Download as pdf or txt
Download as pdf or txt
You are on page 1of 29

SPE-170587-MS

The Use of the Least Squares Probabilistic Collocation Method in Decision


Making in the Presence of Uncertainty for Chemical EOR Processes
A. M. Alkhatib, Saudi Aramco; P. R. King, Imperial College London

Copyright 2014, Society of Petroleum Engineers

This paper was prepared for presentation at the SPE Annual Technical Conference and Exhibition held in Amsterdam, The Netherlands, 27–29 October 2014.

This paper was selected for presentation by an SPE program committee following review of information contained in an abstract submitted by the author(s). Contents
of the paper have not been reviewed by the Society of Petroleum Engineers and are subject to correction by the author(s). The material does not necessarily reflect
any position of the Society of Petroleum Engineers, its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written
consent of the Society of Petroleum Engineers is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations may
not be copied. The abstract must contain conspicuous acknowledgment of SPE copyright.

Abstract
The Least Squares Monte Carlo method is a decision evaluation method that can capture the value of
flexibility of a process. This method was shown to provide us with some insight into the effect of
uncertainty on decision making and to help us capture the upside potential or mitigate the downside effects
for a chemical EOR process. The method is a stochastic approximate dynamic programming approach to
decision making. It is based on a forward simulation coupled with a recursive algorithm which produces
the near-optimal policy. It relies on Monte Carlo simulation to produce convergent results. This incurs a
significant computational requirement when using this method to evaluate decisions for reservoir engi-
neering problems because this requires running many reservoir simulations.
The objective of this study was to enhance the performance of the Least Squares Monte Carlo method
by improving the sampling method used to generate the technical uncertainties used in producing the
production profiles. The probabilistic collocation method has been proven to be a robust and efficient
uncertainty quantification method. It approximates the random input distributions using polynomial chaos
expansions and produces a proxy polynomial for the output parameter requiring a limited number of
model responses that is conditional on the number of random inputs and the order of the approximation
desired. The resulting proxy can then be used to generate the different statistical moments with negligible
computational requirement. By using the sampling methods of the probabilistic collocation method to
approximate the sampling of the technical uncertainties, it is possible to significantly reduce the
computational requirement of running the decision evaluation method. Thus we introduce the least square
probabilistic collocation method.
Both methods are then applied to surfactant-polymer flooding problems using a number of stylized
reservoir models. The technical uncertainties for the surfactant-polymer flooding process considered were
the residual oil saturation to chemical flooding, surfactant and polymer adsorption and the viscosity
multiplier of the polymer. The economic uncertainties considered were the oil price and the surfactant and
polymer price. Both methods were applied using three reservoir case studies: a simple homogeneous
model, the PUNQ-S3 model and a modified portion of the SPE10 model. The results show that using the
sampling techniques of the probabilistic collocation method produced relatively accurate responses
compared with the original method. For instance, it was possible to produce the same output for the
modified SPE10 model by using the second order quadrature nodes, 81 realizations, rather than the 103
2 SPE-170587-MS

realizations used for the least squares Monte Carlo method, thus achieving an order of magnitude
reduction in computational time.
Different possible enhancements were discussed in order to practically adapt the least squares
probabilistic collocation method to more realistic and complex reservoir models. The application was
extended to other chemical EOR processes such as alkaline-surfactant-polymer flooding and polymer
flooding. Uncertainty in heterogeneity was also incorporated using Gaussian and multiple point statisti-
cally generated permeability fields. The results show that using the least squares probabilistic collocation
method produced accurate results when compared with the least squares Monte Carlo method. This was
applied to polymer, surfactant-polymer and alkaline-surfactant-polymer flooding case studies.
Introduction
Chemical enhanced oil recovery (EOR) processes promise significant potential; to achieve this it is
necessary to incorporate uncertainty in the evaluation model. Numerous sources of uncertainty exist and
any decision will need to consider both economic and technical uncertainties. The concept of Value of
Flexibility (VoF) was used in a decision making evaluation method under uncertainty for chemical EOR
by Alkhatib et al. (2013) and Alkhatib and King (2014b) assuming economic and technical uncertainty.
VoF is defined as the value of splitting decisions into multiple decisions over time with the opportunity
to learn between decisions and having the option to respond to that learning (Bratvold and Begg 2010).
The Least-Squares Monte Carlo (LSM) method was used as a decision making method to evaluate the
VoF in the oil and gas industry by Willigers et al. (2009). The LSM method originated from real options
valuation and stochastic optimal control of financial processes where uncertainties are defiend by
stochastic differential equations (Longstaff and Schwartz 2001; Gamba 2002; Choudhury et al. 2007;
Cortazar et al. 2008). This method is based on approximate dynamic programming (Powell 2011) by
performing a forward simulation which is then followed by a recursive process to produce an optimal
policy.
The LSM method evaluates a predefined decision in the presence of uncertainty. The stochastic
behavior is produced by Monte Carlo simulation (MCS) generating a large number of realizations or paths
over which the different alternatives in the decision are evaluated and then the VoF is determined by
taking the difference between the mean of the maximum (optimal) values obtained from the decision
scenario and the mean of the value of a predfined static (or inflexible) scenario. Since the method is based
on MCS, this requires a large number of realizations to produce convergent results. A negligible
computational effort is incurred for generating stochastic paths for the economic parameters as it requires
only the evaluation of discretized stochastic differential equations over a range of time. However, the
propagation of technical uncertainties demands a relatively more significant computational effort because
this requires the sovling of the partial differential equations to produce the recovery profiles of the
reservoir models for each set of realizations of the random input parameters.
Thus the efficiency of this process will depend on the complexity of the reservoir model. Previously,
Alkhatib et al. (2013) used simplistic reservoir models, to the order of 103 to 104 grid cells. These models
were run within a manageable time frame. However, to apply the LSM to more complex reservoir models,
a more efficient approach is required. An approximation method is needed for either approximating the
reservoir model solver or approximating the technical random input parameters in order to avoid the MCS
computational requirement. Common robust approaches for approximating reservoir behavior such as
artificial neural networks (Karambeigi et al. 2011) or experimental design (Cheong and Gupta 2005) have
been used in order to obtain statistical moments in a quick manner. However, one disadvatage of
experimental design methods are that they do not incorporate the full probability distributions of the
random input parameters in a consistent manner when producing the response surface. Experimental
design also assumes inherently that parameters are uniformly distributed because all samples are equally
weighted (Li et al. 2011). Other approaches such as perturbation methods are commonly used; however
SPE-170587-MS 3

they are restricted to systems with relatively small random inputs and outputs, which is a difficult
condition to satisfy for nonlinear problems where small random inputs may result in large random outputs
(Xiu and Hesthaven, 2005).
The probabilistic collocation method (PCM) has been shown to be a robust and efficient uncertainty
quantification method for flow in porous media (Li et al. 2011) and was applied by the authors recently
(Alkhatib and King 2013, Alkhatib and King 2014a) for a chemical EOR process to efficiently quantify
parametric uncertainty. The efficient sampling approach of the PCM method is used to replace the MCS
of the recovery profiles and consequently reduce the computational effort required to run the LSM
method. This new method is known as the Least Squares Probabilistic Collocation Method (LSPCM) is
as a robust and efficient flexibility evaluation method in the presence of technical and economic
uncertainty (Alkhatib and King 2014b).
In this study the LSPCM is applied to evaluate a variety of decisions in implementing chemical EOR
processes on synthetic reservoir models. The LSPCM is applied to surfactant-polymer, polymer and
alkaline-surfactant-polymer flooding processes. The initial decision that was assessed is that of timing the
initiation of the chemical EOR processes taking into consideration economic and technical uncertainties.
The robustness of the LSPCM is measured by using the results obtained from the LSM method as a
benchmark. The complexity of the decision is then enhanced to incorporate additional decisions such as
varying the chemical slug formulation.
The paper is structured as follows: the sources of uncertainty in chemical EOR processes are reviewed.
The case studies and implementation approach are then briefly discussed. The objective function is then
defined, after which the LSM method is introduced with some examples. The PCM is then discussed with
some examples. Incorporating the PCM with the LSM method is then presented and is compared with
reference examples obtained by the LSM method. This is followed by applying some enhancements of the
LSM method which are then applied to the LSPCM in order to evaluate more complex decisions based
on changing control variables. The LSPCM is then applied for other chemical EOR processes to illustrate
its extendability.

