Journal of Petroleum Science and Engineering 165 (2018) 161–170

Contents lists available at ScienceDirect

Journal of Petroleum Science and Engineering

journal homepage: www.elsevier.com/locate/petrol

Dynamic optimization of a continuous gas lift process using a mesh refining

sequential method
Lizandro de Sousa Santos *, Kaio Magno Ferreira de Souza, Marina Ribeiro Bandeira,

n, Fernando Cunha Peixoto, Diego Martinez Prata
Víctor Rolando Ruiz Aho
Department of Chemical and Petroleum Engineering, Universidade Federal Fluminense, Rua Passo da P

atria, 156, CEP: 24210240 Niter
oi, RJ, Brazil

A R T I C L E I N F O A B S T R A C T

Keywords: Dynamic optimization of gas lift process (GLP) aims to compute control optimal trajectories of gas injection flow
Optimal control rate so that the oil production can be maximized. It has been observed that dynamic optimization of GLP is little
Mesh refinement discussed in the recent literature since most of the papers have focused on steady state optimization and multi-
Nonlinear programming variable predictive control applications. In this work, a multiple-objective dynamic optimization of a GLP is
applied with the goal of maximizing the oil production while minimizing the gas lift amount of a particular
process. To this end, a mesh refining sequential method, which transforms the dynamic optimization problem into
a finite-dimensional nonlinear program (NLP), was implemented. The main advantages of this approach are
accelerating the numerical algorithm convergence and improving the quality of the optimal control profiles. The
Pareto curve was obtained from the multiobjective optimization, allowing predicting the set of optimal solutions
for the given problem. Numerical examples evidenced that the oil production can be considerably increased with
minimal gas consumption.

1. Introduction basic definition of the optimal allocation problem. The inclusion of

Production of oil from offshore installations is correlated with high
production costs. To maximize the profit, the production rate of oil must
usually be kept as high as possible (Eikrem, 2006). In the initial stage of
its production lifetime, most oil wells flow naturally. It indicates that the
well pressure is sufficient to exceed the pressure loss due to the flow to
the offshore plant. To be capable of producing from a well that does not
flow naturally, artificial lift needs to be used. This strategy is an artificial
technique able of maintaining or increasing the oil production by
injecting gas in the bottom of the well (Nalum, 2013).
A single gas lift system consists of an annulus volume into which gas is
introduced through the gas lift choke. An injection valve allows the gas to
flow into the tubing. The injected gas can reduce the density of the
produced fluid from the well, consequently reducing the pressure at the
bottom of the well (Eikrem, 2006; Nalum, 2013 and the oil produced.
The continuous method uses gas injection with a steady-state flow for oil
production while the intermittent gas lift method (Ayatollahi et al.,
2004) uses gas injections at regular intervals (Tasmi et al., 2017).
In the absence operating constraints, it is essential to optimally allo-
cate the gas lift amongst the wells to maximize the oil production. It is the
additional operating constraints, e.g., choke control for well-rate man-
agement gives rise to a broader optimization problem definition that can
take into account an economic objective function, by the inclusion of
production cost factors and other constraints (Rashid et al., 2012).
The improvement of gas lift process typically involves the design of
sensors, information systems and control and optimization algorithms of
considerable complexity (Camponogara et al., 2010). In this way, the
quantity of injected gas is a critical variable whereas a lower value can
reduce the production and a higher value can increase the operational
costs with compression and gas usage (de Souza et al., 2010). Particu-
larly, gas lift injection rates can be used as decision variables to control
well flows and to maximize the oil production rate. Thus, a mathematical
model may be employed to predict the effect of the decision variables on
the oil production (Krishnamoorthy et al., 2016).
Some authors have been explored strategies to determine the optimal
conditions of gas lift systems. These works have tackled this problem
from different points of view: (i) formulation of the optimization problem
(dynamic or steady state) and (ii) type mathematical model (linear and
nonlinear). Among optimization works, the most recent contributions
have involved the application of different algorithms for solving together

* Corresponding author.
E-mail address: lizandrosousa@id.uff.br (L.S. Santos).

https://doi.org/10.1016/j.petrol.2018.02.019
Received 17 September 2017; Received in revised form 8 January 2018; Accepted 9 February 2018
Available online 12 February 2018
0920-4105/© 2018 Elsevier B.V. All rights reserved.
L.S. Santos et al. Journal of Petroleum Science and Engineering 165 (2018) 161–170

