Professional Documents
Culture Documents
Faouzi Bouani
Université de Tunis El Manar, Ecole Nationale d’Ingénieurs de Tunis,
LR11ES20 Laboratoire Analyse, Conception et Commande des Systèmes, 1002, Tunis, Tunisie;
d
Email : faouzi.bouani@enit.rnu.tn
ite
Badreddine Bouzouita
ed
Université de Tunis El Manar, Ecole Nationale d’Ingénieurs de Tunis,
LR11ES20 Laboratoire Analyse, Conception et Commande des Systèmes, 1002, Tunis, Tunisie;
py
Université de Sousse, Ecole Nationale d’Ingénieurs de Sousse, 4054, Sousse, Tunisie;
Email : badreddine.bouzouita@enit.rnu.tn
Co
Mekki ksouri
Université de Tunis El Manar, Ecole Nationale d’Ingénieurs de Tunis,
ot
LR11ES20 Laboratoire Analyse, Conception et Commande des Systèmes, 1002, Tunis, Tunisie;
tN
Email : mekki.ksouri@enit.rnu.tn
rip
Abstract
This paper provides on Model Predictive Control (MPC) of fractional order systems. The direct method will be used as internal
sc
model to predict the future dynamic behavior of the process which is used to achieve the control law. This method is based on the
Grünwald-Letnikov’s definition that consists of replacing the non integer derivation operator of the adopted system representation
nu
by a discrete approximation. The performances and the efficiency of this approach are illustrated with practical results on a ther-
mal system and compared to MPC based on integer ARX model.
Ma
Keywords: Model predictive control; fractional order systems; thermal system; Grünwald-Letnikov method.
ed
pt
1. Introduction
ce
The concept of fractional calculus has been defined on the 19th century by Riemann and Liouville. The goal was
Ac
to extend the integration or derivation of the fractional order by using the integer orders as well as fractional orders.
Later, in the 20th century Grünewald-Letnikov introduced the notion of fractional-order discrete difference. In recent
years, non integer order calculus, also known as fractional calculus has attracted the attention of researchers in several
fields such as engineering [1] [2], biology [3], economics [4] etc. It was found that many physical systems have shown
a dynamic behavior of non integer order. Probably, the first dynamic physical system to be widely recognized is the
Aymen Rhouma 1
diffusion of heat into semi-infinite (thermal system) [5]. Other well known fractional systems include the viscoelastic
systems, the electrode-electrolyte polarization, the electromagnetic waves and many others [6]. The fractional system
is commonly used in industrial processes particularly in the application of modeling, identification and control [7-10].
The idea of using the fractional order regulator to control the dynamic systems was proposed by Oustaloup in 1988
[11] [12]. Later, in 1994, Podlubny proposed the non integer order PID controller using mainly integrals and deriva-
tives of fractional order [13]. In [14] authors have used the simple state-space realization and the approximation tech-
nique for non integer order controllers. In [15], Monje and Felui were interested in the non integer order lead compen-
d
ite
sator. In Agrawal [16], a new formulation based on the fractional differential equations as well as the solution scheme
ed
for fractional optimal control problems of fractional systems is proposed. Vinagre et al. [17] have used the non integer
py
order calculus in classic model reference adaptive control. In [18], Dadras and Momeni have used the fractional-order
Co
sliding mode control of systems with uncertainty. Ying et al. [19] have proposed two methods of non integer order
proportional integral controllers of fractional order systems. Hitay et al. [20] developed a method to design PID con-
Moreover, the MPC has become a mature control strategy over the last few years. The reason of this success is at-
rip
tributed to the consideration of different types of constraints on input and output signals. It can also handle a large
sc
class of systems such as open-loop unstable systems, non-minimum phase systems, delayed systems and multivariable
nu
systems [21]. Therefore, the model predictive control is widely encountered in the industrial processes [22]. MPC is a
Ma
control technique that optimizes a cost function by using a model to predict the future behavior of process output.
ed
Indeed, the presence of the model is necessary for the development of the predictive control.
