Preprints of the 8th IFAC Symposium on Advanced Control of Chemical Processes

The International Federation of Automatic Control

Furama Riverfront, Singapore, July 10-13, 2012

Disturbance Estimator based NonLinear

MPC of a Three Phase Separator
Mendes P. R. C. Normey-Rico J. E. Plucenio A.
Carvalho R. L.

Federal University of Santa Catarina, Florian

opolis, Brasil, (e-mail:
paulorcm@hotmail.com, jnormeyrico@gmail.com,
rodrigoleao.carvalho@gmail.com, plucenio@das.ufsc.br).

Abstract: This paper presents a nonlinear predictive controller with a feed-forward action
applied to a three-phase separator used in the petroleum industry. Two tools are used in order
to include the feed-forward action in the NMPC, a disturbance estimator and a disturbance
predictor. Moreover, a supervisory control changes the objective function characteristics of the
NMPC depending on the type of disturbances acting on the process. The separator advanced
controller, which is capable to maintain all process variables in pre-defined zones, uses a practical
nonlinear model predictive controller (PNMPC) developed in the Department of Automation
and Systems (Federal University of Santa Catarina). The performance of the proposed strategy
is compared to a traditional PI Zone controller used in industry and to a different NMPC
based on a Hammerstein model. All the case study simulations are developed with a complete
phenomenological nonlinear model of the separator. The obtained comparative results show that
the proposed strategy gives the best results, allowing good oscillation attenuation under slug
perturbations.
Keywords: Three Phase Separators, Predictive Controllers, Zone control, Process Control
1. INTRODUCTION
Most of the oil wells produce fluid consisting of a mixture
of gas, oil and water, Triggia et al. (2001). In a conventional
system of oil separation there is usually a group of gravity
separators connected in series starting with a three-phase
separator. The three-phase separator is basically designed
to separate the multiphase fluid and also to damp the
load oscillation that comes from well. A proper control
scheme of this process allow to maximize oil production
and to minimize the quantity of residual oil in water,
therefore, improving energy production and minimizing
ambient impact.
Horizontal gravity separators, as the one considered in
this work, are normally used in oil industry at the first
stage of the fluid treatment with two principal objectives:
separation of the different phases and absorbing load fluctuations. According to Nunes (1994) oil wells and risers
may produce with different flow regimes. In certain conditions a regime known as slug flow is established which
is characterized by a flow-rate alternating between high
and low values. This flowing regime is detrimental to the
three phase separation process because of the pressure and
flow variation into the separator. Thus separator control
should be designed to allow good efficiency of separation
and still manage to filter out disturbances caused by slug
flow, sending a more stable flow for downstream processes.
Several control strategies have been used to control this
process. For example, Filgueiras (2005) presents a separation vessel controller that has three SISO PI compensators,
one on each fluid phase. This classical control scheme, although quite common in industry, is not able to adequately
IFAC, 2012. All rights reserved.

101

damp load oscillations. To improve the performance of

the classical system against oscillatory disturbances, a PI
based zone control strategy is presented in Nunes (2004).
In this strategy the water level is controlled allowing
slow variations in a pre-defined zone using two different
PI tuning settings. This strategy is being used in Petrobras Platforms with satisfactory results. Advanced control
strategies based on Model Predictive Control (MPC) have
also been studied, using linear algorithms as in Nunes
(2001) and Silveira (2006) or applying non-linear MPC
(NMPC) strategies as in Mendes et al. (2011). The advantages of MPC strategies are the flexibility in the definition
of the objective function and the possibility of considering
process constraints (Camacho and Bordons (2007)).
The study presented in Mendes et al. (2011) compares
different MPC strategies to analyze the effect on disturbance attenuation, considering hard and soft constraints
(Hovland (2004); Prasath and Jrgensen (2009)) and also
zone control (Maciejowski (2002)) in order to maintain
the controlled variables within specified limits or zones.
For the implementation, a Hammerstein-based nonlinear
predictive control was used, based on the ideas of the practical NMPC (PNMPC) algorithm proposed in Plucenio
et al. (2007) and Plucenio (2010) which allows solving the
nonlinear problem with a simple QP algorithm, therefore
allowing for a better compromise between performance and
computational cost than a linear MPC.
Although the Hammerstein zone-PNMPC presented the
best results in a hard slug flow disturbance scenario, some
improvements in the control algorithm can be performed
and are proposed in this work. The first modification consists in the use of a simplified phenomenological model of

