Professional Documents
Culture Documents
StASaER MOO
StASaER MOO
1
The University of Tulsa
2
Shell International Exploration & Production Inc
Summary
Introduction
Large-scale deployment of CCUS projects is a crucial component in the transition to a decarbonized economy. In subsurface engineering
applications, CCUS involves the injection of CO2 into deep underground formations with structural closure, such as saline aquifers and/
or depleted oil/gas reservoirs. Not only carbon could be permanently geologically sequestered in said scenarios, but, during the injection
process, CO2 could also allow additional hydrocarbons to be recovered and marketed, resulting in more profit to be earned.
During the design of a CO2 injection process with specified well placement, coming up with a set of optimal well controls is important.
In such cases, life cycle production optimization, defined as the minimization or maximization of one or multiple production-driven
objective functions via changing the well controls during a reservoir’s lifetime (Brouwer et al. 2004; Jansen et al. 2005; Chen and
Reynolds 2016; Alpak et al. 2021), would play a critical component. One challenge arisen is to obtain the optimal cash flow (i.e., maxi-
mum life cycle NPV), while also trying and maintain the maximum CO2 storage, which could be quantified either in terms of mass, vol-
ume, or carbon tax credits (CTCs) (Jones and Sherlock 2021; Sun et al. 2021; Nguyen et al. 2023b). This concern in the decision-making
process gives rise to the need for a bi-objective optimization solution (Alpak et al. 2022; Wang et al. 2022; Nguyen et al. 2023b), to bal-
ance the conflicting criteria and objectives.
Many different algorithms are available for solving a general multi-objective optimization problem, and they could be categorized into
two families: derivative-free and gradient-based methods. The algorithms within the first class, such as genetic algorithm (Gen and Cheng
2000), particle swarm optimization algorithms (Kennedy and Eberhart 1995; Shi and Eberhart 1998; Bonyadi and Michalewicz 2017),
and bi-objective mesh adaptive direct search algorithm (Alpak et al. 2022), may avoid the local maxima/minima, given the appropriate
size of population and parameter settings or optimization. However, the computational cost of these derivative-free methods could easily
become infeasible as the dimension of the design vector exceeds a hundred, and/or if the cost of the evaluation of the objective function
is expensive, such as the requirement of running a complex numerical model. One example in petroleum engineering literature is the work
of Isebor et al. (2014), who presented a particle swarm optimization-based strategy to solve a bi-objective optimization problem for gen-
eral oilfield development. However, one of the biggest drawbacks of their method is the requirement on the order of 800,000 reservoir
simulation runs to generate the bi-objective Pareto front, which would be impractical for realistic optimization problems. Employing
techniques like semi-analytical solutions (Nguyen et al. 2020a, 2020b; Zhang et al. 2023) could possibly attenuate the computational
expense. However, their applicability remains limited, especially when dealing with complex reservoir systems as often encountered in
practical scenarios. Generally speaking, derivative-free methods are only computationally efficient on small-scale problems; thus, for
large-scale problems with hundreds or thousands of design parameters, gradient-based methods are much more computationally feasible
and efficient, especially when derivative information is available, whether via the means of adjoint method (Kraaijevanger et al. 2007;
Sarma et al. 2008; Voskov 2013; Kourounis et al. 2014), finite-difference estimation, or stochastic approximation (Spall 1998; Chen 2008;
Chen et al. 2009, 2017; Fonseca et al. 2017; Lu et al. 2017b; Li et al. 2022; Almasov and Onur 2021, 2023; Atadeger et al. 2023; Nguyen
Methodology
In this section, we describe the theory and methodology used to perform the nonlinearly constrained bi-objective production optimization
of CO2 storage.
