Proceedings of the 17th World Congress

The International Federation of Automatic Control

Seoul, Korea, July 6-11, 2008

Oil and Gas Production Optimization; Lost

Potential due to Uncertainty ⋆
Steinar M. Elgsaeter ∗ Olav Slupphaug ∗∗
Tor Arne Johansen ∗∗∗
∗
Norwegian University of Science and Technology, Trondheim, N-7491
Trondheim (Tel: 0047 91806886; e-mail: elgsater@ itk.ntnu.no)
∗∗
ABB, Oslo(e-mail: olav.slupphaug@no.abb.com)
∗∗∗
Norwegian University of Science and Technology, Trondheim,
N-7491 Trondheim (e-mail: torj@ itk.ntnu.no)

Abstract:
The information content in measurements of offshore oil and gas production is often low,
and when a production model is fitted to such data, uncertainty may result. If production is
optimized with an uncertain model, some potential production profit may be lost. The costs and
risks of introducing additional excitation are typically large and cannot be justified unless the
potential increase in profits are quantified. In previous work it is discussed how bootstrapping
can be used to estimate uncertainty resulting from fitting production models to data with
low information content. In this paper we propose how lost potential resulting from estimated
uncertainty can be estimated using Monte-Carlo analysis. Based on a conservative formulation
of the production optimization problem, a potential for production optimization in excess of
2% and lost potential due to the form of uncertainty considered in excess of 0.5% was identified
using field data from a North Sea field.

Keywords: Uncertainty, Monte Carlo simulation, Loss minimization, Process identification,

Process control, Production systems

1. INTRODUCTION Lift gas/

gas injection
Gaslift Wells Separation
Production in the context of offshore oil and gas fields, can
....
be considered the total output of production wells, a mass
....

....
Routing
Water Gas
flow with components including hydrocarbons, in addition injection
Oil Export
to water, CO2 , H2 S, sand and possibly other components.
....

Separation
Hydrocarbon production is for simplicity often lumped Gas
Water
Reservoir(s) ....
into oil and gas. Production travels as multiphase flow g
....

injection
from wells through flow lines to a processing facility for Reinjection/To sea
....

separation, illustrated in Figure 1. Water and gas injection

....

is used for optimizing hydrocarbon recovery of reservoirs.

Gas lift can increase production to a certain extent by
increasing the pressure difference between reservoir and Fig. 1. A schematic model of offshore oil and gas produc-
well inlet. tion.
Production is constrained by several factors, including: On
the field level, the capacity of the facilities to separate rates can often be found by adding different measured
components of production and the capacity of facilities to water rates after separation. To determine the rates of oil,
compress lift gas. The production of groups of wells may gas and water produced from individual wells, the produc-
travel through shared flow lines or inlet separators which tion of a single well is usually routed to a dedicated test
have a limited liquid handling capacity. The production separator where the rate of each separated single-phase
of individual wells may be constrained due to slugging, component is measured. In single-rate well tests, only the
other flow assurance issues or due to reservoir management rates of components with current production settings are
constraints. measured, while in multi-rate well tests rates are measured
for different settings. The total amount of gas and water
Multiphase flows are hard to measure and are usually not which can be separated and processed is constrained by
available for individual flow lines in real-time, however the capacity of facilities, these capacities are themselves
measurements of total single-phase produced oil and gas uncertain. Normally production is at setpoints where some
rates are usually available, and estimates of total water of these capacities are at their perceived constraints, there-
⋆ The authors would like to thank the Research Council of Norway, fore a multi-rate well test cannot be performed without
Norsk Hydro and ABB for funding this work. simultaneously reducing production at some other well,