with steady state models: mixed-integer nonlinear optimization problem Table 1

(MINLP) (Camponogara and Nakashima, 2006; Camponogara et al., Recent papers regarding optimization of continuous gas lift process.
2010), Genetic Algorithm (Saepudin et al., 2007), Newton reduction Worksa Steady State Dynamic Dynamic
method (NRM), (Rashid, 2010). Optimization Optimization Optimization with
According to Peixoto et al. (2015) and Saepudin et al. (2007), the gas with nonlinear modelb
approximated
lift performance curve usually has a maximum flow rate point. However,
model
such a maximum exists for a steady state analysis, where the gas lift flow
rate is maintained constant for a given time horizon. On the other hand, (Kanu et al., 1981) x
(Nishikiori et al., x
for operational transitions, the optimal production can vary with time. In 1989)
this case, a dynamic optimization procedure, which requires knowledge (Camponogara and x
of the dynamic system behavior, becomes necessary. Nakashima, 2006)
The theory of dynamic optimization enables optimization of dynamic (Ray and Sarker, x
2007)
models to produce time-varying control settings, which then can be
(Saepudin et al., x
applied to the gas lift process. Such implementation requires the opti- 2007)
mizer to be highly efficient enough to handle complicated fluid flows and (Rashid, 2010) x
operational constraints. Optimization solutions lacking dynamic models (Camponogara et al., x
can propose infeasible transient operation (Vassiliadis et al., (1994)). To 2010)
(Ghaedi et al., 2013) x
circumvent it, a correct dynamic prediction of the plant must be incor- (Sharma and x
porated into the optimization algorithm. (Hartwich and Marquardt, Glemmestad, 2013)
2010; Santos et al., 2012, 2014). (Nalum, 2013) x
Several authors have been developing dynamic models for GLP with (Krishnamoorthy x
et al., 2016)
different purposes, e.g., intermittent gas lift (Ayatollahi et al., 2004),
a
linear control models (Aamo et al., 2005; Eikrem, 2006; Peixoto et al., Type of the problem solved. Continuous gas lift (CGL) or intermittent gas lift
2015), Model Predictive Control (MPC) (Ribeiro et al., 2016) and (IGL).
b
Computational Fluid Dynamics (CFD) (Kadivar and Lay, 2017). These Optimal control formulation of algebraic-differential equation (DAE)
works have shown that the development of dynamic models allows systems.
interpreting the relationship between the process variables as well as the
transient behavior of the state variables. However, none of these works process behavior, since measured disturbances may be taken into ac-
has implemented dynamic optimization algorithms for improving the count. On the other hand, the authors observed that the size of the
process performance. problem grows exponentially with the number of discrete elements.
Direct methods (single shooting, multiple shooting and simultaneous) Consequently, the number of uncertain parameters becomes larger.
are the main strategies used to solve dynamic optimization problems. In Sharma and Glemmestad (2013) an open loop optimization of the
These are based on approximating the infinite dimensional dynamic distribution of the available gas was performed using the Generalized
optimization problem into a finite dimensional problem by numerical Reduced Gradient (GRG) algorithm (Biegler, 2010). A multi-start search
discretization, in which discrete control (or input) and state parameters routine was developed to ensure that the local optimal solution was
are defined as decision variables (Biegler et al., 2002; Biegler, 2007). The closer to the global solution for five wells. The results showed that the
quality of the solution depends on the chosen discretization resolution, optimal gas allocation could increase the oil flow rate. However, the
principally of the control variables (Schlegel et al., 2005). Some authors authors considered a fixed gas lift flow rate along the time horizon, not
applied the dynamic optimization by using non-linear models: Nalum necessarily a time-dependent profile as a typical optimal control solution.
(2013), Sharma and Glemmestad (2013) and Krishnamoorthy et al. Table 1 summarizes the most recent and relevant papers reported in
(2016). the literature.
Nalum (2013) applied direct dynamic optimization algorithms As observed in Table 1 few works are devoted to use rigorous dynamic
(multiple shooting, singe shooting and simultaneous approaches (Biegler optimization methods to improve the performance of gas lift wells.
et al., 2002)) to perform the dynamic optimization of the process. An The aim of this paper is to apply the multi-objective dynamic opti-
objective function for the well-flowline network was formulated to mization of a gas lift process with the goal to maximize the oil production
reflect income from production and cost of using lift-gas injection. Ac- while minimizing gas lift usage. For this, a mesh refining method was
cording to the authors, an increase of 0.5–3.2% of oil production rate implemented. In this strategy simulation and optimization proceed
could be obtained. Furthermore, the author also concluded that the sequentially and the NLP has only the discretized control variables as
multiple shooting strategy was not particularly efficient concerning decision variables. The main advantages of this technique are minimizing
computation time and suggested to parallelize the algorithm to improve the number of decision variables (control parameters), accelerating the
its numerical performance. On the other hand, the single shooting was numerical convergence and to improving the quality of the optimal
the most efficient algorithm, in terms of computational performance. control profiles. In this work we show that to ensure the maximum oil
However, the number of collocation points of the numerical algorithms production during the transient stage a NLP dynamic optimization is
was specified heuristically, and no previous study was made to specify it. required and therefore, numerical precision is crucial to obtain the
According to the author, the number of collocation points is a crucial optimal control law over the time horizon. Mesh refining algorithms for
aspect. Decreasing the number of points could lead to faster running dynamic optimization has been applied for some authors recently (Bal-
times. On the other hand, its increase improves the numerical precision. sa-canto et al., 2001; Hartwich and Marquardt, 2010; Schlegel et al.,
Krishnamoorthy et al. (2016) developed a control oriented modelling 2005), but no application on gas lift optimization has been proposed
of a gas lifted well production suitable for real-time dynamic optimiza- before.
tion. The authors incorporated possible uncertainties (gas-oil ratio) into Several case studies were simulated, with multiple objectives
the model and solved the dynamic optimization problem by application formulated by the Weighted Sum method (Marler and Arora, 2010). A set
of a simultaneous approach using Radau scheme (Biegler, 2010). The of Pareto (Logist et al., 2012) solutions was then constructed.
goal was finding the gas lift injection rate to maximize the profit, subject The paper is organized as follows: In Section 2 the gas lift model is
to total gas capacity constraints. The results showed that the inclusion of presented. The optimization problem is formalized in Section 3. The re-
possible process uncertainties provides a more robust and conservative sults are presented and discussed in Sections 4 and 5. Finally, in Section 6
solution. Moreover, it may give a more realistic representation of the the conclusions and future implementations are presented.

162
L.S. Santos et al. Journal of Petroleum Science and Engineering 165 (2018) 161–170

Production valve Table 2

Details of independent variables and parameters (Peixoto et al., 2015).
Variable Value Description
Gas-lift valve
M 0.0289 kg/mol Molar mass of the gas
Product Ta 293 K Annulus temperature
La 230.87 m Annulus length
Va 29.012 m3 Annulus volume
Gas-lift ρo 923.9 kg/m3 Specific gravity of the oil.
Tw 293 K Well temperature
Lw 1217 m Length of the tubing above the gas lift valve
Lr 132 m Length from the reservoir to the gas lift valve
Annulus Ar 0.203 m2 Tubing cross-sectional area below the gas lift valve
Aw 0.203 m2 Tubing cross-sectional area above the gas lift valve
Cp 1.655e-3 m2 Production valve constant
Cv 15e-5 m2 Injection valve constant
Tubing Cr 2.623e-4 m2 Reservoir valve constant
Injection valve pr 2.549e7 Pa Reservoir pressure
ps 3.704e6 Pa Pressure in manifold downstream
g 9.81 m/s2 Acceleration due to gravity
Packer R 8.314 J/mol⋅K Universal gas constant

Reservoir
mto
wpo ¼ wpc (8)
mt
Fig. 1. Gas lift production well.

mt ¼ mtg þ mto (9)

