after the observed maximum, while 75 km away at Madison, CT the model predicted 200 ppb at 1600 hrs. The model tracked the multiday buildup of ozone to its peak on July 11 fairly well. For example, in the New York City subregion the model predicted a build up from 142 ppb daily maximum on July 9 to 195 and 200 ppb daily maxima for July 10 and 11 when the observed buildup was from 138 ppb on July 9 to 203 ppb on July 10 and 222 ppb on July 11. In the Baltimore-Washington and New York City subregions, the simulations also tracked the day-to-day changes in maximum observations relatively well. For example, the predicted daily maximum concentration were 179, 148, 179, 172, and 204 ppb for July 7-11 in the Baltimore-Washington subregion when 210, 169, 190, 180, and 218 were observed, respectively. In contrast, the results for the Philadelphia subregion did not track the day-to-day changes in ozone levels as well and showed substantial underprediction (35 to 59 ppb unpaired) throughout the worst episode days. For the Boston subregion, there was good agreement (maximum unpaired ozone within 10 ppb) for July 8, when the highest daily ozone was observed, but overprediction by 20-30 ppb for July 7, 9 and 10. For July 12, the estimated maximum ozone concentrations decreased by 40-70 ppb, which was less than the observed decrease. The model did not simulate the "cleaning out" of the Northeast Corridor well.

The predicted and observed 90th percentile ozone concentration agreed within 10 ppb on only 8 of the 28 subregion-days in the 1988 episode which is fewer days than expected. The upper tail of the predicted ozone concentrations did not agree with observations well in the 1988 episode simulations as in the 1991 and 1995 episode simulations.

When all station-hours with concentrations above the 40 ppb clean air background concentrations were considered, there was negative bias (underprediction) on July 6 and positive bias (overprediction) on July 12 in all subregions. In the Boston subregion, the hourly concentrations were overpredicted on average for July 7-11 (daily mean biases ranging from +5 to +56 percent). The model bias for July 7-10 in Boston exceeded EPA's acceptance criteria for urban-scale modeling. In the New York City subregion, the hourly concentrations were significantly overpredicted for July 7-8 (+25 percent bias on average) and reasonably unbiased for July 9-11 (0 to +10 percent bias on average). In the Philadelphia subregion, the hourly concentrations were underpredicted on average for July 7-11 (daily mean biases ranging from 0 to -17 percent), which was consistent with the bias in peak ozone in this subregion. The predictions in the Baltimore-Washington area were relatively unbiased for July 7-10 (­13 to +4 percent) and positively biased on July 11 (+29 percent) on average. For the entire episode (July 6-12) and the entire Northeast Corridor, the model overpredicted hourly ozone concentrations by 3 ppb (84 vs 81 ppb), which is reasonably good performance. Performance on specific days in specific subregions, especially Boston, was problematic.

The model error for all station-hours varied significantly between subregions and days. Model predictions for July 6 had the lowest error, while those for July 12 had the highest error in most subregions. The mean normalized error in the Boston subregion varies from 24 to 56 percent for July 6-12, with the worst model error occurring on July 9 when the model overpredicted at all stations and hours. The mean normalized error in the New York City subregion varied from 25 to 46 percent for July 6-12, with the worst errors occurring on July 7. In the Philadelphia subregion, the mean normalized error varied from 24 to 47 percent for July 6-12 and was generally smaller than in the other subregions. The model error for the Baltimore-Washington subregion was modest (ranging from 23 to 32 percent) through July 10 and significantly higher on the last two days of the episode (38 and 59 percent). Overall, the UAM-V model predictions had less than 35 percent error on about half of the subregion-days. The errors for this simulation were larger than those for the 1991 and 1995 episodes.

For the 1988 episode, model performance was generally better for the Baltimore-Washington and New York City subregions, which had the greatest concentration of local emission sources, than for Philadelphia, where regional transport (both between urban subregions and into the corridor from the west) was expected to be a greater factor. Both observed and predicted concentrations were generally lower in the Boston subregion than in the rest of the corridor for the 1988 episode (except on July 8) and the model performance in the Boston region did not meet EPA's acceptance criteria on most days.

