Co pyrig ht © IFAC 11th Trie nnial World

Co ngress, Talli nn. ESlOn ia, USS R. 1990

ANTI-RESET WINDUP FOR PID

CONTROLLERS
L. Rundqwist
Department of Automatic Control, Lund Institute of Technology, p .a. Box 11 8,
5-22100 Lund, Sweden

Abstract. This paper describes and compares a number of ways of avoiding

integrator windup for proportional-integral-derivative (PID) controllers. It covers
both a number of ad hoc schemes and a general procedure to avoid windup, which
admits a unification of the ideas. Design rules for anti-windup in general purpose
PID controllers are derived and tested in simulations and experiments.

Keywords: PID control, process control, saturation, implementation

1. Introduction

Most control systems are designed for operation

in the linear range. For large set point changes
and disturbances the control signal will however
be saturated, and then the system operates in
open loop since the feedback path is broken. If
the controller is unstable the breakup of the loop
may give severe consequences. A PID controller
Figure 1. PIn controller with anti-wind up
is a typical example of a controller that may based on tracking.
cause instability or poor transient output during
saturation. suspended. Examples of conditional integration
This paper presents design rules for anti- in PID controllers are found in e.g. Fertik and
wind up compensators in general purpose PID Ross (1967), Gallun et al (1985) and Shinskey
controllers. The rules are based on the con- (1988).
troller's performance for a special disturbance.
They are important for controllers that may be Tracking or Backcalculation
utilized on a variety of processes.
The idea of backcalculation was proposed by Fer-
tik and Ross (1967) for a velocity limited incre-
2. Anti-windup methods for PID mental PI algorithm. The idea is that the stored
value is recomputed so that the controller out-
controllers put is exactly at the saturation limit. In Phe-
lan (1977) the "Intelligent Integrator" is ad-
A few methods for avoiding integrator windup
justed in the described way. During saturation
will be briefly described. They are described in
the controller thus tracks given inputs and out-
more detail in Astrom and Rundqwist (1989).
puts. It was found advantageous not to reset the
integrator in one sampling period but dynami-
Conditional integration methods
cally with a time constant Tt. Figure 1 shows a
In these methods the integrator is updated only block diagram of a PID controller with tracking.
during certain conditions. The essential differ- Controllers with tracking are discussed e.g. in
ences between the methods are the exact con- Astrom (1987) and in Glattfelder and Schaufel-
ditions for suspending and resuming integration, berger (1983).
and how the integral part i~ treated when it is

453
A general method where Xl and X2 are upper and lower tank levels
In a PID controller with filtering there are two respectively. The control signal u is restricted to
states but only the integrator state is adjusted by the interval [0,1]. Parameters a = 0.015 (s-l)
the tracking mechanism. It may however be fa- and f3 = 0.05 (S-l), thus the time constant for a
vorable to adjust both states during saturation. tank is approximately 70 seconds.
The idea can be generalized to n-th order con-
The controller
trollers, see Astrom and Wittenmark (1984). For
a controller on state space form The PID controller has filtered derivative of the
measurement signal y(t) and a proportional part
dx that only acts on a fraction b of the reference
Tt = Fx + GrYr - GlIy (1)
signal Yr, see Astrom and Hagglund (1988). The
v=Hx+DrYr-DIIY controller thus have PI action in the feedforward
path and PID action in the feedback path.
anti-windup is obtained by feedback from the
Parameter b positions one closed loop zero which
difference between desired control signal v and
has great influence on the overshoot after a step
the saturated control signal u = sat( v). The
change in the reference signal. A state space
following controller is then obtained.
model for the controller is
dx
Tt = Fx + GrYr - GlIy + M(u - v)
= (F - MH)x + (G r - MDr)Yr
dx
dt (~ _ ~ ) x + ( ~ ) Yr - ( R) Y
- (G II - MDII)y+ Mu (2)
v = Hx + DrYr - DlIy
u= (1 -KN)x+(Kb)Yr (4)

U = sat(v) ( K(l + N) ) Y

where F - M H has stable eigenvalues. This where Xl is the integral part and X2 is a low pass
controller realization corresponds to an observer, filtered -y. The derivative part is -K N(X2 + y).
and is therefor denoted the observer approach. (4) has the same structure as (1) and then many
The conditioning technique, see Hanus et anti-windup methods can be described with (2).
al (1987), is a special case of the observer ap- Controller parameters are K = 5, Ti = 40
proach, where M = G r D;:l in (2). The con- s, Td = 15 s, N = 5 and b = 0.3, which gives an
troller states x are then not directly affected by overshoot of 10 % and a natural frequency ~ 0.05
the reference signal Yr during saturation. The rad/s for the three dominating closed loop poles.
method requires that D;:l exists. The eigenval-
ues of F-G r D;:l H are equal to the transmission
zeros of the controller.
4. Analysis

3. Process and Controller The three principal methods described in Section

2 will be analyzed with respect to stability and
The properties of different methods will be inves- their performance for a special disturbance.
tigated for a special case. Design rules will also Tracking for (4) corresponds to
be developed.

