TP9 - 4:00

Integrator Windup and How to Avoid It

Karl Johan Astram and Lars Rundqwist
Department of Automatic Control
Lund Institute of Technology, S-22100 Lund, Sweden
Phone +46 46 108781, email kjaOcontrol.lth.se

Abstract. This paper describes the phenomenon of Incremental Algorithms

integrator windup and various ways of avoiding it. It first In the early phase of feedback control integral action
covers a number of ad hoc schema. A general procedure was integrated with the actuator by having a motor drive
to avoid windup which admits a unification of the ideas the valve directly. In this case windup is handled auto-
is given and the results are illustrated on a number of matically because integration stops when the valve stops.
enmple. When controllers were implemented by analog techniques
and later with computers some manufacturers used a con-
figuration that was a direct translation of the old me-
1. Introduction chanical design. This led to the so called incremental
algorithm. In this algorithm the rate of change of the
Most control systems are designed based on linear theory. control signal is first computed and then fed to an inte-
This is simple and works very well for regulation around grator. In some cases this integrator is a motor directly
a set point beause the process behavior can then be de- connected to the actuator. In other cases the integrator
scribed very well by linear models. To control a system is implemented digitally. With this approach it is easy to
over a wide range it is, however, necessary to consider handle mode changes and windup. Windup is avoided by
nonlinear effects. Actuator saturation, which is present stopping the integrtion whenever the output saturates.
in all systems, is particularly important. The effects of If the actuator output is not meured a model which
saturation are clearly noticeable when the controller is an computes the saturated output can be used. The rate of
unstable system This is always the case for a controller change of the control signal is also easy to limit.
with integral action. Since the feedback loop is broken
when the actuator saturates the unstable modes of the Back-calculation and Tracking
regulator may then drift to undesirble values. The conse-
quences are that it may take a long time for the system to The phenomenon of windup was noticed when the di-
reach equilibrium after an upset. The phenomenon which rect connection between the integrtor and the actuator
was first noticed in conventional PID control is therefore was broken. Several tricks were invented to avoid windup.
called integrator windup. It is well known to practitioners They were described under labels like preloading, batch
of automatic control but has not received much attention unit, etc. Although the problem was well understood
from theoretician and it is usually neglected in standard there were often limits imposed due to the analog imple-
courses on control. This paper ews the windup phe- mentations. The ideas were often kept as trade socrets
nomenon and different ways to avoid it. and not much spoken about.
The problem of windup was rediscovered when con-
trollers were implemented digitally and several methods
to avoid windup were presented in the literature. Back-
2. Ad hoc methods
Procs out and set poinm
Let us filst illustrate the phenomaenon of windup. Fig-
ure 1 shows PI control of an integrator with a satu-
rating actuator. The initial set-point change is so lare
that the actuator saturates at the high limit. The inte- 0 0io 40 0 80
gral term increases intially because the error is positive
and it reaches its largest value at ti t = 10 when the
error goes through zero. The output ns saturated
at this point because of the large value of the integral
term. It does not leave the saturation limit until the er-
ror has been negative for sufficiently long time to let the
integral part come dow to a small level. The net effect
is a large overshoot and a damped oscillation where the
control signal flips from one extreme to the other like in
reay oscillation. The output finally cons so close to the
set point that the actuator does not saturate. The system Figure 1. lustration of integrator windup (dashed
then behaves linearly and settles. line) and control with anti-windup (solid line).

