Rga PDF

82 MULTIVARIABLE FEEDBACK CONTROL
quantify the degree of directionality and the level of (two-way) interactions in MIMO systems
are the condition number and the relative gain array (RGA), respectively. We rst consider
the condition number of a matrix which is dened as the ratio between the maximum and
minimum singular values,
4
(G) =
(G)/(G) (3.52)
A matrix with a large condition number is said to be ill-conditioned. For a non-singular
(square) matrix (G) = 1/ (G1 ), so (G) = (G1 ). It then follows from (A.120)
(G)
that the condition number is large if both G and G have large elements.
1
The condition number depends strongly on the scaling of the inputs and outputs. To be
more specic, if D 1 and D2 are diagonal scaling matrices, then the condition numbers of
the matrices G and D1 GD2 may be arbitrarily far apart. In general, the matrix G should be
scaled on physical grounds, e.g. by dividing each input and output by its largest expected or
desired value as discussed in Section 1.4.
One might also consider minimizing the condition number over all possible scalings. This
results in the minimized or optimal condition number which is dened by
(G) = min (D1 GD2 ) (3.53)

D1 ,D2
and can be computed using (A.74).

The condition number has been used as an inputoutput controllability measure, and
in particular it has been postulated that a large condition number indicates sensitivity to
uncertainty. This is not true in general, but the reverse holds: if the condition number is small,
then the multivariable effects of uncertainty are not likely to be serious (see (6.89)).
If the condition number is large (say, larger than 10), then this may indicate control
problems:
1. A large condition number (G) = (G)/(G) may be caused by a small value of

(G), which is generally undesirable (on the other hand, a large value of (G) need not
necessarily be a problem).
2. A large condition number may mean that the plant has a large minimized condition
number, or equivalently, it has large RGA elements which indicate fundamental control
problems; see below.
3. A large condition number does imply that the system is sensitive to unstructured (full-
block) input uncertainty (e.g. with an inverse-based controller, see (8.136)), but this kind
of uncertainty often does not occur in practice. We therefore cannot generally conclude
that a plant with a large condition number is sensitive to uncertainty, e.g. see the diagonal
plant in Example 3.12 (page 89).
3.4 Relative gain array (RGA)

The RGA (Bristol, 1966) of a non-singular square complex matrix G is a square complex
matrix dened as
RGA(G) = (G) , G (G1 )T (3.54)
INTRODUCTION TO MULTIVARIABLE CONTROL 83
where denotes element-by-element multiplication (the Hadamard or Schur product). With

Matlab, we write5
RGA = G.*pinv(G).
The RGA of a transfer matrix is generally computed as a function of frequency (see Matlab
program in Table 3.1). For a 2 2 matrix with elements gij the RGA is

11 12 11 1 11 1
(G) = = ; 11 = g12 g21 (3.55)
21 22 1 11 11 1 g11 g22
The RGA is a very useful tool in practical applications. The RGA is treated in detail at
three places in this book. First, we give a general introduction in this section (pages 8291).
The use of the RGA for decentralized control is discussed in more detail in Section 10.6
(pages 442454). Finally, its algebraic properties and extension to non-square matrices are
considered in Appendix A.4 (pages 526529).
3.4.1 Original interpretation: RGA as an interaction measure

We follow Bristol (1966) here, and show that the RGA provides a measure of interactions. Let
uj and yi denote a particular inputoutput pair for the multivariable plant G(s), and assume
that our task is to use uj to control yi . Bristol argued that there will be two extreme cases:
All other loops open: uk = 0, k 6= j.
All other loops closed with perfect control: yk = 0, k 6= i.
Perfect control is only possible at steady-state, but it is a good approximation at frequencies
within the bandwidth of each loop. We now evaluate our gain yi /uj for the two extreme
cases:

yi
Other loops open: = gij (3.56)
uj uk =0,k6=j

yi
Other loops closed: , gbij (3.57)
uj yk =0,k6=i
Here gij = [G]ij is the ijth element of G, whereas gbij is the inverse of the jith element of
G1
gbij = 1/[G1 ]ji (3.58)
To derive (3.58) we note that

yi
y = Gu = [G]ij (3.59)
uj uk =0,k6=j
and interchange the roles of G and G1 , of u and y, and of i and j to get

uj
u=G y 1
= [G1 ]ji (3.60)
yi yk =0,k6=i
5 The symbol in Matlab gives the conjugate transpose (AH ), and we must use . to get the regular transpose
(AT ).
and (3.58) follows. Bristol argued that the ratio between the gains in (3.56) and (3.57) is a
useful measure of interactions, and dened the ijth relative gain as
gij
ij , = [G]ij [G1 ]ji (3.61)
gbij
The RGA is the corresponding matrix of relative gains. From (3.61) we see that (G) =
G (G1 )T where denotes element-by-element multiplication (the Schur product). This
is identical to our denition of the RGA matrix in (3.54).
Remark. The assumption of yk = 0 (perfect control of yk ) in (3.57) is satised at steady-state

( = 0) provided we have integral action in the loop, but it will generally not hold exactly at other
frequencies. Unfortunately, this has led many authors to dismiss the RGA as being only useful at
steady-state or only useful if we use integral action. On the contrary, in most cases it is the value
of the RGA at frequencies close to crossover which is most important, and both the gain and the phase
of the RGA elements are important. The derivation of the RGA in (3.56) to (3.61) was included to
illustrate one useful interpretation of the RGA, but note that our denition of the RGA in (3.54) is
purely algebraic and makes no assumption about perfect control. The general usefulness of the RGA
is further demonstrated by the additional general algebraic and control properties of the RGA listed on
page 88.
Example 3.8 RGA for 2 2 system. Consider a 2 2 system with the plant model
y1 = g11 (s)u1 + g12 (s)u2 (3.62)
y2 = g21 (s)u1 + g22 (s)u2 (3.63)
Assume that our task is to use u1 to control y1 . First consider the case when the other loop is open,
i.e. u2 is constant or equivalently u2 = 0 in terms of deviation variables. We then have
u2 = 0 : y1 = g11 (s)u1
Next consider the case when the other loop is closed with perfect control, i.e. y 2 = 0. In this case, u2
will also change when we change u1 , due to interactions. More precisely, setting y2 = 0 in (3.63) gives
g21 (s)
u2 = u1
g22 (s)
Substituting this into (3.62) gives

g21
y2 = 0 : y1 = g11 g12 u1
g22
| {z }
b11 (s)
g
This means that our gain changes from g11 (s) to gb11 (s) as we close the other loop, and the
corresponding RGA element becomes
open-loop gain (with u2 = 0) g11 (s) 1

11 (s) = = = g12 (s)g21 (s)
closed-loop gain (with y2 = 0) gb11 (s) 1 g11 (s)g22 (s)
Intuitively, for decentralized control, we prefer to pair variables uj and yi so that ij is close
to 1 at all frequencies, because this means that the gain from uj to yi is unaffected by closing
the other loops. More precisely, we have:
Pairing rule 1 (page 450): Prefer pairings such that the rearranged system, with
the selected pairings along the diagonal, has an RGA matrix close to identity at
frequencies around the closed-loop bandwidth.
However, one should avoid pairings where the sign of the steady-state gain from u j to yi
may change depending on the control of the other outputs, because this will yield instability
with integral action in the loop. Thus, gij (0) and gb11 (0) should have the same sign, and we
have:
Pairing rule 2 (page 450): Avoid (if possible) pairing on negative steady-state
RGA elements.
The reader is referred to Section 10.6.4 (page 438) for derivation and further discussion of
these pairing rules.
3.4.2 Examples: RGA

Example 3.9 Blending process. Consider a blending process where we mix sugar (u 1 ) and water
(u2 ) to make a given amount (y1 = F ) of a soft drink with a given sugar fraction (y2 = x). The
balances mass in = mass out for total mass and sugar mass are
F1 + F 2 = F
F1 = xF
Note that the process itself has no dynamics. Linearization yields
dF1 + dF2 = dF
dF1 = x dF + F dx
With u1 = dF1 , u2 = dF2 , y1 = dF and y2 = dx we then get the model
y1 = u 1 + u 2
1 x x
y2 = u 1 u2
F F
where x = 0.2 is the nominal steady-state sugar fraction and F = 2 kg/s is the nominal amount. The

transfer matrix then becomes

1 1 1 1
G(s) = 1x x =
F
F 0.4 0.1
and the corresponding RGA matrix is (at all frequencies)

x 1 x 0.2 0.8
= =
1x x 0.8 0.2
For decentralized control, it then follows from pairing rule 1 (prefer pairing on RGA elements close to
1) that we should pair on the off-diagonal elements; that is, use u1 to control y2 and use u2 to control
y1 . This corresponds to using the largest stream (water, u2 ) to control the amount (y1 = F ), which is
reasonable from a physical point of view. Pairing rule 2 is also satised for this choice.
Example 3.10 Steady-state RGA. Consider a 3 3 plant for which we have at steady-state

16.8 30.5 4.30 1.50 0.99 1.48
G = 16.7 31.0 1.41 , (G) = 0.41 0.97 0.45 (3.64)
1.27 54.1 5.40 0.08 0.95 2.03
For decentralized control, we need to pair on one element in each column or row. It is then clear that
the only choice that satises pairing rule 2 (avoid pairing on negative RGA elements) is to pair on
the diagonal elements; that is, use u1 to control y1 , u2 to control y2 and u3 to control y3 .
Remark. The plant in (3.64) represents the steady-state model of a uid catalytic cracking (FCC)
process. A dynamic model of the FCC process in (3.64) is given in Exercise 6.17 (page 257).
Some additional examples and exercises, that further illustrate the effectiveness of the steady-
state RGA for selecting pairings, are given on page 443.
Example 3.11 Frequency-dependent RGA. The following model describes a a large pressurized
vessel (Skogestad and Wolff, 1991), for example, of the kind found in offshore oil-gas separations. The
inputs are the valve positions for liquid (u1 ) and vapour (u2 ) ow, and the outputs are the liquid volume
(y1 ) and pressure (y2 ).

0.01e5s 34.54(s + 0.0572) 1.913
G(s) = (3.65)
(s + 1.72 104 )(4.32s + 1) 30.22s 9.188(s + 6.95 104 )
1 4
|12 | = |21 | Diagonal pairing

0.8
3
k Iksum
0.6
|ij |
2
0.4
0.2 PSfrag
|11 | = |22 | replacements 1
PSfrag
Off-diagonal pairingreplacements
0 4 2 0
0 4 2 0
10 10 10 10 10 10
Frequency [rad/s] Frequency [rad/s]
(a) Magnitude of RGA elements (b) RGA number
Figure 3.8: Frequency-dependent RGA for G(s) in (3.65)
The RGA matrix (s) depends on frequency. At steady-state (s = 0) the 2,1 element of G(s) is zero,
so (0) = I. Similarly, at high frequencies the 1,2 element is small relative to the other elements, so
(j) = I. This seems to suggest that the diagonal pairing should be used. However, at intermediate
frequencies, the off-diagonal RGA elements are closest to 1, see Figure 3.8(a). For example, at frequency
= 0.01 rad/s the RGA matrix becomes (see Table 3.1)

0.2469 + 0.0193i 0.7531 0.0193i
= (3.66)
0.7531 0.0193i 0.2469 + 0.0193i
Thus, from pairing rule 1, the reverse pairings is probably best if we use decentralized control and
the closed-loop bandwidth is around 0.01 rad/s. From a physical point of view the use of the reverse
pairings is quite surprising, because it involves using the vapour ow (u 2 ) to control liquid level (y1 ).
and the liquid ow (u1 ) to control pressure (y2 ).
Table 3.1: Matlab program to calculate frequency-dependent RGA

% Plant model (3.65)
s = tf(s);
G = (0.01/(s+1.72e-4)/(4.32*s + 1))*[-34.54*(s+0.0572),....
omega = logspace(-5,2,61);
% RGA
for i = 1:length(omega)
Gf = freqresp(G,omega(i)); % G(j)
RGAw(:,:,i) = Gf.*inv(Gf).; % RGA at frequency omega
RGAno(i) = sum(sum(abs(RGAw(:,:,i) - eye(2)))); % RGA number
end
RGA = frd(RGAw,omega);
Remark. Although it is possible to use decentralized control for this interactive process, see the
following exercise, one may achieve much better performance with multivariable control. If one insists
on using decentralized control, then it is recommended to add a liquid ow measurement and use an
inner (lower layer) ow controller. The resulting u1 is then the liquid ow rate rather than the valve
position. Then u2 (vapour ow) has no effect on y1 (liquid volume), and the plant is triangular with
g12 = 0. In this case the diagonal pairing is clearly best.
Exercise 3.7 Design decentralized single-loop controllers for the plant (3.65) using (a) the diagonal
pairings and (b) the off-diagonal pairings. Use the delay (which is nominally 5 seconds) as a
parameter. Use PI controllers independently tuned with the SIMC tuning rules (based on the paired
elements).
Outline of solution: For tuning purposes the elements in G(s) are approximated using the half rule
to get
s (+2.16)s
0.0823 e 0.01913 e
G(s) e
s
s
e
s
s

0.3022 4.32s+1 0.09188 4.32s+1
For the diagonal pairings this gives the PI settings
Kc1 = 12.1/(c1 + ), I1 = 4(c1 + ); Kc2 = 47.0/(c2 + ), I2 = 4.32
and for the off-diagonal pairings (the index refers to the output)
Kc1 = 52.3/(c1 + + 2.16), I1 = 4(c1 + + 2.16); Kc2 = 14.3/(c2 + ), I2 = 4.32
For improved robustness, the level controller (y1 ) is tuned about 3 times slower than the pressure
controller (y2 ), i.e. use c1 = 3 and c2 = . This gives a crossover frequency of about 0.5/ in
the fastest loop. With a delay of about 5 s or larger you should nd, as expected from the RGA at
crossover frequencies (pairing rule 1), that the off-diagonal pairing is best. However, if the delay is
decreased from 5 s to 1 s, then the diagonal pairing is best, as expected since the RGA for the diagonal
pairing approaches 1 at frequencies above 1 rad/s.
3.4.3 RGA number and iterative RGA

Note that in Figure 3.8(a) we plot only the magnitudes of ij , but this may be misleading
when selecting pairings. For example, a magnitude of 1 (seemingly a desirable pairing)
may correspond to an RGA element of 1 (an undesirable pairing). The phase of the RGA
elements should therefore also be considered. An alternative is to compute the RGA number,
as dened next.
RGA number. A simple measure for selecting pairings according to rule 1 is to prefer
pairings with a small RGA number. For a diagonal pairing,
RGA number , k(G) Iksum (3.67)
P
where we have (somewhat arbitrarily) chosen the sum norm, kAksum = i,j |aij |. The RGA
number for
other
pairings is obtained by subtracting 1 for the selected pairings; for example,
(G) 1 0 for the off-diagonal pairing for a 2 2 plant. The disadvantage with the
0 1
RGA number, at least for larger systems, is that it needs to be recomputed for each alternative
pairing. On the other hand, the RGA elements need to be computed only once.
Example 3.11 continued. The RGA number for the plant G(s) in (3.65) is plotted for the two
alternative pairings in Figure 3.8(b). As expected, we see that the off-diagonal pairing is preferred at
intermediate frequencies.
Exercise 3.8 Compute the RGA number for the six alternate pairings for the plant in (3.64). Which
pairing would you prefer?
Remark. Diagonal dominance. A more precise statement of pairing rule 1 (page 85) would be to prefer
pairings that have diagonal dominance (see denition on page 10.6.4). There is a close relationship
between a small RGA number and diagonal dominance, but unfortunately there are exceptions for plants
of size 4 4 or larger, so a small RGA number does not always guarantee diagonal dominance; see
Example 10.18 on page 441.
Iterative RGA. An iterative evaluation of the RGA, 2 (G) = ((G)) etc., is very
useful for choosing pairings with diagonal dominance for large systems. Wolff (1994) found
numerically that
, lim k (G) (3.68)
k
is a permuted identity matrix (except for borderline cases). More importantly, Johnson and
Shapiro (1986, Theorem 2) have proven that always converges to the identity matrix if G
is a generalized diagonally dominant matrix (see denition in Remark 10.6.4 on page 440) .
Since permuting the matrix G causes similar permutations of (G), may then be used as
a candidate
pairing
choice. Typically,
k approaches
for k between

4 and 8. For example,
for G = 1 1 we get = 0.67 0.33 , = 1.33 0.33 , 3 = 0.07
1 2 0.33 0.67 2 0.33 1.33 1.07
1.07 0.07

and 4 = 0.00
1.00
1.00
0.00
, which indicates that the off-diagonal pairing is diagonally dominant.
Note that may sometimes recommend a pairing on negative RGA elements, even if a

positive pairing is possible.
Exercise 3.9 Test the iterative RGA method on the plant (3.64) and conrm that it gives the diagonally
dominant pairing (as it should according to the theory).
3.4.4 Summary of algebraic properties of the RGA

The (complex) RGA matrix has a number of interesting algebraic properties, of which the
most important are (see Appendix A.4, page 526, for more details):
A1. It is independent of input and output scaling.

A2. Its rows and columns sum to 1.
A3. The RGA is the identity matrix if G is upper or lower triangular.
A4. A relative change in an element of G equal to the negative inverse of its corresponding
RGA element, gij 0
= gij (1 1/ij ), yields singularity.
A5. From (A.80), plants with large RGA elements are always ill-conditioned (with a large
value of (G)), but the reverse may not hold (i.e. a plant with a large (G) may have
small RGA elements).
From property A3, it follows that the RGA (or more precisely I) provides a measure
of two-way interaction.
Example 3.12 Consider a diagonal plant for which we have

100 0
(G) 100
G= , (G) = I, (G) = = = 100, (G) = 1 (3.69)
0 1 (G) 1
Here the condition number is 100 which means that the plant gain depends strongly on the input
direction. However, since the plant is diagonal there are no interactions so (G) = I and the minimized
condition number (G) = 1.
Example 3.13 Consider a triangular plant G for which we get

1 2 1 2 2.41
G= , G1 = , (G) = I, (G) = = 5.83, (G) = 1 (3.70)
0 1 0 1 0.41
Note that for a triangular matrix, there is one-way interaction, but no two-way interaction, and the RGA
is always the identity matrix.
Example 3.14 Consider again the distillation process in (3.45) for which we have at steady-state

87.8 86.4 0.399 0.315 35.1 34.1
G= , G1 = , (G) = (3.71)
108.2 109.6 0.394 0.320 34.1 35.1
In this case (G) = 197.2/1.391 = 141.7 is only slightly larger than (G) = 138.268. The
magnitude sum of the elements in the RGA matrix is kksum = 138.275. This conrms property
A5 which states that, for 2 2 systems, k(G)ksum (G) when (G) is large. The condition
number is large, but since the minimum singular value (G) = 1.391 is larger than 1 this does not by
itself imply a control problem. However, the large RGA elements indicate problems, as discussed below
(control property C1).
Example 3.15 Consider again the FCC process in (3.64) with = 69.6/1.63 = 42.6 and
= 7.80. The magnitude sum of the elements in the RGA is kksum = 8.86 which is close to
as expected from property A5. Note that the rows and the columns of in (3.64) sums to 1. Since
(G) is larger than 1 and the RGA elements are relatively small, this steady-state analysis does not
indicate any particular control problems for the plant.
3.4.5 Summary of control properties of the RGA

In addition to the algebraic properties listed above, the RGA has a surprising number of useful
control properties:
C1. Large RGA elements (typically, 5 10 or larger) at frequencies important for control
indicate that the plant is fundamentally difcult to control due to strong interactions and
sensitivity to uncertainty.
(a) Uncertainty in the input channels (diagonal input uncertainty). Plants with large RGA
elements (at crossover frequency) are fundamentally difcult to control because of
sensitivity to input uncertainty, e.g. caused by uncertain or neglected actuator dynamics.
In particular, decouplers or other inverse-based controllers should not be used for plants
with large RGA elements (see page 251).
(b) Element uncertainty. As implied by algebraic property A4 above, large RGA elements
imply sensitivity to element-by-element uncertainty. However, this kind of uncertainty
may not occur in practice due to physical couplings between the transfer function
elements. Therefore, diagonal input uncertainty (which is always present) is usually
of more concern for plants with large RGA elements.
C2. RGA and RHP-zeros. If the sign of an RGA element changes as we go from s = 0 to
s = , then there is a RHP-zero in G or in some subsystem of G (see Theorem 10.7,
page 446).
C3. Non-square plants. The denition of the RGA may be generalized to non-square matrices
by using the pseudo-inverse; see Appendix A.4.2. Extra inputs: If the sum of the elements
in a column of RGA is small ( 1), then one may consider deleting the corresponding
input. Extra outputs: If all elements in a row of RGA are small ( 1), then the
corresponding output cannot be controlled.
C4. RGA and decentralized control. The usefulness of the RGA is summarized by the two
pairing rules on page 85.
Example 3.14 continued. For the steady-state distillation model in (3.71), the large RGA element of
35.1 indicates a control problem. More precisely, fundamental control problems are expected if analysis
shows that G(j) has large RGA elements also in the crossover frequency range. Indeed, with the
idealized dynamic model (3.93) used below, the RGA elements are large at all frequencies, and we will
conrm in simulations that there is a strong sensitivity to input channel uncertainty with an inverse-
based controller, see page 100. For decentralized control, we should, according to rule 2, avoid pairing
on the negative RGA elements. Thus, the diagonal pairing is preferred.
Example 3.16 Consider the plant

1 s+1 s+4
G(s) = (3.72)
5s + 1 1 2
We nd that 11 () = 2 and 11 (0) = 1 have different signs. Since none of the diagonal elements
have RHP-zeros we conclude from property C2 that G(s) must have a RHP-zero. This is indeed true
and G(s) has a zero at s = 2.
Let us elaborate a bit more on the use of RGA for decentralized control (control property
C4). Assume we use decentralized control with integral action in each loop, and want to
pair on one or more negative steady-state RGA elements. This may happen because this
pairing is preferred for dynamic reasons or because there exists no pairing choice with only
positive RGA elements, e.g. see the system in (10.81) on page 444. What will happen? Will
the system be unstable? No, not necessarily. We may, for example, tune one loop at a time
in a sequential manner (usually starting with the fastest loops), and we will end up with a
stable overall system. However, due to the negative RGA element there will be some hidden
problem, because the system is not decentralized integral controllable (DIC); see page 443.
The stability of the overall system then depends on the individual loops being in service.
This means that detuning one or more of the individual loops may result in instability for the
overall system. Instability may also occur if an input saturates, because the corresponding
loop is then effectively out of service. In summary, pairing on negative steady-state RGA
elements should be avoided, and if it cannot be avoided then one should make sure that the
loops remain in service.
For a detailed analysis of achievable performance of the plant (inputoutput controllability

analysis), one must consider the singular values, as well as the RGA and condition number as
functions of frequency. In particular, the crossover frequency range is important. In addition,
disturbances and the presence of unstable (RHP) plant poles and zeros must be considered.
All these issues are discussed in much more detail in Chapters 5 and 6 where we address
achievable performance and inputoutput controllability analysis for SISO and MIMO plants,
respectively.
3.5 Control of multivariable plants

3.5.1 Diagonal controller (decentralized control)
The simplest approach to multivariable controller design is to use a diagonal or block-
diagonal controller K(s). This is often referred to as decentralized control. Decentralized
control works well if G(s) is close to diagonal, because then the plant to be controlled is
essentially a collection of independent sub-plants. However, if the off-diagonal elements
in G(s) are large, then the performance with decentralized diagonal control may be poor
because no attempt is made to counteract the interactions. There are three basic approaches
to the design of decentralized controllers:
Fully coordinated design
Independent design
Sequential design
Decentralized control is discussed in more detail in Chapter 10 on page 429.
3.5.2 Two-step compensator design approach
?
Gd
r +
- - u - -?
+ -
y
K G
-
6 +
ym
+
+6
n
Figure 3.9: One degree-of-freedom feedback control conguration
Consider the simple feedback system in Figure 3.9. A conceptually simple approach
to multivariable control is given by a two-step procedure in which we rst design a

compensator to deal with the interactions in G, and then design a diagonal controller
using methods similar to those for SISO systems in Chapter 2. Several such approaches are
discussed below.
The most common approach is to use a pre-compensator, W1 (s), which counteracts the
interactions in the plant and results in a new shaped plant:
Gs (s) = G(s)W1 (s) (3.73)
which is more diagonal and easier to control than the original plant G(s). After nding a
suitable W1 (s) we can design a diagonal controller Ks (s) for the shaped plant Gs (s). The
overall controller is then
K(s) = W1 (s)Ks (s) (3.74)
In many cases effective compensators may be derived on physical grounds and may include
nonlinear elements such as ratios.
Remark 1 Some design approaches in this spirit are the Nyquist array technique of Rosenbrock (1974)
and the characteristic loci technique of MacFarlane and Kouvaritakis (1977).
Remark 2 The H loop-shaping design procedure, described in detail in Section 9.4, is similar in that
a pre-compensator is rst chosen to yield a shaped plant, G s = GW1 , with desirable properties, and
then a controller Ks (s) is designed. The main difference is that in H loop shaping, Ks (s) is a full
multivariable controller, designed and based on optimization (to optimize H robust stability).
3.5.3 Decoupling
Decoupling control results when the compensator W1 is chosen such that Gs = GW1 in
(3.73) is diagonal at a selected frequency. The following different cases are possible:
1. Dynamic decoupling: Gs (s) is diagonal at all frequencies. For example, with Gs (s) = I
and a square plant, we get W1 = G1 (s) (disregarding the possible problems involved
in realizing G1 (s)). If we then select Ks (s) = l(s)I (e.g. with l(s) = k/s), the overall
controller is
K(s) = Kinv (s) , l(s)G1 (s) (3.75)
We will later refer to (3.75) as an inverse-based controller. It results in a decoupled nominal
system with identical loops, i.e. L(s) = l(s)I, S(s) = 1+l(s) 1
I and T (s) = 1+l(s)
l(s)
I.
Remark. In some cases we may want to keep the diagonal elements in the shaped plant unchanged
by selecting W1 = G1 Gdiag . In other cases we may want the diagonal elements in W1 to be 1.
This may be obtained by selecting W1 = G1 ((G1 )diag )1 , and the off-diagonal elements of W1
are then called decoupling elements.
2. Steady-state decoupling: Gs (0) is diagonal. This may be obtained by selecting a constant
pre-compensator W1 = G1 (0) (and for a non-square plant we may use the pseudo-
inverse provided G(0) has full row (output) rank).
3. Approximate decoupling at frequency wo : Gs (jo ) is as diagonal as possible. This is
usually obtained by choosing a constant pre-compensator W1 = G1 o where Go is a real
approximation of G(jo ). Go may be obtained, for example, using the align algorithm of
Kouvaritakis (1974) (see le align.m available at the books home page). The bandwidth
frequency is a good selection for o because the effect on performance of reducing
interaction is normally greatest at this frequency.
The idea of decoupling control is appealing, but there are several difculties:
1. As one might expect, decoupling may be very sensitive to modelling errors and
uncertainties. This is illustrated below in Section 3.7.2 (page 100).
2. The requirement of decoupling and the use of an inverse-based controller may not be
desirable for disturbance rejection. The reasons are similar to those given for SISO systems
in Section 2.6.4, and are discussed further below; see (3.79).
3. If the plant has RHP-zeros then the requirement of decoupling generally introduces extra
RHP-zeros into the closed-loop system (see Section 6.6.1, page 236).
Even though decoupling controllers may not always be desirable in practice, they are of
interest from a theoretical point of view. They also yield insights into the limitations imposed
by the multivariable interactions on achievable performance. One popular design method,
which essentially yields a decoupling controller, is the internal model control (IMC) approach
(Morari and Zariou, 1989).
Another common strategy, which avoids most of the problems just mentioned, is to use
partial (one-way) decoupling where Gs (s) in (3.73) is upper or lower triangular.
3.5.4 Pre- and post-compensators and the SVD controller