2. Mathematical model
where wpc is the total wellhead mass flow rate, wpg is the wellhead gas
The dynamic model considered in this paper is based on a model mass flow rate and mt is the total mass flow rate involving oil and gas.
proposed by Eikrem (2006). The model used here is a simplified model of Note that the fluid flow rates through the valves are restricted to be
a subsea production plant. Despite the minor differences, the dynamic positive.
responses of such simple models based on mass balances have been Equations (4)–(6) are derived by assuming that the pressure drop
verified and have been shown to match the results from commercial high through the valves is equal to the kinetic energy pressure drop divided by
fidelity simulators such as OLGA® (dynamic multiphase simulation the square of a drag coefficient. This equation applies for subcritical flow,
software from Schlumberger) (Eikrem, 2006; Lund, 2014) with sufficient which will usually be the case for single-phase liquid flow (Economides
accuracy. Further work may include more advanced models for multi- et al., 2013).
phase flow that describe phenomena such as slug flow and pressure loss. The densities denoted by ρai (gas density in the annulus), ρm (density
According to Fig. 1, the well geometry is modeled as a vertical cyl- of the oil and gas mixture at the wellhead) are given by:
inder, consisting of two concentric inner cylinders being the tubing and
annulus. The injected gas flow is controlled by a gas lift choke at the ρai ¼
M
pai (10)
wellhead. In this way the gas flows along the annulus to the bottom of the RTa
well, through the injection valve and into the tubing, helping to enhance  
the oil production (Nalum, 2013). Liquid fluids oil and water are ρm ¼ mag þ mtg ðLw ⋅Aw Þ (11)
assumed to be incompressible, whereas gas obeys the ideal gas law.
The mass flow of gas in the annulus is written as mag , the mass flow of and pressures pwh (wellhead pressure), pai (the annulus pressure at the
gas in the tubing is mtg and the mass of oil in the tubing is written as mto . injection point of the column), pwi (column pressure at the injection point
Thus, the mass balance equations can be written as: of the column) and pwb (downhole pressure) are described by:
 
m_ ag ¼ wlif  wiv (1) pai ¼
RTa
⋅mag (12)
Va M
m_ tg ¼ wiv  wpg (2)   
RTw mtg
pwh ¼ ⋅ (13)
m_ to ¼ wro  wpo (3) M Lw Aw  Vo mto

where wiv is the flow gas from the annulus to the tubing, wpg is the gas gmt
pwi ¼ pwh þ (14)
flow through the production valve, wro is the oil flow from the reservoir Aw
into the tubing and wpo is the produced oil flow rate.
The algebraic equations can be written as: pwb ¼ pwi þ ρo gLr (15)
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Details of the independent variables and parameters can be checked
wpc ¼ αCp Cp ρm ⋅maxf0; pwh  ps g (4)
in Table 2:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
wiv ¼ αCV CV ρai ⋅maxf0; pai  pwi g (5) 3. Dynamic optimization formulation
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
wro ¼ αCr Cr ρo ⋅maxf0; pr  pwb g (6) 3.1. Dynamic optimization of gas lift process

mtg The theory of dynamic optimization enables optimization of dynamic

wpg ¼ wpc (7)
mt models to produce time-varying control settings, which then can be
applied to the gas lift process. These can be implemented in an open-loop

163
L.S. Santos et al. Journal of Petroleum Science and Engineering 165 (2018) 161–170

manner or in closed-loop by including measurement feedback (Biegler,

2010). In this paper the open-loop dynamic optimization is studied.
The dynamic optimization problem for the gas lift problem can be
formulated as:
  t 
J ¼ x tf ¼ ∫ tf0 ω⋅~
wpo ðtÞ  ð1  ωÞ⋅~
wlift ðtÞ dt (16a)

s.t.

x_ ¼ f ðxðtÞ; yðtÞ; uðtÞ; tÞ (16b)

xðt0 Þ ¼ x0 (16c)

hðxðtÞ; yðtÞ; uðtÞ; tÞ ¼ 0 (16d)

gðxðtÞ; yðtÞ; uðtÞ; tÞ  0 (16e)

uðtÞ 2 ½uL ; uU  (16f)

where ω 2 ½0; 1 is a weight factor that balances normalized oil produc-

tion w~ po ðtÞ and gas lift w ~ lift flow rates. J 2 ℝ is the objective function,
f 2 ℝnx þny , g 2 ℝng and h 2 ℝnh are vector valued functions that represent
the DAE (differential algebraic equation) system, inequality constraints
and equality constraints, respectively. u 2 ℝnu denotes the vector of
control variables, xðtÞ 2 ℝnx and xðtÞ _ 2 ℝnx are the vectors of differential
variables and their time-derivatives, respectively, and yðtÞ 2 ℝny is the
vector of the algebraic variables. DAEs in process engineering are typi-
cally specified as initial value problems with given initial conditions,
xðt0 Þ ¼ x0 , and the final time tf can be either fixed or an unknown de-
cision parameter as well. The sub-indexes L and U represent the lower
and upper bounds, respectively. Fig. 2. Summary of the mesh refining approach.
Notice that the well-flowline model is given as a semi-explicit DAE in
Equations (16b)-(16d). Input control inputs being gas lift flow rates are Table 3
constrained by the path constraints (Equation (16e)) and bounds in Setting for simulation and optimization of GLP.
Equation (16f).
NLP solver SQP (fmincon)
The sequential approach is the method used in this work. In this Integrator solver Dassl (Brenan et al., 1989)
strategy a convenient way to parameterize uðtÞ is by subdividing the time ϕr 10–5
horizon into ns control elements, or elements, and using a polynomial ϕA 10–5
interpolation over each time interval t ¼ {t1, ⋯,tns} with equidistant el- εfun 10–4
εcon 10–4
ements. Thus, a component of uðtÞ can be written as uðtÞ ¼ ℘ðt; υÞ where
εx 10–4
℘ðt; υÞ is a polynomial that interpolates the control parameters tol 10–6
υ(t) ¼ { υ1 ; ⋯; υns1 ; υns }. In this work, it was used a piecewise constant
parameterization for this purpose:
where J1 ¼ ω⋅~ wpo ðtÞ and J2 ¼ ð1  ωÞ⋅~ wlift ðtÞ. Both terms are between
uðti Þ ¼ υi ; ti1 < t  ti ; i ¼ 1; …; ns (17)
0 and 1 because the individual objective function are scaled according to:
~ po ðtÞ ¼ wpo ðtÞ=wpo;max ðtÞ and w
w ~ lift ðtÞ ¼ wlift ðtÞ=wlift;max ðtÞ, in which
3.2. Multiobjective optimization wpo;max ðtÞ and wlift;max ðtÞ are maximum oil and gas lift flow rates.
Notice that Equation (18) is a linear relation of the two objective
In many engineering systems, we need to find solutions in the pres- functions. Setting the weights in front of each objective function, and
ence of conflicting objectives. In such cases, solutions are chosen such maximizing J yields a Pareto point. By changing the weights uniformly,
that there are reasonable trade-offs among different objectives. Pareto say from 0 to 1, we can obtain a series of Pareto points on the Pareto
search is an approach for handling such situations. Instead of generating frontier (Messac, 2015).
a single optimal solution, many solutions are generated that satisfy Par-
eto Optimality Criterion. The set of all Pareto optimal solutions form a
surface known as a Pareto front. The Pareto front helps engineers un- 3.3. Mesh refining algorithm
derstand the nature of trade-offs that need to be made to select suitable
solutions (Lotov et al., 2005). The basic steps of the mesh refining algorithm are shown in Fig. 2 and
One of the most often utilized approaches to generate this Pareto set is are outlined in the following pseudocode:
the Weighted Sum (WS) (Marler and Arora, 2010) of the individual ob- Step 1. Set initial guess (index ℓ ¼ 2) for control variable and choose
jectives. In this method a number of single-objective problems are solved values for the discretization and the integration tolerances (summarized
for a grid of different weights using deterministic or stochastic optimi- in Table 3). Notice that in the first iteration the control variable has four
zation routines. elements (ns ¼ 8). Usually, the initial guess is constant and corresponds to
By considering the maximization of the given objective function: the average of minimum and maximum control bounds;
Step 2. Perform the equidistant parameterization according to
  t
J ¼ x tf ¼ ∫ tf0 ½J1  J2  dt (18) Equation (17): uðti Þ ¼ υi ; ti1 < t  ti ; i ¼ 1; ⋯; ns ;
Step 3. Initial optimization run: compute, using a suitable NLP solver,

