ARTICLE IN PRESS

Signal Processing 86 (2006) 2771–2784

www.elsevier.com/locate/sigpro

Tuning of fractional PID controllers with

Ziegler–Nichols-type rules
Duarte Valério,1, José Sá da Costa
Department of Mechanical Engineering, Instituto Superior Técnico, Technical University of Lisbon, GCAR,
Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal
Received 12 April 2005; received in revised form 29 October 2005; accepted 6 December 2005
Available online 9 March 2006

Abstract

In this paper two sets of tuning rules for fractional PIDs are presented. These rules are quadratic and require the same
plant time–response data used by the first Ziegler–Nichols tuning rule for (usual, integer) PIDs. Hence no model for the
plant to control is needed—only an S-shaped step response is. Even if a model is known rules quickly provide a starting
point for fine tuning. Results compare well with those obtained with rule-tuned integer PIDs.
r 2006 Elsevier B.V. All rights reserved.

Keywords: Fractional controllers; PID; Ziegler–Nichols tuning rules

1. Introduction rules is to have a certain time response of the plant.

The output of PID controllers (proportional—
derivative—integrative controllers) is a linear com-
bination of the input, the derivative of the input and
the integral of the input. They are widely used and
enjoy significant popularity, because they are
simple, effective and robust.
One of the reasons why this is so is the existence
of tuning rules for finding suitable parameters for
PIDs, rules that do not require any model of the
plant to control. All that is needed to apply such
Corresponding author. Tel.: +351 218 419 119.
E-mail addresses: dvalerio@dem.ist.utl.pt (D. Valério),
sadacosta@dem.ist.utl.pt (J.S. da Costa).
URLs: http://mega.ist.utl.pt/
dmov, http://www.gcar.dem.
ist.utl.pt/pessoal/sc/.
1
Partially supported by Fundac- ão para a Ciência e a
Tecnologia, grant SFRH/BPD/20636/2004, funded by POCI
2010, POS_C, FSE and MCTES.
Examples of such sets of rules are those due to
Ziegler and Nichols, Cohen and Coon, and the
Kappa–Tau rules [1].
Actually, rule-tuned PIDs often perform in a non-
optimal way. But even though further fine-tuning be
possible and sometimes necessary, rules provide a
good starting point. Their usefulness is obvious
when no model of the plant is available, and thus no
analytic means of tuning a controller exists, but
rules may also be used when a model is known.
Fractional PIDs are generalisations of PIDs:
their output is a linear combination of the input,
a fractional derivative of the input and a fractional
integral of the input [2]. Fractional PIDs are
also known as PIl Dm controllers, where l and m
are the integration and differentiation orders; if
both values are 1, the result is a usual PID
(henceforth called ‘‘integer’’ PID as opposed to a
fractional PID).

0165-1684/$ - see front matter r 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.sigpro.2006.02.020
 ARTICLE IN PRESS
2772 D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784

Even though fractional PIDs have been increas- It is worth noticing that, when n is positive but non-
ingly used over the last years, methods proposed to integer, operator D still needs integration limits c
tune them always require a model of the plant to and x; in other words, D is a local operator for
control [3,4]. (An exception is [5], but the proposed natural values of n (usual derivatives) only.
method is far from the simplicity of tuning rules for The Laplace transform of D follows rules rather
integer PIDs.) This paper addresses this issue similar to the usual ones:
proposing two sets of tuning rules for fractional
L½0 Dnx f ðxÞ
PIDs. Proposed rules bear similarities to the first rule 8 n
proposed by Ziegler and Nichols for integer PIDs, > s F ðsÞ if np0;
<
and make use of the same plant time response data. ¼ n
P k nk1
n1
>
: s F ðsÞ  s 0 Dx f ð0Þ if n  1onon 2 N:
The paper is organised as follows. Section 2 sums k¼0
up the fundamentals of fractional calculus needed to
understand fractional PIDs. Then, in Section 3, two ð4Þ
analytical methods for tuning fractional PIDs when This means that, if zero initial conditions are
a plant model is available are addressed; these are assumed, systems with a dynamic behaviour de-
used as basis for deriving the tuning rules given in scribed by differential equations involving fractional
Section 4. Section 5 gives some examples of derivatives give rise to transfer functions with
application and Section 6 draws some conclusions. fractional powers of s.
Even though n may assume both rational and
2. Fractional order systems irrational values in (4), the names ‘‘fractional’’
calculus and ‘‘fractional’’ order systems are com-
2.1. Definitions monly used for purely historical reasons. Some
authors replace ‘‘fractional’’ with ‘‘non-integer’’ or
Fractional calculus is a generalisation of ordinary ‘‘generalised’’, however.
calculus. The main idea is to develop a functional Thorough expositions of these subjects may be
operator D, associated to an order n not restricted found in [2,6,7].
to integer numbers, that generalises the usual
notions of derivatives (for a positive n) and integrals 2.2. Integer order approximations
(for a negative n).
Just as there are several alternative definitions of The most usual way of making use, both in
(usual, integer) integrals (due to Riemann, Lebes- simulations and hardware implementations, of
gue, Steltjes, etc.), so there are several alternative transfer functions involving fractional powers of s
definitions of fractional derivatives that are not is to approximate them with usual (integer order)
exactly equivalent. The most usual definition is due transfer functions with a similar behaviour. So as to
to Riemann and Liouville and generalises two perfectly mimic a fractional transfer function, an
equalities easily proved for integer orders: integer transfer function would have to include an
Z x infinite number of poles and zeroes. Nevertheless, it
n ðx  tÞn1
D
c x f ðxÞ ¼ f ðtÞ dt; n 2 N, (1) is possible to obtain reasonable approximations
c ðn  1Þ!
with a finite number of zeroes and poles.
Dn Dm f ðxÞ ¼ Dnþm f ðxÞ; m 2 Z One of the best-known approximations is due to
0 _ n; m 2 N0 . (2)
Manabe and Oustaloup and makes use of a
The full definition of D becomes recursive distribution of poles and zeroes. The
n
approximating transfer function is given by [8]
c Dx f ðxÞ
8 Y
N
1 þ ðs=oz;n Þ
> R x ðx  xÞn1 sn  k ; n40. (5)
>
> f ðxÞ dx if no0;
> c
> GðnÞ 1 þ ðs=op;n Þ
>
> n¼1
<
¼ f ðxÞ if n ¼ 0; The approximation is valid in the frequency range
>
>
>
> Dn ½c Dxnn f ðxÞ if n40; ½ol ; oh . Gain k is adjusted so that both sides of (5)
>
>
>
: shall have unit gain at 1 rad/s. The number of poles
n ¼ minfk 2 N : k4ng:
and zeroes N is chosen beforehand, and the good
ð3Þ performance of the approximation strongly depends
 ARTICLE IN PRESS
D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784 2773