The above pre-compensator approach may be extended by introducing a post-compensator
W2 (s), as shown in Figure 3.10. One then designs a diagonal controller Ks for the shaped
K
- W2 - Ks - W1 -
Figure 3.10: Pre- and post-compensators, W1 and W2 . Ks is diagonal.
plant W2 GW1 . The overall controller is then
K(s) = W1 Ks W2 (3.76)
The SVD controller is a special case of a pre- and post-compensator design. Here
W1 = V o and W2 = UoT (3.77)
where Vo and Uo are obtained from the SVD of Go = Uo o VoT , where Go is a real
approximation of G(jo ) at a given frequency wo (often around the bandwidth). SVD
controllers are studied by Hung and MacFarlane (1982), and by Hovd et al. (1997) who
found that the SVD-controller structure is optimal in some cases, e.g. for plants consisting of
symmetrically interconnected subsystems.
In summary, the SVD controller provides a useful class of controllers. By selecting
Ks = l(s)1 o a decoupling design is achieved, and selecting a diagonal K s with a low
condition number ((Ks ) small) generally results in a robust controller (see Section 6.10).
10
CONTROL STRUCTURE
DESIGN
Most (if not all) available control theories assume that a control structure is given at the outset. They
therefore fail to answer some basic questions, which a control engineer regularly meets in practice.
Which variables should be controlled, which variables should be measured, which inputs should be
manipulated, and which links should be made between them? The objective of this chapter is to describe
the main issues involved in control structure design and to present some of the quantitative methods
available, for example, for selection of controlled variables and for decentralized control.
10.1 Introduction
(weighted) (weighted)
exogenous inputs exogenous outputs
w - -z
- P
u y
manipulated inputs sensed outputs
(control signals)
K
Figure 10.1: General control conguration
In much of this book, we consider the general control problem formulation shown in
Figure 10.1, where the controller design problem is to
Find a stabilizing controller K, which, based on the information in y, generates a control
signal u, which counteracts the inuence of w on z, thereby minimizing the closed-loop
norm from w to z.
We presented different techniques for controller design in Chapters 2, 8 and 9. However, if
we go back to Chapter 1 (page 1), then we see that controller design is only one step, step 9,
in the overall process of designing a control system. In this chapter, we are concerned with
the structural decisions of control structure design, which are the steps necessary to get to
Figure 10.1:
Multivariable Feedback Control: Analysis and Design. Second Edition

S. Skogestad and I. Postlethwaite c 2005, 2006 John Wiley & Sons, Ltd
Step 4 on page 1: The selection of controlled outputs (a set of variables which are to be
controlled to achieve a set of specic objectives).
See Sections 10.2 and 10.3: What are the variables z in Figure 10.1?
Step 5 on page 1: The selection of manipulated inputs and measurements (sets of variables
which can be manipulated and measured for control purposes).
See Section 10.4: What are the variable sets u and y in Figure 10.1?
Step 6 on page 1: The selection of a control conguration (a structure of interconnecting
measurements/commands and manipulated variables).
See Sections 10.5 and 10.6: What is the structure of K in Figure 10.1; that is, how should we
pair the variable sets u and y?
The distinction between the words control structure and control conguration may seem
minor, but note that it is signicant within the context of this book. The control structure (or
control strategy) refers to all structural decisions included in the design of a control system
(steps 4, 5 and 6). On the other hand, the control conguration refers only to the structuring
(decomposition) of the controller K itself (step 6) (also called the measurement/manipulation
partitioning or input/output pairing). Control conguration issues are discussed in more detail
in Section 10.5. The selection of controlled outputs, manipulations and measurements (steps
4 and 5 combined) is sometimes called input/output selection.
One important reason for decomposing the control system into a specic control
conguration is that it may allow for simple tuning of the subcontrollers without the need for
a detailed plant model describing the dynamics and interactions in the process. Multivariable
centralized controllers can always outperform decomposed (decentralized) controllers, but
this performance gain must be traded off against the cost of obtaining and maintaining a
sufciently detailed plant model and the additional hardware.
The number of possible control structures shows a combinatorial growth, so for most
systems a careful evaluation of all alternative control structures is impractical. Fortunately, we
can often obtain a reasonable choice of controlled outputs, measurements and manipulated
inputs from physical insight. In other cases, simple controllability measures as presented
in Chapters 5 and 6 may be used for quickly evaluating or screening alternative control
structures. Additional tools are presented in this chapter.
From an engineering point of view, the decisions involved in designing a complete
control system are taken sequentially: rst, a top-down selection of controlled outputs,
measurements and inputs (steps 4 and 5) and then a bottom-up design of the control
system (in which step 6, the selection of the control conguration, is the most important
decision). However, the decisions are closely related in the sense that one decision directly
inuences the others, so the procedure may involve iteration. Skogestad (2004a) has proposed
a procedure for control structure design for complete chemical plants, consisting of the
following structural decisions:
Top-down (mainly step 4)
(i) Identify operational constraints and identify a scalar cost function J that characterizes
optimal operation.
(ii) Identify degrees of freedom (manipulated inputs u) and in particular identify the ones that
affect the cost J (in process control, the cost J is usually determined by the steady-state).
(iii) Analyze the solution of optimal operation for various disturbances, with the aim of nding
primary controlled variables (y1 = z) which, when kept constant, indirectly minimize the
CONTROL STRUCTURE DESIGN 385
cost (self-optimizing control). (Section 10.3)

(iv) Determine where in the plant to set the production rate.
Bottom-up (steps 5 and 6)
(v) Regulatory/base control layer: Identify additional variables to be measured and controlled
(y2 ), and suggest how to pair these with manipulated inputs. (Section 10.4)
(vi) Advanced/supervisory control layer conguration: Should it be decentralized or
multivariable? (Sections 10.5.1 and 10.6)
(vii) On-line optimization layer: Is this needed or is a constant setpoint policy sufcient (self-
optimizing control)? (Section 10.3)
Except for decision (iv), which is specic to process control, this procedure may be applied
to any control problem.
Control structure design was considered by Foss (1973) in his paper entitled Critique of
chemical process control theory where he concluded by challenging the control theoreticians
of the day to close the gap between theory and applications in this important area. Control
structure design is clearly important in the chemical process industry because of the
complexity of these plants, but the same issues are relevant in most other areas of control
where we have large-scale systems. In the late 1980s Carl Nett (Nett, 1989; Nett and
Minto, 1989) gave a number of lectures based on his experience of aero-engine control at
General Electric, under the title A quantitative approach to the selection and partitioning
of measurements and manipulations for the control of complex systems. He noted that
increases in controller complexity unnecessarily outpace increases in plant complexity, and
that the objective should be to
minimize control system complexity subject to the achievement of accuracy

specications in the face of uncertainty.
Balas (2003) recently surveyed the status of ight control. He states, with reference to the
Boeing company, that the key to the control design is selecting the variables to be regulated
and the controls to perform regulation (steps 4 and 5). Similarly, the rst step in Honeywells
procedure for controller design is the selection of controlled variables (CVs) for performance
and robustness (step 4).
Surveys on control structure design and inputoutput selection are given by Van de Wal
(1994) and Van de Wal and de Jager (2001), respectively. A review of control structure design
in the chemical process industry (plantwide control) is given by Larsson and Skogestad
(2000). The reader is referred to Chapter 5 (page 164) for an overview of the literature on
inputoutput controllability analysis.
10.2 Optimal operation and control

The overall control objective is to maintain acceptable operation (in terms of safety,
environmental impact, load on operators, and so on) while keeping the operating conditions
close to economically optimal. In Figure 10.2, we show three different implementations for
optimization and control:
(a) Open-loop optimization
Objective Objective Objective
? ? ?
Optimizing
Optimizer Optimizer
Controller y
u r u
? +?

z = H(y)
?
-
G G
?
Controller
?
y u H
?
G 6
(a) (b) (c)

Figure 10.2: Different structures for optimization and control. (a) Open-loop optimization. (b) Closed-
loop implementation with separate control layer. (c) Integrated optimization and control.
(b) Closed-loop implementation with separate control layer

(c) Integrated optimization and control (optimizing control)
Structure (a) with open-loop optimization is usually not acceptable because of model
error and unmeasured disturbances. Theoretically, optimal performance is obtained with
the centralized optimizing controller in structure (c), which combines the functions of
optimization and control in one layer. All control actions in such an ideal control
system would be perfectly coordinated and the control system would use on-line dynamic
optimization based on a nonlinear dynamic model of the complete plant instead of, for
example, infrequent steady-state optimization. However, this solution is normally not used
for a number of reasons, including: the cost of modelling, the difculty of controller design,
maintenance and modication, robustness problems, operator acceptance, and the lack of
computing power.
In practice, the hierarchical control system in Figure 10.2(b) is used, with different tasks
assigned to each layer in the hierarchy. In the simplest case we have two layers:
optimization layer computes the desired optimal reference commands r (outside the
scope of this book)
control layer implements the commands to achieve z r (the focus of this book).
The optimization tends to be performed open-loop with limited use of feedback. On the other
hand, the control layer is mainly based on feedback information. The optimization is often
based on nonlinear steady-state models, whereas linear dynamic models are mainly used in
the control layer (as we do throughout the book).
Additional layers are possible, as is illustrated in Figure 10.3 which shows a typical
control hierarchy for a complete chemical plant. Here the control layer is subdivided into
two layers: supervisory control (advanced control) and regulatory control (base control).
We have also included a scheduling layer above the optimization layer. Similar hierarchies
Scheduling
(weeks)
?
Site-wide optimization
(day)

?
U
Local optimization
(hour)

?
Supervisory
control
(minutes)
Control 9
layer +

W
Regulatory
control
(seconds)
Figure 10.3: Typical control system hierarchy in a chemical plant
are found in control systems for most applications, although the time constants and names
of the layers may be different. Note that we have not included any functions related to logic
control (startup/ shutdown) and safety systems. These are of course important, but need not
be considered during normal operation.
In general, the information ow in such a control hierarchy is based on the upper layer
sending setpoints (references, commands) to the layer below, and the lower layer reporting
back any problems in achieving this. There is usually a time scale separation between the
upper layers and the lower layers as indicated in Figure 10.3. The slower upper layer controls
variables that are more important from an overall (long time scale) point of view, using as
degrees of freedom the setpoints for the faster lower layer. The lower layer should take care
of fast (high-frequency) disturbances and keep the system reasonably close to its optimum
in the fast time scale. To reduce the need for frequent setpoint changes, we should control
variables that require small setpoint changes, and this observation is the basis for Section 10.3
which deals with selecting controlled variables.
With a reasonable time scale separation between the layers, typically a factor of ve or
more in terms of closed-loop response time, we have the following advantages:
1. The stability and performance of a lower (faster) layer is not much inuenced by the
presence of upper (slow) layers because the frequency of the disturbance from the upper
layer is well inside the bandwidth of the lower layer.
2. With the lower (faster) layers in place, the stability and performance of the upper (slower)
layers do not depend much on the specic controller settings used in the lower layers
because they only effect high frequencies outside the bandwidth of the upper layers.
More generally, there are two ways of partitioning the control system:
Vertical (hiearchical) decomposition. This is the decomposition just discussed which

usually results from a time scale difference between the various control objectives
(decoupling in time). The controllers are normally designed sequentially, starting
with the fast layers, and then cascaded (series interconnected) in a hierarchical manner.
Horizontal decomposition. This is used when the plant is decoupled in space, and
normally involves a set of independent decentralized controllers. Decentralized control
is discussed in more detail in Section 10.6 (page 429).
Remark 1 In accordance with Lunze (1992) we have purposely used the word layer rather than level
for the hierarchical decomposition of the control system. The somewhat subtle difference is that in
a multilevel system all units contribute to satisfying the same goal, whereas in a multilayer system
the different units have different local objectives (which preferably contribute to the overall goal).
Multilevel systems have been studied in connection with the solution of optimization problems.
Remark 2 The tasks within any layer can be performed by humans (e.g. manual control), and the
interaction and task sharing between the automatic control system and the human operators are very
important in most cases, e.g. an aircraft pilot. However, these issues are outside the scope of this book.
Remark 3 As noted above, we may also decompose the control layer, and from now on when we talk
about control congurations, hierarchical decomposition and decentralization, we generally refer to the
control layer.
Remark 4 A fourth possible strategy for optimization and control, not shown in Figure 10.2, is
(d) extremum-seeking control. Here the model-based block in Figure 10.2(c) is replaced by an
experimenting controller, which, based on measurements of the cost J, perturbs the input in order
to seek the extremum (minimum) of J; see e.g. Ariyur and Krstic (2003) for details. The main
disadvantage with this strategy is that a fast and accurate on-line measurement of J is rarely available.
10.3 Selection of primary controlled outputs

We are concerned here with the selection of controlled outputs (controlled variables, CVs).
This involves selecting the variables z to be controlled at given reference values, z r, where
r is set by some higher layer in the control hierarchy. Thus, the selection of controlled outputs
(for the control layer) is usually intimately related to the hierarchical structuring of the control
system shown in Figure 10.2(b). The aim of this section is to provide systematic methods for
selecting controlled variables. Until recently, this has remained an unsolved problem. For
example, Fisher et al. (1985) state that Our current approach to control of a complete plant
is to solve the optimal steady-state problem on-line, and then use the results of this analysis to
x the setpoints of selected controlled variables. There is no available procedure for selecting
this set of controlled variables, however. Hence experience and intuition still plays a major
role in the design of control systems.
The important variables in this section are:
u degrees of freedom (inputs)
z primary (economic) controlled variables
r reference value (setpoint) for z
y measurements, process information (often including u)
In the general case, the controlled variables are selected as functions of the measurements,
z = H(y). For example, z can be a linear combination of measurements, i.e. z = Hy. In
many cases, we select individual measurements as controlled variables and H is a selection
matrix consisting of ones and zeros. Normally, we select as many controlled variables as the
number of available degrees of freedom, i.e. nz = nu .
The controlled variables z are often not important variables in themselves, but are
controlled in order to achieve some overall operational objective. A reasonable question is
then: why not forget the whole thing about selecting controlled variables, and instead directly
adjust the manipulated variables u? The reason is that an open-loop implementation usually
fails because we are not able to adjust to changes (disturbances d) and errors (in the model).
The following example illustrates the issues.
Example 10.1 Cake baking. The overall goal is to make a cake which is well baked inside and has
a nice exterior. The manipulated input for achieving this is the heat input, u = Q (and we will assume
that the duration of the baking is xed, e.g. at 15 minutes).
(a) If we had never baked a cake before, and if we were to construct the oven ourselves, we might
consider directly manipulating the heat input to the oven, possibly with a watt-meter measurement.
However, this open-loop implementation would not work well, as the optimal heat input depends
strongly on the particular oven we use, and the operation is also sensitive to disturbances; for example,
opening the oven door or whatever else might be in the oven. In short, the open-loop implementation is
sensitive to uncertainty.
(b) An effective way of reducing the uncertainty is to use feedback. Therefore, in practice we use a
closed-loop implementation where we control the oven temperature (z = T ) using a thermostat. The
temperature setpoint r = Ts is found from a cook book (which plays the role of the optimizer).
The (a) open-loop and (b) closed-loop implementations of the cake baking process are illustrated in
Figure 10.2.
The key question is: what variables z should we control? In many cases, it is clear from
a physical understanding of the process what these are. For example, if we are considering
heating or cooling a room, then we should select the room temperature as the controlled
variable z. Furthermore, we generally control variables that are optimally at their constraints
(limits). For example, we make sure that the air conditioning is on maximum if we want to
cool down our house quickly. In other cases, it is less obvious what to control, because the
overall control objective may not be directly associated with keeping some variable constant.
To get an idea of the issues involved, we will consider some simple examples. Let us rst
consider two cases where implementation is obvious because the optimal strategy is to keep
variables at their constraints.
Example 10.2 Short-distance (100 m) running. The objective is to minimize the time T of the race
(J = T ). The manipulated input (u) is the muscle power. For a well-trained runner, the optimal
solution lies at the constraint u = umax . Implementation is then easy: select z = u and r = umax or
alternatively run as fast as possible.
Example 10.3 Driving from A to B. Let y denote the speed of the car. The objective is to minimize
the time T of driving from A to B or, equivalently, to maximize the speed (y), i.e. J = y. If we are
driving on a straight and clear road, then the optimal solution is always to stay on the speed limit
constraint (ymax ). Implementation is then easy: use a feedback scheme (cruise control) to adjust the
engine power (u) such that we are at the speed limit; that is, select z = y and r = y max .
In the next example, the optimal solution does not lie at a constraint and the selection of
the controlled variable is not obvious.
Example 10.4 Long-distance running. The objective is to minimize the time T of the race (J = T ),
which is achieved by maximizing the average speed. It is clear that running at maximum input power is
not a good strategy. This would give a high speed at the beginning, but a slower speed towards the end,
and the average speed will be lower. A better policy would be to keep constant speed (z = y 1 = speed).
The optimization layer (e.g. the trainer) will then choose an optimal setpoint r for the speed, and this is
implemented by the control layer (the runner). Alternative strategies, which may work better in a hilly
terrain, are to keep a constant heart rate (z = y2 = heart rate) or a constant lactate level (z = y3 =
lactate level).
10.3.1 Self-optimizing control

Recall that the title of this section is selection of primary controlled outputs. In the cake
baking process, we select the oven temperature as the controlled output z in the control layer.
It is interesting to note that controlling the oven temperature in itself has no direct relation to
the overall goal of making a well-baked cake. So why do we select the oven temperature as a
controlled output? We now want to outline an approach for answering questions of this kind.
Two distinct questions arise:
1. What variables z should be selected as the controlled variables?
2. What is the optimal reference value (zopt ) for these variables?
The second problem is one of optimization and is extensively studied (but not in this book).
Here we want to gain some insight into the rst problem which has been much less studied.
We make the following assumptions:
1. The overall goal can be quantied in terms of a scalar cost function J.
2. For a given disturbance d, there exists an optimal value uopt (d) (and corresponding value
zopt (d)), which minimizes the cost function J.
3. The reference values r for the controlled outputs z are kept constant, i.e. r is independent
of the disturbances d. Typically, some average value is selected, e.g. r = zopt (d).

In the following, we assume that the optimally constrained variables are already controlled
at their constraints (active constraint control) and consider the remaining unconstrained
problem with controlled variables z and remaining unconstrained degrees of freedom u.
The system behaviour is a function of the independent variables u and d, so we may
formally write J = J(u, d).1 For a given disturbance d the optimal value of the cost function
1 Note that the cost J is usually not a simple function of u and d, but is rather given by some implied relationship
such as
min J = J0 (u, x, d) s.t. f (x, u, d) = 0
u,x
where dim f = dim x and f (x, u, d) = 0 represents the model equations. Formally eliminating the internal state
variables x gives the problem minu J(u, d).
is
Jopt (d) , J(uopt (d), d) = min J(u, d) (10.1)
u
Ideally, we want u = uopt (d). However, this will not be achieved in practice and we have a
loss L = J(u, d) Jopt (d) > 0.
We consider the simple feedback policy in Figure 10.2(b), where we attempt to keep z
constant. Note that the open-loop implementation is included as a special case by selecting
z = u. The aim is to adjust u automatically, if necessary, when there is a disturbance d such
that u uopt (d). This effectively turns the complex optimization problem into a simple
feedback problem. The goal is to achieve self-optimizing control (Skogestad, 2000):
Self-optimizing control is when we can achieve an acceptable loss with constant

setpoint values for the controlled variables without the need to reoptimize when
disturbances occur.
Remark. In Chapter 5, we introduced the term self-regulation, which is when acceptable dynamic
control performance can be obtained with constant manipulated variables (u). Self-optimizing control
is a direct generalization to the layer above where we can achieve acceptable (economic) performance
with constant controlled variables (z).
The concept of self-optimizing control is inherent in many real-life scenarios

including (Skogestad, 2004b):
The central bank attempts to optimize the welfare of the country (J) by keeping a constant
ination rate (z) by varying the interest rate (u).
The long-distance runner may attempt to minimize the total running time (J = T ) by
keeping a constant heart rate (z = y1 ) or constant lactate level (z = y2 ) by varying the
muscle power (u).
A driver attempts to minimize the fuel consumption and engine wear (J) by keeping a
constant engine rotation speed (z) by varying the gear position (u).
The presence of self-optimizing control is also evident in biological systems, which have
no capacity for solving complex on-line optimization problems. Here, self-optimizing control
policies are the only viable solution and have developed by evolution. In business systems,
the primary (economic) controlled variables are called key performance indicators (KPIs)
and their optimal values are obtained by analyzing successful businesses (benchmarking).
The idea of self-optimizing control is further illustrated in Figure 10.4, where we see
that there is a loss if we keep a constant value for the controlled variable z, rather than
reoptimizing when a disturbance moves the process away from its nominal optimal operating
point (denoted d).

An ideal self-optimizing variable would be the gradient of the Lagrange function for
the optimization problem, which should be zero. However, a direct measurement of the
gradient (or a closely related variable) is rarely available, and computing the gradient
generally requires knowing the value of unmeasured disturbances. We will now outline some
approaches for selecting the controlled variables z. Although a model is used to nd z, note
that the goal of self-optimizing control is to eliminate the need for on-line model-based
optimization.
z2 = constant
Cost J
PSfrag replacements
z1 = constant
Loss
Reoptimized Jopt (d)
d
Disturbance d
Figure 10.4: Loss imposed by keeping constant setpoint for the controlled variable. In this case z 1 is a
better self-optimizing controlled variable than z2 .
10.3.2 Selecting controlled outputs: local analysis

We use here a local second-order accurate analysis of the loss function. From this, we derive
the useful minimum singular value rule, and an exact local method; see Halvorsen et al.
(2003) for further details. Note that this is a local analysis, which may be misleading; for
example, if the optimum point of operation is close to infeasibility.
Consider the loss L = J(u, d) Jopt (d), where d is a xed (generally non-zero)
disturbance. We here make the following additional assumptions:
1. The cost function J is smooth, or more precisely twice differentiable.
2. As before, we assume that the optimization problem is unconstrained. If it is optimal
to keep some variable at a constraint, then we assume that this is implemented (active
constraint control) and consider the remaining unconstrained problem.
3. The dynamics of the problem can be neglected when evaluating the cost; that is, we
consider steady-state control and optimization.
4. We control as many variables z as there are available degrees of freedom, i.e. n z = nu .
For a xed d we may then express J(u, d) in terms of a Taylor series expansion in u around
the optimal point. We get
T
J
J(u, d) = Jopt (d) + (u uopt (d))
u opt
| {z }
=0
2
1 J
+ (u uopt (d)) T
(u uopt(d) ) + (10.2)
2 u2 opt
| {z }
=Juu
We will neglect terms of third order and higher (which assumes that we are reasonably close
to the optimum). The second term on the right hand side in (10.2) is zero at the optimal point
for an unconstrained problem. Equation (10.2) quanties how a non-optimal input u u opt
affects the cost function. To study how this relates to output selection we use a linearized
model of the plant
z = Gu + Gd d (10.3)
where G and Gd are the steady-state gain matrix and disturbance model respectively. For a
xed d, we have z z opt = G(u uopt ). If G is invertible we then get
u uopt = G1 (z zopt ) (10.4)
Note that G is a square matrix, since we have assumed that nz = nu . From (10.2) and (10.4)
we get the second-order accurate approximation
1
(10.5)
T
L = J Jopt (z zopt ) GT Juu G1 (z zopt )
2
where the term Juu = ( 2 J/u2 )opt is independent of z. Alternatively, we may write
1
L= z k22
ke (10.6)
2
where ze = Juu G1 (z zopt ). These expressions for the loss L yield considerable insight.
1/2
Obviously, we would like to select the controlled outputs z such that zzopt is zero. However,
this is not possible in practice because of (1) varying disturbances d and (2) implementation
error e associated with control of z. To see this more clearly, we write
z zopt = z r + r zopt = e + eopt (d) (10.7)
where
Optimization error : eopt (d) , r zopt (d)
Implementation error : e,zr
First, we have an optimization error eopt (d) because the algorithm (e.g. the cook book for
cake baking) gives a desired r which is different from the optimal zopt (d). Second, we have
a control or implementation error e because control is not perfect; either because of poor
control performance or because of an incorrect measurement (steady-state bias) n z . If we
have integral action in the controller, then the steady-state control error is zero, and we have
e = nz
If z is directly measured then nz is its measurement error. If z is a combination of several

measurements y, z = Hy, see Figure 10.2(b), then nz = Hny , where ny is the vector of
measurement errors for the measurements y.
In most cases, the errors e and eopt (d) can be assumed independent. The maximum value
of |z zopt | for the expected disturbances and implementation errors, which we call the
expected optimal span, is then
span(z) = max |z zopt | = max |eopt (d)| + max |e| (10.8)

d,e d e
Example 10.1 Cake baking continued. Let us return to the question: why select the oven temperature
as a controlled output? We have two alternatives: a closed-loop implementation with z = T (the oven
temperature) and an open-loop implementation with z = u = Q (the heat input). From experience, we
know that the optimal oven temperature Topt is largely independent of disturbances and is almost
the same for any oven. This means that we may always specify the same oven temperature, say
r = Ts = 190 C, as obtained from the cook book. On the other hand, the optimal heat input Q opt
depends strongly on the heat loss, the size of the oven, etc., and may vary between, say, 100 W and 5000
W. A cook book would then need to list a different value of r = Qs for each kind of oven and would in
addition need some correction factor depending on the room temperature, how often the oven door is
opened, etc. Therefore, we nd that it is much easier to get eopt = Ts Topt [ C] small than to get
eopt = Qs Qopt [W] small. Thus, the main reason for controlling the oven temperature is to minimize
the optimization error. In addition, the control error e is expected to be much smaller when controlling
temperature.
From (10.5) and (10.7), we conclude that we should select the controlled outputs z such that:
1. G1 is small (i.e. G is large); the choice of z should be such that the inputs have a large
effect on z.
2. eopt (d) = rzopt (d) is small; the choice of z should be such that its optimal value zopt (d)
depends only weakly on the disturbances (and other changes).
3. e = z r is small; the choice of z should be such that it is easy to keep the control or
implementation error e small.
4. G1 is small, which implies that G should not be close to singular. For cases with two or
more controlled variables, the variables should be selected such that they are independent
of each other.
By proper scaling of the variables, these four requirements can be combined into the
maximize minimum singular value rule as discussed next.
10.3.3 Selecting controlled outputs: maximum scaled gain method

