Advances in natural gas extraction technology have led to increased activity
in the production and transport sectors in the United States and, as a
consequence, an increased need for reliable monitoring of methane leaks to
the atmosphere. We present a statistical methodology in combination with an
observing system for the detection and attribution of fugitive emissions of
methane from distributed potential source location landscapes such as natural
gas production sites. We measure long (> 500 m), integrated open-path concentrations of atmospheric methane using a dual frequency comb
spectrometer and combine measurements with an atmospheric transport model to
infer leak locations and strengths using a novel statistical method, the
non-zero minimum bootstrap (NZMB). The new statistical method allows us to
determine whether the empirical distribution of possible source strengths for
a given location excludes zero. Using this information, we identify leaking
source locations (i.e., natural gas wells) through rejection of the null
hypothesis that the source is not leaking. The method is tested with a series
of synthetic data inversions with varying measurement density and varying
levels of model–data mismatch. It is also tested with field observations of
(1) a non-leaking source location and (2) a source location where a controlled
emission of 3.1
The combustion of natural gas in high-efficiency power cycles is cleaner and
produces less climate-warming carbon dioxide gas than the combustion of coal
(Environmental Protection Agency, 2015), which has
led to interest in natural gas as a cleaner alternative to coal for energy
generation. Advances in natural gas extraction technology have led to a
35 % increase in total natural gas production between 2005 and 2013 in the
United States (U.S. Energy Information Administration,
2015). Production is expected to increase by 45 % above 2013 levels by the
year 2040 (U.S. Energy Information Administration, 2015).
A caveat to the promise of natural gas as a lower climate impact energy
source, however, is that leaks of methane during extraction and delivery can
result in climate warming. Methane gas has high global warming potential
(GWP): much higher, for example, than carbon dioxide (CH
The current industry practice for leak detection and repair (LDAR) is to perform infrequent (annual or less for most sites) “spot” checks for leaks, for example by visual inspection with an optical gas imaging (OGI) camera. However, recent work has shown that methane concentrations measured by OGI cameras can be drastically underestimated when conditions are not ideal, for example under conditions of lower temperature values or higher wind speeds, or when viewing distances are greater than 50 m (Ravikumar et al., 2016). Furthermore, spot check monitoring is inadequate for detection of leaks, given strong evidence for intermittency of leaks (Allen et al., 2013, 2015a; Mitchell et al., 2015; Subramanian et al., 2015). It has been observed that a small number of facilities leaking at very high rates – so-called “super-emitters” (Brandt et al., 2014; Frankenberg et al., 2016; Rella et al., 2015; Zavala-Araiza et al., 2015b) – can account for a majority of total emissions (Allen et al., 2013, 2015a, b; Brandt et al., 2014). These characteristics underscore the importance of continuous monitoring for leaks over large areas. Field campaigns with sophisticated atmospheric sampling techniques provide valuable snapshots of the state of natural gas development facility leaks (e.g., Brantley et al., 2014; Karion et al., 2013), but it would be too costly to employ such measurement strategies for long-term continuous monitoring of most natural gas sector facilities.
