Você está na página 1de 38

A stru

tured approa h to optimizing oshore oil


and gas produ tion with un ertain models
Steinar M. Elgster, Norwegian University of S ien e and Te hnology,
O.S Bragstads plass 2D, N-7491 Trondheim,Norway
Olav Slupphaug, ABB, Ole Deviks vei 10, N-0666 Oslo, Norway
Tor Arne Johansen, Norwegian University of S ien e and Te hnology,
O.S Bragstads plass 2D, N-7491 Trondheim, Norway

Abstra t
Optimizing oshore produ tion of oil and gas has re eived omparatively little attention despite the large s ale of revenues involved.

The omplexity of

multiphase ow means that any model for use in produ tion optimization must
be tted to produ tion data for a ura y, but the low information ontent of
produ tion data means that the un ertainty in the tted parameters of any
su h model will be signi ant. Due to osts and risk the information ontent in
produ tion data annot be in reased through ex itation unless the benets are
do umented.
A stru tured approa h is suggested whi h iteratively updates setpoints while
do umenting the benets of ea h proposed setpoint hange through ex itation
planning and result analysis. In simulations on an analog whi h mimi s a realworld oil eld and its typi al low-information ontent data the approa h is able
to realize a signi ant portion of the available prot potential while ensuring
feasibility despite large initial model un ertainty.
Keywords: Un ertainty, Produ tion optimization, Parameter estimation, Oil
and gas produ tion, Ex itation Planning, Result Analysis

Computers & Chemi al Engineering,Volume 34, Issue 2, 8 February 2010, Pages


163176, (Author submitted, unformatted version)

1 Introdu tion
The potential for optimizing oshore oil and gas produ tion may be signi ant
as small in reases in prots in relative terms may translate into large gains
due to s ales of revenues involved (Elgsaeter, Slupphaug and Johansen, 2008a),
and as this topi has re eived little attention ompared to optimization and
ontrol of downstream pro essing fa ilities. Both modeling and measuring the
oshore produ tion of oil and gas present hallenges for optimization. Exploiting
the information in what measurements are available to t produ tion models
is ompli ated by the low information ontent of produ tion data (Elgsaeter,
Slupphaug and Johansen, 2007). Pra titioners in the oil and gas industry are
risk averse as the s ale of revenues mean ost and risks of implementing hanges
in setpoints an be signi ant. An approa h to optimization whi h an quantify

these expe ted benets may have in reased likelihood of industry a eptan e.
This paper proposes a stru tured approa h to optimizing produ tion whi h takes
the nature of models, measurements and data into a ount, while the expe ted
monetary benets of ea h setpoint hange are quantied.

Produ tion

in the ontext of oshore oil and gas elds, an be onsidered

the total output of produ tion wells, a mass ow with omponents in luding
hydro arbons, in addition to water,
ponents.

CO2 , H2 S ,

sand and possibly other om-

Hydro arbon produ tion is for simpli ity often lumped into oil and

gas. Produ tion travels as multiphase ow from wells through ow lines to a
pro essing fa ility for separation. Water and gas inje tion is used for optimizing
hydro arbon re overy of reservoirs. Gas lift an in rease produ tion to a ertain
extent by in reasing the pressure dieren e between reservoir and well inlet.
Produ tion is onstrained by several fa tors, in luding: On the eld level,
the apa ity of the fa ilities to separate omponents of produ tion and the
apa ity of fa ilities to ompress gas. The produ tion of groups of wells may
travel through shared ow lines or inlet separators whi h have a limited liquid
handling apa ity. The produ tion of individual wells may be onstrained due to
slugging, other ow assuran e issues or due to reservoir management onstraints.
Multiphase ows are hard to measure and are usually not available for individual ow lines in real-time, however measurements of total single-phase produ ed oil and gas rates are usually available, and estimates of total water rates
an often be found by adding dierent measured water rates after separation.
To determine the rates of oil, gas and water produ ed from individual wells,
the produ tion of a single well is usually routed to a dedi ated test separator at
intervals where the rate of ea h separated single-phase omponent is measured,
a

well test.

Well tests allow biases in models of individual wells to be updated,

and well tests whi h measure rates for dierent settings,

multi-rate

well tests,

also allow responses predi ted by models to be validated.


In the produ tion of oil and gas it is ommon to divide the task of optimization into subproblems on dierent time s ales to limit omplexity, and
to onsider separately

reservoir management,

the optimization of reservoir in-

je tion and drainage on the time s ales of months and years, and

optimization,

produ tion

the maximization of prot from the daily produ tion of reservoir

uids (Saputelli, Mo hizuki, Hut hins, Cramer, Anderson, Mueller, Es or ia,


Harms, Sisk, Pennebaker, Han, Brown, Kabir, Reese, Nuez, Landgren, M Kie
and Airlie, 2003). Reservoir management typi ally spe ies onstraints on produ tion optimization to link these problems.
Produ tion optimization requires

produ tion models,

equations whi h ex-

press the relationship between hange in de ision variables and resulting hange
in produ tion. Produ tion models are usually nonlinear to des ribe signi ant
nonlinear phenomena in produ tion.

Parameter estimation is adjustment of

tted parameters so that predi tions of the produ tion model mat h a set of re ent histori al produ tion data as losely as possible. Parameters of produ tion
models should be tted to produ tion data through parameter estimation to
ompensate for un-modeled ee ts or disturban es and to set reasonable values
for physi al parameters whi h annot be measured dire tly or determined in
the laboratory. Erosion of produ tion hokes is an example of a typi al su h
un-modeled disturban e.

The term

information ontent

in this paper refers

the amount in variation in de ision- and disturban e variables observed in produ tion data.

Fitted parameters may be un ertain if tted to data with low

information ontent, as many parameter estimates may t data equally well for
a given model. If there is su ient information ontent in data a unique parameter estimate an be found robustly whi h mat hes data better than all other
estimates. A previous study on luded that un ertainty in tted parameters is
likely to result when produ tion models are tted to produ tion data des ribing normal operations, due to low information ontent (Elgsaeter et al., 2007).

Planned ex itation

is variation in de ision or disturban e variables introdu ed

deliberately for the purpose of exposing some aspe t of produ tion and redu ing
parameter un ertainty. The on epts of information ontent and ex itation used
in this paper are motivated by the eld of system identi ation (Ljung, 1999).
As planned ex itation in the oshore produ tion of oil and gas requires a temporary redu tion in produ tion, planned ex itation has a signi ant asso iated
ost and ompletely eliminating parameter un ertainty may therefore not be
ost-ee tive.
As a onsequen e of the omplexity of the pro ess onsidered, of measurement di ulties, and of the low information ontent in data, produ tion models
may be subje t to signi ant un ertainty that an be redu ed against a ost or
not at all. As a result more than one setpoint an be the plausible optimum,
whi h raises several pra ti al hallenges. One hallenge is that the result of a
setpoint hange is un ertain, it may even be negative. Another hallenge is that
ensuring that an implemented setpoint is

feasible,

i.e. that it obeys produ tion

onstraints, is non-trivial in the presen e of un ertainty. A third hallenge is


that when un ertainty an be redu ed at a ost, some method of determining
how and when un ertainty redu tion is ost-ee tive may be required.
The approa h to optimization with un ertain models is motivated by urrent pra ti e in the optimization of oshore oil and gas produ tion.

In the

urrent pra ti e, produ tion optimization begins by implementing well tests, a


form of planned ex itation.

Whi h wells and when to test are hosen based

on the subje tive insight of pra titioners. Produ tion models are updated by
tting against the most re ent well test. An target setpoint is then al ulated
by mathemati al programming. If the target setpoint is expe ted to in rease
prots signi antly, a setpoint hange is implemented. The target setpoint is
not implemented instantaneously, instead produ tion is moved toward the target setpoint by an operator in a series of smaller steps to limit transients and
to ensure feasibility. This type of gradual setpoint hange to ensure feasibility
post-optimization will be referred to as an

operational strategy

in this paper.

Produ tion optimization therefore involves ex itation planning, model updating, risk analysis and operational strategy, and urrent industry pra ti e does
not expli itly onsider un ertainty in any of these steps.

This is paradoxi al

as the state of produ tion data and omplexity of multiphase ow di tate that
un ertainty will be signi ant in ea h of these steps. The ontribution of this
paper is to suggest how to un ertainty an be a ounted for in a stru tured
manner in ea h of these steps.

1.1 Prior work


1.1.1 Modeling produ tion and quantifying model un ertainty
Quantifying un ertainty in reservoir models has attra ted mu h interest in re ent years.

Authors have des ribed tting multiple models to data (Griess,

Diab and S hulze-Riegert, 2006), des ribing measurement un ertainties (Little,


Fin ham and Jutila, 2006) and sensitivity analysis using prior knowledge of parameter un ertainty (Costa, S hiozer and Poletto, 2006). Very few referen es
have been found whi h attempt to quantify un ertainty in models for produ tion optimization, whi h seems paradoxi al as produ tion models are tted to
mu h the same produ tion data as reservoir models.

