Sei sulla pagina 1di 6

Prepr. Automation Days, Helsinki, 3{5.5.

1995

LIMITATIONS AND USE OF THE RGA AS A


CONTROLLABILITY MEASURE
Kurt E. Haggblom
Process Control Laboratory, Department of Chemical Engineering,

Abo Akademi University, FIN{20500 
Abo, Finland
Tel: +358{21{2654447, Fax: +358{21{2654479, E-mail: khaggblo@abo.
KEYWORDS: Relative gain array (RGA), controllability, interaction analysis,
multivariable control systems, decentralized control

Abstract. The relative gain array (RGA) is a

frequently used measure of controllability and control


loop interaction in multivariable control system design. Basically, the RGA is used for pairing controlled
and manipulated variables in a decentralized control
system, but it has also been used as a more general
measure of controllability. In this paper it is demonstrated that the RGA has a number of fundamental
shortcomings. A procedure for dealing with some of
these is suggested.

1. INTRODUCTION

Relative gain analysis is a widely used technique in


the design of control systems for multivariable plants.
The method was introduced by Bristol 1] and has
been further developed and popularized especially by
Shinskey 21 22] and McAvoy 19]. The analysis
is based on a \relative gain array" (RGA), which,
according to Bristol's denition, is a matrix of interaction measures for all possible single-input singleoutput (SISO) pairings of the variables considered.
The RGA thus indicates the preferable variable pairings in decentralized (multiloop SISO) control systems
based on interaction considerations. Originally, the
RGA was only dened at steady state, but it is
straightforward to include also dynamics 2 27].
Further developments have shown that the RGA is
much more than a simple measure of interactions.
The RGA provides information about fundamental
properties such as closed-loop stability and robustness
with respect to modelling errors and input uncertainty
9 25 30]. A summary of the use and known properties of the RGA is given by Hovd and Skogestad
16].
Despite its considerable popularity, the RGA has
several shortcomings. Friedly 6] showed that control
loops may interact both in the steady state and
dynamically even if the RGA suggests that there is
little interaction. According to Jensen et al. 18], the
reason for this is that the RGA is not a true measure
of closed-loop interactions. Furthermore, Bristol 3]
has shown that the RGA does not necessarily provide
an answer to the variable pairing problem for systems
larger than 2 2. In particular, this occurs when the
feasibility of a variable pairing is entirely dependent
on other control loops 12 13 22].
178

To deal with these and other deciencies of the


RGA, new independent measures that complement the
RGA have been proposed 4 8 17 26 28 29]. The
purpose of the measures is generally to guarantee
(decentralized) integral controllability of the plant. In
the present paper, a procedure based on the RGA is
suggested for dealing with the case when conventional
application of the RGA does not solve the variable
pairing problem.

2. THE RELATIVE GAIN ARRAY

In this section, the relative gain is dened and some


basic interpretations of its properties are given. Some
limitations are pointed out, but these are discussed in
more detailed in sections 3 and 4. A simple illustration
of the use of the relative gain is given.

2.1 Denition

The interaction measure proposed by Bristol 1 2]


attempts to answer the question how the apparent
transfer function between a given output variable ( i )
and a given input variable ( j ) is aected by control of
the other output variables. The measure is expressed
as the ratio ( ij ) of the transfer function between the
two variables with all other outputs uncontrolled, and
the transfer function between the same variables when
all other outputs are perfectly controlled, that is,




i 
i 
(1)
=
ij
j  k k6=j
j  k k6=i
The relative gains ij for all possible variable pairings
dene a matrix, the relative gain array (RGA),  .
The partial derivatives in Eq. (1) can be related to the
open-loop transfer functions of a system. Consider a
2 2 system described by the model
1 ( ) = 11( ) 1 ( ) + 12( ) 2 ( )
(2)
2 ( ) = 21( ) 1 ( ) + 22( ) 2 ( )
The transfer function between 1 and 1 with 2
uncontrolled (i.e., 2 = 0) is 11 . The corresponding
transfer function when 2 is perfectly controlled is
obtained by elimination of 2 with 2 = 0. This gives
the relative gain
;1
(
)
(
)
12
21
(3)
11 ( ) = 1 ;
11( ) 22( )
y

