close

Вход

Забыли?

вход по аккаунту

?

j.ijggc.2018.06.014

код для вставкиСкачать
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Contents lists available at ScienceDirect
International Journal of Greenhouse Gas Control
journal homepage: www.elsevier.com/locate/ijggc
Toward an adaptive monitoring design for leakage risk – Closing the loop of
monitoring and modeling
T
⁎
Ya-Mei Yanga,c, , Robert M. Dilmorea, Grant S. Bromhala, Mitchell J. Smallb
a
National Energy Technology Laboratory, United States
Carnegie Mellon University, United States
c
Oak Ridge Institute for Science and Education, United States
b
A R T I C LE I N FO
A B S T R A C T
Keywords:
Geologic carbon storage
Leakage risk
Event tree
Risk management
Wellbore monitoring uncertainty quantification
Detection probability
Monitoring is a key component of risk management at geologic carbon storage (GCS) sites, serving both to help
operators understand and manage site performance, and to assure the public and other stakeholders that effective containment is maintained and impacts avoided. Potential leakage of CO2 and/or brine through wellbores, faults, and fractures to potable groundwater resources is a primary risk concern at onshore GCS sites. In
this paper, we present an adaptive methodology for leakage risk-based monitoring design. The methodology uses
a risk event tree to predict the likelihood of leakage occurrence, with detection probabilities of risk events
estimated for multiple monitoring plans. The overall detection probability of a proposed monitoring plan incorporates baseline data, stochastically simulated leakage events, and the likelihood that a set of technologies
will detect the changes in baseline conditions induced by the simulated leakage events. The adaptive monitoring
design methodology is demonstrated with a representative case study of CO2 and brine leaking from a well to a
potable groundwater aquifer using simulated data at the High Plains aquifer in the United States. Groundwater
quality parameters, pH, total dissolved solids and benzene concentrations, were used to calculate the corresponding detection probabilities of conventional groundwater sampling and fixed sensor monitoring for selected
leakage scenarios. The overall detection probability considering all monitoring information was then calculated
to evaluate proposed monitoring plan designs. Finally, a simple optimization problem to maximize detection
probability with constrained monitoring resources was presented as an application example to close the loop of
monitoring and modeling.
1. Introduction
Geologic carbon storage (GCS) – wherein large volumes of anthropogenic CO2 are captured from industrial and power generation
sources, and stored in geological formations – is one of several technically viable climate-change mitigation strategies (IPCC, 2005). Stakeholders, however, including the general public, have expressed concerns about potential environmental impacts associated with large-scale
implementation of GCS. Chief among these concerns is that leakage of
CO2 and displaced brine through natural (e.g., faults and fractures) and
engineered (e.g., wells) leakage pathways could result in contamination
of potable groundwater sources at onshore GCS sites.
Monitoring is an important part of GCS and is included in the risk
assessment, management and communication frameworks recommended by several government agencies and authoritative organizations (IEAGHG, 2009; Harbert et al., 2016). Monitoring and mitigation actions are essential to confirming site performance (Pawar et al.,
⁎
2015), and in practice GCS monitoring plans must be able to demonstrate that the GCS activities meet the regulation requirements
(Hovorka, 2012). Several types of monitoring frameworks have been
proposed to address these management needs in the past decade. Seto
and McRae (2011) proposed a model-based framework for integrated
monitoring design for risk reduction at GCS sites, formulating the
monitoring design problem in a systematic approach in order to consider many factors, such as operational decisions, footprint size, and
uncertainty in monitoring strategies. Yang et al. (2012) proposed a
general framework of Bayesian belief network (BBN) for CO2 leak detection at GCS sites, using leakage simulations and baseline data, and
the BBN framework was illustrated using simulations and measurements from the ZERO Emissions Research Technology (ZERT) study
site. The recent Quest project also used a storage management framework, which includes a measurement, monitoring and verification
(MMV) plan to assess the storage risks and evaluate the monitoring
performance in an adaptive manner to ensure conformance and
Corresponding author at: U.S. Department of Energy National Energy Technology Laboratory, 626 Cochrans Mill Road, P.O. Box 10940, Pittsburgh, PA, 15236-0940, United States.
E-mail addresses: ya-mei.yang@netl.doe.gov, yameiy3104@outlook.com (Y.-M. Yang).
https://doi.org/10.1016/j.ijggc.2018.06.014
Received 25 October 2017; Received in revised form 16 May 2018
Available online 07 July 2018
1750-5836/ © 2018 Elsevier Ltd. All rights reserved.
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
gain information of greatest value, thus forming a monitoring-modeling
loop. This manuscript is organized with a focus on explaining the
general risk-based monitoring design methodology in Section 2.1, followed by a prototype demonstration of that methodology using wellbore leakage simulations for the High Plains aquifer in Section 2.2. The
results of the High Plains aquifer case study are presented and discussed
in Section 3. The last section summarizes the methodology, key points
and potential applications in the future.
demonstrate containment of CO2 storage throughout the project (Shell
Canada, 2012; Bourne et al., 2014; Dean and Tucker, 2017).
Various monitoring technologies have been evaluated using leakage
simulations to understand detectability and spatial resolutions given
different leakage scenarios, such as near-surface soil flux and tracer
measurement (Yang et al., 2012), pressure monitoring at different
depths (Sun et al., 2013a; Wang and Small, 2014; Azzolina et al., 2014;
Keating et al., 2014; Trainor-Guitton et al., 2016; Namhata et al., 2017)
and shallow groundwater monitoring (Dai et al., 2014; Keating et al.,
2014; Yang et al., 2015). To assist monitoring network design, Yang
et al. (2015) used monitoring effectiveness (ME) of a shallow groundwater wellbore leakage event, similar to the overall detection probability used by Keating et al. (2014) for stochastic simulations, to
evaluate monitoring plans for different groundwater quality indicators.
Both studies concluded that groundwater monitoring may not be effective in terms of leakage detection time. Moreover, Sun et al. (2013a)
considered wellbore leakage risk and performed optimization of a
pressure-based monitoring network design for GCS sites, using a binary
integer programming formulation.
The National Risk Assessment Partnership (NRAP) is a collaborative
effort of researchers from contributing U.S. Department of Energy
National Laboratories. The objective of NRAP is to develop tools and
methods to understand and quantify the long-term environmental risk
associated with potential leakage through wells and faults, in the context of system-wide uncertainty. To date, NRAP’s approach to leakage
characterization involves forward modeling of important system components (reservoir, seals, leakage pathways, and overlying receptors of
concern) using physics-based numerical simulations, building reduced
order models that provide representative performance indicators with
significantly less computational burden over the range of the uncertain
parameter space, and integrating those models into a comprehensive
characterization of site-scale whole system performance. NRAP’s
Integrated Assessment Model for Carbon Storage (NRAP-IAM-CS) is one
important tool for stochastic prediction of system-wide leakage performance at carbon storage sites (Pawar et al., 2013, 2016). This tool
builds upon the previous CO2 storage integrated assessment model CO2PENS by Stauffer et al. (2009) and colleagues.
Another aspect of the NRAP effort is consideration of how to
monitor system performance to ensure storage effectiveness, reduce
uncertainty in system behavior, and inform decisions about monitoring
design and mitigation alternatives. Foundational work on the topic of
strategic monitoring has included establishing monitoring detection
thresholds based on interpretation of baseline data (Last et al., 2013),
considering the state-of-art detection resolution of various monitoring
technologies (Harbert et al., 2016), developing optimization methods
and tools to identify optimal monitoring location for a leakage pathway
of known location (Shang and Huang, 2012; Gastelum and Porter,
2014; Yonkofski et al., 2016), and initial evaluation of effective detection sensitivity of different monitoring technologies as a function of
site specific heterogeneities and uncertainties (Yang et al., 2017; Carroll
et al., 2017). In the context of those previous and ongoing efforts, this
study seeks to establish a comprehensive risk-based monitoring assessment methodology that effectively synthesizes site baseline data
and modeling of monitoring responses based on forward modeling results to assess site performance and inform decisions about future
monitoring and mitigation.
The purpose of this paper is to provide a risk assessment-based
framework for adaptive monitoring design at geologic carbon storage
(GCS) sites in an attempt to close the loop of monitoring and modeling.
In this context monitoring consists of site baseline information and
monitoring data to better understand current site conditions and processes under various modes of operation and management. Modeling
involves prediction of reservoir behavior, potential risk event scenarios
and the associated monitoring responses for the risk events. Monitoring
can provide data to construct, update and calibrate the models, while
modeling can guide the ongoing adaptation of monitoring designs to
2. Methods
Key aspects of this methodology are that it (i) incorporates sitespecific monitoring background data for detection thresholds and
monitoring design, (ii) utilizes a leakage event tree for risk analysis, (iii)
uses stochastic simulations to characterize the uncertainty of leakage
risk events and their effects on monitored parameters in the surrounding environment, (iv) applies statistical methods to combine the
results from multiple monitoring techniques for evaluating and optimizing monitoring plans, and (v) accommodates implementation in the
integrated assessment model (IAM) framework of NRAP to close the
monitoring-modeling loop. The methodology uses a leakage event tree
concept as the basis to calculate the combined detection probability of a
monitoring plan that uses various monitoring technologies given a
leakage event at a known or unknown location (Fig. 1). This methodology is illustrated with a case representative of the High Plains
aquifer (Carroll et al., 2014), considering the stochastic wellbore
leakage events simulated using the NUFT reactive transport simulation
code (Nitao, 1998; Hao et al., 2012). The resulting groundwater quality
changes are reflected in three groundwater monitoring parameters (pH,
total dissolved solids (TDS) and benzene concentrations), which are
then used to calculate the corresponding monitoring responses and the
detection probabilities for each simulation, based on the background
distributions and the selected thresholds. The resulting detection
probabilities are then used to evaluate the monitoring well density and
the proposed network design for both known and unknown wellbore
leakage events. Moreover, the proposed monitoring network can be
optimized in an adaptive manner based on the plume or impact area
predictions as they evolve with time.
2.1. Risk-based monitoring design and evaluation methodology
The proposed risk-based monitoring design methodology is based
upon a leakage risk event tree for GCS activities (See Fig. 1). Risk event
trees provide a straightforward and practical structure for risk identification, forming the basis for probabilistic risk assessment and facilitating risk management and communication. In addition to providing
an inventory of leakage risk events (pathways in Fig. 1), leakage event
trees can also be used to quantify the weights of potential leakage risk
scenarios as a function of the probabilities along the tree branches. The
leakage risk event tree starts with the occurrence of a leakage event and
whether the leakage event is known or unknown. The tree then subdivides into types of potential leakage pathways and specific leakage
locations, and connects to how likely it is that this leakage event will be
detected by the monitoring network design. Using leakage occurrence
and detection probability, the monitoring design can then be focused on
detecting more likely risk events and on providing information needed
to enable a more effective mitigation response. For evaluating low
probability - high consequence events (Oldenburg and Budnitz, 2016),
the event tree structure can be further extended with utility or preference functions that consider impacts on human health, ecosystem
damages, social disruption or other outcomes resulting from routine site
operations or accidents. Although the event tree used for the case study
might be simple, it can be augmented with a more detailed fault tree for
site specific risk assessment and management.
The methodology presented herein incorporates a simulation-based
systems approach that bridges the event tree method for risk analysis
126
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
Fig. 1. Leakage risk event tree and detection (blue lines are specific to this case study, dotted red circles are detection probabilities used to evaluate proposed
monitoring plans, and grey lines refers to other possible scenarios allowed for in the proposed methodology, but not implemented in the case study) (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article).
then fed into several subsequent process elements: threshold determination for leakage detection, leakage pathway identification and prior
probability estimation, leakage event simulations and the associated
modeling of monitoring responses, and scoping of the monitoring plan.