thereon: low values result in simpler approxima- e

d
tions, but also cause the appearance of a ripple in + F G* G +
+
-
both gain and phase behaviours; such a ripple may
be practically eliminated increasing N, but the
approximation will be computationally heavier. G′
Frequencies of poles and zeroes in (5) are given by
pffiffiffi
oz;1 ¼ ol Z, ð6Þ
op;n ¼ oz;n a; n ¼ 1 . . . N, ð7Þ
Fig. 1. Block diagram for internal model control.
oz;nþ1 ¼ op;n Z; n ¼ 1 . . . N  1, ð8Þ
a ¼ ðoh =ol Þn=N ð9Þ
d
ð1nÞ=N e
Z ¼ ðoh =ol Þ . ð10Þ
C G +
+ +
The case no0 may be dealt with inverting (5). But if -
jnj41 approximations become unsatisfactory; for
that reason (among others), it is usual to split
fractional powers of s like this: Fig. 2. Block diagram equivalent to that of Fig. 1.
n n d
s ¼s s ; n ¼ n þ d ^ n 2 Z ^ d 2 ½0; 1. (11)
In this manner only the latter term has to be were unity, control would be perfect. Since no
approximated. models are perfect, e will not be exactly the
If a discrete transfer function approximation is disturbance. That is also exactly why F exists and
sought, the above approximation (or any other is usually a low-pass filter: to reduce the influence of
alternative approximation) may be discretised using high-frequency modelling errors. It also helps
any usual method (Tustin, Simpson, etc.). But there ensuring that product FG  is realisable.
are also formulas that directly provide discrete The interconnections of Fig. 1 are equivalent to
approximations. None shall be needed in what those of Fig. 2 if the controller C is given by
follows. See, for instance [9] for more on this FG 
subject. C¼ . (13)
1  FG  G 0
3. Analytical tuning methods Controller C is not, in the general case, a PID or a
fractional PID, but in some cases it will, if
The transfer function of a fractional PID is given K
by G¼ eLs . (14)
1 þ sm T
I Firstly, let
CðsÞ ¼ P þ þ Dsm . (12)
sl
1
In this section, two methods published in the F¼ , ð15Þ
1 þ sT F
literature for analytically tuning the five parameters m
1þs T
of such controllers are given. G ¼ , ð16Þ
K
K
3.1. Internal model control G0 ¼ ð1  sLÞ. ð17Þ
1 þ sm T
The internal model control methodology may, in Note that the delay of G was neglected in G  but not
some cases, be used to obtain PID or fractional PID in G0 , where an approximation consisting of a
controllers. It makes use of the control scheme of truncated McLaurin series has been used. Then (13)
Fig. 1. In that control loop, G is the plant to control, becomes
G  is an inverse of G (or at least a plant as close as
1=KðT F þ LÞ T=KðT F þ LÞ
possible to the inverse of G), G 0 is a model of G and C¼ þ . (18)
F is some judiciously chosen filter. If G0 were exact, s s1m
the error e would be equal to disturbance d. If, This can be viewed as a fractional PID controller
additionally, G were the exact inverse of G and F with the proportional part equal to zero.
 ARTICLE IN PRESS
2774 D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784

