2008 Oil and Gas Production Optimization Lost potential due to uncertainty (Steinar M Elgsæter)

8/3/2019 2008 Oil and Gas Production Optimization Lost potential due to uncertainty (Steinar M Elgsæter)

http://slidepdf.com/reader/full/2008-oil-and-gas-production-optimization-lost-potential-due-to-uncertainty 1/8

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-7491Trondheim (Tel: 0047 91806886; e-mail: elgsater@ itk.ntnu.no)

∗∗ABB, Oslo(e-mail: [email protected])∗∗∗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 isoptimized with an uncertain model, some potential production profit may be lost. The costs andrisks 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 bootstrappingcan be used to estimate uncertainty resulting from fitting production models to data withlow information content. In this paper we propose how lost potential resulting from estimateduncertainty can be estimated using Monte-Carlo analysis. Based on a conservative formulationof 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 identifiedusing field data from a North Sea field.

Keywords: Uncertainty, Monte Carlo simulation, Loss minimization, Process identification,Process control, Production systems

1. INTRODUCTION

Production in the context of offshore oil and gas fields, canbe considered the total output of production wells, a massflow with components including hydrocarbons, in additionto water, CO2, H 2S , sand and possibly other components.Hydrocarbon production is for simplicity often lumpedinto oil and gas. Production travels as multiphase flowfrom wells through flow lines to a processing facility forseparation, illustrated in Figure 1. Water and gas injectionis used for optimizing hydrocarbon recovery of reservoirs.Gas lift can increase production to a certain extent byincreasing the pressure difference between reservoir andwell inlet.

Production is constrained by several factors, including: Onthe field level, the capacity of the facilities to separate

components of production and the capacity of facilities tocompress lift gas. The production of groups of wells maytravel through shared flow lines or inlet separators whichhave a limited liquid handling capacity. The productionof individual wells may be constrained due to slugging,other flow assurance issues or due to reservoir managementconstraints.

Multiphase flows are hard to measure and are usually notavailable for individual flow lines in real-time, howevermeasurements of total single-phase produced oil and gasrates are usually available, and estimates of total water The authors would like to thank the Research Council of Norway,

Norsk Hydro and ABB for funding this work.

Reservoir(s)

....

.

.

.

.

Routing

.

.

.

.

Water injection

.

.

.

.

Gasinjection

.

.

.

.

. . . .

Separation

.

...

Separation

g

....

Gaslift

ExportOil

Gas

Water

Wells

Reinjection/To sea

Lift gas/gas injection

Fig. 1. A schematic model of offshore oil and gas produc-tion.

rates can often be found by adding different measured

water rates after separation. To determine the rates of oil,gas and water produced from individual wells, the produc-tion of a single well is usually routed to a dedicated testseparator where the rate of each separated single-phasecomponent is measured. In single-rate well tests, only therates of components with current production settings aremeasured, while in multi-rate well tests rates are measuredfor different settings. The total amount of gas and waterwhich can be separated and processed is constrained bythe capacity of facilities, these capacities are themselvesuncertain. Normally production is at setpoints where someof these capacities are at their perceived constraints, there-fore a multi-rate well test cannot be performed withoutsimultaneously reducing production at some other well,



which may cause lost production and a cost. There isalso a risk that changes in setpoints during testing maycause some part of the plant to exceed the limits of safeoperation, which may force an expensive shutdown andre-start of production. Well tests are only performed whena need for tests has been identified, due to the costs andrisks involved.

In the context of oil and gas producing systems, real-timeoptimization (RTO) has been defined in Saputelli et al.(2003) as a process of measure-calculate-control cycles ata frequency which maintains the system’s optimal operat-ing conditions within the time-constant constraints of thesystem. Saputelli et al. (2003) suggests dividing real-timeoptimization into subproblems on different time scales tolimit complexity, and to consider separately reservoir man-agement, optimization of injection and reservoir drainageon the time scales of months and years, and productionoptimization, maximization of value from the daily pro-duction of reservoir fluids. Reservoir management typicallyspecifies constraints on production optimization to linkthese problems. We will refer to models for productionoptimization as production models.

