Probably the best simple PID tuning rules in the world

Sigurd Skogestad
Department of Chemical Engineering
Norwegian University of Science and Technology
N–7491 Trondheim Norway

Submitted to Journal of Process Control July 3, 2001

This version: July 17, 2001

Abstract
The aim of this paper is to present analytic tuning rules which are as simple as possible and still
result in a good closed-loop behavior. The starting point has been the IMC PID tuning rules of Rivera,
Morari and Skogestad (1986) which have achieved widespread industrial acceptance. The integral term
has been modified to improve disturbance rejection for integrating processes. Furthermore, rather than
deriving separate rules for each transfer function model, we start by approximating the process by a
first-order plus delay processes (e.g. using the “half method”), and then use a single tuning rule. This is
much simpler and appears to give controller tunings with comparable performance.

1 Introduction
Hundreds, if not thousands, of papers have been written on tuning of PID controllers, and one must question
the need for another one. The first justification is that PID controller is by far the most widely used control
algorithm in the process industry, and that improvements in tuning of PID controllers will have a significant
practical impact. The second justification is that the simple rules and insights presented in this paper may
contribute to a significantly improved understanding into how the controller should be tuned.
The PID controller has three principal control effects. The proportional (P) action gives a change in
the input (manipulated variable) directly proportional to the control errorr. The integral (I) action gives a
change in the input proportional to the integrated error, and its main purpose is to eliminate offset. The
less commonly used derivative (D) action is used in some cases to speed up the response or to stabilize
the system, and it gives a change in the input proportional to the derivative of the controlled variable. The
overall controller output is the sum of the contributions from these three terms. The corresponding three
adjustable PID parameters are most commonly selected to be

 Controller gain Kc (increased value gives more proportional action and faster control)

 Integral time I [s] (decreased value gives more integral action and faster control)

 Derivative time D [s] (increased value gives more derivative action and faster control)
 E-mail: skoge@chembio.ntnu.no; Phone: +47-7359-4154

1
 Although the PID controller has only three parameters, it is not easy, without a systematic procesure,
to find good values (tunings) for them. In fact, a visit to a process plant will usually show that a large
number of the PID controllers are poorly tuned. The objective of this paper is to simple model-based tuning
rules that give insight into how the tuning depends on the process parameters based on very simple process
information. These rules may then be used to assist in retuning the controller if, for example, the production
rate is changed. Another related objective is that the rules should be so simple that they can be memorized.
There has been previous work along these lines; most noteworthy the early paper by Ziegler and Nichols
(1942), the IMC PID-tuning paper by Rivera, Morari and Skogestad (1986), and the book by Smith and
Corripio (1985). The Ziegler-Nichols tunings result in a very good disturbance response for integrating
processes, but are otherwise known to result in rather aggressive tunings (e.g., Tyreus and Luyben (1992)),
and also give poor performance for processes with a dominant delay. On the other hand, the IMC-tunings of
Rivera et al. (1986) are known to result in poor disturbance response for integrating processes (e.g., Chien
and Fruehauf (1990), Horn et al. (1996)), but generally give very good responses for setpoint changes.
d

?
gd

ys
-+ e - c
u - g -+ ?e+ q -y
6-

Figure 1: Block diagram of feedback control system.

In the simulations we consider input “load” disturbances (gd = g).
Notation. The notation is summarized in Figure 1. Here u is the manipulated input (controller output),
d the disturbance, y the controlled output, ys the setpoint (reference) for the controlled output, and y ys
the control error (offset). g s

( )=
y
()

u denotes the process transfer function and c s is the feedback part of
the controller.

2 Summary of method
2.1 Process information
The controller tunings are based on first approximating the process by a first- or second-order plus delay
model with the following model information (see Figure 2):
 Plant gain, k
 Dominant time constant, 1
 Effective time delay, 
 Second-order time constant, 2 (only used for dominant second-order process for which 2 >  ,
approximately)

2
 1

0.9
y(t)

0.8 k = ∆ y(∞) / ∆ u

0.7
0.63
0.6

u(t)
0.5

0.4

0.3

0.2

0.1

θ τ1
0
0 5 10 15 20 25 30 35 40

Figure 2: Step response of first-order with delay system, g s ( ) = ke s

( + 1).
= 1 s

8
If the response is sluggish or integrating, i.e. typically if 1 >  , then the exact value of the time
constant 1 and of the gain k may be difficult to obtain and is also not important for controller design. For
such processes one should instead obtain a good value for the

 Slope, k 0
def
= k=1
1. Obtaining parameters from experimental step response
If the starting point is an experimental step response (response in process output y to a step in the
process input u), then we may obtain the required process information as follows:

 The gain k is the ratio of the steady-state changes for the output and input,

=

uy (t ! 1)
k

Note that since we have normalized by dividing with
u, the gain k represents the output change
in response to an unit (magnitude 1) step input.
 The delay  is approximately the time it takes for the output to start clearly moving in the “right”
direction (towards its new steady-state).
 For a first-order process (2 =0
) we obtain 1 as the additional time until the output has moved
63% of the way to its new steady state.
 For a process which is dominant second order (with a S-shaped step response), we may want to
obtain also the second time constant 2 . It is recommended that numerical curve fitting is used
to obtain 1 , 2 and  in this case.
 For slow or integrating process we may instead of k and 1 obtain the initial slope k 0 of the step
response
k0 =
y=

u

t
where
y is the (maximum) change in the output y over a period
t following the initial delay.
3
k

A0

τ +θ = (A /k)
1 0

τ = 2.71 (A /k)
1 1

A1

τ +θ
1

Figure 3: Area method for determining parameters in first-order plus time delay model from step response

The above method for obtaining 1 and  is sensitive to errors, and the area method (Astrom et
al. 1993) shown in Figure 3 may be used instead.

2. Obtaining parameters from transfer function model

The effective time delay  is an approximation for the remaining high-order dynamics that are not
included in 1 (or 2 ). If the starting point is a detailed transfer function model, then we propose a
very simple method where the effective delay  is taken as the sum of the

+ “true” delay
+ inverse reponse time constant(s)
+ half of the largest neglected time constant (the “half rule”)
+ all smaller high-order time constants

The “other half” of the largest neglected time constant is added to 1 (or to 2 if we choose to use a
second-order model) – for more details see Section 5.

On transfer function form, the resulting model is then

0
( ) = ( s + 1)(k  s + 1) e s = (s + 1= k)( s + 1) e
g s s
(1)
1 2 1 2

2.2 Recommended SIMC-PID tunings

The tunings given below are for the cascade form PID controller:

Cascade PID : c(s) = Kc
I s +s 1
(D s + 1) (2)

I

The reason for using the cascade form is that the PID rules are much simpler in this case. when we have
derivative action. Following the internal model control approach (Rivera et al. 1986) where one specifies a

4
first-order closed-loop response with time constant c , the following SIMC tunings1 are recommended for
the process in (1) (see derivation in Section 3):

= k1  +1  = k10  1+ 
Kc (3)
c c

I = minf1 ; 0
4 g = minf ; 4( + )g
k Kc
1 c (4)

D = 2 (5)
where  < c < 1 is the tuning parameter. The optimal value of c is determined by a trade-off between
1. fast speed of response and good disturbance rejection (which are favored by a small value of c ), and

2. stability, robustness issues and small input usage (which are favored by a large value of c ).

