Sei sulla pagina 1di 17

Automatica 37 (2001) 1739}1755

Stabilization of distributed systems using irreversible


thermodynamics
Antonio A. Alonso , B. Erik Ydstie *
Department of Chemical Engineering, Universidade de Vigo, Apto 874, 36200 Vigo, Spain
Department of Chemical Engineering, Carnegie Mellon University, Pittsburgh PA 15213, USA
Received 29 June 1999; revised 12 March 2001; received in "nal form 20 April 2001

Abstract
We connect thermodynamics and the passivity theory of nonlinear control. The storage function is derived from the convexity of
the entropy and is closely related to the thermodynamic availability. We relate dissipation to positivity of the entropy production. In
this form the supply function is a product of force and #ow variables in deviation form. Feedback signals originate from intensive
variables like temperature, pressure and composition. We show that the physical dimension of the system matters: The larger the
distributed system is, the more di$cult the stationary state may be to stabilize. Any chemical process can be stabilized by distributed
PID control provided that the sensor and actuator locations are suitable. We apply the results to heat conduction and reaction
di!usion equations.  2001 Elsevier Science Ltd. All rights reserved.
Keywords: Chemical reaction; Convex analysis; Distributed system; Heat conduction; Irreversible process; Nonlinear system; Passive system; Process
control; Thermodynamics

1. Introduction
We call things we don't understand complex, but that
means we have not found a good way of thinking about
them.
Tsutomu Shimomuro
The problem we consider in this paper is how to
stabilize stationary solutions of distributed parameter
systems (DPS). Stabilization of DPS constitutes one of
the central problems in chemical process control. For
example, many chemical reactions, chemical vapor deposition, crystallization, aluminum production, glass
production and silicon production, to name a few, are
most naturally modelled as DPS.
One of the "rst comprehensive reviews of distributed
control of chemical processes was given by Ray (1978).
Christo"des (2000) gives a more current review and de
This paper was not presented at any IFAC meeting. This paper
was recommended for publication in revised form by Editor Sigurd
Skogestad.
* Corresponding author.
E-mail addresses: aalvarez@uvigo.es (A. A. Alonso), ydstie@
andrew.cmv.edu (B. E. Ydstie).

velops methods for stabilization and control of transport


reaction processes. The theory of optimal control of DPS
was developed by Lions (1971). Feedback control of
linear distributed systems was developed by Russell
(1978) and more recent developments are discussed by
Curtain and Zwart (1995).
The method we use for stabilization of DPS distinguishes itself from the methods mentioned above in that
we develop stabilizing controllers using dissipation properties derived from thermodynamics. In this respect our
approach is similar to the passivity theory (Desoer
& Vidyasagar, 1975), the theory of dissipative systems of
Willems (1974) and the more recent behavioral approach
to system modelling (Siep & Willems, 1991). These
methods all share common points with thermodynamics
in that stability and dissipation is de"ned in terms of
actions (Coleman & Owen, 1974) on the system instead
of algebraic properties of the state equations describing
the system.
It is well-established that the dynamics of a chemical
processes are restricted*all observed behaviors give positive entropy production. The Clausius}Planck inequality and classical irreversible thermodynamics (CIT) with
Onsager-Casimir relations for the transport mechanisms
allow us to apply this theory to open systems (Jou,

0005-1098/01/$ - see front matter  2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 0 5 - 1 0 9 8 ( 0 1 ) 0 0 1 4 0 - 6

1740

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

Nomenclature
A
B
c
4
e
F
f
g
H
k
K
A

I
L
M
m
M
G
m
G
n
n
A
n
p
P
P
Q
q
R
S
s

t
;
u
<

the available storage


boundary
constant volume heat capacity
total energy
body force
3;n matrix describing #ux through a point
positive function
Hamiltonian
Boltzman's constant
positive proportional feedback gain
positive de"nite (Onsager}Casimir) matrix
n;n identity matrix
positive de"nite matrix
Massieu function
mass of component i
mass fraction of component i
dimension of z
number of components
unit normal pointing outward from a
boundary element
production rate
the pressure tensor
hydrostatic pressure
positive de"nite matrix
heat #ux
gas constant
entropy
entropy density
absolute temperature
time
internal energy
internal energy density
storage function

Casa-Vazquez, & Lebon, 1996). However, positivity of


entropy production is not su$cient for the development
of a passivity theory for process control. One problem is
the fact that the entropy is not bounded from above. The
negative of the entropy can therefore not be used as
a storage function. In Alonso and Ydstie (1996), Ydstie
and Alonso (1997), we developed a connection between
the second law of thermodynamics and the passivity
theory by using thermodynamic availability, instead of
entropy, as a storage function. This quantity is bounded
and passivity does follow. In Ydstie and Viswanath
(1994), Farschman, Viswanath, and Ydstie (1998), we
followed a di!erent line of attack and developed the
inventory control concept. We showed that some classical control schemes based on energy and material balances could be motivated using Lyapunov design and
macroscopic balances. The main idea here was to control
macroscopic quantities like total mass and energy so that

V
u
v
v
vH
w
X
I
x
I
x
y
z

volume element
manipulated input
vector of inventories
#uid velocity at a point
vector of setpoints for inventories
vector of intensive variables in the entropy
formulation
vector of thermodynamic forces in the kth
direction
kth spatial direction
state of "nite dimensional Hamiltonian
system
measured process output
state of distributed system

Greek symbol




Q



G





GH




'

"

Q
G

non-negative constant
constant referring to size of a system
convex extension
supply rate
constant
chemical potential
constant in stability result
constant in stability result
stoichiometric coe$cient of component j is
reaction i
speci"c volume
density
vector of rates of production
positive integral time
positive derivative time
reaction rate of reaction i

these follow given set-points. Such methods have also


been applied in robotics (Slotine & Li, 1992; Ortega,
Loria, Nicklasson, & Sira-Ramirez, 1998; Sepulchre, Jankovic, & Kokotovic, 1998).
In the current paper we de"ne a candidate storage for
passivity design using the di!erence between a convex
thermodynamic potential, like the negative of the entropy, and its supporting hyperplane through the operating point of interest. This function generalizes the idea of
available work (Keenan, 1951) in that it includes chemical transformations in addition to heat and work. The
supply is now a product of force and #ow variables as in
circuit theory and bond graphs. The theory therefore has
intuitive appeal and can be connected with ideas
that play central roles in process modelling. The #ow
variables are net #ows of extensive properties (energy,
mass, component #ow) and the force variables correspond to intensive variables (temperature, pressure,

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

chemical potential). The results of circuit theory can


therefore be applied to design control systems for the
chemical manufacturing industries. An application of
these results for design of control systems for chemical
process networks is described in Co!ey and Ydstie
(2001).
In irreversible thermodynamics the entropy production consists of two terms. The "rst represents entropy
production due to di!usion and heat conduction. The
second term represents entropy production due to chemical reaction. We evaluate these terms relative to the
entropy production at a given, stationary operating
point. We "nd that di!usion and heat conduction always
dissipate energy. Chemical reaction may cause instability. However, the operating point is stable provided the
entropy production corresponding to heat conduction
and di!usion dominate. In such a case we say that we
trade o! dissipative e!ects due to Onsager}Casimir relations, the so-called Rayleigh}Onsager dissipation and
the non-dissipative e!ects that can arise due to chemical
reaction. We use the PoincareH inequality to de"ne this
trade-o! and a critical system size emerges since the
instability due to chemical reaction scales with the volume of the system. We "nd that a large distributed system
is more dizcult to control than a small distributed system.
And, for every reaction diwusion system with a smooth
boundary, there exists a critical size that allows for stabilization using decentralized feedback control. Stationary
points can therefore be stabilized by decentralized PID
control. These results correspond with observed practice
and explain why decentralized PID control has been
successful in stabilizing distributed chemical production
systems.

