TABLE V.

I
616
Characteristics of multivariable control systems
PARTV
Multivariable Control Single-loop characteristics that generally
lead to good control performance in Characteristics unique to multivariable
multivariable systems systems
1. Fast feedback processes (small 9 + x) 1. Interaction between variables influences
control stability and performance.
2. Feedback processes with a small fraction 2. Feasibility of control depends on overall
dead time (0/(0 + r)) and no process, not just individual
inverse response cause-effect relationships.
3. Disturbances with small magnitudes far 3. The source of the disturbance, not just
from the critical frequency the magnitude, must be considered
in designing the control strategy.
4. No limitations encountered in the 4. The pairing of measured variables and
manipulated variable final elements via control
is a design decision.
5. Digital controllers with relatively 5. Some processes have an unequal number
fast execution periods of controlled and manipulated variables.
6. Controllers based on accurate models 6. Some multivariable control designs
are very sensitive to modelling
errors.
7. Controllers using appropriate
enhancements from Part IV

considered. Interaction results from process relationships that cause a manipulated

variable to affect more than one controlled variable. In Figure VI the heating oil
valve position influences both the temperature and, through the reaction rate con
stant, the concentration. This is the major difference from single-loop systems and
has a profound effect on the steady-state and dynamic behavior of a multivariable
system.
Thus, it is not possible to analyze each manipulated-controlled variable con
nection individually to determine its performance; the integrated control system
must be considered simultaneously. A closely related new issue is the distur
bance source, because multivariable systems respond differently to different dis
turbances. For example, the chemical reactor responds differently to disturbances
in feed composition and feed temperature, and, as we shall see, these differences
must be considered in designing a multivariable control system.
Another realistic issue is the number of controlled and manipulated variables,
which may not be equal. Note that the system in Figure V.l has four manipulated
variables, which can be adjusted to control three measured variables. Multivariable
control methods presented in this part are able to utilize all flexibility available in
the process.
There are two basic multivariable control approaches. The first is a straight
forward extension of single-loop control to many controlled variables in a process,
as shown in Figure V.2. This is termed multiloop control and has been applied with
success for many decades. The second main category is coordinated or centralized 617
control, in which a single control algorithm uses all measurements to calculate all
manipulated variables simultaneously, as shown in Figure V.3. Algorithms for this PARTV
approach have been available for several decades and have been widely applied Multivariable Control
for a considerable time in the process industries.
At the conclusion of this part, the unique characteristics of multivariable pro
cess systems and how these characteristics affect process control will have been
presented. The reader is cautioned that this is a complex topic, worthy of an en
tire book, and that the presentation here is introductory. However, it presents the
major issues, along with some of the more common analysis methods and control
approaches.

FIGURE V.1 FIGURE V.2

Multivariable process. Example of multiloop control design.

FIGURE V.3

Example of multivariable control design.

 Multiloop
Control: Effects
of Interaction
20.1 u INTRODUCTION
Multivariable control occurs in nearly all processes, because production rate (flow),
inventory (level and pressure), process environment (temperature), and product
quality are normally controlled simultaneously. The multiloop approach, using
multiple single-loop controllers, was the first approach used for multivariable con
trol in the process industries. Through decades of research and experience, many
successful multiloop strategies have been developed and continue to be used.
One advantage of multiloop control is the use of simple algorithms, which is
especially important when the control calculations are implemented with analog
computing equipment. A second advantage is the ease of understanding by plant
operating personnel, which results from the simplicity of the control structure.
Since each controller uses only one measured controlled variable and adjusts only
one manipulated variable, the actions of the controllers are relatively easy to mon
itor. A third advantage is that standard control designs have been developed for
the common unit operations, such as furnaces, boilers, compressors, and simple
distillation towers. This does not mean that a single control design functions well
for all unit operations of the same type. However, several general structures are in
common use, and selection among alternatives can be based on analysis and expe
rience. Considering these advantages, one could conclude that multiloop designs
will continue to be used extensively, although not exclusively.
An example of multiloop control of a flash process is given in Figure 20.1.
Let us consider the behavior of the system when the feed flow rate increases. An
initial effect is an increase in the amount of vapor entering the drum, although the
620 A - ~ To containment
CHAPTER 20
Multiloop Control: {XI ▶ Vapor
Effects of Interaction

£-^ B-105
Feed

C7
4&r—
&
Q Liquid

Steam
FIGURE 20.1

Example of a multiloop control system.

percentage feed vaporized decreases because of a slight decrease in inlet tempera

ture. The pressure in the drum increases because of the additional vapor; therefore,
the pressure controller PC-1 takes action by increasing the percent opening of the
valve in the vapor line. Another effect is a decrease in the temperature after the
heat exchanger, which is sensed by TC-3. This feedback controller increases the
steam flow to the exchanger, which returns the temperature to its set point and
causes even more feed to be vaporized. This additional vapor causes the pressure
to increase, and the pressure controller has to respond to this change as well. The
increase in feed rate and changes in percent vaporized introduce changes in the
liquid rate into the liquid inventory in the drum. The level controller increases the
opening of the valve in the liquid product line to maintain the level near its set point.
Two important features of this system become clear when observing its dy
namic behavior:

1. The single-loop controllers are completely independent algorithms that do not

communicate directly among themselves.
2. The manipulations made by one controller can influence other controlled vari
ables; that is, there can be interaction through the process among the individual
control loops.

The interaction is the key effect addressed in this chapter, where we will demon
strate that several single-loop controllers on a process should not generally be
analyzed as though each were a single-loop system.
We shall use the following definition of interaction.

A multivariable process is said to have interaction whenprocess input (manipulated)

variables affect more than One^ process output (controlled) variable.

This definition is consistent with the use of the word in the vernacular and will
serve us in the study of multivariable systems. However, the definition does not 621
distinguish between various important properties that will be introduced in this
chapter. Thus, careful attention must be paid to the effects of various types of Modelling and
interaction on control stability and performance. Transfer Functions
In this chapter the basic principles of multiloop control are presented, with
the goal of understanding multivariable systems. As with single-loop control, we
start with the process by reviewing modelling approaches for multivariable pro
cesses and developing models for two sample systems, which will be used in later
examples. Then the concept of interaction is discussed to highlight its effects on
system behavior, and a quantitative measure of interaction is introduced. Finally,
some approaches for tuning multiloop controllers are presented. All of the con
cepts developed in this chapter are employed in the next chapter, which addresses
the performance of multiloop control systems.

20.2 n MODELLING AND TRANSFER FUNCTIONS

Process models for multivariable control can be derived from fundamental prin
ciples or can be estimated based on empirical data. Regardless of the modelling
method used, the analysis, design, and tuning of multiloop controllers will be
based on linear input-output models employing block diagram manipulation, sta
bility analysis, and frequency response. The following two examples demonstrate
the modelling approaches applied to blending and distillation, and the resulting
models will be employed in several subsequent examples.
EXAMPLE 20.1.
Blending is an important unit operation and is employed in a wide variety of indus
tries, as in the production of gasoline (Stadnicki and Lawler, 1985) and cement <5> c%—-
(Sakr et al., 1988). Typically, the controlled variables in a blending process are
production rate and blended product composition. The blending process in Fig
Solvent
^D
ure 20.2 is modelled with the following assumptions: Pure A
c^h ©
1. The inlet concentrations are constant. <2>
2. Mixing where the flows merge is perfect.
FIGURE 20.2
3. The densities of the solvent and component A are equal.
Example blending process.
The overall and component A material balances at the point of mixing are
Fm = FA + FS (20.1)
FmXm = FAXA + F$Xs (20.2)

where Fm — flow rate of mixed liquid (mass/time)

XA = mass fraction of component A in pure A = 1.0
Xs = mass fraction of component A in solvent = 0.0
Xm = mass fraction of component A in the mixed liquid

Equation (20.2) can be linearized about the steady state to give

Fs -Fa
Kit) F'Ait) + Kit) (20.3)
■ i iFs + FA)2]s ViFs + Fta)2 I
with the prime indicating deviation variables. The system is liquid-filled; thus, there
is essentially no delay between a change in a component flow rate and a change
622 in the mixed-product flow rate. It is also assumed that the concentration at the
location of the analyzer is essentially the same as at the mixing point; that is, there
CHAPTER 20 is no transportation dead time. Also, the inlet flow measurements are assumed
Multiloop Control: to be exactly and instantaneously equal to the actual flows, F, = Fs and F2 =
Effects of Interaction FA. The dynamics of the mixed stream flow and concentration sensors are not
instantaneous and are characterized by a first-order-with-dead-time model with
gains of 1.0 and the following dynamic parameters:

Dead time Time constant

Flow 0F xF
Concentration A
e *a

Thus, the measured controlled variables are related to the instantaneous process
variables in equations (20.1) and (20.3) by

dA\jt) = X ' i t - 9 A ) - A \ i t ) (20.4)

"Ca-
dt
xf-dF^jt) = F'mit-eF)-Fdt) (20.5)
dt
Equations (20.1) and (20.3) to (20.5) can be combined to give the following
linearized dynamic model:

\ ~Fa 1 e-*A> e
LiFs + FA)2\s LiFs + FA— * Fds)
Axis) = Fxis) + (20.6)
1 +xAs 1 + xAs

l.Oe"^ 1.0*-''*
F d s ) =1 +^XFS
^ F x i s ) +1 +XpS
^—Fds) (20.7)

Clearly, interaction is present in this process, because each output is affected

by both inputs. Numerical values will be determined for different operating condi
tions later in this chapter.

EXAMPLE 20.2.
The empirical identification procedures described in Chapter 6 can be applied to
the distillation process shown in Figure 20.3. (This design was originally suggested
by McAvoy and Weischedel (1981) and was approximated for constant relative
volatility by Sampath (1991).) The manipulated variables are reflux and reboiler
flow rates, and the controlled variables are distillate and bottoms composition.
Other important variables, such as pressure and levels, are controlled tightly as
shown.
One experiment must be performed for each input variable, and the responses
of all output variables (after 2 min analyzer dead time) are recorded. Either the
process reaction curve or statistical methods can be used to fit parameters in the
transfer functions. The models derived by this empirical procedure are as follows
 Relative volatility 2.4 623
~§ Number of trays
Feed tray
17
9
Analyzer dead times 2 min Modelling and
Transfer Functions
Feed light key XF = 0.50
Distillate light key Xd = 0.98 mole fraction
r\ (\ I7R Bottoms light key *b = 0.02 mole fraction
Feed flow FF = lO.Okgmole/min
■ckj-
^ Reflux flow Fr = 8.53 kgmole/min

-c&»
§ Distillate flow
Reboiler flow
Fd = 5.0 kgmole/min
Fv = 13.53 kgmole/min
Tray holdup H = 1.0 kgmole
Holdup in drums HD = 10.0 kg mole

FIGURE 20.3

Example distillation tower.

with time in minutes:

v , x 0.0747<T3s „ , N 0.0661e-2s „ , v 0.70<r5v
Xois) = ,„_ , , FRis) - —r^—r-Fvis) + , A A , , XFis) (20.8)
125 + 1 15*+ 1 14.45 + 1
v , , 0.1173<r335 „ , x 0.1253*-* „ , x 1.3e-3'
X5(5) = ,,„ , , FRis) - ————-Fv(5) + XF(5) (20.9)
11.75 + 1 10.25 + 1 125+1
Note that the reflux flow iFR) and amount vaporized in the reboiler (Fv) are
potential manipulated variables, and the feed composition iXF) is a disturbance,
because it depends on upstream operations and is assumed not free to be ad
justed.

Finally, the linearized models in Examples 20.1 and 20.2 will be used in
subsequent system analysis examples. When the dynamic responses are determined
via simulation, the linearized distillation model will be used, but the nonlinear
blending model will be used because of the large range of operating conditions
considered in the blending examples.
Linearized models, whether derived from fundamental balances or from exper
iments, can be used to analyze the system with and without control. To understand
the entire system, it is helpful to present the process in a block diagram. The block
diagram of a general 2x2 system, recalling that each process transfer function
relates one input to one output, is shown in Figure 20.4. Each term G,j is) relates
manipulated input j to output i, and the terms Gjds) relate the effects of a dis
turbance on each process output. If more than one important disturbance is to be
considered, additional disturbance transfer functions can be included. Note that if
both Gxds) and G2xis) [or alternatively Gxxis) and G22is)] are zero, the process
has no interaction, because one input affects only one output. In such a case, the
system behaves like two independent processes, and the behavior of each control
loop is independent.
624 MV,(s)
Gnis) ■© ▶© —▶ C V, ( j )

CHAPTER 20
Multiloop Control: Gl2is) Gdxis)
Effects of Interaction
-Dis)
G2iis) G^is)

G22is) -© Kt>—*■ CV2(5)

MV2is)
FIGURE 20.4
Block diagram of 2 x 2 open-loop system.

The set of simultaneous equations relating inputs to outputs in Figure 20.4 are
often presented in matrix form as follows:

\CVxis)] _ \Gxxis) Gxds)] \MVxis)] \Gdds)l 2Q m

[CV2is)\ ~ [G2xis) G22is)\ [MVds) J + [Gd2is) J D{S) W10'
Each element of the matrix is a transfer function relating one input to one
output. Thus:

Linear models for multivariable systems can be developed using the same analytical
and empirical procedures as for single-variable systems.

20.3 n INFLUENCE OF INTERACTION ON THE POSSIBILITY

OF FEEDBACK CONTROL
Previously, some basic requirements were stated for the variables involved in a
single-loop feedback control system. Briefly, the controlled variable should be
closely related to process performance; the manipulated variable should be in
dependently adjustable; there should be a causal relationship between the ma
nipulated and controlled variables; and the dynamics should be favorable. These
guidelines are still useful, but a somewhat more thorough analysis is required for
multivariable systems, because range and controllability are influenced by process
interactions.

Operating Window
The first issue is the control system's range of attainable variable values. The term
operating window will be used for the range of possible (or feasible) steady-state
values of process variables that can be achieved with the equipment available. The
operating window can be sketched using different variables as coordinates; in one
approach, the controlled variables are used to characterize the range of possible
set points, with all disturbances constant. Another common approach is to use
the disturbance variables as coordinates to characterize the range of disturbance
values that can be compensated by the control system (i.e., for which the con-
 + F,Smax
Amax
625

Influence of
Interaction on the
Possibility of
cb F<Smax.
Feedback Control

Feasible

4max

100
Mixed stream concentration

FIGURE 20.5

Operating window for blending with controlled variables as

coordinates.

trolled variables can be maintained at constant set points). The two approaches are
demonstrated in the following examples.
EXAMPLE 20.3.
The component flow rates in the blending example can be adjusted continuously
from zero to maximum rates, FAtmK and FSmax. Draw the operating window of attain
able total flow rate and composition, assuming that the component compositions
<
£
>
Solvent
t&r-
©
remain unchanged.
The attainable total flow F3 and composition Ax are shown in Figure 20.5. The
Pure A
limiting values are easily determined by solving equations (20.1) and (20.2) for
various values of one flow, with the other flow at its maximum value. The interaction t$*
between variables is clear, because the value of one variable influences the range
of the other variable. If the variables were independent and no interaction occurred,
the operating window would be rectangular, which it clearly is not.

EXAMPLE 20.4. <£>

The feed flow rate and composition to the distillation tower in Example 20.2 change
over ranges of 8 to 12 kmole/min and mole fraction 0.4 to 0.6, respectively. Also, the
vapor condensed in the condenser cannot be greater than 15.0 kmole/min. Deter r~\ C
L
"
mine the range of disturbances for which the product qualities can be maintained ■***■ — &
at 0.98 and 0.02 mole fraction. &
The method for calculating the operating window for this example depends on -c&i— -
the equation-solving methods available. A trial-and-error method could be used to
specify the disturbances and simulate the tower with XD and XB at their set points.
This trial-and-error procedure, involving many simulations, would be executed
until the disturbance value that resulted in the maximum overhead vapor flow was ^ % ^
found. A direct method of solving this problem would be to specify XD and XB and
calculate the feed composition XF that resulted in the overhead vapor flow meeting
 1
626 0.6

CHAPTER 20 \
Multiloop Control:
Effects of Interaction \
Feasible
1 \
"3 o
E \
Base case

\
\
0.4
12
Feed rate (kmole/min)
FIGURE 20.6

Operating window for distillation with disturbance

variables as coordinates.

its maximum limit; this approach is possible with a steady-state model solved
using an equation-based approach (Perkins, 1984). The results of the analysis,
performed by either method, are the feasible values of feed rate and composition,
with XD and XB maintained at their desired values; the operating window is given
in Figure 20.6. Again, the interaction is apparent by the shape of the operating
window. The maximum feed rate is attainable with a feed containing the least light
key, because the least amount of distillate product is generated by this feed and
the least distillate requires the minimum overhead vapor.

Controllability
Another important issue in multivariable control is the independence of the input-
output process relationships between selected manipulated variables (MV/s) and
controlled variables (CV,'s); a process in which the relationships are independent
is termed controllable. Many definitions for the term controllability are used in
automatic control (e.g., Franklin et al., 1990); for the purposes of this book we will
use the following definition, which is appropriate for continuously operating plants
that should attain steady-state conditions (a somewhat less restrictive version of
Rosenbrock's (1974) "functionally controllable (f)"):

A system is controllable if the controlled variables can be maintained at their set

points, in the steady state, in spite of disturbances entering the system.

