- Email: [email protected]
Sensitivity of CO2 leak detection using a single atmospheric station
Available online at www.sciencedirect.com
ScienceDirect Energy Procedia 63 (2014) 3907 – 3914
GHGT-12
Sensitivity of CO2 leak detection using a single atmospheric station Padarn Wilsona,b, Andrew Feitza,c, Charles Jenkinsc, Henry Berkoa,c, Zoe Lohc,d, Ashok Luhard, Mark Hibberdd, Darren Spencerd and David Etheridgec,d a Geoscience Australia, GPO Box 378, Canberra, ACT, 2601, Australia b The Australia National University, Canberra, ACT, 0200, Australia c Cooperative Research Centre for Greenhouse Gas Technologies (CO2CRC), GPO Box 463, Canberra, ACT, 2601, Australia d CSIRO Marine and Atmospheric Research, PMB 1, Aspendale, VIC, 3195, Australia
Abstract Atmospheric CO2 perturbations from simulated leaks have been used to determine the minimum statistically significant emissions that can be detected above background concentrations using a single atmospheric station. The study uses high precision CO2 measurements from the Arcturus atmospheric monitoring station in the Bowen Basin, Australia. A statistical model of the observed CO2 signal was constructed, combining both a regression and a time series model. A non-parametric goodness of fit approach using the Kolmogorov-Smirnoff (KS) test was then used to test whether simulated perturbations can be detected against the modelled expected value of the background for certain hours of the day and for particular seasons. The KS test calculates the probability that the modelled leak perturbation could be caused by natural variation in the background. Using pre-whitened data and selecting optimum test conditions, minimum detectable leaks located 1 km from the measurement station were estimated at 22 tpd for an area source of size 100 m x 100 m and 14 tpd for a point source at a KS cutoff defined by using the formal p-value of 0.05. These are very large leaks located only 1 km from the station and have a high false alarm rate of 56%. An alternative p-value could be chosen to reduce the false alarm rate but then the minimum detectable leaks are larger. A long term, single measurement station monitoring program that is unconstrained by prior information on the possible direction or magnitude of a leak, and based solely on detection of perturbations of CO2 due to leakage above a (naturally noisy) background signal, is likely to take one or more years to detect leaks of the order of 10 kt p.a. The sensitivity of detection of a leak above a background signal could be greatly improved through the installation of additional atmospheric monitoring stations or through greater prior knowledge about the location and size of a suspected leak. © by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license © 2014 2013Published The Authors. Published by Elsevier Ltd. (http://creativecommons.org/licenses/by-nc-nd/3.0/). Selection and peer-review under responsibility of GHGT. Peer-review under responsibility of the Organizing Committee of GHGT-12 Keywords: Atmospheric monitoring; leak; emission; CO2; carbon dioxide; geological storage; geosequestration; CCS; model
1876-6102 © 2014 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the Organizing Committee of GHGT-12 doi:10.1016/j.egypro.2014.11.420
3908
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
1. Introduction The Arcturus greenhouse gas (GHG) monitoring station (ARA) operated from July 2010 to June 2014 at a location 50 km southeast of Emerald in Queensland, Australia (Figure 1). The station was part of a collaborative project between Geoscience Australia (GA) and CSIRO Marine and Atmospheric Research (CMAR) to establish and operate a high precision atmospheric monitoring facility for measurement of baseline greenhouse gases in a geological carbon dioxide capture and storage (CCS) region. The primary purpose of the station was to establish newly developed greenhouse gas monitoring technology and demonstrate best practice for regional baseline atmospheric monitoring appropriate for geological storage of carbon dioxide. The GHG records were to be used as a reference for monitoring of the atmosphere at a CO2 storage project (see for example [1] and [2]), providing a baseline to quantify typical variations in the area and a background against which any anomalies in the immediate vicinity of the storage might be detected. Atmospheric monitoring is an effective technique for demonstrating safe and secure storage of GHGs [3]. However, potential leaks from CCS storage operations may be masked by environmental and anthropogenic influences. The Arcturus site and environs is representative of the activities and ecology of Queensland Central Highlands and the GHG signals are influenced by cropping, pasture, cattle production, and gas and coal mining activities. There is a need to determine how large a CO2 leak would have to be before it can be detected above the background CO2 signal. Atmospheric CO2 perturbations, from simulated leaks, have been modelled to determine the minimum statistically significant CO2 emissions that can be detected above the background concentration at Arcturus.
