Sie sind auf Seite 1von 6

1 INTRODUCTION

An apparently appropriate control scheme for a single


reactor or distillation column may actually lead to an
inoperable plant when that reactor or column is connected
to other unit operations in a process with recycle streams
and energy integration. The main objective is to design a
control system that provides basic regulatory control of the
process. Above this regulatory structure, the systems can be
built to improve plant performance: real-time on-line
operations optimization (RTO), planning and scheduling,
and expert systems, among others. But if the basic
regulatory control does not work as the foundation of plant
operation, none of the higher level objectives can be met.
Methods such as singular value decomposition, condition
number analysis, and multivariable plots have their place in
plant-wide control. The model that is being used must
faithfully capture the nonlinearity and the constraints
encountered in the plant under consideration. Any
plant-wide control scheme must be tested on this type of
model because linear, unconstrained models are not
adequate to predict many of the important plant-wide
phenomena, so mathematical modeling and simulation are
vital tools in the solution of the plant-wide control problem.
Fortunately today is dawn of a new era in which the
computer-aided engineering software tools and computer
horsepower permit engineers to assemble a flow sheet
perform the steady-state analysis, and then evaluate the
dynamic performance of the plant. Commercial software

*Corresponding Author. Tel.: +98 611 5551021 (K. Salahshoor).


packages that combine steady-state and dynamic models
represent a major breakthrough in the tools available to the
process engineer and to the control engineer. Actually it is
predicted that in not too many years these two functions
will be combined and will be performed (as they should be)
by the same individual. An appreciation of dynamics is vital
in steady-state design and an appreciation of design is vital
in process control. There is no claim that this control
structure is necessarily the best. But it should provide stable
regulatory control of the plant. Plant-wide process control
involves the systems and strategies required to control an
entire chemical plant consisting of many interconnected
unit operations. While linear methods are very useful to
analyze control concepts, it is strongly believed that the
final evaluation of any plant-wide control structure requires
rigorous nonlinear dynamic simulations, not linear transfer
function analysis. Model predictive control is particularly
useful when several control valves (or manipulators) affect
an output of interest (what is called interaction)and also
when some sort of constraint comes into play either on the
inputs or on some measured variable. Since the controller
itself knows about these interactions and constraints, it can
in theory avoid those perils. Conventional control
approaches also rely on models, but they are usually not
built into the controller itself. Instead the models form the
basis of simulations and other analysis methods that guide
in the selection of control loops and suggest tuning
constants for the relatively simple controllers such as PID,
etc. Conventional control approaches attempt to build the
smarts into the system (the process and the controllers)
rather than only use complex control algorithms [1]. In this
paper for control purpose, HYSYS

is connected to
MATLAB

7.6 using Hysys as Activex Server with some


modifications to have more access to the desired variables.

Plant-wide Control of a Gas Plant Using a Structured Method

Karim Salahshoor
1,
*, Amin Mirzae
2


Department of Instrumentation and Industrial Automation, Petroleum University of Technology
Postal Code: 63431, Ahwaz, Iran
1
E-mail: salahshoor@put.ac.ir
2
E-mail: amin_mirzae@yahoo.com

Abstract: Very few unbiased publications have appeared in the literature comparing control effectiveness using MPC
versus a well-designed conventional control system. Most of the MPC applications reported have considered fairly
simple processes with a small number of manipulated variables. There are no published reports that discuss the
application of MPC to an entire complex chemical plant, with one notable exception. It is found that a good control
structure provides effective control, independent of any particular controller algorithm, while a poor one cannot be
greatly improved with any algorithm (MPC or PID controllers). In this paper, a Gas Plant is simulated in a rather
realistic environment by HYSYS, and the obtained data is in connection with MATLAB

. To have more realistic


simulations, the model of process is simulated in HYSYS

, and the data is transferred (real-time) to MATLAB

for
identification and control purposes.

