Professional Documents
Culture Documents
Power-Shaping of Reaction Systems: The CSTR Case Study: A.Favache, D.Dochain
Power-Shaping of Reaction Systems: The CSTR Case Study: A.Favache, D.Dochain
case study
A.Favache, D.Dochain 1
IMAP, Universite catholique de Louvain, 2 place Sainte Barbe,
B-1348 Louvain-la-Neuve, Belgium
CESAME, Universite catholique de Louvain, 4-6 avenue G. Lemaitre,
B-1348 Louvain-la-Neuve, Belgium
Abstract
The non-isothermal continuous stirred tank reactor (CSTR) is a classical yet complex case study of nonlinear dynamical systems. Power-shaping control is a recent
approach for the control of nonlinear systems based on the physics of the dynamical
system. In this paper we apply the power-shaping control approach to the nonisothermal CSTR study case. A global Lyapunov function is derived for the openloop exothermic CSTR. This Lyapunov function is then reshaped by the means of
a controller in order to stabilize the process at a desired temperature. Some considerations on the local and global convergence to the desired state are presented.
Key words: non-isothermal CSTR, power-shaping control
Introduction
13 October 2009
Power-shaping control
2.1
dx
= P (x) + G (x) uc
dt
(1)
(2)
The function P (x) is called the potential function. In electrical and mechanical systems, the potential function has the units of power and is related to the
dissipated power in the system. In the first one it is related to the so-called
content and co-content of the resistances [18][15]; in the latter it is related to
the Rayleigh dissipation function [14].
Let us now assume the system dynamics is given by the following relation:
dx
= f (x) + g (x) uc
dt
(3)
(4)
(5)
Power-shaping control
Let us assume that the system dynamics can be expressed by using the
Brayton-Moser equations presented before. The desired equilibrium state is
2
is a full-rank left annihilator of g (x) (i.e. g (x) g (x) = 0 with rank g (x) =
N m). Under these conditions, the control input u (x) that re-shapes P (x)
into Pd (x) is the following one:
1
3.1
3.1.1
(8a)
(8b)
(r H)l (T ) rl (T, n)
= Q + Fh +
Cp V
dt
l=1
(9)
xj
=
xi
Nc
X
Nr
X
F in
qik (x)
C i V ni +
il
V
k=1
l=1
Nc
X
"
#!
#!
Nr
X
F in
qjk (x)
Cj V nj +
jl
V
k=1
l=1
"
(10)
The solution of these equations turns out to be quite intricate and we do not
even know if a solution exists. In order to keep the resolution as simple as
possible we shall first work with a simplified version of the CSTR model.
3.1.2
the heat exchange between the reactor and the jacket is proportional to the
temperature difference between them, with the the heat transfer coefficient.
The aim is to control the reactor temperature by acting on the heat transfer,
and more precisely by acting on the cooling fluid flow rate. Changing the
cooling fluid flow rate influences the system dynamics via the heat transfer
coefficient which is dependent on the cooling fluid flow rate. If the jacket
dynamics are fast enough compared to the reactor dynamics, this relation is
time independent. Therefore we shall consider in the sequel that the control
input uc is the quantity CpV . However implementing this control law in practice
requires that the relation between the cooling fluid flow rate and the heat
transfer coefficient is know. The open-loop case corresponds to = 0, i.e. an
adiabatic reactor.
This additional assumptions allows us to rewrite (8) as follows [2][23]:
Nr
X
F in
dni
il rl (T, n)
=
Ci V ni +
dt
V
l=1
(11)
Nr
X
dT
F in
=
l rl (T, n) + uc (Tw T )
T T +
dt
V
l=1
with Tw the (constant) cooling fluid temperature and T in is the inlet temperature.