Monte-Carlo methods

are a lass of omputational algorithms, suitable for

the study of physi al or mathemati al systems with random or un ertain properties, whi h sample a probability distribution using a pseudo-random number
generator with uniform probability and observe the fra tion of the numbers
obeying some property or properties (Metropolis and Ulam, 1949). A MonteCarlo method for planning single-rate well tests under un ertainty was explored
in (Bieker, Slupphaug and Johansen, 2006).

Bootstrapping

is a Monte-Carlo

type method designed to estimate the un ertainty in parameters tted to a set


of data through regression (Efron and Tibshirani, 1993).
Elgsaeter, Slupphaug and Johansen (2008b) onsidered a method for modeling produ tion for produ tion optimization based on lo al valid linear and
nonlinear stationary models, motivated by the on epts of system identi ation.

Bootstrapping methods were used to nd a large number of multiple

parameter estimates whi h des ribe the produ tion data equally well. This set
of multiple parameter estimates expresses the onsequen es of low information
ontent on the tted parameters for the model stru ture hosen. The magnitude of estimated parameter un ertainty of even simple models was found to be
signi ant. Although this paper onsidered bla k-box type system identi ation
models, bootstrapping an be applied regardless of the hoi e of model, as bootstrapping is a omputational approa h to quantifying un ertainty whi h makes
no assumptions about the hosen model stru ture. Elgsaeter et al. (2008a) estimated the lost potential of produ tion optimization aused by un ertainty in
tted parameters to be several per ent of revenue for the ase of a North Sea oil
and gas eld, based on s enario simulations. These earlier ndings motivate the
fo us of this paper on handling quantied parameter un ertainty in produ tion
optimization.

1.1.2 Optimization under un ertainty


Produ tion optimization an be seen as a form of

real-time optimization system,

a losed-loop ontroller whi h attempts to lo ate the optimal setpoint through a


series of steps, ea h onsisting of a smaller setpoint hange. Real-time optimization systems, originally developed for hemi al pro esses, onsider an e onomi
obje tive fun tion maximized subje t to a rigorous steady-state nonlinear pro ess model and pro ess onstraints. Real-time optimization is a form of modelbased ontrol, and obtaining the pro ess model is onsidered the single most
di ult and time- onsuming task in the appli ation and maintenan e of modelbased ontrol and optimization (Andersen, Rasmussen and Jorgensen, 1991)
(Terwies h, Agarwal and Rippin, 1994) (Zhu, 2006).
Some parameters in pro ess models are typi ally tted to data in real-time
optimization. Ogunnaike (1995) has suggested that modeling for ontrol should
be based on riteria related to the a tual end use, and that tting equations
to data may be inadequate in the ontext of ontroller design. A onstraint on
the maximum setpoint hange is often applied in real-time optimization, and

smaller hanges in setpoints implemented iteratively with re-identi ation and


re-optimization between ea h step, a

two-step approa h.

Un ertainty in real-time optimization falls into four main ategories (Zhang,

market un ertainty, the impre ise knowledge of


pro ess un ertainty, the impre ise knowledge of operation
disturban es or un ertain inputs, measurement un ertainty, the

Monder and Forbes, 2002):


pro ess e onomi s,
due to pro ess

impre ise knowledge of measured pro ess variables due to sensor or transmission
errors, and

model un ertainty, plant/model stru tural and parametri mismat h.

Expli it analysis of un ertainty has gained some attention in ontrol design,


the methods of

robust ontrol

address the problem of designing linear multivari-

able ontrollers that adhere to some robust stability and performan e riterion
(Skogestad and Postlethwaite, 1996). Ensuring that a pro ess is designed with
su ient exibility to operate under hanging and un ertain un ertain parameters represented by probability distributions has been suggested, usually under
the assumption of linear models and normal distributions, see for instan e Pistikopoulos and Mazzu hi (1990).
Approa hes to handling un ertainty in optimization an be divided into
sto hasti optimization, sensitivity analysis, ba k-o and robust optimization
methods.

Sto hasti optimization

attempts to dire tly solve a problem given

un ertainty by formulating the sto hasti optimization problem in deterministi form using probabilities (Kall and Walla e, 1994). The

sensitivity analysis

approa h onsists of augmenting the obje tive fun tion of the optimization problem with a penalty term intended to minimize parametri sensitivity (Be ker,
Hall and Rustem, 1994). The

ba k-o

method adds a ve tor onstraint on de i-

sion variables to ensure that operating points are feasible, and the ba k-o ve tor
is omputed at intervals (Loeblein, Perkins, Srinivasan and Bonvin, 1999).

bust optimization

Ro-

optimizes the expe ted value for a hosen performan e index

for a given level of risk, formulated in terms of worst- ase, mean-varian e and so
forth (Darlington, Pantelides, Rustem and Tanyi, 1999) (Mulvey and Vanderbei, 1995). All these approa hes introdu e onservativeness to ensure feasibility
rather than exploiting measurements postoptimization.
Optimization methods whi h attempt exploit measurements fall into the
two main ategories: modied two-step approa hes and approa hes whi h do
not update pro ess models. Roberts and Williams (1981) suggests a modied
two-step approa h whi h add a gradient modi ation term to the ost fun tion
of the optimization problem to ensure that iterates onverge to a point at whi h
the ne essary onditions for optimality are satised.

Methods whi h do not

require model updating an be lassied into model-free and model-xed methods. Some model-free methods mimi various iterative numeri al optimization
algorithms, su h the Nelder-Mead simplex algorithm (Box and Draper, 1969).
Other model-free methods re ast the nonlinear programming problem into a
problem of hoosing de ision variables whose optimal values are approximately
invariant to un ertainty.

Self-optimizing ontrol

onsiders determining whi h

variables to keep at onstant setpoints to keep the pro ess a eptably lose
to optimum (Skogestad, 2000).

Extremum-seeking ontrol

are adaptive ontrol

methods related to self-optimizing ontrol whi h attempt to move pro ess setpoints toward values whi h result in an extreme value (maximum or minimum)
of a measured output without a pro ess model. Extremum-seeking ontrol usually introdu es ample ex itation in setpoints to a hieve this goal, for instan e in
form of a low-frequen y sinusoidal referen e (Krsti , 2000).

NCO-tra king

is a

model-free method whi h uses o-line analysis of the pro ess model to determine
fun tions for setpoint values at whi h the ne essary onditions for optimum are
enfor ed, parameterized in terms of measured variables (Franois, Srinivasan
and Bonvin, 2005).

Fixed-model methods

utilize both the available measure-

ments and a pro ess model for guiding the iterative s heme toward an optimal
operating point, but update onstraint and ost fun tion rather than model at
ea h iteration (Forbes and Marlin, 1994). Methods su h as extremum-seeking or
those whi h mimi iterative numeri al optimization algorithms introdu e many
variations to setpoints. These methods may be unsuitable to the oshore produ tion of oil and gas where the osts and risk of ea h hange in setpoint may
be signi ant. So- alled model-free methods still require models in the design
of ontrollers, and obtaining a urate models of the oshore produ tion of oil
and gas oine may be hallenging.
There is some pre eden e for analyzing the risk of implementing a setpoint
hange. Mileti and Marlin (1998) have proposed

result analysis

using multi-

variable statisti al hypothesis testing to determine whether the predi ted in rease in prot from implementing hanges in setpoints is statisti ally signi ant
or a result of pro ess noise.
A lassi theoreti approa h to un ertainty redu tion is

dual ontrol theory,

whi h deals with the ontroller design for pro esses whi h are initially unknown
(Fel'dbaum, 1961a) (Fel'dbaum, 1961b). The theory is alled dual as the obje tives of su h a ontroller are twofold, rstly to ontrol the system as well as
possible based on urrent system knowledge, se ondly to experiment with the
system so as to better learn how to ontrol it in the future.

The problem of

reinfor ement
learning, an area of ma hine learning on erned with how an agent ought to take

determining un ertainty redu tion is also related to the eld of

a tions in an environment so as to maximize some notion of long-term reward


(Sutton and Barto, 1998), fo used on on-line implementation while making a
tradeo between
knowledge.

exploration

of un harted territory and

exploitation

bat h and semi-bat h hemi al rea tors (Martinez, 2000).

design

of urrent

Reinfor ement learning has seen some appli ation in modeling of

Optimal experiment

for ontrol has fo used on deriving an input signal that minimizes some

ontrol-oriented measure of plant/model mismat h under a onstraint on total


input power.

Optimal experiment design is usually performed with the aim

of a hieving ontrol that is robust to disturban es (Gevers and Ljung, 1986).


Yip and Marlin (2003) suggested in luding ex itation planning in real-time optimization and to weight the ost of the ex itation against benet, under the
assumption that the model parameters are initially known and that the o urren e of a disturban e whi h may ne essitate ex itation to re-t parameters an
be identied by measurements.

1.2 Problem formulation


The aim of this paper is to suggest and study a stru tured approa h to optimizing the oshore produ tion of oil and gas with un ertain produ tion models. An
iterative two-step approa h to optimization ombined with post-optimization
feasibility assuran e through an operational strategy is suggested. Bootstrapping methods are used to quantify parameter un ertainty, whi h is exploited for
ex itation planning and result analysis under un ertainty based on multivariate
Monte-Carlo-like methods.

