Evaluation of the Community Multiscale Air Quality Model for Simulating Winter Ozone Formation in the Uinta Basin

Abstract

The Weather Research and Forecasting (WRF) and Community Multiscale Air Quality (CMAQ) models were used to simulate a 10 day high-ozone episode observed during the 2013 Uinta Basin Winter Ozone Study (UBWOS). The baseline model had a large negative bias when compared to ozone (O₃) and volatile organic compound (VOC) measurements across the basin. Contrary to other wintertime Uinta Basin studies, predicted nitrogen oxides (NO_x) were typically low compared to measurements. Increases to oil and gas VOC emissions resulted in O₃ predictions closer to observations, and nighttime O₃ improved when reducing the deposition velocity for all chemical species. Vertical structures of these pollutants were similar to observations on multiple days. However, the predicted surface layer VOC mixing ratios were generally found to be underestimated during the day and overestimated at night. While temperature profiles compared well to observations, WRF was found to have a warm temperature bias and too low nighttime mixing heights. Analyses of more realistic snow heat capacity in WRF to account for the warm bias and vertical mixing resulted in improved temperature profiles, although the improved temperature profiles seldom resulted in improved O₃ profiles. While additional work is needed to investigate meteorological impacts, results suggest that the uncertainty in the oil and gas emissions contributes more to the underestimation of O₃. Further, model adjustments based on a single site may not be suitable across all sites within the basin.

1. Introduction

The United States (U.S.) Intermountain West is an important source of domestic energy resources, including oil and natural gas (O&G). One of the primary environmental impacts associated with O&G production is emissions of air pollutants, such as nitrogen oxides (NO_x), volatile organic compounds (VOCs), particulate matter (PM), sulfur dioxide (SO₂), greenhouse gases (GHGs), and hazardous air pollutants. In the case of NO_x and VOCs, these pollutants are precursors to the formation of ground-level ozone (O₃). Because exposure to unhealthy levels of O₃ can result in a number of health effects, including chest pain, coughing, throat irri-tation, congestion, and asthma, the U.S. Environmental Protection Agency (EPA) has set a National Ambient Air Quality Standards (NAAQS) for O₃ of 70 ppbv to reduce both public health effects and other ecological impacts of O₃ exposure (U.S. Environmental Protection Agency (EPA), 2015).

Unhealthy levels of O₃ at the surface typically occur during the summer because the photochemical reactions that produce O₃ are accelerated by warm temperatures, higher solar actinic flux, and increased biogenic vola-tile organic compound (VOC) emissions. Episodes of rapid photochemical production of O₃ in winter were first observed by the Wyoming Department of Environmental Quality (WDEQ) in February 2005 (Wyoming Department of Environmental Quality (WDEQ), 2011). These elevated winter O₃ levels were identified in the Upper Green River Basin (UGRB) in western Wyoming (WY), which is a rural area with intensive oil and gas production activity (Schnell et al., 2009). The WDEQ carried out field studies each winter from 2007 to 2016 and found that winter O₃ was caused by VOC and NO_x emissions that were trapped in a shallow inver-sion layer and associated with snow cover that both strengthened the inversion layer and increased surface albedo. This caused increased photolysis rates and photochemical production of O₃ (ENVIRON, 2008). Schnell et al. (2009) also found that snow albedo increased photolysis rates and that peak O₃ production in February in the presence of snow was greater than in June. Carter and Seinfeld (2012) performed box modeling in the UGRB and found that O₃ production rates were highly sensitive to snow albedo and heterogeneous production of nitrous acid. However, the box model simulations did not realistically represent emissions, dispersion, and production of O₃ for multiday inversion episodes. More recently, high ground-level O₃ has also been observed during the winter in rural areas located near O&G operations in western Colorado and Utah (Edwards et al., 2014; Field et al., 2015; Helmig et al., 2014; Martin et al., 2011; Oltmans et al., 2014). Rapid photochemical production of O₃ in the winter has only been observed in conditions with complex terrain and persistent cold air pools (PCAP) that trap emissions of O₃ precursors in a shallow boundary layer (Neemann et al., 2015) and with extensive snow cover that enhances surface albedo and photolysis rates (Edwards et al., 2014; Oltmans et al., 2014; Rappenglück et al., 2014). During the winter of 2013, the Uinta Basin in Utah experienced O₃ levels that exceeded the NAAQS on 29 days, including a maximum 8 h average O₃ of 142 ppbv in March 2013. High O₃ levels were observed during five PCAP episodes, ranging from 3 to 15 days in length.

The goal of this study was to use the Weather Research and Forecasting model (WRF) and the Community Multiscale Air Quality (CMAQ) model to explore the chemical and physical processes that control the elevated levels of surface O₃ associated with O&G emissions in the Uinta Basin during a 10 day high O₃ episode in the winter of 2013. Previous modeling studies have investigated this issue using box models (Carter & Seinfeld, 2012; Edwards et al., 2014) and showed the complexity of the chemistry in this area, the potential heteroge-neity of the O₃ production in an O&G area, and the importance of accurately characterizing the VOCs. Other studies that focused on winter O₃ in the Uinta Basin and used three-dimensional air quality models, including WRF-Chem (Ahmadov et al., 2015) and CAMx (Emery et al., 2015), found that the models performed poorly for the O₃ episode during the Uinta Basin Winter Ozone Study (2013 UBWOS) because VOC emissions from O&G operations were underestimated, while NO_x emissions were overestimated in the basin. Ahmadov et al.(2015) and Emery et al. (2015) focused primarily on VOC and NO_x observations at Horsepool and found that their base model greatly underpredicted observed ozone. Ahmadov et al. (2015) found that the 2011 National Emission Inventory version 1 (2011 NEIv1) O&G NO_x emissions were biased high by about a factor of 4 relative to the observations, while the VOCs were biased low by approximately a factor of 2. Emery et al. (2015) applied a modified O₃ chemical mechanism and similarly found a large underprediction of peak ozone. Even after introducing NO_x reductions, VOC increases, and other modifications, that study was only able to predict about two thirds of the peak O₃ observed at Horsepool.

The work presented here builds on previous modeling assessments in the Uinta Basin to understand how well the most recent regulatory modeling system (WRF, Sparse Matrix Operator Kernel Emissions (SMOKE), and CMAQ) captures the complex combination of PCAPs and emissions that can lead to elevated O₃ during the winter (Baker et al., 2011). Here model predictions are compared to special study measurements taken as part of the 2013 UBWOS campaign, along with routine surface measurements of O₃ and NO_x. These compar-isons are used to characterize basin-wide performance and avoid adjustments made for a singular location that may not reflect heterogeneity in O₃ formation throughout the Uinta Basin. Model sensitivities were performed to investigate how modeled VOC, NO_x, and O₃ respond to changes in emissions, meteorology, and deposition. Improved confidence in this regulatory modeling platform will lead to stronger and more robust regulatory assessments intended to characterize source impacts on air pollution in this area and more efficient, effective, and accurate strategies to reduce harmful air pollutants.

2. Methods

2.1. Domain

The Uinta Basin is a part of the Colorado Plateau in the northeast corner of Utah. The basin is bounded in the north by the Uinta Mountain range, in the south by the Book and Roan Cliffs, in the west by the Wasatch Range, and in the east by Douglas Creek Arch separating it from the Piceance Basin in Colorado. The floor of the basin is at approximately 1,500 m above sea level. Duchesne and Uintah Counties make up most of the basin. The Uinta Basin is a rural area with a population of about 50,000 people primarily located in three main towns (Duchesne, Roosevelt, and Vernal). Oil and gas development (approximately 10,000 producing wells) is widely scattered throughout with associated drilling, processing, compression, and pipeline facilities (Stoeckenius et al., 2014). Based on the data collected from the 2013 UBWOS, about 24 O&G fields accounted for 90% of the 2013 O&G production in the Uinta Basin (Stoeckenius et al., 2014). In addition, most of the natural gas wells are located east of the Green River, while the oil wells are located to the west (Stoeckenius et al., 2014).

2.2. Observational Data Sets and Episode Selection

The 2013 UBWOS took place between January and March 2013 (Stoeckenius & McNally, 2014). Researchers from the National Oceanic and Atmospheric Administration (NOAA), Utah Department of Environmental Quality (UDEQ), and several universities conducted extensive ground-based and airborne measurements of O₃ and other key air quality and meteorological parameters. The monitoring methods have been previously described in detail and included canisters, balloon-borne instruments, ozonesondes, rawinsondes, aircraft measurements, mobile instruments, tower instruments, and other ground-based instruments (Edwards et al., 2014; Oltmans et al., 2014; Schnell et al., 2016; Stoeckenius & McNally, 2014). These measurements collected O₃, NO_x, speciated VOCs, and various meteorological parameters (e.g., wind speed, wind direction, and temperature). Most of the measurements occurred at Horsepool, which is located in the southeastern portion of the basin, although other monitors within the basin were also operational during this campaign (Figure 1). Table S1 in the supporting information presents a list of the data used in this study and the asso-ciated groups that performed the measurements.

