Stationary and Dynamic Flowsheeting in the

Chemical Engineering Industry

Edda Eich-Soellner, M
unchen,
Peter Lory, Regensburg,
Peter Burr,
Andreas Kr
oner, Hollriegelskreuth
Summary. Mathematical models have become indispensable tools in the design and the operation of chemical plants. Practical experience in this field during the last decade has proven modern equationoriented simulation techniques
superior to the classical sequential modular approach. The equationoriented
concept strongly relies on efficient numerical methods for large, sparse algebraic
and differentialalgebraic systems of equations. The aims of the present paper
are to introduce mathematical models for the most important unit operations in
chemical engineering and to survey numerical methods for the solution of large
systems of equations arising from modeling entire chemical production plants.
Special attention is given to numerical problems still encountered in productive
applications in the process industry.
AMS Subject Classifications (1991): 65C20, 65H10, 65L05, 65M20, 80A99.

Keywords and Phrases: Chemical process simulation; Mathematical models; Large nonlinear systems of equations; Sparse matrices; Method of lines;
Differentialalgebraic equations.

Introduction

The last decades have seen an increasing competition in the chemical engineering
industry. This has forced companies to develop and apply mathematical simulation techniques to improve their products. Main issues in this context are
plant economy and safety, minimum feedstock and utilities consumption and
environmental impacts. The pursuit of these goals leads to a rapidly growing
number of applications of computer simulations both in the design phase and
in the operating phase of chemical plants such as refineries, ethylene or gas processing plants or air separators. These examples give some of the applications
the authors have been involved in.
The most important steps in these processes are transformation of the components by chemical reactions and separation of mixtures to obtain certain purities. The effects of the processes are based on different physical and chemical
properties such as volatility, dew and bubble point, density and solubility of the
components in mixtures.
The key tool for modeling a chemical engineering plant is the flowsheet, an
abstract, graphoriented scheme of a planned or existing chemical plant. It
contains in its nodes the relevant single process steps, so called unit operations,
which are modelled independently. A unit operation comprises either an apparatus like a pump, a valve, a distillation column, a heat exchanger or a reactor
as a whole or only one or more distinct parts of a more complex process step.
The key function of a unit operation is to transform the incoming material and
information streams to give new output streams. The directed edges of the
graph represent the ideal flow of material, energy and information among the
unit operations. The flow of material in pipelines is modeled by individual unit
operation only if transportation deadtime and the effects of turbulent/laminar
flow is to be considered.
Figure 1 gives an example for a flowsheet of an ethylene plant. The central
role of flowsheets has made the term flowsheeting a synonym for the simulation
of chemical engineering plants.
Most of the large scale industrial plants operate in a steady-state mode. The
feedstock properties are constant in quantity and composition and - in the ideal
case - the plant continuously yields products at a constant rate and quality.
During the design or process synthesis of such plants an intense feedback from
the customer is indispensable to achieve the goals that he sets. In recent years
customers requirements increased drastically concerning purity of products,
minimization of feedstock, energy consumption, equipment costs and build time.
Additionally, legal constraints for environmental impacts have to be strictly
obeyed. With all major constraints in mind the necessary basic process steps
are determined in a first design phase, called basic engineering. Then, the
flowsheet is refined gradually in the phase of detailed engineering, leading to
the exact dimensions of all unit operations, e.g. the areas of heat exchangers,
volumes of tanks or number of trays in distillation columns. Also, different
2

Figure 1: Flowsheet of an ethylene plant

operating scenarios (summer/winter, day/night, different feedstock qualities)
are investigated. Stationary flowsheeting is an indispensable tool for all of these
tasks. In any cases mentioned, the mathematical problem is to solve a large
system of nonlinear discontinuous equations with a sparse Jacobian matrix.
A more sophisticated approach to meet the above mentioned goals in a systematic way makes use of mathematical methods of optimization. Clearly, the
objective function typically represents feedstock or energy consumption or the
costs for the construction of the plant. The equality constraints are the equations that describe the flowsheet. Finally, inequality constraints typically force
the purity of the products to obey certain limiting requirements of the customer
or guarantee the satisfaction of legal environmental restrictions. The simultaneous optimization of various continuous process parameters together with integer quantities such as column feed or side draw points calls for mixed-integer
nonlinear programming methods. These techniques allow even the systematic
optimization of structural alternatives in the flowsheet network.
The operation of an actual plant poses a set of new tasks to be solved by
optimization methods: First, the plant model has to be fitted to actual plant
measurements by adjusting free parameters using data collected over an extended period of stable operation (parameter tuning). Second, a regression
of measured data on the basis of this improved model gives reliable information
on the exact state of the plant (data reconciliation). Third, the new target
setpoints of the control system are determined by mathematical optimization
on the basis of the previously reconciled plant model (setpoint optimization).

All three steps are repeated in a cyclic manner. The first two tasks are inverse
problems and use suitable versions of the least-squares function as objective
functions. The objective functions for the latter task usually include maximum
current profit and production yield as well as minimum feedstock and utilities
consumption. Equality constraints arise from the equations describing the flowsheet. Inequality constraints typically force the purity of the products to obey
certain minimum requirements of the customer or guarantee the satisfaction
of legal environmental restrictions. An efficient application of all three steps
requires the on-line operation of the control system and the plant model. Online operation in this case comprises data transfer in both directions - from the
process to the plant model and vice versa - at unique user controlled points in
time. The simultaneous optimization of various continuous process parameters
together with integer quantities such as column feed or side-draw points calls
for mixed-integer nonlinear programming methods in the off-line analysis.
In recent years dynamic simulation has become an increasingly important
tool in the process industries. It is used in a variety of off-line applications,
i.e. with no direct interactions with the actual plant, like operability studies or
safety and risk analyses. It allows the investigation of start-up and shut-down
procedures and, together with optimal control techniques, even their systematic
optimization. Most important, dynamic flowsheeting is the basis for the design
of standard and advanced control strategies and a platform for the preliminary
implementation of the control systems. This is decisive for a stable and efficient
operation of a plant.
Furthermore, dynamic simulation models form the heart of operator training
systems. These simulation programs are used long time before the completion
of the plant to implement distributed control systems and to train future operator personnel in monitoring and controlling the design states of operation
as well as emergency situations. These applications make tough demands on
the efficiency and robustness of the integration algorithms. To represent the
behavior of the actual plant the model must be solved in real-time or faster.
Data exchanged with the on-line operating control system and operator input
account for permanent discontinuities in input values.
Dynamic simulation is increasingly used for dynamic on-line optimization
and identification of process information not directly available through measurements. Both applications are an integral part of modern model-based control systems. Mathematically, dynamic process simulation is governed by large
systems of differentialalgebraic equations with discontinuities.
Principally, simulators for the above mentioned tasks follow two different
concepts in setting up and solving the plant simulation model:
the sequential modular approach and
the equationoriented approach.
Both strategies are compared and discussed e.g. in [2, 17, 26, 32, 43]. Practi4

cal experience during the last decade has proven equationoriented simulation
technique superior to the sequential modular approach. It strongly relies on
efficient numerical methods for large, sparse algebraic and differentialalgebraic
systems of equations.
The present paper focuses on these mathematical aspects of stationary and
dynamic flowsheeting. Practical experience of the authors in these fields is
strongly influenced by the development of the equationoriented simulator
OPTISIM1 , an inhouse development of Linde AG [6]. If not stated otherwise,
model equations and numerical solution methods in the subsequent sections
follow the approaches in this software product. It combines stateoftheart
mathematical methods with a huge amount of experience originating from a
daytoday use of various simulators in different areas of flowsheeting. The
paper includes also results from individual research of the authors.
Section 2 presents the mathematical models of the most important units.
For the simulation of a plant, these units must be combined to a flowsheet.
This process is described in Section 3 with the help of a simple, yet instructive
example. Section 4 focuses on stationary flowsheeting. It describes the numerical solution of the steady state equations, especially the treatment of the large,
sparse systems and remedies for typical numerical difficulties. This section also
surveys techniques for steady state optimization. Section 5 describes properties
of differentialalgebraic equations arising in dynamic flowsheeting. Special attention is given to the index of the model equations for typical unit operations
and to the treatment of discontinuities.