978-3-902661-00-5/08/$20.00 © 2008 IFAC 4540 10.3182/20080706-5-KR-1001.2832

17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

which may cause lost production and a cost. There is

also a risk that changes in setpoints during testing may Planned
cause some part of the plant to exceed the limits of safe Production
constraints excitation Disturbances
operation, which may force an expensive shutdown and
re-start of production. Well tests are only performed when Decision
a need for tests has been identified, due to the costs and Production variables
risks involved.
optimization Production
In the context of oil and gas producing systems, real-time
optimization (RTO) has been defined in Saputelli et al. State and model
(2003) as a process of measure-calculate-control cycles at parameters Measurements
Parameter and
a frequency which maintains the system’s optimal operat-
ing conditions within the time-constant constraints of the state estimation
system. Saputelli et al. (2003) suggests dividing real-time
optimization into subproblems on different time scales to Fig. 2. Example 1: Relationship between subproblems in
limit complexity, and to consider separately reservoir man- production optimization.
agement, optimization of injection and reservoir drainage
on the time scales of months and years, and production duction data from actual fields. The field produces oil qo ,
optimization, maximization of value from the daily pro- water qw and gas qg from two wells, each with a gas lift
duction of reservoir fluids. Reservoir management typically rate qgl according to equations
specifies constraints on production optimization to link qoi = ci1 + ci2 (qgl
i l,i
− qgl i
) + ci3 (qgl l,i 2
− qgl ) ∀i ∈ {1, 2} (1)
these problems. We will refer to models for production
optimization as production models. qgi = i
KGOR qOi
(2)
i
The aim of production optimization is to determine set- i KW C
qw = i qo + w1 . (3)
points for a set of chosen decision variables which are op- 1 − KW C
timal by some criterion. These setpoints are implemented l,i
qgl l,1
are local operating points and qgl l,2
= qgl = 1000. Let
by altering the settings of production equipment. Decision
i △ i i i i i
 
variables may be any measured or computed variables as- θ = c1 c2 c3 KW C KGOR w1 and
sociated with the production system which are influenced  ⋆,1  
θ 39.62 0.025 −4 · 10−6 0.173 30 0

by changes in settings, but the number of decision variables θ⋆,2  76.62 0.030 −4 · 10−6 0.173 30 0 
is limited by the number of settings. We may for instance  1 =
 θ̂  39.62 0.033 −6 · 10−6 0.185 30 0.85 (4)

determine the settings of a gas lift choke by deciding a
target lift gas rate, a target annulus pressure or a target θ̂2 76.62 0.032 −12 · 10−6 0.160 30 1.5
gas lift choke opening. On short timescales the flow from
individual wells can be manipulated by production choke The ratio of water to produced liquid (watercut) and the
settings, by gas lift choke settings and possibly by routing gas-oil ratio are assumed rate-independent. True param-
settings. eters (θ⋆,1 , θ⋆,2 ) and fitted parameters (θ̂1 , θ̂2 ) produce
predictions of measured total oil qotot = qo1 + qO 2
, total
Parameters of production models should be fitted to pro- tot 1 2 tot 1
gas qg = qg + qg and total water qw = qw + qw which 2
duction data through parameter estimation to compen- match closely for a typical set of gas lift rates with little
sate for un-modeled effects or disturbances and to set excitation, shown in Figure 3, and which match the single-
reasonable values for physical parameters which cannot be 1
rate well test available for (qgl 2
, qgl ) = (1000, 1000).
measured directly or determined in the laboratory. Erosion
1 2
of production chokes is an example of an un-modeled Assume that at the current setpoint (qgl , qgl ) = (1000, 1000)
disturbance. tot
with oil production qo = 116.1, all gas processing capac-
Planned excitation is some planned variation in one or ity (associated gas + gas lift) and water processing capac-
more decision variables designed to reveal information on ity is utilized. The current setpoint is optimal according
production through measurements, for instance a multi- to numerical optimization with parameters (θ̂1 , θ̂2 ), yet
rate well test. The information flow in production opti- with the parameters (θ⋆,1 , θ⋆,2 ) (qgl
1 2
, qgl ) ≈ (744, 1238) is
tot
mization may be depicted as in Figure 2. optimal with oil production qo = 116.4. The two models
and the setpoints are illustrated in Figure 4.
Lost profit can result if production models are fitted
to production data with a low information content, as This example illustrates that lost profit can result even if
illustrated in Example 1. Throughout this paper variables production optimization is based on a model that fits pro-
with a hat ( ˆ ) denote estimates, variables with bars ( ¯ ) duction data well if production data has low information
denote measurements, and variables with a star ( ⋆ ) are content. 2
true values in some sense.
Example 1. (Uncertainty in production optimization). This 1.1 Prior work
example illustrates a synthetic field which is producing
at its processing constraints, where only single-rate well In Elgsaeter et al. (2007) an analysis of production data
tests are available and where a production model matches from an oil and gas field determined that information
historical production data with a low information content content appeared low and significant uncertainty should
closely. The example is motivated by observations of pro- be expected if production models are fitted to such data.
Handling uncertainty has been identified as a key challenge

4541
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

focus was on uncertainty in gas-oil and water-oil ratios for

120 wells where these ratios do not change when production
tot
qO change and where no gas-lift was applied and only test
110
separator measurements were considered.
100
0 50 100 150 200 250 300 In Bieker et al. (2007) it has been proposed that uncer-
3600 tainty in production optimization should be related to
qtot
G
3400 production profit to allow structured business cases for un-
certainty mitigation to be created. Elgsaeter et al. (2008)
3200 suggested how uncertainty due to low information content
0 50 100 150 200 250 300
28 in production data can be quantified using bootstrapping.
qtot No references have been found which discuss the value of
W
26
information in the context of daily oil and gas production
24
0 50 100 150 200 250 300
optimization.
1500
i
qgl 1.2 Problem formulation
1000

