Methane emissions from oil and gas production sites in Alberta, Canada

Mitigating methane (CH₄) emissions from the global oil and gas system represents a major opportunity to reduce the near-term climatic effects of this powerful greenhouse gas (GHG) (Bousquet et al., 2006; Shoemaker et al., 2013; Saunois et al., 2016; Höglund-Isaksson, 2017). Several countries have announced pledges to reduce CH₄ emissions from this system (Mexico: Presidencia de la Republica, 2016; Climate & Clean Air Coalition, 2016; Environment and Climate Change Canada, 2017a), with Canada and the US taking concrete steps to regulate emissions (US EPA, 2016; Environment and Climate Change Canada, 2017b).

In Canada, the oil and gas system is the major anthropogenic source of CH₄ emissions, accounting for almost half of total CH₄ emissions in the inventory (Environment and Climate Change Canada, 2016). The Canadian government recently announced regulations that support the goal of a 40–45% reduction of CH₄ emissions from the oil and gas system below 2012 levels by the year 2025 (Environment and Climate Change Canada, 2017b). Similarly, Alberta — the largest fossil fuel producing province in Canada — announced a 45% reduction goal and is developing its own set of regulations (Alberta Government, 2016). For federal and provincial goals to be realistically achieved and for the regulations to work, we need to know the current emissions baseline as well as the characteristics of the major emitting sources (e.g., spatial and temporal patterns, emissions distributions).

An extensive body of research has successfully reduced the uncertainty in terms of CH₄ emissions from the US oil and gas system (Allen et al., 2013; Allen, Pacsi, et al., 2015; Allen, Sullivan, et al., 2015; Lamb et al., 2015; Mitchell et al., 2015; Zimmerle et al., 2015; Omara et al., 2016). Relevant lessons include (i) identification of major emission sources across the natural gas supply chain (Allen, Pacsi, et al., 2015; Allen, Sullivan, et al., 2015; Lamb et al., 2015; Zimmerle et al., 2015; Marchese et al., 2015), (ii) proof that high-emitting sources disproportionately account for the majority of emissions (Brandt et al., 2014; Zavala-Araiza, Lyon, et al., 2015; Brandt et al., 2016; Omara et al., 2016; Lyon et al., 2016; Zavala-Araiza et al., 2017), and (iii) examples of smart regulations significantly reducing emissions from different sources (e.g., green completions, replacement of high-emitting pneumatic controllers) (Allen et al., 2013; Allen, Pacsi, et al., 2015). With a limited number of studies reporting empirical observations of oil and gas methane emissions outside the US — and specifically in Canada — (Johnson et al., 2017; Atherton et al., 2017) the question remains about whether the same emission characteristics found in the US apply elsewhere.

The main motivation behind the empirical data collected and analyzed here is to compare the emission patterns from oil and gas production sites of Alberta, Canada, to previously reported datasets of emissions from the US. If we were to find similarities, it would be possible to translate the multiple lessons learned in the US in terms of improving emission inventories, reducing emissions, and establishing effective and efficient regulations.

In this work, we (i) present empirical site-level estimates of CH₄ emissions from oil and gas production sites in Alberta derived from downwind ground-based measurements, (ii) characterize the emissions distribution and discuss its implications, and (iii) re-analyze five additional datasets to present a systematic and consistent comparison of CH₄ emissions from production sites across seven production basins in North America. As described below, we report measurements of production sites near Red Deer, Alberta — a region characterized by older natural gas production sites and light oil production (i.e., average age 15 years; age range: 0.8–55 years) (i.e., average gas production per site: 98 thousand cubic feet per day (Mcf d^–1), average oil production per site: 3.4 barrels per day (bbl d^–1)).

