Evaluating methods to estimate methane emissions from oil and gas

Aug 28, 2018 - ... worth of Chinese imports would raise the prices of countless products. ... Support. Get Help · For Advertisers · Institutional Sale...
1 downloads 0 Views 1MB Size
Subscriber access provided by University of South Dakota

Environmental Modeling

Evaluating methods to estimate methane emissions from oil and gas production facilities using LES simulations Pablo E Saide, Daniel Steinhoff, Branko Kosovic, Jeffrey Weil, Nicole Downey, Doug Blewitt, Steven Hanna, and Luca Delle Monache Environ. Sci. Technol., Just Accepted Manuscript • DOI: 10.1021/acs.est.8b01767 • Publication Date (Web): 28 Aug 2018 Downloaded from http://pubs.acs.org on September 1, 2018

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 29

Environmental Science & Technology

1

Evaluating methods to estimate methane emissions from oil and gas production facilities

2

using LES simulations

3

Pablo E. Saide (1,*, †), Daniel F. Steinhoff (1), Branko Kosovic (1), Jeffrey Weil (1), Nicole

4

Downey (2), Doug Blewitt (2), Steven R. Hanna (3), Luca Delle Monache (1)

5 6 7 8 9 10 11 12

(1) Research Applications Laboratory, National Center for Atmospheric Research, 090 Center Green Drive, Boulder, CO, 80301, United States. (2) Earth System Sciences, LLC, 117 Bryn Mawr Dr. SE Suite 111, Albuquerque, NM, 87106, United States. (3) Hanna Consultants, 7 Crescent Ave., Kennebunkport, ME, 04046, United States.

13

KEYWORDS: Source term estimation, methane emissions, transport and dispersion, stochastic

14

and variational approaches

15

ABSTRACT. Large-eddy simulations (LES) coupled to a model that simulates methane

16

emissions from oil and gas production facilities are used to generate realistic distributions of

17

meteorological variables and methane concentrations. These are sampled to obtain simulated

18

observations used to develop and evaluate source term estimation (STE) methods. A widely used

19

EPA STE method (OTM33A) is found to provide emission estimates with little bias when

20

averaged over six time-periods and seven well-pads. Sixty-four percent of the emissions

21

estimated with OTM33A are within +/-30% of the simulated emissions, showing slightly larger

22

spread than the 72% found previously using controlled release experiments. A newly developed

23

method adopts the OTM33A sampling strategy and uses a variational or a stochastic STE

24

approach coupled to an LES to obtain a better fit to the sampled meteorological conditions and to

25

account for multiple sources within the well-pad. This method can considerably reduce the

26

spread of the emissions estimates compared to OTM33A (92-95% within +/-30% percent error),

ACS Paragon Plus Environment

1

Environmental Science & Technology

Page 2 of 29

27

but it is associated to a substantial increase in computational cost due to the LES. It thus provides

28

an alternative when the additional costs can be afforded to obtain more precise emission

29

estimates.

30

Table of Contents (TOC)/Abstract Art.

31 32 33

INTRODUCTION

34

Uncertainties in methane emissions from oil and gas production 1, 2 and their potential influence

35

on climate change 3 have driven extensive research using regional and local field experiments

36

and theoretical analysis to better constrain estimates of emissions from these sources 4. While

37

source term estimation (STE) is used in multiple disciplines 5, there are a wide variety of

38

methods that have been proposed and used specifically to estimate methane emissions from oil

39

and gas production by using ground-based gas concentrations and meteorological observations 6-

40

12

41

verification of the ground-based methodologies using controlled releases, it is difficult to

42

quantify the accuracy of these methane STE methods for scenarios where factors such as terrain,

, as well as aircraft and remote sensing observations 13-16. Although there has been some

ACS Paragon Plus Environment

2

Page 3 of 29

Environmental Science & Technology

43

emissions from multiple sources with different gas compositions within a well pad, and time-

44

varying emissions representative of typical operations might influence the results.

45

In this work we run a large-eddy simulation (LES) coupled to a model that realistically simulates

46

emissions from an oil field (i.e., an emission simulator) to generate synthetic 4-D (space and

47

time) methane concentration fields, which we then use to test a variety of STE methods for oil

48

and gas production. This technique is known as an Observing System Simulation Experiment

49

(OSSE) in the data assimilation field 17, 18. LES simulations are starting to be included in

50

methane STE research 19-21, but to our knowledge OSSEs using an LES driven by emission

51

simulators have not been applied to test source estimation methodologies at the scale of a single

52

well pad (facility level). Although the word “observing” is used in the term OSSE, it is important

53

to stress that the observations do not correspond to field measurements of methane and

54

meteorology but are in fact simulated by the model, thus we refer to them as “simulated

55

observations”.

56

The Draft EPA Other Test Method (OTM) 33A 6 is one of the most widely-used STE methods

57

for individual facilities 22-24. Assumptions inherent in OTM33A include that methane

58

concentrations and wind observations (three velocity components) are available from a vehicle

59

that takes measurements downwind of the facility, and that the locations of emissions from major

60

sources are known. OTM33A is a simple and easy to employ method that has been evaluated

61

using controlled releases 6, and has been well documented with the code being freely available 25.

62

In this study we compare and contrast the skill in retrieving emissions of three STE methods:

63

OTM33A, a variational-based STE method, and an approach based on Bayesian inference and

64

stochastic sampling. The latter two methods use an LES simulation and constitute a novel

ACS Paragon Plus Environment

3

Environmental Science & Technology

Page 4 of 29

65

contribution to the field of methane emission estimation at the facility level. Their development,

66

advantages, and limitations are also discussed in this study.

67

METHODS

68

LES model configuration

69

The LES configuration is based on idealized simulations of the Weather Research & Forecasting

70

(WRF) LES model 26 version 3.8.1, which includes transport and dispersion of passive tracers 27.

71

WRF-LES is configured with an outer (coarse) domain with dimensions of 10.3 km by 12.5 km,

72

with 30 m horizontal grid-spacing and periodic lateral boundary conditions, and an inner (fine)

73

domain of dimensions 6.9 km by 8.3 km (688 x 829 grid-cells) and 10 m horizontal resolution

74

(Fig. 1a). This resolution and domain size allows the placement of multiple pads in the domain

75

with realistic spacing within each other. It also permits locating observations close-by to the pads

76

while still properly resolving turbulence in between pads and monitoring locations, and keeps the

77

computational and storage costs at a reasonable level. The inner domain does not feed

78

information back to the outer domain (i.e., one-way nesting is adopted in the modeling

79

configuration) to avoid the plumes re-entering the outer domain once they reach the boundary as

80

this would occur due to the periodicity of the boundary conditions. The simulations have 121

81

vertical levels, with grid spacing of ~3 m for the first three layers and ~10 m for the upper layers

82

up to the model top at 2 km height. The outer domain is initialized with southerly 5 m s-1

83

geostrophic winds and a temperature inversion of 8 K over 150 m to cap the boundary layer

84

height at 1 km. Note that due to the LES resolving turbulence and terrain, the winds within the

85

inner domain are not constant but exhibit natural variability similar to field observations (Fig.