500 The objective of this paper is to develop an analytical

0 50 100 150 200 250 300
t[hours] framework to quantify the lost potential that results from
the low information content in production data used for
production optimization, suitable to offshore oil and gas
Fig. 3. Example 1: True (solid) and modeled (dashed) production.
total production rates of oil, gas and water resulting
1 2
from gas lift rates (lower graphs) qgl (dotted) and qgl 2. ESTIMATING THE LOSS DUE TO UNCERTAINTY
(dashdot).
2.1 Prerequisites

Let x be a vector of the flows of each modeled component

1 1 1 of production for each modeled well and let u be a vector
qO qG qW
of decision variables.
In this paper we will consider production optimization
problems on the form
 
q1 q1 q1
û(θ̂) x̂(θ̂) = arg max M (x, u, d) (5)
u,x
gl gl gl
s.t 0 = f (x, u, d, θ̂) (6)
0 ≤ c(x, u, d). (7)
q2 q2 q2
O G W θ̂ is a vector of fitted parameters to be determined. û(θ̂)
and x̂(θ̂) are the decision variable value and vector of
flows, respectively, that are the optimal solution of (5)-
(7) for a given θ̂. d is a vector of modeled and measured
q2 q2 q2 disturbances independent of u. M (x, u, d) is a production
gl gl gl
profit measure which we seek to maximize subject to a
production model (6) and production constraints (7) for
Fig. 4. Example 1: True (solid) and modeled (dashed)
production of oil,gas and water, well 1 (upper graphs) a given θ̂. We only consider instantaneous optimization,
and well 2 (lower graphs). Optimal operating point of determining (û(θ̂), x̂(θ̂)) for the current time.
the true production (square) and optimal operating When measurement uncertainty can be neglected, the
point of modeled production (circle). relationship between the measured production ȳ and x is
given by the binary routing matrix R[t]:
in real-time optimization of oil and gas production in
△
Bieker et al. (2006a). The concept of a loss resulting from ȳ[t] = R[t]x[t]. (8)
uncertainty is well established in control theory, it is for Routing R[t] may change with time, as wells are routed
instance used to select controlled variables in Halvorsen through different processing trains or to test separators.
et al. (2003). The concept of the value of information y may consist of test separator rate measurements, total
has been applied previously in reservoir management. In rate measurements and in some cases multiphase rate
Branco et al. (2005) an analysis of what the authors measurements at some or all wells.
called the value of the new information was used to justify
investments to determine static and dynamic reservoir Let the modeled output for a given parameter θ̂ be
behavior. Monte Carlo simulations have been applied △
ŷθ̂ [t] = R[t]x̂(θ̂)[t] + β̂y , (9)
for analysis of uncertainty in production optimization
previously in Bieker et al. (2006b), where it was applied where β̂y is the vector of measurement biases due to cali-
to well test planning, and in Bieker et al. (2007), where it bration inaccuracies to be determined. A set of production
was used for optimal well management. In both cases, the data spanning N time steps

4542
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

¯ ¯ a distribution in interval away from zero. To allow these


N ȳ[1] d[1] ū[1] ȳ[2] d[2] ū[2] . . .
Z = ¯ ] ū[N ]
 (10) two parts of fLu (L̂u ) and fPo (P̂o ) to be studied separately,
ȳ[N ] d[N
we define the two events
has residuals
H0 : production is optimal (16)
ǫ[t] = ȳ[t] − ŷθ̂ [t] ∀t ∈ {1, . . . , N } . (11)
H1 : production is suboptimal. (17)
Parameter estimation determines θ̂ by minimizing the sum Estimates p̂(H0 ) and p̂(H1 ) can be found from estimated
of the squared residuals for the data set: probability density functions fLu (L̂u ) and fPo (P̂o ).
N
w[t]ǫ[t]22 ,
 