We here derive a very simple method for selecting controlled variables in terms of the steady-
state gain matrix G from inputs u (unconstrained degrees of freedom) to outputs z (candidate
controlled variables).
Scalar case. In many cases we only have one unconstrained degree of freedom (u is a
scalar and we want to select one z to control). Introduce the scaled gain from u to z:
G0 = G/span(z)
Note form (10.8) that span(z) = maxd,e |z zopt | includes both the optimization (setpoint)
error and the implementation error. Then, from (10.5), the maximum expected loss imposed
by keeping z constant is
2
|Juu | maxd,e |z zopt | |Juu | 1
Lmax = max L = = (10.9)
d,e 2 G 2 |G0 |2
Here |Juu |, the Hessian of the cost function, is independent of the choice for z. From (10.9),
we then get that the scaled gain |G0 | should be maximized to minimize the loss. Note that
the loss decreases with the square of the scaled gain. For an application, see Example 10.6 on
page 398.
Multivariable case. Here u and z are vectors. Introduce the scaled outputs z 0 , S1 z and
the scaled plant G0 = S1 G. Similar to the scalar case we scale with respect to the span,
1
S1 = diag{ } (10.10)
span(zi )
where
span(zi ) = max |zi zi,opt | = max ei,opt (d) + max |ei |
d,e d e
From (10.6), we have L = 21 ke z k22 where ze = Juu G1 (z zopt ). Introducing thescaled

1/2
outputs gives ze = Juu G01 (z 0 zopt ). With the assumed scaling, the individualscaled
1/2 0
output deviations zi zi,opt are less than 1 in magnitude. However, the variables
0 0
zi are
generally correlated, so any combinations of deviations with magnitudes less than 1 may
not possible. For example, the optimal values of both z1 and z2 may change in the same
direction when there is a disturbance. Nevertheless, we will here assume that the expected
output deviations are uncorrelated by making the following assumption:
A1 The variations in zi0 zi0opt are uncorrelated, or more precisely, the worst-case
combination of output deviations zi0 zi0opt , with kz 0 zopt
0
k2 = 1, can occur in practice.
Here z = S1 z denotes the scaled outputs.
0
The reason for using the vector 2-norm, and not the max-norm, is mainly for mathematical
comvenience. With assumption A1 and (A.104), we then have from (10.6) that the maximum
(worst-case) loss is
ke
z k2 1 2 1/2 01 1 1
Lmax = max = (Juu G ) = 1/2
(10.11)
0
kz 0 zopt k2 1 2 2 2 2 (G0 Juu )
where G0 = S1 G and the last equality follows from (A.40). The result may be stated as
follows
Maximum gain (minimum singular value) rule. Let G denote the steady-
state gain matrix from inputs u (unconstrained degrees of freedom) to outputs
z (candidate controlled variables). Scale the outputs using S1 in (10.10) and
assume that A1 holds. Then to minimize the steady-state loss select controlled
variables z that maximize (S1 GJuu ).
1/2
The rule may stated as minimizing the scaled minimum singular value, (G 0 ), of the scaled
gain matrix G0 = S1 GS2 , where the output scaling matrix S1 has the inverse of the spans
along its diagonal, whereas the input scaling is generally a full matrix, S2 = Juu .
1/2
This important result was rst presented in the rst edition of this book (Skogestad and
Postlethwaite, 1996) and proven in more detail by Halvorsen et al. (2003).
Example 10.5 The aero-engine application in Chapter 13 (page 500) provides a nice illustration of
output selection. There the overall goal is to operate the engine optimally in terms of fuel consumption,
while at the same time staying safely away from instability. The optimization layer is a look-up table,
which gives the optimal parameters for the engine at various operating points. Since the engine
at steady-state has three degrees of freedom we need to specify three variables to keep the engine
approximately at the optimal point, and six alternative sets of three outputs are given in Table 13.3.2
(page 503). For the scaled variables, the value of (G0 (0)) is 0.060, 0.049, 0.056, 0.366, 0.409 and
0.342 for the six alternative sets. Based on this, the rst three sets are eliminated. The nal choice is
then based on other considerations including controllability.
Remark 1 In the maximum gain rule, the objective function and the magnitudes of the disturbances
and measurement noise enter indirectly through the scaling S1 of the outputs z. To obtain S1 =
1
diag{ span(z i)
} we need to obtain for each candidate output span(zi ) = maxd |ei,opt (d)| + max |ei |.
The second contribution to the span is simply the expected measurement error, which is the
measurement error plus the control error. The rst contribition, e i,opt , may be obtained from a
(nonlinear) model as follows: Compute the optimal values of the unconstrained z for the expected
disturbances (with optimally constrained variables xed). This yields a look-up table of z opt for
various expected disturbance combinations. From this data obtain for each candidate output, the
expected variation in its optimal value, eiopt = (ziopt,max ziopt,min )/2.
Remark 2 Our desire to have (G0 ) large for output selection is not related to the desire to have (G)
large to avoid input constraints as discussed in Section 6.9. In particular, the scalings, and thus the
matrix G0 , are different for the two cases.
Remark 3 We have in our derivation assumed that the nominal operating point is optimal. However,
it can be shown that the results are independent of the operating point, provided we are in the region
where the cost can be approximated by a quadratic function as in (10.2) (Alstad, 2005). Thus, it is
equally important to select the right controlled variables when we are nominally non-optimal.
Exercise 10.1 Recall that the maximum gain rule requires that the minimum singular value of the
(scaled) gain matrix be maximized. It is proposed that the loss can simply be minimized by selecting the
controlled variables as z = y, where is a large number. Show that such a scaling does not affect the
selection of controlled variables using the singular value method.
10.3.4 Selecting controlled outputs: exact local method

The maximum gain rule is based on assumption A1 on page 395, which may not hold for
some cases with more than one controlled variable (nz = nu > 1). This is pointed out by
Halvorsen et al. (2003), who derived the following exact local method.
Let the diagonal matrix Wd contain the magnitudes of expected disturbances and the
diagonal matrix We contain the expected implementation errors associated with the individual
controlled variables. We assume
that the combined disturbance and implementation error
d0
vector has norm 1, k 0 k2 = 1. Then, it may be shown that the worst-case loss
e
is (Halvorsen et al., 2003)
1
" max
# L =
([ Md Me ])2 (10.12)
d0 2
k 0 k 2 1
e
where
1
Md = 1/2
Juu Juu Jud G1 Gd Wd (10.13)
Me = 1/2 1
Juu G We (10.14)

Here Juu = 2 J/u2 opt , Jud = 2 J/ud opt and the scaling enters through the
weights Wd and We .
10.3.5 Selecting controlled outputs: direct evaluation of cost

The local methods presented in Sections 10.3.2-10.3.4 are very useful. However, in many
practical examples nonlinear effects are important. In particular, the local methods may not
be able to detect feasibility problems. For example, in marathon running, selecting a control
strategy based on constant speed may be good locally (for small disturbances). However,
if we encounter a steep hill (a large disturbance), then operation may not be feasible,
because the selected reference value may be too high. In such cases, we may need to use
a brute force direct evaluation of the loss and feasibility for alternative sets of controlled
variables. This is done by solving the nonlinear equations, and evaluating the cost function
J for various selected disturbances d and control errors e, assuming z = r + e where r
is kept constant (Skogestad, 2000). Here r is usually selected as the optimal value for the
nominal disturbance, but this may not be the best choice and its value may also be found
by optimization (optimal back-off) (Govatsmark, 2003). The set of controlled outputs
with smallest worst-case or average value of J is then preferred. This approach may be
time consuming because the solution of the nonlinear equations must be repeated for each
candidate set of controlled outputs.
10.3.6 Selecting controlled outputs: measurement combinations

We have so far selected z as a subset of the available measurements y. More generally, we may
consider combinations of the measurements. We will restrict ourselves to linear combinations
z = Hy (10.15)
where y now denotes all the available measurements, including the inputs u used by the
control system. The objective is to nd the measurement combination matrix H.
Optimal combination. Write the linear model in terms of the measurements y as
y = Gy u + Gyd d. Locally, the optimal linear combination is obtained by minimizing
([ Md Me ]) in (10.12) with We = HWny , where Wny contains the expected
measurement errors associated with the individual measured variables; see Halvorsen et al.
(2003). Note that H enters (10.12) indirectly, since G = HGy and Gd = HGyd depend on
H. However, (10.12) is a nonlinear function of H and numerical search-based methods need
to be used.
Null space method. A simpler method for nding H is the null space method proposed
by Alstad and Skogestad (2004), where we neglect the implementation error, i.e., M e = 0 in
(10.14). Then, a constant setpoint policy (z = r) is optimal if zopt (d) is independent of d,
that is, when zopt = 0 d in terms of deviation variables. Note that the optimal values of the
individual measurements yopt still depend on d and we may write
yopt = F d (10.16)
where F denotes the optimal sensitivity of y with respect to d. We would like to nd z = Hy

such that zopt = Hyopt = HF d = 0 d for all d. To satisfy this, we must require
HF = 0 (10.17)
or that H lies in the left null space of F . This is always possible, provided ny nu +nd . This
is because the null space of F has dimension ny nd and to make HF = 0, we must require
that nz = nu < ny nd . It can be shown that when (10.17) holds, Md = 0. If there are too
many disturbances, i.e. ny < nu + nd , then one should select only the important disturbances
(in terms of economics) or combine disturbances with a similar effect on y (Alstad, 2005).
In the presence of implementation errors, even when (10.17) holds such that M d = 0, the
loss can be large due to non-zero Me . Therefore, the null space method does not guarantee
that the loss L using a combination of measurements will be less than using the individual
measurements. One practical approach is to select rst the candidate measurements y, whose
sensitivity to the implementation error is small (Alstad, 2005).
10.3.7 Selecting controlled outputs: examples

The following example illustrates the simple maximize scaled gain rule (mimimum singular
value method).
Example 10.6 Cooling cycle. A simple cooling cycle or heat pump consists of a compressor (where
work Ws is supplied and the pressure is increased to ph ), a high-pressure condenser (where heat is
supplied to the surroundings at high temperature), an expansion valve (where the uid is expanded to
TH
ph , T h
PSfrag replacements
ph
Mh
u Ws
Ml
pl
pl , T l
TC
Figure 10.5: Cooling cycle
a lower pressure pl such that the temperature drops) and a low-pressure evaporator (where heat is
removed from the surroundings at low temperature); see Figure 10.5. The compressor work is indirectly
set by the amount of heating or cooling, which is assumed given. We consider a design with a ooded
evaporator where there is no super-heating. In this case, the expansion valve position (u) remains as
an unconstrained degree of freedom, and should be adjusted to minimize the work supplied, J = W s .
The question is: what variable should we control?
Seven alternative controlled variables are considered in Table 10.1. The data is for an ammonia
cooling cycle, and we consider yopt for a small disturbance of 0.1 K in the hot surroundings
(d1 = TH ). We do not consider implementation errors. Details are given in Jensen and Skogestad
(2005). From (10.9), it follows that it may be useful to compute the scaled gain G 0 = G/span(z(di ))
for the various disturbances di and look for controlled variables z with a large value of |G0 |. From a
physical point of view, two obvious candidate controlled variables are the high and low pressures (p h
and pl ). However, these appear to be poor choices with scaled gains |G 0 | of 126 and 0, respectively. The
Table 10.1: Local maximum gain analysis for selecting controlled variable for cooling cycle
Variable (y) zopt (d1 ) G = u z |G|
|G0 | = |zopt (d1 )|
Condenser pressure, ph [Pa] 3689 464566 126
Evaporator pressure, pl [Pa] 167 0 0
Temperature at condenser exit, Th [K] 0.1027 316 3074
Degree of sub-cooling, Th T sat (ph ) [K] 0.0165 331 20017
Choke valve opening, u 8.0 104 1 1250
Liquid level in condenser, Mh [m3 ] 6.7 106 1.06 157583
Liquid level in evaporator, Ml [m3 ] 1.0 105 1.05 105087
zero gain is because we assume a given cooling duty QC = U A(Tl TC ) and further assume saturation
Tl = T sat (pl ). Keeping pl constant is then infeasible when, for example, there are disturbances in T C .
Other obvious candidates are the temperatures at the exit of the heat exchangers, T h and Tl . However,
the temperature Tl at the evaporator exit is directly related to pl (because of saturation) and also has
a zero gain. The open-loop policy with a constant valve position u has a scaled gain of 1250, and
the temperature at the condenser exit (Th ) has a scaled gain of 3074. Even more promising is the
degree of subcooling at the condenser exit with a scaled gain of 20017. Note that the loss decreases
in proportion to |G0 |2 , so the increase in the gain by a factor 20017/1250 = 16.0 when we change
from constant choke valve opening (open-loop) to constant degree of subcooling, corresponds to a
decrease in the loss (at least for small perturbations) by a factor 16.0 2 = 256. Finally, the best single
measurements seem to be the amount of liquid in the condenser and evaporator, M h and Ml , with
scaled gains of 157583 and 105087, respectively. Both these strategies are used in actual heat pump
systems. A brute force evaluation of the cost for a (large) disturbance in the surrounding temperature
(d1 = TH ) of about 10 K, conrms the linear analysis, except that the choice z = T h turns out to be
infeasible. The open-loop policy with constant valve position (z = u) increases the compressor work
by about 10%, whereas the policy with a constant condenser level (z = M h ) has an increase of less
than 0.003%. Similar results hold for a disturbance in the cold surroundings (d 2 = TC ). Note that the
implementation error was not considered, so the actual losses will be larger.
The next simple example illustrates the use of different methods for selection of controlled
variables.
Example 10.7 Selection of controlled variables. As a simple example, consider a scalar
unconstrained problem, with the cost function J = (u d)2 , where nominally d = 0. For this problem
we have three candidate measurements,
y1 = 0.1(u d); y2 = 20u; y3 = 10u 5d
We assume the disturbance and measurement noises are of unit magnitude, i.e. |d| 1 and |n yi | 1.
For this problem, we always have Jopt (d) = 0 corresponding to
uopt (d) = d, y1,opt (d) = 0, y2,opt (d) = 20d and y3,opt (d) = 5d
For the nominal case with d = 0, we thus have uopt (d ) = 0 and yopt (d ) = 0 for all candidate
controlled variables and at the nominal operating point we have J uu = 2, Jud = 2. The linearized
models for the three measured variables are
y1 : Gy1 = 0.1, Gyd1 = 0.1

y2 : Gy2 = 20, Gyd2 = 0
y3 : Gy3 = 10, Gyd3 = 5
Let us rst consider selecting one of the individual measurements as a controlled variable. We have
Case 1: z = y1 , G = Gy1
Case 2: z = y2 , G = Gy2
Case 3: z = y3 , G = Gy3
The losses for this example can be evaluated analytically, and we nd for the three cases
L1 = (10e1 )2 ; L2 = (0.05e2 d)2 ; L3 = (0.1e3 0.5d)2
(For example, with z = y3 , we have u = (y3 + 5d)/10 and with z = ny3 , we get L3 = (u d)2 =
(0.1ny3 + 0.5d d)2 .) With |d| 1 and |nyi | 1, the worst-case losses (with |d| = 1 and
|nyi | = 1) are L1 = 100, L2 = 1.052 = 1.1025 and L3 = 0.62 = 0.36, and we nd that
z = y3 is the best overall choice for self-optimizing control and z = y1 is the worst. We note that
z = y1 is perfectly self-optimizing with respect to disturbances, but has the highest loss. This highlights
the importance of considering the implementation error when selecting controlled variables. Next, we
compare the three different methods discussed earlier in this section.
A. Maximum scaled gain (singular value rule): For the three choices of controlled variables we have
without scaling |G1 | = (G1 ) = 0.1, (G2 ) = 20 and (G3 ) = 10. This indicates that z2 is
the best choice, but this is only correct with no disturbances. Let us now follow the singular value
procedure.
p
1. The input is scaled by the factor 1/ ( 2 J/u2 )opt = 1/ 2 such that a unit deviation in each
input from its optimal value has the same effect on the cost function J.
2. To nd the optimum setpoint error, rst note that uopt (d) = d. Substituting d = 1 (the maximum
disturbance) and u = uopt = 1 (the optimal input) into the dening expressions for the
candidate measurements, then gives eopt,1 = 0.1(u d) = 0, eopt,2 = 20u = 20 and
eopt,3 = 10u 5d = 5. Alternatively, one may use the expression (Halvorsen et al., 2003)
eopt,i = (Gyi Juu
1
Jud Gydi )d. Note that only the magnitude of eopt,i matters.
3. For each candidate controlled variable the implementation error is assumed to be n z = 1.
4. The expected variation (span) for z = y1 is |eopt,i | + |ny1 | = 0 + 1 = 1. Similarly, for z = y2
and z = y3 , the spans are 20 + 1 = 21 and 5 + 1 = 6, respectively.
5. The scaled gain matrices and the worst-case losses are

z = y1 : |G01 | = 11 0.1/ 2 = 0.071; L1 = 2|G10 |2 = 100

z = y2 : |G02 | = 21 1
20/ 2 = 0.67; L2 = 2|G10 |2 = 1.1025

z = y3 : |G3 | = 6 10/ 2 = 1.18;
0 1
L3 = 2|G10 |2 = 0.360
We note from the computed losses that the singular value rule (= maximize scaled gain rule) suggests
that we should control z = y3 , which is the same as found with the exact procedure. The losses
are also identical.
B. Exact local method: In this case, we have Wd = 1 and Wei = 1 and for y1

Md = 2 21 (2) 0.11 (0.1) 1 = 0 and Me = 2 0.11 1 = 10 2

([ Md Me ])2 1
L1 = = (
(0 10 2)) = 100
2 2
Similarly, we nd with z2 and z3
1 1
L2 = ( ( 2 2/20)) = 1.0025 and L3 = (
( 2/2 2/10)) = 0.26
2 2
Thus, the exact local method also suggests selecting z = y3 as the controlled variable. The reason
for the slight difference from the exact nonlinear losses is that we assumed d and n y individually
to be less than 1 in the exact nonlinear method, whereas in the exact linear method we assumed that
the combined 2-norm of d and ny was less than 1.
C. Combinations of measurements: We now want to nd the best combination z = Hy. In addition to

y1 , y2 and y3 , we also include the input u in the set y, i.e.
y = [ y 1 y2 y3 u ] T
We assume that the implementation error for u is 1, i.e. nu = 1. We then have Wny = I, where Wny
is a 4 4 matrix. Furthermore, we have
Gy = [ 0.1 20 10 1 ]T Gyd = [ 0.1 0 5 0 ]T
Optimal combination. We wish to nd H such that ([ Md Me ]) in (10.12) is minimized, where

G = HGy , Gd = HGyd , We = HWny , Juu = 2, Jud = 2 and Wd = 1. Numerical optimization
yields Hopt = [ 0.0209 0.2330 0.9780 0.0116 ]; that is, the optimal combination of the
three measurements and the manipulated input u is
z = 0.0209y1 0.23306y2 + 0.9780y3 0.0116u
We note, as expected, that the most important contribution to z comes from the variable y 3 . The loss
is L = 0.0405, so it is reduced by a factor 6 compared to the previous best case (L = 0.26) with
z = y3 .
Null space method. In the null space method we nd the optimal combination without implementation
error. This rst step is to nd the optimal sensitivity with respect to the disturbances. Since u opt = d,
we have
yopt = F d = Gy uopt + Gyd d = (Gy + Gyd ) d
| {z }
F
and thus the optimal sensitivity is
F = [ 0 20 5 1 ]T
To have zero loss with respect to disturbances we need to combine at least n u + nd = 1 + 1 = 2

measurements. Since we have four candidate measurements, there are an innite number of possible
combinations, but for simplicity of the control system, we prefer to combine only two measurements.
To reduce the effect of implementation errors, it is best to combine measurements y with a large
gain, provided they contain different information about u and d. More precisely, we should maximize
([ Gy Gyd]). From this we nd that measurements 2 and 3 are the best, with ([ G y Gyd ]) =
20 0
10 5
= 4.45. To nd the optimal combination we use HF = 0 or
20h2 + 5h3 = 0
Setting h2 = 1 gives h3 = 4, and the optimal combination is z = y2 4y3 or (normalizing the

2-norm of H to 1):
z = 0.2425y2 + 0.9701y3
The resulting loss when including the implementation error is L = 0.0425. We recommend the use
of this solution, because the loss is only marginally higher (0.0425 instead of 0.0405) than that
obtained using the optimal combination of all four measurements.
Maximizing scaled gain for combined measurements. For the scalar case, the maximize scaled gain
rule can also be used to nd the best combination. Consider a linear combination of measurements
2 and 3, z = h2 y2 + h3 y3 . The gain from u to z is G = h2 Gy2 + h3 Gy3 . The span for z,
span(z) = |eopt,z | + |ez |, is obtained by combining the individual spans
eopt,z = h2 eopt,2 + h3 eopt,3 = h2 f2 + h3 f3 = 20h2 + 5h3

and |ez | = h2 |e
2 | + h3 |e3 |. If we assume that the combined implementation errors
are 2-norm
e2 h
bounded, k k 1, then the worst-case implementation error for z is |ez | = k 2 k2 . The
e3 2 h3
resulting scaled gain that should be maximized in magnitude is
G h2 Gy2 + h3 Gy3
G0 = = (10.18)
span |h2 eopt,2 + h3 eopt,3 | + |ez |
The expression (10.18) gives considerable insight into the selection of a good measurement
combination. We should select H (i.e. h2 and h3 ) in order to maximize |G0 |. The null space method
corresponds to selecting H such thateopt = h2 eopt,2 + h3 eopt,3 = 0. This gives h2 = 0.2425
h
and h3 = 0.9701, and |ez | = k 2 k2 = 1. The corresponding scaled gain is
h3
20 0.2425 + 10 0.9701
G0 = = 4.851
0+1
with a loss L = /(2|G0 |2 ) = 0.0425 (as found above). (The factor = Juu = 2 is included
because we did not scale the inputs when obtaining G0 .)
Some additional examples can be found in Skogestad (2000), Halvorsen et al. (2003),
Skogestad (2004b) and Govatsmark (2003).
Exercise 10.2 Suppose that we want to minimize the LQG-type objective function, J = x2 + ru2 ,
r > 0, where the steady-state model of the system is
x + 2u 3d = 0
y1 = 2x, y2 = 6x 5d, y3 = 3x 2d
Which measurement would you select as a controlled variable for r = 1? How does your conclusion
change with variation in r? Assume unit implementation error for all measurements.
Exercise 10.3 In Exercise 10.2, how would your conclusions change when u (open-loop
implementation policy) is also included as a candidate controlled variable? First, assume the
implementation error for u is unity. Repeat the analysis, when the implementation error for u and
each of the measurements is 10.
10.3.8 Selection of controlled variables: summary

When the optimum coincides with constraints, optimal operation is achieved by controlling
the active constraints. It is for the remaining unconstrained degrees of freedom that the
selection of controlled variables is a difcult issue.
The most common unconstrained case is when there is only a single unconstrained
degree of freedom. The rule is then to select a controlled variable such that the (scaled) gain
is maximized.
Scalar rule: maximize scaled gain |G0 |
G = unscaled gain from u to z

Scaled gain G0 = G/span
span = optimal range (|eopt |) + implementation error (|e|)
In words, this maximize scaled gain rule may be expressed as follows:

Select controlled variables z with a large controllable range compared to their

sum of optimal variation and implementation error. Here
controllable range = range which may be reached by varying the inputs (as
given by the steady-state gain)
optimal variation: due to disturbance (at steady-state)
implementation error = sum of control error and measurement error (at steady-
state)
For cases with more than one unconstrained degree of freedom, we use the gain in the most
difcult direction as expressed by the minimum singular value.
General maximum gain rule: maximize the (scaled) minimum singular
value (G0 ) (at steady-state), where G0 = S1 GS2 and S2 = Juu (see
1/2
page 395 for details).