2.3 MODEL PERFORMANCE FOR THE JULY 1991 EPISODE

The UAM-V model was applied to the July 16-21, 1991 episode. The predicted and observed maximum concentrations, 90th percentile concentrations, and mean concentrations as well as model bias and model error for July 18-21 are summarized in Tables 2-4, 2-5, and 2-6. During this episode, ozone concentrations above 160 ppb were observed in the Northeast on three consecutive days, July 18-20. On Wednesday, July 17 and Thursday, July 18, high concentrations occurred primarily in the New York City and Boston subregions, with maximum observed concentrations of 155 and 174 ppb, respectively, at the Connecticut monitors. On July 19, the area with ozone concentrations above the Federal standard extended from Philadelphia to Maine, with a maximum of 161 ppb observed in Connecticut. The episode-maximum observed concentrations occurred on Saturday, July 20, with 178 ppb measured in Virginia and 175 ppb measured on Long Island, NY. On July 21, the observed peak concentration dropped to 135 ppb in Connecticut.

The accuracy of the UAM-V model's predictions for maximum ozone levels in the Northeast Corridor is quite variable for the 1991 episode. The model estimates of peak ozone concentrations (unpaired in space or time) are within 20 ppb for about half of the subregion-days. The results for the other half of the subregion-days show a mixture of under- and overestimation. The model's predictions paired in space or space and time tend to be equal or lower than the observed maximum. On July 18, the peak unpaired predictions show good agreement with observations for the New York City, Philadelphia, and Baltimore-Washington subregions and poor agreement (underprediction) for the Boston subregion. Boston's highest ozone in the 1991 episode occurred on July 18 and the model predicted 109 ppb when 169 ppb was observed. For July 19, the model predicts the New York City and Boston peaks accurately, but underestimates the peak values in the Philadelphia and Baltimore-Washington subregions. On July 20, the unpaired peak concentrations are in good agreement in the Philadelphia subregion where 151 ppb was observed. However, in the New York City and Baltimore-Washington subregions where 175 and 178 ppb were observed, respectively, the model underestimates the peak concentrations by 17-33 ppb. The 1-hr maximum concentration in Boston was also underestimated (124 versus 157 ppb observed) on July 20. For July 21, maximum values were estimated accurately in the Boston subregion and significantly overestimated in the New York City, Philadelphia, and Baltimore-Washington subregions. The UAM-V did not simulate the observed decrease in peak concentrations on the final day of the episode as was seen in the 1988 simulation.

The 90th percentile concentrations often show good agreement between the model and observed values. The predicted and observed 90th percentile concentrations were within 10 ppb for about two thirds of the subregion-days in 1991. Furthermore, the 90th percentile concentrations often show good agreement in cases where the peak values were significantly underpredicted. For example, the predicted and observed 90th percentile concentrations were 117 and 121 ppb on July 20 in the Baltimore-Washington subregion where the 178 ppb observed peak was underestimated by 32 ppb. Similarly, the predicted and observed 90th percentile concentrations were 103 and 105 ppb on July 20 in the New York City subregion where the 175 ppb observed peak was underestimated by 17 ppb. For the entire region and episode, the predicted and observed 90th percentile concentrations were 108 and 105 ppb, which is good agreement.

The mean bias and error for all station-hours with ozone above 40 ppb were significantly smaller for the 1991 episode than for 1988. The mean normalized bias was within 15 percent on 14 of the 16 subregion-days. The mean normalized error was within 35 percent on 15 of the 16 subregion-days. On the majority of subregion-days, the mean normalized bias was within 8 percent and the mean normalized error was within 22 percent, which is considered good performance for photochemical modeling. Overall, for the entire episode and region, the mean bias was 1 ppb or 1 percent, and the mean error was 17 ppb or 24 percent.

The smaller bias and error in this episode simulation compared to those for 1988 indicate the model did a better job of simulating the relevant atmospheric processes, including simulation of ozone transport. This may in part be due to differences in the meteorological models used (the RAMS model for 1991 and the SAIMM for 1988). Overall, the UAM-V model simulations of the July 1991 episode exhibited good performance in the Northeast corridor with the exception of underprediction of some of the maximum concentrations.