Figure 1. Atmospheric GHG monitoring site, Arcturus (ARA), straddling pasture and cropping agricultural land.
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
3909
2. Methods 2.1. Measured CO2 The monitoring station comprises a modified air-conditioned shipping container equipped with gas monitoring instruments, meteorological sensors, and a 10 m fibre-glass mast with air inlets [4]. Two Picarro gas analysers (wavelength scanned cavity ring down spectrometers) continuously monitor GHGs and CO2 isotopes. One unit measures water vapour and isotopic ratios of carbon in CO2 (12C and 13C) while the other unit measures the concentrations of CH4, CO2 and water vapour. Atmospheric composition is also occasionally measured via air samples collected with flask sampling equipment. An automated weather station measures wind speed, wind direction, temperature, humidity and rainfall. A solar powered eddy covariance flux tower was also installed at the site, 250 m south of the main station. The flux tower comprises a LI-7500A LI-COR open-path eddy covariance gas instrument that measures atmospheric concentration of CO2 and H2O and their fluxes. Wind components in the three dimensions are measured using a CSAT3 sonic anemometer (Campbell Scientific Inc). Details on the measurements and associated metadata for the monitoring station and flux tower can be found in Etheridge et al. [5].
Figure 2: Hourly averaged concentration measurements of CO2, CH4 and H2O at Arcturus, from July 2010 to December 2012.
Hourly averaged concentrations of CO2, CH4 and water vapour measured at the site from July 2010 to December 2012 are shown in Figure 2, which shows that the CO2 concentration ranges from 373 to 531 ppm. The variation is largely due to diurnal cycles (e.g. Figure 4), with high CO2 levels at night as respired CO2 is trapped in the stable atmospheric boundary layer. Lower CO2 levels measured during the daytime are due to a combination of
3910
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
photosynthesis and greater atmospheric mixing. The CH4 levels range between a baseline value of about 1760 ppb to 2800 ppb. As expected, the water vapour displays a strong seasonal trend with November–April corresponding to a period of high rainfall. 2.2. Simulations of CO2 leakage The contribution to CO2 concentrations at the monitoring station (ARA) from a simulated leak was computed for the period January to December (2011) using a 3D coupled prognostic meteorological and air pollutant dispersion model, TAPM [6,7]. Simulations were conducted for various emission rates and distances from the monitoring station (1-10 km). TAPM has previously been applied to a variety of regional- and local-scale dispersion problems [8]. The leaks were simulated for an area source (100 m x 100 m) and a point source located SSE of Arcturus; this direction was found to produce the largest perturbations in CO2 concentration at Arcturus. Figure 3 shows that the modeled maximum daily CO2 perturbation at Arcturus during 2011 for a 25 tonnes per day (tpd) area source emission located 1 km upwind is approximately 15 ppm. As expected, there are higher perturbations closer to the source.
Figure 3: Modelled maximum daily CO2 perturbation from a 25 tpd area source (grey area) located 1 km SSE of the Arcturus monitoring station (ARA), January to December 2011.
2.3. Statistical modelling of the background CO2 We developed two statistical models of the observed CO 2 concentrations for use in our detection algorithm (described in the next section): a regression model, and a hybrid statistical model combining both a regression and time series model [9]. The regression model is a time dependent, generalized additive model relating the CO 2 concentration to other observed atmospheric variables (e.g. wind speed, temperature, humidity) and the CH4 concentration, whose variability is controlled my many of the same boundary layer dispersion processes as for CO 2. It accounts for seasonal trends through the inclusion of dummy variables. The time series model is based on a seasonal auto-regressive integrated moving average (ARIMA) model [10], but with the additional complexity of allowing auto-regressive relationships to depend on the time of day. Essentially, the combined model attempts to predict the CO2 concentrations based on presumed causal factors (e.g. atmospheric variables) which were also measured. The methodology is similar to that used, for example, in economic forecasting. The model can be trained on one set of measured data, and then checked against measurements that were not used in the training.
