&VSPQFBO
$POUSPM
$POGFSFODF

&$$


o
4FQUFNCFS

$BNCSJEHF
6,

SYNTHESIS OF ROBUST PID CONTROLLERS FOR TIME

DELAY SYSTEMS
N. Hohenbichler , J. Ackermann

Institute of Automatic Control (IRT), RWTH Aachen University, 52056 Aachen, Germany,
[email protected]

Institute of Robotics and Mechatronics, German Aerospace Centre (DLR) Oberpfaffenhofen, P.O. Box 1116, 82230 Weling,
Germany, [email protected]
Keywords: PID control, time delay system, robust control, parameter space approach, quasipolynomials.

in [11] formulas to calculate some stability boundaries in the

dead time for certain classes of quasipolynomials are given.
However, both approaches are limited to very restricted cases.

Abstract
Tuning rules are frequently used to choose the parameters of
PID controllers. However, they are often based on heuristics,
are limited to systems of a certain class, use dead time
approximations and ignore parameter uncertainties in the
modeled system. This work develops a systematic, universal
and transparent method to design a robust PID controller based
on the parameter space approach, which is extended to cope
with quasipolynomials.

Introduction

2 Problem formulation
Consider a single loop containing a PID controller and a linear
time delay system (see Fig. 1), given by the transfer functions
KI + KP s + KD s2
,
s (1 + TR s)
A(s, q) sL
G(s, L, q) =
e
,
R(s, q)
PID(s, k) =

(1)
(2)

where k = (KI , KP , KD )T are the controller parameters. (TR

ensures feasibility of the controller and lters noise ; it is assumed to be xed prior to the controller design, e.g. by adding
a non dominant pole to R(s, q)). The unknown but constant
plant parameters are the dead time L > 0 and the parameters
in the vector q. They lie in an operation domain

PID is the most common controller in industrial practice [2].

But a high percentage of PID control systems seems to be tuned
badly [6]. One major reason may be that todays tuning methods are limited to very restricted conditions on the plant (concerning model order, pole and zero location, neglected parameter uncertainty) [12]. Time delay systems, especially with uncertain and immeasurable dead time, present one of the most
challenging problems for tuning a PID controller.

where qi and qi+ are specied as the lower and upper limits of
parameter qi in q (analog L and L+ ).

This paper develops a PID tuning method based on the parameter space approach [1]. So far, in [3] the synthesis step is
extended to time delay systems, but important results for the
practical application are still missing. Also, the analysis step is
not developed in the literature and results have not been compared with existing tuning methods.

The problem of designing a robust PID controller is to nd a set

of controller parameters k = k , that meets the specication
for all values of (L, q)T Q. Specications are assumed in the
form of Hurwitz stability (all roots of the characteristic function
are in the open left half plane (LHP)) and -stability (all roots
have a real part smaller then a real number 0 ).

In [15] an analytical solution to nd the Hurwitz stable PID

parameters for a rst order system plus time delay is published;

The characteristic function of the loop in Fig. 1

Q = { (L, q)T | L [L ; L+ ], qi [qi ; qi+ ] },

P (s, k, L, q) = (KI + KP s + KD s2 ) A(s, q)+

+ s (1 + TR s) R(s, q) esL ,




PID(s, k)

A(s,q)
R(s,q)

sL

Figure 1: Single loop with PID controller and time delay system

*4#/


(4)

B(s,q)

d
r

(3)

with polynomials
A(s, q) = a0 (q)+a1 (q) s + . . . + am (q) sm ,
B(s, q) = b0 (q) +b1 (q) s + . . . + bn (q) sn ,

(5)
(6)

with am (q) = 0, bn (q) = 0 belongs to the class of quasipolynomials [5, 13] due to the dead time. (Note that b0 (q) = 0 for



basic case of a PID controller (1). However, later a b0 (q) = 0

may appear through transformations, see section 7.)

5. Compute the -stable region in plant parameter space for

the candidate controller.

The principal term condition [13] requires for Hurwitz stability

that in the case of PID control (KD = 0) the degrees full
n m + 2. In the sequel we treat only this case (i.e. we
assume a proper A(s, q)/R(s, q) for TR = 0.)

