Scholarly article on topic 'Atmospheric tomography to locate CO2 leakage at storage sites'

Atmospheric tomography to locate CO2 leakage at storage sites Academic research paper on "Earth and related environmental sciences"

Share paper
Academic journal
Energy Procedia
{"CO2 sequestration" / "Monitoring and verification" / "Atmospheric techniques"}

Abstract of research paper on Earth and related environmental sciences, author of scientific article — Charles Jenkins, Ray Leuning, Zoe Loh

Abstract Atmospheric monitoring can in principle be used to quantify emissions of leakage from geologically stored CO2, if the locations of the leaks are known. However, large-scale industrial storage sites will have a surface footprint of many tens of square kilometres. Atmospheric tomography can be used to determine the location and area of an unknown source using an array of instruments that measure atmospheric concentrations of the leaking gas or its tracer(s). We have completed a detailed simulation study that uses synthetic concentration perturbations downwind of a source at an unknown location within an array of point samplers to test the feasibility of using ‘atmospheric tomography’ to locate the source. The results of this simulation are extremely encouraging and suggest that the limiting factor in a tomographic method of leakage location detection is largely the need to gather enough wind directions to triangulate the location of the leakage source. This may take some months, depending on weather patterns, but by hypothesis one would be searching for a slow, steady leak; intermittent or fast leaks can be detected in other, easier ways. A field experiment is in progress to verify these conclusions.

Academic research paper on topic "Atmospheric tomography to locate CO2 leakage at storage sites"

 Available online at

W ScienceDirect Energy

l^jBL Procedía

ELSEVIER Energy Procedía 4 (2011) 3502-3509


Atmospheric tomography to locate C02 leakage at storage sites

Charles Jenkinsa'b'*, Ray Leuningab, Zoe Loha,c

a Cooperative Research Centre for Greenhouse Gas Technologies b CSIRO Pye Laboratory, Clunies Ross St, Black Mountain, ACT2601, Australia c CSIRO Marine & Atmospheric Research, 107-121 Station Rd, Aspendale Victoria 3195, Australia


Atmospheric monitoring can in principle be used to quantify emissions of leakage from geologically stored C02, if the locations of the leaks are known. However, large-scale industrial storage sites will have a surface footprint of many tens of square kilometres. Atmospheric tomography can be used to determine the location and area of an unknown source using an array of instruments that measure atmospheric concentrations of the leaking gas or its tracer(s). We have completed a detailed simulation study that uses synthetic concentration perturbations downwind of a source at an unknown location within an array of point samplers to test the feasibility of using 'atmospheric tomography' to locate the source. The results of this simulation are extremely encouraging and suggest that the limiting factor in a tomographic method of leakage location detection is largely the need to gather enough wind directions to triangulate the location of the leakage source. This may take some months, depending on weather patterns, but by hypothesis one would be searching for a slow, steady leak; intermittent or fast leaks can be detected in other, easier ways. A field experiment is in progress to verify these conclusions.

© 2011 Published by Elsevier Ltd.

Keywords: C02 sequestration; monitoring and verification; atmospheric techniques. Introduction

Atmospheric monitoring has often been discussed as a method of detecting leaks from storage sites (e.g. [1]). One may measure the concentration of C02 in the atmosphere near a site, or measure the flux of C02 from the surface. It is sensible to look for evidence of leakage in the atmosphere, as this is exactly where we do not wish the C02 to be. Monitoring in the atmosphere also has the benefit of a large "footprint" - C02 will be carried large distances from a leakage point or area - but equally it will be diluted as it is dispersed. More localized techniques, such as soil flux chambers [2], may have a bigger signal to look for but also have a much smaller footprint. If we do not know exactly where to look, an improbably large number of sensors is soon required if the footprint is small [2]. These issues of relative scales are discussed in [3] and [1]. It now seems clear that there is no single, perfect monitoring method that is applicable in all circumstances. In this paper we discuss ways of using atmospheric monitoring as an initial survey technique when some information is available about where a leak may occur, and where other strategies will be used for follow-up investigation.

The C02 from a leakage source (point, line, or area) is carried downwind. This transport is always turbulent, but more so if the boundary layer is unstable and convection is occurring. The dispersion is quite well described by Monin-0bukhov similarity theory [4] near the ground when conditions are windy or convection is strong. In calm conditions, with little convection or mechanical mixing, the dispersion is difficult to model and is strongly influenced by details of local topography [5]. Unfortunately, the greatest dilution thus occurs in the best understood conditions for atmospheric transport.