86

2a,b). The inner domain covers a region with terrain obtained from USGS at 1/3 arc second (~10

ACS Paragon Plus Environment

4

Page 5 of 29

Environmental Science & Technology

87

m) resolution (http://ned.usgs.gov/) representative of an area located in the Barnett shale region,

88

Texas (33° latitude, -98.2° longitude), where intensive oil and gas operations are active (Fig. 1a).

89

The surface roughness is set at 0.1 m uniformly across the domain (which is representative of an

90

open field with brush/low vegetation) except for the location of tanks and compressor buildings,

91

where a value of 0.6 m is set to model the added turbulence due to these structures, and is chosen

92

as the upper range suggested for roughness produced by scattered settlements 28. For a surface

93

release, sensitivity analysis showed that mean surface concentrations decreased by less than 15%

94

at a distance of 100 m downwind from the source when using 0.6 m versus 0.1 m surface

95

roughness at the source location. The reductions were even lower for a source release from the

96

second model vertical level showing that even this upper range of surface roughness has a

97

limited impact on concentrations relevant for this study. Daytime convective conditions are

98

modeled by assuming a 0.2 K m/s sensible heat flux at the surface 29, which are typical

99

meteorological conditions found in the Barnett shale region. Background methane is set to 1.98

100

ppm and 1.85 ppm in the boundary layer and free-troposphere which correspond to the 2015

101

annual mean for the Southern Great Plains (OK site) and Mauna Loa (MLO site) NOAA flask

102

network measurements (https://www.esrl.noaa.gov/gmd/ccgg/flask.php), respectively. Three

103

hours of simulation are performed, with the first hour used as spin-up, and analysis performed

104

for the last two hours. The WRF-LES time-step for the inner domain is 0.1 s and instantaneous

105

values are saved every second (i.e., every ten time-steps). Additional details on the WRF-LES

106

configuration can be found in the Supplemental Information (SI).

107

Emissions

108

In the current modeling exercise, we simulate hypothetical oil and gas facilities (i.e., multiple

109

well pads) with emission rates and gas compositions based on information from real facilities.

ACS Paragon Plus Environment

5

Environmental Science & Technology

Page 6 of 29

110

Seven well pads containing two wells each and their associated equipment are included in the

111

model domain. The pads, with dimensions of 100 m x 100 m (i.e., 100 grid cells), are arranged in

112

the domain as shown in Figure 1a, which is consistent with the arrangement of actual pads in the

113

Barnett shale region based on satellite imagery. An oil field emission simulator is used to

114

generate emissions time series for individual components on each well pad. The simulator

115

predicts emissions based on operating parameters for the well pad equipment 30. The simulator

116

estimates methane emissions for condensate tanks (flashing), dehydrators, compressors,

117

pneumatic chemical injection (CI) pumps, pneumatic controllers, and liquids unloading at one

118

second temporal resolution. Components are distributed on wells as typically observed for the

119

Barnett shale region. The components, when present, are arranged spatially on each pad

120

following the distribution shown in Figure 1b. Emissions from tanks (flashing and liquids

121

unloading) were assumed to be released from the top of the tank, and thus were put into the

122

second model vertical level (~5 m height). Emissions from compressors were assumed to come

123

from a compressor building and thus were released into the first two model levels (2 and 5 m),

124

distributed in equal parts. The rest of the emissions (CI pump, pneumatic controller and

125

dehydrator) were released into the first model level. Emissions are assumed to be at ambient

126

temperature and without significant speed when emitted. Future studies should consider cases

127

where emissions have significant speed and/or are emitted at high temperature which could

128

affect the ability of the STE methods to properly retrieve emissions. Emissions from different

129

components and pads are assigned to twenty-two simulated tracers in the WRF-LES framework

130

(see assignments in the Table S3) which are then aggregated to build methane concentrations.

131

Ten days of emissions are generated with the simulator from which two hours are selected to be

132

used as input for the LES simulation. These two hours are selected to contain liquids unloading

ACS Paragon Plus Environment

6

Page 7 of 29

Environmental Science & Technology

133

events which can then be included or excluded for analyses of different scenarios as their

134

emission rates are several orders of magnitude greater than the other emission sources (200-400

135

g/s). With respect to the other sources over the two hours, the most frequent simulated emissions

136

come from pneumatic controllers, but their rate is generally below 0.3 g/s. Tank flashing and

137

compressor emissions are also frequent and show larger values on the 0.1-3 g/s range. One

138

dehydrator is simulated in well-pad #2 assuming constant emissions of ~1.5 g/s. CI pump

139

emissions are in the 0.3-3 g/s range and are quite infrequent (less than 5 minutes over 2 hours

140

and all pads). Further details including the emission time series (Fig. S1), emission histograms

141

(Fig. S2), and detailed descriptions of the emission simulator and emissions from each

142

component can be found in the SI text and tables S1-S3.

143

Simulated observations

144

The simulated observations are obtained by extracting data from the LES model output of the

145

inner domain following the OTM33A protocols and then adding a random error representative of

146

instrument uncertainties as described later in this section. The OTM33A measurement protocol

147

consists of ~20 minutes of meteorological and methane sampling from a vehicle parked

148

downwind of the facility at distances ranging from 20 to 200 m. Since turbulence and plume

149

dispersion are better resolved at longer distances in the LES, thena distance of 100 m north (i.e.,

150

downwind) of the largest emission source on a given well pad is selected. Based on this criterion,

151

all potential locations of measurements are shown in Figure 1b with the letter “O”. Although

152

winds are forced to be southerly on the LES outer domain, terrain effects could cause deviation

153

in the inner domain. All 20-minute sampling intervals for all well-pads were reviewed finding

154

that on average winds were southerly for all cases (e.g., see Table S4 and S5 in the SI),

155

confirming that the choice of placing observations directly north is appropriate. Also, given the

ACS Paragon Plus Environment

7

Environmental Science & Technology

Page 8 of 29

156

dimensions of the pads and the distance selected, the sampling location falls outside of the pad

157

area for all pads, where the location may be publicly accessible for measurements. For the

158

measurements to be valid for source estimation using the OTM33A approach, the mean wind

159

conditions should remain constant over the sampling period, which is why the large-scale

160

conditions forcing the LES are kept constant (see LES model configuration section). The LES

161

outputs are sampled every second (instantaneous values) at the first vertical layer (~1.8 m height)

162

producing six unique 20-minute time series for each pad over the two hours of LES simulations.

