Journal of Petroleum Science and Engineering 109 (2013) 1–11

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

journal homepage: www.elsevier.com/locate/petrol

Application of multi-criterion robust optimization

in water-flooding of oil reservoir
Elham Yasari a, Mahmoud Reza Pishvaie a,n, Farhad Khorasheh a, Karim Salahshoor b,
Riyaz Kharrat b
a
Sharif University of Technology, Chemical and Petroleum Engineering Department, Azadi Ave, P.O. Box 11155-9465, Tehran, Iran
b
Petroleum University of Technology, Ahwaz, Iran

art ic l e i nf o a b s t r a c t

Article history: Most of the reported robust and non-robust optimization works are formulated based on a single-
Received 3 April 2012 objective optimization, commonly in terms of net present value. However, variation of economical
Accepted 29 July 2013 parameters such as oil price and costs forces such high computational optimization works to regenerate
Available online 8 August 2013
their optimum water injection policies. Furthermore, dynamic optimization strategies of water-flooding
Keywords: often lack robustness to geological uncertainties. This paper presents a multi-objective while robust
Multi-objective optimization optimization methodology by incorporating three dedicated objective functions. The goal is to determine
Multi-criterion optimization optimized and robust water injection policies for all injection wells. It focuses on reducing the sensitivity
Water flooding to the uncertainty in the model and objective function parameters when no measurement information is
NSGA-II
assumed to be available. This work also, utilizes a derivative-free Evolutionary Multi-objective
Injection rate
Optimization (EMO) procedure in the form of a Non-dominated Sorting Genetic Algorithm (NSGA)
Oil reservoir
Model/parameter uncertainty which attempts to find a robust Pareto-optimal solution without a priori knowledge of the reservoir
dynamic models. Some modifications have been introduced to the original NSGA-II code to handle the
constraints of the optimization problem. The comparative test studies clearly demonstrate superiority of
the proposed methodology to give optimal robust solutions under geological uncertainties with much
less standard deviations and variances. Furthermore, the optimization results demonstrate less
sensitivity to the imposed time-varying economical parameters such as operation costs and oil price,
revealing non-dependency of the introduced multi-objective functions.
& 2013 Elsevier B.V. All rights reserved.

1. Introduction function, known as production or net present value, has been

considered as the objective function in all the previous investiga-
1.1. Water-flooding process tions reported in the literature.

Fundamentally, water-flood involves pumping water through a

well (injector) into the reservoir. The water is forced through the
pore spaces and sweeps the oil towards the producing wells
(producers). The percentage of water in the produced fluids
steadily increases until the cost of removing and disposing of
water exceeds the income from oil production. After this point, it
becomes uneconomical to continue the operation and the water-
flooding is stopped. On the average, about one-third of the original
oil in place (OOIP) is recovered, leaving two-thirds behind after
secondary recovery. Product optimization of water-flooding has
shown a significant potential to increase ultimate recovery
(Brouwer and Jansen, 2004; Jansen et al., 2005; Sarma et al.,
2005; Wang et al., 2007; Sarma et al., 2008). A single objective
n
Corresponding author. Tel.: +98 2166166429.
E-mail address: pishvaie@sharif.edu (M.R. Pishvaie)
1.2. Uncertainties and robust optimization
Dealing with uncertainty is an important topic encountered in
many fields related to modeling and control. Reducing the uncer-
tainty itself, using measurements and reducing the sensitivity to
the uncertainty are two different strategies which are not basically
conflicting with each other (Van Essen et al., 2009).
Beyer and Sendhoff (2007) in their survey classified modeling of
the uncertainties to deterministically, probabilistically, or possibilisti-
cally. The deterministic type defines parameter domains in which the
uncertainties can vary; the probabilistic type defines probability
measures describing the likelihood by which a certain event occurs.
The possibilistic type defines fuzzy measures describing the possibility
or membership grade by which a certain event can be plausible or
believable. However, a publically-accepted remedy for various tech-
nologies that suffer from vast uncertainties of any above-mentioned