The time evolution equations of n and T have the same form and the dynamics
of the system can be written as follows:
F in
dx
=
x x + r (x) + uc [0, ..., 0, Tw T ]t
dt
V
(12)
3.2
F in
x x + r (x)
V
and
t
g (x) = 0, , 0, Tw T
The first step to find the Brayton-Moser form of the dynamics is the solution
of a partial differential equation, i.e. find a matrix Q (x) that fulfills (2) and
6
(4). By using (12) we can write the following relation (for i = 1, . . . , Nc + 1):
(Q (x) f (x))i =
NX
c +1
qik
k=1
F in
xk xk + kl rl (x)
V
"
NX
c +1
k=1
qik
xj
F in
xk xk + kl rl (x)
V
NX
Nr
c +1
X
rl
F
kl
= qji +
qjk
V
xi
l=1
k=1
"
NX
c +1
k=1
qjk
xi
(13)
F in
xk xk + kl rl (x)
V
qik
=0
xj
(14)
NX
Nr
c +1
X
F
rl
= qji +
qjk
kl
V
xi
k=1
l=1
"
"
(15)
The above set of algebraic equations is still complex to solve due to the general form of the kinetics. Therefore, we shall also assume that the kinetics of
reaction l can be written as a product as follows:
rl (x) = kl (T )
Nc
Y
(l)
m (nm )
m=1
(l)
The functions (l)
m (nm ) are smooth functions of their argument and m (0) = 0
if the species nm intervenes in the kinetics of reaction l; otherwise
(l)
m (nm ) = 1. kl (T ) is the kinetic coefficient of the reaction l. kl (T ) is such
lim kl (T ) = 0,
dkl
> 0,
dT
lim kl (T ) = kl > 0
T 0
dkl
= 0,
T 0 dT
dkl
=0
T dT
lim
(16)
lim
Remark 1 This assumption restricts the set of reaction kinetics that are considered since it excludes for example the Michaelis-Menten kinetics for enzymatic reactions. However the considered ones still cover a large number of
practical cases, and among them the reactions with mass action law kinetics.
(15) is written as follows (for i, j = 1, . . . , Nc + 1 with i 6= j):
for i, j 6= Nc + 1:
(l)
Nr
X
j
F
kl (T ) (l) (ni )
qij (ni ) +
i
V
nj
l=1
NX
c +1
qik (ni ) kl
! N
Yc
(l)
m (nm )
m6=i,j
k=1
Nr
X
(17a)
(l)
F
i
kl (T ) (l) (nj )
= qji (nj ) +
j
V
ni
l=1
NX
c +1
qjk (nj ) kl
k=1
! N
Yc
(l)
m (xm )
m6=i,j
for j = Nc + 1:
Nr
X
dkl
F
(l)
i (ni )
qij (ni ) +
V
dT
l=1
"
NX
c +1
qik (ni ) kl
k=1
Nc
Y
(l)
m (nm )
m=1,m6=i
(17b)
(l)
Nr
X
F
i
kl (T )
= qji (T ) +
V
ni
l=1
NX
c +1
qjk (T ) kl
k=1
Nc
Y
(l)
m (nm )
m=1,m6=i
A general solution of (17a) and (17b) has not been found, yet. Despite of this,
some characteristics of the solution (if it exists) can be derived. Indeed (17a)
has to be satisfied for any (n, T ) and in particular for T tending to 0. By
taking the limit of (17a) for T tending to 0, we find that the following relation
4
As already mentioned, a general solution of (17a) and (17b) has not been
found, yet. Furthermore to this stage we do not even know if a solution exists.
Therefore we shall in a first instance work with the most simple case, namely
the case of an irreversible reaction with one reactant.
4.1
F in
dT
=
T T + k (T ) nA + uc (Tw T )
dt
V
(18)
(r H)
. It can be shown that this system can have several equilibcp V
rium points depending on the numerical values of the parameters. In this work
we do not consider the cases where limit cycles can occur. In the endothermic case there can be only one asymptotically stable equilibrium, whereas in
the exothermic case an odd number of equilibria can occur. In this last case,
there is an alternance of asymptotically stable and unstable equilibrium points
with =
[2][8][23]. The most common case which can be observed for instance with
Arhenius type kinetics is the case of three equilibrium points where the temperature of the unstable equilibrium is intermediate to the two other (asymptotically stable) equilibrium temperatures.