163

This layer is selected because is the closest to the ~ 3 m height at which instrumentation is

164

usually mounted (the second layer is at ~5.4 m height). Sensitivity simulations using

165

observations from the second layer were performed showing similar results (see SI section

166

“Sensitivity simulations”). Two scenarios are analyzed: (1) including emissions from liquid

167

unloading events, and (2) excluding emissions from liquid unloading events. Due to the

168

significant emission rate associated with liquids unloading events (more than an order of

169

magnitude larger than the sum of other emissions), liquid unloading events confound emission

170

estimates from nearby well pads.

171

Noise representing instrument uncertainties is added to the time series extracted from LES

172

outputs by assuming a normal distribution with zero mean and standard deviation (SD) based on

173

the precision of instrumentation typically used for OTM33A applications (cavity ring down

174

methane analyzer and ultrasonic anemometers). A sampling error (SD) of 1.5 ppb is used for

175

simulated methane observations, while 2 degrees SD and 1% or 0.05 m s-1 SD (whichever is the

176

largest) is used for wind direction and speed, respectively. An example of real25 and simulated

177

observations is found in Figures 2a and 2b, respectively, which suggest that the methodology

178

produces a similar representation of conditions encountered when sampling according to the

ACS Paragon Plus Environment

8

Page 9 of 29

Environmental Science & Technology

179

OTM33A emission estimation method. While the sampling error applied to methane

180

concentrations has negligible impact as the typical methane enhancements in this study are 1-10

181

ppm (i.e., 3-4 orders of magnitude larger), the errors applied to winds generate some of the noise

182

in the data when plotted as a function of wind spend and direction (e.g., compare Fig. 2b to Fig.

183

S2). Additional statistics on the mean, maximum, and minimum winds speed and direction can

184

be found in Table S4 and S5 of the supplement.

185

STE methods

186

Three STE methods are evaluated: the standard EPA OTM33A method and two new methods, a

187

variational inversion method 31, and a Bayesian inference and stochastic sampling method 32, 33.

188

The OTM33A relies upon measured data while the two new methods additionally require the use

189

of a LES to estimate emissions sources.

190

The OTM33A STE framework, referred to as Point Source Gaussian (PSG) method, is fully

191

described by Brantley et al. 6 and EPA 25 and only a brief overview is provided here. First, the

192

observations are collected by a stationary vehicle downwind of the source being estimated. Then,

193

the observed methane concentrations with the observed background subtracted are binned by

194

wind direction, and a Gaussian function is fitted to the resultant data (Fig. 2c). The maximum

195

value of the Gaussian fit is then converted from concentration to emissions by performing a two-

196

dimensional Gaussian integration that uses the mean wind speed, the distance between the source

197

and the stationary vehicle, and local stability classes derived from wind measurements 6. As

198

multiple sources are found within a well pad, the location used on the Gaussian integration is that

199

of the dominant source (the equipment with the largest emissions) for that time period.

ACS Paragon Plus Environment

9

Environmental Science & Technology

Page 10 of 29

200

The new methods use an LES simulation in place of the Gaussian integration to connect

201

concentrations to emissions providing additional flexibility to the STE methods. For instance, it

202

can facilitate the inclusion of multiple sources (PSG assumes a single source) and deviations

203

from the mean winds can be taken into account (PSG assumes constant winds over the whole

204

sampling period). The WRF-LES configuration is modified for its use in the STE methods to

205

avoid biasing the results due to the use of the same model when generating the simulated

206

observations. These modifications include: a different advection scheme and sub-grid turbulence

207

parameterization (see details in the SI), and changing the boundary wind speed to 5.5 m s-1 (10%

208

increase), and setting the sensible heat flux to 0.16 K m s-1 (20% decrease). Perturbations that

209

both change the internal variability and bias the model are introduced, as this would likely be the

210

case if this method is applied to field data. These changes in configuration generate large

211

differences in simulated wind fields and methane concentrations (Fig. S3 in the SI). These

212

differences make these simulations uncorrelated within each other most of the time (Fig. S4 in

213

the SI), showing correlation coefficients lower than 0.5 99.97% of the time. Therefore, the two

214

simulations are independent, uncorrelated and can be used within an OSSE framework. In

215

addition, emissions from all simulated tracers in the modified LES run are set to a constant 1 g s-

216

1

217

temporal variability in the newly developed STE methods, which is consistent with the PSG

218

estimate assumptions. Simulated and estimated emissions are then compared by averaging

219

emissions over each of the six 20 minute periods of simulated data collection and for each well

220

pad. In the same manner as the PSG method, emissions are estimated for each well pad

221

independently, without knowledge of sources or simulated observations upwind of the well pad

222

being assessed.

(selected as a unit that could be scaled) representing no a-priori knowledge of the source

ACS Paragon Plus Environment

10

Page 11 of 29

Environmental Science & Technology

223

The variational inversion method (simply referred to as variational henceforth) involves

224

minimizing a cost-function (J) which can be expressed in the following form for a passive tracer

225

34

:

 =

1 1  ∗  −    ∗  −  +  −      −   1 2 2

226

The right hand side of (1) has two terms, one with the discrepancies between observations (O,

227

simulated observations in this work) and model estimates (H*E, with H the sensitivity

228

matrix/basis functions) to improve the fit to the observations, and the other with the difference

229

between the updated emission estimate (E) and an initial guess (Ep) that contains prior

230

knowledge about the source. These two terms are weighted by the confidence (i.e., inverse

231

covariance matrices, B-1 and R-1) in the observations errors (R) and emissions errors (B). As the

232

variational method requires a first guess on the emissions (Ep), the PSG emissions estimate is

233

used as prior information on the total emissions from the pad which then needs to be distributed

234

into the sources being estimated within the pad. For this we assume the largest source within a

235

pad dominates the emissions (as this is generally the case), assigning it 90% of the total, with the

236

remaining 10% distributed among the other sources. Note that this assumption implies that all

237

equipment within the well pad could potentially be emitting, while this is often not the case (see

238

Fig. S3 in the SI). As the location of the source is known in the PSG method, a similar

239

assumption is made for the variational method. That is, we assume that the location of all

240

possible emitting sources and the largest source within a pad are known. The square root of the

241

diagonal elements of B (standard deviation) represents the errors in the emission priors and are

242

set to 100% of the emission guess as most OTM33A estimates are within a factor of two of the

243

simulated emissions (see Results section). The diagonal of R represents both the observation

ACS Paragon Plus Environment

11

Environmental Science & Technology

Page 12 of 29

244

error (which in the case of the methane sensor is negligible for this study) and the error in the

245

model representation of the observations. As the latter is hard to estimate, we assume a value of

246

10% of the simulated observation mean and perform sensitivity tests around this value (see

247

Results section). This configuration results in a close fit to the simulated observations with little

248

information retained about the initial guess. Off-diagonal elements of the R and B covariance

249

matrices are set to zero. J is minimized numerically using the algorithm described in Zhu et al. 35,

250

bounding the solution to be positive.

