DYNAMIC MODELLING AND SIMULATION

INTRODUCTION

INDUSTRIAL EXPERIENCE WITH DYNAMIC

SIMULATION

Paul I. Barton
Department of Chemical Engineering
Massachusetts Institute of Technology
Cambridge, MA

copyright Paul I. Barton, July 1997

 Process modeling is the single
technology that has had the biggest impact
on our business in the last decade

Frank Popoff
former CEO, Dow Chemical
April 1996
 APPLICATIONS OF DYNAMIC SIMULATION

Dynamic simulation is useful throughout the entire

lifetime of a plant: from conception to decommissioning.

Plant design:
a) a priori assessment of intrinsic operability and
controllability of a plant - especially for highly
integrated plants.
b) design and testing of regulatory control systems -
selection of control structures, control algorithms,
and initial tuning of loops.
c) design and testing of operating procedures - start-
up, shut-down, feed stock changeover, etc.
d) hazard/safety studies - answer important questions
raised by HAZOP study.
e) design and testing of protective and relief devices -
study plant behaviour and performance of protective
system during major deviations from steady-state
(need good models!)
f) environmental studies - predict emissions generated
during plant upsets/failures.
g) analysis of intrinsically dynamic processes:
batch/semi-continuous processes, periodic processes
(e.g., pressure swing adsorption)
 APPLICATIONS OF DYNAMIC SIMULATION

Plant operation: typically a dynamic simulation is linked

to the plant's real time control and monitoring software:
(a) computer based operator training:

Plant Operator

Real Time
Instructor
Control System

Dynamic Simulation

dynamic simulation runs in real time and mimics

behaviour of real plant.
operator interfaces to control system (not
simulator) more realistic.
instructor monitors and creates scenarios (e.g.,
disturbances, failures)
(b) validation of operating and safety procedures test
control software by dynamic simulation before
implementing on the real plant!
 APPLICATIONS OF DYNAMIC SIMULATION

(c) on-line operator decision support tool (simulation

runs in parallel with real plant):

Plant Operators

Real Time Control System

Dynamic
Plant Simulation

update current state (initial condition) directly from

measurements.
dynamic predictor run simulation faster than real
time to predict hazards or problems (so that early
corrective or preventative action can be taken).

experimental tool predict consequences of proposed

actions.
 CONTROLLABILITY STUDY OF BAYER
DISTILLATION PROCESS
(Zitney et al., 1995)

BACKGROUND

system of eight coupled and heat integrated distillation

columns (nearly 1200 trays)
separates eight crude silanes from a mixture of 40
components
operation characterized by large holdup and time
varying feed composition with no virtually no buffering
between columns and upstream process

OBJECTIVES

Study performance of control system in order to:

increase plant capacity

and at same time improve quality and reduce energy
consumption

APPROACH

develop dynamic model of the overall system

use dynamic model for identification and controllability
experiments - avoid disrupting production
 THE DYNAMIC MODEL

detailed SPEEDUP (Aspen Technology, 1993) model

developed of eight coupled columns

one year effort for 2-3 engineers: process modeled in

stages, so results available throughout this period

tray/reboiler/condenser mass and energy balances,

phase equilibrium modeled by Wilson equation

differential-algebraic system of 75,000 coupled

equations - largest known industrial SPEEDUP
application at the time

simulations performed on CRAY C90 supercomputer -

solution speed increased upto 50-fold by exploiting
vector parallelization of supercomputer

simulation runs approx. 10 times faster than real time

 APPLICATIONS OF THE MODEL

steady-state simulation:
steady-state is just one initial condition for a
dynamic model: always get steady-state model for
free in dynamic study!
optimize plant for different feed concentrations and
loads
identify which equipment must be revamped to
increase capacity

plant must be regulated at new optimal operating

points - use dynamic model for identification and
controllability studies:
sensitivity studies to identify necessary process
measurements
sensitivity studies to define appropriate control
structure (matching of sensors to control elements)
and tuning of control parameters
validation of control system by predicting plant
response to a variety of load and setpoint changes
 MEASUREMENT SELECTION

study sensitivity of process outputs to range of common

disturbances (see figure)
with dynamic model can observe all the plant states

many existing sensors in inappropriate positions

why use dynamic simulation? - necessary information

in steady-state gain matrix
much easier to get steady-state gain matrix from
SPEEDUP (compared to ASPEN PLUS)
think before applying dynamic simulation: effort
must be justified
SENSITIVITY ANALYSIS FOR HEAT INTEGRATED
COLUMN

from Zitney et al., Plantwide Dynamic Simulation on Supercomputers: Modeling a

Bayer Distillation Process, AIChE Symposium Series, 91(304), pp. 313-316, 1995.
Reproduced with permission. Copyright
American Institute of Chemical Engineers
and Computers Aids for Chemical Engineering Education. All rights reserved.
 CONTROLLER DESIGN AND VALIDATION

Given new measurement, completely new control system

was designed.

Use dynamic simulations (rather than experiments with

real plant) for:

identification experiments to determine transfer

function models for controller design

controller design: major goal was disturbance rejection:

match outputs and controls to minimize interaction
tune PID constants using dynamic model

controller validation: test controllers in wide variety of

scenarios before implementation in plant
 CONTROLLABILITY STUDY OF BAYER
DISTILLATION PROCESS

BENEFITS

doubling capacity only required revamping half the

plant

one third of the sensors were no longer necessary, and

outputs more sensitive to disturbances identified

decision made to design completely new control system

controller testing by simulation saved much

troubleshooting in plant, and faster and more successful
startup of new control system

certain time constants so large that operator never sees

consequences of manual operations during their shift
(see figure):
operators instructed to leave process in automatic
operation as much as possible

the act/discipline of developing a dynamic model can

lead to significant new knowledge and insights
concerning the process
EFFECTS OF A DISTURBANCE MAY BE FELT OVER
SEVERAL SHIFTS
(VERY LARGE TIME CONSTANTS)

from Zitney et al., Plantwide Dynamic Simulation on Supercomputers: Modeling a

Bayer Distillation Process, AIChE Symposium Series, 91(304), pp. 313-316, 1995.
Reproduced with permission. Copyright
American Institute of Chemical Engineers
and Computers Aids for Chemical Engineering Education. All rights reserved.
 RETROFIT OF NORSK HYDRO TOPSIDE OFFSHORE
OIL PRODUCTION PROCESS
(Naess et al., 1993)1

BACKGROUND

introduction of flow from Tarbert pipeline may expose

the Osberg topside process to slug regimes

necessary to study process response to slugs without

disrupting production
more sophisticated approach than lets just try it
and see what happens!

is insertion of slugcatcher in front of separation train

necessary?

APPROACH

use simulations with dynamic model to avoid

disrupting production

predict performance of slugcatcher designs

1
this paper is instructive because it applies dynamic simulation to four different processes at four
different stages of their lifetimes. It also compares two dynamic simulators.
 THE DYNAMIC MODEL

dynamic model absolutely necessary as topside process

cannot cope during the transients
the steady-states may be fine, but the path between
them may lead to a plant trip

SPEEDUP (Aspen Technology, 1993) used

rigorous equilibrium models of two- and three-phase

separators with seven components

models of PI-controllers with real life tuning parameters

models identified for control valves including time lag

dynamics of piping neglected

interfaced to in-house physical properties package via

procedures
same physical properties as steady-state simulation
model validated against steady-state model used by
engineering division
 CASE STUDY

use model to predict system response to seven typical

slugs:
production start
maximum gas production
maximum water production

worries about pressure trips unwarranted

gas valve on second stage separator has too little

capacity due to back pressure from the gas train

in worst case (maximum water production) fourth stage

separator will trip due to high-high level after 30
minutes
TYPICAL SLUG THROUGH TARBERT PIPELINE TO
OSBERG

Reprinted from Computers Chem. Engng. 17, Naess et al., "Using Dynamic Process
Simulation from Conception to Normal Operation of Process Plants," pp. 585-600,
Copyright 1993, with kind permission from Elsevier Science Ltd, The Boulevard,
Langford Lane, Kidlington OX5 1GB, UK.
PRESSURE TRANSIENT IN FIRST SEPARATOR IN
RESPONSE TO SLUG

Reprinted from Computers Chem. Engng. 17, Naess et al., "Using Dynamic Process
Simulation from Conception to Normal Operation of Process Plants," pp. 585-600,
Copyright 1993, with kind permission from Elsevier Science Ltd, The Boulevard,
Langford Lane, Kidlington OX5 1GB, UK.
 VALVE STEM POSITION TRANSIENTS IN
RESPONSE TO SLUG

Reprinted from Computers Chem. Engng. 17, Naess et al., "Using Dynamic Process
Simulation from Conception to Normal Operation of Process Plants," pp. 585-600,
Copyright 1993, with kind permission from Elsevier Science Ltd, The Boulevard,
Langford Lane, Kidlington OX5 1GB, UK.
RETROFIT OF NORSK HYDRO TOPSIDE OFFSHORE
OIL PRODUCTION PROCESS

BENEFITS

problems lay elsewhere to what was anticipated: real

value added from simulation!

produce less from Osberg wells during startup of

Tarbert pipeline to avoid level trip

slugcatcher not required due to slow dynamics of slug

DESIGN OF PRODUCT GRADE TRANSITIONS FOR
POLYOLEFIN PROCESSES
(Debling et al., 1994)

BACKGROUND

wide variety of polyolefin products:

polypropylene
low and high density polyethylene
ethylene-propylene rubber

number of catalytic processes

slurry
solution
gas-phase
...each has particular advantages

but, each will be required to produce range of product

grades to meet demand (typical product campaign
varies from less than one day to several weeks)

OBJECTIVE

Design changeover policies that:

conduct grade transitions as swiftly as possible

minimize amount of off-spec product
...market conditions may favour one over the other.
 DESIGN OF PRODUCT GRADE TRANSITIONS FOR
POLYOLEFIN PROCESSES

APPROACH

design changeover policy to move process between two

steady-states: a path between these points in the time
domain

use dynamic simulation to predict response of the

processes to a particular changeover policy, and
experiment with improved changeover policies

Alternative: experiment with a real plant or pilot plant

ADVANTAGES (OF SIMULATION)

safety: path taken by process must satisfy constraints to

avoid, for example, temperature or pressure runaway:
optimal changeover policy subject to path constraints
constraints can be violated by a simulation without
dire consequences: experiment with policies close
to runaway

environment: off-spec or vented material not created by

a simulation!

economic: process not taken out of production for

experiments, can explore more alternatives in hope
finding a better policy, potentially faster (opportunity
cost), etc.
 DESIGN OF PRODUCT GRADE TRANSITIONS FOR
POLYOLEFIN PROCESSES

CASE STUDY 1

grade transitions in polypropylene homopolymer (PP)

and impact copolymer (IPP) studied

processes studied:

gas-phase fluidized bed (Union Carbide/Shell)

gas-phase vertical stirred bed (BASF/ICI)
gas-phase horizontal stirred bed (AMOCO/Chisso)
liquid pool loop (Himont)

grade transition: increase number average molecular

weight from 55,000 to 65,000 (stiffer product)

use POLYRED - simulator specifically designed for

polymer processes developed at the University of
Wisconsin

1 the original paper (Delbing et al., 1994) contains several case studies of different products and processes
 DESIGN OF PRODUCT GRADE TRANSITIONS FOR
POLYOLEFIN PROCESSES

ALTERNATIVE CHANGEOVER POLICIES

In both cases, try to reduce accumulated hydrogen in the

reactor as quickly as possible.

strategies for gas-phase reactors

A: i) step change in hydrogen feed rate

B: i) step change in hydrogen feed rate

ii) open gas vent for first ten minutes to change
gas phase composition quickly

...mitigate large hydrogen residence time in gas phase

reactors

strategies for loop reactor

A: i) step change in hydrogen feed rate

H: i) step change in hydrogen feed rate

ii) reduce recycle rate in loop to narrow residence
time distribution

...narrow residence time distribution for hydrogen.

 GAS-PHASE FLUIDIZED BED PROCESS

from Debling et al., Dynamic Modeling of Product Grade Transitions for Olefin
Polymerization Processes, AIChE Journal, 40(3), pp.506-520, 1994. Reproduced with
permission of the American Institution of Chemical Engineers. Copyright
1994
AIChE. All rights reserved.
 GAS-PHASE VERTICAL STIRRED BED

from Debling et al., Dynamic Modeling of Product Grade Transitions for Olefin
Polymerization Processes, AIChE Journal, 40(3), pp.506-520, 1994. Reproduced with
permission of the American Institution of Chemical Engineers. Copyright
1994
AIChE. All rights reserved.
 GAS-PHASE HORIZONTAL STIRRED BED

from Debling et al., Dynamic Modeling of Product Grade Transitions for Olefin
Polymerization Processes, AIChE Journal, 40(3), pp.506-520, 1994. Reproduced with
permission of the American Institution of Chemical Engineers. Copyright
1994
AIChE. All rights reserved.
 LIQUID POOL LOOP

from Debling et al., Dynamic Modeling of Product Grade Transitions for Olefin
Polymerization Processes, AIChE Journal, 40(3), pp.506-520, 1994. Reproduced with
permission of the American Institution of Chemical Engineers. Copyright
1994
AIChE. All rights reserved.
 NUMBER AVERAGE MOLECULAR WEIGHT
DURING TRANSITION
BASE CASE POLICY (A)

from Debling et al., Dynamic Modeling of Product Grade Transitions for Olefin
Polymerization Processes, AIChE Journal, 40(3), pp.506-520, 1994. Reproduced with
permission of the American Institution of Chemical Engineers. Copyright
1994
AIChE. All rights reserved.
 NUMBER AVERAGE MOLECULAR WEIGHT
DURING TRANSITION
ALTERNATIVE POLICY (B or H)

from Debling et al., Dynamic Modeling of Product Grade Transitions for Olefin
Polymerization Processes, AIChE Journal, 40(3), pp.506-520, 1994. Reproduced with
permission of the American Institution of Chemical Engineers. Copyright
1994
AIChE. All rights reserved.
PERCENT OFF-SPEC PRODUCT DURING TYPICAL
CAMPAIGN

from Debling et al., Dynamic Modeling of Product Grade Transitions for Olefin
Polymerization Processes, AIChE Journal, 40(3), pp.506-520, 1994. Reproduced with
permission of the American Institution of Chemical Engineers. Copyright
1994
AIChE. All rights reserved.
DESIGN OF PRODUCT GRADE TRANSITIONS FOR
POLYOLEFIN PROCESSES

CONCLUSIONS

for gas phase reactors, blow down to reduce hydrogen

accumulation can greatly reduce changeover time
(holdup/accumulation of hydrogen dominates speed of
changeover)

careful design of operating policies can change relative

rankings of processes

gas phase reactors now exhibit natural residence

times

these conclusions reached in the comfort and safety of

the office!
 BATCH PROCESS DEVELOPMENT: SILOXANE
MONOMER PROCESS
(Allgor et al., 1996; Ahmad and Barton, 1994)

BACKGROUND

effective process development critical to success of a

new specialty chemical or synthetic pharmaceutical.
develop efficient process rapidly
environmental problems: use and disposal of
organic solvents

can detailed modeling technology help?

batch processes
dynamic simulation

siloxane monomer process (Allgor et al., 1996):

products A and D required in arbitrary ratio: A
formed in first reaction, byproduct C further
reacted to D.
allyl alcohol (R1 in figure), toluene, methanol and
water used as solvents and reagents
complex reaction kinetics, highly nonideal mixtures

OBJECTIVE
Can detailed dynamic models of a batch process help in
the design of a process with better environmental
performance?2

2 for more lengthy discussion see Ahmad and Barton (1994)

 SILOXANE MONOMER PROCESS FLOWSHEET
R1, R2, Toluene H2O
H2 R1, Methanol, Toluene, E
Methanol
E, A, Toluene
H2O, Methanol, Toluene, E

Pt D, A

Reactor I Reactor II Rectifier I Rectifier II Reactor III Rectifier III

R1 + R2 I1 (Pt catalyzed) C + MeOH E
2 E + H2O D + 2 MeOH
R1 + I1 A
I1 C + H2
C + I1 I2
Pt Pt* Pt, I2
 BASE CASE DESIGN: PROBLEMS

process developed in pilot plant simulated in ABACUSS

to quantify all process streams (on a batch size basis)

organic waste stream: overhead from Rectifier I.

Analyze solvent recovery possible using theory of
simple distillation residue curve maps:

Conclusions:
only toluene can be recovered in pure form for recycle
30% of stream disposed of by incineration.

aqueous waste stream: solvent recovery complicated by

presence of toluene (heterogeneous mixture, azeotropes)

dont let toluene get downstream from Reactor I?

 PROCESS ALTERNATIVE: INSERT RECTIFIER FOLLOWING REACTOR I

H2

R1, R2, Toluene Methanol

H2O
Cut 1: C C, Methanol
Cut 2: R1, I1, Toluene, C
H2O, Methanol
Cut 3: A

Pt E D, A

Reactor I Reactor II Rectifier II Reactor III

Rectifier I Rectifier III
R1 + R2 I1 (Pt catalyzed) C + MeOH E
2 E + H2O D + 2 MeOH
R1 + I1 A
I1 C + H2
C + I1 I2
Pt Pt* Pt, I2
 IS PROCESS ALTERNATIVE FEASIBLE?

analyze separation alternatives for mixture leaving

Reactor I: is recovery of pure C feasible?

Cut 1: C
RI: Cut 2: C-R1-Toluene
Cut 3: C-R1

Cut 2: C
RV: Cut 2: C-R1-Toluene
Cut 3: Toluene

Conclusions:
if mixture created by Reaction I lies in regions I or V
then process is feasible
some C will also be recycled in azeotropes with
solvent and/or reagent to next batch

recycle of C will lead to build up of C in the reactor to a

cyclic steady-state3 concentration:
R1 + R2  I1 (Pt catalyzed)
R1 + I1  A
I1  C + H 2
C + I1
I 2
Pt  Pt *

...encourages formation of undesired oligmer I2.

3 a cyclic dynamic system is said to have reached cyclic steady-state when the variation of the variables
(profiles) over a cycle is the same from one cycle to the next
 DESIGN OPERATING POLICY FOR COUPLED
REACTOR AND COLUMN

DESIGN OBJECTIVES

At cyclic steady-state:
mixture at the end of Reaction I must be in either batch
distillation region I or V
formation of undesired oligmer I2 must be minimized

KEY DESIGN VARIABLES

charge and temperature policy for reactor (amount of

solvent and reagent at end, control conc. of I1)
reflux policy for column (recycled cuts do not need to be
sharp)
 ROLE OF DYNAMIC SIMULATION

rigorous dynamic model of coupled reactor and column

can:
predict cyclic steady-state for given operating
policy
allow experimentation with operating policy to
meet design objectives
cheaper, safer, quicker, better (more alternatives
explored) than pilot plant

complex dynamic model of reactor and column must

reflect:
nonideal thermodynamics: azeotropic mixtures
complex reaction kinetics
sequential operating policies for reactor and
column, and intermediate storage and recycle of
solvent/reagent
physico-chemical discontinuities: reaction mass
boiling, intensive quantities not defined in empty
vessel, etc.

need state-of-the-art combined discrete/continuous

dynamic simulator: ABACUSS
2657 equations (including physical properties)
design of suitable operating policy achieved
MOLE HOLDUPS IN REACTOR I: APPROACH TO
CYCLIC STEADY-STATE
 OPTIMAL CYCLIC STEADY-STATE
DYNAMIC OPTIMIZATION

time

time

Product

Waste
 BATCH PROCESS DEVELOPMENT: SILOXANE
MONOMER PROCESS

BENEFITS

total organic waste reduced by 88% compared to the

base case (with central solvent recovery facility) - only
waste as unavoidable consequence of stoichiometry left
in process recycling of solvent more effective!

reduced load on downstream unit operations: only C

passed to rest of process, selectivity optimized with
respect to A

raw material saved (solvent/reagent)

aqueous waste easier to treat

throughput in same equipment similar

targeted use of dynamic simulation in conjunction with

pilot trials and simpler models (e.g. residue curve maps)
leads to effective problem solving strategy
 SECTION I

DYNAMIC MODELLING

Paul I. Barton
Department of Chemical Engineering
Massachusetts Institute of Technology
Cambridge, MA

copyright Paul I. Barton, July 1997

MODELLING DYNAMIC BEHAVIOUR OF PROCESS
SYSTEMS

Three key ideas:

(a) model system dynamics

System
complex artifact composed of connected and

interacting components.

For example, an entire chemical process:

...or, a complex unit operation - e.g., a batch distillation

column:
 MODELLING DYNAMIC BEHAVIOUR OF
PROCESS SYSTEMS

(b) focus on deterministic first principle based

(mathematical) models
emphasis on lumped parameter systems
model will always rely to a certain extent on
empirically derived parameters (e.g. reaction
kinetic parameters) and correlations (e.g.
distillation tray hydrodynamics).
(c) goal is to predict the evolution of the system
behaviour with respect to time. For example,
dynamic simulation results are typically presented as
time trajectories:
Level, h [m]

Time, t [s]

Note that there is more to a transient than the steady-

states.....
 MODELLING CAVEATS

Any mathematical model (by definition) is an abstraction

of the true system behaviour.1

 does the model exactly mimic the true behaviour?

does the model predict the aspects of system

behaviour (phenomena) of interest with sufficient
accuracy for the current application?

This implies:

a model is only valid within the context and

assumptions it was developed.

extrapolation of the model beyond this context and

assumptions is extremely dangerous - critical
reappraisal of the model required (hence need to
document the model)

verify a model against the real system behaviour

whenever possible.

there are many models for a single system each

represents a different level of abstraction. Key task is
to select the appropriate level of abstraction.

1 in fact, any "perfect model" would exhibit such complexity that it would be
indistinguishable from the real system.
 PRELUDE TO MODELLING

Because of this fundamental nature of models, before and

during any modelling activity it is important to clarify and
document the following information:

(a) identify the system for which a model is required

THE ENVIRONMENT

SYSTEM
Inputs(t) Outputs(t)
State(t)

disturbances
controls

Identify:
boundaries (function of system) - a system is defined
by its boundary
constraints
quantities describing system behaviour: inputs, states,
outputs
assume: inputs
(outputs)
 PRELUDE TO MODELLING

Notes:
we are developing a model for the system. Everything
else is the environment.
inputs define the influence of the environment on the
system.
while in real life the inputs will be further subdivided
into:
(a) controls those inputs that can be manipulated in
order to control the system behaviour.
(b) disturbances those inputs over which we have
no control.
. . . from the point of view of modelling we can play
God: we must define the time variation of all the
inputs in order to pose a fully determined simulation
problem.
inputs
(outputs):
more precisely: we can sufficiently decouple the
influence of the outputs on the inputs (feedback via
the environment) for the purposes of the current
exercise.
pragmatic view: can only model so much at a given
level of detail.
 PRELUDE TO MODELLING

(b) what is the intended application of the model?

