EVALUATION AND OPTIMIZATION OF THREE-PHASE

SEPARATOR VESSEL CONTROLLED BY PI CONTROLS

APPLYING PARTICLE SWARM OPTIMIZATION
Andr e Felipe de Azevedo Dantas

Leandro Luttiane da Silva Linhares

Jan Erik Mont Gomery Pinto

F abio Meneghetti Ugulino de Ara ujo

Andr e Laurindo Maitelli

Automation Laboratory in Oil

Federal University of Rio Grande do Norte, Natal/RN - Brazil
Abstract: The oil platforms are large structures with a large amount of control loops based on
PI controllers. Over time, even when they are tuned correctly at the beginning of its operation,
their performance deteriorates due the changes in the dynamics of the processes, making it
important to evaluate the control loop and tuning PI controllers when necessary. In this paper
we propose a combination of the CPSO and GPSO algorithms to tune PI controllers in a
primary separation system of a primary platform. The tuning and optimization are performed
through the control loop evaluation indices IAE, ISE, ITAE, ITSE and Goodhart.
Keywords: PI controller, Global optimization, Performance indices, Level control
1. INTRODUCTION
The effective separation of produced uids in the oil
extraction has been one of the main challenges in
the oil industry (Behin and Aghajari, 2008). Once
extracted from the wells, the obtained uids are trans-
ferred to a chemical separation system (Silva et al.,
2000). Even with the separation of the componen ts
cointained in the extracted oil occurring at special sta-
tions or in the self production unit it is still necessary
to process and rene the mixture from the reservoir
rock, in order to get their nal subproducts.
The high amount of water in the oil extracted is one
of the main problems faced by the oil industry. The
water needs to receive an adequated treatment to be
discarded in environment, following the specications
dened by gorvernamental regulator organs. There are
several techniques to separate water from oil, the main
applied in practice are otation, gravitational meth-
ods such as three-phase separator vessel and hydro-
cyclones, chemical treatments, methods using mem-
brane, biological treatment and junctions of these cited
techniques.
In this work, part of the separation process realized
into an oil platform is simulated using Simulink am-
bient of the mathematical software MATLAB
R
. The
simulation system has three gas-lift oil wells that
transfer the extracted oil to a three-phase separator
vessel, the rst separation stage. After this, the oil pro-
cessed in the rst stage ows through a series of three
hydrocyclones. Some important parts of this process
is controlled by six PID controllers. They were origi-
nally tuned using the classical Ziegler-Nichols tuning
method.
Due to their simple structure and high stability, PID
controllers are used to various objects and processes
widely. An extensive survey on the regulatory con-
trollers used in rening, chemicals, pulp, and paper in-
dustries reveals that 97% of them utilize PID feedback
control (Desbourough and Miller, 2002). PID control
algorithm is even used at the lowest level of hierarchi-
cal sophisticated control strategies. So the design of
PID controller is important to satisfy the need of good
robustness and near-optimal performance in industrial
processes (Xian and Minrui, 2011).
10
th
Portuguese Conference on Automatic Control
16-18 July 2012
CONTROLO2012
Funchal, Portugal
216
For an ideal control performance by the PID con-
troller, an appropriate PID parameter tuning is nec-
essary. In fact, PID parameter tuning depends on op-
erators know-how, therefore a PID parameter has
not been frequently optimal. The design of PID con-
trollers is considered as a nonlinear optimization prob-
lem (Xian and Minrui, 2011). In order to solve this
problem,many researchers introduce different intelli-
gent optimization algorithms into PID controller de-
sign, such as Genetic Algorithm (GA) (Wang and
Kwok, 1994; Behroozsaranda and Shaei, 2010), Par-
ticle Swarm Optimization (Ma et al., 2010; Mukherjee
and Ghoshal, 2007), and Simulated Annealing (Zhang
and Hu., 2009).
The PSO is an optimization technique inspired by
social behavior of bird ocking and sh schooling in
search of food. The technique was originally designed
and developed by Kennedy and Eberhart (1995). The
prominent features of PSO are its easy implementa-
tion, robustness to control parameters and computa-
tion efciency compared with other existing heuristic
algorithms such as genetic algorithm in a continuous
problem. PSO can be applied to nondifferentiable,
non-linear, huge search space problems and gives the
better results with a good efciency. Here, we presents
a combination of the algorithms GPSO and CPSO,
proposed by Baiquan et al. (2011) and Shi et al.