2.4.1 Model Performance Statistics

The UAM-V model was applied to simulate the relatively severe episode that occurred July 10-18, 1995. It was one of the most severe episodes in the Northeast since the July 1988 episode. Ozone concentrations above the Federal standard occurred on July 12-15 in the Northeast. The predicted and observed maximum concentrations, 90th percentile concentrations, and mean concentrations as well as model bias and model error for July 12-15 are summarized in Tables 2-7, 2-8, and 2-9. The subregion with the highest ozone shifts from Baltimore-Washington on Wednesday, July 12 to New York City on July 13, New York City and Philadelphia on July 14, then back to Baltimore-Washington on Saturday, July 15. The highest concentrations were observed on July 14 and 15.

The model's estimates of the maximum concentrations were of varying accuracy in the 1995 episode simulation. The 1-hr maximum values were predicted within 20 ppb on 10 of the 16 subregion-days (unpaired in space or time). However, the higher observed values were frequently underestimated, and comparisons of predicted maxima paired in space or space and time were all lower than the observed maxima. The model provided reasonably accurate estimates of the maximum ozone on July 12 when 136 ppb was observed in the Baltimore-Washington area and the model predicted 150 ppb. The agreement was closer in the other subregions on July 12. (e.g., 127 ppb predicted versus 125 observed in the Philadelphia subregion). On July 13, the model predicted a 145 ppb peak concentration in the New York City subregion when a maximum of 157 ppb was observed. The maximum ozone (unpaired) on July 13 was estimated within 12 ppb in the Boston, New York City, and Baltimore-Washington subregions, and overestimated by 28 ppb in in the Philadelphia subregion. On July 14, the maximum ozone (unpaired) was accurately estimated in the New York City (175 ppb predicted versus 176 observed), and underestimated in the other three subregions. The largest underprediction on July 14 was in the Philadelphia subregion where the model predicted a 135 ppb maximum when 170 ppb was observed. On July 15, the model underestimated the maximum concentrations by 33 to 46 ppb in the subregions with exceedances. For example, the model predicted 132 and 138 ppb in the New York City and Baltimore-Washington subregions when a maximum of 165 ppb was observed in both subregions. The only region in the Northeast with predictions above 140 ppb on July 15 was over the Atlantic Ocean south of Long Island, which indicates the transport was poorly simulated on this day. Clearly, the model did not capture the peak values well in this episode except in New York on the July 13.

The 90th percentile ozone concentrations agree within 5 ppb on about half of the subregion-days and agree within 10 ppb on 14 of the 16 episode days. During the July 12-15 period, the model slightly underestimated the 90th percentile ozone in the Boston subregion and slightly overestimated the 90th percentile ozone in the New York City, Philadelphia, and Baltimore-Washington subregions. The bias in the 90th percentile ozone was usually directionally consistent with the bias in peak ozone (unpaired) in most cases. For example, the 90th percentile concentrations were underestimated by 6 to 8 ppb on July 15 when the peak values were underestimated by 33 to 46 ppb in the New York City, Philadelphia, and Baltimore-Washington subregions. There was little bias in the 90th percentile or maximum ozone on July 14 in the New York subregion; the predicted and observed 90th percentile ozone were 114 and 112 ppb, respectively.

The mean bias and error statistics showed good model performance on this episode. The bias in hourly ozone concentrations above 40 ppb was within 15 percent on 15 of the 16 subregion-days. The model error was less than 27 percent in all of the subregion-days. On average for July 12-15 in the Northeast Corridor, the model bias was +2 percent and the model error 15 ppb or 21 percent. These performance statistics are better than those for the 1988 simulations, and almost as good as those for the 1991 simulations.

2.4.2 Spatial Patterns and Transport in the 1995 Episode

The July 1995 OTAG episode provides a valuable opportunity for examining model performance. The information available from the 1995 NARSTO-Northeast measurements program makes it possible to assess model performance in greater detail, in the context provided by a conceptual model of regional transport processes. In particular, the detailed meteorological information collected for the NARSTO-Northeast program provide a basis for evaluation of the observed and model-input wind information, in addition to observed and predicted spatial patterns of concentrations. Such comparisons provide a more complete picture of the strengths and limitations of the OTAG modeling approach for estimating regional and local contributions during ozone episodes. Model performance by subregion in the Northeast is discussed below for July 13, 14, and 15, 1995, the three "key days" with the highest observed and predicted ozone concentrations during this episode in the Northeast, with particular attention to predicted and observed spatial patterns and to comparisons of two sets of back trajectories, the first based on observed winds (constructed using the diagnostic CALMET meteorological model), and the second based on the RAMS prognostic meteorological model winds which were used in the UAM-V model simulations.