What questions will be asked about the system?
Begin to identify:
what phenomena are of interest?
what quantities describe the system behaviour?
how detailed should the model be?
what assumptions can be made?
(c) what data concerning the system is available?
or can be obtained . . . imposes constraints on the
phenomena that can be modelled and the accuracy of
the simulation results.
Typically relates to parameters and empirical
correlations:
what is available?
in what range of process conditions are the
predictions valid?
how much uncertainty is there in the predictions?
Examples: non-equilibrium distillation tray models,
reaction kinetics for novel synthesis.
 PRELUDE TO MODELLING

In conclusion: modelling is a fundamental engineering

activity. We must trade the need to obtain an answer
against the accuracy (or even validity) of this answer, and
the resources required to obtain it (time, manpower, skills,
computing resources, data, etc. . . .)

principle of optimum sloppiness make as many

=> simplifying assumptions as reasonable without
throwing out the baby with the bath water (Luyben,
1990)

...modelling is still very much an art (but research

continues).
 STEP 1 HIERARCHICAL DECOMPOSITION

No engineer can grasp simultaneously all the relevant

details of a complex system: e.g., how do we construct a
mathematical model composed of thousands of
simultaneous equations?

Two approaches to manage system complexity in model

building:

A: Top Down Approach

(a) identify a series of components from which the

system is composed, and the connections
between these components (e.g., a flowsheet, an
electrical circuit, organs in the human body, etc.).

(b) once identified, concentrate on modelling and

testing each individual component in isolation as
a separate task.

B: Bottom Up Approach

(a) identify a series of standard primitive components

required to describe a class of systems. Develop
and validate models for these primitive
components (archive in library).

(b) develop models of more complex systems by

connecting these components together to form
structures (systems).
 STEP 1 HIERARCHICAL DECOMPOSITION

In my experience, a hybrid of the top down and bottom up

approaches is used:
system structure is invariably unraveled in a top
down manner, but . . .
model libraries are usually available, and these are
kept in mind when selecting a decomposition.
Hierarchical Submodel Decomposition (Elmquist, 1978)
introduces the notion of recursion into these approaches
a component model can either be:

(a) primitive e.g., it is entirely described in terms

of equations and is not decomposed further.

(b) composite e.g., it is treated as a system itself,

and can be therefore described in terms of a set of
interconnected components.

. . . the introduction of recursion allows the decomposition

of composite models to continue into as many levels as is
necessary - hierarchies of arbitrary depth can evolve.
 HIERARCHICAL SUBMODEL DECOMPOSITION - BATCH COLUMN EXAMPLE

Water Out Tray

Water In
Water In Water Out
Tray
Dump
Batch Tray
Column

Accumulator

Tray Section
Charge Tray

Dump
Tray
Steam Out Steam In
Steam In
Reboiler
Steam Out
Charge

Vapour Liquid

Charge dNi VIN ,i +LIN ,i

Vessel NC =
hL =
xi hi dt - VOUT ,i - LOUT ,i
i =1 Liquid
Q Enthalpy
hi = hi (T ,P ,xi )
Jacket
VLE Vapour
Enthalpy
Steam In Steam Out
 STEP 2 MODELLING THE COMPONENTS

Each component model is itself a system (recursion):

(a) identify the boundaries.
the model may represent a physical artifact e.g.,
a plant section, a unit operation, a vessel, or a
section of a vessel.
the model may represent some phenomenological
abstraction e.g., the set of equations defining a
mixture enthalpy, or a set of reaction rate
expressions.
(b) identify the connections of the model to its
environment - this will enable us to connect the
model to others in a larger structure.
Connections will typically represent either:
(i) fluxes of extensive properties, such as:
a diffusive flux - e.g., heat transfer through a
vessel wall
a convective flux - e.g., a pipe connecting two
vessels (ignoring the holdup of material in the
pipe)
Note fluxes will have a direction implied.
(ii) information transfer - e.g., pressure and/or
voltage signals for control instruments, or
intensive properties (e.g., temperature, pressure)
to determine driving forces for fluxes
 STEP 2 MODELLING THE COMPONENTS

Structures are defined by establishing equivalence

(merging) between the connection of one component
model and the connection of another component
model (obviously, the two connections must be
compatible - e.g., must represent a convective flux)
(c) define the internal behaviour of the component
model - what describes the "state" of this system, and
how is it related to both the inputs and the outputs.
 INTERNAL BEHAVIOUR

Ultimately internal behaviour is represented by a set of:

Inputs
VARIABLES - e.g. States
Outputs
...and these variables are related by a set of:
Mass Balances
Energy Balances
EQUATIONS - e.g. Physical Constraints
Thermodynamic Models
etc.
However, there are several options as to how these two
sets can be defined:
(a) the model is primitive e.g., just define a set of
variables and a set of equations
(b) decompose the model further e.g., define a set
components and their connections. The set of
variables is then the union of the sets of variables in
the sub models, and similarly the set of equations is
the union of the sets of equations in the sub models
and the equations implied by the connections
(c) a hybrid of (a) and (b) e.g., both variables and sub
models
 DERIVING THE EQUATIONS

Once a suitable decomposition has been established, it is

necessary to develop all the primitive models - e.g., derive
a set of variables and equations that describe the dynamic
behaviour of that section of the overall system.

Basically, we must utilize our knowledge of chemical

engineering science at this point - control volume analysis,
mass conservation, energy conservation, etc.

Experience with teaching this material suggests that it is

worthwhile to review the principles frequently used in
dynamic modelling - but this discussion is by no means
exhaustive (nor is my experience!). The approach is
relatively systematic and will allow derivation of
component models for a system that are consistent.
 DERIVING THE EQUATIONS

The following information should be developed at this

stage:
1. define the control volume(s) for balance equations
(again. . .identify the boundaries)
2. list the assumptions employed under what
conditions is the model valid?
Example: phase transitions
3. list the set of variables required to describe the
system e.g., symbol, verbal description, units
(always try to use a consistent set of units).
4. derive the set of equations always check that the
units of each term in an equation are consistent.
5. perform a degree of freedom analysis:
(a) identify the natural set of input variables to the
system.
(b) given values for this subset of the variables it
should be possible to calculate values for all the
other variables e.g. it is necessary that the
number of equations equal the number of
remaining unknown variables (ignore time
derivatives of variables in this analysis).
 DERIVING THE EQUATIONS

We will be applying one or more conservation principles

to a macroscopic control volume of definite size and shape
containing a fluid.

Further, we will assume that the contents of the control

volume are well mixed, so intensive properties are
uniform throughout the control volume and do not vary
with spatial position (or with another independent
variables such as polymer chain length).

=> one or more ordinary differential equations

(ODEs) with time as the independent variable.

These ODEs will be augmented with algebraic equations

(e.g., equations not involving time derivatives) that relate
variables or model phenomena that take place on a much
faster time scale than that of interest, for example:

a pseudo steady-state assumption

see discussion of phase equilibrium models later on

Note that in the absence of a vacuum, material will

expand (possibly with phase change) to completely fill the
control volume.
 MASS CONSERVATION
Mass is always conserved2:
CONTROL
VOLUME

FIN M(t) FOUT

BOUNDARY

Mass balance:

 Rate of 
 accumulation of   Flow of   Flow of 
  =  mass into 
 mass out of 
 mass in control     
   control volume   control volume 
 volume 

 kg s-1   kg s-1   kg s-1 

. . . what comes in must come out (eventually).

Notes:
(a) units of each term must be consistent
(b) terms in balance equations are always extensive
quantities.

2 barring nuclear reactions and relativistic effects

 MASS CONSERVATION EXAMPLE

Buffer tank open to the atmosphere:

FIN
P0

M(t) hMAX
h

P1
FOUT

Control volume: liquid in tank e.g., boundary is not

rigid, so size and shape of control volume will change as
the level rises and sinks

Assumptions:
isothermal system => no need for energy balance
single chemical species => fluid density constant
inlet flow defined as an input (or by an upstream
model) e.g., FIN (t) known.
vessel has uniform cross sectional area
model valid for interval 0
h
h MAX
 MASS CONSERVATION EXAMPLE

Variables (what quantities are we interested in?):

M* mass of fluid in vessel [kg]

density of fluid in vessel [kg m-3]
V* volume of fluid in vessel [m3]
FIN volumetric flow of inlet stream [m3s-1]
FOUT volumetric flow of outlet stream [m3s-1]
P0 atmospheric pressure [Nm-2]
P1 pressure at bottom of vessel [Nm-2]
A cross sectional area of vessel [m2]
h* liquid level in vessel [m]

Note (*): in steady-state models we do not usually worry

about quantities describing the "state" of the control
volume we are only concerned with the quantities
crossing the boundary and the fact that they must balance
at steady-state. This is a major reason why dynamic
models are more complex we now have to relate how
this state changes to the quantities crossing the boundary.
Clearly, the whole point of doing dynamic simulation is to
determine how this state changes with time.
 MASS CONSERVATION EXAMPLE

Equations:

Mass Conservation

dM
=  FIN
 FOUT (1)
dt

Relate volume and mass

V = M (2)

Relate volume and liquid level (constant cross sectional

area)

hA = V (3)

Hydrostatic pressure

P1 = P0 +
gh (4)

g
gravitational acceleration (9.81ms-2)

Flow pressure relationship flow out driven by

hydrostatic pressure in vessel

FOUT = k P1
P0 (5)

k
loss coefficient (value known)

 MASS CONSERVATION EXAMPLE

Degree of freedom analysis:

Total number of quantities = 11

Time invariant quantities (model parameters):

A,
,k,g (= 4)
Natural input set:
FIN (t), P0 (t) (= 2)
Note: these functions are defined by the "environment"
which could be either other models, or the engineer.

Remaining variables:
M, V, h, P1, FOUT (= 5)
Equations = 5

=> degrees of freedom are satisfied given

specification of input set above.

Note: due to the assumptions made above, equation (2)

could be substituted into (1) to eliminate M - leading to a
"volume balance." This is an error that is frequently made
- volume is not a conserved quantity (e.g., consider a non
isothermal and/or multi-component system) whereas
mass is.
 MASS CONSERVATION EXERCISE

Develop a model for the following system:

P0

h M (t)

P1
PUP PDOWN

FIN FOUT

Note:
(a) the inlet pipe has moved to a different physical
location consider how this changes the model.
(b) the control valves can be modelled by the following
equation (see later):

PIN
POUT
F = Cv x
 w
 MASS CONSERVATION EXERCISE

Variables:
F = flow through valve
Cv = valve sizing coefficient
x = valve stem position (control)
PIN = inlet pressure to valve
POUT = outlet pressure from valve

= density of fluid

w = density of water at reference
temperature
 SPECIES BALANCES

Unlike mass, chemical species are not conserved: if a

reaction takes place inside a control volume, reactants will
be consumed and products generated.

FIN,i N i (t) FOUT,i

A species balance can be written for each chemical species

in the system (note we are now using moles rather than
mass):

 Rate of   Flow of   Flow of   Rate of 

 accumulation  species i  species i  generation 
 = 
 + 
 of species i in   into the   out of the   of species i 
       
 system   system   system   by reaction 
[mol s-1 ] [mol s-1 ] [mol s-1 ] [mol s-1 ]

i = 1... NC
 SPECIES BALANCES

Further, we can sum these NC3 species balances to derive a

total mole balance.

 Rate of   Flow of   Flow of   Net rate 

 accumulation   moles   moles out   of generation 
 = 
 + 
 of moles in   into   of the   of moles by 
       
 system   the system   system   reaction 

 mol s-1   mol s-1   mol s-1   mol s-1 

Therefore, we can derive three different balance equations:

(a) NC species balances
(b) a total mole balance
(c) a mass balance
. . . clearly these are not independent the total number
of moles and the mass in the system can be related
algebraically to the number of moles of each species, e.g.:
NC
NT =
Ni
i=1
NC
M =
mi Ni
i=1

mi = molecular weight of species i [kg mol-1]

3 NC will be used to denote the number of chemical species in a system.

 SPECIES BALANCES

In general, it is best to derive a model in terms of the

minimal number of independent species balances (usually
NC) and derive other quantities via algebraic equations.
The rationale behind this statement should become clearer
when we discuss numerical solution of dynamic
simulation problems.
 SPECIES BALANCES EXAMPLE

Consider the same tank as before, but the following liquid

phase first order irreversible isomerization reaction takes
place (i.e. isothermal CSTR):
k
A 

R

B
Species A and B are present in dilute solution (i.e. NC = 3;
A, B, and the solvent).

FIN
C A,IN R
A 

k

B
P0
C B,IN

V(t)
C A (t) C B (t)
FOUT
P1 CA
CB

Assumptions:
vessel contents well mixed
isothermal, species A and B present in dilute solution
=> fluid density constant (e.g., neglect density changes
due to presence of A and B)
model valid for interval 0 < h
h MAX (note difference)
otherwise, same as above
 SPECIES BALANCES EXAMPLE

Variables:

V volume of fluid in vessel [m3]

density of fluid in vessel [kg m-3]
FIN volumetric flow of inlet stream [m3 s-1]
FOUT volumetric flow of outlet stream [m3 s-1]
P0 atmospheric pressure [N m-2]
P1 pressure at bottom of vessel [N m-2]
A cross sectional area of vessel [m2]
h liquid level in vessel [m]
CA concentration of species A in [mol m-3]
vessel
CB concentration of species B in [mol m-3]
vessel
CA,IN concentration of species A in inlet [mol m-3]
stream
CB,IN concentration of species B in inlet [mol m-3]
stream
g gravitational acceleration [m s-2]
k loss coefficient
kR reaction rate constant [s-1]
r reaction rate [mol m-3 s-1]
 SPECIES BALANCES EXAMPLE

Equations:

Mass Conservation

dV
= FIN
FOUT (1)
dt

Species balances for A and B

d (VCA ) dV dCA
= CA +V = FIN CA ,IN
FOUT CA
Vr (2)
dt dt dt

d (VCB ) dV dCB
= CB +V = FIN CB ,IN
FOUT CB + Vr (3)
dt dt dt

Notes:
use of volume in mass balance is only valid for the
assumptions above
concentrations in outlet stream equal to bulk
concentration because vessel contents well mixed
we have derived NC (=3) mass and species balances
this is sufficient to define the state of the system: all
other quantities can be related to {V, CA,CB} via
algebraic relationships. Note that an alternative
would have been to derive a species balance for the
solvent instead of the mass balance.
 SPECIES BALANCES EXAMPLE

Relate volume and liquid level

hA = V (4)

Hydrostatic pressure

P1 = P0 +
gh (5)

Flow pressure relationship

FOUT = k P1
P0 (6)

Reaction rate

r = k RCA (7)

. . . and additional equations to define N A (= CAV ) ,

N B (= CBV) , M (=
V) if desired.

Note: if significant density changes occur due to reaction

(e.g., not sufficiently dilute) then it is better to derive NC
species balances and derive the volume from the molar
volumes in solution.
 SPECIES BALANCES EXAMPLE

Degree of freedom analysis:

Total number of quantities = 16

Time invariant parameters:

A,
, g, k, kR (= 5)
Natural input set:
FIN (t), CA, IN (t),CB, IN (t ), P0 (t) (=4)
Remaining variables:
V, FOUT , P1 ,h, CA ,CB ,r (=7)

Equations = 7

=> degrees of freedom satisfied given specification

of input set above
 ENERGY CONSERVATION

According to the First Law of Thermodynamics, energy is a

conserved quantity. For an open system, this can be
expressed as:

rate of 
rate of work 

accumulation 
rate of addition  done on the 
 = of heat from the  +  +
 of energy in    system by the 
 system   environment  environment 
   
-1
[J s ] [J s-1 ] -1
[J s ]

 rate of energy  rate of energy discharge 

addition to system  from the system by 
 
 
 by material flow   material flow out of the
 into system   
   system 
-1 -1
[J s ] [J s ]

Warning: always start from the First Law when deriving

energy balances!!

Here, energy is the summation of internal energy (e.g.,

that associated with translation, rotation and vibration of
molecules), kinetic energy, and potential energy.

In most process simulation applications, it is usually

reasonable to neglect kinetic and potential energy (or to
perform separate balances for internal energy and these
other forms of energy) but always check this
assumption.
 ENERGY CONSERVATION

Given this assumption, the differential form of the First

Law of Thermodynamics for the following "general" open
system (see Modell and Reid (1983); pp. 39-41 for a
complete derivation) reduces to:

.
W
Fe Fl
he hl
U(t)

.
Q
dU
= Q + W +  Fe he
 Fl hl
dt e l

...where we have introduced time as the independent

variable.
Note that the conserved quantity is the internal energy of
the control volume contents (not the enthalpy!). The work
term is composed of two contributions:
dV
W = W s
P
dt
...the summation of the shaft work done on the system
(e.g., an impeller to keep the contents well mixed) and the
PV work due to changes in volume of the system.
 ENERGY CONSERVATION
Variables:

U internal energy of control volume [J]

contents
Q rate of heat addition to control volume [J s-1]
from environment
W rate of work done on the control volume [J s-1]
by the environment
Fi rate of material addition/discharge [kg s-1] or
[mol s-1]
hi specific enthalpy of material stream [J kg-1] or
[J mol-1]
 ENERGY CONSERVATION
The energy balance can therefore be expressed in two
equivalent forms:
dU dV
= Q + Ws
P +  Fe he
 Fl hl (E1)
dt dt e l

or dU dt can be eliminated by substitution of the diff-

erential form of the definition of enthalpy ( H = U + PV ):
dH dU dV dP
= +P +V
dt dt dt dt
which leads to:
dH dP
= Q + Ws + V +  Fe he
 Fl hl (E2)
dt dt e l

where,
H
enthalpy of the control volume contents [J]
Each form is convenient under certain assumptions. If the
volume of the system is constant, energy balance (E1)
reduces to:
dU
= Q + W s +  Fe he
 Fl hl
dt
and if the pressure of the system is constant, energy
balance (E2) reduces to:
dH
= Q + W s +  Fe he
 Fl hl
dt e l
 ENERGY CONSERVATION

Of course, there are many situations in which neither the

volume or the pressure remains constant over the time
horizon of interest. In this case, we must be able to express
the rate of change of the volume or pressure explicitly, or
have this implied by the algebraic equations (which leads
to difficulties - see discussion of "high index" problems
later).
 ENERGY BALANCE CONSTANT PRESSURE
EXAMPLE

Consider the CSTR again the reaction is now

exothermic and heat is removed by a jacket containing a
vaporizing medium, so an energy balance is necessary.

FIN
C A,IN
C B,IN
k
TIN A 

R

B
P0
hIN
M(t) H(t)
FOUT
NA (t) NB (t)
CA
CB
Q P1
T OUT
h OUT

Assumptions:
vessel contents well mixed
species A and B present in dilute solution
atmospheric pressure is constant (e.g., P0
f (t) )
neglect shaft work of impeller
neglect heat interaction with atmosphere
otherwise, same as above.
 ENERGY BALANCE CONSTANT PRESSURE
EXAMPLE

Variables:

M mass of fluid in vessel [kg]

V volume of fluid in vessel [m3]
NA number of moles of species A in [mol]
vessel
NB number of moles of species B in [mol]
vessel
CA concentration of species A in [mol m-3]
vessel
CB concentration of species B in [mol m-3]
vessel

density of fluid in vessel [kg m-3]
A cross sectional area of vessel [m2]
AJ heat transfer area of jacket [m2]
UJ overall heat transfer coefficient [J m-2 K-1 s-1]
for jacket
H enthalpy of vessel contents [J]
CA,IN concentration of species A in [mol m-3]
inlet stream
CB,IN concentration of species B in [mol m-3]
inlet stream
FIN volumetric flow of inlet stream [m3 s-1]
FOUT volumetric flow of outlet stream [m3 s-1]
TIN temperature of inlet stream [K]
TOUT temperature of outlet stream [K]
P0 atmospheric pressure [Nm-2]
P1 pressure at bottom of vessel [Nm-2]
 ENERGY BALANCE CONSTANT PRESSURE
EXAMPLE

g gravitational acceleration [m s-1]

k loss co-efficient
kR reaction rate constant [s-1]
r reaction rate [mol m-3 s-1]

IN density of inlet stream [kg m-3]

OUT density of outlet stream [kg m-3]
Q heat transfer to fluid in vessel [J s-1]
from jacket
hIN specific enthalpy of inlet stream [J kg-1]
hOUT specific enthalpy of outlet [J kg-1]
stream

E R activation energy [J mol-1]
R universal gas constant [J mol-1 K-1]
h liquid level in vessel [m]
T temperature of vessel contents [K]
TJ temperature of vaporizing [K]
medium
 ENERGY BALANCE CONSTANT PRESSURE
EXAMPLE

Equations:

Mass conservation

dM
= IN FIN
OUT FOUT (1)
dt

...note we are unable to assume
is constant

Species balances for A and B

dN A
= FIN C A, IN
FOUT C A
Vr (2)
dt
dN B
= F IN CB, IN
FOUT C B + Vr (3)
dt

Energy balance (constant pressure formulation e.g., it is

most convenient to use enthalpy in the accumulation
term)

dH
= Q + FIN IN hIN
FOUT OUT hOUT (4)
dt
Reaction rate

r = k Re
ER /RT CA (5)

 ENERGY BALANCE CONSTANT PRESSURE
EXAMPLE

Relate volume and mass

V = M (6)

Relate volume and liquid level

hA = V (7)

Relate mole numbers and concentration

N A = VC A (8)
N B = VC B (9)

Hydrostatic pressure

P1 = P0 +
gh (10)

Flow pressure relationship

FOUT = k P1
P0 (11)

Contents well mixed

=
OUT (12)
T = TOUT (13)
 ENERGY BALANCE CONSTANT PRESSURE
EXAMPLE

Define enthalpy holdup (implies temperature of contents)

H = MhOUT (14)

Physical Properties (abstract functions)

=
(T ) (15)

IN =
(T IN ) (16)
hIN = h(T IN ,C A, IN ,C B, IN ) (17)
hOUT = h(T,C A ,CB ) (18)

Heat transfer

Q = U J AJ (TJ
T) (19)

A commonly asked question is: how is the temperature in

the reactor determined? In fact, the temperature is
determined by the simultaneous solution of the complete
set of implicit relationships above. One can view equation
(4) as determining the extensive enthalpy H of the vessel
contents, equation (14) determining the intensive enthalpy
hOUT, and equation (18) implicitly determining T given hOUT.
ENERGY BALANCE CONSTANT PRESSURE
EXAMPLE

Degree of freedom analysis:

Total number quantities = 33

Time invariant parameters:

A, AJ ,U J , g, K ,k R ,
E R R, P0 (=9)
Natural input set:
F IN (t ), CA, IN (t), C B, IN (t ), T IN (t ), T J (t ) (=5)
Remaining variables:
M,V,N A ,N B ,CA ,CB ,
,H,FOUT ,TOUT ,
P1,r,
IN ,
OUT , Q ,hIN ,hOUT ,h,T (=19)

Equations = 19

=> degrees of freedom satisfied given specification

of input set above.
 ENERGY BALANCE CONSTANT PRESSURE
EXAMPLE
Note: in many textbooks it is common practice to include a
reaction term in the energy balance. The control volume
approach (equation (E1) or (E2)) clearly shows that no
energy crosses the system boundary due to a reaction
taking place, so a reaction term should not appear in the
energy balance.
If a reaction is taking place in an isolated system, the total
energy of the system remains unchanged, but the
distribution of energy between "heat of formation" energy
and "sensible heat" energy changes as the reaction
progresses (e.g., the temperature will rise or drop if the
reaction is exothermic or endothermic).
To reflect this, the constitutive equation defining the
specific enthalpy (18) must include the contribution of
both the heat of formation and the sensible heat of each
species to the total system energy. So, the zero energy
reference state for equation (18) must be defined as the
elements making up the chemical species in their standard
states at some temperature and pressure.
If heats of formation are not included in equation (18), a
heat of reaction term must be added to the energy balance,
and the heat of reaction must be calculated as a function of
temperature.
Overall, I consider it clearer and simpler to work with the
first law and include heats of formation in species
enthalpies (obviously, none of this is necessary if no
chemical reactions occur in the system of interest).
 ENERGY BALANCE EXAMPLE OF SYSTEM IN
WHICH PRESSURE AND VOLUME VARY

Consider a system in which an inert gas at high pressure is

employed to charge liquid rapidly to another vessel:

INERT
GAS

LIQUID

TO
OTHER
VESSEL

System decomposition: system is composed of two

components:

(a) the storage vessel

(b) the control valve

... we must develop models for each.