8th IFAC Symposium on Advanced Control of Chemical Processes

Furama Riverfront, Singapore, July 10-13, 2012

Win

Sg
Gin
Lin
Win

Zone
SUPERVISORY

PNMPC

Wout_sp
Lout_sp
Gout_sp

PI

sw
sl
sg

Lin

Gin

SEPARATOR

Vwflcs

Lweir

hT
Vlfwcs

Vwflcl

hW

hl

DETECTOR
Sw

Wout h
w
Lout h
l
Gout p

Disturbance
rejection
Win_est
Lin_est
Gin_est

ht

DISTURBANCE
ESTIMATOR

Sl

Fig. 2. Proposed Control Scheme

Wout

Lout

Fig. 1. Scheme of the horizontal three-phase separator

the process to compute the predictions instead of the Hammerstein model, simplifying the model construction. The
second one includes a feed-forward action in the PNMPC
strategy based on an estimation of the flow disturbances.
For this purpose a disturbance estimator and a disturbance
predictor are developed. The third and final improvement
consist of the use of two different control objectives in the
PNMPC and a supervisory strategy based on a slug flow
disturbance detector. Although model predictive control
including disturbance estimator feedforward action have
been used in other cases as greenhouse production process (Pawlowski et al. (2011)) and olive oil Mill process
(Bordons and Cueli (2004)), it has not been applied in
separation processes of oil industry.
The rest of the paper is organized as follows. In section 2
a brief process description is given. Section 3 is devoted
to present the complete controller structure, including the
supervisory system, the NMPC strategy, the predictor
model and the disturbance estimator-predictor. Simulation
case studies are analyzed in section 4, where the proposed
strategy is compared to two other controllers. The paper
ends with the conclusions.
2. THREE-PHASE SEPARATOR PROCESS AND
CONTROL OBJECTIVES
As pointed out three-phase separators are designed to
separate the multiphase fluid and also to damp the load
oscillation that comes from well. There are a lot of separator types; each one having different shape and internal devices. The horizontal separator scheme presented in Fig. 1
represents one of the more common units used in industry.
As can be seen in the figure, the input flow of water, oil
and gas (Win , Lin , Gin ) comes into the separator and the
liquid phase goes to the separation chamber, where a set of
parallel plates are installed in order to help in the water-oil
separation. The oil passes to the second chamber (the oil
chamber) and a valve (Sl ) is used to extract it form the
vessel. At the same time, valves Sg and Sw are respectively
used to control the extraction of gas and water. Several
factors affect the separator efficiency and also the fluid
behavior inside the tank (Nunes (1994); Filgueiras (2005)).
The process and model considered here does not take into
account thermal effects, emulsifiers and foam in the oil and
only the effect of parallel plates are considered. Therefore,
the control of the separation is done using Sl , Sg and Sw
102

as manipulated variables to control, respectively, the level

of oil phase (hl ), the vessel pressure (p) and the level
of water phase at separation chamber (hw ). In normal
operation, when the input flow varies slowly and has small
oscillations, the desired separation of the three phases is
obtained maintaining the three controlled variables at the
operating point. However, when high amplitude input flow
oscillations affect the process, the vessel control has to be
used to avoid the transmission of these flow oscillations to
the output, mainly in the water flow Wout , which strongly
affects the efficiency of the downstream processes.
3. PROPOSED CONTROL STRATEGY
Figure 2 shows the proposed control scheme based on
the PNMPC and disturbance estimator-predictor strategy
used to obtain a flexible and efficient control strategy of the
separator. As can be seen in the figure the controller has
three levels. In the local level, three slave PID controllers
are used to manipulate the three valves and maintain
the output flows of water, oil and gas at the set-point
values defined by the master PNMPC. The PNMPC, that
operates at the second level, minimizes a defined objective
function and uses both, feedback and feedforward actions.
For the feedforward action, estimated values of the disturbances computed by the disturbance-estimator block are
used. Finally, the objective function to be passed for the
PNMPC is computed at the third level, which decides if
the controller should work with a narrow or large zone in
the control specifications. This supervisor level uses a slug
flow disturbance detector to define the operation mode.
The different parts of the proposed strategy are detailed
in the following sub-sections.
3.1 Supervisory Control
Depending on the well or riser flowing regime, the fluid flow
entering the separator may be steady or oscillatory. For a
steady fluid inflow the best control option is to require
the process variables to be maintained into a narrow band
around a fixed set-point. For an oscillatory fluid inflow an
overall better control result is to allow for a bigger oscillation of the controlled variables around the operating point,
in such a way that the vessel acts as a filter for the disturbances. The disturbance detector is an important tool for
the implementation of the control strategy. Therefore, the
oscillation range where the controller variables are allowed
to vary, go from a maximum value to a minimum value
depending on the amplitude of the inflow oscillations. This