The spatial ozone pattern predicted for 2:00 p.m. (EST) on July 13 is shown in Figure 2-6. The spatial pattern of observed ozone concentrations showed distinct differences in transport patterns between subregions. The highest observed concentrations occurred in the New York City subregion, with a maximum of 157 ppb in Danbury, CT, about 50 km northeast of New York City. An area with measured peak concentrations above 125 ppb extended from Danbury into south-central Massachusetts, about 150 km from New York City. Peak predicted concentrations in the New York City subregion for July 13 were comparable in magnitude (145 ppb), but the region with predicted exceedances was smaller than observed (only extending about 80 km downwind of New York City) and was farther south, along the Connecticut shore. As shown in Figure 2-7, surface level (20 m) back trajectories starting from Ware, MA and East Hartford, CT derived from observed winds tracked back to Long Island, about 20 km east of the New York City metropolitan area, while the 500 m level CALMET back trajectories for these sites (see Figure 2-8) passed north of New York City and originated in north-central Pennsylvania on the morning of July 13.

In the Baltimore-Washington subregion, the area with observed exceedances (maximum 137 ppb) on July 13 was south of Baltimore and in between the two major metropolitan areas. In contrast, the model's peak ozone (136 ppb maximum) was in Delaware and eastern Maryland, roughly 60 km east-northeast of Baltimore. In the Philadelphia subregion, observed concentrations of 110-120 ppb occurred over a broad region north and northeast of the city with no localized "hot spot", while the predicted pattern showed a more well-defined "urban plume" extending into northern New Jersey, about 75 km northeast of Philadelphia. CALMET back trajectories starting at 1500 from Rutgers (northeast of Philadelphia) indicated transport from the west at 20 m and 500 m, while trajectories for northern Delaware and near Washington, DC indicated transport from the south. The measured winds in the Baltimore-Washington and Philadelphia subregions were much lighter and less organized than those in the New York City subregion.

Analysis of the spatial ozone patterns for July 13 indicate the observed New York City urban plume was transported in the northeasterly direction and the simulated urban plume was transported in an east-northeasterly direction. For the Baltimore-Washington and Philadelphia subregions, transport from southwest to northeast was predicted, but the observed pattern showed little organized transport.

On July 14, the highest ozone concentrations again occurred in the New York City subregion, with the highest values observed along the Connecticut shoreline and in Rhode Island. The spatial ozone pattern predicted for 2 p.m. (EST) on July 14 is shown in Figure 2-9. For this day, the predicted magnitude and spatial pattern of peak ozone concentrations were in excellent agreement with observed values (a maximum observed of 175 ppb in Madison, CT and a maximum predicted of 176 ppb in Groton, CT). The area of high observed ozone extended from east-northeast of New York City along the Connecticut shoreline and inland into eastern Connecticut and Rhode Island. The predicted area of high ozone was slightly displaced to the south from the observed, extending along the Connecticut shoreline and Long Island.

In the Philadelphia subregion, peak ozone concentrations of 170 and 163 ppb were observed at two sites in central New Jersey that were located 50 and 90 km east northeast of the city, respectively. While the peak predicted concentration was only 135 ppb, the location of the peak matched the observations. Both observed and predicted peak concentrations on July 14 were lower for the Baltimore-Washington subregion (144 ppb observed maximum, 121 ppb predicted), with spatial patterns indicating transport from west to east. Back trajectories of surface and 500 m level winds for sites in Connecticut, New Jersey, and Delaware, shown in Figure 2-10, showed persistent transport from the west-southwest on July 14. Excellent agreement is also seen between trajectories based on observed (CALMET) and model-input (RAMS) 500 m level winds (see Figure 2-11), which helps to explain the good air quality model performance for this day.