Uncertainty in Chemical EOR Floodinng


Although chemical EOR processes have significant potential, they represent a small fraction of commer-
cially successful EOR projects. This is mainly due to uncertainty (Lake 1989, Thomas 2006). There are
many sources of uncertainty in chemical flooding processes such as heterogeneity, resident brine salinity
and residual oil saturation to chemical flooding (Sorc) and surfactant, polymer and alkline adsorption
(Kossack and Bilhartz 1976; Brown and Smith 1984; Hankins and Harwell 1996; Zheng et al. 2000;
Anderson et al. 2006; Cheng et al. 2012). Sorc is considered to be the one of the most significant factors
affecting recovery efficiency, while surfactant adsorption is found to have a dominant adverse effect as
well for surfactant-polymer floodig (Brown and Smith 1984; Zhang et al. 2005; Dang et al. 2011; Solairaj
et al. 2012). Salinity effects have been studied extensively and found to influence recovery significantly
for all chemical EOR processes (Gerbacia 1978; Pope et al. 1979; Hirasaki et al. 2011; Puerto et al. 2012;
Lohne and Fjelde 2012; Sheng 2013). As for polymer flooding, the integrity of polymer viscosity is
essential for the effectiveness of the process and its behavior can be uncertain (Weiss et al. 1985; Alsofi
and Blunt 2011).
In this study, the chemical flooding processes used are simulated using Schlumberger’s ECLIPSE. It
is important to note that this sovler might not simulate the detailed chemical interactions and phenomena
that occur in the reservoir (ECLIPSE 2010), however it is considered to be sufficient for this study because
the aim is to demonstrate the applicability of the proposed methods, not the design of a specific chemical
EOR process. Furthermore, the solver does capture the main features of chemical flooding processes
(ECLIPSE 2010).
4 SPE-170587-MS

Table 1—The distribution and probability density function bounds for


the technical uncertain parameters.
Variable Distribution PDF Bounds

Tsorc Uniform [0,0.2]


TDS Log-Normal [-8,1]
TDD Log-Normal [-8,1]
TVD Uniform [1,10]

Figure 1—Decision Flexibility Flowchart

The LSM, PCM and LSPCM methods are first illustrated for a surfactant-polymer flooding process
where four main chemical-related uncertainties are assumed: surfactant and polymer adsorption (Ds and
Dp), residual oil saturation to chemical flooding (Sorc) and the viscosity multiplier effect of the polymer
mobility buffer (Vp). The technical uncertainties are assumed to be defined by standard distributions since
no specific set of lab or field data is used. See Table 1. Surfactant and polymer related assumptions are
shown in Appedix X.
For TDs and TDp, a variable transformation is performed once a realization is obtained:
TD ⫽ ek, where k~N(– 8,1).
Case Studies and Implementation
The proposed method is demonstrated using three synthetic reservoir models. The LSM method and the
LSPCM are used to evaluate the decision of timing the initiation of a chemical EOR process over a
finite-discrete decision set given economic and technical uncertainty. Surfactant-polymer flooding is the
process considered for demonstrating the methods. The decision is defined as having the option, or
flexibility, to initiate surfactant-polymer floodig over a discretized decision space as opposed to not
having this option and initiating the surfactant-polymer flood at the start of the reservoir simulation
horizon. See Fig. 1 which illustrates the initiation decision over four decision nodes or times. The
technical random inputs used are the parameters in Table 1. The economic random inputs are discussed
below. The LSM algorithm is performed by a MATLAB (MATLAB 2012) code that is coupled with
SPE-170587-MS 5

Figure 2—Permeability realization and well placement for the PUNQ-S3 and SPE10M reservoir models used. P1, P2, P3 and P4 are the producers
and I1 is the injector. The permeability scale follows a logarithmic scale.

ECLIPSE (ECLIPSE 2010) which performs the reservoir simulations. As for the LSPCM, the LSM
MATLAB code was extended and coupled with a PYTHON (PYTHON 2012) PCM module (Feinberg
2012) that performs the sampling to produce the collocation nodes.
Homogenous Reservoir Model
The homogeneous model was used earlier by Alkhatib et al. 2013 with a total of 10x10x10⫽ 1000 grid
cells. A slightly water-wet black oil system is used. A quarter-five spot pattern was used with a simulation
horizon of ten years. As mentioned earlier the decision that is evaluated is timing the surfactant-polymer
flood, with a decision set containing four possible initiation times: year four, five, six and seven. These
decision times were selected in order to evaluate whether it is best to start the surfactant-polymer flood
before, during or after water breakthrough.
PUNQ-S3 Reservoir Model
This model uses the permeability field and the geometry of the PUNQ-S3 reservoir model (PUNQ-S3
2012) with a total number of 19x25x8⫽3800 grid cells. This example model uses the same fluid data used
for the homogeneous reservoir model mentioned above. The well pattern used for this reservoir model is
based on placing four producers on a structurally higher position (crestal area) within high permeability
streaks and one injector at a structurally lower position. See Fig. 2. The decision set used for this model
is surfactant-polymer flooding initiation at year five, seven, nine and eleven given a field life of twenty
years. This decision set also revolves around the water breakthrough time of the reservoir model.
Reservoir simulation constraints are present in Appendix A.
Modified SPE10 Reservoir Model (SPE10M)
This reservoir model is based on an upscaled section of the SPE10 model (Christie and Blunt 2001). It
represents the upper lithology (the Tarbet formation) of the original reservoir model, the top 35 layers,
upscaled to a 20x55x7 grid with a total number of 7700 grid cells. The original SPE10 model reservoir
and fluid data are used. The well pattern is assumed to be an inverted five-spot pattern, see Fig. 2. The
6 SPE-170587-MS

decision set assumed is the same as that used for the PUNQ-S3 model mentioned above having a
simulation horizon of twenty years. Simulation constraints are presented in Appendix A.

Objective Function
The objective function used is the net present value (NPV) which is based on calculating net revenue as,
(1)

, where Ro is the oil revenue, Cwi, Cwp and CC are the water injection, water production and chemical
costs, respectively. CAPEXC is the additional capital expenditure incurred for the chemical EOR process.
The economic valuation is based on risk-neutral valuation. In this approach the oil price stochastic
model is adjusted for risk using a risk premium of 3%1. This is a more comprehensive approach for
analyzing cash flows because it adjusts every component of the cash flow for risk and then discounts for
time (at the risk-free rate). Traditionally, a constant discount factor (including both time and risk) is used
that can generate biased results due to the varying risk patterns of the constituent cash flows (Jafarizadeh
and Bratvold 2009). A risk-free rate, r, is then used to discount the cash flows, which is assumed to be
constant over time and equal to 6%. This study assumed continuous discounting. Capital expenditure
includes facility cost, laboratory and design costs and incremental monthly operation costs. The capital
expenditure is assumed to be incurred at the start of the surfactant-polymer flood. The economic data used
can be found in Appendix B.
There are five economic random variables, oil price (Po), surfactant cost (Ps), polymer cost (Pp), cost
of water injection (Pwi) and the cost of produced water (Pwp). Po($/stb), Ps($/stb surfactant) and Pp($/stb
polymer) will be modeled independently while Pwi($/stb) and Pwp($/stb) are correlated with Po. Po is
assumed to follow a risk-neutral Ornstein-Uhlenbeck process (Uhlenbeck and Ornstein 1930). This is a
mean reverting stochastic process. See Appendix B. The polymer price and surfactant price2 (Ps) are
modeled using a uniform distribution, U(1,2) $/lbm and U(1,3) $/lbm respectively. At each time step, a
unique realization is sampled from the distribution. A cost multiplier factor of 0.10 is applied to the oil
price (Po) in order to obtain the water injection cost (Pwi) and water production cost (Pwp). This is assumed
in order to penalize for water injection and production.
The economic random variables used in the LSM algorithm: the oil price (Po), the surfactant cost (Ps)
and polymer cost (Pp) which are summarized by the variable, Pi ⫽ {Po, Ps, Pp}. The technical uncertainties
influence is exhibited in the production profile of the reservoir model which affects the NPV estimate for
each path. These are summarized by the variable Ti ⫽ {TS , TD , TD , TV }. The other economic
orc s p p

parameters which are correlated with Po are incorporated in the NPV calculation.