Figure 1.

(a) Full model domain with black dots indicating the locations of measurements used for the model performance evaluation. (b) Enlargement of model domain with the names and locations of the sites selected for the model evaluation.

The analyses presented in this study covered a high O₃ episode that occurred between 1 February 2013 and 10 February 2013 during the 2013 UBWOS field campaign.

2.3. Description of Air Quality Modeling Platform

2.3.1. Air Quality Model Configuration

The public release of CMAQ, version 5.0.2 (Appel et al., 2013), was used for simulating a 10 day period from 1 February to 10 February 2013. The simulations used a domain with a horizontal resolution of 4 km, covering most of the state of Utah and portions of Wyoming, Colorado, and Idaho (Figure 1). The domain included 41 vertical layers with varying thickness from the surface to the model top at approximately 18 km and matched the vertical layers used in WRF. Table S2 in the supporting information lists the model layer heights used in the model platform. The 4 km simulations were driven by boundary and initial conditions from a similarly configured set of simulations using a 12 km horizontal grid spacing nested within a 36 km simulation cover-ing the entire contiguous United States. The coarser simulations (12 km and 36 km) were initialized in December 2012 to minimize initial-condition influence.

The CMAQ configuration used a version of the CB05 gas phase photochemical mechanism with updated toluene chemistry (Whitten et al., 2010) and explicit treatment of methane (CH₄) from O&G sources for a total of 205 reactions involving 80 chemical species. In CB05, individual VOC species are treated either explicitly or by model species designed to represent certain carbon bond types. Under a carbon bond representation, the reactivity of many VOCs are accounted for by multiple model species, depending on the types of carbon bonds contained in the VOC. For example, one molecule of n-pentane is represented in CB05 by four mole-cules of model species PAR, representing a reactive carbon singly bonded to another carbon, to characterize the four reactive, singly bonded carbons. While this representation results in some loss of detail (e.g., predic-tions of n-pentane concentrations are not available), it significantly reduces the computational burden which allows the chemistry to run efficiently in an air quality model. An updated carbon bond mechanism, CB6, is now available, but it has not been fully analyzed or evaluated in CMAQ for regulatory use. Another study using CB6 showed that CB6 produced slightly less ozone than CB05 in the winter, with a 1–3 ppbv reduction in ozone (Emery et al., 2015). Aerosol species were calculated using the AERO6 aerosol microphysics module (Appel et al., 2013; Reff et al., 2009).

Deposition in CMAQ is modeled as the product of the layer 1 concentration and the deposition velocity cal-culated at a reference height of one half of the first layer height. The deposition velocity is calculated using the resistance paradigm, and details of the model are provided in Pleim and Ran (2011) and Pleim and Xu (2003). Briefly, the deposition velocity for gases, such as O₃, is calculated as the inverse of the sum of the aerodynamic, boundary layer, and surface resistances. The aerodynamic resistance is calculated using a theory similar to the WRF Pleim-Xiu land surface model (P-X LSM) (i.e., grid cell average land surface characteristics) and passed to the deposition velocity calculation in CMAQ. The boundary layer resistance is a chemical-specific parameter which accounts for the diffusion across the quasi-laminar layer close to the surface and depends on the friction velocity. The CMAQ approach differs from other models in the parameterization of the surface resistance. The surface resistance accounts for uptake by vegetation and the ground. The resis-tances are treated separately for wet and dry surfaces, and scaling between these values is done using the canopy wetness fraction which varies between 0 and 1. The model also includes a vegetation fraction, which represents the fraction of the grid cell that is covered by vegetation. An area-weighted value for the vegeta-tion fraction is calculated in the P-X LSM and passed to CMAQ. The vegetation fraction is used to split the resistance pathways into vegetated and nonvegetated. The CMAQ deposition model also contains a unique parameterization for the resistance for snow-covered surfaces. The model was developed based on the work of Bales et al. (1987) and includes consideration of the liquid water fraction of the snow, which is important for soluble species.

Photolysis rates have a significant effect on the rate of O₃ formation and are a function of the amount of solar radiation available. In CMAQ, the photolysis rates are derived for each grid cell as a function of the solar zenith angle, altitude, total O₃ column and aerosol scattering, surface albedo, and cloud cover. Surface albedo in CMAQ is calculated using land use and snow fraction from WRF, along with internal tables that apply surface albedo and snow correction factors for each of the land use classes. The fractional land use is used to weight the surface albedo in the case that a grid cell has multiple land use types. However, the land use in the Uinta Basin was almost entirely of the scrub class (index 32 of the NLCD40 land use data set), so little weighting was done. It should be clarified that the meteorological sensitivity analysis discussed in section 3.2, where surface albedo in WRF was increased from 0.60 to 0.80, did not translate to more photolysis in CMAQ because the algorithm for scrub class was not sensitive to albedo changes of this magnitude. The CMAQ albedo tables for this scrub class could have been adjusted, but the goal was to investigate the impact on the meteorology independent of the photochemistry. The major configuration options used in CMAQ are outlined in Table S3 in the supporting information.

2.3.2. Meteorological Model Platform and Configuration

The meteorological inputs were generated by the standard WRF model version 3.6.1 (Skamarock et al., 2008) for the model domain described in section 2.3.1. The main WRF physics options used for the simulations included the Rapid Radiation Transfer Model Global (RRTMG) for long and shortwave radiation (Iacono et al., 2008), Morrison microphysics (Morrison & Gettelman, 2008; Morrison & Grabowski, 2008), the Kain-Fritsch 2 cumulus parameterization (Kain, 2004), the Pleim-Xiu land surface model (P-X LSM) (Pleim & Gilliam, 2009; Pleim & Xiu, 2003; Xiu & Pleim, 2001), and the Asymmetric Convective Model version 2 (ACM2) (Pleim, 2007a, 2007b). The P-X LSM used the release version of the 2011 National Land Cover Dataset (NLCD) 40-class scheme with fractional land use down to approximately 250 m. As previously indicated, the Uinta Basin is dominated (90–100%) by the scrub land use class, which is open flat terrain with scattered low bushes. The simulations also included impervious surface and canopy fraction data for urban modeling, although these have little impact in this case.

Rather than coarser snow cover data from the North American Mesoscale Forecast System (NAM) 12 km ana-lysis that was used for the 12 km parent domain, a much more detailed daily 1 km snow cover analysis was leveraged in the 4 km modeling. This data set was provided by the NOAA National Weather Service’s National Operational Hydrologic Remote Sensing Center’s (NOHRSC) SNOw Data Assimilation System (SNODAS)(Barrett, 2003). The aim of SNODAS is to provide a physically consistent framework to integrate snow data from satellite, airborne platforms, and ground stations with model estimates of snow cover (Carroll et al., 2001). The daily values of snow cover were assigned at 00 Coordinated Universal Time (UTC) each day with linear interpolation updates every 3 h during the simulations for smoother transition of snow cover through-out the day. Since the P-X LSM does not have physics in place to accumulate/melt/sublimate snow but rather just reads the provided snow analysis, the precision of SNODAS in terms of spatial resolution is key for accu-rate snow modeling.

Four-dimensional data assimilation, or grid nudging, was applied every 3 h above the planetary boundary layer (PBL) using a combination of NAM 12 km analyses and 3 h forecasts. This allows the PBL to evolve with less influence from a model analysis, where that forcing is limited to the free atmosphere. State variables tem-perature, moisture, and wind were nudged with strengths of 1 ×10⁻⁴, 1 × 10⁻⁵,and 1 ×10⁻⁴s⁻¹. These nudging strengths are slightly lower than many 12 km applications. The WRF simulations had a spin-up of 31 days. The WRF model output files were processed using the most current version of the Meteorological Chemistry Interface Processor (MCIP) (Otte & Pleim, 2010), version 4.3, to generate meteorological input files for CMAQ. WRF and MCIP were configured for hourly meteorological inputs to CMAQ. This application may have benefited from a tighter coupling of the meteorology (i.e., two-way WRF-CMAQ model), where fields could be exchanged at higher frequency. However, given the meteorology sensitivities presented in section 3.2, it is highly unlikely that it would change the major conclusions. Other major configuration options used in WRF are outlined in Table S4 in the supporting information.