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
3911
Figure 4 presents an example of the observed CO2 hourly time series and that simulated by a) the regression model and b) the hybrid regression and time series model. Both statistical models, especially the hybrid one, simulate the observed CO2 signal very well by considering dependence on other measured quantities. This approach enables the simulation of CO2 baseline time series, for example for testing and calibrating detection methodologies.
Figure 4: Example plots comparing simulations of (a) the regression model and (b) the hybrid regression and time series model against the observed hourly CO2concentrations at Arcturus.
3. Results and discussion 3.1. Detection of leakage Given the reasonable performance of the regression model at simulating the CO2 background, it can be used to evaluate leak detection limits. The hybrid model and uncertainties associated with the modelling approach is assessed in [9]. We can use our regression model, combined with the dispersion model of Section 2.2, to devise a detection algorithm. Note that we need to use a stochastic simulation because the regression model for the CO2 background only fits the data in an average sense; the difference, model minus data, appears to be a random variable. The regression model can be used to simulate CO2 concentrations, both with and without the contribution from the leak. In the detection algorithm, we look at the probability distribution of the simulated CO2 concentrations, with and without an added CO2 source. The two probability distributions are compared using a standard Kolomogorov-Smirnov (KS) test, and we declare that a source has been detected if the distributions differ at a significance level p = 0.05. Results with and without pre-whitening [10] are shown in Figure 5, where we see that the minimum detectable leaks are rather large, except at the most favorable times of day (e.g. between 10am – 2pm). The minimum sized CO2 leak that could be detected using this model when the leak is about 1 km from a single monitoring station is 22 tpd for an area source of size 100 m x 100 m and 14 tpd for a point source.
3912
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
3.2. Quantification of false alarms Using the regression model we have developed, the false alarm rate of the detection methodology can be evaluated. As before, we simulate data with and without leaks, apply the KS test, and declare a leak at a formal significance level p = 0.05. The formal p-value of the KS test statistic is used to indicate a threshold in the KS test statistic, beyond which a leak is deemed to be detected. , It does not correspond to the exact p-value due to the violation of test assumptions. In the simulation, the number of false detections can be tallied to compute a false alarm rate. If the requirements for the KS test were met, the false alarm rate, by definition, would be 0.05. Because of correlations in the data, the rate is much higher. Initial results suggest that the KS test results in very high false alarm rates (e.g. 56% at the optimal detection hour and season found above) at formal p-values which are useful for detecting leaks of a reasonable size (e.g. p=0.05, leak size = 25 tpd). We may view the KS statistic as simply a pragmatic measure of goodness-of-fit, without placing importance on the derived p-value except as a threshold above which a leak is declared. The computed false alarm rates indicate the actual performance of the algorithm, indicating that large false alarm rates are likely at cutoff values which are useful for detection.
Figure 5: Quantification of the minimum leak detectable using a KS test at different times of day during winter using both the raw and whitened data for an emission 1 km from the monitoring station: (a) area source, (b) point source.
3.3. Future directions Evidently the current detection limits and false alarm rates are rather large, despite the good performance of the predictive model; this reflects the simplicity of the detection algorithm. Probably the largest gains could be achieved if the direction to a possible source were known (Figure 6). As presently implemented, the algorithm is attempting to locate a source in any direction. If the direction to a suspected source were known, one could select measurements from wind directions that would carry CO2 to the sensor and use the remaining measurements to train the model on the baseline. Preliminary work shows that this looks promising and could be further improved by removing periods of high variance (e.g. low wind speed) which typically happen in the early hours of the morning. As a related point, even one additional station would improve detection sensitivity [11,12]. In this case, for a favourable wind direction, the background contribution could be reduced to whatever originated between the two sensors. The predictive model of this background would then have a much less demanding task and much smaller leakage perturbations could be detected.
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
3913
Figure 6: Normalised CO2 observations (CO2 in the radial co-ordinate) with a simulated leak plotted radially against the wind direction: (a) before, and (b) after whitening. This shows the increased potential for filtering to improve signal compared to noise.
