2012 American Control Conference
Fairmont Queen Elizabeth, Montréal, Canada
June 27-June 29, 2012
H∞ Model Matching PID Design for Fractional FOPDT Systems
Fabrizio Padula, Ramon Vilanova and Antonio Visioli
Abstract— This paper deals with the H∞ model match-
ing problem for fractional first-order-plus-dead-time processes.
Starting from the analytical solution of the problem, we show
that a fractional proportional-integral-derivative controller can
be obtained. In this context, the (robust) stability issue is
analyzed. Further, guidelines for the tuning of the controller
parameters are given in order to address practical issues and to
obtain the required performance. Simulation results show that
the design methodology is effective and allows the user (on the
contrary to the integer-order case) to consider processes with
different dynamics in a unified framework.
I. I NTRODUCTION
In the last years many theoretical and practical re-
sults have been achieved in fractional-order control (see,
for example, [1]). In particular, different design methods
for fractional-order-proportional-integral-derivative (FOPID)
controllers have been devised (see [2], [3] and references
therein contained). Despite the large number of results pre-
sented in the literature, fractional control has never been
treated analytically under the point of view of H∞ model
matching, until a methodology for fractional first-order-plus-
dead-time (FFOPDT) systems has been recently presented
in [4] (note that, as pointed out in [5], a general solution
to solve H∞ optimization problem involving fractional-
order plants does not exist yet). In [4], starting from an
Internal Model Control (IMC) controller representation of the
feedback loop, the optimization problem is formulated and,
once the optimal IMC controller is obtained, an equivalent
feedback controller is derived via Youla parametrization [6].
It has also been shown that it is possible to determine a
FOPID feedback controller by approximating the optimal
one. Indeed, the method extends to fractional-order systems
the method proposed in [7] for integer-order ones.
In this paper, after having reviewed the methodology, we
address the stability and robustness issues that arise by
applying it. Further, we discuss the tuning of the parameters
from a practical point of view by giving suitable guidelines
in order to achieve the desired performance. Indeed, dealing
with fractional-order dynamics means that more flexibility
is introduced in the design phase. While this implies that a
better performance can be achieved in principle, from another
point of view the design can be more difficult in practice. It is
therefore essential that the physical meaning of all the design
F. Padula and A. Visioli are with Dipartimento
Ingegneria dell’Informazione, University of Brescia, Italy
{fabrizio.padula,antonio.visioli}@ing.unibs.it
R. Vilanova is with Dept. de Telecomunicació i Enginyeria de Sis-
temes, Universitat Autònoma de Barcelona, Bellaterra (BCN), Spain
Ramon.Vilanova@uab.cat
R. Vilanova acknowledges the support received from the Spanish CICYT
program under grant DPI2010-15230.
978-1-4577-1096-4/12/$26.00 ©2012 AACC
parameters is clearly explained and that a tuning procedure is
devised in order to achieve the desired nominal performance
while guaranteeing the robust stability at the same time.
We will also highlight, by means of simulation examples,
that, on the contrary to the solution of the model matching
problem for integer-order systems [7], the devised technique
allows the user to treat processes with different dynamics
(namely, overdamped or underdamped processes) in a unified
framework, in addition to the capability of dealing explicitly
with fractional-order processes. This is an advantage of the
proposed methodology with respect to the integer-order case.
II. P ROBLEM F ORMULATION
We consider a unity-feedback control loop where the
process G(s) is described by a FFOPDT process model
K
Gnt (s) = e−Ls , (1)
1 + T sα
where K is the gain, T is the time constant, L is the dead
time and α is the fractional derivative order. Note that, when
using α = 1, an integer-order FOPDT is obtained. In order
to apply the H∞ model matching technique, the delay term
has to be approximated by means of a first-order Taylor
expansion e−Ls ≈ 1 − Ls thus yielding
1 − Ls
Gn (s) = K . (2)
1 + T sα
The feedback controller is denoted as K(s), while T (s)
denotes the nominal complementary sensitivity function of
the control-loop:
K(s)Gn (s)
T (s) = . (3)
1 + K(s)Gn (s)
It can be shown that, using the IMC paradigm [6] (see Figure
1), T (s) can be expressed by means of the equivalent IMC
controller C(s) as:
T (s) = Gn (s)C(s). (4)
Once a stable C(s) has been synthesized, the feedback
controller can be easily recovered by means of:
C(s)
K(s) = . (5)
1 − C(s)Gn (s)
Usually, when using the IMC tuning method, the desired
di closed-loop time constant is treated as the tuning parameter
which allows the robustness/aggressiveness trade-off of the
control loop to be handled. Conversely, in this paper, using
the same approach proposed in [7] for integer-order transfer
functions, the IMC formulation is employed in order to set
a min-max (or model-matching) optimization problem on
5513