2.3.3. Emissions Inventory and Modeling Platform

All of the non-O&G emissions inventories were based on the public release of the 2011 U.S. EPA National Emission Inventory version 2 (2011 NEIv2) and the 2013 Electric Generating Unit (EGU) emissions data (U.S. EPA, 2016b). The 2011 NEIv2 contains emissions from EGUs, other point sources, O&G, agriculture, dust, bio-genic, locomotive, marine, residential wood combustion, fires, and mobile source sectors. The 2011 EGU data were updated to use the 2013 data for sources for which continuous emissions monitoring (CEM) data were available. The O&G emissions used a combination of the 2011 NEIv2 emissions inventory and data from the Western Regional Air Partnership (WRAP) Phase III emissions inventory (Friesen et al., 2009). The 2011 NEIv2 nonpoint O&G sector was primarily developed using data supplied to EPA by state air agencies. In the case of nonpoint O&G emissions estimates, EPA developed the 2011 Nonpoint Oil and Gas Emission Estimation Tool to estimate emissions from this category (U.S. EPA, 2016b). This tool generates estimates of emissions of NO_x, VOC, PM, carbon monoxide (CO), ammonia (NH₃), and SO₂ from upstream O&G production activities. Table 1 presents the breakdown of the 2011 NEIv2 emissions and top processes contributing to those emissions for NO_xand VOCs in Duchesne and Uintah counties. The spatial distribution of VOC and NO_xemissions across the domain, including the distribution of O&G-specific emissions, are shown in Figure 2, demonstrating the dom-inance of O&G sources at the observational sites within the domain.

Table 1.

Breakdown of 2011 NEIv2 Emissions and Top Processes Contributing to NO_x and VOCs by Source Sectors in Duchesne and Uintah Counties Between 1 February 2013 and 10 February 2013

	NOx	VOC
Sector	(ton)	(ton)	Top processes
Non-point O&G	471	4,010	Industrial processes; oil and gas exploration and production; onshore gas production
Point EGU	278	1.5	External combustion boilers; electric generation; bituminous/subbituminous coal
Point O&G	89	12	Industrial processes; oil and gas production; petroleum and solvent evaporation;  petroleum product storage at refineries; fixed roof tanks
On road	84	38	Diesel; combination short-haul trucks; gasoline; passenger trucks
Point fire	27	500	Miscellaneous srea sources; other combustion; prescribed burning for forest  management
Nonpoint	10	36	Stationary source fuel combustion; residential; liquified petroleum gas; natural gas;  solvent utilization; miscellaneous nonindustrial: commercial; pesticide application
Nonroad	10	48	Construction and mining equipment; mobile sources; off-highway vehicle gasoline,  two-stroke; recreational equipment
Residential wood  combustion	0.6	6.3	Stationary source fuel combustion; residential; wood
Point non-IPM	0.05	0.19

Figure 2.

Spatial distribution of total emissions across the model domain for (a) NO_x and (b) VOC O&G emissions and all other sources of (c) NO_x and (d) VOC emissions. The sites selected for the model performance evaluation are labeled in the figure, with the identification provided in Figure 2a.

The emissions data were processed from annual, daily, or hourly estimates of emissions at the state or county level to create hourly emissions fluxes on the modeled spatial grid by using source-based temporal, spatial, and chemical allocation profiles in the Sparse Matrix Operator Kernel Emissions (SMOKE) processing model (version 3.6.5) (Houyoux et al., 2000).

Chemical speciation profiles for different emission sources were used to translate from total VOC and total organic gases to the modeled species used in CMAQ.

The 2014 version of the Motor Vehicle Emissions Simulator (MOVES 2014) was used to generate the on-road mobile source emissions. MOVES accounts for the effects of fleet age deterioration, ambient temperature and humidity, activity patterns, fuel properties, and inspection and maintenance programs on emissions from all types of motor vehicles. The National Mobile Inventory Model (NMIM) was used to create the nonroad mobile equipment source emissions on a month-specific basis that accounts for temperature, fuel types, and other variables that vary by month (U.S. EPA, 2016b).

3. Results and Discussion

The WRF and CMAQ models were used in this study, and the model performance was evaluated between 1 February and 10 February 2013, which is one of the high-ozone episodes that occurred during the 2013 UBWOS field campaign in February 2013. The primary sites for the CMAQ evaluation were chosen to repre-sent areas with (a) mostly oil production (White Rocks), (b) mostly natural gas production (Horsepool and Ouray), and (c) population centers with large contributions from mobile sources and solid fuel combustion emissions (Roosevelt). Surface measurements and vertical profiles of O₃, NO , CH , speciated VOCs and various meteorological parameters (e.g., wind speed, wind direction, and temperature) were collected at these sites. The performance of the WRF and CMAQ models was also evaluated at additional monitoring sites within the basin, including the Fantasy Canyon, Fruitland, Mountain Home, Myton, Rabbit Mountain, Rangely, Red Wash, Seven Sisters, Vernal, and Wells Draw sites. Some of the additional model performance results are included in the supporting information. Time series, vertical profiles, spatial plots, and statistical analyses were completed to investigate the performance of the models.

Due to the substantial discrepancies between the standard CMAQ modeled predictions and observations of O₃, NO_x, and speciated VOCs, additional WRF and CMAQ simulations were conducted to better understand the chemical and physical processes potentially controlling the elevated ground-level O₃ in the basin during 2013. The additional WRF simulation included different snow albedo and heat capacity assumptions, and the additional CMAQ simulations included emissions perturbations tests and different surface deposition mechanisms. The results of the tests are discussed in the following sections.

3.1. CMAQ Base Case Simulation

The Base Case model simulation used a standard version of the WRF platform, the public release version of CMAQ, and the 2011 NEIv2 as discussed in section 2.

3.1.1. Ozone Evaluation

Figure 3 compares time series plots of O₃ predicted by the Base Case to observations from the NOAA mea-surements collected at Horsepool, and the Air Quality System (AQS) network (U.S. EPA, 2016a) at Ouray, Roosevelt, and White Rocks. Similar plots for O₃ at six other sites in the basin are shown in Figure S1. Hereafter, figures and tables labeled with “S” are located in the supporting information. Table S5 presents the daily daytime average (10:00 a.m. to 3:00 p.m. local time) model bias of surface O₃ at multiple sites within the basin. The Base Case had a large negative bias for O₃ at all sites except for the Fruitland site (Figures 3 and S1 and Table S5). While almost every site and hour shows a model underprediction for O₃, the degree of underprediction varies greatly among sites, with the worst performance at Horsepool, where nearby NO_x O&G emissions are high. Based on the statistical analysis for O₃ (Table S5), NO_x(Table S6), and VOCs (Tables S7–S10), the model performs better at sites with distances further away from O&G sources. These results can be seen at Fruitland, Rabbit Mountain, and White Rocks (Figures 3 and S1). Not only was the Base Case unable to simulate peak O₃ during daytime hours, but the predictions also showed considerable O₃ decay during the night that was not present in the observations. The negative bias for O₃ from the Base Case is consistent with previous studies that have identified errors in O&G emissions as a potential cause of poor model performance (Ahmadov et al., 2015; Emery et al., 2015; Karion et al., 2013; Warneke et al., 2014).

Figure 3.

Time series of surface O₃ at (a) Horsepool, (b) Ouray, (c) Roosevelt, and (d) White Rocks for the Base Case (red) and observations (black). Note the different y axis scales in the panels.