(2011), respectively, to tune the PID controllers, in
order to minimize classical control loops evaluation
indexes.
In This paper is proposed the control loops evaluation
using the classical methods ITE, ITSE, ITAE, IAE and
Goodhart, they will be associated with an optimizer
whose objective is to minimize the indexes values and
then it will be adjusted the parameters of the PID
controllers from the simulated process. In order to
tune the controllers, a derivation of the Particle Swarm
Optimization (PSO) algorithm will be applied.
2. SEPARATOR VESSEL PROCESS
The separator vessel may be biphasic or three-phase.
It is a widely used equipment which has large dimen-
sions and it performs the separation of the aqueous
phase, oil phase and gas phase, in a way to keep them
within a tolerance limits. Separate these uids requires
a long period of time because of the gravitational ac-
tion being relatively slow. In the three-phase separator
it is important to keep the following variables whitin
tolerable regions, the amount of uid together in the
gas, the amount of water together in the oil and oil
together in the water within tolerable regions ensuring
the quality process (Silveira, 2006).
An important element that difcult the control in the
separator vessel is the gushes. They are generated
when the uid arrives in the separator vessel, which is
formed of dispersed phases of water, oil and gas. The
evolution of the uid ow passes by ow instability
that can occur at certain ow, due an arrangement of
the set line-riser adverse, generally at low ow rates
in relatively long lines (Silveira, 2006). Besides being
used separator vessels can also be used in conjunction
hydrocyclones.
The hydrocyclone is a process of oil/water separation
most commonly used by the oil industry today, be-
cause its compact and efcient in meeting environ-
mental requirements for ejectment of water for the en-
vironment (Filgueiras, 2005). The equipment includes
cylindrical and conical portions juxtaposed, having
tangential inlets for the case where the mixed uid
inlet and a bottom outlet and an upper outlet, where
each one is used by a different uid (Junior, 2008).
Within this context its was developed a process sim-
ulator by UFS (Coordinated by Sotomayor) designed
in order to represent phenomena that occur in a three-
phase separator vessel, hydrocyclones and a gush
model for feeding the process. The system was im-
plemented on le extension .m of the Matlab
R
, inter-
acting, after being drawn with the Simulink
R
program
used for simulations.
Its shown in the Figure 1 a three-phase vessel separa-
tor and three hydrocyclones in series.
In the model proposed the separator vessel receives
the production uid addicted to the gush model. In this
part of the separation process the measured variables
are the oils level (LT), the waters level (LT) and
the pressure (PT) which the gas exerts in the vessel.
The variables related to the oil and the pressure in the
vessel are controlled by the valves S
l
and S
g
, respec-
tively, which are directly connected to the separator
vessel. The waters level is controlled in parallel with
the separation which is performed in hydrocyclones,
through the valves S
o1
, S
o2
and S
u3
. Finally, the valve
S
o3
controls the differencial pressure (P) in the se-
ries of hydrocyclones which is important to control
the hydrocyclones separation efciency. Summaryly,
there are four setpoints to control: oils level (L
l
),
waters level (L
w
), pressure in the separator vessel (P)
and differential pressure (P) in hydrocyclone. And
so does manipulated variables that are six: S
g
(which
controls P), S
l
(which controls L
l
), S
o1
, S
o2
and S
u3
(which controls the L
w
) and S
o3
(which controls P).
3. PARTICLE SWARM OPTIMIZATION
The Particle Swarm Optimization (PSO) algorithmm,
originally introduced by (Kennedy and Eberhart, 1995),
has become one of the most important swarm intelli-
gence-based algorithms. Due to its simple implemen-
tation, minimum mathematical processing and good
optimization capability, PSO has attracted more at-
tentions. As Vassiliadis and Dounias (2009) summa-
rized, PSO may be considered the most popularly and
frequently applied among nature-inspired techniques
according to related publications.
217
Fig. 1. Three-phase separator vessel with three hydro-
cyclones in serie
PSO simulates the behavior of a bird ock. When a
ock of birds forage for food, two simple and im-
portant strategies are: (a) searching for the peripheral
region around the bird that is nearest to the food; (b)
judging the position of the food by its own ying expe-
rience. Inspired by the two strategies, the search space
of optimization problem is regarded as birds