Constrained Bi-Objective Production Optimization Problem. The conventional single-objective deterministic production optimization
problem refers to the estimation of the optimal design variables on predefined control timesteps that minimize the negative NPV, or
equivalently, maximize the NPV of the production life cycle. In this work, like Nguyen et al. (2023a, 2023b), we convert any maximization
into its equivalent minimization form, for the sake of consistency with the theories of convex optimization. The production optimization
problem is subjected to bound constraints on the design variables and operational constraints that are typically represented as nonlinear
state constraints. For a multiphase flow deterministic reservoir under CO2 injection, we define the negative life cycle NPV objective
function as 8̂ 9
2 3>
ˆ
< >
X N t
tn
N
X P XNI =
JNPV u = 4 c o q
N n
u cw q
N n
u cCO prd q
N n
u c CO inj q
N n
u 5 , (1)
ˆ tn o,i w,i 2 CO2 prd,i 2 CO2 inj,j >
>
n=1 :̂ i=1 j=1 ;
1 + b 365
T T
where u = u1 , u2 , : : : , uNu = u1,1 , u1,2 , : : : , uNw ,Nc is the Nu - dimensional column design vector containing all the production and
injection well controls and uk,mdenotes the control of well k at the mth control step, with a total of Nw wells and Nc control timesteps. In
Eq. 1, tn denotes the size of the nth simulation timestep, whereas tn stands for the cumulative time at the end of the nthtimestep. The
number of producers, the number of injectors, and the number of timesteps are denoted by NP , NI , and N t , respectively. The terms qN no,i ,
n n
qN w,i , and qN CO2 prd,i denote the average oil production rate, average water production rate, and average CO2 production rate of the pro-
ducer i over the nth timestep, respectively, whereas qN nCO2 inj,j denotes the average CO2 injection rate of injector j over the nth timestep.
co, cw , cCO2 prd , and cCO2 inj denote the oil price, produced water treatment cost, produced CO2 treatment and capture cost, and CO2
injection cost, respectively. The constant bis the annual discount rate. The first term in the summation in Eq. 1 above corresponds to the
oil revenue. The second term denotes the water production cost. The third term is the CO2 production cost. The fourth term corresponds
to the CO2 injection cost. As such, Eq. 1 describes the subsurface performance-driven elements of project economics and represents a
up
where ulow
i and ui denote the lower and upper bounds of the ith control variable, respectively, for a total of Nu control variables. ci u
denotes the general form of the ith inequality constraint, for a total of Nic inequality constraints enforced in the reservoir model. It is
where Nsim,n denotes the number of simulation timesteps within the control timestep tn. tn,i is the ith simulation time, with its corre-
sponding instantaneous IBHP value IBHPw,i from the simulation output, for tn,i . Note that Eq. 3 represents the computation of the average
value with midpoint Riemann sum (or composite trapezoidal rule). Along with rate constraints such as FLPR and FWPR, IBHP opera-
tional constraints are important and need to be taken into consideration to ensure the formation integrity and to prevent any potential
damage during the life cycle injection/production process. We do not include field gas production rate (FGPR) in this work since the only
gas being produced is CO2 (from the injection streams), which is not significant in this work. There is no hydrocarbon gas in the case study
setup (both from the reservoir’s initial condition and fluid composition) considered in this work, which we will describe in detail later.
In the context of CCUS, especially in the United States, industrial organizations that store CO2 are refunded with CTCs from the fed-
eral government to further incentivize the activities and investment in CCUS projects. Similar to the life cycle NPV as shown in Eq. 1,
the CTCs are also subject to monetary depreciation and inflation. Following Nguyen et al. (2023b), it is important to consider a separate
NPCTC objective function:
2 3
NI
Nt
X tn X
JNPCTC u = 4 rCO2 qN nCO2 inj,j u 5,
tn (4)
n=1 j=1
1 + b 365
where rCO2denotes the CTC (Sun et al. 2021; Nguyen et al. 2023b).
A few important remarks regarding the simplifications considered in our objective functions defined by Eqs. 1 and 4 are as follows:
First, we assume that the infrastructure necessary for handling CO2 is already in place; that is, CAPEX can be ignored if the facilities
required to inject and/or process the CO2 are in place. However, if the CO2-EOR strategy is compared against recovery strategies that do
not require acquiring CO2, then any additional CAPEX associated with putting CO2 infrastructure in place should be considered. Second,
the CO2 source for injection is assumed to be solely external, because produced CO2 would need to undergo several post-production
processes such as separation, impurity treatment, drying, compression, etc., which would result in some downtime before it would be
available for reinjection. Additionally, the case study setup in this work, which we will discuss later in this paper, does not result in very
significant CO2 production. Therefore, our justification of solely externally sourced CO2 injection is valid. Third, we treat a CTC and a