2. Mathematical notation
Z,W
N(V; RL)

CK(V; RL)
HK N(V; RL)
y, uB

Banach space and its dual.


space of Lebesque-measurable functions
z : V C RL with "nite N norm:
g "( V g N V)N where 1)p(R
N
space of functions z : V C RL with continuous derivatives up order m.
Sobolev space of functions u3N(V; RL)
with I z 3N(V) for all k"0,1,2, m.
V
the inner product of two vector valued
functions y and u de"ned on B so that
y, u " B y2u dB
@

3. Passivity and the conservation laws


One of the most important and most general stability
results in mathematical systems theory states that the
feedback connection of a passive system and a strictly
passive system is asymptotically stable (Desoer

1741

& Vidyasagar, 1975). In our paper we link the passivity


theory and process control. It then follows that the passivity theory, which originally was motivated by circuit
analysis, can also be used for control system design for
chemical manufacturing systems with states distributed
both in time and space. We refer to the class of systems
we study as Process Systems (Ydstie & Alonso, 1997).
De5nition 1. A process system with input u, output y and
internal states z , de"ned on a domain V with smooth
boundary B is said to be passive if there exist a nonnegative constant  and a functional < : Z C R so that
>
<(0)"0 and for all t(R the following holds
<(z (t))!<(z (0))

y, uB ds!

z  V L ds.
*  _0


If '0 then the system is called strictly passive.
)

The state where <(z )"0 is called the passive state. The
inner product
Q (u, y)"y, uB
is called the supply rate. We note that it is natural to
introduce passivity via the supply function and the dissipation inequality. This approach is normally taken in
thermodynamics and passivity theory (Coleman
& Owen, 1974; Willems, 1974; Ydstie & Alonso, 1997).
According to the de"nition above a passive system has
the property that the amount of `energya we can take out
is limited by what was originally stored up, what has
been supplied minus whatever has been dissipated. In
this way passivity is related to the second law of thermodynamics which says that the entropy of the system is
always greater or equal to the initial entropy plus
whatever has been supplied. More speci"cally the
Clausius}Planck inequality states that
dS
"pR # , S*0, p *0.
1
1
1
dt
The functional S is called the entropy, pR is the entropy
1
production and  is the entropy supply. For a classical
1
system we have
dQ
 " ,
1

where dQ represents the heat supply and the temperature. We now see that thermodynamics is an input}output theory since the unmeasurable quantity S is
de"ned in terms of the supply function.
A strictly passive system is stable and stable invertible
in the following sense: If either or both of y and u are
equal to zero then the internal states z converge to

1742

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

a passive state. The rate of convergence of the uncontrolled states is dominated by the constant . If two strictly
passive systems can be ordered so that
 '


then system 1 is `easier to controla than system 2 since
the uncontrolled dynamics converge faster. Thus we delineate an approach for comparing boundary controls for
strictly passive DPS.
Choose measurement and control structures so that the
mapping uPy is strictly passive and the uncontrolled
dynamics converge as fast as possible.
The objective is to connect these concepts with
thermodynamics so that process physics can be used for
control system design. However, such an objective
should be viewed as an idealization. It may not always be
economically feasible to achieve passivity and co-location of sensors and actuators in a practical application.
In such cases more advanced control and estimation
theory may be needed to solve the stabilization problem.
The dynamics of a process system can be represented
by a set of inventory balances of the form
dv
"p #Q ,
dt

(1)

where v represents a vector of accumulated quantities


like mass, energy and momentum and Q , p the net rate of
supply and production, respectively. As pointed out in
Ydstie and Alonso (1997), Farschman et al. (1998), inventories can be controlled by PID feedback and feedforward. Unfortunately, this approach does not ensure
stabilization of the DPS. In order to address the stability
issue more completely and develop control which relies
more directly on the natural passivity of a process system
we introduce the density variables z, f,  so that

z dV, p "

v"

 dV, Q "

f ) n dB,

where V is an open subset in R with a smooth boundary


B. Elements x"(x , x , x )2 of V;B refer to a position
  
in physical space at a given time. The functions z, will be
referred to as the xeld and are de"ned on the set
D"(<;B;) with T being the semi-open time interval [0,R). The terms  represent production of z and
f represent the #ux of z through the point (x, t) in the
I
x direction.
I
For n inventories v , i"1,2, n the divergence theG
orem shows that the functions z on D satisfy the following system of partial di!erential equations
z # I f " , i"1,2, n,
(2)
R G
V GI
G
where z refers to the ith component of z. Subscript
G
t denotes partial di!erentiation with respect to time and
subscript x denotes partial di!erentiation with respect
I
to the spatial coordinates with summation implied so

that for each element of z and f

z
 f
z " G , I f " GH .
R G
V GI
t
x
H H
The "rst and second laws of thermodynamics restrict
the dynamic behavior of the process system so that total
energy is conserved while mechanical/electrical/chemical
energy dissipate by generating heat. More precisely, there
exists an energy function e(z) which evolves according to
the constraint
e# I f "0
(3)
R
V CI
and an entropy function, s(z), which evolves according to
the constraint
(4)
s# I f " , with  *0.
Q
Q
R
V QI
The last inequality is the local version of the
Clausius}Planck inequality. It states that the production
of entropy is positive for non-equilibrium transformations.
It is important to note that the functions e(z) and
!s(z) are convex if the co-ordinate frame is suitably
chosen. Convexity allows us to de"ne, in a unique manner a new set of variables
w" s
X
these variables are known as the `intensive variablesa
(Callen, 1985). We note that the function
w"w( (z)
is invertible since s(z) is concave.
The production and #ux vectors are generally expressed as functions of the intensive variables and their
spatial gradients, respectively, so that
"( (w),
f "fK (X )
I
I I
where

(5)
for k"1, 2, 3,

(6)

X " I w for k"1, 2, 3


I
V
is the spatial gradient of w in direction k, called the
thermodynamic force. The functions ( and fK are known as
constitutive relations. In classical irreversible thermodynamics (CIT), the constitutive equations for di!usion
and heat conduction satisfy linear relations
f " X ,
I
I I
where the 's are symmetric, positive de"nite matrices
I
called Onsager}Casimir transport matrices. The Fourier/
Fick/Ohm theory of di!usion of heat, mass and electrons
are good examples of constitutive equations that are
nearly linear for wide ranges of operating conditions
(Lavenda, 1993).
By combining Eq. (4) with Eqs. (2), (5) and (6), using
Lagrange multipliers and linear relations we get (Ydstie

