Professional Documents
Culture Documents
Herring Bone Decision
Herring Bone Decision
Abstract
Downhole inflow control devices allow for the flexible operation of nonconven-
tional wells. In this paper, a method for determining the optimal performance of
smart wells; i.e., wells containing downhole inflow control devices, is presented. The
method entails the use of a gradient based optimization technique in conjunction
with a reservoir simulator. The optimization accounts for uncertain geology as well
as the risk of failure of the control devices. A decision analysis approach is presented
to determine whether or not to deploy smart completions in light of these uncer-
tainties. The overall optimization technique is applied to several example problems
involving geological models with multiple geostatistical realizations and with differ-
ent reliability scenarios for the control devices. It is shown that there is significant
variation in the level of improvement attainable using inflow control devices and
that this improvement varies with both the valve reliability and the particular ge-
ological realization. Decision making techniques are used to quantify the benefits
obtained using optimized smart wells under different risk attitudes.
1 Introduction
One approach for using smart well technology is to react to production prob-
lems (e.g., water coning) and then reset the instrumentation to mitigate them.
This is often referred to as a “reactive control strategy”. A better approach is
to use downhole inflow control devices in conjunction with a predictive reser-
voir model. This allows for the optimization of reservoir performance rather
than just the correction of problems that have already occurred. This “defen-
sive” (rather than reactive) strategy is the method considered here.
2
optimization algorithm that applied optimal control theory. Using this ap-
proach they demonstrated improved sweep and recovery over their previous
method (Brouwer et al., 2001). Brouwer and Jansen (2002) investigated the
effects of smart completions with different well targets and constraints in a
water flooding project using optimal control theory. They found considerable
scope for accelerating production and increasing recovery for wells operating
with rate targets. Sudaryanto and Yortsos (2000) also applied optimal control
theory for the optimization of sweep efficiency. They optimized fluid injection
rates for two-dimensional problems with vertical wells.
Khargoria et al. (2002) studied the impact of valve placement, inflow config-
uration and mode of operation on production performance using a horizontal
well on a synthetic bottom water drive model. They considered both reactive
and defensive modes. They used simulated annealing and conjugate gradient
optimization algorithms to determine the optimum location and control set-
tings of the valves to maximize the cumulative oil production. In their work
the valve settings were not updated in time. The same optimum was achieved
with both of the methods.
3
phase.
2 Methodology
4
by summing the frictional pressure losses and the pressure losses due to the
control device (GeoQuest, 2001):
ρvc2
∆pc = Cu , (2)
2Cv2
and the pressure loss due to friction, ∆pf , is calculated by the standard ex-
pression for homogeneous flow through a pipe (GeoQuest, 2001):
L
∆pf = 2Cu f ρm vp2 . (3)
D
In the above equations ρm represents the density of the fluid mixture in the
segment, Cu = 2.159×10−4 is a unit conversion factor (all units are given in the
Nomenclature), Cv is a dimensionless valve coefficient, vc and vp are mixture
velocities (flow rate divided by area) through the choke and pipe, respectively,
f is the Fanning friction factor, and L and D represent the length and diameter
of the pipe segment. The valve setting (i.e., degree of opening) of the inflow
control device is specified in terms of the area of the constriction Ac . This
area enters Eq. 2 via its effect on vc ; i.e., vc = qc /Ac , where qc is the flow rate
through the constriction.
5
Conjugate gradient methods can be used to find the minimum point of a
quadratic function. Nonlinear CG methods can be applied to minimize any
continuous function f (x) for which the gradient, f 0 , can be calculated. In a
nonlinear CG method, the residual, r, or the direction of steepest descent, is
set to the negative of the gradient (Shewuck, 1994):
³ ´
rk = −f 0 xk . (4)
In this and the following equations the subscript k indicates the iteration level;
k = 0 refers to the initial guess. The search directions, dk , are computed by
k
Gram-Schmidt ³ conjugation´ of the residuals. The value of the step size α that
k k k
minimizes f x + α d is found by ensuring that the gradient is orthogonal
to the search direction.
d0 = r0 = −f 0 (x0 )
³ ´
Find αk that minimizes f xk + αk dk
xk+1 = xk + αk dk
³ ´
rk+1 = −f 0 xk+1 (5)
k+1 k+1 k+1 k
d =r +β d
where
( T
)
k+1 (rk+1 ) (rk+1 −rk )
β = max T ,0 .