3.1.2. NO_xand VOCs Evaluation

The NO_x(NO + NO₂) predictions from the Base Case were evaluated against observations to determine whether emission adjustments were needed to improve the model performance. The CMAQ results from the Base Case were compared to the observed NO_x mixing ratios across the basin. The NOAA Cavity Ring Down Spectroscopy (CRDS) instrument (Fuchs et al., 2009; Wild et al., 2014) was used to measure true NO_x and total NO_y at Horsepool. The NOAA CRDS NO_x instrument is based on measurement of NO₂ by optical extinction at 405 nm and conversion of NO to NO₂ by reaction with excess ozone. It is not susceptible to conversion of other NO_y species, although it does require an approximately 2% correction for the conversion of NO₂ to higher oxides of nitrogen in excess ozone (Fuchs et al., 2009). The NOAA CRDS NO_y uses a 405 nm optical NO₂ detector with a heated quartz reactor (650°C) to convert oxidized reactive nitrogen to NO₂ (Day et al., 2002), and a second reactor with excess O₃ to convert NO to NO₂ (Wild et al., 2014). Chemiluminescence NO_x monitors with heated molybdenum oxide (MoO) converters for NO₂ determination were used at Horsepool by the University of Colorado (CU)-Institute of Arctic and Alpine Research (INSTAAR) group and by other groups at other sites in the basin. These measurements are susceptible to interferences from other oxidized nitrogen species (NO_z, which includes organic nitrates (NTR), peroxyacylnitrates (PANs), nitric acid (FHNO₃), dinitrogen pentoxide (N₂O₅), and nitrate radical (NO₃)) (Grosjean & Harrison, 1985; Lamsal et al., 2008; Steinbacher et al., 2007; Winer et al., 1974). Good agreement between MoO and other NO₂ measurement techniques has been demon-strated for unpolluted to moderately polluted conditions. However, the error in the MoO-NO₂ determination increases with ambient levels of these NO_z species, particularly in polluted atmospheres (Gilge et al., 2013). Because of this interference, the comparison of MoO-NO_x measurements with the sum of modeled NO and NO₂ needs to be done with caution. To account for potential interference in the MoO-NO_x measurements, they were compared to modeled NO_x*, which accounts for the contributions from NO_z (Lamsal et al., 2008):

$$N{O}_x*=N{O}_x+NTR+0.95*PAN+0.35*HN{O}_3$$ (1)

Figure 4 compares the Base Case modeled NO_x and NO_x* (from equation (1)) to the MoO-NO_x at Ouray, Roosevelt, and White Rocks. Other sites within the basin are shown in Figure S2. For the Horsepool site, Figure 4 compares the modeled NO_xand NO_y to the CRDS NO_x and NO_y. As shown in Figure 4, the modeled NO_x and NO_y are very similar at the Horsepool site, and the modeled NO_x and NO_x* are similar at other sites. This indicates very little oxidation of NO_xto NO_z in the model, which is consistent with the very low production of O₃ in the Base Case model. In contrast, observed NO_x is lower than the observed NO_y at Horsepool, indicat-ing substantial conversion of NO_x to NO_z, consistent with the high observed O₃ mixing ratio. The observed NO_y at Horsepool is also similar to the observed MoO-NO_x at the nearby Ouray site, with observations mostly in the range of 20 to 40 ppbv. This would be expected if the MoO-NO_x instrument has interference from NO_z.

Figure 4.

Time series of modeled (red) and observed (black) NO_x (dashed lines) and NO_y (solid lines) at (a) Horsepool. The observations at Horsepool were measured by the CRDS instrument. Time series of modeled NO_x (red) and modeled NO_x* (blue) compared to observed MoO-NO_x (black) at (b) Ouray, (c) Roosevelt, and (d) White Rocks. Model results are from the Base Case. Note the different y axis scales in the panels.

At Horsepool, the modeled NO_x has a strong diurnal variation, with peak values during the day in the range of 10 and 50 ppbv and nighttime peak values between 40 and 100 ppbv. However, the observed NO_x from the CRDS is lower with values mostly below 10 ppbv. The large positive bias in the model surface layer at night is a common feature that has been noted in previous modeling studies and is caused by a tendency for CMAQ to overestimate the stability, thereby underestimating the vertical mixing at night (Pleim et al., 2016). Model performance for NOx is of more interest for this analysis during the day, when photochemistry is active, than at night. At Horsepool, the Base Case is biased high for NOx at night, with smaller positive bias during the day (Figure 4 and Table S6). At the rest of the sites, the model NOx is biased low compared to the observations. The NOz interference in the measurements can be ruled out as the cause of negative model bias because the model NOx* is also biased low at these sites.

Spatial plots comparing the Base Case modeled NO_x* to MoO-NO_x monitored by the AQS network also show high modeled NO_x* at Horsepool and strong spatial gradients in modeled NO_x* in the area between Horsepool and Ouray (Figure S3). These results are consistent with the large modeled NO_x emissions near Horsepool shown in Figure 2 and differences between the predicted NO_x at Horsepool and Ouray (Figure 4). As a result, the model predicts substantially higher NO_x mixing ratios at Horsepool, while the mon-itoring data show similar NO_x mixing ratios at Horsepool and Ouray. The Base Case model NO_x* is generally biased low compared to observed MoO-NO_x (Table S6), suggesting that there may be missing sources of NO_xemissions in the 2011 NEIv2 in some parts of the basin. This could limit O₃ production in those areas of the basin. Based on these analyses, the NO_x emissions in the 2011 NEIv2 may be too high at Horsepool, but need to be increased throughout the rest of the domain. These results are consistent with recent updates to the 2014 NEI (UDEQ, personal communication, 2017) that show lower NO_x emissions near Horsepool and increased NO_x emissions in other areas of the basin.

The NO_x emissions from the Deseret power plant in the Uinta Basin are well characterized due to continuous emissions reporting requirements. Since this facility is a large component of the total basin NO_x emission inventory (Table 1), contribution to surface NOx was estimated with a sensitivity simulation, where the facility was zeroed out in the model emissions input files. The contribution of surface NOx from the Deseret power plant in the southeast corner of the basin can be notable during the daytime in close proximity to the facility but minimally impacts other areas of the basin during this episode (Figure S4). Some of these impacts may be overstated if vertical layer stratification is not well represented in the model during the daytime, but most impacts from Deseret tend to be very localized and ventilate out of the basin to the east and south away from the area of high O&G activity, suggesting that it is not contributing to model performance issues at surface monitors.

Figure 5 compares Base Case modeled and observed PAN, pro-pionyl peroxynitrate (PANX), the sum of acrylyl peroxynitrate and methacrylyl peroxynitrate (OPAN), HNO3,N2O5, and aerosol nitrate at Horsepool. The model is biased low for all forms of NOz in the Base Case simulation. The model is also biased low for aerosol nitrate, and thus excessive model conversion of HNO3 to aerosol nitrate is not the cause of the model negative bias for HNO3 or NOx.

Figure 5.

Time series of the Base Case modeled (red) and observed (black) (a) PAN, (b) OPAN, (c) PANX, (d) aerosol nitrate, (e) HNO₃, and (f) N₂O₅ at Horsepool. Model results from the No O₃ Dep Case (blue) are also included in this figure. Note the different y axis scales in the panels.

Figure 6 compares the sum of a set of observed VOCs, in parts per billion by carbon (ppbC), from the NOAA surface monitors to the comparable modeled species in the Base Case at Horsepool. The figure shows that the Base Case has a large negative bias for VOCs during this high O₃ episode. Speciated VOCs from the can-ister data are also underpredicted when compared to modeled species (Tables S7–S10), but the degree of underprediction varies greatly among sites. The negative bias for VOCs from the Base Case is consistent with another study that also modeled this high O₃ event (Ahmadov et al., 2015).

Figure 6.

Time series of the Base Case (red) total VOC and observed (black) total VOC at Horsepool. The total VOCs include formaldehyde, methanol, C8–C12 aromatics, and acetaldehyde (excludes ethane and methane).

3.1.3. WRF Evaluation

The meteorological model inputs to CMAQ were evaluated using data from the selected research sites illu-strated in Figure 1. Figure 7 presents the performance for several near-surface meteorological variables in the form of a time series averaged across seven of the 14 research sites. The sites used in the statistical analysis included Fruitland, Myton, Rabbit Mountain, Rangely, Red Wash, Roosevelt, and Vernal. The other sites where eliminated due to incomplete observation data. Figure 7 has embedded statistics in the top right corner considers temperature, wind speed, and wind direction from Fruitland, Myton, Rabbit Mountain, Rangely, Red Wash, Roosevelt, and Vernal.

Figure 7.

Time series showing the performance of the base WRF simulation against an average of measurements at seven of the research sites (Fruitland, Myton, Rabbit Mountain, Rangely, Red Wash, Roosevelt, and Vernal). Time series include (a) 2 m temperature, (b) 10 m wind speed, and (c) 10 m wind direction. Statistical measures are provided in top right corner and include the root-mean-square error (RMSE), model bias (BIAS), and Index of Agreement (IOA).

The root-mean-square error (RMSE) of the time series for 2 m temperature, 10 m wind speed and direction were 2.02 K, 0.55 m s⁻¹ and 36°, respectively. A clear diurnal oscillation of wind direction evident in the obser-vations is reasonably replicated by the 4 km WRF simulation, especially for such a complex area under light large-scale flow. This oscillation from a daytime upslope flow caused by warming of the slopes surrounding the basin and an opposite nighttime drainage flow was shown in observations examined by Schnell et al.(2016). Wind speed errors and bias are low because of the light winds, but the model does depict the small variations of these light winds. Temperature has a clear night and daytime warm bias for the first half of the time series, but error levels are reasonable given the conditions. These statistics are generally consistent with model performance in areas with complex terrain (Bowden et al., 2015; McNally, 2009; Neemann et al., 2015).