The process (5)

The process is two identical cylindrical cascaded

in (2). The eigenvalues of F - M H are then
tanks used for basic experiments with automatic
control, see Astrom and Ostberg (1986). The
-l/Tt and -N/Td.
control signal is pump speed, which determines The observer approach for (4) needs selec-
the influent flow rate to the upper tank, and the tion of both the eigenvalues of F - M H. A simple
process output is the level of the lower tank. A choice is to make them equal, in this case -Wo,
linearized state space model for the double tank and then
1S
dx
dt
(-aa _Oa) x+ (~) u (3)
(6)

y= (0 1) x
454
The conditioning technique for (4) gives
a fixed anti-windup which is a special case of
tracking with Tt = bTi . This is due to the
feedforward PI structure of (4). If the reference
signal is differentiated conditioning does not
correspond to tracking.
In conditional integration the integral part
is kept constant until the control error changes
sign or the controller desaturates. Thus Ti = 00
in (4) during saturation. The feedback path is
then a PD controller, i.e. a lead compensator.
10 12 14 16
Stability
A system with a saturating actuator can always Figure 2. C*(iw) for tracking for (from right
be reduced to a standard configuration with a to left) Tt = Ti, 3/4Ti, T d , O.3Ti, Ti/lO.
linear system having a nonlinear feedback. For
time invariant monotonic nonlinearities sufficient
conditions for stability can be obtained from the
off-axis circle criterion (Narendra and Taylor,
1973).
If the nonlinear feedback element is a satu-
ration the linear system G*(s) is given by -6 .... :
'. :
'-;
:\
: \
-8 ".
(7) "'"

............
-10
where Gp is the process transfer function, Gc is
the feedback path of the controller and W de- -1~:-2-'--:--~--':-~--:----,I':-O-~12-~14~~16----l18
scribes the anti-windup. The two latter transfer
functions, which follow from (2), are Figure 3. C* (iw) for conditional integration
(solid line) and for C* = KC p (dashed line).
W = H(sI - F)-l M (8)
Gc = H(sI - F)-lG y + Dy (9) and Theorem 1 is satisfied since G* (s) + 1 is
SPR, see Figure 3. This is due to the positive
phase of the PD feedback path. If the derivative
THEOREM 1
action is also inhibited during saturation then
If the linear system G* (s) has all poles in the G*(s) = KGp(s) still satisfies Theorem 1, see
open left half plane and has nonlinear feedback Figure 3, but G* (s) + 1 is not SPR.
from a saturation the closed loop is absolutely
stable provided that a straight line through the
origin can be given a nonzero slope such that
G* (iw) + 1 is strictly to the right of the line. 0 Approximate disturbance response
Proof: See Narendra and Taylor (1973), p. 169. Another approach to determine tracking time
Tracking: Theorem 1 is satisfied for 0 < Tt < constant Tt or observer bandwidth Wo is to com-
00.G*(s) + 1 is strictly positive real (SPR) for pute the controller's response to a specific distur-
i
o < Tt < 30 s ~ Ti. The N yquist curves for bance and then select Tt or Wo such that certain
some values of Tt are shown in Figure 2. conditions are satisfied. A very critical experi-
ment with the two cascaded tanks is pouring a
The observer approach: Theorem 1 is sat- cup of water in the lower tank. The result is a
isfied for 0.012 rad/s < Wo < 2.9 rad/s. For step change in the measurement y(t). The con-
Wo < 0.012 rad/s G*(s) is conditionally stable. troller's derivative part then roughly produces
G*(s) + 1 is SPR for 0.067 rad/s < Wo < 0.93 an impulse, which is the dominating component
rad/s. in the control signal. The impulse saturates the
Conditional integration: Here controller and then the anti-windup mechanism
passes the impulse to the integrator.
Below the desired control output v(t) is
G*(s) = K(N + l)s + N/Td G (10)
determined based on an approximate sawtooth
s+N/Td p