1893
calculation, proposed by Fertik and Rosa (1967), works T1.5Tm2 13 T1.4
as folows: When the output saturates tie iegl is re-
computed so that its new value gives an output at the
saturation limit. It was found advantageous not to reset
the integator in one sampli peiod but dynamically
with a time constant T7. Figure 2 shows a block diagram
of a PID controller with such a backlulation The dif- O
I. 10 1I 2D O 5 to IS 20
ferece et between the output of the regulatorv and the fI 7,tz.0
actuator output u is fed to the input of the integrator
through a gain 1/7T. The signal et is zero when there
is no saturation. It will thus not have any effect on the QLS/
normal operation when the actuator does not saturate.
When the actuator saturates the feedback gnal prevents O S 1o I' 2' O S 10 15 20
the integrator from winding up. Figure S. The proportina band for a systa with
The rate at which the regulator output is reset is conditional integation and trackigwith dHkcnt trc-
governed by the time constant Tt, which determnes how ing time consnts
quickly the integral is reset. Fertik considered a discrte
system. In his scheme he used a dead-beat reset. This
approach has no exact continuous time equivalent but Conditiona Integration
it can be approximated by maing 2 vr small. It Conditional integration is an alternative to bak calcu-
frequently happens that the actuator output cannot be lation or tracking. The idea behind this method is to
measured. The anti-windup scheme just described can apply itegral action only when certain conditions are
be applied by estimating the actuator output using a fulfilled. The conditions are typically that the control er-
mathematical model of the saturation ror is small and that the actuator does not saturte This
Figure 1 also shbow the superior performance of method has been used for a long ti A recent reerce
the simple anti-windup scheme. The integral term has a is Gallun et al (1985). One posbility to implt con-
negative value during the initial phase when the actuator ditional integration is to update the integral only when
is saturated. This behavior is drastically different from the predicted proca output is in the proportional band.
that of the conventional PI controler where the integral When such a condition is impose it may also be neces-
has a positive value during the initial transient. Some sary to introduce a hystere4si or some similar condition
care must be taken when choosing time constan Tt. This to avoid chattering
is discussed further in Section 4. The proportional bst is an interval suh that the
A controller with back-calulation can be interpreted actuator does not saturate if the instatano s value of
as having two modes, the normal control mode when it the proess output or its predicted value is in the interyal
operates lke an ordinary controllr and a trcking mode For PUD control the control sga is given by
when the integrator is tracked so that it matches given
inputs and outputs. This is discussed farther in kstrbm
(1987). Since a controUler with tacnc operate in u=k(bvr -y) + I-kTd/ (1)
two modes, we may expect that it is necessy to have a
logical signal for mode swit This is not necessary Solving for the predicted process output y, = y + T4
because tracking is automatically inhibited when the we find that the proportional band is (I,,yj) where
tracking signal is equal to the regulator output. This can
be used with great advantage whebilding up complex
system with wlectors and cascade control, see Astr6m z =
byr + Ii
(2)
(1987) and Wittenmark (1989). Yk=bYr+ i

and ,, . are the values of the control gnal such

that the actuator saturates.
The notion of proportional band al helps to un-
derstand several other schemes for anti-windup. Figure 3
shows the proportional band for the system with tracking
for diffe t values of tie tracking time constant T2 and
a system with conditional integrtion. The figure shows
that in this particular cas there is very litte ence in
performance between conditional integration and track-
ing Also notice that the effect of tracking is to move the
proportional band cloer to the process output. There
may be a disadvantage in having it too close because the
Figure 2. Regulator with anti-windup band on back- system may become sensitive to occasional measurement
cakulatiot errOs.

1694
Multivariable sytems this interpretation it is easy to find a remedy simply by
feeding u back to the observer instead of v. If the actua-
The problem of -saturated multivariable systems is trea- tor output is not measured the actuator can be modeled
ted in Kapasouris et al (1988), where a stailizing con- mathematically. The controller can be described by
troller is derived. Here direction preservation of the con-
trol signal is accomplished by introducing a time varying t =Az Bu + K(y-CZ)
gain, which is chosen so that saturation is avoided. (3)
u =sat (um+ L(Zm-i))
Summary
where the function sat is defined as
We have thus given two ad hoc methods for avoiding
windup, tracking or back-calculation and conditional in- ulow u < un.w
tegrtion. These scheme can also be combined. In Howes sat(u)= u ui < u < uig (4)
(1986) it is suggested to explicitly manipulate the propor- UhLih u > Ukigh