The first main process element is leakage detection threshold determination based on the baseline data, and the output is characterization of background distributions of the monitoring parameters, and
the suggested leakage detection thresholds after the background-databased thresholds are compared with the regulation thresholds and
monitoring method detection limits.
In the next main process element, leakage simulation results are
converted to detection probability results for selected monitoring
parameters and technologies, based on the leakage detection thresholds. Two steps are included in this process element: modeling of
monitoring responses, and probabilistic and statistical analysis.
Modeling of monitoring responses refers to the transformation of
leakage simulation results into the corresponding monitored responses
for the monitoring parameter and technology chosen. Then, the probabilistic and statistical analysis step takes the simulation output for
leakage-induced changes, combined with selected background distributions and thresholds, to calculate the detection probability for the
monitoring parameter as a function of site and leakage condition-specific parameters. In the context of this study, monitoring parameters
and monitoring technologies are treated as two distinct attributes, for
example, total dissolved solids (TDS) is a groundwater monitoring
parameter, and the monitoring technology for measuring TDS could be
either conventional purge sampling followed by lab gravimetric analysis or simply an electrical conductivity sensor. By following these
sequential analysis steps, from leakage simulation, monitoring response
and detection probability, the detectability of the simulated leakage
event can be estimated for the monitoring parameter and technology
selected.
After obtaining primary detection probabilities for the known
leakage event simulations, the detection probabilities for the unknown
(randomly simulated) leakage events can be estimated, and the analysis
of monitoring detection efficiency as a function of monitoring density
and environmental systems modeling, using simulation outcomes to
estimate the detection probability of a given monitoring network design. Here we advance a step beyond leakage simulations for risk assessment to consider leakage detection for risk management through
better monitoring design, explicitly using the risk event tree. Although
many other types of GCS risks, such as induced seismicity (White and
Foxall, 2016) or operational safety (Seto and McRae, 2011), can be
incorporated into the event tree structure, here we seek to demonstrate
this approach for a simple well leakage risk scenario.
The presented methodology is analogous, in some ways, to the
concept of Risk Priority Number (RPN) as described in failure mode and
effects analysis (FMEA) (Chen, 2007; Stamatis, 2003), which considers
severity, occurrence and detection. In this study, we aim at proposing a
general framework for integrating leakage detection monitoring with
leakage event modeling for risk management (see Fig. 2) and therefore
focus on illustrating how monitoring interacts with other management
elements, such as the event tree, risk scenario modeling and leakage
detection within the monitoring-modeling loop. We leave the assessment of severity of impact (or change in site utility) for future study.
As shown in Fig. 1, the two primary considerations are leakage
occurrence and leakage detection. Generally, the probabilities of
leakage occurrences in the event tree can be obtained from expert elicitation, site-specific or historical analogue data. The probabilities of
leakage detection, which depend on the monitored signals and the
specific monitoring technologies applied, are often obtained through
field measurements, lab tests or simulations. In this study, hypothetical
occurrence probabilities are assigned to the set of known and unknown
leakage pathways and locations, with the focus directed towards the
estimation and use of leakage detection probabilities that result from
alternative monitoring network designs.
The proposed risk-based monitoring assessment methodology is
summarized in a flowchart shown in Fig. 2, which includes the inputs
from site information and risk identification in the left column, the
analysis processes for risk-based monitoring design in the middle
column and the expected outputs for decision support in the right
column. The flowchart starts with baseline data collection, which is
127
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
Fig. 2. Risk-based monitoring network design flowchart. The case considered in this manuscript provides an example of that portion of the work flow marked by solid
lines.
can be performed. The first step in this process element is generalization of the monitoring response from the known leakage simulations
and/or probability estimations for application to the unknown random
leakage simulations. This generalization step may be comprised of
simple simulations using analytical solutions (Nordbotten et al., 2005)
or for those resulting in a symmetric steady seepage flux (Yang et al.,
2011a), or for comprehensive leakage simulations covering all domain
locations. In our study, the generalization step is made at the detection
probability level, where stochastic leakage detection probabilities of
single or multiple leakage sources are simplified by adoption of a box
approximation technique. These simple box representations are then
applied to estimate the spatially-varying probability of detection
around other leakage points. The second step of this process element
consists of analysis of monitoring spatial density and detection time
based on inference made from the unknown random leaks. Cartesian
monitoring grids of various sizes, representing different monitoring
densities, are evaluated to determine their average probabilities of
detection for the unknown random leakage scenarios, to compare the
general detectability of different monitoring parameters and techniques
chosen. These spatial and temporal analysis results can also serve as
constraints for a subsequent adaptive optimization process element.
The final step is to evaluate proposed monitoring plans that incorporate two or more monitoring technologies using the combined
detection probability over time for leakage events from both known and
unknown leakage pathways. Similarly, the same evaluation procedure
can also be performed in search of optimal monitoring designs given the
number of monitors within a necessary and workable domain at a given
time. Evaluating and updating monitoring designs in a time-phased
manner allows us to implement adaptive monitoring strategies with
updated monitoring data, to minimize the uncertainty of field monitoring and reservoir modeling. Moreover, if applicable, the adaptive
nature also allows the use of Bayesian methods to update each process
element in the flowchart (risk event probabilities, leakage detection
thresholds, monitoring models or reservoir simulations) (Jenkins,
2013). More details about the methodologies in Fig. 2 are described in
Section 2.1 and the applications of our case study using the wellbore
leakage simulations at the High Plains aquifer are explained in Section
2.2.
2.1.1. Baseline data and threshold determination
The proposed methodology accommodates consideration of
thresholds that serve different purposes related to the detectability of a
potentially unwanted change in the engineered natural system, and the
acceptability of that potential change relative to some established
standard (e.g., protective regulatory threshold or engineering performance threshold). Detectability thresholds, typically based on background data, are essential for leakage detection since they allow comparison with leakage-induced monitoring signal perturbation to make
inference about leakage occurrence. Previously, a probabilistic method
(Yang et al., 2011) was used to determine the likelihood of a leak based
on the detection probability that a measured monitoring signal is above
a critical threshold value, βpar-tech, given a background signal distribution, Bpar-tech, of a monitoring parameter and a monitoring technology
(together denoted as the par-tech). The threshold values, βpar-tech, can be
obtained from the background distribution, Bpar-tech, using prediction
intervals (Yang et al., 2011) or tolerance intervals (Last et al., 2013) in
order to find the anomalies indicating leakage. In terms of acceptability,
threshold values may also be derived from governing regulations, such
as the U.S. Environmental Protection Agency (EPA) maximum contaminant levels (MCLs) (U.S. EPA, 2016) or California MCLs (California
EPA, 2016), which are intended to protect the environment and public
health. Engineering performance thresholds are involved with natural
128
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
comprehensive uncertainty considered in IAM, although fixed cases can
also be used for non-stochastic simulations. Therefore, given a monitoring parameter and the technique used to measure that parameter,
(denoted as the par-tech), the probability of leakage detection,
(scenario)
)
PDL(,scenario
, of
par − tech, m , for a monitoring response scenario, MSL
leakage pathway, L, at a monitoring location, m(x,y), and time, t, would
be defined as the probability that a measured monitoring signal (i.e.,
the sum of the background signal, Bpar-tech,m, and the simulated leakageinduced signal change, MSL,par-tech,m) is above a critical threshold value,
βpar-tech,m. The PD for each measurement can be expressed as:
or practical measurement constraints, such as method detection limit or
operational range of a specific technology for a monitoring parameter.
Unlike the fixed regulatory-type thresholds, the background data-based
thresholds may be spatially or temporally varying (i.e., βpar-tech(x,y,z,t),
if significant spatial or temporal patterns are shown). For example, if
spatial data are sufficient, the 2D distribution mapping based on
background measurements can be made using kriging simulations.
Moreover, if a monitoring parameter depends on other variables (e.g.,
temperature or redox conditions), its background distributions and the
data-based thresholds can subsequently be explained and modeled as a
function of these variables.
)
(scenario)
PDL(,scenario
par − tech, m = P (Bpar − tech, m + MSL, par − tech, m > βpar − tech, m )
2.1.2. Leakage simulation and modeling of monitoring response
2.1.2.1. Leakage simulation. To better quantify the overall leakage
scenario uncertainty, stochastic leakage events within a scenario in a
natural system, SL(n)(x,y,z,t|Φ,RL) were simulated, where SL(n) refers to
the nth realization of simulated leakage events for leakage pathway, L,
at location, (x,y,z), and time, t, given geologic parameter set, Φ, such as
permeability and porosity, and leakage rate profile, RL, for CO2 and
brine. Here it is assumed that a unique leakage event can occur with
only one type of leakage pathway so that the simulation results or
reduced order models of different of leakage pathways, like wellbore,
single large fault and small fault zone can be combined in the IAM
structure for systematic evaluation. Moreover, although many
realizations are generated for the purpose of parameter uncertainty
quantification and statistical analysis, fixed case simulations (e.g., using
high, median, and low permeability instead of a permeability
distribution), can also be used in a simpler scenario analysis.
(1)
The sum is calculated directly for parameters with effects that are
additive, such as TDS and benzene in groundwater. For pH the background and leakage effects may be additive in terms of the hydrogen
ion concentration. Alternatively, a more complex chemical mixing
model may be needed to compute the change in the background concentration when the simulated leakage signal arrives, as in our case
study.
)
Next, the PDL(,scenario
par − tech, m for multiple monitoring parameters and
technologies are combined through their joint probability distribution
)
to obtain a combined detection probability, PDL(,scenario
com − par , m . If information
is insufficient to establish a covariance matrix for the joint probability
distribution, as in the case study presented herein, the assumption of
conditional independence between the parameters can be made to estimate the combined probability of leakage detection. This monitoring
data-based probability of detection calculation offers a link to field
monitoring data and potential use of more complicated models of
monitoring data for leakage detection. Thus, it goes beyond conventional methods based on a site-wide binary detection probability that
determines detectability based on measured values of a single parameter at a single location being above or below a specified detection
threshold. Accounting for background distributions of monitoring
parameters not only allows estimation of the probability of detection –
with quantification of false positive and false negative rates of a monitoring network – but also allows for combination of the results from
multiple monitoring parameters to enhance the detection diagnosis. As
a CO2 injection project matures, more monitoring data will be collected
to observe the evolution of the CO2 plume, and to look for evidence of
leakage occurrence in situ, as well as conformance (Jenkins et al.,
2015).
2.1.2.2. Modeling of monitoring response. Modeling of monitoring
responses links the simulated leakage event (leakage-induced
changes) in a natural system to the observed measurements of
monitoring parameters obtained by the technologies used. Some
monitoring parameters, such as CO2 flux (Yang et al., 2012) or
pressure measurement (Sun et al., 2013b), can be viewed as near
direct representations of the natural system disposition with respect to
the monitored parameter because relative little monitoring data
inversion is required (for example, compared to seismic survey). In
this case, it is appropriate to directly use the leakage simulation results
with added measurement errors or noises for monitoring design
application. While other monitoring technologies, such as 3D seismic
survey, vertical seismic profile (VSP), are indirect observations of the
natural system in terms of CO2 leakage with more complicated
sampling and post-processing processes. Therefore, the stochastic
leakage events in a natural system, SL(n)(x,y,z,t|Φ,RL), would be
converted
into
the
corresponding
monitored
signals,
MSL(n)(x,y,z,t|Φ,RL), through either their functional relationships or
appropriate calculation procedures.