0920-4105/$ - see front matter & 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.petrol.2013.07.008
2 E. Yasari et al. / Journal of Petroleum Science and Engineering 109 (2013) 1–11
Nomenclature
b discount rate (%/year)
E expected value operator
f objective functions
g inequality constraints
h equality constraints
J objective function
J modified objective function
k time-step counter
K total number of time steps
N number of wells
NT total number of realizations
NR number of small set of realizations
q flow rate (m3/day)
r risk aversion factor – price ($/m3)
t time (day)
tk time at time step k (day)
Δtk time interval of time step k (day)
x manipulated variable
θd finite number of realizations
classes – yet alone with limited measurements – is the use of a
so-called robust optimization.
Since the uncertainties in upstream petroleum industries are
high, conventional optimization strategies are not amenable to
carry over the optimization task deliberately.
Van Essen et al. (2009) presented an approach to reduce the
impact of geological uncertainties in the field development phase
known as a robust optimization (RO). Their proposed RO scheme
uses a set of realizations that reflect the range of possible geological
structures honoring the statistics of the geological uncertainties. The
associated objective function was NPV in terms of a single objective
with pre-defined costs and oil price. They used a classical gradient-
based optimization method where the gradients were obtained with
an adjoint formulation.
The approach of Alhuthali et al. (2010) relies on equalizing arrival
time of the waterfront at all producers using multiple geologic
realizations. They account for geologic uncertainty using two optimi-
zation schemes: a stochastic form which relies on a combination of
expected value and standard deviation combined with a risk attitude
coefficient. This approach is some sort of scalarization of a bi-objective
optimization problem which can be solved by single objective
optimizer engines. Their approach was the analytical computation of
the gradient and Hessian of the objective function.
Almeida et al. (2010) presented an evolutionary algorithm-
based decision support system able to optimize intelligent well
control, in intelligent oil fields, under technical and geological
uncertainties. A genetic algorithm was used for obtaining a pro-
active control strategy and determining an operation that max-
imized the single objective net present value (NPV).
The objective of Chen and Hoo (2012) is to optimize oil
production by managing the amount of water added to a reservoir.
This management is accomplished by employing an optimal
model-based control framework that includes uncertain para-
meter updating and a particular low-order model identified from
a first-principles model.
Nevertheless, to the best of our knowledge, none of the
reported works consider uncertainty in the objective function
parameters. As it is clear, time varying parameters in the objective
function, especially oil price could force a huge persuasion to
recalculate the time-consuming optimization work. In this paper
by introducing three specific objective functions, a new approach
θR representative realizations set
Θ unknown uncertainty space
τI reference time (365 days)
Σ θd standard deviation
Φ volumetric ratio
Subscripts
c, o, p cumulative oil production
c, w, i cumulative water injection
i, well injection well
l, p liquid production rate
o oil
o, Total total oil in place
p, well production well
RO robust optimization
wi injected water
wp produced waters
w, p water production rate
will be proposed to calculate the robust Pareto-front in the defined
multi-objective optimization problems.
Uncertainty of any geological model due to insufficient data is
an inherent characteristic. To cope with the geological uncertainty
in the reservoir, a number of realizations that are equally probable
and reflect the range of possible geological structures, are gener-
ated and robust optimization can be performed based on these
realizations using the specific objective functions.
Reducing the uncertainty using measurement is known as
history matching. In the present study reducing the sensitivity to
the uncertainty was the main goal and optimization was per-
formed in the absence of measurement. Reducing the sensitivity
for two types of uncertainty has been covered through this work:
uncertainty in the reservoir model and uncertainty in the parameters
of the objective functions.
This paper addresses the secondary recovery phase of a
petroleum reservoir using water-flooding based on a multi-
objective robust optimization scheme.
2. Optimization algorithm
Most efficient methods used in solving optimization problems
rely on explicit knowledge of the underlying simulator equations to
compute the gradient of the objective function. As a result of large
and complicated nature of reservoir models with large number of
unknowns and non-linear constraints, the software for gradient
calculations will be very tedious and time-consuming to create for
practical optimization problems. Yet, another major drawback of the
gradient-based methods using adjoint equations is that it requires
explicit knowledge of the simulation model equations describing
the dynamic behavior of the system. By using derivative-free
methods like Genetic Algorithm (GA), no knowledge of the simu-
lator equations is required and the simulator can be run as a black
box. GA does not require any derivative information and is less
likely to be trapped in local minima. It also has the ability to
optimize discrete (and thus non-differentiable) variables such as the
control settings. GA has therefore been utilized as an influential tool
for solution of various problems in reservoir engineering.
Unlike the gradient-based methods, GA typically converges
slowly and becomes inefficient when a large number of variables
 E. Yasari et al. / Journal of Petroleum Science and Engineering 109 (2013) 1–11
are to be optimized. It is therefore common to use proxy models
such as neural network and capacity resistance model (CRM)
combined with genetic algorithms to reduce the number of calls
to the reservoir simulator. Reducing the number of calls to the
simulator leads to a reduction in the computational time. Though
not employed in our study, parallel processing is another efficient
way to cope with the computational runtime. GA has great
potential for parallel processing as each individual in the popula-
tion can be processed separately.
2.1. Application of genetic algorithm in oil industry
Genetic algorithm (GA) is one of the stochastic optimization
methods which simulates the process of natural evolution. GA is