pt
The originality of this work lies in applying the MPC of fractional order systems. The system is approximated
ce
with a direct method that is based on the numerical evaluation of non integer order operators. This method is founded
Ac
The outline of this paper is organized as follows. In section 2, the basic definitions as well as the modeling of frac-
tional order systems are introduced as a preliminary step, where the Grünwald-Letnikov method used to approximate
the fractional system is detailed. Section 3 is reserved to focus on the necessary steps in finding the optimal control
law of MPC for fractional systems. To illustrate the effectiveness of the proposed approach, some simulation results
Aymen Rhouma 2
are given in section 4. The experimental results on a thermal system are exhibited in section 5. Finally, a conclusion is
given.
d
0
dt
d
D
t0 t 1 =0 (1)
ite
t
(d ) <0
t0
ed
where is the order, R , t and t0 are respectively the upper and lower limits of the operation. There are several
py
Co
definitions of fractional order calculus, two of the most commonly used definitions are Grünwald-Letnikov (G-L) and
1 t t0 / h
Ma
( 1) ( i 1)
where h is the sampling period, R and means:
i i i!
pt
ce
The G-L definition is the most popular one for fractional-order control and its application [24-25] is also used in this
Ac
paper.
The Laplace transform of G-L definition for zero initial conditions can be given as:
In general, a fractional model can be described by a fractional differential equation characterized by the following
form:
Aymen Rhouma 3
Using the Laplace transform in Eqn. (5), the fractional-order system can be represented by the following transfer func-
tion:
M
bm s bm
Y ( s) m 0
G( s) (6)
U ( s) L al
al s
l 0
where al , bm R , al , bm
2
R2
d
ite
The equation (3) can be used to numerically evaluate the derivative or the integral of fractional order by ap-
ed
proximating the sampling rate with a more appropriate value [26]. Hence, by removing the limit, one obtains a dis-
py
crete approximation for the fractional derivative [24-25]:
1 t t0 / h
Co
i
t0 Dt f (t )
(1) f (t ih) (7)
h i 0 i
ot
Summation series is contrasted with a number of terms which increases when the term ‘ h ’ decreases. For a
tN
given approximation error ‘ ’, this series can be reduced to a limited number ‘ N ’ of terms, fixed by using the short
rip
memory principle.
sc
Theorem (principle of short memory) [13]: If a temporal function ‘ f ’ is bounded on an interval t0 , t , there exists a
nu
ML
gives an error such that:
(1 )
pt
1 N
t0 Dt f (t )
(1)i f (t ih) (8)
h i 0 i
Ac
The objective of the next section is to propose the MPC that is based on the use of a fractional order model which
Aymen Rhouma 4
MPC for the fractional systems. Consequently, the G-L definition of fractional system represented in section 2 will be
used to obtain the fractional order model. Therefore, it is assumed that this fractional system is described by the model
d
For obvious reasons and without loss of generality, we will express y (k ) in terms of u(k 1) , and depending on
ite
the input deviation. By considering a noise sequence with zero mean and finite variance, the Eqn. (9) becomes:
ed
1 M bm N i b 1 L a N al
1 m 1 y (k i) e k
py
i
y (k ) bm u (k 1 i) L
l
(10)
L al m 0 h i 0 i al l 0 h l i 1
a
i
al
Co
l 0 h l 0 h al
1 q 1 , is an integral action introduced in order to obtain, in closed loop, a zero steady state error.