For robust tunings it is recommended to use c   .

=
The original IMC tuning rules (Rivera et al. 1986) yield I 1 , that is, the integral time is selected so
as to exactly cancel the dynamics correpsonding to the dominant (first-order) time constant 1 . However,
this gives a very sluggish response to input (load) disturbances for “slow” (1 large) or integrating processes
(Chien and Fruehauf 1990). Therefore, for such processes it is suggested to use a smaller integral time, and
the recommended value I =
4 just avoids the slow oscillations that would otherwise result by using
k Kc 0

“too much” integral action for such a process.

Derivative action is primarily recommended for processes with dominant second order dynamics (with
2 >  ). Here the derivative time is selected so as to cancel the second-largest process time constant. In
addition, derivative action is often needed to stabilize unstable processes, but such processes are not covered
here.

2.3 Tuning for fast response with good robustness

The main limitation on achieving a fast closed-loop response is the time delay. Selecting the desired re-
sponse time equal to the time delay,
SIMC :
c  = (6)
gives a reasonably fast response with moderate input usage and good robustness margins, and results in the
following SIMC-PID tunings which may be easily memorized:

Kc = 0k:5 1 = 0k:50 1 (7)

I = minf1; 8g (8)

D = 2 (9)
Two common robustness measures are the gain margin (GM) and phase margin (PM). Typical minimum
17 30
requirements are GM> : and PM> o (Seborg et al. 1989), but for most control loops in the process
industries larger margins are recommended. Alternative robustness measures are the peak value Ms of the
sensitivity function S = = 1 (1 + )
gc , and the peak value Mt of the complementary sensitivity, for which
2
small values are desirable. For example, Ms < guarantees GM>2 and PM> : o . 29 0
With the SIMC PID-tunings in (7)-(9) the gain margin is typically above 3, the phase margin is about
50o o
60
, and Ms is 1.7 or less (Holm and Butler 1998). Specifically, with I =
1 the system always
1
The S in “SIMC” denotes Skogestad, Simple or Super – pick your choice.

5
has a gain margin (GM) of 3.14, a phase margin (PM) of : o , Ms 61 4 = 1 59
: , and a maximum allowed time
2 14
delay error of :  i.e., the tunings provide time delay error robustness in excess of 200% (see Table 1).
As expected, the robustness margins are somewhat poorer for “sluggish” processes, where we in order
to improve the disturbance response use I =8
 . For example, for an integrating process the suggested
46 9
tunings give a a gain margin of 2.96, a phase margin of : o , and a maximum allowed time delay error of
1 49
: .

Process g s() k
e s k0
e s
Controller gain, Kc 0 5 +1
1 s
: 1
s
051
:

8
k  k0 
Integral time, I 1
Gain margin (GM) 3.14 2.96
Phase margin (PM) 61.4o 46.9o
Allowed time delay error, =
2.14 1.59
Sensitivity peak, Ms 1.59 1.70
Complementary sensitivity peak, Mt 1.00 1.30
Phase crossover frequency, !180
 1.57 1.49
Gain crossover frequency, !c
 0.50 0.51
Table 1: Robustness margins for first-order and integrating delay process using SIMC-tunings in (7) and (8)
=
(c  ). The same margins apply to second-order processes if we choose D 2 . =
Derivation: For the first-order delay process with I = 1 the resulting loop transfer function is L s gc ()= s=
0:5 e s . The

frequency !180 where the phase of L is 180o is then 6 L = !180  2 = ) w180 1

2 . =
The gain of L as a function
05
of frequency is : =! and by evaluating the gain at the frequency !180 we find that GM = 1 ( ) = = 3 14
=jL j!180 j  : . Similarly,
we can show that the frequency !c where jLj =1 =05
is !c =
: = og we find that PM  6 L j!c ( ) = 2 0 5[ ] = 61 4
= : rad : o.
The maximum allowed time delay error is then 
= PM = ( =!C   1) = 2 14
: .

These good margins come at the expense of a somewhat more sluggish time response compared to that
which can be achieved with more aggressive tunings. Note that for the case with I 1 , increasing Kc by =
a factor of 2 (corresponding to choosing c =0 33
), reduces PM from 61o to o and reduces GM from 3.14 to
1.57, which are rather poor robustness margins. Thus, to maintain resonable robustness, the controller gain
should be at most a factor of 2 larger than the value given in (7).

2.4 Tuning for slow response

In many cases the above choice c =
 may be unnecessary “aggressive” and we may want to increase c
or equivalently decrease Kc . In particular, this may be the case for processes with a small (effective) time
delay, for example, a pure first-order process. More generally, there are cases where we want to use as little
control as possible, that is, we want a slow or smooth response. However, there is usally some performance
requirements in terms of the allowed output variation, and this gives a minimum controller gain needed to
achieve satisfactory disturbance rejection. For example, for the case of input “load” disturbances we must
approximately require that (see (37) below):

Kc  ydu (10)
max

Here ymax is the allowed output error (y ys ), and du is the magnitude of the input “load” disturbance. As
expected, tight control with ymax small requires Kc large, as does a large disturbances with du large.

6
 After deciding on a reasonable value for Kc , one may from (3) back-calculate the corresponding value
of c . For cases where the integral time is not equal to 1 one may then modify the integral time according
to (4).
If the “minimum” controller gain given by (10) is larger than the “maximum” the controller gain given
in (7), then the process is not controllable – at least not with PID control with reasonably robust tunings. In
words, the speed of response required for disturbance rejection is faster than what can be achieved with the
given time delay.
Example. Consider a second-order with delay process with time constants 1 = 6 and 2 = 1:2, and time
delay  = 0 25
: :
0:25s
( ) = 4 (6s + e1)(1:2s + 1)
g s (11)

The requirements is that the output deviation should be less than ymax =1
in response to a load disturbance
du =05
: . It is also desirable that the input usage is as smooth as possible.
Tuning for fast response. With c = = 0 25
 : the recommended tunings (7)-(9) for a cascade form

= 0k5  = 3;
PID controller are
: 1
Kc I  = 8 = 2;
D 2 : = =12 (12)
The load disturbance response in Figure 4 is much better than the requirement, with a output deviation in
response to the load disturbance of less than 0.1. However, the input has some overshoot and oscillations.
Tuning for slow response. The above response is unecessary fast. To reject the disturbance we need
a minimum gain, which from (10) is approximately Kc du
ymax
=
0:5
1 = =05
: (corresponding to c : ), = 2 75
and the resulting PID tunings are

Kc = 0:5; I = 1 = 6; D = 2 = 1:2 (13)

The load disturbance response in Figure 4 has an output deviation y ys of about : 15 1=05
: which is
well below 1, and the input is smooth with no overshoot or oscillations. Thus, this tuning is preferred in
practice.
Remark: We may reduce Kc further below 0.5 and still achieve an output deviation less than 1. The
reason why (10) is not tight, is that (1) the expression is derived for sinusoidal disturbances whereas we con-
sider a step disturbance, and (2) the derivative time is quite close to the integral time so that the controller
gain as a function of frequency does not come down to its asymptotic value of Kc .