applicable to multi-objectives optimization and can handle con-
flicts among objectives. Therefore, it is robust where multiple
solutions exist. In addition, it is efficient and easy to use (Velez-
Langs, 2005). Of the stochastic algorithms, GA appears to be the
most commonly used for various petroleum engineering applica-
tions including production optimization and history matching.
Harding et al. (1998) employed GA in production scheduling for
a group of linked oil and gas fields to maximize the total net
present value (NPV). Fichter (2000) applied GA to the portfolio
optimization in oil and gas field. Hajizadeh (2007) employed a GA
based model for viscosity prediction and compared their results
with two Iranian PVT reports for a field located in Khuzestan and
three fluid characterization reports. Anderson (2009) used genetic
algorithms (GAs) and artificial neural networks (ANNs) to perform
and augment production optimization of Brugge case. The Brugge
case is a realistic large-scale model that has been used for
academic purposes. The author has concluded that the developed
schemes are suited for solving optimization problems in the
petroleum industry. Emerick et al. (2009) described a GA approach
for the simultaneous optimization of number, location and trajec-
tory of producer and injector wells. By introducing this approach,
they developed software to be applied and tested on three
full-field reservoir models based on real cases. As a result they
were able to significantly increase the NPV and cumulative oil
production.
2.2. Multi-criterion/multi-objective optimization
Multi-criterion optimization also known as multi-objective
optimization problem (MOOP) requires simultaneous optimization
of several incommensurable and often competitive and conflicting
objectives. Because of the presence of multiple conflicting objec-
tives, it is not possible to find a single optimal solution which
satisfies all goals. Instead, the solutions exist in the form of
alternative trade-offs, also known as the Pareto optimal solutions.
Generation of Pareto optimal solutions is a challenging task and
often difficulties arise in using the conventional methods
(Malekmohammadi et al., 2011). Rahman et al. (2001) formulated
a global objective to maximize production and NPV while mini-
mizing treatment cost by summing four objective functions for
hydraulic fracturing. Ray and Sarker (2006) used non-dominated
sorting genetic algorithm (NSGA-II) to minimize the injected gas
and maximize the produced oil in gas-lift operation. Cardoso
(2009) used a linear weighting approach to solve a multi-
objective problem where he formulated a single objective function
in order to minimize water injection and maximize oil production.
He varied the relative weighting of two terms to generate the
convex portion of the Pareto front. In a recent work, Van Essen
et al. (2010) used lexicographic optimization method to optimize
multiple economic objectives. They have considered NPV and daily
production as a bi-objective optimization problem. In the
3
investigations on multi-objective optimizations mentioned above,
however, uncertainty has not been taken into consideration.
Multi-objective optimization approach has been widely used in
other engineering fields, but its application in petroleum engineer-
ing is still limited.
Traditionally, to cope with multi-objective problems, conven-
tional methods such as weighted sum have been used that have
some disadvantages. The conventional optimization methods such
as weighted sum, Dynamic Programming (DP), Linear Program-
ming (LP), and non-linear programming (NLP) are not suitable for
solving MOOPs since these methods use a point-by-point
approach and their outcome is a single optimal solution. These
methods convert the MOOP into a single objective optimization.
By using a single pair of fixed weights, only one point on the
Pareto front can be obtained. Therefore, if one would like to obtain
the global Pareto optimum, all possible Pareto fronts must first be
derived. Conventional methods may also face problems if the
optimal solution lies on non-convex or disconnected regions of the
function space (Deb, 2001).
Recently, meta-heuristic techniques such as evolutionary algo-
rithms (EAs) and swarm intelligence techniques have become
increasingly popular for solving multi-objective reservoir optimi-
zation problems.
In water flooding operations, the optimization problem with
single-objective function requires maximization of the net present
value or recovery factor by manipulation of some parameters in
the reservoir system, e.g. well location, well bottom-hole pressures
(BHP), liquid production or injection rates or valve setting in an
on-off mode. But, a multi-objective optimization problem may
have a number of conflicting objectives (Deb, 2001):
8
>
> Min=Max ðf 1 ðxÞ; f 2 ðxÞ; …; f M ðxÞÞ;
>
>
< subject to g j ðxÞ Z 0; j ¼ 1; 2; 3; …; J
ð1Þ
>
> h ðxÞ ¼ 0; k ¼ 1; 2; 3; …; K
> Lk
>
: x r xi r xU ; i ¼ 1; 2; 3; …; n
i i
The problem of production optimization hence requires max-
imization or minimization of some objective function fm(x). In this
optimization problem, the objective functions to be maximized are
the NPV and/or cumulative oil production while those to be
minimized are water production and/or water cut. Here, x denotes
a set of controls (decision variables) with upper and lower bounds,
which may include bottom-hole pressures, flow rates, choke size,
etc. These controls may be manipulated in order to achieve an
optimum value at which the objective functions are maximized (or
minimized). The maximum water injection, minimum BHP, max-
imum water cut, etc., represent constraints (gj(x) and hk(x)) in the
optimization problem. A dynamic simulation model is required to
relate the objective functions to the set of controls.
The NSGA-II procedure (Deb, 2001; Deb et al., 2002) is one of
the popularly used Evolutionary Multi-objective Optimization
(EMO) procedures which attempt to find multiple Pareto-optimal
solutions in a multi-objective optimization problem.
This algorithm provides a flexible and modular framework
which is capable of solving various kinds of multi-objective
optimization problems.
NSGA employs several layers of classification for individuals by
sorting the population based on Pareto dominance concept. As a
consequence, all the solutions in a Pareto front are better than the
excluded ones (Deb, 2008).
2.3. Robust optimization
Uncertainty is an inherent characteristic of any geological
model due to insufficient data. Therefore, the unique true
4 E. Yasari et al. / Journal of Petroleum Science and Engineering 109 (2013) 1–11
distribution of reservoir properties will remain unknown. To