 MODEL OF CONTROL VALVE

For a more complete discussion see Luyben (1990), pp.

213-222:

FIN FOUT
TIN TOUT
PIN POUT
h IN hOUT

Assumptions:
fluid density constant
irreversible adiabatic expansion
rate of accumulation of material and energy in piping
negligible in comparison to material and energy flows
=> static (pseudo steady-state) mass and energy
balances
value has linear inherent characteristic
non flashing liquid phase.
no significant time lag of stem position in response to
changes in the control signal

Control valve dynamics are particularly important in

control related studies (see Norsk Hydro example).
 MODEL OF CONTROL VALVE

Variables:

FIN volumetric flow into valve [m3s-1]

CV valve sizing coefficient

density of liquid [kg m-3]

w density of water at reference
temperature [kg m-3]
x valve stem position
FOUT volumetric flow out of valve [m3s-1]
T IN temperature of inlet stream [K]
TOUT temperature of outlet stream [K]
hIN specific enthalpy of inlet stream [J kg-1]
hOUT specific enthalpy of outlet stream [J kg-1]
PIN pressure of inlet stream [N m-2]
POUT pressure of inlet stream [N m-2]
 MODEL OF CONTROL VALVE

Equations:

Mass conservation (static - constant fluid density)

FIN = FOUT (V1)

Energy conservation (static - irreversible adiabatic

expansion)

FIN
hIN = FOUT
hOUT (V2)

Flow pressure relationship (linear characteristic)

PIN
POUT
FIN = CV x (V3)
 / w

Physical properties (assume hIN ,T IN calculated from

equations in upstream unit)

hOUT = h(TOUT ) (V4)

 MODEL OF CONTROL VALVE

Degree of freedom analysis:

Total number of quantities = 12

Time invariant parameters:

CV ,
,
w (=3)

Natural input set:

x(t ), T IN (t), hIN (t ), PIN (t ), POUT (t) (=5)
Remaining variables:
F IN , FOUT ,T OUT ,hOUT (=4)
Equations = 4

=> degrees of freedom satisfied for input set

above.
 MODEL OF STORAGE VESSEL

INERT
GAS

LIQUID
FOUT
TOUT
POUT
hOUT

Control Volumes: control volume for gas phase, and

control volume for liquid phase. Both will vary in size and
shape as liquid is ejected.

Assumptions:
liquid density constant
ideal gas equation of state
adiabatic vessel (in time horizon of interest)
negligible heat and mass transfer between liquid and
inert gas (in time horizon of interest)
uniform cross sectional area
liquid phase well mixed and gas phase well mixed
valid in interval 0 < h < hmax
 MODEL OF STORAGE VESSEL

Variables:

MV mass of gas in vessel [kg]

ML mass of liquid in vessel [kg]

density of liquid in vessel [kg m-3]
FOUT volumetric flow of liquid from
vessel [m3s-1]
UV internal energy of gas phase [J]
UL internal energy of liquid phase [J]
P pressure of gas [Nm-2]
VV volume of gas phase [m3]
VL volume of liquid phase [m3]
hOUT specific enthalpy of outlet stream [J kg-1]
MW molecular weight of gas [kg mol-1]
R universal gas constant [J mol-1k-1]
TV temperature of gas [K]
h liquid level in vessel [m]
A cross sectional area of vessel [m2]
POUT pressure of outlet stream [Nm-2]
g gravitational acceleration [ms-2]
VTOT vessel volume [m3]
TL temperature of liquid [K]
TOUT temperature of outlet stream [K]
hL specific enthalpy of liquid phase [J kg-1]
hV specific enthalpy of gas phase [J kg-1]
 MODEL OF STORAGE VESSEL

Equations

Mass balance on gas phase (mass of gas remains constant)

dM V
=0 (S1)
dt
Mass balance on liquid phase
dM L
=
 FOUT (S2)
dt
Energy balance on gas phase (neither pressure or volume
constant)
dU V dV V
=
P (S3)
dt dt
Energy balance on liquid phase
dU L dV L
=
P
FOUT hOUT (S4)
dt dt
Relate mass and volumes of each phase

V L = M L (S5)
Ideal gas equation of state

VMV
PV = RT V (S6)
MW
 MODEL OF STORAGE VESSEL

Relate liquid volume and liquid level

hA = V L (S7)

Hydrostatic pressure

POUT = P +
gh (S8)

Volume constraint

V TOT = V L + V V (S9)

Well mixed liquid

T OUT = T L (S10)

hOUT = h L (S11)

Define internal energy holdups (implies temperatures of

phases)

U V + PV V = M V h V (S12)
U L + PV L = M L h L (S13)

Physical properties (abstract functions)

h V = h V (T V , P) (S14)
h L = h L (T L , P) (S15)
 MODEL OF STORAGE VESSEL

Degree of freedom analysis:

Total number of quantities = 22

Time invariant parameters:

, MW, R, A, g, VTOT (=6)
Natural input set:
FOUT (t) (=1)
Remaining variables:
M V , M L ,U V ,U L , P, V V ,V L ,hOUT ,
(=15)
V L L V
T ,T ,h, POUT ,T OUT ,h ,h

Equations = 15

=> degrees of freedom satisfied for input set above.

 PUTTING IT ALL TOGETHER
FORMULATING A SYSTEM MODEL

We now have models for each component in our system

these must be connected to form a system model.

Create instances of the storage vessel and control valve

models, and connect the outlet stream of the storage vessel
to the inlet stream of the control valve. This adds the
following equations to the system model:

Vessel. FOUT = Valve. FIN

Vessel. TOUT = Valve.T IN
(C1-4)
Vessel. POUT = Valve. PIN
Vessel. hOUT = Valve. hIN

Note the use of a pathname mechanism:

unit_name.variable_name
to identify uniquely variables in different models (hence
the same variable name may be used in many different
models).

Degree of freedom analysis for system model:

Number of Equations = 23 (V1-4, S1-15, C1-4)

Total number of quantities = 34

Number of time invariant parameters = 9

=> Degrees of freedom = 34 - 9 - 23=2

 PUTTING IT ALL TOGETHER
FORMULATING A SYSTEM MODEL

So, what simulation could we specify?

values for all the time invariant quantities (geometry,
physical characteristics of system)
specify the pressure of the downstream vessel (
Valve. POUT (t) ) and the stem position of the control
valve ( Valve. x(t) )
=> we can solve the system model for the time variation
of all the other quantities e.g., open the valve and
determine the system behaviour as liquid is ejected by
the high pressure gas.

Notes:
(a) although the material flow in this system is
unidirectional from the storage vessel to the valve
(assuming Vessel. P > Valve. POUT ), the information
flow in the model is not: we specify the downstream
pressure and calculate the upstream flow rate. This
means that the models of the valve and the storage
vessel must be solved simultaneously in general,
the decomposition techniques applied to solve
steady-state simulations cannot be applied to
dynamic simulations even when material flow in the
flowsheet is unidirectional.
(b) the introduction of the PV work terms in the energy
balances causes problems with consistent
initialization of this model (see discussion of "high
index" problems later on).
 PUTTING IT ALL TOGETHER
FORMULATING A SYSTEM MODEL

Exercise: what is the qualitative time variation of the

temperature of the gas phase if the liquid is ejected rapidly
from the vessel:
(a) predicted by the model above
(b) predicted by a model in which we neglect the PV
work terms in the energy balances.
...bearing this qualitative behaviour in mind, and the fact
that the vessel is constructed from carbon steel:
(i) what safety concerns would you have about this
operation if we started it with the gas at room
temperature?
(ii) in conducting safety studies, which model above ((a)
or (b)) would you recommend?
 MOMENTUM BALANCE

It is sometimes necessary to model the velocity (or

momentum) of the contents of a control volume. As flows
can in general be three dimensional, velocity is a vector
quantity with components corresponding to the velocity
resolved into the coordinate directions of the chosen
coordinate system. So, in principle we can formulate a
momentum balance for each co-ordinate direction (three
balances).

Applying Newton's Second Law to a control volume, we

obtain:
 Rate at which   Rate at which 
 Rate of accumulation   momentum in   momentum in 
 of momentum in the     
  =  ith direction 
the ith direction 

 ith direction in control     
   flows into   flows out of 
 volume   control volume   control volume 
 kg ms
2   kg ms
2   kg ms
2 

 Sum of forces 
 applied to 
 
+  control volume 
 
 in ith direction 
[N]  [kg ms-2 ]

. . . for each direction i in which material is flowing.

Note: momentum is defined as the product of mass and

velocity. Care should be taken if both the mass and the
velocity of the control volume are changing with respect
to time.
 MOMENTUM BALANCE EXAMPLE

Consider the buffer tank open to the atmosphere, but now

the fluid flows out into a long pipeline:

Control volumes: (a) liquid in tank, (b) liquid in pipeline

Assumptions:
same as for original buffer tank example
one dimensional plug flow in pipeline and
incompressible liquid => velocity uniform throughout
pipeline (macroscopic control volume)
 MOMENTUM BALANCE EXAMPLE

Variables:

M mass of fluid in vessel [kg]

density of fluid in vessel [kg m-3]
FIN volumetric flow of inlet stream [m3s-1]
FOUT volumetric flow of inlet stream [m3s-1]
AP cross sectional area of pipeline [m2]
L length of pipeline [m]
v velocity of fluid in pipeline (uniform) [ms-1]
FH hydraulic force on fluid [N]
FF frictional force resisting flow [N]
g gravitational acceleration [ms-2]
h level of liquid in vessel [m]
V volume of liquid in vessel [m3]
k F constant related to Fanning friction
factor
A cross sectional area of vessel [m2]
 MOMENTUM BALANCE EXAMPLE

Equations:

Mass conservation in tank

dM
=  FIN
 FOUT (1)
dt

. . . mass balance on pipeline unnecessary - incompressible

liquid in fixed volume (Flow in = Flow out)

Momentum conservation on pipeline - axial direction only

d( Ap Lv)
= FOUT v
 FOUT v + F H
FF
dt
dv
 AP L = F H
FF (2)
dt

Relate velocity and volumetric flow

APv = FOUT (3)

Hydraulic force

FH = AP
gh (4)

Frictional force (proportional to the square of the velocity

and the length of the pipe large Reynolds numbers)

FF = kF Lv 2 (5)
 MOMENTUM BALANCE EXAMPLE

Relate volume and mass

V = M (6)

Relate volume and liquid level

hA = V (7)

Degree of freedom analysis:

Total number of quantities = 14

Time invariant parameters:

, Ap , L, g, kF , A (=6)
Natural input set:
FIN (t) (=1)
Remaining variables:
M,FOUT ,v, F H , FF ,h, V (=7)
Equations = 7

=> degrees of freedom satisfied

 MULTI-PHASE SYSTEMS

Many systems encountered in chemical processes can be

abstracted as multiple phases that exchange material and
energy, for example:

MASS
FLUX
PHASE 2

ENERGY FLUX
MASS ENERGY
PHASE 1 FLUX FLUX
MASS
FLUX

ENERGY FLUX PHASE 3

System boundary

It is normally assumed that intensive properties inside

each phase are uniform (so macroscopic balances can be
performed on each phase).

Together, the phases form a simple thermodynamic

system that may be constrained at a known volume or
pressure.
 MULTI-PHASE SYSTEMS

Clearly there are two assumptions we can make about the

conditions in this system:
(a) the phases are in phase equilibrium with each other
leading to "equilibrium" models.
(b) the phases are not in phase equilibrium leading
to "non equilibrium" models.

. . . we will consider both cases.

 EQUILIBRIUM SYSTEMS

When modelling a system composed of multiple phases at

phase equilibrium, the following general comments can be
made:
(a) typically, we are assuming that phase equilibrium is
reached on a much faster time scale than the other
transients of interest.
(b) in general, the conditions for phase equilibrium for
multi-component (NC species) simple system
composed of
phases can be expressed as:

T 1 = T 2 =....= T 
P1 = P 2 =....= P
1i = i2 =....= i
i = 1... NC
where:
T j = temperature in phase j
P j = pressure in phase j
ij = chemical potential of species i in phase j
In other words, the condition of phase equilibrium
enforces ( 
1)( NC + 2) equations in the system
model.
(c) when formulating species and energy balances
consider a control volume encompassing all the
phases in equilibrium. The reasons for this relate to
the index of the resulting model (see Ponton and
Gawthrop (1991) for a full explanation), but the
consequence is that we cannot calculate the mass
and energy fluxes between the phases explicitly.
 EQUILIBRIUM SYSTEM FLASH DRUM

Consider a model of a flash drum in which it is assumed

there are vapour and liquid phases in phase equilibrium
with each other at all times:
V
y
i
T
P
hV
F yi
zi
TIN T,P
PIN
hIN xi
L
x
i
T
P
L
h
Control volume: the entire vessel contents (e.g.
encompassing both phases). Phases constrained to occupy
volume of the vessel.
 EQUILIBRIUM SYSTEM FLASH DRUM

Assumptions:
phases at phase equilibrium
individual phases well mixed
vessel adiabatic (no particular need for this just
makes model development easier)
both phases present at all times e.g. model valid
while T BUBBLE
T
T DEW
bulk flows into and out of vessel specified or
determined by upstream and downstream models.

Variables:

Ni number of moles of species i in vessel [mol]

F total molar flow to vessel [mol s-1]
L total liquid flow from vessel [mol s-1]
V total vapour flow from vessel [mol s-1]
zi mole fraction of component i in feed
stream
xi mole fraction of component i in liquid
phase
yi mole fraction of component i in vapour
phase
U internal energy of vessel contents [J]
hIN molar enthalpy of feed stream [J mol-1]
hV molar enthalpy of vapour phase [J mol-1]
hL molar enthalpy of liquid phase [J mol-1]
NV total number of moles in vapour phase [mol]
 EQUILIBRIUM SYSTEM FLASH DRUM

NL total number of moles in liquid phase [mol]

P system pressure (uniform) [Nm-2]
T system temperature (uniform) [K]
VTOT vessel volume [m3]
vV molar volume of vapour phase [m3mol-1]
vL molar volume of liquid phase [m3mol-1]
PIN pressure of inlet stream [Nm-2]
T IN temperature of inlet stream [K]
 EQUILIBRIUM SYSTEM FLASH DRUM

Equations:

Species balances

dN i
= Fzi  Lx i  Vyi
i = 1...NC (1)
dt

Energy balance (constant volume formulation)

dU
= FhIN
VhV
Lh L (2)
dt

Phase equilibrium (the temperature and pressures of each

phase have already been eliminated) can be expressed in
several different ways (each of which is a specialization of
the general reactions given above):

(a) single chemical species

P = P SAT (T )

(b) ideal-gas and ideal-solution in liquid phase

(Raoult's Law)

yi P = xi Pi SAT (T )
i = 1... NC

Note this reduces to the relationship above for a

single chemical species.
 EQUILIBRIUM SYSTEM FLASH DRUM

(c) for low pressure gases, non-ideal liquid solutions:

yi P =  i (T, x) xi Pi SAT (T )
i = 1...NC

(d) vapour-liquid equilibrium distribution coefficients

(k-values):

yi = ki (T , P, x, y)xi
i = 1... NC (3)

...where the K-values are calculated by any suitable

VLE model.

Definition of molar holdups

Ni = N V yi + N L xi
i = 1... NC (4)

Definition of energy holdup

U + PV TOT = N V h V + N L h L (5)

Summation of mole fractions (implicitly define N L and

NV)
NC

xi = 1 (6)
i=1

NC

yi = 1 (7)
i=1
 EQUILIBRIUM SYSTEM FLASH DRUM

Volume constraint

VTOT = N V vV + N Lv L (8)

Physical Properties (abstract functions)

h V = h V (T, P, y) (9)

h L = h L (T ,P, x ) (10)

v V = v V (T, P, y) (11)

v L = v L (T, P, x) (12)

Degree of freedom analysis:

Total number of quantities = 4NC + 16

Time invariant parameters
VTOT (= 1)
Natural input set:
F(t), zi (t),T IN (t), PIN (t ),hIN (t),V (t), L(t ) (= NC+16)
Remaining variables:

Ni , xi , yi ,U,h V ,h L , N V , N L , P, T,v V ,v L (= 3NC+9)

Equations: 3NC+9
=> degrees of freedom satisfied for this input set.
 EQUILIBRIUM SYSTEM FLASH DRUM

Exercise: derive an alternative model for the equilibrium

flash vessel in which separate species and energy balances
are written for both liquid and vapour phases. With
reference to the discussion of the index of differential-
algebraic equations (see later):
(a) what are the mathematical properties of this model?
(b) what additional information does this model
calculate?
(c) from a practical point of view, which of the two
models is appropriate for implementation in
SpeedUp?

Hint: it is useful to introduce variables in the balance

equations representing the flow of energy and each
chemical species between the two phases.
 EQUILIBRIUM SYSTEM DISTILLATION TRAY

Consider the following abstraction of a distillation tower:

 EQUILIBRIUM SYSTEM DISTILLATION TRAY

Each stage has the following regions of material and

energy accumulation:

V i,p
L i,p+1

VAPOUR ABOVE
FROTH
LIQUID
IN
DOWN VAPOUR FLUX
COMER

FROTH ON PLATE
LIQUID L i,p
FLUX

Vi,p-1

. . . having identified this decomposition, we can focus on

modelling each region in isolation.

Here, we will derive a model for the froth on the plate,

which is presumably designed to approach phase
equilibrium as closely as possible.

Note trays are numbered from the top of the column

downwards.
 EQUILIBRIUM SYSTEM DISTILLATION TRAY

Froth on plate:
LD i,p V i,p

MASS
FLUX

LIQUID VAPOUR
PHASE PHASE
ENERGY
FLUX

L i,p Vi,p+1

Control volume: the froth on the plate (e.g., encompassing

both phases)

Assumptions:
phases at phase equilibrium
individual phases well mixed
adiabatic
both phases present at all times e.g., model valid
while T BUBBLE
T
T DEW
number of moles in vapour phase negligible in
comparison to liquid phase
PV work terms in energy balance negligible
 EQUILIBRIUM SYSTEM DISTILLATION TRAY

Variables (i= species number, p = stage number):

Np number of moles in froth on stage p [mol]

LDp total molar flow of liquid from
down comer of stage p [mol s-1]
V p+1 total molar flow of vapour from
stage below [mol s-1]
Lp total molar flow of liquid leaving froth [mol s-1]
Vp total molar flow of vapour leaving froth [mol s-1]
xi, p mole fraction of component i in
liquid phase of froth
yi, p mole fraction of component i in
vapour phase of froth
xi,Dp mole fraction of component i in
down comer
yi, p+1 mole fraction of component i in
vapour rising from stage below
Up internal energy of froth [J]
hD
p molar enthalpy of liquid in
down comer [J mol-1]
h Vp+1 molar enthalpy of vapour rising
from stage below [J mol-1]
h pL molar enthalpy of liquid phase
of froth [J mol-1]
 EQUILIBRIUM SYSTEM DISTILLATION TRAY

h Vp molar enthalpy of vapour phase of

froth [J mol-1]
Tp temperature of froth [K]
Pp pressure of froth [Nm-2]
v pL molar volume of liquid in froth [m3mol-1]
Pp+1 pressure on stage below [Nm-2]
 EQUILIBRIUM SYSTEM DISTILLATION TRAY

Equations:

Total mole balance

dN p
= LD
p + V p+1
L p
V p (1)
dt

NC-1 species balances

d ( xi, p N p ) dN p dx i, p
= x i, p + Np =
dt dt dt (2)
LD D
p xi, p + V p +1 yi, p +1
L p x i, p
V p yi, p

i = 1... NC  1

Energy balance (both pressure and volume of froth can

vary see assumptions above)

dU p
= LDp h pD + V p+1h Vp+1
L p h pL
V p h Vp (3)
dt

Phase equilibrium

yi, p = ki, p (T p , P p , x p , y p )x i, p
i = 1... NC (4)

Definition of energy holdup

U p + P p N p v pL = N p h pL (5)
 EQUILIBRIUM SYSTEM DISTILLATION TRAY

Summation of mole fractions

NC

xi, p = 1 (6)
i =1

NC

yi, p = 1 (7)
i =1

Flow over weir (e.g., Francis weir formula)

L p = f ( N p ,v Lp ) (8)

Relate vapour flow into plate to pressure drop across plate

(e.g., hydrostatic head and dry plate losses)

Vp+1 = f (Pp , Pp+1 , N p ) (9)

Physical properties

h Vp = h V (T p , Pp , y p ) (10)

h pL = h L (T p ,P p , x p ) (11)

v Lp = v L (T p , P p x p ) (12)

See Pantelides et. al. (1988) for a more thorough discussion

of the properties of this model.
 NON EQUILIBRIUM SYSTEMS

(a) still a developing area - active research at the moment

(b) undoubtedly more appropriate for dynamic models;
in many cases it is not reasonable to assume that
phase equilibrium is reached at a much faster rate
than the time constants of interest.
(c) when deriving a model for a simple system:
(i) in a simple system we can assume the pressures
equilibrate:

P1 = P 2 =....P

i.e. (
-1) equations.
(ii) derive separate species balances for each species
in each phase include species flux terms to the
other phases
(iii) derive separate energy balances for each phase
include energy flux terms to the other phases.
(d) species and energy fluxes between each phase are
determined by multi-component heat and mass
transfer relationships:
(i) these are extremely difficult to generalize must
be derived on a case by case basis considering
geometry, flow patterns, etc.
(ii) usually dependent on empirically derived
correlations which may only be valid in a very
small operating region e.g., at or near steady-
state. If the dynamics take the system outside this
region, model assumptions are no longer valid!
 NON EQUILIBRIUM SYSTEMS

(e) the models get larger and more complicated, and

more detailed information concerning geometry,
internal features of the vessel, etc., must be
considered during model development.
. . . research continues.
 NON EQUILIBRIUM SYSTEM - FLASH DRUM

Consider a model of a flash drum in which it cannot be

assumed that the vapour and liquid phases are in phase
equilibrium:
V
y
i
T
P
hV
F yi
zi
TIN T,P
PIN
h IN xi
L
x
i
T
P
L
h

The vessel can be abstracted as two well-mixed phases

constrained within a fixed volume. Material and energy
fluxes exist between the phases:
 NON EQUILIBRIUM SYSTEM - FLASH DRUM

Control volumes: control volume for gas phase and

control volume for liquid phase. Both will vary in size and
shape as the liquid level rises and falls.