Proposition 2 The dynamics of the CSTR can be expressed in the form of
the Brayton-Moser equations with:
Q (nA , T ) =
q12
2
F
F 2
k(T )
k(T ) +4( V ) +4 V
k(T )+4 V
k(T )
k(T )
(19)
where 0 < 1 and q12 are two real coefficients such that
q12
< 0.
PROOF. It is sufficient to check that the matrix Q (x) given in (19) is nonsingular and fulfills (2) and (17b) 5 . Indeed the determinant of Q (nA , T ) is
non-zero:
4 2 F 2
det Q (nA , T ) = 2 q12
>0
k
V
Concerning (2) it is sufficient to check the sign of the trace and the determinant
of the symmetric part of Q (nA , T ) (the dependence in T of k (T ) has not been
written for the sake of shortness of the notations):
tr Q + Qt = 2q12
k 2 ( 2 + 1) + 4
2
F
V
k 2
+ 4 VF k
6= 0
2
F 2 q12
(1 ) 0
V
k2
As a consequence the condition (17b) is fulfilled [8].
det Q + Qt = 16
2
Z T
CAin V k ( ) +
F
V
k ( )2
T0
+ k ( ) (T in )
(20)
10
Q n
A , T f n
A , T = P n
A , T = 0
Hence the equilibrium points of the system are the critical points of P (nA , T ).
Proposition 3 The asymptotically stable equilibria are local minima of the
function P (nA , T ) and the unstable equilibria are saddle points.
PROOF. First let us determine the nature of the critical points of P (nA , T )
by computing the eigenvalues of the Hessian matrix H (P) at the equilibrium
points:
2
q12 F
F
1 + 2 + 4
tr (H (P)) = V 2 k T
V
k T
"
det (H (P)) = 4
2
q12
3
F
V
2
k T
F
dk
k T + n
A
V
dT
F
dk
+ k T +
n
A
dT
V
!#
Besides the eigenvalues of the linearized system around the equilibria 1 and
2 are given as follows:
1 =
F
V
and 2 = n
A
dk
F
k T
dT
V
4.2
Controller design
The aim is to design a control law uc = uc (nA , T ) such that the closed loop
system has an asymptotically stable equilibrium at the desired state (nA , T ).
Let us apply the methodology of Section 2.2.
11
80
nA [mol]
unstable equilibrium point
70
60
50
40
30
20
10
0
380
separatrix
400
420
440
460
480
500
520
540
T [K]
Fig. 1. Level curves of P (nA , T ) for the case with three equilibrium points. The
low temperature equilibrium lies outside the frame.
F
V
F
V
CAin V
+ k (T )
(21)
PROOF. (nA , T ) is an equilibrium of the closed loop system, if and only if:
dnA
F in
=
CA V nA k (T ) nA = 0
dt
V
Z T
v ( ) d
(22)
T0
{z
z
F
k(T )2 +4( V
) +4 VF k(T )
where v (T ) =
differentiable function.
k(T )2
12
PROOF. The closed-loop dynamics can be written in the form (6) if and
only if there exists a solution uc (nA , T ) to the equation
Pa (nA , T ) = Q (nA , T ) g (nA , T ) uc (nA , T )
This is the case if and only if Pa(nA ,T ) is the solution of the differential
equation (7). In our case g (x) = 0 with R so that (7) becomes as
follows:
q12
Pa Pa
v (T )
=0
det (Q (x))
nA
T
The solution of this partial differential equation is of the form given in (22).
Proposition 6 Let us denote by
uc (Tw T ) =
F in
T T k (T ) nA
V
the equilibrium value of the control input. If the conditions of Lemmas 4 and
5 are fulfilled and if:
dw
= |q12 | uc (Tw T )
dz z
with
(23)
k VF
1 d2 w
=
>
2
q12 dz 2 z
4 VF k + VF k 2
where
(24)
F
dk
= k +
+
n
V
dT T A
then (nA , T ) is an asymptotically stable equilibrium of the closed-loop dynamics.
