You are on page 1of 5

2.

25 Relative Gain Calculations


B. D. CAMPBELL, P. G. FRIEDMANN (1995) F. G. SHINSKEY (2005)

Most processes require multiple control loops to operate effi- This section presents the computation, interpretation, and
ciently, and these loops usually do not act independently. application of relative gain arrays. It also provides examples
Changes made in the output of one controller can be a source that illustrate RGA application to simple processes.
of upset to another loop. That interaction may pass in one
direction only; for example, changing the feed rate to a pro-
cess upsets the composition of the product leaving it. How- THE RELATIVE GAIN ARRAY
ever, there are also processes where the disturbed controller
responds by sending another disturbance back to the first The relative gain for a selected pair of variables is defined
loop. For example, a top-composition controller on a distil- as the ratio of the open-loop gain for that pair with all other
lation column manipulating reflux flow can upset bottom loops open to their open-loop gain when all other loops in
composition, whose controller manipulating heat input reacts the process are closed, with their variables held at set point
by upsetting top composition. This bidirectional interaction by their controllers:
comprises an additional feedback loop passing through both
controllers in series and is a cause of instability to the process. ∂ ci
When two or more manipulated variables affect a single ∂ mj
controlled variable, or when a single manipulated variable λij = m
2.25(1)
∂ ci
affects two or more controlled variables, a question arises of
how best to pair the manipulated and controlled variables ∂ mj
c
together in closed loops. With a two-loop process, only two
possible single-loop structures exist: variable c1 can be con- where λij is the relative gain of any pair of variables ci con-
trolled by manipulating mj. The numerator is the open-loop
trolled by manipulated variable ma and variable c2 controlled
gain determined with all other manipulated variables con-
by mb, or vice versa. If one pairing, say c1 -ma and c2 -mb,
stant, and the denominator is the open-loop gain determined
performs well, the other pairing will not and may even be
with all other controlled variables constant, which literally
unstable. It is also possible that both pairings will perform
applies to a steady state. Therefore the above applies strictly
equally poorly. In some cases, the more-effective pairing may
to steady-state gains.
be obvious, but in others it will not. The number of possible
Because the same dimensions appear in both numerator
single-loop structures increases factorially with the number
and denominator, the relative gain numbers are dimension-
of loops, reaching 120 for a five-loop distillation column.
less. They are also independent of scale and of common terms
The relative gain method of interaction analysis solves this
appearing in both open-loop gains. They are unaffected by
problem, providing a quantitative measure of loop interactions
nonlinearities at any given operating point. If the process is
that is surprisingly easy to use and universally applicable. The
nonlinear, however, the calculated relative gains can be
relative gains calculated are dimensionless numbers arranged
expected to change with the operating conditions.
in an array, which by their values indicate the best single-loop
The RGA is a table of λij values arranged in rows c against
pairings and where decoupling is advisable. The method was
1 columns m:
first proposed by Bristol, and it was then named and applied
2,3 4,5
by Shinskey. Its applications include integrating processes m1 . . mn
and those with unequal numbers of manipulated and controlled
6
variables. It has been extended to include dynamic consider- c1 λ11 . . λ1n
7
ations in the analysis of interactions. Λ= . . . . . 2.25(2)
Much of the Relative Gain Array (RGA) literature is devoted . . . . .
8–10
to distillation, particularly to top and bottom composition cn λn1 . . λnn
control. Relative gain calculation for distillation applications is
covered in Chapter 8 of this book. Other applications include where n is the number of variable pairs in the system. An
power systems, blending, combustion, and heat transfer controls. important property of the RGA is that the numbers in each

318

© 2006 by Béla Lipták


2.25 Relative Gain Calculations 319

of the columns and each of the rows add to 1.0. To fill out a has been configured to be a negative feedback loop while
2 × 2 array, then, only one λ needs to be calculated; the others other loops are open, it will become a positive feedback loop
in the same row and the same column are complementary, when the others are closed. Positive feedback causes the
and the diagonal value is identical. To fill out a n × n array controlled variable to run away from set point rather than
2
then, (n – 1) values must be calculated. An important cor- approach it and so must be avoided in a regulatory system.
ollary to this property is that for every λ that exceeds 1.0, It is possible to reconfigure the loop in question to have
there must be a negative number in the same row and in the negative feedback when the other loops are closed by
same column, and control-loop pairs with negative numbers, reversing the action of one controller. But its dynamic
as described below, must be avoided. response will be poor because it contains additional process
elements, along with other controllers. However, the prin-
Properties of  cipal danger is that the opening of another loop, either by
placing its controller in manual or by its output reaching a
The values of λ cover all possible numbers from –infinity limit, will leave the reconfigured loop as positive feedback.
to + infinity. They can be separated into six distinct regimes, Therefore, loops with negative relative gains should never
each with its own set of properties. be connected.