quantify the uncertainty in a geological model, geostatistical
methods can be used through construction of multiple equally
probable realizations of reservoir properties (Van Essen et al.,
2009). This can be understood as discretizing the unknown
uncertainty space Θ, resulting in a finite number of realizations
θd as
θd ¼ fθ1 ; θ2 ; …; θNT g A Θ
where NT indicates the total number of geostatistical realizations.
Suppose that the performance measure is J, which appears as a
! In this work, {min, P20, …, P80, max} realizations are used such
function of optimization parameters x and geostatistical realiza-
tions θd:
! statistics denotes the value of a variable below which a certain
J ¼ Jð x ; θd Þ ð3Þ
The most straightforward RO objective will be based on the
expected outcome over the set of realizations. Other objectives
may involve incorporating the variance of the outcomes or a
worst-case approach. The objective function in robust optimiza-
tion JRO, can be defined as follows (Van Essen et al., 2009) to
account for the uncertainty represented by multiple realizations:
h i h i
! !
J RO ¼ Eθd Jð x ; θd Þ rΣ θd Jð x ; θd Þ where
h i
! 1 NT !
Eθd Jð x ; θd Þ ¼ ∑ Jð x ; θi Þ ð4Þ
NT i ¼ 1
where Eθd represents expected value over the discretized uncer-
tainty space θd , Σ θd is standard deviation, and r is the risk aversion
factor. Assuming that the realizations are equal probable, the
robust optimization (RO) objective function can be further
written as:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
NT 1 NT !
1 !
maxJ RO ¼ Jr: ∑ ðJð x ; θÞJÞ2 where J ¼ ∑ Jð x ; θi Þ
N T 1 i ¼ 1 NT i ¼ 1
An optimal risk weighted solution that gives good performance
for any realization in a given set is sought by decision-makers in
robust optimization.
Generally speaking, the robust optimization objective consists
of two components: expected value and standard deviation of the
desired objective function.
The variable r is the risk aversion factor that weights the trade-
off between the expected value and the standard deviation
(Howard, 1971). According to the defined objective function (Eq.
(5)), a positive r means that the decision maker is risk averse and a
negative r means risk prone. A zero risk coefficient indicates risk
neutral for decision maker. Risk aversion factor has been consid-
ered positive in this study.
2.4. Realization ranking for robust optimization
Each RO optimization function evaluation requires a simulation
run for each individual realization. Thus, the simulation time for NT
realizations will be NT times longer than the time needed for
optimization based on a single realization. Apparently, it can be
costly to use all the NT realizations in the robust optimization
process.
Ranking scheme provides a useful geostatistical tool that is
being widely used for reservoir analysis where significant varia-
tions are present in the reservoir properties. On the other hand,
randomly choosing a small set of geostatistical realizations will not
accurately represent the uncertainty (Fenik et al., 2009). After
these NT realizations are accurately ranked, a small set of repre-
sentative realizations can be chosen for robust optimization,
nominated as
θR ¼ fθ1 ; θ2 ; …; θNR g A θd ¼ fθ1 ; θ2 ; …; θNT g A Θ ð6Þ
With this small set of representative realizations, θR ¼
fθ1 ; θ2 ; …; θNR g, the robust optimization objective function can be
written as
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð2Þ 1 NR
! 1 NR !
J RO ¼ Jr ∑ ðJð x ; θi ÞJÞ2 where J ¼ ∑ Jð x ; θi Þ ð7Þ
NR 1 i ¼ 1 NR i ¼ 1
that the mean and standard deviation of this set remain as close as
possible to the original set. P stands for percentile which in
percent of observations fall. The solution of these selected realiza-
tions will be propagated to the original set of NT realizations. Such
ranking methods are generally used by Yang et al. (2011) for
steam-assisted gravity drainage (SAGD) and the results are shown
to be acceptable (Fenik et al., 2009).
In our study we assumed that no production data was available
for history matching and updating the model. The main objective
of the present work was to reduce the sensitivity to the uncer-
tainty. In the presence of production data, the goal would be to
reduce the sensitivity itself by model updating. In the absence of
production data, NPV was used as a selection criterion for choos-
ing the representative set. If production data is available, then the
representative set would include the realizations that have the
closest match. The multi-objective robust optimization process
would otherwise be the same. In our optimization work, a multi-
objective optimization of a water-flood process is developed in
dynamic mode which is naturally capable of dealing with time-
varying water injection and liquid production scenarios. The
novelty of our proposed approach is that it can deal with
uncertainty in the geological model and economic parameters
ð5Þ using a multi-objective optimization approach which has not been
addressed in any previous investigation.
3. Methodology
The developed robust and non-robust optimization works are
usually formulated in terms of single-objective optimization
criterion in which net present value is the most usual candidate.
However, due to variation of oil price, this type of optimization
with a pre-defined set of relevant parameters should be renewed
to obtain the up-to-dated optimum water injection polices and
this means losing time and money. To cater for this issue, this
paper presents a methodology to optimize specific objective
functions for a water-flood reservoir which leads to optimized
injected water rates. Three specific objective functions have been
introduced so that relevancy of optimization on the costs and
specially oil price is omitted.
On the other side, one of the major challenges in the develop-
ment of an optimized production strategy concerns the way the
impact of reservoir uncertainty is addressed. Most optimized
strategies are model-based and hence are effective if the model
is able to correctly predict the future reservoir behavior. This work
focuses on reducing the sensitivity to the uncertainty when no
measurement information is assumed to be available. This has
been done by introducing three specific objective functions that
are independent of economical parameters and also account for
geological uncertainty in the objective functions. Furthermore,
incorporating a ranking scheme based on NPV of Base Case, a small
set of realizations is selected in order to implement the developed
robust optimization with lower computation time. Results of
robust optimization based on small set (single and multi) will be
 E. Yasari et al. / Journal of Petroleum Science and Engineering 109 (2013) 1–11 5