@y

@y

@u

@u

s u

s u

s u

s u

s G

s G

Because of the properties of the RGA, the sum of the


elements ij in each row and column is unity 1].
For a 2 2 system this means that it is sucient
to determine only one of the relative gains, because
11 = 22 = 1 ; 12 = 1 ; 21 . Usually 11 for
\diagonal variable pairing" is used. Therefore, the
subscripts are sometimes dropped and is referred
to as the relative gain for the system.
A linear
system may be described by the transfer
function model
y( ) = G( ) u( )
(4)
where u and y are -dimensional vectors of inputs
and outputs, respectively, and G is an
matrix of
transfer functions. From the denition of the relative
gain, Eq. (1), it follows that the RGA for the system
can be expressed as
(5)
( ) = G( ) (G( );1 )T
where is the Hadamard or Schur product, which denotes element-by-element multiplication(i.e., ij ( ) =
Gij ( ) (G( );1 )ji ).
A steady-state RGA is obtained when the transfer
functions are evaluated at = 0 (i.e., steady-state
gains are used), whereas = gives a frequencydependent RGA. A frequency-dependent RGA can
obviously provide more detailed information than
the steady-state RGA, especially regarding control
performance. However, it should be noted that the
steady state also provides much information about
closed-loop properties, in particular concerning integral stabilizability and controllability 9].


j!

2.2 Interpretations

Bristol proposed the RGA as a tool for pairing


controlled and manipulated variables in decentralized
(multiloop SISO) control systems. From the denition
of the relative gain it follows that the open-loop gain
;1
ij between i and j will change by the factor ij
when the other control loops are closed. This implies that variable pairings corresponding to positive
relative gains as close to unity as possible should be
preferred. Negative relative gains or relative gains
much larger that unity should be avoided and large
negative relative gains are particularly undesirable 1].
If 0
1, the control loop will interact with
ij
other loops, and for a 2 2 system, this interaction
will be strongest when ij = 0 5. Usually a 2 2
system with relative gains in this range is relatively
easy to control, either by decentralized controllers or
by the use of decoupling.
The relative gain ij = 1 indicates a \perfect"
variable pairing since interaction does not aect the
open-loop gain between i and j . This is often taken
to mean that the loop i { j does not interact with the
rest of the system. However, this interpretation is not
correct in general, as illustrated in section 3.1.
If ij
1, the open-loop gain between i and
j will decrease when the other loops are closed.
This reduces control eectiveness 22] and the control
performance is similar to single-loop control with too
much derivative action 3]. If the RGA contains
G

< 

<

>

entries ij  1, the closed-loop system is very


sensitive to parameter changes. Such a system may
be \practically uncontrollable" 9 25].
If ij 0, the sign of the open-loop gain between i
and j is changed when the other control loops are
closed. Usually this results in an inverse response of
i to changes in j . Furthermore, it has been shown,
under fairly general assumptions regarding the system
and the controller, that a decentralized control system
with a variable pairing i { j having ij (0) 0,
results in a closed-loop system with at least one of the
following properties 9]: (a) the closed-loop system
is unstable (b) the loop i { j is unstable by itself
(when all other loops are open) (c) the closed-loop
system is unstable if the loop i { j is open. This
means that such a variable pairing gives a closed-loop
system which is conditionally stable, at best.
If ij = 0, the relative gain does not indicate
whether the corresponding variable pairing is feasible
or not control depends entirely on other control loops
12 13 22]. This is not an unusual situation. In
distillation, for example, all so-called material balance
control structures contain variable pairings with this
property. This case is considered in sections 4 and 5.