Secondly, let tude be less than some specified gain:

 
F ¼ 1, ð19Þ  Cðjoh ÞGðjoh Þ 
 
1 þ sm T 1 þ Cðjo ÞGðjo ÞoH. (29)
h h
G ¼ , ð20Þ
K
K 1 (4) So as to reject output disturbances and closely
G0 ¼ m
. ð21Þ
1 þ s T 1 þ sL follow references, the sensitivity function must
Then (13) becomes have a small magnitude at low frequencies; thus
it is required that at some specified frequency ol
1 1=KL T=KL T m
C¼ þ þ 1m þ s . (22) its magnitude be less than some specified gain:
K s s K  
 1 
 
1 þ Cðjo ÞGðjo ÞoN.
If one of the two integral parts is neglectable, (22) (30)
will be a fractional PID controller. (The price to pay l l
for neglecting a term is some possible slight
deterioration in performance.) (5) So as to be robust in face of gain variations of
Finally, if a Padé approximation with one pole the plant, the phase of the open-loop transfer
and one zero is used in G 0 , function must be (at least roughly) constant
F ¼ 1, ð23Þ around the gain-crossover frequency:

1 þ sm T d 
G ¼ , ð24Þ arg½CðjoÞGðjoÞ ¼ 0. (31)
K do o¼ocg
K 1  sL=2
G0 ¼ m
, ð25Þ
1 þ s T 1 þ sL=2 Conditions are five because five are the parameters
then (13) becomes to tune. To satisfy them all the authors proposed the
use of numerical optimisation algorithms, namely of
1 1=KL T=KL T m
C¼ þ s þ 1m þ s . (26) Nelder–Mead’s simplex method as implemented in
2K s s 2K Matlab’s function fmincon (the condition in (27) is
Again, (26) will be a fractional PID if one of the two assumed as the condition to minimise; conditions in
integral parts is neglectable. (28)–(31) are assumed as constraints). This is
Obviously, should m 2 Z, Eqs. (18), (22) and (26) effective but allows local minima to be obtained.
become usual PIDs. In practice most solutions found with this
optimisation method are good enough, but they
3.2. Tuning by minimisation strongly depend on initial estimates of the para-
meters provided. Some may be discarded because
Monje et al. [4] proposed that fractional PIDs be they are unfeasible or lead to unstable loops, but in
tuned by requiring them to satisfy the following five many cases it is possible to find more than one
conditions (C being the controller and G the plant): acceptable fractional PID. In other cases, only well-
chosen initial estimates of the parameters allow
(1) The gain-crossover frequency ocg is to have finding a solution.
some specified value:
jCðjocg ÞGðjocg Þj ¼ 0 dB. (27) 4. Tuning rules

The first Ziegler–Nichols rule for tuning an integer

(2) The phase margin jm is to have some specified
value:
p þ jm ¼ arg½Cðjocg ÞGðjocg Þ. (28)
(3) So as to reject high-frequency noise, the closed-
loop transfer function must have a small
magnitude at high frequencies; thus it is required
that at some specified frequency oh its magni-
PID assumes the plant to have an S-shaped unit-step
response, as that of Fig. 3, where L is an apparent
delay and T may be interpreted as a time constant,
such as the one resulting from a pole. The method
cannot be applied if the unit-step response is shaped
otherwise. The simplest plant with such a response is
K
GðsÞ ¼ eLs . (32)
1 þ sT
 ARTICLE IN PRESS
D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784 2775