Least Squares Monte Carlo Method


The economic random variables are assumed to vary with time. Thus Pi will be depicted with an additional
subscript, yn, which illustrates the time dependency of the economic random variables. yn is the decision
node or time which is an element of the set of decision nodes, Y ⫽ [y1, . . ., yn]. Hence, Pi,yn(x) is the set
of economic random variables at the decision node yn along the x-th simulated realization path. The
technical random variables are assumed to be fixed through time, Ti,y1 (x) ⫽ Ti,y2 (x) ⫽. . .⫽ Ti,yn (x). The
time-subscript is dropped for simplicity,Ti(x).
A payoff function V(yn,Pi,yn (x), Ti(x)) is defined and assumed to be the net present value (NPV) of
the chemical EOR process (this is defiend as the NPV at the end of the simulated field life). The payoff
function is achieved when we exercise the option to initiate the surfactant-polymer flood, dependent on

1
Risk premium value is obtained from the oil price model in Willigers and Bratvold (2009).
2
Surfactant and polymer cost range was based on (Anderson et al 2006; Thomas 2006; Wyatt et al 2008).
SPE-170587-MS 7

random variables realiation, at time yn. This function is also defined as the immediate option exercise
value.
An optimal value function, OV, is defined as the value of the optimal NPV at the end of the simulated
field life dependent on exercising the optimal surfactant initiation time. Then:
(2)

,␶(x) is the optimal policy per realization and (x) 僆 Y.


, is assumed to follow the following optimality condition:
(3)

is then determined using Bellman’s dynamic programing equation(Bellman 1957):


(4)

r is the discount factor and is the expectation function.


A continuation function is defined as follows:
(5)

,
where ␪yn is the information set{Pi, Ti} at time yn.
The decision rule or criteria at time yn along the x-th path is defined as:
(6)

The decision rule is applied recursively from yn ⫽ yN back to yn ⫽ y1 to determine the optimal policy.
When all the optimal policies have been determined along all paths, the value of the decision is obtained
by calculating the average of all optimal values of the objective function:
(7)

The main challenge in applying the decision rule is calculating the value of the continuation function,
⌽ at time y given (Pi,␶(x) (x), Ti(x)). The algorithm assumes that the continuation function value is the
expectation, dependent on the available information at yn, of future optimal values. Thus the continuation
value can be defied as the present value of the future expected payoffs:
(8)

Since ⌽ is an element of a linear vector space, thus belonging to the Hilbert space, the contivuation
value is represtented as:
(9)

is the j-th element in the orthonormal basis and y 僆 [y1, . . ., yn]. Lj(y, Pi,y, Ti) can be powers of the
random variables, Laguerre or Hermite polynomials (Longstaff and Schwartz 2001). For simple decision
problems, the choice of basis functions is straightforward and using different polynomials can produce
similar results because any polynomial can be expressed as a linear combination of other polynomials. For
more complex decision problems using different basis function can affect the results obtained (Moreno &
Navas 2003). It is important to perform a sensitivity analysis of the type of basis functions used. In this
study, Lj(y, Pi,y, Ti) is assumed to be a simple polynomial of the second order without cross products
(Stentoft 2004). See Alkhatib et al. (2013) and Alkhatib and King (2014) for a more detailed discussion
on basis function design for problems similar to those discussed in this study.
can be approximated by assuming a finite set of basis functions, J:
8 SPE-170587-MS

Figure 3—Probability density function estimate of VoF and histogram of optimal policy determined by the LSM method using 103 realizations. Each
row is specific to a reservoir model.

(10)

The unknown functional form of the continuation equation can is assumed to be represented as a linear
combination of a countable set of the random variables measurable basis functions. Thus Øj(y) can then
be represented by a linear least squares regression of ⌽J(y, Pi,y, Ti) onto the basis {Lj(y, Pi,y, Ti)}:
(11)

The estimated continuation value is then defined as:


(12)

After obtaining the VoF can then be determined. VoF is defined as the difference between
the mean of the optimal policy scenario and the mean of the static policy scenario:
(13)

,
where is the average NPV of the static scenarios over all paths. If VoF is positive this implies
that value is created by incorporating flexibility into the decision. If VoF is negative, then incorporating
flexibility does not create value.
The accuracy of the estimated VoF can be enhanced by increasing the number of simulated paths, time
steps and basis functions (Gamba 2002). It is important to note that the policies determined by the
algorithm are near-optimal policies because they are obtained by approximating the continuation value.
Fig. 3 shows results of performing the LSM method for the three reservoir model examples introduced
SPE-170587-MS 9

Figure 4 —Bar chart illustrating the total number of realizations favoring flexibility determined by the LSM method using 103 realizations for the
homogeneous, PUNQ-S3 and SPE10M models. A realization favors flexibility if the value of the optimal policy of the decision scenario is greater than
the value of the static scenario.

above. The decision or flexibility evaluated was whether to initiate the surfactant-flooding process at the
beginning of year four, five, six and seven for the homogeneous model. For the two heterogeneous
reservoir models, the decision evaluated was whether to initiate the process at the start of year five, seven,
nine or eleven of the simulation horizon. The left column shows the probability density functions (PDF)
estimate of the VoF. The right column shows the optimal policy histograms produced by the LSM method.
The results show that a unique response is obtained for each reservoir model due to is unique heteroge-
neity structure, fluid properties and field life. For the PUNQ-S3 and SPE10M reservoir models, the VoF
obtained for most realizations, was positive. This contrasts with the VoF obtained for the homogeneous
reservoir model. Similar trends for the policy histograms were observed for both the PUNQ-S3 and the
SPE10M models.
A general observation on the produced optimal policy histograms is that they show that most
realizations favour initiating surfactant-polyemr flooding at the first or last decision node. One possible
explanation for this behavior is the choice of the decision space discretization, i.e. the number of time
steps at which initiating the EOR process is considered. The decision sets used in the example discussed
above was based on four annual decision nodes or times. This is relatively a coarse discretization. If the
decision set is discretized to a finer scale, for instance, quarterly or monthly decision nodes it is expected
to produce different optimal policy histogram trends. However, by increasing the discretization of the
decision space (increasing the number of time steps), the LSM algorithm would require a better choice of
basis functions along with an increase in the number of these basis functions (Bender and Steiner 2010).
This leads to an increase in the number of realizations required. Glasserman and Yu (2004) show that the
number of realizations required to produce convergent results increases exponentially with the number
and the order of the basis functions used in the regression. Thus in order to produce accurate results for
finer discretization of the decision space, a larger number of basis functions is required leading to an
increase in the number of realizations. This leads to a significant increase in the computational cost of
running the algorithm.
Fig. 4 shows how many realizations favor flexibility for each model. Favoring flexibility means that
the value of the optimal policy of the decision scenario is greater than the value of the static scenario. Once
again, PUNQ-S3 and SPE10M exhibit similar trends as opposed to the homogeneous model. One possible
explanation for this discrepancy is the effect of the heterogeneity of the PUNQ-S3 and the SPE10M
models. This effect can be manifested by the channeling that occurs between the injector well and the
producer wells. Furthrmore, the sweep effieciecy of the surfactant-polymer slug varies significantly based
10 SPE-170587-MS

on the permeability realization of the reservoir model. The choice of the static scenario can also affect of
the mean VoF obtained. If the static scenario was defined as initiating the surfactant-polymer flood at a
later time, closer to the decision set for instance, then the VoF and optimal policy produced could be quite
different.
These results were obtained using 103 realizations of the random parameters (inputs). Convergence of
the results produced by the LSM method for chemical EOR decisions is discussed further in Alkhatib et
al. 2013. Using 103 realizations might not always be sufficient to produce a convergent result, however
it still requires a significant computational effort even if simple reservoir model such as the ones in this
study are used. Thus performing this method on more complex reservoir models would require a more
robust and efficient approach in order to produce a converged result. These results are used later in the
study as reference cases.

Probabilistic Collocation Method


The probabilistic collocation method (PCM) is a widely used uncertainty quantification method for both
parametric and spatial random parameters in different disciplines (Mathelin and Husseini 2003; Loeven
et al. 2007; Li and Zhang 2007; Loeven and Bijl 2008; Li and Zhang 2009; Onorato et al. 2010). For the
purpose of this study, the PCM is used from the for parametric uncertainty quantification. The PCM is
based on polynomial chaos expansion and probabilistic colloation. These are discussed further below in
addition to other considerations with regards to PCM implementation.