The aim of production optimization is to determine set-points for a set of chosen decision variables which are op-timal by some criterion. These setpoints are implementedby altering the settings of production equipment. Decisionvariables may be any measured or computed variables as-sociated with the production system which are influencedby changes in settings, but the number of decision variablesis limited by the number of settings. We may for instancedetermine the settings of a gas lift choke by deciding atarget lift gas rate, a target annulus pressure or a targetgas lift choke opening. On short timescales the flow fromindividual wells can be manipulated by production chokesettings, by gas lift choke settings and possibly by routing

settings.

Parameters of production models should be fitted to pro-duction data through parameter estimation to compen-sate for un-modeled effects or disturbances and to setreasonable values for physical parameters which cannot bemeasured directly or determined in the laboratory. Erosionof production chokes is an example of an un-modeleddisturbance.

Planned excitation is some planned variation in one ormore decision variables designed to reveal information onproduction through measurements, for instance a multi-rate well test. The information flow in production opti-mization may be depicted as in Figure 2.

Lost profit can result if production models are fittedto production data with a low information content, asillustrated in Example 1. Throughout this paper variableswith a hat (ˆ) denote estimates, variables with bars (¯)denote measurements, and variables with a star ( ) aretrue values in some sense.

Example 1. (Uncertainty in production optimization). Thisexample illustrates a synthetic field which is producingat its processing constraints, where only single-rate welltests are available and where a production model matcheshistorical production data with a low information contentclosely. The example is motivated by observations of pro-

Productionoptimization

Production

Plannedexcitation

Productionconstraints

State and modelparameters Measurements

Decisionvariables

Disturbances

Parameter andstate estimation

Fig. 2. Example 1: Relationship between subproblems inproduction optimization.

duction data from actual fields. The field produces oil qo,water qw and gas qg from two wells, each with a gas liftrate qgl according to equations

qio = ci1 + ci2(qigl − ql,igl ) + ci3(qigl − q

l,igl )2 ∀i ∈ {1, 2} (1)

qig = K iGORqiO (2)

qiw =K iWC

1 − K iWC

qo + w1. (3)

ql,igl are local operating points and q

l,1gl = q

l,2gl = 1000. Let

θi=

ci1

ci2

ci3

K iWC K iGOR w1

and⎡

⎢⎢⎣θ,1

θ,2

θ1

θ2

⎤⎥⎥⎦ =

⎡⎢⎢⎣

39.62 0.025 −4 · 10−6 0.173 30 076.62 0.030 −4 · 10−6 0.173 30 039.62 0.033 −6 · 10−6 0.185 30 0.8576.62 0.032 −12 · 10−6 0.160 30 1.5

⎤⎥⎥⎦ (4)

The ratio of water to produced liquid (watercut) and thegas-oil ratio are assumed rate-independent. True param-

eters (θ,1

, θ,2

) and fitted parameters (θ1

, θ2

) producepredictions of measured total oil qtoto = q1o + q2O, totalgas qtotg = q1g + q2g and total water qtotw = q1w + q2w whichmatch closely for a typical set of gas lift rates with littleexcitation, shown in Figure 3, and which match the single-rate well test available for (q1gl , q2gl) = (1000, 1000).

Assume that at the current setpoint (q1gl, q2gl) = (1000, 1000)

with oil production qtoto = 116.1, all gas processing capac-ity (associated gas + gas lift) and water processing capac-ity is utilized. The current setpoint is optimal accordingto numerical optimization with parameters (θ1, θ2), yetwith the parameters (θ,1, θ,2) (q1gl, q2gl) ≈ (744, 1238) is

optimal with oil production qtoto = 116.4. The two models

and the setpoints are illustrated in Figure 4.

This example illustrates that lost profit can result even if production optimization is based on a model that fits pro-duction data well if production data has low informationcontent. 2

1.1 Prior work

In Elgsaeter et al. (2007) an analysis of production datafrom an oil and gas field determined that informationcontent appeared low and significant uncertainty shouldbe expected if production models are fitted to such data.Handling uncertainty has been identified as a key challenge



0 50 100 150 200 250 300100

110

120

qO

tot

0 50 100 150 200 250 3003200

3400

3600

qG

tot

0 50 100 150 200 250 30024

26

28

qW

tot

0 50 100 150 200 250 300500

1000

1500

qgl

i

t[hours]

Fig. 3. Example 1: True (solid) and modeled (dashed)total production rates of oil, gas and water resulting

from gas lift rates (lower graphs) q1gl (dotted) and q2gl(dashdot).