We present and test an atmospheric measurement system coupled with a statistical inversion approach for detecting and quantifying emissions of methane. The statistical approach is focused on limiting the occurrence of false-positive leak detection. The measurement system used to test the statistical approach is composed of a long-range open-path laser situated in the center of a field of well sites and a series of retroreflectors around the perimeter of the field to direct light back to a detector co-located with the laser. The concentration of trace gases along the open beam path (defined as the path between the spectrometer–detector system and a retroreflector) is determined from the species-specific absorption of light (Dobler et al., 2015; Flesch et al., 2004; Groth et al., 2015; Hashmonay et al., 1999; Levine et al., 2016). Many open-path absorption methods for determining species concentration have been demonstrated (Akagi et al., 2011; Dobler et al., 2015; Flesch et al., 2004; Jones et al., 2011; Nikodem et al., 2015; Wagner and Plusquellic, 2016; Wu et al., 2014). Here we use a dual frequency comb spectrometer (DCS): a unique broadband, high-resolution spectrometer that offers very high stability (low drift) and measurement reproducibility of the trace gas measurement so that concentrations can be compared across different conditions and times (Coburn et al., 2018). It was recently demonstrated that two separate dual frequency comb spectrometers stationed side by side and measuring the same 1 km outdoor path showed methane concentration agreement to 0.35 % over a 2-week period under ambient variations in temperature, pressure, and stability (Waxman et al., 2017). In principle, the range of conditions under which two separate dual frequency comb spectrometers should be comparable is much wider than ambient conditions, because the concentration retrieval is largely dependent on the quality of absorption models (which are well-defined under most conditions experienced at Earth's surface). Previous work also demonstrates that this method of atmospheric trace gas measurement does not require regular or traditional calibration (Coburn et al., n.d.; Rieker et al., 2014; Truong et al., 2016; Waxman et al., 2017). Laboratory and initial field measurements made with the dual frequency comb spectrometer indicate extremely high measurement precision (3 ppb or lower) over long (1 km one-way, or 2 km round trip) path lengths (Coburn et al., n.d.; Rieker et al., 2014; Truong et al., 2016; Waxman et al., 2017). The combination of low uncertainty and high stability enables new opportunities for detection and sizing of even very small emissions of methane (Coburn et al., n.d.). Furthermore, the demonstration of sensitive methane measurements over kilometer-scale open paths allows for monitoring methane concentrations over large areas such as natural gas production, processing, and distribution sites. While frequency comb measurements have previously been made in laboratory settings, the recent work of Coburn et al. (n.d.) and the new work shown here demonstrate the viability of dual frequency comb spectroscopy in real-world conditions.
We use the dual frequency comb measurements in a series of synthetic data
and field data tests to demonstrate the utility of the observing system and
a novel statistical method for accurately locating one or more point sources
of methane within a large area (4
Synthetic data tests are performed that assess the effects of increasing
measurement density (4, 8, 16, 32, and 64 beams) and the effects of
increasing model–data mismatch (that is, combined uncertainty in the ability
to simulate observations arising from measurement, transport, and other
sources). Field tests with atmospheric observation data are performed in a
3 km
We define leak identification success as maximizing the incidences of leaks found, with a minimal occurrence of false-positive source identification, enabling quick response to leaks and avoiding costly mobilization of repair teams due to false-positive leak identification. The ability to correctly ascertain the absence of a leak is therefore of equal importance to the ability to find leaks for regulatory compliance applications of this method. With the above tests, we therefore seek to determine (1) whether methane point source emissions can be detected and sized under conditions of observational uncertainty (model–data mismatch) and background variation; (2) whether the absence of a leak can be ascertained in an outdoor field setting; (3) whether the NZMB method allows for leaks to be positively identified under scenarios of greater simulated model–data mismatch uncertainty, compared with the non-bootstrap method; and (4) whether a higher number of observations increases likelihood that the NZMB and non-bootstrap methods can positively identify leaks. The success of the synthetic and field data tests demonstrates the potential of this observing system for continuous monitoring applications, such as for natural gas facilities, and for providing emission source locations and their approximate strengths. The experiments here also demonstrate the potential for this technology to be used for other source estimation and monitoring applications, for example carbon sequestration.
In both the synthetic and real data tests, atmospheric transport is simulated using a Gaussian plume model, using Pasquill–Gifford parameterization of plume dispersion in the lateral and vertical directions (Green et al., 1980; Griffiths, 1994; Hanna et al., 1982). Micrometeorology in the boundary layer is a non-trivial source of uncertainty for characterization of atmospheric flow, and the Gaussian plume model represents a simplified representation of atmospheric transport and dispersion. It is used to characterize the mean state (or steady state) of source–receptor relationships with a point source, as long as the transport time from source to receptor is comparable to the data averaging time (Gifford, 1976; Hirst et al., 2004). More sophisticated plume (e.g., AERMOD) or stochastic Lagrangian dispersion models (e.g., WindTrax) and stability parameterizations would be expected to provide more robust representations of the wind shear and inhomogeneities in turbulence in the atmospheric surface layer (Flesch et al., 1995; Perry et al., 1994; Wilson and Sawford, 1996). We select the simplified and low-computational-cost plume model for assessment of the NZMB method as a baseline test rather than implementing more advanced representations of transport. Future campaigns aimed at quantification of true emissions will benefit from an assessment of the drawbacks inherent in Gaussian plume model characterization of atmospheric transport or use of a more sophisticated model, particularly for measurements made at short range.
