SPE International Student Paper Contest
Real-time Optimization of Smart Wells
Inegbenose Aitokhuehi, Stanford University
This paper was prepared for presentation at the 2004 SPE International Student Paper
Contest at the SPE Annual Technical Conference and Exhibition being held in Houston, Texas,
26-29 September 2004.
This paper was selected for presentation by merit of placement in a regional student paper
contest held in the program year preceding the International Student Paper Contest. Contents
of the paper, as presented, have not been reviewed by the Society of Petroleum Engineers
and are subject to correction by the author. The material, as presented, does not necessarily
reflect any position of the Society of Petroleum Engineers, its officers, or members.
Abstract
Smart wells are wells that have downhole instrumentation,
such as sensors and valves, on the production tubing. These
wells provide the ability for both downhole monitoring and
control. Downhole monitoring can be achieved through the
use of sensors while control is realized with downhole valves.
Once a smart well is deployed, valves can be used to
independently control each segment / branch of the well in a
reactive mode, such as shutting off a zone once it starts
producing water, or in a defensive mode, which requires the a
priori determination of valve settings. Using the latter
approach, which is the method applied in this work, valve
settings are determined through an optimization procedure.
We show with this procedure that well instrumentation can
provide over 50% gain in cumulative oil recovery over the
uninstrumented case for systems considered here in which the
geology is assumed to be known. Because the geology is not
known in real applications, we couple the valve optimization
procedure with history matching techniques, in which we use
idealized sensor data to update the reservoir description. Up
to 90% of the gain attainable with known geology is achieved
for the unconditionally and conditionally generated models
considered. In addition, we show that it is beneficial to use
multiple history-matched models for the optimization in some
cases. This is because multiple history-matched models
capture the geologic uncertainty better than single
history-matched models. We also introduce efficient
alternative procedures to improve the speed of the overall
technique. These include the use of a Levenberg-Marquardt
algorithm for the optimizations.
Introduction
A conventional well is a vertical or slightly deviated well.
Horizontal, highly deviated and multilateral wells are
generally referred to as nonconventional or advanced wells
(NCWs). Smart wells are wells (both conventional and
nonconventional) that have been instrumented with downhole
sensors and valves on the production tubing. This
instrumentation provides the ability to independently monitor
and control each segment (or branch) of the well and can
therefore be used to maximize oil recovery or to minimize the
production of unwanted water or gas. A direct outcome of this
is reduced well intervention, particularly in subsea wells.
Smart wells are especially beneficial for multiple reservoirs
where the main production strategy is to commingle
production. These zones are independently controlled by
inflow control valves but produce into the same production
string. Downhole valves can be appropriately (or optimally)
located within the drainhole of a horizontal well to provide
significant gains. This is especially beneficial for a horizontal
drainhole that contacts both sand and shale. With multilateral
wells, because the laterals are completed as open holes in most
cases, it is appropriate to install the valves at the junction of
the lateral and the mainbore.
Smart Well Control Optimization. Once a smart well has
been deployed, maximizing its recovery (or net present value,
NPV) requires finding the optimum settings of the installed
valves. Different methods exist for finding the optimum valve
settings. Because valves can be operated either in a reactive
mode or in a defensive (or predictive) mode, previous methods
can be classified under these two modes of control.
The reactive mode of control1, 2 is a type of control where
zones (well segments / branches) with high water cut and / or
high gas-oil ratio (GOR) problems are isolated by continuous
reduction or outright closure of valve openings. With this
mode of control, downhole sensor data is used directly to
identify zones producing unwanted water and / or gas. The
static optimization method used by Brouwer et al.3 to
waterflood a reservoir can be classified as a reactive mode of
control. In their approach, the valves were operated in an
on-off manner and were closed once water breakthrough
occurred at the producers. The graphical means used by Gai4
to optimize valve settings for a multilateral multi-zone
completion can also be classified as reactive control. Gai
generated straight line inflow performance relationships and
valve performance relationships to optimally produce from the
laterals. One requirement of this method is a means to update
the inflow and valve performance relationships whenever well
conditions change. Arenas and Dolle5 applied a pressure
cycling principle to waterflood a tight fractured reservoir.
They essentially considered a two-dimensional model
completed with a horizontal producer and injector. The
reservoir contained a set of fractures spanning the two wells.
Valves were installed in the injector only and used in the
reactive mode to optimize recovery. They were operated in an
on-off fashion and were closed or opened once water cut
measured at the producer reached a predefined upper or lower
limit, respectively.
2 SPE Student Paper
Reactive control has been applied in real applications. The
operation of smart well valves to maximize recovery in two
Asian fields (reported by Snaith et al.6) essentially utilized
reactive control. Also, Al-Khodhori7 reported the
implementation of a reactive smart quad-lateral well
technology to enhance production and recovery in an Oman
field.
Instead of reacting to problems occurring at the wells, a
defensive mode of control1, 8 can be adopted. This mode of
control entails an optimization and requires that valve settings
be determined a priori using a predictive reservoir and well
model. The aim of defensive control is an early mitigation of
problems that might otherwise occur at the wells. Dolle et al.9
utilized optimal control theory to dynamically determine the
valve settings. They considered the same reservoir and
rate-constrained wells as in their previous work3. With this
method of defensive control, they were able to improve sweep
and recovery over their previous method of static
optimization.
More recently, Yeten1 and Yeten et al.8 used a nonlinear
conjugate gradient method in conjunction with a commercial
simulator that contains a detailed multi-segment well model10
to optimize valve settings defensively. Their method was
based on a fixed reservoir model during the optimization
process, although sensor data allow for continuous model