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

By adopting the assumption of local equilibrium


(Callen, 1985) we de"ne the entropy by the Gibbs relation

& Alonso, 1997)



 " X2 X #2w*0.
Q
I I I
I

(7)

At equilibrium we have X ,0 and it follows that the


I
production terms, 2w are positive and dissipative close
to equilibrium. This motivated the linearization theory of
CIT (see for instance, Degroot, Mazur, & DeGroot,
1964).
Controls are executed at speci"ed boundary elements
by adjusting the #ux of material and energy. This is
normally achieved by changing the resistance to #ow or
by changing the boundary conditions. We note that we
are interested in control problems where the control
surfaces have non-zero measure in a space with dimension D!1, where D is the dimension of the physical
domain. Thus the control acts on boundary surfaces of
"nite size in 3D problems and lines with "nite length in
2D control problems.
Example 1 (Navier Stokes equations with chemical reaction and energy balance). The microscopic balance equations for a process system are derived as in (2). The
continuity equation is stated in a "xed frame of reference
and the energy, momentum and component balances in
the center of mass frame. This allows us to make
a smooth transition to classical irreversible thermodynamics which relies on the assumption of local equilibrium. In standard notation for a simple system we then
have the following system:

# ) (v)"0,
t


u
# ) q"P : v,
t

m
LP
G # ) f " 
Q , i"1,2, n ,
G
GH G
A
t
H

v
# ) P"F,
t

1743

where  represents the stoichiometric coe$cient and


GH

Q the rate of reaction i. The rest of terms is described in


G
the nomenclature section. The total energy of this system
is given by the expression
e"(u#v ) v)

which is a combination of the internal and mechanical
energies. By substituting the balance equation for the
internal energy and the velocity "eld into Eq. (3) we "nd
that the total energy is conserved as suggested by Eq. (3).

ds"w2 dz.
In terms of the variables above this is written
1
P
LA 
ds" du# dv# G dm .
G

G
By substituting the balance equations into this expression we recover the entropy balance as given by Eq. (4)
and the expression for the entropy production (7) can
then be related to the constitutive equations. In this
reference frame the energy and the negative of the entropy are convex functions and can therefore be used for
passivity design in the light of the theory developed in the
next sections.
Example 2 (Non-equilibrium molecular dynamics). The
symplectic form on RL is given

!I

L ,
I
0
L
where I is the n;n identity matrix. The motions of the
L
molecules in a classical system are described by the
equation
S"

x "SH(x),
where x is the state and the convex Hamiltonian is of the
type
H(p, q)"( p # q ), x"(p, q).

We have
HQ "Hx "0
by the anti-symmetry of S*the energy is conserved.
According to classical statistical mechanics we de"ne the
entropy for the equilibrium cells so that
S"!k P ln P .
G
G
It follows that the negative of the entropy is a convex in
the probabilities P since Boltzman's constant is positive
G
and the logarithm is concave. These equations and generalizations to Hamiltonian systems with dissipation are
described by Evans and Morris (1990).

4. The available storage


According to Friedrichs and Lax (1971) a convex extension is a convex function of state which satis"es a conservation law. The convex extension can therefore be
added to the "eld as a new element which acts as a constraint on the allowable behavior. Physical examples of
convex extensions include the following thermodynamic

1744

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

potentials:

extension will be employed:

!s*the negative of the entropy density,

"! Hs,

u*the internal energy density,

where H is a function of position but independent of time


and s is the entropy. We recover the classical equilibrium
theory with H"!1. The function m is then called
a Massieu function (Callen, 1985) and w is the vector of
intensive variables in the entropy formulation. With the
local thermodynamic state

u!s*the Helmholtz free energy density.


The last function is the Legendre transform of u.
De5nition 2. A continuous (C) function (z) : ZCR is
called a convex extension if the symmetric n;n matrix
M with elements
(z)
M "
GH z z
G H
is positive de"nite for all (x, t)3S.

(9)

w"! H

Let w be a vector in RL. Then we can have


m(w)"sup (w2z!(z), z3Z)
X
m is the Legendre}Fenchel transform of . This de"nition
does not require continuity and a general theory can be
developed for discontinuous extensions by using tools
from non-smooth analysis. However, we will follow
thermodynamics and assume that  is continuously differentiable up to, and including, order two. A simple
geometric argument reveals that w is the directional
derivative of  so that
w2" .
X
The map w( (z): Z C W, where the spaces Z and W are
dual, is invertible since  is convex.
We have found that convex extensions can be used to
de"ne a new class of thermodynamic potentials which we
will call `the available storage relative to stationary
statea, or simply `the availabilitya. The concept is related
to the Keenan's (1951) availability which is the minimum
amount of work needed to perform a certain task using
"xed resources and Carnot engines. A similar terminology was used by Willems (1974) in his paper on dissipative dynamical systems. Willems showed that the class
of all possible storage functions, for a given supply function, forms a convex set and that they need not be unique.
Our de"nition of the availability is unique and corresponds to Willems' minimum available storage. More
precisely, we de"ne the available storage at the state
z relative to a reference state zH as
a(z, zH)"(z)![(zH)#w(zH)2(z!zH)]*0.

z"(u, , m ,2, m A )2

L
we get

(8)

The availability has a simple geometric interpretation.


The last two terms represent the supporting hyperplane
tangent to (z) at z"zH. The extension (z) is convex and
it follows immediately that the availability is positive for
any zOzH. In the next sections, the following convex


 2
1 P 
, ,  ,2, A ,2, LA .

(10)

Thus w is the negative of the vector of intensive variables


multiplied by the temperature at the reference state
(Callen, 1985). In our work we will set H"H equal to
the temperature of the reference state. This choice is not
only convenient, it leads to an intuitive de"nition of the
storage function for passivity design which has units of
energy and which can be identi"ed with the available
work.
In the case of one conservation law for the internal
energy we have s"s( (u), where u is the internal energy
and we get
a(z, zH)"u!Hs#aH

(11)

where aH"HsH#uH (Ydstie & Alonso, 1997). This


storage function provides an upper bound for the storage
function de"ned via Eq. (8) with state dimension n*2.
As an example we take the ideal gas with local state
z"(u, )2.
We then have

H PH 2
,
w"!H s"!
.
X

The entropy of an ideal gas satis"es the relationship
u

3
s!sH" R ln #ln .
uH
H
2
The point (uH,H) is the arbitrary reference state. Convexity of the storage function, as de"ned above, follows from
the concavity of the logarithm. In Fig. 1 this is illustrated
in one dimension. We show the supporting hyperplane
through the point (uH,H) projected onto the (s, u) plane.
Positivity of the availability is evident.
The following results establish connections between
primary variables and dual variables in the deviation
form. First consider the new function
g(z, zH)"(w!wH)2(z!zH).