Controllability is defined for a selected set of manipulated and controlled variables,

and a system may be controllable for one selection and uncontrollable for another
selection. A system's controllability is not always easy to determine by observa-
tion; thus, a quantitative method for determining controllability is presented in this 627
section. There is no general method for nonlinear systems; therefore, the control
lability of the locally linearized system will be analyzed to evaluate the system. As Influence of
a result, the results of the controllability test are strictly valid only at the operating Interaction on the
Possibility of
point at which the linear model is evaluated. Feedback Control
The multivariable dynamic system can be described by a model of the form
given in equation (20.10); only a 2 x 2 system is given, but the extension to
higher orders is straightforward. We will assume that the system begins at steady
state. The definition of controllability will be met if the controlled variables can be
maintained at their set points, so that their deviation variables are zero, by adjusting
the specified manipulated variables in the presence of steplike disturbances, which
achieve a constant value, at least asymptotically. The behavior of the system at
steady state can be determined through the final value theorem. As noted in Chapter
4, the final value theorem can be applied if the output is bounded, which excludes
bounded input-bounded output unstable systems. Applying the final value theorem
to equation (20.10), with CV/(j) = 0 for all /, gives

[SHaMMS]' «»»>
with Kij = lim Gy is) being the steady-state gains.
TJ▶O
—
The system is controllable if there is a solution for this set of linear alge
braic equations for arbitrary nonzero values of Kdx, Kd2, and D' (i.e., all possible
disturbances).

A solution exists for a square system of linear equations (20.11) when an inverse to
the matrix of feedback process gains (K) exists; thus, the system is controllable if
the determinant of the gain matrix is nonzero.

A square physical system (numbers of manipulated and controlled variables are

constant) is not controllable if any of the following conditions occurs:

1. Any two process inputs are linearly dependent (giving dependent columns).
2. Any two process outputs are linearly dependent (giving dependent rows).
3. A process output is not influenced by any input (giving a column of zeros).
4. A process input does not influence any output (giving a row of zeros).

The controllability test is applied to the two processes in the following example to
ensure that they are controllable.
EXAMPLE 20.5.
Evaluate the controllability of the blending and distillation processes.
The gain matrices and their determinants are
f -FA F± -Fs + FA
Blending iFs + FA)2 iFs + FA)2 Determinant: 7^0.0
1.0 1.0 iFs + FA)2
_. ,.„ .. T 0.0747 -0.0667] Determinant: -0.001536 # 0.0
Dist.llat.on ^1173 _0.l253j
628 Since each determinant is nonzero, each process is controllable for the selected
manipulated and controlled variables.
CHAPTER 20
Multiloop Control:
Effects of Interaction
Note that a controllable system indicates that the manipulated variables can
compensate for effects of disturbances on selected controlled variables for some
small region over which the linearization is valid and constraints are not encoun
tered in the manipulated variables. In contrast, the operating window, which is
evaluated using the nonlinear steady-state models including constraints, defines
the entire possible region of operation. Both analyses should be performed to
ensure the possibility of multivariable control.
Finally, the controllability and range of the system are affected by the pro
cess design and operating conditions, along with the selected controlled variables.
Therefore, deficiencies in controllability and range must be compensated through
changes to the equipment or process operating point, not control algorithms.

20.4 n PROCESS INTERACTION: IMPORTANT EFFECTS

ON MULTIVARIABLE SYSTEM BEHAVIOR
We now continue investigating the effects of interaction on multivariable system
behavior, assuming that the process has a controllable input-output selection. The
goal of this section is to demonstrate how the responses of a control system are
influenced by interaction. To simplify the analysis, only relationships for two-
input, two-output systems are considered, but the results obtained can be extended
to control systems of higher order. Insights will be provided in this section through
analyzing several examples and are formalized in the next section.
The first step is to derive the transfer function for the multiloop feedback
control system and determine the main differences from single-loop control. We
begin this procedure by considering the same system (1) without control, (2) with
one controller, and finally (3) with two controllers. First, suppose that a single
controller were to be implemented on the system in Figure 20.4, with the goal
of controlling CVxis) by adjusting MViCs). The transfer function Gu(s) would
have to be considered when tuning the controller, as demonstrated by the transfer
function:

CVxis) = Gxxis) no control

(20.12)
MVxis)

In this case, the control loop could be considered a single-loop system; however,
changes in MVi(j) caused by the controller would affect CV2is) because of in
teraction.
Next, we consider a more complex structure to determine whether it affects
the first loop. The block diagram for a multivariable process with one single-
loop controller is given in Figure 20.7. This example is considered to demonstrate
the effects of interaction on closed-loop systems. The transfer function relating
MVi(j) with CViOs) would have to be considered when tuning the controller
using these measured and manipulated variables. This transfer function follows
 MVAs) 629
Gxxis) <+) H±> CV,(5)

Process Interaction:
C,2(5) GrfiW Important Effects on
Multivariable System
-Dis) Behavior
G21(5) Gjiis)

SP2(*) -K>1 Gcl(s) C22(5) Kt) »®b?r-»» CV2(s)

MV2(i)

FIGURE 20.7
Block diagram of 2 x 2 system with one single-loop controller.

for the case with Gc2is) implemented:

CVi(j) = Gxxis)- Gx2js)G2xis)Gc2js)

Gc2 is) implemented (20.13)
MVxis) 1 + Gc2is)G22is)

This equation differs from the transfer function with no control of CV2is), equa
tion (20.12), by the second term, and the path represented by the second term is
shown as a dashed line in Figure 20.7. Clearly, this path results from the process
interaction and the second controller. The second term on the right-hand side in
equation (20.13) would be zero if either or both Gxds) and G2lis) were zero,
in which case the controller Gc2is) would have no effect on the transfer func
tion for CVxis)/MVxis). The path shown with the dashed line will be referred
to as transmission interaction and will be seen to have an important influence on
stability.

Transmission interaction exists when a change in the set point of a controller affects
its controlled variable through a path that includes another controlled variable and
controller.

Note that it is possible to have process interaction [i.e., only Gxds) or G2x is) non
zero] without having transmission interaction, which requires both to be nonzero.
The control design can be completed by applying two single-loop controllers to
the process, as shown in Figure 20.8. The following closed-loop transfer functions
can be determined from block diagram manipulation. (The results for the other
controlled variable, CV2is), can be obtained by transposing the subscripts.)

CVxis) Gcxis)Gxxis) + Gcxis)Gc2[Gxxis)G22is) - Gl2is)G\2is)]

SPxis) CEis)
(20.14)

CVxis) Gc2is)Gl2is)
(20.15)
SPds) CEis)
630
MV,(s)
CHAPTER 20 SP,(5) Gdis) Gxxis) •Kp—^0-L^ CVAs)
Multiloop Control:
Effects of Interaction Gl2is) Gjxis) •*—
-Dis)
G2lis) Gd-s)

sp2(*) -K> Gc2is) G22is) -<±±> ) M j H - * * CV2is)

MV2is)

FIGURE 20.8
Block diagram of 2 x 2 system with two single-loop controller.

Gd2Gxds)G
Gdds) £^U[1 + Gc2is)G22is)]
CVi(s) [l + Gc2(s)G 22(^)]J
(20.16)
Dis) CEis)
with the characteristic expression CE(s), which is the same for equations (20.14)
through (20.16),

CECs) = 1 + Gcxis)Gxxis) + Gc2is)G22is)

(20.17)
+ Gcxis)Gc2is)[Gxxis)G22is) - Gl2is)G2xis)]

When both interaction terms Gxds) and G2i(.s) are nonzero, the dynamic
response of a single-loop controller between CVi is) and MVi is) depends on all
terms in the closed-loop transfer function. As a result, the stability and performance
of loop 1 depend on the tuning of loop 2. By a similar argument, the stability and
performance of loop 2 depend on the tuning of loop 1. Therefore,

The two controllers must be tuned simultaneously to achieved desired stability and
performance.

Further insight can be obtained by considering the steady-state behavior of the

multivariable system. In particular, the necessary adjustments in the manipulated
variables can be used as an indication of how interaction changes the system's
behavior. The general steady-state relationship for a 2 x 2 system is expressed
here in deviation variables:

CV\ = KnMV\ + Kl2MV2 (20.18)

CV2 = K2xMV\ + K22MV'2 (20.19)

These equations are often written in matrix form as

KHK] *"-[££] (20-20)

Equation (20.20) can be rearranged to give 631
MV Process Interaction:
(20.21)
MV 2J"K Lcv2J Important Effects on
Multivariable System
where K~* is the inverse of the steady-state gain matrix and exists for a controllable Behavior
system. Note that equation (20.21) represents the calculation performed by the
controller with zero steady-state offset. For example, equation (20.21) could be
used to determine the steady-state changes in MV\ and MV2 for any specified
changes in CV\ and CV2 (i.e., set point changes). Several hypothetical systems
are considered first so that the extent of interaction can be changed incrementally
from the base model; then some realistic processes are considered.
The process gain matrices in Table 20.1 represent hypothetical systems with
various extents of interaction: A has no interaction iKx2 = K2x = 0); B has
moderate transmission interaction; C has strong transmission interaction; D is not
controllable (the determinant of the gain matrix is zero), and E has one-way in
teraction (£21 = 0). Thus, their behaviors are expected to vary. In particular,
multivariable control is not possible with system D, because it is not possible to

TABLE 20.1
Summary of manipulated-variable changes for example
systems with differing amounts of interaction
Process Inverse
gain matrix gain matrix cv; = 1.0
System K K"1 cv2 = 0.0
MV, = 1.0
No interaction 1.0 1.0 0.0 MV'2 = 0.0
0.0 .[
0.0 1.0 Same as
single-loop
B MV, = 2.29
Moderate "1.0 0.75] r 2.29 -1.71] MV'2 = -1.71
transmission 0.75 1.0 J [-1.71 2.29J Larger than
interaction single-loop

MV, = 5.26
Strong 1.0 0.90 5.26 -4.74] MV2 = -4.74
transmission 0.90 1.0 -4.74 5.26J Much larger than
interaction single-loop

D Singular;
Not inverse does
fl.O 1.0]
controllable [_1.0 1.0J not exist

E mv; = 1.0
One-way 1.0 1.0 1.0 1.0 MV'2 = 0.0
interaction 0.0 1.0 0.0 1.0 Same as
single-loop
632 control CV^ and CV2 independently. This system is not considered further, be
cause a process design change would be required to control the selected controlled
CHAPTER 20 variables.
Multiloop Control: The changes in the manipulated variables required for the specified CV changes
Effects of Interaction
are given in Table 20.1. The manipulated variable changes for systems B and C with
transmission (two-way) interaction differ from the single-loop values reported as
system A. Also, the differences in manipulated variable behavior from system A
increase with increases in the interaction terms. For cases reported in the table,
the manipulations for systems B and C are greater than those for system A, but
for other specified CV changes, systems B and C could be smaller than system
A. The following points summarize the major differences in steady-state behavior
between single-loop and multivariable systems.

1. The values of the manipulated variables that satisfy the desired controlled vari
ables must be determined simultaneously.
2. Differences between single-variable and multivariable behavior increase as the
transmission (two-way) interaction increases.

Before we conclude this section, two examples of process gain matrices are con
sidered. These examples demonstrate that the behavior shown in Table 20.1 occurs
in realistic chemical processes.

(£>t&- EXAMPLE 20.6.

Solvent
(J The first is the blending system shown in Figure 20.2, where the product flow and
composition are controlled by adjusting the flows of the two component streams.
Pure A
t$y (sf The gains are determined from the linearized model in equations (20.6) and (20.7).
The base conditions are taken to be
0^ Fj = 95.0 kg/min F2 = 5.0 kg/min
(20.22)
Ai = 0.05 wt fraction A F3 = 100 kg/min

The gain matrix and its inverse for these conditions are

Ax -0.0005 0.0095 Fx K- l _
~ |_[-100 0.95]J
100 0.05
Fs ]-[ 1.0 1.0 F2
(20.23)

The gain and inverse matrices have one element that is nearly zero. Thus, the
system is likely to behave similar to system E in Table 20.1. As a result, this system
is not expected to experience very strong departures from the single-variable
^ behavior in manipulated-variable adjustment magnitudes.
mm
/*\ (\_~HEr- EXAMPLE 20.7.
— * & 3—A The second example is the binary distillation tower in Figure 20.3, where the prod
$"
-c£k" uct compositions are controlled by adjusting the reflux and reboiler flows. The
steady-state gains can be taken from the transfer function matrix in equations
(20.8) and (20.9).
;€fc
^ ^ - ^ - Q [0.0747 -0.0667] y _ [81.58 -43.42] (20.24)
|_0.1173 -0.1253J ~ [76.36 -48.63 J
The distillation tower appears highly interactive in the two-way manner similar to 633
systems B and C. To complete this distillation example, steady-state changes in
manipulated variables are calculated for single-loop and multivariable control. In Process Interaction:
both cases, the bottoms mole fraction of light key is to be decreased by 0.01. In The Relative Gain
the first case, only the bottoms mole fraction is specified and the distillate mole Array (RGA)
fraction is not controlled. This is single-loop control, and the necessary change in
vaporization is
Single-loop: AFV = B = ~ ' = 0.0798 kmole/min
Kxb.v —0.1253
Since the bottoms composition is not controlled, AFR = 0 and AXD £ 0. In the
alternative multivariable case, the distillate mole fraction is maintained unchanged
(AXD = 0), while the bottoms composition is changed by -0.01.

Multivariable: [AFS1 = [81.53 -43.42] [ 0 ]_ [0.4343]

[AFV\ [76.36 -48.63J [-0.01 J ~ |_0.4863J
The results demonstrate that the change in the vaporization in the reboiler
was much larger in magnitude for the multivariable system (0.4863 compared with
0.0798 kmole/min), and in addition, a large change in reflux was required. Clearly,
the interaction has strongly affected the steady-state behavior of the system.

In conclusion, interaction can strongly influence the steady-state and dynamic

behavior of multivariable systems. There exists a range of interaction from com
pletely independent through nearly dependent (i.e., nearly singular), with this in
teraction dependent on the process characteristics, not on control. In general, the
closer the system approaches singularity (system D in Table 20.1), the more its
behavior differs from that of independent loops. The final two process examples
demonstrated that real processes can have interaction similar to the range of ex
amples in Table 20.1. In the next section, a quantitative measure of interaction is
introduced.

20.5 □ PROCESS INTERACTION: THE RELATIVE GAIN

ARRAY (RGA)
As shown in the previous section, process interaction is an important factor influ
encing the behavior of multivariable systems. A quantitative measure of interaction
is needed to proceed with a multiloop analysis method, and the relative gain ar
ray, which has proved useful in control system analysis, is introduced to meet
this need. The relative gain array was developed by Bristol (1966) and extended
by many engineers, most notably Shinskey (1988) and McAvoy (1983b). In this
section, the relative gain is defined, special properties and methods for calculation
are given, and interpretations for control analysis are presented. The relative gain
array (RGA) is a matrix composed of elements defined as ratios of open-loop to
closed-loop gains as expressed by the following equation, which relates the yth
input and the ith output.
/ dcy \ / acy \
V
\ J' dM
M VV
i=-
con)
s t , f \c ^ / d^ MJV '- )
O mother
c i loops
i u u iopen
j9 uiJtii /on 0^\
kij —
/ dCVj \ (acy- \
\ d M V j ) CWk=const,k^i \ d M V j ) other loops closed
634 Consistent with prior terminology, the open-loop gain (Jfy) is the change in
output i for a change in input j with all other inputs constant (for stable processes).
CHAPTER 20 By closed-loop gain we mean the steady-state relationship between MV'j and CV,'
Multiloop Control: with all other control loops closed (i.e., in automatic). In this definition, it is
Effects of Interaction
assumed that the controllers have an integral mode so that the steady-state values
of the controlled variables are maintained constant (i.e., CV^ = 0 for those under
feedback control). If the relative gain is 1.0, the process gain is unaffected by
the other control loops and no (transmission) interaction exists. Thus, the amount
that the relative gain deviates from 1.0 indicates, in some sense, the "extent" of
transmission interaction in a quantitative manner.
Before control-relevant interpretations of the relative gain are developed, some
important properties must be noted:

1. The relative gain is scale-independent. This is important because rules for

interpretation do not change when the units of a variable change (e.g., from
percent to parts per million).
2. The expression in equation (20.25) suggests that both open- and closed-loop
data is required to determine the relative gain. However, the relative gain can
be calculated from the open-loop data alone, which can be demonstrated by
rearranging equation (20.25) to give

H1= V3MvJMVi=consU,AacvJ
(dCYi\ / 8 M VA (20.26)
CV*=const,*#i

The procedure for calculating the relative gain array is to evaluate the open-
loop gain matrix K; calculate its inverse transposed (K_1)T; and multiply
them in an element-by-element manner. This type of matrix multiplication
is referred to as the Hadamard product (McAvoy, 1983b). The following
expression gives the result for each element in the relative gain array, with
Kjj being the elements in the gain matrix and KI<;- being the elements of the
inverse of the gain matrix,

Ay — Ku^ji (20.27)

For a 2 x 2 system, the (1,1) element of the relative gain array can be shown
to be
1
Xn = (20.28)
1.0- #12^21
Kx 1 #22

3. The rows and columns of the relative gain array sum to 1.0. This property
enables 2x2 systems to be characterized by the Xxx element, as follows:

MV MV2
cv, All 1-All (20.29)
cv2 1 — A.H An

4. The relative gain calculation can be very sensitive to errors in the gain calcu
lation. As an example, consider the following relative gain for a 2 x 2 process,
and assume that each process gain can be in error by a factor e,y-, which is 1.0
 for no error. 535
1
i K-x02 € x 2 )- i K 2 x €
1 —2 x ) — ' T hP er o c e s sR e l a t iIvnet e r a c tGi o
ani n:
iKxx€xxKK22€22) Array(RGA)
When the relative gain element has a large magnitude, the relative gain can take
widely varying values and can even change sign for small errors in individual
process gains, as shown by the following example cases. In this example, the
actual values for the gains are Kxx = K22 = 1.0 and Kl2 = K2\ = 0.949,
and the erroneous relative gain is shown for a few example sets of gain errors.