tional band for batch control. This is done by introducing for a scalar and
so called cutback points. The high cutback is above the sat ul
set point and the low cutback is below. The integrator is gatU2
clamped when the predicted process output is outside the sat(u) = (5)
cutback interval. Integation is performed with a speci-
fied tracking time constant when the process output is sat U.
between the cutback points. The cutback points are con-
sidered as controller parameters which are adjusted to m- for a vector. The values ujO and uki, are chosen to
fluence the response to large set point changes. A similar correspond to the actuator limitations.
method is proposed in Dreinhoefer (1988), where condi- Notice that the dynamics of the controller is given
tional integation is combined with back-calculation. In by
Shinskey (1967) the integrtor is given a prescribed value det(uI-A + BL + KC) = 0 (6)
i = is during saturation. The value of io is tuned to give when the actuator does not saturate and by
satisctory overhoot at start-up. This approach is also
called preloading. Controlers with tracking are also dis- det(aI - A + KC) = 0 (7)
cussed in Glattfelder and Schaufelberger (1983, 1988).
when the actuator saturates. The dynamics of the con-
troller during saturation is thus given by the observer
3. General Methods dynamics.
The idea based on the observer interpretation can
Som gen techniques to avoid integrator windup in be applied to MIMO as well a SISO systems. For mul-
a controler will now be given. They are based on ideas tivariable systems there are several ways to saturate the
from state space theory. control signal. For syst having strong coupling it may
be useful to saturate the signal so that the direction of
An observer approach the control signal is preserved.
Notice that the controller (3) has high frequency roll-
In state space formulation a feedback controller is viewed off. The technique can be extended to controllers where
as a combination of an observer and a state feedback. the high frequency gain of the controller is constant. Such
The dynamics in the controller is giv only by the ob- a controller can be described by
server, see Figure 4. With such a controller it is easy
to understand what goes wrong when the actuator sat- do = Fz + G,y -
urates and to devise anti-windup schemes. The follow- di Gly(8 (8)
ing method was originally given in Astr6m (1983). It is u = Hz + Dy, - DOy
also described in Astrom and Wittenmark (1984). Let the
controller output be v and the process input be u. When To obtain anti-windup the control algorithm is rewritten
the actuator saturates u is differnt from v. Since the so that the control signal appears explicitly, and a feed-
controller is not aware of the saturation in the actuator back firom the difference between desired control signal v
it computes the state as if the process input is v. With and the saturated control signal u = sat(v) is introduced.
The following controller is then obtained.
xm
do Fz + Gryr - G,y + M(u - v) = (F - MH)z
=
I _ I
+ (Gr MDr)yr (GI -MD)y + Mu
v = Hz + D,y,-D ,y (9)
u = sat(v)

Figure 4. A proc with controller consisting of a where F - MH, which corresponds to the observer dy-
state feedback and an observer. namics, has stable eigenvalues.

195
Conditioning (-xo, -1), describing function analysis indicate a limit
cycle if the Nyquist curve of GpG0 is ts this line
A technique called conditioning was proposed by Bann. segment. Furthermre it follows from the circle cniterion
It is described in Hanns et al (1987). The idea is to that the closed loop system is stable provided that
compute the refece sgnal that would just saturate the
output. For a controller given in state space form (Eqn Re Gp(iw)Gc(iw) + 1 > 0 (16)
(8)) conditioning is equivalent to a special case of the
oberver approach, with M = G-rD?1 in Eqn (9). Hence
Controller wnth Anti-Windup
dt = (F -GrD H) - (G - D-1D, )y Now consider the consequence o using a controBler with
+ GrDr;lU (10) anti-windup. It folows from (11) and (12) that
u = sat(Hz + Dy, - D,y) d
=(A-KC)i (17)
The reference sgn does not influence the state during
saturation. A necessay condition is that the controler where S = a - S. The estimation error i thus goes to
has a direct term from the reference sgnl ,,i.e. Dr $ zeo exponentialy if A - KC is stable When i -0 , it
0. The eigenralues of F -GrDCVH are equal to the folos from (17) that the cloed loop system is described
transmison zeros of the controller. by -
- ~~dz
a = Ax + B &at L(z3. - a) (18)
Analysis
A system with a saturating actuator can always be re- This is a linear sstem with t r funtion
du-ed to a sda cfigation with a linear sysm
having a nonlinear feedback. Apprate analysis of Gw(J) L(JI - A)-lB
= (19)
such ems can be done u decrbi function the-
ory (Atherton, 1975). Suficient conditios for stability with a sturation in the feedb Desribing function
can also be obtained from the circle criterion (Deoer theory then indicates a limit cycle if the Nyquist curve
and Vidyasagr, 1975). Global stability is too strict a re- of G(J() interset the segn (-0, -1) and it folows
quirement if the plt is tale, because an unstable from hypertbility theory that the cosed loop is sable
plant can never be stabilied globally when the control provided that
signal is botuded. Re GC(iw) t I > 0 (20)
Consider a system descibed by We formulate the result as
dz THORE 1
W=A+BS (11) A controller with windup given by (12) gives a stable
D= C closed loop sstem wn cntling the proce (11) with
saturation prided that
with the ctroller
RcL(jI- A)-'B + 1> 0 (21)
(12) fora = iw. C
, =AA + Bl+K(y-CC)
Whe the conrol system design is it is
The traser function of the proces is thus straigh forward to investigte if the cntroerhas
windup. Notice that the condition (21) does not depend
Cp = C(aI- A)1-B (13) on the obserer. Also notice that if L is computed from
liner qdratic theory it fows that
and the controller ham the transfer function
.11 + Gw(iw)l 2! 1 (22)
Ge = L(aI-A--BL-KC)-1K (14)
Assume that the system has one control varible and one
measured and consider the conequences of a saturation 4. Example
of the actuator. The closed loop system is then in the
standard configuration with a linear block having the Properties of different methods for avoding windup will
transfer function now be demonstrated on an example. The proces con-
sists of two identical ca aded tanks, used for basic exper-
G(s) = Gp(a)Gc(a) (15) iments with automatic control, see Astr6m and Ostberg
(1986). The control signal is pump speed, which deter-
and a nonlinearity u = sat(y) in the feedback loop. Since mines the influent flow rate to the upper tank, and the
the describing function of a saturation is the line segment proces output is the level of the lower tank. The level