θ̂ β̂y = arg min (12) 2.3 Estimating the potential of production optimization
θ,βy
t=1
where w[t] is a user-specified weighting function. When An estimate P̂o (θ̂) of (13) for a given parameter estimate
Z N has low information content, it may not be possible to
θ̂ when (x, u, d) is the current operating point is
determine θ̂ uniquely from (12), resulting in an uncertainty
in θ̂ that can be quantified in terms of a probability P̂o (θ̂) = M (x̂(θ̂), û(θ̂), d) − M (x, u, d). (18)
density function fθ (θ̂) using bootstrapping, see Efron and When θ̂ is uncertain, we may exploit an estimate of the
Tibshirani (1993) or Elgsaeter et al. (2008). Bootstrapping probabilty density function fθ (θ̂) to estimate a probability
is a computational approach which generates an estimate density function for (18) using a Monte Carlo method,
of uncertainty by re-solving (12) a number of times to generating suitable random numbers and observing the
fit the model to number of synthetically generated, or fraction of the numbers obeying some property or prop-
“re-sampled”, data sets similar to (10). In Monte-Carlo erties, see Metropolis and Ulam (1949). The approach is
simulations in subsequent sections we sample entire pa- outlined in Algorithm 1 in Appendix A.
rameter vectors rather than sampling parameter values for
each component of θ separately, ensuring that co-variance 2.4 Estimating lost potential due to uncertainty
between terms in θ̂ are implicitly accounted for.
A setpoint û(θ̂) calculated with an uncertain θ̂ may be
2.2 Definitions found to violate production constraints (7) when imple-
mented. Therefore, a strategy for constraint handling is
Let the potential for production optimization Po⋆ be the required. Lost potential due to uncertainty will depend on
difference between the production profit at the optimal the strategy employed to deal with constraint handling,
operating point (x⋆ , u⋆ , d) and the current operating point and to estimate lost potential due to uncertainty we must
(x, u, d): express this strategy as an algorithm. We will refer to such
△
Po⋆ = M (x⋆ , u⋆ , d) − M (x, u, d) ≥ 0. (13) an algorithm as an operational strategy function.
If there were no uncertainty, (5)–(7) would yield û(θ) = u⋆ Let u0 be the setpoint implemented on the field prior to
and x̂(θ) = x⋆ , Po⋆ could be found from (13) and the entire implementing û(θ̂). When altering setpoints from u0 along
potential could be realized by implementing û(θ). some path toward û(θ̂), constraints may be encountered
at an operating point (xc , uc ) due to uncertainties. Lost
When θ̂ is uncertain, the solution to (5)–(7) will be
potential due to uncertainty can be estimated in a Monte
uncertain. Uncertainty may cause estimates û(θ) and P̂o (θ̂) Carlo fashion by determining the profit that is lost when
to differ from u⋆ and Po⋆ , and when implementing û(θ) we
a setpoint û(θ̂f ) is implemented and the true parameter
must expect some of the potential Po⋆ to remain unrealized.
This discussion motivates value is actually θ̂t , drawing estimates (θ̂t , θ̂f ) from the
△ estimated probability density function fθ (θ̂):
expected potential of production optimization(P̂o ) =
L̂u (θ̂t , θ̂f , d) = M (x̂(θ̂t ), û(θ̂t ), d)−
expected realizable potential (P̂o,r ) + (19)
M (xc (θ̂f , θ̂t ), uc (θ̂f , θ̂t ), d).
expected lost potential due to uncertainty (L̂u ).
(14) The first term in (19) is the optimal profit with the true
Let parameter value θ̂t , while the second term is an estimate of
L̂u the profit obtained when θ̂f is erroneously assumed to be
△
Q̂ = , (15) the true parameter. (xc (θ̂f , θ̂t ), uc (θ̂f , θ̂t )) is determined by
P̂o an operational strategy function. The proposed algorithm
be the quotient of potential that is lost due to uncertainty. for determining fLu (L̂u ) is summarized in Algorithm 2 in
In this paper, we will determine probability density func- Appendix A.
tions fP o (P̂o ), fLu (L̂u ), discussed below. In addition to
depending on parameter uncertainty, these probability 2.5 Discussion
density functions will depend on production constraints
and on the operational strategy, defined below. In Algorithms 1 and 2 samples are drawn at random,
Monte-Carlo approach in this paper. It may be pos-
P̂o = 0 and L̂u = 0 can be interpreted as production being sible to improve accuracy and reduce computational
optimal. The probability density functions fLu (L̂u ) and burden through the use of sampling techniques based
fPo (P̂o ) may consist of spikes near P̂o = 0 and L̂u = 0 and on Hammersley-sequences, see Diwekar and Kalagnanam