qgl

1

qO

1

qgl

1

qG

1

qgl

1

qW

1

qgl

2

qO2

qgl

2

qG2

qgl

2

qW2

Fig. 4. Example 1: True (solid) and modeled (dashed)production of oil,gas and water, well 1 (upper graphs)and well 2 (lower graphs). Optimal operating point of the true production (square) and optimal operatingpoint of modeled production (circle).

in real-time optimization of oil and gas production inBieker et al. (2006a). The concept of a loss resulting fromuncertainty is well established in control theory, it is forinstance used to select controlled variables in Halvorsenet al. (2003). The concept of the value of informationhas been applied previously in reservoir management. InBranco et al. (2005) an analysis of what the authorscalled the value of the new information was used to justifyinvestments to determine static and dynamic reservoirbehavior. Monte Carlo simulations have been appliedfor analysis of uncertainty in production optimizationpreviously in Bieker et al. (2006b), where it was appliedto well test planning, and in Bieker et al. (2007), where itwas used for optimal well management. In both cases, the

focus was on uncertainty in gas-oil and water-oil ratios forwells where these ratios do not change when productionchange and where no gas-lift was applied and only testseparator measurements were considered.

In Bieker et al. (2007) it has been proposed that uncer-tainty in production optimization should be related toproduction profit to allow structured business cases for un-certainty mitigation to be created. Elgsaeter et al. (2008)suggested how uncertainty due to low information contentin production data can be quantified using bootstrapping.No references have been found which discuss the value of information in the context of daily oil and gas productionoptimization.

1.2 Problem formulation

The objective of this paper is to develop an analyticalframework to quantify the lost potential that results fromthe low information content in production data used forproduction optimization, suitable to offshore oil and gasproduction.

2. ESTIMATING THE LOSS DUE TO UNCERTAINTY

2.1 Prerequisites

Let x be a vector of the flows of each modeled componentof production for each modeled well and let u be a vectorof decision variables.

In this paper we will consider production optimizationproblems on the form

u(θ) x(θ)

= arg maxu,x

M (x,u,d) (5)

s.t 0 = f (x,u,d, θ) (6)

0 ≤ c(x ,u ,d). (7)

θ is a vector of fitted parameters to be determined. u(θ)

and x(θ) are the decision variable value and vector of flows, respectively, that are the optimal solution of (5)-

(7) for a given θ. d is a vector of modeled and measureddisturbances independent of u. M (x,u,d) is a productionprofit measure which we seek to maximize subject to aproduction model (6) and production constraints (7) for

a given θ. We only consider instantaneous optimization,determining (u(θ), x(θ)) for the current time.

When measurement uncertainty can be neglected, therelationship between the measured production y and x isgiven by the binary routing matrix R[t]:

y[t]= R[t]x[t]. (8)

Routing R[t] may change with time, as wells are routedthrough different processing trains or to test separators.y may consist of test separator rate measurements, totalrate measurements and in some cases multiphase ratemeasurements at some or all wells.

Let the modeled output for a given parameter θ be

yθ

[t]= R[t]x(θ)[t] + β y , (9)

where β y is the vector of measurement biases due to cali-bration inaccuracies to be determined. A set of productiondata spanning N time steps



Z N =

y[1] d[1] u[1] y[2] d[2] u[2] . . .

y[N ] d[N ] u[N ] (10)

has residuals

[t] = y[t] − yθ[t] ∀t ∈ {1, . . . , N } . (11)

Parameter estimation determines θ by minimizing the sumof the squared residuals for the data set:

θ β y

=argmin

θ,βy

N t=1

w[t][t]22, (12)

where w[t] is a user-specified weighting function. WhenZ N has low information content, it may not be possible todetermine θ uniquely from (12), resulting in an uncertainty

in θ that can be quantified in terms of a probabilitydensity function f θ(θ) using bootstrapping , see Efron andTibshirani (1993) or Elgsaeter et al. (2008). Bootstrappingis a computational approach which generates an estimateof uncertainty by re-solving (12) a number of times tofit the model to number of synthetically generated, or“re-sampled”, data sets similar to (10). In Monte-Carlosimulations in subsequent sections we sample entire pa-rameter vectors rather than sampling parameter values foreach component of θ separately, ensuring that co-variancebetween terms in θ are implicitly accounted for.