(12)

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

1745

This inequality is valid for any z so the "rst part of the


result follows.
To prove the second part we use Newton's theorem for
vectorial "elds (Dennis & Schnabel, 1983) so that




M(zH#(z!zH)) do (z!zH)

3[0,1] is a scalar parameter and the integration is
carried out element-wise. Let us now de"ne
w(z)!w(zH)"

Q"

M(zH#(z!zH)) d.
(15)

It follows from Eq. (15) and positivity of M that Q is
positive de"nite. A one to one map between w!wH and
z!zH is established.

Fig. 1. The entropy of an ideal gas and its supporting hyperplane at the
point (uH,H). The available storage is the di!erence between the two.

Lemma 1. Suppose that (z) is strictly convex, then we have

Lemma 2. There exist positive constants q and q so that




0)q z!zH )a(z, zH))q z!zH ,


where equalities hold if and only if z"zH.
Proof. De"ne two positive numbers

1. 0)a(z, zH))g(z, zH)


2. w!wH"Q(z!zH) where Q is a positive dexnite n;n
matrix.
Proof. We "rst show that a(z, zH) is strictly convex itself
with respect to z. First we compute the Hessian for a.
From the de"nition of a we have
a 
" !w(zH)"w(z)!w(zH)
z z

(13)

so that the Hessian satis"es


a
(z)
"
"M'0.
(14)
z z
z z
G H
G H
Convexity of a(z, zH) follows. We now compute a supporting hyperplane at an arbitrary point z and we get from
convexity

 

a(z, zH)*a(z, zH)#

a
z

(z!z).
XXY

Hence
a(z, zH)*a(z, zH)#[w(z)2!w(zH)2](z!z).
This inequality is valid for every z. In particular, it is valid
for z"zH. With this substitution we get, using the de"nition of g
a(zH, zH)*a(z, zH)#[w(z)2!w(zH)2](zH!z)
"a(z, zH)!g(z, zH).
Re-ordering the inequality we have
g(z, zH)*a(z, zH).

x2M(zH#(z!zH))x
inf
x2x
WCW V
x2Qx
q " sup
.

x2x
V
By Taylor expansion we have
1
q "
 2

a(z, zH)*q z!zH 



since a is convex and q was evaluated at the point of

minimum curvature. The upper bound follows since we
established in Lemma 1 that
a)g
and
g)q z!zH .


Note that constants q and q can be computed as the




minimum and maximum eigenvalue, respectively, of
M on the compact -set. From the results above we see
that a and g de"ne equivalent norms for deviations
z!zH.
We now want to develop the conservation law for the
availability. The development here is limited to extensions that are continuously di!erentiable whereas the
development in Ydstie and Alonso (1997) applies to discontinuous, convex extensions. First we use Eq. (8) to
write
a" a z
R
X R
"(w!wH)2 z.
R

1746

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

With zH independent of t we have

The functional < used in the dissipation inequality takes


the form

a"(w!wH)2 (z!zH)
R
R
"w 2 z ,
R
where we introduced the deviation variables
w (z, zH)"w(z)!w(zH)

<"B!

From Eq. (2), written in deviation form with respect to


the stationary reference we get
z # I fM ",
R
V I
where we have deviation variables
(16)

a"w 2(! I fM #)


R
V I
"! I (w 2 fM )#fM 2 I w #w 2.
I
I V
V
It is now convenient to introduce the deviation force
XM "X !XH.
I
I
I
The conservation law for the available storage can then
be written in local form as
a# I (w 2fM )"fM 2XM #w 2 with a*0.
(17)
R
V
I
I
I
For a distributed process system with volume V we
de"ne the total available storage

A(t)"

a(z(t)) dV*0.

(18)

The function A measures the availability relative to


a given reference. By integrating Eq. (17) over B;V and
applying the divergence theorem for the second term on
the left-hand side we get from Eqs. (17) and (18)

AQ #

w 2fM ) n dB"

(XM 2 fM #w 2) dV.


I I
V

(19)

This equation has the same structure as the equation


used to de"ne a passive system if one sets <"A and
makes the following assignments:
fM ) nCu for x3B,
!w Cy for x3B.
More generally, the distributed process system is passive
if there exists a constant *0 so that

(X2 fM #w 2) dV)!


I I

z 2z dV#h(t)

with h(t) being a non-negative function such that

h(s) ds)C(R.

h(s) ds#C*0.


We end up this section with one example that illustrates most of the ideas presented so far.

and z (z, zH)"z!zH.

fM "f !f H "!H.
I
I
I
By combining these expressions we get

(20)

Example 3 (Conduction of heat in a one-dimensional, well


insulated bar). The internal energy balance in Example
1 reduces to
u# f"0, V"[0, ],
R
V
where 0 and correspond to the ends of the bar. We will
assume that the volume expansivity is negligible.
The boundary conditions satisfy
"

at x">,

at x"0\.

The heat #ow at the ends of the bar is given by
"

f n"k ((0, t)! (t)), f n"k ((, t)! (t)). (21)





*
*
*
The de"nitions of the normal gives n"1 at x"0 and
n"!1 at x". The resistances k and k can be

*
varied and considered as gains. From the theory developed above we de"ne a candidate convex extension
"!s(u),
where  is an arbitrary reference. By following Callen
(1985) we get
1
w"! s"! .
S

Hence

 

w 
M" "
.
u  u
The constant volume heat capacity is given by
c " u/ . The second law of thermodynamics states
4
that the heat always #ows from a hot to a cold reservoir,
which implies c '0. Hence
4
 1
M"
'0.
 c
4
Convexity of the extension (u) therefore follows. We can
write
dw"M du
so that by integration we have the relationship between
the extensive and the intensive variables in the deviation
form

1
1
 ! #
"Q(u!uH),
H

(22)

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

variables so that

where

 1
d'0
 c
4

as given in the development above.
Let us now choose a stationary reference temperature
H(x). Using Eq. (17) with "0 and de"ning A as in (18),
we have the di!erential equation for the availability

Q"

AQ #




1
1
!
H(0) (0, t)

!

1
1
!
H() (, t)

fM

 
fM "
*

*
XM fM dV.


We now make the assignments


1
1
!
!
(0, t) H(0)
C y,
1
1
! 
!
(, t) H()



 




fM

 C u.
!fM
*
By integration we get

*
XM fM dx.
(24)


The mapping u C y is passive if we can establish that the
right-hand side of inequality (24) is non-positive. In order
to do this we note that the entropy production for this
system, according to Eq. (7) satis"es

A(t)!A(0)#

1
1
f "f H!K
!
, i"0,1.
G
AG
G

H
G
GK
Here are the temperatures at the ends of the bar,
GK
either (0, t) or (, t).
As an aside we note that we can re-write the control
expression so that
K
AG (H! ), i"0,1.
f "f H!
G
GK
G
H G
GK G

(23)

1747

u2y ds"

dw
 "!fX with X"
Q
dx
since "0. According to classical non-equilibrium
thermodynamics we have
f"!X with '0,
where positivity of the transport coe$cient follows
from the Clausius}Planck inequality (4). We therefore
have
!(X!XH)( f!f H)* (X!XH),

where  "'0. Hence

!XM fM * XM ,

and the system is passive.
We can now follow Desoer and Vidyasagar (1975) and
apply decentralized controllers of the form
u"!K y,
A
where K is a diagonal 2;2 matrix of proportional gains.
A
This control can be expressed in terms of the physical

Thus we have the standard temperature error control


scheme. We also need to "nd a scheme to manipulate the
heat-#ux f . In practice this can be achieved in a number
G
of ways. Depending on application we change heat transfer area, "lm resistance and or the temperature of the
surrounding medium.
We show that the system is strictly passive by applying
the PoincareH inequality (Struwe, 1996)

w  dB# XM  dV*
w  dV,

 V
 B
V
where  and  are positive constants. It follows that we


can write

XM fM dV* 
 

w  dV! 
 

w  dB.