164
L.S. Santos et al. Journal of Petroleum Science and Engineering 165 (2018) 161–170

Table 4 17970
Initial conditions (Peixoto et al., 2015).

Wellhead oil flow rate (bbl/day)

Variable Value Description

mag 4350.1 mass of gas in the annulus (in kg)

17940
mtg 10951 mass of gas in the tubing (in kg)
mto 86038 mass of oil production in the column (in kg)
17911
Weellhead gas flow rate (MMSCF/day)

32 a MMSCF/day 17881
30 13.51
18.03
28
22.52 17852
26 27.03
31.54
24
22 17822
20 14 18 23 27 32
18 Gas lift flow rate (MMSCF/day)
16
Fig. 4. Influence of gas lift flow rate over the oil production.
14

0 1 2 3 4 5 6 routine fmincon (Sequential Quadratic Programming (Biegler, 2010), was

used implemented in a 2.6 GHz workstation with 8 GB memory. Table 3
Time (h)
provides an overview of the algorithmic settings by considering the
above routines.where ϕr and ϕA denote the relative and absolute toler-
23000 MMSCF/day
b ances of de integrator solver, εfun , εcon and εx are the tolerances for
Weellhead oil flow rate (bbl/day)

13.51 objective function, constraints and state variables of the NLP solver,
22000 18.03 respectively and tol is the tolerance of the sequential algorithm.
22.52
21000 27.03
31.54 4. Results of the dynamic simulation
20000
In this section we present some numerical case studies to investigate
19000 the dynamic behavior of gas lift process. The problem has been solved
starting from the initial conditions summarized in Table 4, based on the
18000 work of Peixoto et al. (2015) for a time horizon of 5.56 h (2⋅104 s). The
time horizon was determined to ensure convergence to steady state.
17000 This problem considers the operation of a GLP with constant down-
hole pressure, according to the model presented in Section 3. The main
0 1 2 3 4 5 6
variables to be studied are the wellhead gas, oil flow rates and downhole
Time (h) pressure. The usual units for gas flow rate (MMSCF/day -million standard
cubic feet per day) and oil flow rate (US bbl – barrels of crude oil) were
Fig. 3. Wellhead gas (a) and wellhead oil (b) flow rates for different gas lift
used.
flow rates.
Fig. 3 illustrates the profiles of then wellhead gas (a) and oil (b) flow
rates for five different values of gas lift flow rate. The range of variation of
the optimal control profile uðtÞ corresponding to the control discretiza-
gas lift flow rate was from 13.51 MMSCF/day up to 31.54 MMSCF/day.
tion ns ¼ 2ðℓþ1Þ ; The results of Fig. 3a indicate that the gas flow rate profile is different
Step 4. If ℓ > 2 computing the mesh refining convergence: Φℓ ¼ for each case. For gas lift flow rates below 22.52 MMSCF/day, it is noted
kðJℓ  Jℓ1 Þ=Jℓ1 k. In the case of Φℓ is lesser than a given tolerance, tol,
the algorithm stops. Conversely, if Φℓ  tol the algorithm continues to
next iteration ℓ ¼ ℓ þ 1, going to step 5. (here k⋅k denotes the L2 norm). 1.20x10
7
MMSCH/day
Step 5. Mesh refining by doubling the number os elements, ns ¼
13.51
2ðℓþ2Þ , for the next NLP run. 18.03
Downhole pressure (Pa)

In the sequential method, the DAE model is not present directly in the 22.52
NLP problem, rather it is solved by numerical integration in each function 1.15x10
7 27.03
evaluation step of the NLP solver to determine the state variables for a 31.54
given control input variable and therefore present implicitly. Algorithms
for the solution of this NLP, typically Sequential Quadratic Programming
methods (Biegler, 2010), require gradient information of the constraints 1.10x10
7

and the objective function with respect to the decision variables. In this
paper we consider the explicit solution of the arising sensitivity equation
systems, which is the method adopted in most optimization algorithms
(Feehery and Barton, 1998). 7
1.05x10
The algorithm presented above has been implemented into the soft- 0 1 2 3 4 5 6
ware Scilab® 5.1 (Scilab Enterprises, 2012). For the DAE solution we Time (h)
have used the Dassl (Brenan et al., 1989) solver and for the NLP the
Fig. 5. Downhole pressure for different gas lift flow rates.

165
L.S. Santos et al. Journal of Petroleum Science and Engineering 165 (2018) 161–170

8 elements
1.105x10
7
31.54 16 elements

Gas lift flowrate (MMSCF/day)

32 elements
Downhole pressure (Pa)

27.03 64 elements
7
1.100x10
22.53
7
1.095x10
18.02

1.090x10
7 13.52

9.01
7
1.085x10
4.51
14 18 23 27 32
0 1 2 3 4 5 6
Gas lift flow rate (MMSCF/day)
Time (h)
Fig. 6. Influence of gas lift flow rate over the downhole pressure.
Fig. 7. Optimal gas lift profiles for increasing number of elements (for ω ¼ 0:5Þ.