4543
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

(1997). Algorithms 1 and 2 can also be combined with û(θ̂) = arg max q̂otot (θ̂) + bo (θ̂) (27)
more efficient bootstrap methods, see for instance Gigli u
(1996). To limit the scope and complexity of this paper, s.t. q̂gtot (θ̂) + bg (θ̂) ≤ qgmax (28)
we leave such extensions for further work. tot max
q̂w (θ̂)
+ bw (θ̂) ≤ qw (29)
−Uprc ≤ u ≤ Uprc (30)
3. APPLICATIONS
ui > ul,i ∀i ∈ Wc . (31)
In this section we will study applications of the suggested
(28) is a constraint on gas processing capacity and qgmax
approach, first on a set of synthetic examples, then on data
from a real field. is the processing capacity for gas. (29) is a constraint
max
on water processing capacity and qw the processing
capacity for water. (31) constrains the gas lift rate of wells
3.1 Introduction with indices in the set Wc which may not have their gas-lift
rates decreased due to flow assurance issues. bo (θ̂), bg (θ̂)
We will consider oil, gas and water rates for each well i
as elements in x̂, that is qoi , qgi and qw
i
, respectively. The and bw (θ̂) are biases which may be calculated separately
production of individual wells are assumed independent of for each θ̂. All constraints are considered hard in this
each other, or decoupled, as will typically be the case for paper.
so-called platform wells. We will consider two alternative What model structure describes the relation between u
production models, (20) and (21), and y is itself uncertain. It is reasonable to expect a fitted
q̂pi = max{0, qpl,i fz (z i , z i,l ) 1 + αip ∆qgl
i
} (20) model to be a valid description locally around setpoints
(u, d) similar to those observed in the tuning set. The
q̂pi = max{0, qpl,i fz (z i , z i,l ) 1+ i
αip ∆qgl + i 2
κip (∆qgl ) }, mismatch between modeled outputs ŷ and measurements
(21) ȳ will grow when setpoints outside the range of setpoints
for all wells i and for oil, gas and water p ∈ {o, g, w}, observed in (10) are considered. If production optimization
i considers large changes in u, structural uncertainties may
based on Elgsaeter et al. (2008). qgl is gas lift rate and
i
z ∈ [0, 1] is the relative production valve opening of dominate and this observation motivates (30). Uprc limits
i △ i
qgl l,i u to within Uprc · 100% of its current value and is a design
well i. ∆qgl = l,i − 1, where qgl is the gas lift rate paramter that must be chosen sufficiently small.
qgl
measured at the time of optimization for well i, and qpl,i qgmax and qw max
are themselves uncertain in practice, but
is the measured rate of component p of well i at the identifying or mitigating this uncertainty is beyond the
most recent well test, respectively. fz (z i , z l,i ) is a nonlinear scope of this paper. To avoid overstating qgmax and qw max
,
kernel which expresses the nonlinear relationship between
and thereby overestimating L̂u and P̂o , we assume con-
valve opening and production, which obeys fz (0, z l,i ) = 0 servatively that the field is producing at its constraints in
and fz (z i , z l,i ) = 1. z l,i is the relative valve opening at the
water and gas capacity at the time of optimization. We
most recent well test. In this paper we choose
wish to solve (27)–(31) at time t = N , the end of the
1 − (1 − z i )k tuning set (10). Measurements ȳ have a high-frequency,
fz (z i , z l,i ) = , (22)
1 − (1 − z l,i )k low-amplitude, seemingly random component which is not
and choose k = 5. During optimization, we consider modeled. If we choose qgmax = q̄gtot [N ] and qwmax tot
= q̄w [N ],
u = ∆qgl , d = z. We will consider θ = [αo αg αw ] for solutions of (27)–(31) will depend to some degree on these
(20) and θ = [αo αg αw κo κg κw ] for (21). high-frequency disturbances. For this reason, we choose to
consider tuning sets where no significant changes are visi-
Let ¯ for
ble in ū[t] and no large disturbances visible in ȳ[t], d[t]
nw
△ the interval t ∈ [N − L, N ] where L << N . Capacities are
qotot = qoi (23) estimated as average measured rates during this interval,
i=1 to reduce dependency on high-frequency disturbances:
nw
△
qgtot = qgi + qgl
i
(24) qgmax = avg(q̄gtot [t]) t ∈ [N − L, N ] (32)
max tot
i=1 qw = avg(q̄w [t]) t ∈ [N − L, N ]. (33)
nw
tot
qw
△
= qoi (25) Similarly, we desire q̂otot (θ̂), q̂gtot (θ̂)and tot
q̂w (θ̂)
to equal
i=1
averaged observations over t ∈ [N − L, N ] which motivates
 tot
q̄o [t] − q̂otot (θ̂)[t]
  
where nw is the number of wells. Let q̂otot (θ̂), q̂gtot (θ̂) and bo (θ̂)
tot  bg (θ̂)  = avg(q̄gtot [t] − q̂gtot (θ̂)[t]) t ∈ [N − L, N ].
q̂w (θ̂) be estimates derived by combining (23)–(25) with
tot tot
(20) or (21) for a given θ̂. bw (θ̂) q̄w [t] − q̂w (θ̂)[t]
(34)
In this paper we will consider the following measurement
vector when solving (12) bo (θ̂) is included in the objective function (27) so that (18)
tot T yields P̂o = 0 if (27)–(31) returns û(θ) equal to the imple-
y = qotot qgtot qw
 