If
F 2
F
2
k +
k < 0
4
V
V
then the closed-loop system has no asymptotically stable equilibrium.
(25)
PROOF. The closed-loop system (6) has an asymptotically stable equilibrium at the desired set-point (nA , T ) if and only if (nA , T ) is a local minimum
of Pd (nA , T ). Lemmas 4 and 5 combined with (23) ensure that (nA , T ) is
such that Pd (nA , T ) = 0. To have a minimum the Hessian matrix H (Pd )
of the function Pd (nA , T ) must be positive definite at x = (nA , T ).
H (Pd ) (x ) =
VF
VF v (T )
13
VF
v (T )
q12
H22
The simplest function w (z) that fulfills the above proposition is the following
one:
Pa (nA , T ) = w (z)
h
= |q12 | (z z )2 uc (Tw T ) (z z )
The final expression of the control input is finally deduced from (8):
uc (nA , T ) (Tw T ) =
nA
nA
Z T
T
v ( ) d + u (Tw T )
(26)
The control law has two parameters to be chosen: and . Note that the
possible values are constrained by (24) and (25). Finally let us mention that
the control action u (nA , T ) acts in (18) via the term
g (nA , T ) uc (nA , T ) =
R
T
nA nA T v ( ) d + u (Tw T )
which does not depend on the difference (Tw T ) anymore. This means that
the actual control input is the transferred heat.
4.3
The controller designed in the previous section only guarantees a local convergence to the desired equilibrium point, i.e. the system will converge to desired
point if and only if the initial conditions are close enough to it. Global convergence is ensured if and only if the shaped potential function Pd (nA , T ) does
not have other local minima, i.e. if and only if x is a global minimum of the
function Pd (nA , T ).
By analogy with (21) let us define the function neq
A (T ) by the following relation:
F in
C V
eq
V A
nA (T ) = F
+ k (T )
V
14
F
F in
CA V neq
(Tin T )
A (T ) +
V
V
+ uc (Tw T ) neq
A (T ) nA +
ZT
v ( ) d (27)
{z
z (neq
A (T ),T )z
#
Lemma 7 A state n#
is an equilibrium state of the closed loop system
A, T
(18) if and only if it is contained in the following set:
eq
#
= T # = 0, n#
A = nA T
o
#
PROOF. To prove the only-if part, let n#
be an equilibrium state
A, T
of the closed loop system. Hence we have:
dnA
=0=
dt
dT
=0=
dt
F
V
F
V
#
CAin V n#
n#
A k T
A
T in T # + k T # n#
A
nA
n#
A
Z T#
T
(28)
v ( ) d
+ uc (Tw T )
#
that is contained in .
To prove the if part, let us consider a state n#
A, T
#
By replacing nA by the expression of neq
in the first equation of (28), it
A T
is obvious that
dynamics:
dnA
dt
#
= 0 if n#
. Let us now consider the temperature
A, T
dT
F in
=
T T # + k T # n#
A
dt
V
#
+ uc n#
Tw T #
A, T
F in
#
=
T
T T # + k T # neq
A
V
!
nA
neq
A
+ uc (Tw T )
15
Z T#
T
v ( ) d
dT
dt
= 0.
(T ) k 2 (T )
4 VF k (T ) +
F
V
(29)
2
dk
where (T ) = k (T ) + VF + dT
neq
A (T ), then there exists a lower bound
ZT
is given as follows:
k 2 ( ) ( ) + 4 VF k ( ) +
k 2 ( )
F
V
F
V
2
+ k ( )
F
(T ) k 2
V
>
k 2 (T ) + 4 VF k + VF
(30)
then the controller (26) achieves global stability of the desired equilibrium
point.