Polynomial Chaos Expansion


Polynomial Chaos Expansion (PCE) is a spectral expansion of random variables that expresses stochastic
quantities as orthogonal polynomials. By applying this expasion to differential equations with random
inputs, the expansion coefficients of the PCE expansion are solved for (Xiu, 2005). This approach was
introduced by Ghanem and Spanos (1991). PCE is based on the homogeneous chaos theory of Wiener
(1938). PCE for a specific response, y, is defined as:
(14)

d, is the highest order of the expansion, ak is the deterministic coefficient and Гk(x) is the kth order
multidimensional orthogonal polynomial of the random variables. x ⫽ (x1, x2, . . ., xN)T.
The input distribution of the random inputs determines the type of the orthogonal polynomial used to
represent the random input variable is determined by the input distribution of the random variable. A
generalized chaos expansion was developed by Xiu and Karniadakis (2002) based on the different
polynomials in the Askey scheme for different types of random input variables. For example, uniform
variables are better approximated by Laguerre polynomials. For a specific input distribution, the selected
PCE converges at a better rate than other polynomials expansions (Li et al. 2011). For arbitrary
distributions, the associated orthogonal polynomial expansions must be constructed numerically using, for
example, the algorithms introduced by Gautshci (1994).
One of the advantages of PCE’s is that they guarantee convergence of the approximation distribution
to the true distribution of the random parameter as the order of the PCE increases (Field and Mircea 2007).
To obtain the coefficients for the random output variable PCE, usually Galerkin projection is used. This
results in a coupled system of deterministic equations which in turn results in an intrusive method (Loeven
et al 2007). This is not always desribale or possible to sovle depending on the numerical solver used for
simulating reservoir behavior. On the other hand, the probabilistic collocation method is a non-intrusive
method that can be used to solve for the PCE coefficients.
SPE-170587-MS 11

Probabilistic Collocation
Following the notation in Tatang et al. (1997), we have a model
(15)

Where y can be approximated by using a set of specified functions such as orthogonal polynomials
{gi(x)} as:
(16)

,
where N is the order of the approximation. Because the approximation does not reproduce the actual
model, a residual is thus defined as:
(17)

By requiring each member of {gi(x)} and the residual to be orthogonal, it is possible to obtain the set
of coefficients in the approximation {␣i}:
(18)

The obtained coefficients define the PCE of the output parameter resulting in the polynomial proxy of
the model. There are a number of methods for solving equation (18). Gaussian quadrature is one method
which uses the Golub-Welsch algorithm (Golub and Welsch 1969) to obtain the weights g(x) and the
nodes (x) of the quadrature. Another approach is linear regression which uses, for example, Chebyshev
nodes and then performs a regression of the model response at these nodes onto a basis of the PCE’s.

Collocation Nodes
The choice of collocation nodes strongly influences the efficiency of the PCM (Lin and Tartakovsky
2009). Gaussian quadrature (i.e. full product tensor) collocation nodes are obtained by using the roots of
the next-higher order polynomial expansion of the random input variables depending on desired order of
approximation. As for the Chebyshev-type nodes, a specific type known as Fejer nodes were used because
they are better suited for infinitely bound distributions, such as normal distributions, than the more
common Clenshaw-Curtis sampling because they use the interior Chebyshev extrema only and neglect the
boundary values. The Fejer nodes are used under two variants: boxing and mapping methods. Boxing
refers to scaling the nodes to the distribution bounds of the input parameters. Mapping refers to mapping
points from the unit hypercube to the desired domain using the inverse of the cumulative distribution
function. The boxing method is used to scale the nodes which are generated for (0, 1) domain to the
distribution bounds. The mapping method is used in order to uniformly sample around the high probability
region. The mapping approach is based on the idea of inverse transform sampling. Once the Fejer nodes
are obtained in the (0,1) domain, The value of each of the collocaitn nodes is found by using the inverse
cumulative distribution function of the random input parameters. For multidimensional cases, the
Rosenblatt transform is used.
The Fejer nodes are defined as Chebyshev extrema of the second type over an arbitrary interval (0,1):
(19)

,
where N is the order of the polynomial chaos expansion. See Fig. 5 for an illustration of quadrature
(full tensor product), mapped and boxed Fejer nodes for a two-dimensional random input case.
Fig. 6 presents results for parametric quantification of uncertainty for a surfactant-polymer flood
obtained by using the PCM discussed above assuming the recovery factor as the random output parameter
to be approximated and the parameters in Table 1 are the random input variables used. See Alkhatib and
12 SPE-170587-MS

Figure 5—Two-Dimensional quadrature (full tensor product), mapped and boxed Fejer grids for a 4th order approximation.

Figure 6 —Recovery factor probability density function estimates determined by second, third and fourth order PCM compared with the probability
density function estimates determined by Monte Carlo simulation (MCS). The number of realizations used is shown in parentheses for each sampling
method in the first row. Each row is specific to a reservoir model while each column is specific to the collocation method mentioned at the top of the
column.

King (2013) and Alkhatib and King (2014a) for a detailed discussion on the implementation and accuracy
of the PCM. For the results in Fig. 6, the surfactant-polymer flooding process was initiated at the start of
the simulation using a surfactant concentration of 3% wt. and duration of 10% PV followed by a polymer
SPE-170587-MS 13

slug of 3% wt. and 100% PV. The PDF estimates


produced by the PCM are compared with the PDF
estimates produced by MCS using 103 realizations.
The results show that PCM performed by linear
regression with mapped Fejer sampling produced
the relatively best visual agreement with the MCS
PDF for all reservoir models. Gaussian quadrature
(full tensor product) PCM also achieved a decent
visual agreement. To illustrate the efficiency of
PCM, a second order PCM performed by linear
regression with mapped Fejer nodes on PUNQ-S3
model can produce a similar PDF to the MCS while
only using 33 nodes (i.e. 33 simulation runs).
Least Squares Probabilistic Figure 7—Number of nodes or realizations as a function of the PCM
order for the Fejer and Gaussian quadrature (full tensor product)
Collocation Method sampling methods. The number of realizations used by the LSM (Monte
Carlo simulation) is plotted for comparison.
When the LSM method is performed, the elements
of the vector T are obtaiend by a Monte Carlo
simulation assuming a large number of realizations. The finite difference equations of the simulator are
then solved for each defined strategy for each realization. The forward simulation is therefore the most
computationally costly component of the LSM method. In this section, a more efficient sampling method
is proposed to substitute the MCS while maintaining the same effectiveness of MCS.
This is based on the PCM since it has been shown to robustly quantify uncertainty for different
multidimensional random input surfactant-polymer flooding examples (Alkhatib and King 2013; Alkhatib
and King 2014a). The collocation nodes selection process is suggested to be incorporateed with the LSM
method. This new approach, LSPCM, can significantly reduce the number of reservoir simulations
required.
As mentioned above, the residual in the PCM is defined as the difference between the model and the
approximation:
(20)

,
the set of collocation points {q} is used as the vector of random inputs in place of the MCS generated
values. The three sampling methods discussed above are suggested to be used: the Gaussian quadrature
generated nodes, boxeded Fejer nodes and mapped Fejer nodes.
The Bellman dynamic programming equation then becomes:
(21)

,
where q 僆 Q is the collocation node and Q ⬍⬍ X, where X is the total number of MCS realizations.
By significantly reducing the number of realizations, the efficiency of the LSM method is therefore
enhanced. See Fig. 7.
LSPCM Case Studies and Discussion
The LSPCM algorithm was applied to the same three LSM examples presented above. The aim was to test
the LSCPM for both the homogenous model and well known heterogeneous benchmark reservoir models.
The results of the LSM, mentioned earlier, are used as the reference case to evaluate the performance
of the LSPCM. See Fig. 2. By substituting the Monte Carlo sampling component with PCM sampling
14 SPE-170587-MS

Figure 8 —Probability density function estimates for the VoF determined by the LSPCM using second, third and fourth order collocation nodes
compared with that determined by the LSM method (using MCS samples). The number of realizations used is shown in parentheses for each sampling
method in the first row. Each row is specific to a reservoir model while each column is specific to the sampling method mentioned at the top of the
column.

methods it is desired to produce similar results to the LSM method while using significantly fewer
simulation runs. Fig. 8 presents the results of performing the LSPCM for a second, third and fourth order
approximations using the sampling methods mentioned above. The PDF estimates of the VoF show that
for the homogeneous model, Fejer nodes seem to achieve a relatively better visual match with the LSM
result. The PDF estimates of the PUNQ-S3 model show good visual agreement with the LSM PDF
estimate. Using third order mapped Fejer nodes, relatively the same PDF estimate is produced as that of
MCS while using only 120 nodes. This results in an enhancement of a factor of 8.3 compared with MCS
with respect to the number of realizations. As for the SPE10M model, the best visual match was obtained
by using second order Gaussian quadrature nodes which require 81 realizations thus resulting in an
enhancement of a factor of 12.3.
Fig. 9 presents the optimal policy histograms for the example reservoir mdoels. The histograms
obtained via the LSPCM are scaled to the histogram obtained by the LSM method. For the homogeneous
model, the scaled LSPCM histograms exhibit similar trends to the LSM histogram. For the PUNQ-S3
model, the optimal policy histograms obtained by LSPCM exhibit general agreement with the LSM
produced policy, except for the results obtained by second order boxed Fejer samples, which exhibit a
clear discrepancy. The policy histograms obtained by the LSPCM for the SPE10M model show a
generally good agreement with the LSM policy histogram.
SPE-170587-MS 15

Figure 9 —Optimal policy histograms by the LSPCM using second, third and fourth order collocation nodes and compared with the LSM method.
Each row is specific to a reservoir model while each column is specific to the sampling method mentioned at the top of the column. The frequency
obtained for the LSPCM generated policy is scaled to the frequency of the LSM (using MCS) generated policy.