The last term is related to the boundary conditions and it


decays when we apply feedback control. The "rst term on
the right-hand side then dominates and shows that the
system is strictly passive. This has two important consequences:
First, we can apply and integral action to reduce o!set.
The controller then takes the form

1 R
dy
f "K y#
y ds#
,
G
AG
" dt

' 
where y is the error as de"ned above and the derivative
action is included as well.
Second, and possibly more interesting, we can use
dissipative e!ects, in this case heat conduction, and trade
these o! against non-dissipative e!ects, for example
chemical reaction. This issue is explored in the following
sections where we derive the general theory. All the
ingredients leading up to the main stability results are
contained in the example above, however.
We "nish up this section by relating the thermodynamic theory of heat conduction to Fourier's law
f"!k for x3V.
V
From the developments above we have



1
1
X"!
"
V
 V

1748

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

so that with w"!1/ we have


"k

(25)

which is nonlinear. However, is often large relative to


the total variation in so that the term  does not
change much in the region of interest. The di!erence
between classical irreversible thermodynamics and the
Fourier/Fick/Ohm theory is therefore not signi"cant.

5. Feedback stabilization
In this section we provide su$cient conditions for
passivity of process systems and convergence to stationary solutions. These results allow us to develop boundary
controllers assuring convergence of the outputs to their
respective setpoints and convergence of internal states to
stationary values satisfying
I f ".
(26)
V I
We denote the stationary solutions to (26) by zH. The
boundary conditions can for example be given by
dz
a z#a
"b ,

 dn


(27)

where a , a , b are appropriate functions. The objective


  
of the control system is to adjust the control parameters
f and the boundary conditions so that the solution to the
I
system (2) in the limit converges to zH. Physically this
amounts to adjusting the #ow of material and heat at
boundary elements by controlling transport coe$cients
or states at the boundary as illustrated in Example 3.
De"ne the scalar product of two matrices

f 2X " f X "Tr( f 2X),
I I
G G

where
f"[ f , f , f ]; X"[X , X , X ].
  
  
Assumption A3 (Dissipation conditions).
1. The solutions z(x, t) and zH(x) are members of
H (V; RL) for all t'0.
2. There exist real numbers  and  so that for all


t*0
 fM , XM V #w , V )! XM  V L
I I
 I *  _ 0
# w  V L #h(t)

*  _ 0
with  '0 and h(t) being a non-negative function of

time satisfying:

h(t) dt)C(R.

(28)

Lemma 3. Suppose that the condition A3(1) is satisxed.


Then
 w  B L # XM  V L * w  V L ,

*  _0 
I *  _0 

*  _ 0
where  "c meas(V) and c is a positive constant. If in



addition A3(2) holds, then there exists a function < bounded
from below such that
 
<Q (t)#w , fM ) nB !   w  B L
*  _0 



)!  !  XM  V L .
 
I *  _ 0


Proof. We "rst need to develop the inequality that relates XM  V L to w  V L . To that purpose, we
*  _ 0
*  _ 0
make use of Green's formula
XM , XM V "w , XM B !w , w V ,
(29)
I I
I
where  represents the Laplace operator. Since
w3H (V; RL) we expand w as an in"nite series

w " c  (x , x , x ).
(30)
H H   
H
The 's represent a complete orthonormal system
 H satisfying the Euler}Lagrange equations
H H
! " 
(31)
H
H H
with appropriate boundary conditions. The numbers
 are known as the eigenvalues and the functions  are
H
the eigenfunctions of the operator !. Using Eqs. (30)
and (31) we obtain




w , w V "! c  , c  
"!  c,
H H
H H H V
H H
H
H
H
where the property that the 's are orthonormal has been
used. Since 0( ( for all i(j and
G
H

w , w V " c
H
H
we can conclude that
!w , w V * w  V L .

*  _ 0
Standard results in functional analysis show that  de
pends on the size of the domain (Courant & Hilbert,
1937). Combining the above inequality with (29) we "nd
that there exists a constant  so that

XM  V L *w , XM B #c (meas(V)) w  V L ,
I *  _ 0

*  _ 0
where
 "c meas(V).


We can now re-arrange the inequality above so that it
appears in the required form



w  dB# XM  V L * w  V L .

*  _ 0
I *  _ 0

(32)

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

These results are re-statements of the PoincareH inequalities and details are given for example in Smoller (1983)
and the appendix of Struwe (1996).
To prove the second part of the result we combine Eq.
(19) with the dissipation condition A3(2) so that we get
the equation for the total available storage

AQ #

w 2fM ) n dB)! XM  V L
 I *  _ 0
# w  V L #h(t).

*  _ 0

(33)

The storage function

<"A!

Proposition 1. System (2) with inputs and outputs dexned so


that
u"fM ) n!w ,
y"!w
is strictly passive, in the sense of Dexnition 1 when Assumption 3 holds with   ! and *  . The rate of
 

 
dissipation is given by
"(  ! ) ,
 
 
where the constant  is dexned in the proof below.

Proof. From Lemma 3 we have

h(t)#C

is bounded from below by zero since A is non-negative


according to Lemma 1 and h(t) satis"es (28). In addition,
the last term in Eq. (33) can be bounded from above by
using the PoincareH inequality (32). After substitution we
get

<Q #

1749


w 2fM ) n dB)! XM  V L #  XM  V L

_
0


I
*
I *  _ 0

B

 
# 



For homogeneous boundary conditions we have


w "0 at the boundary and we get the simpler PoincareH
inequality
(34)

<(t)!<(0)!

For a one-dimensional problem with homogeneous


Dirichlet conditions we have



 
 "
,


where is the length of the domain. In this way we see


that the maximum eigenvalue depends on the size of
system, and that it increases as the size gets smaller. The
same expression holds for the sphere if we replace with
the radius R. Similar expressions can be developed for
other domains like cylinders rectangles etcetera. In all
cases we "nd that the parameter  depends on the

size of the domain provided that B is su$ciently regular.
These results can now be compared with dimensionless
groups like the Reynolds, Rayleigh, and Dahmkohler
numbers.

with
 " .






<Q (t)!u, yB )!(  ! ) AM  V L


 

*  _ 0
!(!  )y, yB .
 
Integrating over time and noting that *  we have
 
R
<(t)!<(0)! u, yB ds

R
)!(  ! ) w  V L .
*  _ 0
 


Now w "Qz with Q invertible so that

w  dB

and the result follows.