(rk ) rk
The algorithm is deemed to be converged when the norm of the residual falls
below a specified value, which is typically taken to be a small fraction of the
initial residual. The convergence criteria can thus be specified as:
° ° ° °
° k° ° °
°r ° < ε °r0 ° . (6)
6
where ε is the fractional convergence tolerance, here taken to be 0.01.
where f is the objective function (e.g., the recovery factor or net present value),
x is the vector that holds the valve settings and xi is a component of x. These
valve settings vary continuously between 0 (valve fully closed) and 1 (valve
fully open) and are given by Ac /Amax , where Amax is the maximum constric-
tion area (valve fully open). The gradient vector, f 0 , is calculated numerically
using a forward finite difference approximation:
∂f f (x + h) − f (x)
f0 = = . (8)
∂x h
A step size, h, of 0.05 is used in this study. The f 0 computed by Eq. 8 can
sometimes be a very large positive or negative number. This might cause the
next proposed setting to be too large or too small or even negative, which
is of course unphysical. To avoid this situation the objective function is also
scaled between 0 and 1. If we specify the recovery factor as our objective
function there is no need for this scaling, as the recovery factor is already in
the range 0 ≤ f ≤ 1. When the valve settings approach the upper limit (i.e.,
xi → 1), we replace Eq. 8 with a backward difference approximation to avoid
the unphysical situation of xi + h > 1.
2.2 Implementation
7
but severe water coning appears for t = t1 to t = t2 . The approach described
here avoids this limitation. Note that the control devices are assumed to have
infinite resolution in their settings.
(1) Divide the simulation period into n time periods at which the settings
will be updated to optimize the objective function.
(2) For each period i, optimize the device settings such that they maximize
or minimize the objective function for the remaining simulation period
(i.e., solve Eq. 7).
(3) Restart the simulation from the end of the previous period.
(4) Repeat steps 2 and 3 until the entire simulation period is covered.
The overall process is depicted in Fig. 1. In this figure each color represents an
optimized valve setting for the specific time period. The circles on this figure
represent the restart points which coincide with the end of an optimization
period. The optimization time periods can be determined based on the ac-
tual downhole hardware (i.e., it may not be possible to update valve settings
very frequently). The optimization process clearly becomes more computation-
ally expensive as more periods are considered. For the cases presented below,
O(100) simulations were required for the optimizations. The exact number of
simulations is quite case specific, and could be reduced through the use of
proxy functions such as artificial neural networks and response surfaces.
8
3 Probabilistic Model for Valve Reliability
Downhole valves are relatively new in the industry, and only limited informa-
tion is available on their reliability. Valve failure may, however, greatly affect
the performance of the smart completions. The severity of this effect will de-
pend on the time and the type of valve failure. The particular failure modes
are expected to be related to how the hardware was manufactured.
One failure mode is that the valve sticks in its current setting, which may be
caused by failure in the valve itself, or by the loss of control from the surface.
This type of valve failure will be referred to as failure type 1. Some valves are
manufactured according to the failure-open type, meaning that upon failure,
the valve fully opens automatically. In the remainder of this paper this will be
referred to as failure type 2. Other valves are configured to fully close upon
failure, referred to as failure type 3. In the coming sections the cumulative
distribution function (cdf) for assessing device reliability will be presented.
The probability for a valve to fail is related to its reliability, which can be
expected to change over time. In addition, over the lifetime of a component,
the most likely failure mechanism can also be expected to change. Initially the
highest likelihood of failure for a component may be of the infant mortality
type. In this stage failure is most likely caused by inadequate testing of the
equipment, or due to damage caused during installation. As time progresses
this failure mechanism becomes less likely to occur. Hence the failure likelihood
decreases with time after the infant mortality period. The failure likelihood in
this stage is rather low and approximately constant and is deemed to be due
to random errors. After this period of relatively high reliability the component
begins to age and may fail as a result. The expected trend at this stage is that
the failure likelihood increases with time.