4. Conclusions x
We have created a statistical model to predict the major features of the CO2 background. This regression model allows improved detection rates and opens up options for more sophisticated detection.
x
The minimum sized CO2 leak that could be detected using this background model when the leak is about 1 km from a single monitoring station is 22 tpd for an area source of size 100 m x 100 m and 14 tpd for a point source. There is a high false alarm rate.
x
Improvements in detection could be achieved through additional stations or greater prior information concerning the leak to provide measurements or better estimates of the background.
x
This study suggests that without the additional information suggested above, atmospheric monitoring for CO2 using a single station is more suitable for monitoring and quantification of an identified leak (i.e. the instrument can be located close to the leak and in an optimal location) rather than kilometre-scale CO2 leak detection over vegetated regions.
Acknowledgements We acknowledge funding provided by the Australian Government through the Carbon Capture and Storage – Implementation budget measure and CSIRO to support this research. The authors also acknowledge funding for the research provided by the Australian Government through the CRC program. The authors also wish to acknowledge the assistance of the Sullivan Family and Field Engineering Services at Geoscience Australia. Theo Chiotis is thanked for preparing the figures. AF publishes with the permission of the CEO, Geoscience Australia. References [1] Leuning R, Etheridge DM, Luhar AK, Dunse BL, Atmospheric monitoring and verification technologies for CO2 geosequestration. International Journal of Greenhouse Gas Control 2008; 2:401–414. [2] Etheridge D, Luhar A, Loh Z, Leuning R, Spencer D, Steele P, Zegelin S, Allison C, Krummel P, Leist M, van der Schoot M. Atmospheric
3914
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
monitoring of the CO2CRC Otway Project and lessons for large scale CO2 storage projects. Energy Procedia 2011: 4:3666-3675. [3] Jenkins C, Cook PJ, Ennis-King J, Underschultz J, Boreham C, Dance T, de Catriat P, Etheridge D, Freifeld B, Hortle A, Kirste D, Paterson L, Pevzner R, Schacht U, Sharma S, Stalker L, Urosevic M. Safe storage and effective monitoring of CO2 in depleted gas fields. Proceedings of the National Academy of Science 2012; 109: 35-41. [4] Berko H, Etheridge D, Loh Z, Kuske TJ, Law R, Zegelin S, Gregory R, Spencer D, Feitz, AJ. Installation Report for Arcturus (ARA): An inland baseline station for the continuous measurement of atmospheric greenhouse gases. Record 2012/054. Canberra: Geoscience Australia; 2012. [5] Etheridge D, Loh Z, Schroder IF, Berko H, Kuske TJ, Allison C, Gregory R, Spencer D, Langenfelds R, Zegelin S, Hibberd M, Feitz AJ. Metadata report: Arcturus atmospheric greenhouse gas monitoring. Record 2014/037. Canberra: Geoscience Australia; 2014. [6] Hurley PJ, Physick WL, Luhar AK. TAPM: a practical approach to prognostic meteorological and air pollution modelling. Environmental Modelling and Software 2005; 20: 737–752. [7] http://www.csiro.au/Outcomes/Environment/Population-Sustainability/TAPM; [accessed 10/9/14]. [8] Luhar AK, Hurley PJ. Application of a coupled prognostic model to turbulence and dispersion in light-wind stable conditions, with an analytical correction to vertically resolve concentrations near the surface. Atmospheric Environment 2012; 51: 56–66. [9] Wilson P, Jenkins C, Luhar A, Hibberd M, Etheridge D, Loh Z, Feitz A. A filtered goodness-of-fit approach to the detection of CO2 leaks using a single background monitoring station, International Journal for Greenhouse Gas Control (in prep). [10] Brockwell PJ, Davis RA. Introduction to Time Series and Forecasting. Springer: New York; 1996. [11] Luhar AK, Etheridge DM, Leuning R, Loh ZM, Jenkins CR, Yee E. Locating and quantifying greenhouse gas emissions at a geological CO2 storage site using atmospheric modeling and measurements. Journal of Geophysical Research – Atmospheres 2014 (in press). [12] Humphries R, Jenkins C, Leuning R, Zegelin S, Griffith D, Caldow C, Berko H, Feitz A. Atmospheric tomography: A Bayesian inversion technique for determining the rate and location of fugitive emissions. Environmental Science and Technology 2012; 46: 1739–1746.