To measure the accuracy of the LSPCM, the Root Mean Square Relative Error (RMSRE) of the mean
VoF is used. This based on using the LSM mean VoF as a benchmark. The RMRSE is defined as follows:
(22)

,
where M is the total number of realizations, F̂i is the VoF obtained by the LSPCM for realization i and
F is the benchmark VoF (from the LSM). Fig. 10 shows the RMSRE results. A general trend of increaring
accrucay as LSPCM order increases is observed and it is clear that the LSPCM performance is reservoir
model unique.
Based on the results from Fig. 8 and Fig. 10, the performance of the LSPCM was further analysed with
respect to the LSM results. The following LSPCM results were chosen for this comparison: Fourth order
mapped Fejer sampling for the homogenous model, third order mapped Fejer sampling for the PUNQ-S3
model and second order Gaussian quadrature (full tensor product) sampling for the SPE10M model. The
results of this comparision are present in Fig. 11 which plots the optimal NPV, i.e. the optimal NPV
obtained by exercising the optimal policy, determined by the LSPCM and LSM method. The x-axis for
LSPCM is scaled to the number of realizations (103) used in the LSM method. The LSPCM aproximates
the optimal NPV trends relatively well, with the SPE10M model showing the best visual match.
In summary, the PCM was used to enhance the performance of the LSM method by using efficient
sampling methods. By reducing the number of realizations up to a factor of 12.3, the LSPCM can
16 SPE-170587-MS

Figure 10 —RMSE plots as a function of PCM order for the different reservoir models.

Figure 11—RMSE plots as a function of PCM order for the different reservoir models.

significantly reduce the computational requirement for performing the LSM method. The results of the
LSPCM were reservoir specific, which agrees with the results obtained by the original LSM method. The
optimal policy histograms also show that the LSPCM was able to produce, in general, the same trends that
were generated by the LSM for each reservoir model.
Given that standard probability distributions were used to definee the technical random input param-
eters, in reality these parameters might be better defined using arbitrary or categorical distributions. To
generate collocation nodes for these types of distributions, the PCEs of the random inputs can be
numerically constructed. The efficiency of the sampling method could be further improved by using
sparse grid techniques (Xiu and Hesthaven 2005; Foo et al. 2008) or by using sampling methods that have
a weak dependency on the order of the PCM approximation (Sarma and Xie 2011).
Uncertainty in heterogeneity is an important factor that could be incorporated into the evaluation
method. For instance, using PCM to approximate Gaussian or non-Gaussian generated permeability fields
rather than using MCS (Li and Zhang 2009). Finally, the LSPCM (and the LSM method) is easily
adaptable to other chemical EOR processes. This will only require a well-defined decision scenario and
the definition of the relevant uncertain parameters. The next section will disucss the extension of these
approaches to other chemical EOR processes, namely, polymer and alkaline-surfactant-polymer flooding.
These approaches are also easily scalable to include additional uncertain parameters.
SPE-170587-MS 17

Figure 12—VoF PDF estimates and optimal policy histograms produced by the LSPCM method using third, fourth and fifth order collocation nodes
and compared with that produced by the LSM method.

LSPCM Extension
This section will introduce extension of the LSPCM to other chemical EOR processes and to more
compelx decision scenarios.
Polymer Flooding Example
The LSCPM is applied to a polymer flooding process. The decision evaluated was that of timing the
polymer flood, similar to the decision evaluated above for the surfactant-polymer flooding cases. This
decision was evaluated for the PUNQ-S3 reservoir model using the same fluid data for this model. The
polymer flood consists of injecting a 100% PV polymer slug with a concnetrantion of 1% wt. Two
polymer related uncertian parameters were assumed, Dp (polymer adsorption) and Vp (viscosity multi-
plier). The LSPCM was performed for a third, fourth and fifth order sampling and compared with the LSM
method results. The results are presented in Fig. 12.
Results show that the Fejer sampling procudes a better visual agreement with respect to the PDF
estimates and preserves the same optimal policy trends obtained by the LSM method. This is confirmed
by the RMSE plot, see Fig. 13. The optimal NPV comparison plot (right plot, Fig. 13) was obtained using
the third order mapped Fejer LSPCM and shows that with 20 technical realizations it is possible to
accurately approximate the performance of the LSM method. It is clear that having flexibility in the timing
of the polymer flooding process for this reservoir model creates value. Based on the RMSRE plot, if the
3rd order mapped Fejer LSPCM is used for compaing the optimal NPVs, it is clear that the LSPCM
produces an accurate approximation to the LSM method.
Mutually Exclusive Decisions
In the previous exampels discussed above, a simple decision was evaluated. In reality, decisions can be
high-dimensional. In this section, an extension to more complex decision problems is introduced. This
18 SPE-170587-MS

Figure 13—RMSRE plot (left plot) for mean VoF obtained by the LSPCM for polymer flooding example assuming the mean VoF obtained by the
LSM method as the reference case. Results for optimal policy NPV (right plot) comparing the performance of the LSPCM algorithm with the LSM
algorithm for the PUNQ-S3 reservoir model for the ASP flooding decision example. The LSPCM results were obtained using 3rd order mapped Fejer
nodes. The optimal NPVs are sorted in ascending order and the number of realizations for LSPCM is called to 103 realizations.

extension will value the decision of having an additional flexibility (decision) embedded within the timing
decision. This is defined as a mutually exclusive decision problem.
If there are U available mutually exclusive embedded decisions (alternatives) at any decision time, y,
the optimal decision can be defined as a couple (␶, u), where ␶ is the optimal initiation time and u 僆 U,
is the embedded (additional) decision that is dependent on ␶. The Bellman dynamic programming equation
then becomes:
(23)

,
which expands to:
(24)

The decision rule along path x then becomes:


(25)

(26)

(27)

The continuation value is then defined as:


(28)

,
which is approximated by ⌽J. In order to apply the decision rule OV needs to be estimated. Since
Vu(yn,Pi,y (x),Ti(x)) is known for u ⫽ 1, . . ., U, then we have:
n

(29)

The convergence of the results is maintained because this is a direct extension of the LSM algorithm,
(Gamba 2002).
SPE-170587-MS 19

Table 2—Random input parameters and their distributions


Random Input Distribution Bounds

Residual oil saturation to surfactant flooding, fr Uniform [0,0.2]


Surfactant adsorption, g/g rock Log-Normal (-8,1)
Polymer adsorption, g/g rock Log-Normal (-8,1)
Polymer viscosity multiplier Uniform [1,10]
Alkaline adsorption, g/g rock Log-Normal (-8,1)

Figure 14 —VoF PDF estimate and policy histograms for the ASP flooding decision example for the PUNQ-S3 reservoir model. The first row shows
the PDFs of the VoF. The second row shows the optimal policy histograms of the initiation decision and the third row shows the histograms of the
optimal alkaline policy, the frequency is scaled to 103 realizations in order to be compared with the LSM generated results.

Complex Decision Case for Alkaline-Surfactant-Polymer Flooding


The extended LSPCM described in the previous section is applied to another chemical EOR process;
alkaline-surfactant-polymer flooding. For this example, the decision whether to inject alkaline with the SP
flood is coupled with the timing decision. The advantage of injecting alkaline is to reduce surfactant and
polymer adsorption and thus improve oil recovery (Lake 1989; Green and Willie 1998; Sheng 2011). This
20 SPE-170587-MS

Figure 15—RMSRE plot (left plot) as a function of PCM order for the PUNQ-S3 reservoir model for the ASP flooding decision example. Results for
optimal policy NPV (right plot) comparing the performance of the LSPCM algorithm with the LSM algorithm for the PUNQ-S3 reservoir model for
the ASP flooding decision example. The LSPCM results were obtained using 3rd order mapped Fejer nodes. The optimal NPVs are sorted in ascending
order and the number of realizations for LSPCM is called to 103 realizations.