9
3.2 The Weibull Distribution in Life Testing
A distribution function that is very convenient for representing the life distri-
bution of components where the failure rate varies over time is the Weibull
distribution. The Weibull failure cumulative distribution function is given by:
( µ ¶β )
t
F (t) = 1 − exp − , (9)
α
where β is the shape parameter and α is the scale parameter. Assuming the
valve is operating at a time t1 , the probability that it is still operating at a
later time t2 is given by:
1 − F (t2 )
p12 = . (10)
1 − F (t1 )
The scale parameter α (α > 0) stretches or compresses the pdf. Values of α > 1
act to stretch the pdf. Higher values of α give broader distributions, with the
peak occurring at later times. The peak time is the most likely time of failure.
Values of α < 1 compress the pdf, with the distribution approaching a spike
as α → 0. More detailed information about the Weibull distribution can be
found in Abernathy (2000).
10
β (t) = c + zt, (11)
Inserting β = c + zt into Eq. 9 provides the cdf defining the valve failure prob-
abilities. The pdf can then be obtained by differentiating Eq. 9 with respect
to t, with β now a function of time as given by Eq. 11.
At the start of each optimization step, a random number R is drawn for each
of the valves. This number is compared to the failure probability pf of the
valve. All failure states (types 1, 2 and 3) are considered equally likely to
occur. Therefore the probability of failure at a certain time is divided into
three intervals that represent the failure mode (i.e., if the random number is
0 ≤ R < 31 pf , the valve fails according to type 1, if 31 pf ≤ R < 23 pf , the failure
is of type 2 and if 23 pf ≤ R ≤ pf , the failure is of type 3). If R is larger than
pf , the valve remains intact.
The gradients are then calculated for all valves that are still intact, and the re-
maining production interval is optimized by determining the optimum settings
for these valves.
In this section, the methodology that will be used to assess the effect of un-
certainties on decisions will be discussed.
11
N
X
hu (x)i = u (xi ) · pi . (12)
i=1
where hu (x)i is the expected utility of all N outcomes of xi that occur with
a probability of pi .
• If alternative L1 is preferred to alternative L2 then u (L1 ) > u (L2 ).
where r is called the risk attitude. For r > 0 the decision maker is risk prone,
for r < 0 the decision maker is risk averse, and for r = 0 the decision maker
is risk neutral.
1 − erx
u (x) = . (14)
1 − er
From Eq. 14, it can be shown that u (x) → x as r → 0 (i.e., u (x) = x for a
risk neutral decision maker).
Typical utility curves for different risk attitudes are shown in Fig. 3. The
shape of the curve determines whether the decision maker is risk neutral, risk
averse or risk prone. A risk neutral decision maker has a linear utility function
which is equivalent to basing decisions purely on expected value. A risk averse
decision maker has utility function with u00 (x) < 0, which shows the avoidance
of uncertainty even if it might have the possibility of greater gain. The risk
prone decision maker has u00 (x) > 0 which represents a decision maker who
is willing to take some risk for the chance of greater gain. Once the decision
maker provides the risk attitude coefficient and x is defined, u (x) can be
used to determine the decision maker’s preferences; i.e., the alternative that
maximizes the expected value of the utility function.
The preferences represented by the utility function are unchanged if the func-
tion is subjected to a linear transformation of the form:
12
u∗ (x) = a + bu (x) , b > 0. (15)
In addition the utility values are invertible (see Eq. 14), so preferences can
be transformed back to the original data units. The inverse of the expected
utility is called certain equivalent, CE. “Inverse” in this context, represented
as u−1 , means the value of x that provides a particular value of u. In this
study the decisions will be based on the CE values of the alternatives rather
than on the utilities, though this does not change the decisions themselves.
5 Applications
13
the water-oil contact. The main bore was not perforated. Each lateral was
approximately 2150 ft long. The well location, completion architecture and
instrumentation were the same for each of the realizations.
Initial production was specified at a total liquid rate of 5.0 MSTB/d. There
was a constraint on water production; specifically, the well was shut in if the
water cut exceeded 85%. A minimum bottomhole pressure of 1500 psi was
imposed for lift considerations, although this BHP was not reached during
any of the simulations due to the presence of the aquifer. The simulations
proceeded for 2960 days (∼ 8 years), with valve settings updated at every 185
days (16 optimization steps). The optimizations were based on maximization
of oil recovery.