 = 1.0 The most desirable value of λ is 1.0, which indicates 0 <  < 1.0 Relative gains in the zero to one range indicate
that numerator and denominator of Equation 2.25(1) are equal. the existence of an additional negative feedback loop formed
This means that the closing of other loops has no effect on the by interaction. If λ = 0.5, that loop is as strong as the intended
behavior of the loop being examined. So in evaluating an RGA, single loop, and lower values indicate it is even stronger. To
it is advisable to connect those loops having relative gains close return the intended loop to the same stability it had when
to 1.0. This is not an absolute guarantee of best performance operating alone, the proportional gain of the controller would
because it does not include dynamic considerations, but it does have to be multiplied by λ. The additional feedback loop
guarantee steady-state stability. One-way interaction could still contains more dynamic elements as well, requiring an
exist: if either mj does not upset other controlled variables, or increase in integral time. Two-by-two systems where all the
ci is not affected by other manipulated variables, the relative relative gains are 0.5 are common; an example is control of
gain will be 1.0. But without two-way interaction, no hidden boiler pressure and flue-gas oxygen by fuel and air flows.
feedback loops exist, and hence the stability of the loop in Since neither configuration of single loops is effective, decou-
question is not affected by the closure of other loops. pling is required. In the case of the boiler, flue-gas oxygen
content is controlled by the fuel-to-air ratio.
 = 0 If λ is zero, it is because the numerator in Equation
2.25(1) is zero, but the denominator is not. In this case, the  > 1.0 Relative gains > 1.0 indicate the existence of an
loop in question cannot be closed successfully alone, but additional positive feedback loop formed by interaction. As
could possibly control quite well when the other loops in the that loop increases in strength, λ also increases. (When the
system are closed. This is usually the case when one of the positive feedback loop dominates, λ is negative.) The pos-
other controlled variables is liquid level. A continuous pro- itive feedback loop has 180 degrees less phase shift than
cess must have its level loops closed before any of the others the negative feedback loop and therefore does not signifi-
can function properly. In a complex multivariable process cantly slow the dynamic response of the negative feedback
such as distillation, it is usually necessary to assign the level loop. In the case of two interacting loops having similar
loops first and then evaluate the relative gains of the compo- dynamics and λ = 10 or more, both controller gains need
sition loops. This eliminates zeros from the RGA. This pro- to be halved when both loops are closed, but integral and
cedure is described in detail in Section 8.20. derivative settings may be left as they were for single-loop
operation.
 = ± As the denominator in Equation 2.25(1)
approaches zero, λ will approach infinity, and a plus-infinity
anywhere in the array will require a minus-infinity in the same CALCULATION METHODS
row and in the same column. These loops cannot be controlled
independently because the variables have essentially the same The RGA can be evaluated in a number of different ways,
effects on each other. While calculating a result of infinity is with the choice depending on the availability of information
not always achievable, any relative gain whose absolute value about the process to be controlled.
exceeds 100 qualifies — the variables are essentially dependent
and the loops should not be closed. The Ratio of Partial Derivatives

 < 0 Negative relative gains indicate the presence of con- This method solves Equation 2.25(1) directly for each of the
2
ditional stability because the numerator and denominator in needed (n – 1) elements in the array. The numerator is
Equation 2.25(1) have opposite signs. If the loop in question obtained by a partial derivative of ci with respect to mj in

© 2006 by Béla Lipták


320 Control Theory

AIC applies generally to a class of processes, e.g., two-component


FIC
blending, where it can be used again.
F X
F1 Two Loops Open
AT
Process models are often formulated as a set of linear equa-
FT tions with gains K:
F2
c1 = K11m1 + K12m2
2.25(9)
c2 = K 21m1 + K 22m2
FIG. 2.25a
A simple two-component blender has two interacting loops. If the process is limited to two pairs of variables, a simple
formula applies:
terms of the other manipulated variables held constant. As 1
an example, consider the blending of two streams into a λ11 = 2.25(10)
1 − K 21K12 /K11K 22
single product of composition x flowing at rate F, as shown
in Figure 2.25a. The total flow is the sum of the two manip- This formula gives some important insights. If there are
ulated flows: an odd number of negative signs among the gains, λ11 will
fall between zero and 1.0. Otherwise, λ11 will fall outside of
F = F1 + F2 2.25(3) that range and be negative if the product of the off-diagonal
gains exceeds that of the diagonal gains — dominant positive
where the individual flows are expressed as linear with con- feedback.
troller output. (Relative gain is independent of nonlinearities
and common factors, allowing such simplifications.) Then, Two Loops Closed