2.1.4. Spatial and temporal monitoring resolution analysis
2.1.4.1. Generalization from the known to the unknown. To account for
all potential leakage scenarios within a spatial domain of interest (at the
CO2 storage site scale), it is necessary to account for the possibility of
unwanted fluid migration through pathways (e.g., wells, faults) that are
not identified before injection is initiated (unknown leakage pathways).
In the proposed methodology, leakage from those unknown pathways is
assumed to behave similarly to, and generate a similar response to, that
from comparable known leakage pathways. Details of this conceptual
design, as applied to known and unknown wells, are provided below;
extension to fault leakage would proceed similarly.
In the case of known wellbore leaks, all known wellbore locations in
the model domain are considered as potential leakage locations, (i.e. L
(known-well, xi, yi) where i = 1, …, NK, and NK is the number of known
existing well). Each known wellbore location was assigned a weight (or
wellbore score), w(i), to represent the relative probability of leakage
potential, which was then used to estimate the occurrence probability
at that wellbore. We assume only one leakage event can occur at one
location, and some wells are more likely to become leakage pathways
due to poor wellbore casing or geological fragility. In practice, the
leakage occurrence probability at a known wellbore could be estimated
from wellbore ranking by site experts or be integrated with wellbore
scores or wellbore quality data (Azzolina et al., 2015; Dilmore et al.,
2.1.3. Probabilistic and statistical analysis for risk scenarios
The probabilities of leakage detection (PD) for the monitoring response scenarios were then computed using a modified version of the
methodology developed by Yang et al (2011a), which assumed that the
distance between the leak source and the nearest monitor is used to
determine the detected leakage signal and its probability of detection
by the nearest monitor. In this study, our methodology improvements
are made to: (1) to clearly differentiate transport modeling and modeling of monitoring response steps, and use the modeling of monitoring
response outputs, which are closer to what would be observed in reality, for detection probability calculation and monitoring network design; and (2) to use the probability of detection by the monitor with the
greatest monitoring response, instead of simply the nearest monitor,
since the simulation incorporates spatial variability that yields irregular, anisotropic concentration fields and monitoring signals. Moreover, the methodology is expanded to use the summarized monitoring
response scenarios from the stochastic simulations given more
129
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
2.1.5. Monitoring plan evaluation and optimization
2.1.5.1. Proposed monitoring plan evaluation for both known and unknown
leaks. Finally, a specific proposed monitoring plan can be evaluated
based on the total PD for both known and unknown leakage pathways,
considering the leakage pathway probabilities of the risk event tree.
After we obtained the overall expected estimated detection
(scenario)
probabilities of known wellbore leakage, PDknown
− well, par − tech, mplan , and
2015; Dobossy et al., 2011; Watson and Bachu, 2009). Next, the detection probability for each known wellbore (a potential leakage event)
is calculated given a proposed monitoring location plan, mplan, for
selected monitoring parameters and technologies, par-tech, or combined
detection, com-par. Similar to the approach described previously by
Yang et al (Yang et al., 2011a), we use the two-dimension relative
distance relationship between the leakage location and the monitors,
along with the approximated PD zones, to determine in which detection
probability zone the nearest monitor with maximum response is located, and then return the approximated detection probability by the
monitor. The overall expected detection probability of the proposed
monitoring location plan for a known wellbore leak is then calculated
as the weighted average of the product of the leakage scenario occurrence probability and the detection probability of each known well for
that scenario. In other words, the overall expected estimated detection
probability of a known wellbore leakage event with multiple possible
scenarios is expressed as:
(scenario)
unknown random wellbore leakage, PDunknown
− well, par − tech, mplan , the two
detection probabilities are then further combined through the
multiplication with their prior leakage pathway probabilities (i.e.,
p°known-well and p°unknown-well) to get the final estimated total wellbore
(scenario)
leakage PD, PDwellbore
, par − tech, mplan , given the selected monitoring
parameters and technologies, as expressed in the Eq. (4) below.
∼ (scenario)
(scenario)
0
PDwellbore, par − tech, mplan = pknown
− well * PDknown − well, par − tech, mplan
(scenario)
0
+ punknown
− well * PDunknown − well, par − tech, mplan
The final total PD calculation can also be made using the PDs of
combined detection for known and unknown wellbore leakage pathways to obtain the final total PD of combined detection,
∼ (scenario)
PDwellbore, com − par , mplan , to evaluate the monitoring plan.
(scenario)
PDknown
− well, par − tech, mplan
=
1
NK
∑i = 1 w (i)
NK
∼ (scenario)
*∑i = 1 w (i)* PD L (known well, i), par − tech, mplan
(2)
For unknown wellbore leaks, ten thousand random leak locations
were generated based on an assumed uniform-probability distribution
across the simulation domain; that is, L(unknown-well, xi, yi) where
i = 1, …, NS, and the random sample number NS = 10,000. The generalized PD approximation zones are applied to these random leak locations to represent random wellbore leak events, and the occurrence of
each random leak is assumed equally likely. Similarly, the detection
probability for each unknown random leak is calculated given the
proposed monitoring location plan, mplan, for selected monitoring
parameters and technologies, par-tech, or combined detection, com-par.
Then the overall expected detection probability of the proposed monitoring location plan for unknown random wellbore leak is calculated
by summing the product of the leakage occurrence probability and the
detection probability of each unknown random leak. In other words, the
overall expected estimated detection probability of unknown wellbore
leakage is expressed as:
(scenario)
PDunknown
− well, par − tech, mplan =
(4)
2.1.5.2. Risk-based monitoring design optimization. With the evaluation
procedure of a proposed monitoring location plan established, it can be
applied to any proposed monitoring location plans, allowing selection
of the best plan based on the detection probability and given other
constraints on meeting that objective, and therefore turning this
procedure into a straightforward optimization problem. For a simple
maximum detection probability or maximum coverage problem, the
detection probability would be the evaluation objective, and the
monitoring location plan with the maximum detection probability (or
minimum of its transformation in practice) would be recommended,
given the constraints of time (e.g., monitoring result feedback time and
regulation compliance time), cost (e.g., budget limit and numbers of
monitors) and/or spatial feasibility (e.g., minimum monitoring spacing
and/or a feasible or mandatory monitoring region). To illustrate the
application of optimization in this study, a simple maximum detection
probability/coverage problem example is formulated. We used the final
total detection probability as the evaluation objective and treated the
time as adaptive monitoring stages and the numbers of monitors
representing a cost constraint. A simple monitoring example is
constructed with an initial monitoring location at the injection well
to search for optimal locations for four additional monitoring wells at
different points in time, using the PD of combined diagnosis. This
simple optimization example, derived from Nowak’s expression (Nowak
et al., 2015; Bode et al., 2016), can be formulated as:
NS ∼ (scenario)
1
*∑ PDL (unknown well, i), par − tech, mplan
i=1
NS
(3)
2.1.4.2. Spatial and temporal monitoring density analysis. The same
procedure used for unknown random leaks can also be applied to
assess the monitoring density requirement and the corresponding
detection time, as demonstrated previously for soil flux monitoring
(Yang et al., 2012). That previous work considered Cartesian
monitoring grids of different spacing distances, and then computed
expected detection probabilities for different monitoring grid densities
given one thousand random leakage locations. This allowed estimation
of the monitoring distance requirement for the monitoring parameter
and monitoring technique considered. In the methodology presented
here, the resulting expected PDs of various monitoring densities are
calculated according to the method of Yang et al. (2012), and used to
compare the general detectability of different monitoring parameters
and techniques chosen. The expected PD for each monitoring parameter
can also be integrated through the joint probability to obtain a
combined detection probability. Since both the simulated leakage
plume and the probability of leak detection evolve with time, we can
also obtain the estimated earliest detection time given a minimum
requirement of PD values and a selected monitoring network density.
This general analysis of spatial and temporal monitoring resolution
performed without knowing any specific monitoring locations, allows
consideration of the trade-off between spatial density and detection
time in the planning phase of a GCS project.
(scenario)
(scenario)
dopt
= argmin(f detect
)
d∈D
(5)
∼ (scenario)
(scenario)
where f detect
= 1−PDwellbore , the non-detection probability (nonPD), D represents the space of decision variables (such as monitoring
technologies, costs, time and spatial feasibility, etc.), and dopt represents
the suggested optimal solutions of decision variables. It is noted that the
GCS optimization problem formulation can include any factors
mentioned above and more and use more complicated geophysical
and geochemical model simulations and more detailed management
models, as shown in recent GCS optimization publication (Sun et al.,
2013a; Gastelum and Porter, 2014; Nowak et al., 2015; Bode et al.,
2016; Yonkofski et al., 2016). In this paper our focus is to show where
the monitoring optimization is positioned in the risk-based monitoring
design methodology and how it can be achieved using a simple example
problem.
2.1.6. Closing the monitoring-modeling loop
From Section 2.1.1–2.1.5, we described the use of risk event tree,
monitoring data and leakage and monitoring response simulations to
130
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
method detection limits do not provide information useful for inference
of the background distributions. If the data can be found in other similar sites, Bayesian hierarchical modeling using information from
other similar sites may be considered (Yang et al., 2011a,b).
perform risk-based monitoring design. These descriptions were in the
context of forward design thinking. To form the monitoring-modeling
loop, monitoring data have to feed back to update and calibrate the
reservoir and risk scenario models, and then the updated reservoir and
risk scenario model predictions can guide the monitoring designs in the
next time step. The use of monitoring data to update forward models is
technically an inverse problem for model calibration and parameter
estimation (Sun and Sun, 2015), and the most well-known case of GCS
is the Sleipner project (Singh et al., 2010). As shown in Fig. 2, monitoring data can update not only the reservoir and risk scenario simulations but also the knowledge about the natural systems of the site, the
engineered systems of GCS operations and the occurrence of risk events.
Although describing the methods of model calibration and update is not
in the scope of this paper, we try to illustrate the other direction of the
loop – model prediction guided monitoring design with a simple
adaptive monitoring design optimization example given a no-leakage
assumption. It is noted that monitoring design for model calibrations
(establishing conformance) and monitoring design for leakage detection (verifying containment) and risk management are different. The
data collected for risk management might not be sufficient for model
calibration, and vice versa.
2.2.2. Leakage simulation and modeling of monitoring response
The leakage simulations used for the case were taken from a larger
set of stochastic simulations of the High Plains aquifer used in building
a reduced order model (ROM) of aquifer impacts resulting from wellbore CO2 and brine leakage, as part of the NRAP effort (Carroll et al.,
2014; Mansoor et al., 2014). Stochastic realizations of leakage from a
wellbore and into a heterogeneous representation of the High Plains
aquifer were simulated using NUFT (Nitao, 1998; Hao et al., 2012), a
numerical code for simulating flow and reactive transport in porous
media. Uncertain parameters considered in those stochastic representations include permeabilities in sand and clay, sodium molality,
trace metal and organic constituent molality as well as CO2 and brine
leakage rates. For this study, the groundwater quality changes resulting
from fluid leakage into the groundwater aquifer were represented using
a subset of the simulation output parameters: pH, TDS and benzene
concentrations, which were then used to calculate the detection probabilities. There were 697 successful realizations generated for the High
Plains aquifer ROM, and the first 100 realizations were used to perform
the case study analysis described herein. The simulation domain size is
10 km long (x-axis aligns with the direction of regional groundwater
flow), 5 km wide (y-direction) and 240 m deep (z-direction) with a
single wellbore leakage source located at (x, y) = (2000 m, 2500 m), at
a depth of 240 m. The CO2 leakage rates range from 0.001 to 0.5 kg/s,
while the brine leakage rates range from 0.005 to 0.075 kg/s. Both CO2
and brine leakage periods are between 1 to 200 years, and the
groundwater quality simulation time lasts for 200 years. For more details about the stochastic simulations, please see the original report by
Carroll et al. (Carroll et al., 2014), the related publications (Carroll
et al., 2016; Mansoor et al., 2014) and the supporting information.