We have written at steady-state because the cost usually depends on the steady-state, but
more generally it could be replaced by at the bandwidth frequency of the layer above (which
adjusts the setpoints for z).
10.4 Regulatory control layer

In this section, we are concerned with the regulatory control layer. This is at the bottom of
the control hierarchy and the objective of this layer is generally to stabilize the process and
facilitate smooth operation. It is not to optimize objectives related to prot, which is done
at higher layers. Usually, this is a decentralized control system of low complexity which
keeps a set of measurements at given setpoints. The regulatory control layer is usually itself
hierarchical, consisting of cascaded loops. If there are truly unstable modes (RHP-poles)
then these are usually stabilized rst. Then, we close loops to stabilize the system in the
more general sense of keeping the states within acceptable bounds (avoiding drift), for which
the key issue is local disturbance rejection.
The most important issues for regulatory control are what to measure and what to
manipulate. Some simple rules for these are given on page 405. A fundamental issue
is whether the introduction of a separate regulatory control layer imposes an inherent
performance loss in terms of control of the primary variables z. Interestingly, the answer is
no provided the regulatory controller does not contain RHP-zeros, and provided the layer
above has full access to changing the reference values in the regulatory control layer (see
Theorem 10.2 on page 416).
10.4.1 Objectives of regulatory control

Some more specic objectives of the regulatory control layer may be:
O1. Provide sufcient quality of control to enable a trained operator to keep the plant running
safely without use of the higher layers in the control system.
This sharply reduces the need for providing costly backup systems for the higher layers of
the control hierarchy in case of failures.
O2. Allow for simple decentralized (local) controllers (in the regulatory layer) that can be
tuned on-line.
O3. Take care of fast control, such that acceptable control is achievable using slow
control in the layer above.
O4. Track references (setpoints) set by the higher layers in the control hierarchy.
The setpoints of the lower layers are often the manipulated variables for the higher levels in
the control hierarchy, and we want to be able to change these variables as directly and with as
little interaction as possible. Otherwise, the higher layer will need a model of the dynamics
and interactions of the outputs from the lower layer.
O5. Provide for local disturbance rejection.
This follows from O4, since we want to be able to keep the controlled variables in the
regulatory control system at their setpoints.
O6. Stabilize the plant (in the mathematical sense of shifting RHP-poles to the LHP).
O7. Avoid drift so that the system stays within its linear region which allows the use of
linear controllers.
O8. Make it possible to use simple (at least in terms of dynamics) models in the higher
layers.
We want to use relatively simple models because of reliability and the costs involved in
obtaining and maintaining a detailed dynamic model of the plant, and because complex
dynamics will add to the computational burden on the higher-layer control system.
O9. Do not introduce unnecessary performance limitations for the remaining control
problem.
The remaining control problem is the control problem as seen from the higher layer
which has as manipulated inputs the setpoints to the lower-level control system and the
possible unused manipulated inputs. By unnecessary we mean limitations (e.g. RHP-
zeros, large RGA elements, strong sensitivity to disturbances) that do not exist in the original
problem formulation.
10.4.2 Selection of variables for regulatory control

For the following discussion, it is useful to divide the outputs y into two classes:
y1 (locally) uncontrolled outputs (for which there is an associated control objective)
y2 (locally) measured and controlled outputs (with reference value r2 )
By locally we mean here in the regulatory control layer. Thus, the variables y 2 are the
selected controlled variables in the regulatory control layer. We also subdivide the available
manipulated inputs u in a similar manner:
u1 (locally) unused inputs (this set may be empty)
u2 (locally) used inputs for control of y2 (usually nu2 = ny2 )

We will study the regulatory control layer, but a similar subdivision and analysis could be
performed for any control layer. The variables y1 are sometimes called primary outputs,
and the variables y2 secondary outputs. Note that y2 is the controlled variable (CV) in the
control layer presently considered. Typically, you can think of y1 as the variables we would
really like to control and y2 as the variables we control locally to make control of y1 easier.
The regulatory control layer should assist in achieving the overall operational goals, so if
the economic controlled variables z are known, then we should include them in y 1 . In other
cases, if the objective is to stop the system from drifting away from its steady-state, then the
variables y1 could be a weighted subset of the system states; see the discussion on page 418.
The most important issues for regulatory control are:
1. What should we control (what is the variable set y2 )?
2. What should we select as manipulated variables (what is the variable set u2 ) and how
should it be paired with y2 ?
The pairing issue arises because we aim at using decentralized SISO control, if at all possible.
In many cases, it is clear from physical considerations and experience what the variables
y2 are (see the distillation example below for a typical case). However, we have put the word
clear in quotes, because it may sometimes be useful to question the conventional control
wisdom.
We will below, see (10.28), derive transfer functions for partial control, which are useful
for a more exact analysis of the effects of various choices for y2 and u2 . However, we will
rst present some simple rules that may be useful for reducing the number of alternatives that
could be studied. This is important in order to avoid a combinatorial growth in possibilities.
For a plant where we want to select m from M candidate inputs u, and l from L candidate
measurements y, the number of possibilities is

L M L! M!
= (10.19)
l m l!(L l)! m!(M m)!
A few examples: for m = l = 1 and M = L = 2 the number of possibilities is 4; for

m = l = 2 and M = L = 4 it is 36; and for m = M , l = 5 and L = 100 (selecting 5
measurements out of 100 possible) there are 75287520 possible combinations.
It is useful to distinguish between two main cases:
1. Cascade and indirect control. The variables y2 are controlled solely to assist in achieving
good control of the primary outputs y1 . In this case r2 (sometimes denoted r2,u ) is
usually free for use as manipulated inputs (MVs) in the layer above for the control of
y1 .
2. Decentralized control (using sequential design). The variables y2 are important in
themselves. In this case, their reference values r2 (sometimes denoted r2,d ) are usually
not available for the control of y1 , but rather act as disturbances to the control of y1 .
Rules for selecting y2 . Especially for the rst case (cascade and indirect control), the
following rules may be useful for identifying candidate controlled variables y 2 in the
regulatory control layer:
1. y2 should be easy to measure.
2. Control of y2 should stabilize the plant.

3. y2 should have good controllability; that is, it has favourable dynamics for control.
4. y2 should be located close to the manipulated variable u2 (as a consequence of rule 3,
because for good controllability we want a small effective delay; see page 57).
5. The (scaled) gain from u2 to y2 should be large.
In words, the last rule says that the controllable range for y2 (which may be reached
by varying the inputs u2 ) should be large compared to its expected variation (span). It
is a restatement of the maximum gain rule presented on page 395 for selecting primary
(economic) controlled variables z. The rule follows because we would like to control
variables y2 that contribute to achieving optimal operation. For the scalar case, we should
maximize the gain |G022 | = |G22 |/span(y2 ), where G22 is the unscaled transfer function
from u2 to y2 , and span(y2 ) is the sum of the optimal variation and the implementation error
for y2 . For cases with more than one output, the gain is given by the minimum singular
value, (G022 ). The scaled gain (including the optimal variation and implementation error)
should be evaluated for constant u1 and approximately at the bandwidth frequency of the
control layer immediately above (which adjust the references r2 for y2 ).
Rules for selecting u2 . To control y2 , we select a subset u2 of the available manipulated
inputs u. Similar considerations as for y2 apply to the choice of candidate manipulated
variables u2 :
1. Select u2 so that controllability for y2 is good; that is, u2 has a large and direct effect
on y2 . Here large means that the gain is large, and direct means good dynamics with
no inverse response and a small effective delay.
2. Select u2 to maximize the magnitude of the (scaled) gain from u2 to y2 .
3. Avoid using variables u2 that may saturate.
The last item is the only new requirement compared to what we stated for selecting y 2 .
By saturate we mean that the desired value of the input u2 exceeds a physical constraint;
for example, on its magnitude or rate. The last rule applies because, when an input saturates,
we have effectively lost control, and reconguration may be required. Preferably, we would
like to minimize the need for reconguration and its associated logic in the regulatory control
layer, and rather leave such tasks for the upper layers in the control hierarchy.
Example 10.8 Regulatory control for distillation column: basic layer. The overall control
problem for the distillation column in Figure 10.6 has ve manipulated inputs
T
u = [L V D B VT ]
These are all ows [mol/s]: reux L, boilup V , distillate D, bottom ow B, and overhead vapour
(cooling) VT . What to control (y) is yet to be decided.
Overall objective. From a steady-state (and economic) point of view, the column has only three
degrees of freedom2 With pressure also controlled, there are two remaining steady-state degrees of
freedom, and we want to identify the economic controlled variables y 1 = z associated with these. To do
this, we dene the cost function J and minimize it for various disturbances, subject to the constraints,
which include specications on top composition (xD ) and bottom composition (xB ), together with
upper and lower bounds on the ows. In most cases, the optimal solution lies at the constraints. A very
2 A distillation column has two fewer steady-state than dynamic degrees of freedom, because the integrating
condenser and reboiler levels, which need to be controlled to stabilize the process, have no steady-state effect.
common situation is that both top and bottom composition optimally lie at their specications (y D,min
and xB,max ). We generally choose to control active constraints and then have
y 1 = z = [ x D xB ] T
Regulatory control: selection of y2 . We need to stabilize the two integrating modes associated with
the liquid holdups (levels) in the condenser and reboiler of the column (M D and MB [mol]). In
addition, we normally have tight control of pressure (p), because otherwise the (later) control of
temperature and composition becomes more difcult. In summary, we decide to control the following
three variables in the regulatory control layer:
y2 = [ M D MB p ]T
Note that these three variables are important to control in themselves.

Overall control problem. In summary, we have now identied ve variables that we want to control
y = [xD xB MD MB p ]T
| {z } | {z }
y1 y2
The resulting overall 5 5 control problem from u to y can be approximated as (Skogestad and
Morari, 1987a):

xD g11 (s) g12 (s) 0 0 0 L
xB g21 (s) g22 (s) 0 0 0 V

MD = 1/s 0 1/s 0 0 D (10.20)
M g (s)/s 1/s 0 1/s 0 B
B L
MV (p) 0 1/(s + kp ) 0 0 1/(s + kp ) VT
In addition, there are high-frequency dynamics (delays) associated with the inputs (valves) and outputs
(measurements). For control purposes it is very important to include the transfer function g L (s), which
represents the liquid ow dynamics from the top to the bottom of the column, L B = gL (s)L.
For control purposes, it may be approximated by a delay, g L (s) = eL s . gL (s) also enters into the
transfer function g21 (s) from L to xB , and by this decouples the distillation column dynamics at high
frequencies. The overall plant model in (10.20) usually has no inherent control limitations caused by
RHP-zeros, but the plant has two poles at the origin (from the integrating liquid levels,
M D and MB ),
g11 g12
and also one pole close to the origin (almost integrating) in GLV = originating from
g21 g22
the internal recycle in the column. These three modes need to be stabilized. In addition, for high-
purity separations, there is a potential control problem in that the G LV -subsystem is strongly coupled
at steady-state, e.g. resulting in large elements in the RGA matrices for G LV and also for the overall
55 plant, but fortunately the system is decoupled at high frequency because of the liquid ow dynamics
represented by gL (s). Another complication is that composition measurements (y 1 ) are often expensive
and unreliable.
Regulatory control: selection of u2 . As already mentioned, the distillation column is rst stabilized
by closing three decentralized SISO loops for level and pressure, y 2 = [ MD MB p ]T . These
loops usually interact weakly with each other and may be tuned independently. However, there exist
many possible choices for u2 (and thus for u1 ). For example, the condenser holdup tank (MD ) has
one inlet ow (VT ) and two outlet ows (L and D), and any one of these ows, or a combination,
may be used effectively to control MD . By convention, each choice (conguration) of u2 used for
controlling level and pressure is named by the inputs u1 left for composition control. For example, the
LV -conguration used in many examples in this book refers to a partially controlled system where
u2 = [ D B VT ]T is used to control levels and pressure (y2 ) in the regulatory layer, and we are left
with
u1 = [ L V ] T
VT PC PSfrag replacements
LC
MD
p
L
D yD
F zF
V
MB LC
B xB
Figure 10.6: Distillation column controlled with the LV -conguration
to control composition (y1 ). The LV -conguration is known to be strongly interactive at steady-state,

as can been seen from the large steady-state RGA elements; see (3.94) on page 100. On the other hand,
the LV -conguration is good from the point of view that it is the only conguration where control of
y1 (using u1 ) is nearly independent of the tuning of the level controllers (K2 ). This is quite important,
because we normally want slow (smooth control) rather than tight control of the levels (M D and
MB ). This may give undesirable interactions from the regulatory control layer (y 2 ) into the primary
control layer (y1 ). However, this is avoided with the LV-conguration.
Another conguration is the DV -conguration where u2 = [ L B VT ]T is used to control levels
and pressure, and we are left with
u1 = [ D V ] T
to control compositions. If we were only concerned with controlling the condenser level (M D ) then this
choice would be better for cases with difcult separations where L/D 1. This is because to avoid
saturation in u2 we would like to use the largest ow (in this case u2 = L) to control condenser level
(MD ). In addition for this case, the steady-state interactions from u1 to y1 , as expressed by the RGA,
are generally much less; see (6.74) on page 245. However, a disadvantage with the DV -conguration
is that the effect of u1 on y1 depends strongly on the tuning of K2 . This is not surprising, since using D
to control xD corresponds to pairing on g31 = 0 in (10.20), and D (u1 ) therefore only has an effect on
xD (y1 ) when the level loop (from u2 = L to y2 = MD ) has been closed.
There are also many other possible congurations (choices for the two inputs in u 1 ); with ve inputs
there are ten alternative congurations. Furthermore, one often allows for the possibility of using ratios
between ows, e.g. L/D, as possible degrees of freedom in u1 , and this sharply increases the number
of alternatives. However, for all these congurations, the effect of u 1 on y1 depends on the tuning of
K2 , which is undesirable. This is one reason why the LV -conguration is used most in practice. In the
next section, we discuss how closing a fast temperature loop may improve the controllability of the
LV -conguration.
In the above example, the variables y2 were important variables in themselves. In the
following example, the variable y2 is controlled to assist in the control of the primary
variables y1 .
Example 10.9 Regulatory control for distillation column: temperature control. We will assume
that we have closed the three basic control loops for liquid holdup (M D , MB ) and pressure (p) using
the LV -conguration, see Example 10.8, and we are left with a 2 2 control problem with
u = [ L V ]T
(reux and boilup) and

y 1 = [ x D xB ] T
(product compositions). A controllability analysis of the model GLV (s) from u to y1 shows that there
is (1) an almost integrating mode, and (2) strong interactions. The integrating mode results in high
sensitivity to disturbances at lower frequencies. The control implication is that we need to close a
stabilizing loop. A closer analysis of the interactions (e.g. a plot of the RGA elements as a function of
frequency) shows that they are much smaller at high frequencies. The physical reason for this is that L
and xD are at the top of the column, and V and xB at the bottom, and since it takes some time (L ) for
a change in L to reach the bottom, the high-frequency response is decoupled. The control implication is
that the interactions may be avoided by closing a loop with a closed-loop response time less than about
L .
PSfrag replacements
B
xB
LC
LC
TC
Ts
MB
MD
V
F
zF
p
L
D
yD
PC
VT
Figure 10.7: Distillation column with LV -conguration and regulatory temperature loop
It turns out that closing one fast loop may take care of both stabilization and reducing interactions.
The issue is then which loop to close. The most obvious choice is to close one of the composition
loops (y1 ). However, there is usually a time delay involved in measuring composition (x D and xB ),
and the measurement may be unreliable. On the other hand, the temperature T is a good indicator of
composition and is easy to measure. The preferred solution is therefore to close a fast temperature loop
somewhere along the column. This loop will be implemented as part of the regulatory control system.
We have two available manipulated variables u, so temperature may be controlled using reux L or
boilup V . We choose reux L here (see Figure 10.7) because it is more likely that boilup V will reach
its maximum value, and input saturation is not desired in the regulatory control layer. In terms of the
notation presented above, we then have a SISO regulatory loop with
y2 = T ; u2 = L
and u1 = V . The primary composition control layer adjusts the temperature setpoint r 2 = Ts for
the regulatory layer. Thus, for the primary layer we have
y 1 = [ x D xB ] T ; u = [ u 1 r2 ]T = [ V Ts ]
T
The issue is to nd which temperature T in the column to control, and for this we may use the maximum
gain rule. The objective is to maximize the scaled gain |G022 (j)| from u2 = L to y2 = T .
Here, |G022 | = |G22 |/span where G22 is the unscaled gain and span = optimal range (|eopt |) +
implementation error (|e|) for the selected temperature. The gain should be evaluated at approximately
the bandwidth frequency of the composition layer that adjusts the setpoint r 2 = Ts . For this application,
we assume that the primary layer is relatively slow, such that we can evaluate the gain at steady-state,
i.e. = 0.
In Table 10.2, we show the normalized temperatures y2 = x, unscaled gain, optimal variation for
the two disturbances, implementation error, and the resulting span and scaled gain for measurements
located at stages 1 (reboiler), 5, 10, 15, 21 (feed stage), 26, 31, 36 and 41 (condenser). The gains
are also plotted as a function of stage number in Figure 10.8. The largest scaled gain of about 88 is
achieved when the temperature measurement is located at stage 15 from the bottom. However, this is
below the feed stage and it takes some time for the change in reux (u 2 = L), which enters at the top,
to reach this stage. Thus, for dynamic reasons it is better to place the measurement in the top part of
the column; for example, at stage 27 where the gain has a local peak of about 74.
Table 10.2: Evaluation of scaled gain |G022 | for alternative temperature locations (y2 ) for distillation
example. Span = |y2,opt (d1 )| + |y2,opt (d2 )| + ey2 . Scaled gain |G022 | = |G22 |/span.
Nominal Unscaled Scaled
Stage value y2 G22 y2,opt (d1 ) y2,opt (d2 ) e y2 span(y2 ) |G022 |
1 0.0100 1.0846 0.0077 0.0011 0.05 0.0588 18.448
5 0.0355 3.7148 0.0247 0.0056 0.05 0.0803 46.247
10 0.1229 10.9600 0.0615 0.0294 0.05 0.1408 77.807
15 0.2986 17.0030 0.0675 0.0769 0.05 0.1944 87.480
21 0.4987 9.6947 -0.0076 0.0955 0.05 0.1532 63.300
26 0.6675 14.4540 -0.0853 0.0597 0.05 0.1950 74.112
31 0.8469 10.5250 -0.0893 0.0130 0.05 0.1524 69.074
36 0.9501 4.1345 -0.0420 -0.0027 0.05 0.0947 43.646
41 0.9900 0.8754 -0.0096 -0.0013 0.05 0.0609 14.376
Remarks to example.
1. We use data for column A (see Section 13.4) which has 40 stages. This column separates a binary
mixture, and for simplicity we assume that the temperature T on stage i is directly given by the mole
fraction of the light component, Ti = xi . This can be regarded as a normalized temperature which
ranges from 0 in the bottom to 1 in the top of the column. The implementation error is assumed to
be the same on all stages, namely ey2 = 0.05 (and with a temperature difference between the two
components of 13.5 K, this corresponds to an implementation error of 0.68 K). The disturbances
are a 20% increase in feed rate F (d1 = 0.2) and a change from 0.5 to 0.6 in feed mole fraction zF
(d2 = 0.1).
2. The optimal variation (y2,opt (d)) is often obtained from a detailed steady-state model, but it was
generated here from the linear model. For any disturbance d we have in terms of deviation variables
(we omit the s)
y1 = G12 u2 + Gd1 d
y2 = G22 u2 + Gd2 d
The optimal strategy is to have the product compositions constant; that is, y 1 = [ xD xB ]T = 0.
However, since u2 = L is a scalar, this is not possible. The best solution in a least squares sense
(minimize ky1 k2 ) is found by using the pseudo-inverse, uopt
2 = G12 Gd1 d. The resulting optimal
change in the temperature y2 = T is then
y2opt = (G22 G12 Gd1 + Gd2 )d (10.21)
100
80 Scaled
60
Gain
40
PSfrag replacements
20 5 Unscaled
0
5 10 15 20 25 30 35 40
Stage Number
Figure 10.8: Scaled (|G022 |) and unscaled (|G22 |) gains for alternative temperature locations for the
distillation example
3. As seen from the solid and dashed lines in Figure 10.8, the local peaks of the unscaled and scaled
gains occur at stages 26 and 27, respectively. Thus, scaling does not affect the nal conclusion much
in this case. However, if we were to set the implementation error e to zero, then the maximum scaled
gain would be at the bottom of the column (stage 1).
4. We made the choice u2 = L to avoid saturation in the boilup V in the regulatory control layer.
However, if saturation is not a problem, then the other alternative u 2 = V may be better. A similar
analysis with u2 = V gives a maximum scaled gain of about 100 is obtained with the temperature
measured at stage 14.
In summary, the overall 5 5 distillation control problem may be solved by rst designing a 4 4
stabilizing (regulatory) controller K2 for levels, pressure and temperature
y2 = [ M D MB p T ] T , u2 = [ D B V T L]
T
and then designing a 2 2 primary controller K1 for composition control
y1 = [ xD xB ], u1 = [ V Ts ]
Alternatively, we may interchange L and V in u1 and u2 . The temperature sensor (T ) should be located
at a point with a large scaled gain.
We have discussed some simple rules and tools (maximum gain rule) for selecting
the variables in the regulatory control layer. The regulatory control layer is usually itself
hierarchical, consisting of a layer for stabilization of unstable modes (RHP-poles) and a layer
for stabilization in terms of disturbance rejection. Next, we introduce pole vectors and
partial control, which are more specic tools for addressing the issues of stabilization and
disturbance rejection.
10.4.3 Stabilization: pole vectors

Pole vectors are useful for selecting inputs and outputs for stabilization of unstable modes
(RHP-poles) when input usage is an issue. An important advantage is that the selection of
inputs is treated separately from the selection of outputs and hence we avoid the combinatorial
issue. The main disadvantage is that the theoretical results only hold for cases with a single
RHP-pole, but applications show that the tool is more generally useful.
The issue is: which outputs (measurements) and inputs (manipulations) should be used for
stabilization? We should clearly avoid saturation of the inputs, because this makes the system
effectively open-loop and stabilization is then impossible. A reasonable objective is therefore
to minimize the input usage required for stabilization. In addition, this choice also minimizes
the disturbing effect that the stabilization layer has on the remaining control problem.
Recall that u = KS(r + n d), so input usage is minimized when the norm of KS is
minimal. We will consider both the H2 and H norms.
Theorem 10.1 (Input usage for stabilization) For a rational plant with a single unstable
mode p, the minimal H2 and H norms of the transfer function KS are given as (Havre and
Skogestad, 2003; Kariwala, 2004)
(2p)3/2 |q T t|
min kKSk2 = (10.22)
K kup k2 kyp k2
2p |q T t|
min kKSk = (10.23)
K kup k2 kyp k2
Here up and yp denote the input and output pole vectors (see page 127), respectively, and
t and q are the right and left eigenvectors of the state matrix A, satisfying At = pt and
q T A = q T p.
Theorem 10.1 applies to plants with any number of RHP-zeros and to both multivariable
(MIMO) and single-loop (SISO) control. In the SISO case, up and yp are the elements in the
pole vectors, up,j and yp,i , corresponding to the selected input (uj ) and output (yi ). Notice
that the term (q T t) is independent of the selected inputs and outputs, uj and yi . Thus, for a
single unstable mode and SISO control:
The input usage required for stabilization is minimized by selecting the output
yi (measurement) and input uj (manipulation) corresponding to the largest
elements in the output and input pole vectors (yp and up ), respectively (see also
Remark 2 on page 137).
This choice maximizes the (state) controllability and observability of the unstable mode. Note
that the selections of measurement yi and input uj are performed independently. The above
result is for unstable poles. However, Havre (1998) shows that the input requirement for
pole placement is minimized by selecting the output and input corresponding to the largest
elements in the yp and up , respectively. This property also holds for LHP-poles, and shows
that pole vectors may also be useful when we want to move stable poles.
Exercise 10.4 Show that for a system with a single unstable pole, (10.23) represents the least
achievable value of kKSk . (Hint: Rearrange (5.31) on page 178 using the denition of pole vectors.)
When the plant has multiple unstable poles, the pole vectors associated with a specic
RHP-pole give a measure of input usage required to move this RHP-pole assuming that the
other RHP-poles are unchanged. This is of course unrealistic; nevertheless, the pole vector
approach can be used by stabilizing one source of instability at a time. That is, rst an input
and an output are selected considering one real RHP-pole or a pair of complex RHP-poles
and a stabilizing controller is designed. Then, the pole vectors are recomputed for the partially
controlled system and another set of variables is selected. This process is repeated until all the
modes are stabilized. This process results in a sequentially designed decentralized controller
and has been useful in several practical applications, as demonstrated by the next example.
Example 10.10 Stabilization of Tennessee Eastman process. The Tennessee Eastman chemical
process (Downs and Vogel, 1993) was introduced as a challenge problem to test methods for
control structure design.3 The process has 12 manipulated inputs and 41 candidate measurements,
of which we consider 11 here; see Havre (1998) for details on the selection of these variables
and scaling. The model has six unstable poles at the operating point considered, p =
[ 0 0.001 0.023 j0.156 3.066 j5.079 ]. The absolute values of the output and input pole vectors
are

6.815 6.909 2.573 0.964
0.000 0.001 0.041 0.112
6.906 7.197 2.636 0.246
0.000 0.004 0.169 0.065
0.000 0.000 0.013 0.366 0.148 1.485 0.768 0.044
3.973 11.550 5.096 0.470
0.000 0.001 0.051 0.410
0.012 0.369 0.519 0.356
0.009 0.581 0.488 0.316
0.597 0.077 0.066 0.033
|Yp | =
0.000 0.001 0.041 0.115
|Up |T =
0.135 1.850 1.682

0.110
1.605 1.192 0.754 0.131
22.006 0.049 0.000 0.000
0.000 0.001 0.039 0.108
0.007 0.054 0.010 0.013
0.000 0.001 0.038 0.217
0.247 0.708 1.501 2.021
0.000 0.001 0.055 0.109 0.753
1.485
0.976 1.447
0.000 0.002 0.132 0.272
0.033 0.095 0.201 0.302
where we have combined pole vectors corresponding to a complex eigenvalue into a single column.
The individual columns of |Yp | and individual rows of |Up | correspond to the poles at 0, 0.001,
0.023 j0.156 and 3.066 j5.079, respectively.
When designing a stabilizing control system, we normally start by stabilizing the most unstable
(fastest) pole, i.e. complex poles at 3.066 j5.079 in this case. From the pole vectors, this mode is most
easily stabilized by use of u10 and y10 . A PI controller, with proportional gain of 0.05 and integral
time of 300 minutes, is designed for this loop. This simple controller stabilizes the complex unstable
poles at 3.066 j5.079 and also at 0.023 j0.156. This is reasonable since the pole vectors show that
the modes at 0.023 j0.156 are observable and controllable through y 10 and u10 , respectively. For
stabilizing the integrating modes, the pole vectors can be recomputed to select two additional inputs
and outputs; see Havre (1998) for details.
Note that the different choices of inputs and outputs for stabilization have different effects
on the controllability of the stabilized system. Thus, in some cases, variable selection using
pole vectors may need to be repeated a few times before a satisfactory solution is obtained.
An alternative approach is to use the method by Kariwala (2004), which also handles the case
of multiple unstable modes directly, but is more involved than the simple pole-vector-based
method.
3 Simulink and Matlab models for the Tennessee Eastman process are available from Professor Larry Ricker at the
University of Washington (easily found using a search engine).
Exercise 10.5 For systems with multiple unstable poles, the variables can be selected sequentially
using the pole vector approach by stabilizing one real pole or a pair of complex poles at a time. Usually,
the selected variable does not depend on the controllers designed in the previous steps. Verify this for
each of the following two systems:

G1 (s) = Q(s) 10 2 1
G2 (s) = Q(s) 10 2 1
12 1.5 5.01 12 1 1.61

Q(s) = 1/(s 1) 0
0 1/(s 0.5)
(Hint: Use simple proportional controllers for stabilization of p = 1 and evaluate the effect of change
of controller gain on pole vectors in the second iteration.)
10.4.4 Local disturbance rejection: partial control