4.1. First set of rules

K

t
in
po
A first set of rules is given in Table 1. This is to be

n
io
read as

ct
fle
in
at
P ¼  0:0048 þ 0:2664L þ 0:4982T
output

nt
ge
þ 0:0232L2  0:0720T 2  0:0348TL
n

ð33Þ
ta

and so on. They may be used if

inflection point
0:1pTp50 and Lp2 (34)
and were designed for the following specifications:
0
0 L L+T ocg ¼ 0:5 rad=s, ð35Þ
time 
jm ¼ 2=3 rad  38 , ð36Þ
Fig. 3. S-shaped unit-step response. oh ¼ 10 rad=s, ð37Þ
ol ¼ 0:01 rad=s, ð38Þ
H ¼ 10 dB, ð39Þ
The minimisation tuning method presented above in N ¼ 20 dB. ð40Þ
Section 3.2 was applied to plants given by (32) for Recall that specifications are only approximately
several values of L and T, with K ¼ 1. The parameters verified.
of fractional PIDs thus obtained vary in a regular
manner with L and T. Using the least-squares method,
4.2. Second set of rules
it is possible to translate that regularity into poly-
nomial formulas to find acceptable values of the
A second set of rules is given in Table 2. This
parameters from the values of L and T. These are
second set of rules may be applied if
given in the two following subsections.2
Two comments. Firstly, to implement the mini- 0:1pTp50 and Lp0:5. (41)
misation tuning method the last condition was Only one set of parameters is needed in this case
verified numerically, evaluating argument in (31) at because the range of values of L with which these
two frequencies, equal to ocg =1:122 and 1:122 ocg rules cope with is more reduced. They were designed
1
(this corresponds to 20 of a decade for each side of for the following specifications:
ocg ). It is of course possible to evaluate the
argument at other frequencies around ocg ; actually, ocg ¼ 0:5 rad=s, ð42Þ
the larger the interval where the argument is
constant (or nearly so) the better, and thus using Table 1
Parameters for the first set of tuning rules
more than two points might ensure that. However,
it was verified that such stronger requirements are P I l D m
so difficult to meet that they often prevent a
solution from being found. Parameters to use when 0:1pTp5
1 0:0048 0:3254 1:5766 0:0662 0:8736
Secondly, least-squares method-adjusted formu- L 0:2664 0:2478 0:2098 0:2528 0:2746
las cannot exactly reproduce every change in T 0:4982 0:1429 0:1313 0:1081 0:1489
parameters. This means that fractional PIDs tuned L2 0:0232 0:1330 0:0713 0:0702 0:1557
with the rules presented below never behave as well T2 0:0720 0:0258 0:0016 0:0328 0:0250
as those tuned analytically (including those they are LT 0:0348 0:0171 0:0114 0:2202 0:0323
based upon), neither are they so robust. In other Parameters to use when 5pTp50
words, conditions in Eqs. (27)–(31) will only 1 2:1187 0:5201 1:0645 1:1421 1:2902
approximately be verified. L 3:5207 2:6643 0:3268 1:3707 0:5371
T 0:1563 0:3453 0:0229 0:0357 0:0381
L2 1:5827 1:0944 0:2018 0:5552 0:2208
T2 0:0025 0:0002 0:0003 0:0002 0:0007
2
LT 0:1824 0:1054 0:0028 0:2630 0:0014
These rules were already presented in [10].
 ARTICLE IN PRESS
2776 D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784

Table 2
Parameters for the second set of tuning rules
1.5
P I l D m

1 1:0574 0:6014 1:1851 0:8793 0:2778

L 24:5420 0:4025 0:3464 15:0846 2:1522
1

output
T 0:3544 0:7921 0:0492 0:0771 0:0675
L2 46:7325 0:4508 1:7317 28:0388 2:4387
T2 0:0021 0:0018 0:0006 0:0000 0:0013
LT 0:3106 1:2050 0:0380 1:6711 0:0021 0.5

jm ¼ 1 rad  57 , ð43Þ 0

0 10 20 30 40 50
(a) time / s
oh ¼ 10 rad=s, ð44Þ
ol ¼ 0:01 rad=s, ð45Þ
50
H ¼ 20 dB, ð46Þ

gain / dB
N ¼ 20 dB. ð47Þ 0

Of course, sets of rules other than those two above

-50
might have been found in the same way as these 10-2 10-1 100 101 102
were for different sets of specifications. ω / rad × s -1

0
5. Examples -200
phase / °

-400
In this section the rules from Section 4 are applied
to three different plants. The performance of the -600
results is then asserted and compared to the
(b) 10-2 10-1 100 101 102
performance obtained with integer PIDs tuned with
the first Ziegler–Nichols rule. 0
Two comments. Firstly, as stated above, rules
gain / dB

usually lead to results poorer than those they were -20

devised to achieve. (The same happens with
Ziegler–Nichols rules: they are expected to result -40
in an overshoot around 25%, but it is not hard to
10-2 10-1 100 101 102
find plants with which the overshoot is 100% or ω / rad × s -1
even more.) Secondly, Ziegler–Nichols rules make 0
no attempt to reach always the same gain-crossover
gain / dB

-20
frequency, or the same phase margin. Actually,
these two performance indicators vary widely as L -40

and T vary. This adds some flexibility to Ziegler– -60

Nichols rules: they can be applied for wide ranges of -80
L and T and still achieve a controller that stabilises (c) 10-2 10-1 100 101 102

the plant. Rules from the previous section always 1

Fig. 4. (a) Step response of (51) controlled with (52) when K is 32 ,
aim at fulfilling the same specifications, and that is 1 1 1 1
16, 8, 4, 2, 1 (thick line), 2, 4 and 8. (b) Open-loop Bode diagram
why their application range is never so broad as that when K ¼ 1. (c) Closed-loop function gain (top) and sensitivity
of Ziegler–Nichols rules. function gain (bottom) when K ¼ 1.
Bode diagrams presented are exact; all time-
responses involving fractional derivatives and in- ol ¼ 103 rad=s, ð48Þ
tegrals were obtained with simulations making use
3
of Oustaloup’s approximation described in the oh ¼ 10 rad=s, ð49Þ
subsection dealing with approximations, with N ¼ 7. ð50Þ
 ARTICLE IN PRESS
D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784 2777