carbon credit rather similarly so that royalties and other types of financial constructs could be ignored in the objective functions. Usually,
a tax credit has a known value per ton of CO2, while carbon credits are bought and sold on the financial markets, and their values therefore
depend on supply and demand. Finally, we assume that both NPV and NPCTC depreciate over time with the same discount rate b. It is
Normalization of the Design Variables and Constraints. The magnitudes of the design variables affect the performance of gradient-
based optimization algorithms in general. Therefore, to make the step-size selection process stable, it is necessary to perform the
normalization of the design variables, which rescales the design variables into the interval of 0, 1 as follows:
ui ulow
uN i = up i low , (7)
ui ui
for all i = 1, 2, : : : Nu . Similarly, for the nonlinear state constraints, their values are also rescaled by their corresponding imposed lower or
upper thresholds to prevent any numerical ill-conditioning or instability. For example, instead of directly enforcing the upper-bounded
FLPR constraints given by
cj uN = FLPRj + FLPRup 0, (8)
for each of the control steps j = 1, 2, : : : , Nc , we enforce the following normalized constraints
FLPRj
cNj uN = + 1 0. (9)
FLPRup
Therefore, the post-normalization single-objective NPV optimization problem, as given by Eq. 2, becomes
minimize JNPV uN , (10a)
uN 2R Nu
subject toW 0 uN i 1, i = 1, 2, : : : , Nu , (10b)
cN i uN 0, i = 1, 2, : : : , Nic . (10c)
Likewise, the post-normalization optimization problem of the secondary objective function JNPCTC uN given by Eq. 6 becomes
minimize JNPCTC uN , (11a)
uN 2R u
N
subject toW J NPV u
N JNPV uN , (11b)
0 uN i 1, i = 1, 2, : : : , Nu , (11c)
cN i uN 0, i = 1, 2, : : : , Nic . (11d)
Note that the optimization problems defined by Eqs. 10 and 11, are solved by the bi-objective lexicographic-based LS-SQP workflow as
presented by Nguyen et al. (2022), instead of Eqs. 2a and 6a, Eqs. 2b and 6b, and Eqs. 2c and 6c, respectively.
Reservoir Model. Like Nguyen et al. (2023b), we use the Brugge reservoir model (Chen et al. 2010; Chen and Oliver 2010; Peters et al.
2010; Dehdari and Oliver 2012; Almasov et al. 2022; ) for our CCUS case study, with a compositional fluid model to simulate a depleted
oil reservoir using the CMG-GEM compositional simulator. The reservoir model, as illustrated in Fig. 1, consists of a total of 60,048
gridblocks (44,550 active cells) with 10 injectors and 20 producers. In this study, we only consider the base case (58th realization) for
deterministic production optimization. The initial reservoir pressure is 3,500 psi, and the water-oil contact depth is 5,377 ft. For more
details regarding the compositional fluid data, please refer to Yu et al. (2014) and Nguyen et al. (2023b). The Peng-Robinson equation of
state is used in this work in the compositional fluid model.
Constraint and Optimization Settings. In our study, we consider a 3,600-day life cycle production optimization, with 20 uniform control
steps of 180 days each. Hence, the number of control variables in this bi-objective optimization problem is Nu = 600, and their lower and
upper bounds are documented in Table 1. The enforced nonlinear state constraints, consisting of IBHP, FLPR, and FWPR, are shown in
Table 2. Note that, as the model has 10 injectors and 20 control timesteps, we would have 20 FLPR constraints, 20 FWPR constraints,
and 20 10 = 200IBHP constraints, totaled to 240 constraints. It is important to note that these constraints are enforced externally in the
optimization workflow, rather than being handled internally by the simulator’s field management framework, due to two main reasons.
First, it is to ensure the numerical stability of stochastic gradient estimation, which is particularly true for IBHP because, if the simulator
enforces this operational mode internally for injectors, it could result in multiple control solutions yielding the same objective function
value. Second, the purpose of having “control timesteps” is to operate the wells at a constant schedule within such time periods. As a
result, if the constraints listed in Table 2 are enforced internally by the simulator, then we could observe nonconstant rate/BHP control
schedules within one or more control timesteps. In fact, Liu and Reynolds (2020) show that the computationally viable alternative of
letting the reservoir simulation enforce the nonlinear state constraints using its internal heuristics may yield significantly inferior results.
Table 1—Lower and upper limits of the well controls in this study.
Economical Parameters for Objective Functions. The constants in two objective functions defined by Eqs. 1 and 4 are summarized in
Table 4. For more details regarding the carbon-related parameters specifically, please refer to Schmelz et al. (2020), Jones and Sherlock
(2021), and Smith et al. (2021).
Computational Results. In this section, we present the bi-objective optimization results obtained from applying the LS-SQP framework
by Nguyen et al. (2023b). Additionally, we show the optimization results without bi-objective optimization to further emphasize the
importance of having multiple objective functions.