ot
tN
By using the Eqn. (10), we obtain the predicted output of the system in k 1 :
yˆ (k 1/ k ) yl (k 1) 1u(k )
rip
(11)
1 M bm
where: 1
sc
L al bm
m 0 h
al
h l 0
nu
yl (k 1) y(k ) s1 s2 (12)
1 M bm N i b
s1 1 m u (k i )
ed
L al m 0 h bm i 1 i
l 0 h
pt
a l
1 L al N i a
1 l y (k 1 i )
ce
s2
L a l 0 al i 1
h
l i
Ac
l 0 h a l
where:
1 M bm 1 L al
1 bm
bm ; 2 al
al
L al m 0 h L al l 0 h
al
al
l 0 h l 0 h
Aymen Rhouma 5
1 M b N i b
s3 m 1 m u (k 1 i )
L a m 0 h bm i 2
l i
l 0 h l
a
1 L al N i a
s4 1 l y (k 2 i)
L a l 0 al i 2
h
l i
l 0 h la
d
ite
yˆ (k 2 / k ) 1 2 yl (k 1) 2 u (k ) 1 u (k 1) 2 y (k ) s3 s4 (14)
ed
where: 2 1 2 1 1
py
we set: yl (k 2) 1 2 yl (k 1) 2 y(k ) s3 s4
Co
then:
yˆ (k 2 / k ) yl (k 2) 1u(k 1) 2 u(k ) ( 15)
ot
Consequently, the expression of the j-step ahead predictor yˆ (k j / k ) is as follows:
tN
rip
j
yˆ (k j / k ) j i 1 u(k i 1) yl (k j ) (16)
i 1
sc
The future control sequence over a control horizon Hc is computed by minimizing a cost function which indi-
Ma
cates how well the process follows the desired trajectory. This function can be expressed by the future errors between
ed
output signals and setpoints, and the future incremental control signals. The cost function is given by the following
pt
equation:
ce
Hp Hc1
J ( yˆ (k j / k ) yc(k j )) 2 u (k i) 2 (17)
j 1 i 0
Ac
Y GU Yl (18)
where:
T
Y yˆ (k 1/ k ),..., yˆ (k Hp / k
U u (k ),..., u (k Hc 1)
T
T
Yl yl (k 1),..., yl (k Hp )
Aymen Rhouma 6
where is the weighting factor, Yc is the sequence of set-points on the prediction horizon.
d
ite
By minimizing Eqn. (19), we obtain the optimal control sequence.
ed
(20)
py
4. Simulation results
Co
In this section, the performance of the proposed method of MPC for the fractional systems will be illustrated with
simulation example. Consider the fractional order system given by the following equation [13]:
ot
tN
G (s) 1
(21)
0.8 s 2.2 0.5 s 0.9 1
rip
This transfer function can be expressed by the Eqn. (5) with the following parameters:
L 2; M 0
sc
b0 1; a0 1; a1 0.5; a2 0.8
nu
0; 0; 0.9; 2.2
b0 a0 a1 a2
Ma
Consequently, the j-step ahead predictor can be computed as described in the section 3 with sample time h
equals to 0.1. The efficiency of the proposed MPC is compared with the classic MPC based on Oustaloup method [28-
ed
30]. This method consists of replacing the fractional operator by its integer order approximation. Therefore, an integer
pt
order system is obtained. The generalized Oustaloup method is based on replacing the term s , R , by N linear
ce
Ac
N s wk'
c (22)
k 1 s wk
Aymen Rhouma 7
The precision of the model obtained by the oustaloup method strongly depends on the choice of the value N.
A high value of N leads to a high computational time. After simulation trials we have choose the following values:
In this case, by using the function “ousta_fod” of MATLAB [31], the Eqn. (21) is expressed by the integer order mod-
N ( s)
C s (23)
D( s )
where:
d
N ( s) s12 1.433e004s11 5.45e007 s10 5.792e010s9
ite
2.008e013s8 2.081e015s 7 7.15e016s 6
ed
7.382e017s5 2.528e018s 4 2.587e018s3
8.637e017s 2 8.056e016s 1.995e015
py
D( s) 5.048s14 6.407e004s13 1.881e008s12
Co
1.651e011s11 4.401e013s10 +3.752e015s 9
9.959e016s8 8.669e017s 7 2.586e018s 6
ot
3.764e018s5 4.379e018s 4 3.14e018s3
tN
9.233e017s 2 8.257e016s 2.011e015
rip
Figure 1 shows the step responses of the fractional order system given by the Eqn. (21) using the function
sc
“fotf” of MATLAB [31], the G-L method and the oustaloup method. The errors of these last two methods are plotted
nu
in Fig. 2. From the results, we deduce that the output computed using the G-L method is closer to the fractional sys-
Ma
1.6
Fractional system
1.4 Oustaloup method
pt
G-L method
1.2
ce
1
Ac Amplitude
0.8
0.6
0.4
0.2
-0.2
0 5 10 15 20 25 30
Time(sec)
Fig. 1. Step responses.