XM  V L * w  V L

*  _ 0
I *  _ 0





<Q (t)!u, yB )!  !  X M  V L
 
I *  _ 0

 
! !   y, yB .


By using the PoincareH inequality we get

u, yB ds

)!(  ! )
 
 

z  V L ds
*  _ 0

where
x2Q2Qx
 " inf

x2x
V VZW
is the smallest Rayleigh coe$cient of Q2Q. The result is
then established by setting
"(  ! ) .
 
 

Proposition 2. Suppose that conditions in Proposition


4 hold, so that the mapping K : yCu is passive and dexne
a controller u"!(y) such that
y, uB )!y, yB )0.
In addition, suppose that h(t)) (t) z  V L with

*  _ 0

 (s) ds)c (R.




(35)

1750

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

Then we have
1. A(t))c e\RO with '0 and

2. lim
z "0
R

Proof. From Proposition 4 we get
<Q (t)!u, yB )! z  V L
*  _ 0
since <"A! R h(t)#C and C is a constant, we have

after substitution that the total availability of the system
satis"es the di!erential inequality
AQ )! z  V L #h(t)
*  _ 0
)! z  V L # (t) z .  V L .
*  _ 0

*  _ 0
Using Lemma 1 we have that q z  V L )A)
 *  _ 0
q z  V L so that
 *  _ 0

 (t)
AQ ) ! # 
A
q
q


let "q \ and  " (t)q . Then we have


 
1
eROAQ ) ! # (t) t/ A


so that
d(eROA)
)eRO (t)A.

dt
Hence

eROA(t))A(0)#

eQO (s)A(s) ds.




And, using Gronwall lemma we get

R
t
A(t))A(0)exp ! #  (s) ds



)c eRO

with c "A(0)exp(c ). The second result follows since we


have
z  V L )cA(t)
*  _ 0
for some constant c.
The results above show that stability of a system of
conservation laws with a convex extension can be assured with simple, decentralized control at the boundary
when the dissipation condition is satis"ed and the reference system and its boundary conditions are chosen so
that a stationary solution exist. The constants in the
stability result are optimized by solving an equivalent
Euler}Lagrange problem. Uniform () convergence is
achieved for equations of the parabolic type since di!usion induced su$cient smoothing to apply standard
maximum principles. The techniques we use are similar

to those used to develop viscosity methods for the


numerical solution of hyperbolic partial di!erential
equations. In the latter method, the viscosity/dissipation
is turned o! completely and the solution to a set of
hyperbolic equations is obtained by a limiting process.
The solutions obtained in this fashion are called weak
solutions since they may display discontinuities.

6. The thermodynamic foundation


In the previous section we developed abstract stability
results for systems described by conservation laws that
possess a convex extension. In keeping with the program
outlined in the introduction to the paper we now must
make the connection with thermodynamics. The following two issues need to be dealt with:
1. A physical basis for a convex extension needs to be
established.
2. The dissipation condition in Assumption A3(2) needs
to be related to the second law of thermodynamics.
Classical irreversible thermodynamics extends the
principles of equilibrium thermodynamics to open systems operating out of equilibrium. It is constructed on
the assumption of local equilibrium. In this way, the
system domain V is assumed to be composed by a large
number of cells of size much smaller than V where the
results of classical equilibrium thermodynamics apply.
The inventory balance and the classical theory then applies to each of these cells and this allows us to de"ne
intensive variables like temperature, pressure and chemical potential that we will use for feedback control.
There exist several formulations of non-equilibrium
thermodynamics suited to deal with chemical process
modeling problems. The most important are
1.
2.
3.
4.

Classical irreversible thermodynamics (CIT)


Rational thermodynamics (RT)
Extended irreversible thermodynamics (EIT)
Non-equilibrium molecular dynamics (NEMD)

A succinct review of the methods, their respective


strengths and limitations can be found in the book by
Jou et al. (1996). The perspective developed for nonequilibrium thermodynamics in our paper is most closely
related to the CIT and EIT approaches. Further work is
needed to make the connection with NEMD and Hamiltonian systems.
6.1. Physical basis for the convexity
Following Callen (1985) we have for an equilibrium
cell.

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

E Postulate I: The state of each equilibrium cell of volume V is characterized by a canonical variable set
A
v"(;,V , M ,2, M ,2, M A )2.
A 
G
L
E Postulate II: There exists a function S"SK (v), called the
entropy which is maximized over a manifold of constrained equilibrium cells.
E Postulate III: The entropy is additive so that for any
scalar 
S"SK (v).
E Postulate
( ;/ S)VA

IV:
The
"0.
+G

entropy

vanishes

for

We now need to deal with the issue of whether the


convexity theory can be connected with the thermodynamic potentials. A function of the type described in Postulate III is said to be homogeneous of degree one. From
Euler's theorem
S
S" v.
v
By di!erentiating with respect to v on both sides of this
identity we have
S
S
S
M "!
"!v2MM # , M
,
GH
v
v
v v
G H
whereby we get by multiplying through with v on the
right
M v"0 for all v.
v2M
It follows that !S cannot be strictly convex in the
inventories v. We note that Postulate II implies convexity, however. For example, if we limit ourselves to second-order conditions, we "nd that functions of the type
described here are minimized provided the Hessian of S is
negative. In order to get beyond this impasse we need to
introduce constraints. Following Jou, Casa-Vazquez,
and Lebon (1996) we divide through with the total mass
we de"ne a vector of densities so that
M" LA
G+G
z"v/V "(u,, c ,2, c A ),
A

L
where  is the density. The entropy density s"s( (z) is no
longer a homogeneous function in the variable set z.
However, s is concave for classical systems that calculate
pressure temperature and chemical potentials using the
Gibbs relation.
H ds"!w2 dz,
where the intensive variables w are de"ned as in (10) and
the corresponding matrix M with elements
s
M "!H
GH
z z
G H

1751

is positive de"nite under the standard assumptions of


thermal mechanical and chemical equilibrium. The explicit expression for M can be found in Callen (1985) and
Jou et al. (1996).
The theory developed in the previous section therefore
applies with the strictly convex extension
"!Hs
since s is concave in the classical equilibrium theory.
H"K H(x) is the temperature of the stationary state.
6.2. The dissipation conditions
We now derive the di!erential equation for the extension and relate the dissipation condition to convexity of
the entropy production. The entropy balance is recovered by substituting Hs"! for a in Eq. (17) so
that
H s! I (w2f )"!f 2X !w2.
I
I I
R
V
By comparing with Eq (4) we see that

(36)

 "(!f 2X !w2)/H
(37)
Q
I
I
corresponds to the rate of entropy production and that
f "(!w2f )/H
(38)
QI
I
de"nes the entropy yux.
In CIT, the balance equation for the entropy is related
to intensive variables in the entropy formulation by substituting the energy and component balances, developed
in Example 1, into Eq. (36) and using the Gibbs relation
to de"ne temperature, pressure, and chemical potential.
The following relations are obtained for the entropy #ux
and entropy production in the center of mass framework
LA 
1
f " q ! Gj G .
QI I
AI
G
The entropy production satis"es

 

(39)

 