propagated on NT realizations. This multi-objective robust optimi- Table 1

zation problem is solved using a Non-dominated Sorting Genetic Objective functions definitions.
Algorithm (NSGA-II) optimization method (Deb et al., 2002). Some
Objective Function equations
modifications were implemented to the original code to handle
the constraints of the optimization problem. f1 M j ¼ N
∑j ¼ 1 p;well qo;j
V FOP ¼ ∑ Δt k ð1þb=100Þtk
k¼1
j ¼ N
f2 M ∑j ¼ 1 i;well qwi;j
PVFWI ¼ ∑ Δt k
4. Case-study: synthetic five spot reservoir k¼1
ð1þb=100Þt k
j ¼ N
f3 M ∑j ¼ 1 p;well qwp;j
PVFWP ¼ ∑ Δt k ð1þb=100Þt k
4.1. Reservoir description k¼1
j ¼ N j ¼ N j ¼ N
f4 M ∑j ¼ 1 p;well qo;j r o ∑j ¼ 1 p;well qwp;j r w ∑j ¼ 1 i;well qwi;j r i
NPV ¼ ∑ Δt k ð1þb=100Þt k
A water-flood case study of a 2-dimensional reservoir model k¼1

with five-spot structure, over a time horizon of about 11 years

(4000 days) is considered. The model dimension is 580 m 
553 m  61 m (I  J  K). It is modeled with 525 grid blocks with
approximately dimension of 30  30  61 (DI  DJ  DK). The aver-
age reservoir initial pressure is around 335 bar. The reservoir has a
pay zone of about 61 m with initial oil saturation being equal to
1.00 (just oil).
The reservoir is highly heterogeneous with high and low
permeability zones. To capture geological uncertainties of the
permeability distribution, a set of 100 geological realizations of
the reservoir was generated using a Gaussian geostatistical simu-
lation method. Since these realizations are based on the same set
of variograms and no production data is assumed to be available to
reduce the uncertainty by model updating, all these realizations
are considered to be equi-probable. Fig. 1 shows the horizontal
permeability field for realization 050 where the horizontal perme-
ability range is 10.03–999.04 md.
The total number of decision variables to be adjusted is the
product of the number of the decision variables to be updated in
time (control steps) and the total number of wells in the reservoir
model. In this simulation test study, there are 5 control steps
and 4 injection wells that are planned to be adjusted via their
injection rates, so the total number of controls or optimization
variables is 20.
4.2. Optimization methodology
4.2.1. Objective functions
In this study, robust optimization is used to reduce sensitivity
of the model to the uncertainty. In this regard, no measurement
information is assumed to be available. In order to omit relevancy
of the results to the time-varying price and cost, a multi-objective
optimization method is developed introducing three specific
objectives (f1, f2, f3) as defined in Table 1. To derive these functions
from f4 (NPV equation), costs and oil price were assumed to be
semi-constant during the optimization horizon. These parameters
Fig. 1. Schematic of five-spot water-flood reservoir.
Table 2
Optimization problem constraint.
Constraints Values
Maximum liquid production rate (m3/day) 1750
Minimum BHP of production well (bar) 152
Maximum water injection rate (m3/day) 1591
Maximum BHP of injection well (bar) 345
were, therefore, taken out of each of the summations and the three
specified objective functions were defined using the three parts of
the NPV equation. After the optimization process is performed and
robust water injection profile was determined, NPV was calculated
using each injection policy and given economic parameters.
The results of the robust multi-objective optimization in terms
of the expected values of f1, f2, f3 will be compared to the robust
single-objective problem with NPV (Table 1) as the objective
function, (expected value of f4) with pre-known costs.
4.2.2. Optimization parameters
To optimize the relevant objective functions (i.e., {f1, f2, f3},
and f4), the injection rates for all the individual injector wells are
considered as optimization parameters. Total water injection and
total liquid production represent the constraints that should be
satisfied. Minimum and maximum injection rates and minimum
and maximum production and injection well pressures are
regarded as bounded constraints, that have been summarized in
Table 2.
4.2.3. Robust optimization algorithm
The developed robust optimization algorithm integrates the
approach proposed by van Essen et al. (2009) with a ranking
scheme that has been presented by Yang et al. (2011) for SAGD
operation. Our work is different in that we have introduced three
specific objective functions that are independent of economical
parameters and also include geological uncertainty. Furthermore,
all previous robust optimization works reported in the literature
were based on optimization using all of the generated realizations,
whereas our approach is based on a multi-objective robust
optimization using representative realizations. In this procedure,
three types of test cases have been considered as: Base Case,
Nominal Case, and Robust Case which are briefly explained below.
For both Dynamic Single Objective Robust Optimization (DSORO)
and Dynamic Multi- Objective Robust Optimization (DMORO), the
procedure remains the same.
Base Case: in this case, Base Data (equal injection rates for all
wells) are exercised on NT realizations. Then, NPV objective
function (f4) is calculated for each realization and the obtained
6 E. Yasari et al. / Journal of Petroleum Science and Engineering 109 (2013) 1–11

NPVs are sorted from low to high. The results of this section are
then utilized to choose the Nominal Case.
Nominal Case: the nominal optimization (NO) approach is
based on a single realization. However, because none of the
realizations in the set of 100 is preferred over the others, the
decision of which realization to use in the NO approach becomes
an arbitrary one. The Base Case NPV data has been selected for the
Nominal Case. Finding P50 NPV and its equivalent realization from
sorted data, the Nominal Case is defined. After this, P50 model is
optimized and injection policy of this optimization is propagated
on NT realizations.
Robust Case: the robust optimization (RO) approach has been
reported to work based on optimization of all the generated
realizations. Thus, the time which is needed for robust optimiza-
tion is NT times the time for optimizations done based on one
realization.
In this section, 9 realizations, {min, P20, …, P80, max}, have
been selected and the optimization is then performed using the Fig. 2. NPV histogram of Base Case.
robust optimization equations. The results of the optimization are
exercised on the NT realizations. The main advantage of this kind
of robust optimization, compared to the one proposed by van
Essen et al. (2009), is less run time requirement. Mean and
standard deviation of this smaller set is almost similar to the Base
Set as indicated in Table 3.
Table 3 compares mean and standard deviation of NPV for the
entire NT realizations and the selected 9 representative realiza-
tions. The comparison shows that mean and standard deviation is
close to each other. This indicates that 9 representative realiza-
tions are sufficient for estimating reliable statistics for the entire
NT realizations based on NPV criterion.

5. Results and discussion

5.1. Base Case based on Base Data (BCBD)

In Base Case (BC), flow rates for all injection wells are assumed
to be the same. Injection rate for each individual well is taken to be
equal to 2500 STB/day. It is assumed that the rate of injection does Fig. 3. FOE histogram of Base Case.
not change during the life of reservoir, implying that the injection
conditions remain static. By utilizing the injection conditions on
the NT realizations, the NPV is calculated. Discount rate has been
included in the NPV calculations. NPV, Field Oil Efficiency (FOE),
Volumetric Ratio (Φ) and Field Water Cut (FWCT) are calculated for
each realization and the results are accordingly shown within
histograms. Definitions of these variables are represented in
Eqs. (8)–(10) and the Base Case histograms are shown Figs. 2–4.
qc;w;i
Φ¼ ð8Þ
Pore volume
Σ qc;o;p
FOE ¼ ð9Þ
qo;Total

qw;p
W UC ¼ ð10Þ
ql;p

The preliminary data corresponding to production cost, water

injection rate, and oil price which are needed to calculate the NPV

Table 3 Fig. 4. FWCT histogram of Base Case.

Mean and standard deviation of small and original sets.

Mean (M$) Standard deviation (M$)

have been summarized in Table 4. This table will also be used in
Original set 1128.2 28.2244
the single-objective optimization.
Smaller set 1123.6 37.3340 The volumetric ratio is a criterion to indicate the total injected
water against pore volume. Hence, a value higher than one means
 E. Yasari et al. / Journal of Petroleum Science and Engineering 109 (2013) 1–11 7

Table 4
Cost and price.

Parameters Value

Oil price, $/m3 ($/STB) 565.6820 (90)

Prod. water cost, $/m3 ($/STB) 113.1364 (18)
Inj. water cost, $/m3 ($/STB) 56.5682 (9)
Discount ratio (%/year) 10

Table 5
Single objective optimization properties.