8th IFAC Symposium on Advanced Control of Chemical Processes

Furama Riverfront, Singapore, July 10-13, 2012

strategy ensures the best control performance regardless of

the fluid inflow regime.
A simple method to detect an oscillatory fluid inflow
disturbance consist of the analysis of a set of data in a
moving time window (Kim et al. (2008)). In this window a
number of Nw sampled data are used to compute the signal
variance. If this variance is smaller than a pre-defined value
(in this case 0, 01 was used) the fluid inflow is considered
steady and the level range is set to the minimum value. On
the contrary, when the variance is greater than the predefined value, a oscillatory fluid inflow is assumed and the
level range is set to the maximal value. The maximum and
minimum values of the zone are then computed adding and
subtracting the value of the range to the nominal set-point
(in this case 0.5meters). For the separator used in this
work a 250 sampled data window is used with a sampling
time of 10 seconds.
3.2 PNMPC Control Strategy
The basic algorithm of the PNMPC proposed in Plucenio
et al. (2007) uses the ideas of the linear GPC. Linear
predictive controllers use the representation of the predictions vector Y along the horizon N , as a function of
the incremental control action vector u. This allows to
transform the objective function in a quadratic function
of u and to solve the optimization problem with simple
quadratic programming (QP) algorithms. Although conceptually this idea is still valid for nonlinear models the
obtained optimization problem is not QP type and demands much more computation time. However, for systems
represented by nonlinear models, where the superposition
principle cannot be applied, an approximation of Y can be
obtained which allows solving the nonlinear problem with
a QP algorithm (Plucenio (2010)). This technique differs
from other proposed solutions for NMPC mainly because
the linearized models are independent of the system equilibrium points.
In PNMPC it is assumed that the prediction vector Y can
be written according to (1), where GP N M P C matrix is the
Jacobian of Y and F is the free response.
= F + GP N M P C u
Y
(1)
This representation is exact for systems represented by linear models and a good approximation for those represented
by non-linear models, provided that the states are continuous and differentiable in relation to the inputs. In PNMPC,
vector F and matrix GP N M P C are obtained numerically.
For that, an algorithm can be executed to calculate the
vector with N predictions when given the values of past
inputs and outputs and the vector with Nu increments
of the future input u. Moreover, for the treatment of
modeling error a prediction correction mechanism is used,
which consists to add a correction factor to each prediction
(see Plucenio (2010) for the details).
As pointed out, when the process is affected by slug flow
disturbances the main objective of the control system is
to attenuate the effect of these disturbances in the water
output flow. In this case the exact value of the water
level (the controlled variable) is not important, provided
it remains within specified limits or zones. This means
that the outputs will only be treated as controlled variables
when these prediction values are outside the limits defined
103

for the corresponding zone. In order to incorporate the

zone control to PNMPC, and to also allow the possibility
of controlling the process variables at a desired set-point,
the objective function is expressed as follows:
J=

N
X


y (t + j|t) y b (t + j|t) 2 +
R

(2)