True An eu €]2 621 €22 An calculated with model errors

10 1.0 1.0 1.0 1.0 10.0 No error
10 1.0 1.1 1.0 1.0 100.0
10 1.0 1.2 1.0 1.0 -16.6
10 0.97 1.03 1.03 0.97 -7.8 Only 3% errors

Since the sign of the relative gain is of great importance in control design
decisions, the sensitivity to model errors demonstrated as the foregoing property 4
must be considered, to prevent incorrect results. Thus, great accuracy is required in
the process gains used for calculating the relative gain. Probably the best method
is to derive an analytical model and evaluate the process gains from analytical
derivatives. This can be done for the blending example using the linearized model
and the foregoing property 2:

r ^1 F2
Fi + F2 Fx + F2
[A//] = (20.31)
F2 Fx
Fi + F2 Fx + F2 -
However, few complex industrial processes can be accurately modelled by sets
of equations small enough to be conveniently manipulated analytically by hand,
although advances in algebraic processing by computers could change this situation
in the future. Thus, numerical differentiation using steady-state process simulators
is a common approach to evaluating process gains. In this procedure, a separate
simulation is performed at the base case and at a case with each input MV; changed
a small amount from the base case. The process gains are calculated using the
equation below, and the relative gain array is determined from equation (20.27):
CV/(MV,, MV2, ...,MVj + AMV;,...) - CV,(MVi, MV2,...)
K
AMVj
(20.32)
Special care is required when using this method because of the accuracy re
quired for the relative gain. When numerical differentiation is used, two potential
causes of errors are introduced: the convergence tolerances in solving the equa
tions and the use of approximate rather than exact derivatives. As demonstrated by
McAvoy (1983fc), the convergence tolerances and AMVs used in equation (20.32)
636 in calculating the approximate gain must be reduced until the estimated process
gains are not significantly influenced by further reductions in their values. The
CHAPTER 20 conclusion of this analysis can be stated as follows:
Multiloop Control:
Effects of Interaction

The gains Ktj used for calculating the relative gains must be accurate; the use of
gains from^linearized fundamental models is recommended. Given the typical errors
in empirical model identification, the use of empirically determined process gains
using methods in Chapter 6 is not recommended for calculating relative gains.

Some very useful control-related interpretations based on the RGA are summarized
as follows and will be used in a hierarchical analysis procedure in the next chapter.

*u <o In this case, the open- and closed-loop process gains are of dif
ferent sighs. In a 2 x 2 process, if the single-loop (CV, — MV,)
controller gain were positive for stable feedback control, the same
controller gain would have to be negative for stable multiloop feed
back control. Thus, the sign of the controller gain to retain stability
would depend on the mode of other controllers in the multiloop
system—not a desirable situation.
ku = 0 One situation in which the relative gain is zero occurs when the
open-loop process gain (ACV//AMV/ with the other loops open)
is zero, which indicates no steady-state relationship between the
input and output variables. Thus, the controller with this pairing
can function, if at all, only when other controllers are in automatic.
Again, this is not generally a desirable situation but is acceptable
in special circumstances, as explained in the next chapter.
0 < kn < 1 From equation (20.25), the steady-state loop process gain with the
other loops closed [e.g., equation (20.13)] is larger than the same
process gain with the other loops open.
An = 1 In this situation, there is no transmission interaction, in the sense
that the product of K\2K2x is zero, but either one of the terms may
be nonzero. Thus, a change in MV/ is) is transmitted to CV/ is) only
through Gijis). Note that this does not preclude the possibility that
the manipulated variable might affect another controlled variable
(i.e., one-way interaction).
A; i > 1 From equation (20.25), the steady-state loop process gain with the
other loops closed [e.g., equation (20.13)] is smaller than the same
process gain with the other loops open.
ku = CO When the process gain is zero with the other loops closed, it is not
possible to control the variable in a multiloop system.

As examples, the relative gains for all cases in Table 20.1 and the two process
examples are reported here. (Note that the model for the distillation tower was
developed from very small perturbations in the nonlinear model without noise;
the probability of obtaining an accurate relative gain value from empirical model
fitting is quite small.)
System Relative gain, Xn 637
A 1.0 Process Interaction:
B 2.29 The Relative Gain
C 5.26 Array (RGA)
D CO
E 1.0
Blending kM-F2 = 0.95 Operating conditions in equation (20.22)
Distillation kxD-FR = 6.09 Operating conditions in Figure 20.3

These values are consistent with the previous, qualitative evaluations of in

teraction in that systems with relative gain deviating most from 1.0 deviate most
from single-loop behavior. Note that system E with only one-way interaction has
A.ii = 1.0; in general, the relative gain array is the identity matrix for systems
with a steady-state gain matrix that is lower (or upper) diagonal [i.e., with nonzero
entries only on and below (or above) the diagonal].
Finally, the relative gain can be related directly to the closed-loop transfer
function of a 2 x 2 system. To do this, the definition of relative gain has been
extended by Witcher and McAvoy (1977) to include frequency-dependent terms
by replacing the steady-state gains with the corresponding transfer functions. Thus,
the frequency-dependent relative gain is

1
All is) = (20.33)
1 - Gxds)G2lis)
Gxxis)G22is)
This expression can be used, by setting s = jco, to evaluate the magnitude of the
relative gain elements at various frequencies.
Using the foregoing expression, the characteristic expression (20.17) can be
rewritten as

Gdis)Gcds)Gxxis)G22is)
CEis) = 1 + Gclis)Gxxis) + Gc2is)G22is) +
Au(5)
(20.34)
This analysis demonstrates the fundamental nature of the relative gain and the
close relationship between the relative gain and system stability for 2 x 2 systems.
A summary of the key results for the relative gain array follows:

1. The deviation from single-loop behavior, specifically the transmission interac

tion, is related to the difference of the relative gain element from the value of
1.0.
2. The condition of A,;- < 0 results in multiloop systems that, to maintain accept
able performance, must alter the (CV,- - MVj) controller gain or automatic
status, depending on the status of other controllers.
3. A direct relationship between frequency-dependent relative gain and control
system stability has been demonstrated for 2 x 2 systems.
638 20.6 a EFFECT OF INTERACTION ON STABILITY
AND TUNING OF MULTILOOP CONTROL SYSTEMS
CHAPTER 20 The final major topic in this chapter is controller tuning. Analysis of the closed-
Multiloop Control:
Effects of Interaction loop transfer function demonstrates that interaction influences the characteristic
equation and, therefore, stability; thus, controller tuning must consider interaction
as well as the single-loop feedback process dynamics. The following example
provides further insight into the effect of interaction on stability and tuning.
EXAMPLE 20.8.
A dynamic system with the following model is to be controlled by two PI controllers.
The input-output pairings are 1-1 and 2-2 as shown in Figure 20.8. Determine the
allowable range of tuning constants that yield a stable system.
r \.0el0s 0.15e-l0s-
CVxis) l+2s \+2s
\ MVxis)] (20.35)
CVds) ] ■ 0.15e-l0s \.0e~L0s Lmv2(*)J
L \+2s \+2s J
The example system has transmission interaction, because both off-diagonal
elements are nonzero; thus, it would not be correct to tune each controller inde
pendently. The stability limit is determined by the characteristic expression, given
in equation (20.17). Finding the limiting values of the tuning constants would be an
arduous task because all four controller tuning constants iKc\, Tn, Kc2, and Tn)
appear in the characteristic equation and, therefore, all affect stability simultane
ously. To simplify the calculations and allow graphical presentation of the results,
the integral times of the controllers will be held constant at 3.0 min, which are rea
sonable values, being the sum of the dead time and time constant of each transfer
function. Note that this selection will not necessarily yield the best performance,
but it is a reasonable choice for this example calculation.
With the integral times fixed, the characteristic equation has two remaining
tuning parameters, the controller gains.
CE{s) = 1 + Gods) (20.36)
where

— - - ( ' i H ^ M - i H ^ )
-10s \.0e~L0s 0.15e-h0s0.15e-L0s\
' • ■ K M - i H ^Y s \+2s 1+2* \+2s )
To calculate the stability region, one gain (e.g., Kc2) was given a value, and
the Bode stability analysis was performed to determine the ultimate value of KcX
that defines the stability limit. These calculations involve extensive manipulations
of complex numbers and were therefore performed using a computer program.
The results of the calculations are displayed in Figure 20.9. If there had been no
interaction, the stability region would have encompassed the entire box defined by
values of the controller gains of (0,0) and (3.76,3.76) shown in the figure, because
the tuning of one controller would not have influenced the tuning of the other. As
can be seen, the interactions in this example reduced the allowable values for the
controller gains.
 639

Effect of Interaction
on Stability and
Tuning of Multiloop
Control Systems

1.5 2 2.5 3 3.5

Loop 1 controller gain
K,c\
FIGURE 20.9

Map of stable and unstable controller gain regions for

Example 20.8 with Tn = T,2 = 3.0.

TABLE 20.2
Summary of example tuning for 2 x 2 system
Case Kc\ Tn Kci Tn IAE! IAE2 IAE! + IAE2
Figure 20.10a 0.95 3.0 0.95 3.0 7.22 5.41 12.63
Figure 20.10b 1.40 3.0 0.50 3.0 4.90 10.3 15.2
Figure 20.10c 0.50 3.0 1.40 3.0 13.7 3.67 17.37
Figure 20.11 1.23 1.76 0.89 1.06 3.46 2.46 5.92
iMMWKMUsUyi

EXAMPLE 20.9 i

Although all tuning within the defined region yields a stable system, the control
performance is different for various tunings chosen from within the stable region.
To investigate by example the effect of tuning on performance, three sets of tuning
constants were chosen for the system in Example 20.8 from within the stable area
shown in Figure 20.9. The tuning was selected to have a reasonable gain margin
(i.e., margin from the stability boundary). The simulation results for multiloop PI
controllers responding to a CV, set point change of 1.0 for three different tuning
constants are given in Figure 20.10a through c and tabulated in Table 20.2. Figure
20.10a gives equal weight to both controlled variables. Figure 20.10b gives more
importance to controlled variable 1, whereas Figure 20.10c gives more importance
to controlled variable 2. These results demonstrate that controller tuning influences
multiloop system performance, so tuning can be used as a method for adapting
system performance to conform to specific priorities in the importance of controlled
variables. This result will be exploited in the next chapter.
640

CHAPTER 20
Multiloop Control:
Effects of Interaction

Time Time

Time Time

FIGURE 20.10

Multiloop control: id) with the same gains for both controllers; ib) with loop 1
gain higher.
 641

Effect of Interaction
on Stability and
Tuning of Multiloop
Control Systems

Time Time

Time Time
ic)
FIGURE 20.10 Con't.

Multiloop control (c) with loop 2 gain higher.

Since tuning influences performance, the engineer should be able to use this
flexibility to obtain good control system performance. Three approaches are typi
cally used for tuning multiloop systems, and each is described here.

Trial and Error

Although potentially tedious, a trial-and-error method is often used in practice.
Initial tuning constant values are typically the single-loop values altered for sta
bility, perhaps with the gains reduced by a factor of 2 or more. These initial values
are adjusted through fine tuning, as described in Chapter 9, with trials performed
on a simulation or directly on the process. The final tuning must be conservative
(i.e., not too close to the stability margin) to account for changes in process op
erating conditions that would occur after the trial-and-error procedure has been
completed. Naturally, the success of this approach depends on the expertise of the
engineer, but the approach can reach reasonable results quickly when transmission
interaction is not too strong.

Optimization
An optimization approach, similar to the approach described in Chapter 9 that
optimized a simulated transient response, can be implemented to automate the
642 trial-and-error procedure. This approach would require a computer optimization
of the simulated transient response to obtain good initial values for each control
CHAPTER 20 system (Edgar and Himmelblau, 1988). Optimization is justified when process
Multiloop Control: interaction is strong and the trial-and-error method would be time-consuming or
Effects of Interaction
result in severe process disturbances.
As an example, the tuning for the control system considered in Example 20.9
was optimized for a unit step change in controlled variable 1, assuming equal
importance of the two controlled variables and no other objectives; therefore, the
objective was to minimize the total integral of absolute value of errors (IAEj +
IAE2). The tuning and transient response are given in Figure 20.11 and included
in Table 20.2. The optimization method yielded initial estimates with little engi
neering effort and modest computing resources. The reader is cautioned that the
results in Figure 20.11 are not satisfactory, because of the lack of robustness and
the very aggressive manipulated variable adjustments; a more complete definition
of control performance, including these factors, should be used. However, it does
provide a useful bound for the lowest IAE that can be attained with PI control.

Approximate, Noniterative Approach

A few methods have been proposed for estimating the tuning for multiloop systems
without the time-consuming iterations associated with trial and error or the com
puter computations associated with the optimization approach. The goal of these
methods is to provide initial tuning constants that are much closer than single-loop
tuning constants to the "best" multiloop values. Naturally, fine tuning based on

Time Time

Time Time
FIGURE 20.11
Multiloop control with PI tuning that minimizes £ IAE (tuning is too aggressive).
plant experience is still required. Unfortunately, there is no generally accepted 643
method for quickly estimating multiloop tuning. The method explained here is se mmmmmmmmmm
lected because it provides insight and introduces some key process-related issues. Effect of Interaction
It also provides a useful correlation for many 2x2 systems; however, it is not on Stability and
Tuning of Multiloop
easily extended to higher-order systems. Control Systems
The method takes advantage of simplifications to determine the tuning for
three cases of limiting process dynamics for 2 x 2 systems with PI multiloop
controllers (McAvoy, 1983a, and Marino-Galarraga et al., 1987). In all of these
cases, the relative importances of the controlled variables are considered equal;
this is the most demanding case for tuning, but other situations are considered in
the next chapter as we tailor the performance to control objectives. The general
approach is to establish how much the PI controller tuning must be changed from
single-loop values when applied in a multiloop system.
The basis of the analysis is the closed-loop characteristic expression (20.34)
divided by 1 + Gc2is)G22is), which does not change the stability limit:
\+Gc2is)G22is)/kxxis)~\
CEis) = \ + Gcxis)G xxis) (20.37)
1 -f- Gc2is)G22is) J
As demonstrated in Chapter 10, the closed-loop characteristic expression given by
equation (20.37) determines the stability of the control system. To evaluate poten
tial simplifications, the relative importance of each term must be determined at the
critical frequency of the loop. Since the approach is based on stability analysis,
which considers only the denominator of the closed-loop transfer function, the
same tuning is obtained for all disturbances and set point changes. The method
considers three limiting cases for tuning loop 1: loop 1 much faster than loop 2;
loop 1 much slower; and both loops having the same dynamics. (The following
analysis considers loop 1, but the same results can be obtained for loop 2 by simply
transposing the subscripts.)

LOOP 1 MUCH FASTER THAN LOOP 2. When the loop 1 process is much
faster than loop 2, the term Gc2ijco)G22ijco) is very small at the loop 1 critical fre
quency because of the tendency of processes to have amplitude ratios that decrease
rapidly after the corner frequency (for example, see Figure 10.13b). Assuming that
A.] i is not a strong function of frequency, as is most often true,
1 -\-Gc2ijco)G22ijco)/kxx
% 1.0 (20.38)
1 + Gc2ijco)G22ijco)
which gives
CE(» « 1 + Gcxijco)Gxxij(o) (20.39)

Therefore, the very fast loop 1 in this case can be tuned like a single-loop controller
without interaction.

This result confirms a qualitative argument in which we would consider the inter
action from the slow loop to be a slow disturbance to the very fast loop 1, which
could be tuned using single-loop methods.
644 LOOP 1 MUCH SLOWER THAN LOOP 2. When loop 1 is much slower,
the term for the fast controller, Gc2ijco), would have a very large magnitude at
CHAPTER 20 the critical frequency of loop 1, because the amplitude ratio of the integral mode
Multiloop Control: in Gc2ijco) will have a very high value at a frequency much less than the loop 2
Effects of Interaction
critical frequency (see Figure 10.13/). Therefore, \Gc2ijco)\ » 1.0, which leads
to the following simplification in the characteristic equation:

CE(» * 1 + Gcl(»G,,(ja» rGc2(>)G22aa»)A„1

L Gc2ija))G22ijco) J

C E ( » ^ 1 + G c l ( » Gxxijco) Gxxijco)
\+Gcxijto) (20.41)

As noted, the steady-state relative gain has been used as an approximation

for the frequency-dependent relative gain. In this case, the gain of the process
"seen" by the controller 1 in the multiloop system is changed by 1/Xn from the
single-loop gain iKxx). Therefore,

The slow controller gain can be modified to be the product of the relative gain and the
single-loop tuning, KcML — ikn)iKcSL), to maintain the desired stability margin.
Since the phase lag is not affected, the integral time can retain its single-loop value.

Again, this result seems consistent with a qualitative argument that a very fast
associated loop would "become part of the process" and affect only the closed-
loop process gain.
The tuning result for the slow loop has a potential flaw. When the relative
gain has a value much different from 1.0, the controller gains for the single-loop
and multiloop situations have very different values. Thus, the correct value for the
controller gain depends whether an interacting controller is in automatic or manual!
To ensure that the stability of the slow loop does not depend on the status of the
interacting loop, the slow loop's controller gain is often limited by its single-loop
value, Kcml < Kcsl> If this limit is significantly exceeded, a real-time computer
program could be implemented to monitor the status of the interacting loop and
adjust the controller gain of the slow loop accordingly.

LOOPS 1 AND 2 HAVE THE SAME DYNAMICS. The entire closed-loop

characteristic equation (20.34), as follows, must be considered.
A2is)
CEis)*\+2Ais) + (20.42)

withACs) = Gc\is)Gxxis) — Gc2is)G22is), because the loop dynamics are equal.