Aymen Rhouma 8
0.1
0.05
Amplitude
0
-0.05
-0.1
-0.15
-0.2
0 5 10 15 20 25 30
d
Time(sec)
Fig. 2. Errors for two methods.
ite
ed
Figure 3 compares the closed loop performances of the proposed controller based on the G-L method and the
py
classical MPC based on the Ostaloup method. The control parameters are the following: Hp 4, Hc 1 and =0.5
Co
The choice of these parameters is explained in [32]. From the simulation results, we notice that the new predictive
controller has the advantage of predicting the behavior of the output with respect to changes of the set-point.
ot
tN
rip
Output Signal
2.5
sc
Set-point
2 Oustaloup method
G-L method
nu
1.5
output
Ma
1
ed
0.5
pt
0
20 40 60 80 100 120 140 160
Iteration
ce
Input signal
2.5
Oustaloup method
G-L method
Ac
1.5
Input
0.5
0
20 40 60 80 100 120 140 160
Iteration
Aymen Rhouma 9
5. Practical results
In this section, we proceed by applying the new predictive controller for the fractional order systems developed in
this paper to a thermal system depicted in Fig. 4. The efficiency of proposed controller is compared with classic MPC
based on integer ARX model (Auto Regressive model with eXternal inputs).
PWM Cylinder
d
Converter
oide
ite
Heating resistor
ed
Temperature
sensor
py
Data acquisi- Amplifier
Computer
Co
tion module
ot
Fig. 4. Thermal system.
tN
rip
The studied process is composed as shown in figure 4 by an aluminum cylinder of 2 cm section and 41 cm
nu
length. The cylinder is submitted to a heating resistor and thermally isolated to insure a unidirectional transfer of the
Ma
heat flux. The thermal system is considered as a semi-infinite dimension because its length is more important when
ed
compared to its section. The input signal of this system is a thermal flux which is generated by a heating resistor. It is
fixed in one of the cylinder’s extremities and controlled by a computer with USB data acquisition module. The power
pt
ce
interface separating the controller from the heating resistance is a PWM converter with an input voltage varying from
Ac
0 to 5v. The output signal of the system is the cylinder temperature measured with a distance ‘d’ from the heated sur-
face by an LM35DZ sensor. The sensor signal is amplified and conditioned in a stage realized for these purposes to
In order to demonstrate the non integer behavior of this thermal system, the cylinder is modeled according to the fol-
lowing assumptions:
Aymen Rhouma 10
Losses on the surface, where the thermal flux is applied, are neglected.
Several approaches have been proposed to model the phenomenon of a thermal system [33]. A solution was pro-
posed by Cois [34] it is to show that the model of this phenomenon is of fractional order medium which has a com-
d
5.2. Identification
ite
Firstly, we have applied various step input signals to the thermal system with different amplitudes. The sam-
ed
ple time is chosen equal to 5 sec. The obtained open loop step responses are depicted in Fig. 5.
py
The table 1 shows the static gain (K), the response time (T) and the time delay (D) for each step input signal. We note
Co
that ambient temperature of 15.4°C, corresponding to 0.77v.
80 ot
tN
70 u=4v
60 u=3v
rip
Temperature(C°)
50 u=2.5v
sc
40
u=2v
nu
30
u=1.5v
20
Ma
10
0 500 1000 1500 2000 2500 3000
Iteration
Fig. 5. Open loop step responses.
ed
Based on the results presented on Tab. 1, we deduce that the time response is about 2000 sec and the time de-
lay is about 100 sec. Consequently, we can adopt as sample time 40 sec.
In order to estimate the fractional order model of the process, we have applied to the heating resistor the input
sequence given by the Fig. 6. This last depicts also the evolution of the temperature, at distance of d=15 cm of the
Aymen Rhouma 11
Measured output
10*Input
60
50
40
30
20
10
0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Iteration
d
70
Measured output
ite
Fractional model
60 Integer model
10*Input
ed
50
40
py
30
Co
20
10
0
ot
tN
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Iteration
Fig. 7. Validation data.
rip
The fractional model is established from the data identification using the simplified refined instrumental vari-
sc
0.8623
H1 (s) e100s (24)
506.2843s 1.5
135.3925s 6.3598s 0.5
1
Ma
For comparison purposes, we have identified the thermal system by an integer order model. Hence, by using
ed
the toolbox ident of Matlab, we were able to determine this model which is given by:
pt
0.712
H 2 ( s) e100s
ce
(25)
30.77s 236.84s 1
2
Ac
In order to test the both models performances, we have performed another input excitation sequence and we
have measured the corresponding temperature. As represented in the validation data of Fig. 7, we deduce that the iden-
tified fractional model is closer to the measured output than to the integer model. The Normalized Mean Squared Er-
ror (NMSE) [36] computed on validation data for the fractional order model is NMSE ( H1) 8.8 104 , whereas for the
Aymen Rhouma 12
integer model is NMSE( H2 ) 5.7 103 . Therefore, the fractional identification is more adapted than integer one with this
thermal system.