2.5 Ideal PID controller

The above tunings (Kc ; I ; D ) are for the cascade form PID controller in (2). To derive the corresponding
tunings for the “ideal” PID controller
!
1 + 1s + Kc0  0 0 2 
Ideal PID : 0( )=
c s Kc 0
0  0
D s 0  
I D =s
s  0 0 s
I D +( + ) +1 (14)
I I

we use the formulas

   
Kc0 = Kc 1 + D ; I0 = I 1 + D ; 0
D = 1 +DD (15)
I I I

Note that it is not always possible to do the reverse and obtain cascade tunings from the ideal tunings. This
is because the ideal form is slightly more general as it also allows for complex zeros in the controller. Thus,
if we want to derive PID-tunings for a second-order oscillatory process which has complex poles, then we
should start directly with the ideal PID controller.

7
 1.5

τ =θ=0.25
OUTPUT y
c
1

0.5 τ =2.75
c

0
0 5 10 15 20 25 30 35 40

1.5
INPUT u

0.5

−0.5
0 5 10 15 20 25 30 35 40
time

Figure 4: Responses for process (11) with “fast” PID-tunings in ( 12) (solid line) and “slow” PID-tunings
in( 13) (dashed line)
Load disturbance of magnitude 0.5 occurs at t = 20.
The tuning parameters for the the cascade and ideal forms are identical when the ratio between the
derivative and integral time, D =I , approaches zero, that is, for a PI-controller (D ) or a PD-controller =0
(I 1). =
The SIMC-PID cascade tunings in (7)-(8) correspond to the following SIMC “ideal”PID tunings (c =
 ):
1  8 : = 0k:5 1 + 2 ; I0 = 1 + 2 ; D0 = 1 +2 
Kc0 2
(16)
 1

1  8 : Kc0 =
0:5 1 1 + 2 ;  0 = 8 +  ;  0 = 2
 
k  8 I 2 D
1 + 8 2
(17)

Note that the tuning rules for the ideal form are much more complicated.
Example. Consider the second-order process in (11) with cascade-form PID tunings given in (12). The
corresponding tunings for the ideal PID controller in (14) are
Kc0 = 4:8; I0 = 3:2; 0
D = 0:75
The robustness margins with these tunings are given by the first column in Table 1.

3 Derivation of SIMC tuning rules

We will now derive the above tuning rules.

3.1 Tuning for setpoint response

Our starting point is a second-order with delay model,
s
g s ( ) = k ( s + 1)(
e
 s + 1)
(18)
1 2
8
for which we want to derive analytical PID-settings. We use the direct synthesis approach of Rivera et al.
(1986) where we specify the desired setpoint response. Under feedback control the closed-loop setpoint
response of the system in Figure1 is

ys
y
= gcgc+ 1 (19)

where c is the feedback controller, and we have assumed that the measurement of the output y is perfect.
Following Rivera et al. (1986), we specify that we, after the delay, desire a simple first-order response
!
y
ys
=  s1+ 1
e s (20)
desired c

We have kept the delay in the “desired” response because it is unavoidable. c is the desired closed-loop
time constant, and is the sole tuning parameter for the controller. Combining (19) and (20) and solving with
respect to the controller gives a “Smith Predictor” controller (Smith 1957):

( ) = (1 s + 1)(k 2s + 1) ( s + 11

c s
e s ) (21)
c

To get a PID-controller we introduce in (21) the follwing first-order Taylor approximation for the delay

e s
1 s (22)

( ) = (1 s + 1)(k 2s + 1) ( +1 )s

and derive
c s (23)
c

which is a cascade form PID-controller (2) with

Kc = k1  +1  ; I = 1 ; D = 2 (24)
c

Alternative approximations of the delay

1. We may instead of (22) use the more exact 1st order Pade approximation,

2s + 1
2
= s 
s
= ee 
2s + 1
e =2s 

With the choice c =  this results in the same PID-controller (23) found above, but in addition we
get a term
2s + 1


0:5 2 s + 1 (25)

which may be viewed as an additional derivative term which is effective over only a very small range,
and increases the controller gain by a factor 2 at high frequencies. Simulations show that performance
is only slightly improved by adding this term (at least with the choice c  ; see Figure 5)), and thus =
does not justify the increased complexity of the controller.

2. Original IMC PID tunings for first-order with delay process. Rivera et al. (1986) introduced
the Pade approximation in the process itself, before deriving the controller. By specifying a closed-
loop response y=ys = (=2)s+1 (note that  is denoted " in their notation) and choosing " , =2
"s+1 c
their resulting “(unimproved) IMC PI-tunings” for a first-order with delay process are identical to the

9
 tunings (24) just derived. They also propose some variations. One is the “improved IMC PI-tuning”
where the integral time is changed from 1 to 1 = : + 2
IMC PI : Kc = 1 1 + =2 ;
k "
I = 1 + =2 (26)

17
with "  :  . This improvement has some effect for dominant time delay processes (with 1 =
small), but it is minor and probably does not justify the added complication in the rule.
Rivera et al. (1986) also propose for a first-order with delay process to use an additional derivative
2
term with time constant = resulting in the “IMC PID-tunings”:

IMC cascade PID : Kc = k1 " +1=2 ; I = 1 ; D = 2 (27)

08
with "  :  . With their recommended value " =08
:  (tight control) this gives some improvement
with less overshoot in the setpoint response, but the load disturbance response is almost unchanged.
For larger values of " (more robust tuning corresponding to the SIMC-rules), there is very little
improvement also in the setpoint response; see Figure 5.

1.4

1.2
τ =0
D

τD=θ/2
OUTPUT y

0.8

0.6 Kc = (0.5 / k ) ⋅ (τ1 / θ )

0.4 τI = τ1

0.2

0
0 5 10 15 20 25 30 35 40
t/θ

Figure 5: Introduction of derivative action (solid line) has only a minor effect for first-order with delay
process, g s ( )=
ke s = 1 ( + 1)
Note: Controller gain corresponds to c  in (24) and " =  in (27) =2
Load disturbance of magnitude 0.5 occurs at t = 20.

In summary, the tunings proposed in this paper are similar to the IMC-tunings of Rivera et al. (1986).
Rivera et al. (1986) proposed some modifications to improve the response, which have only a minor effect,
and do not seem worthwhile. However, for a process with 1 = large, there is a significant scope for
improvement when it comes to disturbance rejection (Chien and Fruehauf 1990). This is discussed next.