With the simplification that all transfer functions in the process, Gyis), have
similar dynamics, the effects of interaction on tuning are completely represented
by the relative gain, and the results can be condensed into detuning correlations
in Figure 20.12a and b (Marino-Galarraga et al., 1987). These figures show how
single-loop tuning must be altered for 2 x 2 multiloop control when all input-output
 645

Effect of Interaction
on Stability and
Tuning of Multiloop
Control Systems

2 2.5 3 3.5 4.5

Steady-state relative gain, k},
ia)

0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Steady-state relative gain, X,,
(*)
FIGURE 20.12
Relationship between single-loop (SL) and multiloop (ML) PI controller
tuning when both loops have similar dynamics.

dynamics are similar. The controller gain is reduced by about a factor of 2.0 as
the relative gain changes from l .0. Also, the integral time increases by a factor of
about 2.0 as the relative gain decreases to 0.5.
Two important conclusions for systems with similar dynamics become appar
ent from this plot:

1. The multiloop controllers must be detuned from their single-loop tuning over
the entire range of relative gain.
2. The change in tuning constants is not very large.

Thus, interaction results in controller detuning, which slows feedback action for
most 2x2 multiloop systems. The tuning results for 2 x 2 PI control presented in
this section are summarized in Table 20.3.
Note that these results are appropriate for systems that satisfy the assump
tions employed. At the current time, there is no approximate method for the gen
eral case with very different dynamics of all paths, Gijis). The trial-and-error or
optimization-based methods must be used in these cases. Also, the importances
of the controlled variables have been assumed to be relatively equal; the case for
unequal importances is covered in the next chapter. The next two examples apply
the tuning approach to realistic processes.
646 TABLE 20.3
Summary off example tuning for 2 x 2 system
CHAPTER 20
Multiloop Control: Situation Characteristic expression Interaction effect
Effects of Interaction
General \+Gcxis)Gxxis) + Gc2is)G22is) Transmission interaction
+Gcxis)Gnis)Gc2is)G22is)/kxxis) affects stability

Loop 1 much faster \+Geiis)Giiis) Loop 1 stability is not

strongly affected by
interaction; use single-loop
tuning
Loop 1 much slower* l+Gciis)Gxlis)/kxx Loop 1 stablility is
affected by the change in
close-loop process gain;
multiply single-loop
controller gain by A.n
Both loops with \ + 2Ais) + Ais)2/kxx Loop 1 stability is affected
equal dynamics With Ais) = Gcxis)Gxxis) = Gc2is)G22is) by changes in gain and
phase; use Figure 20.12
*This approach will lead to a very large controller gain for large A. If the interacting controller is switched to
manual, loop 1 could become unstable. Thus, the additional limit (#c)ml 5 (^c)sl is often applied to
ensure stability for both single- and multiloop systems.

EXAMPLE 20.10.

§>& L Determine initial tuning constants for multiloop PI controllers applied to the blend
Solvent

PureA
(g ing system operating at the conditions given in equations (20.22), 5% A in the
product, and the following sensor dynamics:

( ^
C£J— Dead time Time constant
(2> Flow 1 2 sec
Concentration 15 30 sec

Consider first the A\-F2 and F3-F| controlled-manipulated variable pairing.

The basis for the tuning values is the linear transfer function models in equations
(20.6) and (20.7) with gain values from equation (20.23), and any single-loop tuning
method could be used. The dynamics above indicate that this case fits the situation
having one fast and one slow loop. Referring to Table 20.3, and noting that An =
k22 = 0.95 % 1.0, both the fast and slow loops can be tuned very close to their
single-loop values. The tuning results using the Ciancone single-loop correlations
are summarized in Table 20.4.
A transient response of this system, simulated using the linearized equations,
for a set point change of 0.01 in the mixed concentration, is given in Figure 20.13.
This is a reasonably well-behaved response, which could be fine-tuned as needed.
An important result of this analysis is that the tuning for this loop pairing does not
TABLE 20.4 647
Tuning for the blending system with dilute product (x„, = 0.05, X = 0.95)
Effect of Interaction
Ai-F2 (c o n t r o l l e r F3-F] controller on Stability and
(slow loop) (fast loop) Tuning of Multiloop
Control Systems
TUning term Single-loop Multiloop Single-loop Multiloop
Process gain Ku = 0.0095 AT „ A „ = 0 . 0 1 K22 = 1.0 K22 = 1.0
9/iO + x) 0.333 0.333 0.333 0.333
KcKp 1.0 1.0 1.0 1.0
Ti/iO + x) 0.85 0.85 0.85 0.85
Kc (kg/min/wt fraction) 105.0 100.0 1.0 1.0
Ti (sec) 38.0 38.0 2.6 2.6
litW&MWto*^

0.3

**• 6
<
o
o 5.5
f\ 94.5 -

E o

i i i 93.5
50 100 150 200 100 150 200
Time Time
FIGURE 20.13
Set point response for multiloop blending system in Example 20.10.

change significantly from single-loop to multiloop; in other words, the tuning of the
controllers does not depend on the control status (automatic/manual) of the other
controller. This is a good situation.
Now consider the alternative loop pairing, Ai-F| and F3-F2. Again, the system
consists of a fast and slow loop, so that the same approach can be used. How
ever, in this system, the relative gain has a value far from unity, AM = k22 = 0.05.
Therefore, the response of the slow loop iAx-F\), which has an effective process
gain of Ku/ku, is significantly altered by interaction. The results, using the rec
ommendations in Table 20.3 and the Ciancone single-loop tuning correlations, are
summarized in Table 20.5.
648 TABLE 20.5
Tuning for the blending system with dilute product ixm = 0.05, k = 0.05)
CHAPTER 20
Multiloop Control: A , -Fx pairing Fy-F2 pairing
Effects of Interaction (slow loop) (fast loop)
Tuning term Single-loop Multiloop Single-loop Multiloop

Process gain Kxx =-0.0005 Kxx/kxx = -0.0\ K22 = 1.0 £22 = 1.0
6/iB + x) 0.333 0.333 0.333 0.333
KCKP 1.0 1.0 1.0 1.0
Ti/i0 + r) 0.85 0.85 0.85 0.85
Kc (kg/min/wt fraction) -2000.0 -100.0 1.0 1.0
Ti (sec) 38.0 38.0 2.6 2.6

The transient response of the multiloop system with the multiloop tuning given
in Table 20.5 is essentially the same as that for the previous pairing and is not
shown. However, the single-loop and multiloop tunings are very different in Table
20.5, because the relative gain is much different from 1.0. If both loops are in
automatic, the A] controller gain must be the (small) multiloop value given in the
table. When the F3 controller is in manual, the effective process gain for the Ai
controller changes to its single-loop value (which is lower by a factor of about 20).
A summary of the implications of the multiloop system in Table 20.5 follows:

Tuning of Ax Single-loop (A,) system Multiloop system

Single-loop Good performance Unstable system

iKc = -2000)
Multiloop Poor performance Good performance
iKc = -100) (very slow)

Thus, the controller tuning in Table 20.5 must be matched to the status of
the controllers—a situation to be avoided if possible. This complexity in updating
tuning online suggests that the pairing in Table 20.4, which can have the same
tuning for any combination of loop statuses (since k & 1.0), is a much better
choice.

EXAMPLE 20.11.
Determine initial tuning constants for the distillation tower with the pressure and
level controller pairings given in Figure 20.3, resulting in the model in equations
(20.8) and (20.9). Evaluate the dynamic behavior for a step change in the feed
light key of -0.04 mole fraction light key.
€g=^: y£ This process has similar dynamics for both loops, so that the summary in
Table 20.3 recommends the tuning correlations in Figure 20.12. The large value
TABLE 20.6 649
Tuning analysis for distillation control system
Effect of Interaction
Xd-Fr controller XB-FV controller on Stability and
■ uiiiiivj icnii
Tuning of Multiloop
(A = 6.09) Single-loop Multiloop Single-loop Multiloop Control Systems
Process gain 0.0747 K22 = -0.1253
0/(0 + x) 0.20 0.16
KcKp 1.55 1.7
T,/iO + x) 0.60 0.50
Kc 20.75 Kcsl/2 = 10.4 -13.6 Kcsl/2 = —6.8
T, 9.0 9.0 6.1 6.1

0.03

100
Time
FIGURE 20.14

Example disturbance response for multiloop distillation in Example 20.11.

of the relative gain (6.09) indicates that the controller gains must be reduced by
a factor of 2.0 from their single-loop values, and the integral times can remain
unchanged. The results from applying this approach are given in Table 20.6, and
a dynamic response of the multiloop system, using the multiloop tuning from the
table, is shown in Figure 20.14. The response is well behaved, because the con
trolled variables return to their set points reasonably quickly and the manipulated
variables experience moderate adjustments. Thus, the correlations provide ac
ceptable initial tuning, which can be tailored to specific objectives through fine
tuning.
Ial3iSM?M^»^S!^^
650 20.7 m ADDITIONAL TOPICS IN INTERACTION ANALYSIS
The material on interaction in this chapter is only introductory, and a coverage of
CHAPTER 20
much more material would be required for a mastery of the topic. Some of the key
Multiloop Control:
Effects of Interaction additional topics are reviewed briefly in this section.

Modelling
Models for multivariable control should be developed with their ultimate use
in mind. Recent results on model consistency (Skogestad, 1991; Haggblom and
Waller, 1988) give useful relationships that can be used to verify that linearized
models observe fundamental properties of the nonlinear system. Also, new exper
imental designs (Kwong and MacGregor, 1994) could be of use in obtaining better
empirical estimates of process gains, but even with these careful experimental steps
the use of empirical models for calculating relative gains with large magnitudes is
problematical.

Interaction Measures
The important features of systems with transmission interaction discussed in Sec
tion 20.3 can be developed through singular-value analysis for systems of arbi
trary size and dynamics. The relevant matrix, here the process gain matrix K,
can be decomposed into three matrices, which can be used to determine the di
rections in the CVs that cause the manipulated variables to change the "most"
and the "least" (as measured by the root sum of squares of the changes in the
MVs). Also, the ratio of the largest to the smallest changes in these two direc
tions can be determined and is called the condition number. Clearly, the larger
the condition number, the more interaction affects the multiloop system. Also,
the condition number indicates the sensitivity of the calculation to model errors.
The basic mathematics of this analysis is presented in Ortega (1987), and con
trol applications are given in Barton et al. (1991) and Arkun (1984). The rela
tionship between the relative gain and condition number is given by McAvoy
(19836) and Grosdidier et al. (1985). An alternative measure of interaction has
been proposed by Grosdidier and Morari (1987). Finally, the controllability and
relative gain calculations can be extended to systems with pure integration, such
as liquid levels, by replacing the derivative of the variable idL/dt) with a sur
rogate variable £ and proceeding with the standard method thereafter (McAvoy,
19836).

Frequency-Dependent Measures
The material in this chapter on controllability and interaction relied principally on
steady-state measures. The definition of controllability used here involves steady-
state behavior. An alternative frequency-dependent definition involves the ability
to influence the dynamic trajectory of the output variables and requires that det
Gis) ^ 0 (Rosenbrock, 1974). Since this book deals mainly with continuous pro
cesses operated at specified steady-state conditions, the definition of controllability
used here involves steady-state ico = 0) controllability.
 In addition, the effects of interaction should be evaluated near the critical 651
frequencies of the control loops. Frequency-dependent interaction is discussed by
McAvoy (1981). Conclusions

Tu n i n g
Another approach to tuning multiloop PID controllers that seems to have met
with success is presented by Monica et al. (1988). This method can be extended
to higher-order systems with frequency response calculations. The definition of
modelling errors to be considered in tuning multivariable systems is much more
difficult, because errors in the individual transfer functions and parameters within
an individual transfer function are not independent.

20.8 n CONCLUSIONS

Multiloop process control systems have been introduced, and the important con
cept of process interaction defined. Standard modelling methods can be used to
represent the input-output behavior of the process without control. Interaction—
one input affecting more than one output—is seen to influence the behavior of
multivariable systems. Using the convention that the single-loop controllers are
paired on the 1-1 and 2-2 elements in a two-variable process, interaction occurs
when at least one of the interacting terms, Gxds) or G2x is), is nonzero. The pro
cess model can be employed to determine a useful measure of interaction: the
relative gain array.
Requirements of controllability and values for relative gain are really exten
sions of conditions that are required for good single-loop feedback control, as
summarized in the following table.

Required
condition Single-loop system Multiple-loop system

Controllability A causal relationship n independent, causal

exists between the manipulated relationships exist between
and controlled variables, Kp£0 the manipulated and controlled
variables, det K # 0.
>^,:Wk->-V- H

Since the requirements are less obvious in multiloop systems, the rigorous math
ematical tests are provided.
Transmission interaction—the additional connection path between an input
and output through an interacting controller—occurs when both interacting terms
in a 2 x 2 system are nonzero. Transmission interaction can strongly affect the be
havior of a multivariable system. First, depending on the directions of the desired
changes in controlled variables, it can substantially influence the adjustments re
quired in the manipulated variables. Second, it can influence the system's stability
and proper controller tuning.
652 Some of the results introduced in this chapter are general for all multiloop
systems of any order (n x n), while some are restricted to two-variable (2 x 2)
CHAPTER 20 systems. The following summary is provided to help the reader.
Multiloop Control:
Effects of Interaction

n x n systems 2x2 systems only

Modelling Closed-loop transfer function,

Controllability equations (20.14) to (20.17)
Relative gain array definition, Relationship between RGA and
equation (20.25) stability, equation (20.34)
Relative gain calculation, PI tuning, Section 20.5
equation (20.27)
Interpretations of relative
gain in Section 20.5

Finally, an important interpretation concerning control performance can be

reached from these tuning results by considering a system having similar dynamics
and a relative gain much larger than 1.0. (Many important processes have large
relative gains.) In this system, the multiloop process gain is smaller than the single-
loop process gain by a factor of about 1 /k, as shown in equation (20.25). However,
stability and tuning analysis indicated that the controller gain in the multiloop
system must be reduced from its single-loop value, as shown in Figure 20.12! As a
result, the reduction in effective process gain caused by interaction in the multiloop
system cannot be compensated by an increase in the controller gain; if an attempt
is made to increase the controller gain to improve control performance, the system
will become unstable! Thus, the product KpKc can be small (i.e., very much less
than 1.0) in the interactive system, and feedback adjustments in response to some
disturbances can be very slow because of this "detuning" effect of interaction. This
stability limit for multiloop systems accounts for the very slow return to set point
experienced by some processes with large relative gains.
To this point, general interpretations of multiloop system behavior have been
developed. The many useful insights and quantitative expressions for the effects
of interaction on multivariable behavior and stability will be exploited in the next
chapter on multiloop control performance, in which specific methods for tailoring
control design to performance goals are presented.

REFERENCES
Arkun, Y., B., Monousiouthakis, and A. Palazuglu, "Robustness Analysis of
Process Control Systems, A Case Study of Decoupling in Distillation,"
IEC Proc. Des. Devel., 23, 93-101 (1984).
Barton, G., W. Chan, and J. Perkins, "Interaction between Process Design and
Process Control: The Role of Open-loop Indicators," J. Proc. Cont., I,
161-170(1991).
Bristol, E., "On a New Measure of Interaction for Multivariable Process Con
trol," IEEE Trans. Auto. Control, AC-I1, 133-134 (1966).
Edgar, T., and D. Himmelblau, "Optimization of Chemical Processes," McGraw- 653
Hill, New Yo r k , 1988. MM^mmm^mmMmBmi
Franklin, G., J. Powell, and M. Workman, Digital Control of Dynamic Systems References
(2nd ed.), Addison-Wesley, Reading, MA, 1990.
Grosdidier, P., and M. Morari, "The ll Interaction Measure," IEC Res. 26,
1193-1202(1987).
Grosdidier, P., M. Morari, and B. Holt, "Closed-Loop Properties from Steady-
State Gain Information," IEC Fund., 24, 221-235 (1985).
Haggblom, K., and K. Waller, "Transformations and Consistency Relations
of Distillation Control Structures," AlChEJ., 34,10, 1634-1648(1988).
Kwong, C. W, and J. MacGregor, "Identification for Robust Multivariable
Control: The Design of Experiments," Automatica, 30,1541-1554(1994).
Marino-Galarraga, M., T. McAvoy, and T. Marlin, "Short-Cut Operability
Analysis. 2. Estimation of fx Detuning Parameter for Classical Control
Systems," IEC Res. 26, 511-521 (1987).
Marlin, T., McAvoy, T., M. Marino-Galarraga, and N. Kapoor, "A Short-Cut
Method for Process Control and Operability Analysis," in Morari, M., and
T. McAvoy, Chemical Process Control III, Elsevier, Amsterdam, 1986,
pp. 369-419.
McAvoy, T., "Connection between Relative Gain and Control Loop Stability
and Design," AIChE J., 27, 4, 613-619 (1981).
McAvoy, T., "Some Results on Dynamic Interaction Analysis of Complex
Systems," IEC Proc. Des. Devel., 22, 42^19 (1983a).
McAvoy, T., Interaction Analysis, Instrument Society of America, Research
Triangle Park, NC, 19836.
McAvoy, T., and K. Weischedel, "A Dynamic Comparison of Material Balance
Control of Distillation Columns," Proc. Seventh IFAC Congress, Kyoto,
Japan, paper 107.2,1981.
Monica, T., C. Yu, and W Luyben, "Improved Multiloop Single-Input/Single-
Output Controllers for Multivariable Processes," IEC Res., 27, 969-973
(1988).
Ortega, Matrix Theory, A Second Course, Plenum Press, New York, 1987.
Perkins, J., "Equation Oriented Flowsheeting," in Westerberg, A., and H.
Chien (eds.), Proc. Second Int. Conf Found. Computer Aided Process
Design, CACHE, University of Michigan, Ann Arbor, 1984.
Rosenbrock, H., Computer-Aided Control System Design, Academic Press,
New York, 1974.
Sakr, M., A. Bahgat, and A. Sakr, "Computer-Based Raw Material Blending
Optimization in a Cement Manufacturing Plant," Control and Computers,
16, 3, 75-78 (1988).
Sampath, S., course project, McMaster University, 1991.
Shinskey, F. G., Process Control Systems (3rd ed.), McGraw-Hill, New York,
1988.
Skogestad, S., "Consistency of Steady-State Models Using Insight about Ex
tensive Variables," IEC Res., 30, 654-661 (1991).
Stadnicki, S., and M. Lawler, "An Integrated Planning and Control Package
for Refining Product Blending," Contr. Eng. Conf, pp. 315-322 (1985).
Waller, K., K. Haggblom, P. Sandelin, and D. Finnerman, "Disturbance Sensi
tivity of Distillation Control Structures," AIChE J., 34,5,853-858 (1988).
654 Witcher, M., and T. McAvoy, "Interacting Control Systems: Steady-State and
Dynamic Measurement of Interaction," ISA Trans., 16, 3, 35-41 (1977).
CHAPTER 20 Wood, R., and M. Berry, "Terminal Composition Control of a Binary Distil
Multiloop Control: lation Column," Chem. Eng. ScL, 28, 1707-1717 (1973).
Effects of Interaction

ADDITIONAL RESOURCES
The effects of process interaction were investigated earlier by several researchers,
including

Rijnsdorp, J., "Interaction for Two-Variable Control Systems in Distillation

Columns," I, Automatica, 1, 15-29 (1965) and II, Automatica, 1, 29-51
(1965).