In the first case, the five simulation models, each having a different realization
of the geological description, were run with no instrumentation on the well
(base case). Then four control devices on the tubing of the main bore, close
to each of the junctions, were introduced, so that each of the laterals could
be controlled independently. Following this, the optimization algorithm was
applied to each geological realization to maximize the oil recovery for that
particular realization.
Table 2 displays the cumulative oil production attained at the end of the sim-
ulations for the base and the optimized cases. The variation in cumulative
production between realizations, for both the base and controlled cases, is due
to the high degree of geological variation between the realizations. These vari-
ations strongly impact the degree of connectivity of the well to the channels
and to the aquifer. The percentage increase in cumulative oil production due
to optimized downhole inflow control (last column) varies from 4.7% to 51.0%.
The average increase over the five realizations is 24.7% with a standard de-
viation of 17%. This demonstrates the considerable level of improvement and
the variation in oil production that can be achieved through the use of smart
wells.
The higher average in cumulative oil production and the lower standard de-
viation for the optimized instrumented wells suggest that they can partly
compensate for the adverse effects of the geological features, i.e., the well per-
formance becomes somewhat less dependent on geology. In this respect smart
completions can be considered to be a type of “insurance policy” against ge-
ological uncertainty.
Table 2 clearly shows how the uncertainty around the reservoir description
could affect the decision on whether or not to instrument a well (it should be
14
noted that in the case considered here, however, the geostatistical realizations
were not conditioned to any reservoir data and no attempt was made to assess
the number of realizations required for reliable statistics). For the realizations
with low additional recovery the instrumentation might not be economical.
For other realizations, however, significant resources might be lost by not
deploying the control devices.
In a practical setting, results such as those shown in Table 2 might lead one
to proceed in several ways. One approach is to consider a number of relevant
solution variables and their probability distributions and to introduce a utility
function along with a risk attitude coefficient. This would then allow for the
application of a decision analysis in the presence of uncertainty. Alternatively,
one might attempt to better characterize the reservoir in order to narrow the
geological uncertainty. This would require an assessment of the value of the
additional geological information.
Since the optimizations were based on oil recovery, an NPV analysis with zero
discount rate and zero water handling cost was used. The NPV calculated
with this approach, which also accounts for the cost of the smart completions,
is consistent with the objective function of the optimization problem (i.e.,
maximization of oil recovery).
The NPV values calculated for each realization for the base and instrumented
cases and the percentage increase attained by instrumentation over the base
cases are given in Table 3. An oil price of $18 and a valve cost of $500,000
(per valve) were used. The same type of NPV analysis will be used for the
cases discussed in the next sections.
In order to proceed with the decision analysis, the exponential utility function
given in Eq. 14 was used. The decision was based on choosing the alternative
with the larger CE, where CE = hu (x)i−1 . Applying Eq. 12, x represents
scaled NPV (i.e., between 0 and 1), N = 5 is the number of geological realiza-
tions, which are deemed to be equiprobable, giving pi = 0.2. A sketch of the
overall analysis for a risk neutral decision maker (r = 0) is shown in Fig. 5,
using the data from Table 3. In this figure the circles represent the uncer-
tainty nodes and the rectangles show the decision nodes. The branches after
the uncertainty nodes correspond to the geological realizations. The decision
node represents the two alternatives, L1 , not deploying and L2 , deploying the
devices. After calculating the utility of the scaled NPV values for each geo-
logical realization (using Eq. 14 with r = 0), Eq. 12 is used to calculate the
expected utility of each alternative, L1 and L2 . The calculations show that
hu (L1 )i = 0.453 and hu (L2 )i = 0.796. Therefore, alternative L2 is preferred
15
over L1 , indicating that the devices should be deployed.
From this table there are some important conclusions that can be drawn.
The highly risk averse decision maker (r = −10) has a lower CE and the
highly risk prone decision maker (r = 10) has a higher CE than the others
in both the base and instrumented cases. In terms of increase in CE (fourth
column of Table 4), the highly risk averse decision maker expects to have more
benefit than the others. This is because, with optimized smart completions,
the standard deviation is smaller and the mean is larger relative to the base
case. The highly risk averse decision maker therefore expects to gain more,
since less was expected in the base case.