On July 15, the highest observed ozone concentrations in the northeast occurred in the Baltimore-Washington subregion, with a maximum of 184 ppb in northern Delaware, and observed concentrations exceeding 155 ppb at seven stations, four in an area within 40 km north and east of Washington, DC, the remaining three within 60 km east northeast of Baltimore. The spatial ozone pattern predicted for 1:00 p.m. (EST) is shown in Figure 2-12. Peak predictions in the Baltimore-Washington subregion for July 15 were substantially lower, with a maximum of 138 ppb predicted 30 km east of Washington, DC and 120 ppb predicted downwind of Baltimore. Predicted transport is to the east and east-southeast.

In the New York City subregion, the highest observed ozone concentrations on July 15 occurred east of New York City along the Connecticut shoreline (165 ppb at Madison, CT) and on Long Island. The area with the highest ozone predictions was displaced farther south (to southern Long Island and over the Atlantic Ocean), with the predicted maximum at a monitor of 132 ppb. Back trajectories based on observed winds for sites with peak observed ozone concentrations in Delaware and Connecticut (see Figure 2-13) indicated transport from the west and southwest at the surface (i.e., from Baltimore and New York City, respectively) , while trajectories based on RAMS model winds (see Figures 2-14) showed a west-northwesterly surface layer transport. The RAMS winds appeared to transport the ozone offshore rather than along the shoreline on July 15. The observed 500 m winds to the high ozone areas in Delaware and Conneticut were northwesterly earlier in the day and westerly in the afternoon while the RAMS wind were mostly northwesterly (see Figure 2-15).

Clearly, modest differences between the observed and model transport winds can have substantial impact on the source-receptor relationships and the predicted concentrations at specific monitors. Modest differences in transport can easily cause displacement of the location of maximum concentrations by 50 to 100 km. The differences in the transport winds probably contribute more to errors in maximum values than mean values.

2.5 SUMMARY

Each of the OTAG episodes represents a different pattern of regional transport, including both transport into the OTR and interactions between subregions within the OTR. The statistical measures of model performance for the Northeast Corridor were within EPA's acceptance crieria on most simulated days with high ozone. The simulations reproduced the magnitude (within 30 ppb) and approximate locations (within 100 km) of high ozone occurrences in the Northeast for most episode days. The performance was generally better in the 1991 and 1995 episodes than the 1988 episode. The results also pointed out several important shortcomings, including the following:

• The simulated ozone built up and decreased more slowly than observed over the multiday episode. The first few days of an episode showed underpredictions while the last few days of an episode showed overprediction (bias creep). This behavior is of concern because of the potential systematic nature of the bias which may affect emission control simulations.
• The model did not reproduce the detailed spatial patterns as well as one would like, so the maximum predictions paired in space or paired in space and time did not agree with the observed maximum nearly as well as the unpaired maximum prediction. Inaccuracies in the windfields and the coarse spatial resolution of the model probably contributed to this shortcoming.
• The modeling system rarely predicted the peak concentrations accurately in all four subregions on any given day. Usually significant underprediction or overprediction occurred in one or more subregions, which suggested the transport, not just local emissions, may be inaccurate.

^a For all station-hours with observed concentrations above 40 ppb.

^a For all station-hours with observed concentrations above 40 ppb.

^a For all station-hours with observed concentrations above 40 ppb.

3. MODEL PERFORMANCE FOR WESTERN BOUNDARY OF THE URBAN CORRIDOR

Ozone levels within the Northeast Corridor reflect the combined effect of contributions from "within-corridor" emission sources plus regional transport of ozone and ozone precursors. Since near-surface transport within the corridor during ozone episodes generally has a west-to-east component (between south-southwest and northwest), ozone concentrations at the monitoring sites immediately west of the corridor serve to characterize the inflow concentrations and regional transport component. A comparison of observed and predicted ozone concentrations at monitoring locations in Virginia, Pennsylvania, and New York was carried out to assess possible model bias in the inflow region. The locations of the monitoring stations used in this analysis are shown in Figure 3-1. Note, this "western boundary" analysis relies upon a rather simplistic transport scenario which has obvious limitations. Transport from south to north (without a significant westerly component) can occasionally produce impacts from in-corridor sources at a number of these "boundary" sites. Over the course of each episode, however, air quality at these monitors is generally dominated by boundary layer transport from the west (northwest through southwest).

