Dynamic Optimization with Path Constraints

by
William Francis Feehery
Submitted to the Department of Chemical Engineering
in partial fulfillment of the requirements for the degree of
Doctor of Philosophy in Chemical Engineering
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
March 1998

c Massachusetts Institute of Technology 1998. All rights reserved.

Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Department of Chemical Engineering
March 5, 1998

Certified by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Paul I. Barton
Assistant Professor of Chemical Engineering
Thesis Supervisor

Accepted by . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Robert Cohen
St. Laurent Professor of Chemical Engineering
Chairman, Committee on Graduate Students
2
 Dynamic Optimization with Path Constraints
by
William Francis Feehery

Submitted to the Department of Chemical Engineering

on March 5, 1998, in partial fulfillment of the
requirements for the degree of
Doctor of Philosophy in Chemical Engineering

Abstract
Dynamic optimization problems, also called constrained optimal control problems, are of
interest in many areas of engineering. However, numerical solution of such problems is
difficult and thus the application of dynamic optimization in process engineering has been
limited. The dynamic optimization problems of interest in process engineering typically
consist of large systems of differential and algebraic equations (DAEs), and often contain
path equality or inequality constraints on the state variables. The objective of this thesis
was to improve the efficiency with which large-scale dynamic optimization problems may
be solved and to develop improved methods for including path constraints.
The most efficient method for numerical solution of large dynamic optimization problems
is the control parameterization method. The cost of solving the dynamic optimization
problem is typically dominated by the cost of solving the sensitivity system. The efficiency
with which the sensitivity system can be solved is significantly improved with the staggered
corrector sensitivity algorithm which was developed and implemented during the course of
this thesis.
State variable path constraints are difficult to handle because they can cause the initial
value problem (IVP) to be a high-index DAE. An efficient method for numerical solution of
a broad class of high-index DAEs, the dummy derivative method, is described and demon-
strated. Also described is a method for transforming an equality path-constrained dynamic
optimization problem into a dynamic optimization problem with fewer degrees of freedom
that contains a high-index DAE, which may be solved using the dummy derivative method.
Inequality path-constrained dynamic optimization problems present special challenges
because they contain the additional decisions concerning the order and number of inequality
activations and deactivations along the solution trajectory. Such problems are shown to be
equivalent to a class of hybrid discrete/continuous dynamic optimization problems. Exis-
tence and uniqueness theorems of the sensitivities for hybrid systems are derived. Based on
these results, several algorithms are presented for the solution of inequality path-constrained
dynamic optimization problems. One of them, the fluctuating index infeasible path algo-
rithm, works particularly well, and its use is demonstrated on several examples.

Thesis Supervisor: Paul I. Barton

Title: Assistant Professor of Chemical Engineering

3
4
To Lisa

5
6
Acknowledgments
Although this thesis may seem like an individual effort, it reflects an extensive network
of support, advice, and friendship given to me by many people.
I would like to thank Professor Paul Barton for his intellectual guidance and
friendship. As I look back over the last few years, the chance to work with Paul has
been the biggest advantage of my choice to come to MIT. His encouragement, advice,
and (sometimes) criticism have kept my research focused and interesting.
It has been fun to work with the other students in the Barton group. In particular,
I would like to mention Berit Ahmad, Russell Allgor, Wade Martinson, Taeshin Park,
and John Tolsma. In addition, the presence of many visitors to the group during
my time here have made it a stimulating place, especially Julio Banga, Christophe
Bruneton, Santos Galán, Lars Kreul, and Denis Sédès.
My family has been a great source of love, encouragement, and support, probably
even more than they know. My parents have always been there when I needed them,
and my having gotten this far on the educational ladder is an excellent reflection on
the values that they taught me.
Most of all, thanks to my wife Lisa. I will always think of the period I spent at
MIT as the time where I met the most amazing person and we fell in love. Lisa’s
support and understanding during this project have kept me going– especially at the
end as I was writing this thesis.
Finally, I would like to acknowledge financial support in the form of a fellowship
from the National Science Foundation, and additional support from the United States
Department of Energy.

7
8
Contents

1 Introduction 19

2 Dynamic Optimization: Theory and Numerical Methods 25

2.1 DAE Dynamic Optimization Necessary Conditions . . . . . . . . . . . 28
2.1.1 ODE Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.1.2 DAE Example #1 . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1.3 DAE Example #2 . . . . . . . . . . . . . . . . . . . . . . . . 34
2.1.4 DAE Example #3 . . . . . . . . . . . . . . . . . . . . . . . . 37
2.2 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.2.1 Problems with tf fixed . . . . . . . . . . . . . . . . . . . . . . 40
2.2.2 Problems with tf free . . . . . . . . . . . . . . . . . . . . . . . 41
2.2.3 Problems with algebraic variables in ψ . . . . . . . . . . . . . 43
2.2.4 Boundary condition examples . . . . . . . . . . . . . . . . . . 43
2.3 DAEs and Dynamic Optimization Problems . . . . . . . . . . . . . . 47
2.4 Solution algorithms for Dynamic Optimization Problems . . . . . . . 50
2.5 Optimality conditions for direct methods . . . . . . . . . . . . . . . . 55
2.6 A general control parameterization problem . . . . . . . . . . . . . . 58
2.6.1 Lagrange approximation of control variables . . . . . . . . . . 58
2.6.2 The dynamic optimization problem . . . . . . . . . . . . . . . 59
2.6.3 Gradient evaluation . . . . . . . . . . . . . . . . . . . . . . . . 61
2.7 ABACUSS Dynamic Optimization Input Language . . . . . . . . . . 63

3 Parametric Sensitivity Functions for Hybrid Discrete/Continuous

Systems 65
3.1 Mathematical Representation of Hybrid Systems . . . . . . . . . . . . 68
3.2 Sensitivities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.2.1 Consistent initialization . . . . . . . . . . . . . . . . . . . . . 70
3.2.2 Initial sensitivities . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2.3 Sensitivity trajectories . . . . . . . . . . . . . . . . . . . . . . 74
3.3 Transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.3.1 Time of the event . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.3.2 Reinitialization at the transition . . . . . . . . . . . . . . . . . 79
3.3.3 Sensitivity transfer at transitions . . . . . . . . . . . . . . . . 80
3.4 Existence and Uniqueness of the Sensitivity Functions for Hybrid Systems 82

9
 3.4.1 Existence and uniqueness of the sensitivity functions for hybrid
systems with ODEs . . . . . . . . . . . . . . . . . . . . . . . . 82
3.4.2 Existence and uniqueness of the sensitivity functions for hybrid
systems with linear time-invariant DAEs . . . . . . . . . . . . 87
3.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.5.1 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.5.2 Critical values for the parameters . . . . . . . . . . . . . . . . 93
3.5.3 Functions discretized by finite elements . . . . . . . . . . . . . 99
3.5.4 Singular Van der Pol’s equation . . . . . . . . . . . . . . . . . 101
3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

4 Efficient Calculation of Sensitivity Equations 109

4.1 The Staggered Corrector Sensitivity Method . . . . . . . . . . . . . . 111
4.2 Other Sensitivity Methods . . . . . . . . . . . . . . . . . . . . . . . . 115
4.2.1 The staggered direct method . . . . . . . . . . . . . . . . . . . 115
4.2.2 The simultaneous corrector method . . . . . . . . . . . . . . . 116
4.3 Methods for Evaluating Sensitivity Residuals . . . . . . . . . . . . . . 118
4.3.1 Analytic evaluation of ∂f /∂pi . . . . . . . . . . . . . . . . . . 118
4.3.2 Finite differencing for ∂f /∂pi . . . . . . . . . . . . . . . . . . 119
4.3.3 Directional derivatives . . . . . . . . . . . . . . . . . . . . . . 119
4.3.4 Comparison of methods for evaluating sensitivity residuals . . 120
4.4 Cost Comparison of Sensitivity Methods . . . . . . . . . . . . . . . . 122
4.4.1 Staggered direct versus simultaneous corrector . . . . . . . . . 122
4.4.2 Simultaneous corrector versus staggered corrector . . . . . . . 123
4.4.3 Other cost considerations . . . . . . . . . . . . . . . . . . . . . 124
4.5 Description of DSL48S . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.6 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.7 Truncation Error Control on Sensitivities . . . . . . . . . . . . . . . . 133
4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

5 Numerical Solution of High-Index DAEs 137

5.1 Background on DAEs . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.1.1 Reasons for solving high-index DAEs . . . . . . . . . . . . . . 142
5.1.2 Methods for solving high-index DAEs . . . . . . . . . . . . . . 144
5.2 Consistent Initialization of High-Index DAEs . . . . . . . . . . . . . . 153
5.2.1 Description of Pantelides’ algorithm . . . . . . . . . . . . . . . 153
5.2.2 Example of Pantelides’ algorithm . . . . . . . . . . . . . . . . 155
5.2.3 A Modified Pantelides’ algorithm . . . . . . . . . . . . . . . . 157
5.2.4 The dummy derivative algorithm . . . . . . . . . . . . . . . . 161
5.2.5 Example of dummy derivative algorithm . . . . . . . . . . . . 163
5.3 Differentiation in ABACUSS . . . . . . . . . . . . . . . . . . . . . . . 167
5.4 Dummy Derivative Pivoting . . . . . . . . . . . . . . . . . . . . . . . 169
5.4.1 Selection of an equivalent index-1 model . . . . . . . . . . . . 169
5.4.2 Reasons for switching index-1 models . . . . . . . . . . . . . . 169
5.4.3 Deciding when to switch index-1 models . . . . . . . . . . . . 171

10
 5.4.4 Switching index-1 models during integration . . . . . . . . . . 172
5.5 Pendulum Demonstration and Numerical Results . . . . . . . . . . . 174
5.6 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
5.6.1 Fixed-volume condenser with no liquid holdup . . . . . . . . . 179
5.6.2 Standard high-index model . . . . . . . . . . . . . . . . . . . . 184
5.6.3 Continuous stirred-tank reactor . . . . . . . . . . . . . . . . . 185
5.6.4 High-Index dynamic distillation column . . . . . . . . . . . . . 187
5.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188

6 Equality Path Constraints 189

6.1 Review of Methods for Solving Path-Constrained Dynamic Optimiza-
tion Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
6.2 Equality Path Constraints and High-Index DAEs . . . . . . . . . . . 197
6.3 Dynamic Optimization Feasibility . . . . . . . . . . . . . . . . . . . . 204
6.4 Control Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
6.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
6.5.1 Two-dimensional car problem . . . . . . . . . . . . . . . . . . 214
6.5.2 Brachistochrone . . . . . . . . . . . . . . . . . . . . . . . . . . 218
6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

7 Inequality Path-Constrained Dynamic Optimization 223

7.1 Inequality Path Constraints and the Hybrid Dynamic Optimization
Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
7.2 Review of Methods for Handling Inequality Path Constraints in Con-
trol Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
7.2.1 Slack variable approach . . . . . . . . . . . . . . . . . . . . . . 233
7.2.2 Random search technique . . . . . . . . . . . . . . . . . . . . 235
7.2.3 Penalty functions . . . . . . . . . . . . . . . . . . . . . . . . . 235
7.2.4 End-point constraints . . . . . . . . . . . . . . . . . . . . . . . 236
7.2.5 Interior-point constraints . . . . . . . . . . . . . . . . . . . . . 237
7.2.6 Hybrid interior-point constraint and penalty function approach 237
7.3 A Modified Slack Variable Approach . . . . . . . . . . . . . . . . . . 239
7.4 Fluctuating-Index Feasible-Path Method for Dynamic Optimization
with Inequality Path Constraints . . . . . . . . . . . . . . . . . . . . 246
7.4.1 Constraint activation . . . . . . . . . . . . . . . . . . . . . . . 248
7.4.2 Constraint deactivation . . . . . . . . . . . . . . . . . . . . . . 249
7.4.3 Transfer conditions for state variables . . . . . . . . . . . . . . 250
7.4.4 Sensitivity transfer conditions at inequality deactivation . . . 256
7.4.5 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
7.4.6 Critique of the FIFP method . . . . . . . . . . . . . . . . . . 258
7.5 Fluctuating-Index Infeasible-Path Method for Dynamic Optimization
with Inequality Path Constraints . . . . . . . . . . . . . . . . . . . . 262
7.5.1 Transfer conditions for state variables . . . . . . . . . . . . . . 265
7.5.2 Point constraints . . . . . . . . . . . . . . . . . . . . . . . . . 267
7.5.3 Specifying a sequence of constraint events . . . . . . . . . . . 268

11
 7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269

8 Numerical examples 271

8.1 Line-Constrained Brachistochrone . . . . . . . . . . . . . . . . . . . . 273
8.2 Curve-Constrained Brachistochrone . . . . . . . . . . . . . . . . . . . 276
8.3 Constrained Van Der Pol Oscillator . . . . . . . . . . . . . . . . . . . 278
8.4 Constrained Car Problem . . . . . . . . . . . . . . . . . . . . . . . . 281
8.5 Index-2 Jacobson and Lele Problem . . . . . . . . . . . . . . . . . . . 284
8.6 Index-3 Jacobson and Lele Problem . . . . . . . . . . . . . . . . . . . 288
8.7 Pressure-constrained Batch Reactor . . . . . . . . . . . . . . . . . . . 290
8.8 Fed-Batch Penicillin Fermentation . . . . . . . . . . . . . . . . . . . . 294
8.9 Startup of CSTR and Distillation Column . . . . . . . . . . . . . . . 297
8.9.1 CSTR model . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
8.9.2 Reactor cooling jacket . . . . . . . . . . . . . . . . . . . . . . 300
8.9.3 Column tray model . . . . . . . . . . . . . . . . . . . . . . . . 301
8.9.4 Column reboiler model . . . . . . . . . . . . . . . . . . . . . . 303
8.9.5 Column condenser model . . . . . . . . . . . . . . . . . . . . . 304
8.9.6 Enthalpy model . . . . . . . . . . . . . . . . . . . . . . . . . . 305
8.9.7 Vapor-liquid equilibrium . . . . . . . . . . . . . . . . . . . . . 305
8.9.8 Path constraints . . . . . . . . . . . . . . . . . . . . . . . . . 307
8.9.9 Solution of the problem . . . . . . . . . . . . . . . . . . . . . 308

9 Conclusions 315
9.1 Directions for Future Work . . . . . . . . . . . . . . . . . . . . . . . . 319

A ABACUSS input file for reactor and column example 323

References 380

12
List of Figures

1-1 Dynamic Optimization Example . . . . . . . . . . . . . . . . . . . . . 22

2-1 The control parameterization algorithm . . . . . . . . . . . . . . . . . 57

2-2 Example of ABACUSS input file for constrained dynamic optimization 64

3-1 Rectifier circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

3-2 Controlled and autonomous transitions . . . . . . . . . . . . . . . . . 78
3-3 Discontinuity function for the system (3.80–3.81) . . . . . . . . . . . 94
3-4 Sensitivity and state trajectory for reversible transition condition when
p = 2.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3-5 Sensitivity and state trajectory for reversible transition condition when
p = 3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3-6 Sensitivity and state trajectory for nonreversible transition condition
when p = 2.9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
3-7 Sensitivity and state trajectory for nonreversible transition condition
when p = 3.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
3-8 Sensitivities with respect to junction time . . . . . . . . . . . . . . . 101
3-9 State variable trajectories for p = −3 . . . . . . . . . . . . . . . . . . 104
3-10 State space plot for p = −3 . . . . . . . . . . . . . . . . . . . . . . . 104
3-11 Sensitivity trajectory for p = −3 . . . . . . . . . . . . . . . . . . . . . 105
3-12 Sensitivity trajectory for p = −3 . . . . . . . . . . . . . . . . . . . . . 105
3-13 State trajectories for p = 0 . . . . . . . . . . . . . . . . . . . . . . . . 106
3-14 Sensitivity trajectories for p = 0 . . . . . . . . . . . . . . . . . . . . . 106

4-1 DSL48S algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

4-2 Sensitivity incremental cost per parameter . . . . . . . . . . . . . . . 132
4-3 Comparison of x sensitivity of Brachistochrone with and without sen-
sitivity error test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4-4 Comparison of y sensitivity of Brachistochrone with and without sen-
sitivity error test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

5-1 Simple tank model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

5-2 High-index pendulum . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5-3 Pendulum x trajectory with Gear method index reduction . . . . . . . 148
5-4 Pendulum length with Gear method index reduction . . . . . . . . . . 149
5-5 Starting structural graph for index-3 problem . . . . . . . . . . . . . . 155

13
5-6 Graph of index-3 system after one step of Pantelides’ algorithm . . . 156
5-7 Graph of index-3 system after two steps of Pantelides’ algorithm . . . 156
5-8 Condition number of the corrector matrix at different points on the
solution trajectory of the equivalent index-1 pendulum as a function of
the step size h . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
5-9 Length constraint in dummy derivative solution of high-index pendulum 174
5-10 Solution of index-3 pendulum model using ABACUSS . . . . . . . . . 176
5-11 LINPACK estimate of corrector matrix condition number . . . . . . . 177
5-12 ABACUSS index-reduction output for high-index condenser model . . 182
5-13 Dummy derivative Temperature profile for high-index condenser . . . 183
5-14 Dummy derivative mole holdup profile for high-index condenser . . . 183
5-15 State trajectories for index-20 DAE . . . . . . . . . . . . . . . . . . . 184
5-16 Concentration profile for index-3 CSTR example . . . . . . . . . . . . 186
5-17 Temperature profiles for index-3 CSTR example . . . . . . . . . . . . 186
5-18 Reboiler temperature profile for BatchFrac Column . . . . . . . . . . . 187

6-1 State space plot showing the optimal trajectory of the two-dimensional
car problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
6-2 Optimal acceleration trajectory for two-dimensional car problem . . . 216
6-3 The optimal velocity in the x direction for the two-dimensional car
problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
6-4 The optimal velocity in the y direction for the two-dimensional car
problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
6-5 State-space plot for the brachistochrone problem . . . . . . . . . . . . 219
6-6 Optimal force trajectory for the brachistochrone problem . . . . . . . . 220
6-7 Optimal θ trajectory for the brachistochrone problem . . . . . . . . . . 220

7-1 Feasible trajectories in constrained state space . . . . . . . . . . . . . 226

7-2 Autonomous switching of constrained dynamic simulation . . . . . . . 227
7-3 State trajectories for admissible control profile 1 . . . . . . . . . . . . 231
7-4 State trajectory for admissible control profile 1 . . . . . . . . . . . . . 231
7-5 State trajectories for admissible control profile 2 . . . . . . . . . . . . 232
7-6 State trajectory for admissible control profile 2 . . . . . . . . . . . . . 232
7-7 Control variable in modified slack variable method example with ‘a’
approximated by linear finite elements . . . . . . . . . . . . . . . . . . 245
7-8 Control variable in modified slack variable method example with ‘a’
approximated by quadratic finite elements . . . . . . . . . . . . . . . . 245
7-9 Tank and valve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
7-10 Solution of constrained brachistochrone problem . . . . . . . . . . . . 259
7-11 The FIIP method works by offsetting the constraint to the state variable
trajectory at points where the constraint activates . . . . . . . . . . . 263

8-1 Line-constrained brachistochrone control variable trajectory . . . . . . 275

8-2 Line-constrained brachistochrone state variable trajectory . . . . . . . 275
8-3 Curve-constrained brachistochrone control variable trajectory . . . . . 277

14
8-4 Curve-constrained brachistochrone state variable trajectory . . . . . . 277
8-5 Constrained Van Der Pol control variable trajectory . . . . . . . . . . 280
8-6 Constrained Van Der Pol state variable trajectories . . . . . . . . . . 280
8-7 Constrained car problem acceleration profile . . . . . . . . . . . . . . 283
8-8 Constrained car problem velocity profile . . . . . . . . . . . . . . . . . 283
8-9 Original Index-2 Jacobson and Lele control trajectory . . . . . . . . . 286
8-10 Original Index-2 Jacobson and Lele y2 trajectory . . . . . . . . . . . . 286
8-11 Modified Index-2 Jacobson and Lele control trajectory . . . . . . . . . 287
8-12 Modified Index-2 Jacobson and Lele y2 trajectory . . . . . . . . . . . . 287
8-13 Index-3 Jacobson and Lele control trajectory . . . . . . . . . . . . . . 289
8-14 Index-3 Jacobson and Lele y1 trajectory . . . . . . . . . . . . . . . . . 289
8-15 Pressure-constrained reactor control variable trajectory . . . . . . . . 292
8-16 Pressure-constrained reactor pressure trajectory . . . . . . . . . . . . 292
8-17 Pressure-constrained reactor concentration trajectories . . . . . . . . . 293
8-18 Fed-batch penicillin fermentation control variable trajectory . . . . . . 296
8-19 Fed-batch penicillin fermentation state variable trajectories . . . . . . 296
8-20 CSTR and column flowsheet . . . . . . . . . . . . . . . . . . . . . . . 297
8-21 ABACUSS index-reduction output for reactor and column startup model
when the constraint is inactive . . . . . . . . . . . . . . . . . . . . . . 308
8-22 ABACUSS index-reduction output for reactor and column startup model
when the constraint is active . . . . . . . . . . . . . . . . . . . . . . . 309
8-23 The optimal trajectory for the cooling water flowrate . . . . . . . . . . 311
8-24 The temperature profile in the reactor . . . . . . . . . . . . . . . . . . 311
8-25 The molar flowrates of the species leaving the system . . . . . . . . . 312
8-26 The temperature profile in the column . . . . . . . . . . . . . . . . . . 312
8-27 The reboiler duty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
8-28 The condenser duty . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313

15
16
List of Tables

4-1 Notation used in calculation of computational cost . . . . . . . . . . . 113

4-2 Results for integration of DAE . . . . . . . . . . . . . . . . . . . . . . 129
4-3 Results for one parameter sensitivity and DAE system (analytic sensi-
tivity residuals) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4-4 Results for two parameter sensitivity and DAE system (analytic sensi-
tivity residuals) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4-5 Comparison of analytic residuals and directional derivative residuals
for staggered corrector . . . . . . . . . . . . . . . . . . . . . . . . . . 132

5-1 Different Pivoting Strategies . . . . . . . . . . . . . . . . . . . . . . . 175

5-2 Results for Solution of Index-3 Pendulum . . . . . . . . . . . . . . . . 176
5-3 Example problem solution statistics . . . . . . . . . . . . . . . . . . . 178
5-4 Variables in the high-index condenser model . . . . . . . . . . . . . . 180
5-5 Parameters in the high-index condenser model . . . . . . . . . . . . . 180
5-6 Parameters for the CSTR example . . . . . . . . . . . . . . . . . . . 185

6-1 Statistics for the two-dimensional car problem . . . . . . . . . . . . . 215

6-2 Statistics for the brachistochrone problem . . . . . . . . . . . . . . . . 219

7-1 Dynamic optimization results for solution of equations (7.41–7.44) . . 240

7-2 Solution statistics for constrained brachistochrone . . . . . . . . . . . 258

8-1 Statistics for the line-constrained brachistochrone problem . . . . . . . 274

8-2 Statistics for the curve-constrained brachistochrone problem . . . . . . 276
8-3 Statistics for the constrained Van der Pol oscillator problem . . . . . 279
8-4 Statistics for the constrained car problem . . . . . . . . . . . . . . . . 281
8-5 Solution statistics for original index-2 Jacobson and Lele problem . . 285
8-6 Solution statistics for the modified index-2 Jacobson and Lele problem 285
8-7 Solution statistics for the index-3 Jacobson and Lele Problem . . . . . 288
8-8 Solution statistics for the pressure-constrained batch reactor problem . 291
8-9 Solution statistics for the CSTR and column startup problem . . . . . 295
8-10 Parameters used in the CSTR model . . . . . . . . . . . . . . . . . . 300
8-11 Parameters for the enthalpy model . . . . . . . . . . . . . . . . . . . 306
8-12 Parameters for the vapor pressure model (vapor pressure in bar) . . . 306
8-13 Solution statistics for the CSTR and column startup problem . . . . . 311

17
18
Chapter 1

Introduction

Dynamic optimization (also called optimal control ) problems require determining a

set of control variable time profiles for a dynamic system that optimizes a given perfor-
mance measure. Some examples of dynamic optimization problems are determination
of optimal operating policies for chemical plants subject to safety, operational, and
environmental constraints, determination of the control policy for an industrial me-
chanical robot, and solution of the minimum time-to-climb problem for an aircraft
that is required to stay within a specified flight envelope.
Solution of dynamic optimization problems has been a subject of research for hun-
dreds of years. Bernoulli posed the first dynamic optimization problem, the brachis-
tochrone problem, in 1696. The development of the calculus of variations allowed the
derivation of necessary and sufficient conditions for a solution of a dynamic optimiza-
tion problem. However, these conditions are useful for finding an analytic solution
only in restricted cases, and numerical solutions were not attempted until the advent
of modern computers.
Even with modern computers, numerical solution of dynamic optimization prob-
lems is not an easy task. The necessary and sufficient conditions that can be derived
using the calculus of variations indicate if a solution is optimal, but not necessarily
how to find an optimal (or even an improved) solution. However, many numerical
techniques have been developed which utilize information from the necessary optimal-
ity conditions, which have in common that they are computationally costly for all but

19
the most simple problems. Furthermore, these techniques often require a significant
investment of time and effort to set the problem up to be solved.

The focus of much recent process engineering research has been development of
techniques for developing and solving numerically large simulations of dynamic sys-
tems. A typical oil refinery can be modeled with over 100,000 equations, and even
models of complex unit operations can require tens of thousands of equations. En-
gineers often use these simulations to find or improve feasible operating policies for
a given system design. This process is often performed by enumerating various op-
erating policies and comparing them against each other. Such a procedure is time
consuming and provides no guarantee of finding the ‘best’ answer. Finding and im-
proving feasible operating policies can also be done by solving a dynamic optimization
problem. Dynamic optimization problems have the advantage that the solution tech-
niques can often provide an optimal answer, at least locally, but the problems are at
present difficult to set up and solve.

A simple example of a dynamic optimization problem is the brachistochrone prob-

lem that was proposed in 1696 by Johann Bernoulli to challenge the mathematicians
of Europe. The story goes that Bernoulli originally specified a deadline of six months,
but due to a request by Leibniz extended it to one year. The challenge was delivered
to Isaac Newton on January 29, 1697, and before leaving for work the next morning
he had invented the calculus of variations and used it to solve the brachistochrone
problem. Newton published his solution anonymously, but when Bernoulli saw it, he
said “We recognize the lion by his claw”[126]. The problem is to find the shape of a
frictionless wire that causes a bead, initially at rest, to move under the influence of
gravity to a specified point in minimum time.

A dynamic optimization problem that arises in process engineering is finding op-

timal operating policies for batch unit operations. Examples include finding a tem-
perature profile that maximize selectivity to the desired product in a batch reaction
involving competing kinetics, or a reflux and accumulator dump policy for a batch
distillation column that maximizes profit subject to purity and recovery constraints.
In a real plant, the profiles determined a priori by the solution to the dynamic opti-

20
mization problem can serve as set point programs for the control systems. Although
the literature abounds with analytical and numerical treatments of such problems (see
for example [118] for a review), the extension of dynamic optimization to the design
of integrated plant-wide operating policy for an entire batch process has only been
contemplated in recent years. The benefits of plant-wide dynamic optimization of
batch processes were first demonstrated in [10], and more recent work [19, 33] shows
that dynamic optimization of relatively sophisticated plant-wide models involving
thousands of states is possible.

The dynamic optimization problems that arise in process engineering have two
distinct characteristics. First, the dynamic models are composed of sets of differ-
ential and algebraic equations (DAEs). The differential equations typically arise
from dynamic material and energy balances, while the algebraic equations arise from
thermodynamic and kinetic relationships and the physical connectivity of a process
flowsheet. Although numerical solution of the types of DAEs that typically arise in
dynamic simulation of chemical processes has become reliable in recent years, dynamic
optimization problems typically involve (either explicitly or implicitly) more complex
forms of DAEs, called high-index DAEs, for which numerical solution is still an active
research area. Second, dynamic optimization problems that arise in process systems
engineering often contain path constraints on state variables, which result from phys-
ical, safety, and economic limitations imposed on the system. State variable path
constraints are not easily handled by dynamic optimization solution techniques that
are currently available. In fact, it is shown later in this thesis that state variable path
constraints lead to high-index DAEs.

A simple example of a dynamic optimization problem in process engineering is

shown in Figure 1-1. A set of reactions takes place inside a pressure-constrained
batch reactor. The objective is to maximize the amount of the product created in
a fixed amount of time. The controls available are the feed rates of the reactants
and the flow rate of the cooling water through the reactor cooling jacket. Since
the reactions taking place are exothermic, state variable path constraints must be
imposed on the pressure and temperature to keep them within safe limits throughout

21
 FB FA

Reactor

Cooling Jacket

in
Fwater

out
Fwater

Figure 1-1: Dynamic Optimization Example

the entire process.

Although computational techniques for the solution of dynamic optimization prob-

lems has been an active research topic for decades (see for example, [24, 82]), local
solution of large-scale DAE-based dynamic optimization problems has only recently
been demonstrated [33, 142] using a method called control parameterization. How-
ever, the types of dynamic optimization problems that have been solved using control
parameterization have been limited by the requirement that the DAEs not be high-
index. Although this was thought to be sufficient for physically meaningful simulation
problems, it is shown in this thesis that high-index DAEs are common in dynamic op-
timization problems. A related problem is that the methods that have been proposed
to include state path constraints in the control parameterization method are compu-
tationally inefficient and incapable of guaranteeing that the constraints are satisfied
to within a specified tolerance.

The objective of this thesis is to improve the ease with which dynamic optimiza-
tion can be used in process engineering on a regular basis. The three specific areas
of research were improving the computational efficiency of large-scale dynamic op-
timization solution methods, developing numerical solution techniques for the types

22
of high-index DAEs that arise in dynamic optimization problems, and developing
the ability to include state variable path constraints on the dynamic optimization
problem.
The dynamic optimization problem and a review of numerical solution techniques
are presented in Chapter 2. A description of the control parameterization method
is also given in Chapter 2. Control parameterization is dependent on the ability to
calculate the sensitivity of a DAE model to system parameters, which is described
in Chapter 3. A method for efficient numerical sensitivity calculation is given in
Chapter 4. Efficient numerical solution of high-index DAEs is described in Chapter 5.
Chapters 6 and 7 describe a new method for including state variable equality and
inequality path constraints in dynamic optimization problems, and Chapter 8 gives
numerical examples of this method. Chapter 9 contains conclusions and suggestions
for future research.

23
24
Chapter 2

Dynamic Optimization: Theory

and Numerical Methods

The solution of dynamic optimization (also known as optimal control) problems re-
quires the determination of control trajectories that optimize some performance mea-
sure for a dynamic system. The objective of this chapter is to derive the first-order
necessary conditions for the solution to a dynamic optimization problem, and discuss
various numerical methods for solving these problems. The control parameterization
method, which was the method chosen in this thesis to solve dynamic optimization
problems, is described and justified. The dynamic optimization problems discussed in
this chapter do not include so-called additional state variable path constraints, which
are introduced in later chapters. These problems are constrained by the dynamic
system, which is assumed to be a differential-algebraic equation (DAE). The control
parameterization framework developed here is one that may be extended to handle
state variable path constraints later in this thesis.

The following glossary may be useful when reading this chapter. These are brief
definitions intended to clarify the discussion in this chapter, and they are explored in
greater detail later in this thesis.

Differential-Algebraic Equation (DAE) A DAE is a system of equations that

25
 can be written as

f (ż, z, t) = 0 (2.1)

Ordinary differential equations (ODEs) are one class of DAEs. With the excep-
tion of ODEs, one characteristic of DAEs is that there are algebraic constraints
on the state variables z. These constraints may appear explicitly as in

g(ẋ, x, y, t) = 0 (2.2)

h(x, y, t) = 0 (2.3)

∂f
where z = (x, y), or they may appear implicitly due to singularity of ∂ ż
when
it has no zero rows.

Differential Index of a DAE “The minimum number of times that all or part of
(2.1) must be differentiated with respect to t in order to determine ż as a
continuous function of z, t is the index of the DAE” [21]. Reliable numerical
techniques exist for the direct solution of DAEs that have index ≤ 1. The index
is a local quantity– i.e., it is defined at a particular point on the state trajectory
and may change at discrete points along the state trajectory.

High-Index DAE According to common convention a DAE that has index > 1.
The solution of a high-index DAE is constrained by time derivatives of a non-
empty subset of the equations in the DAE. In general, only limited classes of
high-index DAEs may be solved directly using standard numerical techniques.

Sufficient Conditions for a DAE to be at most Index-1 It can be shown that

a sufficient condition for a DAE to at most index-1 is nonsingularity of the
Jacobian of the DAE with respect to the highest order time derivatives of each
state variable in the DAE. Therefore, if the DAE contains some differential
∂f
state variables x and algebraic state variables y and the matrix ∂q
is nonsingular,
where q = (ẋ, y), then the index of the DAE is at most one. Note that this is only

26
 a sufficient condition, and DAEs for which this condition does not hold are not
necessarily high-index. However, this is a very useful criterion for determining
the index of many problems.

Corresponding Extended System A DAE may be transformed into an ODE by

differentiating some or all of the equations in the DAE. The DAE and all of
the time derivatives that were created during this differentiation are collectively
called the corresponding extended system [58].

Consistent Initial Conditions The vectors z(to ), ż(t0 ) are called consistent initial
conditions of (2.1) if they satisfy the corresponding extended system at t0 [139].

Dynamic Degrees of Freedom Variables z or their time derivatives ż which can

be assigned arbitrary initial values and still allow a consistent initialization of
(2.1) are called dynamic degrees of freedom [139]. The number of dynamic
degrees of freedom is less than or equal to the number of state variables, and is
related to the index of the DAE. In general, the higher the index of the DAE, the
greater the number of nonredundant equations in the corresponding extended
system, and the lower the number of dynamic degrees of freedom.

Design Degrees of Freedom The number of unknowns which can be specified ar-
bitrarily as design or input quantities is called the number of design degrees
of freedom [139]. This quantity is different from the dynamic degrees of free-
dom because it involves the input variables, the choice of which can affect the
number of dynamic degrees of freedom [56, 87, 139].

Dummy Derivative Method A method for solving broad classes of high-index

DAEs that was developed in [97] and refined during the course of the work
for this thesis. The method works by deriving underlying index-1 DAEs that
have the same solution set as the high-index DAE and may be solved using
standard numerical techniques.

27
2.1 DAE Dynamic Optimization Necessary Con-
ditions

There are many texts (e.g., [24, 82]) that describe in detail the theory of dynamic
optimization. Most of them derive the necessary conditions for optimality of optimal
control problems under the assumption that the dynamic system is described by
ordinary differential equations (ODEs). In this section, the more general first-order
necessary optimality conditions are presented for dynamic systems that are described
by DAEs. The subject of necessary conditions for optimal control for DAEs was also
addressed in [36, 145], however both papers focus on the derivation of a Maximum
Principle. A Maximum Principle is useful when there are constraints on the control
variable, but in this section the dynamic optimization problem is assumed to have
no constraints on the control variable in order to demonstrate that the necessary
conditions for optimality for a DAE are also DAEs with some interesting properties.

The dynamic optimization problem considered in this section is one where the ini-
tial state for the state variables is given (that is, it is not free to be determined by the
optimization), the control trajectories are unconstrained, and the state trajectories
are constrained only by the DAE.

The dynamic optimization problem described above is expressed mathematically

as:
 tf
min J = ψ (x (tf ) , tf ) + L (x, u, t) dt (2.4)
u(t),tf t0

subject to the DAE:

f (ẋ, x, u, t) = 0 (2.5)

φ (ẋ (t0 ) , x (t0 ) , t0 ) = 0 (2.6)

where J(·), L(·), ψ(·) → R, f (·), φ(·) → Rmx , x ∈ Rmx , and u ∈ Rmu . The state
variables x in this formulation include both differential and algebraic state variables.

28
The DAE (2.5) may have arbitrary differential index v, and (2.6) defines a consistent
set of initial conditions. Since the index of (2.5) is not restricted, this formulation
may include equality path constraints on state variables.

The function ψ in (2.4) may be expressed as:

 tf
dψ
ψ (x (tf ) , tf ) = ψ (x (t0 ) , t0 ) + dt (2.7)
t0 dt

Since it is assumed that the initial time t0 and state condition x(t0 ) are fixed, the
objective function may be expressed as:
 tf
J= L̄ (ẋ, x, u, t) dt (2.8)
t0

where:
 T
dψ ∂ψ ∂ψ
L̄(ẋ, x, u, t) = +L= + ẋ + L (2.9)
dt ∂t ∂x

An augmented objective function is formed by appending the DAE constraints to the

objective function by using the adjoint variables λ (t):
 tf  
J¯ = L̄ (ẋ, u, t) + λT (t) f (ẋ, x, u, t) dt (2.10)
t0

It is convenient to define the Hamiltonian as:

H (ẋ, x, u, λ, t) = L̄ (ẋ, x, u, t) + λT (t) f (ẋ, x, u, t) (2.11)

In order to derive the necessary optimality conditions, it is necessary to define the

variation of a functional. For the functional:
 tf
J¯ = H(ẋ, x, u, λ, t)dt (2.12)
t0

29
the increment of the functional is:
 tf
ΔJ¯ = [H(ẋ + δ ẋ, x + δx, u + δu, λ + δλ, t) − H(ẋ, x, u, λ, t)] dt
t0
 tf +δtf (2.13)
+ H(ẋ, x, u, λ, t)dt
tf

Expanding the increment in a Taylor series around the point (ẋ(t), x(t), y(t)) and
extracting the terms that are linear in δ ẋ, δx, δu, δλ, and δtf gives the variation of
¯
J:
 tf  
∂H ∂H ∂H ∂H
δ J¯ = δ ẋ + δx + δu + δλ dt
t0 ∂ ẋ ∂x ∂u ∂λ (2.14)
+ H(ẋ, x, u, λ, t)δtf

which can be simplified by integrating the first term by parts to obtain:

 tf    
∂H d ∂H ∂H ∂H
δ J¯ = − δx + δu + δλ dt
∂x dt ∂ ẋ ∂u ∂λ
 t0  (2.15)
∂H
+ δx(tf ) + H(ẋ, x, u, λ, t)δtf
∂ ẋ t=tf

Using the following relation:

δx(tf ) = δxf − ẋδtf (2.16)

with (2.15) yields:

   
∂H ∂H
δ J¯ = δxf + H − ẋ δtf
∂ ẋ t=tf ∂ ẋ t=tf
 tf    (2.17)
∂H d ∂H ∂H ∂H
+ − δx + δu + δλ dt
t0 ∂x dt ∂ ẋ ∂u ∂λ

First-order necessary conditions for an optimum can be found by setting the variation

30
of J¯ equal to zero. The conditions are:
   
∂H d ∂H
− =0 (2.18)
∂x dt ∂ ẋ
∂H
=0 (2.19)
∂u
∂H
=0 (2.20)
  ∂λ 
∂H ∂H
δx + H− ẋ δtf = 0 (2.21)
∂ ẋ t=tf ∂ ẋ t=tf

Conditions (2.18–2.21) define a two-point boundary value problem in DAEs, as is

the case with an ODE embedded system. Condition (2.18) is often called the costate
equation. These conditions may be simplified. In the expansion of (2.18), the terms
that include ψ are:
    
∂ ∂ψ ∂ψ d ∂ ∂ψ
ẋ + − ẋ
∂x ∂x ∂t dt ∂ ẋ ∂x
 2   2 
∂ ψ ∂2ψ ∂ ψ ∂2ψ (2.22)
= + ẋ − + ẋ
∂x∂t ∂ 2 x ∂t∂x ∂ 2 x
=0

provided that the second partial derivatives are continuous.

Using (2.11) and (2.22), (2.18–2.21) simplify to:

 
∂L T ∂f T ∂f T d ∂f
+λ − λ̇ −λ =0 (2.23)
∂x ∂x ∂ ẋ dt ∂ ẋ
∂L ∂f
+ λT =0 (2.24)
∂u ∂u
f (ẋ, x, u, t) = 0 (2.25)
     
∂f ∂ψ ∂ψ ∂f
λT + δxf + + L + λT f − λT ẋ δtf = 0 (2.26)
∂ ẋ ∂x t=tf ∂t ∂ ẋ t=tf

These conditions are a generalization of the conditions that have been reported for
dynamic optimization of ODE systems.

31
2.1.1 ODE Example

The following dynamic optimization problem was taken from [24] and is attributed to
Isaac Newton. Essentially, the problem is to find the minimum-drag nose cone shape
in hypersonic flow. The problem is:
 l
1 ru3
min [r(l)]2 + dx (2.27)
u(x) 2 0 1 + u2

subject to:

dr
+u=0 (2.28)
dx
r(0) = r0 (2.29)

where x is the axial distance from the point of maximum radius, r is the radius of
the body, r(0) is the maximum radius, and l is the length of the body.
The optimality conditions (2.23–2.26) reduce in this problem to:

dλ u3
= (2.30)
dx 1 + u2
ru2 (3 + u2 )
λ= (2.31)
(1 + u2 )2
dr
+u=0 (2.32)
dx
r(0) = r0 (2.33)

λ(l) = −r(l) (2.34)

These are the same optimality conditions as those derived in [24] except that the
sign of λ in (2.31) and (2.34) is different. This difference occurs because the usual
derivation of conditions (2.23–2.26) for ODEs assumes that the ODEs have the form
ẋ = f (x, u, t), while the derivation of (2.23–2.26) made the more general assumption
that the DAE has the form f (ẋ, x, u, t) = 0. However, it can be shown that the
solution for the state variable given in [24] also satisfies (2.30–2.34), although the
adjoint variable trajectories are different.

32
 Note that even though the original problem is described by an ODE, the optimality
conditions are an index-1 DAE (when u = 0), since (2.31) contains no time derivative
variables. In general, all dynamic optimization problems described by ODEs have
optimality conditions that are DAEs.

2.1.2 DAE Example #1

It is possible to turn the ODE example of the previous section into an index-1 DAE,
and thus check the validity of the optimality conditions, by writing:
 l
min a(l) + bdx (2.35)
u(x) 0

subject to:

dr
+u=0 (2.36)
dx
1
a = r2 (2.37)
2
ru3
b= (2.38)
1 + u2
r(0) = r0 (2.39)

The optimality conditions for this problem (neglecting the condition (2.26) for the
moment) reduce to (2.36–2.39) and:

dλ1 u 3 λ2 1
=− 2
− λ3 (2.40)
dx 1+u 2
λ2 = −1 (2.41)

λ3 = 0 (2.42)
ru2 (3 + u2 )
λ1 = − λ2 (2.43)
(1 + u2 )2

If (2.41–2.42) are substituted into (2.40) and (2.43), the result is the same as (2.30–
2.31).

33
 The boundary condition (2.26) reduces to:

λ1 (l) = 0 (2.44)

1=0 (2.45)

which apart from being inconsistent, does not agree with the results of the ODE
example. The problem is that the variable a appears in φ but ȧ does not appear in
the DAE (this is further discussed in Section 2.2 below). If (2.37) is replaced with its
time differential, the boundary condition reduces to:

λ1 (l) = r(l)λ3 (l) (2.46)

λ3 (l) = −1 (2.47)

which reduces to the result obtained in the ODE example.

2.1.3 DAE Example #2

The following dynamic optimization contains a high-index (index-3) DAE:

 2
1
min x(tf ) − (2.48)
u(t) 2

subject to:

ẋ = v (2.49)

ẏ = w (2.50)

v̇ + T x = u2 y (2.51)

ẇ + T y = −1 + u2 x (2.52)

x2 + y 2 = 1 (2.53)

x(t0 ) = 0 (2.54)

ẏ(t0 ) = 0 (2.55)

34
 The costate equations are:

λ̇1 = T λ3 − u2 λ4 + 2xλ5 (2.56)

λ̇2 = −u2 λ3 + T λ4 + 2yλ5 (2.57)

λ̇3 = −λ1 (2.58)

λ̇4 = −λ2 (2.59)

xλ3 + yλ4 = 0 (2.60)

and (2.24) simplifies to:

uyλ3 + uxλ4 = 0 (2.61)

The index of the combined system (2.49–2.53) and (2.56–2.61) is 3. A degree of

freedom analysis indicates that there should be two initial time conditions and two
final time conditions. However, (2.26) gives:

λ1 (tf ) = 1 − 2x(tf ) (2.62)

λ2 = 0 (2.63)

λ3 = 0 (2.64)

λ4 = 0 (2.65)

which cannot be correct. As discussed in Section 2.2 the boundary conditions must
be derived using an equivalent non-high index formulation of the DAE.

Another way to find optimality conditions for this problem is to use the methods

35
of Chapter 5 to derive an equivalent index-1 DAE for (2.49–2.53):

ẋ = v (2.66)

ẏ = w (2.67)

mv̇ + T x = u2 y (2.68)

mẇ + T y = −1 + u2 x (2.69)

x2 + y 2 = 1 (2.70)

xẋ + y ȳ = 0 (2.71)

ẋ2 + ȳ 2 + xv̇ + y w̄ = 0 (2.72)

where ȳ and w̄ are dummy algebraic variables that have been introduced in place of
ẏ and ẇ. For this system, the costate equations are:

λ̇1 = T λ3 − u2 λ4 + 2xλ5 + ẋλ6 + v̇λ7 − xλ̇6 − 2ẋλ̇7 (2.73)

−u2 λ3 + T λ4 + 2yλ5 + ȳλ6 + w̄λ7 = 0 (2.74)

λ̇3 = −λ1 − xλ̇7 − ẋλ7 (2.75)

λ2 = 0 (2.76)

xλ3 + yλ4 = 0 (2.77)

λ2 + yλ6 + 2ȳλ7 = 0 (2.78)

λ4 + yλ7 = 0 (2.79)

and (2.24) gives:

uyλ3 + uxλ4 = 0 (2.80)

Interestingly, in this case the combined system (2.66–2.80) is index-2, even though
the DAE (2.66–2.72) is index-1. A degree of freedom analysis indicates that there are
two initial and two final time boundary conditions, as expected.

36
 Applying the boundary condition (2.26) gives:

2x(tf ) − 1 + λ1 (tf ) + x(tf )λ6 (tf ) + 2ẋ(tf )λ7 (tf ) = 0 (2.81)

λ3 (tf ) + x(tf )λ7 (tf ) = 0 (2.82)

which is the correct number of final time conditions. This example leads to a conjec-
ture that it is possible to derive optimality conditions for a high-index DAE directly,
but that the final-time boundary conditions must be derived using an equivalent
non-high index formulation of the high-index DAE.

2.1.4 DAE Example #3

Consider the following DAE dynamic optimization problem:

min tf (2.83)
θ(t),tf

subject to:

ẋ = u (2.84)

ẏ = v (2.85)

u̇ = F sin(θ) (2.86)

v̇ = g − F cos(θ) (2.87)
v
tan(θ) = (2.88)
u
[x(0), y(0), ẋ(0)] = [0, 1, 0]

This problem is a version of the brachistochrone problem due to [14] which is further
discussed in Chapter 6. This problem is used in the next section to demonstrate
the different types of boundary conditions that are possible in dynamic optimization
problems. The index of the DAE in this problem is two if θ is selected as the control.

37
 The optimality condition (2.23) reduces to:

λ̇1 = 0 (2.89)

λ̇2 = 0 (2.90)
v
λ̇3 = −λ1 + λ5 (2.91)
u2
1
λ̇4 = −λ2 − λ5 (2.92)
u
0 = −λ3 sin(θ) + λ4 cos(θ) (2.93)

Note that this condition is also a DAE.

Condition (2.24) reduces to:

λ3 F cos(θ) + λ4 F sin(θ) + λ5 sec2 (θ) = 0 (2.94)

To obtain boundary conditions for this problem, it is necessary to consider both

the index of the DAE (2.84–2.88), and the index of the extended DAE that is formed
by (2.84–2.94) so that the correct dynamic degrees of freedom may be determined.
The dynamic degrees of freedom of the original DAE determines the number of initial
conditions that may be specified, and the dynamic degrees of freedom of the extended
system determine the number of final conditions that may be specified. This result
has not been previously noted in the literature, and it is further discussed in the next
section.

38
2.2 Boundary conditions
The exact form of the boundary condition (2.21) depends on the classes of constraints
imposed on the dynamic optimization problem. This section presents the boundary
conditions for several different classes of constraints. The approach followed is gen-
erally that of [82], but here the results are valid for arbitrary-index DAEs.
The number of boundary conditions depends on the form of (2.5), and is related
to rd , the dynamic degrees of freedom of the DAE. Several observations may be made
about the number of boundary conditions that may be imposed:

• The total number of boundary conditions that may be imposed is roverall + 1,

where roverall is the dynamic degrees of freedom of the extended system (2.23–
2.25). The number of boundary conditions is one more than roverall because one
condition must be imposed to define tf .

• The number of boundary conditions can be dependent on the time at which the
conditions are defined, since roverall can change over the time domain.

• Typically, the boundary conditions consist of initial conditions on the DAE and
end-point conditions on the costate equations. In this case, the number of initial
conditions for the DAE is rd at t = t0 . The number of end-point conditions on
the costate equations is roverall − rd at t = tf .

• It is not possible to make general statements concerning either the dependence

of the index of (2.23–2.25) on the index of (2.25) or the dependence of roverall
on rd .

The final-time boundary conditions are obtained by customizing (2.26) to the par-
ticular type of side conditions in the dynamic optimization problem. It is important
to note that although (2.23–2.25) can be applied to DAEs of arbitrary index, (2.26)
cannot be correctly applied directly to high-index DAEs. Consistent initial conditions
for a high-index DAE can only be obtained with the corresponding extended system
for the high-index DAE, and likewise (2.26) can only be used with the corresponding

39
extended system for the DAE (2.23–2.25). Therefore, for the purposes of this section,
the DAE f is assumed to have index at most one. This restriction is allowable because
the methods of Chapter 5 can be used to derive an equivalent index-1 DAE for the
high-index DAEs of interest in this thesis.

2.2.1 Problems with tf fixed

Fixing tf provides the boundary condition on the final time, and allows δtf to be set
to zero in (2.26). There are several possibilities for the other boundary conditions,
depending on the form of the dynamic optimization problem.

• Final state free: Since the final state is unspecified, δx(tf ) is arbitrary, and
therefore the end-point boundary conditions must satisfy:
 
T ∂f ∂ψ
λ + δxf = 0 (2.95)
∂ ẋ ∂x t=tf

Note that when:

 
T∂f ∂ψ
λ + =0 (2.96)
∂ ẋi ∂xi t=tf

that δxf i is arbitrary.

• Final state constrained : If the final state is constrained to lie on a hypersurface

defined by m(x(tf )) = 0, where m : Rmx → Rnm , 1 ≤ nm ≤ rd − 1, and if it is
assumed that m(x(tf )) is neither inconsistent nor redundant, then δx(tf ) must
be normal to each of the gradient vectors:

∂mi
i = 1 . . . nm (2.97)
∂x

40
 and therefore it can be shown [82] that the boundary conditions are:

  nm
T∂f ∂ψ ∂mi
λ + = νi (2.98)
∂ ẋ ∂x t=tf i=1
∂x

m(x(tf )) = 0 (2.99)

where ν ∈ Rnm is a vector of undetermined multipliers.

• Final state specified : This is a special instance of the preceding case. However,
here the dimensionality of the hypersurface is equal to rd , and therefore the
boundary conditions:

x(tf ) = xf (2.100)

apply. The boundary condition (2.26) applies also, but for every equation in
(2.100), the corresponding δx = 0 in (2.26). Therefore, if roverall > 2rd some of
the boundary conditions will come from (2.26).

2.2.2 Problems with tf free

When tf is free, the assumption that δtf = 0 can no longer be made.

• Final state fixed or final state free: These cases have the same boundary con-
ditions as their counterpart discussed above except that there is an additional
boundary condition:
   
∂ψ T ∂f
+L+λ f −λ
T
ẋ =0 (2.101)
∂t ∂ ẋ t=tf

• Final state constrained to lie on a moving point: In this case, x(tf ) must lie on
the final point defined by θ(tf ). This case is complicated by the fact that no
more than rd state conditions may be simultaneously imposed, and rd will be
less than mx if the DAE is index-1. The dimension of the vector function θ is

41
 equal to rd , and the variations of δxf and δtf are related by:

dθ
P δxf = δtf (2.102)
dt

where P is a permutation matrix which maps Rmx → Rrd . The boundary

condition for this problem is therefore:
      
∂f ∂ψ dθ ∂ψ T ∂f
λT
+ P + +L+λ f −λ
T
ẋ =0 (2.103)
∂ ẋ ∂x dt t=tf ∂t ∂ ẋ t=tf

P x(tf ) = θ(tf ) (2.104)

Note that (2.102) will not define all δx if rd is less than mx . Equation (2.95)
holds for all undefined elements of δx, and may give additional boundary con-
ditions.

• Final state constrained by a surface: Again, the final state is constrained by

the equations m(x(tf )) = 0, m : Rmx → Rnm , 1 ≤ nm ≤ roverall − rd − 1 at
tf . Since the final time is free, the boundary conditions are the same as the
corresponding case with fixed tf with the addition of the boundary condition
(2.101).

• Final state constrained by a moving surface: This is the most complicated case,
where the final state is constrained by the moving surface m(x, t) = 0, m :
Rmx +1 → Rnm , 1 ≤ nm ≤ roverall − rd − 1 at tf . This situation is similar to
the preceding case, except that the vector [δx(tf ) | δtf ] is normal to each of the
gradient vectors [ ∂m
∂x
i
| ∂m
∂t
i
]. Therefore, the boundary conditions are:

  nm
T∂f ∂ψ ∂mi
λ + = νi (2.105)
∂ ẋ ∂x t=tf i=1
∂x

m(x(tf )) = 0 (2.106)
    nm
∂ψ T ∂f ∂mi
+L+λ f −λ
T
ẋ = νi (2.107)
∂t ∂ ẋ t=tf i=1
∂t

42
2.2.3 Problems with algebraic variables in ψ

As shown in Section 2.1.2, if the function ψ in (2.4) contains algebraic state variables,
special treatment is required to derive the boundary conditions with (2.26). The
difficulty is that the time derivatives of algebraic variables do not appear explicitly
in the DAE, but ∂H/∂ ẋ will be nonzero with respect to the time derivatives of any
algebraic variables for which the corresponding elements of ∂ψ/∂x are nonzero.

The boundary condition (2.26) can be correctly applied if a subset of the equations
DAE f (ẋ, x, u, t) is replaced with its time derivatives for the purposes of (2.26). It is
valid to do this because all time derivatives of f (ẋ, x, u, t) must hold at t0 .

2.2.4 Boundary condition examples

The dynamic degrees of freedom of the DAE (2.84–2.88) and the extended DAE
(2.84–2.93) are three and seven, respectively. Three initial conditions were defined,
and therefore, four final time conditions remain to be specified. The implications of
some of the various boundary condition types can be demonstrated by considering
the following cases:

• Case 1: Final condition is x(tf ) = xf :

This is the case where the final state is constrained by a hypersurface. Since the
final time is free, there is an additional degree of freedom that is determined by
the dynamic optimization problem (2.83–2.88). The final time condition (2.101)
reduces to:

v
1 − λ1 u − λ2 v − λ3 F sin(θ) − λ4 F cos(θ) + λ5 tan(θ) − = 0 (2.108)
u t=tf

43
 The optimality conditions (2.98–2.99) reduce to:

λ1 (tf ) = ν1 (2.109)

λ2 (tf ) = 0 (2.110)

λ3 (tf ) = 0 (2.111)

λ4 (tf ) = 0 (2.112)

x(tf ) = xf (2.113)

Note that the undetermined multiplier ν1 has been added to the dynamic opti-
mization problem, and there are now four remaining dynamic degrees of free-
dom.

• Case 2: Final condition is x(tf ) = a(tf ), y(tf ) = b(tf ), v(tf ) = 0:

This is the case where the final state must lie on a moving point. Assuming the
state variable vector for this problem is [x, y, u, v, F ], the permutation matrix
P for this problem is:
⎡ ⎤
1 0 0
⎢ ⎥
⎢ ⎥
⎢0 1 0⎥
⎢ ⎥
⎢ ⎥
P = ⎢0 0 0⎥ (2.114)
⎢ ⎥
⎢ ⎥
⎢0 0 1⎥
⎣ ⎦
0 0 0

and therefore, (2.102) becomes:

δx(tf ) = a (tf )δtf (2.115)

δy(tf ) = b (tf )δtf (2.116)

δv(tf ) = 0 (2.117)

Note that δu and δF can take arbitrary values and still satisfy the boundary

44
 condition. Equations (2.103–2.104) reduce to:

[λ1 a + λ2 b + 1 − λ1 u − λ2 v − λ3 F sin(θ) −
v (2.118)
λ4 F cos(θ) − λ5 tan(θ) − =0
u t=tf
x(tf ) = a(tf ) (2.119)

y(tf ) = b(tf ) (2.120)

ẏ(tf ) = 0 (2.121)

There is still one more final time condition that must be defined. Equation
(2.95) is applied to the undefined elements of δx to obtain the final boundary
condition:

λ3 (tf ) = 0 (2.122)

• Case 3: Final condition is x2 (tf ) + y 2 (tf ) = R(tf )2 :

In this case, the final state is constrained to lie on a circle that is expanding with
time (i.e., R(t) is a scalar valued function of time). Equations (2.105–2.107)
reduce to:

λ1 (tf ) = 2ν1 x(tf ) (2.123)

λ2 (tf ) = 2ν1 y(tf ) (2.124)

λ3 (tf ) = 0 (2.125)

λ4 (tf ) = 0 (2.126)

x2 (tf ) + y 2 (tf ) = R(tf )2 (2.127)

[1 − λ1 u − λ2 v − λ3 F sin(θ)
v ∂R (2.128)
− λ4 F cos(θ) + λ5 (tan(θ) − = −ν1
u t=tf ∂t

• Case 4: Fixed final time with objective function x2 :

45
In this case, the objective function (2.83) is changed to:

min x2 (tf ) (2.129)

θ(t)

and the final time is not a decision variable. Equations (2.89–2.94) apply un-
changed to this problem, and (2.95) gives the boundary conditions:

λ1 + 2x = 0 (2.130)

λ2 = 0 (2.131)

λ3 = 0 (2.132)

λ4 = 0 (2.133)

46
2.3 DAEs and Dynamic Optimization Problems

The previous section demonstrated that it is necessary to know roverall at tf in order

to determine a valid set of end-point boundary conditions. It is also necessary to
know rd at t0 in order to determine consistent initial conditions for the DAE. This
section shows that in fact the necessary conditions for optimality are naturally DAEs,
and therefore the solution of dynamic optimization problems requires the ability to
handle DAEs. This situation is true even for many problems where the dynamic
model is described by an ODE.
The extended system (2.23–2.25) is a DAE with an index of at least one, even if
(2.25) is an ODE, which may be observed in the example (2.27–2.34). The extended
system must be at least index-1 because it contains algebraic variables u.
To show that the index of the extended system can be greater than one, even
for an ODE-embedded system, consider the following system that was discussed in
[24, 90].

min Φ(x(tf )) (2.134)

u

subject to:

ẋ = f (x) + g(x)u (2.135)

This problem is a very simple ODE system which is linear in the control variable u,
and possibly nonlinear in the state variable x. The necessary conditions (2.23–2.26)
simplify to generate the extended system:
   
∂f ∂g
λ̇ = − + u λ (2.136)
∂x ∂x
λg(x) = 0 (2.137)

ẋ = f (x) + g(x)u (2.138)

∂Φ(tf )
λ(tf ) = − (2.139)
∂x

47
The Jacobian of (2.136–2.138) with respect to the highest order time derivatives
(λ̇, ẋ, u) is:
⎡  ∂g  ⎤
1 0 ∂x λ
⎢ ⎥
⎢ ⎥
⎢0 0 0 ⎥ (2.140)
⎣ ⎦
0 1 −g

which is singular. However, if (2.137) is replaced with its time derivative:

∂g
λ ẋ + λ̇g(x) = 0 (2.141)
∂x

the Jacobian of the underlying DAE consisting of (2.136), (2.141) and (2.138) is:
⎡  ∂g  ⎤
1 0 λ
⎢   ∂x ⎥
⎢ ⎥
⎢g ∂x λ
∂g
0 ⎥ (2.142)
⎣ ⎦
0 1 −g

which is structurally nonsingular. Since an algebraic state variable u is still present in

the underlying DAE, the index of the underlying DAE must be at least one. Therefore,
the structural index of the original system (2.136–2.138) was two. However, it can be
shown that matrices of the form of (2.140) are always singular. This singularity may
indicate that the index of the underlying DAE is higher than two, but singularity
of (2.140) is not a sufficient condition for a DAE to be high-index. This fact was
overlooked in [90], however, algebraic rearrangement can be used to show that (2.136–
2.138) is in fact index-3, as reported in that paper. To see this, substitute (2.136)
and (2.138) into (2.141) to get:

∂g ∂f
λ f (x) − λ g(x) (2.143)
∂x ∂x

since the variable u does not appear, a further differentiation of (2.136), (2.141) and
(2.138) is required to derive the corresponding extended system.

It is interesting to see whether (2.26) has given the correct number of final time

48
boundary conditions for this problem. The corresponding extended system for this
DAE consists of seven equations, (2.136–2.138), (2.141), and the time derivatives of
(2.136), (2.141) and (2.138). The variables incident in the corresponding extended
system are ẍ, ẋ, x, λ̈, λ̇, λ, u̇, and u. Since there are eight incident variables and seven
equations in the corresponding extended system, there should be one boundary con-
dition. That boundary condition is the final-time condition given by (2.139), which
implies no initial conditions may be specified for this problem.
The bottom line is that dynamic optimization problems are inherently boundary
value problems in DAEs, not ODEs (even if the dynamic system is an ODE). There-
fore, solving a dynamic optimization problem requires knowledge about the dynamic
degrees of freedom of both the original dynamic system and the DAE defined by
the optimality conditions in order to formulate valid boundary conditions, and the
ability to solve the DAE that is embedded in the two-point boundary value problem
either explicitly or implicitly. This fact has not been formally recognized in any of the
work that has been done on development of methods to solve dynamic optimization
problems. In fact, all of the methods that have been developed handle the DAEs
implicitly, both because these problems were not recognized as DAE problems, and
because advances that allowed direct numerical solution of DAEs have been made
only recently.

49
2.4 Solution algorithms for Dynamic Optimization
Problems

The optimality conditions presented in the previous section are of little practical use
by themselves. They give conditions that must be satisfied at the optimum, but
they are not a method for finding the optimum. In fact, they define a two-point
boundary value problem, which is well known to be difficult to solve numerically.
This section gives a brief overview of the various methods that have been developed
to solve dynamic optimization problems.
There is an extensive literature on numerical methods for the solution of dy-
namic optimization problems, which fall into three classes. The dynamic programming
method was described in [18] and the approach was extended to include constraints
on the state and control variables in [91]. Indirect methods focus on obtaining a
solution to the classical necessary conditions for optimality (2.18–2.21) which take
the form of a two-point boundary value problem. There are many examples of such
methods, for example [24, 37, 41, 42, 80, 100, 111]. Finally, direct methods trans-
form the infinite-dimensional dynamic optimization problem into a finite dimensional
mathematical programming problem, typically a nonlinear program (NLP).
The dynamic programming method is based on Bellman’s principal of optimality,
which can be used to derive the Hamilton-Jacobi-Bellman equation:
 
∂J ∂J
0= + min L + f (x, u, t) (2.144)
∂t u(t) ∂x
J(x(tf ), tf ) = ψ(x(tf ), tf ) (2.145)

which must hold at the optimum, assuming that the DAE is index-0 and may be
written as ẋ = f (x, u, t). This is a partial differential equation and in practice it is
very difficult to solve except in certain fortuitous cases [133].
The indirect methods are so termed because they attempt to find solutions to the
two-point boundary value problem (2.18–2.21), thus indirectly solving the dynamic
optimization problem (2.4–2.6). There are many variations on indirect methods, but

50
they are generally iterative methods that use an initial guess to find a solution to a
problem in which one or more of (2.18–2.21) is not satisfied. This solution is then
used to adjust the initial guess to attempt to make the solution of the next iteration
come closer to satisfying all of the necessary conditions.

One of the common variations on the indirect method is the steepest descent
algorithm (for example, [82]). In this method, the state equation (2.20) is integrated
forward using a guess for the control profile, and then the costate equation (2.18) is
integrated backward. Equation (2.19) is then used locally to find a steepest descent
direction for u at a discrete number of points, and globally as a termination criterion.
The main disadvantage of this method is the need to backward-integrate the costate
equations, which can be inefficient and/or unstable for both DAEs [142] and ODEs.
Also, it is not easy to set up the problem, since the costate equations need to be
generated, and convergence is slow for many problems.