Assumptions:
individual phases are well-mixed.
vessel adiabatic
liquid feed stream well-mixed with the liquid phase
both phases present at all times
bulk flows into and out of vessel specified by user or
determined by upstream and downstream models

These assumptions are the simplest assumptions that can

be made. It is interesting to note that to assume anything
else would require a detailed knowledge of the internal
topology of the vessel and physico-chemical
discontinuities (e.g. the level of the feed pipe relative to
that of the liquid level, the area of heating coil/jacket
immersed in the liquid phase, etc.).
 NON EQUILIBRIUM SYSTEM - FLASH DRUM

Variables:

NiL number of moles of component i

in liquid phase [mol]
NiV number of moles of component i
in vapour phase [mol]
F total molar flow to vessel [mol s-1]
zi mole fraction of component i
in feed stream
A area of phase boundary [m2]

i molar flux of component i from
liquid to vapour phase [mol m-2s-1]
L total liquid flow from vessel [mol s-1]
xi mole fraction of component i
in liquid phase
V total vapour flow from vessel [mol s-1]
yi mole fraction of component i
in vapour phase
HL enthalpy of liquid phase [J]
hIN molar enthalpy of feed stream [J mol-1]
NL total number of moles in liquid [mol]
phase
vL molar volume of liquid phase [m3 mol-1]
P pressure of vessel contents [Nm-2]

H energy flux from liquid to vapour
phase (heat + material flow) [Jm-2s-1]
hL molar enthalpy of liquid [J mol-1]
HV enthalpy of vapour phase [J]
 NON EQUILIBRIUM SYSTEM - FLASH DRUM

NV total number of moles in vapour [mol]

phase
VV molar volume of vapour phase [J mol-1]
hV molar enthalpy of vapour phase [J mol-1]
xi* mole fraction of component i
in liquid phase at phase boundary
yi* mole fraction of component i
in vapour phase at phase boundary
kiL mass transfer coefficient for
component i in the liquid phase [mol m-2s-1]
kiV mass transfer coefficient for
component i in the vapour phase [mol m-2s-1]
k HL heat transfer coefficient in the
liquid phase [J m-2k-1s-1]
k VH heat transfer coefficient in the
vapour phase [J m-2k-1s-1]
TL temperature of liquid phase [K]
TV temperature of vapour phase [K]
T* temperature at phase boundary [K]
VTOT vessel volume [m3]
 NON EQUILIBRIUM SYSTEM - FLASH DRUM

Equations:

Species balance for each species in the liquid phase:

dNiL
= Fzi
A i
Lxi
i = 1 NC (1)
dt

Species balance for each species in the vapour phase:

dNiV
= A i
Vyi
i = 1 NC (2)
dt

Energy balance for the liquid phase (pressure and volume

can vary)

dH L dP
= FhIN + N Lv L
A H
Lh L (3)
dt dt

Energy balance for the vapour phase:

dH V dP
= A H + N V v V
VhV (4)
dt dt
 NON EQUILIBRIUM SYSTEM - FLASH DRUM

Interphase mass and heat transfer relations:

(
 i = kiL xi
x*i )
i = 1 NC (5)

(
 i = kiV y*i
yi )
i = 1 NC (6)

NC
 H = k HL (T L
T )+h * L
 i (7)
i=1

NC
 H = k VH (T *
T V ) + h V  i (8)
i=1

Assume equilibrium at the phase boundary:

( )
yi* = Ki* T * ,P, x * , y* xi*
i = 1 NC (9)

Definitions of molar holdups

NiL = N L xi
i = 1 NC (10)

NiV = N V yi
i = 1 NC (11)

Definitions of enthalpy holdups

H L = N Lh L (12)

H V = N V hV (13)
 NON EQUILIBRIUM SYSTEM - FLASH DRUM

Volume constraint:

VTOT = N V vV + N Lv L (14)

Summation of mole fractions

NC

xi = 1 (15)
i=1

NC

yi = 1 (16)
i=1

Physical Properties (abstract functions)

(
v L = v L T L , P, x ) (17)

(
v V = v V T V , P, y ) (18)

(
h L = h L T L , P, x ) (19)

(
h V = h V T V , P, y ) (20)
 DYNAMIC MODELLING CONCLUSIONS

Dynamic modelling is a difficult task that must be

addressed on a case by case basis:
(i) understand the system under investigation what
are details of the internals of the system:
(a) geometry
(b) internal topology
(c) phases present
(d) when behaviour changes e.g., buffer tank
overflows.
(ii) what are the contributions to the dynamics
control volume analysis: mass, energy, and/or
momentum dynamics.
(iii) apply chemical engineering principles to derive a
complete model.
(iv) develop and maintain full model documentation
never extrapolate model beyond region in which
assumptions are valid.
 DISTRIBUTED SYSTEMS

All the above discussion is based on the premise that it is reasonable to assume that the
macroscopic control volumes employed to derive the model equations are well mixed.
Whether or not this assumption is reasonable, our options are pretty limited by the
capabilities of current process modelling technology. In this section we will briefly
discuss the current options when the above assumption breaks down.

Example: isothermal plug flow tubular reactor with axial dispersion:

0 L
z

CAIN CAOUT

CBIN CBOUT

2A
B

Equations: species balances derived from microscopic control volume analysis:

 CA  2CA C
=D
v A
2kC A2
t z 2
z
 CB  CB
2
C
=D
v B + kC A2
t z 2
dz

Note that these are now partial differential equations (PDEs) rather than ordinary
differential equations, with the following independent variables:

time
axial position along the reactor (in the domain z
[O,L] ).

The concentrations of the species in the reactor are now a function of both time and
spatial position in the above domain (i.e., CA (z,t) ). We need to solve for these functions
(or fields).

Note that this problem is somewhat more difficult than one in reaction engineering
where we would explicitly define the boundary conditions as functions and solve the
reactor model in isolation. Here, we do not know the boundary conditions explicitly
because the inlet and outlet streams are coupled to upstream and downstream units,
and the boundary conditions are determined implicitly by numerical solution of these
other units. As the solution to the reactor model (at least) influences the downstream
boundary condition (and the upstream one if there is a recycle stream), all the units
must be solved simultaneously. This gives rise to a process model composed of a mixed
system of partial differential, ordinary differential, and algebraic equations (PDAEs).
Further, the PDEs in this system may be distributed on multiple spatial domains, each
corresponding to the spatial dimensions of different vessels in the flowsheet. Finally, as
noted above, the boundary conditions of these multiple domains are coupled.

Current technology will not accept a PDE model in the form above. One approach is to
apply what is called the Method of Lines. Here, the PDE model is reduced to a system of
ODEs by discretization of the spatial domain. The system of ODEs is then coded into
the simulation package and the overall process model is solved with general purpose
ODE/DAE solvers.

In the example above, discretization of the spatial domain yields a set of ODEs at a
series of interior (or mesh) points in the domain. At these points, the spatial derivatives
are replaced by a discrete approximation, for example a centered differentiation
formula on a regular mesh:

CA ,i C A,i +1
C A ,i
1


z 2h
 C A ,i C A ,i+1
2C A,i + C A,i
1
2

z 2 h2

where the subscript i denotes the spatial mesh point, and h the distance between mesh
points.

Substituting these into the PDE at each mesh point yields the system of ODEs:

dC A ,i (C A ,i+1  2C A,i + CA ,i 1 ) (C  C A,i 1 )

=D 2
 v A ,i +1  kC A ,i 2
dt h 2h

i = 1Nmesh

Note that in the discretized model we solve for the time variation of the concentrations
at a series of fixed spatial positions (i.e., CA, i (t) ). The above equations have to be
tweaked to take into account the boundary conditions.

The gPROMS simulator (Barton, 1992; Oh, 1995) can now perform this discretization on
a fixed spatial mesh automatically. Note that only the derivation of the ODEs is
automated, so all the observations below also apply to gPROMS.

However, this approach should be applied with extreme caution (i.e., these notes are
not sufficient, study this in more detail). Two important issues are:

an appropriate discretization formula is very dependent on the physics of the

system (e.g., the centered difference formula is appropriate for systems with
dispersion, but not for purely convective systems).
the spatial discretization introduces error into the solution. With a fixed regular
mesh, the number of mesh points must be sufficiently large so that this
discretization error is below the desired tolerance during all transients of interest.
This can lead to very big systems of ODEs.

The above approach is called the finite difference method. There also two widely used
alternatives: finite element methods and spectral methods. Research continues . . .
 SECTION II

DYNAMIC SIMULATION

Paul I. Barton
Department of Chemical Engineering
Massachusetts Institute of Technology
Cambridge, MA

copyright Paul I. Barton, July 1999

 ORDINARY DIFFERENTIAL EQUATIONS
DEFINITIONS

We will be concerned with explicit first-order systems of

ordinary differential equations (ODEs) of the form:

dx 1
= x 1 = f1 (x 1 ,x 2 ...., xn ,t)
dt
dx 2
= x 2 = f2 (x1 ,x 2 ...., x n ,t)
dt



dx n
= x n = fn (x1 , x2 ...., x n ,t)
dt

or in vector notation:

dx
= x = f(x(t),t) (ODE)
dt

Note: in general, f is a n vector of nonlinear functions

involving the dependent (or state) variables x(t) and the
independent variable t (which we will always assume to be
time).

First-order refers to the fact that only first derivatives of the

independent variables appear. Explicit refers to the fact
that each time derivative is expressed individually as an
explicit function of the variables (therefore the Jacobian
with respect to x is the identity matrix I ).
 ORDINARY DIFFERENTIAL EQUATIONS
DEFINITIONS

The more general implicit form is defined in terms of

general functional relationships between x , x , and t , e.g.:

F( x ,x,t) = 0 (IODE)

with the condition that the Jacobian (matrix of partial

derivatives) of these functions with respect to x is
nonsingular everywhere in the time horizon of interest (to
distinguish from a system of differential-algebraic equations
see later).

It is important to realize that system (ODE) is coupled

each function can in general depend on all the dependent
variables. The consequence of this coupling is that the
equations cannot be solved independently and thus must
be solved simultaneously.

(ODE) is also sometimes referred to as a nonlinear state-

space description of a dynamic system.
 ORDINARY DIFFERENTIAL EQUATIONS
DEFINITIONS

The general solution x(t) of the first order system (ODE):

involves, in general, n arbitrary constants.
these constants can be fixed by imposing n side
conditions => a particular solution satisfying these side
conditions.
if these side conditions fix the values for all the
elements of x at the same initial point in time, the
combination of the system (ODE) and the side
conditions constitutes an initial value problem (IVP):

x
= f (x,t) 

x(t 0 ) = x 0 
(IVP)
t
[t 0 ,t f ] 

where the notation x(a) denotes the vector of

variables at time a, t0 is used to denote the initial time,
t f the final time. Note the distinction between system
and problem.

Dynamic simulation of chemical processes gives rise to

initial value problems.
 ORDINARY DIFFERENTIAL EQUATIONS
DEFINITIONS

In system (ODE), if f is independent of t , the problem

(and the system) are said to be autonomous and non-
autonomous otherwise. A non-autonomous system can be
converted to an autonomous system by defining a new
variable:

x n+1 = t
x n+1 = 1

and the autonomous system:

X = F(X)

where,
X = (x 1 ,x 2 ,, x n
xn+1 )T
F = ( f1 , f2 ,, fn
1)T
so, there is no loss of generality in assuming a system is
autonomous.
 ORDINARY DIFFERENTIAL EQUATIONS
DEFINITIONS

Higher order systems (e.g., those involving the second and

higher order derivatives of the dependent variables with
respect to the independent variable) can always be
transformed into a (larger) equivalent first order system.

Example: consider the second order system:

x1 = f 1 (x1, x2 )
x2 = f 2 ( x1, x2 )
define a new variable x3 = x1:

=> x3 = x1

and this leads to the first order autonomous system:

x1 = x3
=> x2 = f 2 ( x1, x 2 )
x3 = f1 ( x1, x2 )

So, there is again no loss of generality in assuming a

system is first-order.
 ORDINARY DIFFERENTIAL EQUATIONS
DEFINITIONS

The first-order system (ODE) is said to be linear (or linear

time varying) if it takes the form:
x
= A(t)x + g(t)

Further if the n
n coefficient matrix A is independent of t

, the system is said to be linear with constant coefficients (or
linear time invariant):

x
= A(t)x + g(t) (LTI)

A closed form general solution can be obtained for the

autonomous linear constant coefficient system:1

x
= Ax

expressed as:
n
x(t) =  ci exp(
it )si
i=1

where
i is an eigenvalue of A , s i is the corresponding

eigenvector, and ci are the arbitrary constants that
characterize the general solution (see above). The ci can
be calculated for a particular solution given values for x at
some initial time (an initial value problem).

1the result is stated assuming A has n distinct eigenvalues - the Jordan form defines
the solution when this does not hold
 ORDINARY DIFFERENTIAL EQUATIONS
DEFINITIONS

System (LTI) is sometimes written:

x = Ax + Bu(t)

where u(t) are the known inputs to the systems. This form,
which is popular with control engineers, is called a linear
time invariant system.

Linear constant coefficient systems are the only general

class of systems for which analytical solutions can always
be found. Process simulation problems give rise to
nonlinear systems, so we must resort to numerical
solution of (IVP) (numerical integration).

Numerical methods generate approximations to particular

solutions satisfying given initial conditions. Obviously, it
is desirable if we guarantee that the results of a numerical
method are within a specified numerical tolerance of the
"true" particular solution => guarantees an accuracy of the
dynamic simulation results (e.g., that they are a solution of
the model equations this does not guarantee that the
model itself is an adequate representation of the true
system behaviour).
 DIFFERENTIAL-ALGEBRAIC EQUATIONS
DEFINITIONS

A generalization of the explicit linear time invariant ODE

(LTI) can be expressed as:

Ax + Bx = g(t) (LTIDAE)

If the matrix A is nonsingular, then this system reduces

trivially to an ODE:

1
1
x =
A Bx + A g(t) = Cx + h(t)

1
on the other hand, is A singular (i.e., A does not exist)
then (LTIDAE) is known as a linear time invariant
differential-algebraic equation (or DAE).

Example: if A has a row of all zeroes (making it singular),

then that row of (LTIDAE) will amount to a completely
algebraic relationship between the state variables. This is
the origin of the term differential-algebraic equation.

See (Campbell, 1980; Campbell, 1982; Brenan et al., 1996)

for detailed discussions of the theory of DAEs.
 DIFFERENTIAL-ALGEBRAIC EQUATIONS
DEFINITIONS

Necessary and sufficient conditions for solvability of

(LTIDAE) are that the family of matrices
A + B defined
by the scalar variable
is nonsingular for some value of

R (the family
A + B is known as a matrix pencil). Such
a matrix pencil is said to be regular.

Example: if B is nonsingular, then the matrix pencil is

regular (i.e., at least for
= 0 ,
A + B is nonsingular).

If the matrix pencil
A + B is regular then there exist

nonsingular matrices P and Q such that:

I 0 
C 0 
PAQ =   and PBQ =  
 0 N   0 I 
   


1
where N is a nilpotent matrix to degree
(i.e., N 0


and N = 0 ).

Note that
is equivalent to the differential index (see

definition below) of (LTIDAE).

This property can be used to construct an analytic solution

for (LTIDAE).
 DIFFERENTIAL-ALGEBRAIC EQUATIONS
DEFINITIONS

Applying the co-ordinate transformations defined by P

and Q to (LTIDAE) (i.e., multiply through by P and define
x = Qy ) combined with the above identities yields:

Ax + Bx = g(t) (LTIDAE)

P AQ y + P BQ y = Pg(t)

y1 + C y1 = h1(t)


N y 2 + y 2 = h2 (t )

The first block row of this transformed and uncoupled

system is just a linear time invariant ODE for which the
analytic solution is already known. The analytic solution
of the second block row can be constructed as follows:

d 
N + I y 2 (t) = h 2 (t)
 dt 

and using a famous identity yields:

 i 
1 (i )
i d i d h2
y 2 (t) =  (
1) N h 2 (t) =  (
1) N
i
  (i) (t)
i=0 dt i=0 dt

the infinite series being truncated by the nilpotency of N .

 DIFFERENTIAL-ALGEBRAIC EQUATIONS
DEFINITIONS

Note that solvability in general will also require sufficient

differentiability of the right hand sides (or forcing
functions). This result also implies that the differential
index must be
2 for time derivatives of the forcing
functions to appear in the solution.
 DIFFERENTIAL-ALGEBRAIC EQUATIONS
EXAMPLES

Example 1:

x 1 + x2 = g1(t)
x 2 = g2 (t)

Hence:
 1 0   0 1 
A=  B= 
 0 0   0 1 
 1
1   1 0 
P=  Q =  
 0 1   0 1 

 1
1   1 0   1 0 
 PAQ =   = 
 0 1  0 0   0 0 

 1
1   0 1   0 0 
 PBQ =   = 
 0 1  0 1   0 1 

=> N = [0 ] C = [0]
 DIFFERENTIAL-ALGEBRAIC EQUATIONS
EXAMPLES

Hence the system is nilpotent to degree 1 (index 1). Only

the zero matrix is nilpotent to degree 1, so index 1 system
does not involve time derivatives of the inputs.

Analytic solution to Example 1:

 1
1   g1   g1
g2 
  = 
 0 1   2   g2
g 

x
1 = g1 (t)
g2 (t)

x = g (t)
2 2

Which is evident from inspection of the original system.

 DIFFERENTIAL-ALGEBRAIC EQUATIONS
EXAMPLES

Example 2:

x 1 + x2 = g1(t)
x1 = g 2 (t)

Hence:

1 0 
0 1 
A=  B= 
 0 0   1 0 

1 0 
0 1 
P=  Q =  
 0 1   1 0 

 1 0  0 1   0 1 

PAQ =   = 
 0 0  1 0   0 0 
 0 1  0 1   1 0 

PBQ =   = 
 1 0  1 0   0 1 

 0 1 

N= 
 0 0 
 DIFFERENTIAL-ALGEBRAIC EQUATIONS
EXAMPLES

0 1 
0 1 
0 0 
N2 =   = 
 0 0  0 0   0 0 

Hence the system is nilpotent to degree 2 (index 2).

Analytic solution to Example 2:

x1 
0 1 
y1 
 =  
 x2   1 0   y2 

 x1   y2 

 = 
 2   1 
x y

 y1   g1   0 1   g1 
 = 
  
 2   2  
y g 0 0 g 
 2 

x1 = g 2 (t)


x 2 = g1(t)
g2 (t)

Which is again evident from inspection of the original

system.
 DIFFERENTIAL-ALGEBRAIC EQUATIONS
DEFINITIONS

The analytical solution above reveals several interesting

properties of DAEs:
the solution can involve the first and higher order time
derivatives of the forcing functions.
not all initial conditions will satisfy the solution (i.e.,
y 2 (t 0 ) cannot be chosen arbitrarily).
the solution must satisfy explicit or implicit algebraic
constraints.

These properties will also be revealed by our later, more

physically based, analysis.

As in the ODE case, linear time varying and nonlinear

DAEs are elusive to general analytical results, so we must
resort to numerical solutions.
 DYNAMIC SIMULATION

A dynamic model (or system of equations) encapsulates a

representation of the time dependent behaviour of a
system in the real world.

Dynamic simulation is one2 activity (or calculation) which

utilizes dynamic models. Specifically, to formulate a
dynamic simulation we must:
specify the model representing the system
the initial state of the system
the time variation of the inputs to the system
mathematically an initial value problem that can be

solved numerically to determine the time variation of
the system from the initial state to some final state.

. . . this clarifies the distinction between a representation of

the system (the model) and the calculations it is used for
(the simulation). Obviously the same model may be used
for many different simulations (different scenarios).

2 but by no means the only calculation.

 DYNAMIC SIMULATION

More formally, we can view a dynamic simulation as an

experiment with a model.

An experiment is composed of three elements:

the object (or artifact, system) under investigation - in
this case we use a model to represent the behaviour
rather than experimenting with a physical system.
the experimental frame - e.g., the circumstances under
which the system is to be investigated.
the data generated by execution of the experiment.
Obviously, the experimental frame is unique to each
experiment and a dynamic simulation experiment will
require specific categories of information in order to be
fully defined.

This second section of the notes is concerned with

specification and solution of dynamic simulation
experiments:
software tools that support specification and solution
of dynamic simulation experiments.
specifying the experimental frame: initial conditions
and input trajectories.
checking if the experiment is well posed.
numerical solution of the experiment.
 DYNAMIC SIMULATION AS AN EXPERIMENT

. . . the most useful model is one that can be disengaged

from the details of individual experiments (it can be
reused).
 SETTING UP THE DYNAMIC SIMULATION
EXPERIMENT - EXAMPLE

Let us consider the model of a non isothermal CSTR from

section I and identify the information required to define
an experimental frame:

(a) assign values to all the time invariant (constant)

parameters:

A, AJ ,UJ ,g, K, K R ,
E R, R, P0

(b) select a subset of the variables to be designated as

inputs (disturbances, controls), e.g.:
FIN (t), CA, IN (t),CB, IN (t ),T IN (t), T J (t)
and specify their time variation explicitly, e.g.:

F IN TJ

t t

Denote a variable that is a known function of time by

FIN (t) - this function is not necessarily continuous or
differentiable (e.g. step change above).
 SETTING UP THE DYNAMIC SIMULATION
EXPERIMENT - EXAMPLE
Notes:
the selection of a set of input variables is analogous
to the selection of the degrees of freedom in a
steady-state simulation.
once we have selected the input variables, we should
have a fully determined set of differential equations
(number of unknown variables = number of
equations) that determine the time variation of these
unknowns uniquely.
from a practical point view, in dynamic simulation
we are very limited in our choice of input variables
see discussion of high index problems later
(numerical methods for solution of dynamic 'design'
problems are only just emerging).
This raises the question of how to select a set of input
variables that will lead to a problem we can solve with
current technology?
 SETTING UP THE DYNAMIC SIMULATION
EXPERIMENT - EXAMPLE
(c) specify the initial state of the system (initial condition
at t = t0 ). This defines where we start from! Note:
i) correct numerical solution of the initial value
problem requires consistent values for all the
unknown variables at t = t0
ii) in general, we are not free to specify initial values
for all the unknowns independently the values
must satisfy the model equations. In this case, for
example, only the values:
M(t0 ), CA (t 0 ),CB (t 0 ),T (t0 )
are sufficient to define the initial state of the
reactor - given these, all the other initial values can
be determined by solving the model equations.
Hence we must address the questions:

for models of the general functional form derived

in section I, what is the number of side conditions
at t = t0 that is sufficient to define uniquely the
initial state of the system?
how are these side conditions expressed?
given the side conditions, how can initial values for
all the unknowns be determined?

(d) the duration of the simulation - e.g. t f .