Since detailed reliability data on inflow control devices are not readily avail-
able, three reliability scenarios were considered. In scenario I the most likely
failure time of the valves was considered to be low with the valves expected to
fail after only a few years. In scenario II the most likely failure time was higher
but the valves were still expected to fail before the end of the production time
interval. In scenario III the most likely failure time was close to the total pro-
duction time, and there existed a possibility of valves remaining intact over
the entire production time interval. For each scenario the failure mechanism
as a function of time was expected to be the same, hence the same values of
c and z (which define β) were used (though α was varied).
For parameter c, in Eq. 11, a value of 0.99 was chosen, indicating that the
failure type was initially infant mortality. For parameter z a value of 0.5 was
selected. Consequently β linearly increased from 0.99 (at t = 0) to 4.99 (at
t = 8 years), thereby including the infant mortality, the random error, and
16
the early and late wear stages. The total production period was divided into
16 intervals of 6 months each, which corresponds to the optimization steps
(i.e., the time at which valve settings were computed).
A different expected life time for each scenario was obtained by selecting a
different value for the scale parameter α. For scenarios I, II, and III, α values
of 3, 5, and 7 were chosen respectively. Fig. 6 shows the failure probability
density functions (pdfs) for these scenarios. In Fig. 6 the probability of failure
is zero beyond 5.5 years in scenario I (α = 3), indicating that all valves will fail
before this time. This indicates a maximum lifetime of 5.5 years for the valves
in scenario I. Similarly, a maximum lifetime of about 7.5 years is observed for
the valves in scenario II (α = 5). In scenario III (α = 7), however, the pdf is
still larger than zero at 8 years, indicating that valves can remain intact over
the entire simulation period.
For each scenario, 40 separate simulation models were run (and optimized).
Valve failure time and type, determined probabilistically as described above,
was different for each of these models. We refer to these runs as “failure re-
alizations”. At each optimization step, the settings of the intact valves were
optimized to maximize oil recovery. The distribution of the valve failure times
for scenario II is shown in Fig. 7. The similarity between Fig. 6 (α = 5) and 7
suggests that the number of failure realizations was sufficient to reproduce the
pdf for the valve failure times.
Fig. 8 shows the valve states as a function of time for 4 of the 40 failure
realizations for scenario II. For run1 all valves remained intact for 2.5 years
after the start of production. Valves 1 and 3 were closed at 4.5 and 2.5 years
respectively due to failures of type 3. Valve 2 became stuck at 4 years due to
a failure of type 1, while valve 4 opened completely at 5 years due to a failure
of type 2.
Fig. 9 shows the differences in cumulative oil production (relative to the unin-
strumented base case) using the smart completions for all three scenarios. For
scenario I the average increase in cumulative production for the 40 runs was
5.9%, for scenario II it was 10.8% and for scenario III, 15.5%. The figure clearly
shows the impact of reliability on the added value of the smart completions.
With increasing reliability (e.g., scenario III), the additional cumulative oil
distribution moves increasingly towards positive values.
The results indicate that apart from the time of failure, the type of failure
is also important. Fig. 10 displays the difference in cumulative oil production
(relative to the base case) versus the time of first valve failure. The figure
shows that failure of type I (valve stuck in existing position) often did not
have a strong negative impact on the cumulative oil production, even when
the valve failed at an early stage. The reason for this is that the valve was likely
17
to have a near-optimal setting before it failed. The strongest negative impact
was found for valves failing according to type III (valve being fully closed).
The earlier this type of failure occurred, the lower the resulting cumulative
production. The impact of failure type II was found to be between the other
two failure modes.
The results are presented in Table 5. It is evident from this table that a highly
risk averse decision maker (r = −10) will not deploy the devices. This is the
case because, in an extreme risk averse situation, the focus is more on avoiding
the potential negative impact (cost and the loss of production due to failure)
of deploying the devices. In less risk averse situations, by contrast, the focus
tends to be more on the benefits that deployment of the devices may bring.