1.5 1.5

1 1
output

output
K
0.5 0.5
K

0 0
0 10 20 30 40 50 0 10 20 30 40 50
(a) time / s (a) time / s

50 50

gain / dB
gain / dB

0 0

-50
-50
10-2 10-1 100 101 102
10-2 10 -1
10 0
101
102
ω / rad × s-1
ω / rad × s-1
0
0
-200
phase / °

-200
phase / °

-400
-400
-600
-600
(b) 10-2 10-1 100 101 102
(b) 10-2 10-1 100 101 102

0
0
gain / dB
gain / dB

-20
-20

-40
-40
10-2 10-1 100 101 102
10-2 10-1 100 101 102
ω / rad × s-1
ω / rad × s-1
0
0
-20
gain / dB

-20
gain / dB

-40
-40
-60
-60
-80
-80 (c) 10-2 10-1 100 101 102
(c) 10-2 10 -1
10 0
10 1
10 2
1
Fig. 6. (a) Step response of (51) controlled with (54) when K is 32 ,
1
Fig. 5. (a) Step response of (51) controlled with (53) when K is 32 , 1 1 1 1
, , , and 1 (thick line). (b) Open-loop Bode diagram when
1 1 1 1 16 8 4 2
, , ,
16 8 4 2, 1 (thick line). (b) Open-loop Bode diagram when K ¼ 1. K ¼ 1. (c) Closed-loop function gain (top) and sensitivity
(c) Closed-loop function gain (top) and sensitivity function gain function gain (bottom) when K ¼ 1.
(bottom) when K ¼ 1.
 ARTICLE IN PRESS
2778 D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784

5.1. First-order plant with delay according to the 10–90% rule and the settling time
is reckoned according to the 5% rule.)
The plant considered was The corresponding open-loop Bode diagrams and
the gains of sensitivity and closed-loop functions are
K 0:1s also given in those figures. They show that the
GðsÞ ¼ e . (51)
1þs desired conditions—given by Eqs. (27)–(31)—are
The nominal value of K is 1. Controllers obtained reasonably—though not exactly—followed. The
with the two tuning rules from the previous section approximations incurred by the least-squares fit
and with the first Ziegler–Nichols rule are are responsible for the conditions being only
approximately verified.
0:5158
C 1 ðsÞ ¼ 0:4448 þ þ 0:2045s1:0202 , ð52Þ
s1:4277 5.2. Second-order plant
1:3106
C 2 ðsÞ ¼ 1:2507 þ 1:1230  0:2589s0:1533 , ð53Þ
s The plant considered was
60:0000 K K
C ZN ðsÞ ¼ 12:0000 þ þ 0:6000 s. ð54Þ GðsÞ ¼  e0:2s
s 4:3200s2 þ 19:1801s þ 1 1 þ 20s
(Note that due to the approximations involved one (55)
of the gains is negative. This will not, however, with a nominal value of K of 1. The approximation
affect results.) Corresponding step responses are stems from the values of L and T obtained from its
given in Figs. 4–6. These show what happens for step response.
several values of K, the plant’s gain, which is Controllers obtained with the two rules given
assumed to be known with uncertainty. It should be above and with the first Ziegler–Nichols rule are
noticed that fractional PIDs can deal with a clearly
6:5185
broader range of values of K. This is likely because C 1 ðsÞ ¼ 0:0880 þ þ 2:5881s0:6957 , ð56Þ
the specifications the integer PID tries to achieve are s0:6751
different: that is why responses are all faster, at the 12:4044
C 2 ðsÞ ¼ 6:9928 þ 0:6000 þ 4:1066s0:7805 , ð57Þ
cost of greater overshoots. More important is that s
the overshoot is fairly constant with fractional 300:0000
C ZN ðsÞ ¼ 120:0000 þ þ 12:0000 s. ð58Þ
PIDs—at least for those values closer to 1. This is s
because fractional PIDs attempt to verify (31), The step responses obtained (together with open-
which the integer PID does not. Data on these loop Bode diagrams and sensitivity and closed-loop
responses are summed up in Table 3. (In this and functions’ gains) are given in Figs. 7–9. Table 4
the following tables, the rise time is reckoned presents data on the step responses. This time, since

Table 3
Data on step responses of Figs. 4–6

K Controller of Eq. (52) Controller of Eq. (53) Controller of Eq. (54)

Rise time Overshoot Settling time Rise time Overshoot Settling time Rise time Overshoot Settling time
(s) (%) (s) (s) (%) (s) (s) (%) (s)

1 22.1 26 94.6 28.1 5 78.2 1.0 23 3.7

32
1 13.8 27 59.1 15.1 5 19.2 0.6 33 3.9
16
1 8.9 28 36.5 8.1 5 10.2 0.4 40 2.8
8
1 5.9 30 22.6 4.4 6 12.4 0.2 45 1.9
4
1 4.0 30 19.8 2.4 8 7.7 0.1 48 1.3
2
1 2.6 27 14.6 1.3 9 4.7 0.1 74 0.7
2 1.7 20 7.4 0.7 8 2.8 – – –
4 0.9 12 5.5 0.3 8 1.4 – – –
8 0.2 7 3.9 – – – – – –
 ARTICLE IN PRESS
D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784 2779

there is no delay, the plant is easier to control and a that is more constant—in spite of the plant having
wider variation of K is supported by all controllers. a structure different from that used to derive the
But fractional PIDs still achieve an overshoot rules.