 SETTING UP THE DYNAMIC SIMULATION
EXPERIMENT - EXAMPLE

The initial value problem is now fully defined submit it

to the computer for numerical solution.

This process will determine the variation of all the

unknowns with respect to time in the interval [ t0 ,t f ] e.g.:

Dynamic Simulation

6

5

Ca [mol m-3]

4

3

2

1

0

0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
6
6.5
7
7.5
8
8.5
9

Time [hr]

Note that we are calculating functions of time, rather than

point (steady-state) values. These are commonly referred
to as the 'trajectories' of the variables. Again, these
functions need not be continuous or differentiable.
Change any of the above information and you have a new
scenario a different dynamic simulation experiment,
with different solution trajectories.
 DYNAMIC SIMULATION TOOLS

A dynamic simulation tool is some form of computer

software that supports the engineer in building dynamic
models and solving the resulting simulation problems
(e.g., something more than coding everything from scratch
in FORTRAN).

There are really three features of a dynamic simulation

tool that are of interest:
(a) how is the model developed and coded?
(b) how are individual simulation experiments
specified?
(c) how is numerical solution of the simulation
experiment conducted?
. . . most dynamic simulation tools offer a mixture of
different features in each category.

Note: in the early days analogue computers were widely

used for dynamic simulation of chemical processes I
don't think anyone is doing this any more! All these
remarks are concerning dynamic simulation on digital
computers.
 CODIFYING THE MODEL

increasing ease of use,

maintenance and
debugging

PROGRAM M ING LANGUAGES

(FORTRAN, PASCAL, C)

automated transformation

EQUATION B ASED
LANGUAGES
(SPEED UP, ASCEND , AB ACUSS)

automated transformation

PHENOM ENOLOGICAL
LANGUAGES
(M OD EL.LA, VED A)

increasing
flexibility
 PROGRAMMING LANGUAGES
The old fashioned approach! Basically, the user codes a
subroutine in some general-purpose programming
language (usually FORTRAN) that returns some specific
information concerning the process model.

All numerical methods for solving ordinary differential

equations require either:
function evaluations consider system (ODE): a
function evaluation will evaluate the vector of
functions f given values for x and t (e.g., evaluate x
given x and t )
residual evaluations a more general approach: the
model equations are rearranged to the general form:
g( x , x,t ) = 0
. . . and the vector of functions g (residuals) is
evaluated given values for x , x and t . Note that the
residuals will only equal the zero vector if the set of
values { x , x , t } are a solution of the differential
equations.

Consider the model of a buffer tank from section I

overleaf, I have coded a FORTRAN subroutine to return
the residuals given values for the unknowns. This
subroutine would (for example) be called repeatedly
during an iterative process to find variable values that
satisfy the differential equations.
 RESIDUAL SUBROUTINE EXAMPLE

SUBROUTINE BUFFER(X,XDOT,U,T,RES)
IMPLICIT DOUBLE PRECISION (A-H, O-Z)
DIMENSION X(5), XDOT(5), U(2), RES(5)
C
C Definition of unknown variables:
C
C X(1) mass of fluid in vessel [kg]
C X(2) volume of fluid in vessel [m3]
C X(3) volumetric flow out of vessel [m3 s-1]
C X(4) pressure at bottom of vessel [N m-2]
C X(5) liquid level in vessel [m]
C XDOT(i) time derivative of X(i)
C
C Definition of input variables:
C
C U(1) volumetric flow into vessel [m3 s-1]
C U(2) atmospheric pressure [N m-2]
C
C Definition of time invariant parameters:
C
C AREA cross sectional area of vessel [m2]
C DENS fluid density [kg m-3]
C COEF loss coefficient
C G gravitational acceleration [m s-2]
C
AREA = 1.767
DENS = 1000
COEF = 0.01
G = 9.81
C
C Evaluate residuals given X, XDOT and U
C
RES(1) = XDOT(1) - (DENS*U(1) - DENS*X(3))
RES(2) = DENS*X(2) - X(1)
RES(3) = X(5)*AREA - X(2)
RES(4) = X(4) - (U(2) + DENS*G*X(5))
RES(5) = X(3) - (COEF*SQRT(X(4) - U(2)))
RETURN
END
 PROGRAMMING LANGUAGES

In addition, other subroutines can be written that return:

Jacobian evaluations many numerical methods work
better if it is possible to evaluate the matrix of partial
derivatives of the functions with respect to the
dependent variables analytically. The user can derive
expressions for the partial derivatives and then write
an appropriate subroutine (extremely tedious and
prone to error).
In fact, there are now automatic differentiation codes
being quite widely used (see Griewank et al., 1991).
These codes can take a FORTRAN residual evaluator
as input, and automatically generate a FORTRAN
routine coding the Jacobian evaluation.
incidence information for medium to large systems
of equations (>20) only a few of the variables will
appear in each equation sparse equation system.
Therefore, most of the Jacobian entries (partial
derivatives) will be zero at all times. Incidence
information encapsulates this an explicit statement
of which variables appear in each equation. This can
be used to great advantage in numerical Jacobian
evaluations and in solving systems of linear
equations (sparse linear algebra). Again, this is tedious
and error prone.

. . . as you can imagine, these become monumental tasks to

do by hand when dealing with models composed of
1,000's of equations!
 PROGRAMMING LANGUAGES

Advantages:
anyone can do it you don't need to buy a dynamic
simulator!
the ultimate in flexibility can express the most
general functional forms.
Disadvantages:
tedious, time consuming, error prone, not a natural
way to represent equations, etc. . . Probably the
biggest problem is human error - one Jacobian entry
that is calculated incorrectly can cause a numerical
method to fail to converge.
difficult to separate model from the mechanics of its
solution model locked into a single application
(calculation).
mathematical information concerning the model is
lost e.g., the only access we have to the model is
the residuals given values for the variables.
inflexible if you want to change the model (or even
the set of input variables) must code a completely
new model.
no way to check automatically if the model has been
coded correctly
requires expertise in modelling, computer science and
mathematics expert users!
 EQUATION BASED LANGUAGES

The simulation software supports an equation based

programming language with which to code process
models.

These languages are:

high level and specialized to the task of specifying
equation based models.
declarative rather than procedural the purpose is
to specify the functional form of the model (e.g., a
system of equations) and is not a series of statements
to be executed in a sequence.
. . . currently, this is the state of the art.

This high level representation can be automatically

translated into FORTRAN subroutines to interface to
standard numerical solvers.

Procedural: x := x^2 + 1 ;

Compute x2 + 1 based on the current value of memory

location x and assign the result to memory location x.

Declarative: x = x^2 + 1 ;

Find solution of quadratic equation x2 + 1 - x = 0.

 EQUATION BASED LANGUAGES EXAMPLE

MODEL Buffer_Tank

###########################################################
#
# ABACUSS model of a buffer tank - Paul I. Barton 7/27/94
#
# NOTE: anything following a # symbol is a comment
#
###########################################################

PARAMETER
Cross_Sectional_Area AS REAL
Grav_Acceleration AS REAL
Loss_Coefficient AS REAL
Density AS REAL

VARIABLE
Mass_Holdup AS Mass
Liquid_Volume AS Volume
Flow_In, Flow_Out AS Volumetric_Flow
Atm_Press, Press AS Pressure
Liquid_Level AS Length

STREAM
Inlet : Flow_In, Atm_Press AS MainStream
Outlet : Flow_Out, Press AS MainStream

EQUATION

# Mass balance - $ is the time differential operator

$Mass_Holdup = Density*Flow_In - Density*Flow_Out ;

# relate liquid volume and mass

Density*Liquid_Volume = Mass_Holdup ;

# relate liquid level and volume

Liquid_Level*Cross_Sectional_Area = Liquid_Volume ;

# hydrostatic pressure
Press = Atm_Press + Density*Grav_Acceleration*Liquid_Level ;

# flow/pressure relationship
Flow_Out = Loss_Coefficient*(Press - Atm_Press)^0.5 ;

END # model buffer_tank

 EQUATION BASED LANGUAGES

Advantages:
declarative model is decoupled from mechanics of
solution. No assumption concerning which variables
are inputs is made during model coding.
incidence information and Jacobian matrix can be
generated automatically using symbolic
differentiation or automatic differentiation technology
(automatic, fast, and error free!)
all mathematical information is retained computer
stores complete functional form of model.
tools that check if a system of equations is
mathematically well posed can be applied
automatically (see appendix).
can embed automatically more sophisticated solution
strategies: e.g., combined discrete/continuous
simulation, dummy derivative method for high index
systems. FORTRAN based approach will just solve an
IVP with smooth inputs unless user adds code for
discontinuity handling herself.
Disadvantages:
although high flexibility, advances are still needed to
express more general functional forms.
user must derive equations => expert users! - but,
user only has to focus on the physics and chemistry.
model documentation only as comments.
 PHENOMENOLOGICAL LANGUAGES

This is really only in the realm of (not very mature)

academic research. . .

Problem: the following paradox:

(a) deriving models from first principles in terms of
equations is a difficult and skilled task requiring
experts - error prone!
(b) it is difficult (or even impossible) to standardize
dynamic models3 each study of a new system will
require some model development activity.4
Idea: can we develop dynamic models in terms of higher
level concepts than equations?
develop a phenomena based modelling methodology
codify this as a software tool
and then generate the equations automatically from
this higher level description

3 in contrast to steady-state simulation where by and large (reactors being an exception)

standardization has been achieved.
4 this is, in my opinion, a major reason why the equation based langauges have become

so popular
 PHENOMENOLOGICAL LANGUAGES

The advantages claimed of such an approach are:

the engineer can work directly with the principles of
physics, chemistry and biology rather than equations
this is perceived as "easier" (but a complete set of
concepts is required and the user must learn the
phenomena based modelling methodology).
guarantees can be made that the model is correct for
the assumptions made (?)
assumptions, domains of validity, etc., can be
generated and documented automatically
qualitative information forms part of the model.

Several ideas have emerged in the recent literature, for

example:
(a) specify the phenomena of interest => balance
equations required; specify the assumption => terms
in these balance equations (Stephanopoulos et al.,
1990a; 1990b)
(b) specify a set of thermodynamic phases exchanging
mass, energy and/or momentum. Connections
between phases are characterized by the transport
phenomena leading to transfer of mass, energy
and/or momentum (Perkins et al., 1994; Marquardt
1994)

. . . research continues.
 DYNAMIC SIMULATION TOOLS
The software tools currently available roughly fall into the
following categories:
(a) subroutine libraries coded in a general-purpose
programming language (typically FORTRAN):
models coded as subroutines - some library models
series of subroutines to conduct numerical
integration, coordinate models, report results, etc.
at least use a standard numerical integration code
(Perkins, 1986) e.g., Harwell Subroutine Library,
IMSL, ODEPACK, DASSL, DSL48S.
Examples: DYFLO (Franks, 1972), DAEPACK (Tolsma
and Barton, 2000).
(b) continuous system simulation languages (CSSLs)
typically derived from the CSSL'67 (Strauss, 1967)
standard:
equation based simulation language
extensively used by control engineers for many
years.
only solve models in state space form (i.e. ODEs but
not DAEs)
not really suitable for large-scale simulations.
no chemical engineering specific concepts (e.g.,
multi-component streams, physical properties, unit
operation libraries).
Examples: ACSL, SIMUSOLV, SIMULINK (Matlab),
etc.
 DYNAMIC SIMULATION TOOLS

(c) operator training tools typically vendor will

develop plant model, simulation and interface.

(d) dynamic flowsheeting packages5 e.g., dynamic

simulation tools specifically designed for chemical
process simulation:
some have equation based simulation languages
facilities to solve large-scale differential-algebraic
models.
support flowsheet paradigm (e.g., concept of
streams, unit operations, etc.)
physical properties, unit operation, model libraries,
etc.
Examples: ACM, gPROMS, ABACUSS II, DIVA,
HYSYS

Most of the simulation tools above use simultaneous

numerical solution methods e.g., the differential
equations are assembled together and solved
simultaneously with a single algorithm.

However, in the development of dynamic flowsheeting

packages two approaches have emerged: modular and
direct. These are introduced in the following pages.

5see Marquardt (1991) for a comprehensive list of software tools and their current
availability.
 MODULAR INTEGRATORS
Effectively, this is an attempt to extend the sequential
modular approach to steady-state process simulation to
dynamic simulation.
Each unit operation model (black box model):
is coded with its own numerical integration routine
(any suitable method)
can solve for the time variation of its outputs given the
time variation of its inputs:

So, given the input trajectories each unit model can be

integrated independently over a time horizon TH (the
numerical algorithms embedded may take several steps in
this time horizon).
Advantage: can tailor algorithm to a particular unit
operation - for example, rapid transients may be localized
to a single unit operation.

Problem: must coordinate solution of individual models if

there is feedback of material or information in the
flowsheet - e.g. inputs of some unit operations not known
in advance. Need to develop a coordinator algorithm.
 MODULAR INTEGRATORS

Solution: tearing concept - guess the time variation of the

recycle streams and iterate until the function guessed
matches the function calculated.

Problem: find x(t), t
[0, TH] such that:

x(t) = g(x(t), t)
t [0, TH]

Note that in dynamic simulation we are solving for a

function rather than a single value. Severe problems with:
integration error control
stability
converging recycles (choice of norm to compare
functions)
extrapolation of variable trajectories
choice of time horizon
treatment of discontinuities
computational expense of co-ordination
Examples: GEPURS, FLOWPACK II, PROSIM
 DIRECT INTEGRATORS

Examples: SPEEDUP, DIVA, DYNSIM, ASCEND,

gPROMS, ABACUSS, POLYRED:
all equations assembled and integrated
simultaneously with a single algorithm
guarantees for stability, and integration error can be
controlled automatically theoretical guarantees on
the accuracy of solution (
errors = modelling errors)
efficient general-purpose codes for large-scale
problems exist
can use equation based modelling language
in practice, more efficient for accurate simulation of
processing systems!
 MATHEMATICAL FORMULATION OF
DYNAMIC SIMULATION PROBLEMS

A simulation tool will be coded to support a specific

mathematical formulation for dynamic simulation
experiments. In principle, the most general
formulation is desirable. However, from a practical
point of view we must work within the constraints of
current technology.

Assuming we only have time as an independent

variable, a dynamic model together with a choice of
input variables will typically give rise to one of the
following (similar) mathematical problems:

(A) systems of explicit first order ordinary differential

equations (ODEs):

x = f ( x,t ) (1)

which is the mathematical formulation supported by

the CSSL-type simulation languages.

More specifically, it is possible to introduce auxiliary

algebraic equations of the form:

y1 = g1 (x)
y 2 = g2 (x, y1 )
 (2)

y m = gm (x, y1
y m
1 )
 MATHEMATICAL FORMULATION OF
DYNAMIC SIMULATION PROBLEMS

In words: algebraic equations are acceptable provided

a suitable calculation sequence can be found so that
each algebraic variable can be determined explicitly
given x and any previously calculated y i - i.e. the
algebraic variables can be eliminated by direct
substitution of the algebraic equations into the
ordinary differential equations.
Note that most of the models derived in section I do
not conform to this formulation without some form of
manipulation (in general, differentiation with respect
to time).

(B) systems of differential-algebraic equations (DAEs

- coupled systems of ordinary differential and
algebraic equations):

F ( z,z,t ) = 0 (3)

where z
R , F : Rn
Rn
R
Rn this general

n

formulation is known as the fully implicit form.

Note: what distinguishes (3) from a system of implicit

ODEs (IODE) is the fact that the Jacobian of F with
respect to z is singular everywhere. In other words,
equations (3) cannot be solved for z given z and t .
 MATHEMATICAL FORMULATION OF
DYNAMIC SIMULATION PROBLEMS

Chemical engineering models frequently give rise to

the semi-explicit form of differential-algebraic
equations, which provides useful insights into the
structure of DAEs:

( )
f x , x, y,t = 0
(4)
g(x, y,t) = 0

where we can categorize:

x
Rn (differential variables)

y
Rm (algebraic variables)
x
Rn (time derivatives of the
differential variables)
f : Rn
Rn
Rm
R
Rn (differential equations)
g : Rn
Rm
R
Rm (algebraic equations)
...so we have (n + m) unknowns and (n + m)
equations.

Note that (4) is just a specialization of (3): z would be

just a n + m vector composed of the components of x
and y .
The Jacobian of equations (4) with respect to the time
derivatives of the unknowns is clearly rank deficient
everywhere ( y do not appear anywhere).
 MATHEMATICAL FORMULATION OF
DYNAMIC SIMULATION PROBLEMS

In the literature some authors advocate that two other

(more restricted) forms are appropriate for chemical
engineering applications, the linear semi-explicit form:

(
x = f x, y,t )
( )
g x, y,t = 0
(5)

or the linear implicit form:

B( x,t ) x = f ( x,t ) (6)

These arguments seem rather pointless because

reliable general-purpose codes for the direct solution
of the general formulation (3) are widely available (if
index = 1). Further, formulation (5) is clearly
unsuitable for some of the models derived in section I
of these notes.

However, theoretical treatment of both (5) and (6) is

easier.
 WHY USE DAEs?

(a) it is the most natural form in which to derive the

model e.g. modelling of physico-chemical
phenomena in chemical processes gives rise to
DAEs (see section I).
Advantage: user can just submit model directly
to the computer without further manipulation.
(b) derivation of ODEs requires (extremely) tedious
differentiation and algebraic manipulation of the
equations. To illustrate this, consider the semi-
explicit form (4):
f ( x , x, y,t) = 0 (7)

g(x, y,t) = 0 (8)

Take the total differential and divide by dt to
yield the time derivatives of these equations:

f =
f x +
f x +
f y +
f = 0

x
x
y
t (9)

g
g
g
g = x + y + =0 (10)

x
y
t


f 

x 0 

g
g
and provided that the matrix:  

x
y 
 WHY USE DAEs?
is nonsingular, then (7) and (10) can be solved
simultaneously for x and y given x , y and t .
Hence the system (7) and (10) is a system of
implicit ODEs (i.e. the relevant Jacobian is
nonsingular).
For example, consider just the energy balance
and enthalpy relations in our CSTR model:
dH
= FIN IN hIN
FOUT OUT hOUT + Q
dt
H = MhOUT
hOUT = h(T,CA ,CB )
...and derive ODEs by differentiation of the
algebraic equations:
dH dM dh
= hOUT + M OUT
dt dt dt
dhOUT 
h  dT 
h  dCA 
h  dCB
= +  + 
dt 
T  C ,C dt 
CA  
CB  T,C dt
A B T,C dt
B A

these are not yet in state space form in

general, algebraic manipulation may not be
able to yield the explicit ODE form.
total differentials of physical property models
are particularly tedious to derive.
In general, derivation of ODEs from DAEs
requires one (or more) differentiations of the
equations (or subsets) with respect to time.
 WHY USE DAEs?
(c) the solution set of the ODEs derived by
differentiation is larger than the solution set of
the original DAEs the solution to the
derivative of an equation is the original equation
plus an arbitrary constant.
Example: the numerical integration of the ODEs
above will treat the enthalpy holdup and
temperature as independent variables
however, we know they are not independent but
related by the physical property model.
Typically, we can ensure the initial conditions
(e.g. H(t 0 ) and T(t0 ) ) are consistent but we
cannot avoid numerical errors (i.e. under
discretization) building up during integration
that will lead to H(t) and T(t) at some later time
that are not consistent with the physical property
model. This may, or may not, be disastrous (see
below).

Direct integration of the DAEs ensures that the

algebraic equations are satisfied at each time step
hence the algebraic variables will remain
consistent with the differential variables.
 WHY USE DAEs?
(d) the numerical integration of a minimal set of
independent variables (e.g. the differential
variables in a DAE system) is more consistent
with the results of classical thermodynamics - e.g.
only a certain minimal set of independent
quantities is required to fully define the 'state' of
a simple system (Modell and Reid, 1983).
 DEGREE OF FREEDOM ANALYSIS

Process models are typically under determined systems

of equations - there are more variables describing the
behaviour of the system than equations relating them.
Degree of freedom analysis is concerned with
identifying a subset of variables to assign values to (or
specify), so that the equations become fully determined
in the remaining unknown (or dependent) variables -
and thus can be solved to determine a set of unique
values for these unknowns.
A necessary condition that a system of NEQ equations
is fully determined in the NUK unknown variables is
that:
NEQ
NUK
So, clearly, we must at least specify values for a
sufficient number of variables so that the number of
unknowns becomes equal to the number of equations.
If the situation arises where so many specifications
have been made that there are more equations than
unknowns (i.e. NEQ > NUK ), the extra equations are
either:
redundant - i.e. they add no extra information, so
they do not constrain the solution further
(example: adding an overall mass balance to a
flowsheet model where we already have mass
balances for each unit)
inconsistent - i.e. there is no solution that will
satisfy all the equations.
 DEGREE OF FREEDOM ANALYSIS

Bearing this in mind, we will only accept a model for

solution if :
NEQ = NUK
...and degree of freedom analysis is concerned with
selecting a subset of variables to specify so that this
condition is satisfied.

When setting up a dynamic simulation experiment

we have to worry about degree of freedom analysis
twice:
first derive a fully determined system of DAEs
from the under determined model by selecting a
set of input variables (in fact, we are assigning
known functions of time rather than values to the
input variables)
next (given a fully determined system of DAEs)
establish the degrees of freedom available to
specify the initial state of the system - e.g. the
number of side conditions required at t = t0 .

...details of each follow.

 INPUT SPECIFICATIONS - GENERAL
FORMULATION

Typically, chemical engineering models give rise to

under determined systems of DAEs we have
degrees of freedom to satisfy before a fully
determined system that can be submitted to the
computer.

Process model:

f ( x , x, y,u,t ) = 0

where:

x
R n (differential variables)
y
R m (algebraic variables)
u
R l (input variables)
f :Rn
R n
R m
R l
R
Rn+m

This is just a restatement of the general formulation

(3) above here, the set of variables has merely been
partitioned into differential, algebraic, and input
variables according to the definitions:
input - assigned a known function of time
differential - whose time derivative appears
algebraic - whose time derivative do not appear
Satisfying the degrees of freedom is merely selecting
which subset of l variables will belong to the vector u,
and assigning functions of time to these variables.
 INPUT SPECIFICATIONS - GENERAL
FORMULATION

This leads to the fully determined DAE system:

f ( x, x, y,u,t) = 0
u = u(t)

Alternatively, we could add extra equations to the

model, reducing the number of input specifications
required by one for each non redundant equation
added.
 SOLVING DYNAMIC SIMULATIONS

Solve initial value problem in differential-algebraic

equations in the interval [ t0 ,t f ]:

This requires solution of three sub problems:

1. Defining and solving for the initial state of the
system at t0 .
2. Numerical integration of the DAEs from t0 to t f .
3. Discontinuity handling for example, correct
processing of discontinuities in the functions
assigned to the input variables (also their time
derivatives):

Details of each follow. . .

DEFINING THE INITIAL STATE OF THE SYSTEM

Mathematically: defining the side conditions for the

initial value problem, and utilizing these to determine
initial values for the unknowns.