Mathematical models of unit operations

This section introduces dynamic simulation models of the most important process steps found in chemical and refinery plants. All models are based on conservation laws for mass, energy and momentum. Since time constants related
to mass and energy holdups are several magnitudes larger than those found in
momentum transport phenomena and dominate the dynamic behavior, mass
and energy balances are formulated dynamically whereas momentum balances
are usually applied in a simplified and quasi-stationary form. Trivially, for directly calculating a steady state solution with the model equations discussed,
accumulation terms on the left hand sides of the balance equations have to be
neglected.

2.1

The flash unit

A basic operation in process engineering is the singlestage thermal separation

of a multicomponent mixture (see Figure 2). The mixture is heated to the
bubblepoint. The ascending vapor is richer in the low boiling, i.e. volatile,
1 OPTISIM
R

is a registered trademark of Linde AG.

V, yj , hV , TV, PV
Vapor

F, z j, h ,T , P

Feed

Heat Q

L, xj , h, T , P

Liquid

Figure 2: Flash unit

components than the liquid. As the vapor is removed continuously, the original
mixture gets poorer in the light, more volatile components. The residue in the
flash drum becomes richer in the heavy, less volatile components.
A single-stage equilibrium separation with neglectable holdup in the vapor
phase is described by the following equations:
d
M
dt
d
(M xj )
dt
nc
X
xj
j=1
nc
X

= F V L ,

(2.1a)

= F z j V yj L x j

for j = 1, . . . , nc 1 ,

(2.1b)

= 1,

(2.1c)

yj

= 1,

(2.1d)

yj
L

= Kj (T, P, x, y) xj
= (M ) ,

j=1

d
(M hL )
dt
hL
P

for j = 1, . . . , nc ,

= F hF V hV L hL + Q ,

(2.1g)

= hL (T L , P L , x) , hV = hV (T V , P V , y) ,

(2.1h)

= P , PL = P , TV = T , TL = T .

(2.1i)

These equations determine the following state variables:

M:

(2.1e)
(2.1f)

molar liquid holdup in the flash drum,

6

L,V :
xj :
yj :
h L , hV :
T ,TL ,TV :
PL ,PV :

molar liquid and vapor flowrates,

mole fraction of component j = 1 . . . nc in the liquid,
mole fraction of component j = 1 . . . nc in the vapor,
molar enthalpies of liquid and vapor,
temperatures in the drum and in the products,
pressures in the products.

The pressure P in the drum, the added heat Q, the molar feed flowrate F as
well as its mole fractions zj and its molar enthalpy hF are given. The number of
components is denoted by nc . The row vectors x and y collect the mole fractions
of the liquid and the vapor, respectively:
x = (x1 , . . . , xnc ) ,

y = (y1 , . . . , ync ) .

(2.2)

Equation (2.1a) is the all over material balance, while (2.1b) are individual
material balances for the first nc 1 components. The last component is determined by (2.1c). This equation as well as (2.1d) expresses the fact, that the
mole fractions in liquid and vapor sum up to 1 . Thermodynamic equilibrium is
ensured by (2.1d) and (2.1e). The latter equations describe the relation between
the mole fractions in the vapor phase and in the liquid phase. The so-called
Kvalues Kj depend on pressure, temperature and mole fractions of the liquid
and vapor phase. They are determined by a physical property system based on
given T , P , x, and y.
It is assumed, that the flowrate L of the liquid leaving the drum is a known
function of the holdup M in the drum (2.1f). Equation (2.1g) is a total
enthalpy balance, with the liquid and vapor molar enthalpies defined by (2.1h).
Both values are also evaluated in a physical property system from T , P , and
x or y, respectively. Equations (2.1i) equate pressures and temperatures in the
product flows to those in the drum. The system consists of 2nc +10 equations for
the same number of unknowns. For variants of the flash unit and its applications,
see [41].

2.2

The distillation column

The extent of separation in a singlestage flash unit is limited to the composition

difference between liquid and vapor phase at a given pressure P and boiling
temperature. To increase the effect of separation and to achieve higher purities
numerous equilibrium stages at different pressures and temperatures are stacked
to give a tray distillation column (Figure 3). Pressure and temperature decrease
from bottom to top. The simulation model of a distillation column is easily
derived from the flash unit model described in the preceding section.
d
Mi
dt

= Li+1 + Vi1 Li Vi + Fi ,

(2.3a)







Figure 3: Distillation column

d
(Mi xi,j ) = Li+1 xi+1,j + Vi1 yi1,j Li xi,j Vi yi,j + Fi zi,j(2.3b)
dt
for j = 1, . . . , nc 1 ,
nc
X
xi,j = 1 ,
(2.3c)
j=1
nc
X

yi,j

= 1,

(2.3d)

yi,j
Li

= Ki,j (Ti , Pi , xi , yi ) xi,j

= (Mi ) ,

j=1

for j = 1, . . . , nc ,

d
L
V
L
V
(Mi hL
i ) = Li+1 hi+1 + Vi1 hi1 Li hi Vi hi
dt
+Fi hF
i + Qi ,
L
L
hi = hi (Ti , Pi , xi ) ,
hVi = hVi (Ti , Pi , yi ) ,
Pi Pi+1 = (Mi+1 , Vi ) .

(2.3e)
(2.3f)
(2.3g)
(2.3h)
(2.3i)

In this system, i = 1, . . . , N denotes the number of the tray, while j = 1, . . . , nc

is the number of the component in the mixture. Similarly to (2.2), the row

vectors xi and yi are defined as

xi = (xi,1 , . . . , xi,nc ) ,

yi = (yi,1 , . . . , yi,nc ) .

(2.4)

The N (2nc + 7) state variables are

Mi :
Li :
Vi :
xi,j :
yi,j :
V
hL
i , hi :
Ti :
Pi :

molar liquid holdup on stage i,

molar liquid flowrates descending from stage i to stage i 1,
molar vapor flowrates ascending from stage i to stage i + 1,
mole fractions of component j in the liquid at stage i,
mole fraction of component j in the vapor at stage i,
molar enthalpies of liquid and vapor at stage i,
temperature at stage i,
pressure at stage i.