3.2 Modifying the integral term for improved disturbance rejection

Above we derived PI- and PID tunings based on considering the setpoint response. We found that we should
effectively cancel the first order dynamics of the process by selecting the integral time I 1 . This is a =
10
robust setting which results in very good responses when it comes to setpoint changes and disturbances
occuring directly at the process output. However, it is well known that for processes where 1 is “large”
(e.g. an integrating processes), this choice results in a long settling time for input “load” disturbances
(Chien and Fruehauf 1990). The reason is that the controller cancels the process dynamics, whereas for a
disturbance occuring at the input we actually want to keep the dynamics. To improve the load disturbance
response we therefore want to reduce the integral time. However, we must not reduce the integral time
too much, because otherwise we will encounter slow oscillations caused by almost having two integrators
in series (one from the slow dynamics in the process and one from the controller). This is illustrated in
Figure 6 for a “slow” process with 1 : = 30
 I = 1 = 30 (IMC) gives slow settling for a load disturbance.
 I = 8 = 8 (SIMC) gives faster settling.

 I = 4 gives even faster settling, but the setpoint response (and robustness) is poorer.

 I = 2 gives poor response with oscillations.

1.8

y(t) τ =2
I
1.6
τ =30
I
4
1.4
8
8
1.2
30 4
ys= 1

2
0.8

0.6

0.4

0.2

0
0 10 20 30 40 50 60
time

Figure 6: Effect of changing the integral time I for PI-control of “slow” process g s ( ) = e s=(30s + 1)
with Kc . = 15
Load disturbance of magnitude 10 occurs at t = 20.
A good trade-off between disturbance response and robustness is obtained by selecting the integral time
such that we just avoid the oscillations (I =8
 in the above example). Let us analyze this in more detail.
First, note that these “slow” oscillations have a different origin and occur at a lower frequency than the
1
usual fast oscillations which occur at about the frequency of the delay, = . Because of this, we neglect the
delay in the model when we analyze the slow oscillations. The process model then becomes

( ) = k  es + 1  k  s1+ 1  ks = ks
s 0
g s
1 1 1
where the second approximation applies since the resulting frequency
 of oscillations
 ! is such that 1 ! 2 ( )
is much larger than 1 (see footnote). With a PI controller c Kc = 1+
1 s the closed-loop characetristic
I

11
equation 1 + gc then becomes I
s 2 + I s + 1
k 0 KC
which is on standard second-order form 02 s2 + 20 s + 1 with
s
0 =
I
; = 1 qk 0 K 
k 0 Kc

2 c I (28)

To avoid slow oscillations we must have a damping coefficient   1. Of course, some oscillations may be
tolerated, but nevertheless a good starting value is to have  = 1 (see also Marlin (1995) page 588), which
gives
Kc I = 4=k 0 (29)
or equivalently
= k04K I (30)
c

which is the value recommended in (4). The choice Kc = 0k:5 1 in (7) gives I = 8 as given in (8). For a
0

first-order with delay process this gives a gain margin better than 2.96 and a phase margin better than 46.9o ;
see Table 1.
We get slow oscillations if the product of the controller gain Kc and the integral time I is reduced
compared to the value given in (29). What is the period P of these oscillations? From a standard analysis
of second-order systems, we have that (e.g. Seborg et al. (1989) page 118)
s
P = 12
p 2 0 > 0  0I


2 =2 k Kc
(31)

where the inequality applies since the presence of oscillations requires   1. With the suggested tuning
I =4 =k 0 Kc (30) this gives
P >  I
(32)
Thus, the “slow” oscillations which result by reducing the controller gain have a period larger than 3 times
the integral time.2. On the other hand, the “usual” fast oscillations that appear by increasing the controller
gain have a period of 6 times the delay. This is illustrated in Figure 7 for a “slow” process with 1 = 30 ,
 =1 =4
and integral time I :

 Kc = 15 gives no detectable oscillations.

 Increasing the controller gain (Kc = 30) gives fast oscillations with a period of about 6 (about 6 times
the delay).

 Decreasing the controller gain (Kc = 3) gives slow oscillations with a period of about 30 (larger than
3 times the integral time).

2
The corresponding normalized frequency of these slow oscillations is 1 ! = 1
2=P  21=I which is larger than 2
since we use I  1 .

12
 2
Kc=30
1.8

1.6
Kc=15 K =3
1.4 c

1.2
OUTPUT y

0.8

0.6

0.4

0.2

0
0 5 10 15 20 25 30 35 40 45 50
time

Figure 7: Effect of changing the gain Kc for PI-control of “slow” process g s ( ) = e s=(30s +1) with I = 4.
Setpoint responses.

3.3 Lower limit on controller gain

In many practical situations we do not require very fast control, and to reduce the use of manipulated
inputs and generally make operation smoother we may want to use lower controller gains. Are there any
lower limits on the controller gain? Yes, there are, and to derive this we will consider the performance
requirements for disturbance rejection.
The linear transfer fuction model with deviation variables is

y = g(s)u + gd(s)d (33)

()
where gd s is the disturbance transfer function model. With feedback control, u = c(s)y, the effect of a
disturbance d on the control output y is
y = S (s)gd(s)d
where S s ( ) = 1 (1 + ( )) ( )
= g s c s is the sensitivity function. Let d denote the disturbance magnitude, and
ymax the allowed output variation. We assume that this requirement applies on a frequency-by-frequency
basis, i.e., for a sinusoidal disturbance with frequency ! [rad/min] and magnitude d, the resulting sinusoidal
output should have a magnitude less than ymax . Since the sinusoidal response is mathematically obtained
=
by setting s j! , the requirement becomes

jy(j!)j = jS (j!)j
jgd(j!)j
d  ymax

j1 + g(j!)c(j!)j  jgd(yj!)j
d
or

max

1
At low frequencies (i.e., within the closed-loop bandwith) we have that jgcj  and we derive the following
lower limit on the frequency-dependent controller gain

d (j! )j
d
jc(j!)j  jgj(gj! )j
y (34)
max

13
At lower frequencies, where this expression applies, we effectively have “perfect control” and y  . From 0
(33) the required input to reject the disturbance (i.e., achieve y ) is ud =0 =( )
gd =g d, and we derive the
following alternative expression
jc(j!)j  uyd(j!) (35)
max

( )
where ud j! is the magnitude of the input change needed to reject the disturbances and ymax is the maxi-
mum allowed output deviation (y ys ). By constructing a controller which just satisfies the bound (34) or
(35), we obtain the “slowest” acceptable controller (this is generally not a PID controller).
For the special case of a load disturbance (distubance du at the input) we have gd g and the require- =
ment (34) becomes
( )  ydu c j! (36)
max

For a P-, PI- and PID-controller the controller gain jc(j! )j has a minimum asymptotic value 3 of Kc , and
we derive the following lower limit on the controller gain,

Kc  ydu (37)
max

From (35) and (37) we derive the following useful rule:

 The minimum controller gain is approximately equal to the expected input change divided by the
allowed output variation.
=
We can rearrange (37) into ymax du =Kc , which in words says that that the maximum output change ymax
in response to a load disturbance is du =Kc . This is for a sinusoidal disturbance, but as illustrated in the
simulations, a step disturbance often results in a similar value. For example, in Figure 6 we see that the
maximum output deviation y ys in response to a step disturbance is about 0.65 (independent of the integral
time) which compares well the value du =Kc = = 10 15 = 0 67
: .