For explicit ODE systems the side conditions are

usually specified by providing initial values for x :

x(t0 ) = x 0

x (t 0 ) = f (x(t0 ),t0 )

In this case, calculating the initial state is trivial and

requires only a function evaluation - i.e. x(t0 )
independent.

However, for most dynamic simulation experiments a

more general approach is both desirable and possible.
 CONSISTENT INITIALIZATION

Consistent initialization: find a consistent set of initial

values for the variables x(t) and x (t) at t = t0 :

Denoted by: { x(t0 ), x (t0 )}

A set of consistent initial values defines one possible
initial state for the system.

Idea: in general, undertake a consistent initialization

calculation in which the model equations at t = t0 , e.g.:

x (t0 ) = f (x(t0 ), t0 )

are solved for the initial state of the system.

The initial values { x(t0 ), x (t0 )} will then be consistent

with the model equations.
 CONSISTENT INITIALIZATION

Example: many dynamic simulations involve starting

from the system from steady-state and then
introducing a disturbance (e.g. step change, ramp,
etc.).

In the ODE example, an initial state of steady-state is

equivalent to defining:

x (t0 ) = 0

...substitute this specification into the ODEs (system

at t = t0 ) to yield:

0 = f (x(t0 ),t0 )

and solve the resulting system of nonlinear algebraic

equations to find consistent initial values for x(t0 ) .

This assumes that the model has a steady-state

solution - not always the case, e.g. a batch reactor.

Note that this is entirely equivalent to a steady-state

simulation. In fact, steady-state simulation is really
only a subset of dynamic simulation: it is merely one
possible consistent initialization calculation.
 DEGREES OF FREEDOM FOR CONSISTENT
INITIALIZATION OF ODEs

For ODEs, degrees of freedom for the initialization

calculation are clear.

Given a fully determined ODE system:

n equations x (t0 ) = f (x(t0 ), t0 )

2n variables { x(t0 ), x (t0 )}

=> n degrees of freedom

Examples of possible specification sets for consistent

initialization:
a) { x(t0 )}
b) { x(t0 )}
c) or some mixture of values for the variables and
their time derivatives
 DEGREES OF FREEDOM FOR CONSISTENT
INITIALIZATION OF DAEs

In direct analogy, to define the initial state of a fully

determined system of DAEs we need to find a set of
consistent initial values for the variables x(t) , x (t) and
y(t ) at t = t0 , e.g.:

{ x(t0 ), x (t0 ), y(t0 )}

Clearly, a necessary condition for a set of consistent
initial values is that the model equations are satisfied
at t = t0 , e.g.:

f ( x(t0 ), x(t0 ), y(t0 ),u,t0 ) = 0

u = u(t0 )

Assuming that this necessary condition is also

sufficient (not always the case - see discussion of high
index problems), degree of freedom analysis gives:

n + m equations f ( x(t0 ), x(t0 ), y(t0 ),u(t0 ), t0 ) = 0

2n + m variables { x(0), x (0), y(0)}

=> n degrees of freedom

Note the degrees of freedom is equal to the number of

differential variables n .
 DEGREES OF FREEDOM FOR CONSISTENT
INITIALIZATION OF DAEs

So, we can choose a n membered subset of variables

out of { x(t0 ), x (t0 ), y(t0 )} to assign initial values to.
This is the formulation supported by SpeedUp.

These values (substituted into the model equations)

will fully define the initial state of the system.

Notes:
a) consistent initialization of DAEs requires the
solution of large sparse systems of nonlinear
algebraic equations from poor initial guesses - in
principle, we can use any technique already
developed for steady-state simulation.
b) as with steady-state simulation, not all subsets
of variables and not all choices of values lead to
a well-posed problem (functionally nonsingular
equations).
c) numerical integration of DAEs should always
be started from consistent initial values -
otherwise the results can be garbage and/or the
integration can fail. The numerical integration
algorithms most commonly used can generate a
solution from arbitrary (possibly inconsistent)
initial values - in fact, standard DAE solver
routines do not usually have facilities for the
consistent initialization calculations described
above! (see Kroner, et al., 1992)
 DEGREES OF FREEDOM FOR CONSISTENT
INITIALIZATION OF DAEs

d) we will also discuss the situation when the

model equations are not necessary and
sufficient to define consistent initial values - in
general, we have at most n degrees of freedom
to specify the initial state of the system.
 CONSISTENT INITIALIZATION - EXAMPLES

Consider possible initial states for our well stirred reactor

system:

a) operate the reactor as a CSTR and start the simulation

form steady-state at t = 0 .

Define steady-state by setting the time derivatives of

all the differential variables to zero (n specifications):

M (0) = 0, N A (0) = 0, N B (0) = 0, H (0) = 0

...and solve the model equations for x(0) and y(0) . In

this case, the equations need to be solved
simultaneously.

b) operate the reactor as a batch reactor - define the initial

charge to the vessel - e.g. at t = 0 :

M(0) = 1000, N A (0) = 100, N B (0) = 0, H(0) = 1E7

so, given x(0) solve equations x (0) and y(0) . Note that
the equations are much easier to solve for this choice of
initial condition!
 CONSISTENT INITIALIZATION - EXAMPLES

c) for a batch reactor, we typically do not know the

enthalpy of a charge - after all this is a value related to
some arbitrary zero energy reference state.

For example, a more realistic operating policy would

specify the charge as 1m3 of solvent with a specified
concentration of A at 300K:

V(0) = 1,C A (0) = 1,C B (0) = 0,T(0) = 300

and we solve the equations for the remaining variables

- x(0) , x (0) and remaining y(0) .

Note that it is often most convenient to specify the

initial condition in terms of values for a subset of the
algebraic variables - hence the immense flexibility of
this approach.

Exercise: why is the specification of:

{V (0), N A (0),CA (0),T(0)}

invalid?
 CONSISTENT INITIALIZATION - GENERAL
FORMULATION

Example: consider a batch distillation column:

Qc

L D

Species: A, B , C

Qr

Typical operating policy:

a) charge the reboiler with mixture of A, B and C

b) apply steam to reboiler and bring to steady-state at

total reflux (D(t) = 0)

c) draw cut rich in A overhead D(t) > 0

d) draw cut rich in B overhead D(t) > 0

Simulation experiment: define initial state as steady-state

at total reflux and simulate the sequence of cuts.

Note: the model required to simulate start-up is much

more complex - neglect this period.
 CONSISTENT INITIALIZATION - GENERAL
FORMULATION

How to define steady-state at total reflux?

a) if we specify steady-state: x (0) = 0
badly posed

problem - no way to determine the initial extent of
material in the reboiler.

b) define the initial condition as:

NTray
N T ,i = N R,i (0) +
N j,i (0)
i = 1 NC
j =1

N i, j (0) = 0
j = 1 NTray

i = 1 NC

where

N T ,i = amount of species i in charge (known)

N R,i = holdup of species i in reboiler

N j ,i = holdup of species i on stage j

i.e. all trays at steady-state, holdup in reboiler

determined by an overall mass balance on column.

Note that this time we have defined an initial condition in

terms of a set of equations rather than an assignment of
values.
 CONSISTENT INITIALIZATION - GENERAL
FORMULATION

The general formulation for consistent initialization

requires the solution of the following system of nonlinear
algebraic equations for the unknowns {x (t 0 ),x( t0 ),y( t0 )}:

f (x (t 0 ), x(t 0 ), y(t 0 ),u(t0 ),t0 ) = 0

c( x (t0 ),x(t0 ),y(t0 ),u(t 0 ),t 0 ) = 0
where:

c : R n
R n
R m
R l
R
Rn

and the functions c may be value assignments or general

functional relationships.

Again, this system of equations must be functionally

nonsingular.

Remarks:

a) we are again faced with the problem of solving

a large set of nonlinear algebraic equations
from poor initial guesses.

b) two competing technologies:

i) sequential modular
ii) equation oriented - assemble all equations and
solve simultaneously with general-purpose solver.
 CONSISTENT INITIALIZATION - GENERAL
FORMULATION

c) the sequential modular approach is superior for

finding steady-state - but what about other initial
states? - little flexibility to define the initial state of a
system.

d) however, direct (equation-oriented) methods superior

for numerical integration of DAEs.

e) unresolved research question, two approaches:

i) improve robustness of equation oriented technology

- ongoing problem: solve large systems of nonlinear
equations from poor initial guesses.

ii) interface sequential modular initial condition to

direct integrator - but problem with flexibility
remains.
 CONSISTENT INITIALIZATION AND
HIGH INDEX DAE SYSTEMS

Some DAE systems pose fundamental mathematical

difficulties.

Example: total condenser for a pure component vapour

(e.g. a steam jacket):

h IN
VAPOUR
F IN
V Q
N (t)
(vapour)

L
(liquid) h OUT

Assumptions:

(i) all liquid formed is immediately removed

from the system (e.g., N L = 0 )
(ii) no sub cooling occurs
(iii) system reaches vapour-liquid equilibrium
instantaneously
(iv) ideal physical properties
 NOMENCLATURE

FIN total molar flow rate to vessel [mol s-1]

(vapour)
hIN molar enthalpy of feed stream [J mol-1]
hOUT molar enthalpy of liquid stream [J mol-1]
L total molar flow rate of liquid [mol s-1]
from vessel
NV number of moles of vapour in [mol]
vessel
P pressure of vessel [Nm-2]
Q cooling load [J s-1]
R gas constant [J mol-1 K-1]
T temperature of vessel contents [K]
U internal energy of gas in vessel [J]
V volume of vessel (constant) [m-3]
 CONSISTENT INITIALIZATION AND
HIGH INDEX DAE SYSTEMS

Material Balance:

dN V
= FIN
L
dt

Energy Balance:

dU
= FIN hIN
LhOUT
Q
dt

Physical Properties:

P=P
SAT
(T ) (vapour-liquid equilibrium)

PV = N V RT (equation of state)

hOUT = h L (T )

N h (T) = U + PV
V V
 CONSISTENT INITIALIZATION AND
HIGH INDEX DAE SYSTEMS

Simulation experiment:

(a) input set R,V ,Fin (t),hin (t),Q(t)

(b) initial condition at t0 = 0
V
solve model for: N (t),T(t), L(t), P(t),hOUT (t),U(t)

Degrees of freedom to specify initial condition:

6 equations
{N V
(0), V (0),U(0),U (0),T (0),
N
8 unknowns
L(0), P(0),hOUT (0)}

2 degrees of freedom for consistent

initialization (?)
V
Try to specify the initial conditions N (0),U(0) :
V
1. given N (0) solve:

P(0)V = N (0)RT(0)
V

P(0) = P SAT (T (0))

for P(0),T(0) .

2. hOUT (0) = h L (T(0))

 CONSISTENT INITIALIZATION AND
HIGH INDEX DAE SYSTEMS

3. this leaves 3 equations:

N V (0) = FIN
L(0)

U (0) = FIN hIN
L(0)hOUT (0)
Q
N V (0)h V (T(0)) = U(0) + P(0)V
V
in the unknowns L(0), N (0),U (0) .

BUT: equations are structurally singular - none of

the unknowns appears in the third equation, e.g. the
incidence matrix is singular for all values of the
unknowns:

N
V (0) U(0)

L(0)
f1  

0

f2  
 0


f3  0 0 0 

badly posed problem.
V
This implies that the differential variables N and U are
not independent. This is a:

high index problem

 CONSISTENT INITIALIZATION AND
HIGH INDEX DAE SYSTEMS

In this case, the problem is a consequence of our

modelling assumptions: from classical thermodynamics
(Duhems theorem) we know that NC + 2 (3 in this case)
independently variable properties fully define that state of
a system at equilibrium (we have assumed the system is
always at equilibrium). We have specified:

(a) the volume of the system

(b) the vapour fraction of the system (=1).
V
Therefore, we can only specify one out of N (0) and U(0)
to fully define the state of the system.

Remarks:
V
(a) if we can only specify one out of N (0) and
U(0) , where do we get an (additional)
specification or equation to satisfy the degrees
of freedom for consistent initialization?

(b) in this case, the high index problem is a

consequence of our modelling assumptions (e.g.
N = 0 ) - a fundamental property of the model.
L

We could relax this assumption, and add a

relationship relating the liquid holdup to the
liquid flow rate.

However, we frequently want to make assumptions like

this due to a lack of data, to reduce stiffness of the model,
for simplicity, etc.
 CONSISTENT INITIALIZATION AND
HIGH INDEX DAE SYSTEMS

Example: dynamic design problem. Consider an

exothermic fixed volume CSTR:

Reaction
A>B

CIN
C
TIN T

COOLING
MEDIUM
TJ

Mass Balance:

dC 1
= (CIN
C )
r
dt 

Energy Balance:

dT 1 
H R   UA 
= (TIN
T ) + 
dt  C r
 C V (T
TJ )
 p
 p

Rate equation:

r = ke
E / RT C
 NOMENCLATURE

A heat transfer area of heating coil [m2]

C concentration of A in reactor [mol m-3]
CIN feed concentration of A [mol m-3]
CP heat capacity [J mol-1 K-1]
K pre-exponential factor [s-1]
T temperature in reactor [K]
TIN feed temperature [K]
TJ jacket temperature [K]
U heat transfer coefficient [J m-2 K-1 s-1]
V reactor volume [m2]

H R heat of reaction [J mol-1]

E activation energy [J mol-1]

density of reactor contents [mol m-3]

residence time [s]
 CONSISTENT INITIALIZATION AND
HIGH INDEX DAE SYSTEMS

1. Simulation Problem

A,C p ,
E,
H R ,k,U, V, , R, (t),

Given:
CIN (t),TIn (t),TJ (t)

Determine: C(t ),T(t ), r(t) .

Straightforward: specify initial conditions C(0),T(0) and

solve equations.

no high index problem

We can assume from this that any high index problems

are not a fundamental property of this model.

2. Dynamic Design Problem (or trajectory tracing

problem): determine variation in coolant temperature
TJ (t) necessary to achieve a specified outlet concentration
profile, e.g.:

C(t)

t
 CONSISTENT INITIALIZATION AND
HIGH INDEX DAE SYSTEMS

Simulation experiment: dynamic design problem:

A,C p ,
E,
H R , k,U,V, , R,

Given:
 (t),C IN (t),TIn (t),C(t)

Determine: T(t),TJ (t),r(t)

Degrees of freedom for consistent initialization:

3 equations
4 unknowns {T (0),TJ (0),r(0), T (0)}

1 degree of freedom for consistent initialization

( C (0) derived by differentiating C(t ) )

Try: C(0) or T(0) :

1. obviously we cannot specify an arbitrary initial value

for C(0) already specified.
2. can we specify an arbitrary initial value for T(0) ?

Substitute rate equation into mass balance:

1
E / RT (0)
C (0) = (C IN
C(0))
ke C(0)


BUT: C(0), C (0) fully specified

T(0) determined by this equations, so it cannot be

specified.
 CONSISTENT INITIALIZATION AND
HIGH INDEX DAE SYSTEMS

In this case, the high index problem arises through a

particular choice of input variables and is not a
fundamental property of the model.

In general, dynamic design problems will lead to high

index problems because you are effectively trying to
specify the time variation of a differential variable
hence it is no longer independent.
 HIGH INDEX PROBLEMS - REMARKS

(a) despite the difficulties with consistent initialization,

both problems have proper mathematical solutions.

(b) common features of high index problems:

(i) the differential variables are not independent

and therefore cannot be given arbitrary initial
values.

(ii) in many problems, a quick way to spot a high

index problem is the fact that the algebraic
equations cannot be solved for the algebraic
variables given the differential variables - e.g.
one or more algebraic variables only appear in
the differential equations (example: the liquid
flow rate from the total condenser).

(iii) it follows from (ii), that high index DAEs cannot

be reduced to ODEs by using the algebraic
equations to eliminate the algebraic variables

differentiation is required.

(c) two important issues with high index problems:

(i) where to get the additional equations to satisfy

the degrees of freedom for consistent
initialization?

(ii) controlling the error of integration in numerical

algorithms the differential variables are not
really independent, even though the
discretization treats them as such.
 CONSISTENT INITIALIZATION OF
HIGH INDEX SYSTEMS

We have seen that problems with consistent initial-ization

are associated with high index problems.

Definition: the index of a DAE system is defined as the

smallest non-negative integer I such that the DAE system
and its first I derivatives with respect to time define x and
y as locally unique functions of x, y,u (and its time
derivatives) and t (Brenan et al., 1989).

Although the above definition is more precise, it is, as is

the case with many mathematical definitions, not very
helpful from a practical point of view.

A more informal statement is that the index of a DAE

system is the maximum number of differentiations of the
original DAEs (or subsets of them) that must be
undertaken to derive an ODE system (implicit or explicit).

all ODE systems are by definition of index 0.

This definition is sometimes called the differential index.

Some authors use a perturbation index with a different
definition, which is equal to or one greater that the
differential index. In general, both indices are local
quantities.
 CONSISTENT INITIALIZATION OF
HIGH INDEX SYSTEMS

Example 1:

x1 = 2 x1 + x 2 + y (1)
x 2 = x1
2x 2
3y (2)
0 = x1 + x 2
y (3)

Consistent initialization at t = 0: specify x1 (0), x 2 (0) :

1. y(0) = x1 (0) + x 2 (0)

2. x1 (0) = 2 x1 (0) + x2 (0) + y(0)
3. x 2 (0) = x1 (0)
2 x2 (0)
3y(0)

Everything fine.

What is the index? Differentiate the algebraic equation

(3):

 0 = x1 + x 2
y
 y = 3x1
x2
2y (3)

Equations (1),(2),(3) form a system of three ODEs in

x1 , x 2 , y :

Index = 1
 CONSISTENT INITIALIZATION OF
HIGH INDEX SYSTEMS

Example 2:

x1 = 2 x1 + x 2 + y (1)
x 2 = x1
2x 2
3y (2)
0 = x1 + x 2 (3)

Note: the algebraic variable y does not appear in the

algebraic equation so we can expect trouble.

Consistent initialization: clearly x1 (0) and x 2 (0) not

independent.

What is the index? Differentiate the algebraic equation (3):

0 = x1 + x 2
 0 = 3x1
x 2
2 y (3)

Equations (1), (2), (3) are still DAEs! Differentiate (3) - to

derive second derivative of (3):

 0 = 3x1
x 2
2 y
5 5

y = x1 + x 2 + 3y (3)
2 2

Equations (1), (2), (3) are an ODE system in x1 , x 2 , y . We

have differentiated (3) twice to derive the ODEs:

Index = 2
 CONSISTENT INITIALIZATION OF
HIGH INDEX SYSTEMS

Each differentiation of (3) (i.e., a subset of equations)

reduces the index by one (1), (2), (3) is an index-1
system.

In our index 2 example, if we cannot specify x1 (0) and

x 2 (0) independently, can we try another specification set
of two variables?

Example: x1 (0) = 1, y(0) = 1:

1. x 2 (0) =
1
2. x1 (0) = 2
3. x 2 (0) = 0

which at first glance seems fine. However, from

differentiation of the algebraic equation (3), we also know
that:

0 = x1 (t) + x 2 (t) (3)

...and this relationship must be satisfied at t = 0.

Clearly the values calculated above are inconsistent with

this equation.

So, what are the necessary and sufficient conditions for a

set of consistent initial values?
 NECESSARY AND SUFFICIENT CONDITIONS
FOR CONSISTENT INITIALIZATION

Necessary and sufficient conditions for a consistent set of

initial values {x(t 0 ), x (t0 ),y(t0 )} are that they satisfy:

(a) the model equations at t = t0 - i.e.:

f (x (t 0 ), x(t 0 ), y(t 0 ),u(t0 ),) = 0

(b) the first and higher order time derivatives of the model
equations at t = t0

Note:

(i) we originally have 2n + m degrees of freedom.

(ii) the model equations take up n + m degrees of

freedom.

(iii) whether or not the derivatives of the equations take

up further degrees of freedom is closely related to the
index of the equations:

(a) if index
2, extra equations derived by

differentiation are always required.

(b) for most index 1 systems, the original equations

are sufficient.

For high index systems it is therefore necessary to

identify which subset of the derivatives take up
further degrees of freedom.
 NECESSARY AND SUFFICIENT CONDITIONS
FOR CONSISTENT INITIALIZATION

Obviously, for each extra equation required we have to

make one less specification on the initial state of the
system. Revisiting our examples:

Example 1:

x1 = 2 x1 + x 2 + y
x 2 = x1
2x 2
3y
0 = x1 + x 2
y

differentiate:

x1 = 2 x1 + x 2 + y
x2 = x1
2 x 2
3 y
0 = x1 + x 2
y

we have introduced 3 new equations and three new

variables ( x1 , x2 , y ) these equations do not constrain our
original variables further.

we have 2 degrees of freedom to specify the

initial condition.
 NECESSARY AND SUFFICIENT CONDITIONS
FOR CONSISTENT INITIALIZATION

Example 2:

x1 = 2 x1 + x 2 + y
x 2 = x1
2x 2
3y
0 = x1 + x 2

differentiate:

x1 = 2 x1 + x 2 + y
x2 = x1
2 x 2
3 y
0 = x1 + x 2

Again, we have introduced three new equations and three

new variables ( x1 , x2 , y ). BUT the third equation does not
involve these variables, it only involves our original
variables and therefore must constrain them further.
Further differentiation yields no further constraints on our
original variables.

Therefore, the necessary and sufficient conditions for

consistent initial values in this example are:

x1 (0) = 2 x1 (0) + x2 (0) + y(0)

x 2 (0) = x1 (0)
2 x2 (0) + 3y(0)
0 = x1 (0) + x 2 (0)
0 = x1 (0) + x 2 (0)

This yields 4 equations in the 5 unknowns, so we have one

degree of freedom to specify the initial condition.
 NECESSARY AND SUFFICIENT CONDITIONS
FOR CONSISTENT INITIALIZATION

Exercise: perform this analysis for the CSTR dynamic

design problem in fact, in this case we have zero
degrees of freedom to specify the initial condition!
 CONSISTENT INITIALIZATION - REMARKS

(a) in general, the degrees of freedom for consistent

initialization is equal to or less than the number of
differential variables.

(b) for large systems of DAEs it is difficult to determine by

inspection which (if any) of the equations must be
differentiated with respect to time to derive the
necessary and sufficient conditions.

An algorithm that uses the structure of the equations

(the incidence matrix) to identify the minimal set of
equations that must be differentiated has been
developed (Pantelides, 1988a). For most practical
chemical engineering problems, the necessary and
sufficient conditions can be derived in this manner
from structural information alone.

However, there are situations in which structural

information is insufficient. For example:

x1 = x1
x 2 + y (1)
x 2 = x1 + x 2
y (2)
x1 + x2 = 1 (3)

the structural algorithm will differentiate (3) to yield:

x1 = x1
x 2 + y (1)
x 2 = x1 + x 2
y (2)
x1 + x 2 = 0 (3)

and terminate.
 CONSISTENT INITIALIZATION - REMARKS

The algorithm terminates because these equations are

structurally nonsingular with respect to { x1 , x 2 , y }.
However, they are functionally singular with respect to
this set a further differentiation is actually required
to derive the necessary and sufficient conditions.

(c) A sufficient condition for the original equations to be

necessary and sufficient to define consistent initial
values is:

Rank fx[ ]
fy = n + m

i.e. the Jacobian of the equations with respect to x and

y is nonsingular.