2.2 Definitions

Let the potential for production optimization P o be thedifference between the production profit at the optimaloperating point (x, u, d) and the current operating point(x ,u ,d):

P o= M (x, u, d) − M (x,u,d) ≥ 0. (13)

If there were no uncertainty, (5)–(7) would yield u(θ) = u

and x(θ) = x

, P

o could be found from (13) and the entirepotential could be realized by implementing u(θ).

When θ is uncertain, the solution to (5)–(7) will be

uncertain. Uncertainty may cause estimates u(θ) and P o(θ)to differ from u and P o , and when implementing u(θ) wemust expect some of the potential P o to remain unrealized.This discussion motivates

expected potential of production optimization(P o)=

expected realizable potential (P o,r) +

expected lost potential due to uncertainty (Lu).

(14)

Let

Q=

Lu

P o, (15)

be the quotient of potential that is lost due to uncertainty.

In this paper, we will determine probability density func-tions f Po(P o), f Lu(Lu), discussed below. In addition todepending on parameter uncertainty, these probabilitydensity functions will depend on production constraintsand on the operational strategy, defined below.

P o = 0 and Lu = 0 can be interpreted as production beingoptimal. The probability density functions f Lu

(Lu) and

f P o(P o) may consist of spikes near P o = 0 and Lu = 0 and

a distribution in interval away from zero. To allow thesetwo parts of f Lu

(Lu) and f P o(P o) to be studied separately,we define the two events

H 0 : production is optimal (16)

H 1 : production is suboptimal. (17)

Estimates ˆ p(H 0) and ˆ p(H 1) can be found from estimated

probability density functions f Lu(Lu) and f P o(P o).

2.3 Estimating the potential of production optimization

An estimate P o(θ) of (13) for a given parameter estimate

θ when (x,u,d) is the current operating point is

P o(θ) = M (x(θ), u(θ), d) − M (x ,u ,d). (18)

When θ is uncertain, we may exploit an estimate of theprobabilty density function f θ(θ) to estimate a probabilitydensity function for (18) using a Monte Carlo method,generating suitable random numbers and observing thefraction of the numbers obeying some property or prop-erties, see Metropolis and Ulam (1949). The approach is

outlined in Algorithm 1 in Appendix A.

2.4 Estimating lost potential due to uncertainty

A setpoint u(θ) calculated with an uncertain θ may befound to violate production constraints (7) when imple-mented. Therefore, a strategy for constraint handling isrequired. Lost potential due to uncertainty will depend onthe strategy employed to deal with constraint handling,and to estimate lost potential due to uncertainty we mustexpress this strategy as an algorithm. We will refer to suchan algorithm as an operational strategy function .

Let u0 be the setpoint implemented on the field prior to

implementing u(ˆθ). When altering setpoints from u0 along

some path toward u(θ), constraints may be encounteredat an operating point (xc, uc) due to uncertainties. Lostpotential due to uncertainty can be estimated in a MonteCarlo fashion by determining the profit that is lost whena setpoint u(θf ) is implemented and the true parameter

value is actually θt, drawing estimates (θt, θf ) from the

estimated probability density function f θ(θ):

Lu(θt, θf , d) = M (x(θt), u(θt), d)−

M (xc(θf , θt), uc(θf , θt), d).(19)

The first term in (19) is the optimal profit with the true

parameter value θt, while the second term is an estimate of

the profit obtained when θf is erroneously assumed to bethe true parameter. (xc(θf , θt), uc(θf , θt)) is determined byan operational strategy function. The proposed algorithmfor determining f Lu

(Lu) is summarized in Algorithm 2 inAppendix A.

2.5 Discussion

In Algorithms 1 and 2 samples are drawn at random,Monte-Carlo approach in this paper. It may be pos-sible to improve accuracy and reduce computationalburden through the use of sampling techniques basedon Hammersley-sequences, see Diwekar and Kalagnanam



(1997). Algorithms 1 and 2 can also be combined withmore efficient bootstrap methods, see for instance Gigli(1996). To limit the scope and complexity of this paper,we leave such extensions for further work.

3. APPLICATIONS

In this section we will study applications of the suggested

approach, first on a set of synthetic examples, then on datafrom a real field.