The suggested approa h is outlined in Se tion 2, a simulation ase-study is


des ribed in Se tion 3 before on lusions are drawn.

2 Produ tion optimization under un ertainty


2.1 Casting un ertainty in produ tion optimization in mathemati al terms
This se tion suggest how un ertainty resulting from low information ontent
in produ tion data an be ast in mathemati al terms to allow a stru tured
treatment.

Let

be a ve tor of internal variables and let

be a ve tor of

de ision variables. This paper will onsider a produ tion optimization problem
on the form


u()


x
() = arg max M (x, u, d)
s.t

where

(1)

u,x

0 = f (x, u, d, )
0 c(x, u, d),

(2)
(3)

u
() and x
() are the
. d is a ve tor of modeled and measured
u. M (x, u, d) is a prot measure whi h is to be

is a ve tor of parameters to be determined.

optimal solution of (1)-(3) for a given


disturban es independent of

maximized subje t to a pro ess model (2) and pro ess onstraints (3) for a
given parameter value

For oshore oil and gas produ tion,

produ tion rates of ea h uid from ea h well,

may be the

may be relative valve openings,

may des ribe onstraints in total water and gas pro essing apa ity and

may express total oil produ tion.


In this paper we will onsider response surfa e or performan e urve
type produ tion models (2), stationary, nonlinear, lo ally valid equations whi h
express produ tion rates of ea h well in terms of gas-lift rates, produ tion hoke
openings, the most re ent well test and tted parameters.
As produ tion data are often lo al in nature, i.e. setpoints are only varied
within a narrow range of values, a model tted to data may only be

lo ally valid,

only a urately able to des ribe produ tion for a narrow range of setpoints. The
lo al nature of data and of any model tted to su h data an be a ounted for
by enfor ing

max{umin , u0 Uprc U } u min{umax , u0 + Uprc U }

(4)

when solving (1)(3) and iteratively re-updating the produ tion model and reoptimizing, an two-step approa h.

U denes the s ale elements of u, u0 is the


Uprc < 1 is a design parameter whi h limits
at ea h step, and umin , umax are minimum

initial value of the de ision variable,


the magnitude of setpoint hange
and maximum values of

u,

respe tively.

instantaneous optimization, determining (


u(), x())
y may
for instan e be measured total rates of oil, gas and water. Let y
(u, d, ) be an
estimate of pro ess measurements based on the model (2). Let the tuning data
set be a set of histori al pro ess data spanning N time steps spanning the time
This paper only onsiders

at the urrent time. Let

be the ve tor of produ tion measurements,

interval

t [1, N ]
Z

y(1)] d(1)
u
(1) y(2)

d(2)
u
(2) . . .


y(N ) d(N ) u(N )

(5)

with residuals

(t, ) = y(t) y(u(t), d(t), )


Parameter estimation determines

t {1, . . . , N } .

(6)

by minimizing the sum of the squared resid-

uals for the tuning data set:

= arg min

where

w(t)

N
X

w(t)k(t, )k22 + Vs (),

s.t.

is a user-spe ied weighting fun tion,

onstraint and

c ()

c () 0

(7)

t=1

an optional onstraint on

Vs ()

is an optional soft-

The pro ess model (2) should

only be tted to histori al pro ess data that are

re ent,

in the sense produ -

tion during the time interval spanned by the tuning set should be onsistent
with urrent produ tion and the produ tion model.

It may be impossible to

with (7) when the information ontent in Z N


Extending the model stru ture to better des ribe more long-term ee ts

determine an a urate estimate


is low.

may enable the use of a longer tuning set, whi h may in turn redu e parameter un ertainty. Therefore parameter un ertainty is linked to other forms of
un ertainty su h as stru tural un ertainty.
Let the

potential for produ tion optimization Po

be the in rease in prot at-

tainable if the produ tion was moved from the initial operating point

to the globally optimal operating point (x (d0 ), u (d0 ), d0 ):

(x0 , u0 , d0 )

Po = M (x (d0 ), u (d0 ), d0 ) M (x0 , u0 , d0 ) 0.


When implementing

u
()

Some of the potential

Po

(8)

may be remain unrealized

due to un ertainty, whi h motivates


potential of produ tion optimization(Po )
realized potential (Po,r )

=
+

(9)

lost potential due to un ertainty (Lu ).

reBy repeatedly solving (7) with bootstrapping methods, un ertainty in


N
sulting from low information ontent in Z
an be quantied. In the inuen e
of un ertainty in

on

setpoints

Carlo simulations of (1)(3).

Elgsaeter et al. (2008a).

u
()

and prots

an be assessed by Monte-

This idea was explored for estimation of

u
L

in

2.2 Motivation for the hosen stru tured approa h to optimization with un ertain models
The strategy onsidered in this paper, motivated by the above dis ussion, is
outlined in Figure 4.

This paper onsiders the operational strategy expli itly as a omponent of


a stru tured approa h to handling un ertainty, as su h strategies are expe ted
to add robustness against some of the un ertainty onsidered.

Operational

strategies an be onsidered a form of un ertainty handling through feedba k.


This paper onsiders planning of ex itation in a single de ision variable,
whi h will redu e only some of the present parameter un ertainty.

The ost

of su h planned ex itation is asso iated with its resulting temporary redu tion
in produ tion and an be estimated in a fairly straightforward manner.

The

benet of a planned ex itation is mu h harder to quantify in advan e, as it will


depend on the parameter value found, on the inuen e of this updated parameter value on the target setpoint found through produ tion optimization, and
on subsequent implementation of the operational strategy while un ertainty in
other parameters still persist.

planned ex itation,

This paper onsiders estimating the

benet of

the marginal in rease in prots that an be expe ted when

re-optimizing after the planned ex itation, and this benet is estimated through
sto hasti simulations. Not all redu tion in parameter un ertainty will be profitable, for instan e it should be intuitively lear that eliminating un ertainty
against whi h the operational strategy is robust has zero benet.
As the hange in prots that result from implementing setpoint hange toward a target is un ertain and an even be negative, this paper onsiders basing
the hoi e of whether to implement setpoint hange on the distribution of simulated prot hanges found through sto hasti simulations of the operational
strategy.
The hoi e of relating de isions to prots both in result analysis and ex itation planning is motivated by a desire to make the ost of un ertainty visible in
business terms as this is what ultimately drives industry de ision making.
Sto hasti simulations were hosen as the onsidered method for their on eptual simpli ity, as they an be applied to a wide variety of operational strategies without introdu ing simplifying assumptions, and as the frequen y of de ision making in produ tion optimization on the order of days or weeks allows
omputationally intensive methods su h as these to remain pra ti al.
A formalization of an operational strategy is dis ussed in detail in Se tion
2.3, result analysis in Se tion 2.4, and ex itation planning in Se tion 2.5.

2.3 Operational strategy


Example 1 illustrates how an operational strategy may operate.

Example 1 Consider
a hypotheti al oshore oil and gas eld, produ ing from


T

, one de ision variable ae ting the produ tion of




1
2 T
qo2 qw
ea h well. Ea h well produ es oil and water at rates x = qo1 qw
whi h depend nonlinearly on u, and the obje tive is to maximize total oil produ tot
1
2
tion qotot = qo1 + qo2 while the total rate of produ ed water qw
= qw
+ qw
must not
ex eed a onstraint. Suppose that a target setpoint has been determined by solving the produ tion optimization problem. Suppose that at the urrent setpoint,
feasible and infeasible regions are as illustrated in Figure 1. The boundary between the feasible and infeasible regions is the set of setpoints at whi h all water
pro essing apa ity is utilized, and parameter un ertainty may ause the position of this boundary and the optimal setpoint to be un ertain. To ensure that
the implemented setpoint remains in the feasible region, an operational strategy

two wells with u = u1

u2

ould alternately de rease u1 and in rease u2 in smaller steps as illustrated in


Figure 1.
In this se tion the on epts illustrated in Example 1 are generalized and an operational strategy is des ribed formally for the purposes of analysis. In pra ti e
a human operator would perhaps not be able to stringently reprodu e the operational strategy as stated in this paper, instead the algorithm ould be used
more as a guideline.
get

The operational strategy attempts to implement hanges in

uT

and monitoring

by sequentially in reasing and de reasing elements of

responses in prots and onstraints.

uT

toward a tar-

may be any desired value of

u,

for in-

stan e that found by solving the produ tion optimization problem numeri ally.
The operational strategy is intended as a postoptimization strategy for the
implementation of the new desired setpoint, and no produ tion optimization or
parameter estimation is performed during the ourse of the operational strategy.
It is assumed that the sign of the response in prots and onstraint utilization
to a hange is setpoints is known. This will usually be the ase in oshore oil
and gas produ tion, where de ision variables may for instan e be hoke openings or gas-lift rates and the response to an in rease in u is usually positive for

be the set of indi es orresponding


the range of setpoints onsidered. Let U
+
the omponents of uT u0 whi h are negative, and let U
be the indi es of
the omponents of orresponding the omponents of

uT

uT u0

whi h are positive.

does not have to be feasible in pra ti e, the operational strategy will ensure

feasibility of implemented setpoints.