Table 3-1 summarizes the mean predicted and observed concentrations and the mean bias and error in the model at these western boundary stations. When the data for hours with observed concentrations above 40 ppb are considered, the results showed that the model consistently underestimated the observed concentrations in the western boundary region. On average, the model underestimated concentrations by 9 ppb in 1988 and 1991, and 5 ppb in 1995. The magnitude of the bias was small on most days, although, it was very consistent. The mean ozone was underpredicted by more than 10 ppb on 5 of the 14 days considered in the analysis. The model error was 25 percent in 1988, 25 percent in 1991, and 23 percent in 1995 on average, which is comparable to elsewhere in the Northeast.

The daily maximum observed and predicted ozone concentrations (unpaired in time) at the western boundary stations are summarized in Table 3-2 for the 1988 episode. The stations are listed from north to south. The average daily maximum ozone concentration over all stations showed a systematic underprediction of 7 to 24 ppb over the critical episode days from July 7 through 11. The bias was particularly large on July 7, when the observed and predicted 12-station average daily maximums were 131 and 107 ppb, and on July 10 when the observed and predicted 12-station average daily maximums were 105 and 89 ppb. Early in the episode, a north-south concentration gradient was evident in both observed and predicted values. The gradient in predicted values persisted through July 11, but was less evident in observed values. In addition, the time of the observed and predicted maxima (not shown) often differed by two to four hours.

The daily maximum observed and predicted concentrations (unpaired in time) at western boundary sites for the 1991 episode are listed in Table 3-3. Average differences for the critical days of July 18-20 were quite similar to those from the 1988 episode, and again showed a systematic underprediction bias. The 14-station average daily maxima were underpredicted by 14 to 21 ppb on these three days. Both observed and predicted concentrations were lower for 1991 than for 1988. For the 1991 episode, however, peak concentrations were more variable across the network of monitors (than in 1988), indicating perhaps a less organized flow field and some impacts from local emission sources in this region.

The maximum concentrations (unpaired in time) for the 1995 episode are listed in Table 3-4. For the key episode days of July 12-15, peak observed and predicted concentrations for the "northern tier" stations (north of the Centre/State College site) showed systematic underprediction bias of 10-15 ppb, with higher predictions and minimal bias at southern boundary sites. On average, the model underestimated peak concentrations in this upwind region by 8 ppb on July 12, 19 ppb on July 13, 1 ppb on July 13, and 4 ppb on July 15 at this group of 14 stations.

These comparisons indicate that regional transport of ozone into the Northeast Urban Corridor from the west was systematically underpredicted by 5 to 10 ppb on average in the OTAG simulations. Relative to a nominal "clean background" ozone concentration of 40 ppb, the difference between a "regional background" of 70-90 ppb, as predicted, versus 80-100 ppb, as observed, represents a 15 to 25 percent underestimate of regional transport.

^a For all station-hours with observed hourly ozone concentrations greater than 40 ppb.

Table 3-4. Comparison of maximum ozone concentrations at western boundary sites for 1995 episodes.

4. MODEL PERFORMANCE ALOFT IN THE NORTHEAST

For four days during the 1995 OTAG episode, aircraft measurements were made in the Northeast as part of the NARSTO-Northeast measurements program. Comparisons of observed and predicted concentrations aloft provide additional insights concerning model performance. In this analysis, the model performance for ozone, NO_y, reactive hydrocarbons (RHC), and CO are evaluated against data collected by aircraft. The general pattern of aircraft flights during the July 12-15, 1995 episode is illustrated in Figure 4-1. On each morning, a flight was made along the "western boundary of the urban corridor" starting in Poughkeepsie, NY and ending at Shenandoah, VA, with spirals to measure vertical profiles (from 100 m above ground level (agl) to 1500 m above sea level (msl)) at Poughkeepsie, Kunkletown, PA, Gettysburg, PA and Shenandoah. On the tranverses between spiral locations, the plane flew at 250 to 650 m agl. The same plane flew an afternoon "in corridor" route starting at Manassas, VA and ending at Poughkeepsie. On July 14, a second plane flew from Poughkeepsie to Brookhaven, NY and along the "northern corridor" up to Portland, ME in the morning and returned in the afternoon.

Comparison of observed and predicted concentrations aloft is somewhat compromised by the coarse vertical resolution used in the OTAG simulations. Model predictions represent "layer average" concentrations for 50-100 m agl (layer 2), 100-250 m agl (layer 3),

250-500 m agl (layer 4), 500-1500 m agl (layer 5), and 1500-2500 m agl (layer 6).