6. If the Q-domain lies entirely in the -stable region, then

the problem is solved.

Parameter space approach

The parameter space approach is used to solve the problem in

two main steps. In the controller synthesis step, we compute
the stable (either Hurwitz or -stable) region in the space of
controller parameters k for several representatives (L , q )T
out of Q (usually the vertices). A candidate for a robust controller k is chosen from the intersection of stable regions.
This controller satises the specication for the representatives.
The second step, the control loop analysis, is applied to test
the robust stability for the continuum of all values in Q. Now
we compute the stable region in the space of plant parameters
(L, q)T with xed controller k . If Q lies entirely in the stable
region, then a solution of the problem is found.
The calculation of a Hurwitz stable region in a parameter space
(either k or (L, q)T ) is based on the fact that the roots of
the quasipolynomial (4) with continuous coefcient functions
ai (q), bi (q) do not jump when the parameters are changed continuously. Thus, a stable quasipolynomial, whose roots all lie
in the LHP, becomes unstable if and only if at least one root
crosses the imaginary axis. The parameter values of the root
crossings form the stability boundaries in the parameter space,
which can be classied into three cases: the real root boundary (RRB), where a root crosses the imaginary axes at the origin (substitute s = 0 in the quasipolynomial), the innite root
boundary (IRB), where a root leaves the LHP at innity (set
|s| ) and the complex root boundary (CRB), where a pair
of conjugate complex roots crosses the imaginary axes (substitute s = j and sweep over all real > 0). These stability
boundaries separate different regions in the parameter space.
To classify a region as Hurwitz stable it sufces to prove stability for one inner test point (e.g. by the Nyquist criterion).

-stability can be reduced to the Hurwitz case by the substitution s = v + 0 which leads to a transformation in parameters
and polynomials (see section 7). So Hurwitz stability is considered rst in the next paragraphs.

5 Controller synthesis
For each xed representative (L , q )T the Hurwitz stability
boundaries of (4) in the k-space are determined. The RRB
turns out to be simply a straight line given by the equation
P (0, k) = KI A(0) + B(0) = 0 KI =

b0
.
a0

(7)

(In the basic case we have b0 = 0 and the RRB is KI = 0.)

More theoretical difculties arise when calculating the IRB.
Quasipolynomials possess an innite number of roots, which
can not be calculated analytically in the general case. However, the asymptotic location of roots far from the origin is well
known [5]. It turns out that innite root boundaries only exist, if the degree equation n = m + 2 is fullled (in case of
KD = 0). These are two straight lines
KD =

bn
.
am

(8)

The calculation of the CRB starts analog to the delay free case
of polynomials [1]. The root condition P (j, k) = 0 can be
separated into a system of two equations for real and imaginary
part


RP (, k)
IP (, k)



KI
RA 2 RA
+
IA 2 IA
KD
  

0
RB KP IA
=
,
+
0
IB + KP RA


(9)

Controller design algorithm

The proposed controller design procedure is summarized in the

following steps:
1. Specify the maximum real part from closed-loop settling time requirements.
2. Compute the -stable regions in controller parameter
space for representatives (usually the vertices) of the Qdomain.
3. Determine the intersection of the -stable regions.
4. Choose a candidate controller out of the intersection.

where B(j)
= B(j) ejL and R, I denote the real and
and P at (j).
imaginary parts of A, B
Clearly, the matrix multiplying (KI , KD )T is singular. Thus,
the key idea is to x KP = KP and to evaluate the CRB in the
(KD , KI )-plane. A solution of (9) exists and only exists for
the real zeros gi of


RA RB KP IA
g() = det
IA IB + KP RA
(10)

2
2
= KP (RA + IA ) + RA IB IA RB .
The zeros of g() are called singular frequencies. For each i
appears a straight line as CRB in the (KD , KI )-plane, ruled by
the equation
2
KI = gi
KD + KI0 (g ),
(11)



20

5
RRB
IRB
CRB

15

4
10

5
K ()
P

K =1

KI

10

P1

15

0
20

10

12

14

16

18

20

1
1.5

Figure 2: Function KP () (solid), its limit function (dashed),

singular frequencies for KP = 1 (x-marks) and stabilizing
KP -interval (dash-dotted) of G1 (s).