The external heat input Qi to stage i, the feed flowrate Fi to stage i as well as
its mole fractions zi,j and its molar enthalpy hF
i are given. The system (2.3a)
(2.3i) consists of N (2nc +7) equations. The following specifications complete the
system: the reflux flowrate LN +1 to the top tray together with its mole fractions
xN +1,j and its molar enthalpy hL
N +1 ; the vapor flowrate V0 to the bottom tray
as well as its composition y0,j and the molar enthaly hV0 ; the external pressure
at the column top PN +1 .
The rationale for these equations is the same as that for a single stage separation (see Sec. 2.1). Only the material and enthalpy balances have to be
modified adequately. Equation (2.3i) expresses the pressure drop between two
consecutive stages in terms of resistance incurred by the vapor flowing through
the holes or valves in the tray, and also through the static height of liquid on
the tray. The exact forms of the liquid flow function and vapor flow function
depend on the type and geometry of the trays used. The pressure drop equation for the top tray i = N is usually formulated in a different way than the
pressure drop across an internal tray since there is no external holdup MN +1 .
Usually, the number N of trays is within 10 N 100 , and the number nc of
components is restricted to nc 50 . For a more detailed discussion, see [18]
and [31].

2.3

The heat exchanger

The main purpose of heat exchangers is to transfer energy among two or more
fluid stream separated through the heat exchanger material. Hot streams leave
the heat exchanger at a lower temperature as they enter the apparatus. The
transferred energy is used to heat up cold streams on their way through the
equipment. Depending on the relative flow direction in the passages of the
exchanger there exist three principal modes of operation: cocurrent, countercurrent, and crosscurrent. The decision for a certain mode of operation depends
9

on the number of passages, the maximum energy transfer rate, or constructive

limitations. In low temperature processes, such as air separators or parts of
an ethylene plant, countercurrent heat exchangers are most efficient and are
therefore used predominantly. For demonstration purposes, Fig. 4 and the following model equations are restricted to two streams. The generalisation to
more streams is straightforward.

v1

stream 1
metal
stream 2

v2

q1

6
q2

x=0

h1 , T1
hm , Tm
h2 , T2
x=L

Figure 4: Heat exchanger

It is assumed that all streams exchange heat with the heat exchanger metal.
The flow velocities v1 and v2 are taken constant. The enthalpy flowrates
h1 (x, t), h2 (x, t) as well as the temperatures T1 (x, t), T2 (x, t) of the two streams
depend on the spatial variable x and on time t . The state of the metal is
described by the enthalpy hm (x, t) per unit length and by the temperature
Tm (x, t) . The energy transferred from a fluid to the metal is given by the heat
transfer equation defining the energy flux per unit length
qi = i i (Ti Tm )

(i = 1, 2).

Here, i = i (Ti , Pi , zi ) is the heat transfer coefficient between the i-th stream
and the heat exchanger wall. It depends on the flow regime and on physical
properties and hence on temperature Ti , pressure Pi and composition zi =
(zi,1 , . . . , zi,nc ) of the stream. The quantity i denotes the heat transfer area
per unit length for the i-th stream.
The effect of heat transport in axial direction due to conduction and dispersion is usually much smaller than that of convection and is therefore neglected
here. Consequently, the model does not contain diffusive terms and is of hyperbolic type:
1
v1
1
v2

h1
t
h2
t

=
=

h1
1 1 (T1 Tm ) ,
x
h2
2 2 (T2 Tm ) ,
x

10

(2.5a)
(2.5b)

hm
t
0
0
0

1 1 (T1 Tm ) + 2 2 (T2 Tm ) ,

(2.5c)

=
=
=

hstream (T1 , P1 , z1 ) h1 ,
hstream (T2 , P2 , z2 ) h2 ,
hmetal (Tm ) hm .

(2.5d)
(2.5e)
(2.5f)

The physical property functions hstream and hmetal depend on the temperature and in the case of a stream also on pressure and composition as indicated
in (2.5d)(2.5f). The concentrations are assumed constant along each heat exchanger tube, the pressure is assumed constant or varying linearly with the
exchanger length. The system is completed by the initial conditions hi (x, 0),
hm (x, 0) and the boundary conditions h1 (0, t), h2 (L, t) .

2.4

The ideal continuous flow stirred tank reactor

cj,F , TF , F
?
'

@
@

&

cj
T
ri
V

cp
%
? cj , T, F

Figure 5: A continuous flow stirred tank reactor

The continuous flow stirred tank reactor consists of a vessel of constant liquid
volume V furnished with inlet and outlet nozzles, a cooling water jacket and
a stirrer assuring a uniform concentration and temperature distribution in the
reactor volume. Inlet and outlet flow rate are assumed constant at rate F . Let
Aj , j = 1, . . . , nc denote the chemical species which are taking part in a system
of M reactions. The equations for these reactions, which are assumed to be
11

independent, may be written as

nc
X

i,j Aj = 0 ,

i = 1, . . . , M .

(2.6)

j=1

Here, i,j is the stochiometric coefficient of component Aj in the i-th reaction. Stochiometric coefficients are taken positive for products and negative for
reactants.
Let cj,F be the molar concentration of Aj in the feed to the reactor, and
cj the molar concentration in the reactor and hence in the outflowing product.
Then a mass balance for Aj gives
V

M
X
dcj
= F (cj,F cj ) + V
i,j ri ,
dt
i=1

j = 1, . . . , nc .

(2.7)

Here, ri is the total rate of reaction for the i-th reaction. For many technical
reaction systems it is given by the power law rate equation
ri = ri (c1 , . . . , cnc , T ) = ki

nc
Y
j=1

cj i,j

ki0

nc
Y

cj i,j .

(2.8)

j=1

The Arrhenius rate constants ki and ki0 of the i-th reaction in forward and
backward direction obey the Arrhenius equation:

E0
E
0
0
,
k = k0 exp
,
(2.9)
k = k0 exp
RT
RT
where T is the temperature in the reactor. The constants k0 , k00 and E, E 0 are
frequency factors and activation energies, respectively. As usual, R is the gas
law constant. The exponents i,j and i,j denote the order of the i-th reaction
with respect to Aj in forward and backward direction, respectively. They obey
the relation
i,j i,j = i,j .
For a reversible reaction each direction must be considered separately, whereas
for an irreversible reaction only the forward direction is relevant, hence ki0 =
i,j = 0 .
Let TF denote the temperature in the feed. Then, under reasonable assumptions (e.g. constant pressure) an enthalpy balance for the reactor gives
V cp

M
X
dT
= F cp (TF T ) + V
(H,i )ri + Q .
dt
i=1

(2.10)

Here, cp is the constant molar heat capacity of the reaction mixture, is the
constant density, H,i is the heat of the i-th reaction, and Q is the rate of heat
entering the reactor. For more details, the reader is referred to [1, 37].
12

2.5

Physical property data

The model equations require physical property data, e.g. density, enthalpy or
volume. They are given by explicit functions or as solutions of highly nonlinear
implicit equations (thermodynamic state equations). In order to obtain the
physical property data asked for by a unit model, these equations must be
evaluated or solved.
The thermodynamic state equations implicitly relate temperature, pressure
and volume. Often these equations, like the cubic state equation, have multiple
solutions for certain variables. So, the choice of the solution to return depends
on the thermodynamic state of the system. The physical property equations
are solved iteratively in the physical property data system each time a unit
operation model is evaluated. This approach is used in almost all steady state
and dynamic simulation systems.

A flowsheet example

In this section the steps on the way to the simulation model of an actual process
or plant are shown. The small example process in Fig. 6 is first explained
and structured in a flowsheet. The model equations of the unit operations in
the flowsheet are then combined to give the complete simulation model of the
process.
F I , cIA,0 , T0I

F II , T0II
Membrane

?
TI
QI- cIA , cIB
VI A B

?
M
Jm,B
M
JQ
, JHM
-

T II
II
cII
Q
B
V II

TC
?
F I , cIA , cIB

?
F II , cII
B

Figure 6: Two phase process

Fig. 6 shows the schematic of a tank reactor to produce component B from
component A. The interior of the tank is separated in a reaction phase (I) and
a product phase (II). The exothermic first order irreversible reaction in phase I

13

is
A

with heat of reaction H,R < 0.