j=1

Nu
X

ku (t + j 1)kQ

j=1

Thus, the zone control strategy is implemented as follows

(Sotomayor et al. (2009)): the variable y b is the set point
of the system output. The prediction at time t + j of each
output i (
yi (t + j|t)) is computed and:
If yi,min yi (t + j) yi,max , the output yi should
be ignored in J (released or removed from the control
calculations) at the moment. Therefore, the parameter of R matrix corresponding to this output should
be set to zero.
If yi (t + j) > yi,max , the output yi should be brought
to its upper limit. Therefore, we do, yib (t+j) = yi,max .
The parameter of R matrix corresponding to this
output is a controller tuning parameter.
If yi (t + j) < yi,min , the output yi should be brought
to its lower limit. Therefore, we do, yib (t + j) = yi,min
and the parameter of R matrix corresponding to this
output is the same as in the previous case.
In the separator, the set-point of the pressure uses always
a zero width band. For the water and oil levels the same
narrow band of 0.0m is used around the normal operation
point in the case of steady disturbances and a band of 0.2m
is considered for the case of oscillatory ones (yi,max = 0.7m
and yi,min = 0.3m).
3.3 Nonlinear Prediction Model and Disturbance Estimator
To compute the output predictions a simple phenomenological model is used in the PNMPC which considers only
a mass balance of the different fluids, therefore disconsidering the unknown characteristics of the fluid inflow.
Therefore, this simple model is described by equations (3),
(4), (5) and (6).
dht
Win + Lin Lweir Wout
p
=
(3)
dt
2 Ccs ht (D ht )
Lweir Lout
dhl
p
=
dt
2 Ccl hl (D hl )
Win Wout
dhw
p
=
dt
2 Ccs hw (D hw )

(4)

(5)

dp
(Gin + Win + Lin Gout Wout Lout ) p
=
. (6)
dt
Vt Vcl Vcs
As can be seen, this model needs the values of the disturbances Gin , Win , Lin that are not measurable. Therefore,
to compute the output predictions, an estimated value of
these variables is needed. Thus, using equations (3), (4),
(5) and (6) it is possible to compute the estimated values
as follows:
i
h
p
cin = Ccs 2 (D hw ) hw dhw + Wout
(7)
W
dt

8th IFAC Symposium on Advanced Control of Chemical Processes

Furama Riverfront, Singapore, July 10-13, 2012

i
b weir + Ccs 2 (D ht ) ht dht
=L
dt
i dh
h
p
w
Ccs 2 (D hw ) hw
dt


hw

0.6
0.4
0.2

time (seconds)

(9)

7
4

x 10

1
0.8
0.6
0.4
0.2

time (seconds)

7
4

x 10

13

12.5

dp c
Vt Vcs Vcl
b in

Win L
(10)
p
dt
+Wout + Lout + Gout
In the controller implementation, discrete equivalent equations are used, where the derivatives are computed using
Euler approximation.
b in =
G

1
0.8

b in
L

(8)

hl

h
i
p
b weir = Ccl 2 (D hl ) hl dhl + Lout
L
dt

3.4 Feedforward action and disturbance predictor

With the previous estimator equations it is possible to
consider a feedforward action in the controller. As the
estimator gives the value of the disturbances at time t, only
a simple feedforward action can be implemented. However,
if the future values of the disturbances were available,
the predictive control law could take advantages of this
information in the computation of the future sequence of
control actions. Note that the prediction equation is now
computed using the vector of future values of the control
actions u and disturbances q:
= f (u, q)
Y
(11)
and the linearized relation between the future values of
the inputs and the prediction vector is computed using
the same procedure explained before.
Clearly, the future values of the disturbances q are not
available but an estimated vector
q can be computed
using past information given by the estimator. Several approaches can be used to obtain the prediction of the future
disturbance vector. Temporal series are used in Pawlowski
et al. (2010) and Reikard (2009), autocorrelated models
are considered in Bordons and Cueli (2004) while Paoli
et al. (2009) presents and algorithm based on artificial
neural nets. In this work a simple linear autoregressive
model shown in equation (12) is proposed which uses na
past values of the estimated disturbances.
(12)
qi (k) = a1 qi (k 1)... anai qi (k na)
The predictor parameters are obtained off-line, for each
one of the three input flows, using a least square method
with a sampling time of 10 seconds and considering a
set of feeding flows obtained through a simulation in the
software Olga (Scandpower (2003)). na was selected using
the period of the disturbance slug flow (na = 250).
4. RESULTS
This section presents several simulation results obtained
using the proposed control strategy and a complete simulation phenomenological model of a separator developed
by Filgueiras (2005). Feeding flows used in control tests
were obtained through a simulation in the software Olga
(Scandpower (2003)).
To illustrate the advantages of the proposed solution, some
104

12

11.5

time (seconds)