1.5 1.5

output
output

1 1

0.5 0.5
K

0 0
0 10 20 30 40 50 0 10 20 30 40 50
(a) time / s (a) time / s

100 100
gain / dB

gain / dB

50 50
0 0
-50 -50
-4 -2 0 2
10 10 10 10 10-4 10-2 100 102
ω / rad × s -1
ω / rad × s -1

0 0
phase / °

-50 50
phase / °

-100 -100

-150 - 150

(b) 10-4 10-2 100 102 (b) 10-4 10-2 100 102

0 0
gain / dB
gain / dB

-20 -20
-40 -40
-60 -60
-80 -80
10-4 10-2 100 102 10-4 10-2 100 102
ω / rad × s-1 ω / rad × s -1

0 0
gain / dB
gain / dB

-50 -50

-100 -100

(c) 10-4 10-2 100 102 (c) 10-4 10-2 100 102
1 1
Fig. 7. (a) Step response of (55) controlled with (56) when K is 32 , Fig. 8. (a) Step response of (55) controlled with (57) when K is 32 ,
1 1 1 1 1 1 1 1
16, 8, 4, 2, 1 (thick line), 2, 4, 8, 16 and 32. (b) Open-loop Bode 16, 8, 4, 2, 1 (thick line), 2, 4, 8, 16 and 32. (b) Open-loop Bode
diagram when K ¼ 1. (c) Closed-loop function gain (top) and diagram when K ¼ 1. (c) Closed-loop function gain (top) and
sensitivity function gain (bottom) when K ¼ 1. sensitivity function gain (bottom) when K ¼ 1.
 ARTICLE IN PRESS
2780 D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784

5.3. Fractional-order plant with delay

1.5
The plant considered was
K K
GðsÞ ¼ pffiffi e0:5s  e0:1s (59)
1
1þ s 1 þ 1:5s
output

with a nominal value of K of 1. The approximation

is derived from the plant’s step response at
0.5 t ¼ 0:92 s. (It might seem more reasonable to base
K the approximation on the step response at t ¼ 0:5 s,
but this cannot be done, since the response has an
infinite derivative at that time instant.)
0
0 10 20 30 40 50 Controllers obtained with the two rules
(a) time / s given above and with the first Ziegler–Nichols
rule are
100
0:6187
gain / dB

50 C 1 ðsÞ ¼ 0:6021 þ þ 0:3105s1:0618 , ð60Þ

0
s1:3646
1:6486
-50 C 2 ðsÞ ¼ 1:4098 þ 1:1011  0:2139s0:1855 , ð61Þ
s
10-4 10-2 100 102 90:0000
C ZN ðsÞ ¼ 18:0000 þ þ 0:9000 s. ð62Þ
ω / rad × s-1 s
0

-50
The step responses obtained (together with open-
phase / °

loop Bode diagrams and sensitivity and closed-loop

-100 functions’ gains) are given in Figs. 10–12. Table 5
-150 presents data on the step responses. The PID
performs poorly because it tries to obtain a fast
(b) 10-4 10-2 100 102 response and thus employs higher gains (and hence
1
the loop becomes unstable if K is larger than 32 ), but
0 that is not what is most relevant. The most relevant
gain / dB

-20 result here is that fractional PIDs still achieve

-40 practically constant overshoots, since, in spite of the
-60 different plant structure, the conditions they were
-80
expected to verify are still verified to a reasonable
10-4 10-2 100 102 degree, as the frequency–response plots show.
ω / rad × s -1
In this case it is possible to find IMC-tuned
0 fractional PIDs to compare results. Using the
parameters of plant (59) (K ¼ 1, T ¼ 1, L ¼ 0:5
gain / dB

50 and m ¼ 0:5; these are not the parameters of

the approximation used with the tuning rules),
100 Eqs. (22) and (26) yield the two following transfer
functions:
(c) 10-4 10-2 100 102
1
Fig. 9. (a) Step response of (55) controlled with (58) when K is 32 , 2 2
1 1 1 1
, , , , 1 (thick line), 2, 4, 8, 16 and 32. (b) Open-loop Bode C IMC ¼ 1 þ þ 1=2 þ s1=2 , ð63Þ
16 8 4 2 s s
diagram when K ¼ 1. (c) Closed-loop function gain (top) and
1 2 2 1
sensitivity function gain (bottom) when K ¼ 1. C IMC ¼ þ þ 1=2 þ s1=2 . ð64Þ
2 s s 2

In none of the two cases is clear which of the two

integral terms is better to discard. By trying both
possibilities, it is found out that it is better to keep
 ARTICLE IN PRESS
D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784 2781