Parameters Values

Optimization variable Water injection rate

Fig. 6. FOE box plot of nominal and Robust Case.
Optimization steps 5
No. optimization variable 20
No. objective functions 1
No. population 60
No. generations 30
Objective function f4
Robust objective function Eq. (7)

Fig. 7. Box plot of nominal and Robust Case.

Fig. 5. NPV (M$) box plot of nominal and Robust Case.

that the total injection of water is higher than the total pore
volume indicating more water is injected. Volumetric ratio in the
BC for the generated realizations is 1.2807. Field oil efficiency (FOE)
represents the total oil production against Total Oil In Place (TOIP).

5.2. Dynamic single objective robust optimization (DSORO) – f4

Characteristics of this type of optimization are summarized in

Table 5. As was mentioned before, P50 realization is optimized and
the results of this optimization are propagated on NT realizations. Fig. 8. FWCT box plot of nominal and Robust Case.
In the next stage, 9 realizations ({min, P20, …, P80, max}) are
chosen out of the entire 100 to carry out the robust optimization In this single objective robust optimization problem according
and the obtained water injection polices are then propagated on to Tables 1, 5 and previous explanation the desired robust
the entire realizations. Figs. 5–8 illustrate the results for the objective function will be Eq. (11)
nominal case in which the Robust Cases have been depicted sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
NR 1 NR
within the box plots. A box plot (also known as a box-and- 1 ! !
f 4RO ¼ f 4 r ∑ ðf ð x ; θi Þf 4 Þ2 where f4 ¼ ∑ f ð x ; θi Þ
whisker diagram or plot) is a convenient way to graphically depict N R 1 i ¼ 1 4 NR i ¼ 1 4
groups of numerical data through their five-number summaries: ð11Þ
the smallest observation (sample minimum), lower quartile (Q1),
median (Q2), upper quartile (Q3), and largest observation (sample where f 4 is defined in Table 1.
maximum). A box plot may also indicate which observations, if In comparison with the Base Case, better results have been
any, might be considered as outliers. obtained in the optimization section. The result in Fig. 5 implies
In the single-objective optimization, the objective function is that the NPV in the robust optimization is better than the Nominal
considered to be NPV (Table 5) that needs to be maximized. Case optimization, especially for the worst-case (minimum NPV).
8 E. Yasari et al. / Journal of Petroleum Science and Engineering 109 (2013) 1–11

follows:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
NR 1 NR
1 ! !
f 1RO ¼ f 1 r ∑ ðf ð x ; θi Þf 1 Þ2 where f1 ¼ ∑ f ð x ; θi Þ
N R 1 i ¼ 1 1 NR i ¼ 1 1
ð12Þ
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
NR 1 NR
1 ! !
f 2RO ¼ f 2 r ∑ ðf 2 ð x ; θi Þf 2 Þ2 where f2 ¼ ∑ f ð x ; θi Þ
N R 1 i ¼ 1 NR i ¼ 1 2
ð13Þ
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
NR 1 NR
1 ! !
f 3RO ¼ f 3 r ∑ ðf ð x ; θi Þf 3 Þ2 where f3 ¼ ∑ f ð x ; θi Þ
N R 1 i ¼ 1 3 NR i ¼ 1 3
ð14Þ

By introducing these specific objective functions the sensitivity

Fig. 9. Total water injection profile (m3/day).
of the optimization problem to the uncertainties in model and
objective function will be reduced.
Table 6 Robust Pareto fronts have been depicted in Fig. 10 after
Multi objective optimization properties.
propagating the water injection polices (water injection profiles)
Parameters Values on the NT realizations in terms of the obtained expected values of
the corresponding objectives. It is noted that the individual f1, f2
Optimization variable Water injection rate and f3 objective functions have already been defined in Table 6.
Optimization steps 5
As can be seen in the Pareto front, moving in direction of
No. optimization variable 20
No. objective functions 3 increasing one objective leads to decrease in the other one. So, this
No. populations 60
No. generations 60
Objective functions (f1, f2, f3) f 1; f 2; f 3
Robust objective functions Eq. (7)

Furthermore, close examinations of FOE (Fig. 6) and Φ values

(Fig. 7) reveal that better results have been obtained via the robust
optimization. The robust optimization has improved the NPV by
adjusting water injection profile. It can be seen from Fig. 7 that the
volumetric ratio ðΦÞ in the Robust Case is lower than the Nominal
Case. This means that higher NPV is not necessarily accomplished
by an increase in the production. There is also lower water cut in
the robust optimization compared with the Nominal Case, as
demonstrated in Fig. 8. Lower variances of the boxplots in the Fig. 10. Pareto front of multi objective optimization. (For interpretation of the
four Robust Cases can clearly be differentiated in all four figures. references to color in this figure, the reader is referred to the web version of this
The results of RO are comparatively better than the NO cases due article.)
to the fact that more than one realization is used in RO
optimization.
Injection profiles for all the three cases (BC, NO, RO) are shown
in Fig. 9.
Having explored the performance of the robust optimization
using single-objective optimization, the proposed multi-objective
robust optimization methodology is examined under the same test
scenarios.

