You are on page 1of 8

Available online at www.sciencedirect.

com

ScienceDirect
Procedia Engineering 148 (2016) 1104 1111

4th International Conference on Process Engineering and Advanced Materials

Control of depropanizer in dynamic Hysys simulation using MPC in


Matlab-Simulink
Truong Thanh Tuana, Lemma Dendena Tufaa,*, Mohamed Ibrahim Abdul Mutaliba,
Abdelraheem Faisal Mohammed Abdallaha
a
Department of Chemical Engineering, Universiti Teknologi PETRONAS, Bandar Seri Iskandar, 32610 Tronoh, Perak, Malaysia

Abstract

Control of depropanizer process is challenging due to the interaction caused by the multivariable nature of the process and
significant non-linearity of the system. In-situ plant tests are mandatory for application of MPC in a real industrial environment
due to the necessity of the model for the design of the controller. However, these tests are costly for most chemical industries due
to process complexities, significant disturbances and unavoidable process interruptions. In this study, Aspen Hysys is used to
simulate the real plant, generate input-output data to develop the plant model and to conduct performance tests. Matlab-Simulink
is used to conduct model identification, design the MPC and implement the multivariable control action. The control objective is
to reduce variation of product purity due to operation constraints. The paper demonstrates how Aspen Hysys and Matlab-
Simulink can be effectively used for control study.

2016
2016TheTheAuthors. Published
Authors. by Elsevier
Published Ltd. Ltd.
by Elsevier This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the organizing committee of ICPEAM 2016.
Peer-review under responsibility of the organizing committee of ICPEAM 2016
Keywords: model predictive control; PRBS closed loop test; ARX model; Matlab-Hysys Interface

1. Introduction

Nomenclature

RCY recycle operator

* Corresponding author. Tel.: +6053687624; fax: +6053656176.


E-mail address: lemma_dendena@petronas.com.my

1877-7058 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the organizing committee of ICPEAM 2016
doi:10.1016/j.proeng.2016.06.601
Truong Thanh Tuan et al. / Procedia Engineering 148 (2016) 1104 1111 1105

TEE flow splitter


LIC level controller
FIC flow controller
TIC temperature controller
XIC propane mole fraction controller
NU number of input,
NY number of output
na order of the polynomials that relate past output values
nb order of the polynomials that relate past input values
r set point
yk+1 output prediction
u input increment
uk input deviation
Q, R, W weighted factors of J
Ts Sample time
R, S Reflux flowrate and boil up flowrate, respectively

In industrial processes, the use of distillation columns has increased rapidly and the demand for better control of
process conditions has led to greater complication in control mechanisms. Controllers are required to make non-
offset or tiny offset between signal outputs and references by tuning to compensate for frequently changing
operating conditions. A substantial research in field of control engineering is directed towards finding a better
strategy control instead of conventional control. Advanced control is the ideal control which should be capable of
compensation for tracking set points. Model predictive control (MPC) is one of the advanced control techniques
which relies on model of system identification and usually applies with linear process. MPC is used to steer a
process to operate at the values given from optimization calculation step. MPC is broadly acknowledged in the
process industry, demonstrated by significant number of successful control technique applications presented in
literature [1] There are three common ways to approach prediction in model predictive control such as state space
model, transfer function model, and finite impulse response (FIR) model [2]. Identification model for use in MPC is
conducted by closed loop or open loop test. Closed loop identification has many advantages over open loop
identification [3] because closed loop tests can effectively analyze behavior of input changing. In this study, the
closed loop test for model identification has been carried out.

The process was simulated in Aspen Hysys which was the modified process based on the depropanizer process
from [4]. The distillation column in the process was used to distillate propane and lighter component from C4+ such
as n-butane, i-butane, n-pentane and other components. The distillation column in simulation consisted of 24 trays
with 100% efficiency for each tray. The distillation column is well-known with nonlinear characteristics. However,
distillation column exhibit linear behavior while controlling narrow manipulated variables band. Frequently, the
column operates in specific spectrum with linear phenomenon. Muddu et al. [5] studied predictive control in
controlling the composition of product with narrow band of manipulated variables. Kon et al. [6] carried out the
application of ARX models to the refinery distillation column also with small margin of manipulated variables.
Seban et al. [7] performed system identification on reactive distillation column with 3% changing for the inputs
simultaneously, reflux ratio and vapor flow were selected as manipulated variables effected varying product
compositions. In numerous case, it is essential, and helpful, to determine a system model by connecting online while
the process is running. In numerous cases, they observed the systems in linear behavior by choosing narrow inputs
band.
In this research, multivariable model predictive control for distillation column is presented with manipulated and
controlled variables in scope of linear behavior. Model identification for MPC is of significant value for succeeding
highest performance which is calculated by ARX model. Data for performing system identification is taken by
conducting Matlab-Hysys interface. Hysys simulates the depropanizer process and Matlab connects with Hysys to
change and receive data.
1106 Truong Thanh Tuan et al. / Procedia Engineering 148 (2016) 1104 1111

2. Problem Statement

In order to implement MPC to control live process such as depropanizer process, plant testing is intensive part to
identify model for carry out in MPC. Classical plant testing can extend many weeks; it derives interruption to plant
operation. Therefore, effective time and financial savings can be achieved by simulating the real process by dynamic
process simulation which have full characteristic of that process. The objectives of this work are (1) simulation of
the depropanizer process in steady-state and then convert it to dynamic; (2) Using SISO PI closed loop controllers
which were set in Simulink (Matlab) to control the multivariable process; (3) Introducing PRBS (Pseudo Random
Binary Sequence) signals for changing both reference signals in the Simulink setup and collecting measured inputs
and outputs data; (4) MPC is apply to control the process by tracking the set points.

