c Goodwin, Graebe, Salgado , Prentice Hall 2000




 apter 11




 


 c Goodwin, Graebe, Salgado , Prentice Hall 2000



opics to be covered
An ubiquitous problem in control is t at all real actuators
ave limited aut ority.  is implies t at t ey are
constrained in amplitude and/or rate of c ange. If one
ignores t is possibility t en serious degradation in
performance can result in t e event t at t e input reac es
a constraint limit.  is is clearly a very important
problem.  ere are two ways of dealing wit it:
(i) reduce t e performance demands so t at a linear controller
never violates t e limits, or
(ii) modify t e design to account for t e limit.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Here we give a first treatment of option (ii) based on

modifying a given linear design.  is will usually
work satisfactorily for modest violations of t e
constraint (up to say 100%). If more serious
violations of t e constraints occur t en we would
argue t at t e actuator as been undersized for t e
given application.
We will also s ow ow t e same ideas can be used
to avoid simple kinds of state constraints.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Wind-Up
One very common consequence of an input itting a
saturation limit is t at t e integrator in t e controller
(assuming it as one) will continue to integrate
w ilst t e input is constrained.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000




  Consider the following nominal plant

model:

Say that the target complementary sensitivity is

It is readily seen that this is achieved with the following

controller.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



J unit step reference is applied at t=1 and a

negative unit step output disturbance occurs at
t=10. The plant input saturates when it is outside
the range [-3, 3]. The plant output y(t) is shown in
Figure 11.1.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.1: ·oop performance with (think line) and without

(thin line) saturation at the plant input.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



£e observe from Figure 11.1 that the plant output

exhibits undesirable transient behavior which is
inconsistent with the linear nominal bandwidth of
approximately 10[rad/s]. This deficiency originates
from the saturation, since a unit step in the reference
produces an instantaneous demanded change of 50
in the controller output and hence saturation occurs,
which a linear design procedure for C(s) does not
take into account.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Anti-Windup Sc eme
 ere are many alternative ways of ac ieving
protection against wind-up. All of t ese met ods rely
on making sure t at t e states of t e controller ave
two key properties; namely
(i) t e state of t e controller s ould be driven by t e actual
(i.e. constrained) plant input;
(ii) t e states of t e controller s ould ave a stable realization
w en driven by t e actual plant input.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 is is particularly easy to ac ieve w en t e

controller is biproper and minimum p ase. Say t at
t e controller as transfer function C(s), t en we
split t is into t e direct feedt roug term C and a
strictly proper transfer function C(s); i.e.

 en consider t e feedback loop s own in  igure

11.2.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.2: Feedback form of biproper controller

 e transfer function from e(t) to u(t) in  igure

11.2 is readily seen to be
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



We next redraw  igure 11.2 as in  igure 11.3.

Figure 11.3: Desired and actual plant input

In t e case of a limited input, all we now need to do

is to ensure t at t e correct relations ip is ac ieved
between t e desired and actual input.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Saturation
 e appropriate function to describe input saturation is:

W ere uÖ(t) is t e unconstrained controller output

and u(t) is t e effective plant input.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Slew Rate Limit

Similarly, we can describe a limit on t e rate of
c ange of t e input (called slew rate) as follows:

A block diagram realization of a slew rate limiter

is s own on t e next slide.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.4: Slew rate limit model

 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.5: Combined saturation and slew rate limit model

A slew rate limiter can be combined wit a saturation

constraint as follows:
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.6: Simplified anti wind-up control loop (C form)

Referring back to  igure 11.3, we can realize an

anti-windup compensated controller by placing t e
appropriate limiter into t e block diagram connecting
t e desired input to t e actual (or allowed) input.  is
leads to t e feedback loop s own below:
 c Goodwin, Graebe, Salgado , Prentice Hall 2000




  Consider the same plant as in

Example 11.1 with identical reference and
disturbance conditions. However, this time we
implement the control loop using anti-windup
protection.
On running a simulation, the results are shown in
Figure 11.7, where the plant output has been plotted.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000


 igure 11.7: ·oop performance with anti wind-up controller
(thick line) compared to performance achieved
with no anti wind-up feature (thin line). The
latter corresponds to the thick line in Figure 11.1
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



A second example aving slew

rate limits is described below.
Consider a plant having a linear model given by

Jssume that a PI controller with KP=0.5 and Tr=1.5[s],

has been tuned for the linear operating range of this
model, i.e., ignoring any nonlinear actuator dynamics.
If the input u(t) cannot change at a rate faster than
0.2[s-1], verify that implementation of the controller as in
Figure 11.6 provides better performance than ignoring
the slew rate limitation.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Solution
We build a control loop wit t e controller structure
s own in  igure 11.6 (see t e next slide) wit Lim
replaced by t e slew rate limiter in  igure 11.4.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.6: Simplified anti wind-up control loop (C form)