<

<

2.3 Example: Simple blending system

Consider a simple blending system, where two streams


with ow rates 1 and 2 , and temperatures 1 and
2 , respectively, are joined to produce a stream with
ow rate and temperature . This system can be
described by the model
= 1+ 2
= ( 1 1 + 2 2)
(6)
It is desired to control
and
by two SISO
controllers manipulating 1 and 2 . How should the
variables be paired?
The process gains required for the relative gain analysis are obtained by linearization of Eqs. (6). This
yields
FF1 = 1
FF2 = 1
(7)
TF1 = ( 1 ; )
TF2 = ( 2 ; )
where FF1 denotes the gain between and 1 , etc.
According to Eq. (3), the relative gain for pairing
with 1 is
(8)
FF1 = ( ; 2 ) ( 1 ; 2 ) = 1
where the last equality is given by Eqs. (6).
It is obvious that FF1 is between zero and unity,
which means that also FF2 for pairing with 2 is
between zero and unity. The pairing with the relative
gain  0 5 should therefore be chosen. From Eq. (8)
it follows that FF1  0 5, and that should be
controlled by 1 , if 1  2 , or equivalently, if 
0 5( 1 + 2 ). These expressions mean that should
be controlled by the larger stream, which also is the
stream whose temperature is closer to .
An experimental illustration of the importance of
choosing the correct variable pairing is given by H!aggblom 11] in a study of conventional and nonlinear
model-based control of a mixing tank with similar
F

T F

T F

=F

T =F

T =F

= T

F =F

179

temp. (C)

50

1.0

45

0.8

40

35

level (cm)

10

20

30

40

50

0.0
0
5

60

20

30

40

50

60

70

80

90

100
b

2
1

10

20

30
time (min)

40

50

0
0

60

variable pairing ( = 0 68).


:

10

20

30

40

50
t

60

70

80

90

change of (a) 1 , (b) 2 . 1 = |, 2 = - -.


y

1.0

0.8

35
y

30

100

Figure 3. Interaction in triangular system. Setpoint

40

25

0.6
0.4
0.2

20
0
40

10

20

30

40

50

0.0
0
5

60

10

20

30

40

50

60

70

80

90

20

100
b

30

10

0
0

10

20

30
time (min)

40

50

incorrect variable pairing ( = 0 38).




properties as the blending system described. The


temperature and the level of the tank are controlled
by two streams of dierent temperature. The relative
gain expression is identical with Eq. (8). Figure 1 illustrates the control performance around an operating
point with = 0 68 using two SISO PI controllers. At
another operating point, the relative gain for the same
variable pairing is 0.38. Figure 2 shows the control
performance obtained by using this variable pairing,
which is the wrong pairing at the operating point in
question. As can be seen, the wrong variable pairing
results in much worse control performance.


The relative gain ij relates the open-loop gain between i and j when other outputs are uncontrolled
to the same open-loop gain when other outputs are
controlled. It is thus an open-loop measure for the
i { j pairing. The RGA does not contain explicit
information on how other inputs k 6= aect i
when it is paired with j , or how j aects other
outputs k 6= when the loop i { j is closed
18]. Therefore, there may be considerable interaction
between control loops even when the relative gain
implies that variable pairings are perfect, that is,
ij = 1 6]. This is illustrated in section 3.1.
It has been proved that the RGA is a rigorous measure
of the sensitivity of the system to modelling errors
9 25 30]. It can be shown that the system becomes


10

20

30

40

50
t

60

70

80

90

100

Figure 4. Control of \almost" triangular system.

Setpoint change of (a) 1 , (b) 2 . 1 = |, 2 = - -.


singular (which means that it becomes uncontrollable)
if a transfer function ij is changed by the amount
; ij ;ij1 30]. If there are perturbations in several
transfer functions, the system may become singular
even for smaller changes 25]. A system with large
entries in the RGA (i.e., j ij j  1) is thus very
sensitive to modelling errors, and may be very dicult
to control 9].
It should be noted that ;ij1 is a measure of the relative
error in ij that will make the system singular. If
ij is close to or equal to zero, a small absolute perturbation of ij may completely change the system
properties even if j ij j is small. See section 3.2.
y

3. LIMITATIONS CONCERNING
INTERACTION AND ROBUSTNESS
y

0
0

60

Figure 2. SISO PI control of mixing tank with

180

20

10

10

Figure 1. SISO PI control of mixingtank with correct

temp. (C)