251

The Bayesian inference (BI) and stochastic sampling method 32 is based on Bayes’s theorem

252

(i.e., BI) and uses a Markov Chain Monte Carlo (MCMC) procedure to sample the parameter

253

space (i.e., emissions from each source estimated within a pad). Through an iterative procedure,

254

the method provides the probability density function (PDF) of the parameters being estimated

255

given their prior distributions, which are selected as uniform between zero and three times the

256

PSG emissions estimate for the pad. This upper limit is based on the performance of the PSG

257

method (see next Results section) so the simulated emissions would be contained in this interval.

258

The BI-MCMC methods uses a likelihood function to assess the agreement between model and

259

observations (simulated observations in this work). Instead of using the logarithm of the

260

observations and model estimates as in Delle Monache et al. 32, this study uses the likelihood

261

function as the 1st term of the right-hand side of Equation 1 (i.e., no logarithm applied). This

262

function is selected to prioritize the fit to larger concentrations rather than to the tails of the

263

simulated observations (Fig. 2c). The concentrations in the tails are less robust due to the lower

264

number of data points as the tails are associated with less frequent wind directions. Also, the tails

265

could be affected to a larger extent by sources outside the pad which are not being accounted for

266

in the estimation procedure. Preliminary testing with the log-functions showed the superiority of

ACS Paragon Plus Environment

12

Page 13 of 29

Environmental Science & Technology

267

the non-log approach for these reasons. The variational method uses Equation 1 instead of a log-

268

normal cost-function 31 for the same reasons. Although it doesn’t apply to the conditions

269

simulated on this work, the tails would likely need to be filtered for cases with low wind speeds

270

where low probability off-axis trajectories are observed. A range of standard deviations (10-

271

40%) of the observation mean used in the likelihood function (i.e., diagonal of R) was tested

272

with the final value set to 30 %, which provides more satisfactory results. The output of this

273

method is a PDF (see example in Figure S5); thus we chose the distribution mean as a point

274

estimate. By doing so the skill in estimating emissions can be compared to the PSG and

275

variational methods using the same framework (see next section).

276

The variational and BI-MCMC STE methods rely on simulated meteorological fields to link the

277

observed concentrations to the sources being estimated (H in Eq. 1). To match these fields to the

278

simulated observations, a new procedure for computing the sensitivities is adopted. This

279

procedure involves finding the winds from the perturbed simulation (U, V, and W components)

280

that have the best fit to the simulated observations over the time-period that it takes for an air

281

parcel to travel from the source to the observation location. Thus, the sensitivities used

282

correspond to the best representation of the simulated observed winds (and thus the transport) by

283

the modified WRF-LES run. The time-period is chosen as 30 seconds because the simulated

284

measurements are 100 m from the sources and mean wind speeds are 3-3.5 m s-1 at 1.8 m height

285

(lower than the 5 m s-1 geostrophic winds due to surface drag). The best fit of the winds is found

286

by minimizing the difference of the squared errors between the 30 second time series of

287

simulated wind observations and all possible 30 second time series of the modified WRF-LES

288

run at the sampling location (i.e., 7,170 possibilities during the two hours of simulation). An

289

empirical scaling factor of 10 is applied to the vertical winds (W) when performing the fitting so

ACS Paragon Plus Environment

13

Environmental Science & Technology

Page 14 of 29

290

that they are of the same order of magnitude and thus have similar weight as the horizontal winds

291

(U, V). Scaling W by 10 leads to improved results relative to results with no factor or when a

292

lower factor is used (not shown). An example of the outcome of the fitting can be found in

293

Figure S6. This procedure is performed for each simulated observation, i.e., every second over

294

the twenty minutes of simulated observations. The simulated methane observations used in these

295

inversion methods correspond to the methane concentrations binned by wind direction after

296

subtracting the background (red circles in Fig. 2c) and are used to have a fair comparison with

297

the PSG estimates; thus, the time series of sensitivities are aggregated in the same way to

298

represent these simulated observations.

299

RESULTS AND DISCUSSION

300

Performance of the emission estimation methods

301

The performance of the three STE methods is assessed in terms of bias and the spread of the

302

scatter plots, as shown in Figure 3, where estimated and simulated methane emissions are

303

compared. The bias is measured by the average over all times and wells (i.e., ensemble mean) of

304

the estimated-to-simulated emission ratio, which is chosen instead of a metric measuring the

305

difference between estimated and simulated emissions to provide similar weight to all data

306

points. On the other hand, the spread is defined in two ways: by the percentage of estimated-to-

307

simulated emission ratios that fall within factor of 2.0 and 1.5, and by the number of data points

308

where the percent error defined as in Brantley et al. 23 ([estimated emission – simulated

309

emissions]/[estimated emission]) falls within +/- 30%. While all points are included for the

310

emission ratio metric, the percent error metric excludes data points where the goodness of fit

ACS Paragon Plus Environment

14

Page 15 of 29

Environmental Science & Technology

311

coefficient (R2) for the Gaussian fit computed in the PSG estimate (e.g., Fig. 2c) is below 0.8 to

312

assure the validity of the Gaussian assumption of this estimate23.

313

Figure 3a shows how the OTM33A emission estimation results compare to the simulated

314

methane emissions when liquids unloading events are excluded (Table S4 shows actual values).

315

The ensemble mean of the estimated-to-simulated methane emission ratio is 1.08 (1.13 when

316

considering estimates with R2>0.8), showing little bias over all time periods and pads, while 93%

317

and 76% of the data fall between factors of 2.0 and 1.5, respectively. Additionally, 64% of the

318

data have +/- 30% percent error, which is comparable but slightly lower than the 72% found by

319

Brantley et al. 23 when assessing the OTM33A performance using controlled outdoor releases.

320

However, these studies are not directly comparable as Brantley et al. 23 used a range of

321

meteorological conditions and observation distances while in this study large scale

322

meteorological conditions and observations distances are kept constant. Despite these

323

differences, we hypothesize that the lower performance in this study could be attributed in part to

324

the time variation in emissions in contrast to the steady emissions used during controlled

325

releases.

326

Figure 3b shows the scenario where emissions from liquids unloading are included (Table S5

327

shows actual values), showing a large overestimation of emissions on pads #5 and #6 that are

328

directly downwind from pads #1 and #4 containing the liquids unloading events (Fig. 1). As pad

329

#4 is downwind of pad #1, the emission overestimation is also found for pad #4 during periods

330

outside of the liquid unloading event of this pad. The methane background estimated by

331

OTM33A is not increased for these pads when they are under the influence of liquid unloading

332

events from other pads (Table S5); thus this outer influence is assigned to the local emissions

333

producing the overestimation. Although some of these estimates can be screened using R2 fit of

ACS Paragon Plus Environment

15

Environmental Science & Technology

Page 16 of 29

334

the PSG lower than 0.8, some would still be accepted. The OTM33A method includes a mapping

335

survey in its protocol consisting of a vehicle driving around the source before the stationary

336

emission quantification. This survey is performed to potentially identify issues or cases such as

337

the influence of nearby sources. The results here suggest the survey may be useful to eliminate

338

some cases.