ScienceDirect Energy Procedia 63 (2014) 3907 – 3914
GHGT-12
Sensitivity of CO2 leak detection using a single atmospheric station Padarn Wilsona,b, Andrew Feitza,c, Charles Jenkinsc, Henry Berkoa,c, Zoe Lohc,d, Ashok Luhard, Mark Hibberdd, Darren Spencerd and David Etheridgec,d a Geoscience Australia, GPO Box 378, Canberra, ACT, 2601, Australia b The Australia National University, Canberra, ACT, 0200, Australia c Cooperative Research Centre for Greenhouse Gas Technologies (CO2CRC), GPO Box 463, Canberra, ACT, 2601, Australia d CSIRO Marine and Atmospheric Research, PMB 1, Aspendale, VIC, 3195, Australia
Abstract Atmospheric CO2 perturbations from simulated leaks have been used to determine the minimum statistically significant emissions that can be detected above background concentrations using a single atmospheric station. The study uses high precision CO2 measurements from the Arcturus atmospheric monitoring station in the Bowen Basin, Australia. A statistical model of the observed CO2 signal was constructed, combining both a regression and a time series model. A non-parametric goodness of fit approach using the Kolmogorov-Smirnoff (KS) test was then used to test whether simulated perturbations can be detected against the modelled expected value of the background for certain hours of the day and for particular seasons. The KS test calculates the probability that the modelled leak perturbation could be caused by natural variation in the background. Using pre-whitened data and selecting optimum test conditions, minimum detectable leaks located 1 km from the measurement station were estimated at 22 tpd for an area source of size 100 m x 100 m and 14 tpd for a point source at a KS cutoff defined by using the formal p-value of 0.05. These are very large leaks located only 1 km from the station and have a high false alarm rate of 56%. An alternative p-value could be chosen to reduce the false alarm rate but then the minimum detectable leaks are larger. A long term, single measurement station monitoring program that is unconstrained by prior information on the possible direction or magnitude of a leak, and based solely on detection of perturbations of CO2 due to leakage above a (naturally noisy) background signal, is likely to take one or more years to detect leaks of the order of 10 kt p.a. The sensitivity of detection of a leak above a background signal could be greatly improved through the installation of additional atmospheric monitoring stations or through greater prior knowledge about the location and size of a suspected leak. © by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license © 2014 2013Published The Authors. Published by Elsevier Ltd. (http://creativecommons.org/licenses/by-nc-nd/3.0/). Selection and peer-review under responsibility of GHGT. Peer-review under responsibility of the Organizing Committee of GHGT-12 Keywords: Atmospheric monitoring; leak; emission; CO2; carbon dioxide; geological storage; geosequestration; CCS; model
1876-6102 © 2014 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the Organizing Committee of GHGT-12 doi:10.1016/j.egypro.2014.11.420
3908
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
1. Introduction The Arcturus greenhouse gas (GHG) monitoring station (ARA) operated from July 2010 to June 2014 at a location 50 km southeast of Emerald in Queensland, Australia (Figure 1). The station was part of a collaborative project between Geoscience Australia (GA) and CSIRO Marine and Atmospheric Research (CMAR) to establish and operate a high precision atmospheric monitoring facility for measurement of baseline greenhouse gases in a geological carbon dioxide capture and storage (CCS) region. The primary purpose of the station was to establish newly developed greenhouse gas monitoring technology and demonstrate best practice for regional baseline atmospheric monitoring appropriate for geological storage of carbon dioxide. The GHG records were to be used as a reference for monitoring of the atmosphere at a CO2 storage project (see for example [1] and [2]), providing a baseline to quantify typical variations in the area and a background against which any anomalies in the immediate vicinity of the storage might be detected. Atmospheric monitoring is an effective technique for demonstrating safe and secure storage of GHGs [3]. However, potential leaks from CCS storage operations may be masked by environmental and anthropogenic influences. The Arcturus site and environs is representative of the activities and ecology of Queensland Central Highlands and the GHG signals are influenced by cropping, pasture, cattle production, and gas and coal mining activities. There is a need to determine how large a CO2 leak would have to be before it can be detected above the background CO2 signal. Atmospheric CO2 perturbations, from simulated leaks, have been modelled to determine the minimum statistically significant CO2 emissions that can be detected above the background concentration at Arcturus.