3. Methodology

The depropanizer process is modelled using Aspen-Hysys 8.0 for steady state simulation and then convert to
dynamic simulation. In this study, Peng-Robinson package is used for whole simulation because depropanizer is a
natural gas liquefaction (NGL) process [8].

The depropanizer process scheme is shown in Fig. 1. The NGL feed stream which consists of variety of
hydrocarbon ranging from C1-C5+ is fed to the Separator to separate light hydrocarbon (with 0.2-0.3 mole % of C5+
material). The mixed gas come out from Separator is sent to sales gas product. The liquid stream from Separator
passes through heat exchanger (E-100) to the depropanizer column (T-100). E-100 is used to retrieve heat energy of
vapour stream from top column to heat up the feedstock to the column. The depropanizer column is used to isolate
the desired propane (xD(C3H8) 0.8, xB(C3H8) 0.15) from the mixed hydrocarbons.

Fig. 1. Steady state depropanizer process simulation.


Truong Thanh Tuan et al. / Procedia Engineering 148 (2016) 1104 1111 1107

Fig. 2. Dynamic depropanizer Process Simulation.

Before transitioning from steady state simulation to dynamic state, sizing of the equipment (Valve, Separator,
Tank, Heat exchanger, Distillation Column) is carried out based the control objectives. In order to avoid vapor feed
to pumps, the level of liquid in Separator, Tank (V-100) is maintained by a controller. Cooler (E-101) is used to
condense volatile component from the top column. The dynamic simulation is shown in Fig.2. The control objective
of the MPC is to attain specifications of top and bottom products (desired propane mole fraction) by manipulating
flow rate of the reflux and the boil up.

3.1. Controller Design Steps

3.2.1. SISO PI Controllers

To conduct test for gathering data for system identification of the MPC, the plant should normally be under
control. For this purpose, PI controllers are designed in Simulink (Matlab) to control the purity of top and bottom
product by changing reflux flow rate and steam flow rate, respectively.

Fig. 3. SISO PI Closed Loop Controllers.


1108 Truong Thanh Tuan et al. / Procedia Engineering 148 (2016) 1104 1111

Fig.3. shows the closed loop PI controllers for Dynamic Process in Hysys. The above PI controller controls the
top product; the other PI controller controls the bottom product in depropanizer distillation column. Proportional
gain, P and integral time, I are determined by manual tuning method on the online dynamic system in Hysys or the
other way by the Ziegler-Nichols (ZN) tuning method. This step prepares the closed loop control for identification
which produces the model using for in MPC.

3.2.2. PI controller detuning

Before introducing PRBS signals to collect the data, PI controllers were detuned to avoid the observation PI
controller behaviors. The detuning parameter was defined in [9]:

I
P D P , I (1)
D
where: the values P*, I* are determined by tuning method. Controllers are detuned with <1.

3.2.3. Introducing PRBS signals and collecting data.

Fig. 4. Introducing PRBS signals in set point

After the dynamic process in Hysys getting steady state at given set points by PI closed loop control. Fig.4 shows
that Pseudo-Random Bit Sequence (PRBS) signals are introduced as a bias to the set points for generating the data.

3.2.4. Model Predictive Control

MPC is the algorithm on predictive control using almost linear models such as ARMAX, ARARMAX and ARX.
System identification is the first step for applying MPC. This work chooses ARX model to detect process model for
MPC controller. The ARX model is fed with input and output values for parameters estimation.
MPC control law calculations:

n
J (e
k 0
2
k 1  O (uk  uss ) 2  P ('uk ) 2 ); ek 1 rk 1  yk 1 (2)
Truong Thanh Tuan et al. / Procedia Engineering 148 (2016) 1104 1111 1109

ek 1Qek 1  (u k )T R(u k )  'u k W 'u k ; uk uk  uss ;


T T
J (3)

Optimum trajectory of future control input depends linearly on future target and past inputs/outputs.

wf wf wf
min J grad J 0; where : grad J ('u k ) ; ; ; (4)
w'u k1 w'u k 2 w'u k n

Solving (4) to find control input value (uk)

3.2.5. MPC controller design

Fig. 5. MPC controller structure.

In Fig.5, MPC controller is designed in Simulink Matlab. MPC control the boil up and reflux streams in Hysys
Dynamic Plant to track the set points which are propane composition in the top (xD) and bottom (xB) products.