The goal is to maintain the temperature of the system measured at 15 cm from the heated surface by adjust-
ment the power of the heating resistor obtained with application of the MPC based on the G-L method that developed
d
in this paper and MPC based on integer model.
ite
In all experiences, the sample time is equal to 40 sec and the control signal is limited between the following values:
ed
0 u(k ) 5v
py
In the case of MPC proposed in this paper, the j-step ahead prediction is expressed by Eqn. (10) with the following
Co
parameters:
ot
h 0.1; L 3; M 0; b0 0.8623
tN
a0 1; a1 6.3598; a2 135.3925; a3 506.2843
0; 0; 0.5; 1; 1.5
b0
rip
a0 a1 a2 a3
Figure 8 exhibits the measured temperature, the setpoint and the control signal when the proposed controller is
sc
In the second experiment, we have increased the prediction horizon from Hp 8 to Hp 12 . The closed loop re-
Ma
Aymen Rhouma 13
60
50
Temperature(C°)
40
30
20
10
Output
set point
0
0 100 200 300 400 500 600
Iteration
Control signal
d
5
ite
4.5
ed
3.5
3
Input(v)
py
2.5
Co
1.5
0.5
Input
0
0 100 200 300
Iteration
ot
400
60
sc
50
Temperature(C°)
40
nu
30
Ma
20
10
Output
ed
set point
0
0 100 200 300 400 500 600
pt
Iteration
Control signal
5
ce
4.5
Ac
3.5
3
Input(v)
2.5
1.5
0.5
input
0
0 100 200 300 400 500 600
Iteration
Aymen Rhouma 14
Based on practical results shown in figures 8 and 9, it is clear that the temperature follows the desired set
points. Consequently, these results show good performances of the proposed approach. But in the first experience
when the prediction horizon is equal to 8, the input/output signals of the thermal system contain oscillations. In fact,
when we have increased the prediction horizon from 8 to 12, the response of the thermal system becomes faster.
In these experiences, the control signal is saturated and attained the maximal value. In order to avoid the control input
saturation, we have introduced a model for the sequence of set-points which is given by:
d
1
Hc s (26)
1 .s
ite
The constant is used to modify the closed loop dynamic system. In this case, we choose this constant equals
ed
py
to 2800 sec. The evolutions of the set point, the control signal and the measured temperature (output signal) obtained
Co
with the proposed MPC with Hp = 12 are represented in Fig. 10. Based in these results, we notice that the measured
temperature meets the desired requirements. We remark also that the control signal not saturated.
Set point and output
ot
tN
70
60
rip
50
sc
Temperature(C°)
40
nu
30
20
Ma
10
Output
set point
ed
0
0 100 200 300 400 500 600
Iteration
Control signal
pt
4.5
ce
4
Ac
3.5
3
Input(v)
2.5
1.5
0.5
Input
0
0 100 200 300 400 500 600
Iteration
Aymen Rhouma 15
A comparison of the closed-loop performances of the proposed approach is established with classic MPC
based on integer order model given by the Eqn. (25). The both predictive controllers are designed with the following
parameters:
Figure 11 presents the outputs and control signals evaluation obtained by the predictive controller based on in-
teger order model. Based on this figure, we deduce that the measured temperature and the control signal present many
fluctuations. We can compute the average and the variance of the control sequence, respectively by 1
M kM1 u (k ) and
d
ite
kM1 u (k ) , with M is the number of iterations.
1 2
M
ed
Based on the results presented on Tab. 2, we notice also that the control sequence obtained with the proposed
py
MPC based on the fractional model, presents the least control variance compared to the control sequence obtained
Co
with the classic MPC founded on the integer model.
ot
tN
Table2. Average and variance of the control sequence.