Fig. 1. The IMC and the equivalent feedback control schemes.
which the controller architecture is based. The problem is
formulated as finding an IMC controller C o (s) that solves
the following min-max optimization problem:
C o (s) = min kW (s)(M (s) − C(s)Gn (s))k∞ ,
C(s)
where M (s) is the reference model for the closed loop
system transfer function, while W (s) is a weighting function.
Then, by means of (5), the equivalent feedback controller
transfer function is obtained. The reference model is chosen
to be a fractional-first-order (FFO) process
1 the optimal interpolation error. In the integer case, the
M (s) = (7)
1 + T m sλ
whose time constant Tm plays the same role of the time
constant of the classical IMC control strategy, i.e., it specifies
the desired closed-loop bandwidth. In this paper, a fractional-
order weighting function is employed:
1 + zsµ
W (s) = γ . (8)
sµ
The weighting function plays a double role because it allows
the user to add degrees of freedom to the controller design
and to include the integral action into the equivalent feedback
controller. Thus, it guarantees, at least for step-like signals,
to achieve a zero steady-state set-point following error and
complete load disturbance rejection. It is worth noting that
a fractional-order integral action s1µ , with µ < 1, as well as
an integer-order one, guarantees zero steady-state error for
step-like signals.
III. S OLUTION OF THE O PTIMIZATION P ROBLEM
In order to solve the model-matching problem, applying
the polynomial division, we slightly reduce the class of
allowable fractional-order weights and fractional-order refer-
ence models to the ones whose exponents can be represented
as the ratio of two natural numbers, that is:
m
1 + zs n
W (s) = γ m , n, m ∈ N
sn
and
1 C o (s) = (W (s)Gn (s))−1 (W (s)M (s) − ρ)
M (s) = l , l, m ∈ N.
1 + Tm s n m
Note that, with an appropriate choice of n, m, and l, the
corresponding real numbers µ and λ can be approximated
with arbitrary precision. Therefore the exponents µ and λ,
from now on, are assumed to be rational numbers.
The min-max optimization problem (6) can be reformulated
as follows: find a stable transfer function C(s) such as the
cost function
J∞ = kE(s)k∞ = kW (s)(M (s) − C(s)Gn (s))k∞ (11)
is minimized, where Gn (s) is defined in (2), M (s) is defined
in (10) and W (s) in (9). All the transfer functions involved
in the model matching problem are stable and minimum
phase, with the exception of Gn (s) that has a right half
plane (RHP) zero at zg = L1 . This implies that the optimal
IMC controller cannot have an unstable pole at zg , because
it would cause internal instability. It is also well known
that this kind of problem can be recast in an equivalent
optimal interpolation one. Thus, the RHP zero introduces
an interpolation constraint since, in order to avoid internal
(6) instability, the model matching error has to satisfy:
E o (zg ) = W (zg )M (zg ) (12)
where E o (s) is the optimal interpolation error:
o
J∞ = kE 0 (s)k∞ = min kE(s)k∞ . (13)
Generalizing, each RHP zero zi generates a constraint for
constraints, together with a theoretical result that states that
E o is an all-pass transfer function (see [8], [9], [10]), give
a general solution to the H∞ interpolation problem.
In the fractional case a similar result does not exist, but, for
the particular (but useful to describe plenty of applications)
problem considered in this paper, it is possible to find an
analytical solution. Indeed, by imposing the constraints (12),
it is easy to see that the following equation holds:
L2µ + zLµ
ρ = |E 0 (zg )| = γ . (14)
Lµ + Tm Lµ−λ
The constraint imposes a lower bound to the ∞-norm of
the optimal interpolation error. Thus, any stable controller
C(s) such as kE(s)k∞ = ρ is optimal. Note that, at this
point, we cannot yet state the opposite, indeed an optimal
controller could make the optimal interpolation error ∞-
norm to be greater than ρ. The following lemma (see the
proof in [4]) is useful to derive the solution to this problem
stated in Theorem 1.
Lemma 1: A proper fractional-order system
Pk ϕi
i=0 bi s
P (s) = mP ψj
, ψi < 1, a0 > 0, aj ≥ 0 j = 1 · · · m;
j=0 aj s
(15)
is always stable.
Theorem 1: The optimal controller that solves the min-
(9)
max problem stated in (6) is
(16)
(10)
provided that µ = < 2, n, m ∈ R
n
Proof. First note that, if the controller defined in (16) is
stable, then E(s) = ρ. This comes directly from (16) together
with (6) and it is a sufficient condition for the optimality,
provided C o (s) to be stable.
Now, to conclude the proof, let us show the stability of
5514
C o (s). It is a fractional transfer function whose expression from C o (s) (IMC controller) and Gn (s) by means of (5).
with respect to the parameters of the transfer functions After some passages we have:
involved in the optimization process is:
1 1 + T sα Q(s)
α K o (s) = ρ µ ρ λ+µ Pn−1 k k .
1 1 + Ts K γs sλ
+ Tm + Tm (z + γ )s
C o (s) = k=0 L s
n n
χ(s) (17)
K (1 + zsµ )(1 + Tm sλ ) (24)
We can notice that the optimal controller has the same
where asymptotic behavior of a filtered FOPID controller.
1 + zsµ − sµ (1 + Tm sλ ) γρ
χ(s) = (18) It is worth stressing that, when using unitary exponents
1 − Ls µ = 1, λ = 1 and α = 1, the optimal controller obtained by
is a commensurate transfer function [11]. If µ < 2 the means of (24) is the same one obtained in [7]. Indeed, the
first part of C o (s) is stable, as a direct consequence of the fractional-order case could be seen as the natural extension of
stability theorem stated in [11]. Now the stability of χ(s) the integer-order one and, viceversa, the integer-order one is
has to be proven. Let us define its integer counterpart: a particular case of a more general solution in the presented
fractional-order systems framework.
1 + zsm − sm (1 + Tm sl ) γρ The optimal controller (24) that solves the model match-
χ̄(s) := . (19) ing problem is very difficult to implement because of its
1 − Lsn
complexity. Thus, it is meaningful to look for a simpler
The denominator dχ̄ (s) of χ̄(s) can be easily factorized as: suboptimal controller. In order to do that we assume that
n−1 the fractional degree µ = m n of (9) is equal to the fractional
k 1
degree λ = nl of (10). For this reason from now on we will
X
dχ̄ (s) = ( L n sk )(1 − L n s) (20)
k=0 only use µ.
First note that, as a consequence of the stability of the
and going back to its fractional counterpart leads to: optimal controller, the rest of the division of nχ̄ (s) by
1
n−1 (1 − L n s) has to be null. Thus, it has to be:
k k 1 1
X
dχ (s) = ( L n s n )(1 − L n s n ). (21) ρ ρ Tm m
k=0 −(z − − λ
)=Ln. (25)
γ γL
Now, using Lemma 1, we can see that the first multiplicative
Now let us consider the optimal feedback controller (24).
term of dχ (s) is stable, while the second one is clearly
By means of (25) and by taking into account the previous
unstable (it can be proven by means of the stability result
assumption it can be rewritten as:
derived in [11]). Let us now focus on the numerator nχ̄ (s)
1
of χ̄(s). If it is divisible by (1 − L n s) then the optimal (1 + T sα )(1 + ρT
Pm−1 k k
γLµ )
m
1 k=0 L s
n n
o
controller C o (s) is stable. It is straightforward to see, by K (s) = ρ µ ρ 2µ Pn−1 k k ,
K ( γ + Tm )s + Tm (z + γ )s k=0 L s
n n
polynomial division, that they divide each other and their (26)
ratio Q(s) is: that can be rearranged in the following form:
m−1 m+l−1
X ρ ρ Tm 1 k
X ρ Tm
1 (1 + T sα )(1 + ρT
γLµ )
m
Q(s) = −(z− − l ) m−k s + sk o
K (s) =
γ γ Ln L n γ L m+l−k
n K ( γ + Tm )sµ + Tm (z + γρ )s2µ
ρ
k=0 k=m
(22) Pn−1 k k Pm−1 k k ! (27)
and going back once again to the fractional counterpart of − k=m L n s n + k=n L n s n
· 1+ Pn−1 k k ,
nχ̄ (s) we can see that the optimal controller can therefore k=0 L s
n n