A large diurnal variation in C02 concentration is observed over vegetated surfaces due to the interaction of atmospheric mixing, photosynthetic uptake of C02 during the day and its release by respiration at night. Figure 1


shows data from our C02 injection demonstration site in the Otway Basin, Australia [6]. This variability constitutes the noise against which one wishes to detect a leakage signal. The variability can be predicted to a quite useful extent from linked meteorological and ecosystem models, so allowing anomalies in C02 concentration to be identified [7].

3 440 a

5420 t_>

380 360

Feb 22 Feb 24 Feb 26 Feb 28

Figure 1 Typical time series of C02 concentration from the Otway site. The strong night-time peaks result from a combination of plant respiration and light winds. These peaks do not appear on the nights of the 23rd, 27th and 28th February, which were windy. Warm days and onshore breezes result in a low, stable C02 background during the day.

High levels of turbulent mixing during the daytime result in fairly constant C02 concentrations (Figure 1). Even without the aid of dispersion modeling, Etheridge et al. [8] detected a synthetic leak 700 m distant at the Otway site from inspection of C02 daytime concentration data. The nature of the "leak" was confirmed by more detailed measurements of CO and 13C02 concentrations, and modelling. Detection of the source was made with a single concentration sensor and was made possible by a favorable wind direction. This result suggests that the unknown position of a source may be discovered using a network of sensors measuring C02 concentrations in air coming from various wind directions. Indeed, Hirst et al [9] used such an approach with measurements of ethane concentrations from various wind directions to locate potential hydrocarbon reservoirs over an area of several hundred km2. 0nce the position of apotential source is located, appropriate measurement techniques can be selected to quantify the leakage rates [1], geophysical surveys may be initiated, pressure data closely scrutinized, and so on.

The essential advantage of an atmospheric measurement is that a C02 source affects concentrations for a large distance downwind (Figure 2). How large depends on how low a concentration can be measured above natural variability, and on the dispersion regime, but we know from experience with the 0tway data that the useful extent must be hundreds of meters along-wind. Even this may not be useful if we have to monitor the entire area in which a C02 leak may lie. A search strategy must make use of prior information to narrow down the places to look, and should also be adaptive and hence mobile. In what follows we will describe a tomographic technique, which uses a network of sensors to localize an anomalous source of C02, and we will then also comment on the search problem -where to put the sensors and the benefits of moving them about.

Current Instrumentation and Models

A wide range of instrumentation is available to measure C02 and other trace gas concentrations [1],as is the instrumentation needed to measure mean wind speed and direction and the Monin-0bukhov (M0) length (which represents atmospheric stability) required to model dispersion modeling.

Data from the 0tway Project illustrate the possibilities. C02 concentrations are measured with a high precision non-dispersive infrared analyzer (LoFlo) [10], [11]. Air is sampled at 10 m height approximately 700 m downwind of the injection and monitoring wells under prevailing wind conditions. Since December 2009, the LoFlo measurements

n / k i ; -W-J4 ! / 1, 1 1 1 ii I n

have been augmented by two wavelength-scanned cavity ring-down spectrometers (Picarro Inc.) which provide simultaneous measurements of C02 plus natural tracers present in the storage reservoir, one unit measures C02, CH4 and H20 and the second 13C02 and H20. The newer tracer measurements add considerable sensitivity to the monitoring effort [11]. Interpretation of these data is achieved by detailed boundary-layer measurements, which quantify the dispersion conditions in the boundary layer (Figure 3) and modeling that integrates large- and small-scale ecological, meteorological, and dispersion information [7].

Also located on the site are three solar-powered autonomous sensor platforms under development by CanSyd [13]. These platforms, named CEGEM, use the Vaisala GMP343 instruments for C02 air concentration and soil flux measurements, and are exploring the possibilities offered by multiple, lower-precision sensors. The CEGEM data are however not used for monitoring and verification at the site.

Figure 2. Examples of the concentration plumes downwind of a source, computed for Monin-0bukhov lengths of, top, 50m (stable conditions) and -20m (unstable conditions). The y-axis is crosswind direction. The contours are at logarithmic intervals of 0.5 in concentration. The calculations were made with the Windtrax software code [13], which implements a Lagrangian stochastic particle dispersion model.