Practical Consequence: if we can solve the original

equations for x (0) and y(0) given x(0) , then we do not
need to differentiate the original equations.

(d) most dynamic simulators1 are therefore limited to

those DAEs that satisfy the condition above. ABACUSS
(Feehery and Barton, 1996) uses a combination of the
structural algorithm and automated symbolic
differentiation of the equations to find consistent initial
conditions for a much broader class of DAE systems.

1
i.e. any simulator apart from ABACUSS
 CONSISTENT INITIALIZATION - REMARKS

(e) the condition above:

[
Rank fx ]
fy = n + m

is also a sufficient (but not necessary) condition for the

DAEs to be index 1. Not all index 1 DAEs satisfy this
condition: for example:

x + 2 x2 = 1 (1)
x1 + x2 = 1 (2)

Clearly x1 (0) and x 2 (0) are not independent one

must derive an extra equation to define a consistent
initial condition. However, the index is still 1 by the
rigorous definition at the beginning of this section:
differentiate the algebraic equations:

x1 + x 2 = 0 (2)

and (1) and (2) yield the ODEs:

x1 =
1
x 2 = 1

One differentiation
Index = 1.

 CONSISTENT INITIALIZATION - REMARKS

Special index 1 problems of the above form arise most

frequently when an index 2 problem has been
transformed into an index 1 problem by eliminating an
algebraic variable that only appears in the differential
equation.

Observe that this transformation does not eliminate the

problem of finding a consistent initial condition!

(f) there are claims in the literature concerning a necessary

and sufficient condition for an index 1 system
(Lefkopoulos and Stadtherr, 1993). These are incorrect:
the condition in question is also only sufficient.

(g) in many cases, high index problems arise due to errors

in model formulation. Typically, the model
formulation implies that a greater number of quantities
are independent than is actually the case. An index 1
model for the system can be derived through
identification of the true number of independent
quantities (Ponton and Gawthrop, 1991).

Clearly it is extremely important for a simulator to

automatically identify these errors, and inform the user
of alternatives.
 CONSISTENT INITIALIZATION - REMARKS

(h) it is sometimes possible to trick a simulator into

finding consistent initial values for a high index
problem by specifying steady-state as the initial
condition. In these cases, the specification of steady-
state is redundant with the additional equations
derived through differentiation, rather than
inconsistent. Therefore, solving the equations with
x (0) = 0 will give the same answer.

There are several examples in the literature of attempts

to solve high index DAEs in this manner it works
provided no discontinuities in the inputs are
introduced (consistent initialization is required at
discontinuities as well - see later).

However, there are also severe problems with the

numerical integration of DAEs with index
2 (see
below) the results of attempting to do this are
invariably failure or garbage!
 NUMERICAL INTEGRATION OF DAEs

Dynamic simulation of chemical processes requires

the numerical solution of large DAE systems that are:

(a) large typically 100s - 10,000s of equations (or

even 100,000s!)

(b) sparse only a few (~5) variables appear in

each equation - special advantage can be taken
in Jacobian evaluations and linear algebra.

(c) stiff a range of phenomena taking place on a

wide range of time scales so the equations
have widely different time constants.

Numerical methods for integration of differential

equations are usually categorized as explicit or
implicit.

Implicit algorithms are used in most cases . . . details

to follow.

See the recommended text books for a comprehensive

treatment of this subject: different integration
algorithms; properties of algorithms: convergence,
consistency, stability (of the solution process, not the
system of equations); integration error control; etc.
 NUMERICAL INTEGRATION

Numerical integration methods approximate the

solution of differential equations by obtaining the
values of the unknowns at a set of distinct points in
time (a process known as discretization):

Distinct time points: t( i)
i = 1... N

Variables at t (i) : x (i )
Derivatives at t (i) : x (i )
The values obtained must satisfy the differential
equations at these points in time:

x ( i ) = f ( x ( i ) ,t ( i ) )
This is a set of n algebraic equations in the 2n
{ }
unknowns x (i ) , x (i ) .

=> we need another n equations to solve for the

unknowns. Different numerical methods suggest
different forms of these n extra equations.
 EXPLICIT METHODS

There are two main objections to explicit methods for

solving DAEs:

(a) it is not any cheaper (in computational effort) to

solve DAEs with explicit methods (vs. implicit
methods).

With ODEs, one advantage of an explicit method

is that it is unnecessary to solve a set of
nonlinear algebraic equations at each time step.

Example: Explicit Euler Method for ODEs uses

the approximation:

x (i+1) = x (i) + x (i) (t (i+1)
t (i) )

This leads to the algorithm:

 EXPLICIT METHODS

for Explicit Euler each time step only requires a

function evaluation of the model to determine
x (i )

For a system of DAEs the algorithm would be:

Note: must be able to solve f ( x , x, y,t) = 0 given x(t ) ;

explicit methods limited to index one problems that
satisfy this condition.

explicit methods for the general DAE

formulation1 require iterative solution of
nonlinear equations.

1
and indeed ODEs in implicit form as well (remember that an ODE in implicit form is the typical
result of trying to derive an ODE from a DAE).
 EXPLICIT METHODS

(b) process models give rise to stiff sets of

differential equations.

For linear time invariant ODEs, i.e.:

x
= Ax

stiffness is characterized by the Stiffness Ratio:

max Re
i
i
SR  1
min Re
i
i

where
i = ith eigenvalue of A .

i.e. the stiffness ratio is the ratio of the

eigenvalues of with the largest and smallest
real components.

Process models tend to have high stiffness

ratios (>1000) i.e., there are phenomena taking
place with widely different time constants.
 EXPLICIT METHODS

Example: jacketed bath with thermowell:

Normalized model:

dT
= 1
T
d  T  
1 0   T  1
 = +
dT M 1 T M
1 
1 
T M
 0

= (T
T M )
d 
The time constant
of the thermowell is related to
ratio of the thermal inertias of the bath and the
thermowell i.e., the smaller the thermowell, the
smaller the time constant
(
<<1) :

1  1
SR = =

1 
...as
the thermowell becomes smaller relative to the
bath, the stiffness ratio increases.
 EXPLICIT METHODS

The maximum time step an explicit method can take

is usually limited by stability criteria i.e., the step
size must always be smaller than some threshold for
the solution process to remain stable.

Note:

the time domain of interest is determined by the

slowest decaying component that with the
largest time constant (smallest eigenvalue).

but, the maximum step size is limited by the fastest

decaying component (largest eigenvalue) to
maintain stability.

As the stiffness ratio increases, the maximum step size

decreases, but the time domain of interest remains
fixed
an excessive number of time steps required.

Because stability criteria severely limit the step size in

stiff problems, explicit methods are not applied very
often to process simulation problems.
 EXPLICIT METHODS

Note: several textbooks that discuss dynamic

simulation of chemical processes advocate the use of
explicit methods. These books were either written
before reliable methods for solving stiff systems of
ODEs had been developed, or recommend that the
engineer determine the stiffness of the problem and
select an appropriate numerical method. In the latter
case, it seems safer to always use an implicit method
because most problems are stiff, and non-stiff
problems are trivial to solve with today's
computational resources.
 IMPLICIT METHODS

Implicit methods: always required to solve a set of

nonlinear equations at each time step.

Advantage: theoretical guarantees for stability

regardless of step size step size limited by the need
to control truncation error rather than stability.

=> we can take much larger time steps when the

problem is stiff and the fast components of the
solution have decayed away.

Example: Implicit Euler method for ODEs:

x (i+1) = x (i) + x (i+1) (t (i+1)
t (i) )

...and substitute:

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

into this to yield Implicit Euler algorithm.

 IMPLICIT METHODS

Predictor-Corrector methods:

the numerical methods used to solve the set of

nonlinear equations at each step require a good
initial guess.
use an explicit method to provide this guess
(predictor step)
then solve nonlinear equations (corrector step)

Example:

Predictor: Explicit Euler step

Corrector: Implicit Euler step using results of
predictor step as initial guess.
 IMPLICIT METHODS FOR DAEs

Most codes for solving DAEs use linear multi-step

methods - in particular, the Backward Differentiation
Formula (BDF) family of methods (also known as
Gear's method) (Gear, 1971).

These methods have particularly good stability

properties and hence are well suited to stiff problems.

Idea: fit qth order polynomial to q+1 values of x :

x(t) = c 0 + c1t + c 2t 2 ++cq t q

d x(t (i)
) q
1
x (i) = = c1 + 2c 2 t (i) ++qc qt (i)
dt

Given q+1 values for t (i) and x (i)

{t (i) , x (i) },{t (i
1) , x(i
1) },,{t (i
q) , x (i
q)}

we can derive the coefficients for these polynomials in
terms of x (i  j)
j = 0q .
 IMPLICIT METHODS FOR DAEs

Therefore, by differentiating the polynomial, the time

derivatives can be expressed as a linear combination
of these values:
q
1
(i)
x =
h (i)   j x (i
j)
j=0

where:

h(i) = current step size

j = scalar multipliers; a function of the order of
integration q and the current and previous
step sizes.

So, we have established a relationship between x (i)

and x (i) and q previous values of the unknowns
(presumed known).

This relationship can be solved simultaneously with

the differential equations to determine x (i) and x (i) .

Note that the order of the integration method is given

by the order of the polynomial q - the number of
previous values used to derive the approximation to
x (i) .

Exercise: show that the first order Gear method is

equivalent to the Implicit Euler method.
 IMPLICIT METHODS FOR DAEs

Gear's method is a predictor corrector method:

a) predictor: extrapolate polynomial derived at

previous time step.
b) corrector: solve polynomial and model equations
iteratively.

Two approaches to solving DAEs with implicit

methods have been proposed:

(A) reduce the DAE system to an explicit ODE

system and solve using a standard BDF code for
ODEs:
 IMPLICIT METHODS FOR DAEs

Objections:

a) VERY expensive - at each step of the ODE

method:

each polynomial approximation must be

solved iteratively.
each function evaluation of this iteration
requires the nested iterative solution of the
DAEs

b) each nested solution to the DAEs is inexact

and subject to numerical noise.

c) again, can only be applied to those index one

DAEs for which we can solve f ( x , x, y,t ) = 0
given x(t ) .
 IMPLICIT METHODS FOR DAEs

(B) Solve the DAE system directly - eliminate x (t)

from f ( x , x, y,t ) = 0 using the polynomial
approximation to yield:

0 (i)
(i) x + d , x , y ,t ) = 0
f( (i) (i) (i) (i)
h

...and solve this set of nonlinear equations

simultaneously for x (i) and y(i) at each time step.

The d (i) are constants defined by:

q
1
d (i) = (i)
h
  j x (i
j)
j =1

This method is much more efficient because the

nested iteration is avoided - note the number of
equations solved is the same as the inner iteration
above, and in the explicit method.

There are several reliable and efficient implement-

ations of the direct approach - e.g., DASSL (Petzold,
1983), DASOLV (Jarvis and Pantelides, 1992), and
DSL48S (Barton et al., 1997).

A version of DASOLV is used by SpeedUp. Within

ABACUSS, the user may either use DSL48S or
DASOLV.
 VARIABLE ORDER, VARIABLE STEP SIZE
METHODS

Typically, rather than requiring the user to specify the

step size and integration order, a good BDF code will
automatically adjust:
the integration step size
and the order of integration
so that estimates of the truncation error satisfy some
user specified tolerance.

Warning: controlling the error of integration (i.e.

choosing step size, order) is a complex task - standard
codes will invariably do a better job than you can.
Fix the step size at your peril!

Clearly Gear's method must always start with a first

order step (Implicit Euler) and then automatically
adjust the order and step size as the number of
previous steps increases and "confidence" concerning
the behaviour of the solution is built up.

Note: some simulators prompt the user for a step size

- always check what this means (for example, in
SpeedUp this is just a reporting interval for the results
and has no effect whatsoever on the actual steps taken
by the numerical methods).
NUMERICAL INTEGRATION OF HIGH INDEX
DAEs

Severe difficulties are encountered when the

standard codes mentioned above are applied for
numerical integration of high index systems.

Example: undamped rigid pendulum modelled in

Cartesian co-ordinates.
y
x

Length L

Problem:
x
mx + =0
L
y
my + =
mg
L
x 2 + y 2 = L2
y(0) = 0
y (0) = 0

Differential index = 3.
NUMERICAL INTEGRATION OF HIGH INDEX
DAEs

The pendulum problem is the standard academic

example. In this case, the index problem arises due
to the assumptions used in model development
(use of Cartesian co-ordinates when rigidity of the
pendulum implies the x and y positions are not
independent).

For comparison purposes, if the model is

reformulated in polar co-ordinates, we get:

g



+ sin(
) = 0
L

where
is the angle of rotation.

The index of this model is 0 (an ODE). Hence, the

numerical solution to this model is a reference by
which numerical solutions to the index 3 model can
be judged.

Note that in the following, only the solution found

by the method of dummy derivatives is virtually
indistinguishable from that of the index 0 model.
NUMERICAL INTEGRATION OF HIGH INDEX
DAEs

CASE 1: SOLVE FROM CONSISTENT INITIAL

CONDITIONS WITH STANDARD CODE

All case studies are solved from the consistent

initial condition defined by the solution to the
equations:

 (0)x(0)
x(0) +
m

=0
L
 (0)y(0)
m

y(0) + =
mg
L
x(0)2 + y(0)2 = L2

+ 2y(0) y
(0) = 0
2x(0) x(0)
x(0) + 2y(0)


2x(0)

y(0) + 2 x(0)

2 + 2 y
(0)2 = 0
y(0) = 0
y
(0) = 0

For this problem, standard code will take a few

steps, then cut the step size to zero and terminate
abnormally.

Since a high index problem is not the only cause of

this type of behaviour, the user is left clueless as to
what to do next.
NUMERICAL INTEGRATION OF HIGH INDEX
DAEs

CASE 2: SOLVE USING METHOD SUGGESTED

BY GEAR AND PETZOLD (1984)

Gear and Petzold (1984) suggested deriving the

underlying index 1 system from a high index
system using differentiation, and then submitting
this index 1 system to a standard solver.

For the pendulum example, this process yields the

index 1 system:

x
mx + =0
L
y
my + =
mg
L
2xx + 2yy + 2 x 2 + 2 y 2 = 0

PROBLEM: solution set of this system is bigger

than original index 1 system. The solution satisfies
the second derivative of the path constraint
2 2 2
x + y = L rather than the path constraint itself.

In practice, even given consistent initial conditions,

rounding and truncation error lead to a disastrous
deviation from the path constraint (see overleaf).
 METHOD OF GEAR AND PETZOLD (1984)
RESULTS

ABACUSS Dynamic Simulation

Values
1.10 X
1.00 Y
0.90
0.80
0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
-0.10
-0.20
-0.30
-0.40
-0.50
-0.60
-0.70
-0.80
-0.90
-1.00
-1.10 Time
0.00 50.00 100.00 150.00 200.00 250.00 300.00

The incredible shrinking pendulum!!

 NUMERICAL INTEGRATION OF HIGH INDEX
DAEs

CASE 3: THE METHOD OF DUMMY

DERIVATIVES
(Mattsson and Soderlind, 1993; Feehery and Barton,
1996)

PROPOSITION: for any high index system there

exists a family of index 1 systems that have an
equivalent solution set to the original high index
system.

IDEA: if we can find one of these equivalent index

1 systems, we can then submit this index 1 system
to a standard code and the results from solution of
this system can be treated as the solution of the
high index system.

Mattsson and Soderlind (1993) proposed an

algorithm that can find a member of this family of
equivalent index 1 systems for a broad class of
nonlinear DAEs of arbitrary index.1

Overleaf we show how their algorithm works when

applied to the pendulum example.

1
specifically, those for which structural criteria can correctly determine the differential index
of the system
NUMERICAL INTEGRATION OF HIGH INDEX
DAEs
1. apply Pantelides (1988) algorithm to derive the
system of equations that must be satisfied by
consistent initial conditions:
x
mx + =0
L
y
my + =
mg
L
x 2 + y 2 = L2
2xx + 2yy = 0
2xx + 2yy + 2 x 2 + 2 y 2 = 0
2. this is an overdetermined dynamic system (more
equations than unknowns). Pick a subset of the
time derivatives to substitute for dummy
algebraic variables.
x
mx  + =0
L
y
my + =
mg
L
x 2 + y 2 = L2
2xx  + 2yy = 0 (EQUIV)
2xx  + 2yy + 2 x 2 + 2 y 2 = 0

... x and x replaced by new algebraic variables x

and x

. Note true dynamic degrees of freedom of
system revealed.
NUMERICAL INTEGRATION OF HIGH INDEX
DAEs

3. system (EQUIV) is a fully determined index 1

system (when x
0 ). Solve with standard code.

In this example there are two members of the

family of equivalent index 1 systems: instead, y and
y could have been substituted (this system is index
1 except when y
0 ).

PROBLEM: members of the family of index 1

models can have their index change locally (to
index 2+) at certain points in state space. If a
numerical method steps in the neighbourhood of
such points, severe difficulties, typically failure, are
encountered (Feehery and Barton, 1996).

Example: the model (EQUIV) above becomes high

index locally when x = 0 (similarly, the other
member of the family becomes high index locally
when y = 0 ).

SOLUTION: since the points in state space where

the index rises locally do not coincide, switch
between members of the family to avoid
approaching such points for the current model.

This is known as dummy pivoting.

Feehery and Barton (1996) have recently developed

an efficient and automated approach for dummy
pivoting.
 THE METHOD OF DUMMY DERIVATIVES IN
ABACUSS

ABACUSS is now capable of automatically solving

a broad class of nonlinear DAEs of arbitrary index2
using the method of dummy derivatives (Feehery
and Barton, 1996).

Give a model coded in ABACUSS, the following

steps are done automatically:

identification of the extra equations that constrain

the initial conditions using Pantelides (1988)
algorithm.

automated derivation of these extra equations

using symbolic or automatic differentiation

generation of statistics on the (structural)

differential index, degrees of freedom for
consistent initialization, etc.

given the degrees of freedom, solution of the

consistent initialization (standard NLE solver)

automated numerical integration using the

method of dummy derivatives, and the dummy
pivoting algorithm of Feehery and Barton(1996).

2
specifically, those for which structural criteria can correctly determine the differential index
of the system
 METHOD OF DUMMY DERIVATIVES
RESULTS

ABACUSS Dynamic Simulation

Values
1.10 X
1.00 Y
0.90
0.80
0.70
0.60
0.50
0.40
0.30
0.20
0.10
0.00
-0.10
-0.20
-0.30
-0.40
-0.50
-0.60
-0.70
-0.80
-0.90
-1.00
Time
0.00 50.00 100.00 150.00 200.00 250.00 300.00

Almost identical to ODE model in polar co-

ordinates. Very small energy loss.

This problem is particularly challenging because

the pendulum is on the limit of stability.
NUMERICAL INTEGRATION OF HIGH INDEX
DAEs
STANDARD CODES

When the standard numerical codes are applied to

DAEs with index
2, the stability, convergence and
error control of the BDF method begin to break
down.
Health Warning: although (given a consistent
initial condition) standard DAE codes may be able
to advance the solution of high index problems -
the results will be garbage (error control breaks
down).
in the pendulum example we were fortunate that
the code gave up
there are examples in the literature (particularly
distillation simulation) where the code has
advanced solution of a high index model: the
results are terrible - for example, predicting
inverse responses where no such physical
behaviour exists.
Standard codes can be modified to solve classes of
nonlinear index 2 and index 3 systems correctly.
However, no such modification exists to solve
nonlinear DAEs of arbitrary index.

Why care about index 3+ systems?

dynamic design problems: specify pressure at the

bottom of a column and the index is NP+1 where
NP is the number of trays
path constrained dynamic optimizations
 COMBINED DISCRETE/CONTINUOUS SIMULATION
EXAMPLE: BUFFER TANK UPSET
y*
dy y* y
dt y

Pup Fovr

h*
Patm
x* k(h h*)

Fin Kv y(t) Pup Patm

dx x* x
x * (t) dt x
h(t)

dN Pdown
Fin Fout Fovr
dt
N Ah
Pbot Patm gh Fout K v x(t) Pbot Pdown
VALVE AUTOMATON
RESPONSE OF TANK TO MAJOR SURGE IN
UPSTREAM PRESSURE
 DISCONTINUITY HANDLING

In most dynamic simulation experiments we must be able

to handle discontinuities in both the functions assigned to
the input variables and discrete changes to the functional
form of the equations that describe the behaviour of the
system. Example: open the feed valve to a tank:

FIN = KV z
P

In this case, we select the input to the system (e.g., the

variable we can control) as the stem position of the control
valve z(t) . We can model the valve opening at time t * by
defining the time variation of the stem position as a
discontinuous function, for example:

DISCONTINUITY HANDLING
Note: assigning a step change to the stem position will
typically be an abstraction. In reality, the valve will
require a finite amount of time to open. Obviously, if this
level of detail is necessary, we could assign the following
function to the stem position:

. . . or even model the valve dynamics and introduce the

step change in the pressure signal to the valve. Remember,
as the modeller you choose the appropriate level of
abstraction.
 DISCONTINUITY HANDLING

If we know in advance when the valve is to be opened, we

can divide the time interval of interest [ t0 ,t f ] into two sub
intervals:

. . . note that in this case the process model in each interval

is the same, but we have assigned different functions to
the input variables in each interval.
What we have done is posed the overall simulation
problem as two initial value problems in DAEs:
(a) given consistent initial conditions at t = t0 , integrate
numerically the first initial value problem to t * . If the
last step goes beyond t * , simply curtail the step to t * .
(b) at t * set up the new initial value problem change
the functions assigned to the inputs.
(c) determine consistent initial conditions for the
variables x , x and y at t * .
(d) integrate numerically this new initial value problem
from t * to t f . Note integration is restarted: a BDF
method must start again with a first order step.
 DISCONTINUITY HANDLING

If there are more discontinuities like this, just divide the

time interval into more sub intervals:

=> dynamic process simulation is a sequence of initial

value problems in DAEs interspersed by
discontinuities (Barton and Pantelides, 1994)

Note:
(a) the end of a sub interval is marked by the
occurrence of an event.
(b) the time variation of the system between events is
smooth.
 CONSISTENT REINITIALIZATION

At the events that mark the end of the sub intervals it is

necessary to determine consistent initial values for the
variables at the beginning of the new interval - this is
called consistent reinitialization.

For a ODE system this reinitialization is trivial:

x (t*) = f (x(t*),u(t*),t*)

. . . it is assumed that the values of the variables x are

continuous at the event, so discontinuities can only appear
in the values of the time derivatives x (the function above
is evaluated with the new functions assigned to the input
variables).

For a DAE system both the time derivatives and the values
of the algebraic variables can be discontinuous.