. (26)
mented u at the time of optimization. (bo (θ̂), bg (θ̂), bw (θ̂))
The objective function is chosen as (27) and production may differ from elements of β̂y , as β̂y is an estimate of the
constraints as (28)-(31): bias over in the tuning set, while bo (θ̂), bg (θ̂) and bw (θ̂) are

4544
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

biases calculated for a short time interval at the end of the

tuning set. f (%)
P |H ] 1
o 1
0.8
We will consider the operational strategy function 0.6
0.4
uc (θ̂f , θ̂t ) = u0 + s(û(θ̂f ) − u0 ) (35) 0.2
c c
0 = f (x (θ̂f , θ̂t ), u (θ̂f , θ̂t ), d, θ̂t ) (36) 0.1 0.2 0.3 0.4 0.5
where s = max ŝ (37) fL |H (%)
u 1 0.2
s.t 0 ≤ ŝ ≤ 1 (38)
u = u0 + ŝ(û(θ̂f ) − u0 ) (39) 0.1
0 ≤ c(x, u, d). (40)
f 0.1 0.2 0.3 0.4 0.5
0 = f (x, u, d, θ̂t ). (41) Q|H
1 1
(35)-(41) simulates moving the setpoint from u0 to-
ward û(θ̂f ) in a straight line until a constraint is met, 0.5
when the system is described by θ̂t . If implementing
(xc (θ̂f , θ̂t ), uc (θ̂f , θ̂t )) results in a decline in production 0.1 0.2 0.3 0.4 0.5
profits, we will assume that the operating point is returned Uprc
to is value prior to optimization when solving (19).
The set of operating points which are feasible is non- Fig. 5. Example 2: Estimated potential and lost potential
convex in general, and the conservative statement of gas for formulation A (plus) and B(circles). Center graphs
and water capacities qgmax and qw max
means that the choice are expected values, upper and lower graphs denote
u0 = 0 may cause (35)–(41) to meet a constraint close to 95% confidence intervals.
u0 , producing a large estimate of loss due to uncertainty
L̂u . Instead, we choose u0 = −Uprc , as simulations have
shown this estimate to produce far more conservative
estimates of L̂u . The real-world analogy to u0 = −Uprc
is a field where production is moved toward a calculated q1
o
q1
g
q1
w
setpoint û(θ̂) as production is re-started after a shutdown.
We will consider two different formulations of the problem
(27)-(31):
1 1 1
qgl qgl qgl
Formulation A: production model (20), a linear pro-
gramming problem
Formulation B: production model (21), a nonlinear pro-
gramming problem
q2 q2 q2
o g w
The applications in the following section were imple-
mented in MATLAB 1 and all optimization problems were
solved with the TOMLAB 2 toolbox. All nonlinear pro-
gramming problems were solved sequentially with global
solvers Jones et al. (1993) and local solvers based on se- q2 q2 q2
gl gl gl

quential quadratic programming, for parameter estimation

based on Huschens (1994), and for production optimiza-
tion based on Schittkowski (1982). In all cases analytical Fig. 6. Example 2: Production models (20) (dotted) and
Jacobians and Hessians were supplied. (21)(dashed) based on parameters derived using boot-
strapping, operating point (xl , ul ) (circle). Vertical
line illustrates span of decision variable values in
3.2 Synthetic example tuning set. The solid line illustrates the true model.
To reduce clutter, only every fifth model based on
We will revisit Example 1 in the following, to illustrate the resampled parameters are plotted.
the suggested methodology on a set of data where the true
model is known. The use of a nonlinear production model in formulation B
Example 2. (Example 1 revisited: fPo (P̂o ) and fLu (L̂u )). resulted in estimates P̂o and L̂u that were less dependent
We apply Algorithms 1 and 2 to the synthetic production of Uprc than those found with formulation A. While
data described in Example 1 with Nt = Nf = 100 for estimates of P̂o remained in the region of Po⋆ = 0.29%
different Uprc . Results are shown in Figure 5. For this ex- with formulation B as Uprc grows, estimates P̂o found
ample, all estimates produced by Algorithm 1 gave P̂o > 0 with formulation A grow without bounds as Uprc grows.
using both formulations A and B, or p̂(H1 ) = 100%. A closer analysis reveals that formulation A tends to find
estimates û(θ) at the constraint (30), while formulation B
1 The Mathworks,Inc., version 7.0.4.365 finds estimates û(θ) within this constraint, and this may
2 TOMLAB Optimization Inc., version 5.5 explain the difference observed.