Figure 1. Atmospheric GHG monitoring site, Arcturus (ARA), straddling pasture and cropping agricultural land.
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
3909
2. Methods 2.1. Measured CO2 The monitoring station comprises a modified air-conditioned shipping container equipped with gas monitoring instruments, meteorological sensors, and a 10 m fibre-glass mast with air inlets [4]. Two Picarro gas analysers (wavelength scanned cavity ring down spectrometers) continuously monitor GHGs and CO2 isotopes. One unit measures water vapour and isotopic ratios of carbon in CO2 (12C and 13C) while the other unit measures the concentrations of CH4, CO2 and water vapour. Atmospheric composition is also occasionally measured via air samples collected with flask sampling equipment. An automated weather station measures wind speed, wind direction, temperature, humidity and rainfall. A solar powered eddy covariance flux tower was also installed at the site, 250 m south of the main station. The flux tower comprises a LI-7500A LI-COR open-path eddy covariance gas instrument that measures atmospheric concentration of CO2 and H2O and their fluxes. Wind components in the three dimensions are measured using a CSAT3 sonic anemometer (Campbell Scientific Inc). Details on the measurements and associated metadata for the monitoring station and flux tower can be found in Etheridge et al. [5].
Figure 2: Hourly averaged concentration measurements of CO2, CH4 and H2O at Arcturus, from July 2010 to December 2012.
Hourly averaged concentrations of CO2, CH4 and water vapour measured at the site from July 2010 to December 2012 are shown in Figure 2, which shows that the CO2 concentration ranges from 373 to 531 ppm. The variation is largely due to diurnal cycles (e.g. Figure 4), with high CO2 levels at night as respired CO2 is trapped in the stable atmospheric boundary layer. Lower CO2 levels measured during the daytime are due to a combination of
3910
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
photosynthesis and greater atmospheric mixing. The CH4 levels range between a baseline value of about 1760 ppb to 2800 ppb. As expected, the water vapour displays a strong seasonal trend with November–April corresponding to a period of high rainfall. 2.2. Simulations of CO2 leakage The contribution to CO2 concentrations at the monitoring station (ARA) from a simulated leak was computed for the period January to December (2011) using a 3D coupled prognostic meteorological and air pollutant dispersion model, TAPM [6,7]. Simulations were conducted for various emission rates and distances from the monitoring station (1-10 km). TAPM has previously been applied to a variety of regional- and local-scale dispersion problems [8]. The leaks were simulated for an area source (100 m x 100 m) and a point source located SSE of Arcturus; this direction was found to produce the largest perturbations in CO2 concentration at Arcturus. Figure 3 shows that the modeled maximum daily CO2 perturbation at Arcturus during 2011 for a 25 tonnes per day (tpd) area source emission located 1 km upwind is approximately 15 ppm. As expected, there are higher perturbations closer to the source.
Figure 3: Modelled maximum daily CO2 perturbation from a 25 tpd area source (grey area) located 1 km SSE of the Arcturus monitoring station (ARA), January to December 2011.
2.3. Statistical modelling of the background CO2 We developed two statistical models of the observed CO 2 concentrations for use in our detection algorithm (described in the next section): a regression model, and a hybrid statistical model combining both a regression and time series model [9]. The regression model is a time dependent, generalized additive model relating the CO 2 concentration to other observed atmospheric variables (e.g. wind speed, temperature, humidity) and the CH4 concentration, whose variability is controlled my many of the same boundary layer dispersion processes as for CO 2. It accounts for seasonal trends through the inclusion of dummy variables. The time series model is based on a seasonal auto-regressive integrated moving average (ARIMA) model [10], but with the additional complexity of allowing auto-regressive relationships to depend on the time of day. Essentially, the combined model attempts to predict the CO2 concentrations based on presumed causal factors (e.g. atmospheric variables) which were also measured. The methodology is similar to that used, for example, in economic forecasting. The model can be trained on one set of measured data, and then checked against measurements that were not used in the training.