updating. As a result, their method was most appropriate for
use as a smart well screening tool. In their well placement
work, Yeten et al.11 modeled the operational reliability of
downhole valves. They applied a utility risk function to model
downhole valve failure and explored its effect on recovery.
History Matching. History matching is an ill-conditioned
inverse problem resulting in nonunique solutions. History
matching (HM) techniques can be grouped into traditional HM
methods and automatic HM methods (AHM). The traditional
methods examine the HM problem in a series of scales. They
first match the pressure by tuning the directional
transmissibilities and the pore volume and then adjust the
relative permeability curves to obtain a saturation match.
These approaches assume a single (or very few) models. They
are very time consuming and rely on experience.
Kabir et al.12 compared the traditional HM method with
two forms of the automatic method using a real oil and gas
field. They used the Gauss-Newton optimization algorithm in
the two AHMs to minimize a least-square type error between
the observed data and the simulated data. Based on their
results, they concluded that the AHM approaches reduced the
time required to complete a HM study even with multiple
parameter adjustments. In addition, they stated that the AHMs
preserved the HM objectivity.
Caers13 proposed another form of AHM, the
probability-perturbation method. The method is facies based
and conditions a model to production data by perturbing an
initial geologic realization iteratively until a match is
achieved. The geologic models are generated with
multiple-point geostatistics using a training image, which is an
unconstrained conceptual image of the reservoir. As a result,
the probability-perturbation method is suitable for diverse
geological structures including channels and fractures. The
method honors the geology because the models are also
conditioned to a specified training image event. Caers
modeled the magnitude of the facies perturbation with a single
parameter determined through a one-dimensional
optimization. Hoffman and Caers14 extended the
probability-perturbation method to handle regional facies
distributions with known within-facies petrophysical
properties. They divided the reservoir model into regions and
perturbed each of them independently to determine the overall
facies distribution that matches the history data. Suzuki15 also
extended the probability-perturbation method to treat channel
systems and developed a method of perturbing the unknown
within-facies petrophysical properties in conjunction with the
facies perturbation.
In this paper, we present the combined use of history
matching and valve optimization to maximize recovery. The
valve optimization procedure follows the approach presented
by Yeten1, 8. It however utilizes the sensor data to update
(history-match) the geological model, which is then used to
optimize the valve settings in real time. Our history matching
procedure utilizes the combined works of Hoffman and
Caers14, with the reservoir specified as a single region, and
that of Suzuki15. We then apply the overall valve optimization
and history matching (VOHM) procedure to several cases.
We first consider a simple variogram based model with a
segmented horizontal well. Next, we consider channelized
models with a quad-lateral smart well. In this case, both
unconditioned and conditioned history matched models are
used. For all of the cases investigated, our results show that
substantial increase in cumulative recovery over the
uninstrumented case can be achieved with the VOHM
procedure. As much as 90% of the gain attainable with known
geology was realized. The implementation of a procedure that
combines valve optimization with history matching has not
been previously reported.
Well Optimization with History Matching
Our method makes use of the downhole sensors and valves in
a smart well. The data from the sensors can be used to
periodically or continuously update the geological model.
Using the updated model, valve settings can be determined via
optimization by maximizing a specified objective function. In
order to evaluate the objective function, it is necessary to
perform simulations that accurately model the smart well.
ECLIPSE16, 17 was used as the function evaluator because of
its detailed multi-segment well option, which includes the
capability to model smart wells10.
Linking Valve Optimization and History Matching. We
link both the valve optimization and history matching
techniques according to the flow diagram shown in Figure 1.
A full description of the overall valve optimization and history
matching algorithm (VOHM) is provided by Aitokhuehi18.
We can summarize the procedure as follows:
a) Divide the entire simulation period into a number of
steps (these steps are different than the simulation
time steps). Specify initial geological model(s). This
can be a single model or an ensemble of models and
can be conditioned to well data.
SPE Student Paper
b) Use this model(s) in the optimization module to find
a single set of valve settings that optimizes recovery
(in an average sense for the ensemble case). The
module returns the valve settings and the optimized
recovery.
c) Apply the valve settings from the optimization
module to wells in a reference model (this would be
the sub-surface reservoir in real applications and is
here assumed to be unknown). Run the reference
model for a period of time with these new valve
settings to generate data. Pass the generated data to
the history matching module.
d) Run the history matching module with the data
obtained from the reference model (using the actual
valve settings) to generate single or multiple history-
matched (or updated) models.
e) Loop through steps (b) to (d) for each optimization
step until the entire simulation period is covered.
Initial Model
Condition to well data
Optimization Module New valve settings
Optimize valve settings
Updated Data Generation
model Data
Apply new valve settings
to well(s) and generate data
History Matching Module
Update geological model
Figure 1: Flow diagram of the overall valve control
optimization and history matching procedure.
The initial model is used only for the first step of the entire
simulation period. Subsequent steps use updated model(s)
obtained from the history matching module. These models are
expected to better reflect the characteristics of the actual
(unknown) model. Multiple initial models can also be
specified. In all cases, a single set of valve settings is
generated. These settings optimize the recovery on average
for the specified models.
The data generation module is used to simulate the
sub-surface reservoir. The target model is treated as unknown.
Valve settings determined through optimization are
implemented in the wells. We run this true model for a period
of time to generate sensor data for use within the history
matching module.
Optimization Module. This module utilizes a nonlinear
conjugate gradient (CG) algorithm1, 8, 18, 19 to determine the
valve settings that maximize cumulative recovery (or that