To illustrate the modeling of monitoring response process, here we
considered two types of groundwater monitoring techniques: conventional groundwater purging sampling and lab analysis vs. fixed borehole sensor measurement. Conventional groundwater purging and lab
analysis procedures (U.S. EPA, 2013) use the groundwater sampled
directly from the observation wells, and the water samples are analyzed
either on site (like pH, TDS or temperature) or in the laboratory (like
TDS, benzene, Cd and As). The water sampling and laboratory analysis
procedure normally requires days to weeks to obtain the results, which
usually represent the average concentrations of the sampled water volume, but the analysis precision is generally better. Another type of
measurement is obtained using fixed borehole sensors, which take insitu measurements and return (or save) the observed values quickly,
such as pH or TDS sensors which measure electrical conductivity as
approximation. However, borehole sensors are often less accurate, and
not all monitoring parameters can be measured by borehole sensors.
Consequently, for conventional groundwater sampling, the modeling of
monitoring response at the wellbore location (x,y) simply corresponds
to calculating the average of simulated concentrations over the saturated depth, z, representing complete mixing of the water sample. For
fixed borehole sensor monitoring, measurements correspond to simulated values at a fixed depth where the sensor is located (in the illustration that follows the sensor is located near the wellbore bottom at a
fixed depth z = 232.9 m. For measurement errors, a normal distribution
with a mean of zero is assumed for all monitoring parameters and
added to the simulated monitoring responses. The standard deviation of
pH error distribution is assumed to be 0.05 pH units (Leito et al., 2002;
Mettler Toledo, 2018) for both monitoring techniques. A relative
standard deviation (i.e. CV) of 2% is assumed for TDS using purging
sampling and lab analysis (Carter et al., 1976; Gilmore and Luong,
2016) and for TDS using sensor approximation method (Hayashi, 2004;
Feng et al., 2014), given the negligible amount of organic constituents
2.2. Prototype methodology demonstration: the High Plains aquifer leakage
case study
In this study, a case representative of the High Plains aquifer wellbore leakage is used to illustrate the methodology described in the
Fig. 2 flowchart, considering leakage-induced changes in three
groundwater monitoring parameters (pH, TDS and benzene concentrations) with two monitoring technologies (conventional purging
sampling and lab analysis vs. fixed borehole sensor). Hypothetical prior
probabilities of the event tree, wellbore scores and the predicted reservoir impact areas are used for this case study.
2.2.1. Baseline data and threshold determination
The background data for pH and TDS were obtained from a study for
no-impact thresholds by Last et al. (2013), while the background benzene data were drawn from a USGS groundwater quality report for the
central High Plains aquifer (Litke, 2001). The background pH and TDS
data were sufficient for characterizing general one dimensional (1D)
background distributions and the data-based tolerance interval
thresholds, recommended for showing no groundwater impact and used
as leakage detection thresholds in this case study. However, the background benzene data were insufficient to allow direct development of
such a distribution because most benzene concentrations were below
reported detection limit (0.5 μg/L) at that time (Litke, 2001). The total
percentage of observations above the detection limit (2%) (Litke, 2001)
was used as the basis to infer the background benzene distribution,
using this percentage as a fixed condition with varying coefficient of
variation (CV) values for an assumed lognormal distribution. Details of
the background distributions are provided in the supporting information.
The leakage detection thresholds for pH, TDS and benzene concentrations are obtained using a tolerance interval (95% confidence and
95% coverage) approach (Last et al., 2013). In the case of pH, the
suggested threshold is the lower bound of the 95% confidence and 95%
coverage tolerance interval, while, for TDS, it is the upper bound of the
95% confidence and 95% coverage tolerance interval. Due to limited
data, the several inferred benzene thresholds were estimated using the
same tolerance interval (95% confidence and 95% coverage) approach
with hypothetical lognormal distributions of several CV values (0.01,
0.1, 1 and 10) given the fixed 98% level at 0.5 μg/L (Litke, 2001). In
such cases where concentrations are below method detectability and
measured data are few, the analytical method detection limit could be
considered as a substitute background threshold, with a degenerate
distribution often assumed for computational purpose. However, the
131
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
in the simulations. A relative standard deviation of 3% is assumed for
benzene using purging sampling and lab analysis (Aguilera-Herrador
et al., 2008; Marotta, 2011). While the complete mixing assumption
might somewhat underestimated the actual monitoring due to the use
of a whole vertical wet length, the fixed level measurement value might
be overestimated due to the neglect of calibration issues. Nevertheless,
the intended contrast of the two simple modeling processes for the two
different groundwater monitoring technologies is also indicative of the
uncertainty in monitoring procedure selection.
well = 0.7 and p°unknown-well = 0.3). For the unknown wellbore leakage
pathway, every unknown location is assumed to be equally likely
(randomly sampled), while the known potential wellbore leakage locations are assigned arbitrary weights for calculating the leakage
probabilities at these locations, as mentioned. To demonstrate how
monitoring plans are evaluated, three example monitoring plans, mplan, of different monitoring well numbers, are presented: (1) all thirtythree available well locations within the simulation domain, (2) sixteen
well locations and (3) eight well locations selected from the thirty-two
available locations, in addition to the injection well used in the original
simulations, i.e. (1) 33 + 1 wells, (2) 16 + 1 wells and (3) 8 + 1 wells.
The three proposed monitoring plans were then evaluated using the
high, median and low monitoring scenarios to show their performance
given the simulation uncertainty. In this evaluation, we showed the
typical trade-off between monitoring density requirement and the early
detection time, and thus propose an adaptive/phased manner in monitoring design. The evaluation results are summarized in the result
section below.
Since the proposed monitoring plans can be evaluated through PD
and detection time, the monitoring plan can be optimized accordingly.
In this case study, a simple adaptive single-objective optimization example is illustrated, which minimizes the non-PD (i.e., maximize the
PD) with design constraints of available monitoring parameters and
technologies (three monitoring parameters and two technologies),
numbers of monitoring wells (the original injection well plus additional
three and four monitoring wells), spatial feasibility (limit to existing
well locations and 100 m minimum spacing between wells) and the
time between monitoring activities (five and ten years). In this example,
most decision variables are treated as fixed cases or constraints, and the
overall response uncertainty is binned into low, median and high scenarios, which allows this example to be treated as a simple integer
programing problem of location selection using the complicated system
modeling results. Here we used the R genetic algorism package,
“genalg” (Willighagen, 2005), to solve the problem for the solution
space, and one example solution is presented in the result section.
2.2.3. Probabilistic and statistical analysis of risk scenarios
In our study, the full set of simulation realizations, SL(n), where
n = 1, …, N, and the corresponding modeled monitored signals, MSL(n),
were summarized into the high (95%), median (50%) and low (5%)
response scenarios, i.e. MSL(high, 95%), MSL(median, 50%) and MSL(low, 5%)
for the three monitoring parameters (pH, TDS and benzene concentrations) and the two groundwater monitoring techniques (conventional groundwater sampling and lab analysis vs. fixed borehole
sensors). These response scenarios were used in monitoring network
design based on probabilities of leakage detection (PD) for each monitoring parameter and its associated monitoring techniques. The PDs for
the monitoring response scenarios were then computed by the methodology modified from Yang et al. (2011a), as detailed in Section 2.1,
and the PD results of selected response scenarios are shown in the result
section.
2.2.4. Spatial and temporal monitoring resolution analysis
The detection probability results estimated from these response simulations were generalized to estimate the effects of leakage occurring
at other well locations within the same domain, assuming that the
leakage plume sizes depend more on the wellbore leakage rates than on
the specific locations. This generalization is made at the detection
probability step, where simplified PD estimation boxes for detection are
drawn around the locations with the detection probability greater than
0.9 and 0.5 for the monitoring parameters and techniques used, and the
PD estimation boxes evolve over time as the leakage plume spreads. The
PD estimation boundaries are provisionally set to 0.9 and 0.5 so that the
whole domain is divided into three zones: the detection zone, the intermediate zone and a non-detection zone with reassigned detection
probabilities approximated as 1, 0.5 and 0, respectively. (For more
detailed explanations about the PD estimation boxes, please see the
supporting information). In this study, the estimated PD estimation
boxes were also calculated based on these summarized low, median and
high response scenarios given selected monitoring parameters and
technologies, and then applied to both known or unknown potential
wellbore leakage locations. Therefore, given a monitoring parameter
and the monitoring technique used to detect that parameter, we can
)
obtain the estimated probability of leakage detection, PDL(scenario
, par − tech, m ,
for a monitoring response scenario of leakage pathway, L, at monitoring
location, m.
As described in Section 2.1, the monitoring grids of different sizes
along with unknown random leaks were used to estimate the general
spatial and temporal monitoring resolutions. Here the sizes of
groundwater monitoring grids range from 20 m, 50 m, 100 m, 200 m,
…, 1000 m, 1500 m, and 2000 m were tested, and the results of the high
response scenarios of pH, TDS, benzene monitoring and the combined
are shown in Section 3, and the results of the low and the high scenarios
are in the supporting information.
3. Case study results
This section presents results based on application of selected aspects
of the proposed methodology detailed in Section 2.1, as applied to the
case study described in Section 2.2. Implications of those results to the
specific case and to the general applicability for CO2 storage site
monitoring and decision support are also briefly discussed after the
presented results in each subsection.
3.1. Background distribution and suggested thresholds
The inferred background distributions of pH, TDS and benzene
concentrations as well as the background data-based and regulation
thresholds are summarized in Fig.3. As seen in the Fig.3, compared to
the no-impact pH threshold (7.02), the EPA pH threshold (6.5) is too far
from the background distribution to be effective in detecting small leaks
(although the false positive rate using the EPA pH threshold would be
nearly zero). In contrast, the EPA TDS threshold (500 mg/L), which is
pretty close to the background mean, is masked by the background
variation, leading to a 37% false positive rate. The inferred possible
benzene leakage detection thresholds were estimated using the same
tolerance interval approach with hypothetical lognormal distributions
of several CV values, given the fixed 98% level at 0.5 μg/L (Litke,
2001). The resulting no-impact threshold values are 0.50, 0.48, 0.37
and 0.22 for CV = 0.01, 0.1, 1 and 10, respectively.
2.2.5. Monitoring plan evaluation and optimization
In this study, we used the PDs of combined detection for different
monitoring scenarios to evaluate the proposed monitoring plan since
combined diagnosis may provide better indication of leakage than
single parameter detection. The prior occurrence probability of known
and unknown potential wellbore leakage pathways are, for illustrative
purposed, assumed to be 0.7 and 0.3 in the example (i.e., p°known-
3.2. Modeling of monitoring and probabilistic analysis for leakage risk
scenarios
The summarized low (5%), median (50%) and high (95%)
132
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
Fig. 3. High Plains aquifer background cumulative distributions and thresholds of (a) pH, (b) TDS and (c) benzene concentrations. (pH is resampled from normal
distribution, where μ = 7.42, σ = 0.12; TDS is resampled from log10-normal distribution, where μ = 2.64, σ = 0.19 (Last et al., 2013); benzene concentrations are
inferred from the percentage above reported method detection limit (Litke, 2001). Please see the supporting information).
of the monitoring technology selected. In the median scenarios, the
leakage-induced monitoring response changes for benzene concentration were not detectable in the year 20 result. This is probably because
benzene concentration decreases within a few years of leakage in response to the combined effect of natural attenuation and dilution
(Lawrence, 2006; Bruce et al., 2010), causing the benzene concentration to decrease below the observable threshold by year 20. Among the
three monitoring parameters, pH is much more responsive than TDS
and benzene concentration. Given the simulation context, pH is more
responsive to CO2 leakage, and TDS is more associated with brine
leakage. This is because the brine leakage rate profiles in modeled
scenarios are several times smaller than CO2 leakage rate profiles
(please see the supporting information), we can expect greater pH
monitoring responses here.