1696
a[\

given
,
u

0
(Vj;

Figure S. The standard

500

exeimn,

disturbances, give large ovrhoots due to integrtor

windup.

control is conducted by a P1D controller. The curves

in the paper are results of simulations. The cor-
responding behavior has however been observed in the
labortory.
A linearized state space
1000
with ddayed

model for the double tank

1.5

0.8

0.41

15

0o
II

OK'

1i-
yr

/
200
200

400
-

400

600
Figure 6. Trck}ing (solid), the obsver approac
(dashed) and conditional integrtion (dotted), with weUl
chosn paramets, re sted on the standard experi-
ment.

Ad=kb-
r
.i,.
\1
- --=-

660 86o

80(0

is
Li
-1
240 4*0 660 sbo
doz
d
=
a C
+
10
0) = Apa + Bpu
8]uA2+B
y= (O 1) z=Cz(

where ml and 2S are upper and lower tank levels respec-

tively, u is control input and y is measurement, all di-
mensionless. Parameters a 0.015 (s-1) and , = 0.05
=
20D 400 600
(-1). The process input u is restricted to the interval Figure 7.Iustration of the effects of different ob-
[0, 1]. The stationary in is $/a 3.33. erver bandwidths for anti-wndup based on the ob-
w

The process is controlled by a continuous P1D con- seveapproach. The bandwidths are = 0.05 nd/s w

troller with a filtered derivative and a proportional part (solid), w = 0.10 rad/s (dashed) and = 0.15 (dotted). w

that only acts on a faction b of the reference signal yr

(Astr6m and Higglund (1988)). The controller parame-
ters areK=5,T =40 T=15s,N =5andb=0.3, ,

which gives an overshoot of 10 % and a natural frequency Anti-windup methods

0.05 rad/s for the three dominating closed loop poles. In this Section some of the described anti-windup meth-
ods are compared, namely
Conditional integration, with integration suspended
Experiments during saturation.
A standard experint is used to test the anti-windup
Tracking,with Tt= T,=40 s.
methods. The process and controller starts at stationarity
with all signals zero. The experiment is as follows:
* The observer approach (Eqn (9)), with natural fre-
quency w = 0.05 rad/s and relative damping C = 1.
1. Start-up: at time t = 0 the reference Y, = 1.
These methods, see Figure 6, give smilar results. Con-
2. Impulse disturbance: at time t = 250 state Z2 is ditional integration gives the same overshoot as tracking
changed to 1.5. This corresponds to quickly pouring with small Tt. The methods are almost identical for load
of water into the lower tank.
a cup
and impulse disturbances.
3. Load disturbance: at time t = 500 a load on the In Figure 7 the observer approach is compared for
upper tank is introduced. three choices of w, with C = 1. w = 0.05 rad/s is approx-

The three parts of the experiment have been designed imately the natural frequency for the three dominati
to saturate the control sgnal in one or both direction. closed loop poles. The two faster observers (w = 0.10
Figure 5 shows the consequences of windup when the rad/s and = 0.15 rad/s) give less overshoot but the re-
control signal is restricted to the interval [0,1] without sponse from the impulse disturbance is very poor. Track-
use of anti-windup. ing with Tt < Ti gives similar results.