minimize the negative of the cumulative recovery) for any
given model(s). This is achieved by maximizing an objective
function defined for the problem as18, 20:
3
N
maximize F (x) = r σ +
xi ∈[0, 1]
∑ λ j f j ( x) (1)
j =1
where, f ∈ [0, 1] is the recovery factor, x is the scaled vector
of valve settings, r is a risk factor and σ is the standard
deviation of the recovery over a given number of models. N is
the total number of realizations considered, and λ ∈ [0, 1] is
N
the weight assigned to each fj such that ∑ λ j = 1 . We note
j =1
that x (the valve settings) is the same for all N realizations.
The case of r = 0 is a risk neutral case and F(x) is then simply
a weighted average of the recovery factors. The case of r = -1
is a risk averse case where F(x) is also a function of σ (in
which case σ will tend to be smaller than in the case of r ≥ 0).
We used a value of r = -1 for the work presented in this paper.
For the initial model specification, λj is set to 1/N (equal
weighting) and subsequently taken as the rank transform of the
inverse objective function value from the history matching
module.
The optimization can be accomplished subject to
production and injection constraints. The gradient of the
objective function, which is computed numerically8, 18, is used
to provide the direction (incrementally opened or closed) of
the valve settings. The procedure is depicted schematically in
Figure 2 below.
Opt im izatio n Ste p
Pas s #
Re s tar t p o in t
Ne w h is to r y-m atch e d m o d e l
Figure 2: Real-time optimization of valve settings with
history matching.
The simulation period is divided into a number of optimization
steps (Figure 2). The valve settings are optimized for the first
step by solving Eq. 1 over that step and the remaining
simulation period (represented by the same color in the
horizontal direction). This is done with the specified initial
model (Figure 1). The optimized valve settings are applied to
wells in the reference model (sub-surface reservoir in real
applications) to generate data for history matching. Once the
model has been updated with the sensor data, it replaces the
existing history matched model (shown as the boxes in the
figure). The simulation is then restarted by running the
updated model with the previously optimized valve settings
from the beginning of the entire simulation to the end of the
last optimized step. The end of the last optimized step marks
the beginning of the next step to optimize with the updated
model. This procedure is repeated for each subsequent
optimization step until the entire simulation period is covered.
4 SPE Student Paper
History Matching Module. The technique used here is the
probability perturbation approach13-15. As stated earlier in the
introduction, this method is based on multiple point (mp)
geostatistics in which the geological model is characterized
via a “training image”. The history matching procedure seeks
to minimize the sum of the squares of the differences between
the observed and predicted production data. This is
accomplished through the gradual modification of the initial
mp permeability field until the production data are honored.
For purposes of the history match, we assume that each of
the downhole sensors provides individual phase flow rates.
For a multilateral well in which each branch is instrumented,
this means that the pressure and phase flow rates in each
lateral are assumed to be known. In current applications,
downhole sensors do not yet provide this information directly
(though flow rates can be estimated from temperature and
pressure measurements), but it is reasonable to assume that
future sensors will provide such data more directly and with
greater degrees of accuracy.
Application of the Overall VOHM Procedure
Variogram Model. Here, we consider an unconditioned two
facies model containing oil with an analytical aquifer acting at
its bottom edge. Independent and unconditional population of
the permeability and porosity within each facies was
performed with SGSIM21. The high and low permeability
sands were of average permeability 2933 md and 5 md
respectively. Porosity was uniform (average 0.19). The
physical size of the system is 3000 ft × 3000 ft × 50 ft. The
simulation models are of dimension 50 × 50 × 5. Mobility
ratio was 6.7.
A single horizontal well located 10 ft below the top of the
reservoir and with a total length of 2400 ft was introduced into
the model. The well and permeability field are presented in
Figure 3. The well has five segments and each is
instrumented with a valve and sensor (shown in Figure 3 as
dots). Constant liquid rate control was specified for the well
with a minimum bottomhole pressure of 1500 psi. The initial /
maximum production was specified at a total liquid rate of
6 MSTB/day. The simulation proceeded for 720 days and the
valve settings were updated every 240 days. This implies that
the entire simulation period was divided into three
optimization and history matching steps.
Figure 3: Permeability field of variogram based model.
We assess the performance of the overall method relative
to two reference cases. These are (1) simulation using the
actual (known) geology but with no valves (uninstrumented
base case) and (2) simulation using the actual geology with
optimized valves. We expect that the results from our method
(optimized valves but unknown geology) will fall between
these two results. The cumulative oil production and water cut
results of this model are shown in Figure 4 and Figure 5,
respectively. These are plotted for the uninstrumented base
case (blue curve), for the case of known geology and
optimized valves (green curve), and for the case using our
overall VOHM procedure (red curve). Considering the known
geology case (blue and green curves), the optimization of the
valves tends to delay water breakthrough, leading to a
substantial increase in oil recovered. Specifically, the
percentage gain in cumulative oil recovered over the
uninstrumented base case was approximately 58%.
Next, we consider the VOHM procedure. It is obvious in
Figure 4 that the procedure (red curve) leads to cumulative oil
results very close to those attained with known geology. The
sharp water cut discontinuity observed in Figure 5 at about
240 days on the HM water cut curve (red curve) occurs
because the initial model used in the first step has little
resemblance to the actual model and was therefore unable to
significantly delay water breakthrough. At the second step,
the history-matched model was used to update the valve
settings causing some of the valves to close thus resulting in
the observed discontinuity. The percentage gain in cumulative
oil recovered over the uninstrumented base case due to
VOHM was approximately 54%. This translates to about 93%
of the gain with known geology, which, for this case, is 58%.
2500
2000
Cum. oil, MSTB
1500
1000
Known geol. w/o valves
500
HM w/valves
Known geol. w/valves
0
0 200 400 600 800
days
Figure 4: Cumulative oil plot from the overall VOHM
procedure and known geology cases of variogram model.
SPE Student Paper 5