455
shaped process output y(t) such that The observer approach Here the desired
control signal v(t) has the form
y(t) = {Yr + 6.y + y't if 0 < t. < T1I (11)
Yr otherwIse v(t) = Vo + vIt + V2e-wot + v3te-wot (20)

where 6.y > 0, y' < 0 and 6.y+y'T1I = O. For t < where v(O) is independent of Wo. Assume T1I ~
o stationarity is assumed. This approximation of WOI. Then
y( t) is only valid as long as the control signal is
saturated at the lower limit. Detailed derivation ky' (N 2N)
of the results below are found in Rundqwist v(TII) = Umin - -2
Wo
-T
d
+ T1•. - T.T
Wo. d
(21)
(1989) .
v(TII) 2:: Umin if
Tracking From (2) with matrices given by (4)
and (5) and y(t) given by (11) the desired control
output v(t) has the form (if N/Td =1= l/Tt ) (22)

v(t) = Vo + vIt + V2e-t/Tt + v3e-tN/Td (12)

~~ (0) however depends on Wo and a sufficient
where all Vi:S are functions of K, Ti, etc. but not condition for ~~ (0) > 0 is
of t. Now, approximate v(t) by assuming
1 (1
Wo 2:: 2(N + 1) Ti + Td ~
N) 1
2Td
(23)
N ~~ and t > 3 Td (13)
Td Tt N
The observer approach with Wo = N /Td gives
giving the same controller realization as tracking with
v(t) = Vo + vIt + V2e-t/Tt (14) Tt = Td/N. This is a too short time constant,
see (17), and thus N /Td is above an acceptable
In v(t) the fast mode is eliminated but not the
upper limit for Wo.
resetting of the integrator at t = O. We also
assume T1I ~ Tt and then RESULT 2
For anti-windup by the observer approach it is
required that
v(O) = Uo + k ( T
Tt d
- 1) 6.y - kTdy , (15)

V_( T1I ) = Umin - ky , Tt2 ( Tt1 - Ti1 ) (16)

o
where Uo is the stationary control signal be-
fore the disturbance and Umin is the lower con- For the controller in this paper 2h
0.033
trol limit. Two reasonable demands are that 1) rad/s, ]., = 0.050 rad/s and ~ = 0.333 rad/s.
V(O) ::; Uo and 2) v(TII) 2:: Umin. A necessary but
not sufficient condition for 1) is

(17)
5. Evaluation

In this section simulations with a linear process

A necessary and sufficient condition for 2) is
model and experiments on a nonlinear process
are used for evaluation of previous results.
(18)
Tracking
In Figure 4 some tracking time constants are
RESULT 1
compared in simulations with start-up and the
For anti-windup by tracking the tracking time
cup of water disturbance. I terms of the IAE
constant is to be chosen such that
(integral-absolute-error) criterion the choice Tt =
VTiTd gives the best result for the disturbance.
(19) Tt = Ti gives an acceptable result with slightly
larger undershoot . When Tt ::; Td (as when us-
o ing the conditioning technique for this particular
For the controller in this paper we get 15 s < Tt ::; controller) the controller desaturates too quickly
40 s. and gives a prolonged period of too high level

456
 o 100 200 300 400 500

0.8

0.4

o~----~~-----'--~~~-------r------~
o 100 200 300 400 500
Figure 4. Simulation of start-up and distur-
bance with anti-windup by tracking. Tracking
time constants are Tt = ..(T;Td (solid), Tt= T, Figure 6. Experiment with start-up and dis-
turbance with anti-windup by tracking, where
(dashed), Tt = Td (dotted) and Tt = bT" i.e. the
conditioning technique (dash-dotted). Reference tracking time constant Tt « T,. Reference sig-
signal is also dotted. nal is dashed.

I.S Y. r

.......
0.5

100 200 300 400 500

0.8

0.4

O~--~~~~--~~LL~~-- __~____~
o 100 200 300 400 500
Figure 7. Simulation of start-up and distur-
Figure 5. Experiments with start-up and dis- bance with anti-windup by the observer ap-
turbance with anti-wind up by tracking. Track- proach. Observer poles are given by Wo =
0.064
ing time constants are Tt = v'T,Td (solid), Tt = rad/s (solid), 0.050 rad/s (dashed), 0.033 rad/s
T, (dashed), Tt = Td (dotted) and Tt = bT, i.e. (dotted) and 0.100 rad/s (dash-dotted). Refer-
the cond:tioning technique (dash-dotted). Ref- ence signal is also dotted.
erence signal is also dotted.
undershoot. For Wo = 0.050 rad/s the result
in the tank. This agrees with that (17) was not is acceptable, but not for Wo = 0.033 rad/s.
a sufficient condition. In Figure 5 experiments Wo = 0.10 rad/s gives an acceptable response
with the same tracking time constants verify the without undershoot, but for higher values of Wo
results. the tank level is too high during a prolonged
The drawback of having too small a constant period. For start-up Wo = 0.14 rad/s (not shown)
Tt is shown in Figure 6. The fast resetting gives the minimal IAE value. Thus wo-values in
of the controller gives two results: 1) too quick the interval 0.05-0.10 rad/s, here corresponding
desaturation of the controller, and 2) saturation to 2/Ti -4/Ti , seem to be reasonable.
at the upper limit, causing a severe performance
deterioration. There is not much difference in Conditional integration
the start-up. The IAE criterion is smallest for
In Figure 8 conditional integration is tested. The
Tt ~ 0.2Ti· Thus Tt = ../TiTd (which often method handles both start-up and the distur-
becomes ~Ti) seems to be a reasonable choice.
bance properly. Conditional integration and the
best choices of tracking time constant Tt and 0 b-
The observer approach
server poles Wo give almost identical result for
In Figure 7 some values of Wo are compared the disturbance. For start-up conditional inte-
for the observer approach. For the disturbance gration gives slightly less overshoot. Thus the
the IAE criterion is smallest for Wo = 0.064 three methods are essentially equal in their anti-
rad/s. The two lower limits in (24) give larger windup capability.