is applied to the PUNQ-S3 reservoir model. An additional uncertian parameter, alkaline adsorption, is
incorporated in the LSM and LSPCM evaluations. It is assumed to follow the same probability distribution
as that of polymer and surfactant adsorption, see Table 2. Because five technical uncertainties are now
considered, this increases the dimensionality of the LSCPM thus limiting the maximum PCM order that
can be used. Therefore only 2nd and 3rd order Fejer nodes and 2nd order Gaussian quadrature (full tensor
product) nodes were used for the LSCPM. ECLISPE is used as the reservoir simulator for this chemical
EOR process. The alkaline is injected with the surfactant slug with a concentration of 1 % wt. The
economic model incorporates the alkaline cost3 as a uniformly distributed parameter~ U[1,2] $/lbm. See
Appendix C for the alkaline related properties used in the reservoir simulator. The VoF is obtained
assuming a static scenario of initiating the ASP flooding process at the start of field life.
The results, see Fig. 14, show reasonable visual matches of the LSPCM VoF PDF for the boxed and
mapped Fejer nodes compared with with the LSM PDF estimate. The same trends were maintained as
those observed for the LSM method with respect to the policy histograms. The results show that for most
realizations, SP flooding was favored as opposed to ASP flooding. Since the VoF is based on an ASP
flooding static scenario, this might indicate that the improvement in oil recovery due to reduced
surfactant/polymer adsorption does not justify the added cost of the alkaline chemical. Other factors
contribute to this such as the value of the alkaline dependent surfactant/polymer adsorption multipliers
used in the numerical simulator.
The RMSRE plot, see Fig. 15 (left plot), shows that all sampling methods used in the LSPCM achieve
a relatively accurate approximation of the LSM method. The third order Mapped Fejer nodes produce the
most accurate approximation. Using this approximation, Fig. 15 (right plot) compares the LSPCM with
the LSM appraoch. The plots show that the optimal policy obtained by these approaches always
maximizes the objective function when compared to the static policy. Furthermore, the comparison
between the optimal policy plot (far right plot) obtained by the LSM method and the LSPCM shows very
good agreement. Thus for a relatively more complex decision which incorporates another EOR process,
the LSM and LSPCM produce meaningful policies while taking into consideration technical and
economic uncertainties.

3
Alkaline cost bounds are obtained from Thomas (2006) and Sheng (2011).
SPE-170587-MS 21

Conclusion
There are many sources of uncertainty in implementing chemical EOR flooding processes. It is necessary
to have the ability and the tools to make decisions in the presence of this uncertainty in order too optimize
the performance of these EOR processes. The purpose of this study was to enhance the efficiency of the
LSM method in evaluating flexibility for chemical EOR flooding processes. The LSM method was
previously used to produce a near-optimal policy in the presence of uncertainty, both economic and
technical. The main advantage of the method is that it is straightforward and easily scalable to multiple
sources of uncertainty. However, the accuracy of the outcome is dependent on the number of realizations
used. The method uses Monte Carlo simulation to generate the stochastic realizations for the random input
parameters. This can incur prohibitive computational cost since reservoir simulation is used to quantify
the effect of the technical uncertainties. A new method was introduced, the LSPCM, which utilizes the
sampling techniques of the PCM to sample the technical state variables efficiently. This approach
improves the performance of the LSM method by using fewer realizations while maintaining the
convergence of the solution.
PCM is a robust ad efficient uncertainty quantification method (Tatang et al. 1997; Li et al. 2011).
Three sampling methods were used which were previously used to successfully quantify uncertainty for
surfactant-polymer flooding (Alkhatib and King 2013). The LSPCM was performed on three reservoir
model case studies: a simple homogeneous reservoir model, the PUNQ-S3 reservoir model and a modified
portion of the SPE10 reservoir model. Both methods were used to value flexibility in timing the
surfactant-polymer flooding as opposed to initiating the process at the start of field life. The accuracy of
the LSPCM results was determined by comparing the statistical outputs with those obtained from the LSM
method. The LSPCM did produce relatively accurate results for the PUNQ-S3 and modified SPE10
models while the results were not as accurate for the homogeneous model. By using the LSPCM, it was
possible to acheive an improvement factor of 12.3 for the modified SPE10 model by using the second
order quadrature nodes rather than the 103 realizations used for the LSM method.
Although it was possible to significantly reduce the computational requirement for applying the LSM
method, greater enhancements are still required in order to apply such methods to more realistic and
complex reservoir models. Improved sampling techniques were discussed.
The LSPCM algorithm was then extended to incorporate mutually exclusive decisions which can be
more representative of realistic decision scenarios. This extension was applied to an example problem that
incorporated an additional decision of injecting alkaline with the surfactant slug along with the timing
decision in order to reduce surfactant and polymer adsorption. This was applied to the PUNQ-S3 model.
The LSPCM was able to produce relatively accurate approximation to the LSM method. The results
showed that this flexibility did not create value and that most realizatios favoured SP flooding. Other
decisions could also be incorporated; for instance having flexibility in deciding between different
chemical slug control variables such as surfactant concentration or slug size. The algorithm can be
extended to incorporate such decision scenarios and convergence of the method is maintained because it
is a direct extension of the LSM approach (Gamba 2002).
Uncertainty in heterogeneity is another important factor that can be incorporated into the method. For
instance, PCM can be used to approximate Gaussian or non-Gaussian generated permeability fields (Li
and Zhang 2009) rather than using MCS approaches. The LSPCM (and the LSM method) can be easily
adaptable to other EOR processes as shown. This requires a well-defined decision scenario and the
selection of the relevant uncertain parameters.
22 SPE-170587-MS

References
Alkhatib,A, Babaei,M. and King,P. 2013. Decision Making under Uncertainty: Applying the Least
Squares Monte Carlo Method in Surfactant Flooding Implementation. SPE Journal 18(4): 721–735.
SPE-154467-PA.
Alkhatib,A. and King,P. 2013. Uncertainty Quantification of a Chemically Enhanced Oil Recovery
Process: Applying the Probabilistic Collocation Method to a Surfactant-Polymer Flood. Paper SPE
164244 presented at the 18th Middle East Oil and Gas Show and Conference, Manama, Bahrain, 10-13
March. http://dx.doi.org/10.2118/164244-MS.
Alkhatib,A. and King,P. 2014a. Robust quantification of parametric uncertainty for surfactant-
polymer flooding. Computational Geosciences 18(1): 77–101.
Alkhatib,A. and King,P. 2014b. An approximate dynamic approach to decision making in the presence
of uncertainty for surfactant-polymer flooding. Computational Geosciences 18(2): 243–263
Alsofi,A. and Blunt,M. 2011. The Design and Optimization of Polymer Flooding under Uncertainty.
Paper SPE 145110 presented at the SPE Enhanced Oil Recovery Conference, Kuala Lumpur, Malaysia,
19-21 July. http://dx.doi.org/10.2118/145110-MS.
Anderson,G., Delshad,M., King,C., Mohammadi,H. and Pope,G. 2006. Optimization of Chemical
Flooding in a Mixed-Wet Dolomite Reservoir. Paper SPE 100082 presented at the 2006 SPE/DOE
Symposium on Improved Oil Recovery, Tulsa, Oklahoma, USA, 22-26 April. http://dx.doi.org/10.2118/
100082-MS.
Bellman,R. 1957. Dynamic Programming. Princeton, New Jersey: Princeton University Press.
Bender,C. and Steiner,J. 2010. Least-squares Monte Carlo for backward SDEs. Numerical Methods in
Finance. 12:257–289.
Bratvold,R. and Begg,S. 2010. Making Good Decisions. Richardson, Texas: Society of Petroleum
Engineers.
Brown,C. and Smith,P. 1984. The Evaluation of Uncertainty in Surfactant EOR Performance Predic-
tion. Paper SPE 13237 presented at the 59th SPE Annual Technical Conference and Exhibition, Houston,
Texas, USA, 16-19 September. http://dx.doi.org/10.2118/13237-MS.
Cheng,H., Shook,G., Taimur,M. and Dwarakanath,V. 2012. Interwell Tracer Tests to Optimize
Operating Conditions for a Surfactant Field Trial: Design, Evaluation, and Implications. SPE Reservoir
Evaluation & Engineering 15(2):229 –242. SPE 144899-PA. http://dx.doi.org/10.2118/144899-PA.
Cheong,Y. and Gupta,R. 2005. Experimental Design and Analysis Methods for Assessing Volumetric
Uncertainties. SPE Journal 10(3): 324 –335. SPE 144899-PA. http://dx.doi.org/10.2118/80537-PA.
Choudhury,A., King,A., Kumar,S. and Sabharwal,Y. 2008. Optimizations in Financial Engineering:
The Least-Squares Monte Carlo method of Longstaff and Schwartz. Paper IEEE presented at the 2008
IEEE International Symposium on Parallel and Distributed Processing, Miami, Florida, USA, 14-18 April.
Christie,M. and Blunt,M. 2001. Tenth SPE Comparative Solution Project: A Comparison of Upscaling
Techniques. SPE Reservoir Engineering and Evaluation 4(4):308 –317. SPE 72469-PA. http://dx.doi.org/
10.2118/72469-PA.
Cortazar,G., Gravet,M. and Urzua,J. 2006. The valuation of multidimensional American real options
using the LSM simulation method. Computers and Operations Research 35(1): 113–129. http://
dx.doi.org/10.1016/j.cor.2006.02.016.
Dang,C., Chen,Z., Nguyen,N., Bae,W. and Phung,T. 2011. Development of Isotherm Polymer/
Surfactant Adsorption Models in Chemical Flooding. Paper SPE 147872 presented at the SPE Asia Pacific
Oil and Gas Conference and Exhibition, Jakarta, Indonesia, 20-22 September. http://dx.doi.org/10.2118/
147872-MS.
Dias,M. 2004, Monte Carlo Simulation of Stochastic Processes, http://www.puc-rio.br/marco.ind/
sim_stoc_proc.html (accessed 17 July 2012).
SPE-170587-MS 23