 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.8: Performance of PI control loop when no slew

rate limitation exists (1), with slew rate
limitation but no compensation (2) and with anti
wind-up for slew rate limitation (3)
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Interpretation in terms of
onditioning
Here we ask t e following question: W at conditioned
set-point r would ave avoided producing an input uÖ
beyond t e limits of saturation in t e first place?
We assume t at C(s) is biproper and can ence be
expanded in terms of its strictly proper and feed-t roug
terms as

Let us assume t at we ave avoided saturation up to t is

point in time by c anging e(t) to e (t ).  en, at t e
current time, we want to c oose e so t at ...
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



learly t is requires t at we c oose e as

 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 e above formula can be represented as in  igure

11.9.

 igure 11.9: Condition equivalent for the

anti wind-up controller
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



o s ow t at t is is equivalent to t e previous design, we

note t at in  igure 11.9

 rom w ere

Also

and

Hence t e sc eme in  igure 11.9 implements t e same

controller as t at in  igure 11.6.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



State Saturation
As a furt er illustration of t e application of anti-
windup procedures, we next s ow ow t ey can be
applied to maintain state limits.
We consider a plant wit nominal model given by
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



We ac ieve state constraints based on switc ing

between two controllers. One of t ese controllers
(t e prime controller) is t e standard controller
aimed at ac ieving t e main control goal, i.e. t at t e
plant output y(t) tracks a given reference, say ry(t).
 e task for t e secondary controller is to keep t e
variable z(t) wit in prescribed bounds.  is is
ac ieved by use of a secondary closed loop aimed at
t e regulation of t e estimated state, zÖ(t ) using a
fixed set point.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Our strategy will be to switc between t e primary

and secondary controller. However, it can be seen
t at t ere is a strong potential for wind-up since one
of t e two controllers, at any one time, will be
running in open loop. We will t us implement bot
controllers in anti-windup form.
 or simplicity of presentation, we assume t at a
bound is set upon |z(t)|, i.e. t at z(t) is symmetrically
bounded.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



ot controllers ave been implemented as in  igure

11.6.  us, t e prime (linear) controller as a
transfer function Cy(s), given by

Analogously, t e secondary (linear) controller as a

transfer function Cz(s), given by

 e final composite controller is s own on t e next

slide.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.10: Switching strategy for state saturation

£ ü is a switc w ic transfers between t e two controllers.

 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Substitutive Switc ing wit

Hysteresis
A simple approac is to transfer t e generation of t e
real plant input, u(t) from one controller to t e ot er,
in suc a way t at, at any time, u(t) is determined by
eit er uy(t) or uz(t).
If we aim to keep |z(t)| bounded by a known constant
zsat > 0, t en t is approac can be implemented using
a switc wit ysteresis, w ere t e switc ing levels
zl and zh, are c osen as 0 < zl < zh < zsat.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Weig ted Switc ing

A switc ing strategy w ic is an embellis ment of
t e one described above is described next. It also
relies on t e use of t e switc ing levels z1 and zh, but
wit t e key difference t at now t e (unsaturated)
plant input u(t) is a linear combination of uy(t) and
uz(t), i.e.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



W ere Ô [0, 1] is a weig ting factor. One way of

determining Ô would be:
 c Goodwin, Graebe, Salgado , Prentice Hall 2000




  Consider a plant with a model

where Y = G0U and Z =G0zU.

The reference is a square wave of unity amplitude
and frequency 0.3[rad/s]. It is desired that the state
z(t) does not go outside the range [-1.5; 1.5].
Furthermore, the plant input saturates outside the
range [-2; 2].
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



For this plant, the primary and secondary controller

are designed to be

The basic guidelines used to develop the above

designs is to have the secondary control loop faster
than the primary control loop, so that the state z(t)
can be quickly brought within the allowable bounds.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.11: Process variable z(t) (i) with state control

saturation (zc) and (ii) without state control
saturation (zu)

ote t at, w ereas t e unconstrained state exceeds

1.5 in magnitude, t e solution wit switc ed controller
leads to z not exceeding t e desired magnitude constraint.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.12: Plant output with (yc) and without (yu) state
control saturation

We observe t at t e effect of imposing t e state

constraint is to cause t e output (yc) to respond more
slowly t an w en t e state is unconstrained (yu).
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



The evolution of the weighting factor Ô(t) is shown in

Figure 11.13.
Figure 11.13 shows that the strategy uses a weight
which does not reach its maximum value, i.e. the
upper level zh is never reached.
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



 igure 11.13: £eighting factor behavior

 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Summary
Ë onstraints are ubiquitous in real control systems
Ë  ere are two possible strategies for dealing wit t em
limit t e performance so t at t e constraints are never
violated
carry out a design wit t e constraints in mind
Ë Here we ave given a brief introduction to t e latter idea
 c Goodwin, Graebe, Salgado , Prentice Hall 2000



Ë A very useful insig t is provided by t e arrangement

s own below:

 igure 11.14: Implicit inversion X-1

 is idea as been used t roug out t is c apter to
generate control strategies incorporating anti-windup
protection.