Let y1 denote the primary variables, and y2 the locally controlled variables. We start by
deriving the transfer functions for y1 for the partially controlled system when y2 is controlled.
We also partition the inputs u into the sets u1 and u2 , where the set u2 is used to control y2 .
The model y = Gu may then be written4
y1 = G11 u1 + G12 u2 + Gd1 d (10.24)
y2 = G21 u1 + G22 u2 + Gd2 d (10.25)
d
?
Gd1 Gd2
u1 +?+ y-
- G11 G12 - 1
u2 +?+
- G21 G22 -
y2
+?
n2
+
-? r2
K2
+
Figure 10.9: Partial control
Now assume that feedback control
u2 = K2 (r2 y2,m )
4 We may assume that any stabilizing loops have already been closed, so for the model y = Gu, G includes the
stabilizing controller and u includes any free setpoints to the stabilizing layer below.
is used for the secondary subsystem involving u2 and y2 , see Figure 10.9, where y2,m =
y2 + n2 is the measured value of y2 . By eliminating u2 and y2 , we then get the following
model for the resulting partially controlled system from u1 , r2 , d and n2 to y1 :

y1 = G11 G12 K2 (I + G22 K2 )1 G21 u1
| {z }
Pu

+ Gd1 G12 K2 (I + G22 K2 )1 Gd2 d
| {z }
Pd
+ G12 K2 (I + G22 K2 )1 (r2 n2 ) (10.26)

| {z }
Pr
Note that Pd , the partial disturbance gain, is the disturbance gain for a system under partial
control. Pu is the effect of u1 on y1 with y2 controlled. In many cases, the set u1 is empty
because there are no extra inputs. In such cases, r2 is probably available for control of y1 ,
and Pr gives the effect of r2 on y1 . In other cases, r2 may be viewed as a disturbance for the
control of y1 .
In the following discussion, we assume that the control of y2 is fast compared to the control
of y1 . This results in a time scale separation between these layers, which simplies controller
design. To obtain the resulting model we may let K2 in (10.26). Alternatively, we may
solve for u2 in (10.25) to get
u2 = G1 1 1
22 Gd2 d G22 G21 u1 + G22 y2 (10.27)
We have assumed that G22 is square and invertible, otherwise we can use a least squares
solution by replacing G1
22 by the pseudo-inverse, G22 . On substituting (10.27) into (10.24)

and assuming y2 r2 n2 (perfect control), we get
y1 (G11 G12 G1 G21 ) u1 + (Gd1 G12 G1 Gd2 ) d + G12 G1 (r2 n2 ) (10.28)

| {z 22 } | {z 22 } | {z 22} | {z }
y2
Pu Pd Pr
The advantage of the approximation (10.28) over (10.26) is that it is independent of K 2 , but
we stress that it is useful only at frequencies where y2 is tightly controlled.
Remark 1 Relationships similar to those given in (10.28) have been derived by many authors, e.g. see
the work of Manousiouthakis et al. (1986) on block relative gains and the work of Haggblom and Waller
(1988) on distillation control congurations.
Remark 2 Equation (10.26) may be rewritten in terms of linear fractional transformations (page 543).
For example, the transfer function from u1 to y1 is
Fl (G, K2 ) = G11 G12 K2 (I + G22 K2 )1 G21 (10.29)
Exercise 10.6 The block diagram in Figure 10.11 below shows a cascade control system where the
primary output y1 depends directly on the extra measurement y2 , so G12 = G1 G2 , G22 = G2 ,
Gd1 = [ I G1 ] and Gd2 = [ 0 I ]. Assume tight control of y2 . Show that Pd = [ I 0 ] and Pr = G1
and discuss the result. Note that Pr is the new plant as it appears with the inner loop closed.
The selection of secondary variables y2 depends on whether u1 or r2 (or any) are available
for control of y1 . Next, we consider in turn each of the three cases that may arise.
1. Cascade control system

Cascade control is a special case of partial control, where we use u2 to control (tightly) the
secondary outputs y2 , and r2 replaces u2 as a degree of freedom for controlling y1 . We would
like to avoid the introduction of additional (new) RHP-zeros, when closing the secondary
loops. The next theorem shows that this is not a problem.
Theorem 10.2 (RHP-zeros due to closing of secondary loop) Assume that n y1 = nu1
+ nu2 and ny2 = nr2 = nu2 (see Figure 10.9). Let the plant G = G11
G21
G12
G22
and the
secondary loop (S2 = (I + G22 K2 ) 1
) be stable. Then the partially controlled plant
PCL = [ G11 G12 K2 S2 G21 G12 K2 S2 ] (10.30)
from [u1 r2 ] to y1 in (10.26) has no additional RHP-zeros (that are not present in the open-
loop plant [ G11 G12 ] from [u1 u2 ] to y1 ) if
1. r2 is available for control of y1 , and
2. K2 is minimum-phase.
Proof: Under the dimensional and stability assumptions, PCL is a stable and square transfer function
matrix. Thus, the RHP-zeros of PCL are the points in RHP where det(PCL (s)) = 0 (also see Remark 4
on page 141). Using Schurs formula in (A.14),
det(PCL ) = det(M ) det(S2 )
where

G11 0 G12 K2
M=
G21 I I + G22 K2
with the partitioning as shown above. By exchanging the columns of M , we have

G11 G12 K2 0
det(M ) = (1)n det
G21 I + G22 K2 I

= det G11 G12 K2

I 0
= det G11 G12 det
0 K2

= det G11 G12 det(K2 )
The second equality follows since the rearranged matrix is block triangular and det(I) = (1) n .
Then, putting everything together, we have that

det(PCL ) = det G11 G12 det(K2 ) det(S2 )
Although the RHP-poles of K2 appear as RHP-zeros of S2 due to the interpolation constraints, these
zeros are cancelled by K2 and thus det(K2 ) det(S2 ) evaluated at RHP-poles of K2 is non-zero.
Therefore, when r2 is available for control of y1 and K2 is minimum-phase, the RHP-zeros of PCL
are the same as the RHP-zeros of [ G11 G12 ] and the result follows. When u1 is empty, the transfer
matrix from r2 to y1 is given as G12 K2 (I +G22 K2 )1 and thus K2 being minimum-phase implies that
the secondary loop does not introduce any additional RHP-zeros. A somewhat more restrictive version
of this theorem was proven by Larsson (2000). The proof here is due to V. Kariwala. Note that the
assumptions on the dimensions of y1 and u2 are made for simplicity of the proof and the conclusions
of Theorem 10.2 still hold when these assumptions are relaxed. 2
For a stable plant G, the controller K2 can usually be chosen to be minimum-phase. Then,
Theorem 10.2 implies that whenever r2 is available for control of y1 , closing the secondary
loops does not introduce additional RHP-zeros. However, note that closing secondary loops
may make the system more sensitive to disturbances if the action of the secondary (inner)
loop overcompensates and thereby makes the system more sensitive to the disturbance. As
an example consider a plant with Gd1 = 1, G12 = 1, G22 = 0.1 and Gd2 = 1. Then
with tight control of y2 , the disturbance gain for y1 increases by a factor 9, from Gd1 = 1
to Pd = Gd1 G12 G1 22 Gd2 = 9. In summary, it follows that we should select secondary
variables for cascade control such that the inputoutput controllability of the new partially
controlled plant PCL = [ G11 G12 K2 S2 G21 G12 K2 S2 ] = [ Pu Pr ] with disturbance
model Pd is better than that of the original plant [ G11 G12 ] with disturbance model Gd1 .
In particular, this requires that
1. ([ Pu Pr ]) (or (Pr ), if u1 is empty) is large at low frequencies.
2.
([ Pd Pr ]) is small and at least smaller than (Gd1 ). In particular, this argument
applies at higher frequencies. Note that Pr measures the effect of measurement noise n2
on y1 .
3. To ensure that u2 has enough power to reject the local disturbances d and track r2 , based
on (10.27), we require that 22 Gd2 ) < 1 and
(G1 22 ) < 1. Here, we have assumed
(G1
that the inputs have been scaled as outlined in Section 1.4.
Remark 1 The above recommendations for selection of secondary variables are stated in terms of
singular values, but the choice
of norm is usually of secondary importance. The minimization of
d
([ Pd Pr ]) arises if k
k 1 and we want to minimize ky1 k2 .
n2 2
Remark 2 By considering the cost function J = mind,n2 y1T y1 , the selection of secondary variables
for disturbance rejection using the objectives outlined above is closely related to the concept of self-
optimizing control discussed in Section 10.3.
2. Sequentially designed decentralized control system

When r2 is not available for control of y1 , we have a sequentially designed decentralized
controller. Here the variables y2 are important in themselves and we rst design a controller
K2 to control the subset y2 . With this controller K2 in place (a partially controlled system),
we may then design a controller K1 for the remaining outputs.
In this case, secondary loops can introduce new RHP-zeros in the partially controlled
system Pu . For example, this is likely to happen if we pair on negative RGA elements
(Shinskey, 1967; 1996); see Example 10.22 (page 447). Such zeros, however, can be moved
to high frequencies (beyond the bandwidth), if it is possible to tune the inner (secondary) loop
sufciently fast (Cui and Jacobsen, 2002).
In addition, based on the general objectives for variable selection, we require that (P u )
instead of ([ Pu Pr ]) be large. The other objectives for secondary variable selection are
the same as for cascade control and are therefore not repeated here.
3. Indirect control
Indirect control is when neither r2 nor u1 are available for control of y1 . The objective is to
minimize J = ky1 r1 k, but we assume that we cannot measure y1 . Instead we hope that y1
is indirectly controlled by controlling y2 . With perfect control of y2 , as before
y1 = Pd d + Pr (r2 n2 )
With n2 = 0 and d = 0 this gives y1 = G12 G1

22 r2 , so r2 must be chosen such that
r2 = G22 G1
12 r1 (10.31)
The control error in the primary output is then
y1 r 1 = P d d P r n 2 (10.32)
To minimize J = ky1 r1 k we should therefore (as for the two other cases) select the
controlled outputs y2 such that kPd dk and kPr n2 k are small or, in terms of singular values,
([ Pd Pr ]) is small. The problem of indirect control is closely related to that of cascade
control. The main difference is that in cascade control we also measure and control y 1 in an
outer loop; so in cascade control we need k [ Pd Pr ] k small only at frequencies outside the
bandwidth of the outer control loop (involving y1 ).
Remark 1 In some cases, this measurement selection problem involves a trade-off between wanting
kPd k small (wanting a strong correlation between measured outputs y 2 and primary outputs y1 )
and wanting kPr k small (wanting the effect of control errors (measurement noise) to be small). For
example, this is the case in a distillation column when we use temperatures inside the column (y 2 ) for
indirect control of the product compositions (y1 ). For a high-purity separation, we cannot place the
measurement close to the column end due to sensitivity to measurement error (kP r k becomes large),
and we cannot place it far from the column end due to sensitivity to disturbances (kP d k becomes large);
see also Example 10.9 (page 409).
Remark 2 Indirect control is related to the idea of inferential control which is commonly used in
the process industry. However, with inferential control the idea is usually to use the measurement of
y2 to estimate (infer) y1 and then to control this estimate rather than controlling y2 directly, e.g. see
Stephanopoulos (1984). However, there is no universal agreement on these terms, and Marlin (1995)
uses the term inferential control to mean indirect control as discussed above.
Optimal stabilizing control in terms of minimizing drift

A primary objective of the regulatory control system is to stabilize the plant in terms of
minimizing its steady-state drift from a nominal operating point. To quantify this, let w
represent the variables in which we would like to avoid drift; for example, w could be the
weighted states of the plant. For now let y denote the available measurements and u the
manipulated variables to be used for stabilizing control. The problem is: to minimize the
drift, which variables c should be controlled (at constant setpoints) by u? We assume linear
measurement combinations,
c = Hy (10.33)
and that we control as many variables as the number of degrees of freedom, n c = nu . The
linear model is
w = Gw u + Gw ew u
dd=G d

ey u
y = Gy u + Gyd d = G d
With perfect regulatory control (c = 0), the closed-loop response from d to w is

w = Pdw d; Pdw = Gw w y 1
d G (HG ) HGyd
Since generally nw > nu , we do not have enough degrees of freedom to make w = 0 (zero
drift). Instead, we seek the least squares solution that minimizes kwk2 . In the absence of
implementation error, an explicit solution, which also minimizes kPdw k2 , is
e w (G
H = (Gw )T G e y ) (10.34)
where we have assumed that we have enough measurements, ny nu + nd .
Proof of (10.34): We want to minimize
J = kwk22 = uT (Gw )T Gw u + dT (Gw T w T w T w
d ) Gd d + 2u (G ) Gd d
Then,
w T ew u
dJ/du = 2(Gw )T Gw u + 2(Gw )T Gw d d = 2(G ) G
d
An ideal self-optimizing variable is c = dJ/du, as then c =0 is always optimal with zero loss (in
the absence of implementation error). Now, c = Hy = H G e y u , so to get c = dJ/du, we would
d
like
HG e y = (Gw )T G
ew (10.35)
(the factor 2 does not matter). Since ny nu + nd , (10.35) has an innite number of solutions, and
the one using the right inverse of G e y is given by (10.34). It can be shown that the use of the right
inverse is optimal in terms of minimizing the effect of the (until now neglected) implementation error
on w, provided the measurements (y) have been normalized (scaled) with respect to their expected
measurement error (ny ) (Alstad, 2005, p. 52). The result (10.34) was originally proved by Hori et al.
(2005), but this proof is due to V. Kariwala.
2
H computed from (10.34) will be dynamic (frequency-dependent), but for practical
purposes, we recommend that it is evaluated at the closed-loop bandwidth frequency of the
outer loop that adjusts the setpoints for r. In most cases. it is acceptable to use the steady-state
matrices.
Example 10.11 Combination of measurements for minimizing drift of distillation column. We
consider the distillation column (column A) with the LV -conguration and use the same data as in
Example 10.9 (page 409). The objective is to minimize the steady-state drift of the 41 composition
variables (w = states) due to variations in the feed rate and feed composition by controlling a
combination of the available temperature measurements. We have u = L, n u = 1 and nd = 2 and we
need at least nu +nd = 1+2 = 3 measurements to achieve zero loss (see null space method, page 397).
We select three temperature measurements (y) at stages 15, 20 and 26. One reason for not selecting the
measurements located at the column ends is their sensitivity to implementation error, see Example 10.9.
By ignoring the implementation error, the optimal combination of variables that minimizes kP dw (0)k2
is, from (10.34),
c = 0.719T15 0.018T20 + 0.694T26
When c is controlled perfectly at cs = 0, this gives (Pdw (0)) = 0.363. This is signicantly
smaller than
(Gd (0)) = 9.95, which is the open-loop deviation of the state variables due to the
w
disturbances. We have not considered the effect of implementation error so far. Similar to (10.28), it can
be shown that the effect of implementation error on w is given by
(G w (Gy ) ). With an implementation
error of 0.05 in the individual temperature measurements, we get (G w (Gy ) ) = 0.135, which is
small.
10.5 Control conguration elements

In this section, we discuss in more detail some of the control conguration elements
mentioned above. We assume that the measurements y, manipulations u and controlled
outputs z are xed. The available synthesis theories presented in this book result in a
multivariable controller K which connects all available measurements/commands (y) with
all available manipulations (u),
u = Ky (10.36)
However, such a big (full) controller may not be desirable. By control conguration
selection we mean the partitioning of measurements/commands and manipulations within
the control layer. More specically, we dene
Control conguration. The restrictions imposed on the overall controller K by
decomposing it into a set of local controllers (subcontrollers, units, elements,
blocks) with predetermined links and with a possibly predetermined design
sequence where subcontrollers are designed locally.
In a conventional feedback system, a typical restriction on K is to use a one degree-of-
freedom controller (so that we have the same controller for r and y). Obviously, this
limits the achievable performance compared to that of a two degrees-of-freedom controller.
In other cases, we may use a two degrees-of-freedom controller, but we may impose the
restriction that the feedback part of the controller (Ky ) is rst designed locally for disturbance
rejection, and then the prelter (K r ) is designed for command tracking. In general, this will
limit the achievable performance compared to a simultaneous design (see also the remark on
page 111). Similar arguments apply to other cascade schemes.
Some elements used to build up a specic control conguration are:
Cascade controllers
Decentralized controllers
Feedforward elements
Decoupling elements
Selectors
These are discussed in more detail below, and in the context of the process industry in
Shinskey (1967, 1996) and Balchen and Mumme (1988). First, some denitions:
Decentralized control is when the control system consists of independent
feedback controllers which interconnect a subset of the output measure-
ments/commands with a subset of the manipulated inputs. These subsets should
not be used by any other controller.
This denition of decentralized control is consistent with its use by the control community.
In decentralized control, we may rearrange the ordering of measurements/commands and
manipulated inputs such that the feedback part of the overall controller K in (10.36) has a
xed block-diagonal structure.
Cascade control arises when the output from one controller is the input to
another. This is broader than the conventional denition of cascade control which
is that the output from one controller is the reference command (setpoint) to
another. In addition, in cascade control, it is usually assumed that the inner loop
(K2 ) is much faster than the outer loop (K1 ).
Feedforward elements link measured disturbances to manipulated inputs.
Decoupling elements link one set of manipulated inputs (measurements) with

another set of manipulated inputs. They are used to improve the performance
of decentralized control systems, and are often viewed as feedforward elements
(although this is not correct when we view the control system as a whole) where
the measured disturbance is the manipulated input computed by another
decentralized controller.
Selectors are used to select for control, depending on the conditions of the
system, a subset of the manipulated inputs or a subset of the outputs.
In addition to restrictions on the structure of K, we may impose restrictions on the way,

or rather in which sequence, the subcontrollers are designed. For most decomposed control
systems we design the controllers sequentially, starting with the fast or inner or lower-
layer control loops in the control hierarchy. Since cascade and decentralized control systems
depend more strongly on feedback rather than models as their source of information, it is
usually more important (relative to centralized multivariable control) that the fast control
loops are tuned to respond quickly.
In this section, we discuss cascade controllers and selectors, and in the following section,
we consider decentralized diagonal control. Let us rst give some justication for using such
suboptimal congurations rather than directly designing the overall controller K.
10.5.1 Why use simplied control congurations?

Decomposed control congurations can be quite complex, see for example Figure 10.13
(page 427), and it may therefore be both simpler and better in terms of control performance to
set up the controller design problem as an optimization problem and let the computer do the
job, resulting in a centralized multivariable controller as used in other chapters of this book.
If this is the case, why are simplied parameterizations (e.g. PID) and control
congurations (e.g. cascade and decentralized control) used in practice? There are a number
of reasons, but the most important one is probably the cost associated with obtaining good
plant models, which are a prerequisite for applying multivariable control. On the other hand,
with cascade and decentralized control the controllers are usually tuned one at a time with
a minimum of modelling effort, sometimes even on-line by selecting only a few parameters
(e.g., the gain and integral time constant of a PI controller). Thus:
A fundamental reason for applying cascade and decentralized control is to save on
modelling effort.
Other benets of cascade and decentralized control may include the following:
easy for operators to understand
ease of tuning because the tuning parameters have a direct and localized effect
insensitive to uncertainty, e.g. in the input channels
failure tolerance and the possibility of taking individual control elements into or out of
service
few control links and the possibility for simplied (decentralized) implementation
reduced computation load
The latter two benets are becoming less relevant as the cost of computing power is
reduced. Based on the above discussion, the main challenge is to nd a control conguration
which allows the (sub)controllers to be tuned independently based on a minimum of model
information (the pairing problem). For industrial problems, the number of possible pairings
is usually very high, but in most cases physical insight and simple tools, such as the RGA,
can be helpful in reducing the number of options to a manageable number. To be able to tune
the controllers independently, we must require that the loops interact only to a limited extent.
For example, one desirable property is that the steady-state gain from ui to yi in an inner
loop (which has already been tuned) does not change too much as outer loops are closed. For
decentralized diagonal control the RGA is a useful tool for addressing this pairing problem
(see page 450).
Remark. We just argued that the main advantage of applying cascade and decentralized control is that
the controllers can be tuned on-line and this saves on the modelling effort. However, in our theoretical
treatment we need a model, for example, to decide on a control conguration. This seems to be a
contradiction, but note that the model required for selecting a conguration may be more generic and
does not need to be modied for each particular application. Thus, if we have found a good control
conguration for one particular applications, then it is likely that it will work well also for similar
applications.
10.5.2 Cascade control systems

We want to illustrate how a control system which is decomposed into subcontrollers can be
used to solve multivariable control problems. For simplicity, we use SISO controllers here of
the form
ui = Ki (s)(ri yi ) (10.37)
where Ki (s) is a scalar. Note that whenever we close a SISO control loop we lose the
corresponding input, ui , as a degree of freedom, but at the same time the reference, ri ,
becomes a new degree of freedom.
It may look like it is not possible to handle non-square systems with SISO controllers.
However, since the input to the controller in (10.37) is a reference minus a measurement, we
can cascade controllers to make use of extra measurements or extra inputs. A cascade control
structure results when either of the following two situations arise:
The reference ri is an output from another controller (typically used for the case of an extra
measurement yi ), see Figure 10.10(a). This is conventional cascade control.
The measurement yi is an output from another controller (typically used for the case of
an extra manipulated input uj , e.g. in Figure 10.10(b) where u2 is the measurement for
controller K1 ). This cascade scheme where the extra input u2 is used to improve the
dynamic response, but is reset to a desired mid-range target value on a longer time scale,
is referred to as input resetting (also known as mid-ranging or valve position control).
10.5.3 Extra measurements: cascade control

In many cases, we make use of extra measurements y2 (secondary outputs) to provide local
disturbance rejection and linearization, or to reduce the effects of measurement noise. For
example, velocity feedback is frequently used in mechanical systems, and local ow cascades
y
+
r1 - + - K2
r2- -1
- K1 u-
Plant
-6 -6
y2
(a) Extra measurements y2 (conventional cascade control)

ru2 + u-
- - K1 1
-6 y
Plant -
+
r - - K2 u-
2
-6
(b) Extra inputs u2 (input resetting)
Figure 10.10: Cascade implementations
are used in process systems. For distillation columns, it is usually recommended to close an
inner temperature loop (y2 = T ), see Example 10.9.
A typical implementation with two cascaded SISO controllers is shown in Figure 10.10(a)
where
r2 = K1 (s)(r1 y1 ) (10.38)
u = K2 (s)(r2 y2 ) (10.39)
u is the manipulated input, y1 the controlled output (with an associated control objective r1 )
and y2 the extra measurement. Note that the output r2 from the slower primary controller
K1 is not a manipulated plant input, but rather the reference input to the faster secondary
(or slave) controller K2 . For example, cascades based on measuring the actual manipulated
variable (in which case y2 = um ) are commonly used to reduce uncertainty and nonlinearity
at the plant input.
r2 d2 d1
+
r1 - - K1 -+ - K2 u- G +?
- + - G +?
- + y-
1
-6 -6 2 1
y2
Figure 10.11: Common case of cascade control where the primary output y 1 depends directly on the
extra measurement y2
In the general case, y1 and y2 in Figure 10.10(a) are not directly related to each other,
and this is sometimes referred to as parallel cascade control. However, it is common to
encounter the situation in Figure 10.11
where y1 depends directly on y2 . This is a special case
G1 G2
of Figure 10.10(a) with Plant = , and it is considered further in Example 10.12
G2
and Exercise 10.7.
Remark. Centralized (parallel) implementation. Alternatively, we may use a centralized
implementation u = K(r y) where K is a 2-input 1-output controller. This gives
u = K11 (s)(r1 y1 ) + K12 (s)(r2 y2 ) (10.40)

where in most cases r2 = 0 (since we do not have a degree of freedom to control y2 ). With r2 = 0
in (10.40) the relationship between the centralized and cascade implementations is K 11 = K2 K1 and
K12 = K2 .
An advantage with the cascade implementation is that it more clearly decouples the design of the
two controllers. It also shows that r2 is not a degree of freedom at higher layers in the control system.
Finally, it allows for integral action in both loops (whereas usually only K 11 would have integral action
in (10.40)). On the other hand, a centralized implementation is better suited for direct multivariable
synthesis; see the velocity feedback for the helicopter case study in Section 13.2.
When should we use cascade control? With reference to the special (but common) case
of conventional cascade control shown in Figure 10.11, Shinskey (1967, 1996) states that the
principal advantages of cascade control are:
1. Disturbances arising within the secondary loop (before y2 in Figure 10.11) are corrected
by the secondary controller before they can inuence the primary variable y 1 .
2. Phase lag existing in the secondary part of the process (G2 in Figure 10.11) is reduced
measurably by the secondary loop. This improves the speed of response of the primary
loop.
3. Gain variations in the secondary part of the process are overcome within its own loop.
Morari and Zariou (1989) conclude, again with reference to Figure 10.11, that the use of an
extra measurement y2 is useful under the following circumstances:
(a) The disturbance d2 (entering before the measurement y2 ) is signicant and G 1 is non-
minimum-phase e.g. G1 contains an effective time delay [see Example 10.12].
(b) The plant G2 has considerable uncertainty associated with it e.g. G2 has a poorly known
nonlinear behaviour and the inner loop serves to remove the uncertainty.
In terms of design, they recommended that K2 is rst designed to minimize the effect of d 2
on y1 (with K1 = 0) and then K1 is designed to minimize the effect of d1 on y1 .
An example where local feedback control is required to counteract the effect of high-order
lags is given for a neutralization process in Figure 5.25 on page 216. The benets of local
feedback are also discussed by Horowitz (1991).
Exercise 10.7 We want to derive the above conclusions (a) and (b) from an inputoutput
controllability analysis, and also explain (c) why we may choose to use cascade control if we want
to use simple controllers (even with d2 = 0).
Outline of solution: (a) Note that if G1 is minimum-phase, then the inputoutput controllability of G2
and G1 G2 are in theory the same, and for rejecting d2 there is no fundamental advantage in measuring
y1 rather than y2 . (b) The inner loop L2 = G2 K2 removes the uncertainty if it is sufciently fast (high-
gain feedback). It yields a transfer function (I + L2 )1 L2 which is close to I at frequencies where K1
is active. (c) In most cases, such as when PID controllers are used, the practical closed-loop bandwidth
is limited approximately by the frequency wu , where the phase of the plant is 180 (see Section 5.8
on page 191), so an inner cascade loop may yield faster control (for rejecting d 1 and tracking r1 ) if the
phase of G2 is less than that of G1 G2 .
Tuning of cascaded PID controllers using the SIMC rules. Recall the SIMC PID
procedure presented on page 57, where the idea is to tune the controllers such that the
resulting transfer function from r to y is T ec s+1 . Here, is the effective delay in G
s
(from u to y) and c is a tuning parameter with c = being selected for fast (and still
robust) control. Let us apply this approach to the cascaded system in Figure 10.11. The inner
loop (K2 ) is tuned based on G2 . We then get y2 = T2 r2 , where T2 ec2 s+1 and 2 is the
2 s
effective delay in G2 . Since the inner loop is fast (2 and c2 are small), its response may be
approximated as a pure time delay for the tuning of the slower outer loop (K1 ),
T2 1 e(2 +c2 )s (10.41)
The resulting model for tuning of the outer loop (K1 ) is then
e 1 = G1 T2 G1 e(2 +c2 )s
G (10.42)
and the PID tuning parameters for K1 are easily obtained using the SIMC rules. For a fast
response from r2 to y2 in the inner loop, the SIMC-rule is to select c2 = 2 . However, this
may be unnecessarily fast and to improve robustness we may want to select a larger c2 . Its
value will not affect the outer loop, provided c2 < c1 /5 approximately, where c1 is the
response time in the outer loop.
Example 10.12 Consider the closed-loop system in Figure 10.11, where