be expressed as:
where each summation is supposed to be null when the
o 1 1 + T sα Q(s) final index is smaller than the initial one. By neglecting
C (s) = Pn−1 k k (23) the summation terms in the previous equation and, once
K (1 + zsµ )(1 + Tm sλ ) k=0 Ln sn
again, by means of (25), we obtain the suboptimal feedback
If the conditions of the theorem are satisfied, the previous controller:
controller is stable and the theorem is proven. 2 +z µ µ
1 (1 + T sα )(1 + Tm LLµ +T s )
Remark 1. The optimal controller does not depend on the K̃(s) = m
. (28)
ρ ρ
+z
coefficient γ, see (9), because it always appears in the ratio K( γ + Tm ) sµ (1 + T ργ sµ )
m +Tm
ρ γ
γ , where ρ is defined in (14). Actually, the weight function
role is to formalize how much we want to consider a certain The previous expression is a FOPID controller in series form
frequency with respect to the others. Thus, a multiplica- (see [3]), filtered by means of a fractional pole. It is worth
tive coefficient that uniformly increases (or decreases) the noting that when using µ = 1 the suboptimal and the optimal
weighting function magnitude does not change the result. solutions coincide, i.e. K̃(s) = K o (s). Indeed, when using
Therefore, without loss of generality, we can assume γ = 1. an integer weight and an integer model reference, the optimal
The equivalent optimal feedback controller K o (s) is obtained controller is already an interacting FOPID controller. Finally