Single-Objective Optimization. In this first subsection, we show the solution obtained by solving the single-objective NPV optimization
problem as defined by Eq. 10. The control variables are uniformly initialized as uN 0i = 0.5 for all i = 1, 2, : : : , Nu . To emphasize the
importance of bi-objective optimization that will be presented later in this section, we also perform a single-objective optimization of the
total life cycle cash flow, which is the sum of the two objective functions, defined as
JTot (u)
N =JNPV (u)
N + JNPCTC (u),N (12)
and we minimize it using Eq. 10 with JTot (u) N . We denote the corresponding optimal control solution as uN . Fig. 2
N instead of JNPV (u)
compares the results based on the sum of the two single-objective functions given by Eq. 12 and the single-objective NPV optimization,
while their corresponding optimal well control schedules are displayed as heatmaps in Fig. 3. Additionally, we added to Fig. 2 the NPV
convergence curves for the single-objective NPV optimization solution without IBHP nonlinear state constraints, indicated by the dark
gray line, to demonstrate how our IBHP-constrained solutions of interest differ from it.
Specifically for the results with IBHP constraints, an optimal solution uN from single-
ˇ objectively optimizing JNPV uN is found after
ˇ
53 SQP iterations, corresponding to 1,403 reservoir simulation calls. The NPV value is ˇJNPV uN ˇ = 3.990 109USD, and the NPCTC
Fig. 3—Optimal single-objective optimization schedule heatmaps for cases with IBHP nonlinear state constraints.
Fig. 4—Comparison of FLPR and FWPR nonlinear state constraints obtained from the maximizations of the sum of the two single-
objective functions (Eq. 12) and the single-objective function of NPV (Eq. 2 with Eq. 1). The green line represents the upper bound.
Orange lines indicate FLPR and FWPR constraint values for single-objectively optimizing NPV without imposing IBHP constraints.
Bi-Objective Optimization. Here, we focus on solving the NPCTC optimization problem as defined by Eq. 11 as the second step of
the lexicographic method. Note that for the initialization, we follow the procedure as stated by Nguyen et al. (2023b) to initialize the
design vector at the optimal solution uN of single-
objective NPV optimization. The negative life cycle NPV state constraint given by Eq.
6b is upper-
bounded using the value of JNPV uN = 3.990 109 USD, while the initial negative NPCTC objective function value is
8
JNPCTC uN = 5.264 10 USD.
Throughout this whole subsection, we show the results using = 0.95 for the state constraint given by Eq. 6b, corresponding to the
maximum
ˇ allowance
ˇ of a 5% decrease in the life cycle NPV. Therefore, upon convergence, it is expected that the NPV cannot be lower
than ˇJNPV uN ˇ = 3.775 10ˇ9USD. Fig. 6 shows the bi-objective optimization results for NPCTC and NPV, in which the green dashed
ˇ
lines represent the bound value ˇJNPV uN ˇ, serving as the lower bound for NPV or upper bound for JNPV uN . The optimization work-
flow achieves convergence after 76 iterations, corresponding to 1,530 reservoir simulation calls. We can see that, at the early stage of the
ˇbi-objective
optimization
ˇ process, the NPV fluctuates and violates a lot and then laterˇ stabilizes. ˇThe optimal NPCTC found is
ˇJNPCTC uN opt ˇ = 6.097 108 USD, which is a 15.6% increase compared with its value ˇJNPCTC uN ˇ as the initial guess of the bi-
ˇ ˇ
objective problem, and a 6.2% increase compared with ˇJNPCTC uN ˇ.
The schedule heatmaps at the bi-objective optimum are shown in Fig. 7, which appear to be quite different from the ones in Fig. 4. To
better demonstrate how the injection rates evolve visually, we also include the results for = 0.99in Fig. 7. For
ˇ the case
with
ˇ = 0.99
specifically, upon convergence, we obtained the NPV of 3.950 109USD, which is equal to the set bound of ˇ0.99JNPV uN ˇ. The opti-
mal NPCTC we obtained is 5.932 108USD, which corresponds to a 12.7% uplift over the initial guess. From Fig. 7, we observe that
the injection rates for both = 0.95and = 0.99are significantly higher (identified by more red colors in the control steps), which trans-
lates directly into higher CO2 storage in the reservoir. The values of the imposed field rate and IBHP nonlinear state constraints are shown
in Figs. 8 and 9, respectively. Unlike the single-objective solutions, here we observe much higher FLPR and FWPR at the beginning,
Fig. 8—Imposed FLPR and FWPR nonlinear state constraints for bi-objective optimization with
= 0.95.
Fig. 9—Imposed IBHP nonlinear state constraints for bi-objective optimization with
= 0.95.