The quantification of site-level CH₄ emissions took place using an instrumented mobile laboratory. During the first week of November 2016, the miniature Aerodyne Mobile Laboratory (minAML) performed measurements in the oil and gas production region near Red Deer, Alberta (Figure 1). The Alberta Energy Regulator (AER) provided site access, which allowed quasi-random site selection (constrained by wind and road conditions) by the sampling team. Prior to the start of the measurements, the sampling team had a spatially-explicit list of active production sites. The list of wells was based on publicly available data from AER, (Alberta Energy Regulator, 2017) and then clustered into production sites as described in Zavala-Araiza et al. (Zavala-Araiza, Lyon, et al., 2015). Data on production site characteristics (e.g., age, gas production, oil production) were obtained from Drillinginfo (DI Desktop, 2017).

The research team conducted two types of experiments to quantify CH₄ emission rates of specific facilities. In the first, a method using controlled releases of tracer gas was employed (tracer flux method, hereafter). The second approach used an atmospheric dispersion algorithm to estimate CH₄ emissions from sites where the tracer release was not conducted (Gaussian dispersion method, hereafter). To deploy these two approaches, the sampling team randomly selected suitable sites for the tracer flux method, and in transit between tracer measurements would capture plume data suitable for posterior analysis using the Gaussian dispersion method. These fortuitous plume measurements have a higher probability of capturing bigger plumes related to sites with higher emissions, and do not capture sites with no emissions. As a consequence, we can characterize the tracer flux sample as a systematic sample, and the Gaussian dispersion sample as biased towards higher emissions (see discussion below).

The details of the deployment platform, the analytical equipment, and the two measurement approaches are briefly summarized below, with more details provided in SM, Text S1, Section 1.

The minAML acted as the deployment platform for the meteorological, GPS and analytical equipment used in these measurements. The design, and operational and analysis protocols used in the minAML are inherited from the Aerodyne Mobile Laboratory deployments (Kolb et al., 2004; Herndon et al., 2005; Roscioli et al., 2015). For this deployment, ambient air was continuously analyzed for CH₄, ethane (C₂H₆), and tracer release gases as well as other trace gas species. A modified pickup truck that accompanied the minAML also quantified meteorological parameters and is referred to as the tracer release vehicle.

An AirMar 200WX anemometer mounted on the sample mast was used to measure wind speed and direction. The position and heading of the minAML was determined by a Hemisphere GPS model V100.

The sample inlet was 2.2 m above ground, on the passenger side, and extended to the front bumper of the vehicle. Ambient air was continuously sampled through a ½” outer diameter (O.D.) Teflon tube. Three flow paths diverge, with 625 standard cubic centimeters per minute (sccm) splitting out to a LiCor-6262A (CO₂), ~500 sccm are directed to a Proton Transfer Reaction Mass Spectrometer (PTRMS) (Knighton et al., 2012) and 6000 sccm to the three infrared laser spectrometers in series (McManus et al., 2015). An upstream pressure controller regulated the cell pressure in the tunable infrared laser spectrometers. The typical mobile laboratory operating protocols and sampling techniques have been described previously (Herndon et al., 2005; Yacovitch et al., 2017).

Three tunable infrared direct absorption spectrometers (TILDAS) quantified nitrous oxide (N₂O), CH₄, C₂H₆, and propane (C₃H₈) as well as carbon monoxide (CO) and water vapor (H₂O) (McManus et al., 2015). A thorough description of the TILDAS instruments is documented elsewhere (McManus et al., 2015), but some brief details of how they were deployed on this project are warranted. The first spectrometer quantified N₂O, H₂O and CH₄ using the rotation-vibration absorption lines between 1300.8 and 1301.7 cm^–1. The second spectrometer was outfitted with two lasers, with the first quantifying CH₄ and C₂H₆ using the lines at 2990 cm^–1, and C₃H₈ at 2965 cm^–1. The final laser spectrometer measured N₂O, CO and H₂O at 2200 cm^–1. Note that there were redundant measurements of CH₄, N₂O and H₂O. Further details of the TILDAS instruments are documented elsewhere (Yacovitch et al., 2014).

Calibrations of the CH₄ mixing ratio were performed using a compressed gas cylinder containing 777 ± 7.8 ppmv CH₄ (1%). Calibrations of the N₂O tracer were performed before and after the measurement project using the procedure described in the supplemental information section. Propagation of uncertainties in the calibration factors suggested that the systematic uncertainty in the molar ratio of CH₄/N₂O was 1.8%.