1

LP
LA
LA

" I
q # I G j G !
Q  A . (40)
Q
V I
V AI
H
HA
G
H A
These equations can now be matched to our framework
through Eqs. (37) and (38).
We will now explore the constraints imposed by the
Clausius}Planck inequality, which in the local form
states that
 *0.
Q
Together with (37) this implies
fK (X )2X )0 ( (w)2w)0
I I
I
since the terms X and w may vary independently. These
I
inequalities impose constraints on the constitutive

1752

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

equations (5) and (6). Close to equilibrium these conditions motivate linear laws and symmetry conditions for
the Onsager transport relations.
We now extend the scope of CIT and include nonlinear
phenomena that may occur far from equilibrium.
De5nition 4 (Purely dissipative systems). The #ux and
production variables satisfy Onsager}Casimir relations.
De5nition 5 (General systems). The #ux variables satisfy
Onsager}Casimir relations and the production variables
are Lipschitz continuous in the dual variables w.
In the "rst case we have the following inequalities
satis"ed:
fM 2XM )! XM  and 2w )! X M ,
I
I
 I
 I
where  '0 and  *0 are related to the eigenvalues of


the Onsager}Casimir transport matrices. The dissipation
condition A3(2) is then satis"ed with h(t)"0 for any
 '0. Stability of a purely dissipative system does not

depend on size.
A system which satis"es dissipation condition A5 but
not A4 does not necessarily have a stable stationary state
for any size domain. However, we can prove stability of
stationary states for domains that satisfy a critical size
constraint. First, let us write
fM 2XM )! XM  and 2w ) w ,
I
I
 I

where  is the Lipschitz constant. The following result,

valid for non-equilibrium systems then follows.
Suppose that  is a strictly convex function in X and
Q
I
Lipschitz continuous in w. The dissipation condition
A3(2) is then satis"ed and the state converges to a stationary, non-equilibrium state characterized by minimum available storage and zero dissipation relative to
the reference state provided the critical size condition is
satis"ed and the boundary #uxes are controlled using
passive feedback.
The convexity conditions above can be related to the
usual Onsager}Casimir-type relations between #uxes
and thermodynamic forces of the form

Fig. 2. The entropy production and dissipation for carbon steel at


203C. The coe$cient of heat conduction for carbon steel at this temperature. The reference point for the available storage is XH"0.6. At this
point the entropy production is 0.303 J/K cm.

where  is the smallest eigenvalue of . In the previous



section we related Rayleigh}Onsager dissipation to strict
dissipativity by using the PoincareH inequality. In the case
of constant Onsager coe$cients the rate of entropy production is a quadratic function as shown in Fig. 2. This
leads us to the following conclusion:
If ' then System 1 is easier to control than


System 2.
The result says that the uncontrolled dynamics decay
more quickly the more viscous the system is. This is
because viscosity causes mechanical energy to dissipate
to heat.
A more general situation can be taken into account in
our theory by assuming a bound of the form
 fM 2XM V )! XM  V L #h(t)
I
I
 I *  _ 0
with h(t) satisfying

0)
f "! X .
I
I I

(41)

This gives  "X2 X '0, which is convex since is


Q
I I I
I
symmetric and positive de"nite.
In the absence of chemical reaction and other production terms we have  "0 in Assumption 3(2) and the

local rate of `thermodynamic dissipationa in Lemma 3 is
given by the Rayleigh}Onsager dissipation on deviation
form
dQ "XM 2 XM * XM ,
I I I
 I

(42)

h(t) dt)C(R.

This condition allows us to go beyond the normal Onsager}Casimir relations for di!usion and heat transfer.
We now develop  fM I , XM V to get
V
I
 fM I , XM V "! I z 2Q2, k I z V #z 2Q2 , k I z V ,
V
I
V
V
V V
where Q stands for the spatial derivatives of matrix
I
Q taken element-wise. On the other hand, let us compute
the time derivative of the function < "q z  V L

 *  _ 0

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

along the "eld (2) to obtain


<Q (t)"!q z , fM I B !q  I z , k I z V .


V
 V
V
Fixing the #uxes at the boundary and using Lemma 2 we
have that the availability is bounded by the storage
function so that A(t))< (t), we have

p(s)"q  I z , k I z V ! I z 2Q2, k I z V *0
V
V
V
 V
and

A(0)#

z 2Q2I , k I z V ds
V
V

)< (0)!


p(s) ds)< (0),




z 2Q2I , k I z V ds
V
V



)< (0)!A(0)

which is true for every time t so we can write

h(s) ds)< (0)!A(0))C(R.




The same line of arguments with < "q z  V L and

 *  _ 0
< )A can be used to show that h(t) is non-negative. If

in addition h(t)) (t) z  V L then exponential stab
*  _ 0
ility can be concluded from Proposition 5. We believe
these methods are new and that they give a complete
derivation of the classical stability results without the
need of relying on linear regimes.

7. Reaction-di4usion equations

I we have
V
z # I fM "(z).
(45)
R
V I
From convexity of the Onsager relations we have a positive constant  so that

Mf 2XM "! XM .
I
I
 I
Assuming that (w) is globally Lipschitz we "nd that
there exists a constant  so that

w 2) w 

and the dissipation conditions are then satis"ed. A
distributed system described by reaction di!usion equations is passive and it can therefore be stabilized with
PID control provided that the size is such that
 "meas (V)' / .

 
For a one-dimensional problem we illustrate the critical size constraint for the di!usion equation with autocatalytic reaction. This situation may represent di!usion
of neutrons in a nuclear chain reaction. We have from
Example 3
u# f".
R
V
The production term on the right accounts for neutron
generation. It is given by the expression
",
where is the temperature. Following the previous developments we use the availability as a storage function.
The theory developed above then applies with the following assignments when we use Fourier's law:

Systems of nonlinear reaction-di!usion equations are


used to model a wide range of processes in chemical
engineering, including plug #ow reactors with wellestablished velocity pro"les, distributed predator}prey
problems, reaction at catalyst surfaces and oscillatory
phenomena (Smoller, 1983). The system takes the form

 "(/),

 "(w)"k,

 "H.

Thus we get strict passivity if

z# I f "(z),
(43)
R
V I
where z represents the vector of density of internal energy
and concentrations of chemical species so that



z"(u, c ,2, c A )2

L
(z) is a smooth function and f the vector of heat and
I
mass di!usive #ux. It takes the form
f "! I w,
I
I V
where w is the vector of dual variables

w"!H

(44)

1 
 2
,  ,2, LA .

The speci"c volume is not included since we are not


considering the convective #ow. Using the results in the
previous sections we "rst rewrite the system on deviation
form relative to a steady state of interest. By linearity of

1753

 
k!H'0.

From this expression it is clear that the stability is local


and that there exists a critical length so that stability is
achieved for every

(

k
.
H

It follows that the decentralized PID control laws stabilizes the system as long as the critical size constraint is
satis"ed.