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
3911
Figure 4 presents an example of the observed CO2 hourly time series and that simulated by a) the regression model and b) the hybrid regression and time series model. Both statistical models, especially the hybrid one, simulate the observed CO2 signal very well by considering dependence on other measured quantities. This approach enables the simulation of CO2 baseline time series, for example for testing and calibrating detection methodologies.
Figure 4: Example plots comparing simulations of (a) the regression model and (b) the hybrid regression and time series model against the observed hourly CO2concentrations at Arcturus.
3. Results and discussion 3.1. Detection of leakage Given the reasonable performance of the regression model at simulating the CO2 background, it can be used to evaluate leak detection limits. The hybrid model and uncertainties associated with the modelling approach is assessed in [9]. We can use our regression model, combined with the dispersion model of Section 2.2, to devise a detection algorithm. Note that we need to use a stochastic simulation because the regression model for the CO2 background only fits the data in an average sense; the difference, model minus data, appears to be a random variable. The regression model can be used to simulate CO2 concentrations, both with and without the contribution from the leak. In the detection algorithm, we look at the probability distribution of the simulated CO2 concentrations, with and without an added CO2 source. The two probability distributions are compared using a standard Kolomogorov-Smirnov (KS) test, and we declare that a source has been detected if the distributions differ at a significance level p = 0.05. Results with and without pre-whitening [10] are shown in Figure 5, where we see that the minimum detectable leaks are rather large, except at the most favorable times of day (e.g. between 10am – 2pm). The minimum sized CO2 leak that could be detected using this model when the leak is about 1 km from a single monitoring station is 22 tpd for an area source of size 100 m x 100 m and 14 tpd for a point source.
3912
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
3.2. Quantification of false alarms Using the regression model we have developed, the false alarm rate of the detection methodology can be evaluated. As before, we simulate data with and without leaks, apply the KS test, and declare a leak at a formal significance level p = 0.05. The formal p-value of the KS test statistic is used to indicate a threshold in the KS test statistic, beyond which a leak is deemed to be detected. , It does not correspond to the exact p-value due to the violation of test assumptions. In the simulation, the number of false detections can be tallied to compute a false alarm rate. If the requirements for the KS test were met, the false alarm rate, by definition, would be 0.05. Because of correlations in the data, the rate is much higher. Initial results suggest that the KS test results in very high false alarm rates (e.g. 56% at the optimal detection hour and season found above) at formal p-values which are useful for detecting leaks of a reasonable size (e.g. p=0.05, leak size = 25 tpd). We may view the KS statistic as simply a pragmatic measure of goodness-of-fit, without placing importance on the derived p-value except as a threshold above which a leak is declared. The computed false alarm rates indicate the actual performance of the algorithm, indicating that large false alarm rates are likely at cutoff values which are useful for detection.
Figure 5: Quantification of the minimum leak detectable using a KS test at different times of day during winter using both the raw and whitened data for an emission 1 km from the monitoring station: (a) area source, (b) point source.
3.3. Future directions Evidently the current detection limits and false alarm rates are rather large, despite the good performance of the predictive model; this reflects the simplicity of the detection algorithm. Probably the largest gains could be achieved if the direction to a possible source were known (Figure 6). As presently implemented, the algorithm is attempting to locate a source in any direction. If the direction to a suspected source were known, one could select measurements from wind directions that would carry CO2 to the sensor and use the remaining measurements to train the model on the baseline. Preliminary work shows that this looks promising and could be further improved by removing periods of high variance (e.g. low wind speed) which typically happen in the early hours of the morning. As a related point, even one additional station would improve detection sensitivity [11,12]. In this case, for a favourable wind direction, the background contribution could be reduced to whatever originated between the two sensors. The predictive model of this background would then have a much less demanding task and much smaller leakage perturbations could be detected.
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
3913
Figure 6: Normalised CO2 observations (CO2 in the radial co-ordinate) with a simulated leak plotted radially against the wind direction: (a) before, and (b) after whitening. This shows the increased potential for filtering to improve signal compared to noise.