The results are presented in Table 6. In this table, the third column provides
the average cumulative oil production of 40 failure realizations (in the instru-
mented case) for each geological realization. The fourth column shows the
average percentage increase over the base case. Comparing Tables 2 and 6, a
reduction in the optimized cumulative productions is evident in the latter set
of runs. This is due to the adverse effects of device failures.
18
5.3.1 Deployment Decision with Overall Uncertainty
In this case the decision tree is more complex than in the other cases. Again
there are two branches emanating from the decision node: do not deploy and
deploy. The do not deploy has an uncertainty node with 5 branches, each cor-
responding to equiprobable realizations of the geology. The deploy node again
has an uncertainty node with 5 branches corresponding to geological realiza-
tions, but now each branch emanating from this node has another uncertainty
node corresponding to the reliability of the devices. As in Case 2, these nodes
have 40 branches, each corresponding to a particular valve failure realization.
A simplified decision tree for this case is shown in Fig. 11.
Table 7 shows the decisions based on the CE values of the alternatives when
all the uncertainties are considered. From this table it is clear that regardless of
the risk attitude coefficient, the decision should be made in favor of deploying
the devices. When the base case and instrumented case certain equivalents
are compared for each decision maker, the more risk averse decision makers
generally expect to gain more than the others. The reasoning is the same in
this case as was observed in Case 1.
6 Conclusions
A general method for the optimization of smart wells with downhole inflow
control devices was presented and applied to various examples. The impor-
tant effects of uncertainty in reservoir description and equipment reliability
were considered. The benefits of these devices were analyzed within a decision
making framework.
• The higher average in cumulative oil production (or net present value) and
the lower standard deviation for the optimized instrumented wells (Cases
19
1 and 3) suggest that downhole control can compensate to some extent
for geological uncertainty, even when the possibility of equipment failure is
included. This demonstrates the insurance value of smart completions.
• The impact of equipment reliability was related to both the timing and
type of failure. Generally, the earlier the valves failed the larger the negative
impact on the cumulative oil production. The impact of failure type 1 (valve
stuck in current position) was often not that severe since the valve may
already have been in an optimized position. The negative impact was largest
for failure type 3 (valve fully closed upon failure).
• In the case when both the geology and the device reliability were uncertain
(Case 3), it was seen that all of the decision makers considered here would
choose to deploy the devices. Some of these decisions would change, how-
ever, if geological uncertainty was eliminated; i.e., if the geology was deemed
to be deterministic. This suggests that the impact of geological uncertainty
dominates that of engineering uncertainty, at least for the scenarios consid-
ered here.
Acknowledgments
We are grateful for financial support from the Industrial Affiliates of the Stan-
ford University Project on the Productivity and Injectivity of Horizontal Wells
(SUPRI-HW) and from the U.S. Department of Energy (under contract no.
DE-AC26-99BC15213). We thank Prof. J. Caers (Stanford University) for use-
ful discussions regarding the Weibull distribution. Portions of this work ap-
peared earlier in preprint form in SPE paper 79031, presented at the 2002
SPE International Thermal Operations and Heavy Oil Symposium and In-
ternational Horizontal Well Technology Conference held in Calgary, Alberta,
Canada, November 4-7.
Nomenclature
20
k permeability, md
L alternative
n number of optimization steps
p pressure, psi, or probability
q flow rate, ft3 /s
r residual in CG formulation
R random number, 0 ≤ R ≤ 1
t time, days or years
v velocity, ft/s
V volume, STB or MSCF
x scaled valve constriction areas, dimensionless
z slope of change of β, years−1
α step size, scale parameter in Weibull distribution, years
β Gram-Schmidt constant, shape parameter in Weibull distribution
ε fractional convergence tolerance
γ specific gravity, dimensionless
φ porosity, fraction
µ viscosity, cp
ρ density, lbm/ft3
σ standard deviation, data units
Subscripts
c constriction
f friction, or failure
I scenario I
II scenario II
III scenario III
m mixture
or residual oil
r random
ro relative to oil
rw relative to water
t total
wc connate water
v valve
21
References
22
ing to Assist Operation of an Intelligent Multilateral Well in the Sherwood
Formation, paper SPE 71828 presented at the SPE Annual Technical Con-
ference and Exhibition, New Orleans, Louisiana, 30 September-3 October.