339

Figures 3c,d show the performance of the variational and BI-MCMC methods. We find that 98%

340

and 95% of the estimated to simulated emission ratios are within the 1:2/2:1 and 1:1.5/1.5:1

341

lines, respectively, for both methods. Also, after considering estimates with R2 > 0.8 (R2 from the

342

PSG Gaussian fit) we find that 92% and 95% of the data fall within +/- 30% percent error for the

343

variational and BI-MCMC methods, respectively. This shows that the emission estimation

344

methods based on LES simulations contain much lower spread than the OTM33A estimate.

345

Sensitivity analysis was performed by changing the standard deviation of the observation errors

346

around the value originally selected (changed by ±5% and ±10% for the variational and BI-

347

MCMC, respectively) obtaining an increase in the spread with 87% and 92% of data falling

348

within +/-30% percent error for the variational and BI-MCMC methods, respectively. Thus,

349

although results are dependent on these parameters, the skill remains significantly higher than the

350

OTM33A. The variational and BI-MCMC methods also show low overall bias, with ensemble

351

mean of the estimated-to-simulated methane emission ratios of 0.93 and 0.97, respectively.

352

Opposite to the OTM33A results, the overall bias is actually improved for these methods (0.96

353

and 1.0 for variational and BI-MCMC, respectively) when only considering emission estimates

354

with R2 of the Gaussian fit over 0.8.

355

Since the OTM33A method is designed to estimate a single source, a sensitivity test with the

356

LES based methods estimating only the largest source from each well pad was performed. The

ACS Paragon Plus Environment

16

Page 17 of 29

Environmental Science & Technology

357

variational and BI-MCMC methods showed similar results, with 95% and 81% of the emission

358

ratios within a factor of 2.0 and 1.5, respectively, and 72% of the data with percent errors within

359

+/- 30%. These results are similar to what is obtained with OTM33A, showing that the strategy

360

developed for fitting the winds using the modified LES simulation, as opposed to observed

361

winds in the OTM33A method, was effective. Also, these results show that estimating the

362

emission rate of the multiple sources within a well pad is critical for the reduction on the spread

363

of the total emission estimates from a well pad.

364

Although the variational and BI-MCMC methods use the same basis functions in their

365

algorithms (based on LES simulations), their performances differ, with the BI-MCMC showing

366

slightly better bias and spread. Variational methods target solutions that do not deviate largely

367

from the initial guess, while the BI-MCMC method samples the whole parameter space without

368

penalizing solutions that significantly deviate from the initial estimate. Thus, when the initial

369

guesses are far from the simulated emissions (i.e., there is not a single dominating source), the

370

variational method is at a disadvantage and the BI-MCMC method produces better results. This

371

is reflected in the variational model producing a poorer fit to the observed data with a slightly

372

lower fraction of R2 values above 0.9 (52% variational vs 57% BI-MCMC) and 0.8 (81%

373

variational vs 86% BI-MCMC), and a poorer bias especially for pads #3 (mean ratio 0.74

374

variational vs 0.83 BI-MCMC) and #4 (mean ratio 0.78 variational vs 0.96 BI-MCMC), where

375

different equipment can have comparable emissions within the pad (Fig. S1). A sensitivity test

376

was performed for the variational inversion considering an initial guess with equal distribution of

377

emissions between components, resulting in a similar overall performance (0.97 mean emission

378

ratio, and 90% of emissions within +/- 30% percent error). This shows that the prior distribution

379

of emissions within components plays a small role given the large uncertainty assumed in the

ACS Paragon Plus Environment

17

Environmental Science & Technology

Page 18 of 29

380

emission errors. As a general conclusion, the use of the BI-MCMC method over the variational

381

method is recommended when there are large a priori uncertainties in the contribution of each

382

source to the total emissions from a well pad.

383

Emission estimates from all methods are biased low for two 20-minute time periods captured at

384

pads #4 and #6 (Fig. 3a, c, d). Analysis of the simulated emissions (Fig. S1) and simulated

385

methane observations (Fig. S7) time series shows that these periods contain tank-flashing events

386

(large emissions over a short period of time). In the case of pad #4 the wind direction has a

387

persistent easterly component for a few minutes and thus the plume from the tank-flashing

388

emissions is not captured in the simulated concentration time series. On the other hand, the tank-

389

flashing plume from pad #6 is captured very weakly (i.e., peak concentrations are not as large

390

and do not last as long compared to similar cases). The LES-based emissions estimates are

391

within a factor of two of the simulated emissions for the pad #6 case which does not occur for

392

OTM33A estimates (Fig. 3), showing that the LES-based methods show better performance than

393

OTM33A for this time period. We hypothesize this is because the fitting of the winds for each

394

data point is able to replicate instantaneous conditions that lead to the emission peak being

395

weakly represented in the simulated observations. On the other hand, the OTM33A method uses

396

average meteorological conditions over the whole observational time period, and thus is not able

397

to resolve the specific time-dependent conditions. Note that tank-flashing emissions occur for

398

these and other pads for other time periods and are well captured by all methods; therefore, the

399

issues mentioned earlier are not related to the constant emission rate assumption of the OTM33A

400

and LES-based methods. Although these issues have little impact on the overall bias (all STE

401

methods show little overall bias), they do increase the spread and thus large temporal variation in

402

the emission time series is a factor that needs to be considered when evaluating STE methods.

ACS Paragon Plus Environment

18

Page 19 of 29

Environmental Science & Technology

403

The fact that emission spikes and the distribution of sources within a well pad can have an

404

influence on the STE results confirms the value of employing an emission simulator on this

405

study.

406

Applicability of LES-based estimation methods for real-case scenarios

407

The application of the new methods proposed here are based on simulations with an LES model,

408

which requires specific modeling expertise and significant computational resources. For instance,

409

one hour of the simulations performed here requires ~24 hours on 1,296 processors on a high-

410

performance computing platform. However, the computing power needed could be substantially

411

reduced by making the LES domain smaller, containing a small area around the pad being

412

estimated as the presence of nearby pads in the simulations is not required for applying these

413

new STE methods. For instance, we estimate that domain lateral dimension would have to be

414

between 0.5-1 km if a single pad was modeled, which implies a reduction of a factor of 50 to 200

415

in the computational requirements. Also, new developments in LES models, including the

416

portability to a Graphic Processing Unit (GPU) -based system, are making LES simulations

417

faster and computationally more affordable 36. Another requirement of the LES is information on

418

vertical profiles of meteorological variables, which could be obtained from nearby soundings as

419

has been done in previous studies 11. A combination of surface weather measurements performed

420

as part of the STE method (in the case of OTM33A a weather station and an ultrasonic

421

anemometer are required) and reanalysis from numerical weather prediction models to constrain

422

the upper level conditions can be used when the nearest sounding is far away and is not

423

representative of the local weather conditions.