Backward integration of the costate equations may be avoided by using a direct

shooting method, in which the initial values of the adjoint variables λ(t0 ) are guessed
initially, and updated iteratively [99]. There are some problems with this method,
including a small region of convergence to the optimum, the need to formulate the
costate equations, and the difficulties in choosing reasonable initial guesses for λ(t0 ),
which may or may not have an obvious physical meaning. The multiple shooting
method was proposed to extend the small convergence region of the direct shooting
method [39, 79, 122]. Multiple shooting algorithms transform the problem into a
multipoint boundary value problem for the state and adjoint variables. Homotopy
methods have also been proposed [144] as a response to the small region of conver-
gence, and automatic symbolic differentiation has been used to generate the costate
equations [129]. Other indirect techniques reported have been the finite difference
method [89] which is essentially a low-order collocation method, and standard collo-
cation techniques [4].

There are two general strategies within the framework of the direct method. In
the sequential method, often called control parameterization, the control variables are
discretized over finite elements using polynomials [23, 105, 104, 115, 120, 128, 142, 143]

51
or in fact any suitable basis functions. The coefficients of the polynomials and the size
of the finite elements then become decision variables in a master nonlinear program
(NLP). Function evaluation is carried out by solution of an initial value problem
(IVP) of the original dynamic system, and gradients for a gradient-based search may
be evaluated by solving either the adjoint equations (e.g., [128]) or the sensitivity
equations (e.g., [142]). In the simultaneous strategy both the controls and the state
variables are discretized using polynomials on finite elements, and the coefficients and
element sizes become decision variables in a much larger NLP (e.g., see [12, 61, 85,
92, 106, 132, 137, 147]). Unlike control parameterization, the simultaneous method
does not require the solution of IVPs at every iteration of the NLP.

Direct methods have been demonstrated to work automatically and reliably for
very large problems. Within direct methods, there has been significant discussion in
the literature about the relative advantages of simultaneous and sequential strategies.
The main areas of contention are the ease of use, the computational cost in obtaining
a solution, and the ability to handle state variable path constraints.

The sequential strategy is easier to use on large problems than the simultaneous
strategy because it uses a numerical IVP solver to enforce (2.5), rather than including
it as a set of constraints in the NLP. Including these constraints in the NLP in the
sequential strategy causes the NLP to grow explosively with the size of the DAE,
the time horizon of interest, and the excitation of high-frequency responses. Solution
of such large NLPs requires special techniques and careful attention to determining
an initial guess for the optimization parameters that leads to convergence of the
NLP algorithm. On the other hand, solution of the IVP in the sequential strategy
can take advantage of the highly refined heuristics available in state-of-the-art DAE
integrators. In fact, it can be shown that if collocation is used for the discretization
in the simultaneous method, the problem is equivalent to performing a fully implicit
Runge-Kutta integration [142], which is not as efficient as the BDF method for DAEs
[21].

It is not clear from the literature discussions which strategy requires less compu-
tational cost. On the one hand, it is expensive to solve the extremely large NLPs that

52
result from the simultaneous method. On the other hand, an efficient search using
the sequential strategy requires gradients of the objective function, which can also
be very expensive to obtain. This issue is addressed in Chapter 4, where an efficient
method for calculating the sensitivities of the state variables is described, which can
then be used to calculate the gradients.
Until now, the simultaneous strategy has had the clear advantage in handling
constraints on the state variables. Such constraints may be handled by including them
directly as additional constraints in the NLP. However, the theoretical properties of
the discretization employed in the simultaneous strategy break down for nonlinear
DAEs which have index > 2 (at least locally; for example, during an activation
of a path inequality constraint) [90]. Attempts to handle state path constraints
in the sequential strategy have relied on including some measure of their violation
indirectly as point constraints in the NLP, rather than directly in the IVP subproblem
[84, 133, 142]. This method was used because it was previously not possible to solve
numerically the high-index DAEs which resulted from appending the path constraints
to the IVP problem. In fact, the use of a dynamic optimization formulation with such
point constraints has been proposed as a method for solving high-index DAEs [76],
although this method is extraordinarily costly.
This work focuses on the control parameterization method. This choice was based
on the fact that many problems in chemical engineering are modeled with large sys-
tems of equations, and the control parameterization method appears to be the most
easily applied and reliable method for these problems. The reasons for this include:

• Direct methods are more easily implemented than indirect methods because
there is no need to generate the additional costate equations. Generation of
the costate equations is a significant problem for large dynamic optimization
problems, and more than doubles the size of the DAE system that must be
solved.

• It is extremely difficult to find numerical solutions to the two-point boundary

value problem that indirect methods are dependent upon if mx is large. In fact,

53
 we were unable to find any examples in the literature of the use of indirect meth-
ods to solve a problem with mx > 50. On the other hand, there are reported
examples of the direct method being used to solve problems with thousands of
state variables [33].

• The sequential direct method results in fairly small dense NLPs, but potentially
large sparse IVPs. Simultaneous direct methods do not require the solution of
an IVP, but require large sparse NLPs. Research into large-scale NLP solution
methods is currently a very active research area but it is still very difficult to
solve large NLPs from poor initial guesses. On the other hand, solution of large
sparse IVPs has become a standard numerical mathematical tool, and it is much
easier to provide good initial guesses for the small number of parameters in the
master NLP.

• The developments of this thesis dramatically increases the efficiency of control

parameterization through the staggered corrector algorithm (see Chapter 4) and
develops novel methods to solve problems that include path constraints.

54
2.5 Optimality conditions for direct methods

The aim of direct methods is to transform the infinite-dimensional problem (2.147–

2.149) into a finite dimensional problem. This is accomplished by constraining the
control functions u(t) to lie in a subspace of function space that is characterized by
a finite number of parameters. In other words, the control trajectories are described
by some function:

u = u(p, t) (2.146)

where p ∈ Rnp is a set of parameters. Hence, as shown in [63] the solution of (2.23–
2.26) found by direct methods will be suboptimal to those found by an analytic
solution to the necessary conditions (2.4–2.6).

The use of control parameterization turns the dynamic optimization problem into
a parameter optimization problem:
 tf
min J = ψ (x (p, tf ) , tf ) + L (x(p, t), u(p, t), t) dt (2.147)
p,tf t0

subject to the DAE:

f (ẋ(p, t), x(p, t), u(p, t), t) = 0 ∀t ∈ [t0 , tf ] (2.148)

φ (ẋ (t0 ) , x (t0 ) , t0 ) = 0 (2.149)

which may be solved as a mathematical program.

The necessary optimality conditions for (2.147–2.149) may be derived using La-
grange multipliers [16]:
  
∂J ∂ ẋ ∂J ∂x ∂J ∂u T ∂f ∂ ẋ ∂f ∂x ∂f ∂u
+ + + λ + + =0 (2.150)
∂ ẋ ∂p ∂x ∂p ∂u ∂p ∂ ẋ ∂p ∂x ∂p ∂u ∂p ∀t∈[t0 ,tf ]

f (ẋ(p, t), x(p, t), u(p, t), t) = 0 ∀t ∈ [t0 , tf ] (2.151)

φ (ẋ (t0 ) , x (t0 ) , t0 ) = 0 (2.152)

55
Note that this problem is still infinite dimensional because (2.150–2.151) are enforced
for t ∈ [t0 , tf ].
The simultaneous direct method approximates (2.150–2.151) at discrete points in
the interval [t0 , tf ]. Doing this requires that the state variables as well as the control
variables are parameterized using:

x = x(p̄, t) (2.153)

ẋ = x (p̄, t) (2.154)

These state variable approximations are then substituted into (2.150–2.151), and the
result evaluated at a set of discrete points along the solution trajectory.
The sequential direct method creates a finite dimensional problem from (2.150–
2.152) by transforming it into the following problem:

∂J ∂ ẋ ∂J ∂x ∂J ∂u
+ + =0 (2.155)
∂ ẋ ∂p ∂x ∂p ∂u ∂p
∂f ∂ ẋ ∂f ∂x ∂f ∂u
+ + = 0 ∀t ∈ [t0 , tf ] (2.156)
∂ ẋ ∂p ∂x ∂p ∂u ∂p
 
∂φ ∂ ẋ ∂φ ∂x
+ =0 (2.157)
∂ ẋ ∂p ∂x ∂p t=t0

f (ẋ(p, t), x(p, t), u(p, t), t) = 0 ∀t ∈ [t0 , tf ] (2.158)

φ (ẋ (t0 ) , x (t0 ) , t0 ) = 0 (2.159)

An IVP solver is used to discretize (2.156–2.159).

A diagram illustrating the general algorithm for solving a dynamic optimization
problem using the sequential direct method with an NLP solver is shown in Figure 2-1.

56
 Objective
Gradients
Function

Sensitivities State Variables

NLP Solver
Sets p
∂x(t)/∂p

x(t)

Control Variables

Integrate
u(p, t)

Dynamic
System
f (ẋ, x, u, t) = 0

Figure 2-1: The control parameterization algorithm

57
2.6 A general control parameterization problem

The control parameterization problem (2.155–2.159) presented in the previous section

was used to describe how the method works. This section presents a more general
control parameterization formulation that allows the solution of a broad range of
dynamic optimization problems.

2.6.1 Lagrange approximation of control variables

The control parameterization method approximates the control profiles using poly-
nomials over a set of NF E finite elements of length hi = ti − ti−1 , i = 1 . . . NF E .
For convenience in bounding the control profiles, the controls are parameterized us-
ing Lagrange polynomials. For control variable uj in finite element i the Lagrange
approximation is:

M +1
(M )  
uij (τ (i) ) = uijk φk τ (i) ∀τ (i) ∈ [0, 1] (2.160)
k=1

i = 1 . . . NF E j = 1 . . . mu (2.161)

where:
⎧
⎪
⎪
⎪
⎪ 0 if M = 0
  ⎨
= 
(M ) (i) M +1
φk τ τ − τl (2.162)
⎪
⎪ if M ≥ 0
⎪
⎩ l=1 τk − τl
⎪
l=k

and τ (i) is the normalized time in the finite element i:

t − ti−1
τ (i) (t) = (2.163)
ti − ti−1

The main advantage of Lagrange polynomials is that

uij (τ (i) ) = uijk if τ (i) (t) = τl (2.164)

58
Therefore, bounding the control parameters uijk is sufficient to bound the control
functions at the points τl . Bounding the profile at these points is sufficient to guar-
antee bounding of the entire control profile for M ≤ 1. It is possible to bound control
profiles with M > 1, but this requires the use of more complicated constraints that are
functions of the parameters. In practice, however, bounding higher order polynomials
can be done by ensuring that NF E is large.

Theoretically, the choice of the set of points τl does not affect the solution because
all polynomial approximations of the same order are equivalent. However, a poor
choice for τl can result in an ill-conditioned approximation of the control function.
Ideally, τ1 = 0 and, if M ≥ 0, τM +1 = 1 so that the control may be more easily
bounded. The other τl may be set at collocation points or at equidistant intervals.
The ‘best-conditioned’ approximation is given by selecting collocation points, but in
practice equidistant points are easier to use and do not seem to give a noticeable
decrease in the performance of the algorithm.

2.6.2 The dynamic optimization problem

The goal here is not to present the most general formulation of the control param-
eterized problem, but rather to show a formulation that is general enough for the
purposes of this thesis. Note that this formulation does not include path constraints
on state variables, which are discussed in Chapter 6.

The control parameterization formulation used in this thesis is:

min J (ẋ(uijk , ti ), x(uijk , ti ), u(uijk , ti ), v, tf ) (2.165)

uijk ,ti

subject to:

59
Initial time point constraints:

φ(ẋ(u111 ), x(u111 ), u(u111 ), t0 ) = 0 (2.166)

Final time point constraints:

r eq (ẋ(uijk , ti ), x(uijk , ti ), u(uijk , ti ), tf ) = 0 r eq (·) → Rnreq (2.167)

r ineq (ẋ(uijk , ti ), x(uijk , ti ), u(uijk , ti ), tf ) ≤ 0 r ineq (·) → Rnrineq (2.168)

Control parameter bounds:

uLijk ≤ uijk ≤ uUijk (2.169)

Time invariant bounds:

vL ≤ v ≤ vU v ∈ Rmv (2.170)

Finite element size constraints:

hmin ≤ ti − ti−1 ≤ hmax (2.171)

tNF E = tf (2.172)

The values of the variables are obtained by solving the DAE:

f (ẋ(uijk , ti ), x(uijk , ti ), u(uijk , ti , t), v, t) = 0 (2.173)

subject to the initial conditions (2.166), the control parameterization (2.160).

Transferring the values of the state variables from one finite element to the next re-
quires junction conditions. On the boundaries between finite elements, these junction

60
conditions have the form:

Φi (ẋ(uijk , tj=i , t− − − −
i ), x(uijk , tj=i , ti ), u(uijk , tj=i , ti ), ti ,
(2.174)
ẋ(uijk , tj=i , t+ + + +
i ), x(uijk , tj=i , ti ), u(uijk , tj=i , ti ), ti ) = 0

where Φ(·) → Rrd .

For most physically meaningful ODEs and index-1 DAEs the junction conditions
(2.174) have the form

−
x+
n = xn ∀n ∈ Γ (2.175)

where Γ is an index set of the state variables which have time derivatives that explicitly
appear in the DAE. However, for high-index DAEs or DAEs where the index fluctuates
between boundaries equation (2.175) is not valid, and methods such as those described
in [141] must be used to determine appropriate forms of the transfer conditions.
The control parameterization formulation presented above is simpler than the
multi-stage formulation given in [142]. The multi-stage formulation allows the time
domain to be divided into subdomains which have end times that may or may not
coincide with finite element boundaries. The functional form of the DAE is permitted
to change from one stage to the next. However, it is worth pointing out that the multi-
stage formulation requires the stage sequence to be fixed, rather than be defined by
implicit state events as in [15, 110]. The significance of this limitation is further
discussed in Chapter 7.

2.6.3 Gradient evaluation

Efficient control parameterization requires the gradients of the objective functions

and NLP constraints with respect to the optimization parameters. Evaluating these
gradients requires values for the sensitivity variables ∂ ẋ
∂p
and ∂x
∂p
, where p = {uijk , ti , v}.
There are three basic methods for obtaining the sensitivity variables: finite dif-
ferences, adjoint equations, and direct solution of the sensitivity equation. Of these,
direct solution of the sensitivity equation is the preferred method because of its accu-

61
racy and the high computational efficiency that may now be achieved when solving
the combined DAE and sensitivity system [49, 94, 123, Chapter 4].
The sensitivity equations associated with (2.173) are obtained by differentiating
(2.173) with respect to the optimization parameters p:

∂f ∂ ẋ ∂f ∂x ∂f ∂u(p, t) ∂f
+ =− − , ∀t = [t0 , tf ] (2.176)
∂ ẋ ∂p ∂x ∂p ∂u ∂p ∂p

which is in itself a DAE that has the initial conditions:

∂φ ∂ ẋ ∂φ ∂x ∂φ ∂u(p, t0 ) ∂φ
+ =− − , t = t0 (2.177)
∂ ẋ ∂p ∂x ∂p ∂u ∂p ∂p

There are also junction conditions for the sensitivity equation that are valid at the
same times ti as the junction conditions for the DAE, which are discussed in Chapter 3.
Solving (2.176–2.177) requires the partial derivatives of the control functions with
respect to the parameters, which are [142]:

∂uij
= φM i
k (τ )δii δjj  k = 1...M + 1 (2.178)
∂ui j  k
⎧
⎪
⎪ ti − t
M +1
dφ
(M )
⎪
⎪− uijk k i for i = i − 1, i > 1
⎪
⎪ (ti − ti−1 )2
⎪
⎪ dτ
∂uji ⎨ k=1
M +1 (M )
= − t − ti−1 dφ
uijk k i for i = i (2.179)
∂ti ⎪
⎪
⎪
⎪ (ti − ti−1 )2 dτ
⎪
⎪
k=1
⎪
⎪
⎩0 otherwise

62
2.7 ABACUSS Dynamic Optimization Input Lan-
guage
State variable path-constrained optimization has been implemented within the ABA-
CUSS1 large-scale equation-oriented modeling system. The input language in ABA-
CUSS has been extended to permit the representation of large-scale path-constrained
dynamic optimization systems in a compact manner.
Examples of the input language are given in Figure 2-2 and Appendix A. Variable
types are declared in the DECLARE block. The equations describing the DAE are
given in the MODEL block. The OPTIMIZATION block is used to define a dynamic
optimization using the MODEL. Note that it is possible to move from a simulation
to an optimization in a seamless manner.
The sections in the optimization block are:

PARAMETER Defines the parameters that are used in the dynamic optimization
that are not defined in the model.

UNIT Defines the MODELs to use in the current dynamic optimization.

VARIABLE Defines variables that are used in the dynamic optimization that are
not defined in the models.

OBJECTIVE Defines the objective function to be MINIMIZEd or MAXIMIZEd in

terms of variables defined in the UNIT or the VARIABLE section.

SET Sets values for any parameters defined in the UNIT or PARAMETER section

INEQUALITY Defines any path inequalities in terms of variables defined in the

UNIT or the VARIABLE section.

CONTROL Indicates that the listed variables are control variables. The first func-
tion in the triple notation defines the initial guess, the second and third define
the lower and upper bounds for the control, respectively.

TIME INVARIANT Indicates that the listed variable are time invariant optimiza-
tion parameters. The first number in the triple notation defines the initial guess,

1
ABACUSS (Advanced Batch And Continuous Unsteady-State Simulator) process modeling soft-
ware, a derivative work of gPROMS software, 1992
c by Imperial College of Science, Technology,
and Medicine.

63
 DECLARE OBJECTIVE
TYPE MINIMIZE Final_Time
NoType = 0.0 : -1E9 : 1E9
SET
END WITHIN kraft DO
G := -1.0
MODEL brachistochrone END
A := -0.40
PARAMETER B := 0.30
G AS REAL
INEQUALITY
VARIABLE WITHIN kraft DO
X AS NoType Y>=A*X-B
Y AS NoType END
V AS NoType
Theta AS NoType CONTROL
Constr AS NoType WITHIN kraft DO
Theta := -1.6+1.6/0.7*TIME: -1.6: 0.0
EQUATION END

$X = V*COS(Theta) TIME_INVARIANT
$Y = V*SIN(Theta) Final_Time := 0.68 : 0.0 : 2.0
$V = G*SIN(Theta)
Constr = -0.40*X-0.30 INITIAL
WITHIN kraft DO
END # brachistochrone X=0.0
Y=0.0
OPTIMIZATION Brach V=0.0
END
PARAMETER
B AS REAL FINAL
A AS REAL WITHIN kraft DO
X=1.1
UNIT END
kraft AS brachistochrone
SCHEDULE
VARIABLE CONTINUE FOR Final_Time
Final_Time AS NoType
END # Brach

Figure 2-2: Example of ABACUSS input file for constrained dynamic optimization

the second and third define the lower and upper bounds for the variable, re-
spectively.

INITIAL Gives additional equations that define the initial condition for the prob-
lem, possibly as a function of the optimization parameters.

FINAL Defines point constraints that must be obeyed at the final time.

SCHEDULE Indicates the time domain of the dynamic optimization problem.

64
Chapter 3

Parametric Sensitivity Functions

for Hybrid Discrete/Continuous
Systems

This chapter is a summary of a paper [55] of the same title that was coauthored
with Santos Galán. The results of this paper are included here because the ability
to handle sensitivity functions for hybrid discrete/continuous systems is necessary
for the development of the methods to handle inequality path-constrained dynamic
optimization problems discussed in Chapter 7.
Parametric sensitivity analysis is the study of the influence of variations in the
parameters of a model on its solution. It plays an important role in design and mod-
eling, is used for parameter estimation and optimization, and is extensively applied
in the synthesis and analysis of control systems.
The variations of the parameters can be differential or finite. Here only the former
case is considered, which approximates linearly the variation of the solution. Two
kinds of sensitivity analysis can be distinguished, depending on whether the variation
is finite-dimensional (lumped) or it is a function (distributed). In the former case,
which is the only one considered in this chapter, the sensitivities are ordinary partial
derivatives. In the second, as in structural systems, functional derivatives are needed.
In process design, sensitivity information provides an elucidation of the influence

65
of design changes, without requiring trial and error. In modeling, sensitivity analysis
reveals the relative importance of every parameter, giving arguments for simplify-
ing the model or guiding new experiments. Sensitivities are also used in parameter
estimation for error analysis, and in gradient computation for dynamic optimization.

There has been great interest in the application of sensitivity information in con-
trol system design. Sensitivity analysis is necessary because the parameters are sub-
ject to inaccuracy in data, models and implementation, and because they can deviate
with time. Furthermore, sensitivity analysis is essential for adaptive systems. Early
reviews in this area are given in [83, 136].

Here, the term “hybrid” will refer to the combined existence and interaction of
continuous and discrete phenomena. The continuous part is usually modeled by
differential-algebraic equations (DAEs) and the discrete behavior by finite automata.
A precise definition of the systems considered in this chapter follows in Section 3.1.
Hybrid systems pose a problem for the calculation of sensitivities because the sensi-
tivities are not defined in general when changing from one continuous subsystem to
another.

The study of the sensitivity of control systems experienced intense development in

the late 1960s, mostly in Eastern Europe. Some authors have dealt with the subject of
sensitivities in discontinuous systems, which are frequent in control systems. Tsypkin
claims in a short note [138] to have obtained variational equations for relay systems
in 1955. Using an obscure variational notation, De Backer [35] derives the “jump
conditions” for the sensitivities with respect to the initial conditions of the initial
value problem (IVP) in ordinary differential equations (ODEs) of an autonomous
system that changes to a second vector field at a time when a state event (i.e., an
event which is a function only of the state variables) is satisfied. With the intention of
extending system identification to systems with discontinuities, Russ [125] rediscovers
De Backer’s result. While Russ broadened De Backer’s result to cover nonautonomous
systems and state events that are functions of time and the parameters, there are
several drawbacks to this methodology: the derivation is unnecessarily complex, the
representation is inconvenient, and hybrid phenomena are not well understood, as

66
evidenced by the imposition of an obviously incorrect anullation of components of
the gradient of the discontinuity function at the switching times.
In a concise paper, Rozenvasser [124] derived the sensitivity functions of discontin-
uous systems modeled by a given sequence of explicit ODE vector fields with explicit
or implicit switching times. Compared to the above works, it seems that the deriva-
tion is a relatively straightforward calculus exercise. This is the most general result
we know for sensitivities of hybrid systems and ironically, until now, has been en-
tirely neglected in the subsequent literature. In fact, we derived Rozenvasser’s results
independently, only subsequently stumbling upon [124] by chance.
This chapter extends Rozenvasser’s results in several ways. First, the discrete
aspects of the system model are significantly generalized in line with modern notions
of hybrid systems. Second, the results are extended to include DAE embedded hybrid
systems. Existence and uniqueness theorems are also presented for the sensitivity
functions of hybrid systems. These theorems shed light on the issue of sequencing
of state events in hybrid systems. Numerical results are given for the calculation of
sensitivity functions for hybrid systems.

67
3.1 Mathematical Representation of Hybrid Sys-
tems
A formalism derived from [6] and [15] is used to model a broad class of hybrid discrete/
 k
continuous phenomena. Consider a system described by a state space S = nk=1 Sk
where each mode Sk which is characterized by:

(k)
1. A set of variables {ẋ(k) , x(k) , y (k) , u(k), p, t}, where x(k) (p, t) ∈ Rnx are the
(k)
differential state variables, y (k)(p, t) ∈ R ny
the algebraic state variables and
(k)
u(k) (p, t) ∈ R nu
the controls. The time invariant parameters p ∈ Rnp and time
t are the independent variables, and the controls u are explicit functions of both
the parameters and time.

2. A set of equations f (k) (ẋ(k) , x(k) , y (k) , u(k) , p, t) = 0, usually a coupled system
(k) (k) (k) (k)
of differential and algebraic equations, f (k) : Rnx × Rnx × Rny × Rnu ×
(k) (k)
Rn p × R → Rn x +ny
. In the mode Sk the specification of the parameters
p (and, consequently, the controls) coupled with a consistent initial condition
(k)
Tk (ẋ(k) , x(k) , y (k) , u(k), p, t) = 0 at t = t0 determines the evolution of the system
(k) (k)
in [t0 , tf ).

3. A (possibly empty) set of transitions to other modes. The set of modes Sj where
a transition from mode Sk is possible is J (k) . These transitions are described
by:

(k)
(a) Transition conditions Lj (ẋ(k) , x(k) , y (k), u(k) , p, t), j ∈ J (k) , determining
the transition times at which switching from mode k to mode j occurs. The
transition conditions are represented by logical propositions that trigger
the switching when they become true. They are described in section 3.3.
Note that discontinuities in the controls are included here.

(b) Sets of transition functions:

(k)
Tj (ẋ(k) , x(k) , y (k) , u(k), ẋ(k+1) , x(k+1) , y (k+1) , u(k+1) , p, t)

68
are associated with the transition conditions which relate the variables in
the current mode Sk and the variables in the new mode Sj at the transition
(k)
time tf . A special case of the transition functions is the set of initial con-
ditions for the initial mode S1 . These initial conditions will be designated
(1)
by T0 .

69
3.2 Sensitivities

The partial derivatives of the variables with respect to the parameters are known as
sensitivity functions. Before discussing the sensitivities, the solution of the hybrid
system in isolation must be examined in more detail.

3.2.1 Consistent initialization

(1)
At t0 , the following system exists:

f (1) (ẋ(1) , x(1) , y (1) , u(1) , p, t) = 0 (3.1)

(1)
T0 (ẋ(1) , x(1) , y (1) , u(1) , p, t) = 0 (3.2)
(1)
t = t0 (3.3)

(1) (1) (1) (1) (1) (1)

where T0 : Rnx × Rnx × Rny × Rnu × Rnp × R → Rnx . For the purposes of this
chapter it is assumed that:

rank ∂f (k) ∂f (k) = nx(k) + ny(k) (3.4)

∂ ẋ(k) ∂y (k)

everywhere. This condition holds for all index 0 systems (ODEs) and most index 1
(1) (1)
systems, and implies that nx side conditions T0 are required to define uniquely a
consistent initial condition [15]. Chapter 5 details methods for deriving an equivalent
index-1 DAE from a high-index DAE that work for a broad class of high-index DAEs.
Equation (3.3) is usual at initialization, but in general the initial time can be
(1)
an implicit or explicit function of other parameters. Actually, t0 is a parameter.
The more general case is considered below when dealing with the sensitivities at
transitions.
(1) (1) (1) (1)
The number of equations nx + nx + ny + 1 and the number of variables nx +
(1) (1)
nx + ny + np + 2 (u are explicit functions of p and t) admit in general np + 1 degrees
(1)
of freedom for specification of the independent variables p and t0 .
A sufficient (implicit function theorem) and practical (numerical solution by a

70
Newton method) condition for the solvability of the system (3.1–3.3) is that the
matrix:
⎡ ⎤
∂f (1) ∂f (1) ∂f (1) ∂f (1) ⎡ ⎤
⎢ ∂ ẋ(1) ∂x(1) ∂y (1) ∂t ⎥ ∂f (1) ∂f (1) ∂f (1)
⎢ ∂T0(1) (1) (1)
∂T0 ⎥
(1)
⎣ ∂ ẋ(1) ∂x(1) ∂y (1) ⎦
⎢ ∂ ẋ(1) ∂T0 ∂T0
⎥ ⇐⇒ (1) (1) (1)
⎣ ∂x (1) ∂y (1) ∂t ⎦ ∂T0 ∂T0 ∂T0
∂ ẋ (1) ∂x (1) ∂y (1)
0 0 0 1

is nonsingular.

The solution of (3.1), (3.2) and (3.3) is represented as:

(1) (1) (1) (1) (1) (1) (1)

x0 (p, t0 ), ẋ0 (p, t0 ), y0 (p, t0 ), t0 (3.5)

and let:

(1) (1)
x(1) (p, t0 , t), y (1) (p, t0 , t) (3.6)

be the solutions of f (1) that satisfy (3.1), (3.2) and (3.3). These solutions are functions
of t that pass through the point (3.5). The following relations can be derived:

(1)
(1) ∂x(1) (p, t0 , t)
ẋ(1) (p, t0 , t) = (3.7)
∂t
(1) (1) (1) (1)
x(1) (p, t0 , t0 ) = x0 (p, t0 ) (3.8)
(1) (1) (1) (1)
y (1) (p, t0 , t0 ) = y0 (p, t0 ) (3.9)

∂x(1) (p, t0 , t) 
(1)
(1) (1) (1) (1) (1)
ẋ (p, t0 , t0 ) =  = ẋ0 (p, t0 ) (3.10)
∂t t=t
(1)
0

71
3.2.2 Initial sensitivities

Now consider sensitivity functions of the above system. Differentiating the system
used for consistent initialization and applying the chain rule yields:
⎡ (1) (1)
⎤
∂ ẋ0 ∂ ẋ0
⎢ ∂p ∂t0 ⎥
(1)
⎢ ∂x(1) (1) ⎥
⎢ 0 ∂x0 ⎥
⎡ ⎤⎢ ∂p
⎢ (1) ∂t0 ⎥
(1)
⎥
∂f (1) ∂f (1) ∂f (1) ∂f (1) ∂f (1) ∂f (1) ∂f (1) ⎢ ∂y0 ∂y0 ⎥
(1)

⎢ ∂ ẋ(1) ∂x(1) ∂y (1) ∂u(1) (1) ⎢

∂t0 ⎥⎢ ∂p
⎥
⎢ ∂T (1)
∂t ∂p
(1) ⎥⎢
∂t0
(1)
⎥
⎢ 0
(1) (1) (1) (1) (1)
∂T0 ⎥ ∂u(1) ⎥
⎢ ∂ ẋ(1)
∂T0 ∂T0 ∂T0 ∂T0 ∂T0
⎢ ∂u(1) ⎥ = 0 (3.11)
∂x(1) ∂y (1) ∂u(1) ∂t0 ⎥⎢ ∂p ∂t0 ⎥
∂t ∂p (1) (1)
⎣ ⎦⎢ ⎥
0 0 0 0 1 0 −1 ⎢ ∂t
⎢ ∂p
∂t ⎥
∂t0 ⎥
(1)
⎢ ⎥
⎢ ⎥
⎢ I 0 ⎥
⎣ ⎦
0 1

From the last row:

∂t ∂t
=0 (1)
=1 (3.12)
∂p ∂t0

(1)
∂f (1) ∂T0
Since (1) = 0 and (1) = 0, the system is:
∂t0 ∂t0

⎡ (1) (1)
⎤
∂ ẋ ∂ ẋ0
⎡ ⎤ ∂p0 ⎡ ⎤
∂f (1) ∂f (1) ∂f (1)⎢ ∂t0 ⎥
(1)
∂f (1) ∂u(1) ∂f (1) ∂f (1) ∂u(1) ∂f (1)
⎢ (1)
∂y (1) ⎦ ⎢ ∂x0
(1) ⎥
∂x0 ⎥ + +
⎣ ∂ ẋ(1) ∂x(1) ⎣ ∂u(1) ∂p
− ∂T ∂p ∂u(1) ∂t ∂t ⎦
∂T0
(1)
∂T0
(1)
∂T0
(1) ⎢ ∂p (1) ⎥= (1)
∂u(1) ∂T0
(1) (1)
∂T0 ∂u(1) ∂T0
(1)

∂ ẋ (1) ∂x (1) ∂y (1)

⎣ (1) ∂t0
(1) ⎦ ∂u
0
(1) ∂p
+ ∂p ∂u (1) ∂t
+ ∂t
∂y0 ∂y0
∂p (1)
∂t0

(3.13)

Note that the sensitivities of the solution of the consistent initialization problem
(3.5) have been defined. The initial sensitivities of the dynamic problem must also
be determined, which may or may not be the same. For the parameters p the initial

72
sensitivities are:

∂x(1) (p, t0 , t) 
(1) (1) (1) (1) (1)
∂x0 (p, t0 ) ∂x(1) (p, t0 , t0 )
= =  (1) (3.14)
∂p ∂p ∂p t=t0

∂y (1) (p, t0 , t) 
(1) (1) (1) (1) (1)
∂y0 (p, t0 ) ∂y (1) (p, t0 , t0 )
= =  (1) (3.15)
∂p ∂p ∂p t=t0

∂ ẋ(1) (p, t0 , t) 
(1) (1) (1) (1) (1)
∂ ẋ0 (p, t0 ) ∂ ẋ(1) (p, t0 , t0 )
= =  (1) (3.16)
∂p ∂p ∂p t=t 0

but for the sensitivities with respect to the initial time:

(1) (1) (1) (1)

∂x0 (p, t0 ) ∂x(1) (p, t0 , t0 )
(1)
= (1)
=
∂t0 ∂t0
 
∂x(1) (p, t0 , t)  ∂x(1) (p, t0 , t) 
(1) (1)

∂t  (1) + (1)  (1) =

t=t0 ∂t0 t=t0

∂x(1) (p, t0 , t) 
(1)
(1) (1) (1)
ẋ (p, t0 , t0 ) + (1)  (1) (3.17)
∂t0 t=t0

Therefore:

∂x(1) (p, t0 , t) 
(1) (1) (1)
∂x0 (p, t0 ) (1) (1)
(1)  (1) = (1)
− ẋ(1) (p, t0 , t0 ) (3.18)
∂t0 t=t0 ∂t0

Similarly,


∂y (1) (p, t0 , t) 
(1) (1) (1)
∂y0 (p, t0 ) (1) (1)
(1)  (1) = (1)
− ẏ (1) (p, t0 , t0 ) (3.19)
∂t0 t=t0 ∂t0


∂ ẋ(1) (p, t0 , t) 
(1) (1) (1)
∂ ẋ0 (p, t0 ) (1) (1)
(1)  (1) = (1)
− ẍ(1) (p, t0 , t0 ) (3.20)
∂t0 t=t0 ∂t0

Hence, ẏ (1) and ẍ(1) must be known in order to determine the initial sensitivities.
The next section shows how ẏ and ẍ can be calculated at any time.

73
3.2.3 Sensitivity trajectories

(k) (k)
From a practical point of view, the trajectory of the variables for t ∈ (t0 , tf ) can
be computed by numerical integration of the system f (k) = 0 from the initial values.
(1)
Hereafter p will include all the parameters (including t0 ).

Differentiating the system with respect to the independent variables p and t gives:
⎡ ⎤
∂ ẋ(k) ∂ ẋ(k)
⎢ ∂p ∂t ⎥
∂f (k) ∂f (k) ∂f (k) ⎢ ∂x(k) ∂x(k) ⎥ ∂f (k) ∂u(k) ∂f (k) ∂f (k) ∂u(k) ∂f (k)
⎢ ∂p ⎥ =− + + (3.21)
∂ ẋ(k) ∂x(k) ∂y (k) ⎣ ∂t ⎦ ∂u(k) ∂p ∂p ∂u(k) ∂t ∂t
∂y (k) ∂y (k)
∂p ∂t

Then, representing the matrix of sensitivities by s, under conditions of continuity of

the time derivative of the state variable sensitivities and independence between t and
p, define:

∂x(k)
sx(k) = (3.22)
∂p
(k)
∂sx ∂ ∂x(k) ∂ ∂x(k) ∂ ẋ(k)
ṡx(k) = = ( )= ( )= (3.23)
∂t ∂t ∂p ∂p ∂t ∂p
(k)
∂y
sy(k) = (3.24)
∂p

Sensitivity trajectories are determined by integrating the differential linear time

varying system:
⎡ ⎤
(k)
ṡx
⎢ ⎥
∂f (k) ∂f (k) ∂f (k) ⎢ (k) ⎥ ∂f (k) ∂u(k) ∂f (k)
⎢ sx ⎥ = − + (3.25)
∂ ẋ(k) ∂x(k) ∂y (k) ⎣ ⎦ ∂u(k) ∂p ∂p
(k)
sy

from the initial sensitivities.

The derivative with respect to time is a linear but not differential system (i.e.ẋ(k)
is known) that allows us to obtain the the values of ẏ (k) and ẍ(k) required in the

74
previous section and below:
⎡ ⎤
(k)
ẍ
∂f (k) ∂f (k) ⎣ ⎦=− ∂f (k) (k)
ẋ + ∂f (k) ∂u(k)
+ ∂f (k) (3.26)
∂ ẋ(k) ∂y (k) ∂x(k) ∂u(k) ∂t ∂t
ẏ (k)

75
 D1 I1 V3 I3 I2 D2

R1L1 R3L3 R2L2

V1 V2

Figure 3-1: Rectifier circuit

3.3 Transitions

(k)
The transition conditions Lj (ẋ(k) , x(k) , y (k), u(k) , p, t) , j ∈ J (k) , are formed by logical
propositions that contain logical operators (e.g.NOT, AND, OR) connecting atomic
propositions (i.e.relational expressions) composed of valid real expressions and the
relational operators {>, <, ≤, ≥}. For example, in the rectifier circuit of Figure 3-1
[31], the condition for both diodes D1 and D2 to be conductive requires the following
logical proposition to be true:

[(i1 > 0) ∨ (v1 > v3 )] ∧ [(i2 > 0) ∨ (v2 > v3 )] (3.27)

Associated with every relational expression is a discontinuity function:

(k) (k)
gji (ẋ(k) , x(k) , y (k), u(k) , p, t) i = 1, . . . , nj

formed by the difference between the two real expressions. For example, in the rectifier
system:

(1)
g2,4 = v2 − v3 (3.28)

Each relational expression changes its value whenever its corresponding discontinuity
function crosses zero. The transition conditions neutralize the only degree of freedom
(k)
(time), determining tf . The set L(k) of all transition conditions for the mode Sk
defines the boundary of Sk that, when intercepted by the trajectory, triggers the
switching to a new mode. Notice that the system can be in the same mode after the

76
application of a transition function.
Several issues arise in connection with the boundary created by the transition
conditions that are beyond the scope of this chapter. Some of them are discussed
in [69]. Here smoothness is assumed in the neighborhood of the transition time and
(k)
that only one relational expression activates at that moment. Hereafter, gj will be
used to designate the discontinuity function that actually determines the transition
from mode Sk to Sj .

3.3.1 Time of the event

In practice, the satisfaction of the transition conditions (events) is monitored during

(k)
numerical integration so that the smallest tf for all the possible transitions in the
current mode, which establishes the actual transition, is easily found. An extensive
treatment of this subject appears in [110].
Figure 3-2 illustrates different classes of events. The simplest condition is that of
the controlled transitions (Π1 ), where only independent variables (necessarily time,
as it is the only degree of freedom) appear. This case is known as a time event.
Traditionally, if the transition functions involve discontinuities in the state variables,
then the transitions are called controlled jumps, usually reverting to the same mode.
If there is a change of mode this is termed controlled switching. Controlled switching
is the case of discontinuities in the controls.
Autonomous transitions are those whose conditions include the dependent vari-
ables. The satisfaction of these conditions is called a state event. A further distinction
is whether time appears explicitly (Π3 ) or not (Π2 ). The same distinction between
switchings and jumps can be made.
The evolution of a hybrid system can be viewed as a sequence of subdomains of
the time interval, where each subdomain is characterized by a continuous evolution,
an event, and a transition to a new mode (which can be the same). For convenience,
the use of the superscript (k) and the subscript (k + 1) will indicate, respectively, the
mode in which the event occurred and the subsequent mode which is determined by
the condition.

77
 x

Figure 3-2: Controlled and autonomous transitions

(k)
The transfer to the new mode k + 1 is described by the transition function Tk+1 .
At the event following system exists:

ẋ(k) = ẋ(k) (p, t) x(k) = x(k) (p, t) y (k) = y (k)(p, t) (3.29)

(k)
gk+1 (ẋ(k) , x(k) , y (k) , u(k), p, t) = 0 (3.30)
(k)
Tk+1 (ẋ(k) , x(k) , y (k), u(k) , ẋ(k+1) , x(k+1) , y (k+1), u(k+1) , p, t) = 0 (3.31)

f (k+1) (ẋ(k+1) , x(k+1) , y (k+1), u(k+1) , p, t) = 0 (3.32)

This system may have multiple solutions, but for simplicity it is assumed that criteria
are given that define a unique solution (for example, limiting the domain of the
transition function).

The equations (3.29) represent the state trajectory resulting from the integration
of f (k) in mode k. In practice they are calculated numerically. The system naturally
decomposes into two structural blocks that can be solved sequentially. The first
(k) (k+1)
two equations determine t = tf = t0 (i.e., time is a dependent variable), and
the corresponding values for ẋ(k) , x(k) , y (k) and u(k) . The last two equations allows
calculation of the initial values of ẋ(k+1) , x(k+1) , y (k+1) , and u(k+1) in the new mode.

Again, a sufficient and practical condition for solving the system formed by (3.29)

78
and (3.30) is the nonsingularity of the Jacobian:
⎡ ⎤
(k)
1 0 0 − ∂ ẋ∂t
⎢ ⎥
⎢ (k) ⎥
⎢ 0 1 0 − ∂x∂t ⎥
⎢ ⎥
⎢ (k) ⎥
⎢ 0 0 1 − ∂y∂t ⎥
⎣ (k) (k) (k) (k) (k)
⎦
∂gk+1 ∂gk+1 ∂gk+1 ∂gk+1 ∂u(k) ∂gk+1
∂ ẋ(k) ∂x(k) ∂y (k) ∂u(k) ∂t
+ ∂t

That is, the following determinant is not zero:

(k) (k) (k) (k) (k)

∂gk+1 ∂ ẋ(k) ∂gk+1 ∂x(k) ∂gk+1 ∂y (k) ∂gk+1 ∂u(k) ∂gk+1
+ + + + = 0 (3.33)
∂ ẋ(k) ∂t ∂x(k) ∂t ∂y (k) ∂t ∂u(k) ∂t ∂t

(k)
The derivatives must exist and the trajectory should not be tangent to gk+1 in the
(k) (k) (k)
neighborhood of tf . If the trajectory is tangent to gk+1 in the neighborhood of tf ,
small variations of the parameters may not lead to a unique solution for the time of
the transition. As shown later in this chapter, the points where (3.33) is not satisfied
play an important role for hybrid systems.

3.3.2 Reinitialization at the transition

The initialization of the new mode is a problem similar to the one described in Sec-
tion 3.2.1. The same assumptions for the transition functions T are made, providing
a condition for well-posed transitions. Again, the nonsingularity condition applies:
⎛⎡ (k) (k) (k)
⎤⎞
∂Tk+1 ∂Tk+1 ∂Tk+1

rank ⎝⎣ ∂ ẋ(k+1) ∂y (k+1) ⎦⎠

(k+1) ∂x(k+1)
= nx(k+1) + nx(k+1) + ny(k+1) (3.34)
∂f ∂f (k+1) ∂f (k+1)
∂ ẋ(k+1) ∂x(k+1) ∂y (k+1)

Using the same notation applied above, the solution of the reinitialization is expressed
as:

(k+1) (k+1) (k+1) (k+1)

x0 (p), ẋ0 (p), y0 (p), t0 (p) (3.35)

79
3.3.3 Sensitivity transfer at transitions

The discontinuity system is differentiated for evaluating the sensitivities. The only
degrees of freedom are the time-invariant parameters. The structural decomposition
is used again to solve the system in two steps.

First, (3.29) and (3.30) are differentiated. Differentiation of (3.29) indicates that
the sensitivities are the ones calculated in the mode k. Differentiation of (3.30) yields
a system that can be solved for the switching time sensitivities:

(k)   (k)   (k)  

∂gk+1 (k) (k) ∂t ∂gk+1 (k) (k) ∂t ∂gk+1 (k) (k) ∂t
ṡx + ẍ + sx + ẋ + sy + ẏ
∂ ẋ(k) ∂p ∂x(k) ∂p ∂y (k) ∂p
(k)   (k) (k)
∂g ∂u(k) ∂u(k) ∂t ∂g ∂g ∂t
+ k+1 (k)
+ + k+1 + k+1 = 0 (3.36)
∂u ∂p ∂t ∂p ∂p ∂t ∂p

(k+1)
Note that in the case of a controlled transition with the logical condition t ≥ t0 ,
the solution is similar to that obtained in the initialization (3.12).

Second, the system formed by (3.31) and (3.32) is differentiated. Applying the
chain rule:
⎡ (k) ⎤T⎡ ⎤
∂Tk+1
0 ∂ ẋ(k) ∂ ẋ(k)
⎢ ∂ ẋ(k)
(k)
⎥ ⎢ ∂p ∂t ⎥
⎢ ∂Tk+1 ⎥ ⎢ ∂x(k) ⎥
⎢ ∂x(k) 0 ⎥ ⎢ ∂p ∂x(k)
⎥
⎢ (k) ⎥ ⎢ ∂t ⎥
⎢ ∂Tk+1 ⎥ ⎢ ∂y(k) ∂y (k) ⎥
⎢ (k) 0 ⎥ ⎢ ∂p ⎥
⎢ ∂y(k) ⎥ ⎢ ∂t ⎥
⎢ ∂Tk+1 ⎥ ⎢ ∂u(k) ⎥
⎢ (k) 0 ⎥ ⎢ ∂p ∂u (k)
⎥
⎢ ∂u ⎥ ⎢ ∂t ⎥⎡ ⎤
⎢ ∂T (k) ⎥ ⎢ ∂ ẋ(k+1)
⎢ k+1 ∂f (k+1) ⎥ ∂ ẋ(k+1) ⎥
⎢ ∂ ẋ(k+1) ⎥ ⎢ ∂p ⎥ I
⎢ ∂T (k) ∂ ẋ (k+1)
⎥ ⎢ ∂t ⎥⎣ ⎦
=0 (3.37)
⎢ k+1 (k+1) ⎥ ⎢ ∂x(k+1) ∂x(k+1) ⎥
⎢ ∂x(k+1)
∂f ⎢ ∂p ⎥ ∂t
∂x(k+1) ⎥ ⎢ ∂t ⎥ ∂p
⎢ (k) ⎥ ⎢ ∂y(k+1)
⎢ ∂Tk+1 ∂f (k+1) ⎥ ∂y (k+1) ⎥
⎢ ∂y(k+1) ⎥ ⎢ ∂p ⎥
⎢ (k) ∂y (k+1)
⎥ ⎢ ∂t ⎥
⎢ ∂Tk+1 ∂f (k+1) ⎥
⎢ ∂u(k+1) ∂u(k+1) ⎥
⎢ ∂u(k+1) ⎥ ⎢ ∂p ⎥
⎢ (k) ∂u(k+1) ⎥ ⎢ ∂t ⎥
⎢ ∂Tk+1 ⎢ ⎥
⎢ ∂p ∂f (k+1) ⎥ ⎥ ⎢ I 0 ⎥
⎣ (k) ∂p ⎦ ⎣ ⎦
∂Tk+1 ∂f (k+1) 0 1
∂t ∂t

80
Reordering to separate known and unknown variables yields:

⎡ (k) ⎤T ⎡ ⎤ ⎡ (k) ⎤T ⎡ ⎤
∂Tk+1 ∂f (k+1) ∂ ẋ(k+1) ∂Tk+1 ∂f (k+1) ∂ ẋ(k+1)
⎢ ∂ ẋ(k+1) ∂ ẋ(k+1) ⎥ ⎢ ∂p ⎥ ⎢ ∂ ẋ(k+1) ∂ ẋ(k+1) ⎥ ⎢ ∂t ⎥
⎢ ∂Tk+1(k)
∂f (k+1) ⎥ ⎢ ∂x(k+1) ⎥ ⎢ ∂Tk+1(k)
∂f (k+1) ⎥ ⎢ ∂x(k+1) ⎥ ∂t
⎢ ∂x(k+1) ⎥ ⎢ ∂p ⎥ = − ⎢ ∂x(k+1) ⎥ ⎢ ∂t ⎥
⎣ (k) ∂x(k+1) ⎦ ⎣ ⎦ ⎣ (k) ∂x(k+1) ⎦ ⎣ ⎦ ∂p
∂Tk+1 ∂f (k+1) ∂y (k+1) ∂Tk+1 ∂f (k+1) ∂y (k+1)
∂y (k+1) ∂y (k+1) ∂p ∂y (k+1) ∂y (k+1) ∂t
⎡ (k)
⎤T ⎡ ⎤
∂Tk+1
0 ⎥ ∂ ẋ(k) ∂ ẋ(k) ∂t
⎢ ∂ ẋ(k) +
⎥ ⎢ ⎥
∂p ∂t ∂p
⎢ (k)
⎢ ⎥
⎢ (k) ∂Tk+1
0 ⎥ ∂x(k) ∂x(k) ∂t
⎢ ∂x ⎥ ⎢⎢ ∂p
+ ∂t ∂p
⎥
⎥
⎢ ∂T (k) ⎥ ⎢
⎢ k+1 ⎥ ∂y (k) (k)
∂t ⎥
⎢ ∂y(k) 0 ⎥ ⎢ ∂p + ∂t ∂p ⎥
∂y
⎢ ∂T (k) ⎥ ⎢⎢ ∂u(k)
⎥
−⎢ ∂u(k) ∂t ⎥
⎢ ∂u(k)
k+1
0 ⎥ ⎥ ⎢ + ∂t ∂p ⎥ (3.38)
⎢ ∂T (k) ⎥ ⎢⎢
∂p ⎥
⎢ k+1 ∂f (k+1) ⎥ ⎢ ∂u(k+1) ∂u(k+1) ∂t ⎥
⎢ ∂u(k+1) + ∂t ∂p ⎥
⎢ (k) ∂u(k+1) ⎥⎥ ⎢ ∂p ⎥
⎢ ∂Tk+1 ⎥ ⎢ ⎥
⎢ ∂p
(k+1)
⎢ ⎥
∂p ⎥
∂f I
⎣ (k) ⎦ ⎣ ⎦
∂Tk+1 ∂f (k+1) ∂t
∂t ∂t ∂p

Again, two conditions are required for the existence of the sensitivity function in a
(k)
hybrid system: differentiability of the transfer functions Tk+1 and (3.34).
From equation (3.38) is obtained the initial sensitivities for the new mode directly
(recall that ẍ and ẏ can be calculated at any time with (3.26)). For clarity, the deriva-
tion is performed in two steps for the initialization (i.e., sensitivity of the initialization
solution and initial sensitivity of the trajectory) but now the differentiation has been
applied taking in account that the time of the event is a function of the parameters.
Note that the solution obtained in both cases is the same.
A procedure similar to the one in 3.2.1 may be followed to find the sensitivities of
the variables at the discontinuities (3.35). The difference is that in this case, when the
parameters change, the time of the event also changes, and the sensitivity includes
the variation due to this factor.

81
3.4 Existence and Uniqueness of the Sensitivity
Functions for Hybrid Systems
Existence and uniqueness theorems for sensitivity functions are intimately related to
theorems governing existence and uniqueness of the embedded differential system [68].
Hence, since no existence and uniqueness theorem exists for general nonlinear DAEs
[21], it is not possible to state an existence and uniqueness theorem for sensitivity
functions of general nonlinear DAEs. However, existence and uniqueness theorems
exist for both nonlinear explicit ODEs and linear time invariant DAEs. Thus, this
section establishes sufficient conditions for the existence and uniqueness of sensitivity
functions for both nonlinear explicit ODE embedded hybrid systems and linear time-
invariant DAE embedded hybrid systems.

3.4.1 Existence and uniqueness of the sensitivity functions

for hybrid systems with ODEs

Let [S1 , S2 , . . . , Sj , . . . , Snj ] with nj ≥ 2 be the ordered succession of modes, (this

sequence is a function of the parameter values), describing the evolution of an ODE
embedded hybrid system characterized by:

1. A system of differential equations (index 0) for every Sj :

ẋ(j) = f (j) (x(j) , p, t) (3.39)

2. The initial conditions for S1 :

(1) (1)
T0 (ẋ(1) , x(1) , p, t0 ) = 0 (3.40)

3. The transitions to the next stage determined by the discontinuity function:

(j) (j) (j)

gj+1(ẋ(j) , x(j) (tf ), p, tf ) = 0 (3.41)

82
 The transition functions:

(j) (j) (j) (j)

Tj+1 (ẋ(j) , x(j) (tf ), ẋ(j+1) , x(j+1) (tf ), p, tf ) = 0 (3.42)

(1) (j) (nj −1)

where [tf , .., tf , .., tf ] are the times when the changes of mode happen.

Theorem 3.1. Suppose that:

(j) (j)
1. For t ∈ [t0 , tf ], j = 1, ..., nj the partial derivatives

∂f (j) ∂f (j)
and
∂x(j) ∂p

exist and are continuous in a neighborhood of the solution x(j) (p, t).

(j) (j)
2. ∀tf , j = 0...nj − 1, the system h(j) (ẋ(j) , x(j) , ẋ(j+1) , x(j+1) , tf ; p) = 0:

 (j)
tf
(j) (j)
x (j)
(tf ) −x (j)
(t0 ) − f (j) (x(j) , p, t)dt = 0 (3.43)
(j)
t0
(j) (j) (j) (j)
gj+1(ẋ(j) (tf ), x(j) (tf ), p, tf ) = 0 (3.44)
(j) (j) (j) (j) (j) (j)
Tj+1(ẋ(j) (tf ), x(j) (tf ), ẋ(j+1) (tf ), x(j+1) (tf ), p, tf ) = 0 (3.45)
(j) (j) (j)
ẋ(j+1) (tf ) − f (j+1) (x(j+1) (tf ), p, tf ) = 0 (3.46)

is a continuously differentiable mapping of an open set

(j) (j+1)
E ⊂ R(2nx +2nx +1)+(np )

(j) (j+1) (j)

(np = 1) into R(2nx +2nx +1)
, such that h(j) (ẋ(j) , x(j) , ẋ(j+1) , x(j+1) , tf ; p) = 0
(j)
for the point (ẋ(j) , x(j) , ẋ(j+1) , x(j+1) , tf ; p) ∈ E. Assume that the submatrix
formed by the columns of the Jacobian matrix corresponding to the variables
(j)
ẋ(j) , x(j) , ẋ(j+1) , x(j+1) , and tf is invertible.

In the special case of the initialization at S1 , (j = 0) there is no equation (3.43)

and no variables x(j) in the system h(0) .

83
 Then,

1. ∀j the partial derivatives:

∂x(j)
∂p

exist, are continuous, and satisfy the differential equations:

# $
(j) (j) (j)
∂ ∂xi ∂fi ∂x(j) ∂fi
= + (3.47)
∂t ∂p ∂x(j) ∂p ∂p

(j) (j)
in (t0 , tf )

(j)
2. At tf the relationship between the right-hand sensitivities of the variables in
mode Sj and the left-hand sensitivities of the variables in mode Sj+1 is deter-
mined by:

⎡ # $−1
(j+1) (j)
∂x ∂Tj+1
= − ⎣f (j+1) +
∂p ∂x(j+1)
# (j) (j) (j) (j)
$%
∂Tj+1 ∂f (j) ∂Tj+1 (j) ∂Tj+1 ∂f (j+1) ∂Tj+1 dt
(j)
+ (j)
f + (j+1) +
∂ ẋ ∂t ∂x ∂ ẋ ∂t ∂t dp
# (j)
$ −1 # (j) (j) (j) (j)
$
∂Tj+1 ∂Tj+1 ∂f (j) ∂Tj+1 ∂x(j) ∂Tj+1 ∂f (j+1) ∂Tj+1
− + + (j+1) +
∂x(j+1) ∂ ẋ(j) ∂p ∂x(j) ∂p ∂ ẋ ∂p ∂p
(3.48)

where

(j) (j) (j)

∂gj+1 ∂f (j) ∂gj+1 ∂x(j) ∂gj+1
+ +
dt ∂ ẋ(j) ∂p ∂x(j) ∂p ∂p
= − (j) (j) (j)
(3.49)
dp ∂gj+1 ∂f (j) ∂gj+1 ∂x(j) ∂gj+1
+ +
∂ ẋ(j) ∂t ∂x(j) ∂t ∂t

Proof. Assumptions 1 and 2 correspond, respectively, to (1) the Gronwall’s theorems

[68, 71] for existence and uniqueness of the derivatives with respect to parameters and
initial conditions, including the initial time, and (2) the implicit function theorem.

84
 For the initialization problem, the implicit function theorem guarantees the ex-
istence of open sets U (1) and W (1) such that to every p ∈ W (1) ⊂ Rnp (np = 1)
(1) (1) (1)
there corresponds a unique (x(1) , ẋ(1) , t0 ; p) ∈ U (1) ⊂ R(nx +nx +1)+(np )
such that
(1)
h(0) (ẋ(1) , x(1) , t0 ; p) = 0, and there is a continuously differentiable mapping of W (1)
(1) (1) (1)
into R(nx +nx +1)
defining (x(1) ,ẋ(1) ,t0 ). Therefore, it is possible to find in a neighbor-
(1)
hood of p a point that produces (x(1) , ẋ(1) , t0 ) within that set where the conditions
of the Gronwall’s theorem apply for system f (1) .
Similarly Gronwall’s theorems assure that the difference:

& (j) &

&x (p + , t) − x(j) (p, t)&

(j) (j)
can be made as small as desired with t ∈ [t0 , tf ] by selecting a sufficiently small .
This is true for j = 1.
In the transition to the mode S2 , or in general from Sj to Sj+1 , the implicit func-
tion theorem is applied again. Therefore, sets U (j) and W (j) can be found such
that to every p ∈ W (j) ⊂ W (j−1) ⊂ Rnp (np = 1) there corresponds a unique
(j) (j) (j+1)
(ẋ(j) , x(j) , ẋ(j+1) , x(j+1) , tf ; p) ∈ U (j) ⊂ R(2nx +2nx +1)+(np )
such that

(j)
h(j) (ẋ(j) , x(j) , ẋ(j+1) , x(j+1) , tf ; p) = 0

(j) (j+1)
and there is a continuously differentiable mapping of W (j) into R(2nx +2nx +1)
defin-
(j)
ing (ẋ(j) , x(j) , ẋ(j+1) , x(j+1) , tf ).
Consequently, by selecting the smallest neighborhood of p from those that apply
for the Gronwall’s theorems and the implicit function theorem for all the modes, and
applying them in a chain, the existence and uniqueness of the sensitivity functions
for the ODE-embedded hybrid system is demonstrated. Expression (3.47) is from
Gronwall’s theorem. Expressions (3.48) and (3.49) are derived by differentiation of
the system h(j) (see Section 3.3.3).

Remark 3.1. Since it is desired to obtain partial derivatives, the theorem is applied
individually to every parameter while the rest are fixed, so the statement of the theorem

85
with only one parameter is applicable to the general multiple parameter case.

Remark 3.2. The controls u(p, t) have been dropped from the formulation of the
theorem, but are implicit since in all the functions p and t appear as variables. If the
transition function is explicit in the state variables x(j+1) and does not depend on the
time derivatives:

(j) (j) (j) (j) (j) (j)

Tj+1 (ẋ(j) (tf ), x(j) (tf ), ẋ(j+1) (tf ), x(j+1) (tf ), p, tf )
(j) (j) (j) (j)
= x(j+1) (tf ) − Φj+1 (x(j) (tf ), p, tf ) = 0 (3.50)

(j) (j) (j)

∂Tj+1 ∂Φj+1 ∂Tj+1
= − =0 (3.51)
∂x(j) ∂x(j) ∂ ẋ(j)
(j) (j)
∂Tj+1 ∂Tj+1
=I =0 (3.52)
∂x(j+1) ∂ ẋ(j+1)
(j) (j) (j) (j)
∂Tj+1 ∂Φj+1 ∂Tj+1 ∂Φj+1
=− =− (3.53)
∂p ∂p ∂t ∂t

Substituting and reordering in (3.48) yields the expression derived by Rozenvasser

[124]:

' # (j) $ %
∂x(j+1) ∂x(j)  (j+1)  ∂Φj+1 ∂Φ
(j)
j+1 dt
− = − f − f (j) + (j)
− I f (j) +
∂p ∂p ∂x ∂t dp
# (j) $ (j)
∂Φj+1 ∂x(j) ∂Φj+1
+ −I + (3.54)
∂x(j) ∂p ∂p

Remark 3.3. Hybrid systems formulated with this model are a sequence of alternat-
ing differential and algebraic systems of equations, and the conditions for solvability,
existence and uniqueness are the union of these conditions for all of the systems.

Remark 3.4. The sensitivity functions are not defined for solutions passing through
points where g or T are not differentiable. In particular, this is true at the points
(k) (k) (k) (k)
where gjα = giβ = 0 or gjα = gjβ = 0,which are transitions to a different mode or
transitions to the same mode with different discontinuity functions.

86
Remark 3.5. In general sensitivities jump even if the states are continuous. This
statement implies:

(j) (j) (j) (j) (j) (j)

∂Tj+1 ∂Tj+1 ∂Tj+1 ∂Tj+1 ∂Tj+1 ∂Tj+1
=I = −I = = = =0 (3.55)
∂x(j+1) ∂x(j) ∂ ẋ(j+1) ∂ ẋ(j) ∂p ∂t

so equation (3.48) becomes:

∂x(j+1) ∂x(j)   dt
− = − f (j+1) − f (j) (3.56)
∂p ∂p dp

As Rozenvasser points out, sensitivities are continuous either if the time derivative is
continuous or if the event time does not depend on the parameter.

3.4.2 Existence and uniqueness of the sensitivity functions

for hybrid systems with linear time-invariant DAEs

In the remaining part of this section define:

⎡ ⎤
x
z=⎣ ⎦
y

Lemma 3.1. Suppose that the linear constant-coefficient DAE system with differen-
tial index = ν:

F (ż, z, t; p) = Aż + Bz − f (p, t) = 0 (3.57)

is solvable ( i.e., λA + B is a regular pencil) for t ∈ [t0 , tf ] and the partial derivatives
 
∂ ∂if
i = 0, . . . , ν − 1 (3.58)
∂p ∂ti

exist and are continuous in a neighborhood of the solution z(p, t).

87
 Then, the partial derivatives:

∂z
s= (3.59)
∂p

exist, are continuous and satisfy the differential equations

∂f
Aṡ + Bs = (3.60)
∂p

Proof. Since the system is solvable, there exist nonsingular matrices p, Q such that:

z = Qw (3.61)

P AQẇ + P BQw = P f (p, t) (3.62)

⎡ ⎤ ⎡ ⎤
I 0 C 0
P AQ = ⎣ ⎦ P BQ = ⎣ ⎦ (3.63)
0 N 0 I

where N is a matrix of nilpotency ν. The resulting system after the change of coor-
dinates is:

ẇ1 + Cw 1 = f1 (3.64)

N ẇ2 + w2 = f2 (3.65)

The first equation is an ODE and Gronwall’s theorems can be applied. The second
equation has only one solution:

ν−1
∂ i f2
w2 = (−1)i N i (3.66)
i=0
∂ti

and the existence and uniqueness of the remaining parametric sensitivities can be
deduced directly from (3.58). The system (3.60) obtained by differentiating (3.57),
has the same matrix pencil as (3.57), assuming the forcing functions are sufficiently
differentiable (3.58). Hence (3.60) is solvable.

Remark 3.6. The sufficient differentiability (3.58) is not required for all the compo-

88
nents of f but only for the ones appearing in the linear combination f2 . If the exact
differentiability required for each component of f is specified that can be deduced from
(3.66), then the conditions of the lemma are necessary too.

Theorem 3.2. Suppose the linear constant coefficient DAE system with differential
index= ν:

F (ż, z, t; p) = A(p)ż + B(p)z − f (p, t) = 0 (3.67)

is solvable for t ∈ [t0 , tf ], that f is (2ν − 1)-times differentiable with respect to time,
and that the partial derivatives:
 
∂ ∂if
i = 0, . . . , ν − 1 (3.68)
∂p ∂ti
∂A(p)
(3.69)
∂p
∂B(p)
(3.70)
∂p

exist and are continuous in a neighborhood of the solution z(p, t).

Then, the partial derivatives:

∂z
s= (3.71)
∂p

exist, are continuous, and satisfy the differential equations:

∂f ∂A(p) ∂B(p)
Aṡ + Bs = − ż − z (3.72)
∂p ∂p ∂p

Proof. The partial differentiation of (3.67) leads to (3.72). This system is solvable if
the right hand side is (ν − 1)-times differentiable with respect to time, which requires
z to be ν-times differentiable with respect to time, (3.68), (3.69) and (3.70). Since
the solution z needs f to be (ν − 1)-times differentiable with respect to time, f must
be (2ν − 1)-times differentiable.

Remark 3.7. As in the lemma, not all the components of f need to be (2ν −1)-times

89
differentiable, since in general not all of them will end up in the rows of the nilpotent
matrix that requires this property.

Theorem 3.3. Given a solution of a hybrid system described by a sequence of modes

[S1 , S2 , .., Sj , .., Snj ], every mode characterized by a set of variables z (j) and a linear
time invariant DAE satisfying (3.4), whose coefficients are functions of the parameter
p:

F (j) (ẋ(j) , x(j) , y (j), u(j) , t; p) = A(j) (p)ż (j) + B (j) (p)z (j) − f (j) (p, t) = 0 (3.73)

(j)
and transitions to the following mode represented by the discontinuity function gj+1
(j)
and the transition functions Tj+1 .
Suppose that:

(j) (j)
1. For t ∈ [t0 , tf ], j = 1, ..., nj , every f (j) satisfies the conditions of Theorem
3.2.