Table 5
Summary of the results of the dynamic model. These behavior was also observed by Ayatollahi et al. (2004), which
Case Gas lift flow rate Oil flow Barrels of crude Oil-gas lift verified that the oil production increases with the pressure depletion
[MMSCF/day] rate oil produced ratio (OGR) along the production well of an interminent gas lift process.
[bbl/day]a [bbl]b The summary of the results, for the different cases, is summarized in
Case 13.51 17840.57 4105.16 10.05 Table 5.
1 According to Table 5, the total produced oil flow rate increases
Case 18.03 17944.21 4193.66 7.70 directly with the gas flow rate. The oil-gas lift ratio (OGR) is used to
2
Case 22.52 17956.74 4248.12 6.24
measure the relation of the rate produced oil and the rate gas lift injected
3 into the well. As observed, the OGR decreases with the increase of gas lift
Case 27.03 17916.51 4275.35 5.23 flow rate, as expected. These results reveal that the Case 5 would be the
4 optimal case, if the goal was to maximize the oil produced, producing
Case 31.54 17843.04 4288.97 4.50
4288.97 bbl. On the other hand, if the objective was to minimize gas
5
consumption, Case 1 would be the best choice, with 13.51 MMSCF/day.
a
Steady state condition. It is important to emphasize that the above results were obtained by
b
Barrels of crude oil produced in the period of 5.56 h. considering a constant gas flow rate, for each case. Nevertheless, these
results can be improved by the application of dynamic optimization of
that the wellhead gas flow rate decreases with time. On the other hand, the process, as will be discussed in the next topic.
for flow rates above 22.52 MMSCF/day the gas flow rate increases from
the starting point. In Fig. 3b it is possible to note that the transient phase 5. Dynamic optimization results
occurs before 3 h with the oil flow rates presenting different trajectories.
However, the steady state flow rate is very close (around 18000 bbl/day) In this section, we discuss on the results of dynamic optimization of
for all profiles. the given GLP. The results are compared with the results of the previous
Fig. 4 illustrates the influence of gas lift over the oil flow rate at the section (base-case). For this, the optimization problem was solved for
steady state point. It can be observed that the oil flow rate first increases, ω ¼ 1, 0.75, 0.5 and 0.25. The bounds of the decision variable were
achieves a maximum value equals to 17958 bbl/day and then decreases specified as uL ¼ 4:51 MMSCF=day and uU ¼ 31:54 MMSCF=day.
again. Such result is according to the results observed in the work of tf
Moreover, for the objective function, in which ∫ to wpo;max ðtÞdt ¼
Peixoto et al. (2015), Eikrem (2006) and Saepudin et al. (2007). It occurs tf
because the reduction of downhole pressure due to gas injection nor- 4310 bbl and ∫ to wlift;max ðtÞ dt ¼ 31:54 MMSCF are maximum oil and gas
mally increases the oil production rate and the gas injection lightens the amounts.
fluid column. On the other hand, injecting too much amount of gas can
increase the downhole pressure, which may decrease the oil production 5.1. Results for ω ¼ 0:5
rate. The change of downhole pressure with time can be seen in Fig. 5.
In Fig. 5 it can be noted that the downhole pressure decreases from its In this section we analyze the performance of the mesh refining al-
initial value, achieving the steady state after approximately 3 h. Clearly, gorithm for ω ¼ 0:5. The optimal control profiles, for each discretization
such transient trajectory depends on the initial conditions, which are not level, are illustrated in Fig. 7. The results evidence that the mesh refining
necessarily at steady state. Another effect that must be considered is the algorithm starts with 8 elements and is equidistantly refined up to the
influence of gas lift flow rate over the downhole pressure. It is observed convergence with 64 control elements. Apparently, the control profile
that there is a minimum pressure value, equals to 1.084  107 Pa, at the with 64 elements look slightly smoother than the other discretization
steady state, being represented by the gas flow rate of 22.52 MMSCF/day. levels, as expected. The control trajectories evidence that at the begin-
Such result can be better visualized in Fig. 6. ning of the process the control variable remains at its lower bound (4.51
Fig. 6 illustrates the steady state downhole pressure change with the MMSCF/day) and gradually grows up to the upper bound (31.54
gas lift flow rates. It is possible to note that the downhole pressure, at MMSCF/day) after 4.6 h. It is also noted that during most of the time the
steady state, presents an opposite profile of Fig. 4. As expected, the lower gas lift flow rate remains lower than the optimal steady state value ob-
the downhole pressure, the more favorable the oil production will be. tained in the las section, which was 31.54 MMSCF/day. The results also
reveal that the discretization of the control variable is crucial to improve

166
L.S. Santos et al. Journal of Petroleum Science and Engineering 165 (2018) 161–170

Table 6 a
Statistical data of the meh refining algorithm (for ω ¼ 0:5Þ. MMSCF/day
24000

Weellhead oil flow rate (bbl/day)

ℓ J Accumulated CPU (min)a ns Φ⋅105
22000 Optimal
20000 13.41
2 0.42798 1.65 8 –
18.03
3 0.44501 6.84 16 0.039 18000
22.52
4 0.44504 13.90 32 6.38 16000 27.03
5 0.44504 52.72 64 0.009
14000 31.54
Referenceb 0.44504 398.52 64 – 12000
a
Minutes. 10000
b
Single iteration with 64 elements. 8000
6000
4000
MMSCF/day 2000
24000 0
Weellhead oil flow rate (bbl/day)

Optimal
13.41 0 1 2 3 4 5 6
22000
18.03 Time (h)
20000 22.52
27.03
18000 31.54 b MMSCF/day
24000

Weellhead oil flow rate (bbl/day)