Example: consider again the inlet valve of our tank. The

flow rate into the tank is related algebraically to the stem
position of the valve by the equation:

FIN = KV z
P

. . . even if the pressure drop remains constant at the

discontinuity, the value of FIN must be discontinuous to
remain consistent with the discontinuous value of the
stem position z(t) at t * .
 CONSISTENT REINITIALIZATION

In general, with a DAE system we need to determine

consistent initial values for the new initial value problem
at each event. We proceed in the following manner:
(a) determine the values of the inputs at t *
u(t*)

according to the new functions.
(b) the necessary conditions for a consistent initial
condition are given by the model equations at t * :

f ( x (t*), x(t*), y(t*),u(t*),t*) = 0

These may have to be augmented with additional

equations derived by differentiation of the original
equations - i.e. when index
2.

For this discussion, we will assume the index = 1 and

the original equations are necessary and sufficient to
define consistent initial values.
(c) we have:
n + m equations
and
2n + m unknowns { x (t*), x(t*), y(t*)}
=> we need n more side conditions to fully define the
initial state of the system in the new sub interval.
 CONSISTENT REINITIALIZATION

In order to obtain these side conditions, the normal

assumption is that the values of the differential variables
are continuous across the boundary between two
intervals, e.g., we solve:
f ( x (t*), x(t*), y(t*),t*) = 0 (R1)
for x (t*) and y(t*) keeping x(t*) constant at the values at
the end of the previous interval (leading to a fully
determined system of equations).

The physical reasoning behind this is that the differential

variables represent conserved quantities (such as mass,
energy, or momentum) or are directly related to them:

=> at an event, the values of input, algebraic or time

derivative variables may be discontinuous, but the
values of the differential variables are continuous.

Note: this poses problems for DAE systems for which the
condition:


Rank  f x
f y  = n + m
 

is not satisfied (i.e. all index
2 and some index = 1).

Clearly, if this condition is not satisfied the calculation
(R1) above is impossible. In these cases, the assumption of
continuity for all the differential variables at the boundary
is either redundant or inconsistent with some of the time
derivatives of the model equations. Research continues...
 EVENTS

Discontinuities1 are triggered by the occurrence of events

during the simulation experiment.

Events belong to one of two categories:

1. Time Events the exact time of occurrence is known
in advance. Solution of the sequence of initial value
problems can proceed to these events in strict time
order:
(a) advance numerical integration in the current
interval.
(b) curtail the final integration step length to the
event time.
(c) restart the simulation with the new initial value
problem.
Time events can be further categorized as:
exogenous if the time of occurrence is known a
priori
endogenous if the time of occurrence is
calculated from some earlier event in the
simulation.
...in either case, the time of occurrence of the next time
event is known when integration starts in a sub
interval.

1discontinuities may occur in the value of variables and/or in the value of their time
derivatives.
 EVENTS

2. State Events the time of occurrence of state events

is not known in advance because it is dependent on
the system fulfilling certain conditions state
conditions.

Numerical integration of the current IVP must be

advanced speculatively until the state condition
becomes satisfied.

Discontinuities as a consequence of state events require

careful handling . . . details to follow.
 STATE EVENTS EXAMPLE

Example: consider a batch reactor:

Reactions
1k
A 
B
A 2k
B 
C

Homogeneous catalytic liquid phase reactions; product B

desired, C waste.

=> operating objective: maximize yield of B

From reaction engineering we know that the maximum

yield of B occurs when:

dCB
=0
dt

i.e. the concentration of B in the reactor reaches a

maximum.
 STATE EVENTS EXAMPLE

Provided we can estimate dCB / dt with some analytical

instrument, we could define the operating policy for the
reactor in the following manner:

# charge the reactor

OPEN INLET VALVE
WAIT 10 MINUTES
CLOSE INLET VALVE

INJECT CATALYST
WAIT UNTIL dCB / dt
0

# discharge the reactor

OPEN OUTLET VALVE
WAIT 10 MINUTES
CLOSE OUTLET VALVE

In order to simulate this operating policy, we must be able

to deal with the state condition defined by:

dCB

0
dt

. . . i.e., advance simulation speculatively until this

condition is satisfied, then open the outlet valve (a
discontinuity).

It is extremely important to locate state events accurately

and deal with them in strict time order the occurrence
of a discontinuity can radically change the way in which
the subsequent process behaviour evolves with time.
 STATE EVENTS
In general, a state event is the time at which a state
condition expressed in terms of classical propositional
logic becomes true, i.e.:
 TRUE
t  t *
l( x, x , y, y,u,t)  

FALSE
t < t *
. . . t in some non zero interval around the event time t * .
Example: steam trap (a common device that only allows a
liquid phase to flow):

IF (h  hMIN )  (
P  0)THEN

v = KV
P
ELSE
v=0
END
where:
h = liquid level in steam trap
v = velocity of liquid through steam trap

= logical AND operator. Note: (h  hMIN )  (
P  0)
 STATE EVENT LOCATION

Simple minded approach: rely on the automatic error

control mechanism of the numerical integration method.

Example: predictor corrector method:

 STATE EVENT LOCATION

Algorithm: Ad Hoc:

(a) Take a step from t k to t k +1. State event occurs at t * in

this interval we do not know t * because we have
to integrate to t * to determine when the condition
becomes satisfied.
(b) Predictor step gives point on old trajectory.
Corrector step gives point on new trajectory.
(c) Error control mechanism uses difference between
predictor and corrector steps to estimate the
truncation error:
=> if discontinuity in step, error estimate is very large!
Alternatively, the initial guess provided by the
predictor may be so poor that the corrector iteration
fails to converge.

(d) Error control will mechanism will reject step to t k +1,

halve the step length and try the step again.
(e) Eventually. . . after repeated truncation error failures
and/or corrector iteration failures, and great
reduction in step length . . . a step is taken close
enough to t * that the error criterion is satisfied, and
integration can then proceed on the new trajectory.
 STATE EVENT LOCATION

Problems with Ad Hoc:

i) extremely inefficient repeated truncation error
failures and cutting the step size both drastically slow
down the simulation.
ii) DAEs require consistent initialization at a
discontinuity - code has to guess that a discontinuity
has occurred. However, a rapid transient can also
cause repeated truncation error failures and step size
reduction - how to distinguish between the two?
iii) as the only way of detecting the state event is the
difference between the predictor and corrector steps,
it is possible to skip over the discontinuity
completely, e.g.,

. . . as a new initial value problem must be started at t *

which may have a radically different solution from the
predicted trajectory, this is extremely dangerous
(particularly in safety related applications).
 DISCONTINUITY LOCKING

Numerical methods for ordinary differential equations

typically make assumptions concerning the continuity of
the solution trajectories:

=> do not expect efficient (or even correct) solutions

along discontinuous trajectories!

Solution: discontinuity locking:

 DISCONTINUITY LOCKING
Algorithm: Discontinuity Locking:
(a) lock the process model in a sub interval: do not allow
equations or input functions to change while
numerical integration is being advanced in a sub
interval.
(b) advance solution on this continuous trajectory until a
converged corrector step beyond the state event is
obtained (e.g. we have bounded the time of
occurrence)
(c) locate the exact time at which the state event occurs
t * within the step just taken (i.e., backtrack using
interpolation over this step).
(d) implement the change to the process model at t * ,
perform consistent reinitialization at t * , and restart
integration in the new sub interval.
Assumption:
the differential equations that make up the old model
have a mathematically defined solution beyond t *
even if the solution is not physically meaningful (e.g.,
negative mole fractions).
counter examples exist: i.e., a mathematical solution
does not exist beyond the event.
Once the time step in which the state event occurs has
been located, a variety of methods can be used to locate
the exact state event time.
 DISCONTINUITY LOCKING - BISECTION
ALGORITHM

15

10

5
Current (A)

0
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05

-5

-10

-15
Time (sec)

I1 I2

110

90

70

50

30
Voltage (V)

10

-10 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05

-30

-50

-70

-90

-110
Time (sec)

V1 V2 V3

Algorithm implemented in SpeedUp (Pantelides, 1998b),

gPROMS.
 DISCONTINUITY LOCKING - HIERARCHICAL
INTERVAL ARITHMETIC BASED ROOT
FINDING

4.50

4.00

3.50

3.00
Current (A)

2.50

2.00

1.50

1.00

0.50

0.00
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
Time (sec)

I1 I2

110

90

70

50

30
Voltage (V)

10

-10 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05

-30

-50

-70

-90

-110
Time (sec)

V1 V2 V3

New algorithm implemented in ABACUSS (Park and

Barton, 1996).

Which answer is correct? Park and Barton algorithm

guarantees correctness.
 COMBINED DISCRETE/CONTINUOUS
SIMULATION1

As the number of discontinuities increases in a dynamic

simulation, we are really performing what is called a
combined discrete/continuous simulation i.e., we are
simulating a system whose dynamic behaviour has both
significant continuous and discrete aspects:

continuous aspects physico-chemical phenomena

modelled by differential equations (material, energy,
momentum balances)

discrete aspects:

(a) discontinuities in physical behaviour e.g.,

phase changes, flow transitions

(b) control actions e.g., open/close valves,

start/stop pumps, open/close control loops.

. . . a dynamic simulation tool must be able to cope with

descriptions of both aspects of process behaviour (see
ABACUSS workshop).

1 for a more comprehensive discussion of this subject see Barton and Pantelides (1994)
 WHEN IS COMBINED SIMULATION IMPORTANT?

Whenever the system state moves far from steady-state:

modelling sequential operations e.g., start-up,

shut-down, feedstock changeover. Must model the
control actions in the operating procedures, and the
physical transitions the system goes through.

major process disturbances model interaction of

process, regulatory control system, and automatic
protective system. Process model can change (e.g.
vessel running dry) and control/safety systems can
impose discrete actions.

periodic processes e.g., pressure swing absorption.

batch processes must model complex operating

policies composed of discrete actions.

We can see that almost all interesting process simulations

are combined discrete/continuous - certainly it provides a
convenient unified strategy for process simulation.
 PROCESS MODEL DECOMPOSITION

Key concept:

Discrete aspects of process behaviour can be modelled by

manipulation of the functional form of the equations
describing the process model, e.g. changing the functional
form of:

f ( x, x, y,u,t) = 0
u = u(t)
 PHYSICO-CHEMICAL DISCONTINUITIES

Any description of the physical behaviour should include

those discrete changes to the process model that occur as a
result of physical phenomena:

(a) so that the model of physical behaviour is compact

and complete.

(b) the model will automatically adjust its functional

form in response to current conditions without
intervention from the user.

General observation:

Physico-chemical discontinuities can be modelled

changing subsets of the equations that make up the
process model in response to changes in process
conditions.

. . . therefore physico-chemical discontinuities occur as a

consequence of state events.
 PHYSICO-CHEMICAL DISCONTINUITIES
EXAMPLE

Example: tank with an overflow pipe:

This can be represented in an equation based modelling

language by:

IF h
hMIN THEN
FOUT = f (h
hMIN )
ELSE
FOUT = 0
END
 PHYSICO-CHEMICAL DISCONTINUITIES
EXAMPLE

Finite automaton representation of the model state a

digraph:

nodes represent model in a particular state

(characterized by a set of equations)
arcs represent discrete transition from one state to
another transition triggered by the occurrence of a
state event.

NOT h  h MIN

FOUT = 0 FOUT = f (h
hMIN )

h  hMIN

. . . IF equations represent symmetric reversible

discontinuities: the condition for the forward transition is
the negation of the condition for the reverse transition.
 PHYSICO-CHEMICAL DISCONTINUITIES
EXAMPLE

Example: pressure vessel fitted with safety relief valve:

FRELIEF

Mechanical safety device: relief valve opens at the set

pressure ( PS ), but closes at a (lower) reseat pressure ( PR ).
 PHYSICAL-CHEMICAL DISCONTINUITIES
EXAMPLE

Finite automaton:

P < PR

FRELIEF = 0 FRELIEF = f (P)

P > PS

The behaviour of this system exhibits hysteresis

(memory): the equations describing the system are a
function of both the current value of the process variables
and the previous behaviour of the system.

For example, if the pressure is between the reseat and set

pressures the system may be in either state the actual
state depends on what has happened in the past. The
initial state of the model must be defined as part of the
initial conditions of the system.
 REFERENCES AND FURTHER READING

Literature Cited

Ahmad B.S. and P.I. Barton, Solvent Recovery Targeting for Pollution Prevention in
Pharmaceutical and Specialty Chemicals Manufacturing, AIChE Symposium Series,
90(303), 59-73, 1994.

Allgor R.J., M.D. Berrera, P.I. Barton and L.B. Evans, Optimal Batch Process
Development, Computers Chem. Engng., 20(6/7), 885-896, 1996.

Aspen Technology, SPEEDUP User Manual, volumes I and II, version 5.4,
Cambridge, Massachusetts, 1993.

Barton, P.I., The Modelling and Simulation of Combined Discrete/Continuous

Processes, Ph.D. Thesis, University of London, 1992.

Barton, P.I. and C.C. Pantelides, Modelling of Combined Discrete/Continuous

Processes, AIChE Journal, 40, 966-979, 1994.

Barton, P.I., R.J. Allgor and W.F. Feehery, DSL48S A Large-Scale Differential-Algebraic
and Parametric Sensitivity Solver, ABACUSS Project Report no. 97/1, 22nd January
1997.

Campbell S.L., Singular Systems of Differential Equations, Pitman, Boston, 1980.

Campbell S.L., Singular Systems of Differential Equations II, Pitman, Boston, 1982.

Elmquist, H., A Structured Model Language for Large Continuous Systems, Ph.D.
Thesis, Lund Institute of Technology, 1978.

Feehery W.F. and P.I. Barton, A Differentiation-Based Approach to Simulation and

Dynamic Optimization with High-Index Differential-Algebraic Equations, In M. Berz,
C. Bischof, G. Corliss and A. Griewank, editors, Computational Differentiation, SIAM,
1996.

Gear C. W., Simultaneous Numerical Solution of Differential-Algebraic Equations,

IEEE Transactions on Circuit Theory, CT-18, p. 89, 1971.

Gear C. W. and L.R. Petzold, ODE Methods for the Solution of Differential/Algebraic
Systems, SIAM J. Numer. Anal., 21(4), 716-728, 1984.

Griewank A. and G.F. Corliss, editors, Automatic Differentiation of Algorithms:

Theory, Implementation, and Application, SIAM, Philadelphia, 1991.

Jarvis, R.B. and C.C. Pantelides, DASOLV a differential-algebraic equation solver,

technical report, Imperial College, London, 1992.
Kroner, A., W. Marquardt and E.D. Gilles, Computing Consistent Initial Conditions for
Differential-Algebraic Equations, Computers Chem. Engng., 16S, pp. S131-S138, 1992.

Lefkopoulos A. and M.A. Stadtherr, Index Analysis of Unsteady-State Chemical

Process Systems - I: An Algorithm for Problem Formulation, Computers Chem. Engng.,
17, pp. 399-413, 1993.

Marquardt W., Trends in computer-aided process modeling, In Proceedings of the 5th

International Symposium on Process Systems Engineering, Kyongju, Korea, 1994.

Mattsson, S.V. and G. Soderlind, Index reduction in differential-algebraic equations

using dummy derivatives, SIAM J. Sci. Stat. Comput., 14, pp. 677-692, 1993.

Modell, M. and R.C. Reid, Thermodynamics and its Applications, 2nd Edition,
Prentice-Hall, 1983.

Oh M., Modelling and Simulation of Combined Lumped and Distributed Processes,

Ph.D. Thesis, University of London, 1995.

Pantelides, C.C., The consistent initialization of differential-algebraic systems, SIAM J.

Sci. Stat. Comput., 9, pp. 213-231, 1988a.

Pantelides, C.C., SpeedUp - recent advances in process simulation, Computers and

Chem. Engng., 12, pp. 745-755, 1988b.

Pantelides, C.C., D. Gritsis, K.R. Morison and R.W.H. Sargent, The mathematical
modelling of transient systems using differential-algebraic equations, Computers Chem.
Engng., 12, pp. 449-454, 1988.

Park T. and P.I. Barton, State Event Location in Differential-Algebraic Models, ACM
Transactions on Modeling and Computer Simulation, 6(2), 1996.

Perkins J.D., R.W.H. Sargent and R. Vasquez-Roman, Computer generation of process

models, In Proceedings of the 5th International Symposium on Process Systems Engineering,
Kyongju, Korea, 1994.

Petzold, L.R., Differential/Algebraic Equations are not ODES, SIAM J. Sci. Stat.
Comput., 3(3), 367-384, 1982.

Petzold, L.R., A description of DASSL: a differential/algebraic system solver, in

Scientific Computing, eds. R.S. Stepleman et. al., North-Holland, Amsterdam, pp. 65-68,
1983.

Ponton, J.W. and P.J. Gawthrop, Systematic Construction of Dynamic Models for
Phase Equilibrium Processes, Computers Chem. Engng., 15, pp. 803-808, 1991.

Strauss J. C., The SCi continuous system simulation language (CSSL), Simulation, 9,
281-303, 1967.
Stephanopoulos G., G. Henning, and H. Leone, MODEL.LA. a modeling language for
process engineering - I. The formal framework, Computers Chem. Engng., 14, pp. 813-
846, 1990.

Stephanopoulos G., G. Henning, and H. Leone, MODEL.LA. a modeling language for

process engineering - II. Multifaceted modeling of processing systems, Computers
Chem. Engng., 14, pp. 847-869, 1990.

Tolsma J.E. and P. I. Barton, DAEPACK: A Flexible Environment for Open Modeling,
AIChE Annual Meeting, Dallas, Texas, 31st October 3rd November, 1999.

Industrial Experience with Dynamic Simulation

Cole J.B. and K.B. Yount, Applications of Dynamic Simulation to Industrial Control
Problems, In Advances in Instrumentation and Control, 48(2), pp. 1337-1344, Chicago,
1993.

Debelak K.A., G. Karsai, G. Chen, S. Zitney and F.O. DeCaria, Application of Dynamic
Simulation to an Industrial Process Control System, AIChE Annual Meeting, Miami,
November 1995.

Debling J.A, G.C. Han, F. Kuijpers, J. VerBurg, J. Zacca, and W.H. Ray, Dynamic
Modeling of Product Grade Transitions for Olefin Polymerization Processes, AIChE
Journal, 40(3), pp. 506-520, 1994.

Grassi V.G., Dynamic Simulation as a Tool to Integrate Process Design and Process
Control, In Advances in Instrumentation and Control, 48(2), pp. 1346-1365, Chicago, 1993.

Longwell E.J., Dynamic Modeling for Process Control and Operability, In Advances in
Instrumentation and Control, 48(2), pp. 1323-1336, Chicago, 1993.

Mani S., S.K. Shoor and H.A. Pedersen, Experience with a Simulator for Training
Ammonia Plant Operators, Plant/Operations Progress, 9(1), pp. 6-10, 1990.

Naess, L., A. Mjaavatten and J.O. Li, "Using Dynamic Process Simulation from
Conception to Normal Operation of Process Plants," Computers Chem. Engng. 17, pp.
585-600, 1993.

Zitney S.E., L. Brull, L. Lang and R. Zeller, Plantwide Dynamic Simulation on

Supercomputers: Modeling a Bayer Distillation Process, AIChE Symposium Series,
91(304), pp. 313-316, 1995.

Modelling of Process Dynamics

Franks R. G. E., Modelling and Simulation in Chemical Engineering, Wiley 1972.

Ingham J, I. J. Dunn, E. Heinzle, J. E. Prenosil, Chemical Engineering Dynamics.

Modelling with PC Simulation, VCH, Weinheim, 1994.
Luyben W. L., Process Modeling, Simulation and Control for Chemical Engineers,
chapters 2 and 3, 2nd edition, McGraw-Hill, 1990.

Review Papers (recent)

Barton, G.W. and J.D. Perkins, "Modelling and simulation in process operation," In G.V.
Reklaitis and H.D. Spriggs, editors, Foundations of Computer Aided Process
Operations, pp. 287-316, Cache-Elsevier, 1987.

Marquardt, W., "Dynamic process simulation recent progress and future challenges,"
Proceedings of CPC IV, Texas, USA, 1991.

Pantelides, C.C. and P.I. Barton, "Equation-Oriented Dynamic Simulation: Current

Status and Future Perspectives," Computers Chem. Engng., 17S, pp. S263-S285, 1993.

Marquardt W., Numerical Methods for the Simulation of Differential-Algebraic

Process Models, In R. Berber, editor, Methods of Model-based Process Control, pp. 41-79,
NATO-ASI Series, Kluver Press, 1995.

Numerical solution of initial value problems in ODEs and DAEs

Ascher, U.M. and L.R. Petzold, "Computer Methods for Ordinary Differential Equations
and Differential-Algebraic Equations", SIAM, Philadelphia, 1998. (this is the best
introductory text to get).

Lambert, J.D., "Numerical Methods for Ordinary Differential Systems the Initial
Value Problem," Wiley, 1991.

Brenan, K.E., S.L. Campbell and L.R. Petzold, "Numerical Solution of Initial-Value
Problems in Differential-Algebraic Equations," SIAM, Philadelphia, 1996.

Hainer, E., S.P. Norsett and G. Wanner, "Solving Ordinary Differential Equations:
Volume 1 - Nonstiff problems; Volume 2 - Stiff and differential-algebraic problems,"
Springer Verlag, 1987-1991.

Extensions of Sequential Modular Approach to Dynamic Simulation

Aylott M.R., J.W. Ponton and D.H. Lott, Development of a Dynamic Flowsheeting
Program, , Proceedings 2nd International Symposium on Process Systems
Engineering, Institute of Chemical Engineers, Rugby, England, 1985.

Cook W.J., J. Klatt, C.B. Brosilow, Simulation of Large Scale Dynamic Systems - II. A
Modular Simulator for the Dynamics of Distillation Systems, Computers Chem. Engng.,
11(3), 255-264, 1987.

Hillestad M. and T. Hertzberg, Dynamic Simulation of Chemical Engineering Systems

by the Sequential Modular Approach, Computers Chem. Engng., 10, 377-388, 1986.
Fagley J.C. and B. Carnahan, The Sequential-Clustered Method for Dynamic Chemical
Plant Simulation, , Computers Chem. Engng., 14(2), 161-177, 1990.

Laganier F., J.M. LeLann, X. Joulia and B. Koehret, Use of the Dynamic Sensitivity
Matrix Concept for the Efficient Solution of Dynamic Process Simulation, Proceedings
4th International Symposium on Process Systems Engineering, III.12.1, Montebello,
Quebec, August 5-9th, 1991.

Liu Y.C. and C.B. Brosilow, Simulation of Large Scale Dynamic Systems - I. Modular
Integration Methods, Computers Chem. Engng., 11(3), 241-253, 1987.

Ponton J.W., Dynamic Process Simulation Using Flowsheet Structure, Computers

Chem. Engng., 7(1), 13-17, 1983.

Ponton J.W. and V. Vasek, A Two-Level Approach to Chemical Plant and Process
Simulation, Computers Chem. Engng., 10(3), 277-286, 1986.

Ponton J.W., E.M. Johnston, J.M. Forsyth, A. Matheson, and F.M. Rutherford, Real
Time Dynamic Simulation Using Networked Computers, Computers Chem. Engng.,
13(11/12), 1245-1253, 1989.

Shinohara T., A Block Structured Approach to Dynamic Process Simulation, In:

Computer Aided Process Operations, G.V. Reklaitis and H.D. Spriggs, eds, Cache-Elsevier,
1987.