Figures 8 and 9 show vertical profiles of air temperature predicted by the Base Case and compared to vertical profiles of data collected at Horsepool and Ouray, respectively. Figure S5 also shows the vertical air tempera-ture profiles at Fantasy Canyon. The comparisons are shown for 1 February 2013 and 3 February 2013 at 12:00 p.m. and 6:00 p.m. The times were selected based on when photochemistry is active and when the data were available. Evening times were selected rather than nighttime periods because very few measurements were collected during the nighttime. The specific time points selected are representative of the model’s general performance. On multiple days, the overall vertical temperature gradients of WRF were similar to observations in the lowest hundred meters, but the model tends to overestimate the temperature throughout the diurnal cycle, with the largest warm bias at night.

Figure 8.

Vertical profiles of air temperature predicted by the Base Case (red) and compared to observations (black) at Horsepool on (a) 1 February 2013 at 12:00 p.m. local time, (b) 1 February 2013 at 6:00 p.m. local time, (c) 3 February 2013 at 12:00 p.m., and (d) 3 February 2013 at 6:00 p.m. local time. The dashed line represents the modeled PBL height. The error bars represent the range of data points collected within the hour that matched the model.

Figure 9.

Vertical profiles of air temperature predicted by the Base Case (red) and compared to observations (black) at Ouray on (a) 1 February 2013 at 12:00 p.m. local time, (b) 1 February 2013 at 6:00 p.m. local time, (c) 3 February 2013 at 12:00 p.m., and (d) 3 February 2013 at 6:00 p.m. local time. Dash line represents the modeled PBL height. The error bars represent the range of data points collected within the hour that matched the model.

In addition to uncertainty in the O&G emissions, bias in predicted O₃, VOCs, and NO_x could be impacted by the performance of the boundary layer description in the WRF model. The model generally predicted warmer temperatures relative to the observations during the middle of the day (12:00 p.m. local time) and evening (6:00 p.m. local time). This suggests that the model has a deeper boundary layer and greater vertical mixing during the day. The model does predict a very different near-surface temperature or stability structure at Horsepool on 1 February 2013 at 12:00 p.m., where the model simulates 200 m deep mixed layer. However, out of all profiles across all sites (not shown), this was an anomaly. The model also predicts a mid-day mixed layer around 200 m to 300 m, while the observations show a capping inversion at around 100 m to 150 m. Colder modeled near-surface temperatures over the period would typically reduce the midday mixed layer closer to the 100 m to 150 m shown in these observed temperature soundings. An example of this is discussed in section 3.2, which tested a higher snow albedo specification as suggested by Neemann et al.(2015) and resulted in cooler daytime temperatures. This systemic issue in the WRF modeling is likely contributing to the negative bias in the predicted NOx and VOC concentration at the surface because the emissions are being diluted in a deeper mixed layer.

The warm bias in the evening is also similar to Neemann et al. (2015). Cold air pools are often associated, as in this case, with low clouds/fog and at these temperatures liquid fog with some ice crystals present. However, numerical model microphysics schemes often struggle to accurately simulate the phase properties during these events (Gultepe et al., 2014; Neemann et al., 2015; Zängl, 2005). In this case, near surface ice phase crys-tals were observed by satellite at night that were mixed with supercooled liquid water (Neemann et al., 2015). The Morrison microphysics scheme used in the Base Case had no ice crystals, only cloud water from the sur-face up to about 150 m that developed overnight and dissipated around 10:00 a.m. local time. This all-liquid fog in WRF more effectively traps longwave radiation than ice crystals, which limits the minimum tempera-ture overnight. This could be a factor in warm bias, but not tested in this study. A final issue for a modeling problem that did result in cooler nighttime temperatures is the modification of the specific heat capacity of snow (additional discussion included in section 3.2). It is directly proportional to snow density, which has a large range of values. The P-X LSM has a static snow density of 250 kg m⁻³, which is not representative of this modeling domain in most midwinter cases. The impact of snow heat capacity on the meteorology and chemistry is explored in section 3.2.

3.2. Sensitivity of CMAQ to WRF Snow Albedo and Heat Capacity

Given the warm bias of the model, a sensitivity test was conducted to investigate the model sensitivity to cooler near-surface air temperatures. This test used the Base Case emissions (public release of the 2011 NEIv2 and baseline CMAQ platform), but the land surface model in WRF was set to a more realistic snow heat capacity that reduced nighttime temperatures. The snow density and subsequent heat capacity of snow were also changed in this simulation to be reflective of the western United States. The initial WRF P-X LSM scheme used in the Base Case simulation assumed to have a static value for the snow density of 250 kg m⁻³. The lower snow density results in an exponentially lower heat capacity because of the larger amount of air space between snowflakes. If saturation is not reached, the lower heat capacity will allow faster cooling in the eve-ning, and subsequently lower nighttime temperatures. Instead of assuming a density of 250 kg m⁻³ for gen-eralized snow conditions, this sensitivity test simulation assumed a value of 50 kg m⁻³. Also, to improve upon a general daytime warm bias, the snow albedo was changed from 0.60 to 0.80 for the basin as observed dur-ing the 2013 UBWOS field campaign (Helmig et al., 2014) and suggested by Neemann et al. (2015). Neemann et al. (2015) showed that a value up to 0.80 for fresh snow over rangeland/shrub improved a warm bias in their modeling. This sensitivity test simulation is referred to as the Base-SNOW Case.

Temperature and O₃ vertical profiles from the surface to about 300 m were examined between 1 February 2013 and 10 February 2013 at Horsepool, Ouray, and Fantasy Canyon (not shown). The vertical profiles were generally collected between 7:00 a.m. and midnight local time. Figures 10 and 11 present examples of a day-time and a nighttime vertical profile of O₃ and air temperature at Horsepool and Ouray, respectively. When the temperature profile of the Base-SNOW Case simulation is improved relative to the Base Case and nearly identical to the observed profile, the predicted O₃ profiles in the Base Case are not significantly impacted by the cooler near-surface profile. During the day when the predicted temperatures match the observations in the lower 100 m, the predicted O₃ from the Base-SNOW Case decreases relative to the Base Case, which is a response to a shallower mixed boundary layer. This cooler daytime temperature and shallow mixed layer are a direct response to the higher albedo, but the improved meteorology actually slightly degrades the O₃ model performance in these cases, which complicates the analysis. This could suggest that larger adjust-ments could be applied to the O&G emissions.

Figure 10.

Vertical profiles of air temperature and O₃ predicted by the Base Case (red) and Base-SNOW Case (blue) com-pared to observations (black) at Horsepool. This figure presents the vertical profiles of (a) temperature and (b) O₃ on 2 February 2013 at 2:00 p.m. local time, and (c) temperature and (d) O₃ on 3 February 2013 at midnight (local time).

Figure 11.

Vertical profiles of air temperature and O₃ predicted by the Base Case (red) and Base-SNOW Case (blue) compared to observations (black) at Ouray. This figure presents the vertical profiles of (a) temperature and (b) O₃ on 1 February 2013 at 8:00 p.m. local time and (c) temperature and (d) O₃ on 4 February 2013 at 12:00 p.m. (local time).

Out of all temperature profiles examined at Horsepool, Ouray and Fantasy Canyon (177 total), the Base-SNOW Case was improved over the Base Case in 119 (68%) of those. These sites were selected based on the available data during this time period. The Base Case was better in 38 (22%) and there was not much difference in 20 (11%). When O₃ was examined for these same profiles only about 5 (3%) showed the Base-Snow Case meteorology actually improved O₃. In 120 or 68% of the profiles, the O₃ performance was degraded by the better temperature profiles. In almost all of those cases, the colder boundary layer reduced O₃, where the model was already low. In 50 cases (29%) the meteorology differences resulted in almost no change in O₃. This informs that the complexity of PCAPs and its impact on CMAQ is not completely understood, except that the sensitivity to the emissions changes is a much stronger driver of improving the significant underpredic-tion of O₃ that exists in both the Base and Base-SNOW cases. Given these results, the model sensitivity tests presented in the following sections used the standard WRF platform (i.e., did not adjust the snow albedo or heat capacity in WRF).