0.4

30

0
0

level (cm)

0.6

0.2

30
0
40

3.1 Interaction in triangular system

Consider the triangular system



 
 
1 10 1
1 = 1
(9)
10 + 1 0 1 2
2
which, at all frequencies, has the RGA


 = 10 01
(10)
Variable pairing along the diagonal is therefore the
\perfect" pairing for decentralized control.
However, the large (1,2) entry in the transfer function
implies that 2 has a strong eect on 1 . This
interaction is not eliminated by a decentralized control
system. Figure 3 illustrates the responses to setpoint
changes using two PI controllers with the indicated
u

y
y

variable pairing. As can be seen, there is a strong


\one-way" interaction from control loop 2 to loop 1.

3.2 Sensitivity of triangular system

The RGA for the system in Eq. (9) tells that the
diagonal
gains have to be changed by the amount
; ii ;ii 1 = ;1  1;1 , that is by ;100 %, to make the
system singular. Furthermore, a change of the gain in
position (1 2) cannot make the system singular since
;10  0;1 = ;1 . It thus appears that the system is
very robust to modelling errors.
However, a change of the zero gain in position (2 1) to
0.1 is enough to make the system singular. This
is not
revealed by the relative gain 21 since 21 ;211 = 0 0.
A change of the gain to 0.09, which appears the be a
small perturbation in relation to the size of the other
gains, changes the system to
 

 
1
10 1
1 = 1
(11)
10 + 1 0 09 1 2
2
and the RGA to


 = ;109 ;109
(12)
G

y
y

This RGA indicates that the \almost" triangular


system in Eq. (11) is quite sensitive to modelling errors
and that it is fundamentally dicult to control. Figure
4 shows the responses to setpoint changes for this
system when the same controllers are used as in the
previous case. The responses are now very sluggish,
as implied by the large relative gains.

4. LIMITATIONS IN VARIABLE PAIRING

For a 2 2 system the RGA always indicates a preferable variable pairing for decentralized control, except
when all relative gains are equal to 0.5. In the latter
case, decoupling should usually be considered, unless
a frequency-dependent RGA indicates otherwise. It
is quite possible that a steady-state RGA suggests
a dierent variable pairing than the RGA evaluated
at higher frequencies 7 19 21 24]. In such cases, it
has to be considered what frequency range is most
important for feedback control.
For 3 3 or larger systems, there is not always
a clearcut choice of variable pairings even if only
a single frequency is considered (usually the steady
state). It may be necessary to select variable pairings
corresponding to zero or negative relative gains if
a decentralized control system is desired 3]. In
such cases, the feasibility of the pairings have to be
decided by other means. Even if all variable pairings
correspond to positive relative gains, decentralized
integral controllability cannot necessarily be guaranteed. Therefore, the RGA is usually used together
with other measures, such as the Niederlinski index
9 20] or the Morari index of integral controllability
26 30].
If there is only one set of variable pairings that results
in a decentralized controllable system, that set of pairings should usually be selected. However, often there
are several sets of feasible variable pairings, and the
task is then to choose the best set of pairings, usually

the set that yields the best control performance. As


shown in section 4.1, the best control performance is
not necessarily obtained by the set of variable pairings
with relative gains closest to unity.
A gain equal to zero in the transfer function model
always gives a relative gain equal to zero for the
corresponding variable pairing. Although counterintuitive, such a variable pairing may become feasible
when other control loops in the system are closed.
In distillation, for example, many well-known control
structures are based on such variable pairings 12 13].
See section 4.2.

4.1 Several feasible variable pairings

Hovd and Skogestad 16] have considered the plant


2
3
1
;
4
19
;
25
96
G( ) = (1(1+;5 ))2 4 6 19 1 ;25 96 5 (13)
1
1
1
:

which has the RGA

2
3
1 5 ;5
 = 4 ;5 1 5 5
(14)
5 ;5 1
at all frequencies.
Since the conventional rule is to prefer variable pairings with ij close to 1, pairing along the diagonal is
indicited in this case. Hovd and Skogestad found that
this set of variable pairings resulted in very sluggish
closed-loop performance with a time constant CL 
1160. For the set of variable pairings corresponding
to ij = 5, signicantly better performance with the
closed-loop time constant CL  220 was obtained.