5.3. Dynamic multi objective robust optimization

(DMORO) – (f1, f2, f3)

Table 6 shows the assumed experimental data for the multi-

objective optimization problem. Three objective functions have
been realized in this case, aiming to omit the relevancy of the
optimization problem to the prices. So, the optimization results
can be treated as generalized solutions without any dependency
on the probable changes due to injected and produced water costs
and time-varying oil price. Therefore, there is no need to carry out
re-optimization under such time-varying conditions, resulting into Fig. 11. Propagated NPV for single (big blue plus, cross and dot) and multi (small
black pluses, crosses and dots) objective optimization (Nominal Case) for three
time and money savings. different oil prices ((a) 188.5607 $/m3, (b) 565.6820 $/m3, (c) 879.9497 $/m3). (For
According to Tables 1 and 6, the objective functions of interpretation of the references to color in this figure legend, the reader is referred
this multi-objective robust optimization strategy are defined as to the web version of this article.)
 E. Yasari et al. / Journal of Petroleum Science and Engineering 109 (2013) 1–11
Fig. 12. Propagated NPV for single (big blue square) and multi (small red crosses)
objective optimization (Robust Case) for oil price 565.6820 $/m3. (For interpreta-
tion of the references to color in this figure legend, the reader is referred to the web
version of this article.)
Fig. 13. Propagated NPV (blue plus), FOE (green circle) and FWCT (red star) for
single (bigger plus, circle and cross) and multi (bigger plus, circle and cross)
objective robust optimization. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)
Fig. 14. Standard deviation of nominal (blue circles) and robust (red crosses) propagated (Expected) NPV for multi objective solutions at oil prices: 188.5607, 565.6820,
879.9497 $/m3. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
9
trade off forms a front and the solutions that are in Pareto front are
the ones that cannot be dominated by the other solutions (non-
dominated solutions). In our multi-objective problem, f1 should
increase while the other two objective functions (f2, f3) should
decrease, approaching towards the higher NPV.
Blue dots in Fig. 10 depict propagation of the objective func-
tions in terms of expected values for the robust optimization that
is based on 9 realizations. These values indicate the case after
propagating the water injection polices (water injection profiles)
on NT realizations.
As noted earlier, the multi-objective optimization results are
free of costs and prices and hence NPV is easily calculated every
time the prices are changed without any need to carry out further
optimization. But, on the contrary, different optimizations are
needed for different costs and prices in single-objective optimiza-
tion with NPV objective function.
The propagated NPV (expected NPV) has been shown in Fig. 11
for Nominal Case due to single- and multi-objective optimizations.
There have been three optimization scenarios for respective oil
prices of 188.5607, 565.6820 and 879.9497 ($/m3), that is 30, 90
and 140 ($/STB) respectively, in single optimization and the results
have been compared with 60 solutions (water injection polices) of
multi-objective optimization just with a very simple algebraic
calculation. As can be seen, there are various solutions (water
injection polices), demonstrating better, identical or close enough
in comparison with the single-objective optimization that can be
chosen without any further need to perform new optimization.
Fig. 12 depicts the comparative results due to oil price 565.6820
($/m3) propagated (expected) NPV for Robust Case in single-
and multi-objective optimization. Examining the outcomes
implies better results of multi-objective optimization and robust
optimization.
The propagated NPV, FOE and FWCT representations for single-
and multi-objective robust optimizations have been illustrated in
Fig. 13 where the NPV has been calculated based on oil price of
565.6820 ($/m3).
One interesting point can be observed in all the illustrated
figures to differentiate between nominal and robust propagated
charts. Robust optimization results represent less sensitivity with
lower variations in different solutions (a smaller range of possible
outputs). This observation coincides with the robust optimization
10 E. Yasari et al. / Journal of Petroleum Science and Engineering 109 (2013) 1–11

Fig. 15. Variance of nominal (blue circles) and robust (red crosses) propagated (Expected) NPV for multi objective solutions at oil prices: 188.5607, 565.6820, 879.9497 $/m3.
(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 16. Propagated (Expected) NPV of nominal (blue circles) and robust (red crosses) for multi objective solutions at oil prices: 188.5607, 565.6820, 879.9497 $/m3. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