ACS Paragon Plus Environment

19

Environmental Science & Technology

Page 20 of 29

424

The results show a large reduction in spread of the emissions estimates when using multiple

425

sources and their potential locations on the well pads as opposed to assuming only a single

426

source as in the PSG formula used by OTM33A. This requires knowledge of the potential

427

emitting sources on the well pad, which can be obtained by multiple methods including visual

428

inspection of the equipment on a specific site, use of infrared cameras, and use of aerial

429

imagery25. The OTM33A method already requires an evaluation of the source location using

430

these methods, thus little additional work is required to perform this assessment for the LES-

431

based methods. A limitation is that locations of unexpected leaks will not be included, but this is

432

a limitation of most STE methods including the OTM 33A as the knowledge of the source

433

location is needed. However, if the leak happens near-by equipment locations then the

434

implications of unexpected leaks could be minimized as those emissions would be assigned to

435

the nearby equipment.

436

A disadvantage of the LES-based STE methods is that the simulations would have to be

437

performed for every pad configuration and meteorological condition, requiring a large number of

438

simulations unless the pads have the same layout and the meteorological conditions do not vary

439

significantly across samples. A way to deal with this limitation is to build a database of LES

440

simulations. The LES would be run for multiple meteorological conditions that generate different

441

mean wind speeds and stability conditions but keeping the wind direction constant as the

442

measurements can be rotated. Terrain would have to be simplified to flat surfaces in the LES

443

database. An idealized circular pad (to account for a rotating squared pad) would be placed in the

444

middle of the domain (to account for LES spin-up upwind of the pad and leave room for placing

445

observations downwind) and tracers would be tagged to each grid-cell within the pad so they can

446

be activated when the potential sources are identified. The observed wind conditions would be

ACS Paragon Plus Environment

20

Page 21 of 29

Environmental Science & Technology

447

matched by scanning the different simulations by applying the procedure of fitting the winds

448

explained in this work. Such a database would only need to be generated once and could be

449

sampled for multiple well locations and campaigns without much extra cost.

450

The OTM33A method has the advantage that it is straightforward to apply with little

451

computational requirements once the concentration and wind data have been collected on site.

452

Here we show that OTM33A can produce nearly unbiased results, but with substantial spread.

453

The methods based on LES simulations developed here can help reduce the spread but to a

454

substantial additional cost, thus they can be used when the more precise emissions are needed

455

and the extra costs can be afforded.

456

The LES-based STE methods presented here use simulated methane observations extracted from

457

the WRF-LES output using the OTM33A protocols. These methods can also be applied to other

458

measurement protocols that have been developed to estimate emissions using mobile in-situ

459

measurements 7, 9, 10, or even with in-situ or remote sensing data collected from aircraft 37, 38, and

460

can even be applied to other type of point sources where different measurement protocols and

461

emission estimation methods have been developed 39. The results presented in this work

462

represent the first such application of a high-resolution synthetic emission data set to evaluate

463

measurement and STE methods at a single well pad. Also, these methods could be evaluated and

464

compared with each other using a similar framework as the one presented here to better assess

465

their skill and limitations.

ACS Paragon Plus Environment

21

Environmental Science & Technology

466

Page 22 of 29

FIGURES

467 468

Figure 1. a) Terrain used in the inner domain including the location of the seven well pads. Axis

469

shows latitude and longitude in degrees while contours show elevation in meters. The scale is

470

shown on the upper right corner. b) Hypothetical layout of the possible components of a pad

471

including tank with flashing emissions (TF), tank with liquids unloading emissions (TL),

472

chemical injection pump (CI), dehydrator (D), pneumatics (P) and compressor (C). The index

473

indicates to which well the components belong to as each pad contains two wells. The letter O

474

shows the potential location of the simulated observations. The solid black line represents the

475

pad outline (100 m by 100 m) and the solid blue lines the LES grid (10 m resolution).

476

ACS Paragon Plus Environment

22

Page 23 of 29

Environmental Science & Technology

477 478

Figure 2. Methane concentrations (ppm) as a function of wind speed and wind direction

479

corresponding to a) field observations from the OTM 33A test files

25

, and b) simulated

ACS Paragon Plus Environment

23

Environmental Science & Technology

Page 24 of 29

480

observations for pad #1 for the first 20-minute period without liquids unloading emissions. c)

481

Methane concentration binned by wind direction and Gaussian fit used in the PSG estimate for

482

the data in b). a), b) and c) are produced using code found in the OTM33A Appendix F1.

483 484

Figure 3. Scatter plots of total methane emissions from a well pad comparing estimates versus

485

simulated emissions. a) shows OTM33A estimates for the no-liquids unloading emission

486

scenario, and b) shows OTM33A estimates when all emissions are considered. c) shows

487

variational inversion and d) BI-MCMC estimates, both for the no-liquids unloading emission

ACS Paragon Plus Environment

24

Page 25 of 29

Environmental Science & Technology

488

scenario. Note the change in scale on b). Colors represent the R2 of the Gaussian fit for OTM33A

489

(a and b) and R2 of the fit to the observations after variational and BI-MCMC estimates (c and

490

d). Black, magenta, and red show R2 over 0.9, in between 0.9 and 0.8, and lower than 0.8,

491

respectively.

492

Supporting Information.

493

Supplemental figures, tables and text are provided in a single PDF

494 495

AUTHOR INFORMATION

496

Corresponding Author and Present Addresses

497

* †Now at UCLA Department of Atmospheric & Oceanic Sciences, 520 Portola Plaza, 7127

498

Math Sciences Building, MS 156505, Los Angeles, CA 90095. Email: [email protected]

499

ACKNOWLEDGMENTS

500

The National Center for Atmospheric Research is supported by the National Science Foundation.

501

This work was funded by the ExxonMobil Upstream Research Company. We acknowledge the

502

contributions from four anonymous reviewers which helped to improve an early version of this

503

manuscript considerably.

504

REFERENCES

505 506 507 508

1. Brandt, A. R.; Heath, G. A.; Kort, E. A.; O'Sullivan, F.; Pétron, G.; Jordaan, S. M.; Tans, P.; Wilcox, J.; Gopstein, A. M.; Arent, D.; Wofsy, S.; Brown, N. J.; Bradley, R.; Stucky, G. D.; Eardley, D.; Harriss, R., Methane Leaks from North American Natural Gas Systems. Science 2014, 343, (6172), 733-735.

ACS Paragon Plus Environment

25

Environmental Science & Technology

Page 26 of 29

509 510 511

2. Alvarez, R. A.; Pacala, S. W.; Winebrake, J. J.; Chameides, W. L.; Hamburg, S. P., Greater focus needed on methane leakage from natural gas infrastructure. Proceedings of the National Academy of Sciences 2012, 109, (17), 6435-6440.