(0.6s + 1) s 1
G1 = e and G2 =
(6s + 1) (6s + 1)(0.4s + 1)
We rst consider the case where we only use the primary measurement (y 1 ), i.e. design the
controller based on G = G1 G2 . Using the half rule on page 57, we nd that the effective delay is
1 = 6/2 + 0.4 + 0.6 + 1 = 5, and using the SIMC tuning rules on page 57, a PI controller is designed
with Kc = 0.9 and I = 9. The closed-loop response of the system to step changes of magnitude 1 in
the setpoint (at t = 0) and of magnitude 6 in disturbance d2 (at t = 50) is shown in Figure 10.12. From
the dashed line, we see that the closed-loop disturbance rejection is poor.
5
4 PSfrag replacements
3 Without Cascade
y1 (t)
2
With Cascade Setpoint
1
Setpoint Change (r1 ) Disturbance Change (d2 )
0
0 20 40 60 80 100
Time [sec]
Figure 10.12: Improved control performance with cascade control (solid) as compared to single-loop
control (dashed)
Next, to improve disturbance rejection, we make use of the measurement y 2 in a cascade

implementation as shown in Figure 10.11. First, the PI controller for the inner loop is designed based
on G2 . The effective delay is 2 = 0.2. For fast control the SIMC rule (page 57) is to use c2 = 2 .
However, since this is an inner loop, where tight control is not critical, we choose c2 = 22 = 0.4,
which gives somewhat less aggressive settings with Kc2 = 10.33 and I2 = 2.4. The PI controller for
the outer loop is next designed with the inner loop closed. From (10.41), the transfer function for the
inner loop is approximated as a delay of c2 + 2 = 0.6 giving G e 1 = G1 e0.6s = (0.6s+1) e1.6s .
(6s+1)
Thus, for the outer loop, the effective delay is 1 = 0.6 + 1.6 = 2.2 and with c1 = 1 = 2.2 (fast
control), the resulting SIMC PI tunings are Kc1 = 1.36 and I1 = 6. From Figure 10.12, we note that
the cascade controller greatly improves the rejection of d2 . The speed of the setpoint tracking is also
improved, because the local control (K2 ) reduces the effective delay for control of y1 .
Exercise 10.8 To illustrate the benet of using inner cascades for high-order plants, consider
Figure 10.11 and a plant G = G1 G2 G3 G4 G5 with
1
G1 = G 2 = G 3 = G 4 = G 5 =
s+1
Consider the following two cases:
(a) Measurement of y1 only, i.e. G = 1
(s+1)5
.
(b) Four additional measurements available (y2 , y3 , y4 , y5 ) on outputs of G1 , G2 , G3 and G4 .
For case (a) design a PID controller and for case (b) use ve simple proportional controllers with gains
with gains 10 (innermost loop), 5, 2, 1 and 0.5 (outer loop) (note that the gain has to be smaller in the
outer loop to avoid instability caused by the effective delay in the inner loop). For case (b) also try using
a PI controller in the outer loop to avoid the steady-state offset. Compare the responses to disturbances
entering before G1 (at t = 0), G2 (t = 20), G3 (t = 40), G4 (t = 60), G5 (t = 80), and for a setpoint
change (t = 100).
10.5.4 Extra inputs

In some cases, we have more manipulated inputs than controlled outputs. These may be used
to improve control performance. Consider a plant with a single controlled output y and two
manipulated inputs u1 and u2 . Sometimes u2 is an extra input which can be used to improve
the fast (transient) control of y, but if it does not have sufcient power or is too costly to
use for long-term control, then after a while it is reset to some desired value (ideal resting
value).
Cascade implementation (input resetting). An implementation with two cascaded SISO
controllers is shown in Figure 10.10(b). We let input u2 take care of the fast control and u1
the long-term control. The fast control loop is then
u2 = K2 (s)(r y) (10.43)
The objective of the other slower controller is then to use input u1 to reset input u2 to its
desired value ru2 :
u1 = K1 (s)(ru2 y1 ), y1 = u2 (10.44)
and we see that the output u2 from the fast controller K2 is the measurement y1 for the
slow controller K1 .
In process control, the cascade implementation with input resetting often involves valve
position control, because the extra input u2 , usually a valve, is reset to a desired position by
the outer cascade.
Centralized (parallel) implementation. Alternatively, we may use a centralized
implementation u = K(r y) where K is a 1-input 2-output controller. This gives
u1 = K11 (s)(r y), u2 = K21 (s)(r y) (10.45)
Here two inputs are used to control one output, so to get a unique steady-state for the inputs u 1
and u2 we usually let K11 have integral control, whereas K21 does not. Then u2 (t) will only
be used for transient (fast) control and will return to zero (or more precisely to its desired
value ru2 ) as t . With ru2 = 0 the relationship between the centralized and cascade
implementation is K11 = K1 K2 and K21 = K2 .
Comparison of cascade and centralized implementations. The cascade implementation
in Figure 10.10(b) has the advantage, compared to the centralized (parallel) implementation,
of decoupling the design of the two controllers. It also shows more clearly that r u2 , the
reference for u2 , may be used as a degree of freedom at higher layers in the control system.
Finally, we can have integral action in both K1 and K2 , but note that the gain of K1 should
be negative (if effects of u1 and u2 on y are both positive).
Exercise 10.9 Draw the block diagrams for the two centralized (parallel) implementations
corresponding to Figure 10.10.
Exercise 10.10 Derive the closed-loop transfer functions for the effect of r on y, u 1 and u2 in the
cascade input resetting scheme of Figure 10.10(b). As an example use G = [ G 11 G12 ] = [ 1 1 ] and
use integral action in both controllers, K1 = 1/s and K2 = 10/s. Show that input u2 is reset at
steady-state.
10.5.5 Extra inputs and outputs

In some cases performance may be improved with local control loops involving both extra
manipulated inputs and extra measurements. However, as always, the improvement must be
traded off against the cost of the extra actuators, measurements and control system.
Example 10.13 Two layers of cascade control. Consider the system in Figure 10.13 with two
manipulated plant inputs (u2 and u3 ), one controlled output (y1 , which should be close to r1 ) and
two measured variables (y1 and y2 ). Input u2 has a more direct effect on y1 than does input u3 (since
there is a large delay in G3 (s)). Input u2 should only be used for transient control as it is desirable that
it remains close to r3 = ru2 . The extra measurement y2 is closer than y1 to the input u2 and may be
useful for detecting disturbances (not shown) affecting G1 .
r1
-? -
- u1
K1
+
- -
r2 - ? u2
- +
-
y2
- - y1
K2 G1 G2
+ +
6
r3
-? -
- u3
-
K3 G3
+
Figure 10.13: Control conguration with two layers of cascade control
In Figure 10.13, controllers K1 and K2 are cascaded in a conventional manner, whereas controllers
K2 and K3 are cascaded to achieve input resetting. The input u1 is not a (physical) plant input, but it
does play the role of an input (manipulated variable) as seen from the controller K 1 . The corresponding
equations are
u1 = K1 (s)(r1 y1 ) (10.46)
u2 = K2 (s)(r2 y2 ), r2 = u1 (10.47)
u3 = K3 (s)(r3 y3 ), y3 = u2 (10.48)
Controller K1 controls the primary output y1 at its reference r1 by adjusting the input u1 , which
is the reference value for y2 . Controller K2 controls the secondary output y2 using input u2 . Finally,
controller K3 manipulates u3 slowly in order to reset input u2 to its desired value r3 .
Typically, the controllers in a cascade system are tuned one at a time starting with the
fastest loop. For example, for the control system in Figure 10.13 we would probably tune the
three controllers in the order K2 (inner cascade using fast input), K3 (input resetting using
slower input), and K1 (nal adjustment of y 1 ).
Exercise 10.11 Process control application. A practical case of a control system like the one in
Figure 10.13 is in the use of a pre-heater to keep a reactor temperature y 1 at a given value r1 . In this
case, y2 may be the outlet temperature from the pre-heater, u2 the bypass ow (which should be reset to
r3 , say 10% of the total ow), and u3 the ow of heating medium (steam). Process engineering students:
Make a process owsheet with instrumentation lines (not a block diagram) for this heater/reactor
process.
10.5.6 Selectors
Split-range control for extra inputs. We considered above the case where the primary input
is slow, and an extra input is added to improve the dynamic performance. For economic
reasons or to avoid saturation the extra input is reset to a desired mid-range target value on
a longer time scale (input resetting or mid-ranging). Another situation is when the primary
input may saturate, and an extra input is added to maintain control of the output. In this
case, the control range is often split such that, for example, u1 is used for control when
y [ymin , y1 ], and u2 is used when y [y1 , ymax ].
Selectors for too few inputs. A completely different situation occurs if there are too few
inputs. Consider the case with one input (u) and several outputs (y1 , y2 , . . .). In this case,
we cannot control all the outputs independently, so we either need to control all the outputs
in some average manner, or we need to make a choice about which outputs are the most
important to control. Selectors or logic switches are often used for the latter. Auctioneering
selectors are used to decide to control one of several similar outputs. For example, such a
selector may be used to adjust the heat input (u) to keep the maximum temperature (max i yi )
in a red heater below some value. Override selectors are used when several controllers
compute the input value, and we select the smallest (or largest) as the input. For example, this
is used in a heater where the heat input (u) normally controls temperature (y 1 ), except when
the pressure (y2 ) is too large and pressure control takes over.
K(s)
r1 +? - u1 y1
- - k1 - -
G(s)
r2 +
- - k2
u2
- -
y2
-
6
Figure 10.14: Decentralized diagonal control of a 2 2 plant
10.6 Decentralized feedback control

10.6.1 Introduction
We have already discussed, in the previous sections on control congurations, the use of
decentralized control, but here we consider it in more detail. To this end, we assume in this
section that G(s) is a square plant which is to be controlled using a diagonal controller (see
Figure 10.14)
k1 (s)
k2 (s)
K(s) = diag{ki (s)} =
..

(10.49)
.
km (s)
This is the problem of decentralized (or diagonal) feedback control.
It may seem like the use of decentralized control seriously limits the achievable control
performance. However, often the performance loss is small, partly because of the benets
of high-gain feedback. For example, it can be proved theoretically (Zames and Bensoussan,
1983) that with decentralized control one may achieve perfect control of all outputs, provided
the plant has no RHP-zeros that limit the use of high feedback gains. Furthermore, for
a stable plant G(s) (also with RHP-zeros), it is possible to use integral control in all
channels (to achieve perfect steady-state control) if and only if G(0) is non-singular (Campo
and Morari, 1994). Both these conditions are also required with full multivariable control.
Nevertheless, for interactive plants and nite bandwidth controllers, there is a performance
loss with decentralized control because of the interactions caused by non-zero off-diagonal
elements in G. The interactions may also cause stability problems. A key element in
decentralized control is therefore to select good pairings of inputs and outputs, such that
the effect of the interactions is minimized.
The design of decentralized control systems typically involves two steps:
1. The choice of pairings (control conguration selection).

2. The design (tuning) of each controller, ki (s).
The optimal solution to this problem is very difcult mathematically. First, the number of
pairing options in step 1 is m! for an m m plant and thus increases exponentially with the
size of the plant. Second, the optimal controller in step 2 is in general of innite order and
may be non-unique. In step 2, there are three main approaches:
Fully coordinated design. All the diagonal controller elements ki (s) are designed
simultaneously based on the complete model G(s). This is the theoretically optimal
approach for decentralized control, but it is not commonly used in practice. First,
as just mentioned, the design problem is very difcult. Second, it offers few of the
normal benets of decentralized control (see page 421), such as ease of tuning,
reduced modelling effort, and good failure tolerance. In fact, since a detailed dynamic
model is required for the design, an optimal coordinated decentralized design offers
few benets compared to using a full multivariable controller which is easier to
design and has better performance. The exception is situations where multivariable
control cannot be used, for example, when centralized cooordination is difcult
for geographical reasons. We do not address the optimal coordinated design of
decentralized controllers in this book, and the reader is referred to the literature (e.g.
Sourlas and Manousiouthakis, 1995) for more details.
Independent design. Each controller element ki (s) is designed based on the corresponding
diagonal element of G(s), such that each individual loop is stable. Possibly, there
is some consideration of the off-diagonal interactions when tuning each loop. This
approach is the main focus in the remaining part of this chapter. It is used when it is
desirable that we have integrity where the individual parts of the system (including each
loop) can operate independently. The pairing rules on page 450 can be used to obtain
pairings for independent design. In short the rules are to (1) pair on RGA elements
close to 1 at crossover frequencies, (2) pair on positive steady-state RGA elements,
and (3) pair on elements that impose minimal bandwidth limitations (e.g., small delay).
The rst and second rules are to avoid that the interactions cause instability. The third
rule follows because we for good performance want to use high-gain feedback, but we
require stable individual loops. For many interactive plants, it is not possible to nd a
set of pairing satisfying all the three rules.
Sequential design. The controllers are designed sequentially, one at a time, with the
previously designed (inner) controllers implemented. This has the important
advantage of reducing each design to a scalar (SISO) problem, and is well suited for
on-line tuning. The sequential design approach can be used for interactive problems
where the independent design approach does not work, provided it is acceptable to have
slow control of some output so that we get a difference in the closed-loop response
times of the outputs. One then starts by closing the fast inner loops (involving the
outputs with the fastest desired response times), and continues by closing the slower
outer loops. The main disadvantage with this approach is that failure tolerance is not
guaranteed when the inner loops fail (integrity). In particular, the individual loops are
not guaranteed to be stable. Furthermore, one has to decide on the order in which to
close the loops.
The effective use of a decentralized controller requires some element of decoupling.
Loosely speaking, independent design is used when the system is decoupled in space (G(s)
is close to diagonal), whereas sequential design is used when the system outputs can be
decoupled in time.
The analysis of sequentially designed decentralized control systems may be performed
using the results on partial control presented earlier in this chapter. For example, after closing
the inner loops (from
u2 to y2 ), the transfer functionfor the remaining outer system (from u1
to y1 ) is Pu = G11 G12 K2 (I + G22 K2 )1 G21 ; see (10.26). Notice that in the general
case we need to take into account the details of the controller K2 . However, when there is
a time scale separation between the layers with the fast loops (K2 ) being closed rst, then
we may for the design of K1 assume K2 (perfect control of y2 ), and the transfer
function for the remaining slow outer system becomes Pu = G11 G12 G1 22 G21 ; see
(10.28). The advantages of the time scale separation for sequential design of decentralized
controllers (with fast inner and slow outer loops), are the same as those for hierarchical
cascade control (with fast lower and slow upper layers) as listed on page 387. Examples
of sequential design are given in Example 10.15 (page 433) and in Section 10.6.6 (page 446).
The relative gain array (RGA) is a very useful tool for decentralized control. It is dened as
= G (G1 )T , where denotes element-by-element multiplication. It is recommended
to read the discussion about the original interpretation of the RGA on page 83, before
continuing. Note in particular from (3.56) that each RGA element represents the ratio between
the open-loop (gij ) and closed-loop (b gij ) gains for the corresponding input-output pair,
ij = gij /b gij . By closed-loop here we mean partial control with the other outputs
perfectly controlled. Intuitively, we would like to pair on elements with ij (s) close to 1,
because then the transfer function from uj to yi is unaffected by closing the other loops.
Remark. We assume in this section that the decentralized controllers k i (s) are scalar. The treatment
may be generalized to block-diagonal controllers by, for example, introducing tools such as the block
relative gain; e.g., see Manousiouthakis et al. (1986) and Kariwala et al. (2003).
10.6.2 Introductory examples

To provide some insight into decentralized control and to motivate the material that follows
we start with some simple 2 2 examples. We assume that the outputs y1 and y2 have
been scaled so that the allowable control errors (ei = yi ri ), i = 1, 2 are approximately
between 1 and 1. We design the decentralized controller to give rst-order responses with
time constant i in each of the individual loops, that is, yi = i s+1 1
ri . For simplicity, the
plants have no dynamics, and the individual controllers are then simple integral controllers
ki (s) = g1ii 1i s ; see the IMC design procedure on page 54. To make sure that we do not use
aggressive control, we use (in all simulations) a real plant, where we add a delay of 0.5
time units in each output, i.e. Gsim = Ge0.5s . This delay is not included in the analytic
expressions, e.g. (10.52), in order to simplify our discussion, but it is included for simulation
and tuning. With a delay of 0.5 we should, for stability and acceptable robustness, select
i 1; see the SIMC rule for fast but robust control on page 57. In all simulations we drive
the system with reference changes of r1 = 1 at t = 0 and r2 = 1 at t = 20.
Example 10.14 Diagonal plant. Consider the simplest case of a diagonal plant

g11 g12 1 0
G= = (10.50)
g21 g22 0 1
with RGA = I. The off-diagonal elements are zero, so there are no interactions and decentralized
control with diagonal pairings is obviously optimal.
Diagonal pairings. The controller
1
0
K = 1 s 1 (10.51)
0 2 s
gives nice decoupled rst-order responses
1 1
y1 = r1 and y2 = r2 (10.52)
1 s + 1 2 s + 1
1
setpoint
0.5
y(t)
y1 y2
PSfrag replacements
0
0 5 10 15 20 25 30 35 40
Time [sec]
(a) Diagonal pairing; controller (10.51) with 1 = 2 = 1
2
y1
y(t)
y2 setpointreplacements
PSfrag
0
0 5 10 15 20 25 30 35 40
Time [sec]
(b) Off-diagonal pairing; plant (10.53) and controller (10.54)
Figure 10.15: Decentralized control of diagonal plant (10.50)
as illustrated in Figure 10.15(a) for the case with 1 = 2 = 1.

Off-diagonal pairings. When considering pairings other than diagonal, we recommend to rst
permute the inputs such that the paired elements are along the diagonal. For the off-diagonal pairing,
we use the permuted inputs
u1 = u2 , u2 = u1
corresponding to the permuted plant (denoted with )
T
0 1 g g11 0 1
G = G = 12 = (10.53)
1 0 g22 g21 1 0
This corresponds to pairing on two zero elements, g11

= 0 and g22

= 0, and we cannot use independent
or sequential controller design. A coordinated (simultaneous) controller design is required and after
some trial and error we arrived at the following design
(0.5s+0.1)
0
K (s) = s
(0.5s+2) (10.54)
0 s
Performance is of course quite poor as is illustrated in Figure 10.15(b), but it is nevertheless workable
(surprisingly!).
Remark. The last example, where a diagonal plant is controlled using the off-diagonal
pairings, is quite striking. A simple physical example is the control of temperatures in two
unrelated rooms, say one located in the UK (Ians ofce) and one in Norway (Sigurds ofce).
The setup is then that Ian gets a measurement of Sigurds room temperature, and based on
this adjusts the heating in his room (in the UK). Similarly, Sigurd gets a measurement of
Ians room temperature, and based on this adjusts the heating in his room (in Norway). As
shown in Figure 10.15(b), such a ridiculous setup (with g11 = 0 and g2 = 0) is actually
workable because of the hidden feedback loop going through the off-diagonal elements
and the controllers (k1 k2 g12 g21 is nonzero), provided one is able to tune the controllers k1
and k2 (which is not trivial as seen it requires a negative sign in one of the controllers). Two
lessons from this example are that (1) decentralized control can work for almost any plant,
and (2) the fact that we have what seems to be acceptable closed-loop performance does not
mean that we are using the best pairing.
Exercise 10.12 Consider in more detail the off-diagonal pairings for the diagonal plant in the
example above. (i) Explain why it is necessary to use a negative sign in (10.54). (ii) Show that the
plant (10.53) cannot be stabilized by a pure integral action controller of the form K (s) = diag( ksi ).
Example 10.15 One-way interactive (triangular) plant. Consider

1 0
G= (10.55)
5 1
for which
1 0 1 0
G1 = and RGA =
5 1 0 1
The RGA matrix is identity, which suggests that the diagonal pairings are best for this plant. However,
we see that there is a large interaction (g21 = 5) from u1 to y2 , which, as one might expect, implies
poor performance with decentralized control. Note that this is not a fundamental control limitation as
1 0
the decoupling controller K(s) = 1
s 5 1
gives nice decoupled responses, identical to those shown
in Figure10.15 (but the decoupler may be sensitive to uncertainty; see Exercise 10.13).
Diagonal pairings using independent design. If we use independent design based on the paired
(diagonal) elements only (without considering the interactions caused by g 21 = 5 6= 0), then the
controller becomes 1
0
K= 1 s
1 (10.56)
0 2 s
with 1 = 2 = 1 (assuming a 0.5 time delay). However, a closer analysis shows that the closed-loop
response with the controller (10.56) becomes
1
y1 = r1 (10.57)
1 s + 1
52 s 1
y2 = r1 + r2 (10.58)
(1 s + 1)(2 s + 1) 2 s + 1
If we plot the interaction term from r1 to y2 as a function of frequency, then we nd that for 1 = 2 it
has a peak value of about 2.5. Therefore, with this controller the response for y 2 is not acceptable when
we make a change in r1 . To keep this peak below 1, we need to select 1 52 , approximately. This is
illustrated in Figure10.16(a) where we have selected 1 = 5 and 2 = 1. Thus, to keep |e2 | 1, we
must accept slow control of y1 .
Remark. The performance problem was not detected from the RGA matrix, because it only measures
two-way interactions. However, it may be detected from the Performance RGA matrix (PRGA), which
for our plant with unity diagonal elements is equal to G1 . As discussed on page 438, a large element
in a row of PRGA indicates that fast control is needed to get acceptable reference tracking. Thus, the
2, 1 element in G1 of magnitude 5, conrms that control of y2 must be about 5 times faster than that
of y1 .
1
y1 y2 setpoint
0
y(t)
PSfrag replacements
1
0 5 10 15 20 25 30 35 40
Time [sec]
(a) Diagonal pairing; controller (10.56) with 1 = 5 and 2 = 1
1
setpoint
y1 y2
0
y(t)
PSfrag replacements
1
0 5 10 15 20 25 30 35 40
Time [sec]
(b) Off-diagonal pairing; plant (10.59) and controller (10.60) with 1 = 5 and 2 = 1
Figure 10.16: Decentralized control of triangular plant (10.55)
Off-diagonal pairings using sequential design. The permuted plant is

T
0 1 0 1
G = G = (10.59)
1 0 1 5
This corresponds to pairing on a zero element g11

= 0. This pairing is not acceptable if we use the
independent design approach, because u1 has no effect on y1 so loop 1 does not work by itself.