4.2 Variable pairing on zero relative gain

A two-product distillation column with a total condenser can be described by an open-loop model of the
form 12 14]
   G ( ) 0  
u( )
y( ) = yu
(15)
z( )
Gzu( ) I ;1 v( )
where y =  D B ]T denotes product qualities (e.g.,T
distillate and bottoms compositions), z =  D B ]
denotes column inventories
(condenser and reboiler
holdups), u = 
]T denotes
internal ows (reux
and boilup), and v = 
]T denotes product ows.
The RGA for this system becomes
 G ( ) 0 
 G ( ) 0 ;1 !T
yu
( ) = Gyu( ) I ;1
Gzu( ) I ;1
zu



(
)
0
yu
=
(16)
0 I
s

L V

D B

where

yu( ) = Gyu( ) (Gyu( );1 )T


is the RGA for Gyu ( ) alone.
s

(17)

According to Eq. (16), the product qualities should


be controlled by the internal ows and the column
inventories by the product ows. Any other set of
pairings would contain pairings corresponding to a
181

LV

DV

LB

10

LV
DB
magnitude of relative gain

zero relative gain both for product quality and column


inventory control. It thus appears that the so-called
-structure, where and are used for product
quality control, is the only feasible control structure
in two-product distillation.
However, it is well known that the
- and structures, where the product qualities are controlled
by and , and and , respectively, are feasible
control structures. Even the
-structure, where
the product qualities are controlled by the product
ows, can be advantageous for certain columns 5 23].
These control structures become feasible when the
inventory control loops are closed. Since inventory
control is necessary for continuous operation, the
dependence of the quality control on the inventory
control is not necessarily a serious drawback.

LS
10

10

RV
RS
RB
LB

DB

5. A SOLUTION: RGA FOR PARTIALLY


CONTROLLED SYSTEM

The feasibility of a variable pairing corresponding to


a relative gain equal to zero is entirely dependent
on other control loops in the closed-loop system.
Therefore, at least some of the other control loops
have to be closed when the feasibility of such variable
pairings are considered. Actually, the denition of the
relative gain is based on this idea, see Eq. (1), but
when the original open-loop transfer function is zero,
the relative gain, which then also is zero, does not
contain any information on how the transfer function
is changed when the other loops are closed.
In distillation, inventory control is necessary to make
continuous operation possible, but apart from that,
the inventory control is of minor importance compared
to product quality control. Therefore, it is natural
to decompose the control task into inventory control
and product quality control, and to consider product
quality control under the assumption that inventory
control loops are closed 13]. As pointed out in section
4.2, this is also necessary if it is desired to evaluate alternatives to the -structure. The remaining openloop system associated with the product qualities can
then be analyzed, for example, by means of relative
gains 10 12 15]. This is illustrated in section 5.1.
Since closing some control loops may give useful
information when a variable pairing corresponding to
a zero relative gain is contemplated, it seems probable
that the same procedure may prove useful also in other
situations, where conventional application of RGA
does not solve the pairing problem. In section 5.2,
relative gain analysis based on such a \partially controlled" system is applied to the example introduced
in section 4.1.
LV

5.1 Variable pairing on zero relative gain

H!aggblom 12] has evaluated 9 dierent control structures for product quality control of a distillation
column based on a model (\Column B") given by
Skogestad et al. 24]. Frequency-dependent relative
gains for diagonal variable pairing were calculated
for the various 2 2 systems obtained by closing
the implied inventory control loops in each control
structure. Figure 5 shows the magnitudes of the
182

-1

10 -5
10

DS
DV
10

-4

10

-3

-2

10
frequency (1/min)

10

-1

10

10

Figure 5. Frequency-dependent relative gains for


dierent distillation control structures.