The operational strategy should have a dened riterion for when to terminate. If the termination riterion onsiders only the sign of prot hange, the
operational strategy may not be able to pro eed if the initial setpoint is lose to
a lo al optimum. However, if the termination riterion onsiders the dieren e
between monitored and predi ted prot, the operational strategy an pro eed
in the fa e of prot de rease, as long as this de rease is in reasonable agreement
with the model used in optimization. If prots de rease below the predi ted values then the de ision to terminate should weigh the magnitude of this de rease
against the predi ted prot in rease at the target. The rst s enario illustrated
in Figure 2 shows how a prot de rease

is

predi ted by the model and it is

reasonable for the operational strategy to ontinue, while in the se ond s enario
the prot de rease

is not

predi ted by the model and it is reasonable for the

operational strategy to terminate and return to the initial setpoint. These onsiderations an be formulated in terms of estimated and measured prot hange
T
JM against a threshold value JM
.
o
An operating strategy whi h is initiated at time ts and terminated at time
toe , in reases the realized potential by:
and evaluated by omparing a ost fun tion

Po,r = Po,r (tos ) Po,r (toe ),

(10)

and in the se tion below it is suggested how a probability density fun tion

fP r (Po,r )

an be found prior to implementing the operational strategy. The

operational strategy an ontinue altering setpoints beyond


step of the operational strategy auses in reased prot.

uT

as long as ea h

As the operational

strategy may overshoot the most protable setpoint, the nal step should be to
assess all steps and return to the most protable one. The suggested operational
strategy is outlined in Algorithm 1.

10

Algorithm 1 (An operational strategy) Given a pro ess that is initially at


a feasible setpoint (u0 , d0 ) and given measured pro ess onstraints c, measured
and de ision variables u
prot M
, a target value uT and probability distributions
fP r (Po,r ), let 0 < S 1 denote step size.

1. Let the measured prot prior to implementing the operational strategy be


0 . Let k be the index of the urrent step, and set k = 0 initially.
M
2. Determine sets (U , U + ) to orrespond with uT u0 .
3. Repeat (3a),(3b),(3 )
(a) k = k+1
(b) de rease ui , i U by an amount proportional to uiT ui0 , i U ,
so that uik = uik1 S (uiT ui0 ), i U ,
( ) in rease ui , i U + by an amount proportional to uiT ui0 , i U +
, as long as all elements of c obey c < 0 and
while observing c and M

u > 0, let the resulting measured prot be Mk , and the resulting


setpoint uk ,
k M
k1 ) or ( kuk u0 k < kuT u0 k
while (kuk u0k > kuT u0 k and M
T

and JM (Mk , fP (Po,r )) > JM ).

4. Implement the setpoint whi h resulted in the highest measured prot in


steps 1-3.
One of the operators' tasks is to ensure that the operational onstraints are
obeyed, therefore it is reasonable to assume that

(u0 , d0 )

rithm 1. It is assumed that in pra ti e ases of infeasible

is feasible in Algo-

(u0 , d0 )

are handled

operationally, i.e. an operator will apply pro ess knowledge instead of mathemati al optimization to move the pro ess to a feasible setpoint.

S is a user preferen e and preferably S << 1. The magniS will also depend on how often produ tion optimization is implemented.
Smaller S may result in an operational strategy whi h is more labor-intensive to
The magnitude of

tude of

implement, but may in some ases ause the operational strategy to terminate
loser to optimum.

2.4 Result analysis


This se tion aims to investigate whether an estimate of the parameter un ertainty an be exploited in result analysis. The approa h hosen is simulating the
prot hange

Po,r

that results from implementing a setpoint hange suggested

by produ tion optimization on the produ tion model with dierent parameter
estimates drawn from the estimated parameter distribution

f ()

in a Monte-

Carlo manner. The approa h is outlined in Algorithm 2

Algorithm 2 (Estimating fP r (Po,r )) Given an operational strategy, the

and
urrent operating point (x0 , u0 , d0 ), Uprc and probability distributions f ()
fu (
u),
repeat Nt times
,
 draw a sample t from f ()
11

 obtain a point estimate of Po,r by simulating the operational strategy


running on a pro ess des ribed by the pro ess model with negligible
stru tural model un ertainty and with parameters equal to t , while
enfor ing the onstraint (4).

The distribution of Nt point estimates Po,r is an estimate of


fP r (Po,r ).
As a opy of the produ tion model annot a ount for stru tural un ertainty

uu0, a onstraint of the form (4) should


Po,r will depend on
onservative hoi e of Uprc should yield onservative

whi h will grow with the magnitude of

be enfor ed in simulations. The magnitude of estimates of

Uprc , but a
Po,r on whi h result

the hoi e of
estimates

analysis ould be based.

The aim of the result analysis is to determine whether to implement the


setpoint hange suggested by produ tion optimization. This de ision ould be
based on the evaluation of a ost fun tion whi h depends on

fP r (Po,r ):

Jr (fP r (Po,r )) JrT .

(11)

Jr (fP r (Po,r )) ould for instan e be the expe ted or worst- ase value. The
T
threshold Jr should ensure that setpoint hanges are only performed when the
resulting in rease in prots is expe ted to be signi ant and to justify the risks
and operational osts whi h are asso iated with any setpoint hange, su h as
the risk of triggering an unplanned shutdown.

2.5 A ost/benet approa h to ex itation planning


The value of implementing a planned ex itation prior to implementing produ tion optimization is that the ex itation may redu e model un ertainty, allowing
produ tion optimization and the operational strategy to nd a more protable
setpoint than otherwise possible. This paper investigates exploiting estimates
of parameter un ertainty

f ()

found from re ent histori al produ tion data to

form the basis of a stru tured approa h to rank the benet of ex iting dierent
de ision variables. For simpli ity this dis ussion is restri ted to wells whi h are

de oupled, i.e.

where implementing hange in the produ tion of one well will not

inuen e the produ tion of other wells, whi h is often a reasonable assumption
for so- alled platform wells as produ tion from su h wells travel through separate
owlines and join rst at produ tion manifold. The dis ussion is restri ted to
ex itation of single de ision variables, although simultaneously ex iting several
variables is also on eivable.
The benet of ex itation

BE

is proposed dened as the dieren e between

the prot whi h would be realized if no ex itation is performed,

Po,r , and
E
Po,r
:

the

prot whi h would be realized after implementing an ex itation,

E
B E = Po,r
Po,r .

(12)

The idea of the benet of ex itation is illustrated in Example 2.

Example 2 Consider a hypotheti al eld produ ing oil and water from two

1
) at rates whi h depend on u1 , well
de oupled wells. Well 1 produ es (qo1 , qw
2 2
2 produ es (qo , qw ) at rates whi h depend on u2 . The relationships between

12

1
2
(qo1 , qw
) and u1 and between (qo2 , qw
) and u2 are un ertain, but a model whi h ex1
2
presses qo1 (u1 , 1 ), qo2 (u2 , 2 ), qw
(u1 , 1 ), qw
(u2 , 2 ) is available. Models for well
1
2
1 depend on parameters

and
models
for


 well 2 on . Two parameter es
timates a = a1 a2 and b = b1 b2 have been determined whi h result
tot
1
= qw
+ qo2 whi h mat h measurein estimates of qotot = qo1 + qo2 and qw
ments equally well for a set of re ent histori al data. Optimization of the
obje tive M = qo1 (u1 , 1 ) + qo2 (u2 , 2 ) subje t to the model and the onstraint
1
2
max
qw
(u1 , 1 ) + qw
(u2 , 2 ) < qw
is desired. If no ex itation is implemented the

(a ) is planned implemented using an operational strategy. If an


target uT = u
ex itation ould be performed whi h would reveal whi h of a and b better des ribe produ tion, what would be the benet of that ex itation?
E
= Po,r . If b des ribes
If a des ribes produ tion better B E = 0 as Po,r
produ tion better B E will depend on how mu h prot in rease the operational
strategy would be able to implement if given the target uT = u
(b ) as opposed to

uT = u
(a ). In the ase that b des ribes produ tion better while uT = u
(a ),
E
B ould be found through simulations on the produ tion models, provided that
stru tural model errors are small.

B E will depend on the intera tion between un er and the ability of the operational strategy to
un ertainty in u
()
. The benet of ex itation will also depend
()
for un ertainty in u

Example 2 illustrates that


tainty in

ompensate

on what target the operational strategy will pursue if no ex itation is performed.


In the oshore produ tion of oil and gas, ex iting all omponents of
ally not feasible, so some method for ranking omponents of

is usu-

to be ex ited is

desirable, as is illustrated in Example 3.