3.1 Introduction

We will consider oil, gas and water rates for each well ias elements in x, that is qio, qig and qiw, respectively. Theproduction of individual wells are assumed independent of each other, or decoupled , as will typically be the case forso-called platform wells. We will consider two alternativeproduction models, (20) and (21),

qi p = max{0, ql,i p f z(zi, zi,l)

1 + αi pΔqigl

} (20)

q

i

p = max{0, q

l,i

p f z(z

i

, z

i,l

)

1 + α

i

pΔq

i

gl + κ

i

p(Δq

i

gl)

2},(21)

for all wells i and for oil, gas and water p ∈ {o ,g ,w},based on Elgsaeter et al. (2008). qigl is gas lift rate and

zi ∈ [0, 1] is the relative production valve opening of

well i. Δqigl=

qigl

ql,i

gl

− 1, where ql,igl is the gas lift rate

measured at the time of optimization for well i, and ql,i p

is the measured rate of component p of well i at themost recent well test, respectively. f z(zi, zl,i) is a nonlinearkernel which expresses the nonlinear relationship betweenvalve opening and production, which obeys f z(0, zl,i) = 0and f z(zi, zl,i) = 1. zl,i is the relative valve opening at themost recent well test. In this paper we choose

f z(zi, zl,i) =1 − (1 − zi)k

1 − (1 − zl,i)k, (22)

and choose k = 5. During optimization, we consideru = Δqgl, d = z. We will consider θ = [αo αg αw] for(20) and θ = [αo αg αw κo κg κw] for (21).

Let

qtoto

=

nwi=1

qio (23)

qtotg

=

nwi=1

qig + qigl (24)

qtotw=

nwi=1

qio (25)

where nw is the number of wells. Let qtoto (θ), qtotg (θ) and

qtotw (θ) be estimates derived by combining (23)–(25) with

(20) or (21) for a given θ.

In this paper we will consider the following measurementvector when solving (12)

y =

qtoto qtotg qtotw

T . (26)

The objective function is chosen as (27) and productionconstraints as (28)-(31):

u(θ) = arg maxu

qtoto (θ) + bo(θ) (27)

s.t. qtotg (θ) + bg(θ) ≤ qmaxg (28)

qtotw (θ) + bw(θ) ≤ qmaxw (29)

−U prc ≤ u ≤ U prc (30)

ui > ul,i ∀i ∈ W c. (31)

(28) is a constraint on gas processing capacity and qmax

gis the processing capacity for gas. (29) is a constrainton water processing capacity and qmax

w the processingcapacity for water. (31) constrains the gas lift rate of wellswith indices in the set W c which may not have their gas-liftrates decreased due to flow assurance issues. bo(θ), bg(θ)

and bw(θ) are biases which may be calculated separately

for each θ. All constraints are considered hard in thispaper.

What model structure describes the relation between uand y is itself uncertain. It is reasonable to expect a fittedmodel to be a valid description locally around setpoints(u, d) similar to those observed in the tuning set. The

mismatch between modeled outputs y and measurementsy will grow when setpoints outside the range of setpointsobserved in (10) are considered. If production optimizationconsiders large changes in u, structural uncertainties maydominate and this observation motivates (30). U prc limitsu to within U prc · 100% of its current value and is a designparamter that must be chosen sufficiently small.

qmaxg and qmax

w are themselves uncertain in practice, butidentifying or mitigating this uncertainty is beyond thescope of this paper. To avoid overstating qmax

g and qmaxw ,

and thereby overestimating Lu and P o, we assume con-servatively that the field is producing at its constraints inwater and gas capacity at the time of optimization. We

wish to solve (27)–(31) at time t = N , the end of thetuning set (10). Measurements y have a high-frequency,low-amplitude, seemingly random component which is notmodeled. If we choose qmax

g = qtotg [N ] and qmaxw = qtotw [N ],

solutions of (27)–(31) will depend to some degree on thesehigh-frequency disturbances. For this reason, we choose toconsider tuning sets where no significant changes are visi-ble in u[t] and no large disturbances visible in y[t], d[t] forthe interval t ∈ [N − L, N ] where L << N . Capacities areestimated as average measured rates during this interval,to reduce dependency on high-frequency disturbances:

qmaxg = avg(qtotg [t]) t ∈ [N − L, N ] (32)

qmaxw = avg(qtotw [t]) t ∈ [N − L, N ]. (33)

Similarly, we desire qtoto (θ), qtotg (θ) and qtotw (θ) to equalaveraged observations over t ∈ [N − L, N ] which motivates⎡⎣

bo(θ)

bg(θ)

bw(θ)

⎤⎦ = avg(

⎡⎣

qtoto [t] − qtoto (θ)[t]

qtotg [t] − qtotg (θ)[t]

qtotw [t] − qtotw (θ)[t]

⎤⎦) t ∈ [N − L, N ].