0.9
We again assess the performance of the overall method
0.8 relative to two reference cases of the actual geology -
0.7 uninstrumented base case and optimized valves. We expect
that the results from our VOHM method (optimized valves but
0.6
unknown geology) will fall between these two results.
Water Cut

0.5
Simulation results for the fluva model are shown in
0.4 Figure 7. Here, we plot (a) the cumulative oil production and
0.3
(b) water cut for the uninstrumented base case, the case of
Known geol. w/o valves known geology and optimized valves, and the case using our
0.2
HM w/valves VOHM procedure. We first consider the improvement
0.1 Known geol. w/valves attained using optimized valves for the case of known
0 geology. It is apparent that the valves lead to significantly
0 200 400 600 800 improved performance – specifically, an increase in
days cumulative oil of about 40% over the uninstrumented base
Figure 5: Water cut plot from the overall VOHM case. This improvement is directly related to a reduction in
procedure and known geology cases of variogram model. water cut, as the model with no valves reaches the water cut
limit at about 420 days, while the optimized valve case does
Channelized Models. We now present results for two not reach this limit over the time period considered. This is
different channelized systems, referred to as fluva and evident from the water cut behavior (Figure 7b).
We next consider the results using the valve optimization
fluvb. The two systems represent unconditional realizations
plus history matching (VOHM) procedure. As is evident in
generated from different training images using the SNESIM
Figure 7a, the procedure leads to results very close to those
software22. Independent and unconditional population of the
obtained when the geology is known. Specifically, the
permeability and porosity within each facies was performed
technique provides an increase in cumulative oil of about 38%
with SGSIM21. For the fluva case, the channel sand was of over the uninstrumented base case, which is nearly as much as
average permeability 436 md (average porosity 0.24) while the the 40% improvement achieved for the case of known
mudstone was of average permeability 10 md (average geology. This demonstrates the potential gains that can be
porosity 0.07). The physical size of the system is 4000 ft × achieved using this methodology. It is interesting to note,
4000 ft × 100 ft (the upper 50 ft represent a gas cap; an however, that the water cut from the optimized case with
analytical aquifer acts at the bottom edge of the model) and unknown geology is quite different than that for the case of
the simulation models are of dimension 20 × 20 × 6. The known geology, indicating that the optimizations were
mobility ratio was slightly less than one. achieved via different paths.
A quad-lateral well located 20 ft above the oil-water
contact was introduced into the model. The mainbore was of a
total length of 3200 ft and each of the laterals was
approximately 2200 ft. Only the laterals (not the mainbore)
were opened to production. The well and permeability field
are shown in Figure 6. Constant total fluid rate control was
specified for the well, subject to a minimum bottomhole
pressure of 1500 psi. Initial production was specified at a total
liquid rate of 10 MSTB/day. The simulation proceeded for
800 days and the valve settings were updated every 200 days
(i.e., the simulation period was divided into four optimization
and history matching steps). The entire well was shut-in if the
producing water cut exceeded 80% or if the oil rate fell below
the economic limit of 200 STB/day.