Key Words: Plant-wide control, Control Structure, Decentralized Control, SVD, RGA, MPC, DMC, HYSYS simulator.
1111 978-1-4244-5182-1/10/$26.00 c 2010 IEEE





















2 PROCESS DESCRIPTION
In this paper a natural gas processin
modeled that uses propane refrigeration t
from the feed and a distillation tower to p
The flow sheet for this process appears
combined feed stream enters an inlet
removes the free liquids. Overhead gas
is fed to the gas/gas exchanger, where
already refrigerated gas. The cooled gas
chiller, where further cooling is acco
exchange with evaporating propane (r
C3Duty stream). In the chiller, which
simply as a Cooler, enough heavier hydro
such that the eventual sales gas meets a
specification. The cold stream is the
low-temperature separator (LTS). The d
to the gas/gas exchanger and then to
condensed liquids are mixed with free liq
separator. These liquids are processed
column to produce a low-propane-conten
A natural gas stream containing N
2
, CO
2
n-C
4
H
10
is being processed in a refrig
remove the heavier hydrocarbons. T
produced will meet a pipeline hydroc
specification. The liquids removed from
processed in a depropanizer column,
product with specified propane content. T
different natural gas streams containing c
methane through n-butane and combin
them in a simple refrigeration system. A
and coolers removes the heavier hydroc
from the natural gas stream, allowing it
dew point specification. The heavier liq
the gas stream is processed in a depr
yielding a liquid product with specified p



ng facility will be
to condense liquids
process the liquids.
s in Figure 1. The
t separator, which
from the Separator
it is pre-cooled by
s is then fed to the
omplished through
represented by the
h will be modeled
ocarbons condense
pipeline dew point
en separated in a
dry, cold gas is fed
o sales, while the
quids from the inlet
in a depropanizer
nt bottoms product.
2
, and CH
4
through
geration system to
he lean, dry gas
carbon dew point
m the rich gas are
yielding a liquid
The plant takes two
carbon dioxide and
nes and processes
series of separators
carbon components
to meet a pipeline
quid component of
ropanizer column,
propane content.


















3 MODELING AND SIM
There are two basic approache
process. On one hand, if the pro
and some experiments can be
behavior can be captured and
model, using identification a
techniques. A model is an imag
system), focused on a predet
image has its limitations, beca
incomplete knowledge of the s
represents the complete reality
under extreme circumstances, w
realize for the true process or
possible to investigate how a
disturbances. It is also possibl
behavior of a system, by
parameters. A model should th
of the reality that we like t
frequently used in science and te
model refers to entities va
descriptions of a process to a re
model is seldom a goal in itself.
helping to solve a problem
mathematical description of
simulations are of value to
academic researchers because t
demonstrations of the merits
areas such as control design, sys
diagnosis. Process simulations
been used in education and ac
years to compare the perform
methods for control, identificati
loop testing involved steps in
available manipulated variable
variations trend at the most inte
for LTS valve are shown in Fig


Fig 1. Gas Plant Flow Sheet.

MULATION
es to getting the model of a
ocess is physically available
e carried out, its dynamic
(partially) represented by a
and parameter estimation
e of the reality (a process or
termined application. This
ause it is usually based on
system and therefore never
y. A model can be tested
which is sometimes hard to
system. It is, for example,
a chemical plant reacts to
le to improve the dynamic
changing certain design
herefore capture the essence
to investigate. Models are
echnology. The concept of a
arying from mathematical
eplica of an actual system. A
. It provides always a tool in
, which benefits from a
the system [2]. Process
o university teachers and
they allow comparisons and
of different approaches in
stem identification and fault
in the public domain have
cademic research for many
mance and applicability of
ion and diagnosis [3]. Open
n the positions of the all
es and observation of the
rested variables. The results
ures 2, 3, where it can be

1112 2010 Chinese Control and Decision Conference














Fig 3. Step change in C3Duty(3% Increase in co
5min).
seen that the steady state gains and th
transients are generally simulated accura
3.1 Linearization model
Many simulation examples from t
presented in the form of a state-space m
of transfer functions. These forms are als
make the Gas Plant model acces
multivariable control design and analysi
linear models are deviations from the o
this paper for ease of use our linear mo
FOPTD(First Order Plus Time Delay)
Plus Time Delay), Because for PID T
dont need more complex model and the
are valid:
us


for this plant dynamic matrix can be easi
example:
g


And from Fig 4 its validity can be observ
4 CONTROL STRUCTURE
From a set of usually ambiguous us
mainly based on qualitative and econom
as well as operational constraints, the d
plant performances for control design
They may concern different properties,
tracking, control decoupling, distu
measurement noise rejection, robustness
the plant (model) or expected disturba
analysis knowledge may help in som
decisions based on a full-size mod
usefulness is limited, and these selection
Fig 2. Plant simula