4. Results and discussion

Fig 6 shows 5000 collected data from using Hysys-Matlab interface (Ts = 0.1). Dynamic process simulation is
connected with Matlab by running modified functions from HYSYSLIB toolbox [10] to set and get values through
spreadsheet cell.
4501 collected data from introducing PRBS signals is divided to 3376 (75% of 4501) for modelling and 1125 for
validation.

After getting the data, ARX model is applied to identify model for the Hysys Dynamic Process. In this process,
the top and bottom purity product of the distillation column are controlled by reflux flowrate and steam flowrate. So,
ARX model is used for MISO1 with 2 inputs (reflux flowrate, R and boil up flowrate, S) and 1 output (xD) and
MISO2 with the same 2 inputs and other output (xB)

R R
xD MISO1u ; xB MISO 2 u ; (5)
S S

Combining MISO1 and MISO2 become one model ( [MISO1;MISO2] ):


1110 Truong Thanh Tuan et al. / Procedia Engineering 148 (2016) 1104 1111

1 0 xD MISO1 R B1 B2 B3 B4
0 1 x MISO 2 u S ; where : MISO1 A ; MISO 2 ; (6)
B 1 A1 A2 A2

Fig. 6. Data collection by introducing PRBS signals

Following term values achieved after validation:


A1 = [-1.1145 0.1483 -0.1341 0.1051]; A2 = [-1.1219 0.2012 -0.1121 0.0861]
B1 = [6.5766 -5.2582 -1.0825 -0.1935]; B2 = [-1.2943 1.1907 0.0577 0.0393]
B3 = [8.8111 -8.9335 0.5029 0.2170]; B4 = [-6.3689 5.3833 1.7734 -1.2160]
Following the retrieval of model, the rest 1125 data is used for validation. Fig. 7 compares the calculated data
with the measured data. Fitting are 82.55%, 96.82% for xD and xB respectively.

Fig. 7. Cross validation ARX model for xD, xB respectively

Fig.8 shows the performance of MPC controller with tracking set points (xD from 0.92 to 0.94, xB from 0.12 to
0.1). The MPC control gives offset 0.0006 for xD and 0.0023 for xB. These offset are seen because of MPC uses a
model that does not completely represent the Hysys Dynamic process. However, the MPC maintains the system
stability and the variances of controlled variables were reduced substantially.
Truong Thanh Tuan et al. / Procedia Engineering 148 (2016) 1104 1111 1111

Conclusion

In this work, the dynamic depropanizer process in Hysys is controlled by MPC, designed in Simulink
environment. Hysys and Matlab are executed simultaneously. With an assumption of dynamic process, similar to
real plant, the MPC applying steps are quite contemplative. The challenging parts in the process are system
identification and tuning MPC controller. The fitting of model with measured data (82.55 and 96.82%) could not get
100%, so the model does not presented totally dynamic process. That answering the question why in this
application, MPC still give the offset when tracking set points. However, improvements in MPC model which is
called offset-free MPC can allow better control dynamics, which is suitably the objective for further research in this
field.

Fig. 8. MPC control performance

Acknowledgements

We acknowledge the support we get from YUTP- 0153AA-A87 to conduct this research.

References

1. Qin Trade Secret Development S. J. and Badgwell T. A. A survey of industrial model predictive control
technology. Control Eng. Pract. 11 (2003) 73364.
2. Rossiter J. A. Model-based predictive control: a practical approach. CRC press, 2013.
3. Zhu Y. and Butoyi F. Case studies on closed-loop identification for MPC. Control Eng. Pract. 10 (2002)
40317.
4. Fikar M., Latifi M. A., Corriou J. P., and Creff Y. CVP-based optimal control of an industrial depropanizer
column. Comput. Chem. Eng. 24 (2000) 90915.
5. Muddu M., Narang A., and Patwardhan S. C. Reparametrized ARX models for predictive control of staged
and packed bed distillation columns. Control Eng. Pract. 18 (2010) 11430.
6. Kon J., Yamashita Y., Tanaka T., Tashiro A., and Daiguji M. Practical application of model identification
based on ARX models with transfer functions. Control Eng. Pract. 21 (2013) 195203.
7. Seban L., Kirubakaran V., Roy B. K., and Radhakrishnan T. K. GOBF-ARMA based model predictive
control for an ideal reactive distillation column. Ecotoxicol. Environ. Saf. 121 (2015) 1105.
8. Yuan Z., Cui M., Song R., and Xie Y. Evaluation of Prediction Models for the Physical Parameters in
Natural Gas Liquefaction Processes. J. Nat. Gas Sci. Eng.
9. Bezergianni S. and Georgakis C. Evaluation of controller performanceuse of models derived by subspace
identification. Int. J. Adapt. Control Signal Process. 17 (2003) 52752.
10. Olaf T. B. A toolbox for using MATLAB as an activeX/COM controller for Hysys, Matlab Central.
http://www.pvv.org/olafb/hysyslib/, 1999., Computer stuff. 2008.

You might also like