Further information on the effects of interaction on multiloop systems is available

in

Despande, P. (ed.), Multivariable Process Control, Instrument Society of

America, Research Triangle Park, NC, 1989.
Shinskey, F. G., Controlling Multivariable Processes, Instrument Society of
America, Research Triangle Park, NC, 1981.
Skogestad, S., and I. Postlethwaite, Multivariable Feedback Control, Design
and Analysis, Wiley, New York, 1996.

A rigorous criterion for the stability of linear multivariable systems is available

in Despande (1989, just cited) and

Luyben, W., Process Modelling, Simulation, and Control for Chemical Engi
neers (2nd ed.), McGraw-Hill, New York, 1990.

The material in the chapter enables the engineer to evaluate the suitability of candi
date processes and variables for multiloop control quantitatively. Specifically, con
trollability and operating window (or range of operation) can be used to establish
the feasibility (or infeasibility) of feedback control for potential process designs.
Interpretations of the relative gain suggest that only variable pairings with ky > 0
for 2 x 2 systems should normally be considered further (but see Chapter 21 for
important exceptions). Also, the effects of interaction on tuning are demonstrated
by some preliminary tuning rules for 2 x 2 systems. The methods in this chapter
enable the engineer to eliminate some candidate designs as infeasible for multiloop
control, so that future effort can be directed toward evaluating the remaining feasible
candidates.

QUESTIONS
20.1. For the blending process in Figure 20.2, design a control system to control
the following three product variables at independent values: id) the total
flow (F3), ib) the mass fraction of component A, and (c) the mass fraction
 of component S. You may assume that both mass fractions can be measured 655
by the analyzer Ai.
20.2. Answer the following questions for two physical processes: (1) the chemi Questions
cal reactor described in Section C.2 of Appendix C and (2) the same chem
ical reactor with no heat of reaction. Both processes have two feedback PI
controllers: T -* Fc and Ca -*• Cao (with F unchanged).
ia) Does process interaction influence the stability of the closed-loop sys
tem? Provide quantitative analysis to support your conclusion.
ib) Does process interaction influence the dynamic behavior of the closed-
loop system? Explain your answer briefly.
20.3. Prove the statements made in this chapter about the relative gain array:
ia) The elements are scale-independent, ib) The sum of values in a row or
column is 1.0. (c) the A,/; in equation (20.27).
20.4. Verify the closed-loop transfer functions in equations (20.12) through (20.17).
20.5. Answer the following question about controllability.
ia) How must the controllability test be modified when a constraint (bound)
is encountered in one or more manipulated variables?
ib) Develop an alternative definition of controllability and develop a math
ematical test for the situation in which the controlled variables must
only achieve specified values at a single point in time. This might be
valid for batch control or for intercepting a missile.
ic) Relate the definition of controllability used in this chapter to the relative
gain array.
id) How would the test for controllability in Section 20.3 be modified if —C&l-i
the control algorithms were implemented via digital calculation?
ie) How far can one extrapolate the conclusions of the controllability test
to other operating conditions?
20.6. Determine the controllability and possible loop pairings (A. > 0) for the
process in Figure Q20.6 for the following two situations. The feed consists
©-
of only solvent and component A. The manipulated variables are the valves,
and the controlled variables are the level and the composition of A, Ca-
ia) The situation without chemical reaction (i.e., a mixing tank).
FIGURE Q20.6
ib) The situation with a single chemical reaction A -> B, r& = —kC^.
20.7. Consider the CSTR in Figure Q20.7 in which solvent and component A in
solvent (Cao) are mixed. The two streams can be at different temperatures.
A single reaction A -▶ B occurs in the reactor. The rate expression is Solvent Component A
rA = —kCp,, and the heat of reaction can be nonzero. The manipulated C%H
variables are the flow rates of the two inlet streams, and the controlled
variables are the temperature and concentration of A in the reactor.
id) Determine under what conditions the system is controllable.
ib) For the conditions which are controllable, if any, determine allowable
©-
loop pairings (X/y > 0). 0-
20.8. Answer the following questions for a 2 x 2 control system with PI con
trollers. FIGURE Q20.7
656 id) Is it possible for tuning values to exist that would yield a stable multi
loop system and an unstable single-loop system for the same process?
CHAPTER 20 ib) Is it possible for tuning values to exist that would yield an unstable
Multiloop Control: multiloop system and a stable single-loop system for the same process?
Effects of Interaction
ic) State the criteria for the single-loop system in Figure 20.7 to be stable.
id) Suggest a manner for using the results in Example 20.8 in tailoring the
dynamic performance to control system goals.
20.9. The following transfer function was provided by Waller et al. (1987) for
a distillation column with the levels and pressure controlled with single-
loop controllers as in Figure 20.3. The product qualities were not measured
directly; they were inferred from tray temperatures (°C) near the top, T*,
and near the bottom, T\$, trays. The manipulated variables are the reflux,
Fr, and reboiler steam, Fs, both in kg/h. Time is in minutes.
-0.045*r0-5* 0.048*?-,-0.5* -|
r Tds) 1 _ 8.15 + 1 11 5 + T FR{is)]
[Tids)} -0.23«-1'5f 0.55e-°-55 Fs is)}
L 8.15+1 105 + 1 J
Answer the following questions for this system.
id) Determine whether the input-output combination is controllable.
ib) Determine if either loop pairing can be eliminated based on the relative
gains ikij > 0).
ic) Determine the initial tunings for PI controllers for all allowable loop
pairings.
id) Estimate whether the interaction affects the magnitude of the manipu
lated variable changes for a set point change between single-loop and
multiloop control.
20.10. The outlet temperature of the process fluid and the oxygen in the flue gas
can be controlled in the fired heater in Figure Q20.10 by adjusting the fuel
Flue gas
pressure (flow) and the stack damper % open. A dynamic model for the
fired heater in Figure Q20.10 was reported by Zhuang et al. (1987) and is
repeated here:
0.6 -0.04
Ti s ) ' 240052 + 855 + 1 300052+905 + 1 P*(s)
Ais) -1.1 0.30 vds)
705 + 1 705 + 1
The inputs and outputs are in percent of the range of each instrument, and
the time is in seconds.
id) Determine whether the input-output combination is controllable.
ib) Estimate whether the interaction changes the magnitude of the manip
ulated variable changes for a set point change between single-loop and
multiloop control.
(c) Determine if either loop pairing can be eliminated based on the relative
Fuel gains ikij > 0).
FIGURE Q20.10 id) Determine the initial tunings for PI controllers for all allowable loop
pairings.
2 0 . 11 . T h r e e C S T R s w i t h t h e c o n fi g u r a t i o n o f S e c t i o n C . 2 a n d w i t h t h e f o l l o w i n g 6 5 7
design parameters are considered in this example; the common data is given mammmmnM^tM
below, and the unique data and steady states are given in Table Q20.11 for Questions
three cases.
F = 1 m3/min, V = 1 m3, CA0 = 2.0 kmole/m3, Cp = 1 cal/(gK),
p= 106g/m3,fc0= 1.0 x 1010 min-1, E/R = 8330.1 KT1
iFe)s = 15 m3/ min, Cpc = 1 cal/(g K), pc = IO6 g/m3, b = 0.5
The controlled variables are Ca and T, and the manipulated variables
are Cao and Fc. Answer the following questions for each chemical reactor
and explain the differences among the designs. (Note that this question
requires the linearized, steady-state model for each case.)
id) Determine whether the input-output combination is controllable.
ib) Estimate whether the interaction changes the magnitude of the manip
ulated variable changes for a set point change between single-loop and
multiloop control.
ic) Determine if either loop pairing can be eliminated based on the signs
of the relative gains.
id) Determine the initial tunings for PI controllers for all allowable loop
pairings.
ie) Evaluate the transient responses for a concentration set point change
of +0.02 kmole/m3.

TABLE Q20.11

Case I II III

-AtfrxnlO^al/fkmole) 130 13 -30

a (cal/min)/K 1.678 x IO6 1.678 x IO6 0.7746 x IO6
T0K 323 370 370
Tcin K 365 365 420 (heating)
TSK 394 368.3 392.7
Cas kmole/m3 0.265 0.80 0.28

20.12. Discuss an empirical method for identifying the inverse of the process gain
matrix directly from experimental data.
20.13. Determine whether K(K)-1 would give the same (correct) result as equa
tion (20.27) for the elements of the relative gain array.
20.14. The process with two series chemical reactors in Example 3.3 is consid
ered in this question. The process flexibility is increased by allowing the
temperatures of the two reactors to be manipulated independently. The two
controlled variables are the concentrations of reactant A in the two reactors.
658 The rate constant can be expressed as 5.87 x io^-5000/7" (with temperature
inK).
CHAPTER 20 id) Determine whether the input-output combination is controllable.
Multiloop Control: ib) Determine if either loop pairing can be eliminated based on the signs
Effects of Interaction
of the relative gains (X,;- > 0).
20.15. The following transfer functions were provided by Wood and Berry (1973)
for a methanol-water separation in a distillation column similar to Figure
20.3. The products are expressed as mole % light key, and the reflux FR
and reboiler steam Fs are in lb/min. Time is in minutes.
12.8e" -18.9g -3s -x
\XDis)]_ 16.1s +1 215 + 1 \FRis)]
L**(*)J 6.6e"7* -\9.4e~3s LFs(s) J
L 10.95 + 1 14.45 + 1 J
id) Determine whether the input-output combination is controllable.
ib) Estimate whether the interaction changes the magnitude of the manip
ulated variable changes for a set point change between single-loop and
multiloop control.
(c) Determine if either loop pairing can be eliminated based on the sign
of the relative gains (X,;- > 0).
id) Determine the initial tuning for PI controllers for all allowable loop
pairings.
ie) The model was determined empirically. Discuss the effects of likely
model errors on the results in parts id) to id).
20.16. A series of nonisothermal CSTRs shown in Figure Q20.16 is analyzed
in this question. The heat transfer is adjustable in each reactor, so that
each reactor temperature can be considered a manipulated variable. The
feed contains only a nonreacting solvent and component A. The potential

-t&H
'AO

-&y
&
^

' f e -t&r-

FIGURE Q20.16
 manipulated variables are Tx, T2, F, V|, V2, and CAo- The variables to be 659
controlled to independent steady-state values are the compositions of B and
C in the effluent from the second reactor. For each of the sets of elementary Questions
reactions given below, determine (1) for which sets of two manipulated
variables the system would be controllable and (2) for the variables selected
in (1), whether either pairing of variables could be eliminated based on the
relative gain.
ia) A^B-^C
ib) A-^-B + C
Assume that the rate constants can be expressed as Arrhenius functions of
temperature and the heat of reaction is zero.
20.17. The mixing tank in Figure Q20.17 has two independent inlet streams of pure Component B Component A
A and B that can be manipulated. The outlet flow cannot be manipulated by C&H H&J
the unit; it is set by a unit of higher priority. The composition, the weight
percent of B, and the level are to be controlled.
h©
ia) Derive a linearized model of the system.
Flow not
ib) Determine whether the system is controllable. controlled
ic) Calculate the relative gain array for this process and make conclusions
about the possible loop pairings for this system. %A
©- at this unit

20.18. A proposal is made to control the temperature (T) and composition (Ca) FIGURE Q20.17
in the chemical reactor in Figure Q20.18 by manipulating the feed flow
and the inlet temperature. The chemical reaction is A -» B, with r^ =
—£Ca and no heat of reaction. The flow in the pipe is laminar, so that
the flow out can be taken to be proportional to level, F = KL. The data
for this system at the base case operation is the same as for Example 3.2;
in addition, the temperature is 323 K and the reaction rate constant is
* = 2.11xlOV500°/7'.
(a) Derive the linearized model for this system in deviation variables.
ib) Determine whether the system is controllable in the steady state.
ic) Derive the four individual single input-output transfer functions.

FIGURE Q20.18
560 id) Evaluate the relative gain, both at steady state and as a function of
i m ^ . : ^ ^ ^ ^ ^ ^ ! . ^ ; ; : 1 ; - ^ ! f r e q u e n c y. E x p l a i n t h e d i f f e r e n c e s .
CHAPTER 20 (e) Select a feasible loop pairing and design a control system.
Multiloop Control:
Effects of Interaction 20.19. Evaluate the controllability and the interaction for the blending and distil
lation processes modelled in Section 20.2. Discuss the differences, if any,
between the steady-state and frequency-dependent results.
20.20. The analysis of multiloop tuning summarized in Figure 20.9 considered
only positive controller gains. Discuss the control performance when one
of the controller gains is allowed to be negative.
 Multiloop
Control:
Performance
Analysis /n-i
21.1 m INTRODUCTION
Multiloop process control systems were introduced in the previous chapter, where
some important effects of interaction on steady-state and dynamic behavior were
explained, and a quantitative measure of interaction—the relative gain—was pre
sented. This understanding of interaction is now applied in the analysis of multiloop
control performance and design. Three main facets of control performance analy
sis are presented and applied to the design of multiloop systems. The first is loop
pairing: deciding the controlled and manipulated variables for each single-loop
controller in a multiloop system. The second facet is controller tuning to achieve
the desired performance, as well as to maintain stability. The third facet involves
enhancements to the PID control calculations that can improve control perfor
mance while retaining the simplicity of the multiloop control strategy in selected
applications.
As in the single-loop case, the first step is to define control objectives thor
oughly. The main aspects of multivariable control performance are presented in
the following list. Several are the same as for single-loop systems; however, items
2, 5, and 6 are new, and item 4 can assume even greater importance.

1. Dynamic behavior of the controlled variables. The control system should

provide the desired control performance for expected disturbances and set
point changes. The performance can be defined by any appropriate measures
presented in Chapter 9 (e.g., IAE and decay ratio).
2. Relative importance among controlled variables. The multiloop control struc
ture should be compatible with the relative importance of various controlled
662 variables, since some controlled variables may be very important and should
be maintained close to their set points, while others may not be as important
CHAPTER 21 and can be allowed to experience larger short-term deviations.
Multiloop Control: 3. Dynamic behavior of the manipulated variables. Feedback control reduces
Performance Analysis
the variability in the controlled variables by adjusting manipulated variables;
however, the variability in the manipulated variables should not be too large.
4. Robustness to model errors. The control system should be robust so that it
performs well in spite of inevitable modelling errors. As with single-loop
systems, this objective requires that feedback controllers be tuned to ensure
stability and give the best feedback performance possible for the expected
model errors. In addition, we shall see that some multivariable control systems
are highly sensitive to model errors and can be applied only when models are
very accurate.
5. Integrity to controller status changes. Each controller should retain reason
able performance for its basic objectives, even if performance is somewhat
degraded, as changes occur in the automatic/manual status of interacting loops.
6. Proper use of degrees of freedom. The control system should be able to adapt
itself to the degrees of freedom available in the process, which can change
when a manipulated variable cannot be adjusted (e.g., because it reaches a
physical limit). This topic is addressed in Chapter 22.

It would be possible to arrive at the best design by simulating all possible

loop pairings and enhancements. However, simulating the numerous candidate
designs would be a time-consuming task, especially since the controllers in every
candidate would have to be tuned. In addition, such a "brute force" simulation
technique would provide little insight into improving performance through changes
in process equipment, operating conditions, or control structure.
The approaches presented here are selected because they address the most
important issues and generally require less engineering effort than simulating all
possibilities. Because these methods build on the results of the previous chapter,
it will be assumed that all systems considered are controllable. The new analysis
method for each major design decision is addressed in a separate section of the
chapter; then, some advanced topics are introduced. Finally, a flowchart is provided
to clarify the integration of major analysis steps in reducing potential candidate
designs and making decisions for multiloop systems. The hierarchical analysis
method eliminates candidates with a minimum of engineering effort and results in
one or a few final designs. Because of assumptions in some of these methods, the
final design selection may still require simulation, but of only a few candidates. Be
fore the methods are covered, a few motivating examples are presented to highlight
some important issues that distinguish multiloop from single-loop performance.