5515
note that, when using α = µ and z = Tm , a fractional-
order PI (FOPI) controller is obtained. Then, the suboptimal
FOPID controller has the same low-frequency asymptotic
behavior as the optimal one while, at high frequencies, the
actual process model (1) (that is a low-pass filter) makes the
mismatches between the open loop transfer functions less
significant.
IV. N OMINAL S TABILITY
When using the nominal process model (2), the optimal
controller (24) guarantees the closed-loop stability for any
choice of Tm , z, µ and λ, provided that the hypothesis of
Theorem 1 (i.e., µ < 2 and λ < 2) is satisfied.
Conversely, when considering the actual process model (1)
and the suboptimal controller (28), it is important to verify
the control system stability. In order to do this, the dual locus
method [12] can be employed. By denoting the open-loop
transfer function as H(s) := K̃(s)Gnt (s), the root locus
method results in the following characteristic equation:
µ
+z µ
(1 + Tm LLµ +T s )
1 + H(s) = 1 + m
e−Ls = 0.
( γρ + Tm )sµ + Tm ( γρ + z)s2µ
(29)
By defining
2µ µ µ 2 |Tn (jω)| < |1/Γ(jω)|. Rewritten in this form, it is easy to
(L + zL + Tm L + Tm )s
H1 (s) := −
(Lµ + Tm ) + Tm (Lµ + z)
(30)
Tm (L2µ + 2zLµ + zTm )s2µ
−
(Lµ + Tm ) + Tm (Lµ + z)
and H2 := e −Ls
, equation (29), taking into account (14),
can be rewritten as
H1 (s) − H2 (s) = 0. (31)
The dual locus technique is now applied: first, the inter-
section point has to be determined by solving the equation
|L1 (s)|s=jω = 1. Once the intersection point frequency ωc
has been obtained, the phases φ1 (ωc ) and φ2 (ωc ) of H1 (jωc )
and H2 (jωc ), respectively, are obtained. The closed-loop
system is stable if φ1 − φ2 < 0. Briefly speaking, the
stability condition is satisfied if the locus of H1 (s) reaches
the intersection point earlier than the one of H2 (s).
Solving (31) analytically could be a complex task, because,
in particular when µ 6= 1, it results in complex trigonometric
equations. Nevertheless, the dual locus method is useful to
study the nominal stability using a worst case approach.
It is worth noting that, as Tm and z decrease, the loop band-
width increases and the control system robustness decreases.
The worst case is thus obtained by setting Tm = 0 and
z = 0. With this choice, (31) becomes −Lµ s − e−Ls = 0,
and it is easy to prove that the stability condition is always
satisfied, provided that µ < 2 π−1π ≈ 1.36. Indeed, the actual
process model and the suboptimal controller guarantee the
closed-loop stability for any choice of the tuning parameter
Tm and z provided that µ < 1.36. On the other hand, when
µ > 1.36, a numerical solution of the dual locus equation
(31) can be easily computed. Anyway, it is quite uncommon
to meet, in practical cases, fractional integrators with order
greater than 1.3.
5516
V. ROBUST S TABILITY AND T UNING G UIDELINES
A. Robust Stability
The closed-loop stability is guaranteed as long as the pro-
cess model Gnt (s) perfectly describes the process dynamics.
However, as mismatches between the actual process and the
model used are unavoidable, it is important to be able to
stabilize a family of plants around the nominal one. Assume
that the process belongs to a family F defined as:
F = {G(s) = Gnt (s)(1 + ∆m (s)) : |∆m (jω)| < |Γ(jω)|} ,
(32)
where ∆m (s) = (G(s) − Gnt (s))/Gnt (s) is the uncertainty
description and Γ(jω) is a weight that defines the plants
family by upper bounding the modeling error. It is well
known that a controller that stabilizes a control system in
the nominal case, also stabilizes the family F of the control
systems such that kΓ(s)Tn (s)k∞ < 1 where Tn (s) is the
stable nominal closed-loop transfer function, namely:
K̃(s)Gnt (s)
Tn (s) = . (33)
1 + K̃(s)Gnt (s)
The robust stability condition is always satisfied if
compute, and the left hand side of this inequality is usually
a low-pass transfer function.
B. Tuning Guidelines
In this work we propose to handle the robust-
ness/aggressiveness trade-off by means of the z parameter.
1
For a fixed desired bandwidth (Tm ) µ , the user can modify
the z parameter until the robust stability constraint is satis-
fied. Of course, it might be possible that the robust stability
constraints are never satisfied, in this case the desired closed-
loop bandwidth is too large and it is necessary to reduce it by
1
increasing (Tm ) µ . Conversely, it might be possible that the
robust stability constraints are always satisfied, this means
that the required performance is easily achievable and it is
possible to speed up the system response by decreasing Tm .
The tuning of the µ parameter is less significant and, in
several cases, the use of an integer-order integrator is the
best solution [3], [13]. Nevertheless, sometimes the system
response becomes sluggish when the system output is close
to the set point value. This behavior mainly depends on
the model mismatch between the actual process and the
process model. In this context the µ parameter can be
slightly modified in order to achieve a fine tuning that tries
to optimize the transient response close to the set point
value. Summarizing, the tuning procedure can be outlined
as follows:
1) set µ = 1;
1
2) select the desired closed-loop bandwidth Tmµ ;
3) compute the process uncertainty and select z to satisfy
the robust stability constraints;
1
4) if such a value of z does not exist then increase Tmµ
and go to 3;
 5) if the transient response is sluggish when the process 10