4 Some special cases and comparison with other tuning methods

We here present some special cases and compare the suggested SIMC tuning rules with the classical “closed-
loop” tuning rules of Ziegler and Nichols (1942), as well as some other tuning methods. We find that the
simple SIMC tunings generally perform very well.
Ziegle-Nichols (ZN) tuning rules. The first step in the Ziegler-Nichols procedure is to generate sus-
tained oscillations with a P-controller, and from this obtain the “ultimate” gain Ku and corresponding
“ultimate” period Pu . Based on simulations, Ziegler and Nichols (1942) recommended the following ZN
tunings:
P control : Kc = 0:5Ku
PI control : Kc = 0:45Ku; I = Pu=1:2
PID control (cascade) : Kc = 0:6Ku; I = Pu=2; D = Pu=8
These tunings were based on pneumatic controllers similar to the cascade form of the PID controller given
in (2) (Shinskey 1998). From (15) this means that for an ideal PID controller the ZN tunings are:

PID control (ideal) : Kc = 0:48Ku; I = Pu=1:6; D = Pu=10

3
1 1
For a PID controller the break frequencies are at =I and =D , and the controller gain as a function of frequency will only
reach its asymptotic minimum value of Kc for cases where the integral time I is significantly larger than the derivative time D .

14
 Tyreus-Luyben modified ZN tuning rules. The ZN tunings were derived to give decay ratio of 1/4.
This is too aggressive for most process control systems, where oscillations and overshoot is usually not de-
sired at all. This lead Tyreus and Luyben (1992) to recommend the following PI-rules for more conservative
loops:
Kc = 0:313Ku; I = 2:2Pu
Regressed analytic tuning rules. Many papers on PID control include comparisons with the tuning
rules of Cohen and Coon (1953) where the tunings are given by analytical functions of k; 1 and  . These
tunings were also derived for a decay ratio of 1/4 and are generally too aggressive, and performance is
usually poor (this is probably why it is popular to compare with them since anyoone can beat them). Later,
there has been many papers along these lines, e.g. Ho et al. (1998).
Astrom PI tuning rules. Schei (1994) argued that in process control applications we usually want a ro-
bust design with the highest possible attenuation of low-frequency disturbances, and suggested to maximize
the low-frequency controller gain
KI = K c (38)
I
subject to given robustness constraints on Ms and Mt . Astrom et al. (1998) showed how to formulate this
as a convex optimization problem for the case with PI control and a constraint on Ms . The value of the
tuning parameter Ms is typically between 1.4 (robust tuning) to 2 (more agressive tuning). To improve the
setpoint performance Astrom et al. (1998) use a “two degrees of freedom controller” where they use only a
=1
fraction b of the propotional action on the setpoint, but we do not use this here (i.e., we set b ).

4.1 Pure time delay process

( ) = ke
g s s
(39)
Note that a pure P-controller is unacceptable for this process, because even with maximum gain (at the limit
to instability) the steady-state offset is 0.5 (50%). Thus integral action will be needed.
SIMC tunings. This is a special case of (1) with 1 =0
and 2 =0
. The rules (3) and (4) give Kc ! 1
= =0
and I 1 . More precicely, the controller becomes

( ) = Kc I s +s 1 ! k ( 1+ ) 1s
c s
I c

( ) = KsI . With the suggested choice c =  the integral gain is

which is a pure integral controller c s

K =
0:5 (40)
I
k
corresponding to GM = 3.14, PM = 61.4o and Ms = 1 59
: . This is not a PI controller, but it may of course
be approximated by a PI controller by choosing I small and using Kc 0k:5
I . =
ZN tunings. For this process, a pure proportial control with gain Ku =1
=k results in persistent
oscillations with period Pu =2 (at the limit to instability). The Ziegler-Nichols tunings rules then give the
following PI-tunings
Kc = 0:k45 ;
= 1:67  I (41)
corresponding to GM = 2.18, PM = 99.5 o and Ms = 1:85. Thus, the robustness is acceptable, but the
simulations in Figure 8 show that reponse with the ZN controller is sluggish. This may explained by the
relatively low integral gain, KI = Kc =I = 0:27=(k ). We therefore conclude that the Ziegler-Nichols
settings are generally poor for a pure time delay process. This may partly explain the myth in the process
industry that a PI controller should not be used for processes with large time delays.

15
 2

1.5
OUTPUT y
Astrom
SIMC
y = 1
s

ZN
0.5

0
0 5 10 15 20 25 30 35 40

1.5
Astrom
SIMC
1
INPUT u

ZN
0.5

0
0 5 10 15 20 25 30 35 40
time

Figure 8: Responses for pure delay process, g s e s. ( )=

SIMC (solid line): I-controller (40) with Kc and KI : =0 =05
ZN (dashed line): PI-controller (41) with Kc : and KI : =06 = 0 27
Astrom (dashed-dot): PI-controller (42) with Kc : and KI = 0 26 = 0:85
Astrom tunings. With Ms = 1:4 (robust tunings), Astrom et al. (1998) derive a PI-controller c(s) =
Kc + KsI with Kc = 0:k16 and KI = 0:4724 . We note that the integral part is almost the same as for the SIMC
integral controller in (40), and the response is also quite similar. With Ms = 2 (more aggressive tunings),
k

Astrom et al. (1998) derive

K =
0:26 ; K = 0:85 (42)
c I
k k
o
corresponding to GM=2.11, PM = 54.2 and Ms =2
. The simulations in Figure 8 show that the perfor-
mance with this latter PI controller is somewhat faster, but the reponse with the SIMC integral controller is
smoother and more robust.

4.2 Integrating process

Consider an integrating process with delay  ,
s
( ) = k0 e s
g s (43)

The corresponding PI tunings for some methods are summarized in Table 2.

SIMC tunings. This is a special case of (1) with 1 ! 1. From (3) and (4) we get with c =  a PI
= k10
 1+  = 0k:50
1
controller with
Kc (44)
c
I = 4(c +  ) = 8 (45)
ZN tunings. For this process a proportial controller with gain Ku  1 1
=
2 k  results in persistent oscilla-
=4
0

tions with period Pu  . The Ziegler-Nichols tunings rules then give the following tunings:

ZN PI : Kc = 0k:710 1 ; I = 3:33 
16



Kc k 0  I = GM PM Ms Mt
=
SIMC (c  ) 0.5 8 2.96 46.9o 1.70 1.30
IMC (c =17
: ) 0.59 1 2.66 56.2o 1.75 1.07
ZN 0.71 3.33 1.86 24.8o 2.85 2.37
Tyreus-Luyben 0.49 7.32 3.00 45.9o 1.70 1.33
Astrom (Ms =14
: ) 0.28 7.0 5.24 47.5o 1.40 1.43
Astrom (Ms =2) 0.49 3.77 2.77 32.8o 2.00 1.81

Table 2: PI tunings for integrating process, g s ( ) = k0 e s

=s

2
IMC
1.8 ZN

1.6

SIMC
1.4

SIMC
1.2
OUTPUT y

ZN
y = 1
s

0.8

0.6

0.4

0.2

0
0 5 10 15 20 25 30 35 40
time

Figure 9: Responses for integrating process, g s e s =s.( )=

SIMC (solid line): PI-controller with Kc : and I =05. =8
IMC (dashed line): P-controller with Kc : and I = 0 59
1. =
ZN (dashed-dot line): PI-controller with Kc : and I = 0 71
: . = 3 33
ZN PID : Kc = 0k:940 1 ; I = 2 ; D = =2
The ZN tunings give (as usual) poor robustness, but as seen in Figure 9 the ZN tunings give considerably
faster settling than the SIMC tunings for load disturbances.
The Astrom tunings with Ms =2
give responses somewhat in between SIMC and ZN.
The Tyreus-Luyben modified (conservative) ZN tunings are almost indentical to the SIMC tunings for
this particular example.
The IMC tunings of Rivera et al. (1986) result in a pure P-controller since I 1 2 ! 1. This

= +
P-controller is acceptable for setpoint changes, but load disturbances integrate and result in steady-state
offset.