Both phases are separated by a selective membrane which is permeable only
for component B. Heat is exchanged between both phases across the membrane.
Component A is fed into phase I in an aqueous solution by the volume
stream F I , with feed concentration cIA0 and feed temperature T0I . The outlet
from the reaction phase is given by volume stream F I with concentration cIA
and cIB of components A and B and temperature T I of the reaction volume V I
. The pure washwater stream F II with the temperature T0II is fed to volume
V II of the product phase. Its outlet temperature is T II and the concentration
of component B is cII
B.
Both phases exchange heat with the cooling water through heat streams QI
and QII .
For reaction systems some additional modeling assumptions are valid [37]:
Densities and heat capacities are constant and equal to those of the pure
solvent.
No phase changes occur; all phases are liquid.
Heat and mass transfer coefficients are constant.
The temperature T C of the cooling water is constant.
To set up a simulation model the process is first divided into smaller units to
give the flowsheet of Fig. 7. The central unit operations in the flowsheet are a
reactor and a tank unit with dynamic material and energy balances modeling the
reaction phase (3.1) and product phase (3.2). The rate of reaction is modeled by
the Arrhenius equation (3.1d). The membrane is assumed to have no storage
capacities (3.5). The mass transfer stream (3.5b) is modeled proportional to
the concentration difference in both phases. The energy exchange comprises
M
two phenomena. The conduction heat flow JQ
(3.5a) is proportional to the
M
temperature difference of both phases. The enthalpy transport JH
(3.5c) is
caused by mass transfer through the membrane. It is proportional to the mass
M
flow Jm,B
and the temperature difference across the membrane.
The heat exchanged with the cooling water is modeled for both phases in
separate elementary exchanger units assuming the heat transfer to be proportional to the temperature difference between the phase temperatures and the
cooling water (3.3,3.4).
dcIA
dt
I
dc
VI B
dt

VI

= F I (cIA0 cIA ) r0 V I

(3.1a)

M
= F I cIB + r0 V I Jm,B

(3.1b)

14

stream source
QI
heat exchanger
QI
6

TC
?

T C

T C-

stream source

heat exchanger
QII

QII

stream source

cIA,0

F II , T0II
T II
II
F I , T0I cI
c
B
B
?
?

II
TI
T
reactor
tank
?- membrane
?
no storage
mass and energy
mass and energy QII
QI
capacity
balances
balances
M
M
M
II
cIA , cIB , T I , r0
JQ
, Jm,B
, JH
cII
B ,T
M
M
JQ
JQ
TI

cIA , cII
B
F I, T I

M
JH
M
Jm,B

M
JH
M
Jm,B

cII
B
F
, T II
?
II

Figure 7: Flowsheet of the two phase process

V I cp

dT I
M
= F I cp (T0I T I ) + QI JQ
JHM r0 V I H,R
dt
E
0 = r0 cIA k0 e RT I

dcII
B
dt
dT II
V II cp
dt
V II

(3.1c)
(3.1d)

M
= F II cII
B + Jm,B

(3.2a)

M
= F II cp (T0II T II ) + QII + JQ
+ JHM

(3.2b)

0 =

QI k I AI (T C T I )

(3.3)

QII k II AII (T C T II )

(3.4)

15

M
JQ
k M AM (T I T II )

(3.5a)

M
Jm,B
AM (cIB cII
B)

(3.5b)

JHM

(3.5c)

cp (T T

II

M
)Jm,B

Source units for information and material streams are included to define
required input information.
The procedure to analyze a plant and to set up the flowsheet is similar
in almost all stationary and dynamic flowsheet oriented simulation packages
[6, 20, 24, 30]. However, they differ significantly in the solution strategies and
hence in the way the system of equations is generated.
In OPTISIM the model equations with their partial derivatives are available
in executable form in the model library. Prior to a simulation, the flowsheet information is analyzed, input and output variables of the unit models are assigned
internally and a calling sequence to the model equation programs is generated.
If a numerical solution algorithm asks for a function evaluation of the simultaneously solved plant model the sequence is interpreted for calls to the equation set
of the individual unit operations. The individual function evaluations are combined to give the entire plant information. Similar strategies are implemented
in DIVA [20].
The resulting state vector associated with the simulation model (3.1)(3.5)
is
II
I
II
M
M
M T
x = (cIA , cIB , T I , r0 , cII
.
B , T , Q , Q , JQ , Jm,B , JH )
Input or control quantities
u = (cIA0 , T0I , F I , T0II , F II , T C )T
of the system are provided by input unit operations.
In SpeedUp [30] model equations are provided and analyzed symbolically.
The program code of the equation system is generated and compiled as a whole
before a simulation can be run. Somewhat different simulation strategies are
pursued in HYSYS [24] and PROTISS2 .

Stationary flowsheeting

Spatial discretization of the heat exchanger equations. The spatial

discretization of the heat exchanger equations is described here for the dynamic version already. Trivially, this includes the stationary case. The set of
coupled partial differential and algebraic equations (2.5) is transformed into a
2 PROTISS

is a product of SIMSCI, UK.

16

differential-algebraic system of the type (5.1) by the method of lines. Here, finite
element approximations
hi (x, t) =

N
X

i,k (t) k (x)

(4.1)

k=1

with basis functions k (x) and time dependent coefficients i,k (t) are used for
the spatial discretization of the partial differential equations (2.5a)(2.5c). Stability requires upwinding within the frame of a PetrovGalerkin method. This
technique takes test functions k (x), which are different from the basis functions, and leads to upwinded, nonsymmetric Galerkin equations (see e.g. [40]).
The first derivative of the physical property function h (enthalpy) is discontinuous at the temperatures T = TBP (BubblePoint) and T = TDP (Dew
Point) due to phase changes. Consequently, the solutions of the system (2.5)
show discontinuities at those locations where the phases change. Naturally, their
positions in the heat exchanger depend on time t. A discretization, that does
not destroy the order of the approximation, requires the knowledge of these positions xBP (t) and xDP (t) . They provide valuable information to the process
engineer, too. Therefore, finite elements with moving nodes [28] are used. The
principle is outlined here for a single phase change location xBP (t) . The key is a
moving coordinate system that tracks xBP (t). For example, a time dependent,
piecewise linear coordinate transformation x = x(; xBP (t)) with the properties

L
x(0; xBP (t)) = 0 ,
x
; xBP (t) = xBP (t) ,
x(L; xBP (t)) = L
2
defines a new coordinate system with the variable , in which the phase change
takes place at the constant position = L/2 (cf. [36]). Assuming that the phase
change takes place in stream 1, the additional variable xBP is determined by
the equation
T1 (xBP (t), t) TBP = 0 .
(4.2)

4.1

Steady state simulation

Large scale chemical plants deliver products at a constant rate with constant
quality, i.e. concentration, temperature, etc., based on feeds available at uniform
conditions. In other words, the plants operate in a steady state mode. The corresponding mathematical task is the solution of a set of nonlinear equations
F (x, p, sign q)

0,

(4.3a)

q(x, p)