4. Conclusions x
We have created a statistical model to predict the major features of the CO2 background. This regression model allows improved detection rates and opens up options for more sophisticated detection.
x
The minimum sized CO2 leak that could be detected using this background model when the leak is about 1 km from a single monitoring station is 22 tpd for an area source of size 100 m x 100 m and 14 tpd for a point source. There is a high false alarm rate.
x
Improvements in detection could be achieved through additional stations or greater prior information concerning the leak to provide measurements or better estimates of the background.
x
This study suggests that without the additional information suggested above, atmospheric monitoring for CO2 using a single station is more suitable for monitoring and quantification of an identified leak (i.e. the instrument can be located close to the leak and in an optimal location) rather than kilometre-scale CO2 leak detection over vegetated regions.
Acknowledgements We acknowledge funding provided by the Australian Government through the Carbon Capture and Storage – Implementation budget measure and CSIRO to support this research. The authors also acknowledge funding for the research provided by the Australian Government through the CRC program. The authors also wish to acknowledge the assistance of the Sullivan Family and Field Engineering Services at Geoscience Australia. Theo Chiotis is thanked for preparing the figures. AF publishes with the permission of the CEO, Geoscience Australia. References [1] Leuning R, Etheridge DM, Luhar AK, Dunse BL, Atmospheric monitoring and verification technologies for CO2 geosequestration. International Journal of Greenhouse Gas Control 2008; 2:401–414. [2] Etheridge D, Luhar A, Loh Z, Leuning R, Spencer D, Steele P, Zegelin S, Allison C, Krummel P, Leist M, van der Schoot M. Atmospheric
3914
Padarn Wilson et al. / Energy Procedia 63 (2014) 3907 – 3914
monitoring of the CO2CRC Otway Project and lessons for large scale CO2 storage projects. Energy Procedia 2011: 4:3666-3675. [3] Jenkins C, Cook PJ, Ennis-King J, Underschultz J, Boreham C, Dance T, de Catriat P, Etheridge D, Freifeld B, Hortle A, Kirste D, Paterson L, Pevzner R, Schacht U, Sharma S, Stalker L, Urosevic M. Safe storage and effective monitoring of CO2 in depleted gas fields. Proceedings of the National Academy of Science 2012; 109: 35-41. [4] Berko H, Etheridge D, Loh Z, Kuske TJ, Law R, Zegelin S, Gregory R, Spencer D, Feitz, AJ. Installation Report for Arcturus (ARA): An inland baseline station for the continuous measurement of atmospheric greenhouse gases. Record 2012/054. Canberra: Geoscience Australia; 2012. [5] Etheridge D, Loh Z, Schroder IF, Berko H, Kuske TJ, Allison C, Gregory R, Spencer D, Langenfelds R, Zegelin S, Hibberd M, Feitz AJ. Metadata report: Arcturus atmospheric greenhouse gas monitoring. Record 2014/037. Canberra: Geoscience Australia; 2014. [6] Hurley PJ, Physick WL, Luhar AK. TAPM: a practical approach to prognostic meteorological and air pollution modelling. Environmental Modelling and Software 2005; 20: 737–752. [7] http://www.csiro.au/Outcomes/Environment/Population-Sustainability/TAPM; [accessed 10/9/14]. [8] Luhar AK, Hurley PJ. Application of a coupled prognostic model to turbulence and dispersion in light-wind stable conditions, with an analytical correction to vertically resolve concentrations near the surface. Atmospheric Environment 2012; 51: 56–66. [9] Wilson P, Jenkins C, Luhar A, Hibberd M, Etheridge D, Loh Z, Feitz A. A filtered goodness-of-fit approach to the detection of CO2 leaks using a single background monitoring station, International Journal for Greenhouse Gas Control (in prep). [10] Brockwell PJ, Davis RA. Introduction to Time Series and Forecasting. Springer: New York; 1996. [11] Luhar AK, Etheridge DM, Leuning R, Loh ZM, Jenkins CR, Yee E. Locating and quantifying greenhouse gas emissions at a geological CO2 storage site using atmospheric modeling and measurements. Journal of Geophysical Research – Atmospheres 2014 (in press). [12] Humphries R, Jenkins C, Leuning R, Zegelin S, Griffith D, Caldow C, Berko H, Feitz A. Atmospheric tomography: A Bayesian inversion technique for determining the rate and location of fugitive emissions. Environmental Science and Technology 2012; 46: 1739–1746.