PROOF. (29) and (30) imply (25) and (24), respectively. Using Proposition 6
the desired equilibrium is locally stable.
16
(T ) [J/min]
increasing
0 (T )
desired set point
T [K]
Fig. 2. Influence of
on (T )
k (T )2 (T ) + 4 VF k +
k (T )2 k (T ) +
F
V
F
V
< 0 for all T > 0. (T ) is monotonConditions (29) and (30) imply that d
dT
ically increasing and hence has a unique zero, namely the desired closed-loop
equilibrium.
Remark 10 The conditions for the global convergence (29) and (30) are very
similar to the ones for the local convergence (25) and (24), respectively, but
instead of being satisfied only at the desired set point temperature T , they
have to be satisfied at any T .
The shape of the function Pd (nA , T ) is shown in Figures 3 and 4. In the
first case, the parameters and were chosen such that local convergence is
achieved (i.e. (24) is fulfilled but not (30)). The function Pd (nA , T ) has two
local minima and consequently the convergence to the desired state is only
local. In the second case the parameters and were chosen such that global
convergence is achieved (i.e. (30) is fulfilled). The function Pd (nA , T ) has a
single minimum and the convergence to the desired state is global.
Remark 11 Even if the conditions of Proposition 9 are satisfied, global convergence can be lost if the input is bounded. Local convergence will however be
kept. It would be interesting to discuss the size of the attraction region with the
proposed controller in case of input constraints compared to other controller
proposed in the literature (e.g. in [24]).
17
nA [mol]
other convergence point
70
#
(n#
A, T )
60
50
40
locus of possible
closed-loop equilibria
30
20
390
400
410
420
430
440
450
460
470
480
490
500
T [K]
nA [mol]
70
60
locus of possible
closed-loop equilibria
50
40
30
20
10
0
420
460
480
500
520
540
560
T [K]
The power-shaping approach has given interesting results on the case study
of a CSTR with an irreversible reaction. It is therefore of major interest to
see if this approach can be extended to more complex systems, and more
particularly to systems with more than one reacting chemical species and/or
with multiple reactions.
We shall now look at some particular cases of more complex reactions to get
18
m
(l)
m (nm ) = (nm )
It has been shown that in some cases (17a) and (17b) have no solution. This
does not mean that the dynamics cannot be put into the Brayton-Moser form.
But if the Brayton-Moser form exists, then the relation (14) is not fulfilled and
the partial differential equation (13) has to be solved. In the other cases a general solution has been found, but the obtained matrix may have a non-negative
symmetric part. Therefore the corresponding potential functions cannot be
used in general as a Lyapunov functions for the open-loop system because
they may be non-decreasing along the system trajectories. As shown in [12],
given a matrix Q (x) that fulfills (4) other solutions to (4) can be built departing from the first one. Applying this methodology on the matrices that have
been found could lead to another matrix Q (x) that would be negative-definite
and hence give a Lyapunov function for the open-loop system.
6.1
General solution
As it has been explained in Section 3.1.1 the solution to (10) is quite intricate.
One of the difficulties is due to the term in VF . This difficulty can be overcome
by also considering the dilution coefficient VF as an input of the system. This
corresponds to the system studied in [22]. The Brayton-Moser formulation
of the CSTR has been found with the same assumptions as in Section 3.1.2.
However, as the authors mention, they did not manage to extend their result
to more general cases.
Contrary to [22] who used the simplified form of the dynamics (12) with mass
action law dynamics we shall use the more general form (8). Let us consider
19
a reversible reaction A
B. Using the notations of (3) we have:
f (nA , nB , U ) =
r (T, nA , nB ) V
r (T, nA , nB ) V
g (nA , nB , U ) =
CAin V
nA
CBin V nB
(nA , nB , U ) 1
V hin H
with uc =
F
V
t
1, 2, 3):
(31)
Z
Z
Z
d
d
dnA dnB +
dnA
dU
dU
Z
dnB + (U )
+
U
Z Z
Q (x) =
P (nA , nB , U ) =
ZZ
20
(nA , U ) dnA
(nB , U ) dnB +
(U ) dU +
where is some real constant. The functions qij (x) can be chosen arbitrarily as
far as Q (x) is non-singular and (2) is fulfilled. Finally let us mention that the
solution to the partial differential equation system has been obtained without
restricting to a particular subset of solutions, contrary to what has been done
in the previous sections.