(j) (j)
2. ∀tf , j = 0, ..., nj −1, the system h(j) (ẋ(j) , x(j) , y (j) , ẋ(j+1) , x(j+1) , y (j+1) , tf ; p) =
0:

(j) (j) (j)

ẋ(j) = ẋ(j) (tf ; p) x(j) = x(j) (tf ; p) y (j) = y (j)(tf ; p) (3.74)
(j) (j)
gj+1(ẋ(j) , x(j) , y (j), tf ; p) = 0 (3.75)
(j) (j)
Tj+1 (ẋ(j) , x(j) , y (j), ẋ(j+1) , x(j+1) , y (j+1), tf ; p) = 0 (3.76)
(j)
f (j+1) (ẋ(j+1) , x(j+1) , y (j+1) , tf ; p) = 0 (3.77)

is a continuously differentiable mapping of an open set:

(j) (j) (j+1) (j+1)

E ⊂ R(2nx +ny +2nx +ny +1)+(np )

(j) (j) (j+1) (j+1)

(np = 1) into R(2nx +ny +2nx +ny +1)
, such that h(j) = 0 for the point:

(j)
(ẋ(j) , x(j) , y (j), ẋ(j+1) , x(j+1) , y (j+1) , tf ; p) ∈ E.

90
 Assume that the submatrix formed by the columns of the Jacobian matrix corre-
(j)
sponding to the variables ẋ(j) , x(j) , y (j), ẋ(j+1) , x(j+1) , y (j+1) , and tf is invertible.

The equations (3.74) represent the state trajectory resulting from the integration
of f (j) in mode j.

In the initialization at S1 , there is no equation (3.74), the transition functions

(1) (1)
are replaced by the initial conditions T0 (ẋ(1) , x(1) , y (1) , t0 ; p) = 0, and there
are no variables from a previous mode in the system h(0) .

Then:

1. ∀j the partial derivatives:

∂z (j)
s(j) =
∂p

(j) (j)
exist, are continuous, and satisfy the differential equations (3.72) in (t0 , tf ).

(j)
2. At tf , the relationship between the right-hand sensitivities of the variables in
mode Sj and the left-hand sensitivities of the variables in mode Sj+1 is deter-
mined by:

(j)   (j)   (j)  

∂gj+1 (j) (j) dt
∂gj+1 (j) (j) dt
∂gj+1 (j) (j) dt
ṡx + ẍ + sx + ẋ + sy + ẏ +
∂ ẋ(j) dp ∂x(j) dp ∂y (j) dp
(j) (j)
∂gj+1 ∂gj+1 dt
+ + = 0 (3.78)
∂p ∂t dp

91
 ⎡ ⎤
⎡ (j) (j) (j)
⎤ ṡ
(j+1)
+ ẍ (j+1) dt
∂Tj+1 ∂Tj+1 ∂Tj+1 ⎢ x dp ⎥
⎣ ∂ ẋ(j+1) ∂x(j+1) (j+1) ⎢
⎦ ⎢s(j+1) (j+1) dt ⎥
∂y
+ ẋ ⎥=
∂f (j+1) ⎣ dp ⎦
x
∂f (j+1) ∂f (j+1)
∂ ẋ(j+1) ∂x(j+1) ∂y (j+1) (j+1) dt
sy + ẏ (j+1) dp
⎡ ⎤
(j) dt
ṡx + ẍ(j) dp
⎢ ⎥
⎡ ⎢
⎤ ⎢s(j) (j) dt ⎥
(j) (j) (j) (j) (j) x + ẋ ⎥
∂Tj+1 ∂Tj+1 ∂Tj+1 ∂Tj+1 ∂Tj+1 ⎢ dp ⎥
∂p ⎦ ⎢ (j) dt ⎥
− ⎣ ∂ ẋ
(j) ∂x(j) ∂y (j) ∂t
⎢sy + ẏ (j) dp ⎥ (3.79)
∂f (j+1) ∂f (j+1) ⎢ ⎥
0 0 0 ⎢ ⎥
∂t ∂p
⎢ dt ⎥
⎣ dp ⎦
1

Proof. Provided Theorem 3.2 that replaces Gronwall’s theorems for the existence and
uniqueness of the sensitivities of a linear time invariant DAE where the coefficients
are functions of the parameters, the same arguments applied in the proof of Theorem
3.1 can be repeated.

Remark 3.8. This theorem can be extended to systems not satisfying (3.4), but then
the number of equations in the transition functions is smaller, as not all the variables
can be specified independently in the new mode.

92
3.5 Examples

3.5.1 Implementation

The equations derived in Sections 3.2 and 3.3 lead to a straightforward numerical
implementation for the combined state and sensitivity integration in the ABACUSS
mathematical modeling environment. The evolution of a hybrid system can be viewed
as a sequence of subproblems, each characterized by a continuous evolution in a mode
terminated by an event and then a solution of the transition functions to initialize
the new mode.

Continuous evolution of the states and sensitivities in a mode is determined by si-

multaneous numerical solution of f (k) and (3.25) using the staggered corrector method
implemented in the code DSL48S (Chapter 4). Correct location of event times is guar-
anteed by the state event location algorithm of Park and Barton [110]. Solution of the
transition conditions (3.31–3.32) is implemented as a sequence of Newton iterations.
However, since the transitions may take the system outside the region of convergence
for Newton’s method, a more robust approach would involve a homotopy continua-
tion procedure in which the system is deformed continuously from conditions of state
continuity to the new state. Finally, the switching time sensitivity (3.36) and then
transfer of the sensitivities to the new mode (3.38) are linear systems that require
the previous solution of another linear system (3.26) for ẍ and ẏ at both sides of the
transition.

3.5.2 Critical values for the parameters

This example illustrates the practical significance of not satisfying the conditions of
the existence and uniqueness theorem.

Consider the following hybrid system with two modes and a reversible transition

93
 Discontinuity function
4
p = 3.1
3 p = 3.0
p = 2.9
2

0
L

-1

-2

-3

-4
0 0.5 1 1.5 2 2.5 3 3.5
x

Figure 3-3: Discontinuity function for the system (3.80–3.81)

condition:
⎧
⎪
⎪ dx(1)
⎪
⎪ = 4 − x(1)
⎪
⎨ dt

S1 : L(1)
2 : −(x ) + 5(x ) − 7x
(1) 3 (1) 2 (1)
+p≤0 (3.80)
⎪
⎪
⎪
⎪
⎪
⎩T2(1) = x(2) − x(1) = 0
⎧
⎪
⎪ dx(2)
⎪
⎪ = 10 − 2x(2)
⎪
⎨ dt
S2 : L(2)
1 : −(x ) + 5(x ) − 7x
(2) 3 (2) 2 (2)
+p>0 (3.81)
⎪
⎪
⎪
⎪
⎪
⎩T1(2) = x(1) − x(2) = 0

The initial mode is S1 with initial condition x(1) (0) = 0. There is only one
parameter p that only appears in the transition condition. Figure 3-3 shows the
discontinuity function as a function of the state x(k) for three different values of p.
When p = 3.1 the discontinuity function activates at a value of x(1) close to 3. If
p = 2.9 then there are:

1. a transition to mode 2 around x(1) = 0.8,

2. a transition to mode 1 around x(2) = 1.2, and

94
 3. a transition to mode 2 around x(1) = 3.

For p = 3 the function is tangent at x(1) = 1 and crosses zero at x(1) = 3. The
first point is singular. It touches 0, switches to the second mode, and immediately
changes to mode 1 again. At this point the gradient of the discontinuity function
equals zero and (3.33) is not satisfied.
If the evolution of the transition point as a function of the parameter p is examined,
it is observed that for values less than 3 there are three transition times that vary
continuously. But at p = 3 there is a nonsmoothness in the event time: now there is
only one transition time and the transition time for the first event has jumped from
the previous case. Actually, at this point the second condition for the existence and
uniqueness of the parametric sensitivity functions is not satisfied. The changes in the
sequence of events can be seen to be related to these ‘critical’ points.
The sensitivity functions for p = 2.9 and p = 3.1 are plotted in Figures 3-4 and 3-
5. In this system, the sensitivity functions are discontinuous at the time of switching.
The expression for transfer of the sensitivities is:
 
1
−3x2 +10x−7
− s−
s+ = s− − (ẋ+ − ẋ− ) (3.82)
ẋ−

In Figures 3-4 and 3-5 the discontinuous effect of the different sequences is per-
ceptible in the trajectory, but almost negligible in the long term. Consider now a
system with a nonreversible transition condition, that is, a system where there is a
switch from S1 to S2 but no switch back to S1 :
⎧
⎪
⎪ dx(1)
⎪
⎪ = 4 − x(1)
⎪
⎨ dt
S1 : L(1)
2 : −(x ) + 5(x ) − 7x
(1) 3 (1) 2 (1)
+p≤0 (3.83)
⎪
⎪
⎪
⎪
⎪
⎩T2(1) = x(2) − x(1) = 0

S2 : dx(2) = 0 (3.84)
dt

Figures 3-6 and 3-7 show the trajectories and sensitivities for this system. Now
the jump in the final values of the state and the sensitivity for the variation of the

95
parameter p is evident.

96
 Reversible Transition Condition (p = 2.9)
5
x
∂x/∂p
4

0
0 1 2 3 4 5
t

Figure 3-4: Sensitivity and state trajectory for reversible transition condition when
p = 2.9

Reversible Transition Condition (p = 3.1)

5
x
∂x/∂p
4

0
0 1 2 3 4 5
t

Figure 3-5: Sensitivity and state trajectory for reversible transition condition when
p = 3.1

97
 Nonreversible Transition Condition (p = 2.9)
0.8
x
0.6 ∂x/∂p
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t

Figure 3-6: Sensitivity and state trajectory for nonreversible transition condition when
p = 2.9

Nonreversible Transition Condition (p = 3.1)

4
3.5 x
∂x/∂p
3
2.5
2
1.5
1
0.5
0
-0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t

Figure 3-7: Sensitivity and state trajectory for nonreversible transition condition when
p = 3.1

98
3.5.3 Functions discretized by finite elements

In many problems the control variables are approximated by a finite set of parameters
that define some generic functions over finite elements that are different for every
element. When the discretization is over time, the problem can be assimilated to a
hybrid one where the modes are the elements.

For example, let the control variables be discretized using Lagrange polynomials
(k)
ψi (t) of order I over K finite elements, as described in Chapter 2. The transition
conditions and functions assuming the usual continuity of the state variable are:

(k) (k)
gk+1 = t − tI (3.85)
(k)
Tk+1 = x(k+1) − x(k) (3.86)

Now consider the transfer of the sensitivities with respect to the junction time
t∗ = tI at the end of element k. Equation (3.36) gives,
(k)

⎧
⎪
⎨1 if p = t∗ ,
∂t
= (3.87)
∂p ⎪ ⎩0 if p = t∗

and from (3.38), for p = t∗ ,

⎡ ⎤
⎡ ⎤ ṡx(k+1) + ẍ(k+1)
⎢ ⎥
⎦⎢ ⎥
0 I 0
⎣ ⎢ sx
(k+1)
+ ẋ(k+1) ⎥=
∂f (k+1) ∂f (k+1) ∂f (k+1)
⎣ ⎦
∂ ẋ (k+1) ∂x (k+1) ∂y (k+1) (k+1) (k+1)
sy + ẏ
⎡ ⎤ (3.88)
⎡ ⎤ (k)
sx + ẋ (k)
−I ⎢ ⎥
⎦ ⎢ ∂u(k+1) + ∂u(k+1) ⎥
⎢
0 0
−⎣ ⎥
(k+1) ∂f (k+1) ⎣ ∂p ∂t ⎦
0 ∂f∂u(k+1) ∂t
1

99
Noting that the derivative with respect to time of the DAE is identically zero, gives:

sx(k+1) = sx(k) − (ẋ(k+1) − ẋ(k) ) (3.89)

∂f (k+1) (k+1) ∂f (k+1) (k+1) ∂f (k+1) (k+1) ∂f (k+1) ∂u(k+1)
ṡ x + s y = − s − (k+1) (3.90)
∂ ẋ(k+1) ∂y (k+1) ∂x(k+1) x ∂u ∂p

For another junction time p = t∗ , always supposing that p does not appear explicitly
in f , the first equation simplifies to:

sx(k+1) = sx(k) (3.91)

Therefore, for stage k = 1 the initial sensitivities are all zero for (3.13), (3.14), (3.15)
and (3.16). The right hand side of (3.25) for element 1 is zero and the sensitivity
functions are all zero in that stage. The system (3.90) and (3.91) transfers zero initial
sensitivities to the new stage and this stage continues until stage k is entered. Here
(k) (k) (k)
the initial values for sx , sy , and ṡx are zero. However, the RHS of (3.25) is not
zero and therefore integration provides the values for the sensitivity functions.

(k)
At t = tI , the transfer is governed by (3.89) and (3.90). Therefore, the sen-
sitivities will in general jump even if the states are continuous. If the controls are
continuous over the junction of the finite elements, then ẋ(k+1) = ẋ(k) , and the sensi-
tivities of the differential variables will be continuous too.

In the following elements, the integration continues from the initial values at
every junction time. The sensitivities are continuous across element boundaries but
in general not differentiable at the boundary if the controls are discontinuous. The
RHS of (3.25) is zero, which indicates a system without excitation.

Figure 3-8 plots the sensitivities of three variables with respect to the junction
(2)
time tI , using five finite elements. These sensitivities exhibit the typical behavior
described by discontinuous controls: zero sensitivity, excitation, jump and decay with
non differentiable junctions.

100
 ABACUSS Sensitivity Analysis
Values x 10-3
dy1/dt2 []
20.00 dy2/dt2 []
18.00 dy3/dt2 []

16.00

14.00

12.00

10.00

8.00

6.00

4.00

2.00

0.00

-2.00

-4.00

-6.00

-8.00

-10.00

-12.00

-14.00

-16.00
Time
0.00 0.20 0.40 0.60 0.80 1.00

Figure 3-8: Sensitivities with respect to junction time

3.5.4 Singular Van der Pol’s equation

Dorodnicyn studied the Van der Pol equation:

ÿ + (y 2 − 1)ẏ + y = 0 (3.92)

with small . If = 0 the problem can be written as a DAE:

ẋ = −y (3.93)
y3
x= −y (3.94)
3

At y = ±2/3, x = ∓1 the solutions cease to exist. For small values of , the solutions
jump with almost ‘infinite’ speed.

This behavior can be approximated with a DAE-embedded hybrid system with

three modes where (3.93–3.94) is the continuous part for every mode. The modes are

101
defined in relation to values of y:
⎧
⎪
⎨y < −1.001
S1 : (3.95)
⎪
⎩L(1) : y ≥ −1.001 (1)
2 T2 :y=2
⎧
⎪
⎨y > 1.001
S2 : (3.96)
⎪
⎩L(2) : y ≤ 1.001 (2)
1 T1 : y = −2
⎧
⎪
⎪
⎪
⎪ −0.999 < y < 0.999
⎪
⎨
S3 : L(3)
1 : y ≥ 0.999
(3)
T1 : y = −2 (3.97)
⎪
⎪
⎪
⎪
⎪
⎩L(3) (3)
2 : y ≤ −0.999 T2 : y = 2

The parameter is the initial value of y. At every event the sensitivity transfer is:

ẏ + −
s+
y = s (3.98)
ẏ − y

Figure 3-9 shows the trajectories of x and y in the time domain and Figure 3-10
shows y versus x for y(0) = p = −3. After some initial time the system begins cycling.
The sensitivities are plotted in Figures 3-11 and 3-12.

When p = 0, the system starts at a fixed point and y(t) = x(t) = 0. The
sensitivities are defined for t in a bounded interval:

sx = −et (3.99)

sy = et (3.100)

For a given bounded time interval, there will exist a neighborhood of parameter
values around zero for which the sensitivity functions are qualitatively similar in the
sense that they do not experience discontinuities. Outside of this neighborhood, the
evolution of the hybrid system is not smooth. On the positive side, the oscillating
solution is ‘phase shifted’ with respect to one starting from the negative side, as shown
in Figures 3-13 and 3-14. The sensitivities are the same in this example for a positive

102
or a negative value of the parameter of the same magnitude, but this is a special case
where there is symmetry in the system. In general, the sensitivity functions could be
different.

103
 4
x
y
2

-2

-4

-6
0 1 2 3 4 5 6 7 8 9 10
t

Figure 3-9: State variable trajectories for p = −3

2
1.5 y

1
0.5
0
-0.5
-1
-1.5
-2
-2.5
-3
-6 -5 -4 -3 -2 -1 0 1 2
x

Figure 3-10: State space plot for p = −3

104
 8
dx/dp
6

-2

-4

-6
0 1 2 3 4 5 6 7 8 9 10
t

Figure 3-11: Sensitivity trajectory for p = −3

1500
dy/dp
1000

500

-500

-1000

-1500
0 1 2 3 4 5 6 7 8 9 10
t

Figure 3-12: Sensitivity trajectory for p = −3

105
 2
x(0−)
1.5 y(0−)
x(0+)
1 y(0+)

0.5

-0.5

-1

-1.5

-2
0 1 2 3 4 5 6 7 8 9 10
t

Figure 3-13: State trajectories for p = 0

25000
20000 dx/dt(0−)
dy/dt(0−)
15000 dx/dt(0+)
dy/dt(0+)
10000
5000
0
-5000
-10000
-15000
-20000
-25000
0 1 2 3 4 5 6 7 8 9 10
t

Figure 3-14: Sensitivity trajectories for p = 0

106
3.6 Conclusions
The parametric sensitivity functions for a broad class of index ≤ 1 DAE-embedded
hybrid systems have been derived and illustrated with examples. Computationally,
calculation of the sensitivity functions is inexpensive compared with solution of the
original system.
In the cases of index-0 DAEs (ODEs) or linear time invariant DAEs, theorems
giving sufficient conditions for existence and uniqueness of sensitivities have been
proven. These theorems imply the existence of ‘critical’ values for the parameters
related to qualitative changes in the evolution of the system, specifically bifurcations
in the sequence of events.
These results have important implications concerning the application of sensitivity
functions for the optimization of hybrid discrete/continuous dynamic systems. Suffi-
cient conditions for smoothness of the master optimization problem require existence
and local continuity of the sensitivity functions. Thus, changes in the sequence of
events at these critical values introduce nonsmoothness, confounding gradient-based
algorithms. The practical issue of the finite nature of numerical methods may aggra-
vate the nonsmoothness.

107
108
Chapter 4

Efficient Calculation of Sensitivity

Equations

Parametric sensitivity analysis of large-scale differential-algebraic equation systems

(DAEs) is important in many engineering and scientific applications. The information
contained in the parametric sensitivity trajectories is useful for parameter estimation,
optimization, process sensitivity studies, model simplification, and experimental de-
sign. For example, the control parameterization approach for numerical solution of
optimal control problems can require sensitivity information with respect to hun-
dreds of parameters. Inexpensive calculation of sensitivities is particularly important
in many areas of chemical, mechanical and electrical engineering, and economics be-
cause the systems of interest are typically modeled with 102 − 105 equations.
Traditionally, the combined DAE and sensitivity system has been handled by
noting that the sensitivities depend on the solution to the DAE system, and thus may
be solved using a staggered direct scheme in which the linear systems for the sensitivity
corrector steps are solved directly after convergence of the nonlinear corrector step [30,
38, 62, 72, 73, 86, 119]. Recently, a method was proposed that substantially reduces
the cost of parametric sensitivity analysis [94] relative to these earlier efforts. This
method solves the DAEs and sensitivities simultaneously, obtaining the numerical
solution from a corrector iteration on the combined nonlinear system. Within the
context of popular DAE solvers such as DASSL [21], this simultaneous corrector

109
method has lower computational cost than the staggered scheme because it minimizes
the number of matrix factorizations that need to be performed along the solution
trajectory.
In this chapter, a novel staggered corrector method for solving DAEs and sensi-
tivities is developed and demonstrated. In particular, the approach exhibits a com-
putational cost that is proven to be a strict lower bound for that of the simultaneous
corrector algorithm described in [94]. The staggered corrector method uses two cor-
rector iteration loops at each step, one for the DAE and a second for the sensitivities.
The computational savings result from fewer Jacobian updates in order to evaluate
the residuals of the sensitivity equations.
Many large DAEs found in engineering applications have Jacobians that are large,
sparse, and unstructured. Although the DAE may contain many equations, the av-
erage number of variables appearing in each equation is much lower than the number
of equations. Standard linear algebra codes exist that exploit this sparsity to reduce
dramatically the computational cost and memory resources required for the solution
of such systems [44]. The method for solving sensitivities described in this chapter
is particularly useful for, but not confined to, large sparse unstructured DAEs, be-
cause the computational time spent updating the Jacobian matrix is a significant
portion of the total solution time. The reason for this counterintuitive situation is
that state-of-the-art BDF integrators factor the corrector matrix infrequently, and
the factorization of large sparse matrices is particularly efficient, while on the other
hand the simultaneous corrector method [94] requires the Jacobian to be updated
frequently in order to calculate sensitivity residuals.
Also present in this chapter is a code for solving large, sparse DAEs and sensi-
tivities. The DSL48S code is based on the DASSL code but contains several novel
features, including the use of the highly efficient staggered corrector method. Exam-
ples are presented using this code demonstrating that the staggered corrector iteration
is a significant improvement over existing methods.
It is assumed that the reader is familiar with the DASSL code and the algorithms
used in that code [21, 94].

110
4.1 The Staggered Corrector Sensitivity Method

Consider the general DAE system with parameters:

f (t, y, y , p) = 0 (4.1)

φ(to , y(to ), y (to ), p) = 0 (4.2)

where y ∈ Rny are the state variables and p ∈ Rnp are the parameters. Consistent
initial conditions are given by (4.2). If (4.1) is a sparse unstructured DAE, each
equation in (4.1) involves on average c variables, where c  ny .

Sensitivity analysis on (4.1) yields the derivative of the state variables y with
respect to each parameter. This analysis adds the following sensitivity system of
ns = np · ny equations to (4.1):

∂f ∂f ∂f
si +  si + =0 i = 1, . . . , np (4.3)
∂y ∂y ∂pi
∂φ ∂φ ∂φ
si (to ) +  si (to ) + =0 i = 1, . . . , np (4.4)
∂y ∂y ∂pi

where si = ∂y/∂pi . Note that consistent initial conditions (4.4) for the sensitivity
system (4.3) are derived from (4.2).

A numerical solution of the nonlinear DAE system (4.1) is obtained at each time
step of the integration by approximating the solution with a suitable discretization,
typically the k-step general backward difference formula (BDF), to yield the nonlinear
system:
# k
$
g(y(n+1) = f t(n+1) , y(n+1) , αj y(n+1−j) , p =0 (4.5)
j=0

where αj are the coefficients of the BDF formula. The solution of (4.5) is performed
iteratively by solving the following linear system at each iteration:

(m) (m+1) (m)

J y(n+1) − y(n+1) = g(y(n+1) ) (4.6)

111
where

∂f ∂f
J = α0 (m)
+ (m)
(4.7)
∂y(n+1) ∂y(n+1)

Note that the corrector matrix J is calculated using information from the Jaco-
bian of (4.1). DAE codes such as DASSL use an approximation to J in order to
minimize the frequency of updates and factorizations, which are typically the most
computationally expensive steps of the integration algorithm. Thus, the corrector
solution is obtained by solving:

(m) (m+1) (m)

Jˆ y(n+1) − y(n+1) = g(y(n+1) ) (4.8)

at each iteration, where Jˆ is an approximation to J that is updated and factored

infrequently.
The staggered corrector method proposed in this chapter uses a separate but
similar Newton iteration to solve the sensitivity system (4.3). Once the solution to
the corrector for the DAE (4.1) has been obtained, the sensitivity system becomes:
⎡ ⎤
si
A⎣ ⎦ = − ∂f (4.9)
si ∂pi

where

 
∂f ∂f
A= 
∂y(n+1) ∂y (n+1)
(4.10)

In order to avoid factoring A, (4.9) may be solved as in (4.8) using the iteration:
⎡ (k ⎤
(m)
α0 si(n+1) + j=1 αj si(n+1−j)
Jˆ si(n+1) − si(n+1) = A ⎣
(m) (m+1) ⎦ + ∂f (4.11)
(m)
si(n+1) ∂pi

Put simply, a quasi-Newton iteration is used to converge a linear system. Note

that the matrix A and the vector ∂f /∂pi need be updated only once per step of

112
 Niter Number of Newton iterations
Np Number of parameters
CSRES Cost of sensitivity residual evaluation (overall)
CRES Cost of DAE residual evaluation
CBS Cost of a backsubstitution
CMF Cost of LU matrix factorization
CMV Cost of a matrix-vector product
CVA Cost of a vector-vector addition
CVS Cost of a scalar operation on a vector
CJU Cost of a Jacobian update
CRHS Cost of an evaluation of ∂f /∂pi

Table 4-1: Notation used in calculation of computational cost

the integrator, after the DAE corrector iteration and before the sensitivity corrector
iteration, because they are dependent only on information from the DAE solution.
This feature is the main reason that the staggered corrector method is attractive
compared with other methods, because the cost of updating A and ∂f /∂pi can be
high. Further, note that solution of the sensitivity corrector iteration does not require
additional factorization of a matrix, since Jˆ is already available as LU factors from
the DAE corrector iteration (4.8).

At each time step of the integrator, the additional cost involved in solving the
combined DAE and sensitivity system using the staggered corrector method versus
solving the DAE without sensitivities is:

CSRES + 2Np Niter CBS (4.12)

where the definitions of the symbols in this equation are given in Table 4-1. The
cost of sensitivity residual evaluation is discussed in later sections.

The incremental cost presented in (4.12) is for the corrector iteration alone and
assumes that Jˆ is a sufficiently good approximation to J for this sensitivity corrector
iteration to converge without additional corrector matrix factorizations. The addi-
tional cost differentials that arise outside the corrector iteration are ignored here but

113
are discussed in a later section.

114
4.2 Other Sensitivity Methods

The staggered corrector sensitivity method is an improvement on two other methods

that have been developed to solve the combined DAE and sensitivity system. In this
section, these methods and how they differ are briefly described.

4.2.1 The staggered direct method

A number of codes solve the sensitivity system directly using a staggered scheme
[30, 77, 86]. This method involves solving the nonlinear DAE system (4.1) by approx-
imating the solution with a BDF method, as in (4.5–4.8).
Once an acceptable numerical solution has been obtained from the corrector it-
eration, the solution to the sensitivity system (4.9) is obtained through the direct
solution of:
⎡ ⎤
si
A⎣ ⎦ = − ∂f (4.13)
si ∂pi

where the BDF discretization is used to eliminate si .

This method is termed the staggered direct method because the DAE corrector is
first solved using a Newton iteration, and then the linear sensitivity system solution
is obtained directly. Although A is based on the same information as the corrector
ˆ it is necessary to update and factor A at every step of the integrator
matrix J,
because Jˆ is only updated occasionally. As shown in Section 4.4, the computational
cost associated with these additional matrix updates and factorizations makes the
staggered direct method unattractive compared to other methods.
A variant on this method was proposed in [88], in which the already factored
approximation to the corrector matrix Jˆ was used in the direct solution of (4.13) for
the sensitivities. The fact that no additional matrix factorizations are required makes
this method highly efficient; however, since Jˆ is not equal to the A that defines the
sensitivity equations locally, the solution to the sensitivity equations thus obtained is

115
in reality a solution to the unknown perturbed sensitivity system:
⎡ ⎤
si
Jˆ ⎣ ⎦ = − ∂f (4.14)
si ∂pi

Therefore, no guarantees may be made concerning the distance between the solution
to this system and the true sensitivity system. For example, in [77] it is noted that
sensitivities obtained using this method are not sufficiently accurate for the master
iteration of control parameterization.

At each step of the integrator, the additional cost involved in solving the combined
DAE and sensitivity system using the staggered direct method versus solving the DAE
without sensitivities is:

CJU + CMF + Np (2CBS + CRHS ) (4.15)

where the symbols are given in Table 4-1.

4.2.2 The simultaneous corrector method

The method described in [94] combines the DAE and sensitivities to form a combined
system which is then solved using a BDF method on each step. The combined system
is:
 
 ∂f ∂f  ∂f ∂f ∂f  ∂f
F = f (t, y, y , p), s1 +  s1 + ,... , sn + s + (4.16)
∂y ∂y ∂p1 ∂y p ∂y  np ∂pnp

where Y = [y, s1 , . . . , snp ].

This combined nonlinear system can then be discretized as in (4.4–4.8):

# k
$
G(Y(n+1) ) = F t(n+1) , Y(n+1) , αj Yn+1−j , p =0 (4.17)
j=0

116
which can be solved using Newton’s method by solving the corrector equation:

(m) (m+1) (m)

M Y(n+1) − Y(n+1) = G(Y(n+1) ) (4.18)

at each iteration, where

⎛ ⎞
J
⎜ ⎟
⎜ ⎟
⎜ J1 J ⎟
⎜ ⎟
⎜ ⎟
M =⎜ J2 0 J ⎟ (4.19)
⎜ ⎟
⎜ .. .. .. .. ⎟
⎜ . . . . ⎟
⎝ ⎠
Jn p 0 . . . 0 J

where J is defined as in (4.7) and

∂J ∂J
Ji = si + (4.20)
∂y ∂pi

In [94] it is shown that if M is approximated by its block diagonal part, (4.19)

achieves 2-step quadratic convergence for nonlinear problems. Thus, this method
allows the factored corrector matrix to be reused for multiple steps.

This method is a significant improvement over the staggered direct method be-
cause the need for additional matrix factorizations in order to solve the sensitivity
system has been eliminated. However, the disadvantage of the simultaneous corrector
method compared to the staggered corrector method is that the former requires the
system Jacobian A be updated at every corrector iteration in order to calculate the
portion of G due to sensitivity residuals. Although this cost is minor compared with
matrix factorization, it is shown below to be a significant cost for large problems.

At each step of the integrator, the additional cost involved in solving the combined
DAE and sensitivity system using the simultaneous corrector method versus solving
the DAE without sensitivities is:

CSRES + 2Niter Np CBS (4.21)

117
4.3 Methods for Evaluating Sensitivity Residuals

There are three practical methods for evaluating the sensitivity residuals that may
be used with either the staggered corrector or the simultaneous corrector methods.
As shown in this section, the choice of method can significantly affect the cost of the
staggered corrector method.

In this section, it is assumed that it is possible to evaluate the Jacobian of the

DAE either analytically or using finite differences, since this evaluation is necessary to
solve the DAE without sensitivities. The concern in this section is with the additional
information that is required to calculate sensitivity residuals and how to best obtain
it.

4.3.1 Analytic evaluation of ∂f /∂pi

One way to evaluate sensitivity residuals is using the equation:

⎡ ⎤
si
ri = A ⎣ ⎦ + ∂f (4.22)
si ∂pi

where ri is the vector of sensitivity residuals, and A is calculated by updating the

Jacobian of the DAE. If ∂f /∂pi is somehow available, sensitivity residual evaluation
is reduced to two matrix-vector products and two vector-vector additions.

This is the most desirable option because it ensures the accuracy of the sensitiv-
ity residuals and experience shows that it is typically less expensive than the other
residual methods described below. In practice, if automatic differentiation is used to
obtain the DAE Jacobian (for example, see [65]), it is not much trouble to extend the
automatic differentiation to produce ∂f /∂pi .

The cost of an analytic sensitivity residual evaluation is:

CJU + Np (CRHS + 2CMV + 2CVA ) (4.23)

118
4.3.2 Finite differencing for ∂f /∂pi

If it is not possible or desirable to evaluate ∂f /∂pi analytically, it is still possible

to calculate the sensitivity residuals as in (4.22). Doing this requires that ∂f /∂pi
are evaluated via a finite difference approximation. For example, a first order for-
ward finite difference approximation gives the following equation for the sensitivity
residuals:
⎡ ⎤
si  
ri = A ⎣ ⎦ + f (t, y, y , p + δi ei ) − f (t, y, y , p) (4.24)
si δi

where δi is a small scalar quantity, and ei is the ith unit vector.

The cost of the finite difference sensitivity residual evaluation given in (4.24) is:

CJU + Np (CRES + 3CVA + CVS + 2CMV ) (4.25)

4.3.3 Directional derivatives

In [94] it is noted that it is possible to determine sensitivity residuals using a direc-

tional derivative finite difference approximation. For example, the sensitivity residuals
could be approximated as:

f (t, y + δi si , y  + δi si , p + δi ei ) − f (t, y, y , p)

ri = (4.26)
δi

The cost of the directional derivative sensitivity residual evaluation given in (4.26)
is:

Np (CRES + CVA + CVS ) (4.27)

Note that directional derivative residuals are less costly than finite difference residuals,
and therefore directional derivatives are the preferred method if analytical derivatives
are unavailable. However, it is demonstrated in Section 4.6 that analytic residuals
are preferred over directional derivative residuals from a cost standpoint.

119
4.3.4 Comparison of methods for evaluating sensitivity resid-
uals

The above expressions for the cost of evaluating sensitivity residuals are not the same
as the CSRES term in equations (4.12), (4.15), and (4.21).
For analytic sensitivities, these costs are:

StCorr
CSRES = CJU + Np (CRHS + Niter (2CMV + 2CVA )) (4.28)

SimCorr
CSRES = Niter (CJU + Np (CRHS + 2CMV + 2CVA )) (4.29)

The above equations show that the staggered corrector method has an advantage
over the simultaneous corrector method in the calculation of sensitivity residuals
because it is not necessary to update the Jacobian and the vector ∂f /∂pi at each
corrector iteration.
For finite difference sensitivities, the sensitivity residual costs are:

StCorr
CSRES = CJU + Np · Niter (CRES + 3CVA + CVS + 2CMV ) (4.30)

SimCorr
CSRES = Niter (CJU + Np (CRES + 3CVA + CVS + 2CMV )) (4.31)

Therefore the staggered corrector method has an advantage over the simultaneous
corrector method when the sensitivity residuals are calculated with finite differencing,
because it needs to update the Jacobian only once.
For directional derivative sensitivities, the sensitivity residual costs are:

StCorr
CSRES = Np · Niter (CRES + CVA + CVS ) (4.32)

SimCorr
CSRES = Np · Niter (CRES + CVA + CVS ) (4.33)

120
 There is no cost advantage for the staggered corrector method when directional
derivative sensitivity residuals are used. Also, since this method requires fewer oper-
ations and no Jacobian update, it is preferred over the finite differencing of ∂f /∂pi
method for calculating sensitivity residuals for either the staggered or simultaneous
corrector method.
In practice, for large problems the Jacobian and residual evaluations are expensive,
and therefore the ability to reuse A and ∂f /∂pi during the corrector iteration results
in significant cost savings.

121
4.4 Cost Comparison of Sensitivity Methods

This section compares the computational costs of the staggered direct method and
the simultaneous corrector method with the staggered corrector method. The cost
differences presented are for one step of the integration, assuming that the step passes
the corrector and truncation error tests, and that (4.22) is used to calculate the
sensitivity residuals.

The cost comparison measures presented in this section have been derived so that
they may be tested in numerical experiments. In the numerical experiments, the
important statistic to compare is the ratio of the additional cost of integrating state
and sensitivity systems simultaneously to the cost of integrating a state system alone.
This ratio is:
 
Additional time for sens. integration τnp − τ0
ψnp = = (4.34)
Time for state integration τ0

where τnp is the time for the integration of the sensitivity and state system with np
parameters, and τ0 is the time for the integration of the DAE without sensitivities.

4.4.1 Staggered direct versus simultaneous corrector

The state integration is dominated by the cost of matrix factorizations, which occur
for many DAEs on average every 5-10 steps. For large sparse systems, the cost
of matrix factorization has been frequently reported to be approximately 90% of the
total solution time, and the numerical results (Section 4.6) indicate that the balance is
dominated by Jacobian evaluations. The staggered direct method factors the matrix
at every step, so that between 5-10 times more evaluations and factorizations are
performed during the integration than would be performed to solve the DAE alone.
Therefore a lower bound on ψnp for the staggered direct method is:

Nsteps − Nfactorizations
ψnp = (4.35)
Nfactorizations

122
where Nsteps is the number of BDF steps and Nfactorizations is the number of corrector
matrix factorizations in the state integration. In the limit as the number of equations
becomes large the staggered direct method can do no better than this ratio (typically
about 4-9).

The same cost difference estimate can be performed for the simultaneous corrector
method with analytic derivatives. The additional cost of sensitivities in this method
is dominated by the need to update the system Jacobian at each iteration of the
corrector. However, the cost of these extra Jacobian updates is dominated by the
cost of matrix factorizations in an asymptotic sense, and a lower bound on ψnp for
the simultaneous corrector method is therefore zero.

4.4.2 Simultaneous corrector versus staggered corrector

As in the simultaneous corrector method, the asymptotic lower bound on ψnp for the
staggered corrector method is also zero. However, if the cost of matrix factorizations
(which is the same in both methods) is ignored, ψnp for both methods is dominated
by the cost of Jacobian updates and backsubstitutions. Therefore, the ratio of ψnp
for the two methods is approximately:

ψnSimCorr
p Niter (CJU + np (CBS + CRHS )
= (4.36)
ψnStCorr
p
CJU + np (CRHS + Niter CBS )

The two methods have the same cost only if Niter = 1, and the cost differential should
decrease as np increases.

The staggered corrector method is also less expensive than the simultaneous cor-
rector method if finite differencing (4.24) is used to calculate the sensitivity residuals.
However, there is little difference in cost if directional derivatives (4.26) are used to
calculate the sensitivity residuals.

123
4.4.3 Other cost considerations

The above analysis considers only the cost differences within the corrector iteration.
However, there are several other considerations that affect the overall cost of the
integration.
The sequence of step sizes taken by the integrator will vary according to which
sensitivity method is used. This observation is due to the fact that each method
is using the corrector iteration to solve a different nonlinear system. The step size
and the choice of when to factor the corrector are dependent upon the errors in the
Newton iteration, which are different in each method. The solution obtained to the
DAE is still correct to within the tolerances specified for all the methods, but the
number of steps and corrector matrix factorizations may vary. In practice, both the
simultaneous corrector and the staggered corrector often factor the corrector matrix
more often than would be done if just solving the DAE, but the difference in number
of factorizations is not typically large.
When the error control algorithm in the integrator is considered, the differences
in cost between the staggered and simultaneous corrector methods is even greater
than the above analysis would indicate. The staggered corrector method is able to
avoid solving the sensitivity system for steps where the corrector iteration or the
truncation error check fail for the DAE, while the simultaneous corrector iteration is
not capable of such a discrimination. In practice, the overall convergence of the two
corrector iterations in the staggered corrector method appears to be more robust than
convergence of the single corrector iteration in the simultaneous corrector method.

124
4.5 Description of DSL48S
The simultaneous corrector method has been implemented in a FORTRAN code called
DSL48S. This code is based on the original DASSL code [21], with modifications to
handle large unstructured sparse DAEs and to add the staggered corrector method
for sensitivities.
The DSL48S code has the following features:

1. The large-scale unstructured sparse linear algebra package MA48 [45] is embed-
ded in DSL48S for solution of the corrector equation. The MA48 package is
especially suitable for the types of problems that arise in chemical engineering,
as well as many other applications.

2. The staggered corrector method described above has been implemented, and
DSL48S offers options for solving the DAE either alone or with sensitivities.

3. The code has been adapted for use within a larger framework for solving a broad
class of high-index nonlinear DAEs [47] and dynamic optimization problems.

4. DSL48S offers the option to identify sensitivities with respect to a subset of

the parameters as identically zero. This option improves the efficiency of the
control parameterization method for dynamic optimization.

5. There is an option to include/exclude the sensitivity variables from the trun-

cation error test. This is included because it may improve the efficiency of
some applications that do not require guaranteed accuracy for the sensitivity
variables.

6. DSL48S is capable of evaluating sensitivity residuals using either an analytic

expression for ∂f /∂pi , finite differencing to obtain ∂f /∂pi , or directional deriva-
tives. When using analytic expressions for ∂f /∂pi , the user must provide an
external subroutine that may be called by DSL48S.

The result is a code that conforms closely with DASSL’s interface and uses its ex-
cellent heuristics for both the DAE and the sensitivity equations. A diagram detailing
the algorithm is shown in Figure 4-1.

125
 Solve the DAE
corrector iteration

Not
Con- Converged Refactor matrix
verged? and/or cut step

Perform error
test on DAE
variables

OK? Error test failure

Update Jacobian
and ∂F/∂pi

Solve the
sensitivity corrector Refactor matrix
equation

Con- Need NO
Not Converged to
verged? refactor?

Error test on
combined DAE and
sensitivity system

Error test failure

OK?

Figure 4-1: DSL48S algorithm

126
 At each corrector iteration on the DAE, if the norm of the update is sufficiently
small, the corrector iteration is considered to have converged. If not, another itera-
tion is performed, possibly after refactoring the matrix or reducing the step size, as
in DASSL. When the corrector has converged, a truncation error test is performed on
the state variables. If the truncation error test fails, the corrector matrix is refactored
and/or the step size is reduced, and the DAE corrector equation is solved again. If
the state variables pass the truncation error test, the Jacobian is updated, and the
sensitivity corrector equation is solved in the same manner as the DAE corrector
equation. Provision is made to refactor the corrector matrix if it is determined that
the corrector iteration is not converging to the predicted values. After the sensitivity
variables have passed the corrector convergence test, a truncation error test is per-
formed on the combined state and sensitivity system. If this test fails, the step size
is reduced and/or the corrector matrix refactored and the step is attempted again.
The algorithm contains several features designed to minimize wasted computations
due to corrector convergence failure or error test failure. An error test is performed on
the DAE before the sensitivity corrector iteration is started because an error failure
on the DAE will usually cause an error failure on the combined system. The error
check on the DAE is inexpensive compared with the wasted work that would be done
in solving the sensitivity system if the error failure was not caught at this stage. It
was found empirically that an approximation to the corrector matrix that is sufficient
to converge the corrector iteration on the DAE on rare occasions may not be sufficient
to converge the corrector iteration on the sensitivities. Provision is therefore made
to update and re-factor the corrector matrix without reducing the step size if the
sensitivity corrector iteration does not converge.

127
4.6 Numerical Experiments

The DSL48S code was tested on all of the example problems in [94], and produced the
same answers as the DASSLSO code presented in that paper. To compare the code
and the staggered corrector iteration, the problem was tested on a large-scale pressure-
swing adsorption problem [13, 81]. Included in the problem are a number of partial
differential equations which are discretized spatially using a backward difference grid.
The expressions in the DAEs of this problem are complicated and lead to a sparse
unstructured Jacobian. The problem is scalable by the number of spatial mesh points
in the adsorbers, and several different problem sizes were chosen. The number of
equations in this system is 30 · N + 9, where N is the number of mesh points in the
backward difference grid. The number of nonzero elements in the Jacobian of this
system is 180 · N + 11, and therefore the average number of variables in each equation
c is approximately 6.

In order to compare the staggered corrector method with the simultaneous cor-
rector method, an option exists in DSL48S to use the simultaneous corrector method
as described in [94]. The code was extensively tested with both the simultaneous and
staggered corrector options.

The results of two different numerical experiments are reported. In the first ex-
periment, the solution times were recorded as the size of the DAE was increased. In
the second experiment, the size of the DAE was fixed, and the solution times were
recorded as the number of parameters was increased.

The performance measures that are reported for the first experiment are ψ1 and
ψ2 /ψ1 for both the staggered and the simultaneous corrector method, as well as the
number of steps (Nsteps ), the number of corrector matrix factorizations (NMF ), and
the number of Jacobian updates (NJU ). In both the simultaneous and the staggered
corrector methods, the incremental cost for solving one sensitivity should be much
higher than the incremental cost for each additional sensitivity because the same
number of Jacobian updates are performed provided that np ≥ 1.

128
 # Equations τ0 Nsteps NMF
309 2.54s 129 36
3009 44.70s 212 33
6009 107.77s 257 34
9009 179.10s 288 35
12009 265.95s 345 30
15009 351.86s 355 33
18009 456.10s 393 30
21009 574.59s 422 32
30009 921.17s 477 31

Table 4-2: Results for integration of DAE

The timing was performed by embedding the DSL48S code within the ABACUSS1
large-scale equation-oriented modeling system. ABACUSS is an example of high level
modeling environment with an interpretative software architecture. Rather than auto-
matically generating FORTRAN or C code which is then compiled and linked to form
a simulation executable, ABACUSS creates data structures representing the model
equations in machine memory, and during numerical solution these data structures
are “interpreted” by numerical algorithms to evaluate residuals, partial derivatives,
etc. Details of the ABACUSS implementation are given in [13].
Solving the combined sensitivity and DAE system requires more Jacobian eval-
uations than solving the DAE alone, and hence the use of automatic differentiation
to evaluate the Jacobian has much more impact on the sensitivity and DAE solution
time than on the DAE solution time. With automatic differentiation techniques, the
cost of a Jacobian evaluation is typically 2-3 times a function evaluation, although
rigorous upper bounds on this ratio are dependent upon the particular algorithm
employed. For the results reported in this chapter, the modified reverse-mode algo-
rithm described in [134] was used, which is particularly well-suited for large sparse
Jacobians.
Timing data comparing the staggered and simultaneous corrector method were

1
ABACUSS (Advanced Batch And Continuous Unsteady-State Simulator) process modeling software, a derivative
work of gPROMS software, 1992
c by Imperial College of Science, Technology, and Medicine.

129
 Staggered Corrector Simultaneous Corrector
# Equations
ψ1 Nsteps NMF NJU ψ1 Nsteps NMF NJU
309 0.79 126 38 126 1.52 130 38 326
3009 0.80 204 37 204 1.45 195 35 453
6009 0.83 246 32 246 1.71 253 31 562
9009 0.86 277 33 277 1.53 286 30 617
12009 0.85 309 31 309 1.77 308 30 682
15009 0.89 337 32 337 1.65 338 29 729
18009 0.89 362 30 362 1.74 363 30 776
21009 0.94 398 29 398 1.79 408 30 873
30009 0.89 455 30 455 1.62 457 28 979

Table 4-3: Results for one parameter sensitivity and DAE system (analytic sensitivity
residuals)

obtained on an HP C160 single processor workstation. The integration tolerances used

were RTOL = ATOL = 10−5 for all sensitivity and state variables. The computational
times obtained were user times for the integration only, and excluded setup times for
the problem. Performance data is reported for a wide range covering the size problems
typically of interest, ranging from 3 · 102 to 3 · 104 state equations. The results for
the integration of the state variables with no sensitivities are given in Table 4-2. The
results for the integration of DAE/sensitivity system with one parameter are given
in Table 4-3, and those for two parameters are given in Table 4-4. The results in
Tables 4-3-4-4 were obtained with analytic derivatives.

Table 4-3 shows a dramatic difference in the computational cost for the staggered
and simultaneous corrector method with one parameter. Over a wide range of prob-
lem sizes, an average of 30% savings in the integration time was achieved with the
staggered corrector method with one parameter. This is largely due to the empirical
observation that matrix factorizations of the corrector matrix for this problem scale
less than cubically. The cost of a Jacobian update is a significant portion of the cost
of a matrix factorization, and thus the ability of the staggered corrector method to
reduce the number of Jacobian updates results in significant cost savings.

The second numerical experiment that was performed compared the solution cost

130
 Staggered Corrector Simultaneous Corrector
# Equations ψ2 ψ2
ψ1
Nsteps NMF NJU ψ1
Nsteps NMF NJU
309 1.13 124 39 124 1.07 131 37 319
3009 1.06 188 33 188 1.10 190 32 441
6009 1.06 237 35 237 0.99 231 31 522
9009 1.08 274 31 274 1.12 274 29 590
12009 1.07 305 31 305 1.00 313 29 677
15009 1.06 325 32 325 1.06 327 29 708
18009 1.09 362 31 362 1.02 366 28 777
21009 1.07 381 31 381 1.02 374 34 815
30009 1.07 444 30 444 1.07 447 29 946

Table 4-4: Results for two parameter sensitivity and DAE system (analytic sensitivity
residuals)

as the number of parameters increases. The same 3009 equation model as was used in
the previous experiment was solved, using np = 1 . . . 20. Figure 4-2 shows the incre-
mental cost of adding additional parameters for both methods. As (4.36) indicates,
the cost differential is more significant for fewer parameters. For ease of compari-
son, these results were obtained with the truncation error control on the sensitivities
turned off.
A comparison was also made of the use of analytical residuals and directional
derivative residuals within the staggered corrector method. Table 4-5 shows the re-
sults of integrating the same 6009 equation pressure swing adsorption system with
an increasing number of parameters. The ψnp statistic is reported for both the ana-
lytical and the directional derivative residual methods. These results show that the
analytical method was favored over directional derivatives for all the tested problems,
and that the relative difference increases as np increases. Note that the results in this
table are not consistent with the results in Tables 4-3 and 4-4 because the parameters
used in the problem were different.

131
 1.6
Staggered Corrector
1.4 Simultaneous Corrector
1.2

1
ψn p
np

0.8

0.6

0.4

0.2

0
0 5 10 15 20
np

Figure 4-2: Sensitivity incremental cost per parameter

Analytical Residuals Directional Derivative Residuals

np
ψnp Nsteps NMF NRES ψnp Nsteps NMF NRES
1 0.85 248 30 568 1.36 248 30 1516
2 0.99 246 35 540 2.03 246 35 1974
3 1.08 240 31 547 2.62 240 31 2379
4 1.33 247 31 664 3.46 247 31 2924

Table 4-5: Comparison of analytic residuals and directional derivative residuals for
staggered corrector

132
4.7 Truncation Error Control on Sensitivities
There has been some uncertainty [77, 94] about whether truncation error control must
be maintained on the sensitivity variables as well as the state variables. The reason
often given for not including the sensitivity variables in the truncation error test is
that as long as the state variables are accurate, the sensitivity variables should be
fairly accurate. Furthermore, many applications that use sensitivity information could
arguably withstand a small amount of inaccuracy in the values of the sensitivities.
For example, in dynamic optimization it may not be necessary to have extremely
accurate sensitivity information when the optimizer is far from the solution.
There are significant computational advantages for skipping the truncation error
test on the sensitivities. The cost of the test itself is a linear function of problem size,
but even more significant cost savings may come from the integrator being able to
take larger steps on many problems. The larger step size will result even for problems
where the sensitivities never fail the truncation error test, since the test is also used
to choose the step size.
It was found during the course of this work that when the sensitivities are solved
using the staggered corrector scheme, the truncation error test should be performed
on the state variables. When the sensitivities are not included in the truncation error
test, the integrator is sometimes able to take large enough steps to miss features of
the dynamics of the sensitivity system. This effect is not limited to the staggered
corrector method, and should be observed in all of the sensitivity methods detailed in
this chapter. The staggered direct and the simultaneous direct methods also use the
BDF formula to approximate the time derivatives of the sensitivity variables, which
is incorrect if the stepsize is too large.
The problem can be seen by looking at the sensitivities of the following problem:

+
ẏ = −2gy sin(γ) (4.37)
+
ẋ = −2gy cos(γ) (4.38)
   
t − tf t − to
γ = p1 + p2 (4.39)
to − tf tf − to

133
where p1 = 0.5, p2 = 0.5, to = 0, and tf = 0.589.
Figures 4-3 and 4-4 show that the sensitivities can contain significant error when
they are excluded from the truncation error test, even for problems such as this one
that are not considered particularly stiff or otherwise hard to solve. The ‘kinks’ in the
trajectories without error control are due to the integrator taking very large steps.
It is very tempting to exclude the sensitivities from the truncation error test
because this results in substantial computational savings for some problems. However,
doing so in this example resulted in the numerical code giving no warning when
the sensitivity trajectories became incorrect, even by large amounts. Therefore, it
is recommended that the truncation error test should always be performed on the
sensitivities in order to avoid this problem.

134
 Sensitivity Error Test Comparison
1.4
Sens. Error Test OFF
1.2 Sens. Error Test ON
1

0.8
∂x
∂p1
0.6

0.4

0.2

-0.2
0 0.1 0.2 0.3 0.4 0.5 0.6
t

Figure 4-3: Comparison of x sensitivity of Brachistochrone with and without sensi-

tivity error test

Sensitivity Error Test Comparison

0
Sens. Error Test OFF
-0.2 Sens. Error Test ON

-0.4

∂y -0.6
∂p1

-0.8

-1

-1.2

-1.4
0 0.1 0.2 0.3 0.4 0.5 0.6
t

Figure 4-4: Comparison of y sensitivity of Brachistochrone with and without sensitiv-

ity error test

135
4.8 Conclusions
The staggered corrector method for numerical sensitivity analysis of DAEs has been
shown to exhibit a strict lower bound on the computational cost for the two other
methods typically used, the staggered direct and the simultaneous corrector methods.
Experience with large sparse problems has shown that the ability of the staggered
corrector method to reduce the number of Jacobian updates leads to significant cost
savings that are especially apparent with large, sparse DAEs, but also for other types
of DAEs.
It is possible to adapt the staggered corrector method for parallel execution. Par-
allelization may be accomplished by using different processors for each sensitivity
calculation and staggering the sensitivity calculation to follow the DAE corrector,
similar to the method described in [77].
The code DSL48S is reliable, easy to use, and efficient for sensitivities of large
sparse DAEs.

136
Chapter 5

Numerical Solution of High-Index

Differential-Algebraic Equations

This chapter addresses the problem of numerical solution of high-index differential-

algebraic equations (DAEs). The need for solving high-index DAEs in practical engi-
neering applications has been somewhat controversial in the literature, since several
authors [56, 87, 95, 117, 139] have advocated reformulating process models to avoid
high-index DAEs when solving dynamic simulation problems. Whatever the merits
of that argument, Chapter 2 has shown that dynamic optimization problems are nat-
urally DAEs, and often high-index DAEs. Furthermore, in Chapter 6 it will be shown
that using the control parameterization method to solve dynamic optimization prob-
lems efficiently requires a direct method for solving high-index DAEs. In addition
to dynamic optimization, the ability to solve high-index DAEs can in fact often be
useful for certain types of process simulation, allowing the use of alternate coordinate
systems and a broader spectrum of modeling assumptions. Therefore, this chapter is
concerned with the development of a numerical method that is capable of solving as
broad a class of high-index DAEs as possible. The method described here is called
the dummy derivative method, which has been developed during the course of this
thesis into a reliable method for solving many high-index DAEs.

137
5.1 Background on DAEs

The focus here is on the numerical solution of general nonlinear high-index differential-
algebraic equations (DAEs) of the form:

f (ẋ, x, y, t) = 0 (5.1)

where x ∈ Rmx are the differential state variables, y ∈ Rmy are the algebraic state
variables, t ∈ T ≡ (to , tf ], and f : Rmx × Rmx × Rmy × R → Rmx +my . The partitioning
of the state variables into algebraic and differential variables does not imply a loss of
generality.

There are several basic types of DAEs. The following definitions are from [21].
Linear constant coefficient DAEs have the form:

Aẋ(t) + Bx(t) = f (t) (5.2)

where A and B are (possibly singular) square matrices. Linear time varying DAEs
have the form:

A(t)ẋ(t) + B(t)x(t) = f (t) (5.3)

Although very few systems in practical applications fall into this class, linear time
varying DAEs are important because there are proofs that exist only for this class
that lead to results and techniques that seem appropriate for more general nonlinear
DAEs.

This thesis is concerned mostly with nonlinear DAEs, which typically occur in
chemical process modeling. These may be either fully-implicit, which have the form:

f (ẋ, x, t) = 0 (5.4)

138
or semi-explicit, which have the form:

f1 (ẋ, x, t) = 0 (5.5)

f2 (x, t) = 0 (5.6)

The definition of differential index was given in the glossary of DAE terms at the
beginning of Chapter 2. Although there are other definitions of indices of a DAE, the
differential index is of most relevance in this thesis, and it shall be referred to simply as
the index. By this definition, the index of an ODE system is zero. Numerical solution
of DAEs is significantly different than solution of ODEs [112, 130], but the solution
of DAEs with index ≤ 1 is relatively straightforward using one of several methods.
These methods include backwards-difference formula (BDF) methods such as those
implemented in DASSL [21, 113] and DASOLV [76], the extrapolation code LIMEX
[40], and implicit Runge-Kutta methods such as RADAU [70] and those reviewed by
[21] and [70]. Of these, it has generally become accepted that the BDF methods are
the most efficient for a broad spectrum of problems [51, 95], and a variant of DASSL
called DSL48S (see Chapter 4) was chosen for the numerical integrations performed in
this thesis. However, the numerical solution of index > 1 DAEs presents complications
[21]. Convergence proofs for the BDF method are given in [21, 78, 96], and generally
apply only to problems with index less than two and very restricted forms of higher-
index problems. This chapter describes the dummy derivative method for the solution
of what are loosely called high-index DAEs, i.e., those with index > 1.

Since it is difficult to solve high-index DAEs numerically, a criterion must be

established to distinguish between high-index DAEs and those that can be solved
using standard numerical techniques. Based on the definition of the index, it can be
shown [21] that a sufficient condition for the index to be at most one is for the Jacobian
of the DAE (5.1) to be nonsingular with respect to the highest order derivatives (ẋ, y)
in the DAE. Note that this is only a sufficient condition, and there are index-1 DAEs
for which this criterion does not hold. In [87] a different sufficient condition for (5.1)
to be index-1 is presented (although the original presentation in [87] incorrectly claims

139
this to be a necessary and sufficient condition.) It is interesting to note that while the
two sets of index-1 DAEs that satisfy each of these necessary conditions respectively
intersect, they do not coincide. In other words, there exist index-1 DAEs that satisfy
both, one, or none of these two conditions. So, in principle, both criteria should be
applied to cover a broader class of index-1 DAEs.
This criterion is still not of much practical use from a numerical point of view
because, due to rounding error, it is very difficult to tell the difference numerically be-
tween a singular matrix and one that simply has an extremely high condition number
[131]. Also, rank detection computations are very expensive for large-scale systems
[9]. Further, singularity of the Jacobian is a local property, but the nonsingularity
condition must hold globally. For these reasons, the concept of structural singularity
is used. The following definition is from [121]:

Definition 5.1 (Structural matrix). The elements of a structural matrix [Q] are
either fixed at zero or indeterminate values which are assumed to be independent of
one another.

Therefore, the entries in structural matrices differentiate only between hard zeros (0)
and nonzeros (∗) [139]. A given matrix Q is called an admissible numerical realization
with respect to [Q] if it can be obtained by fixing all indeterminate entries (∗) of [Q] at
some particular values. Two matrices A and B are said to be structurally equivalent
if both A and B are numerical realizations of the same structural matrix [Q].
In other words, a property of a matrix is a generic or structural property if it holds
in a neighborhood of the nonzero entries of the matrix. A matrix A is structurally
nonsingular if it is an admissible numerical realization of the structural matrix [A]
and there exists a permutation P on the structural matrix such that P [A] has a
nonzero diagonal, which is referred to equivalently as a maximum transversal, an
output set assignment, or an augmenting path. It can be easily demonstrated that a
structurally singular matrix is also singular, but the converse is not necessarily true.
The advantage of using structural properties is that evaluating structural properties
is much less computationally intensive than numerical evaluation and exact in the

140
sense that it is not subject to rounding error.

When the sufficient conditions for a DAE to be at most index 1 are modified
so that structural singularity of the relevant Jacobian is examined, the criteria are
no longer sufficient conditions. That is, there are DAEs which have structurally
nonsingular Jacobians which are singular. On the other hand, if the Jacobian is
structurally singular, the DAE definitely has an index ≥ 1. It is interesting to note
that there are index-1 DAEs for which the Jacobian with respect to the highest order
time derivatives is structurally singular, for example:

ẋ1 + 2ẋ2 = 0 (5.7)

x1 + x2 = 1 (5.8)

Differentiating this class of “special” index-1 DAEs once produces an index-0 DAE
(at least in the structural sense, which provides a means to detect them.

Structural nonsingularity can be checked using algorithms for finding a maximum

transversal [45]. Even though the index may be underestimated for some DAEs due
to numerical singularity, structural nonsingularity of the Jacobians is an attractive
approach to detect index ≤ 1 DAEs because of the ease of using the relevant criteria
from a numerical standpoint and because it has been found that they work well for
the vast majority of DAEs of practical interest.

A particular solution of a DAE is defined by (5.1) and a set of initial conditions

that hold only at t = t0 :

φ(ẋ(t0 ), x(t0 ), t0 ) = 0 (5.9)

and also additional hidden constraints that are obtained by differentiating some of the
equations of the DAE. The initial condition (5.9) must be consistent, which practically
means that it must allow us to find consistent initial values for the state variables and
their time derivatives at t0 so that numerical integrations may be started smoothly.
Consistent initial values are defined as [26]:

141
Definition 5.2 (Consistent Initial Values). Initial values x0 are consistent if they
admit a smooth solution in [0, t] for t → 0 and x(0) = x0 .

Unlike the solution of ODEs, not all initial conditions (5.9) of DAEs admit a smooth
solution. So, from a practical standpoint solving DAEs requires both specifying a
valid (5.9) and then being able to use it to obtain consistent initial values.
In [139], the r components of x(t0 ), ẋ(t0 ) that can be assigned an arbitrary value
and still allow a consistent initialization are defined as dynamic degrees of freedom.
The number of dynamic degrees of freedom is dependent on the index of the DAE,
but not related by any explicit formula. The issue of finding a valid set of initial
conditions is addressed later in this chapter.
Once a consistent initial condition has been obtained, finding consistent initial
values is nontrivial from a practical point of view. The consistent initial values must
satisfy the nonlinear algebraic system formed by (5.1), (5.9) and the nonredundant
additional constraints formed by differentiating (5.1) one or more times with respect
to time; reliable solution of which is difficult with current large-scale root-finding
technology. Moreover, consistent initial values are necessary because weak instabilities
have been shown to occur when the DAE is solved with a BDF or implicit Runge-
Kutta method from inconsistent initial values [96]. There are several methods for
finding consistent initial values (see [95] for a review, and also [21, 93]) but this
problem is not in the scope of this thesis, and it is assumed to be possible to find a
set of consistent initial values for all of the problems discussed.

5.1.1 Reasons for solving high-index DAEs

High-index DAEs arise when the number of truly independent state variables is less
than the number of variables that have time derivatives appearing in the DAE, or
more simply stated, when there are explicit and/or implicit algebraic relationships
among differential state variables. In general, high-index DAEs may result from sev-
eral sources. They may come from modeling assumptions that were made about the
physical system, which in turn may arise from a lack of information about the pa-

142
 Fin

Fout

Figure 5-1: Simple tank model

rameters or rate processes of a phenomenon, or from a desire to reduce the stiffness

of the system to be solved. For example, a simplified model of a total condenser for
a pure component vapor is high-index because of the simultaneous specification of
both the volume and the vapor fraction of the system [108]. Many high-index DAEs
in chemical engineering result from pseudo-steady state, equilibrium, or incompress-
ibility assumptions [56, 87, 108, 117].

More significantly, as shown in Chapter 6, high-index DAEs also arise naturally

in the solution of simulation and dynamic optimization problems that include path
constraints on the state variables [46, 48], which are common in many engineering
applications. In particular, path constraints on ODE embedded systems require either
implicit or explicit treatment of high-index DAEs.

One example of high-index DAEs in engineering applications is prescribed path

control. Prescribed path control problems are those where it is desired to determine
the input variable trajectories that will produce a specified output trajectory. For
example, consider the simple model of a tank shown in Figure 5-1. A simple model
describing the dynamic behavior of such a system is:

dh
= Fout − Fin (5.10)
dt
Fout = ah (5.11)

143
where Fin is the flow into the valve, Fout is the flow out of the tank, and a is a
constant determined by the characteristics of the output orifice. This problem is not
high-index if the design degree of freedom is satisfied with the constraint:

Fin = f (t) (5.12)

On the other hand, a high-index prescribed path control problem can be created by
satisfying the design degree of freedom with the constraint:

¯
h = f(t) (5.13)

Note that the DAE given by (5.10–5.11) and (5.13) is high-index because the differ-
ential state variable h is explicitly constrained by (5.13).
In this thesis the interest in high-index DAEs is primarily to create an algorithm
that can handle path constrained dynamic optimization problems. However, the
above discussion has shown that it is sometimes convenient to deliberately formulate
a high-index model for simulation purposes. Therefore, the interest here is on methods
that can solve general nonlinear DAEs of arbitrary index.