ooler heat flow at time
e dynamics of the
ately.
the literature are
model or as a matrix
so presented here to
ssible for linear
is. Variables in the
operating point. In
odels are limited to
and IPTD(Integral
Tuning purpose we
ese approximations
(1)
ily obtained, for


ved.
ser control goals,
mical requirements
desirable controlled
should be derived.
such as reference
urbance rejection,
against changes in
ances [4]. System
me cases to take
del. However, its
ns should be also













Fig 4. Step response of FOPTD

guided by experience and com
plants, the basic extension o
design, implementation and tu
approach, where structural con
the interaction between variab
equipment and the ease of hand
non-specialist technicians are t
approach, the performance of
structures is usually quite ade
applications. In fact, there h
unbiased data showing that the
sophisticated controller structur
slight improvement is seldom
additional complexity and
implementation and mainten
multi-loop control would have t
Flexibility and fault tolerance
be turned on and off, due to o
without excessive degradation
rest of loops,
Simplicity, allowing indep
regulators, with easier online tu
In developing a model of a proc
very large number of external a
them related to the partial phen
hydraulic, thermal, concentrati
so on. But, certainly, it does not
and we do not need either to mea
one. That means a selection of
Three selections must be carried
Selection of controlled variabl
Selection of measurements,
Selection of manipulated varia
The steps in the procedure are s
ation, step change in C3Duty(3% Increase in cooler heat flow at tim


D model and real process.
mmon sense many industrial
of classical PID controller
uning is the decentralized
ncepts are used to decouple
bles. The use of standard
d-tuning or understanding by
the main advantages of this
f these diagonal controller
equate for process control
has been little quantitative
e performances of the more
res are really any better! The
m worth the price of the
d engineering cost of
nance. In an ideal case,
the advantages of:
e, as independent loops can
operator decision or faults,
on the performance of the
pendent design of SISO
uning.
cess, it may be considered a
and internal variables, all of
nomena we are interested in:
ion, movement, energy and
t need to control all of these
asure or to manipulate every
f variables should be done.
d out:
les,
ables [5].
summarized below:

me 5min).
2010 Chinese Control and Decision Conference 1113

4.1 Select controlled variables
Use primarily engineering judgment
understanding. Engineering judgment is
for deciding what variables to c
understanding of the process leads in mo
choice of what needs to be controlled. C
as economics, safety, constraints, and th
reliability of sensors must be factored in
must control inventories (liquid levels a
product qualities, and production rate. Ho
selection of the appropriate controlled
easy. In this case singular value decomp
powerful tool that tell us which variab
contribution on each singular value, the
correspond to minimum singular value
that column of output matrix U with th
value [6]. SVD involves expressing th
transfer function steady-state gains K
three matrices:
K 0 v



Fig 5. SVD analysis for output m

The lowest singular value is the most dif
to control. In ill-conditioned plants, the
biggest and lowest singular value is lar
greater than 20). The ratio of the larger s
called the condition number (CN):
CN


The higher the CN number, the more diff
decouple a control loop interaction. A ru
when CN 20 the system is nearly singu
is not feasible. Since this case is a non sq
of output greater than input) and its ob
our plant is ill-conditioned and real plant

based on process
s the principal tool
control. A good
st cases to a logical
Considerations such
he availability and
nto this decision. It
and gas pressures),
owever, sometimes
variable is not so
position(SVD) is a
bles have the most
e less important is
and is located on
he biggest absolute
he matrix of plant
as the product of
(3)

matrix.
fficult combination
e ratio between the
rge (for reference,

to the smaller s

is
(4)
fficult it becomes to
ule of thumb is that
ular and decoupling
uare plant (number
vious from (5) that
t rank is 7:

(5)
It means that for decentralized
seven controlled variables, from
eighth output has the biggest e
matrix; this procedure is repeate
4.2 Select manipulated variab
For a given process, select
variables will produce dif
alternatives. These control struc
controller structure, i.e., pairing
multi-loop SISO structure or o
The set of manipulated variab
minimum singular value

matrix is the best. For a perfect


the inverse of the plant. Therefo
be easily inverted (not close
minimum singular value mea
matrix is from being singular,
variables should be based on f
plant with a large minimum sin
of control structure is independ
controller tuning. The minimum
of the inherent ability of the p
choice of manipulated variable
changes in operating condition
effect of scaling on singular valu
the gains of all the plant transfer
form:
k


The same procedure for selectin
is be done, and from Figure
variables has the biggest eleme
this procedure is repeated for c

Fig 6. SVD Analysis

4.3 Eliminate unworkable va
A number of dynamically poor
by inspection. The Niederlins
eliminate some of the pairing
index mean unstable pairings,
tuning. Assume that the system
process matrix G(s) is non-singu
pairing problem defines the co
which of the available plant inp
each of the plant outputs. The
method is the RGA and NI b
relative gain array as:

d control we can only have
m Figure 5 its clear that the
element in column 10 of U
ed for column 9 and 8.
bles
ing different manipulated
fferent control structure
ctures are independent of the
g of variables in a diagonal
ne multivariable controller.
bles that gives the largest

of the steady-state gain


control, controller would be
ore, we want a plant that can
e to singular). Since the
asures how close the gain
the choice of manipulated
finding the set that gives a
ngular value. This selection
dent of variable pairing and
m singular value is a measure
process (with the specified
es) to handle disturbances,
ns, etc. The problem of the
ues is handled by expressing
r functions in dimensionless

(6)
ng the manipulated variables
6 its clear that the fourth
nt in column 9 of V matrix,
column 8.

for input matrix.
ariable pairings
r pairings can be eliminated
ski index can be used to
gs. Negative values of this
, independent of controller
m is open loop stable and the
ular at steady state. The loop
ontrol system structure, i.e.,
puts are to be used to control
most popular loop pairing
based pairing rules. Define
1114 2010 Chinese Control and Decision Conference










this can be calculated by:
u u


where the operatoris the Hadamard p
based pairing criterion is:
The paired RGA elements are closest to
All the paired RGA elements are positi
Large RGA elements should be avoide
NI for this loop pairing is 1.08 .The sign
provides a necessary stability condition
constitutes a complementary tool to the
pairing selection, NI is calculated by:
NI


There are many other methods such as
gain array (ERGA) [7], relative norm
(RNGA) [8], dynamic RGA (DRGA)
steady state gain and time constants of
account for possible amplification or
disturbance response for a particular can
rough disturbance model is availa
disturbance gain (RDG) is also useful.
5 Controller Tuning
There is no absolute right way or, for th
wrong way to tune controller. Controller
what the engineer/operator deems to be
in terms of the desired response to p
process control systems, load distur
frequently encountered; hence, most syst
for regulatory control. An issue that is ve
for level control that are important sin
material inventory that can be used as
dampen disturbances, hence, loosely tun
is also the best choice, although most op
offset [9]. There are many algorithm
controller tuning parameter based on FO
paper some more useful algor
ZieglerNichols, CohenCoon, IMC
Control), ATV (Auto-tune variation
examined on the performance of e
disturbance rejection and set-point trac
was developed for fluid and therm
emphasizes minimizing overshoot. ATV
the preferred technique for process contr
performance of these methods on set-p
load rejection on Figure 8. The settings
anticipated types of disturbance that
encounter. A controller would be tun
stability due to set-point changes (servo
load disturbances (regulatory control).








(7)
(8)
product. The RGA
o 1.0;
ive;
d.
n of NI, i.e. NI > 0,
and consequently,
e RGA in variable
(9)
s effective relative
malized gain array
that consider both
process. To try to
attenuation of the
ndidate pairing, if a
able, the relative
hat matter, absolute
settings depend on
good performance
process upsets. In
rbances are most
tems are optimized
ry important is that
nce level represent
surge capacity to
ned P-only control
perators do not like
for selecting PID
PTD model, in this
rithms such as
C(Internal Model
n technique), are
entire plant on
cking. ATV tuning
mal processes and
V, therefore, is often
rol. Figure 7 shows
point tracking and
also depend on the
the process will
ned differently for
o control) than for