The measurement experiment consisted of driving the tracer release vehicle to a well-site and beginning the steady release of a controlled flow of pure gas. N₂O tracer gas was released using an Alicat Flow Controller (MCR-20). Calibration verification was performed up to 20 SLPM. At the flow rates used in this work, we assessed the systematic uncertainty in the N₂O release rate to be 1.02%. The purity stated by the manufacturer (Praxair) was 99.2%.

The minAML conducted 4 to 20 transects of the tracer plume downwind (typically 50–1500 m). The sites were selected strictly on the basis of wind direction and suitable downwind roadways as well as road access to the well pad for the tracer release vehicle. One of the measurement days during the campaign was done without site access and the tracer release was limited to safe parking locations on public roads.

The tracer flux quantification method involves releasing a controlled amount of tracer gas(es) close to likely sources at the site (Lamb et al., 1995; Roscioli et al., 2015). Geographically resolved transects of the emission gas (e.g., CH₄ and C₂H₆) and tracer could then be measured downwind. Unlike other downwind quantification techniques (Brantley et al., 2014; US EPA, 2014; Yacovitch et al., 2015), the tracer flux method requires no knowledge or simulation of atmospheric dispersion. While wind is measured on-site and on the minAML, this information is only used as a diagnostic to direct where the tracer plume should be sought. The downwind measurement transects are used to probe how the enhancements above background for both the tracer species and emitted CH₄ are spatially coherent. This is a critical data quality indicator for the tracer flux method. Coupled with the measurement of the C₂H₆/CH₄ enhancement ratio, these spatial characteristics of the downwind transects allow attribution of the observed CH₄ emissions to the site. Interfering sources of CH₄ are identified due to the lack of coherence with the tracer gas and are not quantified by the tracer flux method.

For each well pad, the minAML collected a sequence of downwind plume transects (Figure 2). When the data exhibit the extent of correlation shown in this figure, the molar ratio of CH₄ and tracer species determines the CH₄ emission rate together with the known tracer flow rate (Equation 1).

When the data do not exhibit spatial correlation like the data in Figure 2, the area under the CH₄ and tracer plumes are used to determine the molar ratio. An automated algorithm chose the best analysis approach based on R² values for the linear fits of mixing ratio correlation. Thorough descriptions of the analysis protocol and the method choice algorithm is discussed elsewhere (Roscioli et al., 2015). The site-representative CH₄ emission rate was computed as the arithmetic mean of all quantified plumes and uncertainty estimated based on the statistical characterizations of the sampled population (Roscioli et al., 2015).

The experimental data used in these simulations consisted of periods of data, typically 1–5 min, taken in motion, and showing the CH₄ mixing ratio rise above then return to nominal background value. In this technique, the entire cross-section of a plume is transected at a single height, which differs from other methods using sensor array networks or stationary sampling at a single point combined with measurements of wind fluctuations (Thoma et al., 2012; Brantley et al., 2014).

The dispersion modeling described below is based on the standard Gaussian dispersion equations and procedures (Turner, 1994). Equation 2 relates the measured concentration of emissions (C) at a position × meters downwind, y meters off-axis, and z_r meters above ground. The source emission rate (Q), the horizontal wind velocity (u) and the height above ground of the plume centerline (H_e) are required, along with terms for vertical (σ_z) and horizontal (σ_y) dispersion. These dispersion terms are taken from a lookup table based on a chosen Pasquill stability class and are a function of ×, the downwind distance (see Text S1 for detailed description of input parameters as well as an estimation of error).

In practice, the concentrations are computed using a fixed emission magnitude (Q = 1 g/s), and the simulated plume scaled to the experimentally observed mixing ratios. This same scaling factor is applied to Q, yielding the final emission magnitude determination (see example in Figure 3). Source location is an important component of these simulations and is discussed in detail in Text S1.