The C02 data from the 0tway site are an important part of the assurance monitoring programme and, as noted, have been able to detect and quantify anomalous emissions from the vicinity of the injection and monitoring wells. With a single sensor, however, the location and size of the source has to be assumed before further quantification is possible from the data. In a related experiment at our Ginninderra test site [12], the limits on quantification were explored by sampling air upwind and downwind of a source of C02 and CH4 at a known location. This work showed that inverse dispersion modeling could recover emission rates well, provided the prrturbation in concentrations (downwind - upwind) were greater than 1% above background levels (> 4 ppm for C02 and >1.8 ppm for CH4). These conclusions apply only to well-mixed neutral and unstable atmospheric conditions because of the difficulty of modelling atmospheric dispersion in stable boundary layers. Subtracting upwind concentrations from those downwind automatically eliminates the need to model upwind C02 sources and sinks at distances large compared to the size of the sampling array.

The combination of the 0tway and Ginninderra experience has encouraged us to explore a localization technique we call "atmospheric tomography" which we now describe in more detail.

Atmospheric Tomography

The basic concept is quite simple. A source of C02 must exist between two sensors at positions A and B the wind blows from A to B and the C02 concentration at B is greater than at A. The source can be thus be localized by a kind of triangulation if we have several pairs of sensors and several wind directions. As we have argued, localizing a source is valuable information because it means that specific diagnostic techniques can be deployed in the right place as needed. We can do better than this simple concept because we can select unstable conditions, where we know atmospheric dispersion theory works well, to calculate the concentrations for any location downwind of a known source. This extra information increases the potential accuracy of location, and justifies the use of the word "tomography" as we could in principle map the sources in the interior of our array by probing it with various wind directions. It is worth waiting for the right stability conditions and wind directions. Waiting is not an issue because we are concerned with small leaks by design: large ones can be found by other, more direct methods.

Figure 3 (left), the cumulative distribution of measured inverse M-0 length at the Otway site over three years. Small and negative values of this parameter correspond to unstable conditions, which are seen to be common, and small and positive correspond to strongly stable conditions. Right, the wind rose for the same data (dashed lines for summer). Some wind directions are uncommon and localization along those directions will have to wait some time for suitable winds.

Given the location, shape and strength of the source we can predict the concentrations our sensors will measure. However, it is the concentrations we know and we wish to invert these to obtain the position of the source, which as noted is valuable information by itself. To make progress we firstly assume that the sources are small compared to the sensor spacing. This is not essential, although some assumption is necessary. A leaking well would be a point source and this is a likely scenario. Secondly, we assume some knowledge of the source strength to help constrain the inverse solution. Finally, we use prior information as to where to look for a leak. These considerations suggest a Bayesian approach [15] to the inversion problem, which we now describe.

Localization by a Bayesian method

A Bayesian inversion of the concentration data assumes we know three things - the probability distribution of the concentration perturbation, the prior probability on the source location, and the prior probability on the source strength. An additional assumption is the source geometry. An advantage of Bayesian methods is that the use of prior information is made explicit; moreover, prior information stabilizes the inversion but it can be over-ridden by the data if it is incorrect.

Figure 4. Schematic of the simulated sensor array, showing the four sensors at the corners of the square. The plume shows the actual relative concentrations downwind in a moderately unstable boundary layer.

Suppose the plume shape (for a given stability condition in the boundary layer, and wind direction) is described by a function/. For a source of strength S at location (x0,y0) the predicted concentration c at sensor position (x,y) is then given by c(x,y)=Sf(x,y,x0,y0). Each measurement m is the difference in concentration measured by sensors downwind and upwind in the array. Assume for definiteness that these data are normally distributed with variance a2: we can then write the probability density distribution for m as

For a set of measurements taken at different sensor locations and at different times, the combined probability of the data (the likelihood function) will be the product of terms like the RHS of Equation (1). Denote this likelihood function by L(x0,y0,S), to show its dependence on the unknowns. Bayes' Theorem then tells us the posterior probability density for these unknowns:

in which the two prior probabilities on the RHS indicate the region where we expect to find a source, and the probability of it having a particular strength. Thus relevant information is naturally included. The prior on source location, for instance, might be a fairly broad Gaussian centered on the suspected position of some old wells. The prior on strength, without any better information, might be the so-called Jeffreys prior [14]:

prob^S* ) x —

500 -500

prob(x0,y0,S) x L(x0,y0,S)prob(x0,y0)prob(S)

which is a well-known "agnostic" Bayesian prior for scale parameters and a useful starting point for leak strength, favouring small leaks. Notice that the data (condensed into the likelihood) may be processed with a range of priors and so can be examined again if prior information changes. Thus it is straightforward to check the sensitivity of the conclusions to the prior information.