flying
space; every bird is abstracted as a particle with non-
quality and non-volume so as to denote a candidate so-
lution; and the optimal solution to be searched for the
problem is the food to all the particles. Then, the basic
PSO algorithm comprises a swarmof particles moving
in the D-dimensional search space which includes all
possible candidate solutions (Shi et al., 2011).
We denote the ith particle as X
i
= (x
i,1
, x
i,2
, , x
i,d
),
and its ying velocity as V
i
= (v
i,1
, v
i,2
, , v
i,d
).
Depending on two strategies above, when each parti-
cle ies in the search space, P
i
= (p
i,1
, p
i,2
, , p
i,d
)
denotes the personal best position the ith particle has
found so far, and P
g
= (p
g,1
, p
g,2
, , p
g,d
) is the
global best position discovered by the swarm. At each
time step t, both of each particles velocity and po-
sition are updated so that a particle moves to a new
position. The following two equations are employed
to calculate the velocity and position:
V
t+1
i
= V
t
i
+c
1
r
1

P
t
i

X
t
i
+c
2
r
2

P
t
g
X
t
i

(1)
X
t+1
i
= X
t
i
+V
t+1
i
(2)
where c
1
and c
2
are two positive constants (acceler-
ation constants), r
1
and r
2
are two uniform random
numbers in [0, 1]. A constant V
max
is often used to
limit each particles velocity to guarantee that most
of the new positions of the particles will be restricted
in the search space of feasible region at each itera-
tion. Equations 1 and 2 lead to the movement of each
particle to wards its cognitive best position (P
i
) and
social best position (P
g
) with random perturbations
caused by c
1
r
1
and c
2
r
2
, respectively. The quality of
a particles position is measured by its tness value,
namely, the better tness value means the better posi-
tion.
Shi and Eberhart (1998) modied the original PSO
by introducing an inertia weight () to Equation 1
to balance exploitation and exploration. The modied
Equation 1 is as follows:
V
t+1
i
= V
t
i
+c
1
r
1

P
t
i

X
t
i
+c
2
r
2

P
t
g
X
t
i

(3)
According to the numerical tests in (Shi and Eber-
hart, 1998; Shi and Eberhart, 1999), the adoption of
improves the performance of PSO largely, so the com-
bination of Equations 3 and 2 is always considered as
the standard PSO by many researchers.
The advantage of PSO algorithm is quick in the con-
vergence speed, few in the parameters, simple in op-
erations and etc., therefore it is suitable to solve op-
timization problems. The PSO algorithm, however, is
easy to appear premature convergence, especially in
multimodal cases. To overcome these defects, many
scholars proposed the improvement methods on PSO
algorithmThose improvement methods can be sorted
to two kinds. The rst kind is to combine PSO with
other optimization algorithms. The second kind is to
improve the structure of PSO algorithm (speed itera-
tion formula and position iteration formula), specially
speed iteration formula (Baiquan et al., 2012).
In this paper we combine the CPSOouter proposed
by Shi et al. (2011) with the improved speed formula
of the GPSO proposed by Baiquan et al. (2011).
The cellular particle swarm optimization (CPSO) is a
hybridization of cellular automata (CA) and particle
swarm optimization (PSO) for function optimization.
In order to enhance the performance of the interac-
tion, the CPSO employs the idea of CA to explore
the communication structure and information inherit-
ing and diffusing mechanisms of the swarm system
of PSO. So in CPSO, we consider PSO with a CA
model, where individuals can only exchange informa-
tion within their neighborhood. It could help exploring
the search space due to the slow information diffusion
through the population, promoting the preservation of
diversity, and exploiting every cells local information
inside the neighborhood (Shi et al., 2011).
The ith particles position X
t
i
is dened as the cell
state of this particle. When implementing CPSO-
outer, only smart-cells participate in updating process.
A smart-cell constructs its neighborhood by a neigh-
borhood function given in Equation 4.
218
N(i) =