3.3. Sensitivity to CMAQ VOC Emissions Perturbations

Due to the substantial discrepancies between the standard CMAQ predictions and observed speciated VOCs (see Figure 6 and Tables S7–S10), an additional CMAQ simulation (hereafter referred to as “VOC Case”) was conducted to test whether perturbations in O&G emissions could improve the CMAQ predictions of O₃ in the basin. Previous studies have also suggested that VOC emissions are underpredicted in this area contain-ing O&G operations. Emery et al. (2015) increased VOCs by a factor of 2.5 and formaldehyde by a factor of 12.5, and could not produce sufficient O3 in CAMx. Ahmadov et al. (2015) performed WRF-Chem model simulations using the 2011 NEIv1 emissions inventory and found that the O&G CH4 and VOC emissions were underpredicted by factors of about four and two, respectively. Ahmadov et al. (2015) also used top-down emission estimates to define the speciation of the VOC emissions by using the ratio of observed speciated VOC to CH4 at Horsepool, and spatially allocated the emissions using well pad counts. A similar approach was initially tested with CMAQ, where the 2011 NEIv2 CH4 emissions were increased by a factor of four and then the VOC emissions were adjusted by the observed VOC/CH4 ratios collected at Horsepool. While Ahmadov et al. (2015) decreased the NOx emissions, here NOx emissions are not adjusted because the Base Case simulation was biased low at many monitor locations and high at a single location (i.e., Horsepool).

Therefore, this approach only applied the VOC emissions adjustment from the Ahmadov et al.(2015). While exploring this initial approach with CMAQ, 2011 NEIv2 CH₄ mixing ratios were found to be concentrated in the natural gas production area located in the eastern part of the basin, with lower CH₄ mixing ratios in the primarily oil production region in the western side of the basin. Thus, scaling VOC O&G emissions to CH₄ did not result in increased VOC emissions in the western part of the domain. As a result, CMAQ continued to have large negative bias for O₃ in the western side of the basin in these initial tests (not shown).

Given that the model predictions were not significantly and consistently improved when using CH₄ to scale the VOC O&G emissions across the Uinta Basin, an alternative approach was adopted, where the 2011 NEIv2 O&G emissions of individual VOC species were increased across the basin by a species-specific factor selected to match the observed speciated VOC mixing ratios at Horsepool. The factors by which the 2011 NEIv2 VOC emissions for O&G were increased for each modeled hydrocarbon are listed in Table 2. These factors were not based on certain days or times, but were adjusted to obtain model results that had magnitudes similar to the observations given the large discrepancies in the Base Case. For some VOC model species listed in Table 2, the 2011 NEIv2 O&G emissions inventory did not contain any estimated emissions. Therefore, adjustments were made based on the CB05 model species PAR from O&G sources. The adjustments using PAR were applied in a two-step process. First, the total PAR emissions estimated in the 2011 NEIv2 for O&G were increased by a factor of 3.5 to match the observed mixing ratio at Horsepool. The resulting 2011 NEIv2 O&G PAR emissions were then multiplied by a factor to match the observed mixing ratios at Horsepool for the CMAQ species that did not contain any estimated emissions. These factors were applied across the basin. In addition to the adjustment factors applied to the 2011 NEIv2 VOC O&G emissions, Table 2 lists the O&G VOC emission totals from the Base Case and the total VOC emissions resulting from the adjustments.

Table 2.

Oil and Gas VOC Emission Totals and Adjustments (Multiplication Factors) Applied to the 2011 NEIv2 O&G Emissions in Order to Match the Observed Mixing Ratios at Horsepool

		Base case total	Adjusted total
Modeled  species	Description			Adjustment factor	Scaling species
		Episode total tons
CH4	Methane	5,939	17,817	3.0	CH4
ALD2	Acetaldehyde	0.013	0.22	16.4	PAR*3.5
ALDX	C3 + aldehydes	0.004	0.007	1.6	PAR*3.5
ETH	Ethene	0.281	2.76	9.8	PAR*3.5
ETHA	Ethane	708	1,417	2.0	ETHA
ETOH	Ethanol	0.004	0.022	5.3	PAR*3.5
FORM	Formaldehyde	0.774	15.2	19.6	PAR*3.5
IOLE	R-HC=CH-R	0.204	0.2	1.0	IOLE
MEOH	Methanol	1.08	217	200	MEOH
OLE	Alkenes	0.780	2.8	3.5	PAR*3.5
PAR	C-C bond	2,078	7,273	3.5	PAR
TOL	Toluene and similar	72	649	9.0	TOL
UNR	Unreactive	275	275	1.0	UNR
XYL	Xylene and similar	44	654	15.0	XYL
BENZENE	Benzene	8.5	51	6.0	BENZENE

This approach resulted in increases of highly reactive VOC emissions distributed more evenly across both oil and gas production areas within the basin. Some highly reactive VOC species, such as aldehydes, can be emitted directly and also produced as secondary products from other VOC oxidation reactions. Therefore, the use of the observed ratio to PAR at Horsepool may overestimate direct emissions of these species. However, direct emissions of formaldehyde (HCHO) contribute a larger proportion of total HCHO in the proxi-mity of industrial activity and in the winter. Of more concern, the actual ratios of individual VOC species to PAR may vary across the basin, and the use of a constant ratio based on measurements at Horsepool intro-duces significant uncertainty into the estimates of speciated VOC emissions. As a result, this approach may overestimate or underestimate VOC emissions in other parts of the basin and should only be considered as an indication of the magnitude of emissions that may be missing, not a final correction to the inventory. Regardless, the predicted O₃ significantly improved and compared better to the observations across the entire Basin as discussed in more detail below.

The 2013 UBWOS measurement campaign also observed elevated daytime mixing ratios of nitrous acid (HONO) of 1 to 2 ppbv (Roberts et al., 2014). However, in a subsequent Uinta Basin Winter Ozone Study con-ducted in 2014 (2014 UBWOS), further analysis determined that these HONO measurements were subject to interference from peroxynitric acid (HO₂NO₂) (Stoeckenius, 2015; Stoeckenius & McNally, 2014; Wild et al., 2016). Therefore, a heterogeneous chemistry source of HONO from 2014 that matched the average observed daytime (11:00 a.m. to 4:00 p.m. local time) HONO mixing ratio for winter O3 episodes was added to the 2011 NEIv2 emissions as part of the VOC Case because of the issues identified with the 2013 UBWOS measure-ments. However, the additional source of HONO did not appear to impact the model chemistry in the basin.

A CMAQ simulation was conducted using the adjusted O&G VOC emissions (VOC Case). Figure 12 compares the time series of O3 predicted by the VOC Case to the Base Case and observations. Additional time series of O3 at other sites are presented in Figure S6. Table S5 also presents the daily daytime O3 (10:00 a.m. to 3:00 p. m. local time) predicted by the VOC Case at multiple sites. The predicted peak O3 at Horsepool by the VOC Case increased by as much as 70 ppbv compared to the Base Case. While the modeled O3 increased at all sites in the VOC Case, most sites continued to be biased low for peak O3 during the daytime and had larger nega-tive bias at night.

Figure 12.

Time series of surface O₃ at (a) Horsepool, (b) Ouray, (c) Roosevelt, and (d) White Rocks for the Base Case (red), VOC case (blue), and observations (black). Note the different y axis scales in the panels.

Figures 13 and 14 show that the magnitude of CH₄ and speciated VOCs predicted by the VOC Case at Horsepool improves in the VOC Case relative to the Base Case. However, the model tends to be biased low during the day and biased high at night. For example, HCHO is typically underestimated by as much as a fac-tor of two at 11:00 a.m. local time and overestimated by a factor of two or more at night. The negative bias for HCHO and acetaldehyde at midday indicates that the model is still biased low for VOC oxidation chemistry during the day. In addition to the uncertainty in the O&G emissions, the negative bias of the predicted VOCs could be related to the performance of the WRF model. The consistent warm bias over the period could be producing greater vertical mixing during the day causing the emissions to be diluted in a deeper mixed layer. The positive bias in the surface NO_x and VOC at night and surface NO_x during daytime could also be a result of the shallow mixing layer, where the emissions are being trapped near the surface.

Figure 13.

Time series of the Base Case (red) and VOC Case (blue) surface (a) methane, (b) formaldehyde, (c) PAR, and (d) acetaldehyde compared to observations (black) at Horsepool. Note the different y axis scales in the panels.

Figure 14.

Time series of the Base Case (red) and VOC Case (blue) surface (a) benzene, (b) toluene, (c) xylene, (d) ethylene,(e) methanol, and (f) ethane compared to observations (black) at Horsepool. Note the different y axis scales in the panels.