21.2 m DEMONSTRATION OF KEY MULTILOOP ISSUES

In this section, four important multiloop issues are introduced through process
examples that show the key effects of interaction on the dynamic performance of
multiloop control systems. These issues were selected because they often influence
control design for process units and they are unique to, or assume heightened
significance for, multiloop systems. The analysis methods to address these issues
are provided in subsequent sections of this chapter.
EXAMPLE 21.1. Operating conditions 663
The first issue is the effect of operating conditions on multiloop control perfor
mance, which is introduced through consideration of the blending process in Demonstration of Key
Figure 20.2. We begin by considering the same operating conditions previously Multiloop Issues
considered in Table 20.5, which are repeated in Table 21.1 as the base case.
For these operating conditions, the product is very dilute (5% A). Thus, changing
the flow rate of component A by a small amount affects the product composi
tion significantly while affecting the total product flow only slightly. This qualita
tive analysis was substantiated by the quantitative tuning analysis in Example
20.10, which leads to the recommendation of the pairing for the base case in
<5> cSd-
Solvent
Table 21.1.
Next, we investigate whether a different pairing is recommended for an al
ternative operating condition that involves a very concentrated product (95% A).
In this operation, the product concentration is more sensitive to the flow of the
Pure A
t#h (sf
solvent than to the flow of component A, as it was in the base case. The tuning <2>
for proportional-integral controllers is determined by the guidelines for 2 x 2 sys
tems with one fast and one slow loop. For this alternative case the loop pairings
Ax-Fx and F2-F2 provide better control, because the tunings for the controllers in
this configuration are not dependent on the automatic/manual status of the other
controller. From this example, we can conclude:

The proper control loop pairing depends on the operating conditions of the
process.

Thus, it is not possible to specify a single control design for each unit operation,
like blending or two-product distillation. Even though units may appear similar,
at least with respect to equipment structure, their operating conditions and the
resulting dynamic responses must be considered.

TABLE 21.1
Effect of operating conditions on multiloop performance of the blending system
Relative gain
Operating Set points kAl-F2i kAi-pi
condition A, F3 kF3-Fli kp3-F2 Pairing: A\-F2, F3-F1 Pairing: Ax-Fx, F3-F2
Base case 0.05 100 0.95 0.05 Recommended Not recommended
The controller tuning The controller tuning
is essentially the same for depends strongly on the
single-loop and multiloop status of the interacting
control. loop.
Alternative case 0.95 100 0.05 0.95 Not recommended Recommended
The controller tuning The controller tuning is
depends strongly on the essentially the same for
status of the interacting single-loop and multiloop
loop. control.
immi^m',Msmimm^msmimmmw^«^^mm^MM!mmm
664 EXAMPLE 21.2. Transmission interaction
The previous analysis selected the controller pairing that reduces transmission
CHAPTER 21 interaction. In fact, the best controller pairings for the two examples are consistent
Multiloop Control: with selecting the multiloop pairings that yield relative gain values closest to 1.0, as
Performance Analysis verified by the relative gain values in Table 21.1. Given this result, it is tempting to
assume that the multiloop control with relative gains closest to 1.0 always gives the
best performance. This example demonstrates that this assumption is not always
valid and that a more complete analysis is required.

ia)

0.03

0.98

0.95
100 150 200 150 200
Time ib)

50 100 150 200 150 200

Time
FIGURE 21.1
Energy balance distillation control: (a) schematic diagram; ib) transient response to a change
in light key in feed of -0.04.
 This example consists of the two-product distillation tower separating a binary 665
feed considered in Example 20.2. Both top and bottom product compositions are
of equal importance, and the major disturbance is a change in feed composition. Demonstration of Key
Two regulatory loop pairings, which differ only in how the distillate and reflux flow Multiloop Issues
rates are manipulated, are considered. The first, shown in Figure 21.1a, has the
distillate manipulated to control the overhead drum level and the reflux manipu
lated to control the top product composition; this is called energy balance and was
considered in Chapter 20. The second, shown in Figure 21.2b, has the distillate
and reflux pairings interchanged; this is called material balance and is introduced

(§)
D
o & — H&-
ia)
§™~
-C&H-■

0.03

0.98

0.95
200
ib)

0 50 100 150 200 50 100 150 200

Time Time
FIGURE 21.2

Material balance distillation control: ia) schematic diagram; ib) transient response to a
change in light key in feed of —0.04.
666 TABLE 21.2
Tuning and performance data for distillation dynamics
CHAPTER 21
Multiloop Control: Energy balance Material balance
Performance Analysis
kxD-FB 6.09
kxD-FD 0.39
Kcd 10.4 -9.35
I/O 9.0 10.0
Kcb -6.8 -68.7
Tib 6.1 6.7
Feed
composition IAExd 0.17 0.45
disturbance IAExb 0.35 0.31
iAxf = -0.04)
IAExd 0.35 0.0585
SPxD IAE™ 0.34 0.0456
disturbance
(ASPxd = 0.005)

here for the first time. It is important to recognize that the steady-state responses
of these two systems are identical because the process equipment, controlled
variables, and manipulated variables are the same. Only the transient behavior
is different. The linear transfer functions, including 2 min analyzer dead times, for
the two systems follow.

Energy balance.
- 0.0747c"3v -0.0667c"25 "I 0.70c- 5 s -i

12s+ 1 15s + 1 14.4s + 1

[SI- 0.1173c"3-35
- 11.75* -h 1
-0.1253c"25 [S] +
1.3c"35
L 12s+ 1 J
X, (21.1)

10.2s+ 1 -1
Material balance.
r -0.0747*-* 0.008c-25 "I 0.70c -55 -,

\0s + \ 5s+ 1 14.4s -1-1

[SI- L 9s+ 1
-0.1173c-25 -0.008c"25 [S] +
1.3c"35
L 12s+1 J
XF (21.2)

3s+ 1 -1
Tuning for these control systems can be determined by the methods in Chapter
20. The results are reported in Table 21.2.
The transient responses for well-tuned feedback control in response to a feed
composition upset are given in Figures 21.1b and 21.2b, and the control perfor
mances are summarized in the IAE values in Table 21.2. Based on the total IAE
values (0.52 for energy balance and 0.76 for material balance), the performance
of the energy balance control design is better than the material balance controller
for the feed composition disturbance—in spite of the fact that the interaction, as
measured by the relative gain, is much further from 1.0 for the energy balance
controller pairing. Thus, we conclude:
 667
The best-performing multiloop control system is not always the system with
the least transmission interaction (i.e., with relative gain elements closest Demonstration of Key
to 1.0). Multiloop Issues

This result should not be surprising when one considers the closed-loop trans
fer function for a multiloop system, derived in Chapter 20 and repeated here.
Gd2is)Gl2is)Gc2
CVxis)
Gdlis)- f ^ l fl + Gc2is)G22is)]
[1 + Gc2is)G22lis)] J
(21.3)
Dis) CEis)
with
Glds)Gxxis)Gc2is)G22is)
CEis) = 1 + Gcxis)Gxxis) + Gc2is)G22is) +
kxxis)
The dynamic response depends on all elements in the transfer function, so both
numerator and denominator must be considered, especially in multivariable sys
tems. However, the relative gain appears only in the denominator, whereas the
disturbance transfer function appears in the numerator. This result is a bit disap
pointing, since the design of multiloop systems would have been relatively easy
if the pairing were determined completely by the relative gain. Transmission inter
action is important and must be considered, but a simple pairing method based
entirely on the relative gain is not always correct.

EXAMPLE 21.3. Disturbance type.

A further important question concerns the performance of candidate controls for
different disturbances. Specifically, is it true that one candidate control pairing
performs best for all disturbances? This issue is investigated by extending the
study of the two distillation controller pairings for a different disturbance: a set
point change to the distillate controller. The dynamic responses for a set point
change in the top composition controller of +0.005 mole fraction, with the other
set point and all disturbances constant, are given in Figure 21.3a and b. The
results, summarized in Table 21.2, show that the total IAE values are 0.69 for
energy balance and 0.104 for material balance. In this case, the material balance
system performs better. Note that an attempt to "speed" the sluggish response of
the energy balance system through tighter controller tuning will lead to instability.
From this example we conclude:

The relative performance of control designs and the selection of the best
design can depend on the specific disturbahce(s) considered.

This result seems reasonable when considering the following closed-loop transfer
function for the set point change:
CV,0) Gcxis)Gxxis) + Gcxis)Gc2is)[Gxxis)G22is) - G12(s)G21(s)]
SP,(s) CE(s)
(21.4)
668 0.99 0.025

CHAPTER 21
Multiloop Control:
Performance Analysis

50 100 150 200

Time

50 100 150 200

Time
ia)

0.99 0.025

| 0.02
o
CQ

0.015
200

200

ib)
FIGURE 21.3

Transient response of distillation control to +0.005 distillate light key set point
change: (a) energy balance design; ib) material balance design.
The characteristic equation is unchanged from equation (21.3), but the transfer 669
function numerator is different for different disturbances, and thus the control per
formance could be different. The result again demonstrates the difficulty with hav Demonstration of Key
ing a single, standard design for a unit operation, because the types of distur Multiloop Issues
bances a unit most often experiences depend on the entire plant design.

EXAMPLE 21.4. Interactive dynamics.

The examples covered to this point involved interactive systems in which the trans
mission interaction is not faster than the "direct" transfer function between the
manipulated and controlled variables. Assuming that the controller is paired ac
cording to CVi(s)-MVi(s), the systems studied to this point have had
G2xis)Gc2js)Gx2is)
Gn(s) faster than
1 + Gc2is)G22is)
A particularly difficult control challenge can occur when the transmission inter
action is faster than the direct process response. As an example, two systems
are considered; they have the same steady-state gains, but system B2 has fast
transmission dynamics, whereas system B1 has similar dynamics for all transfer
functions in the process model. In Example 20.9, system B1 has been shown to
have "well-behaved" closed-loop dynamics and to be easily tuned.
System B1.
1.0c - 1 . 0 s 0.75c - l . 0 . v - i
rev, on l+2s l+2s MVxis)
(21.5)
|_CV2(s)J 0.75c"1 °5 !.0c"'°5 MV2(s)
L l+2s l+2s J
System B2.
1.0c -3.05 0.75c -OAs
rcv,(s)"|_ l+2s l+2s MVxis)
(21.6)
LCV2(S)J 0.75c"015 1.0c-°l5 MVds)
L l+2s 1 +2s J
System B2 has the same steady-state gains but very different dynamics. To first
acquire some understanding of this system, the dynamic response is determined
for a step change in MV, it) with only the controller for variable 2 in automatic; this is
the process reaction curve for the process MV] (f)-CVj(r) with the other controller
in automatic. The dynamic response in Figure 21.4a shows an inverse response,
because the fast transmission effect produces an initial negative response before
the slower diagonal [Gnis)] effect produces a positive steady-state response.
It is important to recognize that the structure of a multiloop system with interac
tion ensures that parallel paths exist; the parallel paths include the direct transfer
function and transmission interaction, as shown in Figures 20.7 and 20.8. These
parallel paths do not always create complex feedback dynamics such as inverse
response or initial overshoot, but the possibility always exists. In system B2 the
interactive path is faster and has an effect opposite to the direct effect, leading to
the initial inverse response.
A process with an initial inverse response is usually difficult to control; thus,
interaction with fast transmission dynamics can result in poor control performance.
As an example, the control response of system B2 to a set point change in CV,
with PI tunings that yield minimum (IAE| + IAE2) is given in Figure 21 Ab. (Again,
670

CHAPTER 21
Multiloop Control:
Performance Analysis >
V

Time Time

>
1 -

50
Time Time
ia)

>
u

Time Time

>

Time Time
ib)
FIGURE 21.4

System B2: (a) Process reaction curve of MVi-CVi with other loop closed;
ib) multiloop transient response to set point change in CVj.
TABLE 21.3 671
Effect of dynamics on multiloop performance
Multiloop Control
Case Kcl Tn Kc2 Tn IAE, IAE2 IAE,+IAE2 Performance through
Loop Pairing
B1: Uniform 1.23 1.76 0.89 1.06 3.46 2.46 5.92
interactive
dynamics
(Figure 20.11)
B2: Complex 0.71 3.00 4.00 2.97 9.80 1.27 11.07
interactive
dynamics
(Figure 21.45)

this simple measure of control performance is selected for comparison purposes

only.) The feedback controller cannot eliminate the initial inverse response, which
results in a relatively long time during which CV](0 is far from its set point.
The tuning and performance for systems B1 and B2 are compared in Table
21.3. This example clearly demonstrates the importance of interactive dynamics;
recall that both systems B1 and B2 have the same steady-state interaction, but
system B2 has poorer performance.
This example demonstrates:

Multivariable systems with strong interaction and fast transmission dynam

ics can result in complex dynamic responses, involving inverse response
or large overshoot, which can degrade control performance.

The examples considered in this section have demonstrated that the design of a
multiloop control system is a challenging task, involving more complex issues than
single-loop systems, and that the process dynamic responses, operating conditions,
disturbances, and extent of interaction must all be considered. The next three
sections present methods for considering these issues when making the three main
multiloop decisions: loop pairing, tuning, and enhancements.

21.3 o MULTILOOP CONTROL PERFORMANCE

THROUGH LOOP PAIRING

Loop pairing—the selection of controlled and manipulated variables to be linked

through single-loop controllers—is an extremely important design decision. For
the distillation examples in Figures 21.la, the two possible pairings are (1) XD-
FR and XB-Fy and (2) XD-FV and XB-FR. However, for a system with more
manipulated variables, the number of potential designs becomes very large; in fact,
672 the number of initial candidates for a process with n manipulated and controlled
variables is n factorial (n!). For example, there are 125 candidates for a five-
CHAPTER 21 controller, five-manipulated-variable distillation system in Figure 21.1a when the
Multiloop Control: product compositions, pressure, and levels are considered! Clearly, the number
Performance Analysis of candidates must be reduced significantly, or the analysis task will require an
enormous effort to evaluate all candidates. In this section, four separate analyses are
described for eliminating clearly unacceptable pairing candidates and evaluating
the remainder for likely performance. These analyses would be applied only to
process designs that have been verified to be controllable and to have an adequate
Loop Pairing operating window. Also, the four analyses are employed sequentially, with only
those candidates passing the prior steps evaluated at the next step.
Integrity
Dynamics
Performance Integrity
Range An important factor to be considered in multiloop control design is the performance
of the system when a fault or limitation occurs. Here, a fault is assumed to involve
a sensor or final element so that a control loop ceases to function; we will be
considering the situation after a fault has been recognized and the loop with the
fault has been taken out of service. The resulting situation is the same when one
(or more) controller is placed in the manual status, so that it no longer adjusts the
manipulated variable. In such circumstances, interaction influences the stability
and performance of the remaining closed-loop control system. We would like the
system to have integrity.

A system has integrity if, after one or more loops are placed in manual, the remain
ing closed-loop system can be stable without changing the signs of any feedback
controller gains remaining in automatic.

Some very useful results regarding integrity can be determined from the relative
gain.

NEGATIVE RELATIVE GAIN. If a control loop (with integral mode) is paired

using manipulated and controlled variables that have a negative relative gain ele
ment kij, one of the following situations must exist (McAvoy, 1983; Grosdidier et
al. 1985).

1. The multiloop system is unstable with all controllers in automatic.

2. The single-loop system ij is unstable when all other controllers are in manual.
3. The multiloop system is unstable when the 17th controller is in manual and
all other controllers are in automatic.

Since all three situations are undesirable, the general conclusion is that single-
loop designs should avoid pairings with negative relative gains, whenever possible.
Only when essential, fast feedback dynamics can be achieved only by pairing on a
negative relative gain should this design be considered. Industrial experience has
shown that good designs with loop pairings on a negative relative gain occur very
infrequently. An industrially important example of pairing on a negative relative 673
gain is described by Arbel et al. (1996).
Multiloop Control
Performance through
ZERO RELATIVE GAIN. When the relative gain, ku, is zero for a pairing, Loop Pairing
the steady-state gain of the pairing CV/(f) — MVjit) is zero when the other loops
are open, that is, the process gain Ky = 0. Since no causal relationship exists,
the single-loop controller cannot function. However, the multiloop system can
function because of the causal relationship through the interacting process and
the interacting controller. The causal interaction relationship is demonstrated with
equation (20.13), which gives the transfer function between CV] is) and MVj is)
for a 2 x 2 system with loop 2 in automatic.
0

CVxis)/MVxis)=pfxis)-Gxds)G2xis)Gc2is)/[\ + Gc2is)G22is)] (20.13)

Clearly, a nonzero causal relationship exists between MVj is) and CV] when pro
cess interaction occurs [Gxds)G2xis) ^ 0] and the interacting controller is in
automatic [Gc2is) ^ 0] to create a feedback loop via the interaction path. There
fore, successful operation of a control loop paired on a zero relative gain depends
on the status of the interacting loop. Pairing on a zero relative gain should be im
plemented only when essential, fast feedback dynamics are achieved. Industrial
experience indicates that this situation is not common, but occurs occasionally.
In both of these cases, proper functioning of a control loop requires that the
adjustments from other controllers be implemented at the final elements, which
would not be satisfied if an interactive controller (1) were in manual or (2) had
its output saturated at the upper or lower bound. It is not uncommon for these
situations to occur, at least temporarily, and thus, multiloop control designs with
relative gains less than or equal to zero could often fail to provide stable feed
back regulation. To prevent these failures, a real-time computer program could be
prepared to continuously monitor the control system and change controller gains
and automatic/manual statuses depending on the condition of all controllers in the
multiloop system.
To summarize this discussion on integrity:

Pairing a control loop on negative or zero relative gain should be avoided, if pos
sible; such a pairing is implemented only when essential, significant dynamic
advantages can be gained by this design and by no other reasonable process or
control modifications.
When a control design has a loop paired on a negative or zero relative gain, a
program should be executed in real time to monitor the interacting loops and
either warn the operator or take automated actions to prevent unstable systems
when the status of an interacting loop changes from automatic to manual.