Table 4
Data on step responses of Figs. 7–9

K Controller of Eq. (52) Controller of Eq. (53) Controller of Eq. (54)

Rise time Overshoot Settling time Rise time Overshoot Settling time Rise time Overshoot Settling time
(s) (%) (s) (s) (%) (s) (s) (%) (s)

1 31.4 – 45.5 21.7 – 148.0 1.5 75 41.2

32
1 15.6 8 38.2 10.9 5 23.3 1.0 74 25.9
16
1 9.1 19 47.1 6.3 13 19.2 0.7 71 14.0
8
1 5.7 28 32.1 3.8 17 13.3 0.5 66 8.2
4
1 3.8 34 22.1 2.4 19 9.0 0.4 61 4.6
2
1 2.6 36 15.3 1.5 19 5.9 0.3 53 2.0
2 1.7 35 10.5 0.9 16 3.8 0.2 43 1.3
4 1.2 31 7.2 0.5 13 2.5 0.2 31 0.8
8 0.8 23 3.3 0.3 9 1.5 0.1 22 0.8
16 0.5 15 2.5 0.2 6 0.7 0.1 17 0.8
32 0.2 8 1.8 0.1 6 0.3 0.1 15 0.8

the terms with the first derivative:
2 plant is available, provided a suitable time response
C IMC2 ¼ 1 þ þ s1=2 ,
s is; they may be used as a departing point for fine-
1 2 1 1=2
C IMC1 ¼ þ þ s .
2 s 2
Step responses obtained are shown in Fig. 13 and
compare well with those of Figs. 10 and 11. It is seen
that rule-tuned fractional PIDs perform nearly as
well as those found with IMC.
6. Comments and conclusions
In this paper, two analytical methods (among
others published in the literature) for tuning the
parameters of fractional PIDs were reviewed. The
optimisation method of [4] was then used for
developing two sets of tuning rules similar to those
of the first set of Ziegler–Nichols rules. These new
tuning rules make use of two parameters (L and T)
of the unit-step response of the plant, which should
be S-shaped: otherwise they cannot be applied.
The most obvious difference is that the new rules
are clearly more complicated than those of Ziegler–
Nichols: they have to be quadratic (approximations
of lower order were tried but proved unsatisfac-
tory). And the broader the application range of the
rules is to be, the more complicated they become:
the first rule needs two tables of parameters, while
the second, good for a narrower interval of values of
L only, needs only one.
The usefulness of these rules is that of all sets of
rules: they may be applied even if no model of the
ð65Þ
tuning (this is, for instance relevant if the optimisa-
ð66Þ
tion tuning method is used, since its results depend
significantly from the initial estimate provided);
they are easier and faster to apply than analytic
methods. Their drawbacks are also those of all sets
of rules: their performance is often inferior to the
one sought, fine-tuning being often needed; they
perform worse than controllers tuned analytically;
they cannot be applied to all types of plants, but
only to those with a particular sort of time response.
Fractional PIDs tuned with these new rules
compare well with integer PIDs tuned according
to the first Ziegler–Nichols rule, even though the
comparison is made difficult because Ziegler–
Nichols rules achieve different specifications for
different values of T and L while the new rules
attempt to always keep a uniform result. The
advantage fractional PIDs provide is a roughly
constant overshoot when the gain of the plant
undergoes variations. (It is of course likely that
carefully tuned integer PIDs perform better than
rule-tuned fractional PIDs—just as carefully tuned
fractional PIDs are likely to perform better than
rule-tuned integer PIDs.)
It is surely possible to improve these tuning rules.
Rules similar to the second Ziegler–Nichols rule
(making use of a closed-loop response of the plant)
are certainly possible, and are currently being
 ARTICLE IN PRESS
2782 D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784

1.4 1.4

1.2 1.2

1 1

0.8 0.8

output
output

0.6 0.6

0.4 0.4
K
K
0.2 0.2

0 0
0 10 20 30 40 50 0 10 20 30 40 50
(a) time / s (a) time / s

80
80
60
60
gain / dB

40

gain / dB
40
20
20
0
0
-20
-20
10-2 10 -1
10 0
10 1
10 2
10-2 10-1 100 101 102
ω / rad × s-1
ω / rad × s-1
0
0
phase / °

phase / °

-500
-500

-1000
(b) 10-2 10-1 100 101 102 -1000
(b) 10-2 10-1 100 101 102

0
0
gain / dB

gain / dB

-20
-20

-40
-40
10-2 10-1 100 101 102
ω / rad × s-1 10-2 10-1 100 101 102
ω / rad × s -1
0
0
-20
gain / dB