7
4

x 10

Fig. 3. Process Variables Under PI Zone Control

simulations are also presented using the normal control
structure used in industry, a PI Zone controller. Moreover,
using some performance indices we compare the obtained
results with those presented in Mendes et al. (2011) where
a PNMPC algorithm is used based on a Hammerstein
model but without considering the disturbance estimatorpredictor and the supervisory control. As explained, three
local flow rate PI controllers are used in the low level of
the proposed control structure.
The sampling period used to simulate the process and the
local controllers was 1 second. The PI Zone, the Hammerstein PNMPC and the second level of the proposed
PNMPC were simulated with a sampling period of 10
seconds. The simulation test has a duration of 38, 88 hours;
in the first half of the time the separator was controlled
with the PI Zone controller, and from the time 19, 44 hours
to the end, the proposed control strategy was used. To
obtain a soft switch between the two control strategies,
from 13, 88 and 25, 00 hours, a steady inflow was applied
instead of the slug flow disturbance. Also at time 16, 66
and 22, 22 hours a step on the input flow rate has been
applied.
The PI Zone Controller uses two different tuning of the
proportional Kc and integral Ti actions. When the process
variable is inside the defined zone, a slow control action
is applied (big value of Ti and small value of Kc ). On the
contrary, when the variable goes outside the zone a more
aggressive tuning is used (big value of Kc and small value
of Ti ).
As mentioned, in the first part of the test the PI Zone controller was used. The controller tuning parameters inside
and out of the zone are presented in Table 1. Only the oil
and water levels were controlled by zones, the pressure was
controlled with a fixed set point and an aggressive tuning
to keep it constant.
Table 1. Zone PI - tuning parameters
Variables
hw
hl
p

Inside Zone
Kc
Ti
0, 001
125
0, 01
125
0, 03
50

Out of Zone
Kc
Ti
0, 1
50
0, 02
50
0, 03
50

Figure 3 shows the obtained responses of process variables

with their respective zones and set points. Figure 4 shows
the input and output flow rates.
The proposed controller is then used in the second part of
the test using the tuning parameters presented in table 2.
Zone control was applied only to the level of water and oil,
and a fixed set point was used for pressure. The zone values

Win e Wout

8th IFAC Symposium on Advanced Control of Chemical Processes

Furama Riverfront, Singapore, July 10-13, 2012

tests the input and output gas flow rate are the same,
because the control objective was to maintain a constant
pressure inside the separator. On the other hand, for the
oil and water flow rates different results were obtained with
each controller.

0.014
0.0135
0.013
0.0125

Win
Wout

0.012
0.0115
0

time (seconds)

7
4

x 10

Lin e Lout

0.022
0.02
0.018

Table 3. Performance Indices

0.016
Lin
Lout

0.014
0.012
0

time (seconds)

7
4

x 10

Controller

Gin e Gout

0.14
0.13

Flow Rate

Oscillation (%)

Variance

Win
Lin
Gin
Wout
Lout
Gout
Wout
Lout
Gout
Wout
Lout
Gout

20, 45
14, 97
26, 29
6, 06
63, 69
26, 29
3, 03
27, 97
26, 29
3, 03
29, 17
26, 29

2, 3366 108
1, 5443 106
1, 5589 104
3, 0692 109
2, 4813 106
1, 3945 104
1, 7054 108
1, 5958 106
0, 0019

0.12
0.11

Gin
Gout

0.1
0.09

time (seconds)

7
4

x 10

Fig. 4. Input and Output Flow Rates Under PI Zone

Control

PI Zone

Proposed

hw

0.8
0.6
0.4
0.2

0.8

0.9

1.1

1.2

1.3

time (seconds)

Zone PNMPC

1.4
5

x 10

Hammerstein

hl

0.8
0.6
0.4
0.2

0.8

0.9

1.1

1.2

1.3

time (seconds)

Table 4. Performance Indices

1.4
5

x 10

12.04

12.02

Controller

12

Flow Rate

Variance

Wout
Lout
Gout
Wout
Lout
Gout

6, 6002 108
1, 7689 107
2, 4220 105
2, 1663 108
2, 3255 108
2, 4138 105

11.98
11.96

0.8

0.9

1.1

1.2

1.3

time (seconds)

1.4
5

x 10

PI Zone

Fig. 5. Process Variables Under Proposed Controller

Win e Wout

Zone PNMPC
0.014
0.0135
0.013
0.0125
0.012

Win
Wout

0.0115
0.8

0.9

1.1

1.2

1.3

time (seconds)

1.4
5

x 10

Lin e Lout

0.022
0.02
0.018
0.016
0.014
0.012

Lin
Lout
0.8

0.9

1.1

1.2

1.3

time (seconds)

1.4
5

x 10

Gin e Gout

0.14
0.13
0.12
0.11
0.1
0.09

Gin
Gout
0.8

0.9

1.1

1.2

time (seconds)