6.2
Particular solution
2 S
nA nB
=0
(U, nA ) = 0
(U, nB ) = 0
(U ) = 0
q32 = 0
q13 = 0
q23 = 0
where S (nA , nB , U ) is the entropy function. The potential P (x) is the opposite of the entropy function S (nA , nB , U ) with the following matrix:
Q (x) =
B
T rV
T rV
4
B 4
2
T rV
T rV
T rV
1
T rV
0
0
4(1)T rV
(32)
(33)
21
and U = U 0
(34)
Using (34) as the invariant set, S (x) as a Lyapunov function candidate and
(33), the stability is deduced from La Salles theorem [16].
However the power-shaping methodology cannot be applied using the above
proposed matrix Q (x) and potential S (x). Indeed the matrix Q (x) is not
defined at the open-loop equilibria and Q1 (x) is singular at these points.
Contrary to what has been shown in Section 4 the gradient of the potential
function is not zero at the open-loop equilibria.
Remark 12 Note that the availability function proposed by Alonso, Ydstie
and coworkers (e.g. [1]) can also be used as a potential function for the BraytonMoser form of the dynamics:
ref
ref
1
T ref
nref nA
A
ref
nB
nB
U ref U
ref
ref
is a reference state and S ref = S xref . The
where xref = nref
A , nB , U
corresponding matrix Q (x) is very similar to (32) and is also not defined at
the open-loop equilibria.
Conclusion
shown that this methodology needs to be extended in the case of more complex
reaction kinetics. Finally first results for applying the power-shaping approach
in the case of a general CSTR with two control inputs have been presented.
These results show how the power-shaping approach can be related to the
entropy of the system. The work presented in [12] offers promising tools to
extend these results, by instance for finding alternative potential functions
departing from the proposed one.
Acknowledgements
This paper presents research results of the Belgian Network DYSCO (Dynamical
Systems, Control, and Optimization), funded by the Interuniversity Attraction Poles
Programme, initiated by the Belgian State, Science Policy Office. The scientific
responsibility rests with its authors. The work of A. Favache is funded by a grant
of aspirant of Fonds National de la Recherche Scientifique (Belgium).
References
[1] A.A. Alonso, B.E. Ydstie, and J.R. Banga. From irreversible thermodynamics
to a robust control theory for distributed process systems. Journal of Process
Control, 12:507517, 2002.
[2] R. Aris and N.R. Amundson. An analysis of chemical reactor stability and
controlI : The possibility of local control, with perfect or imperfect control
mechanisms. Chemical Engineering Science, 7(3):121131, 1958.
[3] R.K. Brayton and J.K. Moser. A theory of nonlinear networks I. Quaterly of
applied mathematics, 22:133, 1964.
[4] M. Dalsmo and A.J. van der Schaft. On representations and integrability of
mathematical structures in energy-conserving physical systems. Journal on
Control and Optimization, 37(1):5491, 1998.
[5] D. Eberard.
Extension des syst`emes hamiltoniens ports aux syst`emes
irreversibles: une approche par la geometrie de contact. PhD thesis, Universite
Claude Bernard, Lyon 1, 2006.
[6] D. Eberard, B.M. Maschke, and A.J. van der Schaft. Port contact systems
for irreversible thermodynamical systems. In Proc. 44th IEEE Conference
on Decision and Control and European Control Conference, Sevilla (Spain),
December 2005.
[7] D. Eberard, B.M. Maschke, and A.J. van der Schaft. Energy-conserving
formulation of RLC-circuits with linear resistors. In Proc. 17th International
23
24
25