output is close to the set point, reduce µ if the 0

process output tends from above to the set-point value, −10

otherwise increase µ;

magnitude [dB]
else set µ = 1; −20

6) if µ > 1.36 then verify the nominal stability; −30

7) if µ 6= 1 go to 3; −40

8) end.
−50
−1 0 1 2
10 10 10 10

VI. S IMULATION R ESULTS ω

Fig. 2. Plants family (green lines), fractional model (solid line) and integer
As Example 1, we consider the process model (dotted line) – Example 1.

K 30
G(s) = Gnt ,F (s) = e−Ls (34)
T sα + 1 20

and we suppose that the nominal process parameters assume 10

magnitude [dB]
their values in the middle points of following uncertainty z
0
intervals: L ∈ [0.35, 0.65], T ∈ [0.7, 1.3], K ∈ [0.7, 1.3] and
α ∈ [0.4, 1.0]. The frequency response of the corresponding −10

family of plants is shown in Figure 2. It is worth noting −20

that the considered uncertainty is significant. In particular,

−30
uncertainty on the order α implies that the controller is 10
−1 0
10
ω
1
10
2
10

required to deal with a wide range of different dynamics. For Fig. 3. Robust stability boundaries (green lines) and nominal closed-loop
transfer functions for z ∈ {0.1, 0.8} using the fractional model with µ = 1
the sake of comparison, the integer model with the nominal and Tm = (1.2L)µ – Example 1.
parameters and α = 1 has been also considered:
30
1 −0.5s
Gnt ,I (s) = e . (35)
s+1 20