16000 Optimal
22000 13.41
14000 18.03
20000 22.52
12000 27.03
18000 31.54
10000
0 1 2 3 4 5 6 16000
Time (h)
14000
Fig. 8. Optimal oil flow rate (for ω ¼ 0:5Þ.
12000

the quality of the solution, as well as for the optimal value obtained, as 10000
summarized in Table 6. 0 1 2 3 4 5 6
Table 6 shows the computation statistics of the mesh refining algo- Time (h)
rithm. The columns of the table summarize the values function, CPU in
minutes, number of elements and convergence rate, for each iteration ℓ. Fig. 9. Optimal oil production (a) and oil production with constant 31.54
MMSCF/day of gas (b).
The last line of shows the data for the dynamic optimization obtained
using a single iteration, with no mesh refinement, using 64 elements
(reference case). Table 7
Table 6 reveals that the effort to improve the solution quality, Optimal results for ω ¼ 0:5.
regarding the objective function value, grows with number of iterations.
Data Constant gas lift Optimal results Difference (%)
As observed, the CPU required for the reference solution is much larger (13.41 MMSCF/
than the CPU of the proposed algorithm, with comparable accuracy. This day)a
means that a CPU economy of 86% could be obtained by the application Oil produced [bbl] 4105.16 4166.33 þ1.49
of the mesh refining sequential method, considering the given tolerances. Gas lift amount 3.11 2.64 15.11
Fig. 8 compares the optimal wellhead oil flow rate and the profiles [MMSCF]
obtained in the last section. It is possible to verify that the optimal profile J – 0.44 –
is significantly different from the previously obtained profiles, presenting Oil-gas lift ratio (OGR) 10.05 12.02 þ10.60
an oscillatory behavior, with two maximum points at 0.4 h and 5 h, Constant gas lift Optimal Difference
respectively and a minimum value of 12000 bbl/day at the time of 1 h. (31.54 MMSCF/ results (%)
These instabilities coincide with the switching points of the gas flow rate, day)b
as observed in Fig. 7. It is important to emphasize that the final point does Oil produced [bbl] 4288.97 4179.60 2.55
not corresponds to the steady state, since the gas lift flow rate is not Gas lift amount 7.31 2.64 63.88
constant along time. [MMSCF]
Fig. 9 compares the optimal oil production with the production ob- J – 0.44 –
tained with a constant gas lift flow rate of 31.54 MMSCF/day (Case 5). Oil-gas lift ratio (OGR) 4.49 12.02 þ167.71
The area of each graph corresponds to the oil production. It is possible to a
Case 1.
observe that the area of both cases is relatively close (see Table 7), b
Case 5.
indicating the oil production of both cases is relatively close.
Fig. 10 illustrates the gas consumption of both cases. The results re- by 1.49% if compared to Case 1. On the other hand, when the optimal
veals that the area of the optimal case (Fig. 10a) is considerably lower case (with ω ¼ 0:5) is compared with Case 5, the optimal oil production
than the area obtained in Fig. 10b, indicating that gas consumption could was 2.55% lesser. Nevertheless, for both cases, it was possible to reduce
be significantly reduced under these conditions. the gas lift amount by 15.11%, in relation to the Case 1 and by 63.88% in
Table 7 compares the optimal oil produced, gas lift amount and OGR relation to the Case 5. It explains the higher values of OGR observed for
obtained with constant gas lift flow rates (13.41 MMSCF/day and 31.54 both cases: an increase of 10.6, in relation to Case 1 and 167% in relation
MMSCF/day). As observed, it was possible to increase the oil production do Case 5.

167
L.S. Santos et al. Journal of Petroleum Science and Engineering 165 (2018) 161–170

a Table 8
32.00 Optimal results for different values of ω.
Gas lift flowrate (MMSCF/day)

Data ω ¼1 ω ¼ 0:75 ω ¼ 0:5 ω ¼ 0:25

28.00
Oil produced [bbl] 4309.39 4282.17 4179.60 3907.73
Gas lift amount [MMSCF] 5.73 4.22 2.64 1.84
24.00
J 1 0.71 0.44 0.19
Oil-gas lift ratio (OGR) 5.72 7.72 12.02 20.14
20.00
CPU (min) 18.14 49.21 52.72 37.26
16.00 Reference CPU (min) a
98.33 210.43 398.52 124.12
a
12.00 Single iteration with 64 elements.

8.00
8
4.00 Case 5

7
0 2 4 6 Case 4

Gas lift amount (MMSCF)

w=1
Time (h) 6
Case 3
5
Case 2 w = 0.75
36.00 b 4
32.00 Case 1
Gas lift flowrate (MMSCF/day)

3 w = 0.5
28.00
w = 0.25
24.00 2
20.00 1
16.00
0
12.00 1 2 3 4 5 6 7 8 9
8.00 Cases
4.00 Fig. 12. Comparison of gas lift amount (in MMSCF).

0.00
0 2 4 6 0:25; presents a minimum gas flow profile up to the time of 4 h, pre-
senting an increase to the maximum value after this point. On the other
Time (h) hand, for ω ¼ 1; the gas flow rate remains almost the whole period with
Fig. 10. Optimal gas lift amount (a) and gas lift amount with constant 31.54
values above 20 MMSCF/day. These results indicate that for lower ω
MMSCF/day of gas (b). values the optimization algorithm favour the minimization of gas lift
amount. Conversely, higher ω values favour the maximization of the oil
production, as expected. These complete results are summarized in
Table 8.
w=1 It is observed in Table 8 that the oil production decreases with the
w = 0.75
Gas lift flowrate (MMSCF/day)

30.00 reduction of ω. Notice that the maximum oil production was 4309.39 bbl,
w = 0.5
w = 0.25
4400
20.00 w=1
Case 5
Case 4 w = 0.75
Case 3
Case 2
Oil production (bbl)

4200 w = 0.5
10.00
Case 1

0.00 4000
0 1 2 3 4 5 6
w = 0.25
Time (h)
Fig. 11. Optimal gas lift profile for different values of ω.
3800
5.2. Multiobjective dynamic optimization 1 2 3 4 5 6 7 8 9
Cases
The optimal control profiles (with 64 elements) for ω ¼ 0:25; 0:5;
0:75 and 1 are presented in Fig. 11. The trajectory obtained with ω ¼ Fig. 13. Comparison of oil production (in bbl).