We derived an emissions probability density function (pdf) that characterizes the production sites measured during this study using the statistical estimator developed by Zavala-Araiza et al. (Zavala-Araiza, Lyon, et al., 2015). This estimator allowed the integration of a systematic sample and a high-emitter biased sample to derive a facility-wide pdf representative of an entire population of sites (see SM, Text S1, Section 2 for description of the statistical estimator).

In addition to deriving the pdf, we also: (i) explored the relationship between emissions and production site parameters (i.e., oil/gas production, age of wells), (ii) compared measured emissions with reported emissions, and (iii) compared emissions in the Red Deer region with the emissions from other production regions in North America. To complete this final task, we re-analyzed five additional datasets to present a systematic and consistent comparison of CH₄ emissions from production sites across seven production basins in North America. Our analysis produces probability emission distributions, which are needed to assess the presence and effect of super-emitters (i.e., sites with excess of emissions related to abnormal process conditions) (Zavala-Araiza et al., 2017).

We performed a total of 60 measurements: 25 using the tracer flux method and 35 using the Gaussian dispersion method (see File S1 for dataset of measurements and Table S2). Four sites were measured by both methods, but measurements occurred at different times, thus they are considered independent samples. Figure 4 illustrates the distribution of measurements from both methods. The measurements from the Gaussian dispersion method are shifted towards higher emission rates (mean for the tracer flux method is 2.0 kg CH₄ h^–1 and 8.0 kg CH₄ h^–1 for the Gaussian dispersion method). As previously mentioned, this is expected because the tracer flux method is capturing a systematic sample of the population while the Gaussian dispersion method is biased towards bigger plumes (higher emitters). Hereafter we will refer to the tracer flux method sample as the systematic sample, and the Gaussian dispersion method sample as the high-emitter biased sample.

In line with previous studies in the US (Brantley et al., 2014; Lan et al., 2015; Rella et al., 2015; Yacovitch et al., 2015; Zavala-Araiza, Lyon, et al., 2015; Omara et al., 2016; Robertson et al., 2017), our measurements are characterized by skewed distributions: for the systematic sample 20% (5 of 25) of the measurements account for 68% (33 of 49 kg CH₄ h^–1) of total measured emissions, while for the high-emitter biased sample 29% (10 of 35) of the measurements account for 85% of total measured emissions (240 of 280 kg CH₄ h^–1).

For every measured production site, we evaluated the reported emissions during the month of the study, based on vented and flared whole gas volumes reported by industry under AER’s Directive 60 (AER, 2016). Reported vented volumes “(…) are assumed to comprise, where applicable, casing gas venting, waste associated gas flows, treater and stabilizer off-gas and gas volumes discharged during process upsets and equipment depressurization events (i.e., blowdowns).” (Clearstone Engineering Ltd., 2014). CH₄ emissions from these reported volumes were calculated as in Johnson et al. (Johnson et al., 2017) using site specific gas composition estimates derived using an AER data set of 310,000 useable gas samples associated with 120,000 well segments located throughout the province. Flare efficiencies were assumed to be 98%. Venting accounted for the dominant majority (98%) of the CH₄ emissions associated with reported gas volumes (reported emissions per site are included in File S1). Six sites classified as active had no reported production (oil or gas) during the month of the study, but they still had emissions ranging between 0.21 kg CH₄ h^–1 to 38 kg CH₄ h^–1.

Total measured emissions were 15 times higher than total reported emissions (330 kg CH₄ h^–1 vs. 22 kg CH₄ h^–1), and 60% of the measured sites reported zero emissions but emissions were detectable (Figure 5). These results highlight that the majority of the emissions in the Red Deer region are related to currently unreported sources, such as equipment leaks, emissions from pneumatic controllers, and unreported venting.

We used the systematic sample to perform a multiple regression analysis and explore the relationship between emissions and several production parameters at each site (i.e., gas production, oil production, age of wells, well type, and number of wells per site).

In our final regression model, emissions are a function of oil production, gas production, age, and an interaction term between age and gas production. This model explains roughly two-thirds of the variance in CH₄ emissions (R²_adj = 0.640, p < 1 × 10^–4). Of the explained variance, gas production is responsible for 80% of the variance, the interaction term for 13%, oil production 5.0% and age 2.0% (SM, Text S1, Section 4).