(4.3b)

where F <nx are the model equations with all time derivatives set to zero.
The vector x collects state variables such as concentrations, material and energy
17

flow rates, temperatures, pressures, or material and energy holdups. The vector p summarizes the given parameters, e.g. exchanger areas or vessel volumes.
The so-called switching functions q are used to switch among different states of
the system described by different subsets of equations. Changing from one set
of equations to an other causes discontinuities in the solution of (4.3a). Typical reasons for discontinuities in the steady state case are control devices, the
appearance or disappearance of phases, piecewise continuous physical property
correlations or limitations in rate equations.
This section shows how to obtain a steady state solution of a large system
of nonlinear, discontinuous equations by variants of Newtons method taking
into account the special structure and properties, e.g. sparsity, of the problem.
Approaches to obtain optimal steady states are discussed thereafter.
Eq. (4.3a) is solved using an inexact Newton method
F (i)
(x )x(i) = F (x(i) , p),
x

x(i+1) = x(i) + x(i) ,

(4.4)

where (0, 1] is the step length [38]. The system of linear equations in (4.4)
is solved iteratively, see below.
Solution of sparse linear equations. A simulation model for the on-line
optimization of an ethylene plant consists typically of about 40.000 equations,
the Jacobian has about 200.000 non-zero entries. Because of stability and computing time reasons, the high dimension of the systems requires sparse matrix
methods.

Figure 8: A typical sparsity structure

18

Fig. 8 gives a typical structure of non-zero elements in the Jacobian. Unfortunately, the matrices do not exhibit properties that would facilitate their decomposition, like bandedness, symmetry, positive definitness or diagonal dominance. Rather, general sparse matrix alogrithms have to be used for their
factorization [10]. These methods are combined with iterative methods for linear equations in order to minimize effort and to increase accuracy. The latter
feature is decisive because the matrices are often ill-conditioned (see e.g. [38]).
Principally, direct methods for sparse systems differ from the standard Gaussian algorithms in their pivot selection strategy. Sparse matrix methods do not
use only stability criteria, but also criteria to minimize additional fill-in during
the decomposition phase. In general, the solution of a sparse linear system is
structured in three phases:
1. Analyze phase: The structure of non-zeroes is analyzed and a pivot sequence is determined.
2. Decomposition phase: The LU -decomposition is computed based on the
previously selected pivot-sequence.
3. Solution phase: For a given right hand side a solution is determined by
forwardbackward substitution.
The computational effort decreases from step 1 to step 3.
In OPTISIM the computation of the first two steps is avoided as far as
possible by iteratively solving the linear equations (4.4). The equations are preconditioned with a previously decomposed matrix, so that the factorization of
the current Jacobian need not be computed. Only in the case of poor convergence, a change in the structure of the Jacobian, or discontinuities the solution
procedure is restarted with phase 1 or phase 2.
The use of sparse matrix methods in OPTISIM is particular facilitated because all unit models provide analytical derivatives. Consequently, the sparsity
structure is explicitly known and no numerical tests with finite differences are
necessary to decide if an element is 0 or not.
For the solution of the linear system (4.4) iterative methods GMRES (generalized minimal residual) [34, 42], biconjugate gradient [33] or iterative
refinement together with the above mentioned preconditioning are applied. For
a comparative survey, see [35].
Discontinuities. Discontinuities occur because the structure of equations or
parameter values depend discontinuously on values of state variables. If this is
the case, discontinuities are inevitable during the solution of the nonlinear system (4.3a) because the final solution and hence the structure of model equations
and certain parameter values are unknown initially.
A discontinuity occurs whenever a switching function (4.3b) changes sign.
Different impacts on the equation system must be handled:
19

a: Changes in the equation with no effect on the Jacobian,

b: Changes in the values of the Jacobian elements while keeping the sparsity
structure unchanged,
c: Changes in the dependencies of the equations leading to a different sparsity
structure of the Jacobian,
d: Changes in the number of equations and variables.
Each of these discontinuities require different actions w.r.t. convergence control
(a), recomputation, reanalyzation and decomposition of the Jacobian (b, c) or
even reinitialization of the simulation model (d). The OPTISIM program keeps
track of all discontinuities of the equations.

4.2

Steady state optimization

As discussed in the introduction, flowsheeting in practice often requires the

solution of optimization problems:
!

(x, p) =
xmin

F (x, p)
x xmax ,

min ,

(4.5a)

0,
pmin p pmax .

(4.5b)
(4.5c)

x,p

In the optimization problem (4.5), let the vector

h(x, p) =

(x xmin )T ,

(xmax x)T ,

(p pmin )T ,

summarize the inequality conditions and s =

variables. Then the system
!

(s) =

min ,
s

F (s) = 0 ,
h(s) 0

xT ,

pT

(pmax p)T
T

the optimization

(4.6a)
(4.6b)
(4.6c)

has to be solved. This is done by the iteration s(i+1) = s(i) + s(i) starting from
initial values s(0) . The main difference among optimization methods is the way
the search direction s(i) is computed.
Sequentiallinear programming (SLP) is popular in the chemical engineering
community because of its robustness. Here, the objective function and the
constraints are linearized around the current point s(i) . Then the solution s(i)

20

of the linearized problem

(s(i) ) s(i)
F (i)
(s ) s(i)
s
h (i)
h(s(i) ) +
(s ) s(i)
s

F (s(i) ) +

min

s(i)

(4.7a)

= 0

(4.7b)

(4.7c)

is computed. Because this problem may be unbounded, an additional constraint

on the step length is added
ks(i) k (i) .
(4.8)
In order to ensure that the linearization is a good approximation to the nonlinear
problem (i) is controlled. For details about this trust region approach, see
[14].
In contrast to the SLP method, the objective function in sequential quadratic
programming (SQP) is approximated quadratically. Starting from a positive
definite matrix, the Hessian is approximated successively using the BFGSformula [12, 14].
Feasible path and infeasible path methods. Formulation (4.5) suggests the interpretation of the flowsheet equations (4.5b) as optimization constraints and to solve them as part of the optimization problem. Therefore,
optimization, i.e. the determination of p, and the solution of flowsheet equations, i.e. the determination of x for given p, are performed simultaneously in
the infeasible path approach.
This

can be done efficiently only if the sparsity

F
structure of the Jacobian F
,
and of the constraints is exploited. Effix p
cient software for this task is still under development. This is currently the
main reason not to use the above infeasible path approach, but a feasible
path method: x = x(p(i) ) is computed from the flowsheet equations outside
the optimization procedure. The optimization solver can see only the optimization variables p. In general, feasible path methods require more effort,
but they have the additional advantage that in every optimization iteration a
valid solution of the flowsheet is available. This is especially advantageous if
the optimization is not iterated until convergence.
For the feasible path approach, sensitivities x
p are necessary to compute
derivatives of the objective function and constraints w.r.t. p, e.g.
x
d
=
+
.
dp
x p
p
They can be obtained solving the linear system of equations
F x F
+
= 0.
x p
p
21

Numerical results for the example of the two phase process can be found in the
Appendix.