4545
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

As seen in Figure 6, the linear models used in formulation fP (%)

|H
A predicts oil, gas and water rates within a narrow span o 1

as qgli
change compared with the nonlinear models used 10
8
in formulation B. The narrow span in predicted rates 6
4
with formulation A causes comparatively small estimates 2
(L̂u , Q̂) compared with formulation B. Estimates of Q̂ 0.1 0.2 0.3 0.4 0.5
found with formulation A decreases as Uprc increases, f (%)
L |H
u 1
which seems counter-intuitive.
2
This example indicates that nonlinear models on the form
(21) may be more suitable than linear models (20) in 0
0.1 0.2 0.3 0.4 0.5
conjunction with the methods suggested in this paper. 2 f
Q|H
1
0.5
3.3 Case study: North Sea field 0.4
0.3
0.2
0.1
The set of data we will study is from a North Sea field with
0.1 0.2 0.3 0.4 0.5
20 gas lifted platform wells, producing predominantly oil, U
gas and water. Sporadic single-rate well tests are available, prc

and no routing options are considered in u. The operator

of the field requested that all data be kept anonymous, Fig. 7. Case study, North Sea field data: estimated poten-
therefore all variables will be represented in normalized tial and lost potential for formulation A (plus) and
form. We choose a tuning set of 180 days. The sampling B (circles). Center graphs are expected values, upper
time was one hour, and production was not significantly and lower graphs denote 95% confidence intervals.
changed for the last L = 10 hours prior to t = N . ing capacities. In this case, the potential for production
fPo (P̂o ) was estimated with Algorithm 1 and fLu (L̂u ) was optimization would be larger than our estimates.
estimated with Algorithm 2, both with Nf = Nt = 100.
When estimating parameters, bounds 0 < α < 1 and −1 < 4. CONCLUSIONS
κ < 0 based on physical knowledge were applied. Based on
the knowledge that the watercut is rate-independent, soft Based on a conservative formulation of the production
constraints which penalize deviation from αo = αw and optimization problem, we could identify potential for pro-
κo = κw were added to the objective function. A decline duction optimization in excess of 2%, and lost potential
in q̄otot was visible in the tuning set, and we chose to de- due to the form of uncertainty considered in excess of
trend q̄otot and chose w[t] to weigh older measurements less 0.5%. As sales revenues from an oil and gas field may
than newer measurements when solving (12). For further be very large and the costs of mitigating uncertainty are
details on the modeling, refer to Elgsaeter et al. (2008). independent of revenue, we believe that further work may
The set of indices of wells with flow assurance issues was prove that these potentials are sufficient to justify the costs
Wc = {1, 4, 10, 13, 14, 16}. of mitigating uncertainty and optimizing production.
The methodology presented may have extensions to other
Results The distribution of estimates P̂o , L̂u and Q̂ for
processes optimized using models fitted to process mea-
varying Uprc are shown in Figure 7. The models used are
surements with low information content.
shown in Figure 8. For formulation A p(H1 ) = 100% while
p(H1 ) = 95% for formulation B.
Appendix A. ALGORITHMS
Discussion Uprc limits change in decision variables to
a certain range for which structural uncertainties can be Algorithm 1. (Estimating fPo (P̂o )). Given a dataset Z N
neglected, and P̂o and L̂u depend on this design parameter. on the form (10)
As we do not know with certainty when structural uncer-
• determine θ̂ using (12),
tainties become significant, we cannot determine a single
• determine fθ (θ̂), for instance by bootstrapping as
value for Uprc and so when estimating P̂o and L̂u we have
described in Elgsaeter et al. (2007),
considered a range of Uprc . A conservative assumption
of Uprc = 0.1 indicates Po ≈ 2%. The quotient Q is • sample fθ (θ̂) Nt times,
relatively invariant of Uprc , so from the analysis it seems • determine (x̂(θ), û(θ)) using (5)-(7) for each of the Nt
fair to conclude that on average, we should expect to loose samples, and
about 25% of the P̂o due to uncertainty, or L̂u ≈ 0.5% • insert all determined (x̂(θ), û(θ)) in (18) to obtain the
when Uprc = 0.1. Unfortunately, we are unable to verify probability density function fPo (P̂o ).
estimates P̂o and L̂u made in this paper without further Algorithm 2. (Estimating fLu (L̂u )). Given a dataset Z N
experiments, and such experiments are associated with on the form (10), an operational strategy function, the
great cost and risk. However, the methods presented in this current operating point (x, u, d) and a probability density
paper may be used motivate such experiments in further function fθ (θ̂).
work, as the presented methods allow the associated ben-
efits to be estimated. In this paper we have not considered • do Nt times
constraint uncertainty. It is possible that at the time of the · draw a sample θ̂t from fθ (θ̂).
analysis the field was producing below its actual process- · draw Nf samples θ̂f from fθ (θ̂).