For the synthetic data tests, the choice of transport model is largely trivial, given that the transport is considered “perfect”. Field data are collected with a constant methane source to the atmosphere and a measurement averaging time that is comparable to the source-to-receptor travel time, such that the Gaussian plume model is a simplified but appropriate choice of transport model (Gifford, 1976; Hirst et al., 2004). Because the purpose of this study is to confirm or reject the basic methodology and not to investigate the impacts of micrometeorological representation on flux estimation, we find the plume model to be sufficient as a baseline test (see Sect. 6).
Neglecting influence of background methane concentrations, Eq. (1) shows the
relationship between fluxes and atmospheric concentrations
(e.g., Leuning et al., 2008):
Dual frequency comb spectrometer measurements are made by transmitting light
from the spectrometer through open air at a discrete set of wavelengths
where methane absorbs light. The light is transmitted in the direction of a
retroreflector, which can be placed 1
We use the NNLS algorithm in Fortran-90 to
solve for a flux rate (that is, the emission rate from each potential source
location), given atmospheric observations (synthetic or real) and
atmospheric transport influence functions (Lawson and
Hanson, 1995). This algorithm iteratively solves for the best-fit
The non-zero minimum bootstrap analysis, or NZMB, is a statistical test of
the null hypothesis (Hypothesis
For each of
The method for employing the bootstrap analysis is as follows. We first
solve for surface-to-atmosphere fluxes of CH
The next step in the NZMB method is, for each observation,
For the field data, we apply a moving block bootstrap (Künsch, 1989) because residuals of observations made nearer together in time are more likely to be co-representative, whereas residuals of observations made further apart in time are likely to be less representative due to changes in wind conditions and atmospheric stability. We calculate the autocorrelation in time of the residuals resulting from a single non-negative least-squares fit and use for the moving block window length a value 2 times the lag time at which the autocorrelation falls below the 95 % confidence level. As there is no time dimension in the synthetic data case, we do not apply the moving block bootstrap to those cases.
Next, we use NNLS to solve for
After having identified which source locations are non-zero sources to the atmosphere (leaking), the mean leak strength is estimated as the mean of the 1000 bootstrap solutions for that source location. Uncertainty in the strength of the true leak is calculated as the standard deviation of the 1000 bootstrap solutions at the true leak location.
This method requires little additional computational cost over the non-bootstrap NNLS approach, because additional runs of the transport model are not required, only additional NNLS fits using resampling of the observations. The NZMB approach has the benefit of reducing false-positive solutions while also gathering information regarding the parameters of the assumed Gaussian distribution.