frequency-dependent relative gains for the control


structures considered.
In addition to the conventional
-structure and
the material balance structures
,
and
,
a number of control structures involving the ratios
=
and =
as manipulated variables
were considered. As can be seen, the - and structures are very good over the whole frequency
range. The
-structure, which has an innite
relative gain at steady state, is also very good in the
frequency range important;1for feedback control, which
here is = 0 01 . . .1 min . These control structures
contain variable pairings corresponding both to zero
gains and zero relative gains, and the evaluation
required inventory control loops to be closed.
LV

DV

L=D

LB

DB

LB

RB

V =B

DB

5.2 Several feasible variable pairings

According to Hovd and Skogestad 16], the example


in section 4.1 is a counterexample to the conventional
pairing rule. The variable pairing implied by the RGA
is along the diagonal of the transfer function matrix.
However, Hovd and Skogestad found that the pairing
corresponding to ij = 5 gave superior performance.
Consider the eect of closing the loop 3 { 3 . According to Eq. (13), perfect control of 3 requires
3 = ; 1 ; 2 . Elimination of 3 from the system
gives the reduced transfer function matrix


(1
;
)
26
96
21
77
33
(18)
G ( ) = (1 + 5 )2 32 15 26 96
which has the RGA


26
98
;
25
98
33
 = ;25 98 26 98
(19)
at all frequencies. The RGA for the reduced system
still shows that 1 should be paired with 1 and 2
with 2 , if 3 is paired with 3 , but the large relative
gain values imply that control will be very dicult.
Consider now pairing 3 with 1 , which has 31 = 5.
In a similar way as above, the RGA


6
19
;
5
19
31
(20)
 = ;5 19 6 19
is obtained for the remaining subsystem. Since the
positive relative gain values are reasonable small, this
RGA does not imply any particular control problems.


From this it follows that the set of variable pairings


corresponding to = 5 appears to be superior to the
one corresponding
to = 1. The relative gain values
of 33 and 31 correlate well with the closed-loop
time constants reported by Hovd and Skogestad. The
same result would have been obtained by closing any
other of the loops concerned.


6. CONCLUSIONS

It has been shown that the relative gain array has


some fundamental shortcomings. The main reason is
that it is not a true measure of interactions, especially
for the closed-loop system. Even if the RGA has
the perfect value of an identity matrix, there may
be signicant one-way interaction in the system. For
3 3 and larger systems, variable pairing based on
the RGA may fail, because the RGA may be unable
to (a) discriminate between several feasible sets, (b)
nd all feasible sets, (c) nd any feasible set of
variable pairings. A procedure for dealing with these
cases was suggested: the RGA is determined for the
reduced, \partially controlled", system obtained by
closing some of the control loops implied by a given
set of variable pairings. As shown, this may provide
very useful information about the system.

ACKNOWLEDGEMENT

This work was supported by the Academy of Finland.

REFERENCES

1] Bristol, E.H. (1966). On a new measure of interactions for multivariable process control. IEEE Trans.
Autom. Control, AC-11, 133{134.
2] Bristol, E.H. (1978). Recent results on interactions
in multivariable process control. AIChE Annual
Meeting, paper 67b, Miami Beach, Florida, USA.
3] Bristol, E.H. (1979). Linear and nonlinear eects of
interaction. Proc. AIChE Workshop on Industrial
Process Control, pp. 85{91, Tampa, Florida, USA.
4] Chiu, M.-S., and Y. Arkun (1990). Decentralized
control structure selection based on integrity considerations. Ind. Eng. Chem. Res., 29, 369{373.
5] Finco, M.V., W.L. Luyben, and R.E. Polleck (1989).
Control of distillation columns with low relative
volatility. Ind. Eng. Chem. Res., 28, 75{83.
6] Friedly, J.C. (1984). Use of the Bristol array in
designing noninteracting control loops. A limitation
and extension. Ind. Eng. Chem. Process Des. Dev.,
23, 469{472.
7] Gagnepain, J.-P., and D.E. Seborg (1982). Analysis
of process interactions with applications to multiloop
control system design. Ind. Eng. Chem. Process
Des. Dev., 21, 5{11.
8] Grosdidier, P., and M. Morari (1986). Interaction
measures for systems under decentralized control.
Automatica, 22,309{319.
9] Grosdidier, P., M. Morari, and B.R. Holt (1985).
Closed-loop properties from steady-state gain information. Ind. Eng. Chem. Fundam., 24, 221{235.
10] Haggblom, K.E. (1988). Analytical relative gain
expressions for distillation control structures. In:
Consistent Control Structure Modeling with Application to Distillation Control, Dr. Tech. Thesis, 
Abo
Akademi University, 
Abo, Finland.