5.1.2 Methods for solving high-index DAEs

It is possible [21] to solve certain classes of semi-explicit high-index DAEs using

BDF and implicit Runge-Kutta methods. However, these methods are currently
impractical for general use due to the limitations on the functional form of the DAEs,
the need to determine if the DAE has the correct form to be solved using one of these
methods, and the fact that modification of numeric codes to solve such systems is not
straightforward. In fact, this modification has often been done by reducing the error
control tolerances or eliminating the error control entirely on certain variables, which
produces solutions of dubious accuracy. Another issue is that the corrector becomes
poorly conditioned as h → 0. The conditioning of the corrector scales with 1/hv ,
i.e.the algebraic error in the corrector can be so large it will disrupt the truncation

144
error control [1, 25]. To get around some of these problems, several methods have been
proposed that require some form of manipulation of the high-index DAE. Since the
objective is to find a method that works for general arbitrary-index DAEs, attention
is restricted here for the most part to methods that are not limited to a particular
form or index of DAE.
As noted in [21, 60], one method for obtaining numerical solutions to high-index
DAEs is to differentiate a subset of the equations in the DAE until an index-1 or
0 DAE is obtained that may be solved using standard numerical techniques. This
introduces the concept of underlying DAEs (UDAEs):

Definition 5.3 (UDAE). A DAE Ā is said to be an underlying DAE for the DAE
A if for every equation e ∈ A there is a corresponding ē ∈ Ā where ē = dn e/dtn ,
n ≥ 0.

For any DAE that is sufficiently differentiable, there are associated UDAEs that are
ODEs. These ODEs are termed underlying ODEs (UODE). Any solution to the
original DAE must also satisfy all associated UDAEs, including UODEs. However, it
is possible that some equations in the set of UDAEs may be redundant. For example,
consider the index-2 DAE:

ẋ1 = x2 + x1 (5.14)

x1 = f (t) (5.15)

One UDAE for this system may be obtained by differentiating (5.15) once:

ẋ1 = x2 + x1 (5.16)
d
ẋ1 = f (t) (5.17)
dt

and another UDAE may be obtained by differentiating both (5.14) and (5.15) once:

ẍ1 = ẋ2 + ẋ1 (5.18)

d
ẋ1 = f (t) (5.19)
dt

145
and a third UDAE may be obtained by differentiating (5.14) once and (5.15) twice:

ẍ1 = ẋ2 + ẋ1 (5.20)

d2
ẍ1 = 2 f (t) (5.21)
dt

Th UDAEs (5.18–5.19) and (5.20–5.21) are also UODEs. Equations (5.16), (5.17),
(5.18), (5.19), and (5.21) are nonredundant equations that must be satisfied by any
valid numerical solution to (5.16–5.17). In general, the numerical difficulties involved
with solving high-index DAEs are associated with finding a solution that satisfies all
nonredundant equations in the set of all possible UDAEs.

An example of a high-index DAE that will be used throughout this chapter for
illustrative purposes is the following model of a pendulum with a rigid rod derived in
Cartesian coordinates (see Figure 5-2):

λ
mẍ + x=0 (5.22)
L
λ
mÿ + y = −mg (5.23)
L
x2 + y 2 = L2 (5.24)

x(0) = xo ẏ(0) = ẏo

where x and y are the Cartesian coordinates, λ is the tension in the rod, L is the
length of the rod, m is the mass of the pendulum bob, and g is the gravitational
constant.

The DAE (5.22–5.24) is index-3 because the length constraint (5.24) constrains
the x and y differential state variables. An index-1 underlying DAE may be obtained
by differentiating the length constraint (5.24) twice in order to obtain a UDAE. When

146
 x

Figure 5-2: High-index pendulum

this is done, an index-1 UDAE for the high-index DAE is:

λ
mẍ + x=0 (5.25)
L
λ
mÿ + y = −mg (5.26)
L
2xẍ + 2ẋ2 + 2y ÿ + 2ẏ 2 = 0 (5.27)

x(0) = xo ẏ(0) = ẏo x(0)2 + y(0)2 = L2 x(0)ẋ(0) + y(0)ẏ(0) = 0

Note that this UDAE requires the specification of four initial conditions, rather than
the two required for the original DAE.

Figure 5-3 shows the numerical results obtained by solving the reduced-index
model (5.25–5.27) over a large number of oscillations with the initial conditions x0 = 1
and ẏ0 = 0 and the parameters m = 1 and L = 1 using a standard BDF method
integrator. The numerical code produced no error messages or any other sign that
numerical difficulties had been encountered, but it is obvious that there is some
problem because this is a frictionless system but the bob does not reach the same x
point at each oscillation. This can be seen more clearly in Figure 5-4 which shows
the length of the pendulum throughout the simulation, which shows that the length
invariant is not satisfied. Thus, although the analytic solution of (5.25–5.27) must

147
 Gear Method Pendulum x position
1
0.8
0.6
0.4
0.2
0
x

-0.2
-0.4
-0.6
-0.8
-1
0 50 100 150 200 250 300
t

Figure 5-3: Pendulum x trajectory with Gear method index reduction

satisfy (5.24), the numeric solution does not.

This phenomenon is called constraint drift and was noted in [57, 58, 59]. Several
methods have been proposed to handle this problem. In general, nonlinear implicit
constraints are not enforced when the problem is discretized. The simplest solution
is to integrate a UDAE using step sizes that are hopefully small enough to keep the
constraint drift to an acceptable minimum. However, it was shown in [54] that this
strategy may not work because differentiating a nonlinear constraint can affect the
stability properties of the DAE.

An alternative is the constraint stabilization method proposed in [58]. This

method uses the UODE:

ẋ = f (x, t) (5.28)

and the constraints obtained while deriving the UODE from the DAE:

g(x, t) = 0 (5.29)

where g : Rmx ×R → Rng . By introducing new variables μ ∈ Rng the following system

148
 Gear Method Pendulum Length
1.05
1
0.95
0.9
0.85
L

0.8
0.75
0.7
0.65
0.6
0 50 100 150 200 250 300
t

Figure 5-4: Pendulum length with Gear method index reduction

may be obtained:

 T
∂g
ẋ = f (x, t) + μ (5.30)
∂x
g(x, t) = 0 (5.31)

which is a semi-explicit index-2 DAE. The solution of all the algebraic variables in
this problem can be shown to be zero, and the constraints and all of their derivatives
are explicitly enforced. Therefore, the numerical solution does not exhibit constraint
drift and general multistep methods can be used to solve the problem. In [21] this
constraint stabilization technique was used to solve the high-index pendulum (5.22–
5.24), and it was shown to eliminate the problem of constraint drift. However, a
numerical solution to the resulting semi-explicit index-2 DAE was obtained only by
excluding the algebraic variables from the integration error control estimates, and
solutions at very tight tolerances could not be obtained because the algebraic variables
kept the corrector iteration from converging. This could be due to the fact that the
condition number of the corrector matrix of a high-index DAE increases very rapidly
as the stepsize tends toward zero [1, 25]. Constraint stabilization methods were also
discussed and applied to constrained mechanical systems in [54].

149
 Another method for obtaining numerical solutions to high-index DAEs is through
the use of regularizations [17, 114]. A regularization of a DAE is the introduction
of a small parameter into the DAE in such a way that solution of the regularized
DAE approaches the solution of the high-index DAE as the parameter approaches
zero. Regularization is a standard method that has been employed by engineers for
decades, such as the use of a weir equation for a vessel filled with an incompressible
fluid and any “controller” type equation [53]. Essentially regularization creates a
very stiff system where the solution decays very quickly to the solution manifold of
the high-index DAE. Regularizations have been applied with mixed success to high-
index DAEs [21], but their use has not advanced to the point where they may be
easily applied to arbitrary-index general DAEs. Moreover, in [2] it is shown that
the method of dummy derivatives described later in this chapter is more efficient
numerically than a regularization technique, due to the high degree of stiffness of the
regularized system.

There have been several attempts to develop numerical methods for direct solution
of high-index DAEs. Projected Runge-Kutta collocation methods have been shown
to work for some classes of high-index DAEs [4, 90]. Another interesting approach
is the least squares projection technique described in [11, 27, 28, 29]. This method
works by determining a local set of coordinates for x and then solving the derivative
array equations numerically using least-squares for ẏ. Although these approaches
seem promising, neither has been developed to the point where it is as reliable and
easy to use as BDF methods are for index-1 DAEs.

Finally, a method for finding an index-1 DAE with the same solution set as the
high-index DAE was described in [5, 58, 60, 139]. This method is termed the elim-
ination method of deriving an equivalent index-1 DAE. Essentially the elimination
method takes the nonredundant equations in the UDAEs and substitutes them into
the original high-index DAE, eliminating some variables and thus creating an index-1
DAE with a reduced number of degrees of freedom. As an example of this method
applied to the pendulum (5.22–5.24) the nonredundant equations from the UDAEs

150
are the ones created by differentiating the length constraint (5.24) twice:

2xẋ + 2y ẏ = 0 (5.32)

2xẍ + 2ẋ2 + 2y ÿ + 2ẏ 2 = 0 (5.33)

When (5.32–5.33), and (5.22) are used to eliminate ẍ and ẋ, an equivalent index-1
DAE is obtained:
 
y 2 ẏ 2 ẏ 2 y ÿ λ
m − 3 − − + x=0 (5.34)
x x x L
λ
mÿ + y = −mg (5.35)
L
x2 + y 2 = L2 (5.36)

x(0) = xo ẏ(0) = ẏo

This system can be explicitly solved for its highest order time derivatives (λ, ÿ, x)
in terms of (ẏ, y) and the parameters to get:

y 3 mg − ymgL2 + mẏ 2 L2
λ=− (5.37)
L (y 2 − L2 )
y 4mg − 2y 2 mgL2 + ymẏ 2L2 + mgL4
ÿ = (5.38)
L2 (y 2 − L2 ) m
+
x = ± −y 2 + L2 (5.39)

x(0) = xo ẏ(0) = ẏo

These equations no longer uniquely determine the highest order time derivatives when
y → ±L. Therefore, this system is either locally high-index or locally unsolvable when
y = ±L. It can easily be seen that no finite number of differentiations of this DAE
will produce a UDAE that uniquely defines the highest order time derivatives when
y = ±L, since the denominators of all time derivatives of (5.37–5.38) will contain
powers of (y 2 − L2 ). It appears that this system is locally unsolvable when y = ±L,
since the definition of solvability given in [21] would seem to require that the index
be finite. This same solvability phenomena shall be seen again in the discussion of

151
the dummy derivative method.
However, there is another equivalent index-1 DAE that is formed by eliminating ÿ
and ẏ instead of ẍ and ẋ that is solvable when x = 0 but is not solvable when y = 0.
Theoretically the elimination method is valid for solving high-index DAEs without
constraint drift, and it does work for small systems such as the one demonstrated
above, but in practice the algebraic elimination required is extremely computationally
expensive for large nonlinear DAEs.
A more promising method that is the somewhat related method of dummy deriva-
tives [97] which is described in the later sections of this chapter. Although this
method was originally described in [97], it required considerable development to cre-
ate a practical automated algorithm. This method allows for solution of a broad class
of nonlinear arbitrary-index DAEs efficiently and with guaranteed accuracy, and it
does not require expensive algebraic eliminations. Since the method is based on the
ability to find the nonredundant UDAE equations that permit the solution for con-
sistent initial values, the next sections of this chapter are used to describe Pantelides’
algorithm for obtaining a consistent initial condition for high-index DAEs.

152
5.2 Consistent Initialization of High-Index DAEs

In general, all of the UDAEs of a DAE must be satisfied at the initial condition,
however, only a subset of the UDAE equations are nonredundant. The set of all
UDAEs of a DAE is also called the derivative array. All the equations in the derivative
array that are nonredundant and constrain the state variables or their time derivatives
must be found in order to find a numerical solution using the BDF method. This
problem is related to the problem of finding a consistent initial condition, which
requires finding which equations are differentiated and how many times to constrain
independent initial conditions.
The problem of determining which equations in (5.1) and (5.9) need to be dif-
ferentiated at the initial condition was addressed in a structural sense in [107]. In
this section, Pantelides’ structural algorithm is described. Also described are mod-
ifications to the algorithm, and an implementation of the modified algorithm. In
later sections this modified algorithm is incorporated into a general framework for
obtaining numerical solutions to high-index DAEs.

5.2.1 Description of Pantelides’ algorithm

An algorithm for finding a consistent initialization of DAEs using structural criteria

was proposed in [107]. The algorithm detects the presence of a high-index DAE
and indicates which equations must be differentiated to obtain a consistent initial
condition, provided that this can be done from purely structural information.
At each iteration, Pantelides’ algorithm returns the subset of equations in the
DAE that need to be differentiated to produce a UDAE with index one lower than the
current index. The algorithm terminates when it returns a UDAE that is structurally
nonsingular with respect to the highest order time derivatives in the DAE.
Duff [43] has developed an efficient algorithm for finding an augmenting path for
a set of equations and variables. Simply described, the algorithm determines whether
it is possible find a unique assignment for each variable in the set of interest to a
unique equation in the set. It is equivalent to finding a structural permutation P

153
that gives a nonzero structural diagonal of maximum size.

Pantelides’ algorithm uses Duff’s algorithm to find structurally singular subsets

of equations, or those that are not part of an assignment. These structurally singular
equation subsets are then differentiated and replaced by their differentiated subset,
and the algorithm applied again until a complete assignment can be made for the
highest order time derivatives of the variables appearing in a mixed set of undifferen-
tiated and differentiated equations. This algorithm uses a bipartite graph of equation
and variable vertices. In this type of graph, an edge exists between a variable vertex
and an equation vertex when a variable is present in an equation, and an assignment
is a set of edges such that only one variable is assigned to each equation.

Pantelides’ algorithm proceeds by applying Duff’s algorithm to the graph repre-

sentation of the structural Jacobian of the DAE with respect to the highest order time
derivatives. The algorithm terminates if Duff’s algorithm finds an augmenting path.
If an augmenting path does not exist for the current DAE, the structurally singular
subsets of equations are differentiated with respect to time, and the algorithm is then
applied to this UDAE. The algorithm returns nonredundant UDAEs, each of which
must be explicitly satisfied by a consistent initial condition for the DAE. Since it uses
structural criteria, Pantelides’ algorithm provides only a lower bound on the index
and an upper bound on the dynamic degrees of freedom, and may not return the
correct result for systems that arise in certain (mostly pathological) situations.

The advantage of Pantelides’ algorithm is that it both detects structurally high-

index systems and returns a necessary set of equations that must be satisfied by a
consistent set of initial values. If this set of equations is also sufficient, it may then be
used to formulate an equivalent index-1 DAE using the method of dummy derivatives.

154
 Equations Variables

(5.40) v̇

(5.41) F

(5.42) ẋ

Figure 5-5: Starting structural graph for index-3 problem

5.2.2 Example of Pantelides’ algorithm

Consider the following very simple system that was given in [97]:

mv̇ = F (5.40)

ẋ = v (5.41)

x = x̄(t) (5.42)

The solution to this problem may be interpreted as the force F required to make the
mass m follow a given trajectory x(t). Note that the selection of x as the input to the
system has caused it to be high-index. If instead F had been selected as the input
the problem would be an index-0 system of ODEs in state-space form.
Pantelides’ algorithm may be applied to the system (5.40–5.42) to determine any
additional equations that must be explicitly satisfied by the initial condition. For
this example, the structural graph will be shown using circles for the vertices of a
graph, and the lines connecting them as the graph edges. The system (5.40–5.42) is
structurally represented by Figure 5-5.
The bold edges show one possible attempt to find a matching, but no matching
is possible on this graph because (5.42) has no edges connecting it with a variable.
Therefore, this graph is structurally singular with respect to (5.42), which is differ-
entiated, giving:

ẋ = x̄ (t) (5.42 )

155
 Equations Variables

(5.40) v̇

(5.41) F

(5.42 ) ẋ

Figure 5-6: Graph of index-3 system after one step of Pantelides’ algorithm
Equations Variables

(5.40) v̇

(5.41 ) F

(5.42 ) ẍ

Figure 5-7: Graph of index-3 system after two steps of Pantelides’ algorithm

which replaces the (5.42) equation node in Figure 5-5 to produce the graph in Figure 5-
6 for the next step of the algorithm.
It can be seen from Figure 5-6 that no matching exists because (5.41) and (5.42 )
each have only one edge connecting to the same variable, ẋ. In the scenario depicted
in Figure 5-6, the algorithm has assigned ẋ to (5.41), and cannot assign (5.42 ). Since
the two equations cannot both be assigned to ẋ, they are a structurally singular subset
and are differentiated giving:

ẍ = v̇ (5.41 )

ẍ = x̄ (t) (5.42 )

which results in the graph in Figure 5-7.

The algorithm terminates at this step because a matching is found, shown by the
bold edges in Figure 5-7 which connect each equation to a unique variable. The UDAE

156
returned by the algorithm on the last step is index-1 because it contains an algebraic
state variable F . The final UDAE returned by the algorithm must be structurally
index-1, since the algorithm is guaranteed to return a UDAE with structural index at
most one, and it cannot be index-0 since there is an algebraic variable F in the graph.
Therefore, the original DAE (5.40–5.42) was index-3, since (5.42) was differentiated
twice, and the final UDAE returned by the algorithm is index-1. The initial conditions
must satisfy all of the equations in all of the UDAEs produced by the algorithm, i.e.:
(5.40-5.42), (5.41), (5.42 ), and (5.42 ).

5.2.3 A Modified Pantelides’ algorithm

An implementation of Pantelides’ algorithm (PALG) and a sparse implementation

(SPALG) was reported in [139]. These programs return the structural index and the
structural degrees of freedom of a DAE, as well as the numbers of differentiations
that have to be applied to each equation in the DAE to obtain the nonredundant
equations that must be satisfied by a consistent initialization. Another algorithm
(ALGO) was also reported in [139] which is essentially a structural implementation
of the elimination index-reduction method of [5, 58]. The combined use of ALGO and
PALG returns a list of differentiations and algebraic eliminations required to produce
an equivalent index-1 DAE by the elimination method.
A modified version of Pantelides’ algorithm has been implemented that is some-
what different from the ALGO/PALG implementation. Since the method of dummy
derivatives is being used, rather than the elimination method to derive an equivalent
index-1 DAE, the information returned by ALGO is not needed. The implementation
takes Pantelides’ algorithm and combines it with symbolic differentiation technology.
The algorithm takes a symbolic representation of the equations in the DAE, and
returns a symbolic representation of the nonredundant equations that must be sat-
isfied by a consistent initial condition, with the caveat that since the algorithm is
structural in nature, some required differentiations may not be performed. The mod-
ified version of Pantelides’ algorithm was implemented in the ABACUSS large-scale
equation-oriented modeling system. Section 5.3 discusses the symbolic representation

157
of equations and the differentiation approach employed.
Modifications to Pantelides’ algorithm are also necessary to keep the differentiated
DAE from having an order higher than 1, so that the standard DAE solvers may be
used. This criteria is enforced simply by introducing new equations and variables as
necessary. Thus, when:

f1 (ẋ, x, y, t) = 0 (5.43)

is differentiated:

f1 (ȧ, ẋ, x, ẏ, y, t) = 0 (5.44)

a = ẋ (5.45)

is obtained, where (5.45) is a new equation and a is a new variable introduced into the
model unless it is already present. This is somewhat more complicated in practice
than it would seem because of the desire not to introduce a if it is already in the
problem.
A sketch of the index detection and differentiation algorithm REDUCE-INDEX
that has been implemented in ABACUSS is given below. It uses the AUGMENT-
PATH algorithm given in [107]. The inputs to the algorithm are the DAE f , the
initial conditions φ, the time derivatives of the differential state variables ẋ, and the
algebraic state variables y. Upon successful termination, the algorithm returns the
extended DAE F̄ , its initial conditions D, and the structural index of F̄ .

REDUCE-INDEX(f ,φ,ẋ,y)

1. INITIALIZATION: F0 := f , z := (ẋ, y) , D := φ, Index := 0, Q := ∅.

2. IF the DAE F0 is structurally ill-posed, THEN STOP.

3. ∀ {i} ∈ FIndex :

(a) Apply the AUGMENTPATH algorithm to equation {i} and variable set z
to obtain a set of structurally singular equations W ⊆ FIndex .

158
 (b) ∀ {j} ∈ W :

i. For every time derivative variable q̇ ∈ ẋ in equation {j} create a

new equation {k} of the form a = ẋ, where a is a new variable. Set
z := z ∪ ȧ and ẋ = ẋ ∪ a. IF {k} ∩ (FIndex ∪ W ) = ∅ THEN set
FIndex := FIndex ∪ {k} and Q = Q ∪ {k}.

ii. For every algebraic variable q ∈ y that is in equation {j} set ẋ := ẋ∪ q̇,
y := y \ q, and z := (z \ q) ∪ q̇.

iii. Differentiate equation {j} to obtain equation {l}. Set OldIndex :=

Index.

iv. Set DerivNo :=the number of times an equation {m} ∈ F0 was

differentiated to obtain {l}. IF DerivNo > Index THEN Index :=
DerivNo.

v. FIndex := (FOldIndex \ {j}) ∪ {l}, Q := Q ∪ {l}.

4. ∀ {i} ∈ FIndex :

(a) Apply the AUGMENTPATH algorithm to equation {i} and variable set ẋ
to obtain a set of structurally singular equations W ⊆ FIndex .

(b) IF W = ∅ THEN Index := Index + 1, GOTO 6.

5. IF Index > 0 THEN D := D ∪ (FIndex ∩ Q), Q := Q \ (FIndex ∩ Q).

6. F̄ := FIndex−1 ∪ Q.

7. ∀ {i} ∈ G:

(a) Apply the AUGMENTPATH algorithm to equation {i} and variable set
ẑ = z ∪ x to obtain a set of structurally singular equations W ⊆ G.

(b) IF W = ∅ THEN STOP.

8. RETURN(Index, F̄ , D, F0 , . . . , FIndex−1 ).

159
 Like the original algorithm in [107], REDUCE-INDEX will terminate if and only
if the corresponding extended system:

f (ẋ, x, y, t) = 0 (5.46)

hi (ẋ, x) = 0 ∀i = 1 . . . mx (5.47)

is structurally nonsingular with respect to all occurring variables. The exact form of
(5.47) is not important since the condition is structural, but in practice it could be
thought of as a difference formula relating ẋ and x. In [139] the termination condition
(5.46–5.47) is noted to be equivalent to nonsingularity of the structural matrix pencil :
   
∂F ∂F
pat + pat (5.48)
∂ ż ∂z

where pat(A) is denotes the structural realization of matrix A. Structural non-

singularity of this pencil may be regarded not only as a termination condition for
REDUCE-INDEX, but also as a practical solvability condition for the DAE, since
most numerical methods (including the BDF methods that used here) require the
matrix pencil λ(∂F/∂ ż) + (∂F/∂z) to be regular since it is necessary to invert a ma-
trix of that form to solve the corrector iteration. The REDUCE-INDEX algorithm
checks this condition in Step 2.

Upon successful termination, Pantelides’ algorithm returns a DAE with structural

index at most one [107]. Steps 4-5 check the index of the returned DAE by detect-
ing singularity of the matrix (∂F/∂ ẋ). Structural nonsingularity of this matrix is a
necessary condition for the DAE to be an (index-0) ODE. If the algorithm returns
an ODE, it is preferred to ‘backtrack’ in the algorithm to find an index-1 DAE since
its extended DAE F̄ will have fewer equations. Step 5 does this for index-0 DAEs
by taking the equations created on the last step and appending them to the initial
conditions D, rather than the extended DAE F̄ . In this way, the equations created
on the last step will be enforced in a consistent initialization, but they will not be
explicitly enforced during the integration.

160
 Step 7 is a check of the structural nonsingularity of the extended DAE and its
initial condition with respect to (ẋ, x, y) to ensure that the initial conditions are
consistent in a structural sense. This check has proven to be useful in practical
modeling activities, although it by no means ensures that it is easy to obtain a set of
consistent initial values.
Implementation of this algorithm in ABACUSS has proved to be very useful for
several reasons:

• Structurally ill-posed problems are detected before the start of the integration.

• The user is informed that the DAE is high-index, which is important because
the formulation of a high-index DAE could be unintentional.

• The correct equations for initializing a high-index DAE are automatically de-
rived and reported.

• The algorithm is necessary for the dummy derivative method.

5.2.4 The dummy derivative algorithm

Although the REDUCE-INDEX algorithm is useful by itself to determine equations

that must be satisfied by a consistent initialization, it is not possible to solve directly
the DAE F̄ that is returned by REDUCE-INDEX if the original DAE was high-
index. The reason is that F̄ is an overdetermined system; that is, there are more
equations than state variables. The elimination method for deriving an equivalent
index-1 system from F̄ was described above. A somewhat similar but more practically
useful method, the method of dummy derivatives was proposed by [97].
The dummy derivative method works by substituting some (or even all) of the
time derivatives with “dummy” algebraic variables, effectively removing these time
derivatives from the time discretization and resulting in a completely determined sys-
tem. Unlike the elimination method, which eliminates some variables in the original
DAE using the extra constraints of the extended DAE, the dummy derivative method

161
adds variables so that the entire extended system may be solved. The main advan-
tage of the dummy derivative algorithm is that it does not require computationally
expensive algebraic substitutions. The key to the method is to pick a set of time
derivatives to replace by dummy algebraic variables.
The method of dummy derivatives is described in detail in [97], but it is summa-
rized here for the purpose of clarity in the rest of this chapter. For convenience, let
us rewrite the DAE as an operator equation:

Fz = 0 (5.49)

where the dependent variables z may appear algebraically or differentiated up to q

times. It is assumed that for some p > q that z(t) : R → Rn and F z : R → Rn .
Let D = d/dt denote the differentiation operator, and for ν ∈ Nn define the operator
D = diag(D ν1 . . . D νn ) . Assign μ(F ) ∈ Nn such that D μ(F ) are the highest order
derivatives appearing in the DAE. REDUCE-INDEX finds the minimal ν such that
the differentiated problem FIndex = Gz = D ν F z = 0 is structurally nonsingular with
respect to the highest order derivatives D μ(F ) z.
The dummy derivative algorithm has three steps:

1. Use the REDUCE-INDEX algorithm to differentiate equations in the DAE to

reduce the index to one. REDUCE-INDEX returns an overdetermined UDAE
F̄ which can be used to determine both ν and μ(F ). It also returns a UDAE
FIndex . It is assumed that FIndex is in Block Lower Triangular (BLT) form. Let
gi denote the ith block of of FIndex and let zi denote the vector of highest order
derivatives of the unknowns associated with gi .

2. Sort each gi in descending order with respect to the number of differentiations

required to obtain each equation. If gi has m differentiated equations, let hi
denote the m first equations of gi . Define the Jacobian Hi1 = ∂hi /∂zi .

3. Select m columns of Hij to make a (square) nonsingular matrix Mij . Replace

the time derivatives that are associated with the selected columns with dummy

162
 algebraic variables.

4. Increment j and repeat Step 3 using the “predecessor” of hi , which is D −(j−1) hi =

0, thus eliminating the last differentiation. The new candidates for replacement
with dummy derivatives are time derivatives that are of one order less than
those replaced on the previous step. Continue to repeat these two steps until
there are no more candidates for replacement.

5. The equivalent index-1 DAE is constructed by including all original and all
differentiated equations, with dummy algebraic variables substituted for those
variables indicated in Step 3.

5.2.5 Example of dummy derivative algorithm

As an example of the application of the dummy derivative algorithm, consider the

pendulum problem (5.22–5.24) in which substitutions have been made to reduce the
order of the highest order time derivatives to one. When REDUCE-INDEX is applied
to this problem, F̄ consists of:

a = ẋ (5.50)

b = ż (5.51)
λ
mȧ + x=0 (5.52)
L
λ
mḃ + y = −mg (5.53)
L
x2 + y 2 = L2 (5.54)

2xẋ + 2y ẏ = 0 (5.55)

2xȧ + 2a2 + 2y ḃ + 2b2 = 0 (5.56)

Block triangularization results in one block g1 , consisting of (5.50–5.53) and (5.56).

Since the only differentiated equation in g1 is (5.56), the vector of highest order

163
derivatives z1 = (ȧ, ḃ, ẋ, ẏ, λ) and the Jacobian H11 is:

H11 = 2x 2y 0 0 0 (5.57)

There are two choices of columns that would make M11 nonsingular, because neither
x nor y are nonzero for all t. The states x and y are not zero simultaneously because
of the length constraint. When either x nor y is locally zero, the choice of M11 is clear.
When neither x nor y are locally zero, either choice is valid.

If ȧ is chosen as the dummy derivative, z2 = [ẋ]. Even though the ȧ does not
appear to have a time derivative of one order less than ȧ, the substitution ẍ = ȧ
was made when doing the differentiation to obtain (5.56) and therefore ẋ is a time
derivative of one order less than ȧ. The H21 matrix is simply:

H21 = 2x (5.58)

and thus the choice of dummy derivatives is clear.

Depending on the choice of M11 matrices, the following two equivalent index-1
formulations of (5.22–5.24) can be obtained:

a = x̄ (5.59)

b = ẏ (5.60)
λ
mā + x=0 (5.61)
L
λ
mḃ + y = −mg (5.62)
L
x2 + y 2 = L2 (5.63)

2xx̄ + 2y ẏ = 0 (5.64)

2xā + 2x̄2 + 2y ḃ + 2ẏ 2 = 0 (5.65)

164
and:

a = ẋ (5.66)

b = ȳ (5.67)
λ
mȧ + x=0 (5.68)
L
λ
mb̄ + y = −mg (5.69)
L
x2 + y 2 = L2 (5.70)

2xẋ + 2y ȳ = 0 (5.71)

2xȧ + 2a2 + 2y b̄ + 2b2 = 0 (5.72)

where ā, b̄, x̄, ȳ denote dummy derivatives (that is, they are algebraic state variables).

This example is useful because it illustrates several key points about the dummy
derivative algorithm. Since in general the choice of a nonsingular square sub-matrix
at each step of the algorithm is not unique, a general high-index DAE will have a
“family” of equivalent index-1 DAEs. Furthermore, it is important to recognize that
the dummy derivative method relies on numerical nonsingularity of the M matrices,
which in general may be a local property of the DAE. Therefore, it may become nec-
essary to perform dummy derivative pivoting between the family of equivalent index-1
DAEs, in which different M matrices and therefore a different set of dummy deriva-
tives are selected during the solution of the DAE depending on the local properties
of H.

Detecting the need for dummy derivative pivoting would seem to require the rect-
angular H matrices to be factored at every integration step of the simulation, which
is computationally prohibitive. In [97] it is demonstrated that nonsingularity of Mi
implies nonsingularity of Hi−1 , hence only Hi−1 need be monitored. In general, each
matrix M could be as large as the Jacobian matrix of the original high-index system,
although in many practical problems it has been found to be much smaller. Described
below is a strategy that avoids this expensive factorization at each step, which other-
wise would make the dummy derivative method extremely costly in a similar manner

165
to the staggered direct method for sensitivities described in Chapter 4.

166
5.3 Differentiation in ABACUSS

Derivation of the family of equivalent index-1 DAEs required by the method of dummy
derivatives in an automated fashion requires the application of computational differ-
entiation technology. This section describes the differentiation strategy adopted for
implementation of the dummy derivative method in ABACUSS.

Step 3(b)iii of algorithm REDUCE-INDEX above requires derivation of the deriva-

tive with respect to the independent variable time of one element of the vector of
implicit relationships (5.1). This derivative is defined by the sparse inner product
of the partial derivative vector of fi with respect to {ẋ, x, y, t} and the derivative
vector (ẍ, ẋ, ẏ, 1)T (note from (5.44–5.45) that ẍ is immediately eliminated). In many
engineering applications it can normally be assumed that there are on average only a
small number of entries in the partial derivative vector that are not identically zero.
The fact that this derivative must also equal zero results in a new equation that is aug-
mented to the original system. Note that subsequent iterations of REDUCE-INDEX
may require derivation of the time derivative of this new equation. In addition, dif-
ferentiation technology is employed to derive the Jacobian of the augmented system
(which has similar sparsity) for use by an index-1 solver.

ABACUSS is an example of high level modeling environment with an interpre-

tative software architecture. Rather than automatically generating FORTRAN or C
code which is then compiled and linked to form a simulation executable, ABACUSS
creates data structures representing the model equations in machine memory, and
during numerical solution these data structures are “interpreted” by numerical algo-
rithms to evaluate residuals, partial derivatives, etc. This approach offers significant
advantages from the point of view of interactive model development, debugging, and
reporting and diagnosis of problems during solution of the model [116]. Further, it
allows the functional form of the model to be manipulated efficiently in a general
manner throughout the solution process. This latter feature is particularly advanta-
geous for the solution of combined discrete/continuous simulations [15] and inequality
path constrained dynamic optimization problems [48]. In both cases, solution of the

167
IVP is characterized by a sequence of discrete changes to the functional form of the
system (5.1) along the trajectory. For example, in the case of the activation of an
inequality path constraint at a specific time, the active inequality must be augmented
to the DAE and REDUCE-INDEX applied to derive a family of equivalent index-1
models for this new DAE. These changes in the functional form of the DAE poses
problems for the code generation approach: either new code must be generated and
linked each time the DAE changes, or all possible functional forms must be antici-
pated a priori and appropriate code generated for each one. The former approach
is very costly at run time, whereas the latter is in general combinatorial, since, for
example, all admissible combinations of inequality constraint activations must be enu-
merated, even if only a small subset is encountered along the trajectory. On the other
hand, the interpretative approach allows for very efficient manipulation of the DAE
by merely adjusting the relevant data structures and performing differentiation in an
incremental manner as necessary along the trajectory.
Details of the ABACUSS implementation are given in [13]. Sketched here is how
a typical problem is processed. The user defines the model equations in natural
language using the high level ABACUSS input language. In the first phase, this input
is then compiled into an intermediate representation of the model held in appropriate
data structures. In the second phase, a particular calculation is selected and the
ABACUSS simulation executive transforms this intermediate representation into the
run-time data structures for a calculation. In particular, a function fi in (5.1) is stored
in a binary tree using a dynamic data structure. Given values for the unknowns, a
subroutine can then “interpret” this binary tree to evaluate the function.
Derivatives of all of the functions are obtained using automatic differentiation
technology. Automatic differentiation is a method of obtaining derivatives symboli-
cally without resulting in unnecessary and highly inefficient expression “swell” in the
resulting expressions. The automatic differentiation in ABACUSS uses an algorithm
detailed in [135, 134] that is highly efficient for large sparse equations.

168
5.4 Dummy Derivative Pivoting
The key to the dummy derivative method is the selection of dummy derivatives. As
shown in the Section 5.2.5, the choice is in many cases non-unique, and a practical
implementation must select the set of dummy derivatives, monitor the set to see if it
is no longer optimal or valid, and switch (pivot) among equivalent index-1 DAEs as
necessary.

5.4.1 Selection of an equivalent index-1 model

The problem of selecting an appropriate equivalent index-1 DAE from the family
of possible equivalent index-1 DAEs depends on selecting a nonsingular square sub-
matrix Mi1 . Since structural algorithms are needed in the REDUCE-INDEX algo-
rithm, it might seem reasonable to choose any structurally nonsingular submatrix of
Mi1 . However, as the example in Section 5.2.5 demonstrates, structural criteria are
not sufficient to detect the local points at which an equivalent index-1 model is not
valid. A reliable method for selecting nonsingular sub-matrices from a rectangular
matrix (provided that such a matrix exists) is to use Gaussian elimination with full
pivoting. This Gaussian elimination has been implemented using the MA48 code [45],
which is capable of LU factoring large sparse rectangular matrices. Although this step
is fairly expensive computationally, the overall expense may be reduced by minimizing
the number of times that it is done during solution of the DAE, as described below.

5.4.2 Reasons for switching index-1 models

The questions that arise when the example in Section 5.2.5 are examined is what
is happening when the Mi1 matrix of an equivalent index-1 DAE becomes singular,
and whether there are reasons to prefer one equivalent index-1 DAEs over another
provided their respective Mi1 matrices are nonsingular.
Dummy derivative pivoting is required because the current matrices Mi1 may
become locally singular at certain points in state space. The consequence of a singular
Mi1 is that the equivalent index-1 DAE may cease to define uniquely all of its highest-

169
order time derivatives at some points in the solution trajectory. To see this, consider
one of the equivalent index-1 models (5.59–5.65) of the pendulum. When this system
is solved for explicit expressions for the highest-order time derivatives, the following
system is obtained:

a = x̄ (5.73)

b = ẏ (5.74)
L2 ymg − L2 mb2 − y 3mg
λ=− (5.75)
(−y 2 + L2 ) L
yb2 L2 + gy 4 − 2gy 2L2 + gL4
ḃ = − (5.76)
(−y 2 + L2 ) L2
+
x = ± −y 2 + L2 (5.77)
yb
x̄ = ∓ + (5.78)
−y 2 + L2
L2 yg − L2 b2 − y 3 g
ā = ± + (5.79)
L2 −y 2 + L2

This system does not uniquely define the highest order time derivatives when
y = ±L. The situation is similar to that encountered with the elimination method
(5.37–5.39), since no finite number of differentiations of (5.73–5.79) will produce a
UDAE for which there are not powers of (−y 2 + L2 ) in the denominator. In general,
all that is known about points where Mi1 becomes singular is that the highest-order
time derivatives of the equivalent index-1 DAE may not be uniquely defined, which
can result when the system is locally high-index or locally unsolvable.

Interestingly, the corrector matrix used the BDF method (which is essentially the
local matrix pencil h1 (∂F/∂ ż) + (∂F/∂z) of the system (5.73–5.79)) does not become
locally singular when y = ±L. Indeed, high-index models also do not necessarily have
a singular corrector matrix. However, standard codes do experience difficulty when
solving an equivalent index-1 model in the neighborhood of such points (the chances
of the integrator stepping exactly onto one of these points are vanishingly small).
The codes can cut the step-size drastically in the vicinity of the singular point, and
sometimes fail to integrate past the point. It was observed that the corrector matrix

170
 Corrector Condition Number
1e+12
1e+11 x=0.00
x=0.10
1e+10 x=0.25
1e+09 x=0.50
x=0.75
1e+08 x=0.90
1e+07 x=1.00
1e+06
100000
10000
1000
100
10
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
h

Figure 5-8: Condition number of the corrector matrix at different points on the so-
lution trajectory of the equivalent index-1 pendulum as a function of the step size
h

becomes ill-conditioned near such points, which can cause inaccuracy in the solution
to the corrector and trigger step failures [2].
Figure 5-8 shows the LINPACK estimate of the corrector matrix condition number
for (5.59–5.65) as the integrator step size h → 0. The corrector matrix condition
number increases as x → 0. When the corrector matrix becomes very ill-conditioned
(above about 5 · 105 in this case) the corrector matrix may not converge or it may
converge to an incorrect point and trigger a truncation error test failure.

5.4.3 Deciding when to switch index-1 models

A previously unanswered problem for the dummy derivative method is how to detect
automatically when dummy derivative pivoting is necessary during the solution of the
DAE. The brute-force method would be to perform Gaussian elimination on the Hi1
matrices at every integration step and detect whether a different equivalent index-1
model is chosen than the one most recently solved.
The observation made in the previous section that singular Mi1 matrices lead
to ill-conditioned corrector matrices which can trigger truncation error test failures

171
provides a convenient heuristic for checking whether dummy derivative pivoting is
necessary. The heuristic is to check the factorization of the Hi1 matrix whenever
there are repeated corrector failures or truncation error failures and the integrator
calls for a cut in the step size. An example demonstrating the use of this heuristic is
given below.
This heuristic is somewhat too conservative. It may call for dummy derivative
pivoting at points where none is necessary and the step-size is being reduced for
reasons unrelated to the index of the model. On the other hand, the heuristic does
not call for pivoting unless the integrator experiences difficulty, which means that
some ill-conditioned points may be passed without pivoting. Our experience shows
that the heuristic does result in infrequent dummy derivative pivoting checks for
most problems. Other situations where there are ill-conditioned corrector matrices
are analyzed in [2, 1].

5.4.4 Switching index-1 models during integration

When it has been decided to switch between equivalent index-1 DAEs during the
solution, it is unnecessary to restart the integration code. Doing so is undesirable
because it requires small step-sizes, many corrector matrix factorizations, and low
order polynomial interpolation, increasing the cost of solving the DAE.
One issue with dummy derivative pivoting is that, although it is not necessary the-
oretically to restart the integration after the pivoting, in practice the DAE integrator
may be forced to cut the step size drastically. This happens because the integrator is
controlling the truncation error on the differential and algebraic state variables, but
not directly on the time derivatives. When one of the time derivatives becomes an al-
gebraic dummy derivative, accurate predictor information does not exist for this new
algebraic variable, and the solver is forced to cut the step-size, sometimes drastically.
One way to address this is to introduce an extra equation into the system of the form:

a = ẋ (5.80)

172
where a is a new algebraic variable, for all the time derivatives in the current DAE that
have the possibility of becoming dummy derivatives. It is only necessary to include
these equations in the truncation error control, since solution of (5.80) is trivial and
may be performed after the corrector iteration has finished. The DAE solver is able
to take fewer steps after a dummy pivoting operation if it is able to avoid cutting the
step by using extrapolation information obtained from equations like (5.80).

173
 Dummy Derivative Method Pendulum Length
1.00001

1.00001

1
L

0.999995

0.99999
0 50 100 150 200 250 300
t

Figure 5-9: Length constraint in dummy derivative solution of high-index pendulum

5.5 Pendulum Demonstration and Numerical Re-

sults
The use of ABACUSS to solve high-index problems is demonstrated using the pendu-
lum problem (5.22–5.24) with parameters L = 1, g = 1, and m = 1, initial condition
x = 1 and ẏ = 0, and integration tolerances RT OL = AT OL = 10−5 . Figure 5-9
shows that, unlike when this problem was solved using the Gear Method (Figure 5-4),
the dummy derivative solution to the problem satisfies the length constraint to within
the integration tolerances over many oscillations of the pendulum.
To demonstrate the use of the dummy pivoting heuristic, the index-3 pendulum
problem (5.22-5.24) was solved for t ∈ (0, 10]. The different pivoting strategies tested
were:

1. Check for pivot when the DAE solver calls for a cut in the step size.

2. Pivot to minimize the condition number of the corrector matrix.

3. Pivot whenever factoring the H matrix at each step leads to a different M

matrix.

The heuristic proposed in Section 5.4.3 is used in Strategy 1. Strategy 2 was

174
 Table 5-1: Different Pivoting Strategies
Pivot Pivot Trunc.
Pivots Steps f-evals Jac ΔL ΔE
Strat. Checks Fail
1 7 5 256 772 110 7 1.4 · 10−7 4.7 · 10−7
2 n/a 6 221 656 103 4 6.3 · 10−8 6.2 · 10−6
3 215 6 221 656 103 0 6.3 · 10−8 6.2 · 10−6

implemented by monitoring the condition number of the corrector matrix associated

with both pendulum equivalent index-1 DAEs. Solution statistics for each of these
strategies are given in Table 5-1.

The headings in Table 5-1 refer to the number of pivot checks, the number of piv-
ots, the number of integration steps, the number of function evaluations, the number
of evaluations and factorizations of the corrector matrix, and truncation error test
failures. There were no corrector convergence failures under any of the strategies.
The headings ΔL and ΔE refer to the maximum deviation in the total energy and
length constraints during the time interval. The total energy was calculated after the
integration using the formula:

1
E = m(ẋ2 + ẏ 2 ) + mgy (5.81)
2

The deviation in the energy was measured relative to the value of the energy on
the first step of the integration. The length constraint is present explicitly, and the
drift in this constraint was maintained below the integrator tolerances for all of the
strategies. The energy constraint is an implicit constraint, which is why it was not
enforced as closely as the length constraint, but the drift is still very small.

Strategies 2 and 3 have identical solution statistics because they result in exactly
the same pivot times. Strategy 2 checks the pivot selection at every step except
immediately following dummy derivative pivots, which accounts for the discrepancy
in the number of steps and the number of pivot checks. Table 5-1 shows that the
policy of checking the pivots only when the integrator cuts the step (strategy 1) carries
a price in terms of the number of steps, the number of residual evaluations, and the

175
 Dummy Derivative Method for Index-3 Pendulum

1.5 x
y
Pivot Times
1

0.5

-0.5

-1
0 2 4 6 8 10
Time

Figure 5-10: Solution of index-3 pendulum model using ABACUSS

number of corrector matrix factorizations. However, this price is small compared to

the cost of calculating the condition number of the two corrector matrices on every
step (strategy 2) or checking the factorization of the H matrix on each step (strategy
3).
Using strategy 1, Figure 5-10 shows the times at which ABACUSS performed
dummy derivative pivoting on the pendulum. Figure 5-11 shows that the dummy
derivative pivoting was indeed performed at times at which the corrector matrix was
ill-conditioned. The “stepwise” appearance of Figure 5-11 is due to the fact that
efficient BDF codes do not update the corrector matrix at every solution step.
Table 5-2 gives (strategy 1) results for the solution of the model both with and
without the addition of equations (5.80). The integration statistics show a clear
benefit from including the extra equations so that the DAE solver is not forced to cut
the step after the DAE is changed during dummy derivative pivoting.

Table 5-2: Results for Solution of Index-3 Pendulum

Steps f-evals Jac Corr. Fail Trunc. Fail
Pendulum w/ pivot equations 256 772 110 0 7
Pendulum w/o pivot equations 297 924 165 0 7

176
 Dummy Derivative Method for Index-3 Pendulum
1e+07
Corrector Condition Number
1e+06 Pivot Times

100000

10000

1000

100

10
0 2 4 6 8 10
Time

Figure 5-11: LINPACK estimate of corrector matrix condition number

177
 Table 5-3: Example problem solution statistics
Batch
Problem Condenser Standard CSTR
Column
Model Equations 5 20 4 297
Additional Eqns. Derived 2 0 5 145
Input Functions 1 4 0 23
Input Funcs. Derived 0 19 0 2
Initial Conditions 1 0 0 9
Integration Steps 363 63 65 96
Residual Evaluations 825 128 144 201
Jacobian Factorizations 62 7 16 28
Corrector Conv. Failures 0 0 0 0
Truncation Test Failures 17 1 2 3
Pivot Checks 17 0 0 3
Pivots 0 0 0 2

5.6 Numerical Examples

This section provides several examples of the use of the dummy derivative method
on problems of interest in chemical engineering. The problems used in this chapter
have been discussed in the literature under the context of high-index DAEs, but no
numerical solutions for most of the problems have been reported. Table 5-3 gives the
solution statistics that were obtained by solving these problems using the dummy
derivative method implemented in ABACUSS.

178
5.6.1 Fixed-volume condenser with no liquid holdup

The following simple model of a condenser with fixed volume and negligible liquid
holdup was discussed in [108, 139].

Ṅ = F − L (5.82)

NCp Ṫ = F Cp (Tin − T ) + LΔH + kA(Tc − T ) (5.83)

P V = NRT (5.84)
 
101325 B
P = exp A − (5.85)
760 C −T

where (5.82) represents a material balance, (5.83) the energy balance, (5.84) the
vapor-liquid equilibrium, and (5.85) an equation of state for the vapor. It was noted
in [139] that the assumptions used to derive this model are somewhat questionable,
since the vapor phase is modeled as both an ideal gas (5.83) and a saturated vapor
(5.85).

The heat capacity was modeled using the relation:

Cp = α + β(T − 273.15) + γ(T − 273.15)2 + μ(T − 273.15)3 (5.86)

The feed F is forced with the periodic function:

π
F = 9000 + 1000 sin( t) (5.87)
2

The definitions of the variables and parameters are given in Tables 5-4 and 5-5.
The parameters for the heat capacity equation, vapor pressure equation, and heat of
vaporization are for water and are taken from [50].

The combined model (5.82–5.87) is index-2, and ABACUSS must differentiate

(5.84–5.85) to produce an equivalent index-1 DAE. There is one dynamic degree of
freedom, and therefore an initial condition must be specified for any one of the state
variables (T, L, P, N). Figure 5-12 shows the output that ABACUSS produces for this
high-index model. Figures 5-13 and 5-14 show solution trajectories for T and N on the

179
 Table 5-4: Variables in the high-index condenser model
Variable Units Description
N mols Molar Holdup in the vessel
T K Temperature in the vessel
F (mols/hr) Feed flow rate
P Pa Pressure
L (mols/hr) Liquid flow rate out of vessel
Cp (J/mol · K) Vapor heat capacity

Table 5-5: Parameters in the high-index condenser model

Parameter Value Description
A 7.96681 Antoine Equation Coefficient
B 1668.21 Antoine Equation Coefficient
C 228 Antoine Equation Coefficient
Tin 513 K Feed Temperature
Tc 280 K Coolant Temperature
V 100 m3 Vessel volume
U 10000 J/(hr · m2 · K) Heat transfer coefficient
A 50 m2 Heat transfer Area
ΔH −40656 J/mol Heat of Condensation
R 8.3145 J/(mol · K) Universal gas constant
α 33.46 Heat Capacity Correlation Coefficient
β 0.688 · 10−2 Heat Capacity Correlation Coefficient
γ 0.7604 · 10−5 Heat Capacity Correlation Coefficient
μ −3.593 · 10−9 Heat Capacity Correlation Coefficient

180
time interval (0, 10 hrs] that are obtained from an initial condition of T (0) = 400K.

181
 The following are the equations in the model:
Input Equation # 1
CONDENSER.F = 9000 + 1000*SIN(3.14286E+00*TIME/2) ;

Equation # 2
CONDENSER.$N = CONDENSER.F - CONDENSER.L ;

Equation # 3
CONDENSER.N*CONDENSER.CP*CONDENSER.$T = CONDENSER.F*CONDENSER.CP*
(5.13150E+02 - CONDENSER.T) + -4.06560E+04*CONDENSER.L -
5.00000E+05*(CONDENSER.T - 2.80000E+02) ;

Equation # 4
CONDENSER.P*1.00000E+02 = CONDENSER.N*8.31450E+00*CONDENSER.T ;

Equation # 5
CONDENSER.P = 1.33322E+02*EXP(7.96681E+00 - 1.66821E+03/
(2.28000E+02 + (CONDENSER.T - 2.73150E+02))) ;

Equation # 6
CONDENSER.CP = 3.34600E+01 + 6.88000E-02*(CONDENSER.T - 2.73150E+02)^2
+ 7.60400E-06*(CONDENSER.T - 2.73150E+02)^3 + -3.59300E-09*
(CONDENSER.T - 2.73150E+02)^4 ;

Analyzing the structure of the model...

This model is structurally high-index.

Proceeding with index reduction...

Equation # 7 created by differentiating equation # 5

-1*(-1*(-1*(1.66821E+03/(2.28000E+02 + (CONDENSER.T - 2.73150E+02))^2))
*EXP(7.96681E+00 - 1.66821E+03/(2.28000E+02 +
(CONDENSER.T - 2.73150E+02)))*1.33322E+02)*CONDENSER.$T +
1*CONDENSER.$P = 0 ;

Equation # 8 created by differentiating equation # 4

-1*(CONDENSER.N*8.31450E+00)*CONDENSER.$T + -1*(8.31450E+00*
CONDENSER.T)*CONDENSER.$N + 1.00000E+02*CONDENSER.$P = 0 ;

Before differentiation, this model was index 2.

Equations 1 through 8 form an index=1 model.

Equation 9 is an initial condition.
The following are the specified initial conditions:
Equation # 9
CONDENSER.T = 400 ;

RESULTS FROM ABACUSS STRUCTURAL ANALYSIS

Number of Variables in model : 6

Number of additional variables created : 0
____
Total Variables : 6

Number of Model Equations : 5

Number of Additional Model Equations Derived : 2
Number of Input Functions Specified : 1
Number of Additional Input Functions Derived : 0
Number of Initial Conditions Specified : 1
Number of Additional Initial Conditions Derived : 0
____
Total Equations : 9

Total Unknowns : 9

Figure 5-12: ABACUSS index-reduction output for high-index condenser model

182
 Temperature Profile for High-Index Condenser
500

498

496
T (K)

494

492

490
0 1 2 3 4 5 6 7 8 9 10
t(hr)

Figure 5-13: Dummy derivative Temperature profile for high-index condenser

Mole Holdup Profile for High-Index Condenser

234

232

230
N (mol)

228

226

224

222

220
0 1 2 3 4 5 6 7 8 9 10
t(hr)

Figure 5-14: Dummy derivative mole holdup profile for high-index condenser

183
 Index-20 DAE
1
0.8 x17
x18
0.6 x19
x20
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
0 1 2 3 4 5 6 7 8
t

Figure 5-15: State trajectories for index-20 DAE

5.6.2 Standard high-index model

A ‘standard high-index’ model was proposed and solved in [76]. The model is:

ẋi = xi+1 i = 1...N − 1 (5.88)

xi = f (t) (5.89)

The index of this system is N, and hence the model is primarily interesting because
the index is a function of a parameter. An index-20 version of this problem was
solved, with f (t) = sin(t), which required ABACUSS to differentiate 19 equations.
Some of the state trajectories are shown in Figure 5-15.

184
 Table 5-6: Parameters for the CSTR example
Parameter Value
C0 5 mol/L
T0 273 K
K1 20 s−1
K2 3000 (K · L)/mol
K3 1500 s−1
K4 3000 K

5.6.3 Continuous stirred-tank reactor

The following model of a well-mixed continuous stirred-tank reactor (CSTR) in which

an exothermic reaction takes place:

Ċ = K1 (C0 − C) − R (5.90)

Ṫ = K1 (T0 − T ) + K2 R − K3 (T − Tc ) (5.91)
 
−K4
R = K3 exp C (5.92)
T

where C is the reactant concentration, T is the temperature in the reactor, C0 and T0

are the feed concentration and temperature, R is the rate of reaction, Tc is the coolant
temperature, and K1 , K2 , K3 , K4 are parameters. This model was also discussed in
[108], but no numerical solution was given. The values of the parameters used to
obtain a numerical solution are given in Table 5-6.
Figures 5-16 and 5-17 show the state variable trajectories obtained from the ABA-
CUSS numerical solution to the problem. An interesting feature of this problem is
that it is index-3, and contains no dynamic degrees of freedom, which means that it
is not possible to specify any initial conditions. Also, since the H11 matrix for this
problem has only one row and one column, ABACUSS is able to determine that pivot
checks are unnecessary and thus the numerical solution is very efficient.

185
 CSTR Reactor
5
4.9 C
4.8
Concentration(mol/L)

4.7
4.6
4.5
4.4
4.3
4.2
4.1
4
0 1 2 3 4 5 6 7 8 9 10
t(hr)

Figure 5-16: Concentration profile for index-3 CSTR example

CSTR Reactor

540 T
Tc
520
T emperature(K)

500

480

460

440

420

400
0 1 2 3 4 5 6 7 8 9 10
t(hr)

Figure 5-17: Temperature profiles for index-3 CSTR example

186
 Batch Frac Column
375.5
Reboiler T
375

T (K) 374.5

374

373.5

373

372.5
0 5000 10000 15000 20000 25000
t

Figure 5-18: Reboiler temperature profile for BatchFrac Column

5.6.4 High-Index dynamic distillation column

An interesting high-index example is the BatchFrac model [20]. As written the model
is index-2 due to assumptions about the holdup on the trays given in equation (6)
of [20]. The fact that the problem is index-2 was not specifically noted in [20],
but the problem was solved by replacing the equations for the holdup and enthalpy
derivatives with finite difference approximations (see equations (13) and (14) in [20]).
The method of dummy derivatives eliminates the need for this approximation. The
BatchFrac model was solved in ABACUSS for a five component mixture using ideal
thermodynamic assumptions with one tray in the column.
The original index-2 DAE system for this problem consisted of 155 equations,
and an additional 75 equations were automatically derived to form the equivalent
index-1 system. This model is large because it includes thermodynamic properties
as equations, not procedures. The reboiler temperature profile obtained from the
solution of the 1-tray system is given in Figure 5-18.

187
5.7 Conclusions
This chapter has shown that it is possible to reliably solve a large class of arbitrary-
index DAEs using the dummy derivative method. The advantages of the dummy
derivative method over other methods that have been proposed are that it directly
enforces all of the implicit constraints and it may be easily automated and combined
with automatic differentiation technology to solve high-index DAEs of practical en-
gineering interest. The dummy derivative method does not work for all high-index
DAEs, notably those for which the index and/or dynamic degrees of freedom can-
not be determined correctly using structural criteria. However, our experience has
shown that numerical solutions to many high-index engineering problems are easily
obtained.
Implementation of the dummy derivative algorithm in ABACUSS required sig-
nificant development of the details of the algorithm, as evidenced in the REDUCE-
INDEX algorithm and the dummy derivative pivoting heuristic. This implementation
appears to be the first example of a computational environment capable of easily solv-
ing high-index DAEs. The ability to solve high-index DAEs directly is necessary for
the development of algorithms to solve state-constrained dynamic optimization prob-
lems, which are discussed in the next chapters.

188
Chapter 6

Equality Path Constraints

This chapter is concerned with the solution of equality path-constrained dynamic op-
timization problems. Equality path constraints (as distinguished from inequality path
constraints, which are the subject of the next chapter) constrain the state variables
of the DAE in the dynamic optimization problem. To date, this class of problem
has not been satisfactorily handled within the control parameterization framework.
However, it is shown in this chapter that equality path-constrained problems contain
high-index DAEs which may be solved efficiently using the dummy derivative method
detailed in Chapter 5.
An equality path-constrained subset of the dynamic optimization formulation
given in Chapter 2 is considered in this chapter:
 tf
min J = ψ (x (tf ) , y (tf ) , tf ) + L (x, u, t) dt (6.1)
u(t),tf to

subject to the DAE:

f (ẋ, x, y, u, t) = 0 (6.2)

h(x, y) = 0 (6.3)

φ (ẋ (to ) , x (to ) , y (to ) , to ) = 0 (6.4)

In this formulation, x are differential state variables, y are algebraic state variables

189
and u are control variables. The path constraint (6.3) is assumed to be a vector
function with h(·) → Rnh . The DAE (6.2) is assumed to have index ≤ 1, although
the results are easily extended if (6.2) is a high-index DAE for which an equivalent
index-1 DAE may be derived using the dummy derivative method. Note that (6.1–
6.4) does not include point constraints at times other than t0 , which are irrelevant for
this chapter, or inequality path constraints on the state variables, which are discussed
in Chapter 7.
Our goal is to solve (6.2–6.3) simultaneously as an IVP within the control pa-
rameterization method, since the efficiency of the control parameterization method
increases as more of the constraints are handled by the IVP solver rather than the NLP
solver. Since (6.3) constrains the states of the system (6.2), it can be expected that
the combined system may be high-index. However, the combined system (6.2–6.3) is
also overspecified, and thus not all of the control variables u are truly independent.
The two problems that must be addressed are:

1. Is the state path-constrained problem feasible?

2. How should the overspecified nature of the problem be handled?

The first issue is related to controllability of nonlinear systems and solvability of

nonlinear high-index DAEs. The latter issue leads to the control matching problem.
Since not all controls are independent, a subset of the controls are allowed to be
determined by the solution of the high-index DAE, thus removing the overspecification
of the problem. The control matching problem determines which controls are in the
subset defined by the high-index DAE, and which ones are free to be determined by
the optimization.
This chapter begins with a review of existing methods for solving equality path-
constrained dynamic optimization problems and a discussion of the limitations of
the current methods. Section 6.2 then presents a series of theorems which describe
the circumstances which produce high-index DAEs in the combined system (6.2–6.3).
Some issues relating to feasibility of the constrained dynamic optimization problem
are discussed in Section 6.3, and the problem of control matching is addressed in

190
Section 6.4. Finally, examples are presented in Section 6.5.

191
6.1 Review of Methods for Solving Path-Con-
strained Dynamic Optimization Problems

Several methods have been developed to date for solving path-constrained dynamic
optimization problems within the control parameterization framework. Since it was
not possible to solve high-index DAEs directly at the time that these methods were
developed, the emphasis was on finding methods that discretize the path constraints so
that they could be included either in the objective function or as a set of NLP equality
constraints. Therefore, these methods all handle the path constraints indirectly, in
the sense that they are included in the master NLP rather than the IVP subproblem.
One method for enforcing path constraints is to modify the objective function to
include a measure of the constraint violation [24]:

nh  tf
J¯ = J + Ki h2i dt (6.5)
i=1 t0

where K is a vector of large numbers. The path constraints are satisfied exactly
only if K → ∞. The problem with this approach is that it has been shown to cause
numerical difficulties with the NLP master problem because it modifies the shape of
the objective function, possibly making it more nonlinear or introducing additional
locally optimal points.
Another method [128] is to replace the path constraints with a single end-point
constraint:
 tf
ϕ= hT hdt = 0 (6.6)
t0

or by a set of end-point constraints:

 tf
ϕi = h2i dt = 0 i = 1 . . . nh (6.7)
t0

Since more information is provided to the optimizer by (6.7) than by (6.6), it is

generally preferred. However, even (6.6) lumps the violation of the constraint over

192
the entire state trajectory into a single measure, and thus it provides only limited
information about how to modify the input functions to achieve feasibility. Also,
both (6.6) and (6.7) have gradients with respect to the optimization parameters that
are zero at the optimum, which reduces the efficiency of gradient-based optimization
methods.

A somewhat more sophisticated method was proposed in [142, 143], where the
relationship between path-constrained dynamic optimization problems and high-index
DAEs was noted. It was still not possible at the time of that research to solve most
classes of high-index DAEs, so a method was proposed to append directly to the
DAE any state variable constraints that did not cause the resulting system to be
high-index. The other state variable constraints were transformed into a set of NLP
equality constraints. To form these NLP equality constraints, a hybrid method was
used which combines both global constraints like (6.7) and point constraints like:

ρi (tj ) = hi (x(tj ), y(tj )) = 0 (6.8)

i = 1 . . . nh j = 1 . . . npt tj ∈ (t0 , tf ]

These point constraints may be evaluated either by sampling the state variables as the
simulation progresses, or by interpolating the state trajectories after the simulation
finishes. The point constraints provide local information along the trajectory to the
optimizer, while the global constraint attempts to prevent the constraint from being
violated at times other than the points where (6.8) were evaluated. Similar methods
were proposed in [63], although the relationship between the path constraints and
high-index DAEs was not noted in that work.

In [76] a similar method was described in which high-index DAEs were solved
via a transformation to a path-constrained dynamic optimization problem. In this
method, a high-index DAE:

¯ ẋ, x, y, t) = 0
f( (6.9)

193
is partitioned into two sets of equations:

f¯(1) (ẋ, x, y, t) = 0 (6.10)

f¯(2) (ẋ, x, y, t) = 0 (6.11)

 
such that the matrix (∂f (1) /ẋ(1) ) (∂f (1) /y (1) ) is nonsingular, and the state variables
 T  T
are partitioned into x = x(1) : x(2) and y = y (1) : y (2) . The solution to the high-
index DAE may be obtained by solving the following dynamic optimization problem:

& (2) &

min &f¯ (ẋ, x, y, t)& (6.12)
x(2) ,y (2) ,t∈[t0 ,tf ]

subject to:

f¯1 (ẋ, x, y, t) = 0 (6.13)

where (·) is an appropriate norm on the domain [t0 , tf ]. The dynamic optimization
method for solving high-index DAEs requires the following steps:

1. Solve the IVP numerically using the given control profiles.

2. Evaluate the integral constraint violation (objective function). If the constraint

violation is acceptably small, then stop.

3. Adjust the control profiles x(2) , y (2) to try to minimize the constraint violation
on the next step. Goto Step 1.

This method is interesting here because it demonstrates the problems with all methods
that solve path-constrained dynamic optimization problems by including the path
constraints in the master NLP.
There are problems with this method include:

• Since the most expensive step is Step 1, this method is much less efficient than
direct methods for solving high-index DAEs (such as the dummy derivative

194
 method described in Chapter 5), which requires the solution of only one (albeit
larger) IVP.

• Any method for choosing a time discretization of the controls will be somewhat
ad hoc compared to the step size selection algorithm built into the numerical
DAE solver that handles the IVP. Time discretization of the controls is ar-
guably not a problem for unconstrained dynamic optimization problems, where
the space of implementable control functions is often limited by physical con-
siderations. However, it is a problem here because state trajectories are being
discretized for which the functional form is not constrained by implementable
control functions.

• There is no guarantee on the accuracy of the control trajectories, since the

discretization chosen may not be sufficient to capture all the features of those
trajectories. Furthermore, there is no measure of the error that would indicate
when a solution is unacceptable.

• Independent of the discretization, the NLP has more equations and decision
variables than necessary. Some of the decision variables in an equality path-
constrained optimization are actually completely determined by the solution to
the path constrained DAE. Therefore, it is more efficient to find the solution of
these variables directly, using the IVP solver, rather than indirectly, using the
NLP solver.

• Any control discretization as complex as the discretization typically chosen by

a BDF integrator would create large-scale dense NLPs, which are difficult to
solve using current NLP solution methods.

In short, the combination of inefficiency and uncontrollable accuracy makes the use of
control parameterization unattractive to solve equality path-constrained DAEs with
all the above methods.
Although the problems described above have not been explicitly recognized by
other authors, there have been some methods proposed for solving high-index dy-

195
namic optimization problems directly. In [66, 67] dynamic optimization of index-2
DAEs was described, and in [109] a method similar to the dummy derivative method
was proposed to derive an equivalent index-1 DAE for a given arbitrary-index DAE.
However, neither implementation nor numerical results were reported in [109], and
the method described requires detection of numerical singularity of matrices, which
is problematic both practically and theoretically in the nonlinear sense. In both of
these works it was assumed that (6.2) was high-index, and neither proposed a method
for directly appending (6.3) to (6.2) to form a high-index system.