Example 3 Consider again the eld des ribed in Example 2. Consider that
our aim is to hoose whether to introdu e ex itation in u1 , in u2 or not at
all. Assume that an ex itation of u1 would allow us to distinguish between a1
and b1 , while 2 would still be un ertain, and that an ex itation of u2 would
allow us to distinguish between a2 and b2 while 1 would still be un ertain.
With the assumptions made it an be expe ted that an ex itation of well 1 will
([a1 , a2 ]) if produ tion is des ribed by a , or
either result in implementing uT = u
([b1 , a2 ]) if produ tion is des ribed by b , as it was planned
implementing uT = u
to implement uT = u
([a1 , a2 ]) if no ex itation was performed, and 2 would still
be un ertain after ex itation of u1 . The out ome on the produ tion models an
be simulated in both ases provided that stru tural model errors are small, whi h
E . A stru tured approa h to planning ex itation
would give two estimates of B
E.
ould then be to ex ite the well asso iated with the highest positive average B
Examples 2 and 3 illustrate the prin iples of a Monte-Carlo simulation approa h
E
to estimating B
for the simplest possible ase of two wells, two parameter estimates and one onstraint. In our ase study the same prin iples are applied to
E for a larger number of wells, onstraints, and parameter estimates
estimate B

Let

nominal parameter estimate, the parameter estimate found


B E will onsider the ase when the
(n ) if no ex istrategy attempts to implement the target uT = u

be the

by solving (7) dire tly. The estimation of


operational

tation is performed. Note that although produ tion of wells are de oupled, the
omponents of

u
()

as al ulated by (1)(3) are oupled, so updating models

13

des ribing one well may inuen e suggested settings for other wells, and this
E . Simulations on the produ tion
will need to be ree ted when estimating B
model assume that stru tural model un ertainty is negligible, and, as lo ally
valid models are onsidered, this assumption is reasonable as long as hange
in de ision variables is small, whi h motivates enfor ing the onstraint
E .
B

(4)

in

simulations, and whi h should result in onservative estimates


The approa h onsidered is outlined in Algorithm 3.

Algorithm 3 (Benet of ex itation, ui ) Given that it is planned to imple-

, an
(n ) if no ex itation is performed. Given Nt , the distribution f ()
ment u
operational strategy, a pro ess model whi h depends on parameters , Uprc , and
that the aim is to estimate the benet of ex iting ui , a single de ision variable
in the ve tor u.
repeat Nt times
.
 draw a sample t from f ()
 determine Po,r (t ) by simulating the operational strategy using uT =
u(n ) on the pro ess model with parameters t while setpoint hange
is limited by (4).
 let ti be parameters of t ex ited by ui . Let E (n , ti ) be a ve tor of
parameters that is equal to ti for parameters ex ited by ui and equal
to n for all other parameters.

E,i
(t ) by simulating implementing the operational strat determine Po,r

egy using uT = u
(E (n , ti )) on the pro ess model with parameters
t while setpoint hange is limited by (4).
E,i
(t ) Po,r (t ).
 let B E,i (t ) = Po,r
E,i is an estimate of the probability
the distribution of simulated values B
E,i )
density fun tion for the benet of ex iting ui , fBE (B
A stru tured approa h to ex itation planning ould be to ex ite those de ision
E,i on average.
ui whi h have highest B

variables

u
()
u
( ( , ti ))

as al ulated by (1)(3) are oupled, ex itation of


u
(n ) in omponents other than ui .
(E (n , ti ))
If un ertainty for non-ex ited wells is signi ant, implementing u
n

may inadvertently ause lower prots than u


( ), and in su h ases estimates
E,i may be negative. Negative B
E,i is an indi ation that the prot in rease
B
i
resulting from ex itation of u is variant to un ertainty in models des ribing

ui

As omponents of
E n

an ause

to dier from

other wells.
Ex ept during startup it will in most ases be reasonable to assume that
produ tion is initially utilizing at least one pro essing apa ity fully, so that
any ex itation will require produ tion to ba k o from full apa ity for a period
E
of time, in urring a ost. It is on eivable to simulate the ost of ex itation C

. To limit the s ope of this


in a Monte-Carlo fashion for dierent samples of f ()
E
paper, estimation of C
will not be onsidered. It is proposed that andidate
ex itations an be ranked by the value of a fun tion

E ), f E (C E ) J E,T ,
J E (fBE (B
C
14

(13)

where

J E,T

is a user-spe ied threshold and

E ), f E (C E )
J E (fBE (B
C

is a userE
spe ied metri su h as for instan e the expe ted value of the dieren e B
E

C .

3 Case study
This se tion des ribes the appli ation of the suggested approa h to a ase-study
modeled on the North Sea oil and gas eld and the set of real-world produ tion

data onsidered previously in Elgsaeter et al. (2007), Elgsaeter et al. (2008b)


and Elgsaeter et al. (2008a).

3.1 Field des ription


The ase study onsidered in this paper is motivated by a North Sea oil and
gas eld with 20 gas-lifted platform wells produ ing predominantly oil, gas and
water.

The eld has a layout as depi ted in Figure 3, with one produ tion

separation train and one test separator.

Measurements of the total rates of

produ ed oil, gas and water are available. The operator of the eld requested
that all data be kept anonymous, therefore all variables will be presented in
normalized form.

The produ tion data are hara terized by little variation

in de ision variables.

The aim of produ tion optimization on the eld is to

distribute available lift gas so as to maximize total oil produ tion while keeping
total produ ed water and gas rates below apa ity onstraints.

3.2 Method
The ost and risks of implementing a trial of the suggested approa hes on an
a tual eld are signi ant, and it may be di ult to ompare strategies by
implementation on an a tual eld as produ tion will vary with time due to
disturban es. This motivates the hoi e of studying the suggested approa h in
simulations.
Due to un ertainty there may be many plausible des riptions of produ tion,
and an approa h to optimization under un ertainty should ideally perform well
for all su h plausible des riptions. This paper will onsider simulating optimization on a model that is plausible in the sense that it onforms with produ tion
data and refer to this model as the

produ tion analog.

When optimizing pro-

du tion, the produ tion analog is onsidered unknown, produ tion optimization
an only infer knowledge of the produ tion analog through measurements.
When modeling the response to hange in gas-lift rate, a rst-order linear
kernel with a single tted parameter for ea h uid and phase with

dim() = 63

(formulation A) will be ompared with a se ond-order nonlinear kernel with two


tted parameters for ea h uid and phase and

dim() = 123

(formulation B).

By omparing these formulations the signi an e of model stru ture and parameterization on the suggested methodology an be assessed. Physi al knowledge
is in luded through onstraints and regularization on parameter estimation.
Both for formulations A and B parameters were re-tted and produ tion
re-optimized a ording to the workow outlined in Figure 4 in four iterations.
Algorithms 2 and 3 were run for ea h iteration with

Nt = 200,

but no ex i-

tations were implemented and setpoint hange was implemented regardless of

15

risk-reward estimates. For omparison this pro ess was repeated for two hoi es
of the design parameter

Uprc = 0.2

and

Uprc = 0.5.

Further details on the simulation ase study are given in Appendix A.

3.3 Results
A omparison of the predi tions of the produ tion analog against eld data
is shown in Figure 5.

The produ tion analog is shown along with models of

formulation A and B at iteration 1 in Figures 6 and 7.

The distribution of

found through Monte-Carlo simulation prior to iteration 1 are ompared for

the two models in Figure 8.


that between

30%

and

80%

Po,r of
Uprc = 0.2 and Uprc = 0.5, indi ating

Figure 9 ompares the realized prots

formulations A and B at ea h iteration, for

of the prot potential was realized.

Figure 10 illustrates the hanges in setpoint implemented by the operational


strategy for formulation B. Figures 11 and 12 ompare the simulated benet of
ex itation with estimates found with Algorithm 3 for iterations 1 and 4. Figure
13 ompares

fP r (Po,r )

found with Algorithm 2 with realized prot for ea h

iteration of formulation A and B. Figure 14 ompares target setpoints before


and after ex itation for a parti ular strategy, well and iteration.

3.4 Dis ussion


The suggested optimization method an be lassied as a two-step approa h
and the operational strategy as postoptimization feasibility assuran e. In the
simulation ase onsidered, model un ertainty was signi ant, as illustrated in
Figures 6 and 7, and as a result un ertainty in the setpoint

u
()

suggested

by produ tion optimization was highly un ertain, as illustrated in Figure 8.


Despite this signi ant un ertainty, the suggested approa h was able to realize
a signi ant prot in rease with either model formulation and with either hoi e
of

Uprc ,

as is illustrated in Figure 9.

(1)(3) is a nonlinear optimization problem and onvergen e of its solution


toward the global optimum annot be guaranteed in general unless (1)(3) an
be shown to be a onvex optimization problem and is solved with methods
of onvex optimization (Boyd and Vandenberghe, 2004).