McNamee, P. and Celona, J. (1990). Decision Analysis with Supertree. The
Scientific Press, San Francisco, CA, second edition.
Nyhavn, F., Vassenden, F., and Singstad, P. (2000). Reservoir Drainage with
Downhole Permanent Monitoring and Control Systems. Real - Time Inte-
gration of Dynamic Reservoir Performance Data and Static Reservoir Model
Improves Control Decisions, paper SPE 62937 presented at the SPE Annual
Technical Conference and Exhibition, Dallas, Texas, 1-4 October.
Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P. (1999).
Numerical Recipes in C. Cambridge University Press, Cambridge, UK, sec-
ond edition.
Shewuck, J. R. (1994). An Introduction to the Conjugate Gradient Method
Without the Agonizing Pain. Internal Report. School of Computer Sci-
ence, Carnegie Mellon University, Pittsburgh, Pennsylvania (available at
http://www.cs.cmu.edu/∼quake-papers/painless-conjugate-gradient.ps).
Sinha, S., Vega, L., Kumar, R., and Jalali, Y. (2001). Flow Equilibration to-
ward Horizontal Wells Using Downhole Valves, paper SPE 68635 presented
at the SPE Asia Pacific Oil and Gas Conference and Exhibition, Jakarta,
Indonesia, 17-19 April.
Sudaryanto, B. and Yortsos, Y. C. (2000). Optimization of Fluid Front Dy-
namics in Porous Media Using Rate Control. I. Equal Mobility Fluids.
Physics of Fluids, vol. 12, pp 1656–1670.
Valvatne, P. H., Serve, J., Durlofsky, L. J., and Aziz, K. (2003). Efficient
Modeling of Nonconventional Wells with Downhole Inflow Control Devices.
Journal of Petroleum Science and Engineering, vol. 39, pp 99–116.
Veneruso, A. F., Hiron, S., Bhavsar, R., and Bernard, L. (2000). Reliability
Qualification Testing for Permanently Installed Wellbore Equipment, paper
SPE 62955 presented at the SPE Annual Technical Conference and Exhibi-
tion, Dallas, Texas, 1-4 October.
Yeten, B. and Jalali, Y. (2001). Effectiveness of Intelligent Completions in
a Multiwell Development Context, paper SPE 68077 presented at the SPE
Middle East Oil Show, Bahrain, 17-20 March.
Yu, S., Davies, D. R., and Sherrard, D. W. (2000). The Modelling of Advanced
“Intelligent” Well, An Application, paper SPE 62950 presented at the SPE
Annual Technical Conference and Exhibition, Dallas, Texas, 1-4 October.
23
Table 1
Simulation Model Properties
drainage area 5000 × 5000 ft2
oil thickness 50 ft
φ 0.25
crock 3.0×10−5 psi−1
kro 1.00 at Swc = 0.15
krw 0.57 at Sor = 0.15
γ at 14.7 psi
oil 0.85
water 1.0
µ, cp
oil 4.0 at 14.7 psi
water 0.30 at 4000 psi
B, V /V
oil 1.0 at 14.7 psi
water 1.02 at 4000 psi
Table 2
Case 1 - Comparison of Cumulative Oil Production
Geological Uncontrolled Instrumented Additional
Realization Base Case Case Recovery
# (MMSTB) (MMSTB) %
1 3.34 5.04 51.0
2 5.38 5.63 4.7
3 4.61 5.52 19.7
4 5.07 6.07 19.7
5 4.23 5.43 28.1
Average 4.53 5.54 24.7
Std. Dev. 0.79 0.37 17.0
24
Table 3
Case 1 - Comparison of NPV
Geological Uncontrolled Instrumented Incremental
Realization Base Case Case NPV
# (MM$) (MM$) %
1 60.1 88.7 47.7