Numerical Construction of the Bi-Objective Pareto Front. By repeatedly solving the bi-objective optimization problem as defined
by Eq. 11 with different values of , we can approximate the bi-objective Pareto front, which represents the trade-off between the two
competing objectives, as demonstrated in Nguyen et al. (2023b). Lower values of means more relaxation on the NPV constraint,
resulting in a larger feasible region for the second objective function JNPCTC uN to further decrease. One different solution we have
shown in the previous subsection is with = 0.99. Here, we construct the Pareto front using = f0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95, 0.99g,
as shown in Fig. 10. For more details, please refer to Table 5. Again, decreasing values of is equivalent to being willing to sacrifice more
NPV in exchange for more gain in NPCTC. The selection of the maximum value = 0.99, rather than exactly equal to 1 in the formal
formulation of the lexicographic method (Marler and Arora 2004; Nguyen et al. 2023b), is for the purposes of allowing a small tolerance
window in NPV state constraint to ensure convergence and stability of the optimization algorithm, due to the inaccuracy in the estimation
ˇ ˇ
of stochastic gradients. Note that, on each datapoint generated, the final NPV value does not need to be equal to the bound ˇJNPV uN ˇ
as specified in the state constraint given by Eq. 6b. We observe a steep improvement in the NPCTC between = 0.99and = 0.9.
We also include in Fig. 10 the solution uN from optimizing the total cash flow, as shown in the first subsection, to emphasize the
necessity of performing bi-objective optimization, especially with the usage of stochastic gradients. By comparing the optimum uN
(indicated by the red triangular marker in Fig. 10) with the = 0.99solution from the Pareto front, notice that, although the NPV values
are similar, the NPCTC values are very different, specifically it is much higher at = 0.99. This indicates that separating the two objective
functions instead of lumping them helps avoid this suboptimum, especially as the NPV value is observed to be significantly larger than
NPCTC in terms of magnitude. Also from the Pareto front, we can conclude that for the case study considered here, we observe that there
is not a significant improvement of NPCTC for 0.75. Fig. 11 shows visualization of the CO2 entrapment at the end of the simulation
for various investigated cases. It is worth noting that the results shown in Fig. 11 represent the case where we set the BHP for all producer
wells at its lower bound of 1,000 psi and the injection rates at the upper bound of 15 MMscf/D, as requested by one of the technical
reviewers. For this case, the NPV value is 2.109 109USD, and the NPCTC value reaches the maximum possible value at 6.321 108
USD. We can see that the primary objective (NPV) of such a case is inferior compared with any of the results we have presented so far.
One important point to note is that although in this work (and in Nguyen et al. 2023b), there is quite a big difference between the mag-
nitudes of NPV and NPCTC, specifically with NPV being more dominant. However, this is consistent with practical settings, as although
industrial companies are pushing for low-carbon solutions with a major focus on CCUS, their main revenues are still coming from oil and
gas operations. In other words, carbon tax cannot replace oil and gas revenues in terms of monetary value, but it is still something that
must be done. We will not go into the details of the real market of carbon credits, as it is very complex and focusing on it is not the scope
of our study, but we convey our point of NPCTC being much more important than just its monetary value.
Discussion
In this section, we provide a clear explanation as to why the two objective functions (negative NPV given by Eq. 1 and negative NPCTC
given by Eq. 4) should not be combined into a single-objective (“NPV”) function. Here, we further discuss why a single-objective func-
tion is not appropriate to solve the multi-objective optimization problem considered in this study.
As we stated previously, negative NPV (Eq. 1) and negative NPCTC (Eq. 4) present two conflicting objectives in the sense that we
want to maximize NPV due to oil production via injection of CO2 while maximizing the net CTC by sequestering as much CO2 as possible
in the reservoir. The optimal solution to these conflicting objectives calls for multi-objective optimization, so we envision the problem as
a multi-objective optimization problem. As stated in many previous works focused on solving multi-objective optimization problems [for
example, see Liu and Reynolds (2016a, 2016b, 2016c) and Alpak et al. (2022)], the solution to the multi-objective optimization problem
is the Pareto front (a hyper-surface in the multi-objective optimization problem). Like in our study, when only two objective functions are
considered, the Pareto front is a curve on a 2D plot that represents the tradeoff between the two objectives. The defining characteristic of
the Pareto front is that, given any specific point on the Pareto front, it is impossible to find another point on the Pareto front or another
feasible point that yields a greater value of all objective functions. In other words, the Pareto front serves as the boundary of a region that
represents all possible solutions in the search space. If we simply combine the two conflicting objective functions as a single-objective
function, then the optimal solution of this combined problem will not represent an optimal solution on the Pareto front but a suboptimal
solution.