To discuss a process with conventional and zero relative gain pairing, we

begin by considering the fired heater process in Figure 21.5. The process fluid
flows through a pipe (termed a coil) and is heated by radiant and convective heat
transfer from the combustion of fuel. The variables to be controlled are the process
674

CHAPTER 21
Multiloop Control: Feed
Performance Analysis

Fuel oil
FIGURE 21.5
Furnace multiloop control pairing on variables with
X>0.

fluid flow rate and the process fluid outlet temperature, and the two manipulated
valves are in the process fluid ivx) and fuel iv2) lines. When no feedback controllers
are present, the process fluid flow rate is influenced directly only by v\, and the
outlet temperature is influenced by both vx and v2. Thus, the 2 x 2 gain matrix has
a zero, and as shown in Chapter 20, the relative gain array has ones in the diagonal
elements and zeros in the off-diagonal elements. There is only one pairing with
nonzero relative gain values, and this pairing is shown in Figure 21.5, which is the
common loop pairing used in most industrial designs.
The guideline for eliminating pairings on nonpositive relative gains conforms
to theory and common industrial practice; however, there are a few cases where the
rule is violated and pairings with zero relative gains are used. These unconventional
designs are employed, in spite of their recognized drawbacks, to achieve specific
advantages—typically, very fast feedback dynamics for a particularly important
controlled variable. An example of an exception is given in Figure 21.6. In this
case, the tight control of the coil outlet temperature is very important, and the
dynamic response between the process flow valve vx and the temperature can be
very fast when the fluid residence time in the coils is short. Since the open-loop
gain between valve v2 and the process fluid flow is zero, the proper functioning of
the flow controller in this case requires the operation of the temperature controller.
This design is used industrially only when the temperature is of especially great
importance, feed flow control need not be controlled tightly, and other steps to
improve control performance are not possible or are extremely costly.

Dynamics
If one or a few controlled variables are much more important, the control loop pair
ing should be selected to give good performance for the most important variables.
As demonstrated in discussions on single-loop control, control performance is
 675

Multiloop Control
Performance through
Loop Pairing

FIGURE 21.6
Furnace multiloop control pairing on variables with X = 0.

much better when the feedback process dynamics involve a fast process with small
fraction dead time. Thus, the second loop-pairing guideline is stated as follows:

Very important controlled variables should be paired with manipulated variables

that provide fast feedback dynamics with small dead times and time constants and
negligible inverse response.

As an example of this guideline, consider the simplified system in Figure

21.7 in which two gases are mixed, as might occur where the heating value of
the mixed gaseous fuel stream is to be controlled. The sources of the feeds are a
gas stream L (lower heating value) and a vaporizer for the stream H (higher heat
ing value). The controlled variables are the pressure and the composition in the
pipe after mixing, and both manipulated variables affect both controlled variables.
Generally, the pressure is of greatest importance, because variations could lead to
unsafe conditions; short-term composition variations, while not desirable, can be
more easily tolerated. Therefore, the pressure is controlled by manipulating the
fast-responding gas feed, while the composition is controlled by manipulating the
more slowly responding vaporization process. Since the pressure is most important,
this pairing would be used as long as the gas feed valve has the flexibility range to
control pressure—in other words, as long as it does not go fully opened or closed in
response to disturbances—regardless of the interaction effects on the composition.
<§>C%r
EXAMPLE 21.5.
Solvent
(§)_
Pure A
Evaluate the two possible loop pairings for the blending example process with ( ^
base-case conditions in Table 21.1 according to the relative gain and dynamic D^
responses. 0
676

CHAPTER 21
Multiloop Control:
Performance Analysis

FuelL
Heating medium
FIGURE 21.7
Fuel gas control system with key pressure variable
paired with fast manipulated variable.

The relative gain array for the blending process with dilute product (5% A)
can be evaluated from the steady-state gains to be
F2
Relative gain array: 0.05 0.95
Ax
Ft 0.95 0.05
Since none of the elements is less than or equal to 0.0, both possible pairings
are allowed based on the first guideline. Also, the data reported in Example 20.10
show the same dynamic responses for both pairings, since the dominant dynamics
are due to the sensors. Therefore, neither pairing has an advantage regarding
dynamics. Finally, since the two guidelines do not exclude either pairing, the results
in Table 21.1 give strong evidence for preferring the Ax-F2 and F3-F\ pairing, since
the tuning of each controller does not depend on the automatic/manual status of
the other.

EXAMPLE 21.6.
Evaluate the two possible composition control loop pairings for the distillation ex
~ ^ ample in Figure 20.3 according to the relative gain and dynamic responses.
The relative gain array can be evaluated from the steady-state gains in equa
tion (20.24), giving
r*\ FR
5\«—t£l * t&l—r—
^ Relative gain array: 6.09 -5.09
-&+- XD
XB •5.09 6.09
Since only the pairing XD-FR and XB-FV has positive relative gains, only this
« S ^§ ^ ^ pairing is allowed by the first guideline; this is the design in Figure 21.1a. The loop
dynamics for the allowed pairing are not slower, and are even slightly faster, than
the disallowed pairing, which indicates that there is no significant disadvantage
to this design based on feedback dynamics.
Performance Measure 677
The third analysis addresses the remaining candidate pairings, involving control
Multiloop Control
lable systems with positive relative gains, similar feedback dynamics, and con Performance through
trolled variables of equal importance, by investigating the control performance for Loop Pairing
specific disturbances. If only a few candidates remained at this point, one could
simulate the systems for the important disturbance(s) to select the best design, as
was done for the distillation tower in Examples 21.2 and 21.3. Here a shortcut
method is outlined that provides a quick estimate of control performance and is
useful in reducing the pairing candidates that can yield good control performance.
Equally important, it provides insight into the effects of disturbances, specifically
how interaction can be favorable or unfavorable in multiloop control (Stanley et al.,
1985). The approach is introduced for 2 x 2 systems; however, it can be extended
to higher-order systems (Skogestad and Morari, 1987a). In spite of its advantages,
the method does not provide a definitive recommendation, because of the assump
tions required; thus, some care is required in its application, and the results may
have to be verified through dynamic simulation.
The method takes advantage of a simple estimate of control performance that
can be determined directly from the closed-loop transfer function. The control
performance measure used here is integral error, which can be obtained directly
by using the following relationship (see Appendix D):
/•OO /»OC

/ Eit) dt = lim / Eit)e' dt = Eis)\ 5 = 0 (21.7)

Jo s~+QJo
This relationship demonstrates that the integral of a variable, specifically the error,
can be obtained from the transfer function of a stable system without solving for the
complete transient response (Gibilaro and Lee, 1969). Naturally, much detailed in
formation about the transient response is lost, but a useful single measure of control
performance is easily obtained. A large integral error indicates poor performance
and a pairing candidate that should be eliminated. A small integral error can result
from good performance, and the pairing should be retained for further evaluation.
However, large positive and negative errors occurring during the transient could
cancel in this calculation (this is not the IAE!), so a small value of integral error does
not definitely prove good control performance. Thus, the final selection requires
further evaluation, such as a simulation, to determine the transient behavior.
The closed-loop disturbance response transfer function for a 2 x 2 system
is given in equation (21.3). The relationship in equation (21.7) can be applied to
equation (21.3) with Dis) = \/s, resulting, after some rearrangement, in

.Jo
Exit)dt
J ML [f
Jo
E\it)dt
JSL
(/l.tuneXRDG,) (21.8)

where Integral error under multiloop control =

j°°Exit)dt\ ML

Integral error under single-loop control = j Ex it) dt \ =

Isl KidKci)sL
(21.9)
JKc\/Th)sl
Detuning factor for multiloop control = /i,lune =
iKcx/Tn)wL
(21.10)
678
1 - J2 l2 ) (21.11)
CHAPTER 21 (K KdlK22/
K \
Multiloop Control:
Performance Analysis

The multiloop control performance calculation in equation (21.8) is arranged to be

the product of three factors so that separate facets of multiloop control are repre
sented in each factor: (1) a factor for the single-loop performance, (2) a factor for
tuning adjustment, (3) a factor accounting for interaction and disturbance. The first
factor represents the single-loop performance that would be achieved if the other
control loop were not in operation (e.g., in manual). This term again demonstrates
that aspects of single-loop control performance, which are summarized in Chapter
13, also influence the controlled variables in a multiloop system. For example, fast
feedback dynamics and small disturbance magnitudes are beneficial in multiloop
systems.
The final two factors represent the change in control performance due to the
multiloop structure. The detuning factor /,une represents the effects of detuning the
PI controllers for multiloop control. The values of the multiloop tuning constants
can be estimated using methods in Chapter 20 or alternative methods cited in
the references. By applying the tuning method recommended in Chapter 20 for
2x2 systems with equal dynamics for the two loops, the detuning factor can be
determined from the relative gain, as shown in Figure 21.8. Since the relative gain
in most properly designed control systems is greater than about 0.7, the correlation
shows that the detuning factor is usually bounded between 1.0 and 2.0 for 2 x 2
systems (Marino-Galarraga et al., 1987a).
Thus, the effect of multivariable control is usually dominated by the third
term, which is called the relative disturbance gain, RDG. The relative disturbance
gain is the product of the relative gain and a disturbance factor. Recall that the
relative gain is an inherent property of the feedback process, independent of the
type of disturbance. In contrast, the RDG depends on the type of disturbance; for

1.5 2 2.5 3 3.5 4 4.5 5

Steady-state relative gain, An
FIGURE 21.8
Correlation between detuning factor /tune and
relative gain for 2 x 2 system with equal
input-output dynamics.
example, it has different values for feed composition and set point changes to a 679
distillation tower.
The influence of the RDG is first analyzed from a mathematical, then a process Multiloop Control
point of view. The RDG is the product of two values, and its magnitude is small Performance through
when control performance is good. The first factor is the relative gain; if the relative Loop Pairing

gain has a large value, its contribution will be to degrade control performance,
because the integral error will tend to increase. The second factor represents the
effect of the disturbance type, and because it is the difference of two values, it can
have a magnitude ranging from zero to very large. A small magnitude of this factor
indicates that the multiloop performance could be much better than the single-loop
performance. This situation would occur when the term (1 — Kd2Kx2/Kjx K22) has
a value near zero, which is interpreted as favorable interaction. The other result,
with a large disturbance contribution and much poorer multiloop performance, is
also possible and is interpreted as unfavorable interaction.

The combined effects of inherent process interaction and disturbance type determine
the dominant difference between single-loop and multiloop control performance.
These effects are reflected in the magnitude of the relative disturbance gain (RDG).

This clearly demonstrates that multiloop control performance can be better or

worse than single-loop performance for some disturbances.
A key element in determining the effect of interaction in multiloop systems
is the manner in which a disturbance affects both controlled variables, sometimes
referred to as the "direction" of the disturbance. Thus, it is worthwhile considering
the basis for favorable interaction. Favorable interaction occurs when controller 2,
in correcting its own deviation from set point, makes an adjustment that improves
the performance of controller 1, CVi it). The net effect must consider the effects
of the disturbances on both controlled variables iKjx and Kj2), the manipulation
taken to correct the CV2(0 deviation (characterized by 1/^22) and the interaction
term (£12). All of these parameters are in the interaction factor of the relative
disturbance gain.

EXAMPLE 21.7.
For the distillation towers in Figures 21.1 and 21.2, evaluate the relative distur
bance gain and provide an interpretation of the effect of interaction on the control
performance of the distillate composition, XD, for a disturbance in the feed com
position.
The effect of interaction on control performance is predicted by equation
(21.8), and the calculations are summarized in Table 21.4 for both distillation con
trol designs. This analysis predicts that the energy balance performs better for
feed composition disturbances, because its sum of values of /,une x RDG,- for the
two compositions is smaller than for the material balance system. This conclusion
is confirmed by the simulation results in Figures 21.1b and 21.2b and in Table 21.2.
The physical interpretation of the favorable interaction is considered here
for the control design in Figure 21.1a. The initial effect of increased light key in
the feed (before the analyzer controllers respond) results in the top and bottom
products having too much light key. In response, the bottom controller increases
680 TABLE 21.4
Summary calculations of predicted control performance for the distillation
CHAPTER 21 tower in Examples 21.7 and 21.8
Multiloop Control:
Performance Analysis Energy balance design Material balance design
in Figure 21.1a in Figure 21.2a
Data and
calculated variable xD XB
Kfr 0.0747 0.1173
Kfd 0.0747 -0.1173
Kfv -0.0667 -0.1253 0.008 -0.008
k 6.09 0.39
/tune 2.0 5.0
Feed Kd 0.70 1.3 0.70 1.3
composition RDG 0.071 0.94 1.11 0.06
disturbance /tune • RDG 0.14 1.88 5.55 0.30
Set point Kd 1.0 0.0 1.0 0.0
RDG 6.09 * 0.39 *
change iXD)
12.2 * 1.53 *
j'^e RDG

* Predicted / Edt is finite, although RDG is infinite, due to cancellation of Kd2 (which is zero) in numerator
and denominator.

the heating flow rate (i.e., reboiler duty). This adjustment by the bottom controller
has the effect of decreasing the light key in the top product, exactly what the top
controller is doing itself! The top controller must also take action by increasing the
reflux; however, the (reinforcing) interaction from the bottoms controller improves
the overall control performance. Therefore, the energy balance control pairing
has favorable interaction and good multiloop performance for the top controller
in response to a feed composition disturbance. The reader should repeat this
thought experiment for the material balance system to confirm that the interaction
is unfavorable for XD.

EXAMPLE 21.8.
For the distillation towers in Figures 21.1 and 21.2, evaluate the relative disturbance
gain for a change in the distillate composition controller set point and select the
better design for XD.
The analysis method, summarized in Table 21.4, correctly predicts that the
material balance performs better for set point changes in the distillate controller,
as was found by simulations in Figure 21.3b. Note that equation (21.3) can be
used to represent a set point change by setting Gdlis) = 1.0 and Gd2is) = 0.0,
and in this case the RDGi is equal to kn.

In summary, equation (21.8) provides the basis for estimating the major effect
of multiloop control on the performance of each controlled variable. ITie infor-
mation required to perform this calculation involves process gains in the feedback 681
path Ky and the open-loop disturbance gains Kji, which can be easily determined
from a steady-state analysis. One should consider the likely errors in the values of Multiloop Control
the gains, as well as in the simplifications in linearizing the process model, when Performance through
Loop Pairing
interpreting the results. Small differences (10-20%) in predicted integral error
should be considered within the accuracy of the information, and the candidate
loop pairings should be considered indistinguishable.
This subsection introduced the consideration of disturbance type, which should
be considered in all analyses of multiloop systems. However, it is necessary to re
peat a caution concerning the use of the integral error, which can be small because
of cancellations of large positive and negative errors. Thus, while large values
of |RDG|| AD\K<t\ definitely indicate poor control performance, small values do
not necessarily indicate good performance. The best recourse to determine the
effects of complex dynamics at this time is to perform a dynamic simulation. Note
that the procedures described here are useful in substantially reducing the number
of candidates for simulation, as well as providing insight into the importance of
disturbance type (or "direction") on control performance.

Control Range
The method for determining controllability in Chapter 20 is valid for the linearized
model at the point of linearization. For most processes that are not highly non
linear, the results can be extended in a region about the point. However, there is
no guarantee that the results can be extrapolated, especially when a manipulated
variable encounters a constraint while attempting to make the change required
by the controller. The method for identifying difficulties with range in achievable
steady-state behavior is to determine the operating window of the process. Even if
all steady states are feasible, manipulated variables may reach limits during tran
sients; dynamic simulation would be required to determine the importance of a
temporary saturation of a manipulated variable.

This section demonstrated a stepwise method for evaluating candidate multi

loop control designs:

1. Use the relative gain to eliminate some pairings which lack integrity.
2. Use dynamic models to select pairings with fast dynamics for important vari
ables.
3. Use approximate control performance analysis—the relative disturbance gain
(RDG)—for specific disturbances to evaluate systems with controlled vari
ables of equal importance.

Note that step 1 requires only steady-state information, which means that it is
easy to perform with limited modelling information. Also, steps 2 and 3 require
approximate dynamic information to identify where major differences in feedback
dynamics are present. This approximate dynamic modelling information is also
generally easy to obtain. If the effects of interactive dynamics are not easily pre
dicted, so that the methods here cannot provide conclusive recommendations, the
final design could be simulated to determine its performance.
682 21.4 ® MULTILOOP CONTROL PERFORMANCE
THROUGH TUNING
CHAPTER 21
The tuning of PID feedback controllers should be matched with the control ob
Multiloop Control:
Performance Analysis jectives. Prior to tuning, the first steps presented in the previous section should be
applied, to eliminate inappropriate pairings by the use of the relative gain and to
select pairings with fast feedback dynamics for the important controlled variables.
In all cases, controllers for the most important controlled variables should be tuned
tightly. The tuning of the controllers of lesser importance depends on the type of
interaction present: favorable or unfavorable.
For systems with unfavorable interaction, as predicted by the relative distur
bance gain, the effect of interaction degrades the performance of other loops; this
degradation can be reduced through judicious controller detuning, consistent with
the control objectives. Thus, the controllers for the important variable(s) would
be tuned tightly, as close as possible to single-loop tuning. To ensure stability and
prevent unfavorable interaction, the controllers for the less important variables
would usually be detuned.
If the interaction is favorable, as indicated by a small relative disturbance gain,
interaction improves the performance of other loops and should be maintained by
proper tuning. In this case, the interacting loop, even if not of great importance
itself, should be tuned as tightly as possible to enhance the favorable interaction.
There are no exact guidelines for how the less important controllers should
be tuned. When interaction degrades control performance, a starting approach
is to tune the important loops close to their single-loop values and detune the
less important loops by decreasing their controller gains. Normally, all feedback
controllers would retain an integral mode to return the controlled variables to their
set points (albeit very slowly for some variables) after disturbances. When both
are to be tightly tuned, the method in Chapter 20 would give initial values. An
example of how differences in control performance in the same process can be
induced through different tuning is given in the results in Table 20.2.