where KI0 (g ) can be easily determined from the rst or second row of (9).
Thus, the stability boundaries RRB, IRB and CRB are straight
lines in the (KD , KI )-plane and partition the plane into convex
polygons. (Additionally, for each boundary line the side can be
determined which possesses the lower number of stable poles,
see [3].)
The singular frequencies may be determined by a graph of
KP () =

RA IB + IA RB
.
2 + I2 )
(RA
A

(12)

Graphically, the singular frequencies for a xed KP are the

abscissa values of the intersections between the KP ()-plot
and the (KP = KP )-line. Due to the dead time the number of
singular frequencies is innite. Algorithms for the automatic
calculation of the singular frequencies can be found in [8].

0.5

0
KD

0.5

1.5

Figure 3: Stability boundaries in (KD , KI )-plane for KP = 1

of G1 (s). The side of the lines with more unstable poles is
shaded.

frequencies which lie in a frequency range where (14) and

(12) are practically identical result in CRB lines which do
not contribute to boundaries of the stable polygon. So, it
sufces to compute and evaluate only a nite number of
low singular frequencies, given by the comparison of (14)
and (12).
2. The question of determining the interval of KP which
leads to a stable polygon is analytically solved only for
a rst order system plus time delay [15]. Following work
hypothesis for the general system (2) is stated:
Work hypothesis.
Provided that there is a stabilizing PID controller, a Hurwitz stable polygon exists in the
(KD , KI )-plane, if KP lies in the interval of KP that produces the highest possible number of singular frequencies.
The result of the hypothesis for G1 (s) is depicted in Fig. 2.

The function KP () and the resulting boundaries in the

(KD , KI )-plane for a xed KP = 1 are demonstrated in Fig. 2
and Fig. 3 for the example system (with ideal PID controller
TR = 0)
1
G1 (s) =
es .
(13)
s+1

The entire stable region in k-space can be computed by gridding KP in its stabilizing interval and extracting the stable
polygon for each gridded KP (see Fig. 4).

Stability checks of test points prove that the polygon around

P1 is stable (and it is the only one). Two questions arise: How
many singular frequencies have to be evaluated and which KP values lead to a stable polygon?

Now we determine the Hurwitz stable region in plant parameter

space by mapping the stability boundaries of the quasipolynomial
(15)
P (s, L, q) = D(s, q) + B(s, q) esL ,

6 Control loop analysis

where
1. For high frequencies, function (12) tends to a much simpler limit function of the form

KP () trig(L ),

(14)

where and are real constants ( 1) and trig is either

the sin or cos function. It can be shown, that singular

2
s ) A(s, q)
D(s, q) = (KI + KP s + KD

= d0 (q) + d1 (q) s + . . . + dm (q) sm

(16)

The case of mapping the boundaries to a plane of L and one

additional parameter q is treated. If there are more uncertain
parameters, then they have to be gridded.



1.5

P2

0.5

K 1
P

P1

stable

0
0.5

k=0
k=1
k=2
k=3
k=4

1
0

0.5

0
1

1.5
KI

2.5

1
3.5

1.5
5

10

15

20

25

30

Figure 4: Hurwitz stable controller parameters k of G1 (s).

Figure 5: Stability boundaries RRB (dashed) and CRB (solid,
k = 0, .. , 4) of P2 (s, L, K).
The RRB are lines q = q , where q are the solutions of the
equation
d0 (q) + b0 (q) = 0.
(17)


If the degrees of B(s, q) and D(s, q) fulll n = m , then the

solutions of the following equation lead to IRB
dm (q) bn (q) = 0.

(18)

The CRB condition P (j, L, q) = 0 can be split into a modulus

and a phase equation


 D(j, q) 


(19)
 B(j, q)  = 1,
1
L(, q, k) = (arg(D(j, q)) arg(B(j, q))+

+ (2k 1)), k . (20)

The modulus equation does not depend on L and can be transformed to a polynomial in 2 . Thus, the equation can be solved
by nding the roots of a polynomial, but it is especially convenient if (19) can be solved symbolically for q (e.g. if q is
the dc-gain). With (20) the corresponding values for L can be
added to plot the CRB curves. The phase equation carries the
uncertainty k (only k 0 generate nonnegative L), so there is
an innite number of CRB. However, as k grows, the curves
move due to (20) to higher dead times.
Fig. 5 shows the boundaries and the stable region for the example quasipolynomial
P2 (s, L, K) = K + (s2 + 0.25 s + 1) esL .