196
6.2 Equality Path Constraints and High-Index
DAEs
The method proposed in this chapter for solving equality path-constrained dynamic
optimization problems is to append the path constraint (6.3) directly to the DAE
(6.2), allowing a subset of the control variables to be determined by the solution of
the resulting combined IVP. Then the dummy derivative method is used to derive
an equivalent index-1 DAE for the high-index DAE. For the purposes of this section,
it is assumed that a control matching can be found, and concentrate instead on the
properties of the resulting combined system.
Reference to the following definition which was presented in [21] was made in
Chapter 2:

Definition 6.1 (Differential Index of a DAE). The minimum number of times

that all or part of a DAE must be differentiated with respect to time to determine ẋ(t)
and ẏ(t) as continuous functions of x(t), y(t), u(t), and t is the differential index i
of a DAE.

By this definition, ordinary differential equations (ODEs) have i = 0. The term

high-index DAE is commonly used to refer to systems with i ≥ 2. As mentioned in
Chapter 5, standard numerical integration codes experience difficulty when applied
to systems with i ≥ 2, and only limited classes of high-index problems may be solved
directly at present.

Definition 6.2 (Structurally Singular DAEs). A DAE is structurally singular if

the Jacobian of the DAE with respect to the highest-order time derivatives in the DAE
( e.g., ẋ and y in (6.1)) is structurally singular.

Definition 6.3 (Structural Index of a DAE). The structural index is of a DAE

is the maximum number of times that any subset of equations in the DAE is differ-
entiated using Pantelides’ algorithm.

Theorem 6.1. The index and the structural index of a DAE are related by i ≥ is .

197
Proof. Consider a modification to Pantelides’ algorithm described in Chapter 5, in
which the variables a and equations:

a = ẋ (6.14)

are appended to the underlying DAE as necessary at each step of the algorithm so
that time derivatives of order greater than one do not appear in the underlying DAE
obtained through the next step. Since the new variable a is uniquely determined by
(6.14), the augmented system has the same index as the original DAE.
Define ẑis as the set of highest order time derivatives in the final underlying DAE
obtained with this modified Pantelides’ algorithm, f (is ) . Define Jˆis as the Jacobian
of f (is ) with respect to ẑis .
There are four cases:

i = is : If Jˆis is nonsingular and ẑis does not contain any algebraic variables, then
f (is ) has i = is = 0, and according to the Implicit Function Theorem, all time
derivatives are uniquely determined given the state variables.

i = is + 1 : If Jˆis is nonsingular and ẑis does contain some algebraic variables, then
all time derivatives are not uniquely determined by this final underlying DAE.
Since by assumption:

Jˆis = ∂f (is ) ∂f (is ) (6.15)

∂ ẋ ∂y

is nonsingular, where ẋ and y are respectively the time derivatives and algebraic
variables in ẑis , one further differentiation of f (is ) produces:

∂f (is ) ∂f (is ) ∂f (is ) ∂f (is ) ˙ ∂f (is )

f (is +1) = ȧ + ẋ + ẏ + û + =0 (6.16)
∂ ȧ ∂x ∂y ∂ û ∂t
ȧ = ẍ (6.17)

where û = {u, (du/dt), . . . , (d(is ) u/dt(is ) )}. Because Jˆis is nonsingular, the time
derivatives ȧ and ẏ are uniquely determined by f (is ) .

198
i > is : If Jˆis is singular, then by the properties of Pantelides’ algorithm Jˆis must be
numerically but not structurally singular. Since the general DAE is nonlinear,
no information about whether ẑis is uniquely determined or not is conveyed by
the numerical singularity of Jˆis . It is possible that additional differentiations
must be performed to uniquely determine ẑis , and therefore the differential index
may be larger than the structural index.

i < is : Since Pantelides’ algorithm terminates with the first underlying DAE that
possesses a structurally nonsingular Jˆis , and since structural nonsingularity of
Jˆis is a necessary condition for f (is ) to uniquely define the highest order time
derivatives ẑis , it is not possible for any previous f (is −j) , j = 1, . . . , is to uniquely
determine ẑis −j . Therefore, the structural index is a lower bound on the differ-
ential index.

Definition 6.4 (Structurally Well-Behaved DAEs). A structurally well-behaved

DAE is one for which Jˆis is nonsingular for the entire time horizon of interest.

Note that for structurally well-behaved DAEs, i = is or i = is + 1. Although the

structural index is a global property of DAEs, the differential index is only a local
property. DAEs that are neither structurally high-index nor structurally well-behaved
may or may not be high-index at the points where Jˆis is singular. That is, even though
a nonsingular Jˆis indicates that the DAE is not high-index, a singular Jˆis does not
necessarily indicate that the DAE is high-index.
The method for solving equality path-constrained dynamic optimization problems
described in this chapter applies only if the resulting high-index DAE is structurally
well-behaved. The focus is limited to this class of DAEs because of the practical dif-
ficulties in detecting numerical singularity of matrices, and the theoretical difficulties
of using numerical singularity in the nonlinear or linear time-varying classes of DAEs
to determine the true differential index of the DAE. Furthermore, almost all problems
of practical interest are structurally well-behaved. For the remainder of this chapter,
the DAEs are assumed to be structurally well-behaved.
In general, the strongest statement that can be made is that appending (6.3) to

199
an index-one DAE may result in a high-index DAE. There are cases in which the
index of the augmented DAE remains unchanged. For example, consider the index-1
DAE:

ẋ + y + u = 0 (6.18)

y−u−x= 0 (6.19)

where u is a control variable, and the equality path constraint:

2x + y − 5 = 0 (6.20)

where x and y are state variables and u is a control variable. When u is treated as
an algebraic variable, the augmented DAE (6.18–6.20) also has index = 1.

However, there are some important classes of DAE for which appending equality
path constraints to the DAE in the manner described above will result in high-index
DAEs.

Theorem 6.2. Appending ng ≤ nu state constraints of the form:

g(x) = 0 (6.21)

where there is some χ ⊆ x, χ ∈ Rng such that (∂g/∂χ) is nonsingular, to an explicit

ODE:

ẋ − φ(x, u, t) = 0 (6.22)

yields a DAE with i ≥ 2, assuming the DAE is solvable.

Proof. Appending the path constraints requires ng controls to become algebraic state
variables denoted by y ⊆ u in the augmented DAE. Define χ̄ = x \ χ and ū = u \ y.

200
A suitable partitioning of (6.22) yields:

χ̄˙ − φ1 (χ, χ̄, y, ū, t) = 0 (6.23)

χ̇ − φ2 (χ, χ̄, y, ū, t) = 0 (6.24)

By assumption, χ is uniquely determined by (6.21) so it is not available to determine

ẋ and ẏ. Differentiate (6.23),(6.24), and (6.21) with respect to time to yield:

∂φ1 ∂φ1 ∂φ1 ∂φ1 ∂φ1

¨−
χ̄ χ̇ − χ̄˙ − ẏ − ū˙ − =0 (6.25)
∂χ ∂ χ̄ ∂y ∂ ū ∂t
∂φ2 ∂φ2 ∂φ2 ∂φ2 ∂φ2
χ̈ − χ̇ − χ̄˙ − ẏ − ū˙ − =0 (6.26)
∂χ ∂ χ̄ ∂y ∂ ū ∂t
∂g ∂g
χ̇ + χ̄˙ = 0 (6.27)
∂χ ∂ χ̄

χ̄˙ is uniquely determined by (6.23) and χ̇ by (6.27), but (6.24) is unavailable to

determine ẏ. Differentiating (6.27) with respect to time yields:

 T  T
∂gj T ∂ 2 gj 2
T ∂ gj
2
T ∂ gj ∂gj
χ̈ + χ̇ χ̇ + 2χ̄˙ χ̇ + χ̄˙ χ̄˙ + ¨=0
χ̄ (6.28)
∂χ ∂χ2 ∂ χ̄∂χ ∂ χ̄2 ∂ χ̄

for all j = 1, . . . ng .

Nonsingularity of the matrix:

⎡ ⎤
∂φ1
I 0
⎢ ∂y ⎥
⎢ ⎥
⎢ 0 I ∂φ2
⎥ (6.29)
⎣ ∂y ⎦
∂g ∂g
∂ χ̄ ∂χ
0

¨ , χ̈, and ẏ uniquely. Since (6.21)

is a sufficient condition for (6.25–6.26) to determine χ̄
has been differentiated twice, according to Definition 6.1 the index is at least two.
Since there is no guarantee that (6.29) has full rank, the index may be greater than
two.

201
Theorem 6.3. Appending ng ≤ nu state constraints of the form:

g(x, u) = 0 (6.30)

to an explicit ODE of the form (6.22) yields a DAE with i ≥ 1, assuming the DAE
is solvable.

Proof. Since (6.30) does not contain any time derivatives, the augmented system is
structurally rank-deficient with respect to the time derivatives and hence must be at
least index-1.

A consequence of Theorems 6.2 and 6.3 is that optimal control of ODE systems
subject to state path constraints requires either implicit or explicit treatment of DAEs.
For cases where the DAE is not an explicit ODE, the index may stay the same or
increase when the DAE is augmented with state path constraints. In problems where
the index increases, as the theorems indicate, the new index may indeed rise by more
than one. An example of the index rising by more than one is the index-0 DAE:

ẋ1 = u0 − 5x1 (6.31)

ẋ2 = 5x1 − 3x2 (6.32)

ẋ3 = 3x2 − 4x3 (6.33)

where u0 is a control variable. When this system is augmented with:

x3 = 10 (6.34)

u0 becomes an algebraic variable and the augmented problem is index-4.

The case where (∂g/∂x) is rank-deficient yields a useful insight for diagnosing
poorly posed problems. Rank deficiency of (∂g/∂x) implies that the state path con-
straints are at least locally either redundant or inconsistent. The inconsistent case
can be excluded because inconsistent path constraints cannot be satisfied simulta-
neously. Hence, only the redundant case is possible. Therefore, if (∂g/∂x) becomes

202
rank deficient, a subset of the state path constraints can be ignored at least locally.

203
6.3 Dynamic Optimization Feasibility

An important consideration when specifying path-constrained dynamic optimization

problems is to ensure that the problem is feasible with respect to the path constraints.
In this section, it is assumed that (6.2) is a solvable, structurally well-behaved DAE.
The purpose of this section is to determine conditions that (6.3) must satisfy for the
dynamic optimization problem to be feasible.

When equality path constraints are appended to the DAE, one control variable
must become a state variable in the resulting augmented DAE to prevent it from
being overdetermined. The rationale is simply that the number of state variables in
a properly specified DAE must equal the number of equations in the DAE. One of
the implications is that if the number of control variables is less than the number
of constraints, the problem is either over-constrained or some of the constraints are
redundant. Therefore, the number of control variables must be greater than or equal
to the number of equality path constraints for the dynamic optimization problem to
be solvable.

If the DAE has an equal number of control variables and equality path constraints,
it is still possible that the resulting system is unsolvable, which is to say, there is no
state trajectory for a given set of input trajectories. Assuming that the original
DAE is well-posed and solvable, the augmented system could be unsolvable because
the augmented DAE is uncontrollable, or because the constraints are inconsistent or
redundant. For example, no control trajectory u for the system:

1
ẋ = + u2 (6.35)
x

can match the equality path constraint:

x=0 (6.36)

Likewise, any dynamic optimization problem that includes the equality path con-

204
straints:

x1 + x2 = 5 (6.37)

x1 − x2 = 1 (6.38)

x1 + 5x2 = 100 (6.39)

is unsolvable since the constraints are inconsistent with each other.

Three necessary requirements that must be met for the constrained dynamic op-
timization problem to be feasible are that the equality path constraints must be con-
sistent and nonredundant, and the original DAE must be controllable with respect to
the equality path constraints. A system of equations is consistent if there exists at
least one solution to the system. The requirement that the constraints be nonredun-
dant stems from the fact that one control variable is turned into an algebraic variable
for every equality path constraint, and the presence of redundant equations would
cause the loss of too many degrees of freedom. Controllability is defined according to
[121]:

“Roughly speaking, a dynamic system is ‘controllable’ if its state vector

can be caused, by an appropriate manipulation of system inputs, to behave
in a desirable manner.”

Note that consistency, nonredundancy, and controllability are all local properties of
nonlinear systems, and therefore it is difficult in practice to detect the presence or
absence of these conditions.
However, as in the solution of nonlinear high-index DAEs, some information can be
gained from the structural properties of the augmented system. A necessary condition
for a set of nonredundant constraints to be consistent is that an assignment exists
for every equation to a unique variable in the system. If this condition is not met,
the equations must be either inconsistent or redundant. In the example (6.37–6.39)
no assignment exists for each equation to a nonredundant variable because there are
three equations but only two variables.

205
 A precondition for controllability is input connectability, which exists for a system
“if the system inputs are able to influence all the state variables” [121]. Input con-
nectability is defined in a graph-theoretic sense if a path exists for each state variable
to at least one control variable in the digraph of the system. The digraph of a DAE
was defined in Section 5.2.1.

To see an example of input connectability, consider the system:

ẋ1 = x2 (6.40)

ẋ2 = x2 + u (6.41)

ẋ3 = 4 (6.42)

where u is a control variable, with the path constraints:

x21 = 3 (6.43)

x3 = 9 (6.44)

The input variable u directly influences (6.40), which in turn influences x2 through
(6.41). Therefore, the state variable x2 in the path constraint (6.43) is input con-
nectable with u. However, the variable x3 in the state path constraint (6.44) is not
influenced by the control variable u; therefore x3 is not input connectable and no
feasible solution exists to a dynamic optimization problem that contains (6.40–6.44)
which satisfies (6.44).

Input connectability of all the state variables is not a condition for solvability of
the DAE. For example, the DAE (6.40–6.43) is solvable (and a dynamic optimization
problem would be feasible) even though the state variable x3 is not input connectable.
However, at least one state variable in each of the state variable path constraints must
have the ability to be influenced by an input variable in the unconstrained DAE or
else the constraint cannot be satisfied. Therefore, the entire augmented DAE does
not have to be input connectable, but the state variable constraints do.

Also, there exist input-connectable systems which are still infeasible. For example,

206
consider the system:

ẋ1 = x2 (6.45)

ẋ2 = x2 + u1 (6.46)

ẋ3 = 4 + u2 (6.47)

where u1 and u2 are control variables, with the path constraints:

x21 = 3 (6.48)

x2 = sin(t) (6.49)

Although both (6.48) and (6.49) are input connectable to u1 , the control variable u1
cannot satisfy them both simultaneously. However, if (6.48) is replaced with:

x1 = cos(t) (6.50)

the control u1 could simultaneously satisfy the state path constraints. Even though
the constraints (6.49–6.50) do not appear to be redundant when examined in isola-
tion, they are in fact redundant when coupled with the DAE (6.45–6.47). Therefore,
the nonredundancy consideration requires each state path constraint to be input con-
nectable to a unique control variable in the unconstrained DAE.
The result of the discussion in this section is the following theorem:

Theorem 6.4. Assuming that (6.3) is nonredundant, if an augmented DAE formed

by appending (6.3) to (6.2) and allowing nh control variables to become algebraic
state variables does not satisfy all of the following conditions, the associated path-
constrained dynamic optimization problem (6.1–6.4) is infeasible. The conditions
are:

1. The number of control variables must be greater than or equal to nh .

2. The equations (6.3) must be structurally nonsingular with respect to a nonempty

subset of the state variables in (6.3).

207
 3. There exists at least one transversal of the Jacobian of the constraints (6.3) with
respect to the state variables such that a path exists in the graph of (6.2–6.3)
from every member of the set of nh control variables û ⊆ u that have become
algebraic state variables in (6.2–6.3), to a unique member of the set of state
variables in the transversal.

Proof. All of the following statements are based on the assumption that the con-
straints are nonredundant. If Condition 1 is not true, then the dynamic optimization
problem does not have enough degrees of freedom to satisfy its constraints. If Condi-
tion 2 is not true, then (6.3) must be inconsistent, and therefore no state trajectories
exist that will satisfy all of the constraints simultaneously. If Condition 3 is not true,
then the combined system is either not input connectable, or it is structurally singular
with respect to its state variables.

The assumption that the constraints are nonredundant is difficult to check, but
is a reasonable assumption since it is usually obvious to the person specifying the
problem as to which nonredundant constraints may be imposed. Theorem 6.4 is
useful in practice, since all of the conditions may be easily checked using structural
criteria. However, the theorem is, like the other structural criteria used in this thesis,
a set of necessary conditions, and there are sets of state path constraints for which
this theorem is true which are not feasible because they do not satisfy consistency or
input connectability requirements at local points.

208
6.4 Control Matching

The results presented in this chapter thus far have assumed that it is possible to find
a valid subset of the control variables of the unconstrained problem to make into
algebraic variables in the constrained problem. This section assumes that a feasible
path constrained dynamic optimization problem has been posed as described in the
previous section, and describes how to determine which control variables become
algebraic variables in the path constraint augmented DAE.
The equality path-constrained dynamic optimization problem may have more con-
trols than state variables. In this case, a valid subset of the control variables must
be selected to become algebraic state variables in the augmented DAE. Such a subset
is termed a control variable matching. If Theorem 6.4 is satisfied and there are ex-
actly the same number of controls as state path constraints, there is no uncertainty
about which controls are in the control matching, but it is still necessary to determine
whether the resulting augmented DAE is solvable.
The most general statement that may be made about a control variable matching
is that it must result in a solvable DAE. Solvability criteria for nonlinear high-index
DAEs are discussed in [21, 29], but are of little practical use. Rather, since the
dummy derivative method is dependent upon Pantelides’ algorithm, a related but
more relevant question is to find a control variable matching that will lead to a DAE
for which Pantelides’ algorithm will terminate. If Pantelides’ algorithm terminates,
the DAE is solvable in a structural sense, and the corrector iteration of the BDF
method will be structurally nonsingular.
The following definitions and theorem were presented in [107]:

Definition 6.5 (DAE Extended System). Given the DAE system (6.51), the cor-
responding extended system is:

f (ẋ, x, y, u) = 0 (6.51)

υi (xi , ẋi ) = 0 i = 1 . . . nx (6.52)

209
Definition 6.6 (Structurally inconsistent DAE). The DAE (6.51) is said to be
structurally inconsistent if it can become structurally singular with respect to all oc-
curring variables state variables (x, y) by the addition of the time differentials of a
(possibly empty) subset of its equations.

Theorem 6.5. Pantelides’ algorithm terminates if and only if the extended system
(6.51–6.52) is structurally nonsingular. Furthermore, if (6.51–6.52) is structurally
singular, then the DAE system (6.51) is structurally inconsistent.

A proof of Theorem 6.5 is given in [107]. As an example, consider the system:

f1 (x, y1 , y2 ) = 0 (6.53)

f2 (x, ẋ, u1) = 0 (6.54)

f3 (x, u2 ) = 0 (6.55)

with control variables u1 and u2 , which was given in [107]. The extended system is
formed by appending:

υ(ẋ, x) = 0 (6.56)

to (6.53–6.55). The extended system is structurally singular with respect to ẋ, x,

y1 , and y2 , and therefore Pantelides’ algorithm cannot terminate. This system is not
solvable because (6.53) is not sufficient to define both y1 and y2 , which do not appear
elsewhere in the system. Theorem 6.4 is actually a requirement for the extended sys-
tem to be structurally nonsingular, but Theorem 6.5 is not as useful as Theorem 6.4
when attempting to specify a feasible path-constrained dynamic optimization prob-
lem.
However, Theorem 6.5 can be used to find control matchings that lead to struc-
turally consistent DAEs. Essentially, Duff’s algorithm is used to extend the augment-
ing path of a rectangular structural matrix formed by the addition of nh rows and nu
columns to the structural Jacobian of the unconstrained DAE.
It is assumed that (6.2) is structurally consistent, and that application of the

210
AUGMENTPATH algorithm (see [43, 44, 107] and Chapter 5) to the extended system
of (6.2) has resulted in a vector ASSIGN of length 2nx + ny . The equality path
constraints and the control variables are then appended to the extended system.
The AUGMENTPATH algorithm is then applied to each equality path constraint,
hopefully resulting in a new vector ASSIGN of length 2nx + ny + nh . Any control
variable ui for which ASSIGN(2nx + ny + i) is nonzero is defined to be part of a
control matching. If AUGMENTPATH returns PATHFOUND=FALSE for any of
the equality path constraints, no control matching is possible for this system.

For example, the structural Jacobian with respect to the state variables and time
derivatives for the following extended system:

f1 (ẋ1 , x2 , u1 ) = 0 (6.57)

f2 (ẋ1 , ẋ2 , x1 , u1 , u2) = 0 (6.58)

f3 (x1 , y, u3) = 0 (6.59)

υ1 (ẋ1 , x1 ) = 0 (6.60)

υ2 (ẋ2 , x2 ) = 0 (6.61)

where the x variables are states and the u variables are controls is:

ẋ1 ẋ2 x1 x2 y
f1 × ⊗
f2 ⊗ × ×
(6.62)
f3 × ⊗
υ1 × ⊗
υ2 ⊗ ×

where the symbols ‘×’ and ‘⊗’ denote structural nonzeros, and the set of ‘⊗’ de-
notes an augmenting path found by successive application of the AUGMENTPATH

211
algorithm to each equation. When the following equality path constraints are added:

h1 (x1 , x2 ) = 0 (6.63)

h2 (y) = 0 (6.64)

to form an augmented DAE, the constraints and the u variables are added to the
structural Jacobian, giving:

ẋ1 ẋ2 x1 x2 y u1 u2 u3
f1 × ⊗ ×
f2 × × × × ⊗
f3 × × × ⊗
(6.65)
υ1 ⊗ ×
υ2 ⊗ ×
h1 ⊗ ×
h2 ⊗

when the AUGMENTPATH algorithm is applied to the new equations. The control
matching consists of the variables u2 and u3 , which become algebraic state variables in
the augmented DAE. Note that since the structural singularity of (6.62) is typically
checked when solving a DAE in a modeling environment such as ABACUSS, the
additional computational effort required to obtain the control matching is minimal.

It is possible that the control matching found with this method is nonunique. In
the previous example, the augmenting path for the extended Jacobian could have

212
been:

ẋ1 ẋ2 x1 x2 y u1 u2 u3
f1 × ⊗ ×
f2 × × × ⊗ ×
f3 × × × ⊗
(6.66)
υ1 ⊗ ×
υ2 ⊗ ×
h1 ⊗ ×
h2 ⊗

and therefore the control matching could be the variables u1 and u3 . There is no
structural information that would indicate that one choice is preferred over the other,
and in many cases it may be that all possible control matchings are acceptable. One
consideration may be that the controls that are not in the control matching will be
subject to the control parameterization, while the functional form of those in the
control matching is unrestricted. If, for example, the implementable function space
was limited for some controls but not others, the control matching choice could be
guided by the desire to limit the functional form of some of the controls. Also, since
structural criteria were used to determine the possible control matchings, it may be
that some choices are in fact not permissible because they lead to DAEs which are
numerically unsolvable.

The control matching algorithm presented in this section produces an augmented

DAE that is solvable in the structural sense. Like all of the structural criteria that
have been used in this thesis, the requirement that the extended augmented DAE be
structurally nonsingular is a necessary condition for the DAE to be solvable, but it
is not a sufficient condition. However, it is a sufficient condition if the augmented
system is structurally well-behaved.

213
6.5 Examples

This section gives numerical results for several equality path-constrained dynamic
optimization problems. For other examples of path constrained dynamic optimization
problems in this thesis, see Chapter 7 and Chapter 8. There are very few examples
of equality path-constrained dynamic optimization problems in the literature (there
are examples of inequality path constrained problems, but those are discussed in
the next chapter). One of the few problems for which a numerical solution has been
reported is the high-index pendulum, which was solved in [76] and [142] as an equality
path-constrained dynamic optimization problem. However, the high-index pendulum
was solved in Chapter 5 using a single IVP integration using the dummy derivative
method, and so it has not been included here.

6.5.1 Two-dimensional car problem

In this problem a car which is able to move in two dimensions (x and y) under the
influence of two controls, the acceleration angle θ and magnitude a. The objective
is to find the control profiles that cause the car to move from point A to point B in
minimum time. An equality path constraint is imposed by requiring the car to remain
on a road of a given shape. Mathematically, this dynamic optimization problem is:

214
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−5 10−3 23 26.85 3.88s

Table 6-1: Statistics for the two-dimensional car problem

min tf (6.67)
a(t),θ(t),tf

subject to:

v̇x = ax (6.68)

ẋ = vx (6.69)

v̇y = ay (6.70)

ẏ = vy (6.71)

ax = a cos(θ) (6.72)

ay = a sin(θ) (6.73)

y = x2 (6.74)

x(0) = 0 x(tf ) = 300

vx (0) = 0 vx (tf ) = 0

The equality path constraint (6.74) can be matched to θ. The resulting augmented
DAE is index-3, and the dummy derivative algorithm differentiates (6.74) twice to
obtain an equivalent index-1 DAE. This problem was solved using two constant finite
elements to approximate a. The control was bounded by −500 ≤ a ≤ 500, and the
initial guess was a = 10. Solution statistics are given in Table 6-1 and Figures 6-1 to
6-4 show the solution trajectories. The solution found satisfies the path constraint to
within the integrator tolerances over the entire state trajectory.

215
 Two-Dimensional Car Problem
90000
80000 y(t)

70000
60000
50000
40000
30000
20000
10000
0
0 50 100 150 200 250 300
x(t)

Figure 6-1: State space plot showing the optimal trajectory of the two-dimensional car
problem

Two-Dimensional Car Problem

600
a
400

200

-200

-400

-600
0 5 10 15 20 25 30
t

Figure 6-2: Optimal acceleration trajectory for two-dimensional car problem

216
 Two-Dimensional Car Problem
20
vx

15

10

0 5 10 15 20 25 30
t

Figure 6-3: The optimal velocity in the x direction for the two-dimensional car prob-
lem

Two-Dimensional Car Problem

7000
vy
6000

5000

4000

3000

2000

1000

-1000
0 5 10 15 20 25 30
t

Figure 6-4: The optimal velocity in the y direction for the two-dimensional car problem

217
6.5.2 Brachistochrone

For a description of the brachistochrone problem, see Section 8.1, where it is solved
with an inequality path constraint. There are several possible formulations of the
brachistochrone problem [24, 84], including the following one [14]:

min tf (6.75)
θ(t),F (t),tf

subject to:

ẋ = u (6.76)

ẏ = v (6.77)

u̇ = F sin(θ) (6.78)

v̇ = g − F cos(θ) (6.79)

x(tf ) = 1

where u and v are respectively the horizontal and vertical velocities, θ is the angle at
which the bead is currently heading, and F is the normal contact force. Equations
(6.76–6.79) describe the forces acting on the bead, so a path constraint defining the
shape of the wire must be added:

v
tan(θ) = (6.80)
u

Interestingly, in this problem the index is dependent on the control matching. Either
F or θ can be matched to the constraint (6.80). If F is matched to (6.80), the problem
is index-2, while if θ is matched to (6.80) the problem is index-1. However, in the
latter case ẋ(0) = 0, ẏ(0) = 0 is not a valid initial condition because (6.80) is not
capable of uniquely defining θ at that point. Therefore, the index-2 formulation was

218
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−5 10−3 10 1.772 1.33s

Table 6-2: Statistics for the brachistochrone problem

Brachistochrone Problem
1
y(t)
0.9

0.8

0.7

0.6

0.5

0.4

0.3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x(t)

Figure 6-5: State-space plot for the brachistochrone problem

specified, and the initial condition:

x(0) = 0 (6.81)

y(0) = 1 (6.82)

ẋ(0) = 0 (6.83)

was imposed.
The dummy derivative algorithm calls for differentiating (6.80) once. This dy-
namic optimization problem was solved using a single linear element to approximate
θ. Solution statistics are given in Table 6-2, a state-space plot is shown in Figures 6-5,
and the F and θ trajectories are given in Figures 6-6 and 6-7. The solution agrees
with numerical solutions obtained for other (lower-index) formulations of the brachis-
tochrone problem.

219
 Brachistochrone Problem
0
-0.2 F (t)
-0.4
-0.6
-0.8
-1
-1.2
-1.4
-1.6
-1.8
-2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
t

Figure 6-6: Optimal force trajectory for the brachistochrone problem

Brachistochrone Problem
0.2
0 θ(t)

-0.2
-0.4
-0.6
-0.8
-1
-1.2
-1.4
-1.6
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
t

Figure 6-7: Optimal θ trajectory for the brachistochrone problem

220
6.6 Conclusions
This chapter has demonstrated that equality path-constrained dynamic optimization
problems require the solution of high-index DAEs. Our method of appending the
equality path constraints to the DAE and solving the resulting augmented system
directly is superior to indirect methods for handling equality path constraints which
rely on including constraint violation measures in the NLP. Not only is our method
more efficient, but it eliminates problems that the indirect methods have associated
with the error control of the equality path constraints. Also, it sheds light on the
difficulties that are involved in specifying an feasible equality path-constrained dy-
namic optimization problem. Since the DAE is nonlinear, very few general statements
can be made about whether it is solvable, and, by extension, whether the dynamic
optimization problem is feasible.
The problem of control matching is also very difficult for nonlinear DAEs. How-
ever, the structural control matching algorithm is efficient and consistent with the
structural criteria that we use to detect and solve high-index DAEs. In the next
chapter, we extend these results for inequality path-constrained dynamic optimiza-
tion problems, where the state variables may not be path-constrained at all points on
the state trajectories.

221
222
Chapter 7

Inequality Path-Constrained
Dynamic Optimization

An inequality-constrained subset of the dynamic optimization problem given in Chap-

ter 2 is considered in this chapter:
 tf
min J = ψ (x (tf ) , tf ) + L (x, u, t) dt (7.1)
u(t),tf to

subject to the DAE:

f (ẋ, x, u, t) = 0 (7.2)

g(x) ≤ 0 (7.3)

φ (ẋ (to ) , x (to ) , to ) = 0 (7.4)

In this formulation, x are state variables and u are control variables. The DAE (7.2)
is assumed to have index ≤ 1, although the results of this chapter are easily extended
for equality path-constrained or high-index DAEs for which an equivalent index-1
DAE may be derived using the dummy derivative method.
As in Chapter 6, the focus of this chapter is restricted to direct dynamic op-
timization methods that use control parameterization to approximate the control
variable trajectories. Chapter 6 showed how to include equality path-constraints in

223
the dynamic optimization problem, but did not address how to include inequality
path-constraints, which add an additional layer of complexity to the problem.
Two substantial issues must be addressed to solve problems of the form (7.1–7.4)
using control parameterization. First, the DAE (7.2) can become high-index along
any trajectory segments on which (7.3) is active, and the trajectories of some subset
of u are prescribed during those segments. Therefore, the index of the DAE and the
dynamic and design degrees of freedom can fluctuate between segments where the
constraints are not active and segments where constraints become active. Second,
the sequence of constraint activations and deactivations is unknown a priori and
thus inequality state-constrained dynamic optimization problems can be viewed from
the perspective of hybrid discrete-continuous dynamic optimization problems. This
fact has not been previously recognized, but it is a very significant insight because it
means that there are potentially combinatorial aspects to the solution of inequality
constrained dynamic optimization problems.
The first part of this chapter demonstrates the connection between hybrid discrete-
continuous and inequality path-constrained dynamic optimization problems. Exist-
ing methods for handling these problems using control parameterization are then
reviewed. The later sections of the chapter describe three new methods developed
in the course of this thesis which take advantage of the high-index nature of state
constrained dynamic optimization problems. Of these, the fluctuating index infeasi-
ble path method has proven the most robust method at present, and examples of its
use are provided in Chapter 8. The other two methods, the modified slack variable
method and the fluctuating index feasible path method are not useful for as broad a
class of problems, but are described since they are instructive in illustrating the issues
involved in solving state inequality constrained dynamic optimization problems. Fur-
ther, if the nonsmoothness of the resulting master NLP can be dealt with effectively,
the fluctuating index feasible path method may be useful in the future because it can
handle the sequencing aspect of the problem without introducing integer variables.

224
7.1 Inequality Path Constraints and the Hybrid
Dynamic Optimization Problem

To illustrate the hybrid discrete/continuous formulation for path-constrained dynamic

optimization, consider the set of feasible trajectories between two points in state
space, subject to inequality constraints on the path between these points. Two such
trajectories are illustrated for a two-dimensional state space in Figure 7-1. Each
feasible trajectory connecting the two points is composed of a series of constrained
and unconstrained segments, where a constrained segment tracks one or more active
path constraints.

Trajectories are characterized by a sequence of constraint activations and deacti-

vations at boundaries between constrained and unconstrained segments. A trajectory
may be discontinuous at such points (often referred to as corners in the optimal con-
trol literature [24]). When an unconstrained trajectory intersects one of the path
constraints, feasibility may be enforced by augmenting the DAE (7.2) that describes
the unconstrained dynamics with the set of active constraints:

gj (x) = 0 (7.5)

where gj is the set of all locally active constraints in (7.3). Each equation in gj
takes up one degree of freedom in the augmented DAE, and therefore for every gi a
control is released and permitted to be implicitly determined by the active inequality.
However, an IVP in which the degrees of freedom are fluctuating along the solution
trajectory is problematic from the point of view of analysis and solution.

To avoid problems with fluctuating degrees of freedom, the inequality path-con-

strained problem can be reformulated in a higher dimensional space by relabeling the
controls u in (7.2–7.4) as uc , and introducing a new set of variables us ∈ Rmu . For
example, consider a problem with a single control and a single path constraint. In
unconstrained portions of the trajectory (i.e., when g(x) < 0), the dynamics of the

225
 Inequality
path
constraint
Final Point
Constrained
segment Unconstrained
Feasible segment
trajectories

Inequality
path
Initial Point constraint

Figure 7-1: Feasible trajectories in constrained state space

system are described by:

f (ẋ, x, us , t) = 0 (7.6)

us = uc (7.7)

uc = uc (t) (7.8)

and the constrained portions (i.e., when g(x) = 0) as:

f (ẋ, x, us , t) = 0 (7.9)

g(x) = 0 (7.10)

uc = uc (t) (7.11)

Note that (7.7) has been replaced with (7.10), but the degrees of freedom for the
overall dynamic system remain unchanged. Thus, during unconstrained portions of
the trajectory the control that actually influences the state (us ) is equivalent to the
forcing function for the system (uc ), whereas in constrained portions of the trajectory
us is determined implicitly by the active path constraint (7.10) and is unrelated to uc .
In a control parameterization context, uc would be prescribed by the control param-

226
 g (x) = 0

x
_ = f (x us  t) x
_ = f (x us  t)
s c
u = u g(x) = 0
c c c c
u = u (t) u = u (t)

g(x) < 0
Figure 7-2: Autonomous switching of constrained dynamic simulation

eters over the entire time horizon, but during constrained portions of the trajectory
it would not influence the solution trajectory.
Equations (7.6–7.11) correspond to a hybrid dynamic system that experiences au-
tonomous switching in response to state (or implicit) events [110]. This behavior is
illustrated in Figure 7-2 using the finite automaton representation introduced in [15].
Further, as shown below, the differential index of the constrained problem will in
general differ from that of the unconstrained problem.
Given these preliminaries, it is now evident that a path-constrained dynamic op-
timization with a single control and a single inequality is equivalent to the following
hybrid discrete/continuous dynamic optimization problem:

 tf
min J = ψ(x(tf ), tf ) + L(x, us , t)dt (7.12)
uc (t),tf t0

subject to:

f (ẋ, x, us , t) = 0 (7.13)
⎧ ⎫
⎪
⎨us = uc ⎪
∀t ∈ T : g(x) < 0 ⎬
(7.14)
⎪
⎩g(x) = 0 ⎪
∀t ∈ T : g(x) = 0 ⎭

227
Provided each inequality is matched with a unique control, this problem formulation
can be extended to multiple controls and inequalities.
The extension to multiple controls and inequalities is evident, provided each in-
equality is matched with a unique control. Note that this hybrid dynamic optimiza-
tion problem exhibits the following properties:

• Autonomous switching is defined by implicit events.

• The number and order of implicit events at the optimum is unknown a priori.

• The index of the system can fluctuate at the events.

The inequality path-constrained dynamic optimization is but one form of a hybrid

discrete/continuous dynamic optimization problem. To illustrate the issues involved
in solving such problems, consider the problem, posed by Santos Galán, of a SCUBA
diver who would like to come up to the surface in the minimum safe time.
One of the major hazards of diving is decompression sickness, which is caused when
the nitrogen (N2 ) in the diver’s blood and body tissues (which has been dissolved at
high pressure during the dive) becomes supersaturated and forms bubbles as the
diver ascends from the dive. To avoid decompression sickness, the diver must make
“decompression stops” during the ascent to allow the N2 dissolved in the blood and
body tissues to equilibrate with the ambient pressure before the bubbling occurs. At
the beginning of this century, J. S. Haldane was appointed by the British Admiralty
to develop safer decompression tables for Navy divers, which give the number, depth,
and length of the decompression stops required. Haldane experimented with goats
and proposed a model describing the uptake and elimination of N2 as a first order
system where the time constants for different tissues are represented by ‘half times’.
Overall, this yields the formulation:

min Pt (tf ) (7.15)

u(t),tf ,tb

228
subject to:

ln(2)
Ṗ5 = (0.79P − P5 ) (7.16)
5
q
Ṗt = − P (7.17)
Vt
Ṗ = u (7.18)
⎧
⎨ State A : −1 ≤ u(t) ≤ 3 Switch to B if P ≥ 1.58P
5
(7.19)
⎩ State B : u(t) = 0 Wait for 4 min then switch to A

Pt ≥ P ∀t ∈ [0, tf ] (7.20)

Pt (0) = 200 (7.21)

P (0) = 1 (7.22)

P (tb ) = 6 (7.23)

P (tf ) = 1 (7.24)

0 ≤ tb ≤ tf (7.25)

In this example, only a tissue with a five minute half-time is modeled, yielding equa-
tion (7.16), where P5 is the partial pressure of N2 in the tissue, and P is the sur-
rounding pressure.

The control for this problem is the rate of ascent/descent u, and since pressure
P is equivalent to depth, u is equivalent to the rate of change of P (7.18). Diving
tables establish decompression stops at different depths depending on the depth-
time profile of the dive. These tables follow conceptually the theoretical principles
developed by Haldane combined with more recent developments and data. In this
simplified example, it is assumed that when the partial pressure in the tissue becomes
twice the environmental pressure, the diver must make a decompression stop of four
minutes, which effectively constrains the admissible control profile (7.19). Note that
discontinuities in the control profile occur at points in time determined by the state
trajectory P5 (t) crossing the state trajectory 1.58P (t), and therefore they are not
known a priori but rather are determined implicitly by the solution of the model
equations (this type of implicit discontinuity is commonly called a state event in the

229
simulation literature [110]). Hence, the dynamic optimization has to determine the
number and ordering of these implicit discontinuities along the optimal trajectory. In
addition, the descent and ascent rates are constrained (at 30 m/min and 10 m/min
respectively).
Imagine a scenario in which the diver wishes to collect an item from the ocean
floor at 50 m with minimum total consumption of air. The volume of a tank of
air is Vt = 15 L and it is charged to Pt (0) = 200 bar. Average air consumption
is q = 8 L/min, but the mass consumed varies with the depth (pressure), yielding
(7.17). A parameter tb is introduced to model the time at which the diver reaches
the bottom, and point constraints (7.23–7.24) force the diver to travel to the bottom
and then surface.
In the first approach, two admissible control profiles are examined:

1. Dive as quickly as possible to the bottom, then ascend at 5 m/min, making

decompression stops as necessary (Figures 7-3 and 7-4).

2. Dive as quickly as possible to the bottom, then ascend at 10 m/min, making

decompression stops as necessary (Figures 7-5 and 7-6).

Observe that the different admissible controls yield different sequences of decom-
pression stops. The consequences of these control policies are that in the first case
the diver is forced to make two decompression stops, augmenting the time and air
consumption of the dive, whereas in the second case a more rapid ascent rate only
requires one decompression stop. Further, if only constant ascent rates are admitted,
then if the ascent rate is dropped further, only one stop becomes necessary again.
In general, this need to determine the number and ordering of implicit discontinu-
ities along the optimum trajectory appears to be the most difficult issue with the
optimization of discontinuous dynamic systems.

230
 v = 5m/min
7
P
6 P5
P 5/P
5

0
0 2 4 6 8 10 12 14 16 18 20
t

Figure 7-3: State trajectories for admissible control profile 1

v = 5m/min
200
PT
195

190

185

180

175

170
0 2 4 6 8 10 12 14 16 18 20
t

Figure 7-4: State trajectory for admissible control profile 1

231
 v = 10m/min
7
P
6 P5
P 5/P
5

0
0 2 4 6 8 10 12
t

Figure 7-5: State trajectories for admissible control profile 2

v = 10m/min
200
PT
198

196

194

192

190

188

186

184
0 2 4 6 8 10 12
t

Figure 7-6: State trajectory for admissible control profile 2

232
7.2 Review of Methods for Handling Inequality
Path Constraints in Control Parameterization

Several methods have been proposed for handling state variable path constraints
within the control parameterization framework. Vassiliadis [142] also reviews such
methods. There are four classes of methods: slack variables, random search, penalty
functions, and interior-point constraints.

7.2.1 Slack variable approach

Although the slack variable approach [74, 75, 140] (often known as Valentine’s method)
was not developed for use with control parameterization, its extension to this frame-
work is straightforward. The terminology and approach detailed in [74] is somewhat
ad hoc because the paper predates all of the work on DAEs. However, this work can
be put in the context of modern DAE theory.

The slack variable approach described in [74] is valid when the DAE (7.2) is an
ODE:

ẋ = f (x, u, t) (7.26)

The method proceeds by appending the state variable inequality (7.3) to (7.26) using
a slack variable a(t):

1
g(x) − a2 = 0 (7.27)
2

where it is assumed here that g(·) → R. Squaring the slack variable has the effect of
enforcing the inequality constraint for all admissible trajectories of a.

233
 Equation (7.27) is then differentiated:

g (1) (x) − aa(1) = 0 (7.28)

g (2) (x) − (a(1) )2 − aa(2) = 0 (7.29)

..
.
1 dk 2
g (k) (x) − a (t) = 0 (7.30)
2 dtk

di g di a
where g (i) = dti
and a(i) = dti
. The differentiation is carried out p times until the
system formed by (7.26–7.30) may be solved for some ui ∈ u. The control ui is then
eliminated from the problem, a(k) becomes a control variable, and the time derivatives
a . . . a(k−1) become differential state variables. Note that the a(k−1) are essentially
being enforced as dummy derivatives, and in fact this method is very similar to the
elimination index-reduction method of [22].
There are several significant problems with this method:

• As noted by [74], the dynamic optimization problem contains a singular arc

whenever a(t)=0. In control parameterization this singular arc has the effect of
providing no gradient information to the optimizer whenever a(t)=0, and thus
significantly slowing down or preventing convergence of gradient-based NLP
methods.

• Also observed in [74] is that there is no obvious extension of this method for
problems that contain more state path inequality constraints than control vari-
ables.

• A fundamental characteristic of this method is that stationarity of the slack

variable underlying index-1 problem is only a necessary condition for stationar-
ity of the original problem. Thus it is possible that a stationary solution of the
slack variable problem is not a stationary solution of the original problem.

• It can be difficult to solve (7.26–7.30) to eliminate ui, particularly if (7.26) is

large and nonlinear.

234
7.2.2 Random search technique

A random search dynamic optimization solution technique was proposed in [7, 8, 32].
Although this method uses control parameterization, it does not make use of gradient
information in the NLP optimizer. Instead, normal probability distributions are used
to generate different decision parameter vectors, which are then used to solve the
DAE and evaluate the objective function.

Random search techniques handle state variable path constraints by discarding

any iteration for which the path constraint is violated. The obvious disadvantage of
this method is the large cost associated with solving numerous infeasible subproblems.
However, the method has been used with some success on extremely nonconvex and
nonsmooth dynamic optimization problems for which gradient-based optimization
techniques experience difficulties.

7.2.3 Penalty functions

Another method that has been used to enforce inequality path constraints is the use
of a penalty function (for example [24, 146]). One way to use penalty functions is to
augment the objective function (7.1) as:
 tf
J˜ = J + K r T (t)r(t)dt r ∈ Rn g (7.31)
t0

or as:

ng  tf
Jˆ = J + Ki ri (x(t))dt (7.32)
i=1 t0

where K ∈ R+ is a large positive number and r(t) ∈ Rng is a measure of the constraint
violation, e.g.:

ri (x(t)) = max (gi(x(t)), 0) i = 1 . . . ng (7.33)

235
This approach can cause numerical difficulties because it requires K → ∞ to sat-
isfy the constraint exactly. In addition, many authors that have used operators such
as max in (7.33) have not recognized that such operators introduce implicit discon-
tinuities that require special treatment during integration [110] and calculation of
sensitivities [55].

7.2.4 End-point constraints

An inequality path constraint can be transformed into an end-point constraint Ψ(t) ∈

Rng (for example [128]):
 tf
Ψi (x(t)) = ri (x(t))dt = 0 i = 1 . . . ng (7.34)
t0

This approach also causes numerical difficulties because the gradients of the end-point
constraints are zero at the optimum, which reduces the rate of convergence near the
solution.
Penalty function methods and end-point constraint methods suffer from problems
due to the selection of a suitable r(t). In [133] computational experience was reported
which showed that using r(t) of the form (7.33) rarely converges for nontrivial dy-
namic optimization problems. The primary reason for these difficulties is that (7.33)
is nondifferentiable when gi (t) = 0. Such constraints also introduce implicit disconti-
nuities in the simulation subproblem, which must also be handled carefully to avoid
numerical difficulties [110]. In [133], a different r(t) was proposed of the form:
⎧
⎪
⎪
⎪
⎪gi (t) if gi (t) >
⎪
⎨
ri (t) = (−gi (t) − )2 /4 if − ≤ gi (t) ≤ i = 1 . . . ng (7.35)
⎪
⎪
⎪
⎪
⎪
⎩0 if gi (t) ≤ −

This smooth function is differentiable when ri (t) = 0 and contains no implicit discon-
tinuities. However, (7.35) does reduce the feasible state space region since the true
inequality path constraints can only be active in the limiting case when = 0.

236
7.2.5 Interior-point constraints

Another method proposed to deal with inequality path constraints is to discretize

them via interior-point constraints (for example [142, 143]):

ρ(g(ti )) ≤ i i = 0 . . . npc (7.36)

where ρ(·) : Rng → R is a function that provides a measure of the violation of the
constraints at a given ti , the parameter i ∈ R+ is a small positive number, and npc
is the number of point constraints.

The problem with this method is that npc → ∞ is necessary to guarantee that
the state path constraint is not violated during any portion of the optimal trajectory.
Since each point constraint adds a dense row to the Jacobian of the constraints of
the NLP, large numbers of point constraints can create NLPs that cannot be solved
using currently available numerical algorithms. If only a few point constraints are
used (one per finite element is typical) there is no guarantee that state trajectories
will satisfy the state constraint to within the tolerances that were used to obtain a
numerical solution of the IVP.

7.2.6 Hybrid interior-point constraint and penalty function

approach

A hybrid approach was used in [33, 142] that combines both the penalty function
and interior point constraint methods. This approach transforms the state variable
inequality path constraint into an end point inequality constraint:

χi (tf ) ≤ i (7.37)

237
where i is a small positive number, which is evaluated by appending the following
equation to: (7.2)

χ̇i = [max(0, gi(x))]2 (7.38)

χ(0) = 0 (7.39)

and a set of interior point inequality constraints:

gi (x(tk )) ≤ 0 k = 1 . . . NF E (7.40)

where tk are the boundaries of the control finite elements. Although the constraint
(7.37) is in principle sufficient to ensure that the original inequality constraint is not
violated, it provides no information to the NLP solver when gi (x) = 0. The point
constraints (7.40) are included to provide some information to the optimizer when
gi (x) = 0.
The disadvantages of this hybrid method are:

• The state variable path constraints are satisfied to within known tolerances
only at the points during the state trajectory where NLP point constraints
have been imposed. The state variable path constraints may not be satisfied to
within acceptable tolerances at other points along the trajectory.

• The value of the control that causes the constraint to be active is implicitly de-
fined, and therefore the NLP solver must iterate to find the control whenever an
inequality path constraint is active. Such iteration decreases the computational
efficiency of the method. This method can be viewed in some sense as solving
a high-index DAE by iterating on an index-1 DAE until all of the equations of
the high-index DAE are satisfied.

• The use of the max operator in (7.38), coupled with the discrete nature of (7.40)
can cause the method to converge slowly if the constraint violation is occurring
at points along the solution trajectory other than at the end of the control finite
elements.

238
7.3 A Modified Slack Variable Approach

This section describes a modification to the slack variable approach that is a significant
improvement over the original slack variable method. As mentioned in the previous
section, there are several difficulties with the original slack variable approach that
have kept it from being widely used. An example of these difficulties can be seen by
considering the following problem which was discussed in [66, 74, 90, 142]:
 1  2 
min V = x1 + x22 + 0.005u2 dt (7.41)
u(t) 0

subject to:

ẋ1 = x2 (7.42)

ẋ2 = −x2 + u (7.43)

x2 − 8(t − 0.5)2 + 0.5 ≤ 0 (7.44)

x1 (0) = 0 x2 (0) = −1

t ∈ [0, 1]

A slightly different form was used in [66, 142] where (7.43) was replaced by ẋ2 = x2 +u.

When the method of [74] is applied to this problem (7.41–7.44), (7.44) is replaced
by:

1
x2 − 8(t − 0.5)2 + 0.5 + a2 = 0 (7.45)
2

which is then differentiated, producing:

ẋ2 − 8(t − 0.5)2 + 0.5 + aa1 = 0 (7.46)

where a1 =(da/dt). Equation (7.46) is used to eliminate u, producing the following

239
unconstrained problem:
 1  
min V = x21 + x22 + 0.005(x2 + 8(t − 0.5) − aa1 )2 dt (7.47)
a1 (t) 0

subject to:

ẋ1 = x2 (7.48)

ẋ2 = −aa1 + 8(t − 0.5) (7.49)

ȧ = a1 (7.50)
√
x1 (0) = 0 x2 (0) = −1 a(0) = 5

t ∈ [0, 1]

The results of a numerical solution of this problem using the ABACUSS control
parameterization algorithm are shown in the first row of Table 7-1.

Table 7-1: Dynamic optimization results for solution of equations (7.41–7.44)

Solution Objective Major Total Opt. IVP # Finite Lagrange
Method Function Iter. IVP Tol. Tol. Elements Approx.
Slack Variable 0.171042 89 100 10−5 10−7 10 linear
Mod. Slack Var. 0.170586 40 46 10−5 10−7 10 linear
Mod. Slack Var. 0.170243 31 38 10−5 10−7 10 quadratic
Mod. Slack Var. 0.169875 74 81 10−7 10−9 10 quadratic

Although a solution to the dynamic optimization was found, convergence of the

optimizer was very slow due to the presence of the singular arc. In addition, the alge-
braic manipulation required to find (7.48–7.50) is not general enough to be applicable
to all problems, and it is very difficult to do automatically for large problems.
The original motivation for performing these algebraic manipulations was to elim-
inate u from (7.46) so that the resulting problem (7.48–7.50) is an ODE. Since it is
now possible to solve high-index DAEs directly using the dummy derivative method
described in Chapter 5, this limitation no longer applies, and the algebraic manipu-
lations can be avoided by allowing the DAE integrator to solve for u. Thus, using the

240
dummy derivative method on (7.41–7.44) gives:
 1  2 
min V = x1 + x22 + 0.005u2 dt (7.51)
a1 (t) 0

subject to:

ẋ1 = x2 (7.52)

ẋ2 = −x2 + u (7.53)

1
x2 − 8(t − 0.5)2 + 0.5 + a2 = 0 (7.54)
2
ẋ2 − 16(t − 0.5) + aa1 = 0 (7.55)

x1 (0) = 0 x2 (0) = −1

t ∈ [0, 1]

However, H11 matrix for this problem that is used in the dummy derivative method
is:

H11 = [a] (7.56)

and therefore no equivalent index-1 DAE can be selected when a = 0. As described in

Chapter 5, the BDF corrector matrix for (7.52–7.55) becomes highly ill-conditioned
when a = 0, and the numerical solver is incapable of integrating past this point.

A modified slack variable method is proposed here to handle both the problems
with algebraic manipulations and problems with singular arcs. The modification is
to make a the control variable, and to apply the method of dummy derivatives to the
resulting high-index DAE.

Thus, the state inequality-constrained problem given by (7.1), (7.2), and (7.3)
may be transformed into a state equality-constrained problem using a slack variable

241
a:
 tf
min J = ψ (x (tf ) , tf ) + L (x, u, t) dt (7.57)
a(t),tf to

subject to the DAE:

f (ẋ, x, u, t) = 0 (7.58)
1
g(x) = − a2 (7.59)
2
φ (ẋ (to ) , x (to ) , to ) = 0 (7.60)

provided that the dimension of g does not exceed that of u, and that the resulting
(possibly high-index) DAE is solvable using the dummy derivative method.

Using this modified slack variable method on the problem (7.41–7.44) gives:
 1  2 
min V = x1 + x22 + 0.005u2 dt (7.61)
a(t) 0

subject to:

ẋ1 = x2 (7.62)

ẋ2 = x2 + u (7.63)
1
x2 − 8(t − 0.5)2 + 0.5 = − a2 (7.64)
2
√
x1 (0) = 0 a(0) = 5

t ∈ [0, 1]

Then, the dummy derivative method can be applied to (7.62–7.64) to obtain an

242
equivalent index-1 DAE:

ẋ1 = x2 (7.65)

x̄2 = x2 + u (7.66)
1
x2 − 8(t − 0.5)2 + 0.5 = − a2 (7.67)
2
x̄2 = 16(t − 0.5) = −aā (7.68)
d
ā = a(t) (7.69)
dt

where x̄ and ā are (dummy derivative) algebraic state variables.

This problem does not have a singular arc when a = 0, and the corrector matrix for
the equivalent index-1 formulation (7.65–7.69) does not become ill-conditioned when
a = 0. Solution statistics for this problem are also shown in Table 7-1. They show
that the modified slack variable approach required significantly fewer iterations to
find the optimum, and actually found a slightly better value for the optimal objective
function.

Figure 7-7 shows the optimal trajectory for the control variable u. The u trajectory
contains discontinuities at the control finite element boundaries, which is due to the
choice of control parameterization of a. Since a controls an equation that involves
only a state variable, u is related to the derivative of a. Since a has been control
parameterized with linear finite elements with C0 continuity, a variable such as u that
is algebraically dependent upon the derivative of a can not be C0 continuous at the
finite element boundaries. The discontinuities in u can be eliminated if higher-order
polynomials are used in the control parameterization of a. Figure 7-8 corresponds to
the last row of Table 7-1, and shows the trajectory of u when quadratic finite elements
are used to approximate a.

The ability to solve high-index DAEs has allowed the creation of a modified slack
variable method with performance characteristics that are superior to the original
method. The modified version is interesting because it clearly shows the link between
state inequality constrained dynamic optimization problems and high-index DAEs.

243
However, there are a number of problems with this method that keep it from becoming
generally useful. They are:

• It does not appear to be possible to extend the method for problems that
contain more inequalities than control variables. In general, it is not possible
to have more independent active state variable constraints at any point in time
than there are control variables. However, there exist many problems that have
more state inequality constraints than control variables, but that never have
more active constraints than they have control variables, which could not be
solved using this method.

• The problem becomes high-index for its entire trajectory, even if a state inequal-
ity constraint never becomes active. This phenomenon requires us to solve a
equivalent index-1 DAE that may be much larger than the DAE of the problem
without constraints, thus decreasing the efficiency of the overall optimization.

• The choice of control parameterization can be very tricky. Since a is being

parameterized, this essentially forms an approximation for a state variable tra-
jectory. There may be no information a priori that allows a suitable choice for
this approximation. Also, there is no way to simultaneously restrict the trajec-
tory of u, either to impose path constraints or to restrict the allowable function
space.

244
 Modified Slack Variable Method: Linear a approximation
14
12 u(t)

10
8
6
4
2
0
-2
-4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t

Figure 7-7: Control variable in modified slack variable method example with ‘a’ ap-
proximated by linear finite elements

Modified Slack Variable Method: Quadratic a parameterization

16
14 u(t)
12
10
8
6
4
2
0
-2
-4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t

Figure 7-8: Control variable in modified slack variable method example with ‘a’ ap-
proximated by quadratic finite elements

245
7.4 Fluctuating-Index Feasible-Path Method for
Dynamic Optimization with Inequality Path
Constraints

A novel method for solving inequality path-constrained dynamic optimization prob-

lems within the control parameterization framework is described in this section. The
fluctuating-index feasible path (FIFP) method enforces state variable path constraints
through direct solution of high-index DAEs during the state-constrained portions of
the solution trajectory. Thus, this method is a feasible path method from the stand-
point of the inequality path constraints, which are never violated during intermediate
iterations of the NLP solver.

The inequality path constraints are not violated during intermediate iterations
because the FIFP method uses an implicit event constrained dynamic simulation al-
gorithm. That is, the hybrid discrete/continuous problem given by (7.13–7.14) is
solved directly. The activation and deactivation of the inequality path constraints
are implicit state events that are detected during the integration. Inequality path
constraint activation and deactivation events are termed constraint events. Implicit
event constrained dynamic simulation requires detecting the constraint events, deter-
mining which controls cease to be degrees of freedom in the constrained portions of
the trajectory, and integrating numerically the resulting (possibly) fluctuating-index
DAE.

246
 The implicit event constrained dynamic simulation problem is:

f (ẋ, x, us , t) = 0 (7.70)

uc = uc (t) (7.71)
ng

αi (usi − uci ) + βij gj (x) = 0 ∀i = 1 . . . nu , ∀t ∈ T (7.72)

j=1
ng

αi + βij = 1 ∀i = 1 . . . nu , ∀t ∈ T (7.73)
j=1
nu
βij ≤ 1 ∀j = 1 . . . ng , ∀t ∈ T (7.74)
i=1

αi = αi (t) ∈ {0, 1} βij = βij (t) ∈ {0, 1}

This formulation shows how the equations in the active DAE change upon activation
and deactivation of path constraints. The variables α and β are determined by the
solution of the implicit event constrained dynamic simulation problem. When a con-
straint gj becomes active, some βij = 1 and αi = 0. The βij and αi is determined
by matching controls ui to constraints gj , which is obtained using the method de-
scribed in Chapter 6. It is assumed that all path constraints are inactive at the initial
condition, and therefore that α(t0 ) = 1 and β(t0 ) = 0.
Assuming that a standard BDF integration method is used to solve the DAE, the
algorithm performed at each step of the integration is:

1. Determine whether any of the inequality constraints g(x) activated or deacti-

vated in the last time step. If no, go to Step 5.

2. Detect the earliest constraint activation or deactivation event that occurred

during the last time step:

(a) For every active gj (x), set one βij = 1 and αi = 0. Set all other β\{βij } = 0
and α \ {αi } = 1.

(b) Reset the simulation clock to the time of the event.

247
 3. Apply the method of dummy derivatives to derive an equivalent index-1 formu-
lation for the current problem.

4. Consistently initialize the equivalent index-1 formulation.

5. Take one integration step with the current index-1 formulation of the problem,
performing dummy pivoting as necessary. If not at the end of the simulation,
go to Step 1.

From a practical standpoint, there are several aspects of the above algorithm that need
to be described in detail. Among these are detecting of activation and deactivation of
constraints and transferring initial conditions between constrained and unconstrained
portions of the trajectory. These details are discussed in the following subsections.

7.4.1 Constraint activation

An inequality constraint activates at the earliest point in time at which the previously
inactive constraint becomes satisfied with equality. These points in time are often
termed state events, because they are defined implicitly by the evolution of the system
state and are not known a priori. The state events marking the constraint activation
may be identified algorithmically by introducing a new algebraic variable zj for each
currently inactive constraint:

zj = gj (x) (7.75)

These discontinuity functions are appended to the current equivalent index-1 for-
mulation of the problem, and constraint activations are detected by advancing the
simulation speculatively until a zero crossing occurs in zj (t). This situation corre-
sponds to a special case of the state events handled by the algorithm described in
[110], which handles equations of the form (7.75) efficiently, and guarantees that all
state events are processed in strict time order.

248
7.4.2 Constraint deactivation

Detecting deactivation of inequality constraints during a simulation is somewhat more

complicated than detecting activation because any uci which has a corresponding
αi = 0 does not influence the current state variable trajectories. During solution of
the equivalent index-1 system using the method of dummy derivatives, the corrector
iteration assures that any active constraints remain satisfied within the relevant nu-
merical tolerance, and therefore an additional condition must be considered to detect
deactivation of the inequality.

An inequality state constraint is considered to have deactivated when switching

some αi from 0 to 1 would cause the state trajectories to move into the feasible region.
This situation is detected by introducing a new set of equalities:

ng

αi (usi − uci ) + βij (gj (x) − j ) = 0 i = 1, . . . nu (7.76)

j=1

(
Equation (7.76) contains a set of parameters { j } such that when j βij = 1 and
j = 0, the corresponding constraint gj is active; and if j ≤ 0, the corresponding
constraint has been “tightened” and the feasible state space decreased. As shown
below, (7.76) is not actually solved; rather it allows the calculation of the sensitivity
of the system to { j }, which is used to detect inequality deactivation.

Associated with each new equality (7.76) is the sensitivity system:

∂f ∂ ẋ ∂f ∂x ∂f ∂us
+ + s = 0 ∀j (7.77)
∂ ẋ ∂ j ∂x ∂ j ∂u ∂ j
ng  T ng
∂usi ∂gk ∂x
αi + βik = βik δjk ∀i, j (7.78)
∂ j k=1
∂x ∂ j
k=1

where δjk is the Krönecker delta function. Note that, although a large number of
sensitivity equations are defined by (7.77–7.78), all sensitivities with respect to j are
identically zero if gj (x) is inactive. Provided that all βij = 0 at t0 , all sensitivities with
respect to may be initialized to zero. Transfer conditions for these sensitivities when
constraints become active are the same as the general sensitivity transfer conditions

249
given in Chapter 3.

Deactivation of the inequality is detected by solving this sensitivity system to

find (∂usi /∂ j ) and detecting the simulation times that are implicitly defined by zero
crossings of the discontinuity functions:

qia = uci − usi (7.79)

∂usi
qijb = (7.80)
∂ j
∀(i, j) : βij = 1

These algebraic equations are appended to the DAE and solved as detailed in [110].

Constraint deactivation is detected using the following theorem:

Theorem 7.1. If at any state event time defined by a zero crossing of (7.80), the
condition:
    s  
∂usi ∂ui
< 0 ∧ (ui > ui ) ∨
c s
> 0 ∧ (ui < ui )
c s
(7.81)
∂ j ∂ j

is true, the constraint gj (x) is considered to have deactivated.

Proof. A negative value of j corresponds to tightening the constraint, and thus mov-
ing the solution trajectory into the feasible region. If the sensitivity (∂usi /∂ j ) is
negative, then uci must increase for the solution trajectory to move away from the
inequality and into the feasible region. Therefore, if uci becomes greater than usi at
some point in the time domain, the specified control will inactivate the inequality.
The reasoning for the case in which (∂usi /∂ j ) > 0 follows the same logic.

7.4.3 Transfer conditions for state variables

The constraint events are transition conditions as defined in Chapter 3. Transition

functions are also needed that relate the state variables before the constraint event

250
to the state variables after the constraint event. The transition function is:

T (ẋ(−) , x(−) , u(−) , ẋ(+) , x(+) , u(+) , p, t) = 0 (7.82)

where the “−” denotes the values before the constraint activation, and the “+” de-
notes the values after activation. In some cases, (7.82) could describe a loss function
upon activation of the constraint, i.e., frictional losses due to an object interacting
with a surface. However, the usual method for determining this transition function is
to specify some subset of the state variables x and y which remain continuous across
the constraint event. The dimension of the vector function T is equal to the dynamic
degrees of freedom of the DAE that is active after the constraint event.

In discrete/continuous dynamic simulation it is often assumed that all differential

state variables x are continuous across a discontinuity in the input functions u [15].
This assumption has been proven to be valid for ODEs and for most physically mean-
ingful index-1 DAEs (in fact, all DAEs with is = 0)[22], but several authors [22, 64, 93]
have noted that the assumption is not true for some index-1 and higher-index DAEs.

Continuity conditions for arbitrary-index DAEs are stated in [64], assuming that
the index and the vector field is the same on both sides of the discontinuity. The
condition is that state variables are continuous across a discontinuity in the controls
provided that the corresponding underlying ODE does not depend on any derivatives
of the controls that are causing the discontinuity. A successive linear programming
(SLP) algorithm was proposed in [64] to find the dependencies of the variable deriva-
tives on the derivatives of the controls that are causing the discontinuity.