Analyti al methods

(Ljung, 1999) an only guarantee that the solution of (7) is the best des ription
of produ tion if

approa hes innity, if the reservoirs, wells and pro essing


Z N is su iently informative to

fa ilities an be onsidered stationary, and if

distinguish between all solutions of (7) . Typi ally bootstrap estimates are not
exa t but have an inherent error, while the bias in estimates is often small the
varian e an often be quite large due to the nite amount of data and the nite
number of resamples (Efron and Tibshirani, 1993).
For the reasons listed above, no guarantees an be given that the implemented setpoint will onverge toward global optimum in pra ti e when the
suggested approa h is applied to pro esses su h as these with low information ontent data and limits on the planned ex itation. We argue that la k of
guaranteed onvergen e is a result of the properties of the pro ess onsidered
and not of the proposed solution, and pro esses with these properties are nevertheless interesting for a ademi study. It is be ause of the la k of guaranteed
onvergen e that this paper has fo used on identifying setpoint hange whi h
in rease prots with some measure of onden e.

16

The methods suggested would remain sensible in the spe ial ase that all the
requirements stated above are satised and a unique parameter estimate

an

be found from (7) and (1)(3) is able to return the globally optimum setpoint.
In this ase, bootstrapping (7) would return the same parameter estimate for
all resamples, all estimates of the benet of ex itation found with Algorithm 3
would be zero, and and Algorithm 2 would predi t a single prot hange

Po,r

with high onden e.


The operational strategy suggested will always ensure that produ tion is
feasible as long as produ tion is initially at a feasible setpoint, onstraints are
measured, transients are negligible and as long as the sign of hange in onstraint
utilization to hange in de ision variables are known. Constraints were enfor ed
at all times during the simulation ase study, an example is shown in Figure
10. The assumption of negligible transients may be relaxed by expanding the
operational strategy method, for instan e by model-predi tive ontrol.
Ex itation planning and result analysis suggested employ omputational,
Monte-Carlo-like methods.

These methods an be onsidered a form of non-

parametri multivariate analysis, as a sample of the entire parameter ve tor

is drawn at ea h iteration, and o-varian e between parameters is therefore

a ounted for in the analysis. The fa t that no assumptions about linearity of


models or types of probabilities are made are a strength of the methods, and
this trait alone separates these methods from the majority of rival approa hes
found in the literature.
The method for ex itation planning suggested in this paper deals with identifying the de ision variable whi h when ex ited will in rease the prot attained
by produ tion optimization the most. This is done by analyzing the inuen e of
the modeled un ertainty on prot fun tion by sto hasti simulations. The suggested method dier from optimum experiment design in that the obje t is not
to a hieve ontrol that is robust to disturban e, but rather to in rease prots by
as mu h as possible by targeting whi h de ision variable to ex ite. In this ase
the a tual benet of ex itation ould be omputed, as shown in Figures 11 and
E and large a tual B E is visible.
B

12, and a orrelation between large mean

This nding of the simulation ase study supports the assertion that testing
E is a sensible ex itation planning approa h, provided
wells with large mean B
estimated benets are signi ant ompared to estimated osts.

Further work

ould onsider estimating the benet of simultaneously ex iting a small number of de ision variables simultaneously, as su h ex itation ould be designed to
ause less temporary redu tion in produ tion and hen e in ur lower osts.
The suggested method for result analysis uses sto hasti simulations to determine a distribution for the hange in prot that will result from hanging
de ision variables toward a target by means of the operational strategy. While
methods have been found in the literature whi h assess the signi an e of estimated prot hange to noise, no referen e was found to result analysis in the
presen e of un ertainty in all tted parameters, either with or without postoptimization feasibility assuran e. Figure 13 illustrates that for Formulation B,
the potentially more a urate model with a higher number of parameters, estimates of realized potential

Po

Po

are omparable to the a tual realized potential

and estimates de line with ea h iteration just as do a tual

iteration estimates

Po

are lower than the a tual

Po

Po .

For the rst

due to the onservativeness

in the design of Algorithm 2, but the algorithm orre tly identies a signi ant above-zero potential, whi h would have supported the de ision to update

17

setpoints toward the suggested target.


Mu h further work on design of alternative algorithms for quantifying un ertainty, ex itation planning, result analysis and operational strategies whi h
t the framework suggested here is possible. An element of this paper has not
onsidered in detail is how a tive de ision variables an be hosen to manage
un ertainty, this is left for further work.

4 Con lusion
An approa h for asting in mathemati al terms the un ertainty in produ tion
optimization arising from low information ontent in data has been suggested.
A stru tured approa h to handling this un ertainty by ombining an iterative
two-step approa h to optimization and post-optimization feasibility assuran e
was suggested and by un ertainty estimation, result analysis and ex itation
planning based on multivariate Monte-Carlo-like methods.
In the simulation ase study the method suggested realized between

80%

30% and

of the available prot potential while feasibility was ensured at all times,

despite that models were tted to data with low information ontent similar to
that found on real-world oil elds.

A knowledgments
The authors wish to a knowledge the Norwegian Resear h Coun il, StatoilHydro
and ABB for funding this resear h.

A Details of simulation ase study


To ensure that the produ tion analog is a plausible des ription of the eld
onsidered, parameters in (15) are estimated with bootstrapping methods. (15)
ombined with one of the plausible parameters determined in this manner is
hosen as the the produ tion analog for setpoint values similar to those observed
in the tuning set.
max,i
Let qgl
be the maximum values observed in the tuning set. To simulate
our la k of knowledge about pro ess behavior for values of u outside those
observed in the tuning set, the kernel fun tion

i
fui (qgl
)=

fgl

is repla ed in (15) with

1 i i
i,max 2
i,max
i
c (qgl qgl
) + bi (qgl
qgl
) + ai
2

(14)

max,i
i
i
i
i
qgl
> qgl
. a and b are hosen so that q (u) is smooth and ontinuous
i
for u around umax , and the urvature oe ient c is hosen as a random negal,i,a
of the produ tion analog are hosen so that the
tive value. Operating points q
when

tuning set an be des ribed by the produ tion analog with bias terms. Figure 5
illustrates that the produ tion analog is a plausible des ription of produ tion,
as it mat hes produ tion data well.
Produ tion modeling of the eld onsidered based on the on epts of system

identi ation has been onsidered previously in (Elgsaeter et al., 2008b), and
i
i
i
similar methods are applied here. The oil, gas and water rates qo , qg , qw of ea h

18

i are modeled as the produ t of two kernel fun tions, one des ribing the
i
ee ts of hanges in produ tion valve opening z , and one des ribing the ee ts
i
of hanges in gas lift rates qgl . Models are lo al around the most re ent well
test.
well

The produ tion model

l,i
i
i
qpi = max{0, qpl,i fzi (z i , z l,i ) fgl
(qgl
, qgl
)},

(15)

is onsidered, whi h is intended to be valid lo ally around the most re ent well
l,i
i = 1, . . . , nw p {o, g, w}. z i [0, 1] is the relative valve
test rates qp
i
i
i qgl
opening of well i, qgl is the gas-lift rate of well i and qgl = l,i 1 is the
q
gl

normalized, relative gas lift rate for well

fzi (z i , z l,i ) =

i.

Kernels are hosen as

1 (1 z i )k
,k = 5
1 (1 z l,i )k

(16)


i
i
i 2
fgl
(z i , z l,i ) = 1 + ip qgl
+ ip (qgl
)
for all wells

i.

y(t) = qotot (t) qgtot (t)
so that estimates

The measurement ve tor


attempt is made to nd

tot
qw
(t)

T

(17)

is onsidered and an

y(, t) = 1
x(, t) + y ,
t measurements as lose as possible for the tuning set, where

(18)

is the ve tor

of measurement biases due to alibration ina ura ies to be determined and

is a matrix of ones.
In addition an upper and lower onstraint on

u of the form (4) was enfor ed.

Some wells are in danger of slugging if gas-lift is de reased, and on these wells
a onstraint whi h prohibits gas-lift from being de reased was implemented.
Based on the knowledge that the water ut is rate-independent, soft onstraints whi h penalize deviation from o = w and o = w were added to
otot was visible in the tuning set, and it
the obje tive fun tion. A de line in q
tot
is hosen to de-trend q
o and weigh older measurements less than newer ones
using the weighting-term

w(t) in (7). In formulation B, and for all wells and


y are onsidered part of the ve tor of tted

for oil, gas and water, as well as

parameters. Only the rst-order term in (17) is onsidered in formulation A.


The produ tion optimization problem onsidered was of the form

u
() = arg max
u

(19)

s.t.

(20)

iIa

l,i
qgl

qoi (u, d, ) + bo ()

i
qgl
i IMGS

max{umin , u0 Uprc U } u min{umax , u0 + Uprc U }


X
i
qgi (u, d, )
qgl
+ bg () qgtot,c

iIa

(21)
(22)
(23)

iIB

i
tot,c
qw
(u, d, ) + bw () qw
.

iIa

19

(24)

Ia is the indi es
bo (), bg (), bw ()

of all wells onsidered in produ tion optimization.


were determined for ea h

produ tion mat h at the time of optimization. For wells with indi es

{1, 4, 10, 13, 14, 16} gas-lift ould not be


i
slugging. q
oi , qgi , qw
are modeled rates.

Biases

so that modeled and measured

IMGS =

de reased without the risk of triggering

The ase study was implemented in

MATLAB 1 .

Linear least-squares pa-

rameter estimation problems and linear real-time optimization problems were


solved using the

TOMLAB 2

solver

lssol.