512 513 514 515 516 517 518

3. Stocker, T. F.; Qin, D.; Plattner, G.-K.; Tignor, M.; Allen, S. K.; Boschung, J.; Nauels, A.; Xia, Y.; Bex, V.; Midgley, P. M. Climate Change 2013. The Physical Science Basis. Working Group I Contribution to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change-Abstract for decision-makers; Groupe d'experts intergouvernemental sur l'evolution du climat/Intergovernmental Panel on Climate Change-IPCC, C/O World Meteorological Organization, 7bis Avenue de la Paix, CP 2300 CH-1211 Geneva 2 (Switzerland): 2013.

519 520 521 522 523

4. Zavala-Araiza, D.; Lyon, D. R.; Alvarez, R. A.; Davis, K. J.; Harriss, R.; Herndon, S. C.; Karion, A.; Kort, E. A.; Lamb, B. K.; Lan, X.; Marchese, A. J.; Pacala, S. W.; Robinson, A. L.; Shepson, P. B.; Sweeney, C.; Talbot, R.; Townsend-Small, A.; Yacovitch, T. I.; Zimmerle, D. J.; Hamburg, S. P., Reconciling divergent estimates of oil and gas methane emissions. Proceedings of the National Academy of Sciences 2015, 112, (51), 15597-15602.

524 525 526

5. Hanna, S. R.; Young, G. S., The need for harmonization of methods for finding locations and magnitudes of air pollution sources using observations of concentrations and wind fields. 2017, 148, (Supplement C), 361-363.

527 528 529

6. Brantley, H. L.; Thoma, E. D.; Squier, W. C.; Guven, B. B.; Lyon, D., Assessment of Methane Emissions from Oil and Gas Production Pads using Mobile Measurements. Environmental Science & Technology 2014, 48, (24), 14508-14515.

530 531 532

7. Rella, C. W.; Tsai, T. R.; Botkin, C. G.; Crosson, E. R.; Steele, D., Measuring Emissions from Oil and Natural Gas Well Pads Using the Mobile Flux Plane Technique. Environmental Science & Technology 2015, 49, (7), 4742-4748.

533 534 535 536 537

8. Lamb, B. K.; McManus, J. B.; Shorter, J. H.; Kolb, C. E.; Mosher, B.; Harriss, R. C.; Allwine, E.; Blaha, D.; Howard, T.; Guenther, A.; Lott, R. A.; Siverson, R.; Westburg, H.; Zimmerman, P., Development of Atmospheric Tracer Methods To Measure Methane Emissions from Natural Gas Facilities and Urban Areas. Environmental Science & Technology 1995, 29, (6), 1468-1479.

538 539 540

9. Day, S.; Dell’Amico, M.; Fry, R.; Tousi, H. J., Field measurements of fugitive emissions from equipment and well casings in Australian coal seam gas production facilities. CSIRO: Australia, 2014.

541 542 543

10. Yacovitch, T. I.; Herndon, S. C.; Pétron, G.; Kofler, J.; Lyon, D.; Zahniser, M. S.; Kolb, C. E., Mobile Laboratory Observations of Methane Emissions in the Barnett Shale Region. Environmental Science & Technology 2015, 49, (13), 7889-7895.

544 545 546

11. Lan, X.; Talbot, R.; Laine, P.; Torres, A., Characterizing Fugitive Methane Emissions in the Barnett Shale Area Using a Mobile Laboratory. Environmental Science & Technology 2015, 49, (13), 8139-8146.

ACS Paragon Plus Environment

26

Page 27 of 29

Environmental Science & Technology

547 548 549

12. Omara, M.; Sullivan, M. R.; Li, X.; Subramanian, R.; Robinson, A. L.; Presto, A. A., Methane emissions from conventional and unconventional natural gas production sites in the Marcellus Shale Basin. 2016, 50, (4), 2099-2107.

550 551 552 553 554

13. Karion, A.; Sweeney, C.; Kort, E. A.; Shepson, P. B.; Brewer, A.; Cambaliza, M.; Conley, S. A.; Davis, K.; Deng, A.; Hardesty, M.; Herndon, S. C.; Lauvaux, T.; Lavoie, T.; Lyon, D.; Newberger, T.; Pétron, G.; Rella, C.; Smith, M.; Wolter, S.; Yacovitch, T. I.; Tans, P., Aircraft-Based Estimate of Total Methane Emissions from the Barnett Shale Region. Environmental Science & Technology 2015, 49, (13), 8124-8131.

555 556 557 558

14. Nathan, B. J.; Golston, L. M.; O’Brien, A. S.; Ross, K.; Harrison, W. A.; Tao, L.; Lary, D. J.; Johnson, D. R.; Covington, A. N.; Clark, N. N.; Zondlo, M. A., Near-Field Characterization of Methane Emission Variability from a Compressor Station Using a Model Aircraft. Environmental Science & Technology 2015, 49, (13), 7896-7903.

559 560 561 562

15. Lavoie, T. N.; Shepson, P. B.; Cambaliza, M. O. L.; Stirm, B. H.; Karion, A.; Sweeney, C.; Yacovitch, T. I.; Herndon, S. C.; Lan, X.; Lyon, D., Aircraft-Based Measurements of Point Source Methane Emissions in the Barnett Shale Basin. Environmental Science & Technology 2015, 49, (13), 7904-7913.

563 564 565

16. Hirst, B.; Jonathan, P.; González del Cueto, F.; Randell, D.; Kosut, O., Locating and quantifying gas emission sources using remotely obtained concentration data. 2013, 74, (Supplement C), 141-158.

566 567

17. Arnold, C. P.; Dey, C. H., Observing-Systems Simulation Experiments: Past, Present, and Future. Bulletin of the American Meteorological Society 1986, 67, (6), 687-695.

568 569

18. Hoffman, R. N.; Atlas, R., Future Observing System Simulation Experiments. Bulletin of the American Meteorological Society 2015, 97, (9), 1601-1616.

570 571 572 573

19. Conley, S.; Faloona, I.; Mehrotra, S.; Suard, M.; Lenschow, D. H.; Sweeney, C.; Herndon, S.; Schwietzke, S.; Pétron, G.; Pifer, J.; Kort, E. A.; Schnell, R., Application of Gauss's theorem to quantify localized surface emissions from airborne measurements of wind and trace gases. Atmos. Meas. Tech. 2017, 10, (9), 3345-3358.

574 575 576 577

20. Viatte, C.; Lauvaux, T.; Hedelius, J. K.; Parker, H.; Chen, J.; Jones, T.; Franklin, J. E.; Deng, A. J.; Gaudet, B.; Verhulst, K.; Duren, R.; Wunch, D.; Roehl, C.; Dubey, M. K.; Wofsy, S.; Wennberg, P. O., Methane emissions from dairies in the Los Angeles Basin. Atmos. Chem. Phys. 2017, 17, (12), 7509-7528.