4.1 COMPARISON WITH OZONE ON SELECTED TRAVERSES AND SPIRALS

Comparison of observed and predicted ozone concentrations aloft for the morning "western boundary" flights showed consistent underprediction bias on all four episode days. Figures 4-2, 4-3, 4-4, 4-5, 4-6, 4-7, 4-8, and 4-9 display results for July 14 and 15. On July 14, the spiral at Poughkeepsie showed a layer with 90 to 125 ppb of ozone between 600 and 1100 m agl, which was substantially higher than the model prediction of 75 ppb for layer 5. At Kunkletown, PA (Figure 4-3), the observed ozone concentrations were 90-100 ppb below 1000 m agl, compared to layer 5 predictions of about 50 ppb, while the layer 6 predicted concentration was 100 ppb, nearly double the observed. Thus, the predicted vertical ozone concentration profile was inverted and the opposite of the observed profile. At Gettysburg, PA, observed ozone was

30-40 ppb higher than predicted from 200 to 1000 m agl. The traverse between Kunkletown and Gettysburg (Figure 4-5) showed a consistent difference of about 30 ppb between observed and predicted values at 600 m agl. Results for July 15 showed similar bias, except for the spiral at Kunkletown (Figure 4-7) where observed and predicted ozone concentrations were similar above 500 m agl.

These results suggest that the model is underpredicting regional transport into the OTR from the west. The measured ozone concentrations also provide additional measurements at several "western boundary" sites which were discussed in Section 3. The peak ground-level concentrations observed at numerous western boundary sites (e.g., Gettysburg, PA) for

July 14 and 15 (see Table 3-4) were comparable to (or less than) the peak concentrations measured aloft near these sites in the morning. This suggests that the observed surface ozone concentrations at the western boundary of the urban corridor may be mostly due to transport rather than local emissions and local ozone formation.

Measurements from "in corridor" afternoon flights showed little prediction bias. Figure 4-10, 4-11, 4-12, 4-13, and 4-14 displays results from the July 15 afternoon flight. The spiral at Manassas, VA (Figure 4-10) showed a well-mixed layer of observed ozone of 90-100 ppb up to 1000 m agl, about 15 ppb higher than predicted. The traverse from Manassas to Lums Pond, DE (Figure 4-11) showed increasing observed and predicted concentrations, with an observed peak concentration above 150 ppb over Chesapeake Bay. The spiral above Lums Pond (where the peak ground-level concentration a few hours later reached 184 ppb) showed observed concentrations of about 110 ppb up to 600 m agl, then a sharp drop, while predicted concentrations were well mixed up to 1500 m agl. The spiral at Kunkletown, west of the corridor, showed comparable observed and predicted ozone up to 1200 m agl; predicted concentrations remained constant into layer 6, while observed concentrations decreased. Over the Atlantic Ocean south of Long Island, observed ozone concentrations (see Figure 4-14) showed a complicated vertical structure, including a shallow layer with high ozone concentrations below 200 m agl. The predicted vertical profile showed higher concentrations aloft.

Data from three additional flights with interesting features are shown in Figures 4-15, 4-16, and 4-17. The early morning traverse on July 14 from Poughkeepsie to Brookhaven showed a "spike" in observed ozone near Poughkeepsie and nearly constant ozone of 80 ppb as the flight path crossed the New York City urban plume. The predicted ozone at 300 to 500 m agl dropped from 65 to 20 ppb, and then increased to 70 ppb as the flight path cross the New York City urban plume. The model most likely overestimated the vertical extent of mixing since its predictions show ozone scavenging in the urban plume and the data do not show this effect. On the afternoon of July 14, (Figure 4-16), the flight from Ware, MA to New Haven, CT encountered the New York City urban plume, with a peak in both observed and predicted concentrations (130-150 ppb) as the flight approached New Haven. The afternoon spiral on July 14 (Figure 4-17) over the Atlantic Ocean south of Long Island showed observed ozone reaching 150 ppb between 300 and 800 m agl, then dropping to 70 ppb at 1500 m agl, while the predicted "layer average" concentration was about 115 ppb. Again, predicted vertical gradients in ozone were smaller in magnitude and in the opposite direction than the observed vertical gradient.