Finally the posterior probability can be condensed into a probability distribution for location alone by marginalizing out the source strength; this is a common Bayesian approach to eliminating so-called nuisance parameters [14].

prob(x0,y0) œ JdSL(x0,y0,S)prob(x0,y0)prob(S)

A similar integration could return the posterior probability distribution of source strength alone. The posterior can be normalized by integrating over all three variables. These integrations are needed for the calculation of Bayes Factors, which could be used in more sophisticated Bayesian techniques to answer questions such as, what are the posterior odds on the source being at position A compared to it being at position B [14].

Figure 5. The relative posterior probability density of Equation 3 is plotted in the central region of the simulated area. The number of independent measurements increases by 40 in each panel. At 80 measurements a peak near the correct location is apparent and at 160 there are two clear peaks at (50, 50) and (0, 100). The ridges in the probability reflect the fact that the method can only give information roughly along lines joining the sensors.

To explore the method, we performed Monte Carlo simulations, based on our experiences at the Otway and Ginninderra test sites. Figure 4 illustrates the concept: we simulate four concentration sensors, at the corners of a square 800m on a side. A source is located 50m from the centre of the square in each co-ordinate. Each measurement was then created by

1) drawing a M-O length from the coarsely-binned histogram of Otway data, and thus allocating a plume shape for that measurement. We used only 5 bins to simplify the number of plumes. Each plume used was simulated in WindTrax [13] and the results were fitted to simple analytical forms for faster computation in the simulations. The relative probability of the various plumes obeyed the binned-up Otway statistics.

2) We then drew a wind direction from the Otway wind rose. Note that each successive "measurement" represents a separate drawing from the wind rose. Of course the stability conditions may change over this time but might well be similar during several successive daytime periods, for example.

3) Given the wind direction, plume shape, and source strength, we calculate a concentration value at each sensor, adding Gaussian random noise to take account of measurement error and error resulting from imperfect background correction. We took a value of o = 1.5 ppm which we believe is easily achievable.

4) The source strength was fixed at a constant leak rate of 400 t C02 per year.

This set of synthetic data can then be inverted using the Bayesian method we have described. Inputs to the method are the plume shape, wind direction and measurement noise level, as well as the synthetic data. We used both a Jeffreys and log-normal prior on source strength, with similar results. The spatial prior was a Gaussian of standard deviation 100 m, centered at the origin.

The results of a typical simulation are shown in Figure 5. While a small number of "hits" is sufficient to locate the source accurately, a large number of data are required with a sparse array because many plumes miss the sensors altogether. Thus the main determinant of the location accuracy is the number of wind directions that can be used. Knowing weather patterns might allow planned surveys to minimize the time required to obtain a range of wind directions.

Implementation and Future Work

We have identified two key areas for investigation to achieve a useful implementation of this concept. One is to check whether the downwind-upwind background subtraction works in real terrain and ecosystems, so that this noise source averages away over the long periods we envisage for measurements. The second, related question is to discover if networks can be constructed using individual sensors (e.g. GMP343, Vaisala, Helsinki, Finland). If we are looking for small differential signals over long periods then to exploit a large number of sensors in a network we will have to develop a reliable self-calibration method, to establish and maintain a common calibration scale across the network. To explore these issues we are currently running an experiment at our Ginninderra site, where in collaboration with Geoscience Australia a controlled release site is being constructed. We have deployed 8 GMP343 sensors in a circle of 40 m diameter around a release point. These sensors are backed up by the University of Wollongong FTIR system [15] in which air from each of our 8 sensor locations is sampled sequentially and pumped to a central point for analysis. The array has been collecting data for several weeks with no C02 release taking place. The preliminary data indicate that, at this site and over these small distances, the downwind-upwind C02 signal is essentially white noise out to timescales of many days. This suggests that background subtraction in an array can be very effective. Preliminary results from a continuing small N20 release (where the background is stable) are encouraging.

If relatively large numbers of sensors can be deployed in arrays at reasonable cost and with stable calibrations, the next step is to examine if these arrays could be deployed in an adaptive fashion. The essential idea is that if we believe there is a leak somewhere in a well-defined area and we search for it without success in some sub-area, then the probability of it being elsewhere must be increased. The location of the next search would be dictated by the position of the largest, updated probability. This is a Bayesian idea, as Bayes' theorem gives the formalism for the update just described. Bayesian search theory is a well-developed area [17], particularly in search and rescue. It was famously employed in the location in the deep Atlantic of the wreck of the US nuclear-armed submarine Scorpion in 1968. The theory can be used to derive optimal search strategies subject to constraints such as cost or time.