However, with the sequential design approach, we may rst close the loop around y 2 (on the element

g22 = 5). With the IMC design approach, the controller becomes k2 (s) = 1/(g22
2 s) = 1/(52 s)
and with this loop closed, u1 does have an effect on y1 . Assuming tight control of y2 gives (using the
expression for perfect partial control in (10.28))

g g 1
y1 = g11 12 21 u1 = u1
g22 5
The controller for the pairing u1 -y1 becomes k1 (s) = 1/(g11

1 s) = 5/(1 s) and thus
5
0
K = 1 s 1 (10.60)
0 52 s
The response with 1 = 5 and 2 = 1 is shown in Figure 10.16(b). We see that performance is only
slightly worse than with the diagonal pairings. However, more seriously, we have the problem that if
control of y2 fails, e.g. because u2 = u1 saturates, then we also lose control of y1 (in addition, we get
instability with y2 drifting away, because of the integral action for y1 ). The situation is particularly bad
in this case because of the pairing on a zero element, but the dependence on faster (inner) loops being
in service is a general problem with sequential design.
Exercise 10.13 . Redo the

simulations
in Example 10.15 with 20% diagonal input uncertainty.
1.2 0
Specically, add a block 0 0.8
between the plant and the controller. Also simulate with the

1 0
decoupler K(s) = 1
s 5 1
which is expected to be particularly sensitive to uncertainty (why?
see conclusions on page 251 and note that I (G) = 10 for this plant).
Example 10.16 Two-way interactive plant. Consider the plant

1 g12
G= (10.61)
5 1
for which

1 1 g12 1 1 5g12
G1 = and RGA =
1 5g12 5 1 1 5g12 5g12 1
The control properties of this plant depend on the parameter g 12 . The plant is singular (det(G) =
1 5g12 = 0) for g12 = 0.2, and in this case independent control of both outputs is impossible,
whatever the controller. We will examine the diagonal pairings using the independent design controller
1
0
K = 1 s 1 (10.62)
0 2 s
The individual loops are stable with responses y1 = (1 s+1)

1
r1 and y2 = (2 s+1)
1
r2 , respectively. With
both loops closed, the response is y = GK(I + GK) r = T r, where
1

1 2 s + 1 5g12 g12 1 s
T =
(1 s + 1)(2 s + 1) 5g12 52 s 1 s + 1 5g12
We see that T (0) = I, so we have perfect steady-state control, as is expected with integral action.
However, the interactions as expressed by the term 5g12 may yield instability, and we nd that the
system is closed-loop unstable for g12 > 0.2. This is also expected because the diagonal RGA elements
are negative for g12 > 0.2, indicating a gain change between the open-loop (g ii ) and closed-loop (b
gii )
transfer functions, which is incompatible with integral action. Thus, for g 12 > 0.2, the off-diagonal
pairings must be used if we want to use an independent design (with stable individual loops).
We will now consider three cases, (a) g12 = 0.17, (b) g12 = 0.2 and (c) g12 = 1, each with the
same controller (10.62) with 1 = 5 and 2 = 1. Because of the large interactions given by g21 = 5,
we need to control y2 faster than y1 .
(a) g12 = 0.17. In this case,

6.7 1.1 6.7 5.7
G1 = and RGA =
33.3 6.7 5.7 6.7
The large RGA elements indicate strong interactions. Furthermore, recall from (3.56) that the
RGA gives the ratio of the open-loop and (partially) closed-loop gains, g ij /bgij . Thus, in terms of
decentralized control, the large positive RGA elements indicate that gbij is small and the loops will
tend to counteract each other by reducing the effective loop gain. This is conrmed by simulations
in Figure 10.17(a).
(b) g12 = 0.2. In this case,

0.5 0.1 0.5 0.5
G1 = and RGA =
2.5 0.5 0.5 0.5
The RGA elements of 0.5 indicate quite strong interactions and show that the interaction increases
the effective gain. This is conrmed by the closed-loop responses in Figure 10.17(b).
2
y2
1
setpoint
y(t)
0 y1
PSfrag replacements
1
0 5 10 15 20 25 30 35 40
Time [sec]
(a) g12 = 0.17; controller (10.62) with 1 = 5 and 2 = 1
1
y1 y2 setpoint
y(t)
1
0 5 10 15 20 25 30 35 40
Time [sec]
(b) g12 = 0.2; controller (10.62) with 1 = 5 and 2 = 1
2
y1
1
setpoint
y(t)
0 y2 PSfrag replacements
1
0 5 10 15 20 25 30 35 40
Time [sec]
(c) g12 = 1; controller (10.62) with 1 = 5 and 2 = 1
2
y2
1
y1 setpoint
y(t)
1
0 5 10 15 20 25 30 35 40
Time [sec]
(d) g12 = 1; controller (10.62) with 1 = 21.95 and 2 = 1
Figure 10.17: Decentralized control of plant (10.61) with diagonal pairings

(c) g12 = 1. In this case,

0.17 0.17 0.17 0.83
G1 = and RGA =
0.83 0.17 0.83 0.17
The RGA indicates clearly that the off-diagonal pairings are preferable. Nevertheless, we will
consider the diagonal pairings with 1 = 5 and 2 = 1 (as before). The response is poor as seen in
Figure 10.17(c). The closed-loop system is stable, but very oscillatory. This is not surprising as the
diagonal RGA elements of 0.17 indicate that the interactions increase the effective loop gains by a
factor 6 (= 1/0.17). To study this in more detail, we write the closed-loop polynomial in standard
form
(1 s + 1)(2 s + 1) 5g12 = 2 s2 + 2 s + 1
with r
1 2 1 1 + 2 1
= and =
1 5g12 2 1 2 1 5g12
We note that we get oscillations (0 < < 1), when g12 is negative and large. For example,
g12 = 1, 1 = 5 and 2 = 1 gives = 0.55. Interestingly, we see from the expression for
that the oscillations may be reduced by selecting 1 and 2 to be more different. This follows because
1 +2
1
2 1 2
is the ratio between the arithmetic and geometric means, which is larger the more different
1 and 2 are. Indeed, with g12 = 1 we nd that oscillations can be eliminated ( = 1) by selecting
1 = 21.952 . This is conrmed by the simulations in Figure10.17(d). The response is surprisingly
good taking into account that we are using the wrong pairings.
Exercise 10.14 Design decentralized controllers for the 3 3 plant G(s) = G(0)e 0.5s where G(0)
is given by (10.80). Try both the diagonal pairings and the pairings corresponding to positive steady-
" #T
0 1 0
state RGA elements, i.e. G = G 1 0 0 .
0 0 1
The above examples show that in many cases we can achieve quite good performance
with decentralized control, even for interactive plants. However, decentralized controller
design is more difcult for such plants, and this, in addition to the possibility for improved
performance, favours the use of multivariable control for interactive plants.
With the exception of Section 10.6.6, the focus in the rest of this chapter is on
independently designed decentralized control systems, which cannot be analyzed using the
expressions for partial control presented earlier in (10.28). We present tools for pairing
selections (step 1) and for analyzing the stability and performance of decentralized control
systems based on independent design. Readers who are primarily interested in applications of
decentralized control may want to go directly to the summary in Section 10.6.8 (page 449).
10.6.3 Notation and factorization of sensitivity function

G(s) denotes a square m m plant with elements gij . With a particular choice of pairings
we can rearrange the columns or rows of G(s) such that the paired elements are along the
diagonal of G(s). We then have that the controller K(s) is diagonal (diag{ki }). We introduce

g11
g22
e , diag{gii } =
G ..

(10.63)
.
gmm
as the matrix consisting of the diagonal elements of G. The loop transfer function in loop i is
denoted Li = gii ki , which is also equal to the ith diagonal element of L = GK.

1
Se , (I + GK)
e 1 = diag and Te = I Se (10.64)
1 + gii ki
contain the sensitivity and complementary sensitivity functions for the individual loops. Note
that Se is not equal to the matrix of diagonal elements of S = (I + GK)1 .
With decentralized control, the interactions are given by the off-diagonal elements G G.
e
The interactions can be normalized with respect to the diagonal elements and we dene
e G
E , (G G) e 1 (10.65)
The magnitude of the matrix E is commonly used as an interaction measure. We will

show that (E) (where is the structured singular value) is the best (least conservative)
measure, and will dene generalized diagonal dominance to mean (E) < 1. To derive
these results we make use of the following important factorization of the overall sensitivity
function S = (I + GK)1 with all loops closed,
S =
|{z} Se
|{z} (I + E Te)1 (10.66)
| {z }
overall individual loops interactions
Equation (10.66) follows from (A.147) with G = G e and G0 = G. The reader is encouraged
to conrm that (10.66) is correct, because most of the important results for stability and
performance using independent design may be derived from this expression.
A related factorization which follows from (A.148) is
e ES S)
S = S(I e 1 (I ES ) (10.67)
where
e 1
ES = (G G)G (10.68)
(10.67) may be rewritten as
e I))1 S
S = (I + S( e (10.69)
where is the performance relative gain array (PRGA),
e
(s) , G(s)G 1
(s) (10.70)
is a normalized inverse of the plant. Note that ES = I and E = 1 I. In Section

10.6.7 we discuss in more detail the use of the PRGA.
These factorizations are particularly useful for analyzing decentralized control systems
based on independent design, because the basis is then the individual loops with transfer
function S.
e
10.6.4 Stability of decentralized control systems

We consider the independent design procedure and assume that (a) the plant G is stable and
(b) each individual loop is stable by itself (Se and Te are stable). Assumption (b) is the basis
for independent design. Assumption (a) is also required for independent design because we
want to be able to take any loop(s) out of service and remain stable, and this is not possible if
the plant is unstable.
To achieve stability of the overall system with all loops closed, we must require that the
interactions do not cause instability. We use the expressions for S in (10.66) and (10.69) to
derive conditions for this.
Theorem 10.3 With assumptions (a) and (b), the overall system is stable (S is stable):
(i) if and only if (I + E Te)1 is stable, where E = (G G) e 1 ,
e G
(ii) if and only if det(I + E T (s)) does not encircle the origin as s traverses the Nyquist
e
D-contour,
(iii) if
(E Te(j)) < 1, (10.71)
(iv) (and (10.71) is satised) if
(Te) = max |e
ti | < 1/(E) (10.72)
i
The structured singular value (E) is computed with respect to a diagonal structure (of Te).
Proof: (Grosdidier and Morari, 1986) (ii) follows from the factorization S = S(I e + E Te)1 in (10.66)
and the generalized Nyquist theorem in Lemma A.5 (page 543). (iii) Condition (10.71) follows from
the spectral radius stability condition in (4.110). (iv) The least conservative way to split up (E Te) is to
use the structured singular value. From (8.92) we have (E Te) (E) (T ) and (10.72) follows. 2
Theorem 10.4 With assumptions (a) and (b) and also assuming that that G and G e have no
RHP-zeros, the overall system is stable (S is stable):
(i) if and only if (I ES S(s))
e 1
is stable, where ES = (G G)G
e 1 ,
(ii) if and only if det(I ES S)
e does not encircle the origin as s traverses the Nyquist D-
contour,
(iii) if
e
(ES S(j)) < 1, (10.73)
(iv) (and (10.73) is satised) if
e = max |e
(S) si | < 1/(ES ) (10.74)
i
The structured singular value (ES ) is computed with respect to a diagonal structure (of S).
e
Proof: The proof is similar to that of Theorem 10.3. We need to assume no RHP-zeros in order to get
(i). 2
Remark. The -conditions (10.72) and (10.74) for (nominal) stability of the decentralized control
system can be generalized to include robust stability and robust performance; see equations (31a-b)
in Skogestad and Morari (1989).
In both the above Theorems, (i) and (ii) are necessary and sufcient conditions for stability,
whereas the spectral radius condition (iii) is weaker (only sufcient) and the -condition
condition (iv) is even weaker. Nevertheless, the use of is the least conservative way of
splitting up the spectral radius in condition (iii).
Equation (10.72) is easy to satisfy at high frequencies, where generally (Te) 0.

Similarly, (10.74) is usually easy to satisfy at low frequencies since
(S(0)) = 0 for systems
e
with integral control (no steady-state offset). Unfortunately, the two conditions cannot be
combined over different frequency ranges (Skogestad and Morari, 1989). Thus, to guarantee
stability we need to satisfy one of the conditions over the whole frequency range.
Since (10.72) is generally most difcult to satisfy at low frequencies, where usually
(Te) 1, this gives rise to the following pairing rule:
Prefer pairings with (E) < 1 (diagonal dominance) at frequencies within the closed-
loop bandwidth.
Let denote the RGA of G. For an n n plant ii (0) > 0.5 i is a necessary condition
for (E(0)) < 1 (diagonal dominance at steady state) (Kariwala et al., 2003). This gives the
following pairing rule: Prefer pairing on steady-state RGA elements larger than 0.5 (because
otherwise we can never have (E(0)) < 1).
Since (10.74) is generally most difcult to satisfy at high frequencies where e 1, and
( S)
since encirclement of the origin of det(I ES S(s))
e is most likely to occur at frequencies up
to crossover, this gives rise to the following pairing rule:
Prefer pairings with (ES ) < 1 (diagonal dominance) at crossover frequencies.
Gershgorin bounds. An alternative to splitting up (E Te) using , is to use Gershgorins

theorem, see page 519. From (10.71) we may then derive (Rosenbrock, 1974) sufcient
conditions for overall stability, either in terms of the rows of G,
X
|e
ti | < |gii |/ |gij | i, (10.75)
j6=i
or, alternatively, in terms of the columns,

X
|e
ti | < |gii |/ |gji | i, (10.76)
j6=i
This gives the important insight that it is preferable P to pair on Plarge elements in G,
because then the sum of the off-diagonal elements, j6=i |gij | and j6=i |gji |, is small. The
Gershgorin bounds, which should be small, are the inverse of the right hand sides in (10.75)
and (10.76),
The Gershgorin conditions (10.75) and (10.76), are complementary to the -condition in
(10.72). Thus, the use of (10.72) is not always better (less conservative) than (10.75) and
(10.76). It is true that the smallest of the i = 1, . . . m upper bounds in (10.75) or (10.76)
is always smaller (more restrictive) than 1/(E) in (10.72). However, (10.72) imposes the
same bound on |e ti | for each loop, whereas (10.75) and (10.76) give individual bounds, some
of which may be less restrictive than 1/(E).
Diagonal dominance. Although diagonal dominance is a matrix property, its denition
has been motivated by control, where, loosely speaking, diagonal dominance means that the
interactions will not introduce instability. Originally, for example in the Inverse Nyquist Array
method of Rosenbrock (1974), diagonal dominance was dened in terms of the Gershgorin
bounds, resulting in the conditions kEki1 < 1 (column dominance) and kEki < 1
(row dominance), where E = (G G) e Ge 1 . However, stability is scaling independent,
and by optimally scaling the plant using DGD 1 , where the scaling matrix D is diagonal,
one obtains from these conditions that the matrix G is (generalized) diagonally dominant
if (|E|) < 1; see (A.128). Here (|E|) is the Perron root of E. An even less restrictive
denition of diagonal dominance is obtained by starting from the stability condition in terms
of (E) in (10.72). This leads us to propose the improved denition below.
Denition 10.1 A matrix G is generalized diagonally dominant if and only if (E) < 1.
Here the term generalized diagonally dominant means can be scaled to be diagonally
dominant. Note that we always have (E) (|E|), so the use of is less restrictive than
the Perron root. Also note that (E) = 0 for a triangular plant.5 It is also possible to use
(Es ) as measure of diagonal dominance, and we then have that a matrix is generalized
diagonally dominant if (E) < 1 or if (ES ) < 1.
Example 10.17 Consider the following plant where we pair on its diagonal elements:
" # " # " #
5 1 2 5 0 0 0 0.5 0.33
G= 4 2 1 ; e=
G 0 2 0 ; e G
E = (G G) e 1 = 0.8 0 0.167
3 2 6 0 0 6 0.6 1 0
The -interaction measure is (E) = 0.9189, so the plant is diagonally dominant. From (10.72),
stability of the individual loops eti guarantees stability of the overall closed-loop system, provided
we keep the individual peaks of |e ti | less than 1/(E) = 1.08. This allows for integral control with
t(0) = 1. Note that it is not possible in this case to conclude from the Gershgorin bounds in (10.75)
e
and (10.76) that the plant is diagonally dominant, because the 2, 2 element of G (= 2) is smaller than
both the sum of the off-diagonal elements in row 2 (= 5) and in column 2 (= 3).
Iterative RGA. An iterative computation of the RGA, k (G), gives a permuted identity
matrix that corresponds to the (permuted) generalized diagonal dominant pairing, if it exists
(Johnson and Shapiro, 1986, Theorem 2) (see also page 88). Note that the iterative RGA
avoids the combinatorial problem of testing all pairings, as is required when computing (E)
or the RGA number. Thus, we may use the iterative RGA to nd a promising pairing, and
check for diagonal dominance using (E).
Exercise 10.15 For the plant in Example 10.17 check that the iterative RGA converges to the
diagonally dominant pairings.
Example 10.18 RGA number. The RGA number, k Iksum , is commonly used as a measure
of diagonal dominance, but unfortunately for 4 4 plants or larger, a small RGA number does not
guarantee diagonal dominance. To illustrate this, consider the matrix G = [1 1 0 0; 0 0.1 1
1; 1 1 0.1 0; 0 0 1 1]. It has has RGA= I, but (E) = (ES ) = 10.9 so it is far from
diagonally dominant.
Triangular plants. Overall stability is trivially satised for a triangular plant as described
in the theorem below.
Theorem 10.5 Suppose the plant G(s) is stable and upper or lower triangular (at all
frequencies), and is controlled by a diagonal controller. Then the overall system is stable
if and only if the individual loops are stable.
A triangular
plant mayhave large off-diagonal elements, but it can be scaled to be diagonal. For example
5

d1 0 g11 0 1/d1 0 g11 0 g11 0
= d2 which approaches for |d1 | |d2 |.
0 d2 g21 g22 0 1/d2 g
d1 12
g22 0 g22
Proof: For a triangular plant G, E = (G G) e Ge 1 is triangular with all diagonal elements zero, so it
follows that all eigenvalues of E Te are zero. Thus det(I + E Te(s)) = 1 and from (ii) in Theorem 10.3
the interactions can not cause instability. 2
Because of interactions, there may not exists pairings such that the plant is triangular at
low frequencies. Fortunately, in practice it is sufcient for stability that the plant is triangular
at crossover frequencies, and we have:
Triangular pairing rule. To achieve stability with decentralized control,
prefer pairings such that at frequencies around crossover, the rearranged
plant matrix G(j) (with the paired elements along the diagonal) is close to
triangular.
Derivation of triangular pairing rule. The derivation is based on Theorem 10.4. From the spectral
radius stability condition in (10.74) the overall system is stable if (SE e S (j)) < 1, . At
low frequencies, this condition is usually satised because Se is small. At higher frequencies, where
Se = diag{e si } I, (10.74) may be satised if G(j) is close to triangular. This is because E S
and thus SE
e S are then close to triangular, with diagonal elements close to zero, so the eigenvalues of
e S (j) are close to zero. Thus (10.74) is satised and we have stability of S. The use of Theorem 10.4
SE
assumes that G and G e have no RHP-zeros, but in practice the result also holds for plants with RHP-zeros
provided they are located beyond the crossover frequency range. 2
Remark. Triangular plant, RGA= I and stability. An important RGA-property is that the RGA
of a triangular plant is always the identity matrix ( = I) or equivalently the RGA number is zero;
see property 4 on page 527. In the rst edition of this book (Skogestad and Postlethwaite, 1996), we
incorrectly claimed that the reverse is also true; that is, an identity RGA matrix ((G) = I) implies
that G is triangular. Then, in the rst printing of the second edition we incorrectly claimed that it holds
for 3 3 systems or smaller, but actually it holds only for 2 2 systems or smaller as illustrated by the
following 3 3 counterexample (due to Vinay Kariwala):

g11 0 0
G = g21 g22 g23 (10.77)
g31 0 g33
has RGA= I in all cases (for any nonzero value of the indicated entries g ij ), but G is not triangular.
On the other hand, note that this G is diagonally dominant since (E) = 0 in all cases. However, more
generally RGA= I does not imply diagonal dominance as illustrated by the following 4 4 matrix 6

1 1 0 0
0 1 1
G= 1 1 0
(10.78)
0 0 1 1
which has RGA= I for any nonzero value of and , but G is not triangular and not always diagonal
dominant. For example, (E) = 3.26 (not diagonally dominant) for = = 0.4. Also, for this plant
stability of the individual loops does not necessarily give overall stability. For example, Te = s+1 1
I
(stable individual loops) gives instability (T unstable) with = when || = || < 0.4. Therefore,
RGA= I and stable individual loops do not generally guarantee overall stability (it is not a sufcient
stability condition). Nevertheless, it is clear that we would prefer to have RGA= I, because otherwise
the plant cannot be triangular. Thus, from the triangular pairing rule we have that it is desirable to select
pairings such that the RGA is close to the identity matrix in the crossover region.
6 (10.78) is a generalization of a counterexample given by Johnson and Shapiro (1986). On our books home page a
physical mixing process is given with a transfer function of this form.
10.6.5 Integrity and negative RGA elements

A desirable property of a decentralized control system is that it has integrity, that is, the
closed-loop system should remain stable as subsystem controllers are brought in and out of
service or when inputs saturate. Mathematically, the system possesses integrity if it remains
stable when the controller K is replaced by EK where E = diag{i } and i may take on the
values of i = 0 or i = 1.
An even stronger requirement (complete detunability) is when it is required that the
system remains stable as the gain in various loops is reduced (detuned) by an arbitrary factor,
i.e. i may take any value between 0 and 1, 0 i 1. Decentralized integral controllability
(DIC) is concerned with whether complete detunability is possible with integral control.
Denition 10.2 Decentralized integral controllability (DIC). The plant G(s) (corre-
sponding to a given pairing with the paired elements along its diagonal) is DIC if there
exists a stabilizing decentralized controller with integral action in each loop such that each
individual loop may be detuned independently by a factor i (0 i 1) without introducing
instability.
Note that DIC considers the existence of a controller, so it depends only on the plant G and
the chosen pairings. The steady-state RGA provides a very useful tool to test for DIC, as is
clear from the following result which was rst proved by Grosdidier et al. (1985).
Theorem 10.6 Steady-state RGA and DIC. Consider a stable square plant G and a
diagonal controller K with integral action in all elements, and assume that the loop transfer
function GK is strictly proper. If a pairing of outputs and manipulated inputs corresponds
to a negative steady-state relative gain, then the closed-loop system has at least one of the
following properties:
(a) The overall closed-loop system is unstable.
(b) The loop with the negative relative gain is unstable by itself.
(c) The closed-loop system is unstable if the loop with the negative relative gain is opened
(broken).
This can be summarized as follows:
A stable (reordered) plant G(s) is DIC only if ii (0) 0 for all i. (10.79)
Proof: Use Theorem 6.7 on page 252 and select G0 = diag{gii , Gii }. Since det G0 = gii det Gii and
det Gii
from (A.78) ii = giidet G
we have det G0 / det G = ii and Theorem 10.6 follows. 2
Each of the three possible instabilities in Theorem 10.6 resulting from pairing on a negative
value of ij (0) is undesirable. The worst case is (a) when the overall system is unstable,
but situation (c) is also highly undesirable as it will imply instability if the loop with the
negative relative gain somehow becomes inactive, e.g. due to input saturation. Situation (b)
is unacceptable if the loop in question is intended to be operated by itself, or if all the other
loops may become inactive, e.g. due to input saturation.
The RGA is a very efcient tool because it does not have to be recomputed for each
possible choice of pairing. This follows since any permutation of the rows and columns of
G results in the same permutation in the RGA of G. To achieve DIC one has to pair on a
positive RGA(0) element in each row and column, and therefore one can often eliminate
many candidate pairings by a simple glance at the RGA matrix. This is illustrated by the
following examples:
Example 10.19 Consider a 3 3 plant with

" # " #
10.2 5.6 1.4 0.96 1.45 1.41
G(0) = 15.5 8.4 0.7 and (0) = 0.94 0.37 0.43 (10.80)
18.1 0.4 1.8 0.90 0.07 1.98
For a 3 3 plant there are six possible pairings, but from the steady-state RGA we see that there is only
one positive element in column 2 (12 = 1.45), and only one positive element in row 3 (33 = 1.98),
and therefore there is only one possible pairing with all RGA elements positive (u 1 y2 , u2 y1 ,
u3 y3 ). Thus, if we require to pair on the positive RGA elements, we can from a quick glance at the
steady-state RGA eliminate ve of the six pairings.
Example 10.20 Consider the following plant and RGA:

" # " #
0.5 0.5 0.004 1.56 2.19 4.75
G(0) = 1 2 0.01 and (0) = 3.12 4.75 6.88 (10.81)
30 250 1 0.56 1.56 3.12
From the RGA, we see that it is impossible to rearrange the plant such that all diagonal RGA elements
are positive. Consequently, this plant is not DIC for any choice of pairings.
Example 10.21 Consider the following plant and RGA:

" # " #
(s + 1) 1 4.19 25.96 1 5 5
G(s) = 2
6.19 1 25.96 and (G) = 5 1 5
(5s + 1) 1 1 1 5 5 1
Note that the RGA is constant, independent of frequency. Only two of the six possible pairings give
positive steady-state RGA elements (see pairing rule 2 on page 450): (a) the (diagonal) pairing on all
ii = 1 and (b) the pairing on all ii = 5. Intuitively, one may expect pairing (a) to be the best since
it corresponds to pairing on RGA elements equal to 1. However, the RGA matrix is far from identity,
and the RGA number, k Iksum , is 30 for both pairings. Also, none of the pairings are diagonally
dominant as (E) = 8.84 for pairing (a) and (E) = 1.25 for the pairing (b). These are larger than 1,
so none of the two alternatives satisfy pairing rule 1 discussed on page 450, and we are led to conclude
that decentralized control should not be used for this plant.
Hovd and Skogestad (1992) conrm this conclusion by designing PI controllers for the two cases.
They found pairing (a) corresponding to ii = 1 to be signicantly worse than (b) with ii = 5, in
agreement with the values for (E). They also found the achievable closed-loop time constants to be
1160 and 220, respectively, which in both cases is very slow compared to the RHP-zero which has a
time constant of 1.
Exercise 10.16 Use the method of iterative RGA (page 88) on the model in Example 10.21, and
conrm that it results in recommending the pairing on ii = 5, which indeed was found to be the
best choice based on (E) and the simulations. (This is partly good luck, because the proven theoretical
result for iterative RGA only holds for a generalized diagonally dominant matrix.)
Exercise 10.17 (a) Assume that the 4 4 matrix in (A.83) represents the steady-state model of a
plant. Show that 20 of the 24 possible pairings can be eliminated by requiring DIC. (b) Consider the
3 3 FCC process in Exercise 6.17 on page 257. Show that ve of the six possible pairings can be
eliminated by requiring DIC.
Remarks on DIC and RGA.