Nonlinear produ tion optimization

problems were solved using the sequential quadrati programming solvers based
on S hittkowski (1983).

i
The simulated benet of ex itation of u was found by repla ing tted pai
rameters ex ited by variation of u with those that best des ribe the produ tion
analog, and simulating produ tion optimization and implementation with the
operational strategy.

Nomen lature
Subs ript

o,g

and

indi ate oil, gas and water, respe tively. Numbered super-

s ripts indi ate well indi es, the supers ript

indi ates a lo al operating point.

Bars (
.) indi ate measured variables, while hats (.) indi ate estimated variables.

u
x
d

de ision variables
internal variables
modeled and measured disturban es independent of

parameters to be determined through parameter estimation

M
c
y

prot measure
produ tion onstraints
measured variables exploited in parameter estimation

t
N
ZN

Dy
w
Vs
c
umin
umax
Uprc
U
x0 , u0 , d0
f
Po
Po,r

time
the number of time steps in the tuning data set
a tuning data set spanning

time steps

residuals in parameter estimation


normalization matrix for

weighting of residuals in parameter estimation


soft onstraints on parameters

hard onstraints on parameters

the lower end of the range of possible values for

u
u

the higher end of the range of possible values for

a limit on the size of hange in de ision variables


s aling of u

x, u, d

at the time of optimization

probability fun tion for parameters

the potential for produ tion optimization


realizable potential for produ tion optimization

1 The Mathworks,In ., version 7.0.4.365


2 TOMLAB Optimization In ., version 5.5

20

Lu
uT
U

lost potential of produ tion optimization


a target value for an operational strategy
the set of indi es of

whi h are to be de reased as


uT from u0

an operational strategy approa hes

U+

the set of indi es of

whi h are to be in reased as


uT from u0

an operational strategy approa hes

JM

a ost fun tion against whi h measured prot hange


during implementation of an operational strategy is
to be ompared to determine whether to terminate
the operational strategy before the target is rea hed

T
JM
o
ts
toe

a threshold value for

JM

the time at whi h the operational strategy is initiated


the time at whi h the operational strategy is terminated

fP r
S

probability fun tion for

fu
Jr

probability fun tion for

JrT
BE
CE
n

a threshold value for

Po,r

the user-spe ied size of setpoint hange at ea h iteration of an operational strategy

a ost fun tion for the expe ted prot attainable


from moving setpoints toward a new target

JM

the benet of ex itation


the ost of ex itation
the nominal parameter estimate, in the ontext of
benet of ex itation

JE
J E,T
qgl
z
Ia

the ost-benet tradeo of an ex itation


E
a threshold value for J
gas-lift rate
relative produ tion hoke opening
the set of indi es of all wells onsidered in optimization

b
q tot,c
IMGS
y
,

bias al ulated over interval

[N L, N ]

a apa ity for total produ ed rates


indi es of wells with lower onstraints on gas-lift rate
bias estimated over interval

[1, N ]

tted parameters in gas-lift models

Referen es
Andersen, H. W., Rasmussen, K. H. and Jorgensen, S. B. (1991), Advan es
in pro ess identi ation,

in

`Fourth International Conferen e on Chemi al

Pro ess Control', South Padre Island, Texas.


Be ker, R., Hall, S. and Rustem, B. (1994), `Robust optimal de isions with
sto hasti nonlinear e onomi systems',

and Control 18,

Journal of E onomi Dynami s

125147.

Bieker, H. P., Slupphaug, O. and Johansen, T. A. (2006), Optimal well-testing


strategy for produ tion optimization: A monte arlo simulation approa h,

in

`SPE Eastern Regional Meeting', Canton, Ohio. SPE 104535.

21

Box, G. and Draper, N. (1969),

pro ess improvement,

Evolutionary operation; A statisti al method for

Wiley, New York, NY.

Boyd, S. and Vandenberghe, L. (2004),

Convex Optimization,

Cambridge Uni-

versity Press, Cambridge, UK.


Costa, A. P. A., S hiozer, D. J. and Poletto, C. A. (2006), Use of un ertainty
analysis to improve produ tion history mat hing and the de ision-making
pro ess,

in

`SPE Europe /EAGE Annual Conferen e and Exhibition', Vi-

enna, Austria. SPE 99324.


Darlington, J., Pantelides, C. C., Rustem, B. and Tanyi, B. A. (1999), `An
algorithm for onstrained nonlinear optimization under un ertainty',

tomati a 35(2),

Au-

217228.

Efron, B. and Tibshirani, R. J. (1993),

An Introdu tion to the Bootstrap,

Chap-

man & Hall.


Elgsaeter, S. M., Slupphaug, O. and Johansen, T. A. (2007), Challenges in
parameter estimation of models for oshore oil and gas produ tion,

in

`2007 International Petroleum Te hnology Conferen e', Dubai,U.A.E.


Elgsaeter, S. M., Slupphaug, O. and Johansen, T. A. (2008a), Oil and gas
produ tion optimization; lost potential due to un ertainty,

in `IFAC World

Congress 2008', Seoul, Korea.


Elgsaeter, S. M., Slupphaug, O. and Johansen, T. A. (2008b), Produ tion optimization; system identi ation and un ertainty estimation,

in

`Pro . SPE

Intelligent Energy Conferen e 2008', Amsterdam, The Netherlands.


Fel'dbaum, A. A. (1961a), `Dual ontrol theory, parts i',

mote Control 21(9),

Automation and Re-

874880.

Fel'dbaum, A. A. (1961b), `Dual ontrol theory, parts ii',

mote Control 21(11),

Automation and Re-

10331039.

Forbes, J. F. and Marlin, T. E. (1994), `Model a ura y for e onomi optimizing


ontrollers: the bias update ase',

Ind.Eng.Chem.Res. 33,

19191929.

Franois, G., Srinivasan, B. and Bonvin, D. (2005), `Use of measurements for enfor ing the ne essary onditions of optimality in the presen e of onstraints
and un ertainty',

Journal of Pro ess Control 15(6),

701712.

Gevers, M. and Ljung, L. (1986), `Optimal experiment designs with respe t to


the intended model appli ation',

Automati a 22(5),

543554.

Griess, B. K., Diab, A. and S hulze-Riegert, R. (2006), Appli ation of global


optimization te hniques for model validation and predi tion s enarios of
a north afri an oil eld,

in

`SPE Europe /EAGE Annual Conferen e and

Exhibition', Vienna, Austria. SPE 100193.


Kall, P. and Walla e, S. W. (1994),

Sto hasti programming,

NY.

22

Wiley, New York,

Krsti , M. (2000), `Performan e improvement and limitations in extremum seeking ontrol',

Systems & Control Letters 39(5),

313326.

Little, A. J., Fin ham, A. and Jutila, H. A. (2006), History mat hing with
produ tion un ertainty eases transition into predi tion,

in

`SPE Eu-

rope /EAGE Annual Conferen e and Exhibition', Vienna, Austria. SPE


100206.
Ljung, L. (1999),

System identi ation: Theory for the user,

2nd edn, Prenti e

Hall, Cambridge.
Loeblein, C., Perkins, J. D., Srinivasan, B. and Bonvin, D. (1999), `E onomi
performan e analysis in the design of on-line bat h optimization systems',

Journal of Pro ess Control 9(1),

6178.

Martinez, E. C. (2000), `Bat h pro ess modeling for optimization using reinfor ement learning',

Computers and Chemi al Engineering 24(2),

1187

1193.
Metropolis, N. and Ulam, S. (1949), `The monte arlo method',

Ameri an Statisti al Asso iation 44(247),

Journal of the

335341.

Mileti , I. and Marlin, T. E. (1998), `On-line statisti al result analysis in realtime operations optimization',

Ind. Eng. Chem. Res. 37,

36703684.

Mulvey, J. M. and Vanderbei, R. J. (1995), `Robust optimization of large-s ale


systems',

Operations resear h 43(2),

264281.

Ogunnaike, B. A. (1995), `A ontemporary industrial perspe tive on pro ess


ontrol theory and pra ti e',

A. Rev. Cont 20,

18.

Pistikopoulos, E. and Mazzu hi, T. A. (1990), `A novel exibility analysis approa h for pro esses with sto hasti parameters',

14(9), 9911000.

Computers Chem. Engng.

Roberts, P. D. and Williams, T. W. (1981), `On an algorithm for ombined


system optimization and parameter estimation',

Automati a 17(1),

199

209.
Saputelli, L. A., Mo hizuki, S., Hut hins, L., Cramer, R., Anderson, M. B.,
Mueller, J. B., Es or ia, A., Harms, A. L., Sisk, C. D., Pennebaker, S., Han,
J. T., Brown, A., Kabir, C. S., Reese, R. D., Nuez, G. J., Landgren, K. M.,
M Kie, C. J. and Airlie, C. (2003), Promoting real-time optimization of
hydro arbon produ ing systems,

in `2003 SPE Oshore Europe Aberdeen',

Aberdeen. SPE 83978.


S hittkowski, K. (1983), `On the onvergen e of a sequential quadrati programming method with an augmented lagrangian line sear h fun tion',

Operations s h. u. Statist., Ser. Optim. 14,

Math.