The continuity conditions given in [64] are not directly applicable to the FIFP
method because the FIFP method constraint events cause the DAE after the con-
straint event to have a different vector field and possibly a different index. It is
possible, however, to define transition functions for the FIFP method by noting that
there must be a corresponding explicit input-jump simulation for any valid implicit
event simulation. That is, there must be some input function ū(t) for which the

251
simulation:

f (ẋ, x, us , t) = 0 (7.83)

us = ū(t) (7.84)

produces the same state variable trajectories as the implicit-event simulation (7.70–
7.74). Note that ū(t) is permitted to contain discontinuities. The existence of a
corresponding explicit input-jump simulation is guaranteed because a valid implicit
event simulation must define a unique trajectory us which can then be used to set
ū(t) in the explicit input-jump simulation. The corresponding explicit input-jump
simulation is not necessarily unique since there could be more than one ū(t) which
produce the same state trajectories.

The transition conditions in [64] are valid for the corresponding explicit input-
jump simulation. Since the implicit event and explicit input-jump simulations produce
the same state trajectories, the transition conditions for the state variables must also
be the same for the two DAEs. In practice, defining the transition conditions for the
implicit event simulation is even easier than the SLP algorithm in [64] would imply
when the explicit input-jump simulation has index ≤ 1, since then it is possible to
assume continuity of the variables that are differential state variables in the explicit
input-jump simulation.

Since the index of the DAE in the implicit event simulation may be greater than
the index of the DAE in the explicit input-jump simulation, there may be more con-
tinuity assumptions for state variables than there are dynamic degrees of freedom for
the implicit event simulation. In general, the number of degrees of freedom available
will be less than or equal to the number of continuity conditions required for (7.2).
For example, if (7.2) has is = 0, then the number of conditions on the continuity
of x must be equal to the dimensionality of x, whereas the number of differential
state variables in the augmented equivalent index-1 system derived by the method of
dummy derivatives will be equal to or less than the dimensionality of x. Hence, in
general only that subset of the continuity conditions corresponding to the differential

252
state variables remaining in the equivalent index-1 DAE describing the new segment
can be satisfied by the implicit event constrained dynamic simulation algorithm. The
additional continuity assumptions may or may not be automatically enforced by the
DAE itself.

To see an example of continuity assumptions for the FIFP method, consider the
DAE:

ẋ1 = x1 + x2 (7.85)

ẋ2 = u + sin(t) (7.86)

u = u(t) (7.87)

x1 + x2 ≤ 10 (7.88)

The continuity assumptions for (7.85–7.87) upon a jump in u(t) are that x1 and x2 are
continuous. However, an equivalent index-1 system for the implicit event simulation
after activation of (7.88) is:

x̄1 = x1 + x2 (7.89)

ẋ2 = u + sin(t) (7.90)

x1 + x2 = 10 (7.91)

x̄1 + ẋ2 = 0 (7.92)

which has only one dynamic degree of freedom. However, either of the transition
(+) (−) (+) (−)
functions x1 = x1 or x2 = x2 may be chosen, since whichever continuity
assumption is not explicitly defined will in this case be implicitly enforced by (7.91).

Trajectories generated by the implicit event constrained dynamic simulation algo-

rithm will potentially not satisfy all the continuity conditions required for an optimal
trajectory. However, this complication can be handled by allowing the continuity
conditions to be violated by the IVP subproblems at intermediate iterations, and
then adding the continuity conditions at constraint activation/deactivation through
the use of point constraints in the master NLP. Thus, the master NLP will adjust the

253
 u(t)

Figure 7-9: Tank and valve

control profiles in a manner such that at the optimum, the extra continuity condi-
tions that cannot be enforced by the implicit event constrained dynamic simulation
algorithm are satisfied by the NLP. Thus, while the method is always feasible with
respect to path constraints, it is not necessarily feasible with respect to continuity
conditions at intermediate iterations.

As a simple physical example, consider a constrained dynamic optimization in

which the liquid level in the tank shown in Figure 7-9 may not exceed a given thresh-
old. The valve stem position controlling the flow into the tank is a differential state
in the unconstrained DAE and exhibits a first order response to the control signal.

ḣ = (Fin − Fout )/A (7.93)

Fout = f (h) (7.94)

Ḟin = g(u − u0 ) (7.95)

where Fin and Fout are the flowrates in and out of the tank, A is the area of the
tank, h is the liquid height in the tank, and f and g are functions that describe,
respectively, the hydraulics of the flow out of the tank and the flow response to the
control signal.

Since the liquid height is constrained to be lower than hmax , there is an inequality

254
constraint on h:

h ≤ hmax (7.96)

When this inequality is active the control u becomes an algebraic state variable in
the resulting high-index augmented DAE. Applying the dummy derivative method to
the augmented DAE results in the following equivalent index-1 DAE:

h̄ = (Fin − Fout )/A (7.97)

h̃ = (F̄in − F̄out )/A (7.98)

Fout = f (h) (7.99)

F̄out = f  (h) (7.100)

F̄in = g(u − u0 ) (7.101)

h = hmax (7.102)

h̄ = 0 (7.103)

h̃ = 0 (7.104)

where h̄, h̃, F̄in , and F̄out are dummy algebraic variables (h̃ eliminates ḧ).
Equations (7.97) and (7.103) indicate that the flow into the tank must equal the
flow out to maintain the level at its threshold, and hence the stem position u required
to achieve this is also prescribed algebraically. Since for an arbitrary control signal
profile there is no guarantee that the stem position immediately before the constraint
activation will equal that prescribed at the beginning of the constrained segment
of the trajectory, it must be allowed to jump, which is clearly non-physical and
forbidden by the continuity conditions for the unconstrained DAE. In fact, in this case
the equivalent index-1 DAE has no dynamic degrees of freedom to define the initial
condition, so both the stem position and the level could potentially jump. However,
continuity of the liquid level will always be redundant with the liquid level inequality,
so only the stem position can actually jump. Although the constrained dynamic
simulation subproblems allow jumps, the point constraints in the master NLP will

255
force the optimal control profile to be one in which the stem position does not jump.
If the unconstrained DAE (7.2) is already high index (e.g., if it includes equality path
constraints), then the continuity conditions stated in [64] must be enforced by point
constraints in the master NLP.

7.4.4 Sensitivity transfer conditions at inequality deactiva-

tion

The points along the solution trajectory where inequality deactivations occur are
defined by implicit events. Therefore, the sensitivity variable transfer conditions are
defined by the relations given in Chapter 3. However, the deactivation event (7.81)
requires special attention because it contains a sensitivity variable.
For the deactivation condition (7.81) to be true, one of (7.79–7.80) must hold at
the constraint activation time t∗ . The transfer conditions for the sensitivity functions
follow the derivation given in Chapter 3. Obtaining the sensitivity transfer functions
requires differentiating the discontinuity function, which if (7.79) is currently true
gives:

∂uci − ∂t∗ ∂usi − ∂t∗

+ u̇ci − = + u̇si − (7.105)
∂p ∂p ∂p ∂p

for all αi = 0.
If the discontinuity function is instead (7.80), the sensitivity transfer function is:

∂ 2 usi − ∂ u̇si − ∂t∗

+ =0 (7.106)
∂ j ∂p ∂ j ∂p

This condition has the added complication that it requires second-order sensitivity
information, namely (∂ 2 usi /∂ j ∂p). Obtaining this second order sensitivity informa-
tion requires the solution of the second-order analog to (7.77–7.78). It is possible
to extend the algorithm described in Chapter 4 to handle these second order sen-
sitivities. Although the second order sensitivities are more costly to calculate than
first-order sensitivities, only a few are required and hence the additional cost should

256
not be significant.
There are two possible behaviors for the control variable upon deactivation of an
inequality. The state and parameterized trajectories of the control may cross, or else
the special sensitivity may cross zero, making the control jump. The latter case is
more costly to handle in a numerical algorithm because of the need for second order
sensitivities.

7.4.5 Example

An example demonstrating the use of the fluctuating-index feasible-path method for

dynamic optimization is the constrained brachistochrone problem which was presented
both in [24] and in a slightly different form in [84]. The problem is to find the shape
of a wire on which a bead will slide between two points under the force of gravity in
minimum time.
The formulation of the constrained brachistochrone problem considered here is
[84]:

min tf (7.107)
θ(t),tf

subject to:

ẋ = v cos θ (7.108)

ẏ = v sin θ (7.109)

v̇ = g sin θ (7.110)

y ≥ ax − b (7.111)

x(tf ) = xf (7.112)

[x(t0 ), y(t0), v(t0 )] = [0, 0, 0] (7.113)

where tf is the final time, g is the gravitational constant, and γ and b are given.
Equation (7.111) defines an inequality constraint on the state variables x and y,
(7.112) defines a final-point constraint, and (7.113) gives the initial conditions.

257
 Table 7-2: Solution statistics for constrained brachistochrone
Initial Guess QP LS CPU
θ(t) = −1.558 + 2.645t 2 3 0.38s
θ(t) = −1.6 + 2.2857t 3 3 0.47s
θ(t) = 0.0 3 5 0.52s
θ(t) = −1.6 5 5 0.67s
θ(t) = −1.0 6 6 0.75s

This problem was solved using the feasible-path control parameterization method
described in this section. The discretization chosen for the control θ(t) was a single
linear element. There were therefore three parameters in the NLP (the two for the
control variable and tf ). The parameters g, a, b, and xf were set to −1.0, −0.4,
0.3, and 1.1 respectively. The integrator relative and absolute tolerances (RTOL and
ATOL) were both set to 10−7 and the optimization tolerance was 10−5 . The optimal
objective function value was 1.86475. The ABACUSS input file is shown in Figure 2-
2, and Table 7-2 gives solution statistics from different initial guesses for the control
variable. In this table, “QP” refers to the number of master iterations of the NLP
solver, “LS” refers to the number of line searches performed, and “CPU” gives the
total CPU time for the solution on an HP C160 workstation. These statistics compare
very favorably with those reported for other similarly-sized problems [142].
Figure 7-10 shows the state and control variable solution of the constrained bra-
chistochrone problem described above using the constrained dynamic optimization
algorithm. The ‘kink’ in the control variable profile is due to the control becoming
determined by the high-index DAE during the constrained portions of the trajectory.

7.4.6 Critique of the FIFP method

The FIFP method is interesting because it is the first method proposed that explicitly
addresses the two issues that have been shown to be characteristic of inequality state-
constrained dynamic optimization: the high-index segments of the state trajectory
and the hybrid discrete/continuous nature of the problem. FIFP appears to solve the

258
 ABACUSS Dynamic Optimization
Y x 10-3
CONSTRAINT
0.00 Y

-50.00

-100.00

-150.00

-200.00

-250.00

-300.00

-350.00

-400.00

-450.00

-500.00

-550.00

-600.00

-650.00

-700.00

-750.00
X
0.00 0.20 0.40 0.60 0.80 1.00

ABACUSS Dynamic Optimization

Theta
THETA
0.00

-0.10

-0.20

-0.30

-0.40

-0.50

-0.60

-0.70

-0.80

-0.90

-1.00

-1.10

-1.20

-1.30

-1.40

-1.50
Time
0.00 0.50 1.00 1.50

Figure 7-10: Solution of constrained brachistochrone problem

problems associated with the modified slack variable approach, since it has the ability
to handle more inequality constraints than controls (provided that there is no point
in the problem where there are more active inequality constraints than controls), and
very little extra work is performed when constraints are inactive.

However, experience with the FIFP method has shown that it suffers from a
pathological problem. Recall from Chapter 3 that part of the sufficient conditions
for existence and uniqueness of sensitivity functions was that the sequence of state
events remains qualitatively the same: the location of the state events in state space
may change, but existence and uniqueness cannot be guaranteed if the ordering of the
events changes. This problem was demonstrated in Figures 3-3, 3-4, and 3-5, where
a differential change in a parameter produced a discrete jump in the values of both

259
the state and sensitivity functions.
The fact that an infinitely small movement in the value of a parameter can produce
a finite jump in the value of a state variable implies that the dynamic optimization
master NLP is not smooth. Nonsmooth NLPs often prevent convergence of gradient-
based NLP solution techniques (such as SQP algorithms). Because such methods
are generally based on assumptions of smoothness, nonsmooth problems can prevent
convergence or slow it to an unacceptably slow speed.
The FIFP method does work for some problems like the simple constrained bra-
chistochrone problem shown above because the FIFP problems are “smooth” in some
neighborhood of the optimum. That is, small movements of the parameter values away
from their optimal values generally do not change the sequence of events and thus
the objective function does not contain discontinuities. However, larger movements
of the parameters from their optimal values do cause the state variables to jump. For
example, in the constrained brachistochrone problem, some values of the constraints
produce state trajectories for which the inequality path constraint does not become
active. Fortunately for the constrained brachistochrone problem presented above, the
region around the optimum for which the sequence of events remains constant is fairly
large and convergence is fairly robust.
In a nutshell, the problem with the FIFP method is that it has an unpredictable
and potentially very small convergence region. That is, gradient based NLP algo-
rithms can be expected to converge to a local optimum (if one exists) from a starting
guess in the neighborhood of the optimum. This convergence neighborhood is defined
as the set of parameter values that produce the same constraint event sequence as
the locally optimal parameter values do. Since the constraints are defined implic-
itly, there is no general method at present for defining the convergence neighborhood
without actually solving hybrid IVPs.
Note that the convergence problem with the FIFP algorithm exists only when
FIFP is used with gradient-based NLP algorithms. Nonsmoothness of the NLP may
not be a problem for non-gradient based (e.g., stochastic) optimization algorithms
(see, for example [8]). Since the FIFP method generates trajectories that are always

260
feasible with respect to the path constraints, stochastic methods can then avoid the
need to discard many infeasible trajectories, and will be significantly more efficient.

261
7.5 Fluctuating-Index Infeasible-Path Method for
Dynamic Optimization with Inequality Path
Constraints

The fluctuating-index infeasible path (FIIP) method for inequality constrained dy-
namic optimization avoids the convergence difficulties that the FIFP method expe-
riences when the constraint event ordering changes during intermediate iterations
by requiring that each dynamic simulation subproblem follows the same sequence of
constraint activations and deactivations. This requirement is imposed by explicitly
defining the constraint activation and deactivation times. The method is an infeasi-
ble path method from the standpoint of the state variable inequality path constraints
because these constraints are not required to be satisfied for all time at intermediate
iterations of the FIIP NLP.

Since the inequality path constraints may be violated during intermediate itera-
tions, an explicit event constrained dynamic simulation algorithm can be used to solve
the IVP subproblem. The activations and deactivations of the state variable inequal-
ity path constraints occur at explicitly defined points in the solution trajectory that
are determined by the master NLP solver. During segments of the state trajectory
where an inequality constraint gk (x) is defined to be active, the equality:

gk (x) = k gk ⊆ g (7.114)

is added to the DAE using the method described in Chapter 6. The parameter k is
used because at the time when the constraint is defined explicitly to be active, the
values of the state variables may not satisfy the constraint. Therefore, the constraint
is offset by k such that it is active at the current time, which is shown in Figure 7-11.
The overall NLP contains a constraint that forces k = 0 at the optimum. Note that
the overall state variable inequality path constraint gk (x) = 0 is active when k = 0.

262
 State Inequality Constraint

State Trajectory

Constraint Offset

Unconstrained Element

Constrained Element

Unconstrained Element

Figure 7-11: The FIIP method works by offsetting the constraint to the state variable
trajectory at points where the constraint activates

The explicit event constrained dynamic simulation algorithm is:

f (ẋ, x, us , t) = 0 (7.115)

uc = uc (t) (7.116)
ng
(k) (k) (k)
αi (usi − uci ) + βij gj (x) − j =0 (7.117)
j=1
(k)
j = gj (x(tk ), y(tk )) (7.118)
ng
(k) (k)
αi + βij = 1 (7.119)
j=1
ng
(k)
βij ≤ 1 (7.120)
j=1
(k) (k)
αi , βij ∈ {0, 1}

i = 1 . . . nu k = 1 . . . NF E

As in the implicit event dynamic simulation method (7.70–7.74), the α and β vari-
ables are used to express the matching between active state path constraints and
the control variables that are not free to be independently specified in the resulting
high-index DAE. However, in this case, the α and β variables are not determined by

263
the occurrence of an implicitly defined state event, but instead are defined explicitly
for each control variable finite element. Therefore, the α and β are functions of the
finite element index, rather than the simulation time. The start and end times of
the finite elements are in general optimization parameters that are determined by the
NLP solver.
The j variable in (7.117) permits continuity assumptions for the state variables
when the constraint activates by allowing the inequality path constraint to be vio-
lated. Essentially, j is used to offset the constraint so that the state variables do
not jump when it activates. The amount of constraint violation is measured at the
point where the constraint activates, and remains constant throughout the entire con-
strained segment. The inequality constraint is satisfied at the optimum by imposing
a constraint in the master NLP such that j = 0.
Assuming a standard BDF method is used to solve the DAE, and that a set of α
and β variables have been defined, the steps of the explicit event dynamic simulation
method are:

1. Solve the IVP until the end of the next control finite element is encountered at
t = tk .

2. If tk = tf then stop.

(mu (k)
3. For every j where i=1 βij = 1, determine j by solving:

(k)
gj (x) = j (7.121)

at t = tk .

4. Apply the method of dummy derivatives to derive an equivalent index-1 formu-

lation of the current problem.

5. Consistently initialize the equivalent index-1 formulation.

6. Set k = k + 1. Go to Step 1.

264
The FIIP method is less complicated than the FIFP method because there are no
implicit constraint activations and deactivations. However, there are still two aspects
of the algorithm which require further explanation: how to specify the transfer con-
ditions for the state variables when constraints activate, and how to determine the
sequence of activations and deactivations.

7.5.1 Transfer conditions for state variables

As with the FIFP method, transition conditions must be defined for the state variables
when a constraint becomes active. Even though the constraint event time is explicitly
defined in the FIIP method, special consideration is required to define the FIIP
transition conditions because the dynamic degrees of freedom and possibly the index
of the DAE will change after a constraint event.
The principle that was used to define the state transition conditions in Sec-
tion 7.4.3 for the FIIP method is useful for the FIFP method as well. There is
an explicit input-jump simulation that corresponds to the explicit event constrained
dynamic simulation, and the continuity conditions for the explicit input-jump sys-
tem can be used to derive continuity assumptions for the explicit event simulation.
The use of the offset parameter with the active constraints removes the ambiguity
that would be associated with constraint activation. For example, suppose that the
constraint:

x1 + x2 + x3 ≤ 0 (7.122)

became active when the values of the state variables were x1 = −1, x2 = −2, and
x3 = −1, the corresponding explicit input jump simulation indicated that all of
the state variables in the constraint remain continuous, but the dynamic degrees
of freedom for the DAE after the constraint activation indicated that any two of
the three state variables can be specified to remain continuous. Then, one of the
state variables would have to experience a jump. However, allowing one of the state
variables to jump would violate the corresponding explicit input-jump simulation

265
continuity assumptions. Also, it is ambiguous as to which state variable would jump.
If however the active constraint is defined as:

x1 + x2 + x3 = (7.123)

where = −4, then no state variable will jump, even though continuity is enforced
on only two of the three state variables.

To illustrate the state variable transfer conditions under the FIIP method, consider
the DAE:

ẋ = x + y (7.124)

y = 3x + u (7.125)

u = u(t) (7.126)

The usual way to reinitialize this DAE at a time t∗ when a discontinuity in u is

encountered is to assume that the differential state variable x is continuous across the
discontinuity. Therefore, a consistent initial condition for the system that is active
after the discontinuity would be obtained by solving the algebraic system consisting
of (7.124–7.125) and

x(+) = x(−) (7.127)

u = u(t∗ ) (7.128)

to find values for x(+) , y (+) , and ẋ(+) . Now suppose that (7.124–7.125) is the DAE
part of a dynamic optimization problem that is being solved using the FIIP method
and the constraint:

x≤5 (7.129)

becomes active at t(∗) . After application of the dummy derivative method, the system

266
solved to find a consistent initialization at t(∗) is:

x̄(+) = x(+) + y (+) (7.130)

y (+) = 3x(+) + u(+) (7.131)

x(+) = 5 + (x(−) − 5) (7.132)

x̄(+) = 0 (7.133)

where x̄ is a dummy derivative and = (x(−) − 5) in this case. Note that even
though this system does not require specification of additional transfer conditions,
the differential state variable x remained continuous when the constraint activated.
The same state trajectories could have been obtained with the appropriate jump in
u in the system (7.124–7.125).

7.5.2 Point constraints

Theoretically, there is no need for additional constraints in the NLP to handle in-
equality path constraints in the FIIP method. However, including a point constraint
representation of the inequality path constraints improves the robustness of the FIFP
method. If the point constraints are not included, the FIFP method sometimes finds
locally optimal solutions for which the inequality path constraints are satisfied dur-
ing finite elements where they are defined to be active, but violated during finite
elements where the constraint is not supposed to be active. Including a measure of
the inequality constraint violation at a discrete number of points as an inequality con-
straints in the NLP tends to solve this problem. However, since these NLP inequality
constraints increase the size of the NLP, it is desirous not to include too many. In
practice, one NLP inequality constraint is included per inequality path constraint per
finite element.

267
7.5.3 Specifying a sequence of constraint events

Specifying the sequence of constraint events can be difficult in the FIIP method.
The FIIP method decouples the inequality path-constrained dynamic optimization
problem into discrete and continuous subproblems. The discrete subproblem is to
find the optimal sequence of constraint activations and deactivations. The continuous
subproblem is, given a fixed sequence of constraint activations and deactivations,
to find the optimal control trajectories. The use of the explicit event constrained
dynamic simulation algorithm allows us to solve the continuous subproblem.
Solution of the discrete subproblem is combinatorial in nature, yielding a mixed-
integer dynamic optimization subproblem (MIDO) [1, 3]. Putting the FIIP NLP in
the form of the MIDO formulation given in [3] yields:
 tf
min φ(x(tf ), u(tf ), v, tf ) + L(x(t), u(t), v, t)dt (7.134)
u(t),v,α,β,tf 0

subject to:

f (ẋ, x, us , t) = 0 (7.135)

uc = uc (t) (7.136)
ng
(k) (k) (k)
αi (usi − uci ) + βij gj (x) − j =0 (7.137)
j=1
(k)
j = gj (x(tk ), y(tk )) (7.138)
(k)
j = 0 (7.139)
ng
(k) (k)
αi + βij = 1 (7.140)
j=1
ng
(k)
βij ≤ 1
j=1
(k) (k)
αi , βij ∈ {0, 1}

i = 1 . . . nu k = 1 . . . NF E

At present, there exist no rigorous strategies for solution of general MIDOs. Further,

268
in the worst case where there are many controls, inequality constraints, and control
finite elements, the only rigorous strategy at present that will guarantee an optimal
answer is to enumerate explicitly all of the possible constraint event sequences and
compare the optimal values obtained by solving the associated continuous subprob-
lems.
Most dynamic optimization problems with state inequality constraints that have
been solved in the optimal control literature to date contain only one inequality
constraint and one control variable. These problems may be easily solved using the
FIIP method. The best way is to solve the unconstrained problem and determine the
finite elements during which the constraint is violated. Then, use the FIIP method
to solve the constrained problem where those same finite elements are the ones where
the constraint is defined to be active. An alternate method for single-control, single-
inequality problems is to specify alternating constrained and unconstrained finite
elements. Since the end points of the finite elements are optimization parameters,
the optimization problem can determine where the constraint activation takes place
by shrinking the length of some elements to zero if necessary.
For problems with multiple inequality constraints, one ad hoc strategy is to solve
the unconstrained problem and determine the sequence of constraint violations, and
then solve the constrained problem with that sequence. This strategy does not guar-
antee an optimal solution and some enumeration of constraint event orderings is
usually required. On the other hand, other methods for handling inequality path
constraints generally have problems with multiple constraints, and the FIIP method
does provide an extremely efficient method (see Chapter 8) to obtain the optimal
solution when the constraint activation order is known.

7.6 Conclusions

The advantage of the three methods described in this chapter over other methods
for solving inequality path constrained dynamic optimization problems is that these
guarantee that the inequality path constraints are not violated during an optimal

269
solution. In addition, these methods explicitly recognize the high-index nature of
path constrained problems, and employ the dummy derivative method to solve them
efficiently.
This chapter has clearly shown the hybrid discrete/continuous nature of inequality
constrained dynamic optimization problems that arises because of the need to make
decisions concerning the ordering of constraint events. Although the modified slack
variable method is capable of transforming some inequality path-constrained problems
into completely continuous optimization problems, the range of problems that it can
handle is limited.
The advantage of the FIFP method over other solution methods is that the discrete
nature of the inequality constrained problem is explicitly recognized, and solution is
possible if the constraint activation sequence is known. Although it is difficult at this
time to use the FIFP method to solve dynamic optimization problems with many con-
straints because of the combinatorial nature of these problems, dynamic optimization
problems with multiple inequality constraints are in fact very difficult to solve by any
currently existing method. Furthermore, as shown in the next chapter, the method
works well for a broad range of constrained dynamic optimization problems.

270
Chapter 8

Numerical examples

In this chapter, the FIIP algorithm is applied to several interesting examples in order
to demonstrate the ease with which it can be used to solve state-constrained dynamic
optimization problems that are considered difficult to solve using other methods.
In particular, the abilities of FIIP to track nonlinear constraints, to guarantee that
constraints are not violated at the optimum, and to solve problems without the use
of complex control basis functions and many finite elements are demonstrated.
An implementation of FIIP in ABACUSS was used to solve these problems. The
Harwell code VF13 is used to solve the master NLPs. VF13 uses a sequential quadratic
programming method with a gradient-based line search technique, and hence all of
the IVP subproblems are required to solve the combined DAE and sensitivity system.
This combined system is solved using the DSL48S code detailed in Chapter 4. ABA-
CUSS uses the efficient automatic differentiation algorithm of [135, 134] to obtain
the Jacobian of the DAE and right hand sides of the sensitivity equations. ABA-
CUSS does sacrifice some computational efficiency compared to methods where the
equations and derivatives are hard-coded and compiled, because it creates and stores
the equations and derivatives as memory structures. However, this computational
inefficiency is more than compensated by the modeling efficiency that the ABACUSS
input language and architecture provide.
There are several solution statistics reported for each problem. The integration
tolerance was used for both the relative and absolute tolerance for the numerical IVP

271
solver. The optimization tolerance was used for the NLP solver. The number of IVP
solutions is a measure of how difficult the NLP was to solve, since most of the solution
time is spent in numerical integration of the DAE and sensitivity equations. The CPU
times reported were obtained on an HP 9000 C160 single-processor workstation. The
number of IVP solutions and the CPU time statistics are of only limited usefulness
since they are highly dependent on the initial guess employed for the control variables,
however they are included for use in comparing some of the examples with results
obtained in other works that used the same initial guesses.
For most of the problems, results for several tolerance levels are shown. It is
common with control parameterization for the number of IVP subproblems required to
solve the problem to increase significantly as the tolerances are tightened. This effect
occurs because the objective function is often relatively flat near the optimum with
respect to the exact positions of the finite element boundaries. When the tolerances
are tightened, the optimizer continues to move the finite element boundaries, resulting
in small improvements in the objective function value. The solution trajectories shown
for each example correspond to the results for the tightest tolerances that were tried.
However, the solution statistics for most of the problems show that the work done
increases immensely at tight tolerance levels, and the improvements to the solution
are very small. Reasonably accurate solutions may be obtained without resorting to
the strictest tolerances that were used here.

272
8.1 Line-Constrained Brachistochrone

The constrained brachistochrone problem was presented both in [24] and in a slightly
different form in [84]. The formulation used here is that of [84], who solved the prob-
lem using a discretized constraint control parameterization algorithm. The problem
is to find the shape of a wire on which a frictionless bead will slide between two points
under the force of gravity in minimum time. The brachistochrone is constrained by
requiring it to lie on or above a line drawn in state space.

The formulation of the constrained brachistochrone problem that was considered

is [84]:

min tf (8.1)
θ(t),tf

subject to:

ẋ = v cos θ (8.2)

ẏ = v sin θ (8.3)

v̇ = g sin θ (8.4)

y ≥ ax + b (8.5)

x(tf ) = xf (8.6)

[x(t0 ), y(t0), v(t0 )] = [0, 0, 0] (8.7)

where tf is the final time, g is the gravitational constant, and a and b are given
constants. Equation (8.5) defines an inequality constraint on the state variables x
and y, (8.6) defines a final-point constraint, and (8.7) gives the initial conditions.

Following [84], the parameters used were g = −1.0, a = −0.4, b = 0.2, and xf =
1.0. The starting guess for the control was θ(t) = −π/2 and for the final time was tf =
2.0. The control was discretized using three linear finite elements, and the inequality
constraint was active during the second finite element. The optimal trajectories for
the states and controls are shown in Figure 8-1 and Figure 8-2. Solution statistics

273
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−7 10 −5
8 1.8046 0.59s
10−9 10−7 18 1.795220 2.06s
10−11 10 −9
22 1.79522045 3.98s

Table 8-1: Statistics for the line-constrained brachistochrone problem

are given in Table 8-1.

Our results compare favorably with the results reported in [84], which required 26
IVP solutions to solve the problem from the same initial guess using an optimization
tolerance of 10−6 , even though only three first-order control finite elements were used
for a total of 7 parameters, compared to the ten first-order control finite elements for a
total of 31 parameters used in [84]. The method used in [84] discretizes the inequality
constraint using NLP point constraints, and hence must use a large number of finite
elements to ensure that one lies near the point on the state trajectory at which the
inequality constraint activates. Since the FIIP method does not have to have a point
constraint near where the constraint activates, the method is more efficient because
it requires fewer sensitivity equations and smaller NLPs.

274
 Line Constrained Brachistochrone: Control Variable
0.2
0 θ(t)

-0.2
-0.4
-0.6
-0.8
-1
-1.2
-1.4
-1.6
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
t

Figure 8-1: Line-constrained brachistochrone control variable trajectory

Line Constrained Brachistochrone: State Variables

0
y(t)
-0.1 constraint

-0.2

-0.3

-0.4

-0.5

-0.6

-0.7
0 0.2 0.4 0.6 0.8 1
x(t)

Figure 8-2: Line-constrained brachistochrone state variable trajectory

275
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−7 10−5 10 1.8230 0.96s
10−9 10−7 34 1.812723 5.38s
10−11 10 −9
35 1.81272311 23.55s

Table 8-2: Statistics for the curve-constrained brachistochrone problem

8.2 Curve-Constrained Brachistochrone

An interesting variation on the previous problem is to constrain the brachistochrone
with a curve. This may be done by replacing (8.5) with

y ≥ ax2 + bx + c (8.8)

This problem was solved with the parameters in (8.8) set to a = −0.375, b =
−0.245, and c = −0.4. The initial guesses, parameter values, and control discretiza-
tion were the same as those used for the line-constrained brachistochrone. The opti-
mal trajectories for the states and controls are shown in Figure 8-3 and Figure 8-4.
The solution statistics are given in Table 8-2. The statistics show that the CPU
time grows at a faster rate than the number of IVP solutions as the tolerances are
tightened. This effect occurs because the number of integration steps and Jacobian
factorizations taken during the solution of each IVP increases as the tolerances are
tightened.

276
 Curve Constrained Brachistochrone: Control Variable
0.2
0 θ(t)

-0.2
-0.4
-0.6
-0.8
-1
-1.2
-1.4
-1.6
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
t

Figure 8-3: Curve-constrained brachistochrone control variable trajectory

Curve Constrained Brachistochrone: State Variables

0
y(t)
-0.1 constraint

-0.2

-0.3

-0.4

-0.5

-0.6
0 0.2 0.4 0.6 0.8 1
x(t)

Figure 8-4: Curve-constrained brachistochrone state variable trajectory

277
8.3 Constrained Van Der Pol Oscillator

This problem was taken from [142]. The problem is:

min y3 (tf ) (8.9)

u(t)

subject to:

y˙1 = (1 − y22 )y1 − y2 + u (8.10)

y˙2 = y1 (8.11)

y˙3 = y12 + y22 + u2 (8.12)

−0.3 ≤ u ≤ 1.0 (8.13)

y1 ≥ −0.4 (8.14)

[y1 (t0 ), y2(t0 ), y3 (t0 )] = [0, 1, 0] (8.15)

The problem was solved using ten linear finite elements to discretize the control,
and the inequality constraint was active during the third finite element. The initial
guess for the control variable was u(t) = 0.70. Solution statistics are given in Table 8-
3 and the optimal trajectories for the states and controls are shown in Figure 8-5 and
Figure 8-6. The results shown are those corresponding to the tighter tolerance levels.

It may be noted that the CPU times are much longer for this problem than for the
brachistochrone examples, even though all of the models contain similar numbers of
equations. The reason for the increased solution time for the Van der Pol problem is
that it contains ten finite elements, and the integrator must restart at the beginning
of each finite element. The BDF method can be inefficient while it is starting because
it takes small steps, and therefore a large number of integration restarts can exact
a significant toll on the overall computational efficiency. A modification to the BDF
method was proposed in [2] which allows the integrator to take larger steps as it is
starting.

Several slightly different answers to this problem were presented in [142]. The ob-

278
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−7 10 −5
25 2.9548 11.37s
10−9 10−7 65 2.954756 44.88s
10−11 10 −9
78 2.9547566 80.39s

Table 8-3: Statistics for the constrained Van der Pol oscillator problem

jective function values reported in that work vary from 2.95678 to 2.95421, depending
on the scheme used to measure the violation of the state path constraint. The FIIP
method converged significantly faster, although the best objective function value the
algorithm found was slightly larger than the lowest reported in [142]. One explanation
of this very small difference is that the results presented in [142] may have actually
allowed small violations of the state path constraint, whereas our method guarantees
that the solution does not violate this constraint, and hence solves a more constrained
problem. Alternatively, this problem, like many control parameterization problems,
is multimodal, and the algorithm may not have converged to the same local optimum
that was found in [142]. However, the FIIP solution is guaranteed not to violate the
inequality path constraint.

279
 Constrained Van Der Pol Oscillator: Control Variable
1
u(t)
0.8

0.6

0.4

0.2

-0.2

-0.4
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t

Figure 8-5: Constrained Van Der Pol control variable trajectory

Constrained Van Der Pol Oscillator: State Variables

3
y1 (t)
2.5 y2 (t)
y3 (t)
2

1.5

0.5

-0.5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
t

Figure 8-6: Constrained Van Der Pol state variable trajectories

280
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−5 10−3 17 35.000 0.36s

Table 8-4: Statistics for the constrained car problem

8.4 Constrained Car Problem

The constrained car problem is a classic optimal control problem (see for example,
[90], [142]). The problem is to find the minimum time for a car to travel between two
points, subject to constraints on the velocity and acceleration. The model used was:

min tf (8.16)
a(t),tf

subject to:

v̇ = a (8.17)

ẋ = v (8.18)

v ≤ 10 (8.19)

−2 ≤ a ≤ 1 (8.20)

[x(t0 ), v(t0 )] = [0, 0] (8.21)

[x(tf ), v(tf )] = [300, 0] (8.22)

where x is the position of the car, v is the velocity, and a is the acceleration. The
initial and final point constraints require the optimal solution to be one where the
car starts and finishes with zero velocity. There is a speed limit that constrains the
velocity to be less than 10.
The problem was solved using three constant finite elements, and the inequality
constraint was active during the second element. The optimal trajectories for the
acceleration and velocity are shown in Figure 8-7 and Figure 8-8. They match the
numerical solution given in [142] and the analytical solution given in [90]. Solution
statistics are given in Table 8-4. Statistics for only one tolerance level are reported

281
because no improvement is possible on the solution found, and tighter tolerances do
not increase the number of IVPs or improve the objective function.

282
 Constrained Car Problem: Control Variable
1
a(t)
0.5

-0.5

-1

-1.5

-2
0 5 10 15 20 25 30 35 40
t

Figure 8-7: Constrained car problem acceleration profile

Constrained Car Problem: State Variable

12
v(t)
10

0
0 5 10 15 20 25 30 35 40
t

Figure 8-8: Constrained car problem velocity profile

283
8.5 Index-2 Jacobson and Lele Problem

This problem was originally presented in [74], and was later solved using different
methods in [66, 90, 98, 106, 142]. The problem is:

min y3 (tf ) (8.23)

u(t)

subject to:

y˙1 = y2 (8.24)

y˙2 = −y2 + u (8.25)

y˙3 = y12 + y22 + 0.005u2 (8.26)

y2 ≤ 8(t − 0.5)2 − 0.5 (8.27)

−3.0 ≤ u ≤ 15.0 (8.28)

[y1 (t), y2(t), y3 (t)] = [0, −1, 0] (8.29)

There are two versions of this problem. The original form of the problem presented
in [74] and solved in [34, 106, 98] is shown above. A modified version was used in [66]
and [142], where (8.25) was changed to:

y˙2 = y2 + u (8.30)

The latter problem is unstable. Hence, the application of numerical integration algo-
rithms to this problem should be treated with suspicion.
The original form of the problem was solved using twelve finite elements, with the
inequality constraint active on the seventh element. The initial guess for the control
profile was u(t) = 6.0. The optimal trajectories for the control and constrained
state variable are shown in Figure 8-9 and Figure 8-10. Solution statistics are given
in Table 8-5. The value that was obtained for the objective function is very close
(within 0.1%) to the value reported in [90]).
A sequenced initial guess method was used in this problem. That is, the solution

284
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−5 10−3 11 0.217 2.9s
10−7 10−4 13 0.17982 6.0s
10−9 10 −7
61 0.1698402 39.8s
10−11 10−9 54 0.169832150 49.9s

Table 8-5: Solution statistics for original index-2 Jacobson and Lele problem

Integration Optimization IVP Objective

CPU
Tolerance Tolerance Solutions Function
10−7 10−5 26 0.18421 7.6s
10−9 10−7 73 0.1800561 27.5s
10−11 10 −9
60 0.180021318 31.5s

Table 8-6: Solution statistics for the modified index-2 Jacobson and Lele problem

was obtained for the tighter tolerances by using as the initial guess the answer from
the problem with the next loosest tolerances. This method often decreases the time
to solve control parameterization problems because tight integration tolerances do
not need to be enforced when the NLP is far from the optimal solution.
The modified version of the problem was also solved using a control parameteri-
zation of seven finite elements with the constraint enforced during the fourth finite
element. The solution statistics for the modified problem are given in Table 8-6, and
the control and constrained state variable trajectory are shown in Figure 8-11 and
Figure 8-12. The optimal value for the objective function are slightly better than the
one reported in [142].
The optimal trajectories for the control and constrained state variable for the
original form are shown in Figure 8-9 and Figure 8-10. The same variables are shown
for the modified form in Figure 8-11 and Figure 8-12.

285
 Original Index-2 Jacobson and Lele Problem: Control Variable
12
u(t)
10

-2

-4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t

Figure 8-9: Original Index-2 Jacobson and Lele control trajectory

Original Index-2 Jacobson and Lele Problem: Constrained State Variable

1.5
y2 (t)
constraint
1

0.5

-0.5

-1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t

Figure 8-10: Original Index-2 Jacobson and Lele y2 trajectory

286
 Modified Index-2 Jacobson and Lele Problem: Control Variable
16
14 u(t)
12
10
8
6
4
2
0
-2
-4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t

Figure 8-11: Modified Index-2 Jacobson and Lele control trajectory

Modified Index-2 Jacobson and Lele Problem: Constrained State Variable

1.5
y2 (t)
constraint
1

0.5

-0.5

-1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t

Figure 8-12: Modified Index-2 Jacobson and Lele y2 trajectory

287
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−7 10−5 71 0.75156 35.4s
10−9 10−7 37 0.7514907 25.2s
10−11 10 −9
251 0.75144685 210.9s

Table 8-7: Solution statistics for the index-3 Jacobson and Lele Problem

8.6 Index-3 Jacobson and Lele Problem

Another interesting variation on the Jacobson and Lele problem is to replace (8.27)
with

y1 ≤ 8(t − 0.5)2 − 0.5 (8.31)

This change has the effect of making the problem index-3 when the constraint is
active, rather than index-2 when the problem is constrained as above with (8.27).
This problem was solved using twelve linear finite elements, with the constraint
enforced on the third element. A discontinuity in the control was permitted after the
second element. The finite element size of all the finite elements except the constrained
element was bounded from below by 0.05. The statistics shown in Table 8-7 use the
sequenced initial guess method. The optimal trajectories for the control and the
constrained state variable in Figure 8-13 and Figure 8-14 show that the constraint is
active only at one point. The objective function values reported here are very close
to those reported in [90], even though a much simpler control discretization than the
one used in that work was used.

288
 Index-3 Jacobson and Lele Problem: Control Variable
16
14 u(t)

12
10
8
6
4
2
0
-2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t

Figure 8-13: Index-3 Jacobson and Lele control trajectory

Index-3 Jacobson and Lele Problem: Constrained State Variable

0.1
y1 (t)
0 constraint

-0.1

-0.2

-0.3

-0.4

-0.5

-0.6
0 0.2 0.4 0.6 0.8 1
t

Figure 8-14: Index-3 Jacobson and Lele y1 trajectory

289
8.7 Pressure-constrained Batch Reactor

The following gaseous reactions take place isothermally and simultaneously in a

pressure-constrained batch reactor:

k1
A  2B (8.32)
k2
k
A + B 3 D (8.33)

This simple problem is useful for showing the ability of the high-index DAE to find
a highly nonlinear control that tracks a constraint.
The dynamic optimization problem is:

max CD (tf ) (8.34)

F

subject to:

ĊA = −k1 CA + k2 CB CB + F/V − k3 CA CB (8.35)

ĊB = k1 CA − k2 CB CB − k3 CA CB (8.36)

ĊD = k3 CA CB (8.37)

N = V (CA + CB + CD ) (8.38)

P V = NR T (8.39)

P ≤ 340000 (8.40)

0 ≤ F ≤ 8.5 (8.41)

[CA (0), CB (0), CD (0)] = [100, 0, 0] (8.42)

where Ci is the concentration of species i, P is the pressure in the reactor, N is

the total number of moles in the reactor, V is the reactor volume, R is the gas
constant, F is the feed rate of pure A, and k1 and k2 are the rate constants for the
two reactions. The rate constants were set to k1 = 0.8 hr−1 , k2 = 0.02 m3 /(mol · hr),
k3 = 0.003 m3 /(mol · hr), the volume was set to V = 1.0 m3 , the temperature was

290
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−7 10−5 10 11.7284 0.99s

Table 8-8: Solution statistics for the pressure-constrained batch reactor problem

T = 400 K, the gas constant was R = 8.314 J/(mol · K). The initial guess for the feed
rate of species A was F (t) = 0.5 mol/hr.
This problem was solved using two constant finite elements to approximate the
control, and the constraint was active during the second element. The optimal tra-
jectories for the control, the pressure, and the concentrations are shown in Figures 8-
15–8-17. Solution statistics are given in Table 8-8. During the constrained portion
of the trajectory, the control is highly nonlinear. If this problem were solved using
a penalty function approach, the control would require a complex discretization to
approximate this nonlinear optimal control trajectory. However, the FIIP method is
able to exploit the fact that the control is not independent when the constraint is
active, and a simple control discretization may be used.

291
 Pressure-Constrained Reactor Problem: Control Variable
9
F (t)
8.5

8
mol/hr

7.5

6.5

6
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
t(hr)

Figure 8-15: Pressure-constrained reactor control variable trajectory

Pressure-Constrained Reactor Problem: State Variable

341000
340000 P (t)

339000
338000
337000
Pa

336000
335000
334000
333000
332000
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
t (hr)

Figure 8-16: Pressure-constrained reactor pressure trajectory

292
 Pressure-Constrained Reactor Problem: Control Variable
100
CA (t)
CB (t)
80 CD (t)

60
mol/m3

40

20

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
t(hr)

Figure 8-17: Pressure-constrained reactor concentration trajectories

293
8.8 Fed-Batch Penicillin Fermentation

The determination of optimal control profiles for fed-batch bioreactors is an inter-

esting chemical engineering problem that has received significant attention in recent
years. These problems are considered difficult because of the nonlinear system dy-
namics and the existence of constraints on both the state and control variables.

A model of a fed-batch penicillin fermentation reactor was described and solved

in [127]. The problem is:
 150
min J = −θXV + 0.0103P V + 0.0744μX + 0.00102XV + 0.00691358 (8.43)
S(t) 0

subject to:

FX
Ẋ = μX − (8.44)
V
FP
Ṗ = θX − 0.01P − (8.45)
V
0.11S
μ= (8.46)
S + 0.1X
0.004
θ= (8.47)
1 + 0.0001/S + S/0.1
V̇ = F (8.48)

0.001 ≤ S ≤ 0.5 (8.49)

X ≤ 41 (8.50)
 
[X(0), P (0), V (0)] = 1, 0, 2.5 · 105 (8.51)

The objective function in this problem takes into account the revenue from the prod-
uct, the cost due to product deactivation, the cost of the substrate, and the daily
cost. The state variables in the problem are the biomass concentration X (g/L),
the amount of penicillin product present P (activity/L), the substrate concentration
S (g/L), the reactor volume V (L), the specific product formation rate θ, and the
specific growth rate μ. The feed rate parameter F was set to 1666.67 L/h.

This problem is somewhat sensitive to the initial guess for the control trajectory.

294
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−7 10−5 21 −1.1230 · 106 8.94s

Table 8-9: Solution statistics for the CSTR and column startup problem

The initial guess used in this work was a step function starting out at S(t) = 0.4 and
falling after the first finite element to S(t) = 0.01 at t = 40.0.
The solution statistics for this problem are shown in Table 8-9. The optimal
trajectories for the control and selected state variables are shown in Figure 8-18 and
Figure 8-19. The trajectories and the objective function value agree closely with those
reported in [127].

295
 Fed-Batch Penicillin Fermentation Problem: Control Variable
0.5
S(t)

0.4

0.3
(g/L)

0.2

0.1

0
0 20 40 60 80 100 120 140
t(hr)

Figure 8-18: Fed-batch penicillin fermentation control variable trajectory

Fed-Batch Penicillin Fermentation Problem: State Variables

45
40 X(t) (g/L)
P (t) (act./L)
35
30
25
20
15
10
5
0
0 20 40 60 80 100 120 140
t(hr)

Figure 8-19: Fed-batch penicillin fermentation state variable trajectories

296
 A

B
Reactor R
Qcondenser
Cooling Jacket

Column

W aterin

Fout

Reboiler

W aterout Qreboiler

Figure 8-20: CSTR and column flowsheet

8.9 Startup of CSTR and Distillation Column

This example demonstrates how constraining the dynamic optimization problem can
reduce the number of controls in the dynamic optimization problem, thus making it
easier to solve. It also demonstrates the ability of FIIP to handle problems that are
modeled by large-scale DAEs.

Consider the flowsheet containing a continuous stirred tank reactor (CSTR) and
distillation column shown in Figure 8-20. There are two simultaneous reactions which
take place in the liquid phase in the CSTR:

k
A + B →1 C (8.52)
k
B + C →2 D (8.53)

The desired product is C, but there is a side reaction which is favored at higher tem-
peratures which consumes the desired product and produces the undesired product
D.

The reactor contents must be kept below 400K at all times to avoid vaporization

297
of the reacting mixture. The reactor temperature may be controlled by adjusting the
flowrate of water to a cooling jacket which surrounds the CSTR.

The reaction mixture leaves the reactor at a prescribed flowrate and flows into a
stripper column. The column is intended to strip off the lighter reactants A and B,
and the top product from the column is recycled back to the reactor.

The dynamic model for this example is detailed below. The model and its param-
eters are loosely based on one given in [33].

8.9.1 CSTR model

The reactor material balance on each species i is:

NR
Ṅi = xin
i F
in
+V νji rj − xri F out ∀i = 1 . . . NC (8.54)
j=1

where Ni is the molar holdup of species i, V is the of the reactor contents, rj is

the rate of reaction of reaction j, νji is the stoichiometric coefficient of species i in
reaction j, xin r
i is the mole fraction of species i in the feed, xi is the mole fraction of

species i in the reactor, F in and F out are respectively the mole flows into and out of
the reactor, NR is the number of independent reactions, and NC is the number of
components.

The reaction rate is modeled by:


NC
β
rj = k j Ci ij ∀j = 1 . . . NR (8.55)
i=1

where Ci is the molar concentration of species i, βij is the order of reaction j with
respect to species i, and kj is the rate constant of reaction j given by the Arrhenius
equation:
 
Ej
kj = Aj exp − ∀j = 1 . . . NR (8.56)
RT

298
 The total enthalpy H r of the mixture in the reactor is:

NC
r
H = Ni Ĥir (8.57)
i=1

where Ĥir is the specific enthalpy of species i in the reactor, and an energy balance
on the entire reactor gives:

NC
r in
Ḣ = F xin in
i Ĥi
i=1
NC
(8.58)
+ F out i Ĥi − Q̇jacket
xout r

i=1

where Ĥ in is the specific enthalpy of species i in the reactor feed and Q̇jacket is the
heat removed by the reactor cooling jacket.

The volume of the reactor contents is:

NC
Ni
V = (8.59)
i=1
ρi

where ρi is the molar density of species i.

The reactor concentration and mole fraction are respectively given by:

Ni
Ci = ∀i = 1 . . . NC (8.60)
V
N i
xri = ∀i = 1 . . . NC (8.61)
Ntot

The total reactor molar holdup is:

NC
Ntot = Ni (8.62)
i=1

The parameters used in this model are given in Table 8-10. Both reactions are
first order (therefore βij = 1 ∀(i, j)), and the gas constant is 8.314 (kJ/kmol · K).

299
 Parameter Value
NR 2 Parameter Value
NC 4 A1 100.0 m3 /(hr · kmol)
νA1 -1 A2 120.0 m3 /(hr · kmol)
νA2 0 E1 17000 kJ/kmol
νB1 -1 E2 20000 kJ/kmol
νB2 -1 ρA 11.0 kmol/m3
νC1 1 ρB 16.0 kmol/m3
νC2 -1 ρC 10.0 kmol/m3
νD1 0 ρD 10.4 kmol/m3
νD2 -1

Table 8-10: Parameters used in the CSTR model

8.9.2 Reactor cooling jacket

The heat removed by the reactor cooling jacket Qjacket is:

Q̇jacket = UA (Tr − Tout

w
) (8.63)

w
where Tr is the reactor temperature, Tout is the outlet temperature of the cooling
water, U is the heat transfer coefficient of the jacket, and A is the heat transfer
surface area.

The heat is transferred to the cooling water, which gives the temperature rela-
tionship:

Q̇jacket = Cpw (Tout

w
− Tin
w
)Fw (8.64)

w
where Tout is the inlet temperature of the cooling water and Cpw is the heat capacity
of the cooling water.

The values of the parameters used for the cooling jacket were U = 3000 (kW/m2 ·
K), A = 30 m2 , and Tin
w
= 288 K. The heat capacity of water was assumed to be
independent of temperature at 4200 (kJ/tonne · K).

300
8.9.3 Column tray model

The column has NS stages, but the first stage is the condenser, and the last is the
reboiler, so there are NS − 2 equilibrium trays. The overall material balance on each
tray is:

Ṁk = Lk−1 − Lk + Vk+1 − Vk + Fk ∀k = 2 . . . NS − 1 (8.65)

and the material balance for each species is:

Ṁk xik + Mk ẋik =Lk−1 xik−1 − Lk xik +

∀k = 2 . . . NS − 1 (8.66)
i
Vk+1 yk+1 − Vk yki + Fk zki

where Mk is the molar holdup on stage k, xik and yki are respectively the liquid and
vapor mole fractions of species i on stage k, Fk is the feed flow rate to stage k, and
zki is the composition of the feed on stage k (assumed to be liquid).
Assuming fast energy dynamics and adiabatic operation, the energy balance on
each tray is:

l
0 =Lk−1 Ĥk−1 − Lk Ĥkl +
∀k = 2 . . . NS − 1 (8.67)
v
Vk+1Ĥk+1 − Vk Ĥkv + Fk Ĥkf

where Ĥkl , Ĥkv , and Ĥkf are respectively the molar specific enthalpies of the liquid,
vapor, and feed on stage k.
The vapor mole fractions on each tray are normalized:

NC
yki = 1 k = 2 . . . NS (8.68)
i=1

and, since thermodynamic equilibrium is assumed on each tray, the liquid phase
composition must satisfy:

yki = Kki xik ∀k = 2 . . . NS (8.69)

301
where Kki is the vapor/liquid equilibrium distribution coefficient for species i on tray
k.

The liquid overflow on each tray is related to the tray holdup by the Francis weir
equation:

NC  1.5
hk
Lk xik Z i = ρ̄lk WL (8.70)
i=1
750

where Z i is the molecular weight of species i, WL is the weir length in meters, hk is

the height of the liquid above the weir in millimeters, and ρ̄lk is the mass density of
the liquid. The molecular weights of species A, B, C, and D are 76, 52, 128, and 180
respectively. The mass densities of the liquid and vapor is given by:

# NC $−1
xik
ρ̄lk = (8.71)
i=1
Z i ρi
NC
Pk
ρ̄vk = (yki Z i ) (8.72)
RTk i=1

The volume of the liquid on the tray is:

NC
Mk
Vk = xki Zlk (8.73)
ρ̄lk i=1

and therefore hk is:

Vk
hk = − WH (8.74)
Atray

where Atray is the area of the tray, and WH is the height of the weir.

302
 The pressure drop on each tray is related to the vapor flow by:

tray
Pk = Pk−1 + ΔPk−1 (8.75)

ΔPktray = ΔPkstatic + ΔPkdry (8.76)

Vk
ΔPkstatic = ρ̄l g (8.77)
Atray k
# ( C i+1 i $2
Vk+1 N y Z
ΔPkdry = ρ̄vk i=1 k
Af ree ω (8.78)
ρ̄vk

where g is the gravitational constant, and ω is a constant dependent on the tray

geometry.
The column model has NS = 10 and the feed is on stage 5 (therefore, Fk = 0; ∀i =
5).

8.9.4 Column reboiler model

The overall mass balance on the reboiler is:

ṀN S = LN S−1 − B − VN S (8.79)

where B is the bottoms flowrate. The material balance for each species is:

ṀN S xiN S + MN S ẋiN S = LN S−1 xiN S−1 − BxiN S − VN S yN

i
S (8.80)

Assuming fast energy dynamics, the energy balance is:

0 = LN S−1 ĤNl S−1 − B ĤNl S − VN S ĤNv S + Q̇R (8.81)

where Q̇R is the reboiler heat duty.

The vapor mole fractions are normalized:

NC
i
yN S = 1 (8.82)
i=1

303
and the liquid phase composition must satisfy:

i i i
yN S = KN S xN S (8.83)

The flow out of the reboiler is related to the reboiler holdup by the proportional
control law:

B = 0.1(MN S − MN∗ S ) (8.84)

where MN∗ S is the set-point level.

8.9.5 Column condenser model

The column uses a total condenser with a constant holdup, therefore:

V2 = D + L1 (8.85)

xi1 = y2i (8.86)

V1 = 0 (8.87)

where D is the molar distillate flow.

The amount refluxed back to the top tray obeys the relation:

V2 r = L1 (8.88)

where r is the normalized reflux ratio.

The heat required to condense the vapor into the condenser Q̇C is given by:

Q̇C = V2 Ĥ2v − (L1 + D)Ĥ1l (8.89)

304
 The liquid mole fractions are normalized:

NC
y1i = 1 (8.90)
i=1

8.9.6 Enthalpy model

The specific enthalpy of a pure species in the vapor phase is assumed to be a function
of temperature:
 T
Ĥiv = ΔHif (Tref ) + ai dT (8.91)
Tref

where Tref is a reference temperature.

The specific enthalpy of a pure species in the liquid phase is assumed to obey the
relationship:

Ĥil = Ĥiv − ΔHivap (8.92)

The mixing of species is ideal, therefore:

NS
l
Ĥ = xi Ĥil (8.93)
i=1
NS
Ĥ l = yiĤiv (8.94)
i=1

The values of the parameters for the enthalpy model are given in Table 8-11. The
reference temperature was set to Tref = 298K.

8.9.7 Vapor-liquid equilibrium

Raoult’s law is assumed, and therefore the vapor/liquid equilibrium distribution co-
efficients are independent of composition, and are given by:

Pivap
Ki = (8.95)
P

305
 Parameter Value
aA 172.3 kJ/(kmol · K)
aB 200.0 kJ/(kmol · K)
aC 160.0 kJ/(kmol · K)
aD 155.0 kJ/(kmol · K)
ΔHAvap 31000 kJ/kmol
ΔHBvap 26000 kJ/kmol
ΔHCvap 28000 kJ/kmol
vap
ΔHD 34000 kJ/kmol
ΔHAf −30000 kJ/kmol
ΔHBf −50000 kJ/kmol
ΔHCf −20000 kJ/kmol
ΔHD f
−20000 kJ/kmol

Table 8-11: Parameters for the enthalpy model

where P is the pressure, and Pivap is the vapor pressure of species i.

The vapor pressure is assumed to obey the following function of temperature:

αi
ln Pivap = − + bi (8.96)
T

The parameters for the vapor pressure model are given in Table 8-12. The vapor
pressure is given in bar with these parameters. The pressure in the condenser was 1
atm.

Parameter Value
αA 4142.75 K
αB 3474.56 K
αC 3500.00 K
αD 4543.71 K
bA 11.7158
bB 9.9404
bC 8.9000
bD 11.2599

Table 8-12: Parameters for the vapor pressure model (vapor pressure in bar)

306
8.9.8 Path constraints

The reactor and column model that has been specified thus far has five controls (cor-
responding to the valves shown in Figure 8-20). Dynamic optimization problems with
many control variables are difficult to solve using any method. In control parameter-
ization, the size of the NLP increases with the number of controls, and in practice,
the degree of nonlinearity of the NLP increases. However, there are some additional
constraints on the state variables that may be specified which, due to the results of
Chapter 6, reduce the number of independent control variables. Note that the re-
boiler and condenser duties, listed as controls in Figure 8-20, are in fact completely
determined by the model that has been specified.
The path constraints that have been explicitly imposed on this model are:

1. Constant Reactor Holdup.

2. Constant ratio between A and B entering the reactor: The flowrate of B was
assumed to be 15% of the flowrate of A.

3. Constant flowrate of material out of the reactor. The flowrate was set to 15
kmols/hr.

The only independent control variables remaining in this problem are the column
reflux ratio and the flow rate of the cooling water to the reactor jacket. The reactor
temperature constraint causes the flow rate of the cooling water to be specified when-
ever the constraint is active. The flow rate of the cooling water was approximated
using a piecewise continuous linear control discretization. The column reflux ratio was
made a time-invariant optimization parameter. Hence, the complexity of the dynamic
optimization problem has been reduced to a manageable level using path constraints
which prescribe the values of some of the controls. The index of the DAE is two when
the inequality path constraint is inactive. The index is also two when the inequal-
ity path constraint is active, but a larger set of equations must be differentiated in
order to derive the equivalent index-1 DAE. Figures 8-21-8-22 show the ABACUSS

307
 Before differentiation, this model was index 2.

Equations 1 through 689 form an index=1 model.

RESULTS FROM ABACUSS STRUCTURAL ANALYSIS

Number of Variables in model : 680

Number of additional variables created : 0
____
Total Variables : 680

Number of Model Equations : 676

Number of Additional Model Equations Derived : 8
Number of Input Functions Specified : 5
Number of Additional Input Functions Derived : 0
Number of Initial Conditions Specified : 0
Number of Additional Initial Conditions Derived : 0
____
Total Equations : 689

Total Unknowns : 689

Figure 8-21: ABACUSS index-reduction output for reactor and column startup model
when the constraint is inactive

index reduction output for the model both without and with the inequality constraint
active.

8.9.9 Solution of the problem

The objective function for the dynamic optimization problem was:

min C̄(tf ) − 0.1D̄(tf ) (8.97)

F w ,R

308
 Before differentiation, this model was index 2.

Equations 1 through 705 form an index=1 model.

RESULTS FROM ABACUSS STRUCTURAL ANALYSIS

Number of Variables in model : 680

Number of additional variables created : 0
____
Total Variables : 680

Number of Model Equations : 677

Number of Additional Model Equations Derived : 24
Number of Input Functions Specified : 4
Number of Additional Input Functions Derived : 0
Number of Initial Conditions Specified : 0
Number of Additional Initial Conditions Derived : 0
____
Total Equations : 705

Total Unknowns : 705

Figure 8-22: ABACUSS index-reduction output for reactor and column startup model
when the constraint is active

309
where:

C̄˙ = BxC
NS (8.98)
˙ = BxD
D̄ (8.99)
NS

C̄(0) = 0 (8.100)

D̄(0) = 0 (8.101)

tf = 100hr (8.102)

This objective function measures the relative cost associated with producing C and
D during the startup (i.e., producing an extra kmol of C offsets the disadvantage
associated with producing 10 kmols of the undesired species D). The final time was
chosen to be long enough so that the system will be close to steady state at the end
of the simulation.
The condenser pressure was set to 1 atmosphere for the entire startup procedure.
The feed makeup temperature is 300K. The setpoint for the control equation in the
column reboiler was set to MN∗ S = 22 kmol At the initial time, the reactor contains
50 kmols of A at 300K, the total holdup in the column is 46.586382 kmols of A, and
each of the trays is filled to the weir height with A at its bubble point. Note that
since one of the path constraints fixes the ratio between A and B being fed to the
reactor, the flow of B into the reactor undergoes a step function at t0 .
The control variable F water (t) was approximated by four linear finite elements,
and the inequality path constraint was active during the second element. The initial
guess for the control was F water (t) = 3tonne/hr, and for the time invariant parameter
was R = 0.5. The control was bounded by 0 ≤ F water ≤ 5, and the reflux ratio was
bounded by 0.4 ≤ R ≤ 0.83. The size of the DAE was 741 equations, and there
were 13 optimization parameters, which brought the size of the combined DAE and
sensitivity system to 10374 equations. Solution statistics for this problem are given
in Table 8-13. Optimal solution profiles are shown in Figures 8-23–8-28.

310
 Integration Optimization IVP Objective
CPU
Tolerance Tolerance Solutions Function
10−5 10−3 91 153.9297 2455s

Table 8-13: Solution statistics for the CSTR and column startup problem

Reactor and Column Startup Problem

0.1535
Cooling Water Flow
0.153

0.1525
F w (ton/hr)

0.152

0.1515

0.151

0.1505

0.15
0 10 20 30 40 50 60 70 80 90 100
t (hr)

Figure 8-23: The optimal trajectory for the cooling water flowrate

Reactor and Column Startup Problem

420
Temperature in Reactor
400

380
T (K)

360

340

320

300
0 10 20 30 40 50 60 70 80 90 100
t (hr)

Figure 8-24: The temperature profile in the reactor

311
 Reactor and Column Startup Problem
2.5
A
2 B
C
D
1.5
(kmol/hr)

0.5

-0.5
0 20 40 60 80 100
t (hr)

Figure 8-25: The molar flowrates of the species leaving the system

Reactor and Column Startup Problem

400
395 T10
T9
390 T8
T7
385 T6
380 T5
T4
(K)

375 T3
T2
370 T1
365
360
355
350
0 20 40 60 80 100 120
t (hr)

Figure 8-26: The temperature profile in the column

312
 Reactor and Column Startup Problem
1.5e+07
1.45e+07 Reboiler Duty

1.4e+07
1.35e+07
Q̇r (kJ/hr)

1.3e+07
1.25e+07
1.2e+07
1.15e+07
1.1e+07
1.05e+07
0 20 40 60 80 100
t (hr)

Figure 8-27: The reboiler duty

Reactor and Column Startup Problem

1.446e+06
Condenser Duty
1.444e+06

1.442e+06
Q̇c (kJ/hr)

1.44e+06

1.438e+06

1.436e+06

1.434e+06

1.432e+06
0 20 40 60 80 100
t (hr)

Figure 8-28: The condenser duty

313
314
Chapter 9

Conclusions

This thesis has demonstrated several developments which facilitate reliable numerical
solution of the types of dynamic optimization problems found in process engineering.
Specifically, advances were made in three interrelated areas: sensitivity analysis, nu-
merical solution of high-index DAEs, and solution of dynamic optimization problems
with state variable path constraints.
The staggered corrector algorithm for numerical sensitivity analysis of DAEs has
been demonstrated to be significantly more efficient than other existing methods.
Also, it has been implemented in the DSL48S numerical solver for large-scale sparse
DAEs which is well-implemented, extensively tested, and stable. Therefore, efficient
numerical sensitivity analysis is now an easily available tool in process engineering,
and can be used, for example, in designing control strategies, finding the response of
a system to disturbances, and determining the design variables to which a system is
most sensitive.
The use of efficient sensitivity analysis makes the control parameterization method
more attractive relative to indirect methods for solving dynamic optimization prob-
lems. Use of indirect methods typically requires forward-integration of the DAE and
then backward-integration of the adjoint equations, and it does not appear possible
to take advantage of the fact that the Jacobians of both the adjoint and DAE sys-
tems are based on similar information. By using sensitivities instead of adjoints, the
staggered corrector method is able to minimize both the number of Jacobian factor-

315
izations and the number of Jacobian updates by taking advantage of the similarities
between the DAE and its sensitivity system.