Modeled VOCs were also compared to canister samples for ethylene, ethane, benzene methanol, toluene, xylene, olefins, and total alkanes collected between 7:00 and 9:00 a.m. local time at six sites (Horsepool, Seven Sisters, Fruitland, Vernal, Wells Draw, and Roosevelt) within the basin on 1, 3, 6, and 8 February 2013 (Lyman et al., 2014; Lyman & Tran, 2015). Figure 15 shows the spatial distribution of total VOCs predicted by the Base and VOC cases and measured by the canisters on 1 February 2013 and 3 February 2013. The spatial distribution of the averaged predicted and observed mixing ratios of individual VOCs (BENZ, ETH, ETHA, IOLE, OLE, PAR, TOL, XYL) on the days collected by the canisters are also shown in Figures S7–S14 and Tables S7–S10. The model results are averaged between 7:00 a.m. and 9:00 a.m. to compare against the canister data. The Base Case was generally biased low compared to the canister data, and the model performance for the VOC Case generally improved with increasing the VOC emissions. While there remains significant uncertainty in VOC emissions across the basin, the VOC Case resulted in improved CMAQ performance across the basin for both VOCs and O3. Therefore, the model sensitivity simulations presented in section 3.4 were based on the VOC Case.

Figure 15.

Spatial distribution of averaged predicted and measured total VOCs between 7:00 a.m. and 9:00 a.m. local time by the (a) Base Case and (b) VOC Case on 1 February 2013 and the (c) Base Case and (d) VOC Case on 3 February 2013. Total VOCs are as listed in Table S7.

3.4. Sensitivity to Surface Deposition Mechanisms

Additional CMAQ simulations were conducted to better understand the physical processes potentially controlling elevated ground-level O₃ in the basin during 2013. The additional model simulations included different dry surface deposition approaches. Recent research on snowpack processes and atmosphere-snow gas exchange has demonstrated that chemical and physical interactions between the snowpack and the overlaying atmosphere have a substantial impact on the composition of the lower troposphere. Observations suggest that O₃ surface exchange with the snowpack is widely variable, may be upward or downward, and possibly depends on the quantity and composition of previously deposited trace gases, solar irradiance, snow temperature and substrate below the snowpack (Bocquet et al., 2011; Helmig, Apel, et al., 2009; Helmig, Cohen, et al., 2009; Helmig et al., 2007; Helmig, Seok, et al., 2009, Helmig et al., 2014; Wesely & Hicks, 2000). Measured O3 fluxes from snow range from 3 to 2 cm s1, although most data are within 0 and 0.2 cm s1 (Bocquet et al., 2011; Helmig, Apel, et al., 2009; Helmig, Cohen, et al., 2009; Helmig et al., 2007; Helmig, Seok, et al., 2009, Helmig et al., 2014; Wesely & Hicks, 2000). In comparison, midsummer values for lush vegetation surfaces have a much greater O3 deposition velocity of 0.2 cm s1 to 1.0 cm s1 (Wesely & Hicks, 2000).

The land surface in the Uinta Basin generally consists of dry, sandy soil and low-lying sage brush. During the 2013 UBWOS campaign, flux measurements were made at Horsepool (Helmig et al., 2014), which included sur-face measurements during both the snow-covered and snowmelt (exposed dirt surfaces) periods. Between 25 January and 19 February 2013, snow cover was less than a half a meter deep and sufficiently covered the soil. However, the leafless and dry sage brush still protruded well above the snowpack. The median measured daytime O₃ deposition velocity at Horsepool was 0.003 cm s⁻¹, with a range from −0.071 to 0.079 cm s⁻¹ (Helmig et al., 2017). Ozone deposition velocities increased steadily during snowmelt and reached a maximum of 0.14 cm s⁻¹ (day) with an overall median velocity of 0.002 cm s⁻¹ (day and night) for dry soil after all the snow had melted in the area (Helmig et al., 2017). In CMAQ, leaf area index (LAI) and vegetation fraction play an important role in the calculations of deposition velocities. The model sets the LAI of the land use types in the basin (i.e., scrub, dwarf scrub, barren) during winter to the default value of 1.0 and the vegetation fraction to 0.5. These values are generalized values for all sites and are likely to be high for this site based on observa-tions. This could lead to model overpredictions of depositions velocities of O₃ and other species.

Two model sensitivity simulations with varying surface deposition mechanisms were conducted to investi-gate the impact of surface deposition and to better represent the observed deposition velocities in the stan-dard model for this area. In one simulation, the LAI and vegetation fraction were set to 0.001 for grid cells where the dominant land use type was dwarf scrub, scrub or barren and were snow covered (hereafter referred to as the LAI Case). This case reflects the lower values of LAI and the sparse vegetation reported in Helmig et al. (2014) and affects the deposition of all depositing species. These changes reduce the stomatal uptake by the vegetation, which is the major pathway for O3 deposition, as well as the uptake by the cuticles. The deposition velocity of the Base Case ranged from about zero to 0.21 cm s1, while the LAI Case ranged from about zero to 0.1 cm s1.

An additional simulation was also performed where the O₃ dry deposition velocity was set to zero in CMAQ (hereafter referred to as the No O₃ Dep Case). This case was motivated by the distribution of observed deposition velocities (Helmig et al., 2014), which shows both positive and negative deposition velocities and is centered near zero. The No O₃ Dep Case also included these deposition velocity adjustments applied in the LAI Case to better represent the observed velocities for all species defined in the model. The deposition sensitivity simulations are intended to provide a robust model response to removing the dry deposition sink for O₃ recognizing that some O₃ will deposit to the surface. Additional studies are needed to explore the implementation of mechanistic approaches to improve O₃ dry deposition velocities in this area under these conditions. Both of these sensitivity studies included the increased VOC emissions described previously in Section 3.3.

Figure 16 compares the observed O₃ to the predicted O₃ from both deposition sensitivity tests and the VOC Case at Horsepool, Ouray, Roosevelt, and White Rocks. Comparisons of these model results for other sites within the basin are shown in Figure S15. Table S5 also presents the daily daytime O₃ (10:00 a.m. to 3:00 p.m. local time) predicted by the No O₃ Dep Case at multiple sites. While the LAI and vegetation fraction adjustments in the LAI Case made little difference in the overall model performance, the model was sensitive to the O3 deposition, especially at night. Along with applying the adjustments made in the LAI Case, setting the O3 dry deposition velocity to zero in the No O3 Dep Case increased the nighttime O3 substantially, counteracting the O3 decay at night that caused disparity between the model and observations. The Horsepool site most likely showed less sensitivity to this case than other sites because of the significantrole that NO plays in O3 decay (relative to dry deposition at Horsepool). The No O3 Dep Case also impacted nighttime PAN, OPAN, PANX, HNO3,N2O5, and aerosol nitrate at Horsepool (see Figure 5), while the remaining VOC and NOx species were not significantly impacted by having no O3 deposition. As discussed in section 3.1.2, the Base Case was biased low for all forms of NOz. However, setting the O3 deposition velocity to zero (i.e., No O3 Dep Case) improves the model performance for these species (see Figure 5).

Figure 16.

Time series of surface O₃ at (a) Horsepool, (b) Ouray, (c) Roosevelt, and (d) White Rocks for the VOC Case (red), LAI Case (blue), No O₃ Dep Case (green), and observations (black). Note the different y axis scales in the panels.

Figure 17 shows the spatial distribution of O₃ predicted by the LAI and No O₃ Dep cases on 3 February 2013 at 12:00 p.m. and 6:00 p.m. local time. Relative to the LAI Case, setting the O₃ deposition to zero generally impacted the center portion of the basin, where snow cover is more dominant.

Figure 17.

Spatial distribution of surface O₃ predicted by the (a) LAI Case and (b) No O₃ Dep Case on 3 February 2013 at 12:00 p.m. local time and the (c) LAI Case and (d) No O₃ Dep Case on 3 February 2013 at 6:00 p.m. local time.