∂F Some process models solve for values of the manipulated


= 1.0 2.25(4)
∂ F1 m variables that will produce designated values of the controlled
variables:
To formulate the denominator, the partial differentiation
has to be repeated in terms of the other controlled variable, m1 = H11c1 + H12c2
2.25(11)
in this example, x. Let stream 2 contain none of the measured m2 = H 21c1 + H 22c2
ingredient, and stream 1 have a concentration of xmax. Then
the concentration in the blend will be Again, for two pairs of variables, we have the same
formula:
x max F1
x= 2.25(5) 1
F λ11 = 2.25(12)
1 − H 21H12 /H11H 22
Now the partial may be evaluated in terms of x:
The Two-Slope Method
∂F x
= max 2.25(6) Open-loop testing can provide still another way to calculate
∂ F1 x x
relative gains in a two-loop system. From a steady state, m1
can be stepped with m2 constant, and the resulting changes
Dividing the first partial by the second gives the relative
in the two controlled variables divided. The ratio of their
gain:
changes (∆c1/∆c2)m2 is also the ratio K11/K21. Then m2 is
stepped with m1 constant, with these changes divided,
λF1 = x/ x max = x ′ 2.25(7) (∆c1/∆c2)m1 being the same as the ratio K12/K22. Substitution
of those ratios into Equation 2.25(10) gives
The RGA then looks like this:
1
F1 F2 λ11 = 2.25(13)
1 − (∂ c1/∂ c2 )m1/(∂ c1/∂ c2 )m 2
Λ=F x′ 1 − x′ 2.25(8)
x 1 − x′ x′ This formula applies particularly well to distillation com-
position loops, where the product compositions can be plotted
This method gives the advantage of a solution in terms against one another while various manipulated variables are
of the variables themselves, in this case x. The solution then held constant. This is the method used in Section 8.20.

© 2006 by Béla Lipták


2.25 Relative Gain Calculations 321

The Matrix Method Reducing the Array

Equations 2.25(9) and 2.25(11) are sets of linear simultaneous Arrays larger than 3 × 3 may present misleading information,
equations that can be expanded to any number of variables. principally due to dynamic effects. In general, fast loops can
When the number exceeds two pair, however, it is easier to upset slow loops, but slow loops cannot upset fast ones. In
express the relationships using matrix algebra. Accordingly, this light, the pairing of the fast loops in a system can be
Equation 2.25(9) could be expressed as arbitrarily assigned, reducing the array to include only the
slower loops. This is the approach used to evaluate relative gains
c = Km 2.25(14) for distillation in Section 8.20. A typical column has five
loops: two compositions, two levels, and one pressure. Arbi-
where c is a vector of n controlled variables, m is a vector
trary assignment of the level and pressure loops reduces the
of n manipulated variables, and K is the n × n matrix of open-
RGA to the two composition loops. However, there are several
loop gains relating them. Similarly, Equation 2.25(11) could
of these 2 × 2 RGAs to be evaluated, one for each assignment
be expressed as
of the level and pressure loops. Each of these RGAs needs to
m = Hc 2.25(15) be evaluated to determine which has the most favorable rela-
tive gains.
Matrix H is the inverse of matrix K (but its elements are
not the inverse of the elements in the K matrix):
DECOUPLED RELATIVE GAINS
−1
H=K 2.25(16)
Relative gain analysis is not limited to evaluating interactions
If one matrix is available, the other may be obtained by among single loops. Controller outputs can drive mathemat-
inverting it using standard matrix inversion procedures. Then ical combinations of manipulated variables, such as their
each relative gain element λij may be found by multiplying sums, differences, or ratios. Controller inputs can also be
the element Κij by the corresponding element Hji: mathematical combinations of controlled variables. Selection
of these functions should be logical and meaningful to those
λij = Kij ⋅ H ji 2.25(17) who operate the loops.