4.3 Integrating process with delay and lag

s
( ) = k0 s(es + 1)
g s (46)
2

17
SIMC tunings. This results in the same tunings and responses as for the process (43), but we must add
derivative action to counteract the lag,

PID cascade : Kc = k10
 1+  ; I = 4(c + ); D = 2 (47)

c

If the time constant 2 for the lag is small, then one may approximate the process as k 0 e (+2 )s =s and
derive a PI-controller by using the rules for the integrating proces with delay in (43), but with  replaced by
+
 2 .
If the time constant 2 for the lag is large, such that we in effect have a double integrating process, then
the load response is poor, and the controller needs (47) to be modified. This is discussed next.

4.4 Double integrating process

s
( ) = k00 es2
g s (48)

SIMC tunings. By letting 2 ! 1 and introducing k 00 = k 0 =2 the PID-controller (47) obtained for the
process (46) approaches a PD-controller with

Kc = k100
4( 1+ )2 ; D = 4(c + ) (49)

c

This controller gives good setpoint responses for the process (48), but results in steady-state offset for load
disturbances occuring at the input, see Figure 10. To remove this offset, we need to reintroduce integral
= 4( + )
action, and as before propose to use I c  . With the choice c =
 the resulting SIMC-PID
parameters are
PID cascade : Kc = 0:0625
k 00
1;
2
I = 8; D = 8 (50)
It should also be noted that derivative action is required in order to stabilize this process if we use integral
action in the controller.
ZN tunings can not be derived for this process because we get sustained oscillations with P-control even
with Kc =0 .

5 The half rule: Obtaining the effective delay

In this paper we base the process information on a first-order or second-order plus delay process. This
may seem restrictive, but more complex models can be handled by approximating the remaining high-order
dynamics by an effective delay.
The problem of obtaining the effective delay can be set up as a parameter estimation problem, for
example, by making an least squares approximation of the open-loop step response. However, our goal is
to use the resulting effective delay to obtain controller tunings, so a better approach would be to find the
approximation which for a given tuning method results in the best closed-loop response (here “best” could,
for example, by the in terms of the minimum integrated absolute error (IAE)).
However, our objective is not “optmality” but “simplicity”, so we choose to use a much simpler approach
where we simply add all the “neglected” small time constants to the effective delay, except for the largest
which we distribute evenly to the delay and the time constant using the “half method”. This extremely
simple rule has been applied to numerous examples, and leads to very good final PID tunings.

18
 3

SIMC−PI
2.5

2
OUTPUT y

1.5 SIMC−PID

0.5

0
0 10 20 30 40 50 60 70 80
time

Figure 10: Responses for double integrating process, g s ( )=e s

=s2 .
SIMC-PID with Kc : , I = 0 0625 = 8
and D . =8
SIMC-PD with Kc : = 0 0625 =
, I 1 and D . =8
Disturbance of magnitude 0.1 occurs at t = 40.

The starting point is a model on the following standard form

Qm Qm
g0 s( )= o s j =1 Tj 0 s
Qn
( + 1) =
k 0 e o s j =1 Tj 0 s
Qn
( + 1)
i=1 (i0 s + 1) s + 1=10 i=2 (i0 s + 1)
ke (51)

where Tj 0 are the time constants for overshoot (or inverse response for the case when Tj 0 is negative), and
i0 are the lag time constants. We want to approximate (51) by a first- or second-order plus delay model,

s s
( ) = k ( s + 1)(
g s
e
 s + 1)
= k0
e
(s + 1= s)( s + 1) (52)
1 2 1 2
Here  is the “effective” delay, and we select 2 = 0 if a first-order approximation is desired.
Rules for obtaining the effective delay
1. Approximation of lags i0 . The largest of the neglected time constants i0 is evenly distributed to the
remaining time constant and to the delay (“the half rule”), whereas all the smaller time constants are
added to the delay.
That is, to obtain a first-order model (2 = 0) we choose
20 20 X
1 = 10 + ;  = 0 + + i0
2 2 i3
and to obtain a second-order delay model we choose
X
1 = 10 ; 2 = 20 + 230 ;  = 0 + 230 + i0
i4

19
 2. Approximation of small or negative Tj 0 as effective delay. Let 1 be the effective delay obtained so
= 2
far, and consider a numerator term with Tj 0 T . For T < 1 = (approximately) we simply subtract
T from the delay
 = 1 T

A special case is a process with an inverse reponse (T negative), which then yields a larger effective
delay (for example, the term ( 3 + 1)
s gives an effective delay  =3)

3. Cancellation of large Tj 0 by reducing time constant. For T > 1 =2 (i.e. for large positive values
of T ) we cannot subtract T from the delay, so we instead cancel it be subtracting it from a larger time
+1 
constant, e.g. T ss+1 1
( T )s+1 .
The rules are best understood by considering some examples. Simulations show that the subsequent
application of the SIMC tuning rules result in good responses in all cases.
Example 1. The second-order process

g0 s
1
( ) = 20 (10s + 1)(s + 1)
(53)

= 0) with
is approximated as a first-order delay process (2

1 = 10 + 0:5 = 10:5;  = 0:5

The corresponding SIMC-PI controller tunings are Kc = 0k:5  = 0:525 and I = 8 = 4. The model (53)
1

is already second-order and the SIMC-PID tunings give I = 10 and D = 1, and since there is no delay we
may in theory use an infinite controller gain and achieve perfect responses (and perfect stability margins).
However, in practice Kc will be limited due to uncertainty, unmodelled dynamics and limited input usage.
Example 2. The process

g0 s( ) = k (2s + 1)(1s +( 1)(0

0:3s + 1)(0:08s + 1)
:4s + 1)(0:2s + 1)(0:05s + 1)3
(54)

is approximated as a first-order delay process with

1 = 2 + 1=2 = 2:5;  = 1=2 + 0:4 + 0:2 + 3
0:05 + 0:3 0:08 = 1:47

or as a second-order delay process with

1 = 2; 2 = 1 + 0:4=2 = 1:2;  = 0:4=2 + 0:2 + 3
0:05 + 0:3 0:08 = 0:77

The corresponding tuning parameters for this process are given in Table 3. The responses with the SIMC
tunings are very good as shown in Figure 11.