(34)

bo(θ) is included in the objective function (27) so that (18)

yields P o = 0 if (27)–(31) returns u(θ) equal to the imple-

mented u at the time of optimization. (bo(θ), bg(θ), bw(θ))

may differ from elements of β y, as β y is an estimate of the

bias over in the tuning set, while bo(θ), bg(θ) and bw(θ) are



biases calculated for a short time interval at the end of thetuning set.

We will consider the operational strategy function

uc(θf , θt) = u0 + s(u(θf ) − u0) (35)

0 = f (xc(θf , θt), uc(θf , θt), d, θt) (36)

where s = max s (37)

s.t 0 ≤ s ≤ 1 (38)u = u0 + s(u(θf ) − u0) (39)

0 ≤ c(x,u,d). (40)

0 = f (x,u,d, θt). (41)

(35)-(41) simulates moving the setpoint from u0 to-

ward u(θf ) in a straight line until a constraint is met,

when the system is described by θt. If implementing(xc(θf , θt), uc(θf , θt)) results in a decline in productionprofits, we will assume that the operating point is returnedto is value prior to optimization when solving (19).

The set of operating points which are feasible is non-convex in general, and the conservative statement of gas

and water capacities qmaxg and qmax

w means that the choiceu0 = 0 may cause (35)–(41) to meet a constraint close tou0, producing a large estimate of loss due to uncertaintyLu. Instead, we choose u0 = −U prc, as simulations haveshown this estimate to produce far more conservativeestimates of Lu. The real-world analogy to u0 = −U prcis a field where production is moved toward a calculatedsetpoint u(θ) as production is re-started after a shutdown.

We will consider two different formulations of the problem(27)-(31):

Formulation A: production model (20), a linear pro-gramming problem

Formulation B: production model (21), a nonlinear pro-gramming problem

The applications in the following section were imple-mented in MATLAB 1 and all optimization problems weresolved with the TOMLAB 2 toolbox. All nonlinear pro-gramming problems were solved sequentially with globalsolvers Jones et al. (1993) and local solvers based on se-quential quadratic programming, for parameter estimationbased on Huschens (1994), and for production optimiza-tion based on Schittkowski (1982). In all cases analyticalJacobians and Hessians were supplied.

3.2 Synthetic example

We will revisit Example 1 in the following, to illustrate thethe suggested methodology on a set of data where the truemodel is known.

Example 2. (Example 1 revisited: f P o(P o) and f Lu(Lu)).

We apply Algorithms 1 and 2 to the synthetic productiondata described in Example 1 with N t = N f = 100 fordifferent U prc. Results are shown in Figure 5. For this ex-

ample, all estimates produced by Algorithm 1 gave P o > 0using both formulations A and B, or ˆ p(H 1) = 100%.

1 The Mathworks,Inc., version 7.0.4.3652 TOMLAB Optimization Inc., version 5.5

0.1 0.2 0.3 0.4 0.5

0.20.40.60.8

1f P

o|H

1](%)

0.1 0.2 0.3 0.4 0.5

0.1

0.2f L

u|H

1

(%)

0.1 0.2 0.3 0.4 0.5

0.5

1f Q|H

1

Uprc

Fig. 5. Example 2: Estimated potential and lost potentialfor formulation A (plus) and B(circles). Center graphs

are expected values, upper and lower graphs denote95% confidence intervals.

qgl

1

qo

1

qgl

1

qg

1

qgl

1

qw

1

qgl

2

qo

2

qgl

2

qg

2

qgl

2

qw

2

Fig. 6. Example 2: Production models (20) (dotted) and(21)(dashed) based on parameters derived using boot-strapping, operating point (xl, ul) (circle). Verticalline illustrates span of decision variable values intuning set. The solid line illustrates the true model.

To reduce clutter, only every fifth model based onresampled parameters are plotted.

The use of a nonlinear production model in formulation Bresulted in estimates P o and Lu that were less dependentof U prc than those found with formulation A. While

estimates of P o remained in the region of P o = 0.29%

with formulation B as U prc grows, estimates P o foundwith formulation A grow without bounds as U prc grows.A closer analysis reveals that formulation A tends to findestimates u(θ) at the constraint (30), while formulation Bfinds estimates u(θ) within this constraint, and this mayexplain the difference observed.