(21)

In [11] this example serves for the calculation of some dead

time stability limits (corresponding to k = 1), which agree
with these presented here. However, Fig. 5 presents the stable
region completely for L  22.

7 PID design using -stability

The concept of -stability can be used to speed up the transient
responses robustly. Concerning the synthesis step, this case
can be reduced to the Hurwitz case by substituting s = v + 0 .
Following transformations result
KI = KI + KP 0 + KD 0 2 ,
KP = KP + 2KD 0 ,

KD
= KD ,

(22)

A (v) = A(v + 0 ),

B  (v) = e0 L B(v + 0 ).
Following the parameter space approach presented in section 3,
the smallest possible value for 0 is determined for a given system (2) and an operation domain (3), s. t. the intersection of the
stable regions belonging to the vertices of Q is not empty. This
0 is approximated by an iterative approach: Beginning with
zero, 0 is stepwise reduced and the function KP () is plotted for each vertex, until the work hypothesis reveals that there
is no interval of KP that stabilizes simultaneously all vertices.
With the last 0 having such a interval, the stable k-regions are
computed for all vertices, and a controller k is taken out of the
intersecting region.
The analysis step can be reduced to he Hurwitz case, if the second parameter q enters in form of a dc-gain into the quasipolynomial. In that case, the transformations are
q  = q e0 L ,
L = L,

D (v) = D(v + 0 ),

B  (v) = B(v + 0 ).

(23)



Robust PID controller

1.4

1.6

Modulus optimum
Chien track. aperiodic
Latzel 20%
Latzel 10%
Ziegler oscillation
Tsumrule normal
Chien track. 20%
Chien disturb. aperiodic
Pessen
Chien disturb. 20%
Cohen
Ziegler step response

1.2
1
y(t)

1.4

0.8
0.6
L=1.0 ISE=1.34
L=0.7 ISE=1.11
L=1.3 ISE=1.64

0.4
0.2
0
0

1.2

Robust IMC controller

0.8

1.4
1.2
1
y(t)

0.6

0.2
0
0

0.4
0.06

L=1.0 ISE=1.17
L=0.7 ISE=0.946
L=1.3 ISE=1.55

0.4

0.04

Figure 6: Step responses of PID and IMC controlled loop of

G3 (s, L) for nominal, maximal and minimal L.

0.6

0.8

0.02

10

15

KD

Figure 7: Intersection (I ) of the -stable regions (0 =

0.05) of the vertices of Q4 for G4 (s, L, K) and tuning rule
controllers. The tuning rules are sorted by the settling time after a reference step for the nominal plant, see [14, 4].

Application to literature examples

We apply the proposed method to design PID controllers for

example systems of the literature and compare the results with
other PID tuning methods or control concepts. Whenever possible, we respect the following comparison laws: Copy the
plant model and the controller with all parameters. Prefer robust control methods and use the same operation domain. Apply the same criteria to evaluate the results. Take only literature
examples whose results can be reconstructed.
First the comparison is illustrated by the following system and
operation domain
G3 (s) =

1
esL ,
1+s

Q3 : L [0.7 ; 1.3].

(24)

which serve as an example in [10] to develop an Internal Model

Control (IMC) controller.
The robust PID controller parameters are KI = 0.58, KP =
0.82, KD = 0.26 with TR = 0.1 and 0 = 0.8. For a
comparison, in Fig. 6 the step responses for nominal, maximal
and minimal dead times are shown for the robust PID and IMC
controller. Clearly, by using the PID controller the overshoot,
settling time and oscillations are reduced, while the Integral
Square Error (ISE) is slightly higher.

in Fig. 7. (The tuning rules are applied to the nominal plant

L = 15, K = 1). Clearly, the the tuning rules leading to
the lowest settling time lie close to the intersection region. In
[14, 4] it is stated that it depends heavily on the plant which
tuning rule produces the fastest settling time. However, the
proposed method nds a fast controller universally for all systems in the class (2) and guarantees robustness in the whole
operating domain.
Additionally, in [8] the proposed method is applied to a great
variety of systems (up to fourth order, stable and unstable,
with real and complex poles, with zeros) in different parameter
dependencies, which serve as benchmark problems for single
loop control strategies in the literature (IMC [10], tuning rules
etc. [14, 4, 9] and genetic optimization [7]). In all considered cases we achieve superior or at least similar results: The
overshoot of the robust PID controller is low, the transient responses are robustly fast, there are only small oscillations and
the responses for different operation points resemble largely.