-20
gain / dB

-40
-40
-60
-60
-80
(c) 10-2 10 -1
10 0
10 1 2
10 -80
(c) 10-2 10-1 100 101 102
Fig. 10. (a) Step response of (59) controlled with (60) when K is
1 1 1 1 1 Fig. 11. (a) Step response of (59) controlled with (61) when K is
32, 16, 8, 4, 2, 1 (thick line) and 2. (b) Open-loop Bode diagram 1 1 1 1 1
when K ¼ 1. (c) Closed-loop function gain (top) and sensitivity 32, 16, 8, 4, 2
and 1 (thick line). (b) Open-loop Bode diagram when
function gain (bottom) when K ¼ 1. K ¼ 1. (c) Closed-loop function gain (top) and sensitivity
function gain (bottom) when K ¼ 1.
 ARTICLE IN PRESS
D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784 2783

Settling time
1.4

1.2

3.1 s
(s)

–
–
–
–
–
–
1

0.8
output

Overshoot
0.6

(%)
Controller of Eq. (54)

31
0.4

–
–
–
–
–
–
0.2

Rise time
0
0 10 20 30 40 50

0.6 s
(a) time / s

(s)

–
–
–
–
–
–
80
60
gain / dB

40

Settling time
20
0

86.6
47.7
27.1
16.0
9.6
5.6
-20

(s)

–
10-2 10-1 100 101 102
ω / rad × s -1

Overshoot
phase / °

-500 (%)
Controller of Eq. (53)

7
7
8
9
9
8
–
-1000
(b) 10-2 10-1 100 101 102
Rise time

0
26.5
14.7
8.2
4.6
2.4
1.1
gain / dB

(s)

-20 –

-40
Settling time

10-2 10-1 100 101 102

ω / rad × s-1
105.5
65.8
41.2
26.1
17.2
11.9
8.6

0
(s)
gain / dB

-20
-40
-60
Overshoot
Data on step responses of Fig. 10–12.

-80
(c) 10-2 10-1 100 101 102
(%)
Controller of Eq. (52)

26
27
28
29
27
23
17

Fig. 12. (a) Step response of (59) controlled with (62) when K is
1
32.(b) Open-loop Bode diagram when K ¼ 1. (c) Closed-loop
function gain (top) and sensitivity function gain (bottom) when
K ¼ 1.
Rise time

25.1
15.8
10.3
6.9
4.4
2.7
1.5
(s)

developed. Rules specific for non-minimum phase

Table 5

plants (with which Ziegler–Nichols rules do not

properly deal) may also be of interest.
K

32

16
1

1
2
1
8
1
4
1
2
 ARTICLE IN PRESS
2784 D. Valério, J.S. da Costa / Signal Processing 86 (2006) 2771–2784

1.4 References
1.2
[1] T. Hägglund, K. Åström, Automatic tuning of PID
controllers, in: W.S. Levine (Ed.), The Control Handbook,
1
CRC Press, Boca Raton, FL, 1996, pp. 817–826.
0.8 [2] I. Podlubny, Fractional Differential Equations: An Intro-
output

duction to Fractional Derivatives, Fractional Differential

0.6 Equations to Methods of their Solution and Some of their
Applications, Academic Press, San Diego, 1999.
K
0.4 [3] R. Caponetto, L. Fortuna, D. Porto, Parameter tuning of a
non integer order PID controller, in: Fifteenth International
0.2 Symposium on Mathematical Theory of Networks and
Systems, Notre Dame, 2002.
0 [4] C.A. Monje, B.M. Vinagre, Y.Q. Chen, V. Feliu, P. Lanusse, J.
0 10 20 30 40 50
(a) time / s Sabatier, Proposals for fractional PIl Dm tuning, in: Fractional
Differentiation and its Applications, Bordeaux, 2004.
1.4 [5] Y.Q. Chen, K.L. Moore, B.M. Vinagre, I. Podlubny, Robust
PID controller autotuning with a phase shaper, in:
1.2 Fractional Differentiation and its Applications, Bordeaux,
2004.
1 [6] K.S. Miller, B. Ross, An Introduction to the Fractional
Calculus and Fractional Differential Equations, Wiley, New
0.8 York, 1993.
output

[7] S.G. Samko, A.A. Kilbas, O.I. Marichev, Fractional Integrals

0.6 and Derivatives, Gordon and Breach, Yverdon, 1993.
K [8] A. Oustaloup, La commande CRONE: commande robuste
0.4
d’ordre non entier, Hermès, Paris, 1991.
0.2
[9] B.M. Vinagre, I. Podlubny, A. Hernández, V. Feliu, Some
approximations of fractional order operators used in control
0 theory and applications, Fractional Calculus & Appl. Anal.
0 10 20 30 40 50 3 (2000) 231–248.
(b) time / s [10] D. Valério, J. Sá da Costa, Ziegler–Nichols type tuning rules
for PID controllers, in: Fifth ASME International Con-
Fig. 13. (a) Step response of (59) controlled with (65) when K is ference on Multibody Systems, Nonlinear Dynamics and
1 1 1 1 1
32, 16, 8, 4 and 2. (b) Step response of (59) controlled with (66) Control, Long Beach, 2005.
1 1 1 1 1
when K is 32 , 16, 8, 4, 2 and 1 (thick line).