1.3

1.4
5

x 10

Fig. 6. Input and Output Flow Rates Under Proposed

Controller
are the same as the ones used in the PI Zone Controller
case.
Table 2. Parameters of Proposed Zone PNMPC
Parameter
N
Nu
R
Q
a

Value
[20; 20; 20]
[5; 5; 5]
[20; 3; 20]
[10000; 1000; 1]
[0; 0; 0]

Figure 5 show the obtained responses of process variables

with their respective zones and set points. Figure 6 shows
the input and output flow rates.
Table 3 shows the maximum flow rate oscillation for each
one of the inputs and outputs of the three-phase separator
for each control test. This table contains the informations
about the controllers presented in this paper and the best
controller described in Mendes et al. (2011), named as
Zone PNMPC Hammerstein. As expected, in the three
105

The higher damping of the oscillations in the water output

flow rate is obtained by the Proposed controller and the
Zone PNMPC Hammerstein with a factor of 6, 75 against
3, 37 of the PI Zone controller. Although the two predictive
controllers obtain almost the same damping, the proposed
controller gives a small value of the water output flow
variance. In figure 6 it can be noted that from the instant
22, 51 h the water output flow remains constant until
the end of the test, therefore giving 10, 82 hours of no
oscillatory flow.
Table 4 shows the flow rate variance for each one of
the outputs of the three-phase separator for each control
test. This table contains the informations about the part
of the test when no oscillatory disturbances are acting
on the process. The indices on the table show that the
proposed controller gives better results than the PI Zone,
moreover, it maintains the output flow rates more constant
and stabilizes the process variables with smaller settling
time, as can be observed in figures 4 and 6.
Note that there is an amplification of the oscillations in
the oil output flow rate for all the controllers. As the
water control level allows oscillations of hw to attenuate
the output flow variations, an increase in the oscillation
flow rate in the weir is induced, which causes a major
disturbance to the oil level. This disturbance added to the
input flow disturbance causes higher oscillations on the oil
output flow rate than the ones produces only due to the
input flow.
As pointed out in Filgueiras (2005) and Silveira (2006)
among others, the most important control objective is the
damping in water output flow rate, as this water feeds
the input of the hydrocyclones and its oscillations directly
influences their behavior. If the flow disturbances are

8th IFAC Symposium on Advanced Control of Chemical Processes

Furama Riverfront, Singapore, July 10-13, 2012

passed to the hydrocyclones, its separation efficiency will

be deteriorated. On the other hand, as normally the threephase separator is the first stage of the oil processing plant,
the oil output flow oscillations are not so fundamental.
According to Nunes et al. (2010) it is usual to use a
two phase oil/gas separator in series with the oil output
of three-phase separator, therefore this last equipment is
responsible for damping load oscillation and for providing
a constant flow to the electrostatic treater that is in series.
5. CONCLUSION
In this work a three level control strategy based on a zone
nonlinear predictive controller is proposed to control the
operation of a three-phase separator. A comparative study
of the proposed system with other control strategies is presented to show its advantages. The goals of the proposed
scheme are: (i) the use of a simple phenomenological model
to compute the predictions; (ii) to include a feed-forward
action in the controller based on a disturbance estimatorpredictor; (iii) to consider a flexible objective function
which enables the change of the zone width of the predictive strategy according the feed characteristics, obtained
using a disturbance detector. According to the obtained
simulation results one can conclude that the use of the
proposed strategy allow for better disturbance damping
than other controllers and also gives better performance
when operating in steady conditions. As for the implementation only a simple model of the separator and a QP
algorithm are needed, the use of the proposed controller
in real practice can be considered promissory. Future work
includes studies with controlled variables corrupted by
noise, tests in a real separator unit and the coupling of a
hydrocyclone model with the separator in order to analyze
the effect of this control strategy in the hydrocyclone
separation efficiency.
ACKNOWLEDGEMENTS
The authors thank Petrobras for providing financial support to the research project in which they are inserted.
Julio E. Normey-Rico also thanks CNPq and Agencia Nacional de Petr
oleo, Gas Natural e Biocombustveis (ANP)
for the financial support received from Programa de Recursos Humanos PRH-34 ANP/MCT.
REFERENCES
Bordons, C. and Cueli, J.R. (2004). Predictive controller
with estimation of measurable disturbances. application
to an olive oil mill. Journal of Process Control, 14, 305
315.
Camacho, E. and Bordons, C. (2007). Model Predictive
Control. Springer, Berlin.
Filgueiras, N.G.T. (2005). Modelagem, An
alise e Controle