168
L.S. Santos et al. Journal of Petroleum Science and Engineering 165 (2018) 161–170

w = 0.25 trade-off between the produced oil ( J2 ) and gas lift amount ( J1 ) is
20 plotted for different values of ω. Notice that J1 was considered negative
because the optimization problem aims minimizing J1 . The trade-off is
clearly visible, i.e., focusing on J1 decreases J2 and vice versa.
A central difficulty in multi-objective optimization is that there is no
15
definite method to define the concept of a most preferred solution.
w = 0.5 Furthermore, selecting a set of weights that indicates the importance of
Case 1 one objective over another is not an easy task. According to our nu-
OGR

10 merical experience, the previous results indicate that the range between
Case 2 w = 0.75 0.5 and 0.75 guarantees a high level of oil production and low gas con-
Case 3 w=1 sumption. However, if the gas consumption has no restrictions, a weight
Case 4 factor tending to the unit value is more profitable.
Case 5
5
Finally, the results of this section show that it would not be appro-
priate to maintain a constant lift gas flow during a transient operating
phase to avoid excessive consumption of gas lift or low oil production.
0
1 2 3 4 5 6 7 8 9 6. Conclusions

Cases
In this work the dynamic optimization of an oil production process via
Fig. 14. Comparison of OGR. gas lift was proposed. The problem was formulated to maximize the oil
production and to minimize the gas lift usage in a given transient oper-
ation. The optimization algorithm is based on the mesh refining
sequential approach where the gas lift flow rate is used as control
variable.
In the first part of the work a sensitivity analysis was performed in
1.00 w=1
relation to the gas lift flow rate, showing that the oil production is
directly proportional to the gas lift flow rate. Results also revealed that an
optimal flow rate can be obtained (for the steady state operation).
0.75 w = 0.75 In the second part of the work the dynamic optimization was applied,
evidencing that a considerable amount of gas lift could be minimized
J2

0.50 w = 0.5
w = 0.25
0.25
-1.00 -0.98 -0.96 -0.94 -0.92 -0.90
-J1
Fig. 15. Weighted Sum for Pareto Frontier (using linear interpolation).
for ω ¼ 1, while the required gas lift was 5.73 MMSCF. Nevertheless,
even so, the gas lift consumption was inferior to the gas lift amount
required in Case 5, which was 7.31 MMSCF. On the other hand, for ω ¼
0:25 a significant reduction of gas lift amount could be obtained, with
1.84 MMSCF of gas. However, the oil production was the lowest of all
cases, equals to 3907.73 bbl.
Table 8 also evidences that the gas consumption can be considerably
reduced by the implementation of an optimal control policy instead of
making heuristic decisions, as we have done in Section 4. These results
are confirmed by the OGR values, which grows considerably with ω.
Obviously, such results are valid for this particular system, however, they
are sufficient to prove that the implementation of dynamic optimization
can bring operational and economic benefits.
Finally, the CPU time, shown in Table 8, evidences that it could be
considerably reduced, in comparison with the reference case, by the
application of the mesh refining algorithm.
Figs. 12–14 summarize the bar-graph of gas lift amount, oil produc-
tion and OGR, respectively, regarding all cases previously discussed. We
can observe in Fig. 12 that the main advantage of the dynamic optimi-
zation was to allow significant gas savings compared to previous cases. In
Fig. 13, the maximum oil production was obtained with ω ¼ 1:0. On the
other hand, a substantial amount of gas was used, as shown in Fig. 13.
Fig. 15 displays the obtained Pareto set for the above results. The
while, on the other hand, the oil production could be maximized, ac-
cording to the objectives. Since it is a multi-objective, the balance be-
tween maximization and minimization must be adjusted a priori in order
to prioritize one or another goal. The results evidenced that the relation
between oil produced and gas lift amount could be improved for different
cases.
Finally, the work showed that the mesh refining sequential approach
is suitable to calculate the optimal gas lift trajectories and that is also
efficient enough to be used in low-cost computing platforms with sig-
nificant competitive computation times.
Our future research includes the integration of the optimization
software with process control systems. We also intend to implement
adaptive techniques, such as wavelets to be embedded into the mesh
refining approach, in order to reduce the CPU time.
Acknowledgment
We thank the Brazilian Agency CAPES (Coordenaç~ao de Aperfeiçoa-
mento de Pessoal de Nível Superior) and FAPERJ (Fundaç~ao de Amparo a
Pesquisa do Rio de Janeiro) (number 15/2015) for supporting this work.
References
Aamo, O.M., Eikrem, G.O., Siahaan, H.B., Foss, B.A., 2005. Observer design for
multiphase flow in vertical pipes with gas-lift - Theory and experiments. Model.
Identif. Control 26, 65–80. https://doi.org/10.1016/j.jprocont.2004.07.002.
Ayatollahi, S., Narimani, M., Moshfeghian, M., 2004. Intermittent gas lift in Aghajari oil
field, a mathematical study. J. Petrol. Sci. Eng. 42, 245–255. https://doi.org/
10.1016/j.petrol.2003.12.015.
Balsa-canto, E., Banga, J.R., Alonso, A.A., Vassiliadis, V.S., 2001. Dynamic optimization of
chemical and biochemical processes using restricted second-order information.
Comput. Chem. Eng. 25, 539–546.
Biegler, L., 2010. Nonlinear Programming: Concepts, Algorithms, and Applications to
Chemical Processes.
Biegler, L., Cervantes, A., W€achter, A., 2002. Advances in simultaneous strategies for
dynamic process optimization. Chem. Eng. Sci. 57, 575–593.
Biegler, L.T., 2007. An overview of simultaneous strategies for dynamic optimization.
Chem. Eng. Process. Process Intensif. 46, 1043–1053. https://doi.org/10.1016/
j.cep.2006.06.021.