Moreover, we refer the readers to an example from Alpak et al. (2022); see Fig. 21 of the cited reference. The authors concluded that
it is tedious and time-consuming to find the correct weighting factors for potentially conflicting objectives in an aggregated weighted-sum
objective function approach. Indeed, Alpak et al. (2022) as well as Liu and Reynolds (2016c) reported that only a (small) relevant portion
of the Pareto front could be determined with the weighted-sum method. Even when proper weights are determined, and with four times
more total objective function evaluations (compared with their proposed bi-objective workflow), the weighted-sum approach could only
construct six (red) points on the Pareto front (blue points) in Fig. 21 of Alpak et al. (2022).
Coming back to our paper, we presented our single-objective optimization results in Figs. 2 and 3. As can be seen from Fig. 2 and the
corresponding discussions that follow Figs. 2 and 3, by performing single-
objective optimization on the NPV only, we can evaluate the
value of the combined single-objective function JTot = JNPV + JNPCTC at the optimal (4.515 109 USD, with the NPV and NPCTC
being 3.990 109USD and 5.264 108USD, respectively). When we perform single-objective optimization on JTot instead, we obtained
9 9 8
JTot = 4.517 10 USD, with the NPV and NPCTC being 3.943 10 USD and 5.740 10 USD, respectively. We did observe that, in
both cases, JTot is close in monetary value (with a small difference due to numerical errors), but the values in the constituting objective
functions are different. It is important to note that the single-objective NPV solution does not directly contribute to the bi-objective Pareto
front but rather serves as the starting point to construct the Pareto front in the bi-objective optimization process.
As we proceeded to show in Fig. 10 in the manuscript and discussed in detail in the section “Numerical Construction of the Bi-
Objective Pareto Front,” by comparing the optiml uN (indicated by the red triangular marker in Fig. 10) with the γ = 0.99 solution from
the Pareto front, notice that, although the NPV values are similar, the NPCTC values are very different; specifically, it is much higher at
γ = 0.99. We emphasized the necessity of separating the two objective functions instead of lumping them to avoid such suboptimum,
especially as the NPV value is observed to be significantly larger than NPCTC in terms of magnitude. So, when the two conflicting objec-
tive functions are within different ranges of values (could even be some orders of magnitude different, as in our discussion after Table 5
and before Fig. 11), combining the two objectives is not recommended as the larger-valued objective would dominate in the solution-
finding process. This is especially even more important with the use of stochastic gradients in optimization. Although, in this work, we
did not perform different combinations of weights, such as Liu and Reynolds (2016c) and Alpak et al. (2022), for our combined objective
function approach, as it is very time-consuming (especially with a compositional model), clearly, with the result presented, the optimal
solution from the combined approach is not lying on the Pareto front. To be specific, the red triangular marker is nowhere near the green
marker with γ = 0.99 in Fig. 10. Again, an exhaustive search could be performed, but it would be too computationally inefficient.
Conclusions
In this work, we use the StoSAG gradient-based LS-SQP workflow (Nguyen et al. 2023a) coupled with the lexicographic method, as
described in Nguyen et al. (2023b) to solve the bi-objective production optimization design for CCUS on the Brugge model. Besides
Acknowledgments
This research was conducted under the auspices of Tulsa University Petroleum Reservoir Exploitation Projects (TUPREP). The first and
second authors thank TUPREP member companies for their support and acknowledge the CMG for making available multiple GEM
licenses. Some portions of this paper constitute Quang Minh Nguyen’s PhD dissertation.
References
Almasov, A. and Onur, M. 2021. Life-Cycle Optimization of the Carbon Dioxide Huff-n-Puff Process in an Unconventional Oil Reservoir Using Least-
Squares Support Vector and Gaussian Process Regression Proxies. SPE J. 26 (4): 1914–1945. https://doi.org/10.2118/201721-PA.
Almasov, A. and Onur, M. 2023. Life-Cycle Production Optimization of the CO2-Water-Alternating-Gas Injection Process Using Least-Squares Support-
Vector Regression Proxy. SPE J. 28 (2): 715–736. https://doi.org/10.2118/210200-PA.
Almasov, A., Nguyen, Q., and Onur, M. 2022. Nonlinearly Constrained Life-Cycle Production Optimization With a Least-Squares Support-Vector
Regression Proxy. Paper presented at the ECMOR 2022, The Hague, The Netherlands, 5–7 September. https://doi.org/10.3997/2214-4609.202244034.
Alpak, F., Jain, V., Wang, Y. et al. 2022. Biobjective Optimization of Well Placement: Algorithm, Validation, and Field Testing. SPE J. 27 (01): 246–273.
https://doi.org/10.2118/203960-PA.