8. Conclusions
In this paper we have established that irreversible
thermodynamics and the passivity theory of nonlinear

1754

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755

control can be connected. We started with the de"nition


of a passive system and then we showed that a classical
non-equilibrium system is passive when we use the local
equilibrium hypothesis and Onsager}Casimir-type relations for di!usive/conductive-type phenomena. We presented a new thermodynamic potential which can be
related to the thermodynamic availability. In practical
terms available work is the ideal work that can be extracted from a process using Carnot engines. However, our
potential is more general since, in addition to heat and
mechanical energy, it includes work available from chemical reaction. We showed that chemical processes described by di!usion and heat conduction are dissipative.
Such processes are therefore open loop stable. Exothermic chemical reactions may introduce instabilities. We
have shown that there corresponds a critical system size
which can be expressed in terms of a dimensionless
(Dahmkohler) number which embodies the stability
characteristics. If the system is smaller than the
critical size then it can be stabilized with passive
boundary control. Examples involving chemical reaction
and heat conduction have been developed to illustrate
the application of the theory. Finally, we note that
the framework can be used to de"ne Lyapunov functions when the systems are modeled using "nite, instead of in"nite dimensional systems of di!erential
equations.

Acknowledgements
The authors acknowledge generous support from the
Petroleum Research Fund (PRF C 31498-AC9), the
National Science Foundation (NSF C CTS-9726115),
the European Community (project CT96-1192) and
a PostDoc fellowship (AAA) from the government of
Spain.

References
Alonso, A. A., & Ydstie, B. E. (1996). Process systems, passivity and the
second law of thermodynamics. Comp. Chem. Eng., 20, S1119.
Callen, H. B. (1985). Thermodynamics and introduction to thermostatistics.. New York: Wiley.
Christo"des, P. D. (2000). Nonlinear and robust control of PDE systems..
Boston: Birkhauser.
Co!ey, D., & Ydstie, B. E. (2001). Process networks: Passivity, stability
and feedback control, AIChE Journal, submitted.
Coleman, B. D., & Owen, D. R. (1974). A mathematical foundation for
thermodynamics. Archives Rational Mech. Anal., 20, 1}104.
Courant, R., & Hilbert, D. (1937). Methods of mathematical physics..
New York: Wiley.
Curtain, R. F., & Zwart, H. J. (1995). An introduction to linear inxnite
dimensional systems.. New York: Springer.
Dennis, J. E., & Schnabel, R. B. (1983). Numerical methods for unconstrained optimization and non-linear equations.. Englewood Cli!s,
NJ: Prentice-Hall.

Degroot, S. R., Mazur, P., & DeGroot, S. R. (1964). Non-equilibrium


thermodynamics.. New York: Dover Pubns.
Desoer, C. A., & Vidyasagar, M. (1975). Feedback systems: Input output
properties.. New York: Academic Press.
Evans, D. J., & Morris, G. P. (1990). Statistical mechanics of nonequilibrium liquids.. New York: Academic Press.
Farschman, C. A., Viswanath, K., & Ydstie, B. E. (1998). Process
systems and inventory control. AIChE Journal, 44, 1841.
Friedrichs, K. O., & Lax, P. D. (1971). Systems of conservations equations with a convex extension. Proceedings of the National Academy
of Science, USA, 68(8), 1686.
Jou, D., Casa-Vazquez, J., & Lebon, G. (1996). Extended irreversible
thermodynamics.. New York: Springer.
Keenan, J. H. (1951). Availability and irreversibility in thermodynamics. British Journal of Applied Physics, 2, 183.
Lavenda, B. H. (1993). Thermodynamics of irreversible processes..
New York: Dover.
Lions, J. L. (1971). Optimal control of systems described by partial
diwerential equations.. Berlin: Springer.
Ortega, R., Loria, A., Nicklasson, P. J., & Sira-Ramirez, H. (1998).
Passivity-based control of Euler}Lagrange systems.. New York:
Springer.
Ray, W. H. (1978). Some recent applications of distributed parameter
systems theory*a survey. Automatica, 14, 281}287.
Russell, D. L. (1978). Controllability and stabilizability for partial
di!erential equations. SIAM Review, 20, 639}739.
Sepulchre, R., Jankovic, M., & Kokotovic, P. (1998). Passivity-based
control of Euler}Lagrange systems.. New York: Springer.
Siep, W., & Willems, J. C. (1991). Dissipative dynamical systems in
behavioural context. Mathematical Models and Methods in Applied
Sciences, 1, 1}25.
Slotine, J. J., & Li, S. (1992). Applied nonlinear control.. Englewood
Cli!s, NJ: Prentice-Hall.
Smoller, J. (1983). Shock waves and reaction-diwusion equations.. New
York: Springer.
Struwe, M. (1996). Variational methods. applications to nonlinear partial
diwerential equations and Hamiltonian systems.. New York: Springer.
Willems, J. C. (1974). Dissipative dynamical systems, Part I. Arch. Rat.
Mech, 45, 321}351.
Ydstie, B. E., & Alonso, A. A. (1997). Process systems and passivity via
the Clausius}Planck inequality. Systems & Control Letters, 30, 253.
Ydstie, B. E., & Viswanath, K. (1994). From thermodynamics to process
control. Symposium on process systems engineering, Kyongju, Korea.
B. Erik Ydstie was born in White Plains
New York in 1952 and he grew up in
Norway. He holds a BS degree in Chemistry from the Norwegian University of
Science and Technology and a Ph.D. in
Chemical Engineering from Imperial College in London. His research topic was
predictive self-tuning controllers with variable forgetting factors. He taught Chemical Engineering at the University of
Massachusetts at Amherst from 1982 till
1992. He is presently a Professor in Chemical Engineering at Carnegie
Mellon in Pittsburgh. From 1999 to 2000 Erik Ydstie was Director of
R&D with ELKEM ASA in Oslo, Norway. During his tenure he
initiated new research programs and restructured the R&D organization. In 2001 Erik Ydstie was engaged as Professor II of Materials
Technology at Norwegian University of Science and Technology. He
has held visiting appointments at the Czech Academy of Sciences, the
University of New South Wales in Newcastle, Ecole Nationale
Superieure des Mines de Paris and Imperial College. His current areas
of research are process modeling and control, adaptive systems, supply
chain management and solar cell production processes. In his spare
time he likes to be with his family, read ski or sail.

A. A. Alonso, B. E. Ydstie / Automatica 37 (2001) 1739}1755


Antonio A. Alonso was born in Ribas de Sil,
Spain in 1966. He received a BS degree in
Industrial Chemistry from the University of
Santiago de Compostela in 1990 and obtained the Ph.D. degree in Chemical Engineering from the same University in 1993.
After a one year post-doctoral period at
Carnegie Mellon University, he joined the
Department of Chemical Engineering at the
University of Vigo. From 1995 to 1998 as
Assistant Professor and since then to present as Associate Professor. Currently he is engaged as a Tenured

1755

Researcher in the area of process engineering at the CSIC


(Spanish Council for Scienti"c Research). He has held visiting appointments at Carnegie Mellon and Princeton Universities. His research
interests include dynamic analysis and control of distributed (convection-di!usion) nonlinear (reaction) process systems, modeling and optimal control of biochemical networks and process design and control
integration.

Potrebbero piacerti anche