169
L.S. Santos et al.
Brenan, K.E., Campbell, S.E., Petzold, L.R., 1989. Numerical Solution of Initial Value
Problems in Differential-algebraic Equations.
Camponogara, E., Nakashima, P.H.R., 2006. Solving a gas-lift optimization problem by
dynamic programming. Eur. J. Oper. Res. 174, 1220–1246. https://doi.org/10.1016/
j.ejor.2005.03.004.
Camponogara, E., Plucenio, A., Teixeira, A.F., Campos, S.R.V., 2010. An automation
system for gas-lifted oil wells: model identification, control, and optimization.
J. Petrol. Sci. Eng. 70, 157–167. https://doi.org/10.1016/j.petrol.2009.11.003.
de Souza, J.N.M.N.M., de Medeiros, J.L.L., Costa, A.L.H.L.H., Nunes, G.C.C., 2010.
Modeling, simulation and optimization of continuous gas lift systems for deepwater
offshore petroleum production. J. Pet. Sci. Eng. 72, 277–289. https://doi.org/
10.1016/j.petrol.2010.03.028.
Economides, M.J., Hill, A.D., Ehlig-Economides, C., Zhu, D., 2013. Petroleum Production
Systems. Prentice Hall, NY.
Eikrem, G.O., 2006. Stabilization of Gas-Lift Wells by Feedback Control.
Feehery, W.F., Barton, P.I., 1998. Dynamic optimization with state variable path
constraints. Comput. Chem. Eng. 22, 1241–1256. https://doi.org/10.1016/S0098-
1354(98)00012-X.
Ghaedi, M., Ghotbi, C., Aminshahidy, B., 2013. Optimization of gas allocation to a group
of wells in gas lift in one of the Iranian oil fields using an efficient hybrid genetic
algorithm (HGA). Pet. Sci. Technol. 31, 949–959. https://doi.org/10.1080/
10916466.2010.535081.
Hartwich, A., Marquardt, W., 2010. Dynamic optimization of the load change of a large-
scale chemical plant by adaptive single shooting. Comput. Chem. Eng. 34,
1873–1889. https://doi.org/10.1016/j.compchemeng.2010.02.036.
Kadivar, A., Lay, E.N., 2017. A computation fluid dynamic model for gas lift process
simulation in a vertical oil well. J. Theor. Appl. Mech. 47, 49–68. https://doi.org/
10.1515/jtam-2017-0004.
Kanu, E.P., Mach, J., Brown, K.E., 1981. Economic approach to oil production and gas
allocation in continuous gas lift (includes associated papers 10858 and 10865 ).
J. Pet. Technol. 33, 1887–1892. https://doi.org/10.2118/9084-PA.
Krishnamoorthy, D., Foss, B., Skogestad, S., 2016. Real-time optimization under
uncertainty applied to a gas lifted well network. Processes 4, 52. https://doi.org/
10.3390/pr4040052.
Logist, F., Vallerio, M., Houska, B., Diehl, M., Van Impe, J., 2012. Multi-objective optimal
control of chemical processes using ACADO toolkit. Comput. Chem. Eng. 37,
191–199. https://doi.org/10.1016/j.compchemeng.2011.11.002.
Lotov, A.V., Kamenev, G.K., Berezkin, V.E., Miettinen, K., 2005. Optimal control of
cooling process in continuous casting of steel using a visualization-based multi-
criteria approach. Appl. Math. Model. 29, 653–672. https://doi.org/10.1016/
j.apm.2004.10.009.
Lund, T., 2014. Non-linear Model Predictive Control for an Oil Production Network Based
on Gas-lift.
170
Journal of Petroleum Science and Engineering 165 (2018) 161–170
Marler, R.T., Arora, J.S., 2010. The weighted sum method for multi-objective
optimization: new insights. Struct. Multidiscip. Optim. 41, 853–862. https://doi.org/
10.1007/s00158-009-0460-7.
Messac, A., 2015. In: Cambridge (Ed.), Optimization in Practice with MATLAB®: for
Engineering Students and Professionals. Cambridge University Press, New York.
Nalum, K., 2013. Modeling and Dynamic Optimization in Oil Production.
Nishikiori, N., Redner, R.A., Doty, D.R., Schmidt, Z., 1989. An improved method for gas
lift allocation optimization. SPE Annu. Tech. Conf. Exhib. 105–110. https://doi.org/
10.2118/19711-MS.
Peixoto, A.J., Pereira-dias, D., Xaud, A.F.S., Secchi, A.R., 2015. Seeking oil oil gas control
required is required. In: IFAC-PapersOnLine, pp. 21–26.
Rashid, K., 2010. Optimal allocation procedure for gas-lift optimization. Ind. Eng. Chem.
Res. 49, 2286–2294. https://doi.org/10.1021/ie900867r.
Rashid, K., Bailey, W., Cout, B., 2012. A survey of methods for gas-lift optimization.
Model. Simul. Eng. 2012, 1–16. https://doi.org/10.1155/2012/516807.
Ray, T., Sarker, R., 2007. Genetic algorithm for solving a gas lift optimization problem.
J. Petrol. Sci. Eng. 59, 84–96. https://doi.org/10.1016/j.petrol.2007.03.004.
Ribeiro, C.H.P., Miyoshi, S.C., Secchi, A.R., Bhaya, A., 2016. Model Predictive Control
with quality requirements on petroleum production platforms. J. Pet. Sci. Eng. 137,
10–21. https://doi.org/10.1016/j.petrol.2015.11.004.
Saepudin, D., Soewono, E., Sidarto, K.A., Gunawan, A.Y., Siregar, S., Sukarno, P., 2007.
An investigation on gas lift performance curve in an oil-producing well. Int. J. Math.
Math. Sci. https://doi.org/10.1155/2007/81519.
Santos, L.S., Secchi, A.R., Biscaia, E.C., 2012. Wavelet-Threshold Influence in Optimal
Control Problems, 22 Eur. Symp. Comput. Aided Process Eng.
Santos, L.S., Secchi, A.R., Biscaia, E.C., 2014. A comparative study between wavelet-
adaptive multiple shooting and single hooting implemented in a MATLAB-EMSO
environment. Eng. Optim. https://doi.org/10.1201/b17488-3.
Schlegel, M., Stockmann, K., Binder, T., Marquardt, W., 2005. Dynamic optimization
using adaptive control vector parameterization. Comput. Chem. Eng. 29, 1731–1751.
https://doi.org/10.1016/j.compchemeng.2005.02.036.
Scilab Enterprises, 2012. Scilab: Free and Open Source Software for Numerical
Computation.
Sharma, R., Glemmestad, B., 2013. On generalized reduced gradient method with multi-
start and self-optimizing control structure for gas lift allocation optimization.
J. Process Contr. 23, 1129–1140. https://doi.org/10.1016/j.jprocont.2013.07.001.
Tasmi, T., Rahmawati, S.D., Sukarno, P., Soewono, E., 2017. Applications of line-pack
model of gas flow in intermittent gas lift injection line. J. Petrol. Sci. Eng. 157,
930–940. https://doi.org/10.1016/j.petrol.2017.08.010.
Vassiliadis, V.S., Pantelides, C.C., Sargent, R.W.H., 1994. Optimization of discrete charge
batch reactors. Comput. Chem. Eng. 18, S415–S419. https://doi.org/10.1016/0098-
1354(94)80068-5.