To prepare synthetic data testing of the NZMB method, we randomly distribute
20 possible leak source locations within a theoretical 2 km
Synthetic test observation area: 2 km
The meteorological data used for synthetic data tests include many wind
directions and a variety of wind speeds during the sampling of each beam in
the domain, representing an ideal scenario for the generation of as many
independent measurements of the leak strength as possible. Leak strengths
are constant through time, such that the time dimension of the meteorology
does not need to be considered. This approach assumes that enough time has
passed for all meteorological conditions to have occurred during the
sampling of each beam, a condition that eliminates complications in
comparing synthetic cases with different beam orientations. The idealized
meteorological field applies 216 unique wind conditions to all beams: three
wind speeds (2, 3, and 6 m s
The “synthetic” atmospheric measurements are simulated based on the dual
frequency comb spectrometer observing system described in Sect. 2.2. The
spectrometer is located in the center of the domain, at
Map view of synthetic tests, with 20 source locations shown as
black dots and 16 beams shown as gray lines that extend from the
spectrometer (circle at
The vector of “true” atmospheric methane concentrations,
The dimensions of
Model–data mismatch is the difference between the true atmospheric CH
A range of model–data mismatch values are tested with the expectation that
both the NZMB and non-bootstrap models will be more likely to locate and
source leaks when lower model–data mismatch is added to the data. To
simulate different possible magnitudes of model–data mismatch, the simulated
true atmospheric concentrations,
The first measurements from a field-deployed dual frequency comb
spectrometer are from the NOAA/ESRL Table Mountain Test Facility, 10 km
north of Boulder, Colorado (Fig. 3; Coburn et al.,
n.d.). The spectrometer is located near the center of a large (
Map view of observation test site at Table Mountain, Colorado (upper left inset shows geographic location of test site), with two source locations (location 1, in red, between beams 1 and 2; location 2, in green, between beams 2 and 3) and three beams shown as white lines that extend from the spectrometer (blue square) to retroreflectors (white triangles, labeled 1–3).
For the field experiments at Table Mountain, a cylinder of compressed
methane gas is placed roughly 528 m away from the spectrometer (Fig. 3) with
the gas outlet 1 m a.g.l. The methane cylinder is outfitted
with a regulator and an Alicat mass flow controller (MC-20SLPM-D). The flow
controller is set to release methane in a controlled flow of 3.1
The field tests are arranged so as to approximate the synthetic tests as
closely as possible: to emulate the “perfect” background condition of the
synthetic tests, the background methane concentration for each source
location is measured directly by an upwind beam
(Crenna et al., 2008; Flesch et al., 2009). Because
the background is assumed to be unique for each source location, each
inversion includes only that source location in its solution for fluxes.
That is, one inversion is performed for source location 1, and a separate
inversion is performed for fluxes at source location 2. The dimensions of
Three corner-cube retroreflectors are located near source locations 1 and 2 at Table Mountain (see Fig. 3). At their nearest points, the lateral distances between beams 1 and 2 and source location 1 are 11 and 6 m, respectively. The minimum lateral distances between leak location 2 and beams 2 and 3 are 12 and 8 m, respectively. The horizontal distance from the spectrometer to each retroreflector is 584, 585, and 588 m, respectively, for retroreflectors 1, 2, and 3. All retroreflectors are positioned 1 m a.g.l.
Wind speed and wind direction are measured directly with a 3D Sonic
Anemometer (RM Young 81000 Ultrasonic 3D Anemometer with
manufacturer-specified accuracy of
We test the bootstrap methodology using measurements taken over the course
of 1 day in January 2017. We test the ability of the bootstrap approach to
both disprove the null hypothesis (i.e., to correctly ascertain the presence
of a non-zero methane emission) and to prove the null hypothesis (i.e., to
correctly ascertain the absence of a leak) by gathering measurements along
beam paths that bound (1) source location 1, where methane is released in a
controlled flow rate of 3.1
In the field tests, the dimensions of
To most closely approximate the synthetic data testing framework in the
field environment, we directly sample background CH
We calculate solutions for
Summary of synthetic data test results. Top 5 rows show results of non-bootstrap inversions and bottom 5 rows show results of NZMB inversions for the 4, 8, 16, 32, and 64 beam cases. Columns indicate results for different values of model–data mismatch added as noise to the synthetic measurements. Color coding of cells indicates summary of model success, as detailed by the legend.