2 96.8 99.3 2.6
3 83.0 97.4 17.3
4 91.3 107.3 17.6
5 76.2 95.7 25.5
Average 81.5 97.7 22.1
Std. Dev. 14.3 6.7 16.5
Table 4
Case 1 - Decision on Deploying Devices under Geological Uncertainty
Risk Attitude CE, MM$ CE, MM$ Increase in
Coefficient Base Case Inst. Case in CE, % Decision
-10 67.5 94.4 39.8 Deploy
-6 71.2 95.6 34.1 Deploy
-2 77.8 96.9 24.6 Deploy
0 81.5 97.7 19.9 Deploy
2 84.6 98.5 16.4 Deploy
6 88.5 100.0 12.9 Deploy
10 90.7 101.3 11.7 Deploy
25
Table 5
Case 2 - Decision on Deploying Devices with Uncertain Reliability
Risk Attitude
Coefficient Scenario I Scenario II Scenario III
-10 Do not deploy Do not deploy Do not deploy
-6 Do not deploy Deploy Deploy
-2 Deploy Deploy Deploy
0 Deploy Deploy Deploy
2 Deploy Deploy Deploy
6 Deploy Deploy Deploy
10 Deploy Deploy Deploy
Table 6
Case 3 - Comparison of Cumulative Oil Production
Geological Uncontrolled Instrumented Additional
Realization Base Case Case Recovery
# (MMSTB) (MMSTB) %
1 3.34 4.48 34.3
2 5.38 5.31 -1.2
3 4.61 5.11 10.8
4 5.07 5.61 10.7
5 4.23 4.97 17.4
Average 4.53 5.10 14.4
Std. Dev. 0.79 0.42 13.0
26
Table 7
Case 3 - Decision on Deploying Devices with Uncertain Geology and Reliability
Risk Attitude CE, MM$ CE, MM$ Increase in
Coefficient Base Case Inst. Case CE, % Decision
-10 67.2 78.1 16.3 Deploy
-6 69.9 83.0 18.7 Deploy
-2 76.4 87.8 14.9 Deploy
0 81.5 89.5 9.8 Deploy
2 86.0 90.6 5.4 Deploy
6 90.3 92.2 2.0 Deploy
10 91.7 93.1 1.6 Deploy
Table 8
Case 3 - Decision on Deploying Devices with Uncertain Reliability but Known Ge-
ology
Risk Attitude Geological Realization
Coefficient #1 #2 #3 #4 #5
-10 Deploy Do not deploy Do not deploy Do not deploy Deploy
-6 Deploy Do not deploy Deploy Do not deploy Deploy
-2 Deploy Do not deploy Deploy Do not deploy Deploy
0 Deploy Do not deploy Deploy Do not deploy Deploy
2 Deploy Do not deploy Deploy Do not deploy Deploy
6 Deploy Do not deploy Deploy Do not deploy Deploy
10 Deploy Do not deploy Deploy Deploy Deploy
27
Fig. 1. Optimization of Valve Settings for Each Time Step
Failure pdf
0.9
0.8
0.7
0.6
u(x)
0.5
0.4
0.3
0.2
Risk Neutral
0.1 Risk Averse
Risk Prone
0
0 0.2 0.4 0.6 0.8 1
x
28
10,000 md
1 md
α=3
α=5
0.30
α=7
0.25
0.20
pdf
0.15
0.10
0.05
0
0 1 2 3 4 5 6 7 8 9
Time, years
29
25
20
Occurences
15
10
0
0 1 2 3 4 5 6 7 8
Failure time, years
run1 run2
1 1
Valve #
2 2
3 3
4 4
run3 run4
1 1
Valve #
2 2
3 3
4 4
1 2 3 4 5 6 7 8
Time, years
no fail
fail stuck
fail open
fail closed
30
Occur.
20
10
0
−30 −25 −20 −15 −10 −5 0 5 10 15 20 25
Scenario I: Increase in cum oil, %
Occur.
20
10
0
−30 −25 −20 −15 −10 −5 0 5 10 15 20 25
Scenario II: Increase in cum oil, %
Occur.
20
10
0
−30 −25 −20 −15 −10 −5 0 5 10 15 20 25
Scenario III: Increase in cum oil, %
15
Diff. in cum. oil production, %
10
−5
−10
Fig. 10. Relation between the First Time of Failure and the Difference in Cumulative
Oil Production (Results for Scenario II)
31
Fig. 11. Simplified Version of the Decision Tree Used in Case 3
32