The interaction term in the model (gas production and age) underscores how emissions as a function of gas production increase differently as a site gets older. We see a higher increase in emissions as we transition to older sites (SM, Text S1, Figure S-3). This could be explained by the fact that older sites are more prone to situations that could cause excess emissions (e.g., malfunctions, or failures from older equipment). If any of these issues manifest at an old site, higher gas production translates into a higher volume of gas that could potentially be released to the atmosphere. It is expected that for any given site, gas production will decrease as a function of age (Baihly et al., 2010; Zavala-Araiza, Allen, et al., 2015). Nonetheless, the interaction term in our model captures the aggregate effect of the population of sites where we observe a significant number of older sites (age > 30 years) with substantial gas production (>150 Mcf d^–1).

On top of the relationship between age, gas production and emissions, our model suggests that oil producing sites tend to have higher emissions than sites without oil production: higher emissions are expected at sites with higher oil production (e.g., venting from condensate tanks (Lyon et al., 2016; Lavoie et al., 2017)).

Even though the explained variance is high enough to consider the relationships relevant, we acknowledge that it is still low in terms of appropriate predictive power. We hypothesize that the other one-third of unexplained variance in CH₄ emissions could be related to abnormal process conditions that have been reported in previous studies and would escape the routine behavior characterized by production levels and age (Zavala-Araiza et al., 2017). For example, Mitchell et al. (Mitchell et al., 2015) reported measurements of natural gas gathering and processing facilities in the US and as part of their analysis they tested similar regression models to the one that we present here. In addition, through a partnership with the operators, they were able to identify high-emission episodes at some of the measured facilities. The predictive power of their model increased in a statistically significant way when a categorical variable was included for those sites that were flagged as sites with abnormal process conditions (e.g., substantial venting).

Using two variations of the statistical estimator we derived pdfs that follow lognormal distributions:

• Using the systematic samples only; which relies on a measurement method with lower uncertainties but uses a small sample size (N = 25, µ = –0.31 σ = 1.5).

• Integrating the systematic sample and the high-emitter biased sample into a single distribution; which allows us to assess the impact of the low probability, high emitters that have emissions greater than the systematic sample maxima (N = 60, µ = –0.31 σ = 1.7).

The emission factors (EF) that characterize production sites in the region are 2.2 kg CH₄ h^–1 (95% confidence interval [CI]: 1.0, 5.4 kg CH₄ h^–1) for (i), and 2.9 kg CH₄ h^–1 (CI: 1.3, 6.8 kg CH₄ h^–1) for (ii) (see SM, Text S1, Section 2 for discussions about the difference between EFs and the statistical estimator.)

Plots in Figure 6 illustrate the disproportionate contribution of high-emitting sites to total emissions. While 50% of the sites with lower emissions only contribute 4–7% of total emissions, the 20% of sites with highest emissions contributes 74–79% of total emissions. Similarly, the 10% of sites with highest emissions is responsible for 58–65% of total emissions. In terms of absolute emission rates, 9.7–22% of sites have emissions greater than 5.0 kg CH₄ h^–1, and 3.9–12% of sites have emissions in excess of 10 kg CH₄ h^–1.

Using a bootstrap analysis (Efron and Tibshirani, 1993), we draw a random mean from the distribution of each of the two generated EFs with their associated confidence interval (assuming Gaussian distribution), leading to a central estimate of 2.6 kg CH₄ h^–1 (CI: 0.84, 4.3 kg CH₄ h^–1). We use this central estimate to characterize regional emissions.

Johnson et al. report a top-down regional airborne-based estimate of CH₄ emissions from oil and gas operations in a 50 km by 50 km box in the Red Deer region (SM, Text S1, Section S5 and Figure S-4) of 3,100 kg CH₄ h^–1 (CI: 900; 5,300 kg CH₄ h^–1) (Johnson et al., 2017). Even though only 40% of the ground-based measurements are inside the source region surveyed by the aircraft, the flight boundaries were selected to be representative of the activity in the Red Deer region which is dominated by upstream gas and oil production sites. If we use our central estimate to extrapolate emissions for the entire box (which contained 1,845 production sites), total emissions are 4,800 kg CH₄ h^–1 (CI: 1,500; 7,900 kg CH₄ h^–1). Thus, the ground-based estimate and the airborne-based estimate agree within confidence intervals.