EXAMPLE 21.9.
The effects of tuning the composition controllers on the control performance of
! $ the energy balance distillation control design in Figure 21.1a are investigated.
For this example (only), the distillate product composition is assumed to be much
more important than the bottoms composition, so the bottoms composition will
be allowed to experience larger short-term variation about its set point. Since no
a-i—c^i—* t&-r-- strict guidelines exist for this tuning, the extent of detuning used in this example
■S" represents exploratory results.
-t&- The effects of tuning, as determined by simulating the entire response, are
given in Table 21.5. For a set point change in XD, the interaction is unfavorable,
as demonstrated by the large magnitude of RDG • f^ (12.2) in Table 21.4. There
^fc^-^" ^ <£) fore, tight tuning of the distillate composition controller, along with detuning the
bottoms loop, reduces interaction and improves the performance of the distillate
composition controller (reducing the IAE from 0.71 to 0.35). As expected, the vari
ation in the bottoms composition (IAE) increased as the bottoms controller was
detuned.
For the feed composition disturbance, the interaction is favorable, as demon
strated by the small magnitude of RDG • /tune (0.14) in Table 21.4. Therefore, the
TABLE 21.5 683
The effects of tuning on performance for Example 21.9
Multiloop Control
Tuning Performance Performance through
Enhancements:
Input change KCXD Tixd KCXB TlXB IAE*fl IAE*B Decoupling
Set point, 10.4 9.0 -6.8 6.1 0.71 0.68
iASPXD=0.0\) 20.75 9.0 -3.4 6.1 0.35 1.37
Feed 10.4 9.0 -6.8 6.1 0.17 0.35
composition,
iAXB = -0.04) 10.4 9.0 -2.0 6.1 0.36 1.18
W#WPPI

control performance in the case with both controllers tightly tuned has better dis
tillate composition performance (IAE of 0.17) than the case with the bottoms con
troller detuned (IAE of 0.36), since detuning reduces the favorable interaction.
fi S K S S * ^ ^

The discussion in this section and the results of Example 21.9 reinforce the
importance of considering the effects of the disturbances in control design and
tuning.

Multiloop tuning should be chosen to retain favorable interaction and to reduce

unfavorable interaction.

21.5 n MULTILOOP CONTROL PERFORMANCE

THROUGH ENHANCEMENTS: DECOUPLING
When the previous analyses are complete, it is possible to arrive at a design with two
(or more) equally important controlled variables, which may not have the desired
performance even with the best pairing and tuning. Often, the limiting factor is
unfavorable interaction, which is indicated by a large magnitude of the relative
disturbance gain (| RDG |). When poor control performance stems from unfavorable
interaction, a potential solution involves reducing interaction through an approach
called decoupling, which has the theoretical ability to improve performance in
some loops without degrading performance in others.
Decoupling reduces interaction by transforming the closed-loop transfer func
tion matrix into (an approximate) diagonal form, in which interaction is reduced or
eliminated. There are at least three different decoupling approaches: (1) altering the
manipulated variables, (2) altering the controlled variables, and (3) retaining the
original variables but altering the feedback control calculation. Each is presented
briefly in this section.
684 Manipulated Variables
The first decoupling approach involves changing the control structure to affect dif
CHAPTER 21
ferent manipulated secondary variables in a cascade structure, with the same final
Multiloop Control:
Performance Analysis elements. This approach will be introduced by reconsidering the blending in Exam
ple 20.1, in which both manipulated variables influence both controlled variables.
The goal is to control the same variables (Aj and F3) with altered manipulated
variables so that the altered system's gain matrix is diagonal or nearly diagonal.
This goal is usually achieved through process insight. The restructured dynamic
model can be developed from equations (20.1) and (20.2) without linearizing.
dA i ( o r Fdt-eA)
■ca Axit) = MVxit-9A)-Axit)
dt Fxit-9A) + F2it -BA)\
(21.12)
dFdt) = Fxit -9F) + Fdt - 9F) - F3(f) = MVdt - 6F) - F3(0 (21.13)
xF- dt
From this model it becomes clear that the two controlled variables would be inde
pendent if the manipulated variables were defined as follows:
Manipulated variable number 1 = MVi = F2/iFx + F2)
Manipulated variable number 2 = MV2 = Fx + F2
With this modification, the system in equations (21.12) and (21.13) has been altered
to two independent input-output relationships, and as a side benefit the altered
system is linear. Thus, standard single-loop control methods can be used to tune
the controllers in this decoupled system.
The control strategy can be implemented using real-time calculations and
cascade principles, as shown in Figure 21.9, because Fx and F2 are measured and
respond essentially instantaneously to changes in the valve positions. For example,
when the mixed flow (F3) set point is increased, the initial response of controller

Fl+F2

FIGURE 21.9
Manipulated-variable decoupled control of blending.
F3 is to increase the total flow (Fi + F2) set point; this is achieved by adjusting 685
vx. This changes the flow ratio and is quickly followed by an adjustment by the
flow ratio controller to increase v2 to maintain the proper ratio F2/iFx+ F2); this Multiloop Control
adjustment is made without feedback from the analyzer composition controller. Performance through
These adjustments continue until the desired values of the total flow and ratio Enhancements:
Decoupling
are achieved. By similar analysis, it can be shown that the analyzer controller
output affects only the product composition, not the total flow. Thus, the interac
tions have been eliminated. As an added advantage, the decoupled control system
is also easily understood by plant operating personnel. Naturally, the feedback
controllers remain to account for small inaccuracies in the flow measurements,
manipulated-variable calculations, and disturbances. Many similar strategies are
used industrially to minimize unfavorable interactions and are the basis for the
common water faucet design in which the total water flow and the ratio of hot to
cold can be adjusted independently.

Controlled Variables
Another decoupling approach alters the controlled variables by replacing measured
variables with calculated variables based on process output measurements. Again,
the proper calculation is designed with knowledge of the process dynamics. As a
simple example, the two-tank level control system in Figure 21.10 is considered;
the levels are to be controlled by manipulating the set points of the flow controllers.
If the goal were to design two decoupled controllers for maintaining the desired
levels, calculated variables which yield independent equations would be sought in
the basic linearized model of the process.

dL\ = Fiin - F'x ~ KxdL\ - L2)

(21.14)
~aT
dV2
I T — F2\n — F2 + KxdL\ - L'7) (21.15)

A decoupled system can be derived by noting that the sum of the levels depends
on the sum of the manipulated variables, whereas the difference between the levels
depends on the difference between the manipulated variables. This is easily shown

'lin Fr,2in

£><f-
/
Partially open
FIGURE 21.10
Level process.
686 by adding and subtracting equations (21.14) and (21.15) to give

CHAPTER 21 Ad(L\ + V2) = (f,^ + pL) _ (f, + p,} (2U6)

Multiloop Control:
Performance Analysis
Ad(L''~L'2) = (F'Vm - FiJ - 2Kn(L\ - L'2) - (F[ - F2') (21.17)

Thus, a control design in which (L i + L2) and (L i — L2) are controlled by adjusting
(Fj + F2) and (Fj — F2), respectively, is decoupled. Note that (Lj + L2) is non-
self-regulatory, whereas (Li — L2) is a first-order system. A process application
of this principle to distillation reboiler level and composition control is given by
Shinskey (1988).
This approach is not as widely applied as the approach based on manipu
lated variables, because it uses measured process output values in calculating the
controlled variables. For this approach to function properly, all measured output
variables should respond to adjustments in all manipulated variables with nearly
the same dynamics so that the calculations are "synchronized." This criterion is
easily satisfied for the example in Figure 21.10, because levels respond rapidly,
but it is not commonly satisfied for complex units. Control designs for distillation
composition using these concepts have been reported (Weber and Gaitonde, 1985;
Waller and Finnerman, 1987).

Explicit Decoupling Calculations

The third approach to decoupling is to retain the original manipulated and con
trolled variables and alter the control calculation, while retaining the multiloop
structure. There are two common implementations of this approach. The "ideal"
decoupling compensates for interactions while leaving the input-output dynamic
relationships for the feedback controllers unchanged from their single-loop be
havior, Gais). While the concept is attractive, since controller tuning would not
be affected by decoupling, experience has shown that the resulting system is very
sensitive to modelling errors and generally does not perform well (Arkun et al.,
1984; McAvoy 1979); thus, it is not considered further.
The "simplified" decoupling method presented here achieves a diagonal sys
tem by calculations that result in the interaction relationships between the controller
outputs and controlled variables all being zero. Since it is not possible to eliminate
the process interaction Gjjis), the decouplers are designed to provide compen
sating adjustments that cancel the process effects of manipulations in MVjis) on
CVjis) for i ^ j and thus yield independent, single-loop systems. The system is
shown in Figure 21.11, with the decoupling transfer functions £>,-; is) given by the
following relationships:
Gijjs)
Decoupler: Dijis) = - (21.18)
Gids)
The reader may recognize the decoupler as similar to the feedforward controller,
which compensates for measured disturbances; here the measured disturbance is
the manipulated variable adjusted by an interacting feedback controller. The reader
is referred to Chapter 15 on feedforward control for the derivation of this equation
and a discussion of the possibility of the decoupler being unrealizable.
 687

SP,(5)-^6 ^Q-^+H-^ cv,(5)

I I Multiloop Control
Performance through
Enhancements:
Decoupling
Dis)

SP2is) CV2is)

FIGURE 21.11
Block diagram of explicit decoupling.

When the process behavior can be modelled by first-order-with-dead-time

transfer functions, the decoupler in equation (21.18) becomes

Kjj 1 + TjjS _(oi:_

(0ij-Ou)s
Duis) = - Ku 1 + XnS (21.19)

Again, this is the same form as feedforward controllers. The decoupling calcu
lations in equation (21.19) can be implemented in digital form through the same
procedures used with feedforward controllers in Chapter 15.
The explicit decoupler completely eliminates interaction only when the model
is perfect. The resulting transfer function can be derived through block diagram
manipulation assuming perfect decoupling, equation (21.18). The perfectly de
coupled system is shown in Figure 21.12. Clearly, the "effective process" being
controlled has changed because of the decoupling, and the controller tuning must
be changed from single-loop values. Since the change in the "feedback process"
transfer function is the inverse of the relative gain, the controller gain for the de
coupled system should be taken as (approximately) the product of the single-loop
controller gain, calculated using Guis), and the relative gain. This will maintain
the Gods), product of the controller and the "process" [A.uGCi is)][G\\is)/ku]t
nearly constant, as a first approximation.
Errors in the models used in the decouplers affect the accuracy of the de
coupling and, more seriously, affect the stability of the multiloop system. The
sensitivity can be determined from an analytical expression of the performance as
a function of the decoupler errors. The procedure to calculate the integral error in
equation (21.7) can be applied to the closed-loop transfer function for the decou
pled system with modelling errors. To simplify the analysis, only the decoupler
gains have errors, with e,- being a multiplicative error in the decoupler controller
gain, Koij. The resulting expression for the performance is

r^iT/iir i
Exit)dt = kxxk<xe2 + j€x — l)Kd2Kx2 (21.20)
Jo Kd\K22
688

CHAPTER 21 ■w
MV,(5)
Multiloop Control: SP,(J) Gclis) Guis)/kuis) <+H ▶ CV,(*)
Performance Analysis

Gw,(5)
— Dis)

GtfG*)

SP2is) O
_ ii
Gc2is)
MV2(5)
G22is) I k22is) <+)—T^ CV2(5)

FIGURE 21.12

Consolidated block diagram explicit decoupling with perfect models. (Reprinted

by permission. Copyright © 1983, Instrument Society of America. From
Interaction Analysis.)

1
where k = kxx = kfi = ^•<Me2 =
KxxK22 \ - K 1 — €jK 1 ~€x€2K

Gjjjs)
Dtjis) = -€( €i = (1 for perfect model)
Guis)
Clearly, the error relative gain, k€i€2, plays a key role. As the decoupler errors in
crease, this factor and the integral error can become very large and the performance
very poor. For processes with relative gains significantly greater than 1, even small
decoupling errors can lead to very poor performance. For example, a small (5%)
model error of e/ = 1.05 in a decoupler applied to the distillation example with
energy balance control (A, = 6.09, k — 0.836) would increase the integral error by
about 100% over perfect decoupling! Thus, caution should be used when applying
decoupling, since it requires model accuracies nearly impossible to achieve for
real process systems with large relative gains. Similar results have been presented
by McAvoy (1979), Shinskey (1988), and Skogestad and Morari (19876) using
different analysis methods.
Several simplifications are possible in this decoupling approach. First, the
dynamic decouplers in equation (21.18) can be approximated by the gains when
this is sufficient for good control. Typically, the steady-state approximation is
acceptable when Dyis) has a small dead time and nearly equal lead (numerator)
and lag (denominator) dynamics. Note that this simplification does not reduce the
sensitivity to model gain errors shown in equation (21.20).
Also, decoupling can be simplified by using only one-way decoupling, with
one Dijis) = 0. This approach would be applied to improve the performance of
the more important controlled variable. Sensitivity analysis shows that one-way
decoupling is much less sensitive to model gain errors than full decoupling, which
presumably leads to its more frequent successful application in practice (McAvoy,
1979).
EXAMPLE 21.10. 689
Determine the performance with decoupling for the energy balance distillation
control system in Figure 21.1. The disturbance is a set point change of +0.01 to Multiloop Control
the top composition controller. Performance through
The first question the engineer should ask is "Will error-free decoupling im Enhancements:
Decoupling
prove the control performance?" Recall that the magnitude of RDG • /tune indicates
the effects of interaction on multiloop controllers. Decoupling removes the effects
of interaction, and the integral error will be the same as for a single-loop controller
(i.e., with the other controllers in manual). Therefore, unfavorable interaction oc
curs when RDG • /,une > 1.0, and decoupling can be used in such cases to remove
the unfavorable interaction. The information required is given in Table 21.4, which
gives the values of 12.2 for XD and 0.0 for XB. Since the value for XD is so large,
decoupling should be considered.
The values for the decoupler can be determined from the linear model of the
energy balance system and are as follows:
10.2$ + 1 ,-,->->,
Dris) = 0.893 ig \ g-(2-3.3).v (not realizable)
15^ + 1
10.2$ + 1
0.893 (physically realizable)
155 + 1
10.25 + 1 ,-U.i
D2Xis) = 0.930
11.755 + 1
A dynamic response for this decoupled system to a set point change of 0.01
in the top composition is given in Figure 21.13a, and the tuning values and per
formance are summarized in Table 21.6. This theoretically best decoupling per
formance is quite good, with a much lower IAE than the multiloop case reported
in Table 21.2 (energy balance), although in this example the set point change
has twice the magnitude. Note that both manipulated variables changed imme
diately when the set point was changed. The immediate change in MVi is from
the controller Gc\, while the immediate change in MV2 is from GcxDl2, so that
the decoupler acts before the controlled variable XB is disturbed. Again, the

t 1 1 1 1 1 1 1 r

Fv J

j i L
300 300
Time
ia)
FIGURE 21.13
Explicit decoupling in distillation control, Example 21.6: ia) based on a perfect model; ib) with 15%
gain errors in decouplers. (Scales: One tick = 0.02 for XD and XB, 0.50 for FR, 0.30 for Fv.)
690 TA B L E 21.6
msmmmmmm^ Summary of decoupling Example 21.10
CHAPTER 21
Multiloop Control: Case Kcl Tn Kc2 Tn Kmi KD2l IAE, IAE2
Performance A n -a l y s i s - ■ 0 . 0 0 6■
Exact gains 60 9 -50 6.1 0.893 0.930 0.118
15% gain 60 9 -50 6.1 1.027 1.07 Unstable
errors

similarity to feedforward is apparent, because the decoupler bases an adjustment

in a process input on another process input.
However, the engineer must also consider the sensitivity to modelling errors.
This decoupled system will become unstable for errors of about 10% in both de
coupler gains; an example with 15% errors is given in Figure 21.13b, which shows
the instability. No amount of detuning (short of Kc2 = 0) in the feedback controllers
will stabilize this response. Although the decoupler theoretically could improve
performance, it is doubtful that sufficient model accuracy is generally available to
use simplified (two-way) decoupling for processes with large relative gains.

With perfect decoupling, it is theoretically possible to improve control per

formance by reducing unfavorable interaction through decoupling as well as to
degrade control performance by misapplying decoupling to a system that has fa
vorable interaction. Decoupling should be considered only after an analysis of
the relative disturbance gain has established that interaction is unfavorable for
the expected disturbances and that performance with decoupling is not extremely
sensitive to model errors.

• Decoupling improves control performance only when process interaction is

unfavorable, so favorable interaction should not be reduced by decoupling.
• The stability and performance of full decoupling can be very sensitive to model
errors when the relative gain is greater than 1. One-way decoupling has much
lower sensitivity to model errors.

An important observation is that greater control system complexity does not

always lead to better performance!

21.6 H MULTILOOP CONTROL PERFORMANCE THROUGH

ENHANCEMENTS: SINGLE-LOOP ENHANCEMENTS
Many enhancements were presented in Part IV to improve the performance of
single-loop control systems. These methods are also widely applied to the control
of multiloop systems, as will be covered in more depth in Part VI, but a brief
example is presented here to complete the methods for achieving good multiloop
performance. The distillation tower in Figure 21.14 has multiloop control of the