X
t
i
+A(i)R
3
V
t
i
iffit(X
t
i
) = fit(P
t
g
) 0
X
t
i
+B(i)R
3
V
t
i
iffit(X
t
i
) = fit(P
t
g
) < 0
X
t
i
+C(i)R
3
V
t
i
iffit(X
t
i
) = fit(P
t
g
) 0
X
t
i
+C(i)R
3
V
t
i
iffit(X
t
i
) = fit(P
t
g
) < 0
(4)
A(i) =
fit(P
t
g
fit(X
t
i
)
B(i) = |
fit(X
t
i
fit(P
t
g
)
|
C(i) = (
exp fit(P
t
g
)
exp fit(X
t
i
)
2
)
(5)
where R
3
is a 1 d matrix composed by d uniform
random numbers in [1, 1], and is the operation
symbol of Hadamard product. The transition rule of
CA in CPSO-outer is designed as follows:
f() = min(fit(N(i),
fit(N(i +
i
)),
, fit(N(i +
m
),
, fit(N(i +
l
)))
(6)
where
=

i x iff() = fit(N(i))
i +
m
f() = fit(N(i +
m
))
(7)
S
t+1
i
= S
t

(8)
Equation 6 means that l neighbors of the ith particle
generated by Equation 8 are evaluated, and the neigh-
bor with the best tness value is chosen to replace
the ith particle. The transition rule could give particles
the power to make a wise jump, and it could help ex-
ploring the search space in a local competition neigh-
borhood and enhance the diversity of the swarm. So
CPSO-outer could have greater potential to search for
global optimum in the search space (Shi et al., 2011).
In the algorithm applied in this work the standard
position update formula was used. To each particle,
neighbors were generated using the CPSO-concept.
The particle tness is compared with the its tness
neighbors an replaced if a better neighbor is found.
To update the particle velocity, the improved velocity
formula proposed by GPSO was used. It is described
as in Equations 9, 10 and 11.
V
t+1
i
=

V
t
i
+c
1
r
1
(P
i
X
t
i
)
+c
2
r
2
(P
g
X
t
i
)
(9)

V
t+1
i
= k
0
V
t+1
i
+r
3
(

V
t
i
V
t+1
i
) (10)
X
t+1
i
= X
t
i
+

V
t+1
i
(11)
where adjustment coefcient k
0
is a constant in the
range [0, 1], r
3
is a random number with uniform
distribution on the range [0, 1], the other parameters
are the same with the standard PSO.
4. METHODOLOGY AND RESULTS
In this paper, it is intended to analyze the optimization
method efciency tuning using CGPSO in order to
compare it to the tuning performed with both, the
step response method used in the work of Teixeira
(2010) and the CGPSO. Evaluation performance in-
dexes were implemented to analise the system. The
implemented algorithms are ITE, ITA, ISE, ITSE and
Goodhart. Thus, the sum of these functions is used
in the optimization process being minimized by the
CGPSO algorithm.
First of all, there were developed the algorithms to
evaluate the meshes (using ITE, ITA, ISE, ITSE and
Goodhart). These algorithms have been added to the
model (with three-phase separators) in a way to be-
came possible to evaluate each individual mesh, to-
talizing 6 mesh evaluated. Afterward, PI controllers
were designed following the methodology of Teixeira
(2010) and it was reproduced the same result at tuning
the PI controllers. In other words, the PI controllers
are designed using the Ziegler-Nichols step response
method and the indexes of the model using these
controllers were compared with the controllers of the
indexes obtained after the optimization. The tuned pa-
rameters comparison may be viewed in the Table 1.
Table 1. Tuned parameters comparison
PIs Tuning PSO Original Tuning
P I P I
PI (S
u
) 270,68 -19,18 366.22 -0.51
PI (S
g
) 1039,31 -132,73 366.22 -2.77
PI (S
o3
) 20,71 -0,37 16.66 -0.10
PI (S
o2
) 3,06 0,32 8 0.05
PI (S
o1
) 472,20 -103,13 366.22 -2.77
Pi (S
l
) 493,50 -4,61 366.22 -2.77
Then, the control loops were evaluated in a global way,
with the methods listed in order to obtain the following
results in Table 2.
Table 2. Control loops evaluation for origi-
nal and optimized PI parameters
Evaluation Methods Original Optimized
iae 0,0545 0,0409
ise 0,0099 0,0075
itae 769,2202 528,5356
itse 135,2689 92,8278
goodhart 0,0523 0,0500
In order to achieve the shown results the PSO algo-
rithms displayed in the PSO section was implemented
from particles near the original tuning. So, it was con-
sidered in this algorithm that the initial particles would
be at a point around the original tuning shown and so
the optimization starts. To perform the optimization
was also assumed that the system would be acting
as regulatory, at a same setpoints shown in Table 3.
In addiction, the system was performed at a different
operating point in order to analyze the behavior of the
system.
219
Table 3. Setpoint values of the controlled
variables
Output Setpoint 1 Setpoint 2
L
l
0, 5 0, 6
L
w
0, 5 0, 6
P 9, 4806 9, 5806
P 1, 25 1, 35
Concluding the results, a gure showing the difference
between the graphics optimized system and the not
optimized system from the perspective of output and
control signal can be viewed in Figures 2 and 3.
0 5000 10000 15000
0
0.2
0.4
0.6
0.8
1
Oil Out
Simulation Time (s)
L
e
v
e
l