The proportional loss rate (CH₄ emissions normalized by CH₄ production) is another commonly used metric to characterize emissions from natural gas production sites (Mitchell et al., 2015; Zavala-Araiza, Lyon, et al., 2015; Omara et al., 2016; Robertson et al., 2017). This metric can simplify the intercomparison among production regions with similar characteristics (e.g., similar gas/oil production volumes). In addition, it can sometimes be an effective way to identify sites that have an excess of emissions related to abnormal process conditions (e.g., malfunctions or equipment issues) (Zavala-Araiza, Lyon, et al., 2015; Zavala-Araiza et al., 2017).

We also replicated the analysis performed in Zavala-Araiza et al. (Zavala-Araiza, Lyon, et al., 2015); where we generated the joint density between emissions and production to estimate the pdf of proportional loss rates (Figure 7, and SM, Text S1, Section 6). From this analysis we find that 82% of sites emit >1% of their production, and they are responsible for 91% of the emissions. Similarly, 21% of sites emit >10% of their CH₄ production, accounting for 35% of emissions. Half the sites have proportional loss rate >3.3% and they account for 67% of emissions. It is important to note that the distribution of loss rates is skewed. Although the loss rates are potentially high, this is largely an effect of the small gas production volumes at most sites in the region. Thus, one needs to be cautious in directly comparing the loss rates in the Red Deer region to those observed in other regions with different production characteristics (and in particular, those with lower ratios of oil to gas wells).

To contextualize the Red Deer emissions distribution, we compared our results with several US oil and gas production regions where similar ground-based measurements of production sites have been performed. In this analysis we include the measurements reported for: Barnett Shale, Texas (Rella et al., 2015; Zavala-Araiza, Lyon, et al., 2015); Denver-Julesburg, Colorado (Brantley et al., 2014; Robertson et al., 2017); Fayetteville, Arkansas (Robertson et al., 2017); Marcellus Shale, Pennsylvania and West Virginia (Omara et al., 2016); Uintah, Utah (Robertson et al., 2017); and Upper Green River; Wyoming (Robertson et al., 2017).

With the exception of the Barnett Shale dataset— for which the statistical estimator approach was originally developed and applied— different estimation methods were used to derive EFs and characterize the other basins. To consistently compare the emission profiles, we took the site-level measurements for each region and replicated the statistical estimator approach (SM, Text S1, Section 7, Table S-4).

Figure 8 shows the pdf of each production region. In all cases we find skewed distributions. Previous work has looked at the presence of high-emitting sites across different producing regions in the US (Lyon et al., 2016), highlighting their dynamic behavior (in space and time) as well as their unpredictability. While a combination of production characteristics and operation practices could be related to the different levels of skewedness, we want to underscore that the Red Deer production region seems to fit into a wider group of North American production regions where high-emitting sites are drivers of emissions.

We also look at the relationships between emissions and production parameters (Figure 9). When we compare Upper Green River and Red Deer, we see similar EFs (Upper Green River: 2.4 kg CH₄ h^–1, Red Deer: 2.2 kg CH₄ h^–1), but gas production and oil production are 25× and 20× higher than in Red Deer. This could be explained by the higher skewedness in the Red Deer emissions distributions (Figure 8, and Gini coefficient in Figure 9) – illustrating greater influence of high emitting sites.

As noted above, the median loss rate of 3.3% in Red Deer (Figure 9) should be used cautiously, since average gas production is significantly lower than in the other production regions analyzed to date (the much smaller denominator makes it very sensitive to small changes in emissions– particularly from oil production). Focusing specifically on the 19 sites classified as gas sites (within the 25 sites from the systematic sample), the mean loss rate was ~3% with a 95% confidence interval of 2–5% (calculated via a bootstrapping analysis). This high rate may be heavily impacted by the low production volume at gas wells in this region. However, this gas-site specific loss rate is potentially directly comparable to natural gas wells in other regions with similar production characteristics, and is indicative of the GHG-intensity of natural gas production in the Red Deer region.