Alpak, F., Wang, Y., Gao, G. et al. 2021. Benchmarking and Field-Testing of the Distributed Quasi-Newton Derivative-Free Optimization Method for Field
Development Optimization. Paper presented at the SPE Annual Technical Conference and Exhibition, Dubai, UAE, 21–23 September. https://doi.org/
10.2118/206267-MS.
Antoniou, A. and Lu, W. 2007. Practical Optimization: Algorithms and Engineering Applications. New York, New York, USA: Springer Science &
Business Media.
Atadeger, A., Onur, M., Sheth, S. et al. 2023. Deep Learning-Based and Kernel-Based Proxy Models for Nonlinearly Constrained Life-Cycle Production
Optimization. Paper presented at the SPE Reservoir Characterisation and Simulation Conference and Exhibition, Abu Dhabi, UAE, 24–26 January.
https://doi.org/10.2118/212690-MS.
Bonyadi, M. R. and Michalewicz, Z. 2017. Particle Swarm Optimization for Single Objective Continuous Space Problems: A Review. Evol Comput 25 (1):
1–54. https://doi.org/10.1162/EVCO_r_00180.
Brouwer, D. R., Nœvdal, G., Jansen, J. D. et al. 2004. Improved Reservoir Management Through Optimal Control and Continuous Model Updating. Paper
presented at the SPE Annual Technical Conference and Exhibition, Houston, Texas, USA, 26–29 September. https://doi.org/10.2118/90149-MS.
Chen, Y. 2008. Efficient Ensemble Based Reservoir Management. PhD dissertation, University of Oklahoma, Norman, Oklahoma, USA.
Chen, C. 2011. Adjoint-Gradient-Based Production Optimization with the Augmented Lagrangian Method. PhD Dissertation, University of Tulsa, Tulsa,
Oklahoma, USA.
Chen, C., Li, G., and Reynolds, A. C. 2012. Robust Constrained Optimization of Short- and Long-Term Net Present Value for Closed-Loop Reservoir
org/10.2118/112873-PA.
Chen, B. and Reynolds, A. C. 2016. Ensemble-Based Optimization of the Water-Alternating-Gas-Injection Process. SPE J. 21 (3): 0786–0798. https://doi.
10.2118/173217-PA.
Chen, B. and Reynolds, A. C. 2018. CO2 Water-Alternating-Gas Injection for Enhanced Oil Recovery: Optimal Well Controls and Half-Cycle Lengths.
Computers & Chemical Engineering 113: 44–56. https://doi.org/10.1016/j.compchemeng.2018.03.006.
Chen, B., Fonseca, R.-M., Leeuwenburgh, O. et al. 2017. Minimizing the Risk in the Robust Life-Cycle Production Optimization Using Stochastic Simplex
10.1007/s10596-010-9181-7.
Das, I. and Dennis, J. E. 1998. Normal-Boundary Intersection: A New Method for Generating the Pareto Surface in Nonlinear Multicriteria Optimization
Problems. SIAM J Optim 8 (3), 631–657. https://doi.org/10.1137/S1052623496307510.
Dehdari, V. and Oliver, D. S. 2012. Sequential Quadratic Programming for Solving Constrained Production Optimization–Case Study From Brugge Field.
SPE J. 17 (3): 874–884. https://doi.org/10.2118/141589-PA.
doi.org/10.1002/nav.3800020106.
Gen, M. and Cheng, R. 2000. Genetic Algorithms and Engineering Optimization. New York, USA: John Wiley & Sons.
Isebor, O. J., Durlofsky, L. J., and Echeverría Ciaurri, D. 2014. A Derivative-Free Methodology with Local and Global Search for the Constrained Joint
2005002.
Jones, A. C. and Sherlock, M. F. 2021. The Tax Credit for Carbon Sequestration (Section 45Q). Congressional Research Service Research Report.
Congressional Research Service.
Kennedy, J. and Eberhart, R. 1995. Particle Swarm Optimization. Paper presented at the ICNN’95 - International Conference on Neural Networks, Perth,
Australia, 27 November–1 December. https://doi.org/10.1109/ICNN.1995.488968.
Kourounis, D., Durlofsky, L. J., Jansen, J. D. et al. 2014. Adjoint Formulation and Constraint Handling for Gradient-Based Optimization of Compositional
Reservoir Flow. Comput Geosci 18 (2): 117–137. https://doi.org/10.1007/s10596-013-9385-8.