(
m
)
0 5000 10000 15000
0.2
0.4
0.6
0.8
1
Water Out
Simulation Time (s)
L
e
v
e
l

(
m
)
0 5000 10000 15000
9.2
9.4
9.6
9.8
Pressure Out
Simulation Time (s)
P
r
e
s
s
u
r
e

(
k
g
f
/
c
m
2
)
0 5000 10000 15000
0
0.5
1
1.5
2
2.5
Hidrocyclone Diferential Pressure Out
Simulation Time (s)
D
i
f
e
r
e
n
t
i
a
l

P
r
e
s
s
u
r
e

(
d
P
0
/
d
P
u
)


Not Optimized
Optimized
Fig. 2. Comparing the outputs of the optimized con-
trols and the original tuning
In the Figures 2 and 3 displayed, simulation was per-
formed for the system regulation, as well as changes
were made in the references around the operating
point, extrapolating in the practical the limits of re-
tuning.
The Figures 2 and 3 show that optimization performed
is compatible with the reduction of the indexes shown,
which meant an improvement in the regulatory system
and besides the improvement of the regulatory system
was also perceptible that for a small variation in the
reference, the system still tends to have consistent
performance for optimization performed.
5. CONCLUSIONS
A method of controllers PIs tuning was shown, using
like loss function standard methods of evaluation of
0 5000 10000 15000
0
0.5
1
Vlve Sl
Simulation Time (s)
%
0 5000 10000 15000
0
0.5
1
Vlve Su
Simulation Time (s)
%
0 5000 10000 15000
0.2
0.3
0.4
Vlve Sg
Simulation Time (s)
%
0 5000 10000 15000
0
0.5
1
Vlve So1
Simulation Time (s)
%
0 5000 10000 15000
0.2
0.3
0.4
Vlve So2
Simulation Time (s)
%