Dixit,A. and Pindyck,R. 1994. Investment under Uncertainty. Princeton, New Jersey: Princeton
University Press.
ECLIPSE Reservoir Engineering Software. 2010. Schlumberger, http://www.slb.com/content/
services/software/recent/.
ECLIPSE Technical Description, Version 2010.1. 2010. Schlumberger.
Energy Information Agency (EIA). 2012. NYMEX Futures Price Data. http://www.eia.gov/dnav/pet/
pet_pri_fut_s1_d.htm (accessed 10 July 2012).
Feinberg,J. 2012. Probabilistic Collocation Method Module (polychaos), https://bitbucket.org/jonathf/
polychaos (accessed 7 July 2012).
Field,R. and Mircea,D. 2007. Convergence Properties of Polynomial Chaos Approximations for L2
Random Variables, public report, Sandia National Laboratories.
Foo,J., Xiaoliang,W. and Karniadakis,G. 2008. The Multi-Element Probabilistic Collocation Method:
Error Analysis and Applications. Journal of Computational Physics 227(22): 9572–9595.
Gamba,Andrea. 2002. An Extension of Least Squares Monte Carlo Simulation for Multi-options
Problems. Paper presented at the 6th Annual Real Options Conference, Paphos, Cyprus. July 4-6, 2002.
Gautschi,W. 1994. Algorithm 726: ORTHPOL–a package of routines for generating orthogonal
polynomials and Gauss-type quadrature rules. ACM Transactions on Mathematical Software 20(1):
21–62.
Gerbacia,W. 1978. The Evaluation of Surfactant Systems for Oil Recovery Using Statistical Design
Principles and Analysis. Paper SPE 7070 presented at the SPE Symposium on Improved Methods of Oil
Recovery, Tulsa, Oklahoma, USA, 16-17 April. http://dx.doi.org/10.2118/7070-MS.
Ghanem,R., Spanos,P. 1991. Stochastic Finite Elements: A Spectral Approach. New York: Springer-
Verlag.
Glasserman,P. and Yu,B. Number of paths versus number of basis functions in American option
pricing, Ann. Appl. Probab. 14 (2004) 2090 –2119.
Golub,G. and Welsch,J. 1969. Calculation of Gauss Quadrature Rules. Mathematics of Computation
23(106): 221–230.
Hankins,N. and Harwell,J. 1996. Case Studies for the Feasibility of Sweep Improvement in Surfactant-
Assisted Waterflooding. Journal of Petroleum Science and Engineering 17(1):41–62. http://dx.doi.org/
10.1016/S0920-4105(96)00055-1
Hirasaki,G., Miller,C. and Puerto,M. 2011. Recent Advances in Surfactant EOR. SPE Journal
16(4):889 –907. SPE 115386-PA. http://dx.doi.org/10.2118/115386-PA.
Jafarizadeh,B. and Bratvold,R. B. 2009. Taking Real Options into the Real World: Asset Valuation
through Option Simulation. Paper SPE 124488 presented at the 2009 SPE Annual Technical Conference
and Exhibition, New Orleans, Louisiana, USA, 4-7 October. http://dx.doi.org/10.2118/124488.
Karambeigi,M., Zabihi,R. and Hekmat,Z. 2011. Neuro-simulation modelling of chemical flooding.
Journal of Petroleum Science and Engineering 78(2):208 –219. http://dx.doi.org/10.1016/
j.petrol.2011.07.012.
Kossack,C. and Bilhartz,H. 1976. The Sensitivity of Micellar Flooding to Reservoir Heterogeneities.
Paper SPE 5808 presented at the SPE Improved Oil Recovery Symposium, Tulsa, Oklahoma, USA. 22-24
March. http://dx.doi.org/10.2118/5808-MS.
Lake,L. W. 1989. Enhanced Oil Recovery. Englewood Cliffs, New Jersey: Prentice Hall.
Li,H. and Zhang,D. 2006. Probabilistic collocation method for flow in porous media: Comparisons
with other stochastic methods. Water Resources Research 43(9). W09409, http://dx.doi.org/doi:10.1029/
2006WR005673.
Li,H. and Zhang,D. 2009. Efficient and Accurate Quantification of Uncertainty for Multiphase Flow
with the Probabilistic Collocation Method. SPE Journal 14(4): 665–679. http://dx.doi.org/doi:10.2118/
114802-PA
24 SPE-170587-MS

Li,H., Sarma,P. and Zhang,D. 2011. A Comparative Study of the Probabilistic-Collocation and
Experimental-Design Methods for Petroleum-Reservoir Uncertainty Quantification. SPE Journal 16(2):
429 –439. doi: 10.2118/140738-PA
Lin,G. and Tartakovsky,A. 2009. An Efficient, High-Order Probabilistic Collocation Method on
Sparse Grids for Three-Dimensional Flow and Solute Transport in Randomly Heterogeneous Porous
Media. Advances in Water Resources 32(5): 712–722.
Loeven,G. and Bijl,H. 2008. Probabilistic Collocation Used in a Two-Step Approach for Efficient
Uncertainty Quantification in Computational Fluid Dynamics. Computer Modelling in Engineering and
Science 36(3): 193–212.
Loeven,G., Witteveen,J. and Bijl,H. 2007. Probabilistic collocation: an efficient non-intrusive ap-
proach for arbitrarily distributed parametric uncertainties. Paper AIAA 2007-317 presented at the 45th
AIAA Aerospace Sciences Meeting and Exhibit, Reno, NA, 8-11 January.
Lohne,A. and Fjelde,I. 2012. Surfactant Flooding in Heterogeneous Formations. Paper SPE 154178
presented at the Eighteenth SPE Improved Oil Recovery Symposium, Tulsa, Oklahoma, USA, 14-18
April. http://dx.doi.org/10.2118/154178-MS.
Longstaff,F. A. and Schwartz,E. S. 2001. Valuing American Options by Simulation: A Simple
Least-Squares Approach. The Review of Financial Studies 14(1): 113–147. http://dx.doi.org/10.1093/rfs/
14.1.113.
Mathelin,L. and Hussaini,M. 2003. A stochastic collocation algorithm for uncertainty analysis.
National Aeronautics and Space Agency, http://ntrs.nasa.gov/archive/nasa/casi.ntrs.nasa.gov/
20030016674_2003020380.pdf (accessed 29 June 2012).
Matlab®, version 7.14. 2012. Natick, Massachusetts: The Mathworks, Inc.
Moreno,M. and Navas,J. 2003. On the Robustness of Least-Squares Monte Carlo (LSM) for Pricing
American Derivatives. Review of Derivatives Research 6(2):107–128.
Onorato,G., Loeven,G., Ghornaiasl,G., Bijl,H. and Lacor,C. 2010. Comparison of intrusive and
non-intrusive polynomial chaos methods for CFD applications in aeronautics. Paper 1809 presented at the
European Conference on Computational Fluid Dynamics, ECCOMAS CFD, Lisbon, Portugal, 14-17 June.
Pope,G., Wang,B. and Tsaur,K. 1979. A Sensitivity Study of Micellar/Polymer Flooding. SPE Journal
19(6):357–368. SPE 7079-PA. http://dx.doi.org/10.2118/7079-PA.
Powell,W. B. 2011. Approximating Value Functions. In Approximate Dynamic Programming: Solving
the Curses of Dimensionality. Chap 8. Princeton, New Jersey: John Wiley and Sons.
Puerto,M., Hirasaki,G., Miller,C. and Barnes,J. 2012. Surfactant Systems for EOR in High-
Temperature, High-Salinity Environments. SPE Journal 17(1):11–19. SPE 129675-PA. http://dx.doi.org/
10.2118/129675-PA.
PUNQ-S3 Reservoir model data set. http://www3.imperial.ac.uk/earthscienceandengineering/
research/perm/punq-s3model/onlinedataset (accessed 2012).
PYTHON, version 2.7.3. 2012. Python Software Foundation, http://www.python.org/download/
releases/2.7.3/.
Sarma,P. and Xie,J. 2011. Efficient and robust uncertainty quantification in reservoir simulation with
polynomial chaos expansions and non-intrusive spectral projection. Paper SPE 141963 presented at the
SPE Reservoir Simulation Symposium, The Woodlands, TX, 21-23 February.
Smith,W. 2010. On the Simulation and Estimation of the Mean-Reverting Ornstein-Uhlenbeck
Process. http://commoditymodels.files.wordpress.com/2010/02/estimating-the-parameters-of-a-mean-
reverting-ornstein-uhlenbeck-process1.pdf (downloaded 9 January 2012).
Solairaj,S., Britton,C., Kim,D., Weerasooriya,U. and Pope,G. 2012. Measurement and Analysis of
Surfactant Retention. Paper SPE 154247 presented at the Eighteenth SPE Improved Oil Recovery
Symposium, Tulsa, Oklahoma, USA, 14-18 April. http://dx.doi.org/10.2118/154247-MS.
SPE-170587-MS 25