578 579 580 581

21. Caulton, D. R.; Li, Q.; Bou-Zeid, E.; Lu, J.; Lane, H. M.; Fitts, J. P.; Buchholz, B.; Golston, L. M.; Guo, X.; McSpiritt, J.; Pan, D.; Wendt, L.; Zondlo, M. A., Improving Mobile Platform Gaussian-Derived Emission Estimates Using Hierarchical Sampling and Large Eddy Simulation. 2017, 2017, 1-39.

582 583 584

22. Foster-Wittig, T. A.; Thoma, E. D.; Albertson, J. D., Estimation of point source fugitive emission rates from a single sensor time series: A conditionally-sampled Gaussian plume reconstruction. Atmospheric Environment 2015, 115, (Supplement C), 101-109.

ACS Paragon Plus Environment

27

Environmental Science & Technology

Page 28 of 29

585 586 587 588

23. Brantley, H. L.; Thoma, E. D.; Eisele, A. P., Assessment of volatile organic compound and hazardous air pollutant emissions from oil and natural gas well pads using mobile remote and on-site direct measurements. Journal of the Air & Waste Management Association 2015, 65, (9), 1072-1082.

589 590 591 592

24. Robertson, A. M.; Edie, R.; Snare, D.; Soltis, J.; Field, R. A.; Burkhart, M. D.; Bell, C. S.; Zimmerle, D.; Murphy, S. M., Variation in Methane Emission Rates from Well Pads in Four Oil and Gas Basins with Contrasting Production Volumes and Compositions. Environmental Science & Technology 2017, 51, (15), 8832-8840.

593 594 595

25. EPA, Other Test Method (OTM) 33 and 33A Geospatial Measurement of Air PollutionRemote Emissions Quantification-Direct Assessment (GMAP-REQ-DA). https://www.epa.gov/emc/emc-other-test-methods, 2016.

596 597 598

26. Moeng, C.-H.; Dudhia, J.; Klemp, J.; Sullivan, P., Examining Two-Way Grid Nesting for Large Eddy Simulation of the PBL Using the WRF Model. Monthly Weather Review 2007, 135, (6), 2295-2311.

599 600 601

27. Nunalee, C. G.; Kosović, B.; Bieringer, P. E., Eulerian dispersion modeling with WRFLES of plume impingement in neutrally and stably stratified turbulent boundary layers. Atmospheric Environment 2014, 99, 571-581.

602

28.

603 604

29. Ruiz-Columbie, A., Atmospheric boundary layer evening transitions over West Texas. 2008.

605 606 607

30. Smith, N.; Blewitt, D.; Hebert, L. B. In Impact of routine episodic emissions on the expected frequency distribution of emissions from oil and gas production sources, AGU Fall Meeting, San Francisco, USA, 2015, 2015; San Francisco, USA, 2015.

608 609 610 611 612 613

31. Saide, P. E.; Peterson, D.; da Silva, A.; Anderson, B.; Ziemba, L. D.; Diskin, G.; Sachse, G.; Hair, J.; Butler, C.; Fenn, M.; Jimenez, J. L.; Campuzano-Jost, P.; Perring, A. E.; Schwarz, J. P.; Markovic, M. Z.; Russell, P.; Redemann, J.; Shinozuka, Y.; Streets, D. G.; Yan, F.; Dibb, J.; Yokelson, R.; Toon, O. B.; Hyer, E.; Carmichael, G. R., Revealing important nocturnal and dayto-day variations in fire smoke emissions through a multiplatform inversion. Geophysical research letters 2015, 2015GL063737.

614 615 616 617 618

32. Delle Monache, L.; Lundquist, J. K.; Kosović, B.; Johannesson, G.; Dyer, K. M.; Aines, R. D.; Chow, F. K.; Belles, R. D.; Hanley, W. G.; Larsen, S. C.; Loosmore, G. A.; Nitao, J. J.; Sugiyama, G. A.; Vogt, P. J., Bayesian Inference and Markov Chain Monte Carlo Sampling to Reconstruct a Contaminant Source on a Continental Scale. Journal of Applied Meteorology and Climatology 2008, 47, (10), 2600-2613.

619 620 621

33. Chow, F. K.; Kosović, B.; Chan, S., Source Inversion for Contaminant Plume Dispersion in Urban Environments Using Building-Resolving Simulations. Journal of Applied Meteorology and Climatology 2008, 47, (6), 1553-1572.

Oke, T., Boundary layer climates. 2nd. Methuen, 289p 1987.

ACS Paragon Plus Environment

28

Page 29 of 29

Environmental Science & Technology

622 623 624

34. Saide, P. E.; Bocquet, M.; Osses, A.; Gallardo, L., Constraining surface emissions of air pollutants using inverse modelling: method intercomparison and a new two-step two-scale regularization approach. Tellus B 2011, 63, (3), 360-370.

625 626 627

35. Zhu, C.; Byrd, R. H.; Lu, P.; Nocedal, J., Algorithm 778: L-BFGS-B: Fortran subroutines for large-scale bound-constrained optimization. ACM Transactions on Mathematical Software (TOMS) 1997, 23, (4), 550-560.

628 629 630

36. Schalkwijk, J.; Jonker, H. J. J.; Siebesma, A. P.; Meijgaard, E. V., Weather Forecasting Using GPU-Based Large-Eddy Simulations. Bulletin of the American Meteorological Society 2015, 96, (5), 715-723.

631 632 633 634

37. Caulton, D. R.; Shepson, P. B.; Santoro, R. L.; Sparks, J. P.; Howarth, R. W.; Ingraffea, A. R.; Cambaliza, M. O. L.; Sweeney, C.; Karion, A.; Davis, K. J.; Stirm, B. H.; Montzka, S. A.; Miller, B. R., Toward a better understanding and quantification of methane emissions from shale gas development. Proceedings of the National Academy of Sciences 2014, 111, (17), 6237-6242.

635 636 637 638 639

38. Frankenberg, C.; Thorpe, A. K.; Thompson, D. R.; Hulley, G.; Kort, E. A.; Vance, N.; Borchardt, J.; Krings, T.; Gerilowski, K.; Sweeney, C.; Conley, S.; Bue, B. D.; Aubrey, A. D.; Hook, S.; Green, R. O., Airborne methane remote measurements reveal heavy-tail flux distribution in Four Corners region. Proceedings of the National Academy of Sciences 2016, 113, (35), 9734-9739.

640 641 642

39. Weitkamp, E. A.; Lipsky, E. M.; Pancras, P. J.; Ondov, J. M.; Polidori, A.; Turpin, B. J.; Robinson, A. L., Fine particle emission profile for a large coke production facility based on highly time-resolved fence line measurements. 2005, 39, (36), 6719-6733.

643

ACS Paragon Plus Environment

29