Kraaijevanger, J. F. B. M., Egberts, P. J. P., Valstar, J. R. et al. 2007. Optimal Waterflood Design Using the Adjoint Method. Paper presented at the SPE
Reservoir Simulation Symposium, Houston, Texas, U.S.A, 26–28 February. https://doi.org/10.2118/105764-MS.
Li, Y. and Onur, M. 2023. INSIM-BHP: A Physics-Based Data-Driven Reservoir Model for History Matching and Forecasting with Bottomhole Pressure
https://doi.org/10.1007/s00158-003-0368-6.
Nguyen, Q. M., Onur, M., and Alpak, F. O. 2022. Multi-Objective Optimization of Subsurface CO2 Capture and Storage Using Sequential Quadratic
2214-4609.202244033.
Nguyen, Q. M., Onur, M., and Alpak, F. O. 2023a. Nonlinearly Constrained Life-Cycle Production Optimization Using Sequential Quadratic Programming
(SQP) With Stochastic Simplex Approximated Gradients (StoSAG). Paper presented at the SPE Reservoir Simulation Conference, Galveston,
Texas, USA, 28–30 March. https://doi.org/10.2118/212178-MS.
Nguyen, Q. M., Onur, M., and Alpak, F. O. 2023b. Multi-Objective Optimization of Subsurface CO2 Capture, Utilization, and Storage Using Sequential
Quadratic Programming with Stochastic Gradients. Comput Geosci 1. https://doi.org/10.1007/s10596-023-10213-6.
Nguyen, K. H., Zhang, M., and Ayala, L. F. 2020a. Transient Pressure Behavior for Unconventional Gas Wells with Finite-Conductivity Fractures. Fuel
266: 117119. https://doi.org/10.1016/j.fuel.2020.117119.
Nguyen, K., Zhang, M., Garcez, J., and Ayala L. F. 2020b. Multiphase Transient Analysis of Flowback and Early Production Data of Unconventional Gas
Wells. Paper presented at the SPE Annual Technical Conference and Exhibition, Virtual, October 2020. https://doi.org/10.2118/201303-MS.
Nocedal, J. and Wright, S. J. 2006. Numerical Optimization. New York: Springer.
Peters, E., Arts, R. J., Brouwer, G. K. et al. 2010. Results of the Brugge Benchmark Study for Flooding Optimization and History Matching. SPE Res Eval
& Eng 13 (3): 391–405. https://doi.org/10.2118/119094-PA.
Sarma, P., Chen, W. H., Durlofsky, L. J. et al. 2008. Production Optimization With Adjoint Models Under Nonlinear Control-State Path Inequality
Constraints. SPE Res Eval & Eng 11 (2): 326–339. https://doi.org/10.2118/99959-PA.
Schmelz, W. J., Hochman, G., and Miller, K. G. 2020. Total Cost of Carbon Capture and Storage Implemented at a Regional Scale: Northeastern and
Midwestern United States. Interface Focus 10 (5). doi:https://doi.org/10.1098/rsfs.2019.0065.
Shi, Y. and Eberhart, R. 1998. A Modified Particle Swarm Optimizer. Paper presented at the 1998 IEEE International Conference on Evolutionary
1998.699146.
Smith, E., Morris, J., Kheshgi, H. et al. 2021. The Cost of CO2 Transport and Storage in Global Integrated Assessment Modeling. Int J Greenh Gas Control
109: 103367. https://doi.org/10.1016/j.ijggc.2021.103367.
Spall, J. C. 1998. Implementation of the Simultaneous Perturbation Algorithm for Stochastic Optimization. Paper presented at the IEEE Transactions on
Aerospace and Electronic Systems, Vol. 34, 817–823. https://doi.org/10.1109/7.705889.
Sun, Z., Xu, J., Espinoza, D. N. et al. 2021. Optimization of Subsurface CO2 Injection Based on Neural Network Surrogate Modeling. Comput Geosci 25
(6): 1887–1898. https://doi.org/10.1007/s10596-021-10092-9.
Voskov, D. V. 2013. Advanced Strategies of Forward Simulation for Adjoint-Based Optimization. Paper presented at the SPE Reservoir Simulation
Symposium, Woodlands, Texas, USA, 18–20 February. https://doi.org/10.2118/163592-MS.
TAC.1963.1105511.
Zhang, M., Nguyen, K., Wang, Z., and Ayala L. F.. 2023. Flowback and Early-Time Production Modeling of Unconventional Gas Wells Using an Improved
Semi-Analytical Method. Pet Sci. https://doi.org/10.1016/j.petsci.2023.06.003.