To illustrate the effectiveness of the two monitoring technologies,
the PDs of the high leakage response scenarios in the year 20 are shown
in Fig. 5. As seen in Fig. 5, due to averaging effect (complete mixing of
water sample), the pH and TDS detectability results by purging method
are less significant compared to those by the fixed sensor placed at an
advantageous location close to the leak point. Therefore, the leakage
detection made by purging method indicates that the actual max point
monitoring response scenarios of the two monitoring techniques for the
pH, TDS and benzene concentrations using the original leakage simulations in the year 20 are shown in Fig. 4. This figure shows that the
simulated monitoring observations are lower in concentration for
conventional groundwater sampling technique than for a fixed sensor
sampling technique for all three parameters considered. The results
represent the difference between the vertical mixing effect of conventional groundwater sampling and the wellbore bottom sensors which
report a localized monitoring response, and illustrate the uncertainty
that can be introduced by chosen monitoring technology/method. The
contrast in response of the two monitoring techniques might be higher
than expected due to applied modeling of monitoring process, plus the
selection of near well bottom sensor location, which is close to simulated leakage source though quiet intuitive in application. It is also
noted that the results of fixed wellbore sensor for benzene are notional,
since currently available sensor technology for benzene measurement is
not reliable, and the method detection limit (0.03 μg/L) threshold refers
to a limit of laboratory analysis that is well beyond the state of art for in
situ benzene sensor technology (Adhikari et al., 2016).
Leakage-induced monitoring parameter changes were generally
difficult to identify in all low monitoring response scenarios, regardless
133
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
Fig. 4. Modeling of monitoring response scenarios with uncertainty using conventional groundwater purging sampling method and fixed borehole sensor sampling:
low (5%), median (50%) and high (95%) scenarios for (a) pH, (b) TDS and (c) benzene concentrations, in the 20th year after initiation of leak.
threshold (Carroll et al., 2016), i.e. benzene analytical method detection limit, illustrates this type of thinking – detect any possible change
with high priority even it could be just a false alarm.
Moreover, the probability gradients of pH and TDS are small, suggesting relatively less variable background distributions. To be able to
signal would be bigger than the observed average. To increase leakage
detectability for purging method, precautionary principle could be
applied using tighter threshold values, such as 10% or even 25% of the
estimated distribution of pH, despite the trade-off of higher false positive rates. Similarly, the use of recommended benzene no-impact
134
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
Fig. 5. Probability of detection for pH, TDS and benzene concentrations using the monitoring response scenarios with tolerance interval thresholds and modeled
monitoring responses for two monitoring techniques (top: conventional groundwater purging sampling & bottom: fixed sensor at well depth 232.5 m). *Tolerance
interval thresholds use 95% confidence and 95% coverage upper tolerance intervals, which are also the no-impact thresholds for pH and TDS. The recommended noimpact threshold for benzene is the method detection limit (Last et al., 2013; Carroll et al., 2016).
detect benzene, a less variable background distribution assumption is
also preferred, and thus the distribution with CV = 0.01 is used. With
that distribution, the same 95% confidence and 95% coverage tolerance
interval threshold, 0.5 μg/L, is applied to estimate the benzene PD and
compare with the PDs of pH and TDS. Additionally, the recommended
no-impact threshold (i.e., method detection limit, 0.03 μg/L) is also
used to estimate benzene PD. The contour line of PD equivalent to 0.9
estimated by method detection limit is also added in Fig. 5 for comparison. This figure shows that benzene concentration is less indicative
of leakage than either pH or TDS, though it is a very important parameter for groundwater quality protection. On the other hand, the presence of benzene could enhance the leakage evidence provided by TDS
and pH. The interaction between thresholds and detectability of benzene concentrations can be found in the supporting information.
monitoring has a similar grid spacing range of 20 m–100 m with 2-year
minimum detection time. Benzene groundwater monitoring, aiming
more at water resource safety than at leakage detection, has a similar
grid spacing range 20 m–100 m with a minimum detection time of
1 year, but couldn’t cover large area due to degradation. Therefore, the
combined detection is dominated by pH monitoring, and that suggests a
minimum well spacing constraint of 100 m for the median scenario. The
PDs for the median monitoring response scenario (median leakage
level) are generally lower for all parameters with a grid spacing range
of 20 m to approximately 200 m with 1 year minimum detection time
(not shown here, in the supporting information), while the PDs for low
response scenario (low leakage level) are much lower for all parameters
with a grid spacing range up to 100 m and a minimum detection time
more than 10 years (please see the supporting information).
3.3. Spatial and temporal resolution analysis – groundwater monitoring well
example
3.4. Monitoring plan evaluation and adaptive optimization
After selecting the monitoring parameters and leakage detection
thresholds, the next step in the workflow is to evaluate monitoring
location designs. Here the three proposed monitoring plans, (1) all
thirty-three available wells, (2) sixteen selected wells and (3) eight
selected in addition to the original injection well, were evaluated using
the high, median and low monitoring response scenarios to show the
leakage detection performance. The monitoring well locations and the
evaluation results for the scenarios considered are shown in Fig. 7 and
Table 1, following the same monitoring techniques and detection
threshold assumptions used in the previous section. The first monitoring design is optimistic and conservative, in that it uses all available
wells to maximize the probability of observing leakage from known
wellbores (assuming the monitoring signal is sufficiently strong to detect). Table 1 shows that, even in this conservative case, the PDs did not
initially reach 1, because the event tree prescribes a 30% chance of
unknown random leak occurrence. This conservative all-well
The estimated expected PDs of various monitoring densities using
random leaks provide an unbiased way to understand the overall detectability of different monitoring parameters and techniques at the
selected thresholds. Fig. 6 shows one example of the expected PDs of
pH, TDS, benzene concentration and the combined PD, given the high
monitoring response scenario, with the selected monitoring technique
and threshold for each monitoring parameter. The estimated PDs
change with time in response to the evolution and dissipation of the
leakage plume. Given the leakage simulation scenario considered, pH
shows the best monitoring resolution in both space and time for the
median response scenario, followed by TDS and benzene. Using 90%
averaged PD as the coverage criterion for resolution (shown in the blue
lines in Fig. 6), the pH fixed senor monitoring exhibits grid spacing
range from 20 m to approximately 100 m with 1 year minimum detection time (also the minimum simulation timestep). TDS fixed sensor
135
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
Fig. 6. Spatial and temporal monitoring density evaluation for random leaks for (a) pH, (b) TDS (c) benzene monitoring and (d) combined diagnosis, given the high
response scenario using probability of detection (PD) with selected thresholds.
wellbore locations (Yang et al., 2017) for maximizing spreadness. The
PDs of the second plan are lower than those in the conservative case, as
expected, with roughly half the PDs of the conservative case in the
beginning. The same roughly half reduction in PDs also happens from
the second plan to the third plan. This less-conservative, more realistic
monitoring design can potentially still detect leakage under the high
monitoring design could be more likely to detect leakage for the median
and high leakage response scenarios, but is less successful at detecting
leakage for the low leakage response scenario after year ten. The second
monitoring network design uses only half of the available wells, with
the monitoring locations based on selection of one well from each of
sixteen clusters from a k-means clustering analysis of the existing
Fig. 7. Combined PD evaluation for three proposed monitoring location plans using existing well location: (1) optimistic all available wells (2) half of the available
wells and (3) one fourth of the available wells, considering both known and unknown leakage sources.
136
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
Table 1
Combined PD evaluation for the proposed monitoring plans, considering both known and unknown leakage sources, using low, median and high leakage monitoring
response scenarios. The bold figures indicate the PD values equal to or greater than 0.5.
Scenarios
Monitoring Plan - 1
Monitoring Plan - 2
Monitoring Plan - 3
Year
Low
Median
High
Low
Median
High
Low
Median
High
1
2
3
4
5
10
20
30
40
50
100
0.00
0.70
0.70
0.00
0.70
0.70
0.00
0.00
0.00
0.00
0.00
0.70
0.70
0.70
0.70
0.70
0.71
0.71
0.71
0.71
0.71
0.72
0.70
0.71
0.71
0.72
0.72
0.77
0.85
0.89
0.92
0.95
0.92
0.00
0.37
0.37
0.00
0.37
0.37
0.00
0.00
0.00
0.00
0.00
0.37
0.37
0.37
0.37
0.37
0.37
0.37
0.37
0.37
0.37
0.39
0.37
0.37
0.37
0.41
0.41
0.45
0.64
0.71
0.75
0.78
0.79
0.00
0.20
0.20
0.00
0.20
0.20
0.00
0.00
0.00
0.00
0.00
0.20
0.20
0.20
0.20
0.20
0.20
0.20
0.20
0.20
0.20
0.21
0.20
0.20
0.20
0.21
0.21
0.26
0.40
0.50
0.54
0.59
0.55
the data collected. Despite the fact that the example coverage goal using
groundwater monitoring plan alone is unrealistic, the coverage issue
could be easily solved by adding other large domain monitoring technologies like 3D seismic survey or surface deformation. In addition to
groundwater monitoring, monitoring applications that can target different depths (e.g., well-based pressure observations) and larger areal
domains (e.g., 3D seismic surveys, surface deformation measurements)
could provide indirect yet full site wide monitoring to compensate the
limited domain covered by groundwater monitoring wells.
response scenario, where the signals are significant, and the accumulative leakage results in a larger impact domain over time.
In Table 1, the two spikes at the early stage of the low monitoring
response scenario using leakage simulations indicate the leakage detection contributed by two different monitoring parameters. The first
spike comes from pH detection, and the second spike comes from
benzene detection, illustrating the importance of combined diagnosis of
multiple parameters. Finally, if we set up a coverage target, for example
PD > = 0.5 for a proposed monitoring plan, then the earliest detection
time can be obtained. In this case, the earliest detection time of the
optimistic all-well plan would be the first year for the median and high
monitoring response scenarios, while the earliest detection time for the
half-well plan would be after the 10th year, and for the one-fourth-well
plan would be after the 20th year for the high response scenario. Although these analyses are helpful, they are static with respect to time.
In the adaptive nature of the monitoring-modeling loop, the monitoring
plan evaluation (and optimization) would be performed through different time steps with updated information, as shown in Fig. 2 and illustrated in the next section.
Closing the loop of monitoring and modeling requires adoption of
an adaptive approach of risk assessment, monitoring evaluation and
optimization, which forms the dynamic relationships between system
components. To illustrate the concept of this sequence, a simple optimization problem in two time steps is formulated with the single-objective of maximizing the probability of detection, assuming no-leak
condition. The high leakage response scenario described in the previous
section was again used to illustrate the adaptive monitoring concept,
since it is easier to find meaningful optimization solutions with more
detection responses with the high scenario than the low scenario (i.e.,
we are not searching solutions on all zero space, where any solution
works, but that is not meaningful). The optimization monitoring was
performed between two time steps – year five and year ten, with an
initial monitoring well at the injection well. In the first time step, a
smaller search domain was used, representing a hypothetical impact or
review area, which could be obtained by reservoir simulations or
seismic surveys; in the second stage, the hypothetical search domain
was expanded to a larger domain, representing a larger area of potential
impact as the injected CO2 plume grows. The two time-step adaptive
optimization would be to search for optimal placement solutions for the
first three monitoring wells in the first stage, and then again to search
for location solutions for an additional four monitoring wells, given the
monitoring/simulation feedback (i.e., a larger domain in our case) in
the second stage. One example solution for the two stages along with
the domain setting is presented in Fig. 8, in an attempt to illustrate
adaptive monitoring design based on the feedback loop. This no-leak
example also shows the monitoring plan could evolve with other needs
in reality, for example, adding additional assurance monitoring for
public concern or cutting insensitive monitoring technologies based on
4. Discussion
In addition to threshold selection, the variability of the background
data and the magnitude of leakage-induced change are important factors affecting the detectability of a monitoring parameter (Osenberg
et al., 1994; Macmillan and Creelman, 2005). If the background
variability of a monitoring parameter is unacceptably large, it could
mask leakage-induced signals of interest, leading to low detection
sensitivity. In some cases, a portion of this background variability may
be attributed to underlying phenomena that can be modeled by underlying explanatory variables, and effectively improve detectability.