Note that, in any case, it fits the family of plants (see Figure magnitude [dB] 10

2). The desired closed-loop bandwidth is initially selected as 0

z

Tm = (1.2L)µ = 0.6. Then, it can be seen that by using −10

z = 0.1, both in integer and in the fractional case the robust

stability is not achieved (see Figures 3 and 4). Using z = 0.8 −20

the fractional process model guarantees the robust stability −30

10
−1 0
10
1
10
2
10

whereas, for the integer one, it is required to use z = 1.5.

ω
Fig. 4. Robust stability boundaries (green lines) and nominal closed-loop
The step responses plotted in Figures 5 and 6 show that the transfer functions for z ∈ {0.1, 0.8, 1.5} and µ = 1 (solid lines) and
for z = 1.55 and µ = 1.1 (dashed line), using the integer model with
use of a fractional model allows the user to increase the speed Tm = (1.2L)µ – Example 1.
of the response without loosing the robust stability with
respect to the same family of plants. Finally, in order improve 2
process variable

1.5
the response when using the integer model, the integrator 1

degree has been increased to µ = 1.1 (and consequently, to 0.5

preserve the desired bandwidth, Tm = (1.2L)µ = 0.57), 0

0 2 4
time
6 8 10

increasing also z to 1.55 in order to satisfy the robust 3

stability constraints (see Figure 4). Of course, the price to

control variable

pay in order to control a family of plants different from the 1

0
nominal one used for controller design is a degradation of the −1
0 2 4 6 8 10
system response that could become oscillatory or sluggish time

depending on the actual process. It is worth noting that
the fractional model allows the user to guarantee the robust
stability with a more aggressive tuning (i.e., it improves the
robustness/performance trade-off). On the other hand, be-
cause of the larger bandwidth, the performance degradation
is more significant when the actual process is very different
from the nominal model. Nevertheless, the stability is always
guaranteed.
As Example 2, consider the following (underdamped)
second-order-plus-dead-time process:
1 by the tuning method. Conversely, a FFOPDT model can
G(s) = 2 e−0.6s . (36)
s + 0.4s + 1
Fig. 5. Step-responses using the fractional model with Tm = (1.2L)µ :
plant family response with µ = 1 and z = 0.8 (green lines) and nominal
closed-loop step response with µ = 1 and z = 0.8 – Example 1.
Once again, a FOPDT model of the process has been
identified by means of the MATLAB function pem, resulting
in:
1
Gnt ,I (s) = e−0.9s . (37)
0.566s + 1
Of course, modeling a strongly underdamped process with a
FOPDT model is not sensible, but in this case it is required
also capture an oscillatory behavior. The delay has been