Figure 6: Permeability field of fluva channel reservoir.

6 SPE Student Paper

3000 average sense, over sets of three or five models. In order to

(a) present the results using this procedure more concisely, we
2500
introduce a parameter (∆N) that quantifies the improvement in
cumulative oil recovery (Np) relative to that attained for the
2000
Cum. oil, MSTB

case of known geology:

1500
N ptarget model, w/valve − N pknown geology, no valve
1000 ∆N = (2)
Known geol. w/o valves
N pknown geology, w/valve − N pknown geology, no valve
HM w/valves
500
Known geol. w/valves
If ∆N = 0, the cumulative oil recovered from the target model
0
0 200 400 600 800 1000
is the same as that of the uninstrumented base case with
days known geology. On the other hand, if ∆N = 1 (presumably the
maximum attainable), the cumulative oil recovered from the
0.9
target model is the same as that of the optimized case with
0.8 (b) known geology. Higher values of ∆N are of course desirable;
0.7 for reservoir fluva considered above, we obtained ∆N ≈
0.6 0.94.
The results achieved by optimizing over multiple
Water cut

0.5
realizations are displayed in Table 1. From the table, we see
0.4
that ∆N ≈ 0.4 if we optimize over one history-matched (HM)
0.3 model or if we optimize over multiple models that have not
0.2 Known geol. w/o valves been history-matched. Optimizing over three or five history-
0.1
HM w/valves
matched models, however, leads to significant improvement in
Known geol. w/valves
recovery; i.e., ∆N ≈ 0.85. This clearly demonstrates the
0
0 200 400 600 800 1000 benefit of minimizing the error inherent in any single history-
days matched model by considering multiple such models.
Figure 7: (a) cumulative oil and (b) water cut for the Table 1: Impact of optimizing over multiple unconditioned
overall VOHM procedure and known geology cases of history-matched models (fluvb).
fluva.

The second channelized case simulated is referred to as Number of ∆N ∆N

fluvb. For this case, the channel sand was of higher average
permeability than for fluva (1660 md), though the model
and well were otherwise the same as in the previous example.
Following the same procedure as above, we observed an
increase of 28% in cumulative oil production (relative to the
uninstrumented base case) when optimized valves were used
with known geology. However, when we applied our VOHM
procedure to this case, the results were disappointing; i.e., we
did not achieve results very close to the optimized case with
known geology, as we did previously. We then tried several
different initial geological models and applied the procedure
using these models one at a time. We observed on average
about a 12% improvement in cumulative oil recovery relative
to the uninstrumented base case, much less than the 28%
improvement observed when the geology was known. This
12% improvement is about the same as was achieved using
optimization without history matching; i.e., by simply
optimizing on the initial realizations.
The problem here appears to derive from the inherent
nonuniqueness (and thus uncertainty) in the history-matched
model. The inaccuracy in any of the individual history-
matched models renders them incapable of providing
optimization results comparable to those achieved for known
geology. We address this issue by optimizing the valve
settings over multiple history-matched models. This is
accomplished by determining settings that are optimal, in an
HM models without HM with HM
1 0.393 ± 0.508 0.438 ± 0.273
3 0.417 ± 0.372 0.852 ± 0.165
5 0.358 ± 0.410 0.844
Also reported in the table are the standard deviations in
∆N. These values were determined by simulating multiple
(five or more) groups of one, three, and five models (both with
and without history matching). Note that we do not report
standard deviations for the case of five history-matched
models because only two such groups were considered, due to
the computational costs of these runs. The standard deviations
in ∆N with history-matching demonstrate that the use of
multiple history-matched models also leads to lower
variations; i.e., less sensitivity to the particular set of initial
realizations used for the history matching.
We also considered optimizing over multiple history-
matched models for the fluva case considered above. The
results obtained are shown in Table 2. Again, we did not
perform enough runs for the five history-matched case to
report the standard deviation. These results indicate that the
use of more than one model act to improve ∆N and decrease
the standard deviation. However, in some cases, such as this
SPE Student Paper 7