Control parameterization is also more attractive than the simultaneous direct

methods that have been proposed for solving dynamic optimization problems. From
a practical standpoint, control parameterization is able to take advantage of highly
developed and efficient numerical solvers for the IVP and sensitivity subproblem,
and the master NLPs are kept to a size easily manageable with current technology.
The one supposed advantage of simultaneous methods, the ability to handle path con-
straints directly, is negated by methods like FIIP which handle the constraints directly
within the control parameterization framework. For these reasons, we believe that the
control parameterization method is the best currently available method for solving
the types of dynamic optimization problems encountered in process engineering.

An important result of this thesis is the clarification of the fact that path-con-
strained dynamic optimization problems are naturally high-index DAEs, and therefore
all methods that solve path constrained dynamic optimization problems are capable
of solving high-index DAEs. However, most dynamic optimization methods handle
these high-index DAEs indirectly. This thesis shows that there are advantages to
be gained by directly solving high-index DAEs, and vice-versa. This observation led
to the development in this thesis of the dummy derivative method as a practical
technique for solving a broad class of arbitrary-index DAEs. When used with con-
trol parameterization, the dummy derivative method allows the path constraints to
be included in the IVP subproblem, rather than in the Master NLP. The result is
that fewer IVP subproblems are required to solve many dynamic optimization prob-
lems, and the accuracy of the path constraints is guaranteed over the entire solution
trajectory to within the integration tolerances.

Including path constraints directly in the IVP raises a set of issues concerning
dynamic degrees of freedom and feasibility of the dynamic optimization problem.
Whenever a path constraint is appended to the system, a control variable becomes
determined by the solution to the augmented DAE, and thus is not free to be in-
dependently specified. A control matching algorithm was proposed that is able to

316
determine in many problems which, if any, controls cease to be design degrees of free-
dom in the augmented DAE. Furthermore, it was shown that the constrained dynamic
optimization problem can be feasible only if the augmented DAE is well-posed.

Inequality path constraints on dynamic optimization problems raise an additional

set of issues. The index of the DAE can change between segments of the trajectory
where the constraint becomes active or inactive. The transfer conditions for state and
sensitivity variables at these points require special attention. In Chapter 3 the trans-
fer conditions for the sensitivity variables of hybrid systems (which include systems
where inequalities activate and deactivate) were derived. A key result is that sensitiv-
ities of hybrid systems exist and are unique provided that the sequence of state events
does not change. This result is important for the development of methods for solv-
ing inequality path-constrained dynamic optimization and hybrid discrete/continuous
dynamic optimization problems.

One of the most interesting results of this work is that there are several strate-
gies which allow inequality path constraints on dynamic optimization problems to
be enforced by the IVP solver within the control parameterization framework. The
slack variable method transforms the inequality constraints into equality constraints.
The FIFP method transforms the inequality constrained problem into a hybrid dis-
crete/continuous optimization problem. The FIIP method transforms the inequality
constrained problem into a mixed-integer dynamic optimization problem (MIDO).

Each of these methods has its own set of advantages and disadvantages. The slack
variable method avoids problems with sequencing decisions for constraint activations
and deactivations. However, the slack variable method is not capable of handling
dynamic optimization problems with more inequality constraints than control vari-
ables, and it creates a high-index DAE whether or not the inequality constraint is
active. The FIFP method does not have restrictions on the relative number of control
variables and inequality constraints, and it handles the constraints efficiently, but it
creates nonsmooth NLPs with small regions of convergence. Such problems cannot be
easily solved with standard gradient-based optimization methods. The FIIP method
has all the advantages of the FIFP method and it creates smooth NLPs, but it requires

317
that the sequence of constraint activation and deactivation events be determined a
priori.
Of these methods, the FIIP method proved to be the most successful. To use
the method, a sequence of constraint events must be specified, and the index of the
DAE is permitted to fluctuate between unconstrained and constrained segments. The
path constraints are permitted to be violated during intermediate iterations in order
to provide transition conditions for the state variables between unconstrained and
constrained segments. The FIIP method was successfully demonstrated on several
examples. FIIP is efficient and it guarantees that the path constraints are satisfied
over the optimal solution trajectory.
Finding the optimal sequence of constraint activations and deactivations for the
FIIP method can be difficult. If the dynamic optimization problem contains few
constraints, as is the case with almost all problems that have been solved in the lit-
erature, it is usually possible to find the sequence by examining the solution of the
unconstrained dynamic optimization problem. In the case where the dynamic opti-
mization problem contains many constraints, finding the optimal sequence becomes
a combinatorial problem.
The methods developed in this thesis are capable of solving inequality constrained
dynamic optimization problems that cannot be solved using other methods, and prob-
lems that can be solved using other methods with greater accuracy and efficiency.

318
9.1 Directions for Future Work

An interesting area for future research is to find methods to solve the classes of
high-index DAEs that cannot be solved using the dummy derivative algorithm. The
dummy derivative method is not currently capable of solving problems for which
the index cannot be detected structurally. Detection of the index using numerical,
rather than structural methods is problematic from both theoretical and practical
standpoints. Theoretically, singularity of the Jacobian of a nonlinear DAE with
respect to the highest order time derivatives does not necessarily indicate that the
DAE is high-index. Practically, it is difficult to differentiate between a singular matrix
and one that is highly ill-conditioned.

The computationally expensive part of the dummy derivative method is dummy

derivative pivoting. Although the pivoting operation is not a significant expense
for most problems, detecting the need to pivot can be expensive. The heuristics
for minimizing the number of pivot checks proposed in this thesis perform well for
most problems. However, the heuristics could be improved if there were a method
for differentiating between truncation error failures that occur because of a need for
dummy pivoting, and those that occur for other reasons.

Structural criteria are also used in the methods developed in this thesis to deter-
mine feasibility and find control matchings for path-constrained dynamic optimization
problems. Since the structural criteria do not work for all problems, there are dy-
namic optimization problems which according to these structural criteria are feasible
and/or have a valid control matching which are in fact unsolvable. Also, structural
criteria do not indicate the ‘best’ control matching for those problems where there
are several possible control matchings, and solution of such problems requires further
research. However, these difficulties are related to the problem of nonlinear control-
lability, which has been an outstanding problem in the literature for many years.

The algorithm for numerical sensitivity analysis described in this thesis has sig-
nificantly improved the computational efficiency of numerical solution of dynamic
optimization problems. However, a number of interesting research issues remain.

319
The size of the dynamic optimization problem solvable with the control parameter-
ization method is limited by the number of parameters in the master NLP, rather
than by the number of state variables in the DAE. Numerical solution and sensitivity
analysis of sparse DAEs containing 10,000-100,000 variables is possible with current
technology, but solution of relatively large (1,000-10,000 parameters) dense master
NLPs is problematic. Such master NLPs arise in dynamic optimization problems
with large numbers of controls in the plant-wide model, higher-order basis functions,
or many finite elements. Furthermore, in such problems partial derivatives of the
point constraints with respect to all of the parameters can exist, which means that
strategies such as range and null space decomposition SQP will not work. Therefore,
methods are needed which can solve large dense NLPs for which the explicit func-
tional form of the constraints and objective is not known. A further major challenge
is that these large-scale master NLPs are often multi-modal [8].

Another interesting avenue of research is developing parallel algorithms for solu-

tion of dynamic optimization problems. At the top level, careful problem formulation
(e.g., decoupling of batch units through intermediate storage) can lead to a number
of independent IVP subproblems that can be solved in parallel. Within each IVP
subproblem, the sensitivity equations can be solved in parallel [77, 94]. Finally, at
the lowest level parallel algorithms could be used for the linear algebra and parallel
automatic differentiation techniques could be used for the function and derivative
evaluations.

This thesis has shown that inequality path-constrained dynamic optimization

problems may be posed as either MIDOs or as hybrid dynamic optimization prob-
lems. Solution of both types of problems is currently an active area of research, and
both have applications other than inequality path-constrained dynamic optimization
problems. For example, operating procedure synthesis problems can also be posed
as either MIDOs or hybrid dynamic optimization problems. In operating procedure
synthesis, it is desired to find the optimal controls that transform a process from one
state to another subject to economic, environmental, and safety constraints. The de-
cision variables are typically both discrete (e.g., open or close valves) and continuous

320
(e.g., controller setpoints).
Hybrid dynamic optimization problems are difficult to solve because they often
cause the master NLP to be nonsmooth, which creates problems with gradient-based
optimization methods. Sub-gradient optimization techniques are capable of handling
some classes of nonsmooth optimization problems, and it may be possible to find a
method that transforms the inequality path-constrained dynamic optimization prob-
lem into a hybrid problem that has an NLP of the appropriate form. On the other
hand, stochastic or random search methods are relatively insensitive to nonsmooth-
ness, so these methods offer an immediate and practical way to make sequencing de-
cisions in dynamic optimization problems. However, a more sophisticated approach
that exploits the evident structure of such problems (see Chapter 4) would ultimately
be more desirable.
The difficulty with solution of MIDO problems is finding a ‘link’ between the con-
tinuous and discrete aspects of the problem. That is, there is no satisfactory method
for determining how to vary the discrete variables based on information obtained from
the solution to the continuous problem. Several approaches have been attempted thus
far. In [102, 101], an attempt was made to solve MIDOs by using orthogonal colloca-
tion on finite elements to transform the MIDO into a finite-dimensional mixed integer
nonlinear program (MINLP). More recently, in [103] an analog of the control param-
eterization method for ordinary dynamic optimization problems was proposed. In
this scheme, numerical integration is used to obtain gradient and objective function
information for a master MINLP. It is shown that the adjoint variables of the dy-
namic optimization problem can be used to find the dual information necessary for
MINLP algorithms. However, the problem with these approaches is that the NLPs
are nonconvex, and therefore general methods for solving MINLPs (for a review, see
[52]) do not work [3] (in fact, the methods may converge to points that are not even
local extrema).

321
322
Appendix A

ABACUSS input file for reactor

and column example

DECLARE

TYPE # Default Min Max UNITS

Concentration = 0 : -1e20 : 1E20 unit = "mol/m^3"
Energy = 0 : -1E20 : 1E20 unit = "KJ"
EnergyFlowRate = 0 : -1E20 : 1E20 unit = "KJ/hr"
MolarEnthalpy = 1 : -1E20 : 1E20 unit = "KJ/mol"
LiquidMolarVolume = 20 : -1E20 : 1E20 unit = "cm3/mol"
VaporMolarVolume = .0224: 0 : 1E20 unit = "m3/mol"
MassFlowRate = 1 : 0 : 1E20 unit = "g/s"
KgMassFlowRate = 1 : 0 : 1E20 unit = "kg/s"
MolarVolume = .01 : 0 : 1E20 unit = "m^3/mol"
Fraction = .5 : -1E20 : 1E20 unit = "dimensionless"
MolarFlowRate = 0 : -1e20 : 1E20 unit = "mol/s"
Length = .5 : -1E20 : 1E20 unit = "m"
MolarHoldup = 1 : -1e20 : 1E20 unit = "mol"
PositiveValue = .75 : 0 : 1E20 unit = "dimensionless"
Pressure = 1.013: 0 : 1E20 unit = "bar"
Power = 0 : -1E20 : 1E20 unit = "GJ/s"
ReactionRate = 0 : -1e20 : 1E20 unit = "mol/m3/s"
ReducedQuantity = .5 : 0 : 1E20 unit = "dimensionless"
SmallValue = .02 : -10 : 1E20 unit = "dimensionless"
Temperature = 298 : 0 : 1E20 unit = "K"
Value = 4 : -1e20 : 1e20 unit = "dimensionless"
Velocity = 2 : -1e20 : 1e20 unit = "m/s"
Volume = 1 : 0 : 1E20 unit = "m^3"
VolumeFlowRate = 1e-9 : 0 : 1E20 unit = "m^3/hr"
PosValue = 1 : 1e-2 : 1e40 unit = "dimensionless"
Molar_Concentration = 1.0 : -1E-2: 1.0e20 unit = "mol/m3"
Reaction_Rate = 0.0 : -1e-2: 1e20 unit = "mol/(m3 s)"

323
 WaterFlowRate = 0 : -1e-2 : 1E20 unit = "ton/hr"

END # Declarations
#===============================================================================

MODEL LiquidEnthalpy

PARAMETER
NC AS INTEGER # number of components
RGAS AS REAL
TREF AS REAL
ENTHA AS ARRAY(NC) OF REAL
ENTHB AS ARRAY(NC) OF REAL
ENTHC AS ARRAY(NC) OF REAL
ENTHD AS ARRAY(NC) OF REAL
HVAP AS ARRAY(NC) OF REAL
Heat_Formation AS ARRAY(NC) OF REAL

VARIABLE
Specific_Enthalpy_Liquid AS ARRAY(NC) OF MolarEnthalpy
Specific_Enthalpy_Vapor AS ARRAY(NC) OF MolarEnthalpy
Temp AS Temperature

EQUATION
Specific_Enthalpy_Liquid=
Specific_Enthalpy_Vapor-HVAP ;

Specific_Enthalpy_Vapor=Heat_Formation+
ENTHA*(Temp-TREF)+ENTHB*(Temp-TREF)^2+
(1/2)*ENTHC*(Temp-TREF)^3+(1/3)*ENTHD*(Temp-TREF)^4 ;

END #LiquidEnthalpy

#===============================================================================

MODEL RD1_Reactions

PARAMETER
A, B, C, D AS INTEGER # identifiers for the components
NR AS INTEGER # number of reactions
NC AS INTEGER # number of components
MOLAR_VOLUME AS ARRAY(NC) of REAL
PRE_EXP_FACTOR AS ARRAY(NR) of REAL
RGAS AS REAL
ACTIVATION_ENERGY AS ARRAY(NR) of REAL
STOICH_COEFF AS ARRAY(NC,NR) of INTEGER
#Enthalpy_Reaction AS ARRAY(NR) OF REAL
HVAP AS ARRAY(NC) OF REAL

VARIABLE
Concentration AS ARRAY(NC) of Molar_Concentration
no_Mols AS ARRAY(NC) of MolarHoldup
ReactionRate AS ARRAY(NR) of Reaction_Rate
temp AS Temperature

324
 feed_A AS MolarFlowRate
feed_B AS MolarFlowRate
feed_C AS MolarFlowRate
feed_D AS MolarFlowRate
TotalFeed AS MolarFlowRate
FlowOut AS MolarFlowRate

volume AS Volume

QJacket AS EnergyFlowRate

TotalMols AS MolarHoldup
X AS ARRAY(NC) OF Fraction
Enthalpy AS Energy
Specific_Enthalpy AS ARRAY(NC) OF Energy
FeedEnthalpy AS ARRAY(NC) OF Energy
FeedTemp AS Temperature

EQUATION

# Material Balances for feed species [kmols]

$no_Mols(A) = feed_A + volume * SIGMA(STOICH_COEFF(A,) * ReactionRate)
- FlowOut*X(A);
$no_Mols(B) = feed_B + volume *SIGMA(STOICH_COEFF(B,) * ReactionRate)
- FlowOut*X(B);
$no_Mols(C) = feed_C + volume *SIGMA(STOICH_COEFF(C,) * ReactionRate)
- FlowOut*X(C);
$no_Mols(D) = feed_D + volume *SIGMA(STOICH_COEFF(D,) * ReactionRate)
- FlowOut*X(D);

TotalFeed = feed_A + feed_B + feed_C + feed_D ;

# Define the reaction rates for each reaction

ReactionRate(1)
= PRE_EXP_FACTOR(1)*EXP(-ACTIVATION_ENERGY(1)/(RGAS*temp))
*Concentration(A)*Concentration(B) ;

ReactionRate(2)
= PRE_EXP_FACTOR(2)*EXP(-ACTIVATION_ENERGY(2)/(RGAS*temp))
*Concentration(B)*Concentration(C) ;

# Define the the volume and the concentrations

volume = SIGMA(MOLAR_VOLUME*no_mols) ;
Concentration*volume = no_mols ;

TotalMols=SIGMA(no_Mols) ;

X=no_Mols/TotalMols ;

Enthalpy=SIGMA(No_Mols*Specific_Enthalpy);

$enthalpy = (FeedEnthalpy(A)+FeedEnthalpy(B)+FeedEnthalpy(C)+FeedEnthalpy(D)

-FlowOut*SIGMA(X*Specific_Enthalpy)

325
 -QJacket);

END # RD1_Reactions

#===============================================================================

MODEL Reactor_Jacket

PARAMETER
Heat_Transfer_Coeff AS REAL
Jacket_Area AS REAL
CP_water AS REAL
Twater_in AS REAL

VARIABLE
QJacket AS EnergyFlowRate
TWater_out AS Temperature
Temp_Reactor AS Temperature
Flow_Water AS WaterFlowRate

EQUATION
#Heat removed by jacket
QJacket=Heat_Transfer_Coeff*Jacket_Area*(Temp_Reactor-Twater_out) ;

#Outlet temperature of water

QJacket=CP_water*(Twater_out-Twater_in)*Flow_Water ;
END

MODEL ColumnVLE

PARAMETER
NSTAGE AS INTEGER
NC AS INTEGER
VPA AS ARRAY(NC) OF REAL
VPB AS ARRAY(NC) OF REAL
A, B, C, D AS INTEGER # identifiers for the components
TREF AS REAL
ENTHA AS ARRAY(NC) OF REAL
ENTHB AS ARRAY(NC) OF REAL
ENTHC AS ARRAY(NC) OF REAL
ENTHD AS ARRAY(NC) OF REAL
HVAP AS ARRAY(NC) OF REAL

VARIABLE
X AS ARRAY(NSTAGE,NC) OF Fraction
Y AS ARRAY(NSTAGE,NC) OF Fraction
K AS ARRAY(NSTAGE,NC) OF Value
VAPORENTHALPY AS ARRAY(NSTAGE) OF MolarEnthalpy
LIQUIDENTHALPY AS ARRAY(NSTAGE) OF MolarEnthalpy
VAPORPRESSURE AS ARRAY(NSTAGE,NC) OF Pressure
P AS ARRAY(NSTAGE) OF Pressure
Temp AS ARRAY(NSTAGE) OF Temperature
Specific_Enthalpy_Liquid AS ARRAY(NSTAGE,NC) OF MolarEnthalpy
Specific_Enthalpy_Vapor AS ARRAY(NSTAGE,NC) OF MolarEnthalpy

326
 EQUATION
# Vapor Pressure
FOR I:=1 TO NSTAGE DO
VAPORPRESSURE(I,)=EXP(-VPA/Temp(I)+VPB) ;
END

FOR I:=1 TO NSTAGE DO

Specific_Enthalpy_Liquid(I,)=
Specific_Enthalpy_Vapor(I,)-HVAP ;

Specific_Enthalpy_Vapor(I,)=
ENTHA*(Temp(I)-TREF)+ENTHB*(Temp(I)-TREF)^2+
(1/2)*ENTHC*(Temp(I)-TREF)^3+(1/3)*ENTHD*(Temp(I)-TREF)^4 ;

LIQUIDENTHALPY(I) = SIGMA(X(I,)*Specific_Enthalpy_Liquid(I,));
VAPORENTHALPY(I) = SIGMA(Y(I,)*Specific_Enthalpy_Vapor(I,));
END

# Ideal Phase equilibrium constant

FOR I:=1 TO NSTAGE DO
K(I,)=VAPORPRESSURE(I,)/P(I) ;
END

END

MODEL Column

PARAMETER
NSTAGE AS INTEGER
FEEDSTAGE AS INTEGER
NC AS INTEGER
A, B, C, D AS INTEGER # identifiers for the components
MOLAR_VOLUME AS ARRAY(NC) of REAL
MW AS ARRAY(NC) of REAL
STAGE_AREA AS REAL
WEIR_LENGTH AS REAL
WEIR_HEIGHT AS REAL
ORIFCON AS REAL
RGAS AS REAL
AFREE AS REAL
ALINE AS REAL
KLOSS AS REAL

VARIABLE
M AS ARRAY(NSTAGE) OF MolarHoldup
L AS ARRAY(NSTAGE) OF MolarFlowRate
V AS ARRAY(NSTAGE) OF MolarFlowRate
TotalHoldup AS MolarHoldup
VAPORENTHALPY AS ARRAY(NSTAGE) OF MolarEnthalpy
LIQUIDENTHALPY AS ARRAY(NSTAGE) OF MolarEnthalpy
X AS ARRAY(NSTAGE,NC) OF Fraction
Y AS ARRAY(NSTAGE,NC) OF Fraction
K AS ARRAY(NSTAGE,NC) OF Value

327
 DISTILLOUT AS MolarFlowRate
QR AS EnergyFlowRate
QC AS EnergyFlowRate
REFLUXRATIO AS Value
P AS ARRAY(NSTAGE) OF Pressure
DPSTAT AS ARRAY(NSTAGE) OF Pressure
DPDRY AS ARRAY(NSTAGE) OF Pressure
DPTRAY AS ARRAY(NSTAGE) OF Pressure
T AS ARRAY(NSTAGE) OF Temperature
FEED AS ARRAY(NSTAGE) OF MolarFlowRate
BOTTOMS AS MolarFlowRate
XFEED AS ARRAY(NC) OF Fraction
FEEDENTHALPY AS MolarEnthalpy
VOLUMEHOLDUP AS ARRAY(NSTAGE) OF VOLUME
HEAD AS ARRAY(NSTAGE) OF Length
LIQHEIGHT AS ARRAY(NSTAGE) OF Length
LIQDENS AS ARRAY(NSTAGE) OF Molar_Concentration
VAPDENS AS ARRAY(NSTAGE) OF Molar_Concentration
LIQMDENS AS ARRAY(NSTAGE) OF Molar_Concentration
VAPMDENS AS ARRAY(NSTAGE) OF Molar_Concentration
# FeedTemp AS Temperature
MWV AS ARRAY(NSTAGE) OF Value
MWL AS ARRAY(NSTAGE) OF Value
EQUATION

# Condenser Model

# Material Balance
V(2)=DISTILLOUT+L(1) ;
V(2)*REFLUXRATIO=L(1) ;
X(1,)=Y(2,) ;

# No holdup and no vapor

V(1)=0 ;
M(1)=0 ;

# Energy Balance
QC=V(2)*VAPORENTHALPY(2)-(L(1)+DISTILLOUT)*LIQUIDENTHALPY(1) ;

# Phase Equilibrium
Y(1,)=K(1,)*X(1,);
SIGMA(Y(1,))=1 ;

# Liquid and vapor densities

LIQDENS(1)=SIGMA(X(1,)/MOLAR_VOLUME);
VAPDENS(1)=(P(1)*100000/(RGAS*T(1)))/1000 ;
LIQMDENS(1)=SIGMA(X(1,)*MW/MOLAR_VOLUME);
VAPMDENS(1)=VAPDENS(1)*SIGMA(Y(1,)*MW) ;
MWV(1)=SIGMA(Y(1,)*MW);
MWL(1)=SIGMA(X(1,)*MW);

# Relate liquid overflow rate to holdup

VOLUMEHOLDUP(1)=M(1)/LIQDENS(1);
LIQHEIGHT(1)=0;

328
 HEAD(1)=0 ;
((V(2)*MWV(2)/3600)/VAPMDENS(1)*ALINE)^2=2.0*DPTRAY(1)/VAPMDENS(1)/KLOSS;

# No static or dry pressure loss

DPSTAT(1)=0;
DPDRY(1)=0;

# Tray Model
FOR I:=2 TO NSTAGE-1 DO

# Material Balance
$M(I)=L(I-1)-L(I)+V(I+1)-V(I)+FEED(I) ;
$M(I)*X(I,)+M(I)*$X(I,)=
L(I-1)*X(I-1,)-L(I)*X(I,)+
V(I+1)*Y(I+1,)-V(I)*Y(I,)+
FEED(I)*XFEED ;

# Energy Balance
#$M(I)*LIQUIDENTHALPY(I)+M(I)*$LIQUIDENTHALPY(I)=
0=
L(I-1)*LIQUIDENTHALPY(I-1)-L(I)*LIQUIDENTHALPY(I)+
V(I+1)*VAPORENTHALPY(I+1)-V(I)*VAPORENTHALPY(I)+
FEED(I)*FEEDENTHALPY;

# Phase Equilibrium
Y(I,)=K(I,)*X(I,);

# Normalize liquid mole fraction

SIGMA(Y(I,))=1;

# Liquid and vapor densities

LIQDENS(I)=SIGMA(X(I,)/MOLAR_VOLUME);
VAPDENS(I)=(P(I)*100000/(RGAS*T(I)))/1000 ;
LIQMDENS(I)=SIGMA(X(I,)*MW/MOLAR_VOLUME);
VAPMDENS(I)=VAPDENS(I)*SIGMA(Y(I,)*MW) ;
MWV(I)=SIGMA(Y(I,)*MW);
MWL(I)=SIGMA(X(I,)*MW);

# Relate liquid overflow rate to holdup

VOLUMEHOLDUP(I)=M(I)/LIQDENS(I);
LIQHEIGHT(I)=VOLUMEHOLDUP(I)/STAGE_AREA;
HEAD(I)=(LIQHEIGHT(I)-WEIR_HEIGHT)*1000 ;
L(I)*MWL(I)/3600=LIQMDENS(I)*WEIR_LENGTH*(HEAD(I)/750)^1.5 ;

# Relate vapor flow to plate pressure drop

DPSTAT(I)=LIQHEIGHT(I)*9.81*LIQMDENS(I)/100000;
DPDRY(I)=(1/100000)*(0.50031*VAPMDENS(I))*
((V(I+1)*MWV(I)/3600)/VAPMDENS(I)*AFREE*ORIFCON)^2 ;
DPTRAY(I)=DPSTAT(I)+DPDRY(I)+(12.5*9.81)/100000 ;
P(I)=P(I-1)+DPTRAY(I-1) ;
END

#Reboiler Model
# Material Balance

329
 $M(NSTAGE)=L(NSTAGE-1)-L(NSTAGE)-V(NSTAGE) ;
BOTTOMS=L(NSTAGE) ;
$M(NSTAGE)*X(NSTAGE,)+M(NSTAGE)*$X(NSTAGE,)=
L(NSTAGE-1)*X(NSTAGE-1,)-L(NSTAGE)*X(NSTAGE,)-
V(NSTAGE)*Y(NSTAGE,) ;

# Energy Balance
#$M(NSTAGE)*LIQUIDENTHALPY(NSTAGE)+M(NSTAGE)*$LIQUIDENTHALPY(NSTAGE)=
0=
L(NSTAGE-1)*LIQUIDENTHALPY(NSTAGE-1)-L(NSTAGE)*LIQUIDENTHALPY(NSTAGE)
-V(NSTAGE)*VAPORENTHALPY(NSTAGE)+QR;

# Phase Equilibrium
Y(NSTAGE,)=K(NSTAGE,)*X(NSTAGE,);
SIGMA(Y(NSTAGE,))=SIGMA(X(NSTAGE,));

# Liquid and vapor densities

LIQDENS(NSTAGE)=SIGMA(X(NSTAGE,)/MOLAR_VOLUME);
VAPDENS(NSTAGE)=(P(NSTAGE)*100000/(RGAS*T(NSTAGE)))/1000 ;
LIQMDENS(NSTAGE)=SIGMA(X(NSTAGE,)*MW/MOLAR_VOLUME);
VAPMDENS(NSTAGE)=VAPDENS(NSTAGE)*SIGMA(Y(NSTAGE,)*MW) ;
MWV(NSTAGE)=SIGMA(Y(NSTAGE,)*MW);
MWL(NSTAGE)=SIGMA(X(NSTAGE,)*MW);

# Relate liquid overflow rate to holdup

VOLUMEHOLDUP(NSTAGE)=M(NSTAGE)/LIQDENS(NSTAGE);
LIQHEIGHT(NSTAGE)=0;
HEAD(NSTAGE)=0 ;

# Relate vapor flow to plate pressure drop

DPSTAT(NSTAGE)=0;
DPDRY(NSTAGE)=0;
DPTRAY(NSTAGE)=0 ;
P(NSTAGE)=P(NSTAGE-1)+DPTRAY(NSTAGE-1) ;

#Total Column holdup

TotalHoldup=SIGMA(M);

END # Column

MODEL Flowsheet

PARAMETER
A, B, C, D AS INTEGER # identifiers for the components
NR AS INTEGER # number of reactions
NC AS INTEGER # number of components
MOLAR_VOLUME AS ARRAY(NC) of REAL
MW AS ARRAY(NC) of REAL
PRE_EXP_FACTOR AS ARRAY(NR) of REAL
RGAS AS REAL
ACTIVATION_ENERGY AS ARRAY(NR) of REAL
STOICH_COEFF AS ARRAY(NC,NR) of INTEGER
#Enthalpy_Reaction AS ARRAY(NR) OF REAL
Heat_Formation AS ARRAY(NC) OF REAL

330
 Heat_Transfer_Coeff AS REAL
Jacket_Area AS REAL
CP_water AS REAL
Twater_in AS REAL

TREF AS REAL
ENTHA AS ARRAY(NC) OF REAL
ENTHB AS ARRAY(NC) OF REAL
ENTHC AS ARRAY(NC) OF REAL
ENTHD AS ARRAY(NC) OF REAL
HVAP AS ARRAY(NC) OF REAL

NSTAGE AS INTEGER
FEEDSTAGE AS INTEGER
STAGE_AREA AS REAL
WEIR_LENGTH AS REAL
WEIR_HEIGHT AS REAL

VPA AS ARRAY(NC) OF REAL

VPB AS ARRAY(NC) OF REAL
AFREE AS REAL
ORIFCON AS REAL
ALINE AS REAL
KLOSS AS REAL

SET
NR := 2 ;
NC := 4 ;

A := 1;
B := 2;
C := 3;
D := 4;

MOLAR_VOLUME(A) := 1.0/11.0 ;
MOLAR_VOLUME(B) := 1.0/16.0 ;
MOLAR_VOLUME(C) := 1.0/10.4 ;
MOLAR_VOLUME(D) := 1.0/10.0 ;

MW(A):= 76 ;
MW(B):= 52 ;
MW(C):= 170 ;
MW(D):= 264 ;

PRE_EXP_FACTOR(1) := 100.0 ;
PRE_EXP_FACTOR(2) := 120.0 ;

ACTIVATION_ENERGY(1) := 17000 ;
ACTIVATION_ENERGY(2) := 20000 ;

RGAS := 8.314 ;

331
 # R1 R2
STOICH_COEFF(A,) := [-1, 0] ;
STOICH_COEFF(B,) := [-1, -1] ;
STOICH_COEFF(C,) := [ 1, -1] ;
STOICH_COEFF(D,) := [ 0, 1] ;

# Enthalpy_Reaction(1) := -60000 ;
# Enthalpy_Reaction(2) := -50000 ;

Heat_Formation(A) := -30000;
Heat_Formation(B) := -50000;
Heat_Formation(C) := -20000;
Heat_Formation(D) := -20000;

Heat_Transfer_Coeff := 3000 ;
Jacket_Area := 30 ;
CP_water := 4200 ;
Twater_in := 298 ;

TREF := 298 ;
ENTHA(A) := 172.3 ;
ENTHA(B) := 200.0 ;
ENTHA(C) := 160.0 ;
ENTHA(D) := 155.0 ;

ENTHB(1:NC) := 0.0 ;
ENTHC(1:NC) := 0.0 ;
ENTHD(1:NC) := 0.0 ;

HVAP(A) := 31000 ;
HVAP(B) := 26000 ;
HVAP(C) := 28000 ;
HVAP(D) := 34000 ;

NSTAGE := 10 ;
FEEDSTAGE := 5 ;
STAGE_AREA := 1.05;
AFREE := 0.1*STAGE_AREA ;
ALINE := .03;
KLOSS := 1 ;
ORIFCON :=0.6 ;
WEIR_LENGTH :=1 ;
WEIR_HEIGHT :=0.25;

VPA(A) := 4142.75;
VPA(B) := 3474.56;
VPA(C) := 3500;
VPA(D) := 4543.71;

VPB(A) := 11.7158;
VPB(B) := 9.9404;
VPB(C) := 8.9;
VPB(D) := 11.2599;

332
END #Flowsheet

MODEL ReactorFlowsheet INHERITS Flowsheet

UNIT
Reactor AS RD1_Reactions
Jacket AS Reactor_Jacket
LiqEnthalpy AS LiquidEnthalpy
ReactorFeedEnthalpy AS LiquidEnthalpy

EQUATION
Reactor.Temp IS Jacket.Temp_Reactor ;
Reactor.Temp IS LiqEnthalpy.Temp ;
Reactor.Specific_Enthalpy IS LiqEnthalpy.Specific_Enthalpy_Liquid ;
Reactor.QJacket IS Jacket.QJacket ;
Reactor.FeedTemp IS ReactorFeedEnthalpy.Temp ;
Reactor.FeedEnthalpy IS ReactorFeedEnthalpy.Specific_Enthalpy_Liquid ;
Reactor.$TotalMols=0 ;
Reactor.Feed_A=Reactor.Feed_B;
END # ReactorFlowsheet

MODEL ColumnFlowsheet INHERITS Flowsheet

UNIT
Column AS Column
VLE AS ColumnVLE
ColumnFeedEnthalpy AS LiquidEnthalpy

EQUATION
Column.X IS VLE.X ;
Column.Y IS VLE.Y ;
Column.K IS VLE.K ;
Column.T IS VLE.Temp ;
Column.P IS VLE.P ;
Column.VAPORENTHALPY IS VLE.VAPORENTHALPY ;
Column.LIQUIDENTHALPY IS VLE.LIQUIDENTHALPY ;
#Column.$TotalHoldup=0 ;
#Column.TotalHoldup=43.911382-17.325 ;
Column.BOTTOMS=0.1*(Column.M(Column.NSTAGE)-22) ;
#Column.BOTTOMS=0.1*Column.M(Column.NSTAGE) ;
#Column.FeedTemp IS ColumnFeedEnthalpy.Temp;
Column.FeedEnthalpy=
SIGMA(Column.XFEED*ColumnFeedEnthalpy.Specific_Enthalpy_Liquid)*
SIGMA(Column.FEED);
Column.$TotalHoldup=0 ;
END # ColumnFlowsheet

MODEL BothFlowsheet INHERITS Flowsheet

UNIT
Column AS Column
VLE AS ColumnVLE
Reactor AS RD1_Reactions
Jacket AS Reactor_Jacket
LiqEnthalpy AS LiquidEnthalpy

333
 MakeupEnthalpy AS LiquidEnthalpy

VARIABLE
BFraction AS Fraction
Makeup AS ARRAY(NC) OF MolarFlowRate

EQUATION
WITHIN Column DO
X IS VLE.X ;
Y IS VLE.Y ;
K IS VLE.K ;
T IS VLE.Temp ;
P IS VLE.P ;
VAPORENTHALPY IS VLE.VAPORENTHALPY ;
LIQUIDENTHALPY IS VLE.LIQUIDENTHALPY ;
BOTTOMS=0.1*Column.M(Column.NSTAGE);
$TotalHoldup=0 ;
END

WITHIN Reactor DO
Temp = Jacket.Temp_Reactor+0 ;
Temp= LiqEnthalpy.Temp+0 ;
Specific_Enthalpy IS LiqEnthalpy.Specific_Enthalpy_Liquid ;
QJacket IS Jacket.QJacket ;
#FeedTemp IS ReactorFeedEnthalpy.Temp ;
#FeedEnthalpy IS ReactorFeedEnthalpy.Specific_Enthalpy_Liquid ;
FeedTemp IS MakeupEnthalpy.Temp;
FeedEnthalpy =
(Column.DISTILLOUT*Column.X(1,)*VLE.SPECIFIC_ENTHALPY_LIQUID(1,)+
Makeup*MakeupEnthalpy.Specific_Enthalpy_Liquid);
$TotalMols=0 ;
Feed_B=BFraction*Feed_A;
# Need to change 0.01 in the figure below to 1
Feed_A=Column.DISTILLOUT*Column.X(1,A)+Makeup(A);
Feed_B=Column.DISTILLOUT*Column.X(1,B)+Makeup(B);
Feed_C=Column.DISTILLOUT*Column.X(1,C);
Feed_D=Column.DISTILLOUT*Column.X(1,D);
Makeup(C)=0 ;
Makeup(D)=0 ;
END

WITHIN Column DO
# FeedTemp IS Reactor.Temp;
FeedEnthalpy=
SIGMA(XFEED*Reactor.Specific_Enthalpy)*SIGMA(FEED);
XFEED IS Reactor.X ;
FEED(1)=0 ;
FOR I:=2 TO FEEDSTAGE-1 DO
FEED(I)=0 ;
END
FEED(FEEDSTAGE)= Reactor.FlowOut;
FOR I:=FEEDSTAGE+1 TO NSTAGE-1 DO
FEED(I)=0 ;
END

334
 FEED(NSTAGE)=0 ;
END

END # BothFlowsheet

SIMULATION RunReactor

UNIT
System As ReactorFlowSheet

INPUT
WITHIN System.Reactor DO
FlowOut := 10.0 ;
FeedTemp := 300 ;
Feed_C := 0.0 ;
Feed_D := 0.0 ;
END
WITHIN System.Jacket DO
Flow_Water := 1 ;
END
INITIAL
WITHIN System.Reactor DO
No_Mols(A) = 22.27 ;
No_Mols(B) = 36.81 ;
No_Mols(C) = 0.0 ;
No_Mols(D) = 0.0 ;
Temp=300 ;
END

SCHEDULE
CONTINUE FOR 0

END

SIMULATION RunColumn

UNIT
System As ColumnFlowSheet

INPUT
WITHIN System.Column DO
P(1) :=1.01325 ;
REFLUXRATIO :=0.5 ;
FEED(1):=0 ;
FOR I:=2 TO FEEDSTAGE-1 DO
FEED(I):=0 ;
END
FEED(FEEDSTAGE):=(10/6+8*10/6) ;
FOR I:=FEEDSTAGE+1 TO NSTAGE-1 DO
FEED(I):=0 ;
END
FEED(NSTAGE):=0 ;
XFEED(1) := 1 ;
XFEED(2) := 0 ;

335
 XFEED(3) := 0 ;
XFEED(4) := 0 ;
# FEEDTEMP := 300 ;
#TotalHoldup:=3.6586382E+01+10 ;
END

PRESET
##############################
# Values of All Active Variables #
##############################

SYSTEM.COLUMNFEEDENTHALPY.TEMP := 3.0000000E+02 ;
SYSTEM.COLUMNFEEDENTHALPY.SPECIFIC_ENTHALPY_VAPOR(1) := 3.4460000E+02 ;
SYSTEM.COLUMNFEEDENTHALPY.SPECIFIC_ENTHALPY_VAPOR(2) := 4.0000000E+02 ;
SYSTEM.COLUMNFEEDENTHALPY.SPECIFIC_ENTHALPY_VAPOR(3) := 3.2000000E+02 ;
SYSTEM.COLUMNFEEDENTHALPY.SPECIFIC_ENTHALPY_VAPOR(4) := 3.1000000E+02 ;
SYSTEM.COLUMNFEEDENTHALPY.SPECIFIC_ENTHALPY_LIQUID(1) := -3.0655400E+04 ;
SYSTEM.COLUMNFEEDENTHALPY.SPECIFIC_ENTHALPY_LIQUID(2) := -2.5600000E+04 ;
SYSTEM.COLUMNFEEDENTHALPY.SPECIFIC_ENTHALPY_LIQUID(3) := -2.7680000E+04 ;
SYSTEM.COLUMNFEEDENTHALPY.SPECIFIC_ENTHALPY_LIQUID(4) := -3.3690000E+04 ;
SYSTEM.COLUMN.MWV(1) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(2) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(3) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(4) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(5) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(6) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(7) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(8) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(9) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(10) := 7.6000000E+01 ;
SYSTEM.COLUMN.BOTTOMS := 2.2005890E+00 ;
SYSTEM.COLUMN.VAPORENTHALPY(1) := 9.6490429E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(2) := 9.6493013E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(3) := 9.7613871E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(4) := 9.8713496E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(5) := 9.9792593E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(6) := 1.0094036E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(7) := 1.0206730E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(8) := 1.0317422E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(9) := 1.0426188E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(10) := 1.0533100E+04 ;
SYSTEM.COLUMN.LIQMDENS(1) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(2) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(3) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(4) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(5) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(6) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(7) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(8) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(9) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(10) := 8.3600000E+02 ;
SYSTEM.COLUMN.VAPMDENS(1) := 2.6164666E+00 ;
SYSTEM.COLUMN.VAPMDENS(2) := 2.6165852E+00 ;
SYSTEM.COLUMN.VAPMDENS(3) := 2.6684553E+00 ;

336
SYSTEM.COLUMN.VAPMDENS(4) := 2.7201430E+00 ;
SYSTEM.COLUMN.VAPMDENS(5) := 2.7716449E+00 ;
SYSTEM.COLUMN.VAPMDENS(6) := 2.8272809E+00 ;
SYSTEM.COLUMN.VAPMDENS(7) := 2.8827759E+00 ;
SYSTEM.COLUMN.VAPMDENS(8) := 2.9381325E+00 ;
SYSTEM.COLUMN.VAPMDENS(9) := 2.9933533E+00 ;
SYSTEM.COLUMN.VAPMDENS(10) := 3.0484406E+00 ;
SYSTEM.COLUMN.QR := 6.4072745E+06 ;
SYSTEM.COLUMN.DPSTAT(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.DPSTAT(2) := 2.0759991E-02 ;
SYSTEM.COLUMN.DPSTAT(3) := 2.0726226E-02 ;
SYSTEM.COLUMN.DPSTAT(4) := 2.0689946E-02 ;
SYSTEM.COLUMN.DPSTAT(5) := 2.2496062E-02 ;
SYSTEM.COLUMN.DPSTAT(6) := 2.2483939E-02 ;
SYSTEM.COLUMN.DPSTAT(7) := 2.2471846E-02 ;
SYSTEM.COLUMN.DPSTAT(8) := 2.2459782E-02 ;
SYSTEM.COLUMN.DPSTAT(9) := 2.2447747E-02 ;
SYSTEM.COLUMN.DPSTAT(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.DPTRAY(1) := 5.0229571E-05 ;
SYSTEM.COLUMN.DPTRAY(2) := 2.1986243E-02 ;
SYSTEM.COLUMN.DPTRAY(3) := 2.1952478E-02 ;
SYSTEM.COLUMN.DPTRAY(4) := 2.1916198E-02 ;
SYSTEM.COLUMN.DPTRAY(5) := 2.3722479E-02 ;
SYSTEM.COLUMN.DPTRAY(6) := 2.3710350E-02 ;
SYSTEM.COLUMN.DPTRAY(7) := 2.3698251E-02 ;
SYSTEM.COLUMN.DPTRAY(8) := 2.3686181E-02 ;
SYSTEM.COLUMN.DPTRAY(9) := 2.3674140E-02 ;
SYSTEM.COLUMN.DPTRAY(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(1) := -2.1350957E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(2) := -2.1350699E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(3) := -2.1238613E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(4) := -2.1128650E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(5) := -2.1020741E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(6) := -2.0905964E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(7) := -2.0793270E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(8) := -2.0682578E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(9) := -2.0573812E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(10) := -2.0466900E+04 ;
SYSTEM.COLUMN.LIQHEIGHT(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.LIQHEIGHT(2) := 2.5313481E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(3) := 2.5272311E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(4) := 2.5228073E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(5) := 2.7430342E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(6) := 2.7415560E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(7) := 2.7400814E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(8) := 2.7386104E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(9) := 2.7371429E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.HEAD(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.HEAD(2) := 3.1348103E+00 ;
SYSTEM.COLUMN.HEAD(3) := 2.7231056E+00 ;
SYSTEM.COLUMN.HEAD(4) := 2.2807310E+00 ;
SYSTEM.COLUMN.HEAD(5) := 2.4303419E+01 ;
SYSTEM.COLUMN.HEAD(6) := 2.4155598E+01 ;

337
SYSTEM.COLUMN.HEAD(7) := 2.4008142E+01 ;
SYSTEM.COLUMN.HEAD(8) := 2.3861041E+01 ;
SYSTEM.COLUMN.HEAD(9) := 2.3714286E+01 ;
SYSTEM.COLUMN.HEAD(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.K(1,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(1,2) := 1.1186637E+00 ;
SYSTEM.COLUMN.K(1,3) := 3.6783189E-01 ;
SYSTEM.COLUMN.K(1,4) := 2.0422134E-01 ;
SYSTEM.COLUMN.K(2,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(2,2) := 1.1186548E+00 ;
SYSTEM.COLUMN.K(2,3) := 3.6782906E-01 ;
SYSTEM.COLUMN.K(2,4) := 2.0422232E-01 ;
SYSTEM.COLUMN.K(3,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(3,2) := 1.1147884E+00 ;
SYSTEM.COLUMN.K(3,3) := 3.6660608E-01 ;
SYSTEM.COLUMN.K(3,4) := 2.0464705E-01 ;
SYSTEM.COLUMN.K(4,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(4,2) := 1.1110220E+00 ;
SYSTEM.COLUMN.K(4,3) := 3.6541455E-01 ;
SYSTEM.COLUMN.K(4,4) := 2.0506307E-01 ;
SYSTEM.COLUMN.K(5,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(5,2) := 1.1073514E+00 ;
SYSTEM.COLUMN.K(5,3) := 3.6425316E-01 ;
SYSTEM.COLUMN.K(5,4) := 2.0547070E-01 ;
SYSTEM.COLUMN.K(6,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(6,2) := 1.1034745E+00 ;
SYSTEM.COLUMN.K(6,3) := 3.6302637E-01 ;
SYSTEM.COLUMN.K(6,4) := 2.0590358E-01 ;
SYSTEM.COLUMN.K(7,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(7,2) := 1.0996952E+00 ;
SYSTEM.COLUMN.K(7,3) := 3.6183030E-01 ;
SYSTEM.COLUMN.K(7,4) := 2.0632791E-01 ;
SYSTEM.COLUMN.K(8,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(8,2) := 1.0960090E+00 ;
SYSTEM.COLUMN.K(8,3) := 3.6066354E-01 ;
SYSTEM.COLUMN.K(8,4) := 2.0674404E-01 ;
SYSTEM.COLUMN.K(9,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(9,2) := 1.0924118E+00 ;
SYSTEM.COLUMN.K(9,3) := 3.5952480E-01 ;
SYSTEM.COLUMN.K(9,4) := 2.0715229E-01 ;
SYSTEM.COLUMN.K(10,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(10,2) := 1.0888996E+00 ;
SYSTEM.COLUMN.K(10,3) := 3.5841286E-01 ;
SYSTEM.COLUMN.K(10,4) := 2.0755297E-01 ;
SYSTEM.COLUMN.L(1) := 1.2799411E+01 ;
SYSTEM.COLUMN.L(2) := 1.0700897E+01 ;
SYSTEM.COLUMN.L(3) := 8.6636325E+00 ;
SYSTEM.COLUMN.L(4) := 6.6407080E+00 ;
SYSTEM.COLUMN.L(5) := 2.3099594E+02 ;
SYSTEM.COLUMN.L(6) := 2.2889167E+02 ;
SYSTEM.COLUMN.L(7) := 2.2679899E+02 ;
SYSTEM.COLUMN.L(8) := 2.2471775E+02 ;
SYSTEM.COLUMN.L(9) := 2.2264778E+02 ;
SYSTEM.COLUMN.L(10) := 2.2005890E+00 ;

338
SYSTEM.COLUMN.$M(2) := 1.2242624E-18 ;
SYSTEM.COLUMN.$M(3) := -4.6540611E-19 ;
SYSTEM.COLUMN.$M(4) := -5.9937802E-19 ;
SYSTEM.COLUMN.$M(5) := 1.1802807E-18 ;
SYSTEM.COLUMN.$M(6) := 1.9926115E-19 ;
SYSTEM.COLUMN.$M(7) := -3.8043140E-19 ;
SYSTEM.COLUMN.$M(8) := -8.5739929E-19 ;
SYSTEM.COLUMN.$M(9) := -4.9483541E-19 ;
SYSTEM.COLUMN.M(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.M(2) := 2.9237071E+00 ;
SYSTEM.COLUMN.M(3) := 2.9189519E+00 ;
SYSTEM.COLUMN.M(4) := 2.9138424E+00 ;
SYSTEM.COLUMN.M(5) := 3.1682045E+00 ;
SYSTEM.COLUMN.M(6) := 3.1664972E+00 ;
SYSTEM.COLUMN.M(7) := 3.1647940E+00 ;
SYSTEM.COLUMN.M(8) := 3.1630950E+00 ;
SYSTEM.COLUMN.M(9) := 3.1614000E+00 ;
SYSTEM.COLUMN.M(10) := 2.2005890E+01 ;
SYSTEM.COLUMN.LIQDENS(1) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(2) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(3) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(4) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(5) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(6) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(7) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(8) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(9) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(10) := 1.1000000E+01 ;
SYSTEM.COLUMN.VAPDENS(1) := 3.4427192E-02 ;
SYSTEM.COLUMN.VAPDENS(2) := 3.4428753E-02 ;
SYSTEM.COLUMN.VAPDENS(3) := 3.5111254E-02 ;
SYSTEM.COLUMN.VAPDENS(4) := 3.5791355E-02 ;
SYSTEM.COLUMN.VAPDENS(5) := 3.6469012E-02 ;
SYSTEM.COLUMN.VAPDENS(6) := 3.7201064E-02 ;
SYSTEM.COLUMN.VAPDENS(7) := 3.7931261E-02 ;
SYSTEM.COLUMN.VAPDENS(8) := 3.8659638E-02 ;
SYSTEM.COLUMN.VAPDENS(9) := 3.9386227E-02 ;
SYSTEM.COLUMN.VAPDENS(10) := 4.0111061E-02 ;
SYSTEM.COLUMN.P(2) := 1.0133002E+00 ;
SYSTEM.COLUMN.P(3) := 1.0352865E+00 ;
SYSTEM.COLUMN.P(4) := 1.0572390E+00 ;
SYSTEM.COLUMN.P(5) := 1.0791551E+00 ;
SYSTEM.COLUMN.P(6) := 1.1028776E+00 ;
SYSTEM.COLUMN.P(7) := 1.1265880E+00 ;
SYSTEM.COLUMN.P(8) := 1.1502862E+00 ;
SYSTEM.COLUMN.P(9) := 1.1739724E+00 ;
SYSTEM.COLUMN.P(10) := 1.1976465E+00 ;
SYSTEM.COLUMN.DISTILLOUT := 1.2799411E+01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(2) := 2.6579155E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(3) := 2.6535926E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(4) := 2.6489477E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(5) := 2.8801859E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(6) := 2.8786338E-01 ;

339
SYSTEM.COLUMN.VOLUMEHOLDUP(7) := 2.8770855E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(8) := 2.8755409E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(9) := 2.8740000E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(10) := 2.0005354E+00 ;
SYSTEM.COLUMN.T(1) := 3.5400141E+02 ;
SYSTEM.COLUMN.T(2) := 3.5400291E+02 ;
SYSTEM.COLUMN.T(3) := 3.5465344E+02 ;
SYSTEM.COLUMN.T(4) := 3.5529164E+02 ;
SYSTEM.COLUMN.T(5) := 3.5591793E+02 ;
SYSTEM.COLUMN.T(6) := 3.5658408E+02 ;
SYSTEM.COLUMN.T(7) := 3.5723813E+02 ;
SYSTEM.COLUMN.T(8) := 3.5788057E+02 ;
SYSTEM.COLUMN.T(9) := 3.5851183E+02 ;
SYSTEM.COLUMN.T(10) := 3.5913233E+02 ;
SYSTEM.COLUMN.V(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.V(2) := 2.5598822E+01 ;
SYSTEM.COLUMN.V(3) := 2.3500308E+01 ;
SYSTEM.COLUMN.V(4) := 2.1463044E+01 ;
SYSTEM.COLUMN.V(5) := 1.9440119E+01 ;
SYSTEM.COLUMN.V(6) := 2.2879535E+02 ;
SYSTEM.COLUMN.V(7) := 2.2669108E+02 ;
SYSTEM.COLUMN.V(8) := 2.2459840E+02 ;
SYSTEM.COLUMN.V(9) := 2.2251716E+02 ;
SYSTEM.COLUMN.V(10) := 2.2044719E+02 ;
SYSTEM.COLUMN.DPDRY(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.DPDRY(2) := 1.8679046E-09 ;
SYSTEM.COLUMN.DPDRY(3) := 1.5277952E-09 ;
SYSTEM.COLUMN.DPDRY(4) := 1.2295567E-09 ;
SYSTEM.COLUMN.DPDRY(5) := 1.6714726E-07 ;
SYSTEM.COLUMN.DPDRY(6) := 1.6085789E-07 ;
SYSTEM.COLUMN.DPDRY(7) := 1.5486202E-07 ;
SYSTEM.COLUMN.DPDRY(8) := 1.4914136E-07 ;
SYSTEM.COLUMN.DPDRY(9) := 1.4367911E-07 ;
SYSTEM.COLUMN.DPDRY(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(1,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.X(1,2) := -1.4211639E-30 ;
SYSTEM.COLUMN.X(1,3) := -7.7119210E-33 ;
SYSTEM.COLUMN.X(1,4) := -6.9770011E-33 ;
SYSTEM.COLUMN.X(2,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(2,1) := -1.1837221E-27 ;
SYSTEM.COLUMN.X(2,2) := -1.2704223E-30 ;
SYSTEM.COLUMN.$X(2,2) := 5.7989273E-33 ;
SYSTEM.COLUMN.X(2,3) := -2.0966046E-32 ;
SYSTEM.COLUMN.$X(2,3) := -2.6702840E-35 ;
SYSTEM.COLUMN.X(2,4) := -6.1284288E-32 ;
SYSTEM.COLUMN.$X(2,4) := -1.6449487E-34 ;
SYSTEM.COLUMN.X(3,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(3,1) := 1.2030481E-26 ;
SYSTEM.COLUMN.X(3,2) := -1.2080724E-30 ;
SYSTEM.COLUMN.$X(3,2) := 5.5717889E-33 ;
SYSTEM.COLUMN.X(3,3) := -3.8197066E-32 ;
SYSTEM.COLUMN.$X(3,3) := -1.0199658E-34 ;
SYSTEM.COLUMN.X(3,4) := -8.1682874E-32 ;
SYSTEM.COLUMN.$X(3,4) := -2.0340822E-34 ;

340
SYSTEM.COLUMN.X(4,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(4,1) := -1.3084953E-26 ;
SYSTEM.COLUMN.X(4,2) := -1.0440124E-30 ;
SYSTEM.COLUMN.$X(4,2) := 5.5665390E-33 ;
SYSTEM.COLUMN.X(4,3) := -6.9179846E-32 ;
SYSTEM.COLUMN.$X(4,3) := -1.7839640E-34 ;
SYSTEM.COLUMN.X(4,4) := -6.5639146E-32 ;
SYSTEM.COLUMN.$X(4,4) := -1.6350563E-34 ;
SYSTEM.COLUMN.X(5,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(5,1) := -1.9839325E-24 ;
SYSTEM.COLUMN.X(5,2) := -9.9188031E-31 ;
SYSTEM.COLUMN.$X(5,2) := 5.6461486E-33 ;
SYSTEM.COLUMN.X(5,3) := -9.4816353E-32 ;
SYSTEM.COLUMN.$X(5,3) := -2.4050174E-34 ;
SYSTEM.COLUMN.X(5,4) := -4.3865365E-33 ;
SYSTEM.COLUMN.$X(5,4) := -1.1121583E-35 ;
SYSTEM.COLUMN.X(6,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(6,1) := 7.4938318E-24 ;
SYSTEM.COLUMN.X(6,2) := -8.9965879E-31 ;
SYSTEM.COLUMN.$X(6,2) := 4.9062142E-33 ;
SYSTEM.COLUMN.X(6,3) := -2.6625428E-31 ;
SYSTEM.COLUMN.$X(6,3) := -6.7514852E-34 ;
SYSTEM.COLUMN.X(6,4) := -1.2628664E-32 ;
SYSTEM.COLUMN.$X(6,4) := -3.2429701E-35 ;
SYSTEM.COLUMN.X(7,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(7,1) := -2.4396441E-24 ;
SYSTEM.COLUMN.X(7,2) := -8.2074687E-31 ;
SYSTEM.COLUMN.$X(7,2) := 4.4714835E-33 ;
SYSTEM.COLUMN.X(7,3) := -1.0224736E-30 ;
SYSTEM.COLUMN.$X(7,3) := -1.8627965E-33 ;
SYSTEM.COLUMN.X(7,4) := -5.2858763E-32 ;
SYSTEM.COLUMN.$X(7,4) := -1.3644363E-34 ;
SYSTEM.COLUMN.X(8,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(8,1) := 1.9593269E-24 ;
SYSTEM.COLUMN.X(8,2) := -7.5079031E-31 ;
SYSTEM.COLUMN.$X(8,2) := 4.0864429E-33 ;
SYSTEM.COLUMN.X(8,3) := -2.3610164E-30 ;
SYSTEM.COLUMN.$X(8,3) := -4.3205815E-33 ;
SYSTEM.COLUMN.X(8,4) := -2.4917836E-31 ;
SYSTEM.COLUMN.$X(8,4) := -6.4403095E-34 ;
SYSTEM.COLUMN.X(9,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(9,1) := -4.8662107E-23 ;
SYSTEM.COLUMN.X(9,2) := -6.8857471E-31 ;
SYSTEM.COLUMN.$X(9,2) := 3.7443424E-33 ;
SYSTEM.COLUMN.X(9,3) := -6.1456693E-30 ;
SYSTEM.COLUMN.$X(9,3) := -1.1199828E-32 ;
SYSTEM.COLUMN.X(9,4) := -1.2057443E-30 ;
SYSTEM.COLUMN.$X(9,4) := -3.1172520E-33 ;
SYSTEM.COLUMN.X(10,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(10,1) := 6.2732799E-24 ;
SYSTEM.COLUMN.X(10,2) := -6.3306961E-31 ;
SYSTEM.COLUMN.$X(10,2) := 3.4394567E-33 ;
SYSTEM.COLUMN.X(10,3) := -1.6841651E-29 ;
SYSTEM.COLUMN.$X(10,3) := -3.0684627E-32 ;

341
SYSTEM.COLUMN.X(10,4) := -5.7768554E-30 ;
SYSTEM.COLUMN.$X(10,4) := -1.5124539E-32 ;
SYSTEM.COLUMN.Y(1,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(1,2) := -1.5898045E-30 ;
SYSTEM.COLUMN.Y(1,3) := -2.8366905E-33 ;
SYSTEM.COLUMN.Y(1,4) := -1.4248525E-33 ;
SYSTEM.COLUMN.Y(2,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(2,2) := -1.4211639E-30 ;
SYSTEM.COLUMN.Y(2,3) := -7.7119210E-33 ;
SYSTEM.COLUMN.Y(2,4) := -6.9770011E-33 ;
SYSTEM.COLUMN.Y(3,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(3,2) := -1.3467451E-30 ;
SYSTEM.COLUMN.Y(3,3) := -1.4003277E-32 ;
SYSTEM.COLUMN.Y(3,4) := -1.6716159E-32 ;
SYSTEM.COLUMN.Y(4,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(4,2) := -1.1599208E-30 ;
SYSTEM.COLUMN.Y(4,3) := -2.5279322E-32 ;
SYSTEM.COLUMN.Y(4,4) := -1.3460165E-32 ;
SYSTEM.COLUMN.Y(5,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(5,2) := -1.0983600E-30 ;
SYSTEM.COLUMN.Y(5,3) := -3.4537156E-32 ;
SYSTEM.COLUMN.Y(5,4) := -9.0130471E-34 ;
SYSTEM.COLUMN.Y(6,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(6,2) := -9.9275052E-31 ;
SYSTEM.COLUMN.Y(6,3) := -9.6657325E-32 ;
SYSTEM.COLUMN.Y(6,4) := -2.6002870E-33 ;
SYSTEM.COLUMN.Y(7,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(7,2) := -9.0257137E-31 ;
SYSTEM.COLUMN.Y(7,3) := -3.6996193E-31 ;
SYSTEM.COLUMN.Y(7,4) := -1.0906238E-32 ;
SYSTEM.COLUMN.Y(8,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(8,2) := -8.2287294E-31 ;
SYSTEM.COLUMN.Y(8,3) := -8.5153253E-31 ;
SYSTEM.COLUMN.Y(8,4) := -5.1516140E-32 ;
SYSTEM.COLUMN.Y(9,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(9,2) := -7.5220712E-31 ;
SYSTEM.COLUMN.Y(9,3) := -2.2095205E-30 ;
SYSTEM.COLUMN.Y(9,4) := -2.4977270E-31 ;
SYSTEM.COLUMN.Y(10,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(10,2) := -6.8934926E-31 ;
SYSTEM.COLUMN.Y(10,3) := -6.0362642E-30 ;
SYSTEM.COLUMN.Y(10,4) := -1.1990035E-30 ;
SYSTEM.COLUMN.FEEDENTHALPY := -4.5983100E+05 ;
SYSTEM.COLUMN.MWL(1) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(2) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(3) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(4) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(5) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(6) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(7) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(8) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(9) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(10) := 7.6000000E+01 ;
SYSTEM.COLUMN.TOTALHOLDUP := 4.6586382E+01 ;

342
SYSTEM.COLUMN.$TOTALHOLDUP := 0.0000000E+00 ;
SYSTEM.COLUMN.QC := 7.9357010E+05 ;
SYSTEM.VLE.VAPORENTHALPY(1) := 9.6490429E+03 ;
SYSTEM.VLE.VAPORENTHALPY(2) := 9.6493013E+03 ;
SYSTEM.VLE.VAPORENTHALPY(3) := 9.7613871E+03 ;
SYSTEM.VLE.VAPORENTHALPY(4) := 9.8713496E+03 ;
SYSTEM.VLE.VAPORENTHALPY(5) := 9.9792593E+03 ;
SYSTEM.VLE.VAPORENTHALPY(6) := 1.0094036E+04 ;
SYSTEM.VLE.VAPORENTHALPY(7) := 1.0206730E+04 ;
SYSTEM.VLE.VAPORENTHALPY(8) := 1.0317422E+04 ;
SYSTEM.VLE.VAPORENTHALPY(9) := 1.0426188E+04 ;
SYSTEM.VLE.VAPORENTHALPY(10) := 1.0533100E+04 ;
SYSTEM.VLE.TEMP(1) := 3.5400141E+02 ;
SYSTEM.VLE.TEMP(2) := 3.5400291E+02 ;
SYSTEM.VLE.TEMP(3) := 3.5465344E+02 ;
SYSTEM.VLE.TEMP(4) := 3.5529164E+02 ;
SYSTEM.VLE.TEMP(5) := 3.5591793E+02 ;
SYSTEM.VLE.TEMP(6) := 3.5658408E+02 ;
SYSTEM.VLE.TEMP(7) := 3.5723813E+02 ;
SYSTEM.VLE.TEMP(8) := 3.5788057E+02 ;
SYSTEM.VLE.TEMP(9) := 3.5851183E+02 ;
SYSTEM.VLE.TEMP(10) := 3.5913233E+02 ;
SYSTEM.VLE.LIQUIDENTHALPY(1) := -2.1350957E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(2) := -2.1350699E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(3) := -2.1238613E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(4) := -2.1128650E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(5) := -2.1020741E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(6) := -2.0905964E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(7) := -2.0793270E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(8) := -2.0682578E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(9) := -2.0573812E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(10) := -2.0466900E+04 ;
SYSTEM.VLE.K(1,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(1,2) := 1.1186637E+00 ;
SYSTEM.VLE.K(1,3) := 3.6783189E-01 ;
SYSTEM.VLE.K(1,4) := 2.0422134E-01 ;
SYSTEM.VLE.K(2,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(2,2) := 1.1186548E+00 ;
SYSTEM.VLE.K(2,3) := 3.6782906E-01 ;
SYSTEM.VLE.K(2,4) := 2.0422232E-01 ;
SYSTEM.VLE.K(3,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(3,2) := 1.1147884E+00 ;
SYSTEM.VLE.K(3,3) := 3.6660608E-01 ;
SYSTEM.VLE.K(3,4) := 2.0464705E-01 ;
SYSTEM.VLE.K(4,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(4,2) := 1.1110220E+00 ;
SYSTEM.VLE.K(4,3) := 3.6541455E-01 ;
SYSTEM.VLE.K(4,4) := 2.0506307E-01 ;
SYSTEM.VLE.K(5,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(5,2) := 1.1073514E+00 ;
SYSTEM.VLE.K(5,3) := 3.6425316E-01 ;
SYSTEM.VLE.K(5,4) := 2.0547070E-01 ;
SYSTEM.VLE.K(6,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(6,2) := 1.1034745E+00 ;