4546
17th IFAC World Congress (IFAC'08)
Seoul, Korea, July 6-11, 2008

H.P. Bieker, O. Slupphaug, and T.A. Johansen. Well

qoi qgi i
qw management under uncertain gas or water oil ratios. In
1 SPE Digital Energy Conference and Exhibition, 2007.
SPE 106959.
2 C.C.M Branco, A.C.C. Pinto, P.M.B. Tinoco, P.M.F.
Vieira, A.M. Sayd, R.L.A. Santos, and F. Prais. The
3 role of the value of information and long horizontal wells
in the appraisal and development studies of a brazilian
4 offshore heavy-oil reservoir. In SPE/PS-CIM/CHOA
5
International Thermal Operations and Heavy Oil Sym-
posium, 2005. SPE 97846.
6 U.M. Diwekar and J.R. Kalagnanam. Efficient sampling
technique for optimization under uncertainty. AIChE
7 Journal, 43(2):440–447, 1997.
B Efron and R.J. Tibshirani. An Introduction to the
8 Bootstrap. Chapman & Hall, 1993.
S.M. Elgsaeter, O. Slupphaug, and T.A. Johansen. Un-
9 certainty in model-based production optimization; chal-
lenges. In Proc. SPE International Petroleum Technol-
10
ogy Conference 2007, 2007.
11 S.M. Elgsaeter, O. Slupphaug, and T.A. Johansen. Pro-
duction optimization; system identification and uncer-
12 tainty estimation. In Proc. SPE Intelligent Energy
Conference 2008, 2008.
13 A. Gigli. Efficient bootstrap methods; a review. Statistical
Methods and Applications, 5(1):99–127, April 1996.
14 I.J Halvorsen, S. Skogestad, J. Morud, and V. Alstad.
Optimal selection of controlled variables. Ind. Eng.
15
Chem. Res., 42(14):3273–3284, 2003.
16 J. Huschens. On the use of product structure in secant
methods for nonlinear least squares problems. SIAM
17 Journal of Optimization, 4(1):108–129, Jan 1994.
D.R. Jones, C.D. Perttunen, and B.E. Stuckman. Lip-
18 schitzian optimization without the lipschitz constant.
Journal of Optimization Theory and Applications, 79(1):
19 157–181, Oct 1993.
N. Metropolis and S. Ulam. The monte carlo method.
20 Journal of the American Statistical Association, 44
(247):335–341, September 1949.
i i i
qgl qgl qgl L.A. Saputelli, S. Mochizuki, L. Hutchins, R. Cramer,
M.B. Anderson, J.B. Mueller, A. Escorcia, A.L. Harms,
Fig. 8. Case study, North Sea field data: Production mod- C.D. Sisk, S. Pennebaker, J.T. Han, C.S. Brown,
els (20) (dotted) and (21)(dashed) based on parame- A. Kabir, R.D. Reese, G.J. Nuez, K.M. Landgren, C.J.
ters derived by bootstrapping, operating point (xl , ul ) McKie, and C. Airlie. Promoting real-time optimiza-
(circle). Vertical lines illustrate the range of decision tion of hydrocarbon producing systems. In 2003 SPE
variable values in the tuning set.Lower left bound of Offshore Europe Aberdeen, 2003. SPE 839781.
all plots is (0,0).Well indices along left margin. K. Schittkowski. On the convergence of a sequential
quadratic programming method with an augmented la-
· for each of the Nf samples of θ̂f and the sample grangian line search function. Technical Report, System
θ̂t obtain a point estimate of Lu by solving (19) Optimization Laboratory, Stanford University, Stan-
ford,CA, 1982.
• the distribution of Nt × Nf point estimates of L̂u is
an estimate of fLu (L̂u )

REFERENCES
H.P Bieker, O. Slupphaug, and T.A Johansen. Real-time
production optimization of offshore oil and gas produc-
tion systems: Technology survey. In SPE Intelligent
Energy Conference and Exhibition, 2006a. SPE 99446.
H.P. Bieker, O. Slupphaug, and T.A. Johansen. Optimal
well-testing strategy for production optimization: A
monte carlo simulation approach. In SPE Eastern
Regional Meeting, 2006b. SPE 104535.

4547