Another important factor is the relative magnitude of leakage-induced
change of a monitoring parameter, which is used to select a responsive
monitoring parameter. The magnitude of leakage-induced change can
be estimated by leakage simulation and modeling of monitoring response, field experiments, or a combination thereof.
Calculating the probability of detection at given monitoring locations considers the two factors (background variability and observed
signal magnitude), and this monitoring location-centered detection
probability can be easily updated at the operational stages (using real
field observations) as well as be used at the planning stage (using simulated monitoring responses). The use of background distribution to
estimate the PD allows the calculation of false positive and false negative rates, and thus enables application in probabilistic risk assessment methods, including not only the event tree discussed here, but
also simulation-based Bayesian belief networks, site performance utility
functions, etc. This is the primary advantage over the more conventional single threshold method, which estimates detection probability
using zero (below threshold) or one (above threshold) without use of
the background distributions.
No matter what monitoring technologies are used, both baseline
data and modeling of monitoring responses (i.e., forward modeling of
monitoring response based on the results of transport process modeling)
are essential to demonstrate the CO2 and brine plume conformance and
containment effectiveness, particularly the agreement between the reservoir or system model predictions and the monitoring observations
(Chadwick and Noy, 2015). Disagreement between the predicted and
the observed plume extent does not necessarily indicate increased risk
137
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
Fig. 8. Example two-stage monitoring design solution using the pH high scenario, given one initial monitoring well, three monitors at the first stage (year five) with a
smaller search domain and the additional four monitors at the second stage (year ten) with a larger search domain, based on hypothetical plume predictions.
site data are collected and stochastic models developed from these data
to characterize the spatial distribution of soil and rock properties such
as porosity and permeability. Even if the distribution type and its
parameters are known for different strata at a site (e.g., a lognormal
distribution with its parameters, along with the spatial correlation
function from a variogram or a similar model), the spatial distribution
of porosity and permeability would still be subject to aleatory uncertainty due to differences in the large number of possible random
fields that are consistent with the stochastic model, conditioned on a
limited number of imperfect measurements at the site. However, this
random aleatory uncertainty is just the first tier of the hierarchy. The
underlying model structure and parameters that describe the spatial
variability of porosity and permeability are themselves uncertain. This
scientific uncertainty is epistemic in nature, and it can be reduced with
improved science and models.
As the operations at a site evolve over time, the collection of more
site data will serve to reduce the aleatory uncertainty of the spatial
distribution of model parameters, for example, porosity and permeability, since these data will further constrain the particular realization
of these properties that is present at the site. These same data and the
associated model building, testing and updating that ensue also help to
reduce scientific (epistemic) uncertainty in appropriate process and
statistical models that support porosity and permeability estimation,
both at the site in question and for others. The assessment and simulation approaches proposed here are modeled after this hierarchical
approach to uncertainty, learning, and adaptive response, enabled by
closing the loop of monitoring and modeling.
This type of aleatory uncertainty in parameters has been commonly
seen in the recent GCS leakage simulations (Celia and Nordbotten,
2009; Kopp et al., 2010; Siirila-Woodburn et al., 2017; Sun et al.,
2013b; Dai et al., 2014; Carroll et al., 2014), including NRAP’s integrated assessment models (Pawar et al., 2016). Despite the importance of parameter uncertainty of simulations, a broader check of all
possible scenarios of potential risk events and the detectability of these
events is necessary in practice from the perspective of risk management.
That is what this methodology framework seeks to provide, using a risk
event tree to address the scenario uncertainty from the source and the
monitoring detection probabilities of risk events to include the uncertainties in the monitoring process at the field site. Of course, parameters also have epistemic uncertainty since our knowledge about
parameters is increasing through the project, and this epistemic uncertainty in parameters can be quantified using Bayesian inference
(Morgan et al., 1992; O’Hagan, 2004), which facilitates the adaptive
of leakage or a hazardous event. It does provide new information about
the geologic system and its response to CO2 injection that can be used as
feedback to adaptive monitoring design and to update risk assessment
and risk management (Shell Canada, 2010). For example, data from a
3D seismic survey can suggest presence of preferential flow paths that
can be incorporated to iteratively improve geologic characterization
used in reservoir simulation; the same information can also be used to
inform decisions about locating new monitoring wells.
In addition to the parameter uncertainty reflected by the stochastic
leakage simulations and the model scenario uncertainty of different
monitoring technologies, there are other sources of uncertainty in the
proposed framework. For the inputs from site information and risk
identification as shown in Fig. 2, scenario uncertainty exists, not only in
the simulations of possible geologic risk events and monitoring processes, but also in the identification of potential risk events through
expert knowledge and judgment. During a GCS project, our knowledge
of potential risk scenarios increases, from roughly knowing what events
could happen to better understanding of detailed processes of how
these events happen. This type of uncertainty, which is due to lack of
knowledge, can be reduced through gathering more information and
data, is usually called epistemic uncertainty or subjective scientific
uncertainty (Morgan et al., 1992; Winkler, 1996; O’Hagan, 2004). Similarly, model uncertainty exists in the models for risk event simulation, the processes that affect monitoring responses, probability models
used for parameter characterization, as well as various methods for
approximation, including expert elicitation. The model functional
structure and modeling approaches, such as software, code and algorithm selection or numerical implementation, would be adjusted and
refined, and thus the overall model uncertainty is usually reduced
through the project life cycle. One noted study in GCS model uncertainty is the Sim-SEQ project (Mukhopadhyay et al., 2015), which
attempted to understand and quantify model uncertainties for geologic
carbon sequestration through an international model comparison.
Moreover, the parameter uncertainty addressed via the stochastic simulations in our case study is usually regarded as aleatory uncertainty,
which exists in most data due to random variability and measurement
errors. (For more details about the analysis of this stochastic simulation
data, see the publications by Carroll et al. (2016) and Mansoor et al.
(2014) and the supporting information).
Simulation models for geophysical processes, similar to many environmental systems models, require a hierarchical set of judgments
and analyses to characterize variability (most often reflective of aleatory uncertainty) and uncertainty (most often epistemic). For example,
138
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
concordance to predicted behavior (Oldenburg, 2018), and allows
quantitative assessment of the relative effectiveness of different monitoring scenarios. Site-specific, quantitative, risk based approaches
present a more defensible means of considering tradeoffs between
monitoring design robustness and frugality than qualitative, heuristicsbased approaches. More work is needed, however, to develop and validate such comprehensive site-scale monitoring design and evaluation
structures and workflows.
part of the monitoring and modeling loop. For a detailed review of
potential sources of uncertainty in the proposed framework, please see
the supporting information.
Developing full physics models that consider the full range of geologic parameter data, model, and scenario uncertainty (including uncertainty in leakage pathway types and potential leakage locations)
would be extremely computationally expensive, to the point of being
impractical. Therefore, several strategies can be applied to reduce this
computational burden, such as prioritizing the model parameters to
consider only those that are most important; performing focused simulations at the locations most representative of important geological
properties, with greatest leakage likelihood, or presenting important
safety or environmental concerns; or simplifying complicated geologic
structure, physical and/or chemical interactions with reduced physics
models or reducing computational cost with reduced order models. Our
strategy is to use a set of focused and representative stochastic simulations of a single wellbore leakage event, and then to preserve the simulation uncertainties in summarized low, median and high monitoring response scenarios. Next we convert the monitoring response
results into probabilities of detection (PD) and perform the generalization step at the PD level. Through working on the PD results where
the values are between 0 and 1, the issue of fine-tuning full physics
models could be less concerned (e.g., simulated pH value 5.1 and 5.6 all
have the same PD value 1). Given the framework of IAMs, this generalization step can be regarded as a reduced order model of monitoring
responses in terms of leakage detectability.
In practice, the feedback to monitoring design could incorporate
information from sources other than physics-based systems simulations
and site monitoring. The information feedback loop can also take into
account information that can come from expert judgment (e.g., using
expert judgment to assess prior probabilities of leakage pathways and
locations to populate the event tree, as shown in the flowchart), and
other factors (e.g., cost, labor requirements, regulatory constraints, land
accessibility, and receptor vulnerability). The initial monitoring design,
then, provides a framework into which new information collected in the
first time step can be synthesized, and used to update the physical
systems simulations, the monitoring models and risk assessments in
subsequent time step. This, in turn, provides the basis for comparison
with new monitoring information used in adaptive monitoring design in
the next time step. On a higher level, the system of monitoring-modeling feedback can be treated as a systems dynamics model by incorporating utility models for decision making, such as the cost of
leakage impact consequences or the benefit of successful storage
(Oldenburg et al., 2011; Bielicki et al., 2014; Trainor-Guitton et al.,
2013).
Besides the detection probability for maximum coverage, other
variables can also be used as evaluation objectives, such as the
minimum time for detection (Gastelum and Porter, 2014; Yonkofski
et al., 2016; Nowak et al., 2015, 2015; Bode et al., 2016), minimum
operational cost (Nowak al., 2015; 2015; Bode et al., 2016), or
minimum environmental impacts or leakage volume (Sun et al., 2013a).
Multiple evaluation criteria may also be considered together using
multi-objective optimization methods (Sun et al., 2013a; Nowak et al.,
2015), including utility functions (Bode et al., 2016). When multiple
decision criteria are considered, and the site specific geospatial data are
sufficient to support spatially varying simulations, optimization of
plume monitoring problems can become more complicated and computationally intensive. Given this complexity, some geo-statistical
methodologies can be applied and appropriate optimization algorithms
can be studied (De Gruijter et al., 2006; Dhar and Datta, 2007; Prakash
and Datta, 2013; Helle and Pebesma, 2015).
Methodologies such as that presented herein offer a clear, objective
framework for quantitative assessment of the effectiveness of networks
of monitoring technologies applied over a large spatial domain. It allows both the ability to update monitoring design in response to new
observations of site response to storage, and assessment of system
5. Conclusions
A risk-based monitoring assessment methodology is proposed for an
adaptive monitoring design considering leakage risk, focused on integrating the risk assessment process into monitoring design workflow
and providing a framework to close the loop of monitoring and modeling. To demonstrate how this risk-based monitoring assessment assists
the monitoring design, a simple wellbore leakage risk event tree for
known and unknown wellbore leakage was built and demonstrated
using a set of stochastic groundwater wellbore leakage simulations representative of the High Plains aquifer. The probabilities of detection of
proposed monitoring designs for known and unknown wellbore leakage
were estimated by incorporating background information, detection
thresholds and simulation of the monitoring parameters and techniques
selected, given the summarized low, median and high monitoring response scenarios. The estimated detection probabilities were linked to a
probabilistic leakage risk event tree to complete the overall detection
probability estimation structure for the monitoring design.
Additionally, to evaluate the general detectability for selected monitoring parameters, monitoring techniques and selected thresholds,
monitoring spatial and temporal resolution analyses were performed
using detection probability for random leaks. With the High Plains
aquifer leakage simulations, groundwater pH is the dominant monitoring parameter in leakage detection, as compared with TDS and
benzene concentrations, given selected monitoring techniques. A combined probability of detection was calculated to incorporate results
from all monitoring parameters to benefit leakage detection diagnosis.