The bottom half of Fig. 4 shows the results of the same tests, using instead the NZMB method for locating leaks. The results show that success in leak detection is much higher using NZMB compared with the non-bootstrap tests. Indeed, none of the NZMB tests result in the occurrence of a false-positive leak location, and only tests with low numbers of beams relative to the number of source locations (four- and eight-beam cases) fail to find both of the true leaks. The four-beam case results in positive identification of both leaks up to a model–data mismatch threshold of 2 ppb, above which one true leak is found. One leak is consistently found up to a threshold of 5 ppb, and above 5 ppb model–data mismatch no true leaks are identified (but no false positives are generated either). The eight-beam case results in accurate location of both true leaks up to a model–data mismatch threshold of 3.5 ppb, above which 1 true leak is found (with no false positives). One leak is consistently found up to the maximum testing point of 10 ppb. In order to reliably locate both true leaks with no false-positive results under all model–data mismatch scenarios, 16 or more beams are needed for the set of cases that are tested here. Alternate configurations of “true” leaks at well sites other than 6 and 9 are not tested; however, given that meteorological conditions are simulated equally from all directions, we would not expect a different set of results from a different set of “true” leaks.
Top left panel shows well site numbers (
Top left panel shows well site numbers (
A subset of the results for the eight-beam NNLS without bootstrap and the NNLS with NZMB cases are shown in Figs. 5 and 6 (for conciseness; all results are shown in the Supplement). It is evident from Fig. 5 that, even with very low model–data mismatch noise (0.5 ppb), the non-bootstrap model results in well sites other than the two true leak locations being erroneously identified as sources of methane. It is evident from Fig. 5 that, as model–data mismatch increases, the strength of incident false-positive results also increases. By contrast, no false-positive leaks are identified in the NZMB case shown in Fig. 6, at any level of model–data mismatch noise. Above a model–data mismatch threshold of 4 ppb, only one of two true leaks is found in the eight-beam case using NZMB. As Fig. 4 shows, 16 or more beams are necessary to consistently find both true leaks at higher thresholds of model–data mismatch uncertainty using the NZMB method, given the hub-and-spoke beam placement scheme tested here. More complex placement of beams (for example placing beams closer to known well sites) would likely result in even better ability to locate leaks with fewer beams.
Synthetic data tests of the new bootstrap methodology presented here show high success in leak location, with zero incidence of false-positive leak detections. Figure 6 shows the maximum and minimum values of 1000 bootstrap operations for each model–data mismatch test case for the eight-beam configuration. At low levels of model–data mismatch uncertainty (0.1–0.5 ppb), the maximum and minimum solutions bound a small range that is close to the true leak strength. As higher levels of model–data mismatch noise are added to observations, the maximum and minimum values diverge. However, even as the maximum and minimum solutions diverge, most cases include the true leak strength within the maximum and minimum bounds.
NZMB solutions for leak strength of true leaks, given 2 ppb model–data mismatch uncertainty, for each beam configuration.
Using the NZMB method, all beam cases (even the four-beam case) correctly identify that both well sites 6 and 19 are emitting methane when model–data mismatch is 2 ppb or lower (Fig. 4). At that level of model–data mismatch, higher numbers of beams and observations tend to lead to lower standard deviation around the mean estimated leak strength and a more accurate estimate of true leak strength (Table 1). An exception is at well site 19, where the eight-beam case did not perform as well as the four-beam case. It may be that both cases were inadequate for accurately sizing leaks, and that 16 beams are necessary in a dense field of wells such as is tested here. The failure of the eight-beam case to accurately predict the leak rate at well site 19 is also evident from histograms of bootstrap operations, shown for each beam case with model–data mismatch of 2 ppb in Fig. 7.
Histograms of source strength, with mean
Histograms of the results for the 16, 32, and 64 beam cases with 10 ppb model–data mismatch are shown in Fig. 8. It is clear from Fig. 8 that, even with very high model–data mismatch uncertainty, simple hub-and-spoke configurations of between 16 and 64 beams are able to locate and estimate leak flow rates to within reasonable bounds of uncertainty.
Atmospheric observations were made over the course of 1 day on 26 January 2017 at the Table Mountain site. A set of three retroreflectors
created
long-range open-path beams of
Histograms of source strength, with mean
On 26 January 2017, measurements are made throughout the day, including during a 6.5 h controlled release of methane at source location 1. At adjacent source location 2, no methane release is emitted. A series of three retroreflectors is oriented such that each source region is monitored independently from the other; one beam on either side of each source location serves as a “background” measurement. We examine the results of two separate inversion tests: (1) a day-long set of observations of source location 1 (with the controlled release) that is situated between retroreflectors 1 and 2 and (2) a day-long set of observations of non-leaking source location 2 that is situated between retroreflectors 2 and 3. These tests are performed simultaneously, such that contamination from source location 1 could result in background contamination for monitoring of source location 2.