9 Conclusions

(25)

The parameter space approach offers convincing results in the

synthesis of robust PID controllers for time delay systems. The
developed tuning method is systematic, universal and transparent and leads to superior or similar results than literature examples. Exact stability (Hurwitz or -stability) regions can be determined in the space of controller and plant parameters while
treating the dead time without approximation.

is used to compare the proposed method with tuning rules

that are applied to the system in [14, 4]. Notice the intersections of the -stable regions and the tuning rule controllers

The development of an interactive graphical software package

based on the stated algorithm seems very promising to be a
helpful tool in daily engineers work. So the engineer would be

The second example,

G4 (s, L, K) =

K
esL ,
(1 + 5s)3
Q4 : L [12 ; 18], K [0.9 1.1],



able to re-tune the great amount of existing PID loops at low

cost in industry.

[14] T. Schaar. Vergleich von verschiedenen Einstellregeln fur

PID-Regler. Fachhochschule Koln, Fachbereich Elektrische Energietechnik, 1998. Diplomarbeit.

References

[15] G. J. Silva, A. Datta, and S. P. Bhattacharyya. New results

on the synthesis of PID controllers. IEEE Transactions on
Automatic Control, 47(2):241252, Februar 2002.

[1] J. Ackermann, P. Blue, T. Bunte, L. Guvenc, D. Kaesbauer, M. Kordt, M. Muhler, and D. Odenthal. Robust
Control. Springer, New York, 2002.
om and T. Hagglund. PID Controllers: Theory,
[2] K. Astr
Design and Tuning. Instrument Society of America, 1995.
[3] N. Bajcinca, R. Koeppe, and J. Ackermann. Design of
robust stable master-slave systems with uncertain dynamics and time-delay. In Proceedings IFAC (International
Federation of Automatic Control) 15th World Congress,
Barcelona, 2002.
[4] N. Becker, W. M. Grimm, and U. Piechottka. Vergleich
verschiedener PID-Regler. atp, (41):3946, 1999.
[5] R. E. Bellman and K. L. Cooke. Differential-Difference
Equations. Academic Press, New York, 1963.
[6] David B. Ender. Process control performance: Not as
good as you think. Control Engineering, page 180,
September 1993.
[7] K. Hirata, Y. Yanase, T. Kawabe, and T. Katayama.
A minimax design of robust I-PD controller for timedelay systems with parametric uncertainty. In Proceedings IFAC (International Federation of Automatic Control) 14th World Congress, volume C, pages 259264,
Peking, 1999.
[8] N. Hohenbichler. Auslegung robuster PID-Regler fur Totzeitsysteme. TU Munchen, Lehrstuhl fur Steuerungs- und
Regelungstechnik, 2002. Diplomarbeit.
[9] Yongho Lee, Jeongseok Lee, and Sunwon Park. PID controller tuning for integrating and unstable processes with
time delay. Chemical Engineering Science, 55(17):3481
3493, 2000.
[10] M. Morari and E. Zariou. Robust Process Control.
Prentice-Hall, Englewood Cliffs, 1989.
[11] S.-I. Niculescu. Delay Effects on Stability. Number 269
in LNCIS. Springer, 2001.
[12] A. ODwyer.
PI and PID controller tuning rules
for time delay processes: A summary.
Technical report, School of Control Systems and Electrical Engineering, Dublin Institute of Technology, 2000.
http://citeseer.nj.nec.com/dwyer00pi.html.
[13] L. S. Pontryagin. On the zeros of some elementary transcendental functions, volume 2, pages 95110. American Mathematical Society Translation, 1955. Englische

Ubersetzung.