Finally, to illustrate the adaptive monitoring design optimization process, a simple adaptive optimization problem of maximum detection
probability for a hypothetical simulated impact domain with feedback
through two time steps was presented as a simple application example
to close the loop of monitoring and modeling.
The application of this risk-based monitoring assessment methodology can be expanded in many aspects. In addition to groundwater
monitoring shown here, this assessment methodology can be applied to
any monitoring parameters provided that monitoring responses can be
modeled as a function of simulated leakage event. The probabilistic
design of this methodology accommodates full risk assessment including calculation of false positive and negative rates and allows the
integration of various utility models (e.g., environmental impact, operational cost, etc.) to provide more comprehensive decision support.
Moreover, other leakage detection diagnosis techniques, such as machine learning methods, can be used rather than the basic signal and
detection threshold distribution approach presented here. Estimation of
leakage detection probability can be considered with real monitoring
data to evaluate CO2 plume conformance. Instead of the simple adaptive optimization example (single-objective over two time steps), multicriteria decision analysis and optimization with utility models can also
be performed, considering more decision variables with a more complicated feedback time schedule. Besides wellbore leakage, additional
leakage pathways can be considered, such as fault and cap rock, and
site specific expert survey of prior leakage pathway probability can be
performed to accommodate more comprehensive risk analysis. The
adaptive nature of this methodology allows for implementation of
Bayesian methods in leakage detection. The resolution and complexity
of monitoring models can be easily tuned to match the system scope
and resolution needed to support decision making. Additionally, the
139
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
Bielicki, J.M., Pollak, M.F., Fitts, J.P., Peters, C.A., Wilson, E.J., 2014. Causes and financial consequences of geologic CO2 storage reservoir leakage and interference with
other subsurface resources. Int. J. Greenh. Gas Control 20, 272–284.
Bode, F., Nowak, W., Loschko, M., 2016. Optimization for early-warning monitoring
networks in well catchments should be multi-objective, risk-prioritized and robust
against uncertainty. Transp. Porous Media 114 (2), 261–281.
Bourne, S., Crouch, S., Smith, M., 2014. A risk-based framework for measurement,
monitoring and verification of the Quest CCS Project, Alberta, Canada. Int. J. Greenh.
Gas Control 26, 109–126.
Bruce, L., Kolhatkar, A., Cuthbertson, J., 2010. Comparison of BTEX attenuation rates
under anaerobic conditions. Int. J. Soil Sediment Water 3 (2) Article 11.
California EPA, 2016. California Regulations Related to Drinking Water. Accessed April
2018. http://www.waterboards.ca.gov/drinking_water/certlic/drinkingwater/
documents/lawbook/dwregulations-2016-09-23.pdf.
Carroll, S.A., Keating, E., Mansoor, K., Dai, Z., Sun, Y., Trainor-Guitton, W., Brown, C.,
Bacon, D., 2014. Key factors for determining groundwater impacts due to leakage
from geologic carbon sequestration reservoirs. Int. J. Greenh. Gas Control 29,
153–168.
Carroll, S.A., Bianchi, M., Mansoor, K., Zheng, L., Sun, Y., Spycher, N., Birkholzer, J.,
2016. Reduced-Order Model for Estimating Impacts from CO2 Storage Leakage to
Alluvium Aquifers: Third-Generation Combined Physical and Chemical Processes,
NRAP-TRS-II-009-2016, NRAP Technical Report Series. U.S. Department of Energy,
National Energy Technology Laboratory, Morgantown, WV.
Carroll, S., Mansoor, K., Yang, X., Buscheck, T.A., Sun, Y., 2017. Framing monitoring
needs to detect leakage from Wells to the overburden. Energy Procedia 114,
3628–3635.
Carter, M.J., Huston, M., Logsdon II, O.J., 1976. Micromethods for the determination of
nonfilterable and filterable residues. Water Pollut. Control Fed. 48, 652–659.
Celia, M.A., Nordbotten, J.M., 2009. Practical modeling approaches for geological storage
of carbon dioxide. Groundwater 47 (5), 627–638.
Chadwick, R.A., Noy, D.J., 2015. Underground CO2 storage: demonstrating regulatory
conformance by convergence of history-matched modeled and observed CO2 plume
behavior using Sleipner time-lapse seismics. Greenh. Gases Sci. Technol. 5 (3),
305–322.
Chen, J.K., 2007. Utility priority number evaluation for FMEA. J. Fail. Anal. Prev. 7,
321–328.
Dai, Z., Keating, E., Bacon, D., Viswanathan, H., Stauffer, P., Jordan, A., Pawar, R., 2014.
Probabilistic evaluation of shallow groundwater resources at a hypothetical carbon
sequestration site. Sci. Rep. 4, 4006.
De Gruijter, J.J., Bierkens, M.F.P., Brus, D.J., Knotters, M., 2006. Sampling for Natural
Resource Monitoring. Springer Berlin Heidelberg, Berlin, Heidelberg.
Dean, M., Tucker, O., 2017. A risk-based framework for measurement, monitoring and
verification (MMV) of the goldeneye storage complex for the Peterhead CCS project,
UK. Int. J. Greenh. Gas Control 61, 1–15.
Dhar, A., Datta, B., 2007. Multiobjective design of dynamic monitoring networks for
detection of groundwater pollution. J. Water Resour. Plann. Manage. 133 (4),
329–338.
Dilmore, R.M., Sams, J.I., Glosser, D., Carter, K.M., Bain, D.J., 2015. Spatial and temporal
characteristics of historical oil and gas wells in Pennsylvania: implications for new
shale gas resources. Environ. Sci. Technol. 49 (20), 12015–12023.
Dobossy, M.E., Celia, M.A., Nordbotten, J.M., 2011. An efficient software framework for
performing industrial risk assessment of leakage for geological storage of CO2. Energy
Procedia 4, 4207–4214.
Feng, L., Zhang, W., Liang, D., Lee, J., 2014. Total dissolved solids estimation with a fiber
optic sensor of surface plasmon resonance. Opt. - Int. J. Light Electron Opt. 125,
3337–3343.
Gastelum, J.A., Porter, E.A., 2014. Optimizing monitoring designs under alternative objectives. Energy Procedia 63, 3516–3522.
Gilmore, K.R., Luong, H.V., 2016. Improved method for measuring total dissolved solids.
Anal. Lett. 49, 1772–1782.
Hao, Y., Sun, Y., Nitao, J., et al., 2012. Overview of NUFT: a versatile numerical model for
simulating flow and reactive transport in porous media. In: Zhang (Ed.), Groundwater
Reactive Transport Models. Bentham Science Publishers, pp. 213–240.
Harbert, W., Daley, T.M., Bromhal, G., Sullivan, C., Huang, L., 2016. Progress in monitoring strategies for risk reduction in geologic CO2 storage. Int. J. Greenh. Gas
Control 51, 260–275.
Hayashi, M., 2004. Temperature-electrical conductivity relation of water for environmental monitoring and geophysical data inversion. Environ. Monit. Assess. 96 (1–3),
119–128.
Helle, K.B., Pebesma, E., 2015. Optimising sampling designs for the maximum coverage
problem of plume detection. Spatial Stat. 13, 21–44.
Hovorka, S.D., 2012. In quest of robust and commercial CO2 monitoring. Greenh. Gases
Sci. Technol. 2 (3), 145–147.
IEAGHG, 2009. A Review of the International State of the Art in Risk Assessment
Guidelines and Proposed Terminology for Use in CO2 Geological Storage, Technical
Study, Report Number: 2009-TR7. IEA Greenhouse Gas R&D Programme.
IPCC, 2005. In: Metz, B., Davidson, O., de Coninck, H.C., Loos, M., Meyer, L.A. (Eds.),
Intergovernmental Panel on Climate Change Special Report on Carbon Capture and
Storage, Prepared by Working Group III of the Intergovernmental Panel on Climate
Change. Cambridge University Press, Cambridge, United Kingdom and New York,
NY, USA 442 pp.
Jenkins, C., 2013. Statistical aspects of monitoring and verification. Int. J. Greenh. Gas
Control 13, 215–229.
Jenkins, C., Chadwick, A., Hovorka, S.D., 2015. The state of the art in monitoring and
verification - Ten years on. Int. J. Greenh. Gas Control 40, 312–349.
Keating, E., Dai, Z., Dempsey, D., Pawar, R., 2014. Effective detection of CO2 leakage: a
classic event tree concept could be easily extended to incorporate other
types of risk and liability management problems, such as geologic
carbon storage with brine displacement for pressure management and
plume control.
Future work will focus on implementing this risk-based methodology into NRAP’s IAM framework to allow feedback to monitoring
network design and risk management decision support. More comprehensive system modeling including geologic carbon storage system with
more monitorable parameters/intervals and more detailed leakage
pathway representations will be tested using other real and synthetic
site data. Finally, the full risk assessment including false positive and
negative rates for more detailed monitoring design and the addition of
utility functions in terms of human and environmental aspects (in addition to operation cost) will be explored. Going forward, development
of a multi-criteria decision analysis tool linked to integrated assessment
model framework and application of this methodology to actual field
data will also be pursued.
Disclaimer
This paper was prepared as an account of work sponsored by an
agency of the United States Government. Neither the United States
Government nor any agency thereof, nor any of their employees, makes
any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that
its use would not infringe privately owned rights. Reference therein to
any specific commercial product, process, or service by trade name,
trademark, manufacturer, or otherwise does not necessarily constitute
or imply its endorsement, recommendation, or favoring by the United
States Government or any agency thereof. The views and opinions of
authors expressed therein do not necessarily state or reflect those of the
United States Government or any agency thereof.
Acknowledgements
This work was completed as part of the U.S. Department of Energy’s
National Risk Assessment Partnership. This paper contains work performed at the National Energy Technology Laboratory, with support by
Oak Ridge Institute for Science and Education, for the United States
Department of Energy. The authors thank Kayyum Mansoor, Susan
Carroll and Yunwei Sun at Lawrence Livermore National Laboratory for
generously sharing the leakage simulations. In particular, we thank
Susan and Kayyum for their guidance and explanations on the simulation results and detailed review for an earlier draft. In addition, we
want to thank Prof. Brian Junker at Carnegie Mellon University for
useful methodology discussions, Dr. Thomas Daley and Dr. Abdula
Cihan at Lawrence Berkeley National Laboratory and Dr. Joshua White
at Lawrence Livermore National Laboratory for their valuable suggestions. We also want to thank anonymous reviewers for their valuable
comments.
References
Adhikari, P., Bender, F., Josse, F., Ricco, A.J., 2016. Polymer-plasticizer coatings for shear
horizontal-surface acoustic wave sensors for long-term monitoring of BTEX analytes
in liquid-phase. In: 2016 IEEE International Frequency Control Symposium (IFCS).
IEEE. pp. 1–2.
Aguilera-Herrador, E., Lucena, R., Cárdenas, S., Valcárcel, M., 2008. Ionic liquid-based
single-drop microextraction/gas chromatographic/mass spectrometric determination
of benzene, toluene, ethylbenzene and xylene isomers in waters. J. Chromatogr. A
1201, 106–111.
Azzolina, N.A., Small, M.J., Nakles, D.V., Bromhal, G.S., 2014. Effectiveness of subsurface
pressure monitoring for brine leakage detection in an uncertain CO2 sequestration
system. Stoch. Environ. Res. Risk Assess. 28 (4), 895–909.