Dynamic flowsheeting

As can be seen from the examples in Sec. 2 and 4 dynamic models of chemical processes result in large systems of linearly implicit differential-algebraic
equations (DAEs) of the form

A1 (y, z, p, u) A2 (y, z, p, u)
y
f (y, z, p, u)
=
.
(5.1)
0
0
z
g(y, z, p, u)

y
Let x =
be the state variables with x <nx , y, f <ny , z, g <nx ny ,
z
and A1 <ny ny nonsingular. Parameters p <np are known constants
whereas controls u <nu are given functions in time.
Besides pure dynamic simulation, control and optimal control problems, frequency analysis and real-time training simulation are important tasks to be
solved. The main focus of the following section is on the properties of the
DAE system (5.1) and on aspects of the numerical solution of DAE systems not
covered by standard DAE solution algorithms.

5.1

Typical properties of differential-algebraic equations

occurring in Chemical Engineering

In practical applications, flowsheet DAE systems which consist of thousands

of equations are common. For example, a rigorous model of an air separation
plant involves about 1.000 differential and 3.000 algebraic equations, whereas
ethylene plant models consist of some 30.000 equations.
The various unit operations show a wide span of time constants which leads
to a high degree of stiffness in the model equations. Fast components are controllers, reactors or machinery like compressors. In contrast, distillation columns
as well as storage tanks with large holdups account for large time constants.
Stiffness also results from the spatial discretization of partial differential equations in the dynamic simulation of heat exchanger models. The discretization of
a steady state heat exchanger model with streams of widely different flowrates
leads to stiffness w.r.t. to space.
An even more crucial property than stiffness for measuring difficulties in the
theoretical and numerical treatment of differential-algebraic equations, is their
index. Roughly spoken, the index is the number of times all or parts of the
equations must be differentiated totally w.r.t. time to obtain a set of ordinary
differential equations (5.2a) for all variables [3, 7, 13]
x =

(x, p, u)
22

(5.2a)

(x, p, u)

(5.2b)

In addition to (5.2a) a set of constraint equations (5.2b) is obtained defining

the solution manifold. Both sub-systems (5.2) are called the extended system
of the DAE system (5.1). A different approach to DAEs is found in [16] leading
to the perturbation index.
DAEs in chemical engineering may show an arbitrarily high index. In case
of the index exceeding 2, they cannot be treated directly and reliably with
modern standard integration methods [3]. To reduce the index to 1 or 0 the
DAE system must be analyzed to determine the number of total differentiations
of each equation.
For large scale nonlinear simulation models it is often not practicable to
analyze the DAEs analytically and to reduce the index symbolically. Structural
analysis algorithms have been proposed in [5, 29, 39] to determine bounds on
the index, the number of dynamic degrees of freedom and the number of total
differentiations of each equation to reduce the index to 1 or 0. The actual
index reduction is then carried out by automatic differentiation [15] or numerical
differentiation [22].
Both, automatic and numerical differentiation of equations with the aim to
reduce the index of DAE cause a loss of stability in the numerical solution.
This can be avoided using a projection method, see [11]. Another remedy is an
appropriate reformulation of the model.
The model equations of Sec. 2 and 3 have been analyzed for their index:
The flash unit model is of index 2. The difficulty is to determine a differential
equation for the vapor stream V . To avoid an index 2 problem the model
could be reformulated. If the pressure P in the drum and the vapor
pressure P V are used as state variables and the first equality in (2.1i) is
substituted by a valve equation relating the internal and external pressure
to the vapor stream V

V = (P,
PV )
an index 1 DAE system is obtained.
The distillation column model is an index 1 DAE system. However, if different specifications are assumed in the model higher index problems are
obtained [31].
The heat exchanger model has after discretization also index 1, both with
and without moving nodes.
The two phase process from Sec. 3 implemented as a dynamic simulation
problem with given input values has index 1 because each of the algebraic
equations can be solved for the corresponding variable. For some associated dynamic design cases where output variables are specified and input
variables are released, Table 1 gives the corresponding indices.
23

given
time function
cIA
T II
TI
cII
B

unknown
time function
cIA0
FI
cIA0
TC

Index
2
3
3
4

Table 1: Design cases for two phase process and their index

Example:
Another example illustrating the index problem in chemical engineering are two
coupled tanks as shown in Fig. 9.
F0

h1

F1

F2
2

R
A1

A2

Figure 9: Communicating tanks

Both tanks are connected through a valve with flow resistance R. The dynamic behaviour is governed by
A1 h 1
A2 h 2
F1 R
F2

=
=
=
=

F0 F1
F1 F2
h1 h2
f (h2 ) .

If R and the input stream F0 are given, the unknowns, i.e. the liquid heights
h1 , h2 and the flows F1 , F2 can be computed. The areas A1 , A2 are assumed
to be given. For R 6= 0 this is an index 1 system. For R = , i.e. the valve
is closed completely, the flow F1 becomes zero and the index remains 1. If the
flow resistance R is decreased, the stiffness of the system increases. For R = 0,
the levels in both tanks adjust instantaneously and the index of the system is
2. It is important to see that a direct algebraic coupling of balanced quantities
increases the index of the system.
24

Fig. 10 gives the simulation results for

R0 ( 12 + 12 cos( tt0 ))
R(t) =
0

t < t0
,
t t0

with initial values h1 (0) = 1, h2 (0) = 1, f (h2 ) = h2 , t = 8, F0 = 2.

Two tank simulation

3
h1
2.5

2
h2

1.5

1
R
0.5

t0
0
0

10

15

Figure 10: Dynamic simulation of two communicating tanks

If a design problem is to be solved the index increases. Two design cases are
investigated:
Case a: Determine R such that

h2 = const.

h 2 = 0

F1 F2 =

h1 h2
.
F2

h1 h2
F2 ,
R

Thus, the index is 2.

Case b: Determine F0 such that

F2 = const.

f
f F1 F2
h2 = 0 =
,
h2
h2 A2

Two cases must be distinguished:

R 6= 0:
0=

f
h2

h1 h2
R

F2

A2

Differentiation leads to an equation involving h 1 and thus F0 . Hence,

this is an index 3 system.
25

 R = 0:
0=

f F1 F2
.
h2 A2

This is an index 2 system.

F0
h1

A1

F1

h2
R1

A2

F2
R2

F
h3
A3

...

FN-1
R

N-1

hN

FN

AN

Figure 11: N Communicating tanks

The situation may even become worse if more tanks are coupled. The dynamic
behavior is governed by the system of equations
A1 h 1

A2 h 2

h2 h1
,
R1
h2 h1
h3 h2
F1 F2 =
+
,
R1
R2
F0 F1 = F0 +

..
.
Ai h i

Fi1 Fi =

hi+1 hi
hi hi1
+
,
Ri1
Ri

..
.
AN h N

FN 1 FN =

hN hN 1
f (hN )
| {z }
RN 1
:=chN

together with a specification: Determine F0 such that f (hn ) = const. This

leads to a system of Hessenberg form of size N + 1 and thus the index is N + 1.

0=

A1
A2
..
0

.
AN
0

...
26

h1
h 2

..
.

h N
0
F 0

R11
1
R1

0
0
0

1
R1

R11
..
.
...
...

0
1
R2

...

1
R2

..

..
0

1
RN 1 c 0
c
0
0

.
RN11
0

h1
h2
..
.
hN
F0

Tanks are only a special case of holdup devices with the output being an integral
of the input. If the input to the first holdup is to satisfy a specification for the
last output, the index is always one more than the number of storages in series.
Some unit models such as distillation columns and flowsheets in general may
contain numerous coupled holdups. Consequently, the index in a design task
may be arbitrarily high.