Kc k
I D GM PM Ms Mt
SIMC-PI 0.85 2.5 0 3.37 57.9o 1.66 1.04
SIMC-PID 1.30 2 1.2 2.84 57.5o 1.74 1.05
ZN-PID 2.56 2.65 0.66 1.84 30.8o 1.79 2.13

Table 3: Example 2. Tunings for process (54)

20
 SIMC−PI
1.5
ZN SIMC−PI
SIMC−PID
1
ZN
OUTPUT y

0.5
SIMC−PID

−0.5
0 5 10 15 20 25 30 35 40

4
ZN
3
INPUT u

−1
0 5 10 15 20 25 30 35 40
time

Figure 11: Example 2. Responses for process (54) with tunings from Table 3

Example 3. The process

( ) = k (10(6ss ++ 1)(8
g0 s
1)(3s + 1)e 0:3s
s + 1)(s + 1)
(55)

is approximated as a first-order delay process with

= 10 + 8 6 3 + 0:5 = 9:5;  = 0:5 + 0:3 = 0:8

1

The corresponding SIMC-PI controller tunings are Kc = 5:94=k and I = 6:4. A second-order model (and
thus the use of a PID controller) is not recommended here because the initial response is overall first order
(with a pole excess of one).
Example 4. The process
g0 s( ) = k ( s 1+ 1)4 (56)
0
is approximated as a first-order delay process with

1 = 1:50 ;  = 2:50
or as a second-order delay process with (here we interchange 1 and 2 since we want 1 > 2 )

1 = 1:50 ; 2 = 0 ;  = 1:50
The corresponding SIMC PI- and PID-controller tunings are given in Table 4. In this case 2 <  and the
use of derivative action has little effect.
Example 5. The process (Astrom et al. 1998)

( ) = (s + 1)(0:2s + 1)(0:104s + 1)(0:008s + 1)

g0 s (57)

21
 Kc k
I =0 D =0 GM PM Ms Mt
SIMC-PI 0.3 1.5 0 4.95 62.4o 1.46 1.00
SIMC-PID 0.5 1.5 1 6.73 62.5o 1.43 1.00

Table 4: Example 4. Tunings for process (56)

is approximated as a first-order delay process with

1 = 1:1;  = 0:148
or as a second-order delay process with
1 = 1:0; 2 = 0:22;  = 0:028
The corresponding tunings are given in Table 5.
As seen in Figure 12 the Ziegler-Nichols tunings almost give instability for this process, whereas the
SIMC tunings give nice closed-loop responses. A PID controller gives a significant improvement for this
process, which is expected since for the process is dominant second order with 2 : much larger than= 0 22
= 0 028
: .
Kc I D GM PM Ms Mt
SIMC-PI 3.72 1.1 0 6.69 51.1o 1.59 1.16
ZN-PI 13.6 0.47 0 1.30 5.5o 11.3 10.9
Astrom-PI(Ms = 2) 4.13 0.59 0 4.90 35.1o 2.00 1.66
SIMC-PID 17.9 1.0 0.22 7.54 50.9o 1.58 1.16

Table 5: Example 5. Tunings for process (57)

Example 6. The process (Astrom et al. 1998)

( ) = s(s +(01):172(0s :+0281)s + 1)

2
g0 s (58)

s
is approximated as an integrating process, e =s, with

 = 2
1 + 0:028 2
0:17 = 1:69

or as an integrating process with lag, e s =s(2 s + 1), with
2 = 1 + 1=2 0:17 = 1:33;  = 1=2 + 0:028 0:17 = 0:358

The corresponding SIMC PI- and PID-controller tunings are given in Table 6. The corresponding closed-
loop responses (Figure 13) are again very good, especially for the PID-controller.
In summary, these examples illustrate that the simple SIMC tuning rules used in combination with
the simple half-rule for estimating the effective delay, result in good and robust tunings. The method for
approximating a first-order with delay model (“half rule”) and the PID tuning rules are not “optimal” in any
mathematical sense, but they are simple and give surprisingly good robust performance. Furthermore, the
reason for using a PID controller is simplicity, and if high performance control is desired, then one would
not use PID control in the first place.
A large number of additional comparisons have been performed, and there has been few cases (if any)
where the proposed SIMC tuning rules perform poorly. In cases where there were large differences, the
SIMC tunings could usually be improved by adjusting the tuning parameter c .

22
 2

ZN−PI
1.8

1.6
SIMC−PI
Astrom−PI
1.4

1.2
OUTPUT y

SIMC−PID
0.8

0.6 SIMC−PI

ZN−PI
0.4
Astrom−PI (M =2)
s
0.2
SIMC−PID

0
0 2 4 6 8 10 12 14 16 18 20
time

Figure 12: Example 5. Responses for process (57) with tunings from Table 5
Load disturbance of magnitude 2 occurs at t = 10.
Kc I D GM PM Ms Mt
SIMC-PI 0.296 13.52 0 16.6 48.8o 1.48 1.29
SIMC-PID 1.397 2.894 1.33 1 52.4o 1.23 1.30
Astrom (Ms = 2) 0.47 7.01 0 8.2 33.1o 2.00 1.77

Table 6: Example 6. Tunings for process (58)

6 Insights
6.1 Guidelines for retuning
The tuning rules presented in this paper, see (3)-(5), give invalueable insights, for example, into how we
must change the tuning parameters in response to changes in the process model:

1. An increase in the process gain k is counteracted by reducing the controller gain Kc such that Kc k
remains constant. (The integral time is kept constant, and the closed-loop response will remain un-
changed unless there is also a change in the disturbance transfer function).

2. An increase in the process time constant 1 is counteracted by increasing Kc such that Kc =1 remains
constant. For a “fast” process where we use I =
1 , we also need to incrase the integral time (the
closed-loop response will then remain unchanged). For a “slow” process where we use I = 4( + )
c  ,
we keep I unchanged (but the closed-loop response will change somewhat in this case).

3. In many cases there is a direct correlation between the gain and the time constant such that the initial
=
slope k 0 k=1 remains constant. In this case we should keep Kc constant. For “fast” processes
=
where we use I 1 we should increase the integral time. For “slow” processes where we use
I = 4( + )
c  we should keep the integral time constant.

23
 SIMC−PI

OUTPUT y 2

1
SIMC−PID

0
0 5 10 15 20 25 30 35 40

1.5

1
INPUT u

0.5

−0.5

−1
0 5 10 15 20 25 30 35 40
time

Figure 13: Example 6. Responses for process (58) with tunings from Table 6
Load disturbance of magnitude 2 occurs at t = 10.

4. Note that for a “slow” process, the tunings only depend on the initial response as expressed by k 0 =
k=1 and  , whereas for a “fast” process the steady-state gain k is also of importance.

5. An increase in the delay  is counteracted by a corresponding decrease in Kc in order to maintain the

=
same robustness. For “fast” processes with I 1 the integral time is kept unchanged, whereas for
“slow” processes with I = 4( + )
c  the integral time is increased.

6. For a second order process the derivative time increases when the second order time constant 2 is
increased.

When retuning the controller based on experimental responses the following guidelines for PI control
may prove helpful. The basis for these guidelines is the disturbance response.