one, a single history-matched model may suffice for purposes history-matched models. Nonetheless, it would be desirable to
of the optimization. accelerate the overall procedure to enable frequent model
updates and valve optimizations.
Table 2: Impact of optimizing over multiple unconditioned
history-matched models (fluva). Alternative optimization procedures
The current procedure, which we will call the global
Number of ∆N ∆N procedure, optimizes the valve settings for each optimization
HM models without HM with HM step such that the objective function is maximized over the
entire remaining simulation period. The disadvantage of this
1 0.519 ± 0.255 0.903 ± 0.179 is that the optimization module becomes very time consuming,
3 0.552 ± 0.252 0.939 ± 0.043 particularly for multiple history matched models. To alleviate
this problem and accelerate the VOHM procedure, we propose
5 0.587 ± 0.234 0.928 the local r0 procedure and the r1 procedure. Rather than
maximize the objective function at each optimization step over
the entire remaining simulation period, the local r0 procedure
Conditioned Models. We next apply the overall valve maximizes the objective function at each step by considering
optimization – history matching procedure to channelized that step only. It disregards what may happen at later times
reservoir cases in which the permeability field was and focuses only on what occurs at the current step. The r1
conditioned to well data. We assumed that the facies type was procedure is an extension of the local r0 procedure. It
known along the mainbore and along all of the laterals (as maximizes the objective function by considering the current
could be achieved using LWD). Results for ∆N (using a optimization step plus the next optimization step when
single model) are shown in Table 3. These results represent determining the valve settings for the current step. Figure 8
averages over many (eight or more) models used one at a time. shows the schematic representation of each of the procedures,
With conditioning and history matching, the use of a single similar to that presented in Figure 2 above.
realization provides improved results relative to the case with
Local r0 Procedure
no history matching or conditioning for the fluvb model (∆N Pass #

= 0.645 compared to 0.393). Multiple history-matched models Global Procedure

were also considered for fluvb. In this case, using three Optimization Step

history-matched models (six such sets were considered), ∆N = Pass #

0.83±0.10, which is quite similar to what we obtained using

history matching without conditioning (Table 1). It is Local r1 Procedure
possible that conditioning does not have more of an impact on Pass #

∆N in some cases because there is a degree of redundancy in

the facies and production data (so the benefit of conditioning
is not that great). This issue requires further investigation.

Table 3: Impact of conditioning on optimization results.

Figure 8: Different valve optimization procedures.
∆N ∆N ∆N
Model w/o HM, w/o HM, w/HM, In order to more comprehensively compare the procedures,
w/o condition w/condition w/condition we introduce two new models: a layered system presented
fluva elsewhere8 and a third channel reservoir that we refer to as
0.519 ± 0.255 0.581 ± 0.174 0.881 ± 0.062
fluvc (similar to fluva and fluvb). A water cut
fluvb 0.393 ± 0.508 0.543 ± 0.270 0.645 ± 0.173 constraint of 90% was defined for fluvc and the simulation
was run for 1200 days with an optimization step of 240 days.
Computational Cost. The method presented in this paper is The actual models were optimized directly without HM and
very computationally intensive. For the cases considered here the result is shown in Table 4.
(four valve updates and four history matches), the valve
optimizations required a total of about 100 simulations for the
gradient calculations. The history matching algorithm
required many more simulations, as many as 200 runs each
time the model was updated. These computational
requirements are even greater when multiple realizations are
considered. It should be noted, however, that we could have
used many fewer history-matching runs (40-50) if we used a
less stringent convergence tolerance. It is likely that this
would have a relatively minor impact on the results,
particularly in the case of optimizing over multiple
8 SPE Student Paper

Table 4: Comparisons of the various procedures 3000

(a)
Cumulative oil, MSTB 2500
Model
Global r0 r1 No-valve
2000

Cum. oil, MSTB

Layered 2365 2309 2395 1733
1500
Variogram 2252 2254 2243 1422
Fluva 2596 2659 2462 1870 1000