5517
 1.5
process variable
1 30
0.5
0
0 2 4 6
time
control variable 3 0
2 −10
1 −20
0 −30
0 2 4 6
time
Fig. 6. Step-responses using the integer model with Tm = (1.2L)µ : plant
family response µ = 1.1 and Tm = 1.55 (green lines), nominal closed-
loop response with µ = 1.1 and Tm = 1.55 (solid line) and with µ = 1
and Tm = 1.5 (dashed line) – Example 1.
identified by considering the time interval when the process
step response crosses the threshold of 1% of the steady-state
value. Then, considering that the step response of FFOPDT
model is described by means of the two parameters Mittag-
Leffler function [14], by imposing that that first peak of
the model step response coincides with the first peak of the
process step response, the FFOPDT model has been obtained
as:
1 0
Gnt ,F (s) = e−0.74s . (38)
1.05s1.702 + 1
The uncertainty has been computed as the difference between
the process model and the actual process, normalized with
respect to the process model, that is, Γ(jω) = (G(jω) −
Gnt (jω))/Gnt (jω) Then, Tm = 1.5 has been fixed in order
to set the desired bandwidth. Because of the accurate model,
in the fractional case z can be reduced to 0 preserving the
robust stability. Thus, the selection of z can be done just with
the purpose to speed up or slow down the system response.
Conversely, in the integer case it is necessary to set z = 10
in order to achieve the robust stability. In Figure 7 the robust
stability boundaries and the nominal closed-loop transfer
function frequency responses are shown both in the integer
and in the fractional case: as expected, the robust stability
constraints are much more severe when using an integer
model, because of the non capability of a FOPDT model
to capture underdamped dynamics. Indeed, the capability
of treating in a unified framework (that is, using the same
process model structure) overdamped and underdamped pro-
cesses is a great advantage that the fractional model offers.
The different step responses shown in Figure 8 for different
values of z confirm the role of the parameter z in speeding
up or slowing down the transient response and the major
improvement obtained by using a fractional model. As the
response is satisfactory, there is no need to modify the value
of µ.
VII. C ONCLUSIONS
The proposed fractional H∞ optimal model matching
technique is suitable for FFOPDT processes and allows the
designer to use both fractional-order weighting function and
fractional-order model reference. The nominal and robust
stability issue has been discussed as well as the tuning of
the user-defined parameters. Indeed, the physical meaning
of the parameters has been clarified, thus showing that the
40
20
magnitude [dB]
8 10 10
8 10 −1 0 1
10 10 10
ω
Fig. 7. Robust stability boundaries with Tm = 1.5L and µ = 1: fractional
case (dashed line) and integer case (dash-dot line). Nominal closed-loop
transfer functions: fractional case z = 0.1 (solid line) and integer case
z = 10 (dotted line) – Example 2.
1.5
process variable
1
0.5
0
0 10 20
time
30 40 50
1.5
control variable
1
0.5
0 10 20 30 40 50
time
Fig. 8. Step-responses with Tm = 1.5L and µ = 1: integer model (dotted
line z = 10) and the fractional model for different values of z (dash-dot
line z = 10, dashed line z = 1 and solid line z = 0.1) – Example 2.
methodology can be easily applied in practical cases. It has
also been shown that the additional flexibility in the design
because of dealing with fractional-order systems plays a key
role in improving the control system performance.
R EFERENCES
[1] C. A. Monje, B. M. Vinagre, V. Feliu, and Y. Q. Chen, “Tuning and
auto-tuning of fractional order controllers for industry applications,”
Control Engineering Practice, vol. 16, pp. 798–812, 2008.
[2] D. Valerio and J. Sa da Costa, “A review of tuning methods for frac-
tional PIDs,” in Preprints IFAC Workshop on Fractional Differentiation
and its Applications, Badajoz (E), 2010.
[3] F. Padula and A. Visioli, “Tuning rules for optimal PID and fractional-
order PID controllers,” J. of Process Control, vol. 21, pp. 69–81, 2011.
[4] F. Padula, R. Vilanova, and A. Visioli, “H∞ optimal interpolation
for fractional first order plus dead time systems,” in Proceedings
Symposium on Fractional Signals and Systems, Coimbra (P), 2011.
[5] D. Valerio, “Fractional robust system control,” Ph.D. dissertation,
University of Lisbon (P), 2005.
[6] M. Morari and E. Zafirou, Robust Process Control. Englewood Cliffs
(NY): Prentice-Hall, 1989.
[7] R. Vilanova, “IMC based robust PID design: tuning guidelines and
automatic tuning,” J. of Process Control, vol. 18, pp. 61–70, 2008.
[8] B. Francis, A Course in H∞ Control Theory. Springer Verlag, 1987.
[9] B. Chen, “Controller synthesis of optimal sensitivity: multivariable
case,” in IEEE Proceedings, 1984, pp. 547–551.
[10] G. Zames and B. Francis, “Feedback, min-max sensitivity and optimal
robustness,” IEEE Trans. on Aut. Cont., vol. 28, pp. 585–601, 1983.
[11] D. Matignon, “Stability results for fractional differential equation with
applications to control processing,” Computational Engineering in
System Application, 1996.
[12] S. Alcantara, “Analytical design of feedback compensators based on
robustness/performace and servo/regulator trade-offs,” Ph.D. disserta-
tion, Universitat Autonoma de Barcelona (E), 2011.
[13] F. Padula and A. Visioli, “Tuning of fractional PID controllers for
integral processes,” in Proceedings 4th IFAC Workshop on Fractional
Differentiation and Its Applications, Badajoz (E), 2010.
[14] I. Podlubny, “Fractional-order systems and PIλ Dµ controllers,” IEEE
Transactions on Automatic Control, vol. 44, pp. 208–214, 1999.
5518