1. If the maximum output deviation is too large then the controller gain should be increased - recall (37).

2. If the settling time is too large then the integral time should be reduced.

3. If the oscillations are too large and these have a period shorter than the integral time I , then the gain
should be reduced or the integral time increased - recall Figure 7.

4. If the oscillations are too large and these have a period more than about three times longer than the
integral time I ), then the iproduct of the controller gain and integral time should be increased, recall
(29).

6.2 Retuning the controller for integrating process

Many control loops for integrating processes, including many liquid level control systems, have oscillations
because the controller gain is too low, or alternatively, the integral time is too short. Here we show how to
retune the controller in such cases.

24
 Consider a PI controller with (initial) tunings Kc0 and I 0 which results in “slow” oscillations with
3
period P0 (by slow we mean that P0 is larger than about I 0 ). Then we most likely have an integrating
process
s
g s ( ) = k0 e s
for which the product of the controller gain and integral time (Kc0 I 0 ) is too low. Assuming  2 << 1
(significant oscillations), (31) gives the following approximate expression for P0
s
I 0 1
2
P0  0  =20 k Kc0
(59)

Thus, from (59) the product of the controller gain and integral time is approximately
 2
Kc0 I 0 = (2)2 k10 I 0
P0

To avoid oscillations (  1) we must from (30) require

Kc I  4 k10
that is, we must require that
Kc I 1  2
 2
P0 (60)
Kc0 I 0 0
i

1
Here = 2  0:10, so we have the rule:
To avoid “slow” oscillations the product of the controller gain and integral time should be
0 1(
increased by at least a factor f  : P0 =I 0 2 . )
The application of this simple rule should guarantee you immediate success and respect among plant oper-
ators.
Example. A real industrial case study of a reboiler level control loop is shown in Figure 14. Here
y is the reboiler level and u is the bottoms flow valve position. The PI tunings had been kept at their
default setting (Kc = 05 : and I =1 min) since start-up several years ago, and resulted in an oscillatory
response as shown in the top part of the Figure. The control of the level (y ) itself was acceptable, but the
bottoms flowrate (input u) showed large variations, and because it is the feed to the downstream column
this caused poor temperature control in the downstream column.
From a closer analysis of the “before” response we find that the period of the slow oscillations is
P0 = 0 85: h = 51 min. Since I =1 min, we get from the above rule we should increase Kc
I by a
factor f= 0 1 (51) = 260
:
2 to avoid the oscillations. The plant personnel were somewhat sceptical to
authorize such large changes, but eventually accepted to increase Kc by a factor 7.7 and I by a factor 24,
7 7 24 = 185
that is, Kc I was increased by :
. The much improved response is shown in the “after” plot in
Figure 14. There is still some minor oscillations, but these may be caused by disturbances outside the loop.
In any case the control of the downstream distillation column was much improved, and the plant personnel
were very impressed by what the fresh engineer had learned in her control course in Trondheim.

7 Conclusion
The first step is to approximate the process as a first or second order process with effective delay, and the
half rule is simple to use and gives good results. Based on this model with parameters k; 1 and  , the

25
 Figure 14: Industrial case study of retuning reboiler level control system

following SIMC tunings are suggested:

Kc = k10  1+  ; I = minf1; 4(c + )g

c

If the process is second order (with 2 >  , approximately) and derivative action is acceptable we choose

Cascade PID : D = 2
The parameter c is the only tuning parameter, and a reasonably fast response with good robustness is
=
obtained with c  . This gives robust (conservative) tunings when compared with most other tuning rules.
If the response is too slow, then one may decrease the value of c , and possibly further reduce the integral
time.
However, one may also want to increase c to get a slower and smoother response. This results in a
smaller controller gain Kc , but we must require

Kc  du=ymax
(approximately) in order to keep the output deviation less than ymax in response to a load disturbance of
magnitude du .

Acknowledgement
Discussions with Professor David Clough from the University of Colorado at Boulder are gratefully acknowledged.

26
Simulations
In all simulations we have used a cascade PID controller with derivatice action effective over one decade ( = 0:1)
and without taking the derivative of the setpoint:
I s + 1  D s + 1 
u(s) = Kc ys(s) y(s)
D s + 1
(61)
I s
However, note that stability margins etc. are computed with  = 0. In most cases we use a PI controller, that is
D = 0, and the controller becomes
I s + 1
u(s) = Kc (ys(s) y(s)) (62)
I s
 Zt 
1
or in the time domain

u(t) = Kc (ys(t) y(t)) + I 0 (ys( ) y( ))d (63)

In the simulations a unit setpoint change ys = 1 is introduced at time t = 0, and an input “load” disturbance of
magnitude du = 0:5 occurs at t = 20 (unless otherwise stated).

References
Astrom, K.J., H. Panagopoulos and T. Hagglund (1998). Design of PI controllers based on non-convex
optimization. Automatica 34(5), 585–601.

Astrom, K.J., T. Hagglund, C.C. Hang and W.K Ho (1993). Automatic tuning and adaptation for PID
controllers - A survey. Control Eng. Practice 1(4), 699–714.

Chien, I.L. and P.S. Fruehauf (1990). Consider IMC tuning to improve controller performance. Chemical
Engineering Progress pp. 33–41.

Cohen, G.H. and G.A. Coon (1953). Theoretical consideration of retarded control. Trans. ASME 75, 827–
834.

Ho, W.K., K.W. Lim and W. Xu (1998). Optimal gain and phase margin tuning for PID controllers. Auto-
matica 34(8), 1009–1014. See Automatica.

Holm, O. and A. Butler (1998). Robustness and performance analysis of PI and PID controller tunings.
Technical report. 4th year project, Department of Chemical Engineering. Norwegian University of
Science and Technology, Trondheim. http://www.chembio.ntnu.no/users/skoge/diplom/reports/pid98-
holm-butler/.

Horn, I.G., J.R. Arulandu, J. Gombas, J.G. VanAntwerp and R.D. Braatz (1996). Improved filter design in
internal model control. Ind. Eng. Chem. Res. 35(10), 3437–3441.

Marlin, T.E. (1995). Process Control. McGraw-Hill.

Rivera, D.E., M. Morari and S. Skogestad (1986). Internal model control. 4. PID controller design. Ind.
Eng. Chem. Res. 25(1), 252–265.

Schei, T.S. (1994). Automatic tuning of PID controllers based on transfer function estimation. Automatica
30(12), 1983–1989.

Seborg, D.E., T.F. Edgar and D.A. Mellichamp (1989). Process Dynamics and Control. John Wiley & Sons.

27
Shinskey, F.G. (1998). Personal communication.

Smith, C.A. and A.B. Corripio (1985). Principles and Practice of Automatic Process Control. John Wiley
& Sons.

Smith, O.J. (1957). Closer control of loops with dead time. Chem. Eng. Prog. 53, 217.

Tyreus, B.D. and W.L. Luyben (1992). Tuning PI controllers for integrator/dead time processes. Ind. Eng.
Chem. Res. pp. 2628–2631.

Ziegler, J.G. and N.B. Nichols (1942). Optimum settings for automatic controllers. Trans. of the A.S.M.E.
64, 759–768.

28