0 5000 10000 15000
0
0.5
1
Vlve So3
Simulation Time (s)
%

Not Optimized
Optimized
Fig. 3. Comparing the signal controls of the optimized
controls and the original tuning
meshes in the literature. For the three-phase separa-
tion system were compared with the simulated tuning
presented in a literature and the optimized with CPSO
and GPSO. It was possible to see that using as input
the evaluation methods mentioned, there was a signif-
icant improvement in the tuning shown, representing
a regulatory less sensitive to external disturbances as
well as allowing small changes in setpoints, although
it has been tuned to optimal mode regulatory.
6. ACKNOWLEGMENTS
The authors would like to thank Petrobras, FINEPI
and Coordination for the Improvement of Higher
Level or Education Personnel (CAPES) for the nan-
cial support.
7. REFERENCES
Baiquan, L., G. Gaiqin and L. Zeyu (2011). The block
diagram method for designing the particle swarm
optimization algorithm. Journal of Global Opti-
mization 52(4), 689710.
Baiquan, L., G. Gaiqin and L. Zeyu (2012). The block
diagram method for designing the particle swarm
optimization algorithm. Journal of Global Opti-
mization 52(4), 689710.
Behin, J. and M. Aghajari (2008). Inuence of water
level on oil-water separation by residence time
220
distribution curves investigations, separation and
purication technology. Separation and purica-
tion technology 64, 4855.
Behroozsaranda, A. and S. Shaei (2010). Optimal
control of amine plant using non-dominated sort-
ing genetic algorithm-ii. Journal of Nature Gas
and Engineering 2(6), 284292.
Desbourough, L. and R. Miller (2002). Increasing
customer value of industrial control performance
monitoring-honeywells experience. AIChE Sym-
posium Series 98(326), 172191.
Filgueiras, N.G.T. (2005). Modelagem, an alise e con-
trole de um processo de separac ao oleo/ agua.
Masters thesis. UFRJ. Rio de Janeiro - Brazil.
Junior, C. A. Corr ea (2008). Desenvolvimento de
modelo computacional de previs ao de quebra de
gotas em simulador de separac ao de oleo e agua
em um hidrociclone. Masters thesis. UERJ. Rio
de Janeiro.
Kennedy, J. and R.C. Eberhart (1995). Particle
swarm optimization. In: Proceedings of IEEE
International Conference on Neural Networks.
pp. 19421948.
Ma, C. X., L. Qian, L. Wang, M. I. Menhas and M. R.
Fei (2010). Determination of the pid controller
parameters by modied binary particle swarm
optimization algorithm. In: Proceedings of the
22th Chinese Control and Decision Conference.
pp. 26892694.
Mukherjee, V. and S. P. Ghoshal (2007). Intelligent
particle swarm optimized fuzzy pid controller
for avr system. Electric Power Systems Research
77(12), 16891698.
Shi, Y. and R.C. Eberhart (1998). A modied particle
swarm optimizer. In: Proceedings of IEEE Inter-
national Conference on Evolutionary Computa-
tion. pp. 6673. Anchorage, Alaska.
Shi, Y. and R.C. Eberhart (1999). Empirical study of
particle swarm optimization. In: Proceedings of
IEEE International Conference on Evolutionary
Computation. pp. 19451950.
Shi, Y., H. Liu, L. Gao and G. Zhang (2011). Cellu-
lar particle swarm optimization. Information Sci-
ences, Special Issue on Interpretable Fuzzy Sys-
tems 181(20), 44604493.
Silva, C.B.C., M.J.B. Filho and J.A. Pinheiro (2000).
Medic ao de vaz ao e propriedades em escoa-
mento multif asico: Soluc ao econ omica para
diferentes atividades industriais. In: Boletim
T ecnico PETROBRAS. Vol. 43. Petrobr

As. Rio de
Janeiro. pp. 4561.
Silveira, M.A.C.R. (2006). Controle de um processo
de tratamento prim ario de petr oleo. Masters the-
sis. UFRJ. Rio de Janeiro - Brazil.
Teixeira, A.I.T. (2010). Controle de oscilac oes em
sistemas de elevac ao articial de petr oleo por
injec ao contnua de g as. Masters thesis. UFS.
S ao Cristov ao - Brazil.
Vassiliadis, V. and G. Dounias (2009). Nature-inspired
intelligence: a review of selected methods and
applications. International Journal on Articial
Intelligence Tools 18(4), 487516.
Wang, P. and D. P. Kwok (1994). Optimal design of
pid process controllers based on genetic algo-
rithms. Control Engineering Practice 2(5), 641
648.
Xian, C. and F. Minrui (2011). A multiple tuning
method for pid controllers using constrained par-
ticle swarm optimization. In: Proceedings of the
30th Chinese Control Conference. Yantai, China.
pp. 37413746.
Zhang, Y. C. and Y. M. Hu. (2009). On pid controllers
based on simulated annealing algorithm. In: Pro-
ceedings of the 27th Chinese Control Conference.
pp. 398404.
221