Azzolina, N.A., Small, M.J., Nakles, D.V., Glazewski, K.A., Peck, W.D., Gorecki, C.D.,
Bromhal, G.S., Dilmore, R.M., 2015. Quantifying the benefit of wellbore leakage
potential estimates for prioritizing long-term MVA well sampling at a CO2 storage
site. Environ. Sci. Technol. 49 (2), 37–42.
140
International Journal of Greenhouse Gas Control 76 (2018) 125–141
Y.-M. Yang et al.
uncertainty. Int. J. Greenh. Gas Control 52, 175–189.
Prakash, O., Datta, B., 2013. Sequential optimal monitoring network design and iterative
spatial estimation of pollutant concentration for identification of unknown groundwater pollution source locations. Environ. Monit. Assess. 185, 5611–5626.
Seto, C.J., McRae, G.J., 2011. Reducing risk in basin scale CO2 sequestration: a framework
for integrated monitoring design. Environ. Sci. Technol. 45, 845–859.
Shang, X., Huang, L., 2012. Optimal designs of time-lapse seismic surveys for monitoring
CO2 leakage through fault zones. Int. J. Greenh. Gas Control 10, 419–433.
Shell Canada, 2012. Shell Quest Carbon Capture and Storage Project: Measurement,
Monitoring and Verification Plan.
Siirila-Woodburn, E.R., Cihan, A., Birkholzer, J.T., 2017. A risk map methodology to
assess the spatial and temporal distribution of leakage into groundwater from geologic carbon storage. Int. J. Greenh. Gas Control 59, 99–109.
Singh, V., Cavanagh, A., Hansen, H., Nazarian, B., Iding, M., Ringrose, P., 2010. Reservoir
modeling of CO2 plume behavior calibrated against monitoring data from sleipner,
Norway. In: SPE Annu. Tech. Conf. Exhib. Florence, Italy, 19-22 September.
Stamatis, D.H., 2003. Failure Mode and Effect Analysis: FMEA from Theory to Execution.
ASQ Quality Press.
Stauffer, P.H., Viswanathan, H.S., Pawar, R.J., Guthrie, G.D., 2009. A system model for
geologic sequestration of carbon dioxide. Environ. Sci. Technol. 43 (3), 565–570.
Sun, N.-Z., Sun, A.Y., 2015. Model Calibration and Parameter Estimation: For
Environmental and Water Resource Systems. Springer.
Sun, A.Y., Nicot, J.P., Zhang, X., 2013a. Optimal design of pressure-based, leakage detection monitoring networks for geologic carbon sequestration repositories. Int. J.
Greenh. Gas Control 19, 251–261.
Sun, A.Y., Zeidouni, M., Nicot, J.P., Lu, Z., Zhang, D., 2013b. Assessing leakage detectability at geologic CO2 sequestration sites using the probabilistic collocation method.
Adv. Water Resour. 56, 49–60.
Trainor-Guitton, W.J., Ramirez, A., Yang, X., Mansoor, K., Sun, Y., Carroll, S., 2013. Value
of information methodology for assessing the ability of electrical resistivity to detect
CO2/brine leakage into a shallow aquifer. Int. J. Greenh. Gas Control 18, 101–113.
Trainor-Guitton, W.J., Sun, Y., Mansoor, K., Carroll, S., 2016. Merits of pressure and
geochemical data as indicators of CO2/brine leakage into a heterogeneous, sedimentary aquifer. Int. J. Greenh. Gas Control 52, 237–249.
U.S. EPA, 2013. SESDPROC-301-R3, Groundwater Sampling. Accessed May 2018.
https://www.epa.gov/sites/production/files/2015-06/documents/GroundwaterSampling.pdf.
U.S. EPA, 2016. Drinking Water Contaminants – Standards and Regulations. Accessed
May 2018. https://www.epa.gov/dwstandardsregulations.
Wang, Z., Small, M.J., 2014. A Bayesian approach to CO2 leakage detection at saline
sequestration sites using pressure measurements. Int. J. Greenh. Gas Control 30,
188–196.
Watson, T.L., Bachu, S., 2009. Evaluation of the potential for gas and CO2 leakage along
wellbores. SPE Drill. Complet. 24, 115–126.
White, J.A., Foxall, W., 2016. Assessing induced seismicity risk at CO2 storage projects:
recent progress and remaining challenges. Int. J. Greenh. Gas Control 49, 413–424.
Willighagen, E., 2005. Genalg: R Based Genetic Algorithm. R Package Version 0.1, 1.
Yang, Y.M., Small, M.J., Junker, B., Bromhal, G.S., Strazisar, B., Wells, A., 2011. Bayesian
hierarchical models for soil CO2 flux and leak detection at geologic sequestration
sites. Environ. Earth Sci. 64, 787–798.
Yang, Y.M., Small, M.J., Ogretim, E.O., Gray, D.D., Bromhal, G.S., Strazisar, B.R., Wells,
A.W., 2011a. Probabilistic design of a near-surface CO2 leak detection system.
Environ. Sci. Technol. 45, 6380–6387.
Yang, Y.M., Small, M.J., Ogretim, E.O., Gray, D.D., Wells, A.W., Bromhal, G.S., Strazisar,
B.R., 2012. A Bayesian belief network (BBN) for combining evidence from multiple
CO2 leak detection technologies. Greenh. Gases Sci. Technol. 2, 185–199.
Yang, C., Hovorka, S.D., Treviño, R.H., Delgado-Alonso, J., 2015. An integrated framework for assessing impacts of CO2 leakage on groundwater quality and monitoringnetwork efficiency: case study at a CO2-EOR site. Environ. Sci. Technol. 49,
8887–8898.
Yang, Y.M., Dilmore, R., Mansoor, K., Carroll, S., Bromhal, G., Small, M., 2017. Riskbased monitoring network design for geologic carbon storage sites. Energy Procedia
114, 4345–4356.
Yonkofski, C.M.R., Gastelum, J.A., Porter, E.A., Rodriguez, L.R., Bacon, D.H., Brown, C.F.,
2016. An optimization approach to design monitoring schemes for CO2 leakage detection. Int. J. Greenh. Gas Control 47, 233–239.
comparison of groundwater sampling and pressure monitoring. Energy Procedia 63,
4163–4171.
Kopp, A., Binning, P.J., Johannsen, K., Helmig, R., Class, H., 2010. A contribution to risk
analysis for leakage through abandoned wells in geological CO2 storage. Adv. Water
Resour. 33 (8), 867–879.
Last, G.V., Brown, C.F., Murray, C.J., Bacon, D.H., Qafoku, M.P., Sharma, M., Jordan,
P.D., 2013. Determining Threshold Values for Identification of Contamination
Predicted by Reduced Order Models. Pacific Northwest National Laboratory.
Lawrence, S.J., 2006. Description, Properties, and Degradation of Selected Volatile
Organic Compounds Detected in Ground Water — A Review of Selected Literature.
US Geological Survey, Atlanta, Georgia Open-File Report 2006-1338.
Leito, I., Strauss, L., Koort, E., Pihl, V., 2002. Estimation of uncertainty in routine pH
measurement. Accred. Qual. Assur. 7 (6), 242–249.
Litke, D.W., 2001. Historical Water-Quality Data for the High Plains Regional GroundWater Study Area in Colorado, Kansas, Nebraska, New Mexico, Oklahoma, South
Dakota, Texas, and Wyoming, 1930-98. U.S. Geological Survey Water–Resources
Investigations Report 00-5254.
Macmillan, N.A., Creelman, C.D., 2005. Detection Theory: A User’s Guide, 2nd ed.
Lawrence Erlbaum Associates Publishers.
Mansoor, K., Carroll, S.A., Sun, Y., 2014. The role of wellbore remediation on the evolution of groundwater quality from CO2 and brine leakage. Energy Procedia 63
(2014), 4799–4806.
Marotta, L., 2011. The Determination of Low Level Benzene, Toluene, Ethyl Benzene, and
Xylenes (BTEX) in Drinking Water by Headspace Trap GC/MS. PerkinElmer, Inc.,
Shelton, CT 06484 USA Accessed April 2018. https://www.perkinelmer.com/labsolutions/resources/docs/APP_HeadSpaceTrapGCMSLowBenzine.pdf.
Mettler Toledo, 2018. pH Buffer Accuracy – Comparison Study. Accessed April 2018.
https://www.mt.com/us/en/home/library/tips-and-tricks/lab-analyticalinstruments/comparison-pH-buffer-accuracy.html.
Morgan, M.G., Henrion, M., Small, M., 1992. Uncertainty: A Guide to Dealing With
Uncertainty in Quantitative Risk and Policy Analysis. Cambridge university press.
Namhata, A., Zhang, L., Dilmore, R.M., Oladyshkin, S., Nakles, D.V., 2017. Modeling
changes in pressure due to migration of fluids into the above zone monitoring interval
of a geologic carbon storage site. Int. J. Greenh. Gas Control 56, 30–42.
Nitao, J.J., 1998. User’s Manual for the USNT Module of the NUFT Code, Version 2 (NPPhase, NC-Component, Thermal); UCRL-MA-130653. Lawrence Livermore National
Laboratory.
Nordbotten, J.M., Celia, M.A., Bachu, S., 2005. Injection and storage of CO2 in deep saline
aquifers: analytical solution for CO2 plume evolution during injection. Transp.
Porous Media 58 (3), 339–360.
Nowak, W., Bode, F., Loschko, M., 2015. A multi-objective optimization concept for riskbased early-warning monitoring networks in well catchments. Procedia Environ. Sci.
25, 191–198.
O’Hagan, T., 2004. Dicing with the unknown. Significance 1, 132–133.
Oldenburg, C.M., 2018. Are we all in concordance with the meaning of the word conformance, and is our definition in conformity with standard definitions? Greenh.
Gases Sci. Technol. 8, 210–214.
Oldenburg, C.M., Budnitz, R.J., 2016. Low-Probability High-Consequence (LPHC) Failure
Events in Geologic Carbon Sequestration Pipelines and Wells: Framework for LPHC
Risk Assessment Incorporating Spatial Variability of Risk. Report Number: LBNL1006123.
Oldenburg, C.M., Jordan, P.D., Nicot, J.P., Mazzoldi, A., Gupta, A.K., Bryant, S.L., 2011.
Leakage risk assessment of the in Salah CO2 storage project: applying the certification
framework in a dynamic context. Energy Procedia 4, 4154–4161.
Osenberg, C.W., Schmitt, R.J., Holbrook, S.J., Abu-Saba, K.E., Flegal, A.R., 1994.
Detection of environmental impacts: natural variability, effect size, and power analysis. Ecol. Appl. 4, 16–30.
Pawar, R., Bromhal, G., Dilmore, R., Foxall, B., Jones, E., Oldenburg, C., Stauffer, P.H.,
Unwin, S., Guthrie, G., 2013. Quantification of risk profiles and impacts of uncertainties as part of US DOE’s national risk assessment partnership (NRAP). Energy
Procedia 37, 4765–4773.
Pawar, R.J., Bromhal, G.S., Carey, J.W., Foxall, W., Korre, A., Ringrose, P.S., Tucker, O.,
Watson, M.N., White, J.A., 2015. Recent advances in risk assessment and risk management of geologic CO2 storage. Int. J. Greenh. Gas Control 40, 292–311.
Pawar, R.J., Bromhal, G.S., Chu, S., Dilmore, R.M., Oldenburg, C.M., Stauffer, P.H.,
Zhang, Y., Guthrie, G.D., 2016. The national risk assessment partnership’s integrated
assessment model for carbon storage: a tool to support decision making amidst
141
Документ
Категория
Без категории
Просмотров
1
Размер файла
2 805 Кб
Теги
014, ijggc, 2018
1/--страниц
Пожаловаться на содержимое документа