11] Haggblom, K.E. (1993). Experimental comparison of


conventional and nonlinear model-based control of a
mixing tank. Ind. Eng. Chem. Res., 32, 2653{2661.
12] Haggblom, K.E. (1994). Analytical relationships
between frequency-dependent relative gains for distillation control structures. Proc. Int. Symp. on
Process Systems Engineering (PSE '94), pp. 977-982,
Kyongju, Korea.
13] Haggblom, K.E. (1994). Modeling of control structures for partially controlled plants. Prepr. IFAC
Workshop on Integration of Process Design & Control
(IPDC '94), pp. 183{188, Baltimore, Maryland, USA.
14] Haggblom, K.E., and K.V. Waller (1991). Modeling of distillation control structures. Report 91{4,
bo Akademi University,
Process Control laboratory, A

Abo, Finland.
15] Haggblom, K.E., and K.V. Waller (1992). Control
structures, consistency, and transformations. In:
Practical Distillation Control, W.L. Luyben, Ed., pp.
192{228, Van Nostrand Reinhold, New York.
16] Hovd, M., and S. Skogestad (1992). Simple frequencydependent tools for control system analysis, structure
selection and design. Automatica, 28, 989{996.
17] Hovd, M., and S. Skogestad (1994). Pairing criteria
for decentralized control of unstable plants. Ind. Eng.
Chem. Res., 33, 2134{2139.
18] Jensen, N., D.G. Fisher, and S.L. Shah (1986).
Interaction analysis in multivariable control systems.
AIChE J., 32, 959{970.
19] McAvoy, T.J. (1983). Interaction Analysis. Instrument Society of America, Research Triangle Park,
USA.
20] Niederlinski, A. (1971). A heuristic approach to
the design of linear multivariable control systems.
Automatica, 7, 691{701.
21] Shinskey, F.G. (1979). Process-Control Systems, 2nd
ed. McGraw-Hill, New York.
22] Shinskey, F.G. (1984). Distillation Control, 2nd ed.
McGraw-Hill, New York.
23] Skogestad, S., E.W. Jacobsen, and M. Morari (1990).
Inadequacy of steady-state analysis for feedback control: distillate-bottom control of distillation columns.
Ind. Eng. Chem. Res., 30, 2339{2346.
24] Skogestad, S., P. Lundstrom, and E.W. Jacobsen
(1990). Selecting the best distillation control conguration. AIChE J., 36, 753{764.
25] Skogestad, S., and M. Morari (1987). Implications
of large RGA elements on control performance. Ind.
Eng. Chem. Res., 26, 2323{2330.
26] Skogestad, S., and M. Morari (1992). Variable
selection for decentralized control. Modeling, Identi cation and Control, 13, 113{125.
27] Witcher, M.F., and T.J. McAvoy (1977). Interacting
control systems: Steady state and dynamic measurement of interaction. ISA Transactions, 16, 35{41.
28] Yu, C.-C., and M.K.H. Fan (1990). Decentralized
integral controllability and D-stability. Chem. Eng.
Sci., 45, 3299{3309.
29] Yu, C.-C., and W.L. Luyben (1986). Design of
multiloop SISO controllers in multivariable processes.
Ind. Eng. Chem. Fundam., 25, 498{503.
30] Yu, C.-C., and W.L. Luyben (1987). Robustness with
respect to integral controllability. Ind. Eng. Chem.
Res., 26, 1043{1045.

183

Potrebbero piacerti anche