In summary, this comparison with detailed aircraft data indicates that the OTAG simulations predict afternoon ozone concentrations in the corridor for the 1995 episode without much bias, despite systematic underprediction of morning ozone concentrations aloft (between 200 and 1500 agl) by 20-30 ppb. Observed ozone profiles of ozone concentrations offshore over the ocean also show a pronounced, complex vertical structure, which the model is unable to simulate, given the limited horizontal and vertical resolution of the OTAG grid.
4.2 COMPARISON OF VERTICALLY INTEGRATED OZONE, NO_y, RHC, AND CO DATA

The aircraft collected numerous vertically integrated hydrocarbon and CO samples on vertical spirals on the July 13-15, 1995 flights. The high time resolution ozone and NO_y, data have been averaged to correspond to the hydrocarbon data in order to concurrently evaluate the model predictions aloft for these species. Figures 4-18, 4-19, 4-20, and 4-21 show scatter plots for the predicted and observed vertically integrated ozone, NO_y, RHC, and CO concentrations. The individual data are shown in Tables 4-1 and 4-2. The data are mostly integrated over 50 to 500 m agl, with some samples collected at higher elevations. For purpose of comparing the modeled and observed hydrocarbons, only the reactive hydrocarbons (RHC) are considered. The measured reactive hydrocarbon concentrations were estimated from the nonmethane hydrocarbon, ethane, propane, and benzene concentrations as follows:

RHC = NMHC - 0.8*[Ethane] - 0.5*[Propane] - 0.8*[Benzene]

The models' reactive hydrocarbons included paraffin bonds (PAR), ethene (ETH), olefins (OLE), toluene (TOL), and xylenes (XYL)..

The vertically integrated results show some bias for ozone aloft. For the morning samples, the average predicted and observed vertically integrated ozone concentrations were 60 and 70 ppb, respectively (or -13 percent bias). Comparison of the afternoon samples indicate better agreement, with the model predicting 87 ppb when 91 ppb was observed on average (-5 percent bias). Overall, the average predicted vertically integrated ozone was 73 ppb when 81 ppb was observed. The scatter plots show cases with over- and underpredictions. The predicted and observed ozone are reasonably well correlated

(R² = 0.66, even with two outliers included). These results are consistent with the point-by-point comparison presented above and the analysis of all of the 1995 aircraft data reported by Adamski (1997), in that they show more ozone underestimation bias in the morning than in the afternoon.

The scatter plot for NO_y shows significantly more scatter than the corresponding ozone plot, nevertheless, the predicted and observed NO_y are correlated (R² =0.46). The average vertically integrated NO_y concentration is 8.3 ppb, which is slightly above the 6.6 ppb observed. The NO_y concentrations and bias are similar in the morning and afternoon samples. Note, however, that the model's estimate of NO_y should be biased low (by 1 to 10 percent) because not all of the NO_y species were archived in the simulations. The bias in the NO_y predictions aloft is similar to that found in the surface data in the Northeast (OTAG, 1996b).

The scatter plot for RHC shows the model seriously underestimates the RHC concentrations aloft in this region. On average, the model estimates 27 ppbC RHC compared to 90 ppbC observed in the aircraft spirals. The RHC concentrations and bias are similar in the morning and afternoon samples. The discrepancy for RHC aloft is somewhat larger than the discrepancy found in evaluation of surface RHC data at rural NARSTO-Northeast stations (OTAG, 1996b), although it is directionally similar. In addition, there is no correlation (R² =0.01) between predicted and observed RHC aloft. This comparison is based on a relatively small number of observations. If the underestimation of RHC is widespread, it would suggest the VOC emissions are underrepresented in the modeling system and that the model may be getting the right ozone for the wrong reasons.

The scatter plot for CO shows greater variability in the observed CO aloft than the predicted CO. The dynamic range of the predicted CO is narrower (125 to 312 ppb) than the observed range (144 to 497 ppb). The average vertically integrated CO concentration aloft is 209 ppb, which is substantially lower than the 338 ppb observed aloft. The bias in CO predictions aloft is smaller than the bias for RHC, and probably not of great concern for the modeling.