Stentoft,L. 2004. Assessing the Least Squares Monte-Carlo Approach to American Option Valuation.
Review of Derivatives Research 7(2):129 –168.
Tatang,M., Pan,W., Prinn,R. and McRae,G. 1997. An efficient method for parametric uncertainty
analysis of numerical geophysical models. Journal of Geophysical Research 102(18): 21925–21932.
Thomas,S. 2006. Chemical EOR: The Past – Does It Have a Future? Paper SPE 108828 based on a
speech presented as an SPE Distinguished Lecture during the 2005-2006 season.
Uhlenbeck,G. and Ornstein,L., 1930. On the Theory of the Brownian motion. Physical Review 36(5):
823–841. http://link.aps.org/doi/10.1103/PhysRev.36.823
Weiss,W. and Baldwin,R. 1985. Planning and Implementing a Large-Scale Polymer Flood. Journal of
Petroleum Technology 37(4): 720 –730. http://dx.doi.org/10.2118/12637-PA.
Wiener,N. 1938. The homogeneous Chaos. American Journal of Mathematics 60(4): 897–936.
Willigers,B. J. A. and Bratvold,R. B. 2009. Valuing Oil and Gas Options by Least-Squares Monte
Carlo Simulation. SPE Projects, Facilities & Construction 4(4): 146 –155. http://dx.doi.org/10.2118/
116026-PA.
Wyatt,K., Pitts,M. and Surkalo,H. 2008. Economics of Field Proven Chemical Flooding Technologies.
Paper SPE 113126 presented at the SPE/DOE Symposium on Improved Oil Recovery, Tulsa, Oklahoma,
USA, 20-23 April. http://dx.doi.org/10.2118/113126-MS.
Xiu,D. and Hesthaven,J. 2005. High-order collocation methods for differential equations with random
inputs. SIAM Journal on Scientific Computing 27(3): 1118 –1139.
Xiu,D. and Karniadakis,G. 2002. The Wiener-Askey polynomial chaos for stochastic differential
equations. SIAM Journal on Scientific Computing 24(2): 619 –644.
Zhang,J., Delshad,M., Sepehnoori,K. and Pope,G. 2005. An Efficient Reservoir-Simulation Approach
To Design and Optimize Improved Oil-Recovery-Processes With Distributed Computing. Paper SPE
94733 presented at the SPE Latin American and Caribbean Petroleum Engineering Conference, Rio de
Janeiro, Brazil, 20-23 June. http://dx.doi.org/10.2118/94733-MS.
Zheng,C., Gall,B., Gao,H. and Bryant,R. 2000. Effects of Polymer Adsorption and Flow Behavior on
Two-Phase Flow in Porous Media. SPE Reservoir Evaluation and Engineering 3(3): 216 –223. http://
dx.doi.org/10.2118/64270-PA.
26 SPE-170587-MS

Appendix A
Case Studies Simulation Constraints and Surfactant Properties

Simulation constraints used for the PUNQ-S3 and SPE10M reservoir models are shown in Table 3. The producer oil rate and
bottom hole pressure (BHP) constraints are applied to each of the four producers in these two models. The dead pore space
factor used for the polymer was 0.16 and a residual resistance factor of 1.5 was assumed for all models. The adsorption of the
surfactant and polymer were assumed to be irreversible.

Table 3—Simulation Constraints for PUNQ-S3 and SPE10M Case Table 5—Surfactant/Oil Surface Tension5
Studies
Surfactant concentration Surfactant Water/Oil Surface
Constraint PUNQ-S3 SPE10M (lb/stb) Tension (lb/in)

Producer Oil Rate, Bbl/d 200 150 0 100


Injector Water Rate, Bbl/d 1750 2500 0.1 50
Water Cut, % None None 0.2 10
BHP min, psia 250 250 0.5 0
Max Injection Pressure, psia 4500 8000

Table 4 —Surfactant Viscosity4 Table 6 —Capillary De-Saturation Curve6


Surfactant Concentration Surfactant Water Log10 (Capillary Number) Miscibility Function
(lb/stb) Viscosity (centipoise)
-9 0
0 0.4 2 0
0.1 0.5 5 1
0.2 0.6 10 1
0.8 0.8

4
Data source is ECLIPSE Technical Description 2010.1.
5
Data source is ECLIPSE Technical Description 2010.1.
6
Data source is ECLIPSE Technical Description 2010.1.
SPE-170587-MS 27

Appendix B
Economic Uncertainty Modelling

Pois modelled as a risk-neutral Ornstein-Uhlenbeck process (Uhlenbeck and Ornstein 1930) which is a mean reverting
stochastic process:
(30)

Where:
Wt, is a Brownian-Motion,
␭ is the measure of the speed of mean reversion.
␮ is the long term mean that the process reverts to.
␴ is the measure of the process volatility.
␦ is the risk premium.
To simulate the oil price using the above equation it is necessary to discretize it with respect to time, the following provides
an exact discrete-time expression as:
(31)

This discrete-time version of the process allows an exact discretization in that accuracy is not reduced if a larger time step
is used (Dias, 2005). A more detailed discussion of why the above discretization of a mean-reverting process is chosen can be
found in Dixit and Pindyck (1994). Mean reverting stochastic processes are well suited for modelling commodity prices
because they reflect the economic argument that when prices are too high, supply will increase and demand will reduce. When
prices become too low, the opposite effect occurs, pushing prices towards the long term mean (Smith 2010). Fig. 16 shows the
price model. The parameters in Table 7 were used for the homogeneous reservoir model. For the other two models, the same
parameter values were also used except for the value of t which is 240 in this case because the simulation horizon was twenty
years.
The capital expenditure is adjusted for each reservoir model by scaling to the maximum water injection rate occurring in
the simulation.

Figure 16 —Ornstein-Uhlenbeck Price Model, showing the mean of 1000 paths and the corresponding standard deviation.
28 SPE-170587-MS

Table 7—Ornstein-Uhlenbeck Process Parameters7 Table 8 —Capital Expenditure Data8


t, months 120 Facility Cost ($) 1,500,000
dt, months 1 Laboratory and Engineering Design ($) 700,000
␭ 0.0012 Incremental Operating Cost ($/Month) 10,000
␮ $/Bbl 66.67 Water Injection Rate Reference (Bbl/day) 5,000
␴ 1.81
␦ 0.03
P0o, $/Bbl 90

7
The ␭, ␮ and ␴ parameters were obtained using the Least Squares method to estimate parameters of an
observed Ornstein-Ulhenbeck Process (Smith, 2010). NYMEX futures prices from April 1983-July 2012
were used as the basis of the estimation (EIA, 2012).
8
The capital expenditure data was obtained from Wyatt et al. (2008).
SPE-170587-MS 29

Appendix C
Alkaline Properties

Table 9 —Surfactant/Polymer Adsorption Multipliers as a function of Table 10 —Water/Oil Surface Tension Multipliers as a Function of
Alkaline Concentration Alkaline Concentration
Alkaline Alkaline Alkaline-Surfactant
Concentration Alkaline-Surfactant Alkaline-Polymer Concentration (% wt.) Adsorption Multiplier
(% wt.) Adsorption Multiplier Adsorption Multiplier
0 1.0
0 1.0 1.0 0.028 0.90
0.028 0.99995 0.99995 0.057 0.90
0.057 0.99500 0.99500 0.14 0.90
0.28 0.50 0.50

You might also like