197216.

Skogestad, S. (2000), `Self-optimizing ontrol: the missing link between steadystate optimization and ontrol',

Comput.Chem.Eng. 24(2),

Skogestad, S. and Postlethwaite, I. (1996),


Wiley & Sons, Chi hester, England.

23

569575.

Multivariable Feedba k Control, John

Sutton, R. S. and Barto, A. G. (1998),

Reinfor ement learning; an introdu tion,

MIT, Cambridge, Ma.


Terwies h, P., Agarwal, M. and Rippin, D. W. T. (1994), `Bat h unit optimization with imperfe t modeling - a survey',

4(4), 238258.

Journal of Pro ess Control

Yip, W. S. and Marlin, T. E. (2003), `Designing plant experiments for real-time


optimization systems',

Contr. Eng. Pra ti e 11,

837845.

Zhang, Y., Monder, D. and Forbes, J. (2002), `Real-time optimization under


parametri un ertainty:

Pro ess Control 12(3),

a probability onstrained approa h',

Journal of

373389.

Zhu, Y. (2006), `System identi ation for pro ess ontrol: Re ent experien e
and outlook'.

24

u2
te

Infeasible

ts

Feasible

u1

o
Figure 1: Example 1: Illustration of an operational strategy initiated at ts and
o
terminating at te . Solid line illustrates the unknown border between feasible
and infeasible regions, the dotted lines illustrate the un ertainty of this border,
while the dashed line illustrates how an operational strategy might move setpoints in losed-loop. The ir le illustrates the setpoint
un ertain al ulated optimal setpoints.

25

u0 , and rosses illustrate

^
M0+P
o,r

M(t)

Scenario 1

M0
t
tso

^
M0+Po,r

M(t)

ta

teo

Scenario 2

negative profit bias


t

M0

ta
tso
teo
increasing process model uncertainty
Figure 2:

The two plots illustrate how measured prot (solid lines) may

evolve ompared to a al ulated predi tion interval (dotted lines) in two dif-

M0 is the prot as the operational strategy is implemented, and


M0 + Po,r (dashed line) is the predi ted in rease in prot from implementing

ferent ases.

the operational strategy.

26

Lift gas/
gas injection
Gaslift Platform
wells

Separation

Subsea
Gaslift wells

Gas
Oil

Export

Water

....

....

Reservoir(s)

....

....

....
Gas
injection

Routing

....

....

Water
injection

Separation

Reinjection/To sea

Gaslift

....
Figure 3: A s hemati model of oshore oil and gas produ tion.

27

Wait until new data


becomes avialable

Update model:
Estimate parameters
and parameter
uncertainty

Perform production
optimization
Optionally: select
active decision
variables

Perform excitation
planning

Is the cost/benefit
tradeoff of any
planned excitation
favorable?

Yes Implement
planned
excitation

No
Perform result
analysis

No
Is result analysis
favorable?

Yes
Implement setpoint
change suggested
by production
optimization
according to
operational strategy

Figure 4:

Flow hart of the proposed stru tured approa h to optimization of

oshore oil and gas produ tion with un ertain models.

28

1.2

qotot

0.8

0.6
0

500

1000

1500

2000

2500

500

1000

1500

2000

2500

500

1000

1500

2000

2500

PSfrag repla ements

1.4

qgtot

1.2

0.8

tot
qw

0.8

t[hours

Figure 5:

Case study:

Produ tion data (dotted),

the produ tion analog

(dashed), estimates with nominal nonlinear produ tion model (solid).

29

qgi

qoi

i
qw

i=1
i=2
i=3
i=4
i=5
i=6
i=7
i=8
PSfrag repla ements
i=9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
i = 20
i
qgl
Figure 6: Case-study: Iteration

i
qgl

i
qgl

1, formulation A. Ea h row shows modeled rates

of oil, gas and water for dierent parameter estimates found through bootstrapping (dashed) ompared with modeled rates for the nominal parameter estimate
(dotted) and the produ tion analog (solid), global optimum ( rosses) and urrently implemented setpoint ( ir les).
30

qgi

qoi

i
qw

i=1
i=2
i=3
i=4
i=5
i=6
i=7
i=8
PSfrag repla ements
i=9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
i = 20
i
qgl
Figure 7: Case-study: Iteration

i
qgl

i
qgl

1, formulation B. Ea h row shows modeled rates

of oil, gas and water for dierent parameter estimates found through bootstrapping(dashed) ompared with modeled rates for the nominal parameter estimate
(dotted) and the produ tion analog (solid), global optimum ( rosses) and urrently implemented setpoint ( ir les).
31

Formulation B
Formulation A

i=1
i=2
i=3
i=4
i=5
i=6
i=7
i=8
PSfrag repla ements

i=9

i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
i = 20
i
qgl

i
qgl
Figure 8: Case-study: Distribution
ea h well after iteration

fu (
u)

for formulations A and B shown for

(bars). Constraints on

(stem), and optimal setpoint ( rosses).

32

(dotted),

u0

( ir le) and

uT

100

80

60

Po,r

Po,r

100%

40

20

PSfrag repla ements

t[iterations

Po,r
that was
(squares) for Uprc = 0.2

Figure 9: Case-study: The per entage of the initial potential


realized in simulations of formulation A ( irles) and B
(line) and

Uprc = 0.5

(dashed).

33

0
0.1
0.2

1.03

0.3
1.02
0.4
qotot
0.5 1.01
0.6
0.7

0.8
0.9
1 0.995
tot
qg 0.99

1
2

0.985

0.98

4
5

0
0.1

0.995
0.2
tot
qw
0.3 0.99

6
7

0.4

0.5 0.985
0.6

0.7

10

0.8
0.9

11

0.6

12
0.4

13
14

0.2

qgl
l
qgl

15
16

17
18

-0.2

19
-0.4

20

-0.6

-0.8
1

t
Figure 10:

Case study:

[iterations

An example of how the operational strategy imple-

ments setpoint hanges while obeying pro ess onstraints for formulation B.
tot
tot
Top graphs show normalized prots (qo ), gas apa ity utilization (qo ) and
tot
water apa ity utilization (qw ), lower graphs show normalized relative hanges
in gas lift rates

qgl

for all wells.


34

Formulation A

i=1
i=2
i=3
i=4
i=5
i=6
i=7
i=8
i=9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
i = 20
-10

Figure 11:
iteration

Formulation B

0.69

0.29

1.18

1.16

0.10

0.49

-0.01

-0.01

-0.16

0.08

0.38

0.56

0.97

1.13

0.03

0.42

1.17

0.24

0.15

0.27

0.28

1.64

2.78

1.44

0.03

-0.10

2.14

0.97

0.69

-0.01

0.00

-0.00

0.24

-0.24

1.23

0.70

-0.11

0.17

0.86

-0.36

10

B E (%)

Case-study:

-10

20

Distribution of

E
B

B E (%)

of ex iting well

of the simulation, found with Algorithm 3 (bars),

10

in isolation at
B E found by

simulating ex itation on the produ tion analog (stems). Numbers show average
E as predi ted by Algorithm 3. All values are expressed in per ent of total
B
unrealized potential.
35

Formulation A

i=1
i=2
i=3
i=4
i=5
i=6
i=7
i=8
i=9
i = 10
i = 11
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
i = 20
-20

Figure 12:
iteration

Formulation B

0.88

1.80

-0.22

6.40

0.25

0.38

0.03

0.18

-0.03

-0.14

0.85

3.46

0.83

9.80

0.62

11.98

1.41

11.64

0.28

2.37

0.81

15.25

5.16

6.51

-0.12

1.66

2.50

2.71

0.59

-0.01

0.00

0.12

0.02

0.82

0.79

18.32

-0.07

2.88

1.19

1.35

B E (%)

Case-study:

-50

20

Distribution of

E
B

50

B E (%)

of ex iting well

of the simulation, found with Algorithm 3 (bars),

100

in isolation at
B E found by

simulating ex itation on the produ tion analog (stems). Numbers show average
E as predi ted by Algorithm 3. All values are expressed in per ent of total
B
unrealized potential.
36

Formulation A

Formulation B

it:1

it:2
PSfrag repla ements

it:3

it:4

Po,r (%)

Figure 13: Case-study: Estimated

Po,r (%)

f (Po,r )

for formulations A and B, one row

for ea h iteration. The in rease in prots that was implemented in pra ti e is


shown as stems.

37

i=1
i=2
i=3
i=4
i=5
i=6
i=7
i=8
i=9
i = 10
i = 11
PSfrag repla ements
i = 12
i = 13
i = 14
i = 15
i = 16
i = 17
i = 18
i = 19
i = 20
i
qgl
Figure 14:

Case-study:

Formulation B, iteration 4, andidate ex itation

Upper and lower onstraints on

u9 .

(dashed), urrently implemented setpoint

( ir les), the target setpoint that would be implemented if no ex itation is per9


9
formed of u (stems with ir les) and target setpoint after an ex itation of u
is simulated (stems with squares).
38

Você também pode gostar