As seen in Figure 6, the linear models used in formulationA predicts oil, gas and water rates within a narrow spanas qigl change compared with the nonlinear models usedin formulation B. The narrow span in predicted rateswith formulation A causes comparatively small estimates(Lu, Q) compared with formulation B. Estimates of Qfound with formulation A decreases as U prc increases,which seems counter-intuitive.

This example indicates that nonlinear models on the form(21) may be more suitable than linear models (20) inconjunction with the methods suggested in this paper. 2

3.3 Case study: North Sea field

The set of data we will study is from a North Sea field with20 gas lifted platform wells, producing predominantly oil,gas and water. Sporadic single-rate well tests are available,and no routing options are considered in u. The operatorof the field requested that all data be kept anonymous,therefore all variables will be represented in normalizedform. We choose a tuning set of 180 days. The sampling

time was one hour, and production was not significantlychanged for the last L = 10 hours prior to t = N .f P o(P o) was estimated with Algorithm 1 and f Lu

(Lu) wasestimated with Algorithm 2, both with N f = N t = 100.When estimating parameters, bounds 0 < α < 1 and −1 <κ < 0 based on physical knowledge were applied. Based onthe knowledge that the watercut is rate-independent, softconstraints which penalize deviation from αo = αw andκo = κw were added to the objective function. A declinein qtoto was visible in the tuning set, and we chose to de-trend qtoto and chose w[t] to weigh older measurements lessthan newer measurements when solving (12). For furtherdetails on the modeling, refer to Elgsaeter et al. (2008).The set of indices of wells with flow assurance issues was

W c = {1, 4, 10, 13, 14, 16}.

Results The distribution of estimates P o, Lu and Q forvarying U prc are shown in Figure 7. The models used areshown in Figure 8. For formulation A p(H 1) = 100% while p(H 1) = 95% for formulation B.

Discussion U prc limits change in decision variables toa certain range for which structural uncertainties can beneglected, and P o and Lu depend on this design parameter.As we do not know with certainty when structural uncer-tainties become significant, we cannot determine a singlevalue for U prc and so when estimating P o and Lu we haveconsidered a range of U prc. A conservative assumptionof U prc = 0.1 indicates P o ≈ 2%. The quotient Q isrelatively invariant of U prc, so from the analysis it seemsfair to conclude that on average, we should expect to looseabout 25% of the P o due to uncertainty, or Lu ≈ 0.5%when U prc = 0.1. Unfortunately, we are unable to verify

estimates P o and Lu made in this paper without furtherexperiments, and such experiments are associated withgreat cost and risk. However, the methods presented in thispaper may be used motivate such experiments in furtherwork, as the presented methods allow the associated ben-efits to be estimated. In this paper we have not consideredconstraint uncertainty . It is possible that at the time of theanalysis the field was producing below its actual process-

0.1 0.2 0.3 0.4 0.52468

10

f P

o|H

1

(%)

0.1 0.2 0.3 0.4 0.50

2

f L

u|H

1

(%)

0.1 0.2 0.3 0.4 0.5

0.10.20.30.40.5

f Q|H

1

Uprc

Fig. 7. Case study, North Sea field data: estimated poten-tial and lost potential for formulation A (plus) andB (circles). Center graphs are expected values, upper

and lower graphs denote 95% confidence intervals.ing capacities. In this case, the potential for productionoptimization would be larger than our estimates.

4. CONCLUSIONS

Based on a conservative formulation of the productionoptimization problem, we could identify potential for pro-duction optimization in excess of 2%, and lost potentialdue to the form of uncertainty considered in excess of 0.5%. As sales revenues from an oil and gas field maybe very large and the costs of mitigating uncertainty areindependent of revenue, we believe that further work mayprove that these potentials are sufficient to justify the costs

of mitigating uncertainty and optimizing production.

The methodology presented may have extensions to otherprocesses optimized using models fitted to process mea-surements with low information content.