5.2

Integration algorithms for DAE systems

A wide range of general purpose integration algorithms for the direct solution of
DAEs with index 1 and 2 have been published in recent years. Implementations
of the different approaches are:
Backward differential formula methods DASSL[3] and DAESOL[11],
Runge-Kutta methods RADAU5 [16],
Extrapolation methods LIMEX [9] and MEXX [23].
Engineering approaches to solving higher index DAEs are found in [8, 19, 27].
The method of choice in OPTISIM is an implementation of a variable stepsize,
variable order BDF method exploiting specific structures of the model equations
available in the OPTISIM model library.

5.3

Treatment of discontinuities in the dynamic case

Compared to the previously discussed sources of discontinuities in steady state

simulation, dynamic simulation shows even more cases for discontinuities: interactive user input, discrete measurements or controls, model switching due to
different operation conditions and delays in the model equations. The latter
depend not only on actual values of xT = (y, z) but also on past values. Delays
are used to model transport deadtimes in pipelines.
Discontinuities in dynamic simulation are also treated efficiently and reliably
using switching functions. Rewriting (5.1) with switching function results in

A1 A2
y
f (y, z, p, u, sign q)
=
(5.3a)
0
0
z
g(y, z, p, u, sign q)
q

27

q(y, z, p, u)

(5.3b)

with Ai = Ai (y, z, p, u, sign q), i = 1, 2. A point where a discontinuity occurs

is described as a zero of a switching function (5.3b).
The main steps of a discontinuity algorithm in dynamic simulation are:
1. After every integration step the occurrence of a discontinuity is checked
by testing the sign of the switching function.
2. If no discontinuity has occurred, the integration is continued. Otherwise
the switching point is determined iteratively.
3. The right hand side of the equations are changed according to the new set
of signs of the switching functions.
4. The discretization is adapted. In general, this involves a restart of the
integration. For discontinuities in higher derivatives it might often be
sufficient to reduce the order of the integration scheme.
For details of a switching algorithm, see e.g. [11].
For the iterative determination of the switching point, values of the switching
function and thus of the values of the state variables x are necessary not only
at the discretization points but also at intermediate points. They are computed
using a continuous error-controlled representation P x (t) of the solution in the
last integration step which can be obtained from the integration scheme (here:
the BDF-polynomial) [11]. In the case of an ODE, the switching point t is given
as solution of3 q(P x (t)) = 0. However, in the DAE case the situation is more
complicated: a root t of the switching function
q(P y (t), P z (t)) = 0
does in general not imply that the algebraic equations are fulfilled at t. Thus,
the values are not consistent initial values for the next interval. It has been
proposed4 to obtain t and the algebraic variable z from
g(P y (t), z) =
q(P y (t), z) =

0
0,

which leads to consistent values. However, this method is restricted to semiexplicit DAE systems with index 1. For DAEs with index 2 we suggest to
treat the integration step size (or the right point of the integration interval t) as
a variable. Assume for example an implicit Euler discretization for the system
(5.3) (with A1 = I, A2 = 0 for simplicity), the system
y

0 =
0 =

yn + (t tn )f (y, z)
g(y, z)
q(y, z)

3 Parameters

p and controls u(t) are dropped for ease of notation.

Thomsen, talk at the conference Differential-algebraic equations: Theory and applications in technical simulation, Oberwolfach June 13-19, 1993.
4 P.G.

28

can be solved for (

y , z, t). This is an (ny + nz + 1) (ny + nz + 1) system
in contrast to the one dimensional root finding which has to be carried out in
the ODE case using interpolants. In order to limit computational costs a good
initial guess for y, z, t is obtained using the interpolant.

5.4

Consistent initialization

Standard integration methods with order and stepsize control require consistent
initial values (y0 , z0 ), (y 0 , z0 ), i.e. the initial values must satisfy not only the
original DAE (5.1) but also the extended system (5.2). If initial values are
inconsistent, e.g. at the very beginning of a new simulation experiment or after
a discontinuity has occurred, the solution contains an error portion that cannot
be controlled by conventional error control strategies. For the initialization of
DAEs general approaches have been published [21, 22, 29].
To avoid inconsistent values in the solution after a discontinuity, smoothed
functions, e.g. a ramp or a half cosine wave, are often used instead of step
changes in parameters or input variables. However, in the DAE context nondifferentiable solutions may occur even though the input function is continuously
differentiable. The reinitialization of DAEs after discontinuities has been studied
in [4, 25].

References
[1] R. Aris. Elementary chemical reactor analysis. Butterworth Publishers,
Boston, 1989.
[2] L. T. Biegler. Chemical process simulation. Chemical Engineering Progress,
85/10, 1989.
[3] K.E. Brenan, S.L. Campbell, and L.R. Petzold. The numerical solution
of initial value problems in ordinary differential-algebraic equations. North
Holland Publishing Co., 1989.
[4] L. Br
ull and U. Pallaske. On consistent initialization of differentialalgebraic equations with discontinuities. In Proceedings of the Fourth European Conference on Mathematics in Industry, pages 213217, Stuttgart,
1991. B.G. Teubner.
[5] P. Bujakiewicz and P.P.J. van den Bosch. Determination of perturbation
index of a DAE with maximum weighted matching algorithm. In Proc. of
IEEE/IFAC Symposium CACSD 94, Tucson, AZ, 1994.
[6] P. Burr. The design of optimal air separation and liquefraction processes
with the OPTISIM equation oriented simulator and its application to online and off-line plant optimization. In Proc. AIChE Spring National Meeting, April 1991, Houston, TX, 1991.
29

[7] S.L. Campbell and C.W. Gear. The index of general nonlinear DAEs.
Numer. Math., 72:173196, 1995.
[8] Y. Chung and A. W. Westerberg. Solving a class of near index problems
as perturbations of index problems. Ind. Eng. Chem. Res., 31:25932603,
1992.
[9] P. Deuflhard, E. Hairer, and J. Zugck. One-step and extrapolation methods
for differential-algebraic systems. Numer. Math., 51:501516, 1987.
[10] I. S. Duff, A. M. Erisman, and J. K. Reid. Direct methods for sparse
matrices. Clarendon Press, Oxford, 1986.
[11] E. Eich. Projizierende Mehrschrittverfahren zur numerischen L
osung der
Bewegungsgleichungen technischer Mehrk
orpersysteme mit Zwangsbedingungen und Unstetigkeiten. PhD thesis, Institut f
ur Mathematik, Universitat Augsburg, 1991. VDI-Fortschrittsberichte, Reihe 18, Nr. 109,
VDI-Verlag, D
usseldorf, 1992.
[12] R. Fletcher. Practical methods of optimization. John Wiley & Sons, 2nd
edition, 1987.
[13] C. W. Gear. Differential-algebraic equation index transformation. SIAM
J. Sci. Stat. Comp., 7:3947, 1988.
[14] P.E. Gill, W. Murray, and M.H. Wright. Practical optimization. Addison
Wesley, 1981.
[15] A. Griewank and G.F. Corliss, editors. Automatic differentiation of algorithms: Theory, implementation and application, Philadelphia, 1991.
SIAM.
[16] E. Hairer and G. Wanner.
Springer, 1991.

Solving ordinary differential equations II.

[17] M. Hillestad and T. Hertzberg. Dynamic simulation of chemical engineering systems by the sequential modular approach. Comput. chem. Engng.,
10:377388, 1986.
[18] C.D. Holland and A.I. Liapis. Computer methods for solving dynamic seperation problems. McGraw-Hill, New York, 1983.
[19] R.B. Jarvis and C.C. Pantelides. A differentiation-free algorithm for solving
high-index DAE systems. In Proc. AIChE 1992 Annual Meeting, Miami
Beach, 1.-6. November 1992 1992. American Institute of Chemical Engineers. Paper 146g.