de um Processo de Separaca
o Oleo/
agua. Dissertacao de
Mestrado, COPPE/UFRJ, Rio de Janeiro.
Hovland, S. (2004). Soft Constraints in Explicit Model Predictive Control. Msc. Thesis - The Norwegian University
of Science and Technology.
Kim, M., Domanski, S.H.Y.P.A., and Payne, W.V. (2008).
Design of a steady-state detector for fault detection and
diagnosis of a residential air conditioner. International
Journal of Refrigeration, 31, 790799.
106

Maciejowski, J. (2002). Predictive Control with Constraints. Prentice Hall, New Jersey.
Mendes, P.R.C., Carvalho, R.L., and Normey-Rico, J.E.
(2011). Nonlinear predictive control applied to load
oscillations damping in a three phese separator. Rio
Pipeline 2011, Rio de Janeiro: Brasil.
Nunes, G.C. (1994). Modelagem e Simulaca
o Din
amica do

Separador Agua/
Oleo/
G
as. Dissertacao de Mestrado Universidade Federal do Rio de Janeiro.
Nunes, G.C., Medeiros, J.L., and Ara
ujo, O.Q.F. (2010).
Modelagem e Controle na Produca
o de Petr
oleo.
Blucher, Sao Paulo: Brasil.
Nunes, G. (2001). Design and analysis of multivariable
predictive control applied to an oil-water-gas separator:
A polynomial approach. PhD Thesis, University of
Florida.
Nunes, G. (2004). Controle por bandas: Conceitos basicos
e aplicacao no amortecimento de oscilacoes de carga em
unidades de producao de petr
oleo. Technical report,
Comunicacao Tecnica - PETROBRAS.
Paoli, C., Voyant, C., Muselli, M., and Nivet, M.L. (2009).
Solar radiation forecasting using ad-hoc time series
preprocessing and neural networks. In D.S. Huang,
K.H. Jo, H.H. Lee, H.J. Kang, and V. Bevilacqua
(eds.), Emerging Intelligent Computing Technology and
Applications, volume 5754, 898907. Springer Berlin /
Heidelberg.
Pawlowski, A., Guzman, J.L., Rodrguez, F., Berenguel,
M., and Normey-Rico, J.E. (2011). Predictive control
with disturbance forecasting for greenhouse diurnal temperature control. 18th IFAC World Congress, Milano:
Italy, 17791784.
Pawlowski, A., Guzman, J.L., Rodrguez, F., Berenguel,
M., and Sanchez, J. (2010). Application of time-series
methods to disturbance estimation in predictive control.
IEEE Symposium on Industrial Electronics, Bari: Italy,
409414.
Plucenio, A. (2010). Desenvolvimento de Tecnicas de
Controle N
ao Linear para Elevaca
o de Fluidos Multif
asicos. Ph.D. thesis, PhD thesis PPGEAS - Universidade
Federal de Santa Catarina, Florianopolis-SC-Brasil.
Plucenio, A., Pagano, D.J., Bruciapaglia, A.H., and
Normey-Rico, J.E. (2007). A practical approach to predictive control for nonlinear processes. NOLCOS 2007
- 7th IFAC Symposium on Nonlinear Control Systems.
Prasath, G. and Jrgensen, J.B. (2009). Soft constraints
for robust mpc of uncertain systems. International
Symposium on Advanced Control of Chemical Processes.
Reikard, G. (2009). Predicting solar radiation at high
resolutions: A comparison of time series forecasts. Solar
Energy, 83, 342349.
Scandpower (2003). Users Manual OLGA 2000. Kjeller:
Norway.
Silveira, M.A.C.R. (2006). Controle de um Processo
de Tratamento Prim
ario de Petr
oleo. Dissertacao de
Mestrado - Universidade Federal do Rio de Janeiro.
Sotomayor, O.A.Z., Odloak, D., and Moro, L.F.L. (2009).
Closed-loop model re-identification of processes under
mpc with zone control. Control Engineering Practice,

17, 551U563.
Triggia, A.A., Correia, C.A., and Filho, C.V. (2001).
Fundamentos de Engenharia de Petr
oleo. Interciencia,
Rio de janeiro.