approach in which it is aimed to simultaneously optimize perfor-
mance of a design and minimize its sensitivity (i.e., maximize
robustness), resulting into a trade-off between these two
objectives.
Figs. 14 and 15 illustrate standard deviation and variance of the
results corresponding to the two cases (NO, RO) for multi-
objective optimization at three different oil prices. These figures
show lower standard deviation and variance for the Robust Case
compared with the Nominal case which is often demanding in
these charts. A low standard deviation indicates that the data
points tend to be very close to the mean, whereas high standard
deviation indicates that the data are spread out over a large range
of values. The variance is hence used as a measure of how far a set
of numbers are spread out from each other.
The NPV for the three levels of oil prices together with all the
multi-objective optimization solutions have been shown in Fig. 16.
It can be clearly observed that the robust optimization results are
improved in general, especially for higher oil price and the worst-
case (minimum NPV).
6. Summary and conclusion
As a result of large and complicated nature of reservoir models
with large number of unknowns and non-linear constraints, the
optimization approach based on gradient calculations will be very
tedious and time consuming to be realized for practical optimiza-
tion problems. Yet, the explicit knowledge requirement of the
simulation model equations, describing the dynamic system beha-
vior, imposes another major drawback of the gradient-based
methods using adjoint equations.
 E. Yasari et al. / Journal of Petroleum Science and Engineering 109 (2013) 1–11
This paper addresses an interesting approach in which the
dynamic multi-objective robust optimization problem is solved
using a derivative-free EMO procedure in the form of a revised
NSGA, called as NSGA-II, to find multiple Pareto-optimal solutions
without a priori knowledge of the reservoir dynamic models.
Reservoir uncertainty introduces a challenging issue in develop-
ment of an optimized production strategy. Most of the optimiza-
tion strategies are model-based and hence are effective if their
assumed models are able to correctly predict the future reservoir
behavior. This imposes practically a challenging objective to be
maintained. That is why the paper aims to focus on reducing the
sensitivity to the uncertainty when no measurement information
was assumed to be available. For this purpose, a robust optimiza-
tion methodology was configured to incorporate three specific
objectives functions in a multi-objective optimization framework.
The developed robust and non-robust optimization works are
usually formulated in terms of single-objective optimization
criterion in which net present value (NPV) is the most usual
candidate. However, due to variation of oil price, this type of
optimization with pre-defined parameters should be renewed to
obtain updated optimum water injection polices and this means
losing time and money. Time horizon of the optimization problem
also could be a degree of freedom to specify the life time of the
multi-objective optimization problem to encounter with the
uncertainties in the objective function parameters. Defining spe-
cific objective functions in this work eliminated dependency of the
optimization results on the time-varying parameters such as costs
and oil prices. The introduced multi-objective scheme has been
compared with a dynamic single-objective optimization problem,
leading to better improvement for Robust Case in terms of less
sensitivity and better performance aspects. Moreover, the pro-
posed methodology incorporated a scheme based on NPV ranking
of the realizations, resulting into selection of a small set of
realizations. As a consequence, optimization time becomes less
than the conventional robust optimization approaches.
References
Alhuthali, A.H., Datta-Gupta, A., Yuen, B., Fontanilla, J.P., 2010. Optimizing smart
well controls under geologic uncertainty. J. Pet. Sci. Eng. 73, 107–121.
Anderson, M.G., 2009. Reservoir Production Optimization Using Genetic Algorithms
and Artificial Neural Networks. Department of Computer and Information
Science, NTNU. (Master's thesis).
Almeida, L.F., Vellasco, M.M.B.R., Pacheco, M.A.C., 2010. Optimization system for
valve control in intelligent wells under uncertainties. J. Pet. Sci. Eng. 73,
129–140.
11
Beyer, H.G., Sendhoff, B., 2007. Robust optimization—a comprehensive survey.
Comput. Methods Appl. Mech. Eng. 196, 3190–3218.
Brouwer, D.R., Jansen, J.D., 2004. Dynamic optimization of water flooding with
smart wells using optimal control theory. SPE J. 9, 391–402.
Cardoso, M.A., 2009. Reduced-order models for reservoir simulation. In: Proceed-
ings of the SPE International Student Paper Contest, Annual Technical Con-
ference and Exhibition, 4–7 October. New Orleans, Louisiana, USA.
Chen, Y., Hoo, K.A., 2012. Model parameter uncertainty updates to achieve optimal
management of a reservoir. Control Eng. Pract. 20 (10), 1042–1057.
Deb, K., 2001. Multi-Objective Optimization using Evolutionary Algorithms. John
Wiley & Sons, London.
Deb, K., 2008. A robust evolutionary framework for multi-objective optimization.
In: Proceedings of the 10th Annual Conference on Genetic and Evolutionary
Computation-GECCO, pp. 633–640.
Deb, K., Agrawal, S., Pratap, A., Meyarivan, T., 2002. A fast and elitist multi-objective
genetic algorithm: NSGA-II. IEEE Trans. Evol. Comput. 6 (2), 182–197.
Emerick, A., Silva, E., Messer, B., Almeida, L., Szwarcman, D., Pacheco, M., Marley, M.,
2009. Well Placement Optimization Using a Genetic Algorithm with Nonlinear
Constraints.SPE 118808-MS.
Fenik, D.R., Nouri, A., Deutsch, C.V., 2009. Criteria for Ranking realizations in the
investigation of SAGD reservoir performance. In: Presented at the Canadian
International Petroleum Conference (CIPC).
Fichter, D., 2000. Application of Genetic Algorithms in Portfolio Optimization for
the Oil and Gas Industry. SPE 62970-MS.
Hajizadeh, Y., 2007. Viscosity prediction of crude oils with genetic algorithms. SPE
106763-MS.
Harding, T., Radcliffe, N., King, P., 1998. Hydrocarbon Production Scheduling With
Genetic Algorithms. SPE 36379-PA.
Howard, R.A., 1971. Proximal decision analysis. Manage. Sci. 17 (9), 507–541.
Jansen, J.D., Brouwer, D.R., Naevdal, G., van Kruijsdijk, C.P.J.W., 2005. Closed-loop
reservoir management. First Break 23, 43–48.
Malekmohammadi, B., Zahraie, B., Kerachian, R., 2011. Ranking solutions of multi-
objective reservoir operation optimization models using multi-criteria decision
analysis. Expert Syst. Appl. 38, 7851–7863.
Rahman, M.M., Rahman, M.K., Rahman, S.S., 2001. An integrated model for multi-
objective design optimization of hydraulic fracturing. J. Pet. Sci. Eng. 31, 41–62.
Ray, T., Sarker, R., 2006. Multi-objective evolutionary approach to the solution of
gas lift optimization problems. In: Proceedings of the IEEE Congress on
Evolutionary Computation, July 16–21. Vancouver, Canada.
Sarma, P., Aziz, K., Durlofsky, L.J., 2005. Implementation of adjoint solution for
optimal control of smart wells. In: Proceedings of the SPE Reservoir Simulation
Symposium. Woodlands, Texas, U.S.A. SPE 92864.
Sarma, P., Chen, W., Durlufsky, L., Aziz, K., 2008. Production optimization with
adjoint models under non-linear control-state path inequality constraints. SPE
Reservoir Eng. Eval. J. 11 (2), 326–339.
Van Essen, G.M., Zandvliet, M.J., Van den Hof, P.M.J., Bosgra, O.H., Jansen, J.D., 2009.
Robust water-flooding optimization of multiple geological scenarios. SPE J. 14
(1), 202–210.
Van Essen, G., Van den Hof, P.M.J., Jansen, J.D., 2010. Lexicographic optimization of
multiple economic objectives in oil production from petroleum reservoirs. In:
Proceedings of the American Control Conference (ACC), pp. 1223–1228.
Velez-Langs, O., 2005. Genetic algorithms in oil industry: an overview. J. Pet. Sci.
Eng. 47, 15–22.
Wang, C., Li, G., Reynolds, A.C., 2007. Production Optimization in Closed Loop
Reservoir Management. SPE 109805, S.
Yang, C., Card, C., Nghiem, L., Fedutenko, E., 2011. Robust optimization of SAGD
operations under geological uncertainties. In: SPE Reservoir Simulation Sym-
posium, 21-23 February 2011, The Woodlands, Texas, USA, 141676-MS.