Appendix A. ALGORITHMS

Algorithm 1. (Estimating f P o(P o)). Given a dataset Z N

on the form (10)

• determine θ using (12),

• determine f θ(θ), for instance by bootstrapping asdescribed in Elgsaeter et al. (2007),

• sample f θ(θ) N t times,• determine (x(θ), u(θ)) using (5)-(7) for each of the N t

samples, and• insert all determined (x(θ), u(θ)) in (18) to obtain the

probability density function f P o(P o).

Algorithm 2. (Estimating f Lu(Lu)). Given a dataset Z N

on the form (10), an operational strategy function, thecurrent operating point (x,u,d) and a probability density

function f θ(θ).

• do N t times· draw a sample θt from f θ(θ).

· draw N f samples θf from f θ(θ).



qio

1

qig qiw

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

qigl

20

qigl qigl

Fig. 8. Case study, North Sea field data: Production mod-els (20) (dotted) and (21)(dashed) based on parame-ters derived by bootstrapping, operating point (xl, ul)(circle). Vertical lines illustrate the range of decisionvariable values in the tuning set.Lower left bound of all plots is (0,0).Well indices along left margin.

· for each of the N f samples of θf and the sample

θt obtain a point estimate of Lu by solving (19)

• the distribution of N t × N f point estimates of Lu is

an estimate of f Lu(Lu)

REFERENCES

H.P Bieker, O. Slupphaug, and T.A Johansen. Real-timeproduction 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. Optimalwell-testing strategy for production optimization: Amonte carlo simulation approach. In SPE Eastern Regional Meeting , 2006b. SPE 104535.

H.P. Bieker, O. Slupphaug, and T.A. Johansen. Wellmanagement under uncertain gas or water oil ratios. InSPE Digital Energy Conference and Exhibition , 2007.SPE 106959.

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. Therole of the value of information and long horizontal wellsin the appraisal and development studies of a brazilian

offshore heavy-oil reservoir. In SPE/PS-CIM/CHOAInternational Thermal Operations and Heavy Oil Sym-posium , 2005. SPE 97846.

U.M. Diwekar and J.R. Kalagnanam. Efficient samplingtechnique for optimization under uncertainty. AIChE Journal , 43(2):440–447, 1997.

B Efron and R.J. Tibshirani. An Introduction to theBootstrap. Chapman & Hall, 1993.

S.M. Elgsaeter, O. Slupphaug, and T.A. Johansen. Un-certainty in model-based production optimization; chal-lenges. In Proc. SPE International Petroleum Technol-ogy Conference 2007 , 2007.

S.M. Elgsaeter, O. Slupphaug, and T.A. Johansen. Pro-duction optimization; system identification and uncer-

tainty estimation. In Proc. SPE Intelligent Energy Conference 2008 , 2008.

A. Gigli. Efficient bootstrap methods; a review. Statistical Methods and Applications, 5(1):99–127, April 1996.

I.J Halvorsen, S. Skogestad, J. Morud, and V. Alstad.Optimal selection of controlled variables. Ind. Eng.Chem. Res., 42(14):3273–3284, 2003.

J. Huschens. On the use of product structure in secantmethods for nonlinear least squares problems. SIAM Journal of Optimization , 4(1):108–129, Jan 1994.

D.R. Jones, C.D. Perttunen, and B.E. Stuckman. Lip-schitzian optimization without the lipschitz constant.Journal of Optimization Theory and Applications, 79(1):157–181, Oct 1993.

N. Metropolis and S. Ulam. The monte carlo method.Journal of the American Statistical Association , 44(247):335–341, September 1949.

L.A. Saputelli, S. Mochizuki, L. Hutchins, R. Cramer,M.B. Anderson, J.B. Mueller, A. Escorcia, A.L. Harms,C.D. Sisk, S. Pennebaker, J.T. Han, C.S. Brown,A. Kabir, R.D. Reese, G.J. Nuez, K.M. Landgren, C.J.McKie, and C. Airlie. Promoting real-time optimiza-tion of hydrocarbon producing systems. In 2003 SPE Offshore Europe Aberdeen , 2003. SPE 839781.

K. Schittkowski. On the convergence of a sequentialquadratic programming method with an augmented la-grangian line search function. Technical Report, SystemOptimization Laboratory, Stanford University, Stan-

ford,CA, 1982.

2008 Oil and Gas Production Optimization Lost potential due to uncertainty (Steinar M Elgsæter)

Documents

Transcript of 2008 Oil and Gas Production Optimization Lost potential due to uncertainty (Steinar M Elgsæter)