1. DIC was introduced by Skogestad and Morari (1988b) who also give necessary and sufcient
conditions for testing DIC. A detailed survey of conditions for DIC and other related properties
is given by Campo and Morari (1994).
2. DIC is also closely related to D-stability, see papers by Yu and Fan (1990) and Campo and Morari
(1994). The theory of D-stability provides necessary and sufcient conditions (except in a few
special cases, such as when the determinant of one or more of the submatrices is zero).
3. Unstable plants are not DIC. The reason for this is that with all i = 0 we are left with the
uncontrolled plant G, and the system will be (internally) unstable if G(s) is unstable.
4. For i = 0 we assume that the integrator of the corresponding SISO controller has been removed,
otherwise the integrator would yield internal instability.
5. For 2 2 and 3 3 plants we have even tighter RGA conditions for DIC than (10.79). For 2 2
plants (Skogestad and Morari, 1988b)
DIC 11 (0) > 0 (10.82)
For 3 3 plants with positive diagonal RGA elements of G(0) and of Gii (0), i = 1, 2, 3 (its three
principal submatrices), we have (Yu and Fan, 1990)
p p p
DIC 11 (0) + 22 (0) + 33 (0) 1 (10.83)
(Strictly speaking,
p as pointed
p out by Campo
p and Morari (1994), we do not have equivalence for the
case when 11 (0) + 22 (0) + 33 (0) is identically equal to 1, but this has little practical
signicance.)
6. One cannot in general expect tight conditions for DIC in terms of the RGA (i.e. for 4 4 systems
or higher). The reason for this is that the RGA essentially only considers corner values, i = 0
or i = 1, for the detuning factor, that is, it tests for integrity. This is clear from the fact that
det Gii
ii = giidet G
, where G corresponds to i = 1 for all i, gii corresponds to i = 1 with the
other k = 0, and Gii corresponds to i = 0 with the other k = 1. A more complete integrity
(corner-value) result is given next.
7. Determinant condition for integrity (DIC). The following condition is concerned with whether it
is possible to design a decentralized controller for the plant such that the system possesses integrity,
which is a prerequisite for having DIC. Assume without loss of generality that the signs of the
rows or columns of G have been adjusted such that all diagonal elements of G(0) are positive,
i.e. gii (0) 0. Then one may compute the determinant of G(0) and all its principal submatrices
(obtained by deleting rows and corresponding columns in G(0)), which should all have the same
sign for integrity. This determinant condition follows by applying Theorem 6.7 to all possible
combinations of i = 0 or 1 as illustrated in the proof of Theorem 10.6.
8. The Niederlinski index of a matrix G is dened as
NI (G) = det G/i gii (10.84)
A simple way to test the determinant condition for integrity, which is a necessary condition for DIC,
is to require that the Niederlinski index of G(0) and the Niederlinski indices of all the principal
submatrices Gii (0) of G(0) are positive.
The original result of Niederlinski, which involved only testing NI of G(0), obviously yields
less information than the determinant condition as does the use of the sign of the RGA elements.
det Gii
This is because the RGA element is ii = giidet G
, so we may have cases where two negative
determinants result in a positive RGA element. Nevertheless, the RGA is usually the preferred tool
because it does not have to be recomputed for each pairing. Let us rst consider an example where
the Niederlinski index is inconclusive:
" # " #
10 0 20 4.58 0 3.58
G1 (0) = 0.2 1 1 and (G1 (0)) = 1 2.5 2.5
11 12 10 4.58 3.5 2.08
Since one of the diagonal RGA elements is negative, we conclude that this pairing is not DIC.
On the other hand, NI (G1 (0)) = 0.48 (which is positive), so Niederlinskis original condition

10 0
is inconclusive. However, the NI of the three principal submatrices , 10 20
and
0.2 1 11 10

1 1
12 10
are 1, 1.2 and 2.2, and since one of these is negative, the determinant condition correctly
tells us that we do not have DIC.
For this 4 4 example the RGA is inconclusive:

8.72 2.81 2.98 15.80 0.41 0.47 0.06 0.17
G2 (0) =
20.79
(G2 (0)) =
6.54 2.92 2.50 0.20 0.45 0.32 0.44
5.82 0.99 1.48 7.51
and 0.40 0.08 0.17 0.35
7.23 2.92 3.11 7.86 0.39 0.001 0.57 0.04
All the diagonal RGA values are positive, so it is inconclusive when it comes to DIC. However, the
Niederlinski index of the gain matrix is negative, NI (G2 (0)) = 18.65, and we conclude that this
pairing is not DIC (further evaluation of the 3 3 and 2 2 submatrices is not necessary in this
case).
9. The above results, including the requirement that we should pair on positive RGA elements, give
necessary conditions for DIC. If we assume that the controllers have integral action, then T (0) = I,
and we can derive from (10.72) that a sufcient condition for DIC is that G is generalized diagonally
dominant at steady-state, i.e.
(E(0)) < 1
This is proved by Braatz (1993, p. 154). Since the requirement is only sufcient for DIC, it cannot
be used to eliminate designs.
10. If the plant has j-axis poles, e.g. integrators, it is recommended that, prior to the RGA analysis,
these are moved slightly into the LHP (e.g. by using very low-gain feedback). This will have no
practical signicance for the subsequent analysis.
11. Since Theorem 6.7 applies to unstable plants, we may also easily extend Theorem 10.6 to unstable
plants (and in this case one may actually desire to pair on a negative RGA element). This is shown
in Hovd and Skogestad (1994). Alternatively, one may rst implement a stabilizing controller and
then analyze the partially controlled system as if it were the plant G(s).
10.6.6 RHP-zeros and RGA: reasons for avoiding negative RGA

elements with sequential design
So far we have considered decentralized control based on independent design, where we
require that the individual loops are stable and that we do not get instability as loops are
closed or taken out of service. This led to the integrity (DIC) result of avoiding pairing on
negative RGA elements at steady state. However, if we use sequential design, then the inner
loops should not be taken out of service, and one may even end up with loops that are unstable
by themselves (if the inner loops were to be removed). Nevertheless, for sequential design we
nd that it is also generally undesirable to pair on negative RGA elements, and the purpose
of this section is primarily to illustrate this, by using some results that link the RGA and
RHP-zeros.
Bristol (1966) claimed that negative values of ii (0) imply the presence of RHP-zeros,
but did not provide any proof. However, it is indeed true as illustrated by the following two
theorems.
Theorem 10.7 (Hovd and Skogestad, 1992) Consider a transfer function matrix G(s) with
no zeros or poles at s = 0. Assume that lims ij (s) is nite and different from zero. If
ij (j) and ij (0) have different signs then at least one of the following must be true:
(a) The element gij (s) has a RHP-zero.
(b) The overall plant G(s) has a RHP-zero.

(c) The subsystem with input j and output i removed, Gij (s), has a RHP-zero.
Theorem 10.8 (Grosdidier et al., 1985) Consider a stable transfer function matrix G(s)
with elements gij (s). Let gbij (s) denote the closed-loop transfer function between input uj
and output yi with all the other outputs under integral control. Assume that: (i) gij (s) has
no RHP-zeros, (ii) the loop transfer function GK is strictly proper, (iii) all other elements of
G(s) have equal or higher pole excess than gij (s). We then have:
If ij (0) < 0, then for gbij (s) the number of RHP-poles plus RHP-zeros is odd.
Note that gbij (s) in Theorem 10.8 is the same as the transfer function Pu from u1 to y1 for
the partially controlled system in (10.26).
Sequential design and RHP-zeros. We design and implement the diagonal controller by
tuning and closing one loop at a time in a sequential manner. Assume that we end by pairing
on a negative steady-state RGA element, ij (0) < 0, and that the corresponding element
gij (s) has no RHP-zero. Then we have the following implications:
(a) If we have integral action (as we normally have), then we will get a RHP-zero in
gbij (s) which will limit the performance in the nal output y i (follows from Theorem 10.8).
However, the performance limitation is less if the inner loop is tuned sufciently fast (Cui
and Jacobsen, 2002), see also Example 10.22.
(b) If ij () is positive (it is usually close to 1, see pairing rule 1), then irrespective of
integral action, we have a RHP-zero in Gij (s), which will also limit the performance in the
other outputs (follows from Theorem 10.7).
In conclusion, for performance we should avoid ending up by pairing on a negative RGA
element.
Example 10.22 Negative RGA element and RHP-zeros. Consider a plant with

1 4 4
G(s) = (s) = 1 2
s + 10 2 1 2 1
Note that the RGA is independent of frequency for this plant, so 11 (0) = = 1. We want to illustrate
that pairing on negative RGA elements gives performance problems. We start by closing the loop from
u1 to y1 with a controller u1 = k11 (s)(r1 y1 ). For the partially controlled system, the resulting
transfer function from u2 to y2 (outer loop) is
k11 (s)g21 (s)g12 (s)
gb22 (s) = g22 (s)
1 + g11 (s)k11 (s)
With an integral controller k11 (s) = KI /s, we nd, as expected from Theorem 10.8, that
s2 + 10s 4KI
gb22 (s) =
(s + 10)(s2 + 10s + 4KI )
always has a RHP-zero. For large values of KI , the RHP-zero moves further away, and is less limiting
in terms of performance for the outer loop. With a proportional controller, k 11 (s) = Kc , we nd that
s + 10 4Kc
gb22 (s) =
(s + 10)(s + 10 + 4Kc )
has a zero at 4Kc 10. For Kc < 2.5, the zero is in the LHP, but it crosses into the RHP, when
Kc exceeds 2.5. For large values of Kc , the RHP-zero moves further away, and does not limit the
performance in the outer loop in practice. The worst value is Kc = 2.5, where we have a zero at the
origin and the steady-state gain gb22 (0) changes sign.
10.6.7 Performance of decentralized control systems

Consider again the factorization
e I))1 S
S = (I + S( e
in (10.69) where = GG e 1 is the performance relative gain array (PRGA), The diagonal
elements of the PRGA matrix are equal to the diagonal elements of the RGA, ii = ii , and
this is the reason for its name. Note that the off-diagonal elements of the PRGA depend on the
relative scaling on the outputs, whereas the RGA is scaling independent. On the other hand,
the PRGA also measures one-way interaction, whereas the RGA only measures two-way
interaction. At frequencies where feedback is effective (Se 0), (10.69) yields S S
e Thus,
large elements in the PRGA () (compared to 1 in magnitude) mean that the interactions
slow down the overall response and cause performance to be worse than for the individual
loops. On the other hand, small PRGA elements (compared to 1 in magnitude) mean that the
interactions actually improve performance at this frequency.
We will also make use of the related closed-loop disturbance gain (CLDG) matrix, dened
as
Ge d (s) , (s)Gd (s) = G(s)G
e 1
(s)Gd (s) (10.85)
The CLDG depends on both output and disturbance scaling.
In the following, we consider performance in terms of the control error
e = y r = Gu + Gd d r (10.86)
Suppose the system has been scaled as outlined in Section 1.4, such that at each frequency:
1. Each disturbance is less than 1 in magnitude, |dk | < 1.
2. Each reference change is less than the corresponding diagonal element in R, |r j | < Rj .
3. For each output the acceptable control error is less than 1, |ei | < 1.
Single disturbance. Consider a single disturbance, in which case Gd is a vector, and let
gdi denote the ith element of Gd . Let Li = gii ki denote the loop transfer function in loop i.
Consider frequencies where feedback is effective so S e is small (and (10.89) is valid). Then
for acceptable disturbance rejection (|ei | < 1) with decentralized control, we must require
for each loop i,
|1 + Li | > |e
gdi | (10.87)
which is the same as the SISO condition (5.77) except that Gd is replaced by the CLDG, gedi .
In words, gedi gives the apparent disturbance gain as seen from loop i when the system is
controlled using decentralized control.
Single reference change. We can similarly address a change in reference for output j of
magnitude Rj and consider frequencies where feedback is effective (and (10.89) is valid).
Then for acceptable reference tracking (|ei | < 1) we must require for each loop i
|1 + Li | > |ij | |Rj | (10.88)
which is the same as the SISO condition (5.80) except for the PRGA factor, | ij |. In other
words, when the other loops are closed the response in loop i gets slower by a factor
|ii |. Consequently, for performance it is desirable to have small elements in , at least at
frequencies where feedback is effective. However, at frequencies close to crossover, stability
is the main issue, and since the diagonal elements of the PRGA and RGA are equal, we
usually prefer to have ii = ii close to 1 (see pairing rule 1 on page 450).
Proofs of (10.87) and (10.88): At frequencies where feedback is effective, Se is small, so

e I) I
I + S( (10.89)
and from (10.69) we have

e
S S (10.90)
The closed-loop response then becomes
e = SGd d Sr SeG
e d d Sr
e (10.91)
and the response in output i to a single disturbance dk and a single reference change rj is
ei sei gedik dk sei ik rk (10.92)
where sei = 1/(1 + gii ki ) is the sensitivity function for loop i by itself. Thus, to achieve |ei | < 1
for |dk | = 1 we must require |e si gedik | < 1 and (10.87) follows. Similarly, to achieve |ei | < 1 for
|rj | = |Rj | we must require |si ik Rj | < 1 and (10.88) follows. Also note that |si ik | < 1 will imply
that assumption (10.89) is valid. Since R usually has all of its elements larger than 1, in most cases
(10.89) will be automatically satised if (10.88) is satised, so we normally need not check assumption
(10.89). 2
Remark 1 Relation (10.90) may also be derived from (10.66) by assuming Te I which yields
(I + E Te)1 (I + E)1 = .
Remark 2 Consider a particular disturbance with model gd . Its effect on output i with no control is gdi ,
and the ratio between gedi (the CLDG) and gdi is the relative disturbance gain (RDG) (i ) of Stanley
et al. (1985) (see also Skogestad and Morari (1987b)):
e 1 gd ]i /[gd ]i
i , gedi /gdi = [GG (10.93)
Thus i , which is scaling independent, gives the change in the effect of the disturbance caused by
decentralized control. It is desirable to have i small, as this means that the interactions are such that
they reduce the apparent effect of the disturbance, such that one does not need high gains |L i | in the
individual loops.
10.6.8 Summary: pairing selection and controllability analysis for

decentralized control
When considering decentralized diagonal control of a plant, one should rst check that the
plant is controllable with any controller, see Section 6.11.
If the plant is unstable, then it recommended that a lower-layer stabilizing controller is rst
implemented, at least for the fast unstable modes. The pole vectors (page 412) are useful
in selecting which inputs and outputs to use for stabilizing control. Note that some unstable
plants are not stabilizable with a diagonal controller. This happens if the unstable modes
belong to the decentralized xed modes, which are the modes unaffected by diagonal
feedback control (e.g. Lunze (1992)). A simple example is a triangular plant where the
unstable mode appears only in the off-diagonal elements, but here the plant can be stabilized
by changing the pairings.
10.6.9 Independent design

We rst consider the case of independent design, where the controller elements are designed
based on the diagonal (paired) elements of the plant such that individual loops are stable.
The rst step is to determine if one can nd a good set of inputoutput pairs bearing in
mind the following three pairing rules:
Pairing rule 1. RGA at crossover frequencies. Prefer pairings such that the
rearranged system, with the selected pairings along the diagonal, has an RGA
matrix close to identity at frequencies around the closed-loop bandwidth.
To help in identifying the pairing with RGA closest to identity, one may, at the bandwidth
frequency, compute the iterative RGA, k (G); see Exercise 10.6.4 on page 441.
Pairing rule 1 is to ensure that we have diagonal dominance where interactions from other
loops do not cause instability. Actually, pairing rule 1 does not ensure this, see the Remark on
page 442, and to ensure stability we may instead require that the rearranged plant is triangular
at crossover frequencies. However, the RGA is simple and only requires one computation,
and since (a) all triangular plants have RGA = I and (b) there is at most one choice of
pairings with RGA = I at crossover frequencies, we do nothing wrong in terms of missing
good pairing alternatives by following pairing rule 1. To check for diagonal dominance of a
promising pairing (with RGA = I) one may subsequently compute (ES ) = (PRGAI))
to check if it is smaller than 1 at crossover frequencies.
Pairing rule 2. For a stable plant avoid pairings that correspond to negative
steady-state RGA elements, ij (0) < 0.
This rule follows because we require integrity (DIC) with independent design (page 443), and
also because we would like to avoid the introduction of RHP-zeros with sequential design
(page 446).
Remark. Even if we have ii (0) = 1 and ii () = 1 for all i, this does not necessarily mean that
the diagonal pairing is the best, even for a 2 2 plant. The reason for this is that the behaviour at
intermediate bandwidth frequencies is more important. This was illustrated in Example 3.11, where
we found from the frequency-dependent RGA in Figure 3.8 (page 86) that the off-diagonal pairing is
preferable, because it has RGA close to identity at the bandwidth frequencies.
Pairing rule 3. Prefer a pairing ij where gij puts minimal restrictions on the
achievable bandwidth. Specically, the effective delay ij in gij (s) should be
small.
This rule favours pairing on variables physically close to each other, which makes it
easier to use high-gain feedback and satisfy (10.87) and (10.88), while at the same time
achieving stability in each loop. It is also consistent with the desire that (j) is close to I at
crossover frequencies. Pairing rule 3 implies that we should avoid pairing on elements with
high order, a time delay or a RHP-zero, because these result in an increased effective delay;
see page 58. Goodwin et al. (2005) discuss performance limitations of independent design,
in particular when pairing rule 3 is violated.
When a reasonable choice of pairings has been found (if possible), one should rearrange
G to have the paired elements along the diagonal and perform a controllability analysis as
follows.
1. Compute the PRGA ( = GG e 1 ) and CLDG (G e d = Gd ), and plot these as functions

of frequency. For systems with many loops, it is best to perform the analysis one loop at
a time. That is, for each loop i, plot |e
gdik | for each disturbance k and plot |ij | for each
reference j (assuming here for simplicity that each reference is of unit magnitude). For
performance, see (10.88) and( 10.87), we need |1 + Li | to be larger than each of these
Performance : |1 + Li | > max{|e

gdik |, |ij |} (10.94)
k,j
To achieve stability of the individual loops one must analyze gii (s) to ensure that the
bandwidth required by (10.94) is achievable. Note that RHP-zeros in the diagonal elements
may limit achievable decentralized control, whereas they may not pose any problems for
a multivariable controller. Since with decentralized control we usually want to use simple
controllers, the achievable bandwidth in each loop will be limited by the effective delay
ij in gij (s).
2. In general, see rule 5.13 on page 207, one may check for constraints by considering the
elements of G1 Gd and making sure that they do not exceed 1 in magnitude within the
frequency range where control is needed. Equivalently, one may plot |gii | for each loop i,
and the requirement is then
To avoid input constraints : |gii | > |e

gdik |, k (10.95)
at frequencies where |e
gdik | is larger than 1 (this follows since G
e d = GGe 1 Gd ). This
provides a direct generalization of the requirement |G| > |Gd | for SISO systems.
The advantage of (10.95) compared to using G1 Gd is that we can limit ourselves to
frequencies where control is needed to reject the disturbance (where |e
gdik | > 1).
If the plant is not controllable with any choice of pairings, then one may consider another
pairing choice and go back to step 1. Most likely this will not help, and one would need to
consider decentralized sequential design, or multivariable control.
If the chosen pairing is controllable then the analysis based on (10.94) tells us directly how
large the loop gain |Li | = |gii ki | must be, and this can be used as a basis for designing the
controller ki (s) for loop i.
10.6.10 Sequential design

Sequential design may be applied when it is not possible to nd a suitable set of pairings for
independent design using the above three pairing rules. For example, with sequential design
one may choose to pair on an element with gii = 0 (and ii = 0), which violates both
pairing rules 1 and 3. One then relies on the interactions to achieve the desired performance,
as loop i by itself has no effect. This was illustrated for the case with off-diagonal pairings
in Example 10.15 on page 434. Another case with pairing on a zero element is in distillation
control when the LV -conguration is not used, see Example 10.8. One may also in some
cases pair on negative steady-state RGA elements, although we have established that to avoid
introducing RHP-zeros one should avoid closing a loop on a negative steady-state RGA (see
page 447).
The procedure and rules for independent design can be used as a starting point for nding
good pairings for sequential design. With sequential design, one also has to decide the order
in which the loops are closed, and one generally starts by closing the fast loops. This favours
starting with a pairing where gij has good controllability, including a large gain and a small
effective delay. One may also consider the disturbance gain to nd which outputs need to be
tightly controlled. After closing one loop, one needs to obtain the transfer function for the
resulting partially controlled system, see (10.28), and then redo the analysis in order to select
the next pairing, and so on.
Example 10.23 Application to distillation process. In order to demonstrate the use of the
frequency-dependent RGA and CLDG for evaluation of expected diagonal control performance, we
again consider the distillation process used in Example 10.8. The LV -conguration is used; that is, the
manipulated inputs are reux L (u1 ) and boilup V (u2 ). The outputs are the product compositions yD
(y1 ) and xB (y2 ). Disturbances in feed ow rate F (d1 ) and feed composition zF (d2 ) are included in
the model. The disturbances and outputs have been scaled such that a magnitude of 1 corresponds to a
change in F of 20%, a change in zF of 20%, and a change in xB and yD of 0.01 mole fraction units.
The ve state dynamic model is given in Section 13.4.
Initial controllability analysis. G(s) is stable and has no RHP-zeros. The plant and RGA matrix at
steady-state are
87.8 86.4 35.1 34.1
G(0) = 108.2 109.6
(0) = 34.1 35.1
(10.96)
The RGA elements are much larger than 1 and indicate a plant that is fundamentally difcult to control
(recall property C1, page 89). Fortunately, the ow dynamics partially decouple the response at higher
frequencies, and we nd that (j) I at frequencies above about 0.5 rad/min. Therefore if we can
achieve sufciently fast control, the large steady-state RGA elements may be less of a problem.
1
10
PSfrag replacements
Magnitude
0
10
gd22 gd21
gd11 gd12
1
10 3 2 1 0 1
10 10 10 10 10
Frequency [rad/min]
Figure 10.18: Disturbance gains |gdik | for assessing the effect of disturbance k on output i
2
10
ged21
1
ged11 PSfrag replacements
Magnitude
10
ged22
0
10
ged12
1
10 3 2 1 0 1
10 10 10 10 10
Frequency [rad/min]
Figure 10.19: Closed-loop disturbance gains |e

gdik | for assessing the effect of disturbance k on output i
The steady-state effect of the two disturbances is given by

7.88 8.81
Gd (0) = 11.72 11.19
(10.97)
and the magnitudes of the elements in Gd (j) are plotted as functions of frequency in Figure 10.18.
From this plot the two disturbances seem to be equally difcult to reject with magnitudes larger than
1 up to a frequency of about 0.1 rad/min. We conclude that control is needed up to 0.1 rad/min. The
magnitude of the elements in G1 Gd (j) (not shown) are all less than 1 at all frequencies (at least up
to 10 rad/min), and so it will be assumed that input constraints pose no problem.
Choice of pairings. The selection of u1 to control y1 and u2 to control y2 corresponds to pairing on
positive elements of (0) and (j) I at high frequencies. This seems sensible, and is used in the
following.
Analysis of decentralized control. The elements in the CLDG and PRGA matrices are shown as
functions of frequency in Figures 10.19 and 10.20. At steady-state we have

35.1 27.6 e d (0) = (0)Gd (0) = 47.7 0.40
(0) = 43.2 35.1
, G 70.5 11.7
(10.98)
In this particular case, the off-diagonal elements of RGA () and PRGA () are quite similar. We note
that Ge d (0) is very different from Gd (0), and this also holds at higher frequencies. For disturbance 1
(rst column in G e d ) we nd that the interactions increase the apparent effect of the disturbance, whereas
they reduce the effect of disturbance 2, at least on output 1.
2
10
1
Magnitude
10
PSfrag replacements
21 11 = 22
0 12
10
1
10 3 2 1 0 1
10 10 10 10 10
Frequency [rad/min]
Figure 10.20: PRGA elements |ij | for effect of reference j on output i
We now consider one loop at a time to nd the required bandwidth. For loop 1 (output 1) we consider
11 and 12 for references, and ged11 and ged12 for disturbances. Disturbance 1 is the most difcult, and
we need |1+L1 | > |b gd11 | at frequencies where |b
gd11 | is larger than 1, which is up to about 0.2 rad/min.
The magnitudes of the PRGA elements are somewhat smaller than |e gd11 | (at least at low frequencies),
so reference tracking will be achieved if we can reject disturbance 1. From ged12 we see that disturbance
2 has almost no effect on output 1 under feedback control.
Also, for loop 2 we nd that disturbance 1 is the most difcult, and from ged12 we require a loop gain
larger than 1 up to about 0.3 rad/min. A bandwidth of about 0.2 to 0.3 rad/min in each loop is required
for rejecting disturbance 1, and should be achievable in practice.
Observed control performance. To check the validity of the above results we designed two single-
loop PI controllers:
1 + 3.76s 1 + 3.31s
k1 (s) = 0.261 ; k2 (s) = 0.375 (10.99)
3.76s 3.31s
The loop gains, Li = gii ki , with these controllers are larger than the closed-loop disturbance gains,
|ik |, at frequencies up to crossover. Closed-loop simulations with these controllers are shown in Figure
10.21. The simulations conrm that disturbance 2 is more easily rejected than disturbance 1.
0.4
0.2 y2
0
PSfrag replacements
0.2
y1
0.4
0 10 20 30 40 50 60 70 80 90 100
Time [min]
Figure 10.21: Decentralized PI control. Responses to a unit step in d 1 at t = 0 and a unit step in d2 at
t = 50 min.
In summary, there is an excellent agreement between the controllability analysis and the
simulations, as has also been conrmed by a number of other examples.
10.6.11 Conclusions on decentralized control

In this section, we have derived a number of conditions for the stability, e.g. (10.72) and
(10.79), and performance, e.g. (10.87) and (10.88), of decentralized control systems. The
conditions may be useful in determining appropriate pairings of inputs and outputs and
the sequence in which the decentralized controllers should be designed. Recall, however,
that in many practical cases decentralized controllers are tuned off-line, and sometimes
on-line, using local models. In such cases, the conditions may be used in an inputoutput
controllability analysis to determine the viability of decentralized control.
Some exercises which include a controllability analysis of decentralized control are given
at the end of Chapter 6.
10.7 Conclusion
Control structure design is very important in applications, but it has traditionally received
little attention in the control community. In this chapter, we have discussed the issues
involved, and we have provided some results and rules, dos and donts, which we believe
will be helpful in practice. However, there is still a need for improved tools and theory in this
important area.

Rga PDF

Uploaded by

Copyright:

Available Formats

Rga PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Rga PDF

Uploaded by

Copyright:

Available Formats

82 MULTIVARIABLE FEEDBACK CONTROL

(G) = min (D1 GD2 ) (3.53)

and can be computed using (A.74).

1. A large condition number (G) = (G)/(G) may be caused by a small value of

3.4 Relative gain array (RGA)

where denotes element-by-element multiplication (the Hadamard or Schur product). With

3.4.1 Original interpretation: RGA as an interaction measure

and interchange the roles of G and G1 , of u and y, and of i and j to get

Remark. The assumption of yk = 0 (perfect control of yk ) in (3.57) is satised at steady-state

open-loop gain (with u2 = 0) g11 (s) 1

3.4.2 Examples: RGA

transfer matrix then becomes

and the corresponding RGA matrix is (at all frequencies)

|12 | = |21 | Diagonal pairing

(a) Magnitude of RGA elements (b) RGA number

Figure 3.8: Frequency-dependent RGA for G(s) in (3.65)

Table 3.1: Matlab program to calculate frequency-dependent RGA

3.4.3 RGA number and iterative RGA

positive pairing is possible.

3.4.4 Summary of algebraic properties of the RGA

A1. It is independent of input and output scaling.

3.4.5 Summary of control properties of the RGA

Example 3.16 Consider the plant

For a detailed analysis of achievable performance of the plant (inputoutput controllability

3.5 Control of multivariable plants

3.5.2 Two-step compensator design approach

Figure 3.9: One degree-of-freedom feedback control conguration

to multivariable control is given by a two-step procedure in which we rst design a

3.5.4 Pre- and post-compensators and the SVD controller

Figure 3.10: Pre- and post-compensators, W1 and W2 . Ks is diagonal.

plant W2 GW1 . The overall controller is then

W1 = V o and W2 = UoT (3.77)

Figure 10.1: General control conguration

Multivariable Feedback Control: Analysis and Design. Second Edition

cost (self-optimizing control). (Section 10.3)

Bottom-up (steps 5 and 6)

minimize control system complexity subject to the achievement of accuracy

10.2 Optimal operation and control

Objective Objective Objective

(a) (b) (c)

(b) Closed-loop implementation with separate control layer

Figure 10.3: Typical control system hierarchy in a chemical plant

Vertical (hiearchical) decomposition. This is the decomposition just discussed which

10.3 Selection of primary controlled outputs

10.3.1 Self-optimizing control

Self-optimizing control is when we can achieve an acceptable loss with constant

The concept of self-optimizing control is inherent in many real-life scenarios

10.3.2 Selecting controlled outputs: local analysis

u uopt = G1 (z zopt ) (10.4)

z zopt = z r + r zopt = e + eopt (d) (10.7)

If z is directly measured then nz is its measurement error. If z is a combination of several

span(z) = max |z zopt | = max |eopt (d)| + max |e| (10.8)

10.3.3 Selecting controlled outputs: maximum scaled gain method

From (10.6), we have L = 21 ke z k22 where ze = Juu G1 (z zopt ). Introducing thescaled

10.3.4 Selecting controlled outputs: exact local method

10.3.5 Selecting controlled outputs: direct evaluation of cost

10.3.6 Selecting controlled outputs: measurement combinations

where F denotes the optimal sensitivity of y with respect to d. We would like to nd z = Hy

10.3.7 Selecting controlled outputs: examples