Atmospheric monitoring is a mature and well-understood area of science. It offers the possibility of both quantification and localization of leaks from sequestration sites. There are advantages in focusing on localization as prior information can be effectively used, and large areas can be searched. 0ur modeling work, based closely on actual field experience, suggests that the time scale for a search is naturally quite long, because of the need to sample a range of wind directions to triangulate sources. As long as sensors are stable, it follows that noise levels on the measurement are not a determining factor in the accuracy of localization. Bayesian methods offer an effective method of inverting data to obtain a source position. If mobile and stable arrays of concentration sensors can be

proven, intelligent and adaptive search methods could close in on a leak much more rapidly than random or grid-based searches.


We acknowledge the funding provided by the Commonwealth of Australia through the CRC Programme to support CO2CRC research.


[1] Leuning, R., Etheridge, D. M., Luhar, A., and Dunse, B. (2008). Atmospheric monitoring and verification technologies for CO2 geosequestration. International Journal of Greenhouse Gas Control, 2000; 2: 401-14.

[2] Lewicki, J. L., Hilley, G. E., and Oldenburg, C. M. An Improved Strategy to Detect CO2 Leakage for Verification of Geologic Carbon Sequestration. Geophysical Research Letters, 2005; 32: L19403,

doi: 10.1029/2005GL024281

[3] Oldenburg, C. M., Lewicki, J. L., and Hepple, R P. Near-surface monitoring strategies for geologic carbon dioxide storage verification. . Report LBNL-54089, Lawrence Berkeley National Laboratory, 2003.

[4] Kaimal, J. C. and Finnigan, J. J.. Atmospheric Boundary Layer Flows: Their Structure and Management. Oxford University Press, New York, 1994.

[5] Mahrt, L. Stratified Atmospheric Boundary Layers. Boundary-Layer Meteorology, 1999; 90: 375-396.

[6] Sharma, S., Cook, P.J., Berly, T., Anderson, C., Australia's first geosequestration demonstration project - the CO2CRC Otway Basin Pilot Project. APPEA Journal, 2007; 47: 257-268.

[7] Luhar A., Etheridge, D., Leuning, R., Steele P., Spencer D, Hurley P., et al., in 19th International Clean Air and Environment Conference, Clean Air Society of Australia and New Zealand, 2009. Available at

[8] Etheridge et al, these Proceedings.

[9] Hirst, B., Gibson, G., Gillespie, S., Archibald, I., Podlaha, O., Skeldon, K. D., Courtial, J., Monk, S., and Padgett, M. Oil and Gas Prospecting by Ultra-Sensitive Optical Gas Detection With Inverse Gas Dispersion Modelling. Geophysical Research Letters, 2004; 31; L12115, doi:10.1029/2004GL019678.

[10] Da Costa, G. A. and Steele, L. P. (1999). A low-flow analyser system for making measurements of atmospheric CO2. In: Report of the ninth WMO Meeting of Experts on Carbon Dioxide Concentration and Related Tracer Measurement Techniques WMO/TD no. 952 (ed. Francey, R. J.), 16-20

[11] Francey, R. J. and Steele L. P. Measuring atmospheric carbon dioxide - the calibration challenge. Accredit. Qual. Assure. 2003; 8: 200-204[12] Loh, Z., Leuning, R., Zegelin, S., Etheridge, D., Bai, M., Naylor, T., and Griffith, D. ,Testing Lagrangian Atmospheric Dispersion Modelling to Monitor CO2 and CH4 Leakage From Geosequestration, Atmospheric Environment, 2009; 43 , 2602-2611.

[13] Bernardo,C., Vries,, Permanent shallow subsoil CO2 flux chambers for monitoring of onshore CO2 geological storage sites. Int. J. Greenhouse Gas Control, 2010; doi:10.1016/j.ijggc.2010.05.011

[14] WindTrax software, Thunder Beach, Scientific, Nanaimo, Canada

[15] Sivia, D., Skilling, J., Data Analysis: A Bayesian Tutorial, Oxford, 2006.

[16] Griffith, D. W. T. (1996). FT-IR Measurements of Atmospheric Trace Gases and their Fluxes, in Handbook of Vibrational Spectroscopy, John M. Chalmers and Peter R. Griffiths (Editors), Wiley 2002.

[17] Stone, L.D., The Process of Search Planning: Current Approaches and Continuing Problems, Operations Research, 1983; 31: 207-233