Figures 18 and 19 show the vertical profiles of O₃, NO_x, and total VOC predicted by the model and captured by tethered balloons at Horsepool on 1 February 2013 and 3 February 2013, respectively. Vertical profiles of O₃ at Ouray and Fantasy Canyon are also presented in Figures S16 and S17, respectively. Adjusting the O&G VOC emissions and setting the O₃ deposition to zero (i.e., No O₃ Dep Case) generally improved the predicted O₃ at the surface and higher layers at these sites (Horsepool, Ouray, and Fantasy Canyon) on the selected days. In addition to the overprediction of the surface NO_x at Horsepool, as discussed in section 3.3 and shown in Figures 18 and 19, the predicted NO_x at Horsepool is also overpredicted at layers above the surface for the Base Casen and No O₃ Dep Case. However, the overprediction of VOCs for the No O₃ Dep Case further suggests that applying a constant adjustment factor to the O&G emissions introduces additional uncertainty into the estimates of speciated VOC emissions. As a result, this approach in increas-ing the VOC O&G emissions may overestimate or underestimate VOC emissions. Regardless, the predicted O₃ significantly improved and compares better to the observations across the entire Basin in the No O₃ Dep Case.

Figure 18.

Vertical profiles of O₃ predicted by the Base Case (red) and No O₃ Dep Case (blue) and measured (black) at Horsepool on 1 February 2013 at (a) 12:00 p.m. and (b) 6:00 p.m. local time. Vertical profiles of predicted NO_x* and measured MoO-NO_xat Horsepool on 1 February 2013 at (c) 12:00 p.m. and (d) 6:00 p.m. local time. Vertical profiles of predicted and measured total VOCs at Horsepool on 1 February 2013 at (e) 12:00 p.m. and (f) 6:00 p.m. local time. Observed total VOCs are discussed in Helmig et al. (2014) and compared to similar modeled species. The error bars represent the range of data points collected within the hour that matched the model, and the squares represent the median. Note the different y axis scales in the panels.

Figure 19.

Vertical profiles of O₃ predicted by the Base Case (red) and No O₃ Dep Case (blue) and measured (black) at Horsepool on 3 February 2013 at (a) 12:00 p.m. and (b) 6:00 p.m. local time. Vertical profiles of predicted NO_x* and measured MoO-NO_xat Horsepool on 3 February 2013 at (c) 12:00 p.m. and (d) 6:00 p.m. local time. Vertical profiles of predicted and measured total VOCs at Horsepool on 3 February 2013 at (e) 12:00 p.m. and (f) 6:00 p.m. local time. Observed total VOCs are discussed in Helmig et al. (2014) and compared to similar modeled species. The error bars represent the range of data points collected within the hour that matched the model, and the squares represent the median. Note the different y axis scales in the panels.

In general, the Base Case and observation vertical structures of these pollutants are more similar for each indi-vidual site (Figures 18,19, S16, and S17). The Base Case and No O₃ Dep Case vertical structures are also similar at each site. The similarities between the Base Case and No O₃ Dep Case at each site are likely a result of the WRF air temperature stratification to stability (dT/dz) being similar to most of the observed profiles (section 3.2), even though the actual temperature is overestimated by as much as 3 to 5°C. In addition to the uncertainty in the O&G emissions, the overestimate or underpredictions of these pollutants could be related to the performance of the WRF model. The consistent warm bias could be producing greater vertical mixing during the day, causing the emissions to be diluted in a deeper mixed layer. The positive bias in NO_x and VOC near the surface at night could also be a result of the shallow mixing layer, where the emissions are being trapped near the surface.

4. Conclusions

Standard application of the CMAQ model with the 2011 NEIv2 emissions was not capable of reproducing the observed values of O₃, NO_x, and many VOC species in the Uinta Basin for a February 2013 inversion episode. Several emissions and model process adjustment simulations were performed to investigate O₃ formation in the basin. The WRF model generally performed well for the February 2013 inversion episode. However, the warm temperature bias and overestimated the daytime height of the PBL on some days may have contribu-ted to underestimates of daytime VOC, NO_x, and O₃ in the CMAQ simulations. As mentioned previously, CMAQ was biased high for surface NO_x at Horsepool but low for NO_x at most other sites across the basin. This finding is consistent with the direction of recent revisions by UDEQ in the NO_x emissions inventory. New model simulations with the updated 2014 NEI are needed to determine the magnitude of NO_x under-predictions and if these updates improve model performance for NO_x. The results also suggest that VOC emissions are generally underestimated at Horsepool and at other sites within the basin and that increases in magnitude generally improve predictions of O₃. The results further suggest that applying a constant adjustment factor based on measurements at Horsepool across the model domain introduces additional uncertainty into the estimates of speciated VOC emissions. Regardless, the predicted O₃ significantly improved and compared better to the observations across the entire Basin. The vertical mixing of emissions in CMAQ at night was low and mostly confined to the first few layers of the model (i.e., within 300 m). This caused nocturnal overpredictions of NO_x and VOCs at multiple sites within the basin. An analysis of a more realistic snow heat capacity in WRF to improve the warm bias and vertical mixing resulted in temperature profiles that compared better to observations, although the improved temperature profiles rarely resulted in improved O₃ profiles. While additional work is needed to investigate the meteoro-logical impacts, the results suggest that the sensitivity to the emissions changes contributes more to the underestimation of O₃ in the basin. The analysis of model surface deposition algorithms showed that the model is sensitive to various descriptions of the O₃ deposition velocity, particularly during the night. More research is needed to improve the treatment of deposition for this area in the model and to better under-stand the model’s mechanisms that improved the performance. There is a great deal of heterogeneity in the model behavior across the domain. The large differences in NOx and VOC emissions across the domain, particularly the locally large NOx emissions seen at Horsepool, result in variation of the relative importance of different chemistry processes across the basin. However, these mod-eled differences may be caused by errors in the emissions inventory and may not accurately represent ambi-ent conditions within the basin. Several key aspects of the NOx cycling process are highly uncertain in the model and would benefit from additional laboratory and theoretical data. Of particular concern is the char-acterization of the heterogeneous chemical processes, including reactions of N2O5 on tropospheric aerosols and the potential for chemical reactions on snow and ice surfaces. Critical gas phase processes that impact the NOx cycling have been better studied overall but are less certain under the winter conditions seen in the basin. This includes the fate of organic nitrate in winter and the types of nitrates most likely to be formed in areas with dense O&G operations, and the role of PANs as both a reservoir and source of transport of NOx within the basin during inversion episodes.

Additional simulations with a new version of CMAQ and the CB6r3 chemical mechanism are currently being evaluated to determine the impact of these updates on ozone pre-dictions. This mechanism includes comprehensively updated reaction rates, as well as additional HO_x and NO_x cycling pathways. While the chemistry updates are not expected to radically change model predictions, it should provide a more up-to-date characterization of PAN and alkylnitrate production and decay, as well as better representation of propane, which is a large component of O&G emissions.

The results from this study are useful for understanding the factors important for predicting air quality impacts associated with oil and natural gas production in the Uinta Basin and identifying where future research should be focused to improve air quality models for studying these conditions. Consistent with Ahmadov et al. (2015), this study found that standard photochemical models need to be adjusted to repre-sent the effects of snow albedo on photolysis rates and O₃ deposition velocities to snow. With these updates, photochemical models have the ability to predict wintertime O₃ in the area. The key input parameters and model configurations to consider for this area include the oil and gas emissions, treatment and values for gas deposition velocities, and the treatment of boundary layers and ice fog processes. While relatively simple adjustments to the emissions and deposition can improve the performance, it is important that any updates be supported by thorough scientific research and evaluated over a large variety of conditions. To further improve O₃ wintertime modeling in this domain, future work should concentrate on improving VOC and NO_x emissions estimates, in terms of the spatial allocation of emissions, magnitudes, and speciation profiles for the O&G sector. Additional work is also needed to improve deposition velocities to snow for species that influence the O₃ concentration, including the deposition of PAN, O₃, and HNO₃. Future work should also eval-uate the role of snow as a reservoir and possible source of heterogeneous chemistry of VOC and NO_x(Helmig, Cohen, et al., 2009, Helmig, Apel, et al., 2009, Helmig, Seok, et al., 2009, Helmig et al., 2013, 2014; Williams et al., 2009). Additional investigation of the parameterizations used to represent boundary layers and ice fog processes in numerical models are needed in order to obtain improved cold air pool simulations (Gultepe et al., 2014; Neemann et al., 2015; Zängl, 2005).

Improving the ability of regulatory air quality models to predict elevated O₃ under all conditions, including winter and summer, is important in identifying source contributions to air quality problems and designing effective strategies to reduce harmful air pollutants. Addressing the common uncertainties identified by all air quality models, as in the Uinta Basin in winter, requires coordinated community participation.

• Model underestimated Uinta Basin winter ozone levels because oil and gas emissions appear to be underestimated in national inventory

• Ozone predictions are strongly sensitive to the treatment of deposition to snow at night

• Improvements to warm bias of modeled surface temperatures caused less mixing and more ozone titration and seldom resulted in improved ozone