Allan deviation plot showing changes in measurement precision with averaging time from field data collected at Table Mountain on 26 January 2017.
On 26 January 2017, mean wind speeds are 2.1 m s
At source location 1 (Fig. 10a), during the period when the
controlled release is on (non-zero flow), the downwind retroreflector (Retro
1) shows a clear enhancement above the concentration measured on the upwind
retroreflector (Retro 2), except during the middle of the day when the winds
shift briefly to the south (Fig. 10c). The mean of all CH
Line-integrated atmospheric CH
At source location 2, no leak is released during the period of study, and
throughout the course of the day, both retroreflectors 2 and 3 measure
similar changes in atmospheric CH
The beams stationed upwind of each source location provide estimates of the
background CH
We examine measurements at source location 2, where no leak is present, in order to estimate model–data mismatch in the field, for comparison with the model–data mismatch values applied in the synthetic data tests. By examining the difference between measurements made on different retroreflectors (retroreflectors 2 and 3) at similar points in time (within 5 min), we obtain an approximation of the combined contributions to model–data mismatch arising from measurement uncertainty, representation uncertainty, background construction (the method of background estimation), and background sampling (the method of sampling background concentrations). We find a standard deviation of 5 ppb. This value differs from the standard deviation of the enhancement for the entire time series (reported above in Sect. 4.3) because it compares differences in upwind and downwind concentrations measured at approximately the same time. For estimation of total model–data mismatch, we add (in quadrature) an estimate of the transport uncertainty that includes uncertainties in measurement of wind speed and wind direction, atmospheric stability parameterization, and placement of the sonic anemometer relative to the leak location (see Supplement for detail). Transport uncertainty estimation is for a plume that interacts with any location along the beam and therefore requires knowledge of the mean distance between the leak point and each segment of the beam. The estimated transport uncertainty, calculated in this way, is 0.8 ppb. If, for example, the wind direction is perfectly perpendicular to the beam for the entirety of the measurement period (which does not occur on 26 January 2017), then the leak-to-beam distance used in the calculation should collapse to the minimum lateral distance between the leak and the beam. Using that value instead, transport uncertainty is 12.2 ppb. The overall value of model–data mismatch (reflecting combined measurement, background, and transport uncertainty), estimated in this way, is therefore 5.1 ppb with a maximum range of 13.2 ppb, which suggests that the range of model–data mismatch values tested in the synthetic data experiments are appropriate. The Allan deviation in Fig. 9 shows a similar level of measurement uncertainty, suggesting that most of the uncertainty observed in our record is captured in this estimate of model–data mismatch, which includes effects of atmospheric variability. Precision could be improved by averaging data over a shorter time span (70 s), but those gains would be minimal (Fig. 9).
Controlled methane release flow rates and 1 standard deviation for each field experiment, including local time that leak was turned on and off.
n/a: not applicable.
Both the non-bootstrap and the NZMB approaches accurately predict the
presence of methane emissions at source location 1 (Table 2). The average
bootstrapped flux value is within 2
Histogram of NZMB estimated source strength at source location 1,
with dashed line showing the bootstrap mean and thin dotted lines showing
The results of this study demonstrate success of the new observing system in finding one or more leaks of methane in a field of wells, using synthetic and field data for confirmation. The methods presented here for locating and sizing leaks of methane in a field of natural gas production facilities succeeds not only in identifying the location of a leak, but it also does so with no incidences of “false-positive” leak detection in either the synthetic or field data tests.