Our work indicates that high-emitting sites in Red Deer disproportionately account for the majority of emissions. This finding is consistent with several studies in the US that report CH₄ emissions from oil and gas production sites. Using the definition proposed in Zavala-Araiza et al. (Zavala-Araiza et al., 2017), we hypothesize that a significant fraction of these high-emitting sites are super-emitters: sites with large unintended emissions related to abnormal process conditions (e.g., malfunctions, equipment issues). Zavala-Araiza et al. showed that emissions related to routine operating conditions were not enough to explain sources at high-emitting sites, thus the high-emitting sites are most likely related to abnormal process conditions. A similar conclusion was also reached by an infrared camera survey of 8,000 gas-producing sites across the US (Lyon et al., 2016). On-site, high-frequency measurements are needed to find the root cause of high-emitting sites, nonetheless, it is expected that – as observed in other North American production regions with skewed emission distributions – abnormal process conditions are responsible for high emitting sites in the Red Deer region.

When super-emitting sites are present, different sites may experience such abnormal conditions at different points in time (e.g., the same site will not always have the same malfunction, or an equipment issue could manifest at different sites at varying times). Thus mitigating emissions effectively requires frequent monitoring with the time between inspections short enough to minimize the duration of spatio-temporally dynamic super-emitting sites (Lyon et al., 2016; Lavoie et al., 2017).

In addition to the mitigation opportunity related to the super-emitters, a significant fraction of sites had detectable emissions. In other words, other sources of unreported emissions (e.g., pneumatic controllers, fugitive emissions) are likely causes of detected emissions at a large fraction of production sites; regulatory action should address these sources through a combination of replacement of high-emitting equipment as well as comprehensive and more frequent leak detection and repair programs.

We also showed that emissions are related to age of sites as well as oil and gas production. When we add these relationships on top of the stochastic nature of super-emitters we end up with a complex landscape where it is hard to predict where emissions will appear. As a consequence, widespread inspection and control strategies are critical to reducing emissions until it is possible to accurately predict where and when emissions are likely to occur, something that should be possible once more data is available. With a probabilistic understanding of the sources of emissions they can be prophylactically avoided through routine maintenance and/or observation.

We acknowledge that the Red Deer production region might not be representative of other production regions in Alberta or Canada. However, this first characterization illustrates important similarities between production sites in North America. Future work should focus on collecting additional empirical observations at other regions where different geologic characteristics, operation practices, and age of sites might affect emission patterns.

Site IDs, method, methane emission rates with 95% confidence interval, gas production, oil production, age, wells per site, reported emissions, and methane content in gas are uploaded as SM Dataset S1 in Microsoft Excel. xlsx format.

We appreciate the assistance received from Gerald Palanca at Alberta Energy Regulator (AER) as well as Mike D’Antoni’s support during the measurements. We are very grateful to Conner Daube and Jason Curry at Aerodyne Research, Inc.

Funding for this work was provided by the generous supporters of the Environmental Defense Fund (EDF).

The authors declare no competing financial interest. DZA and MO work for Environmental Defense Fund (EDF), a non-profit organization funded by philanthropic contributions, which does not accept donations from the oil and gas industry in accordance with its corporate donations policy (https://www.edf.org/approach/partnerships/corporate-donation-policy).

• Contributed to conception and design: DZA, SCH, JRR, TIY, MRJ, DRT

• Performed measurements: SCH, JRR, TIY

• Contributed to analysis and interpretation of data: DZA, SCH, JRR, TIY, MRJ, DRT, MO, BK

• Drafted and/or revised the article: DZA, SCH, JRR, TIY, MRJ, DRT, MO, BK

• Approved the submitted version of the article: DZA, SCH, JRR, TIY, MRJ, DRT, MO, BK