Fig 7. Set point tracking for di

Fig 8. Disturbanc
6 Model Predictive Cont
There are very few design techn
stabilize processes in the pres
constraints. MPC refers to a
algorithms that control the fu
through the use of an explici
control interval, the MPC algori
sequence of manipulated variab
optimize future plant behavio
optimal sequence is injected in
optimization is repeated at su
[10]. An industrial type of this c
response information of the
Control (DMC), DMC algorith
response to model
the process:
y

t k g

ut k

And in the matrix form:


Y u0



ifferent tuning algorithms.

ce rejection.
trol
niques that can be proved to
sence of nonlinearities and
class of computer control
future behavior of a plant
it process model. At each
ithm computes an open-loop
ble adjustments in order to
or. The first input in the
nto the plant, and the entire
ubsequent control intervals
controllers that use only step
plant is Dynamic Matrix
hm uses the open-loop step
i ft k
0 F

(11)
2010 Chinese Control and Decision Conference 1115

Where

is future manipulated va
step response matrix and

as future

is the predicted output and is the


and is control horizon. Where f is
response:
ft k g

ut i g

t i
g

t k i

t k i

k p
The general quadratic function can be u
reach a control action as below:
I y

t j y

t j

The optimal value for control signal in fu



0 u

u I

F
Also for smooth reference tracking the re
is defined:
y

t k y

t k
y

t y

t

And:
it k Besiieu Setpoin
6.1 Reference tracking
The Figure 9 shows that MPC controller
reference tracking.
Fig 9. MPC set point tracking (for different

6.2 Disturbance rejection
A severe increase in Feed1 tempe
measureable disturbance (d1) and 50%
flow as second measureable disturb
illustrated that the MPC controller
disturbances that enter to this nonlinear
controller can find the change and t
Figure10 shows that because of high nonl
and also nonlinear relation between

ariables and


free response and
prediction horizon
s the process free
g

t i
i

t

used and finally we
ut j


(13)
uture is obtained:
F (14)
eference trajectory
it k
;

(15)
nt value

(16)

r has high power in

t and = 0.7).
erature (50%) as
increase in Feed1
bance (d2), it is
can reject this
r process. So MPC
try to remove it.
linearity of process
disturbances and
output, MPC controller cann
disturbance completely. Also th
in fact negligible.

Fig 10. MPC disturbance rejection (for
variables and bottoms a

7 CONCLUSION
It is found that a good control
control, independent of any par
while a poor one cannot be g
algorithm (MPC or PID control

REFERENCES

[1] W.L.Luyben,B.D.Tyreus,M.L.Lu
McGraw-Hill.
[2] B.Roffel, B.Betlem, Process Dy
Sons,Ltd.
[3] N.F.Thornhill, S.C.Patwardhan, S
heater simulation model with appl
18 (2008) 347360.
[4] P.Albertos, A. Sala,Multivariable
[5] M.L.Luyben, W.L.Luyben, E
MCGRAW-HILL.
[6] W.L.Luyben,Evaluation of criteri
trays in distillation columns, Jou
115134.
[7] Q.Xiong, W.J.Cai, M.J.He,A pr
multivariable processes, Journal
741747.
[8] M.J He, W.J. Cai, W.Ni, L.H X
configuration for multivariable pro
19 (2009) 10361042.
[9] W.Y.Svrcek, D.P.Mahoney, A
Control, John Wiley& Sons,Ltd.
[10] M.M. Arefi, A.Montazeri,
Wiener-neural identification and
realistic plug-flow tubular reacto
138(2008)274282.

not remove effect of this
his change is very small and


=0.0001; Top Figures are output
are control signals).
structure provides effective
rticular controller algorithm,
greatly improved with any
llers).
uyben,Plantwide Process Control,
ynamics and Control,John Wiley&
S.L.Shah, A continuous stirred tank
lications,Journal of Process Control
e Control Systems, Springer.
Essentials of Process Control,
ia for selecting temperature control
urnal of Process Control 16 (2006)
ractical loop pairing criterion for
l of Process Control 15 (2005)
Xie, RNGA based control system
ocesses, Journal of Process Control
Real-Time Approach to Process
J.Poshtan, M.R.Jahed-Motlagh,
d predictive control of a more
or, Chemical Engineering Journal
1116 2010 Chinese Control and Decision Conference

Das könnte Ihnen auch gefallen