The results of the synthetic data tests demonstrate how the observing system
tested in the field for a single source location can be expanded for
simultaneous monitoring of many source locations. We find that synthetic
tests performed without the NZMB methodology failed to identify the presence
of leaks as reliably as synthetic tests performed with the NZMB method,
demonstrating the improved robustness of this new statistical method for
leak detection. In the non-bootstrap tests, all synthetic data cases
resulted in false-positive solutions (Fig. 4). By contrast, the NZMB method
succeeds in correctly identifying two leaks of strength 3.0
Histogram of NZMB estimated source strength at source location 2,
with dashed line showing the bootstrap mean and thin dotted lines showing
Field data testing of the NZMB method corroborates the synthetic data
findings: that the new atmospheric observing system presented here results
in high accuracy of leak detection without false-positive results. The
ability of the dual frequency comb spectrometer to identify a very small
leak (3.1
An important caveat to the methodology presented here is the short length of
the measurement averaging time, which presents a mismatch with the ideal
application of most dispersion models (for which practice is generally to
use averaging times longer than 2 min). This requirement in our
methodology is due to two factors: the first is that rapid scanning for
potential leaks is an important feature in areas where many sites must be
monitored and leaks can be intermittent. The second factor is that
background methane concentrations can vary with high frequency (order
minutes) in proximity to areas of oil and natural gas production
(Dlugokencky et al., 1995). We attempt to mitigate
uncertainties arising from using dispersion parameters developed for longer
timescale modeling over a 2 min period in several ways. First,
The focus of this study is to show the powerful potential of the combination
of a new statistical method with dual frequency comb spectroscopy for the
location and sizing of point source emissions. The synthetic and field tests
presented here rely on near-perfect (in the synthetic data tests) or
well-constrained (in the field data tests) background concentration
estimation. Future studies are needed to address the potential complications
of more complex background conditions and meteorological conditions under
which it is not possible to obtain sequential “upwind” and “downwind”
samples. Similarly, the tests here rely on the assumption of constant leak
rates, which may not be a realistic assumption that can be made for methane
emissions from oil and gas operations. Future work to address these
complexities will be necessary. Future studies are also needed to examine
the gains that can be made from optimization of beam configurations for
improved leak detection given variable wind and background conditions. In
particular, previous work has shown the critical impact that sensor
placement can have on the conditioning of the source–receptor relationship
matrix (
A notable aspect of the micrometeorological modeling used here to demonstrate the NZMB methodology is the simple representation of atmospheric transport (the Gaussian plume model). The choice to use a simple model that is familiar to the broader scientific community is intentional, but its use belies the complex nature of turbulent mixing and dispersion in the atmospheric surface layer. What is gained in simplicity and in providing a baseline for the most basic performance of the methodology in a field setting may come at the cost of recommending a model that may not ultimately be well-suited for such an endeavor. The Gaussian plume model neglects important aspects of atmospheric mixing such as wind shear and the height dependence of eddy diffusivity, and better models exist for simulation of atmospheric flow at this scale. It is assumed that more sophisticated models of atmospheric dispersion could, therefore, lead to better flux estimation. We suggest that future applications in field settings of the methodology presented here consider their use. Importantly, despite its drawbacks, the Gaussian plume model proves sufficient in the tests here for the accurate identification (and, importantly, avoidance of misidentification) of controlled, field-based methane leaks. Future studies of the best transport model for the application of DCS measurements and the NZMB method for leak detection is warranted.
The initial work presented here demonstrates the promising potential of dual frequency comb spectroscopy for detection of leaks in the natural gas supply chain and the valuable gains that can be provided by using the NZMB method over the NNLS fitting technique alone.
Data are accessible at the following URL:
SG, KP, and CA implemented the statistical NZMB technique. CA, SC, RW, CS, KP, SG, and GR designed the experiments and SC, RW, and CA carried them out. AK provided expert guidance and experimental design input. CA prepared the manuscript with contributions from all co-authors.
The authors declare that they have no conflict of interest.
The information, data, or work presented herein was funded in part by the Advanced Research Projects Agency-Energy (ARPA-E), US Department of Energy, under Award Number DE-AR0000539. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. Edited by: Dietrich G. Feist Reviewed by: two anonymous referees