457
 8. References

~:l cr ~:-
ASTRi}M, K. J. (1987): "Advanced control
!;:.
methods - Survey and assessment of pos-
j
,::
sibilities," in H. M. Morris, E . J. Kompass
and T. J. Williams (Eds.): Advanced Con-
0 50 100 150 200 250 300 350 400
trol in Computer integrated Manufacturing.
• •

Proc. 13th Annual Advanced Control Con-

~:l ts:;; ;;;j

ference Purdue University., Purdue Research
Foundation. West Lafayette, Indiana.
ASTRi}M, K. J., and T. HXGGLUND (1988): Au-
tomatic Tuning of PID Regulators, Instru-
o 50 100 150 200 250 300 350 400
ment Society of America, Triangle Research
Park, NC.
Figure 8. Experiment with start-up and dis-
turbance with anti-windup by conditional inte- ASTRi}M, K. J. and L. RUNDQWIST (1989):
gration. Reference signal is dashed.
"Integrator Windup and How to A void It,"
Proc. American Contr. Coni, Pittsburgh, PA.
6. Conclusions ASTRi}M, K. J. and B. WITTENMARK (1984):
Computer-Controlled Systems Theory and De-
All three anti-windup methods give stable closed
sign, Prentice-Hall, Englewood Cliffs, N J.
loops for the chosen process and controller. For
tracking the results from the disturbance analysis ASTRi}M, K. J. and A.-B. ()STBERG (1986):
are shown to be reliable both in simulations "A modern teaching laboratory for process
and in experiments. The results also agree with control," IEEE Contr. Sys. Mag., 6, No. 5,
well-known rules of thumb. Similar results were 37-42.
obtained for the observer approach. All three
FERTIK, H. A and C. W. Ross (1967): "Direct
anti-windup methods have, with well chosen
digital control algorithms with anti-wind up
parameters, almost identical performance.
feature," ISA Trans., 6, No. 4, 317- 328.
The cup of water disturbance is a very spe-
cial and strongly exciting disturbance for a con- GALLUN, S . E . , C. W. MATTHEWS, C. P. SEN-
troller with large high frequency gain, e.g. a con- YARD AND B. SLATER (1985): "Windup pro-
troller with derivative action. If such a controller tection and initialization for advanced digi-
may receive this type of disturbance it needs tal control," Hydrocarbon Processing, June,
careful anti-windup. This explains the fairly re- 63-68.
strictive results for the choice of tracking time GLATTFELDER, A. H. and W. SCHAUFELBER-
constant and observer poles. These results are
GER (1983): "Stability Analysis of Single
important for general purpose PID controllers,
Loop Systems with Saturation and An-
where the manufacturer cannot assume anything tireset-Windup Circuits," IEEE Trans. Aut.
about the type of disturbances in an application. Contr., AC-28, 1074- 1081.
For special purpose PID controllers which will
not receive this type of disturbances faster anti- HANUS, R., M. KINNAERT and J.-L. HENROT-
wind up is feasible. TE (1987): "Conditioning technique, a general
It was also demonstrated that a set point anti-windup and bump less transfer method,"
. oriented anti-windup method, e.g. the condition- Automatica, 23, 729-739 .
ing technique, may result in fairly poor perfor- NARENDRA, K. S. and J . H. TAYLOR (1973):
mance for disturbances when the feedback and Frequency Domain Criteria for Absolute Sta-
feedforward paths of the controller are too dif-
bility, Academic Press, New York.
ferent.
PHELAN, R. M. (1977): Automatic Control Sys-
tems, Cornell University Press, Ithaca, NY.
7. Acknowledgements RUNDQWIST, L. (1989): "Anti-reset wind up for
PID controllers," Report (to appear).
The author wishes to thank Karl Johan Astrom
for fruitful discussions and suggestions. SHINSKEY, F. G. (1988): Process-Control Sys-
tems (3:rd ed.), McGraw-Hill, New York.

458