The subscripts are reversed because Hji relates ci and mj, as Partial Decoupling
does Kij.
As an example, return to the blending system of Figure 2.25a.
Where a zero appears in the K matrix, one will also
The composition of the blend is obviously determined by the
appear in the RGA. Other than that observation, the calcula-
ratio of the two feed flows, F1/F2. The output of the compo-
tions need to be done to determine the RGA. A numerical
sition controller then ought to manipulate that ratio, rather
example follows. Given that
than one of the individual flows. This is done by inserting a
multiplier [×] in the output of the composition controller, as
 2.2 0.80 0.40  shown in Figure 2.25b. The other input to the multiplier is
 
K = 1.2 1.6 0 
 0 0.70 0.50 
 X
F1 FY
Then
mx = F1
F2 F2

 0.495 −0.074 −0.396  FIC AIC


 
H =  −0.371 0.681 0.297 
 0.520 −0.9953 1.584  F X

F1
AT
and

FT
 1.089 −0.297 0.208 
 
Λ =  −0.089 1.089 0  F2
 0.792 
 0 0.208
FIG. 2.25b
The best pairing appears to be along the diagonal—not sur- Partial decoupling is enough to drive relative gains to one and zero
prising, given the two off-diagonal zeros. and protect the costly variable against upset.

© 2006 by Béla Lipták


322 Control Theory

11
F2, and the output of the multiplier is the driven value of F1. gerous feedback loop and complicates operation of the
Calculating F1 from F2 in this way means that the controller system.
output mx = F1/F2. It will be observed that a partial decoupler is imple-
To evaluate the effectiveness of this configuration, it is mented in the same way as a feedforward loop. There is no
only necessary to calculate one gain. Beginning with Equa- real distinction between them. Feedforward uses an indepen-
tion 2.25(5), we substitute for F1 —which is no longer an dent load variable and the output of a feedback controller to
independent manipulated variable—the product mxF2: calculate its manipulated variable, whereas a decoupler uses
another manipulated variable in place of the load. Both pro-
F1 mx F2 mx tect the controlled variable from disturbances, but the decou-
x = x max = x max = x max
F1 + F2 mx F2 + F2 mx + 1 pler also breaks a hidden feedback loop.
2.25(18)
As can be seen, the second manipulated variable F2 drops References
out of the equation, leaving composition x unaffected by it:
1. Bristol, E. H., “On a New Measure of Interaction for Multivariable
∂x Process Control,” IEEE Transactions on Automatic Control, January
=0 2.25(19) 1966, pp. 133–134.
∂ F2 mx 2. Shinskey, F. G., Process Control Systems, New York: McGraw-Hill,
1967, pp. 189–198.
With its numerator at zero, λx2 = 0. Then the RGA 3. Shinskey, F. G., Process Control Systems, 4th ed., New York:
becomes McGraw-Hill, 1996, pp. 264–278.
4. Wolverton, P., “How to Use Relative Gain Analysis in Systems with
mx F2 Integrating Variables,” InTech, September 1980, pp. 63–65.
5. Arkun, Y., and Downs, J., “A General Method to Calculate Input/
Λ=F 0 1 2.25(20) Output Gains and the Relative Gain Array for Integrating Processes,”
x 1 0 Computers Chem. Engng., 10, 1990, pp. 1101–1110.
6. Chang, J.-W., and Yu, C.-C., “The Relative Gain for Non-Square
The system is now decoupled. Multivariable Systems,” Chemical Engineering Science, 5, 1990,
However, the configuration in Figure 2.25b is only a pp. 1309–1323.
7. McAvoy, T. J., Interaction Analysis, Research Triangle Park, NC: ISA,
partial decoupler. While composition is no longer upset by 1983.
changes in the output of the flow controller, total flow is 8. Shinskey, F. G., Distillation Control, 2nd ed., New York: McGraw-
changed every time the composition controller moves its Hill, 1984.
output. This is not usually considered a problem because the 9. Skogestad, S., Lundstrom, P., and Jacobsen, E. W., “Selecting the Best
flow loop is so much faster than the composition loop. The Distillation Control Configuration,” AIChE, May 1990, pp. 753–764.
output of the composition controller will tend to move so 10. Shinskey, F. G., Process Control Systems, 4th ed., New York:
McGraw-Hill, 1996, pp. 347–350.
slowly that any upset in total flow will be small and easily
11. Shinskey, ibid., pp. 267–270.
corrected by the faster flow controller.
Partial decouplers are usually preferred to full decouplers in
that they accomplish the principal objectives of eliminating the Bibliography
third feedback loop — which created the stability problem — and
of protecting the slower and more costly variable from upsets. McAvoy, T. J., Interaction Analysis, Research Triangle Park, NC: ISA, 1983.
A full decoupler, which would only provide the (unnecessary) Perron, S., “The New Generation Language for Controls Design,” Control
protection of the faster variable, also adds a potentially dan- Engineering, February 1992.

© 2006 by Béla Lipták

You might also like