Fluvb 2751 2775 2732 2159 CG w/valves & HM

500 LM w/valves & HM
Fluvc 3066 3273 3010 2344 Initial model w/valves no HM

0
0 200 400 600 800 1000
days
The best result for each case is shown in bold. The
extended r1 procedure appears to be the best for the layered
1
system, while any of the procedures may be used for the CG w/valves & HM
0.9 LM w/valves & HM
variogram based model. For the channel systems, the local r0 (b) Initial model w/valves no HM

procedure appears to be the preferred option. Therefore, 0.8

rather than optimize the valve settings over the entire 0.7

simulation period at each optimization step, we can optimize 0.6

Water cut
only over the current optimization step, leading to 0.5
computational savings. 0.4
The optimization engine used for the VOHM results was a
0.3
conjugate gradient (CG) procedure. We now consider the use
0.2
of Levenberg-Marquardt (LM) instead of CG. The advantage
0.1
of LM is that it is more computationally efficient than CG.
For the LM formulation we modified the objective function as 0
0 200 400 600 800 1000
the square of F(x):
days

maximize V ( x ) = F 2 ( x ) (3) Figure 9: (a) cumulative oil and (b) water cut comparison
xi ∈[ 0 , 1] of LM and CG for fluva model with VOHM.

The gradient, G, and Hessian, H, of V(x) are given by:

1
G ( x) = V ′(x) = F (x) F ′(x)
2 (4)
1 (CG is applied for the optimization of this model). As
H (x) = V ′′(x) = F F ′′ + F ′ F ′T ≈ F ′ F ′T
2 observed in Figure 9a, it is apparent that both the CG and LM
In the LM procedure, we neglect F F ′′ relative to F ′ F ′T . The
LM transformation also introduces a stronger diagonal:
H LM = µ I + H (5)
where I is the identity matrix. The purpose of the µ in Eq. 5 is
to combine both steepest descent and Gauss-Newton methods.
If µ is large, the matrix becomes diagonally dominant and the
LM method is similar to a steepest descent type algorithm.
The LM algorithm used here follows that of Press et al.19.
We demonstrate the VOHM procedure with the LM
algorithm by applying it to the fluva channel reservoir. We
use the global procedure and consider one conditioned history
matched model to determine the valve settings. For
comparison purposes, we also show results with the CG
algorithm.
Figure 9 shows (a) the cumulative oil recovery and (b) water
cuts for the LM and CG algorithms. Also included in the
figure is the cumulative recovery and water cut from the
initially specified geologic model without history matching
curves produced approximately the same final recovery. The
water cut results of the two algorithms, shown in Figure 9b,
differ considerably. This indicates that the two algorithms can
produce valve settings that are significantly different. Similar
results were obtained when we considered the fluvb model.
The LM method was three to four times faster than the CG
method. This becomes important when multiple models are
considered. The VOHM procedure will run even faster when
the local r0 procedure is used in conjunction with LM.
Summary
In this paper, we used a combined history matching – valve
optimization (VOHM) procedure for maximizing production
from smart wells. This technique is capable of accounting for
geological uncertainty by optimizing over multiple geological
models. Results demonstrate the clear improvement in
cumulative oil recovery attainable using this method (up to
90% of gain obtained with actual geology). We also
developed and investigated efficient alternative procedures to
SPE Student Paper
reduce the substantial computational requirements of this
technique. By using the local ro procedure in conjuction with
the Levenberg-Marquardt algorithm, substantial computational
savings can be achieved with the VOHM method.
Nomenclature
F objective function
f recovery factor
G gradient vector
H Hessian matrix
N total number of realizations (or models)
Np cumulative oil production, MSTB
∆N dimensionless cumulative oil difference
r risk factor
V modified objective function
x scaled vector of valve settings
λ weights
µ Levenberg-Marquardt transformation factor
σ standard deviation, data units
Acknowledgments
I am grateful to my advisor, Prof. Lou Durlofsky, for directing
this work. I appreciate Dr. Burak Yeten for providing me with
his valve optimization algorithm. I also thank Prof. J. Caers
and S. Suzuki for providing me with history matching
algorithms and for assisting me in their use. Financial support
from Chevron Nigeria Limited and from the industrial
affiliates of the Stanford Project on the Productivity and
Injectivity of Advanced Wells (SUPRI-HW) and the U.S.
Department of Energy (contract number DE-AC26-
99BC15213) is deeply appreciated
References
1. Yeten, B.: Optimum Deployment of Nonconventional
Wells, Ph.D dissertation, Department of Petroleum
Engineering, Stanford University, California (2003).
2. Khargoria, A., Zhang, F., Li, R., and Jalali, Y.:
“Application of Distributed Electrical Measurements and
Inflow Control in Horizontal Wells under Bottom-Water
Drive”, paper SPE 78275 presented at the 2002 SPE
European Petroleum Conference, Aberdeen, UK, 29-31
October.
3. Brouwer, D. R., Jansen, J. D., Van Der Starre, S., Van
Kruijsdijk, C. P. J. W., and Berentsen, C. W. J.:
“Recovery Increase through Water Flooding with Smart
Well Technology”, paper SPE 68979 presented at the
2001 SPE European Formation Damage Conference, The
Hague, Netherlands, 21-22 May.
4. Gai, H.: “Downhole Flow Control Optimization in the
World’s 1st Extended Reach Multilateral Well at Wytch
Farm”, paper SPE 67728 presented at the 2001
SPE/IADC Drilling Conference, Amsterdam,
Netherlands, 27 February-1 March.
5. Arenas, E. and Dolle, N.: “Smart Waterflooding Tight
Fractured Reservoirs Using Inflow Control Valves”,
paper SPE 84193 presented at the 2003 SPE Annual
9
Technical Conference and Exhibition, Denver, Colorado,
5-8 October.
6. Snaith, N., Chia, R., Narayasamy, D., and Schrader, K.:
“Experience with Operation of Smart Wells to Maximize
Oil Recovery from Complex Reservoirs”, paper SPE
84855 presented at the 2003 SPE Asia Pacific
International Improved Oil Recovery Conference, Kuala
Lumpur, Malaysia, 20-21 October.
7. Al-Khodhori, S. M.: “Smart Well Technologies
Implementation in PDO for Production & Reservoir
Management & Control”, paper SPE 81486 presented at
the 2003 SPE Middle East Oil Show and Conference,
Bahrain, 5-8 April.
8. Yeten, B., Durlofsky, L. J., and Aziz, K.: “Optimization
of Smart Well Control”, paper SPE 79031 presented at the
2002 SPE International Thermal Operations and Heavy
Oil Symposium and International Horizontal Well
Technology Conference, Calgary, Canada, 4-7 November.
9. Dolle, N., Brouwer, D. R., and Jansen, J. D.: “Dynamic
Optimization of Water Flooding with Multiple Injectors
and Producers Using Optimal Control Theory”, paper
presented at the XV International Conference on
Computational Methods in Water Resources, Delft,
Netherlands, 23-28 June, 2002.
10. Holmes, J. A., Barkve, T., and Lund, Ø.: “Application of
a Multi-segment Well Model to Simulate Flow in
Advanced Wells”, paper SPE 50646 presented at the 1998
SPE European Petroleum Conference, The Hague,
Netherlands, 20-22 October.
11. Yeten, B., Brouwer, D. R., Durlofsky, L. J., and Aziz, K.:
“Decision Analysis Under Uncertainty for Smart Well
Deployment,” Journal of Petroleum Science and
Engineering (2004) (to appear).
12. Kabir, C. S., Chien, M. C. H., and Landa, J. L.:
“Experiences With Automated History Matching”, paper
SPE 79670 presented at the 2003 SPE Reservoir
Simulation Symposium, Houston, Texas, 3-5 February.
13. Caers, J., 2003, “History Matching Under
Training-Image-Based Geological Model Constraints”,
SPE Journal (Sept. 2003) 218-226.
14. Hoffman, T. B. and Caers, J.: “History Matching Using
the Regional Probability Perturbation Method”, Stanford
SCRF Group (2003) Report 16.
15. Suzuki, S.: Determining Petrophysical Properties of
Facies Using a Hierachical History Matching Method,
Master’s Report, Department of Petroleum Engineering,
Stanford University, California (2003).
16. GeoQuest Schlumberger, ECLIPSE Reference Manual
2001A (2001).
17. GeoQuest Schlumberger, ECLIPSE Technical
Description 2001A (2001)
18. Aitokhuehi, I.: Real-time Optimization of Smart Wells,
MS thesis, Stanford University, CA (2004).
19. Press, W. H., Teukolsky, S. A., Vetterling, W. T., and
Flannery, B. P.: Numerical Recipes in C++: The Art of
10 SPE Student Paper

Scientific Computing. Cambridge University Press,

Cambridge, UK, second edition (2002).
20. Yeten, B., Durlofsky, L. J., and Aziz, K., 2003,
“Optimization of Nonconventional Well Type, Location
and Trajectory”, SPE Journal (Sept. 2003) 200-210.
21. Deutsch, C. V. and Journel, A. G.: GSLIB User’s Manual.
Applied Geostatistics Series, Oxford University Press,
New York, second edition (1998).
22. Strebelle, S.: Sequential Simulation: Drawing Structures
from Training Images, Ph.D dissertation, Department of
Petroleum Engineering, Stanford University, California
(2000).