30

[20] A. Kroner, P. Holl, W. Marquardt, and E.-D. Gilles. DIVA An open

architecture for dynamic simulation. Comput. chem. Engng., 14:12891295,
1990.
[21] A. Kroner, W. Marquardt, and E.D. Gilles. Computing consistent initial
conditions for differential-algebraic equations. Proc. ESCAPE-1, May 1992,
Elsinore, Denmark. Comput. chem. Engng., 16, Suppl.:131138, 1992.
[22] B. Leimkuhler, L. R. Petzold, and C. W. Gear. Approximation methods
for the consistent initialization of differential-algebraic equations. SIAM J.
Numer. Anal, 28:205226, 1991.
[23] Ch. Lubich. h2 extrapolation methods for differential-algebraic systems of
index 2. Impact Comp. Sci. Eng., 1:260268, 1990.
[24] D.P. Mahoney and O. Santollani. HYSYS an integrated system for process
engineering and control. Hyprotech, Ltd., Calgary, Alberta, 1994.
[25] C. Majer, W. Marquardt, and D. Gilles. Reinitialization of DAEs after discontinuities. Proc. ESCAPE-5, June 1995, Bled, Slovenia. Comput. chem.
Engng., 19, Suppl.:507512, 1995.
[26] W. Marquardt. Dynamic process simulation - Recent progress and further
challenges. In Proc. Fourth International Conference on Chemical Process
Control, pages 131180, Padre Island, Texas, 1991.
[27] S. E. Mattsson and G. Soderlind. Index reduction in differential-algebraic
equations using dummy derivatives. SIAM J. Sci. Stat. Comput., 14:677
692, 1993.
[28] K. Miller and R.N. Miller. Moving finite elements, Part I. SIAM J. Numer.
Anal., 18:10191032, 1981.
[29] C. C. Pantelides. The consistent initialisation of differential-algebraic systems. SIAM J. Sci. Stat. Comput., 9:213231, 1988.
[30] C. C. Pantelides. SpeedUp - Recent advances in process simulation. Comput. chem. Engng., 12:745755, 1988.
[31] C.C Pantelides, D. Gritsis, K.R. Morison, and R.W.H. Sargent. The mathematical modelling of transient systems using differential-algebraic equations. Comput. chem. Engng., 12:449454, 1988.
[32] J.D. Perkins. Equation-oriented flowsheeting. In A.W. Westerberg and
H.H. Chien (eds.): Foundations of computer aided chemical process design.
CACHE publications, Austin, pp. 309-367, 1983.
[33] W.T H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery.
Numerical recipes. Cambridge University Press, 1992.
31

[34] Y. Saad and M. H. Schultz. GMRES: a generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM J. Sci. Stat. Comput.,
7:856869, 1986.
[35] W. Schmid, M.Paffrath, and R.H.W. Hoppe. Application of iterative methods for solving nonsymmetric linear systems in the simulation of semiconductor processing. Surv. Math. Ind., 5:126, 1995.
[36] N. L. Schryer. Designing software for one-dimensional partial differential
equations. ACM Trans. Math. Software, 16:7285, 1990.
[37] J. M. Smith. Chemical engineering kinetics. McGraw-Hill, Singapore, third
edition, 1981.
[38] J. Stoer and R. Bulirsch. Introduction to numerical analysis. Springer, New
York, 2nd edition, 1993.
[39] J. Unger, A. Kroner, and W. Marquardt. Structural analysis of differentialalgebraic equation systems Theory and applications. Comput. chem.
Engng., 19:867882, 1995.
[40] R. Wait and A. R. Mitchell. Finite element analysis and applications. John
Wiley & Sons, Chichester, 1985.
[41] S.M. Walas. Phase equilibria in chemical engineering. Butterworth Publishers, Boston, 1985.
[42] H. F. Walker. Implementation of the GMRES method using Householder
transformations. SIAM J. Sci. Stat. Comput., 9:152163, 1988.
[43] S. E. Zitney and M. A. Stadtherr. Computational experiments in equationbased chemical process flowsheeting. Comput. chem. Engng., 12, 1988.
Authors addresses:
Prof. Dr. Edda Eich-Soellner, Fachbereich Informatik/Mathematik, Fachhochschule M
unchen, Lothstr. 34, D80323 M
unchen, Germany.
Prof. Dr. Peter Lory, Institut f
ur Wirtschaftsinformatik, Universitat Regensburg, D93040 Regensburg, Germany.
Dr. Peter Burr and Dipl.-Ing. Andreas Kroner, Linde AG, Process Engineering and Contracting Division, Dr.-Carl-von-Linde-Str. 6-14, D82049 Hollriegelskreuth, Germany

A
A.1

Numerical data for the two phase problem

The parameter set

The constant parameters and input variables of the two phase process are:

32

AI =101
AII =101
AM =101

=102
I
cA0 =0.2
cp
=83.2
E
=12000
H,R =1.93 105
F I =2 104
F II =1 104
k0 =1.8 105
kI
=10
k II =10
k M =10.0
R
=8.314

=51.0
T0I =293
T0II =293
T C =293
V I =102
V II =102

cooling area
cooling area
membrane area
mass transfer coefficient
feed concentration of A
molar heat capacitiy
activation energy
heat of reaction
flowrate through phase I
flowrate through phase II
frequency factor
heat transfer coefficient
heat transfer coefficient
heat transfer coefficient
Gas law constant
molar density
feed temperature
feed temperature
cooling water temperature
volume
volume

A.2

m2
m2
m2
m
s
kmol
m3
kJ
kg k
kJ
kmol
kJ
kmol
m3
s
m3
s
1
s
kW
m3 K
kW
m3 K
kW
m3 K
kJ
kmol K
kg
m3

K
K
K
m3
m3

The steady state solution

To compute a steady state solution of this process equations (3.1d),(3.3)-(3.5)

together with the steady state versions of equations (3.1a),(3.1b),(3.1c), (3.2)
have to be solved. The following solution is obtained:
cIA =2.9059 106
cIB =0.1375
cII
B =0.125
T I =296.17
T II =294.31
M
Jm,B
=1.25 105
I
Q =3.168
QII =1.308
M
JQ
=1.8607
JHM =1.93 103
r0 =4 103

33

kmol
m3
kmol
m3
kmol
m3

K
K

kmol
s

kW
kW
kW
kW

kmol
m3 s .

A.3

The stationary optimization solution

The two phase process is optimized w.r.t. costs. Optimization parameters are
the cooling temperature T C , exchange areas AM , AI , AII , volumes V I , V II , and
the input concentration cIA0 (i.e. p = (AM , V I , V II , AI , AII , cIA0 )). The difference
between investment and feed costs, resp., and proceeds is minimized:
(x, p)
with constraints
T II
T II
105
105
105
105

100(V I + V II ) + AI + AII + AM + 100cIA0 1000cII

B

300,
300,
V I 101 ,
V II 101 ,
AI 10,
AII 10 .

The solution of this optimization problem is AM = 101 m2 , V I = 4 104 m3 ,

V II = 105 m3 , AI = 1.54 m2 , AII = 105 m2 , cIA0 = 0.3 kmol
m3 .

34