Average Variance
rip
Aymen Rhouma 16
60
50
Temperature(C°)
40
30
20
10
Output
Set point
0
0 100 200 300 400 500 600
Iteration
Control signal
5
d
4.5
ite
4
ed
3.5
3
Input(v)
py
2.5
Co
1.5
0.5
0
0 100 200 300
ot
400 500
Input
600
tN
Iteration
6. Conclusion
sc
In this paper, a new method of MPC has been introduced to fractional order systems. This method is based on the
nu
Grünwald-Letnikov’s definition. Therefore, the output deviation approach is used to design the j-step ahead output
Ma
predictor and the control law is obtained by solving a quadratic cost function. Experimental results on a thermal sys-
ed
tem show that the predictive controller using the Grünwald-Letnikov method exhibits good performance and it is more
pt
REFERENCES
[1] Bagley, R., and Calico R., 1991, Fractional order state equations for the control of viscoelastically damped structures, Journal
of Guidance, Control, and Dynamics, 14, pp.304–311.
[2] Zhang, Y., Tian, Q., Chen, L., and Yang J., 2009, Simulation of a viscoelastic flexible multibody system using absolute nodal
coordinate and fractional derivative methods, Multibody System Dynamics, 21, pp.281–303.
[3] Yuste, S., Abad, E., and Lindenberg K., 2010, Application of fractional calculus to reaction-subdiffusion processes and
morphogen gradient formation, Arxiv preprint arXiv, 1006.2661.
[4] Mainardi, F., Raberto, M., Gorenflo, R., and Scalas, E., 2000, Fractional calculus and continuous-time finance II: the
waiting-time distribution. Physica A: Statistical Mechanics and its Applications, 287, pp.468–481.
Aymen Rhouma 17
d
a fundamental idea from a new process control strategy, 12th IMACS World Conf., Paris, pp.203–208.
ite
[12] Oustaloup, A., 1991, La commande CRONE (Commande Robuste d’Ordre Non Entier). Paris, Hermès.
ed
[13] Podlubny, I., 1999, Fractional Differential Equations, Academie Press, New York.
[14] Raynaud, H.F., and Zergainoh A., 2000, State-space representation for fractional order controllers. Automatica, 36, pp.1017–
py
1021.
[15] Monje, C. A., and Feliu, V., 2004, The fractional-order lead compensator, IEEE International Conference on Computational
Co
Cybernetics, Vienna, Austria.
[16] Agrawal, O.P., 2004, A general formulation and solution scheme for fractional optimal control problems, Nonlinear
Dynamics, 38, pp.323–337. ot
tN
[17] Vinagre, B.M., Petras, I., Podlubny, I., and Chen Y.Q., 2002, Using fractional-order adjustment rules and fractional-order
reference models in model reference adaptive control, Nonlinear Dynamics, 29, pp.269–279.
rip
[18] Dadras, S., Momeni, H.R., 2012, Fractional terminal sliding mode control design for a class of dynamical systems with
uncertainty, Communications in Nonlinear Science and Numerical Simulation, 17, pp.367-377.
sc
[19] Ying, L., Yang, Q.C., Chun, Y.W., and You, G.P., 2010, Tuning fractional order proportional integral controllers for
fractional order systems. Journal of Process Control, 20, pp. 823–831.
nu
[20] Hitay, O., Catherine, B., and André R.F., 2012, PID controller design for fractional-order systems with time delays, Systems
& Control Letters, 61, pp.18–23.
Ma
[21] Fukushima, H., Kim, T., and Sugie T., 2007, Adaptive model predictive control for a class of constrained linear systems
based on comparison model, Automatica, 43, pp.301–308.
ed
[22] Camacho, E.F., and Bordons C., 2004, Model Predictive Control. Springer-Verlag, Berlin.
[23] Tavazoei, M.S., 2010, A note on fractional-order derivatives of periodic functions, Automatica, 46, pp.945–948.
pt
[24] Hajiloo, A., and Nariman-zadeh, N., and Moeini A., 2012, Pareto optimal robust design of fractional-order PID controllers
ce
Aymen Rhouma 18
d
[36] Gabano, J.D., and Poinot T., 2011, Fractional modelling and identification of thermal systems. Signal Processing, 91, pp.531-
ite
541.
ed
Figure Captions List
py
Fig. 1 Step responses
Co
Fig. 2 Errors for two methods
ot
tN
Fig. 3 Closed-loop responses
rip
Aymen Rhouma 19
d
ite
ed
py
Co
ot
tN
rip
sc
nu
Ma
ed
pt
ce
Ac
Aymen Rhouma 20