1697
 A number of simulations have been done. Some ri, E. J. Kompass and T. J. Williaas (Eds.): Ad-
observations are that the impulse disturbance is the most vaced Control in Computer integrated Manufactur-
critical one, that tuning the anti-windup for a good lug. Proc. 13th Annual Advanced Control Confer-
set point response may give a very bad response for ence Purdue University., Purdue Research Founda-
an impulse disturbance, and that if the respone of an tion. West Lafayette, Indiana.
impulse disturbance is good then the other responses ar Astr6m, K. J., and T. Hilund (19N8): Automatic
good. Tuning of PD) Regulators, Instrument Society of
For tracking, Tt = Ti or slightly 1 (O.8Ti), gives America, Triangle Research Park, NC.
good response from the impuls disturbance. Similarly,
the natural frequency of the observer should be slightly Astr8m, K. J. and B. Wittenmark (198): Com-
larger than the natural frequency of the closed loop. puter-Controled Syste Theory and Design, Pren-
Higher natural frequency or shorter Tt deteriorates the tice-Hall, Enwood Cliffi, NJ.
response of the impulse disturbance. The results from the Astr5m, K. J. and A.-B. Ostberg (1986): 'A modern
simulations agree well with the experimental results. teahing laboratory for proes control" IEEE Contr.
Sys. Mag., 6, No. 5, 37-42.
Desoer, C. A. and M. Vidyasagar (1975): Feedback Sys-
5. Conclusions temr: Input-Output Properties, Academic Press, New
York.
A number of techniques for avoid integrator windup
have been investigpted. Conditional integration is easy to Dreinhoefer, L. H. (1988): "Controller tuning for a slow
apply to most controlers. The key difficulty is to find ap- nonlin process, IEEE Contr. Sys. Meg., 8, No. 2,
propriate conditions for swtching off the tegration and
to avoid chattering. Tracking is another good method Fertik, H. A and C. W. Ros (1987): "Direct digital
for avoidig Windup. This method requires one parame- control algorithms with anti-windup r SA
ter, the tracking time constant, to be chose but there Trans., 6, No. 4, 317428.
is no risk of chattering. Trakings c eonv t to use
for systems with selectors and cascade control Selection Gallun, S. E., C. W. Matthew., C. P. Senyard and
B. Slates (1985): "Wmdup protection and iniitialisa-
of the paraters in the anti-windup schemes are impor- tion for advanced digital control," Hydrocabon Pro-
tant.The tunling is different for disturbances and set point cessng, June, 63-68.
changes. The obsere approach is a good general tech-
nique. It unifies many approaches and can be extended to G ia, C. E. and A. M. Morshedi (1986): "Quadratic
many controllers of different types. An advantage of the Pro m ng Solution of Dynamic Matrix Control
observer approach is that with a complete linear design (QDMC)," Chemical E .n ComWmnicationa,
no additional parameter have to be chosn The method 46, 73-87.
also applies diretly to multivarlable syste. Glattfelder, A. H. and W. Schaufelberger (1983): "Sta-
It should also be pointed out that control design bility Analysis of Single Loop Systeis with Saturation
can be formulated as an optimisation problem. LQG and Antireset-Windup Circuts," lEER Trans. Aut.
is a populr approach which leads to an unconstrained Contr., AC-28, 1074-1081.
quadratic optimisation problem. A natural extension is
to include actuator saturation. This lead to a quadratic Glattfelder, A. H. and W. Schaufelberger (1988): "Sta-
programmi problea There are very good numerical bility of Discrete Override and Cascade-Limiter Sin-
algonrthms available for solving such problems. This gle-Loop Control Systems," EE Tzans. Aut. Contr.,
is incorporated in quadratic dynamic matrix control AC-3U, 532-540.
(QDMC), see Garcia and Morshedi (1986), which gives Hanus, R., M. Kinnaert and J.-L. Henrotte (1987):
a nice solution to the problem of windup at the cost of "Conditioning technique, a general anti-windup and
increased computations. bumpless transfer method," Automatic&, 23, 729-739.
Howes, G. (1986): "Control of overhoot in plas-
tics-extruder barrel sones," El Technolog, No. 3, Eu-
6. References rotherm International, Brighton, England, 16-17.
Atherton, D. P. (1975): Nonlinar Control nee Kapasouris, P., M. Atha and G. Stein (1988): "Design
van Nostrnd Reinhold, London. of Feedback Control Systems for Stable Plants with
Astrom, K. J. (1983): "Practical aspects on digital Satuting Actuators," Prepzints for 27th IEEE Cont
on Dec. and Contr., Austin, Texas.
implementation of control laws," in AGAR.D Lecture
Series No. 128: Computer-Aided Desig and Analysis Shin key, F. G. (1988): Process-Contro Systes (3:rd
of Digital Guidance and Control Systems, AGARD, ed.), McGraw-Hill, New York.
Neully sur Seine, France. Wittenmark, B. (1989): "Integtors, Nolinties and
Astron3, K. J. (1987): "Advanced control methods - Sur- Anti-reset Windup for Different Control Structures,"
vey and assessment of possibilitie," in H. M. Mor- Proc. Amezican Contr. Conf, Pittsburgh, PA.

1698