343
SYSTEM.VLE.K(6,3) := 3.6302637E-01 ;
SYSTEM.VLE.K(6,4) := 2.0590358E-01 ;
SYSTEM.VLE.K(7,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(7,2) := 1.0996952E+00 ;
SYSTEM.VLE.K(7,3) := 3.6183030E-01 ;
SYSTEM.VLE.K(7,4) := 2.0632791E-01 ;
SYSTEM.VLE.K(8,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(8,2) := 1.0960090E+00 ;
SYSTEM.VLE.K(8,3) := 3.6066354E-01 ;
SYSTEM.VLE.K(8,4) := 2.0674404E-01 ;
SYSTEM.VLE.K(9,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(9,2) := 1.0924118E+00 ;
SYSTEM.VLE.K(9,3) := 3.5952480E-01 ;
SYSTEM.VLE.K(9,4) := 2.0715229E-01 ;
SYSTEM.VLE.K(10,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(10,2) := 1.0888996E+00 ;
SYSTEM.VLE.K(10,3) := 3.5841286E-01 ;
SYSTEM.VLE.K(10,4) := 2.0755297E-01 ;
SYSTEM.VLE.VAPORPRESSURE(1,1) := 1.0132500E+00 ;
SYSTEM.VLE.VAPORPRESSURE(1,2) := 1.1334860E+00 ;
SYSTEM.VLE.VAPORPRESSURE(1,3) := 3.7270566E-01 ;
SYSTEM.VLE.VAPORPRESSURE(1,4) := 2.0692728E-01 ;
SYSTEM.VLE.VAPORPRESSURE(2,1) := 1.0133002E+00 ;
SYSTEM.VLE.VAPORPRESSURE(2,2) := 1.1335331E+00 ;
SYSTEM.VLE.VAPORPRESSURE(2,3) := 3.7272127E-01 ;
SYSTEM.VLE.VAPORPRESSURE(2,4) := 2.0693853E-01 ;
SYSTEM.VLE.VAPORPRESSURE(3,1) := 1.0352865E+00 ;
SYSTEM.VLE.VAPORPRESSURE(3,2) := 1.1541254E+00 ;
SYSTEM.VLE.VAPORPRESSURE(3,3) := 3.7954232E-01 ;
SYSTEM.VLE.VAPORPRESSURE(3,4) := 2.1186832E-01 ;
SYSTEM.VLE.VAPORPRESSURE(4,1) := 1.0572390E+00 ;
SYSTEM.VLE.VAPORPRESSURE(4,2) := 1.1746158E+00 ;
SYSTEM.VLE.VAPORPRESSURE(4,3) := 3.8633050E-01 ;
SYSTEM.VLE.VAPORPRESSURE(4,4) := 2.1680067E-01 ;
SYSTEM.VLE.VAPORPRESSURE(5,1) := 1.0791551E+00 ;
SYSTEM.VLE.VAPORPRESSURE(5,2) := 1.1950039E+00 ;
SYSTEM.VLE.VAPORPRESSURE(5,3) := 3.9308568E-01 ;
SYSTEM.VLE.VAPORPRESSURE(5,4) := 2.2173476E-01 ;
SYSTEM.VLE.VAPORPRESSURE(6,1) := 1.1028776E+00 ;
SYSTEM.VLE.VAPORPRESSURE(6,2) := 1.2169973E+00 ;
SYSTEM.VLE.VAPORPRESSURE(6,3) := 4.0037366E-01 ;
SYSTEM.VLE.VAPORPRESSURE(6,4) := 2.2708645E-01 ;
SYSTEM.VLE.VAPORPRESSURE(7,1) := 1.1265880E+00 ;
SYSTEM.VLE.VAPORPRESSURE(7,2) := 1.2389034E+00 ;
SYSTEM.VLE.VAPORPRESSURE(7,3) := 4.0763366E-01 ;
SYSTEM.VLE.VAPORPRESSURE(7,4) := 2.3244654E-01 ;
SYSTEM.VLE.VAPORPRESSURE(8,1) := 1.1502862E+00 ;
SYSTEM.VLE.VAPORPRESSURE(8,2) := 1.2607240E+00 ;
SYSTEM.VLE.VAPORPRESSURE(8,3) := 4.1486630E-01 ;
SYSTEM.VLE.VAPORPRESSURE(8,4) := 2.3781482E-01 ;
SYSTEM.VLE.VAPORPRESSURE(9,1) := 1.1739724E+00 ;
SYSTEM.VLE.VAPORPRESSURE(9,2) := 1.2824613E+00 ;
SYSTEM.VLE.VAPORPRESSURE(9,3) := 4.2207220E-01 ;
SYSTEM.VLE.VAPORPRESSURE(9,4) := 2.4319108E-01 ;

344
SYSTEM.VLE.VAPORPRESSURE(10,1) := 1.1976465E+00 ;
SYSTEM.VLE.VAPORPRESSURE(10,2) := 1.3041169E+00 ;
SYSTEM.VLE.VAPORPRESSURE(10,3) := 4.2925192E-01 ;
SYSTEM.VLE.VAPORPRESSURE(10,4) := 2.4857510E-01 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,1) := 9.6490429E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,2) := 1.1200282E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,3) := 8.9602256E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,4) := 8.6802185E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,1) := 9.6493013E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,2) := 1.1200582E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,3) := 8.9604655E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,4) := 8.6804509E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,1) := 9.7613871E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,2) := 1.1330687E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,3) := 9.0645499E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,4) := 8.7812827E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,1) := 9.8713496E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,2) := 1.1458328E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,3) := 9.1666624E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,4) := 8.8802042E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,1) := 9.9792593E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,2) := 1.1583586E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,3) := 9.2668688E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,4) := 8.9772791E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,1) := 1.0094036E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,2) := 1.1716815E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,3) := 9.3734520E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,4) := 9.0805316E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,1) := 1.0206730E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,2) := 1.1847626E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,3) := 9.4781009E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,4) := 9.1819102E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,1) := 1.0317422E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,2) := 1.1976113E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,3) := 9.5808907E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,4) := 9.2814879E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,1) := 1.0426188E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,2) := 1.2102365E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,3) := 9.6818924E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,4) := 9.3793332E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,1) := 1.0533100E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,2) := 1.2226466E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,3) := 9.7811726E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,4) := 9.4755110E+03 ;
SYSTEM.VLE.P(1) := 1.0132500E+00 ;
SYSTEM.VLE.P(2) := 1.0133002E+00 ;
SYSTEM.VLE.P(3) := 1.0352865E+00 ;
SYSTEM.VLE.P(4) := 1.0572390E+00 ;
SYSTEM.VLE.P(5) := 1.0791551E+00 ;
SYSTEM.VLE.P(6) := 1.1028776E+00 ;
SYSTEM.VLE.P(7) := 1.1265880E+00 ;
SYSTEM.VLE.P(8) := 1.1502862E+00 ;
SYSTEM.VLE.P(9) := 1.1739724E+00 ;
SYSTEM.VLE.P(10) := 1.1976465E+00 ;

345
SYSTEM.VLE.X(1,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(1,2) := -1.4211639E-30 ;
SYSTEM.VLE.X(1,3) := -7.7119210E-33 ;
SYSTEM.VLE.X(1,4) := -6.9770011E-33 ;
SYSTEM.VLE.X(2,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(2,2) := -1.2704223E-30 ;
SYSTEM.VLE.X(2,3) := -2.0966046E-32 ;
SYSTEM.VLE.X(2,4) := -6.1284288E-32 ;
SYSTEM.VLE.X(3,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(3,2) := -1.2080724E-30 ;
SYSTEM.VLE.X(3,3) := -3.8197066E-32 ;
SYSTEM.VLE.X(3,4) := -8.1682874E-32 ;
SYSTEM.VLE.X(4,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(4,2) := -1.0440124E-30 ;
SYSTEM.VLE.X(4,3) := -6.9179846E-32 ;
SYSTEM.VLE.X(4,4) := -6.5639146E-32 ;
SYSTEM.VLE.X(5,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(5,2) := -9.9188031E-31 ;
SYSTEM.VLE.X(5,3) := -9.4816353E-32 ;
SYSTEM.VLE.X(5,4) := -4.3865365E-33 ;
SYSTEM.VLE.X(6,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(6,2) := -8.9965879E-31 ;
SYSTEM.VLE.X(6,3) := -2.6625428E-31 ;
SYSTEM.VLE.X(6,4) := -1.2628664E-32 ;
SYSTEM.VLE.X(7,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(7,2) := -8.2074687E-31 ;
SYSTEM.VLE.X(7,3) := -1.0224736E-30 ;
SYSTEM.VLE.X(7,4) := -5.2858763E-32 ;
SYSTEM.VLE.X(8,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(8,2) := -7.5079031E-31 ;
SYSTEM.VLE.X(8,3) := -2.3610164E-30 ;
SYSTEM.VLE.X(8,4) := -2.4917836E-31 ;
SYSTEM.VLE.X(9,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(9,2) := -6.8857471E-31 ;
SYSTEM.VLE.X(9,3) := -6.1456693E-30 ;
SYSTEM.VLE.X(9,4) := -1.2057443E-30 ;
SYSTEM.VLE.X(10,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(10,2) := -6.3306961E-31 ;
SYSTEM.VLE.X(10,3) := -1.6841651E-29 ;
SYSTEM.VLE.X(10,4) := -5.7768554E-30 ;
SYSTEM.VLE.Y(1,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(1,2) := -1.5898045E-30 ;
SYSTEM.VLE.Y(1,3) := -2.8366905E-33 ;
SYSTEM.VLE.Y(1,4) := -1.4248525E-33 ;
SYSTEM.VLE.Y(2,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(2,2) := -1.4211639E-30 ;
SYSTEM.VLE.Y(2,3) := -7.7119210E-33 ;
SYSTEM.VLE.Y(2,4) := -6.9770011E-33 ;
SYSTEM.VLE.Y(3,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(3,2) := -1.3467451E-30 ;
SYSTEM.VLE.Y(3,3) := -1.4003277E-32 ;
SYSTEM.VLE.Y(3,4) := -1.6716159E-32 ;
SYSTEM.VLE.Y(4,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(4,2) := -1.1599208E-30 ;

346
SYSTEM.VLE.Y(4,3) := -2.5279322E-32 ;
SYSTEM.VLE.Y(4,4) := -1.3460165E-32 ;
SYSTEM.VLE.Y(5,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(5,2) := -1.0983600E-30 ;
SYSTEM.VLE.Y(5,3) := -3.4537156E-32 ;
SYSTEM.VLE.Y(5,4) := -9.0130471E-34 ;
SYSTEM.VLE.Y(6,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(6,2) := -9.9275052E-31 ;
SYSTEM.VLE.Y(6,3) := -9.6657325E-32 ;
SYSTEM.VLE.Y(6,4) := -2.6002870E-33 ;
SYSTEM.VLE.Y(7,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(7,2) := -9.0257137E-31 ;
SYSTEM.VLE.Y(7,3) := -3.6996193E-31 ;
SYSTEM.VLE.Y(7,4) := -1.0906238E-32 ;
SYSTEM.VLE.Y(8,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(8,2) := -8.2287294E-31 ;
SYSTEM.VLE.Y(8,3) := -8.5153253E-31 ;
SYSTEM.VLE.Y(8,4) := -5.1516140E-32 ;
SYSTEM.VLE.Y(9,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(9,2) := -7.5220712E-31 ;
SYSTEM.VLE.Y(9,3) := -2.2095205E-30 ;
SYSTEM.VLE.Y(9,4) := -2.4977270E-31 ;
SYSTEM.VLE.Y(10,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(10,2) := -6.8934926E-31 ;
SYSTEM.VLE.Y(10,3) := -6.0362642E-30 ;
SYSTEM.VLE.Y(10,4) := -1.1990035E-30 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,1) := -2.1350957E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,2) := -1.4799718E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,3) := -1.9039774E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,4) := -2.5319781E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,1) := -2.1350699E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,2) := -1.4799418E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,3) := -1.9039535E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,4) := -2.5319549E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,1) := -2.1238613E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,2) := -1.4669313E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,3) := -1.8935450E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,4) := -2.5218717E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,1) := -2.1128650E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,2) := -1.4541672E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,3) := -1.8833338E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,4) := -2.5119796E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,1) := -2.1020741E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,2) := -1.4416414E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,3) := -1.8733131E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,4) := -2.5022721E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,1) := -2.0905964E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,2) := -1.4283185E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,3) := -1.8626548E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,4) := -2.4919468E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,1) := -2.0793270E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,2) := -1.4152374E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,3) := -1.8521899E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,4) := -2.4818090E+04 ;

347
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,1) := -2.0682578E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,2) := -1.4023887E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,3) := -1.8419109E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,4) := -2.4718512E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,1) := -2.0573812E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,2) := -1.3897635E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,3) := -1.8318108E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,4) := -2.4620667E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,1) := -2.0466900E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,2) := -1.3773534E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,3) := -1.8218827E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,4) := -2.4524489E+04 ;

INITIAL
WITHIN System.Column DO
FOR I:=2 TO NSTAGE DO
X(I,D)= 0.0 ;
X(I,B)= 0.0 ;
X(I,C)= 0.0 ;
END
TotalHoldup=4.6586382E+01 ;
M(2)=2.9237071E+00 ;
M(3)=2.9189519E+00 ;
M(4)=2.9138424E+00 ;
M(5)=3.1682045E+00 ;
M(6)=3.1664972E+00 ;
M(7)=3.1647940E+00 ;
M(8)=3.1630950E+00 ;
M(9)=3.1614000E+00 ;
FOR I:=2 TO NSTAGE DO
SIGMA(X(I,))=1 ;
END

END

SCHEDULE
CONTINUE FOR 1000

END

SIMULATION RunBoth

UNIT
System As BothFlowSheet

INPUT
WITHIN System.Reactor DO
FlowOut := 15 ;
FeedTemp := 300 ;
END
WITHIN System.Jacket DO
Flow_Water := 3 ;
END

348
 WITHIN System.Column DO
P(1) :=1.01325 ;
REFLUXRATIO :=0.5 ;
END
WITHIN System DO
BFraction:=0.15 ;
END

PRESET
##############################
# Values of All Active Variables #
##############################

SYSTEM.MAKEUP(1) := 7.6341614E-02 ;
SYSTEM.MAKEUP(2) := 7.5000000E+00 ;
SYSTEM.MAKEUP(3) := 0.0000000E+00 ;
SYSTEM.MAKEUP(4) := 0.0000000E+00 ;
SYSTEM.COLUMN.MWV(1) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(2) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(3) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(4) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(5) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(6) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(7) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(8) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(9) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(10) := 7.6000000E+01 ;
SYSTEM.COLUMN.BOTTOMS := 2.2005890E+00 ;
SYSTEM.COLUMN.VAPORENTHALPY(1) := 9.6490429E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(2) := 9.6493013E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(3) := 9.7613871E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(4) := 9.8713496E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(5) := 9.9792593E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(6) := 1.0094036E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(7) := 1.0206730E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(8) := 1.0317422E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(9) := 1.0426188E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(10) := 1.0533100E+04 ;
SYSTEM.COLUMN.LIQMDENS(1) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(2) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(3) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(4) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(5) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(6) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(7) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(8) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(9) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(10) := 8.3600000E+02 ;
SYSTEM.COLUMN.VAPMDENS(1) := 2.6164666E+00 ;
SYSTEM.COLUMN.VAPMDENS(2) := 2.6165852E+00 ;
SYSTEM.COLUMN.VAPMDENS(3) := 2.6684553E+00 ;
SYSTEM.COLUMN.VAPMDENS(4) := 2.7201430E+00 ;
SYSTEM.COLUMN.VAPMDENS(5) := 2.7716449E+00 ;
SYSTEM.COLUMN.VAPMDENS(6) := 2.8272809E+00 ;

349
SYSTEM.COLUMN.VAPMDENS(7) := 2.8827759E+00 ;
SYSTEM.COLUMN.VAPMDENS(8) := 2.9381325E+00 ;
SYSTEM.COLUMN.VAPMDENS(9) := 2.9933533E+00 ;
SYSTEM.COLUMN.VAPMDENS(10) := 3.0484406E+00 ;
SYSTEM.COLUMN.QR := 6.4072745E+06 ;
SYSTEM.COLUMN.DPSTAT(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.DPSTAT(2) := 2.0759991E-02 ;
SYSTEM.COLUMN.DPSTAT(3) := 2.0726226E-02 ;
SYSTEM.COLUMN.DPSTAT(4) := 2.0689946E-02 ;
SYSTEM.COLUMN.DPSTAT(5) := 2.2496062E-02 ;
SYSTEM.COLUMN.DPSTAT(6) := 2.2483940E-02 ;
SYSTEM.COLUMN.DPSTAT(7) := 2.2471846E-02 ;
SYSTEM.COLUMN.DPSTAT(8) := 2.2459782E-02 ;
SYSTEM.COLUMN.DPSTAT(9) := 2.2447747E-02 ;
SYSTEM.COLUMN.DPSTAT(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.XFEED(1) := 1.0000000E+00 ;
SYSTEM.COLUMN.XFEED(2) := 0.0000000E+00 ;
SYSTEM.COLUMN.XFEED(3) := 0.0000000E+00 ;
SYSTEM.COLUMN.XFEED(4) := 0.0000000E+00 ;
SYSTEM.COLUMN.DPTRAY(1) := 5.0229571E-05 ;
SYSTEM.COLUMN.DPTRAY(2) := 2.1986243E-02 ;
SYSTEM.COLUMN.DPTRAY(3) := 2.1952478E-02 ;
SYSTEM.COLUMN.DPTRAY(4) := 2.1916197E-02 ;
SYSTEM.COLUMN.DPTRAY(5) := 2.3722479E-02 ;
SYSTEM.COLUMN.DPTRAY(6) := 2.3710350E-02 ;
SYSTEM.COLUMN.DPTRAY(7) := 2.3698251E-02 ;
SYSTEM.COLUMN.DPTRAY(8) := 2.3686181E-02 ;
SYSTEM.COLUMN.DPTRAY(9) := 2.3674140E-02 ;
SYSTEM.COLUMN.DPTRAY(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(1) := -2.1350957E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(2) := -2.1350699E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(3) := -2.1238613E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(4) := -2.1128650E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(5) := -2.1020741E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(6) := -2.0905964E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(7) := -2.0793270E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(8) := -2.0682578E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(9) := -2.0573812E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(10) := -2.0466900E+04 ;
SYSTEM.COLUMN.LIQHEIGHT(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.LIQHEIGHT(2) := 2.5313481E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(3) := 2.5272311E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(4) := 2.5228073E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(5) := 2.7430342E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(6) := 2.7415560E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(7) := 2.7400814E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(8) := 2.7386104E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(9) := 2.7371429E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.HEAD(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.HEAD(2) := 3.1348139E+00 ;
SYSTEM.COLUMN.HEAD(3) := 2.7231082E+00 ;
SYSTEM.COLUMN.HEAD(4) := 2.2807273E+00 ;
SYSTEM.COLUMN.HEAD(5) := 2.4303420E+01 ;

350
SYSTEM.COLUMN.HEAD(6) := 2.4155602E+01 ;
SYSTEM.COLUMN.HEAD(7) := 2.4008139E+01 ;
SYSTEM.COLUMN.HEAD(8) := 2.3861039E+01 ;
SYSTEM.COLUMN.HEAD(9) := 2.3714286E+01 ;
SYSTEM.COLUMN.HEAD(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.K(1,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(1,2) := 1.1186637E+00 ;
SYSTEM.COLUMN.K(1,3) := 3.6783189E-01 ;
SYSTEM.COLUMN.K(1,4) := 2.0422134E-01 ;
SYSTEM.COLUMN.K(2,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(2,2) := 1.1186548E+00 ;
SYSTEM.COLUMN.K(2,3) := 3.6782906E-01 ;
SYSTEM.COLUMN.K(2,4) := 2.0422232E-01 ;
SYSTEM.COLUMN.K(3,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(3,2) := 1.1147884E+00 ;
SYSTEM.COLUMN.K(3,3) := 3.6660608E-01 ;
SYSTEM.COLUMN.K(3,4) := 2.0464705E-01 ;
SYSTEM.COLUMN.K(4,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(4,2) := 1.1110220E+00 ;
SYSTEM.COLUMN.K(4,3) := 3.6541455E-01 ;
SYSTEM.COLUMN.K(4,4) := 2.0506307E-01 ;
SYSTEM.COLUMN.K(5,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(5,2) := 1.1073514E+00 ;
SYSTEM.COLUMN.K(5,3) := 3.6425316E-01 ;
SYSTEM.COLUMN.K(5,4) := 2.0547070E-01 ;
SYSTEM.COLUMN.K(6,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(6,2) := 1.1034745E+00 ;
SYSTEM.COLUMN.K(6,3) := 3.6302637E-01 ;
SYSTEM.COLUMN.K(6,4) := 2.0590358E-01 ;
SYSTEM.COLUMN.K(7,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(7,2) := 1.0996952E+00 ;
SYSTEM.COLUMN.K(7,3) := 3.6183030E-01 ;
SYSTEM.COLUMN.K(7,4) := 2.0632791E-01 ;
SYSTEM.COLUMN.K(8,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(8,2) := 1.0960090E+00 ;
SYSTEM.COLUMN.K(8,3) := 3.6066354E-01 ;
SYSTEM.COLUMN.K(8,4) := 2.0674404E-01 ;
SYSTEM.COLUMN.K(9,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(9,2) := 1.0924118E+00 ;
SYSTEM.COLUMN.K(9,3) := 3.5952480E-01 ;
SYSTEM.COLUMN.K(9,4) := 2.0715229E-01 ;
SYSTEM.COLUMN.K(10,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(10,2) := 1.0888996E+00 ;
SYSTEM.COLUMN.K(10,3) := 3.5841286E-01 ;
SYSTEM.COLUMN.K(10,4) := 2.0755297E-01 ;
SYSTEM.COLUMN.L(1) := 1.2799411E+01 ;
SYSTEM.COLUMN.L(2) := 1.0700915E+01 ;
SYSTEM.COLUMN.L(3) := 8.6636449E+00 ;
SYSTEM.COLUMN.L(4) := 6.6406918E+00 ;
SYSTEM.COLUMN.L(5) := 2.3099596E+02 ;
SYSTEM.COLUMN.L(6) := 2.2889172E+02 ;
SYSTEM.COLUMN.L(7) := 2.2679894E+02 ;
SYSTEM.COLUMN.L(8) := 2.2471772E+02 ;
SYSTEM.COLUMN.L(9) := 2.2264778E+02 ;

351
SYSTEM.COLUMN.L(10) := 2.2005890E+00 ;
SYSTEM.COLUMN.FEED(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(2) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(3) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(4) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(5) := 1.5000000E+01 ;
SYSTEM.COLUMN.FEED(6) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(7) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(8) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(9) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.M(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.$M(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.M(2) := 2.9237071E+00 ;
SYSTEM.COLUMN.$M(2) := -5.8431630E-05 ;
SYSTEM.COLUMN.M(3) := 2.9189519E+00 ;
SYSTEM.COLUMN.$M(3) := 1.9291184E-05 ;
SYSTEM.COLUMN.M(4) := 2.9138424E+00 ;
SYSTEM.COLUMN.$M(4) := 8.9519949E-05 ;
SYSTEM.COLUMN.M(5) := 3.1682045E+00 ;
SYSTEM.COLUMN.$M(5) := -1.1165674E-04 ;
SYSTEM.COLUMN.M(6) := 3.1664972E+00 ;
SYSTEM.COLUMN.$M(6) := -1.0041827E-04 ;
SYSTEM.COLUMN.M(7) := 3.1647940E+00 ;
SYSTEM.COLUMN.$M(7) := 3.0944158E-04 ;
SYSTEM.COLUMN.M(8) := 3.1630950E+00 ;
SYSTEM.COLUMN.$M(8) := -5.6979389E-05 ;
SYSTEM.COLUMN.M(9) := 3.1614000E+00 ;
SYSTEM.COLUMN.$M(9) := -7.1432749E-05 ;
SYSTEM.COLUMN.M(10) := 2.2005890E+01 ;
SYSTEM.COLUMN.$M(10) := -1.9333939E-05 ;
SYSTEM.COLUMN.LIQDENS(1) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(2) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(3) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(4) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(5) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(6) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(7) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(8) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(9) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(10) := 1.1000000E+01 ;
SYSTEM.COLUMN.VAPDENS(1) := 3.4427192E-02 ;
SYSTEM.COLUMN.VAPDENS(2) := 3.4428753E-02 ;
SYSTEM.COLUMN.VAPDENS(3) := 3.5111254E-02 ;
SYSTEM.COLUMN.VAPDENS(4) := 3.5791355E-02 ;
SYSTEM.COLUMN.VAPDENS(5) := 3.6469012E-02 ;
SYSTEM.COLUMN.VAPDENS(6) := 3.7201064E-02 ;
SYSTEM.COLUMN.VAPDENS(7) := 3.7931261E-02 ;
SYSTEM.COLUMN.VAPDENS(8) := 3.8659638E-02 ;
SYSTEM.COLUMN.VAPDENS(9) := 3.9386227E-02 ;
SYSTEM.COLUMN.VAPDENS(10) := 4.0111061E-02 ;
SYSTEM.COLUMN.P(2) := 1.0133002E+00 ;
SYSTEM.COLUMN.P(3) := 1.0352865E+00 ;
SYSTEM.COLUMN.P(4) := 1.0572390E+00 ;

352
SYSTEM.COLUMN.P(5) := 1.0791551E+00 ;
SYSTEM.COLUMN.P(6) := 1.1028776E+00 ;
SYSTEM.COLUMN.P(7) := 1.1265880E+00 ;
SYSTEM.COLUMN.P(8) := 1.1502862E+00 ;
SYSTEM.COLUMN.P(9) := 1.1739724E+00 ;
SYSTEM.COLUMN.P(10) := 1.1976465E+00 ;
SYSTEM.COLUMN.DISTILLOUT := 1.2799411E+01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(2) := 2.6579155E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(3) := 2.6535926E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(4) := 2.6489476E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(5) := 2.8801859E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(6) := 2.8786338E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(7) := 2.8770855E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(8) := 2.8755409E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(9) := 2.8740000E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(10) := 2.0005354E+00 ;
SYSTEM.COLUMN.T(1) := 3.5400141E+02 ;
SYSTEM.COLUMN.T(2) := 3.5400291E+02 ;
SYSTEM.COLUMN.T(3) := 3.5465344E+02 ;
SYSTEM.COLUMN.T(4) := 3.5529164E+02 ;
SYSTEM.COLUMN.T(5) := 3.5591793E+02 ;
SYSTEM.COLUMN.T(6) := 3.5658408E+02 ;
SYSTEM.COLUMN.T(7) := 3.5723813E+02 ;
SYSTEM.COLUMN.T(8) := 3.5788057E+02 ;
SYSTEM.COLUMN.T(9) := 3.5851183E+02 ;
SYSTEM.COLUMN.T(10) := 3.5913233E+02 ;
SYSTEM.COLUMN.V(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.V(2) := 2.5598822E+01 ;
SYSTEM.COLUMN.V(3) := 2.3500268E+01 ;
SYSTEM.COLUMN.V(4) := 2.1463017E+01 ;
SYSTEM.COLUMN.V(5) := 1.9440153E+01 ;
SYSTEM.COLUMN.V(6) := 2.2879531E+02 ;
SYSTEM.COLUMN.V(7) := 2.2669097E+02 ;
SYSTEM.COLUMN.V(8) := 2.2459850E+02 ;
SYSTEM.COLUMN.V(9) := 2.2251722E+02 ;
SYSTEM.COLUMN.V(10) := 2.2044721E+02 ;
#SYSTEM.COLUMN.FEEDTEMP := 3.0000000E+02 ;
SYSTEM.COLUMN.DPDRY(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.DPDRY(2) := 1.8678982E-09 ;
SYSTEM.COLUMN.DPDRY(3) := 1.5277914E-09 ;
SYSTEM.COLUMN.DPDRY(4) := 1.2295610E-09 ;
SYSTEM.COLUMN.DPDRY(5) := 1.6714719E-07 ;
SYSTEM.COLUMN.DPDRY(6) := 1.6085773E-07 ;
SYSTEM.COLUMN.DPDRY(7) := 1.5486215E-07 ;
SYSTEM.COLUMN.DPDRY(8) := 1.4914144E-07 ;
SYSTEM.COLUMN.DPDRY(9) := 1.4367913E-07 ;
SYSTEM.COLUMN.DPDRY(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(1,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.X(1,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(1,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(1,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(2,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(2,1) := 1.0816707E-21 ;

353
SYSTEM.COLUMN.X(2,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(2,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(2,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(2,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(2,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(2,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(3,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(3,1) := -1.5264713E-22 ;
SYSTEM.COLUMN.X(3,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(3,2) := 2.4869401E-23 ;
SYSTEM.COLUMN.X(3,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(3,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(3,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(3,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(4,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(4,1) := 2.1141297E-21 ;
SYSTEM.COLUMN.X(4,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(4,2) := -2.1385140E-23 ;
SYSTEM.COLUMN.X(4,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(4,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(4,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(4,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(5,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(5,1) := -1.2429451E-21 ;
SYSTEM.COLUMN.X(5,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(5,2) := -3.2446314E-24 ;
SYSTEM.COLUMN.X(5,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(5,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(5,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(5,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(6,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(6,1) := 8.0887492E-22 ;
SYSTEM.COLUMN.X(6,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(6,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(6,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(6,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(6,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(6,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(7,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(7,1) := -6.1828866E-22 ;
SYSTEM.COLUMN.X(7,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(7,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(7,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(7,3) := 4.1248869E-24 ;
SYSTEM.COLUMN.X(7,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(7,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(8,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(8,1) := 1.2278487E-21 ;
SYSTEM.COLUMN.X(8,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(8,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(8,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(8,3) := -4.1271025E-24 ;
SYSTEM.COLUMN.X(8,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(8,4) := 0.0000000E+00 ;

354
SYSTEM.COLUMN.X(9,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(9,1) := 1.0411440E-21 ;
SYSTEM.COLUMN.X(9,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(9,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(9,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(9,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(9,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(9,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(10,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(10,1) := -8.9506882E-23 ;
SYSTEM.COLUMN.X(10,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(10,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(10,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(10,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(10,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(10,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(1,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(1,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(1,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(1,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(2,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(2,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(2,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(2,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(3,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(3,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(3,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(3,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(4,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(4,2) := 3.3822172E-24 ;
SYSTEM.COLUMN.Y(4,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(4,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(5,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(5,2) := 5.2878473E-25 ;
SYSTEM.COLUMN.Y(5,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(5,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(6,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(6,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(6,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(6,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(7,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(7,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(7,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(7,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(8,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(8,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(8,3) := 5.8123352E-26 ;
SYSTEM.COLUMN.Y(8,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(9,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(9,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(9,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(9,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(10,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(10,2) := 0.0000000E+00 ;

355
SYSTEM.COLUMN.Y(10,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(10,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEEDENTHALPY := -4.5983100E+05 ;
SYSTEM.COLUMN.MWL(1) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(2) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(3) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(4) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(5) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(6) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(7) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(8) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(9) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(10) := 7.6000000E+01 ;
SYSTEM.COLUMN.TOTALHOLDUP := 4.6586382E+01 ;
SYSTEM.COLUMN.$TOTALHOLDUP := 0.0000000E+00 ;
SYSTEM.COLUMN.QC := 7.9357010E+05 ;
SYSTEM.MAKEUPENTHALPY.TEMP := 3.0000000E+02 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_VAPOR(1) := 3.4460000E+02 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_VAPOR(2) := 4.0000000E+02 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_VAPOR(3) := 3.2000000E+02 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_VAPOR(4) := 3.1000000E+02 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_LIQUID(1) := -3.0655400E+04 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_LIQUID(2) := -2.5600000E+04 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_LIQUID(3) := -2.7680000E+04 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_LIQUID(4) := -3.3690000E+04 ;
SYSTEM.REACTOR.TOTALMOLS := 5.0000000E+01 ;
SYSTEM.REACTOR.$TOTALMOLS := 0.0000000E+00 ;
SYSTEM.REACTOR.TEMP := 3.0000000E+02 ;
SYSTEM.REACTOR.ENTHALPY := -1.5327700E+06 ;
SYSTEM.REACTOR.$ENTHALPY := -4.3344954E+06 ;
SYSTEM.REACTOR.TOTALFEED := 1.5000000E+01 ;
SYSTEM.REACTOR.QJACKET := 4.8152866E+04 ;
SYSTEM.REACTOR.NO_MOLS(1) := 5.0000000E+01 ;
SYSTEM.REACTOR.$NO_MOLS(1) := -7.5000000E+00 ;
SYSTEM.REACTOR.NO_MOLS(2) := 0.0000000E+00 ;
SYSTEM.REACTOR.$NO_MOLS(2) := 7.5000000E+00 ;
SYSTEM.REACTOR.NO_MOLS(3) := 0.0000000E+00 ;
SYSTEM.REACTOR.$NO_MOLS(3) := 0.0000000E+00 ;
SYSTEM.REACTOR.NO_MOLS(4) := 0.0000000E+00 ;
SYSTEM.REACTOR.$NO_MOLS(4) := 0.0000000E+00 ;
SYSTEM.REACTOR.VOLUME := 4.5454545E+00 ;
SYSTEM.REACTOR.CONCENTRATION(1) := 1.1000000E+01 ;
SYSTEM.REACTOR.CONCENTRATION(2) := 0.0000000E+00 ;
SYSTEM.REACTOR.CONCENTRATION(3) := 0.0000000E+00 ;
SYSTEM.REACTOR.CONCENTRATION(4) := 0.0000000E+00 ;
SYSTEM.REACTOR.FEED_A := 7.5000000E+00 ;
SYSTEM.REACTOR.FEED_B := 7.5000000E+00 ;
SYSTEM.REACTOR.FEED_C := 0.0000000E+00 ;
SYSTEM.REACTOR.X(1) := 1.0000000E+00 ;
SYSTEM.REACTOR.X(2) := 0.0000000E+00 ;
SYSTEM.REACTOR.X(3) := 0.0000000E+00 ;
SYSTEM.REACTOR.X(4) := 0.0000000E+00 ;
SYSTEM.REACTOR.FEED_D := 0.0000000E+00 ;
SYSTEM.REACTOR.FEEDENTHALPY(1) := -5.0553546E+05 ;

356
SYSTEM.REACTOR.FEEDENTHALPY(2) := -1.9395435E+05 ;
SYSTEM.REACTOR.FEEDENTHALPY(3) := -2.0971314E+05 ;
SYSTEM.REACTOR.FEEDENTHALPY(4) := -2.5524695E+05 ;
SYSTEM.REACTOR.SPECIFIC_ENTHALPY(1) := -3.0655400E+04 ;
SYSTEM.REACTOR.SPECIFIC_ENTHALPY(2) := -2.5600000E+04 ;
SYSTEM.REACTOR.SPECIFIC_ENTHALPY(3) := -2.7680000E+04 ;
SYSTEM.REACTOR.SPECIFIC_ENTHALPY(4) := -3.3690000E+04 ;
SYSTEM.REACTOR.REACTIONRATE(1) := 0.0000000E+00 ;
SYSTEM.REACTOR.REACTIONRATE(2) := 0.0000000E+00 ;
SYSTEM.VLE.VAPORENTHALPY(1) := 9.6490429E+03 ;
SYSTEM.VLE.VAPORENTHALPY(2) := 9.6493013E+03 ;
SYSTEM.VLE.VAPORENTHALPY(3) := 9.7613871E+03 ;
SYSTEM.VLE.VAPORENTHALPY(4) := 9.8713496E+03 ;
SYSTEM.VLE.VAPORENTHALPY(5) := 9.9792593E+03 ;
SYSTEM.VLE.VAPORENTHALPY(6) := 1.0094036E+04 ;
SYSTEM.VLE.VAPORENTHALPY(7) := 1.0206730E+04 ;
SYSTEM.VLE.VAPORENTHALPY(8) := 1.0317422E+04 ;
SYSTEM.VLE.VAPORENTHALPY(9) := 1.0426188E+04 ;
SYSTEM.VLE.VAPORENTHALPY(10) := 1.0533100E+04 ;
SYSTEM.VLE.TEMP(1) := 3.5400141E+02 ;
SYSTEM.VLE.TEMP(2) := 3.5400291E+02 ;
SYSTEM.VLE.TEMP(3) := 3.5465344E+02 ;
SYSTEM.VLE.TEMP(4) := 3.5529164E+02 ;
SYSTEM.VLE.TEMP(5) := 3.5591793E+02 ;
SYSTEM.VLE.TEMP(6) := 3.5658408E+02 ;
SYSTEM.VLE.TEMP(7) := 3.5723813E+02 ;
SYSTEM.VLE.TEMP(8) := 3.5788057E+02 ;
SYSTEM.VLE.TEMP(9) := 3.5851183E+02 ;
SYSTEM.VLE.TEMP(10) := 3.5913233E+02 ;
SYSTEM.VLE.LIQUIDENTHALPY(1) := -2.1350957E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(2) := -2.1350699E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(3) := -2.1238613E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(4) := -2.1128650E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(5) := -2.1020741E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(6) := -2.0905964E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(7) := -2.0793270E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(8) := -2.0682578E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(9) := -2.0573812E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(10) := -2.0466900E+04 ;
SYSTEM.VLE.K(1,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(1,2) := 1.1186637E+00 ;
SYSTEM.VLE.K(1,3) := 3.6783189E-01 ;
SYSTEM.VLE.K(1,4) := 2.0422134E-01 ;
SYSTEM.VLE.K(2,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(2,2) := 1.1186548E+00 ;
SYSTEM.VLE.K(2,3) := 3.6782906E-01 ;
SYSTEM.VLE.K(2,4) := 2.0422232E-01 ;
SYSTEM.VLE.K(3,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(3,2) := 1.1147884E+00 ;
SYSTEM.VLE.K(3,3) := 3.6660608E-01 ;
SYSTEM.VLE.K(3,4) := 2.0464705E-01 ;
SYSTEM.VLE.K(4,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(4,2) := 1.1110220E+00 ;
SYSTEM.VLE.K(4,3) := 3.6541455E-01 ;

357
SYSTEM.VLE.K(4,4) := 2.0506307E-01 ;
SYSTEM.VLE.K(5,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(5,2) := 1.1073514E+00 ;
SYSTEM.VLE.K(5,3) := 3.6425316E-01 ;
SYSTEM.VLE.K(5,4) := 2.0547070E-01 ;
SYSTEM.VLE.K(6,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(6,2) := 1.1034745E+00 ;
SYSTEM.VLE.K(6,3) := 3.6302637E-01 ;
SYSTEM.VLE.K(6,4) := 2.0590358E-01 ;
SYSTEM.VLE.K(7,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(7,2) := 1.0996952E+00 ;
SYSTEM.VLE.K(7,3) := 3.6183030E-01 ;
SYSTEM.VLE.K(7,4) := 2.0632791E-01 ;
SYSTEM.VLE.K(8,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(8,2) := 1.0960090E+00 ;
SYSTEM.VLE.K(8,3) := 3.6066354E-01 ;
SYSTEM.VLE.K(8,4) := 2.0674404E-01 ;
SYSTEM.VLE.K(9,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(9,2) := 1.0924118E+00 ;
SYSTEM.VLE.K(9,3) := 3.5952480E-01 ;
SYSTEM.VLE.K(9,4) := 2.0715229E-01 ;
SYSTEM.VLE.K(10,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(10,2) := 1.0888996E+00 ;
SYSTEM.VLE.K(10,3) := 3.5841286E-01 ;
SYSTEM.VLE.K(10,4) := 2.0755297E-01 ;
SYSTEM.VLE.VAPORPRESSURE(1,1) := 1.0132500E+00 ;
SYSTEM.VLE.VAPORPRESSURE(1,2) := 1.1334860E+00 ;
SYSTEM.VLE.VAPORPRESSURE(1,3) := 3.7270566E-01 ;
SYSTEM.VLE.VAPORPRESSURE(1,4) := 2.0692728E-01 ;
SYSTEM.VLE.VAPORPRESSURE(2,1) := 1.0133002E+00 ;
SYSTEM.VLE.VAPORPRESSURE(2,2) := 1.1335331E+00 ;
SYSTEM.VLE.VAPORPRESSURE(2,3) := 3.7272127E-01 ;
SYSTEM.VLE.VAPORPRESSURE(2,4) := 2.0693853E-01 ;
SYSTEM.VLE.VAPORPRESSURE(3,1) := 1.0352865E+00 ;
SYSTEM.VLE.VAPORPRESSURE(3,2) := 1.1541254E+00 ;
SYSTEM.VLE.VAPORPRESSURE(3,3) := 3.7954232E-01 ;
SYSTEM.VLE.VAPORPRESSURE(3,4) := 2.1186832E-01 ;
SYSTEM.VLE.VAPORPRESSURE(4,1) := 1.0572390E+00 ;
SYSTEM.VLE.VAPORPRESSURE(4,2) := 1.1746158E+00 ;
SYSTEM.VLE.VAPORPRESSURE(4,3) := 3.8633050E-01 ;
SYSTEM.VLE.VAPORPRESSURE(4,4) := 2.1680067E-01 ;
SYSTEM.VLE.VAPORPRESSURE(5,1) := 1.0791551E+00 ;
SYSTEM.VLE.VAPORPRESSURE(5,2) := 1.1950039E+00 ;
SYSTEM.VLE.VAPORPRESSURE(5,3) := 3.9308568E-01 ;
SYSTEM.VLE.VAPORPRESSURE(5,4) := 2.2173476E-01 ;
SYSTEM.VLE.VAPORPRESSURE(6,1) := 1.1028776E+00 ;
SYSTEM.VLE.VAPORPRESSURE(6,2) := 1.2169973E+00 ;
SYSTEM.VLE.VAPORPRESSURE(6,3) := 4.0037366E-01 ;
SYSTEM.VLE.VAPORPRESSURE(6,4) := 2.2708645E-01 ;
SYSTEM.VLE.VAPORPRESSURE(7,1) := 1.1265880E+00 ;
SYSTEM.VLE.VAPORPRESSURE(7,2) := 1.2389034E+00 ;
SYSTEM.VLE.VAPORPRESSURE(7,3) := 4.0763366E-01 ;
SYSTEM.VLE.VAPORPRESSURE(7,4) := 2.3244654E-01 ;
SYSTEM.VLE.VAPORPRESSURE(8,1) := 1.1502862E+00 ;

358
SYSTEM.VLE.VAPORPRESSURE(8,2) := 1.2607240E+00 ;
SYSTEM.VLE.VAPORPRESSURE(8,3) := 4.1486630E-01 ;
SYSTEM.VLE.VAPORPRESSURE(8,4) := 2.3781482E-01 ;
SYSTEM.VLE.VAPORPRESSURE(9,1) := 1.1739724E+00 ;
SYSTEM.VLE.VAPORPRESSURE(9,2) := 1.2824613E+00 ;
SYSTEM.VLE.VAPORPRESSURE(9,3) := 4.2207220E-01 ;
SYSTEM.VLE.VAPORPRESSURE(9,4) := 2.4319108E-01 ;
SYSTEM.VLE.VAPORPRESSURE(10,1) := 1.1976465E+00 ;
SYSTEM.VLE.VAPORPRESSURE(10,2) := 1.3041169E+00 ;
SYSTEM.VLE.VAPORPRESSURE(10,3) := 4.2925192E-01 ;
SYSTEM.VLE.VAPORPRESSURE(10,4) := 2.4857510E-01 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,1) := 9.6490429E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,2) := 1.1200282E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,3) := 8.9602256E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,4) := 8.6802185E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,1) := 9.6493013E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,2) := 1.1200582E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,3) := 8.9604655E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,4) := 8.6804509E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,1) := 9.7613871E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,2) := 1.1330687E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,3) := 9.0645499E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,4) := 8.7812827E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,1) := 9.8713496E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,2) := 1.1458328E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,3) := 9.1666624E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,4) := 8.8802042E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,1) := 9.9792593E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,2) := 1.1583586E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,3) := 9.2668688E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,4) := 8.9772791E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,1) := 1.0094036E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,2) := 1.1716815E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,3) := 9.3734520E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,4) := 9.0805316E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,1) := 1.0206730E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,2) := 1.1847626E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,3) := 9.4781009E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,4) := 9.1819102E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,1) := 1.0317422E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,2) := 1.1976113E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,3) := 9.5808907E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,4) := 9.2814879E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,1) := 1.0426188E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,2) := 1.2102365E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,3) := 9.6818924E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,4) := 9.3793332E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,1) := 1.0533100E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,2) := 1.2226466E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,3) := 9.7811726E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,4) := 9.4755110E+03 ;
SYSTEM.VLE.P(1) := 1.0132500E+00 ;
SYSTEM.VLE.P(2) := 1.0133002E+00 ;
SYSTEM.VLE.P(3) := 1.0352865E+00 ;

359
SYSTEM.VLE.P(4) := 1.0572390E+00 ;
SYSTEM.VLE.P(5) := 1.0791551E+00 ;
SYSTEM.VLE.P(6) := 1.1028776E+00 ;
SYSTEM.VLE.P(7) := 1.1265880E+00 ;
SYSTEM.VLE.P(8) := 1.1502862E+00 ;
SYSTEM.VLE.P(9) := 1.1739724E+00 ;
SYSTEM.VLE.P(10) := 1.1976465E+00 ;
SYSTEM.VLE.X(1,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(1,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(1,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(1,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(2,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(2,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(2,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(2,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(3,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(3,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(3,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(3,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(4,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(4,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(4,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(4,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(5,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(5,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(5,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(5,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(6,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(6,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(6,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(6,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(7,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(7,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(7,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(7,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(8,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(8,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(8,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(8,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(9,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(9,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(9,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(9,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(10,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(10,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(10,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(10,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(1,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(1,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(1,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(1,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(2,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(2,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(2,3) := 0.0000000E+00 ;

360
SYSTEM.VLE.Y(2,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(3,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(3,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(3,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(3,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(4,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(4,2) := 3.3822172E-24 ;
SYSTEM.VLE.Y(4,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(4,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(5,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(5,2) := 5.2878473E-25 ;
SYSTEM.VLE.Y(5,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(5,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(6,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(6,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(6,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(6,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(7,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(7,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(7,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(7,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(8,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(8,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(8,3) := 5.8123352E-26 ;
SYSTEM.VLE.Y(8,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(9,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(9,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(9,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(9,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(10,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(10,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(10,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(10,4) := 0.0000000E+00 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,1) := -2.1350957E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,2) := -1.4799718E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,3) := -1.9039774E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,4) := -2.5319781E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,1) := -2.1350699E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,2) := -1.4799418E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,3) := -1.9039535E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,4) := -2.5319549E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,1) := -2.1238613E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,2) := -1.4669313E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,3) := -1.8935450E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,4) := -2.5218717E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,1) := -2.1128650E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,2) := -1.4541672E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,3) := -1.8833338E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,4) := -2.5119796E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,1) := -2.1020741E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,2) := -1.4416414E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,3) := -1.8733131E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,4) := -2.5022721E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,1) := -2.0905964E+04 ;

361
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,2) := -1.4283185E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,3) := -1.8626548E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,4) := -2.4919468E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,1) := -2.0793270E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,2) := -1.4152374E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,3) := -1.8521899E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,4) := -2.4818090E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,1) := -2.0682578E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,2) := -1.4023887E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,3) := -1.8419109E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,4) := -2.4718512E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,1) := -2.0573812E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,2) := -1.3897635E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,3) := -1.8318108E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,4) := -2.4620667E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,1) := -2.0466900E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,2) := -1.3773534E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,3) := -1.8218827E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,4) := -2.4524489E+04 ;
SYSTEM.JACKET.TWATER_OUT := 2.9946497E+02 ;
SYSTEM.JACKET.QJACKET := 4.8152866E+04 ;
SYSTEM.JACKET.TEMP_REACTOR := 3.0000000E+02 ;
SYSTEM.LIQENTHALPY.TEMP := 3.0000000E+02 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_VAPOR(1) := 3.4460000E+02 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_VAPOR(2) := 4.0000000E+02 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_VAPOR(3) := 3.2000000E+02 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_VAPOR(4) := 3.1000000E+02 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_LIQUID(1) := -3.0655400E+04 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_LIQUID(2) := -2.5600000E+04 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_LIQUID(3) := -2.7680000E+04 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_LIQUID(4) := -3.3690000E+04 ;

INITIAL
WITHIN System.Reactor DO
# No_Mols(A) = 22.27 ;
# No_Mols(B) = 36.81 ;
No_Mols(A) = 50 ;
No_Mols(B) = 0 ;
No_Mols(C) = 0.0 ;
No_Mols(D) = 0.0 ;
Temp=300 ;
END
WITHIN System.Column DO
FOR I:=2 TO NSTAGE DO
X(I,D)= 0.0 ;
X(I,B)= 0.0 ;
X(I,C)= 0.0 ;
END
TotalHoldup=4.6586382E+01 ;
M(2)=2.9237071E+00 ;
M(3)=2.9189519E+00 ;
M(4)=2.9138424E+00 ;
M(5)=3.1682045E+00 ;
M(6)=3.1664972E+00 ;

362
 M(7)=3.1647940E+00 ;
M(8)=3.1630950E+00 ;
M(9)=3.1614000E+00 ;
FOR I:=2 TO NSTAGE DO
SIGMA(X(I,))=1 ;
#SIGMA(K(I,)*X(I,))=1 ;
END
END

SCHEDULE
CONTINUE FOR 100

END

OPTIMIZATION OptBoth

UNIT
System As BothFlowSheet

VARIABLE
Mols_A_OUT AS MolarFlowRate
Mols_B_OUT AS MolarFlowRate
Mols_C_OUT AS MolarFlowRate
Mols_D_OUT AS MolarFlowRate
ChangeMols_A_Out AS Value
ChangeMols_B_Out AS Value
ChangeMols_C_Out AS Value
ChangeMols_D_Out AS Value
Total_C_OUT AS MolarHoldup
Total_D_OUT AS MolarHoldup
#Final_Time AS Value

OBJECTIVE
MAXIMIZE Total_C_OUT-0.1*Total_D_OUT;
#MAXIMIZE Total_C_OUT;
#MINIMIZE Final_Time ;
#MINIMIZE 100000*(ChangeMols_A_Out^2 +ChangeMols_B_Out^2+
# ChangeMols_C_Out^2 +ChangeMols_D_Out^2);
#MINIMIZE Final_Time+ 100000*(ChangeMols_A_Out^2 +ChangeMols_B_Out^2+
# ChangeMols_C_Out^2 +ChangeMols_D_Out^2+
# System.Reactor.$No_Mols(1)^2+System.Reactor.$No_Mols(2)^2+
# System.Reactor.$No_Mols(3)^2+System.Reactor.$No_Mols(4)^2);

EQUATION

ChangeMols_A_Out= $Mols_A_OUT ;
ChangeMols_B_Out= $Mols_B_OUT ;
ChangeMols_C_Out= $Mols_C_OUT ;
ChangeMols_D_Out= $Mols_D_OUT ;

Mols_A_OUT=System.Column.BOTTOMS*System.Column.X(System.Column.NSTAGE,1) ;

363
 Mols_B_OUT=System.Column.BOTTOMS*System.Column.X(System.Column.NSTAGE,2) ;

Mols_C_OUT=System.Column.BOTTOMS*System.Column.X(System.Column.NSTAGE,3) ;

Mols_D_OUT=System.Column.BOTTOMS*System.Column.X(System.Column.NSTAGE,4) ;

$Total_C_OUT=Mols_C_OUT;

$Total_D_OUT=Mols_D_OUT;

INEQUALITY
System.Jacket.Temp_Reactor<=400 ;

INPUT
WITHIN System.Reactor DO
FlowOut := 15 ;
FeedTemp := 300 ;
END
WITHIN System.Column DO
P(1) :=1.01325 ;
# REFLUXRATIO :=0.5 ;
END
WITHIN System DO
BFraction := .15 ;
END

CONTROL
WITHIN System.Jacket DO
Flow_Water := 3 :0 :5 ;
END

TIME_INVARIANT
WITHIN System.Column DO
REFLUXRATIO:=0.5 : 0.4 : 0.83;
END
#Final_Time:=90:71:500 ;

PRESET
##############################
# Values of All Active Variables #
##############################

SYSTEM.MAKEUP(1) := 7.6341614E-02 ;
SYSTEM.MAKEUP(2) := 7.5000000E+00 ;
SYSTEM.MAKEUP(3) := 0.0000000E+00 ;
SYSTEM.MAKEUP(4) := 0.0000000E+00 ;
SYSTEM.COLUMN.MWV(1) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(2) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(3) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(4) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(5) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(6) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(7) := 7.6000000E+01 ;

364
SYSTEM.COLUMN.MWV(8) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(9) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWV(10) := 7.6000000E+01 ;
SYSTEM.COLUMN.BOTTOMS := 2.2005890E+00 ;
SYSTEM.COLUMN.VAPORENTHALPY(1) := 9.6490429E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(2) := 9.6493013E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(3) := 9.7613871E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(4) := 9.8713496E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(5) := 9.9792593E+03 ;
SYSTEM.COLUMN.VAPORENTHALPY(6) := 1.0094036E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(7) := 1.0206730E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(8) := 1.0317422E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(9) := 1.0426188E+04 ;
SYSTEM.COLUMN.VAPORENTHALPY(10) := 1.0533100E+04 ;
SYSTEM.COLUMN.LIQMDENS(1) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(2) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(3) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(4) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(5) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(6) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(7) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(8) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(9) := 8.3600000E+02 ;
SYSTEM.COLUMN.LIQMDENS(10) := 8.3600000E+02 ;
SYSTEM.COLUMN.VAPMDENS(1) := 2.6164666E+00 ;
SYSTEM.COLUMN.VAPMDENS(2) := 2.6165852E+00 ;
SYSTEM.COLUMN.VAPMDENS(3) := 2.6684553E+00 ;
SYSTEM.COLUMN.VAPMDENS(4) := 2.7201430E+00 ;
SYSTEM.COLUMN.VAPMDENS(5) := 2.7716449E+00 ;
SYSTEM.COLUMN.VAPMDENS(6) := 2.8272809E+00 ;
SYSTEM.COLUMN.VAPMDENS(7) := 2.8827759E+00 ;
SYSTEM.COLUMN.VAPMDENS(8) := 2.9381325E+00 ;
SYSTEM.COLUMN.VAPMDENS(9) := 2.9933533E+00 ;
SYSTEM.COLUMN.VAPMDENS(10) := 3.0484406E+00 ;
SYSTEM.COLUMN.QR := 6.4072745E+06 ;
SYSTEM.COLUMN.DPSTAT(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.DPSTAT(2) := 2.0759991E-02 ;
SYSTEM.COLUMN.DPSTAT(3) := 2.0726226E-02 ;
SYSTEM.COLUMN.DPSTAT(4) := 2.0689946E-02 ;
SYSTEM.COLUMN.DPSTAT(5) := 2.2496062E-02 ;
SYSTEM.COLUMN.DPSTAT(6) := 2.2483940E-02 ;
SYSTEM.COLUMN.DPSTAT(7) := 2.2471846E-02 ;
SYSTEM.COLUMN.DPSTAT(8) := 2.2459782E-02 ;
SYSTEM.COLUMN.DPSTAT(9) := 2.2447747E-02 ;
SYSTEM.COLUMN.DPSTAT(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.XFEED(1) := 1.0000000E+00 ;
SYSTEM.COLUMN.XFEED(2) := 0.0000000E+00 ;
SYSTEM.COLUMN.XFEED(3) := 0.0000000E+00 ;
SYSTEM.COLUMN.XFEED(4) := 0.0000000E+00 ;
SYSTEM.COLUMN.DPTRAY(1) := 5.0229571E-05 ;
SYSTEM.COLUMN.DPTRAY(2) := 2.1986243E-02 ;
SYSTEM.COLUMN.DPTRAY(3) := 2.1952478E-02 ;
SYSTEM.COLUMN.DPTRAY(4) := 2.1916197E-02 ;
SYSTEM.COLUMN.DPTRAY(5) := 2.3722479E-02 ;

365
SYSTEM.COLUMN.DPTRAY(6) := 2.3710350E-02 ;
SYSTEM.COLUMN.DPTRAY(7) := 2.3698251E-02 ;
SYSTEM.COLUMN.DPTRAY(8) := 2.3686181E-02 ;
SYSTEM.COLUMN.DPTRAY(9) := 2.3674140E-02 ;
SYSTEM.COLUMN.DPTRAY(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(1) := -2.1350957E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(2) := -2.1350699E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(3) := -2.1238613E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(4) := -2.1128650E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(5) := -2.1020741E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(6) := -2.0905964E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(7) := -2.0793270E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(8) := -2.0682578E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(9) := -2.0573812E+04 ;
SYSTEM.COLUMN.LIQUIDENTHALPY(10) := -2.0466900E+04 ;
SYSTEM.COLUMN.LIQHEIGHT(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.LIQHEIGHT(2) := 2.5313481E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(3) := 2.5272311E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(4) := 2.5228073E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(5) := 2.7430342E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(6) := 2.7415560E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(7) := 2.7400814E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(8) := 2.7386104E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(9) := 2.7371429E-01 ;
SYSTEM.COLUMN.LIQHEIGHT(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.HEAD(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.HEAD(2) := 3.1348139E+00 ;
SYSTEM.COLUMN.HEAD(3) := 2.7231082E+00 ;
SYSTEM.COLUMN.HEAD(4) := 2.2807273E+00 ;
SYSTEM.COLUMN.HEAD(5) := 2.4303420E+01 ;
SYSTEM.COLUMN.HEAD(6) := 2.4155602E+01 ;
SYSTEM.COLUMN.HEAD(7) := 2.4008139E+01 ;
SYSTEM.COLUMN.HEAD(8) := 2.3861039E+01 ;
SYSTEM.COLUMN.HEAD(9) := 2.3714286E+01 ;
SYSTEM.COLUMN.HEAD(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.K(1,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(1,2) := 1.1186637E+00 ;
SYSTEM.COLUMN.K(1,3) := 3.6783189E-01 ;
SYSTEM.COLUMN.K(1,4) := 2.0422134E-01 ;
SYSTEM.COLUMN.K(2,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(2,2) := 1.1186548E+00 ;
SYSTEM.COLUMN.K(2,3) := 3.6782906E-01 ;
SYSTEM.COLUMN.K(2,4) := 2.0422232E-01 ;
SYSTEM.COLUMN.K(3,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(3,2) := 1.1147884E+00 ;
SYSTEM.COLUMN.K(3,3) := 3.6660608E-01 ;
SYSTEM.COLUMN.K(3,4) := 2.0464705E-01 ;
SYSTEM.COLUMN.K(4,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(4,2) := 1.1110220E+00 ;
SYSTEM.COLUMN.K(4,3) := 3.6541455E-01 ;
SYSTEM.COLUMN.K(4,4) := 2.0506307E-01 ;
SYSTEM.COLUMN.K(5,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(5,2) := 1.1073514E+00 ;
SYSTEM.COLUMN.K(5,3) := 3.6425316E-01 ;

366
SYSTEM.COLUMN.K(5,4) := 2.0547070E-01 ;
SYSTEM.COLUMN.K(6,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(6,2) := 1.1034745E+00 ;
SYSTEM.COLUMN.K(6,3) := 3.6302637E-01 ;
SYSTEM.COLUMN.K(6,4) := 2.0590358E-01 ;
SYSTEM.COLUMN.K(7,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(7,2) := 1.0996952E+00 ;
SYSTEM.COLUMN.K(7,3) := 3.6183030E-01 ;
SYSTEM.COLUMN.K(7,4) := 2.0632791E-01 ;
SYSTEM.COLUMN.K(8,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(8,2) := 1.0960090E+00 ;
SYSTEM.COLUMN.K(8,3) := 3.6066354E-01 ;
SYSTEM.COLUMN.K(8,4) := 2.0674404E-01 ;
SYSTEM.COLUMN.K(9,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(9,2) := 1.0924118E+00 ;
SYSTEM.COLUMN.K(9,3) := 3.5952480E-01 ;
SYSTEM.COLUMN.K(9,4) := 2.0715229E-01 ;
SYSTEM.COLUMN.K(10,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.K(10,2) := 1.0888996E+00 ;
SYSTEM.COLUMN.K(10,3) := 3.5841286E-01 ;
SYSTEM.COLUMN.K(10,4) := 2.0755297E-01 ;
SYSTEM.COLUMN.L(1) := 1.2799411E+01 ;
SYSTEM.COLUMN.L(2) := 1.0700915E+01 ;
SYSTEM.COLUMN.L(3) := 8.6636449E+00 ;
SYSTEM.COLUMN.L(4) := 6.6406918E+00 ;
SYSTEM.COLUMN.L(5) := 2.3099596E+02 ;
SYSTEM.COLUMN.L(6) := 2.2889172E+02 ;
SYSTEM.COLUMN.L(7) := 2.2679894E+02 ;
SYSTEM.COLUMN.L(8) := 2.2471772E+02 ;
SYSTEM.COLUMN.L(9) := 2.2264778E+02 ;
SYSTEM.COLUMN.L(10) := 2.2005890E+00 ;
SYSTEM.COLUMN.FEED(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(2) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(3) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(4) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(5) := 1.5000000E+01 ;
SYSTEM.COLUMN.FEED(6) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(7) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(8) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(9) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEED(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.M(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.$M(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.M(2) := 2.9237071E+00 ;
SYSTEM.COLUMN.$M(2) := -5.8431630E-05 ;
SYSTEM.COLUMN.M(3) := 2.9189519E+00 ;
SYSTEM.COLUMN.$M(3) := 1.9291184E-05 ;
SYSTEM.COLUMN.M(4) := 2.9138424E+00 ;
SYSTEM.COLUMN.$M(4) := 8.9519949E-05 ;
SYSTEM.COLUMN.M(5) := 3.1682045E+00 ;
SYSTEM.COLUMN.$M(5) := -1.1165674E-04 ;
SYSTEM.COLUMN.M(6) := 3.1664972E+00 ;
SYSTEM.COLUMN.$M(6) := -1.0041827E-04 ;
SYSTEM.COLUMN.M(7) := 3.1647940E+00 ;

367
SYSTEM.COLUMN.$M(7) := 3.0944158E-04 ;
SYSTEM.COLUMN.M(8) := 3.1630950E+00 ;
SYSTEM.COLUMN.$M(8) := -5.6979389E-05 ;
SYSTEM.COLUMN.M(9) := 3.1614000E+00 ;
SYSTEM.COLUMN.$M(9) := -7.1432749E-05 ;
SYSTEM.COLUMN.M(10) := 2.2005890E+01 ;
SYSTEM.COLUMN.$M(10) := -1.9333939E-05 ;
SYSTEM.COLUMN.LIQDENS(1) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(2) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(3) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(4) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(5) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(6) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(7) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(8) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(9) := 1.1000000E+01 ;
SYSTEM.COLUMN.LIQDENS(10) := 1.1000000E+01 ;
SYSTEM.COLUMN.VAPDENS(1) := 3.4427192E-02 ;
SYSTEM.COLUMN.VAPDENS(2) := 3.4428753E-02 ;
SYSTEM.COLUMN.VAPDENS(3) := 3.5111254E-02 ;
SYSTEM.COLUMN.VAPDENS(4) := 3.5791355E-02 ;
SYSTEM.COLUMN.VAPDENS(5) := 3.6469012E-02 ;
SYSTEM.COLUMN.VAPDENS(6) := 3.7201064E-02 ;
SYSTEM.COLUMN.VAPDENS(7) := 3.7931261E-02 ;
SYSTEM.COLUMN.VAPDENS(8) := 3.8659638E-02 ;
SYSTEM.COLUMN.VAPDENS(9) := 3.9386227E-02 ;
SYSTEM.COLUMN.VAPDENS(10) := 4.0111061E-02 ;
SYSTEM.COLUMN.P(2) := 1.0133002E+00 ;
SYSTEM.COLUMN.P(3) := 1.0352865E+00 ;
SYSTEM.COLUMN.P(4) := 1.0572390E+00 ;
SYSTEM.COLUMN.P(5) := 1.0791551E+00 ;
SYSTEM.COLUMN.P(6) := 1.1028776E+00 ;
SYSTEM.COLUMN.P(7) := 1.1265880E+00 ;
SYSTEM.COLUMN.P(8) := 1.1502862E+00 ;
SYSTEM.COLUMN.P(9) := 1.1739724E+00 ;
SYSTEM.COLUMN.P(10) := 1.1976465E+00 ;
SYSTEM.COLUMN.DISTILLOUT := 1.2799411E+01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(2) := 2.6579155E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(3) := 2.6535926E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(4) := 2.6489476E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(5) := 2.8801859E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(6) := 2.8786338E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(7) := 2.8770855E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(8) := 2.8755409E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(9) := 2.8740000E-01 ;
SYSTEM.COLUMN.VOLUMEHOLDUP(10) := 2.0005354E+00 ;
SYSTEM.COLUMN.T(1) := 3.5400141E+02 ;
SYSTEM.COLUMN.T(2) := 3.5400291E+02 ;
SYSTEM.COLUMN.T(3) := 3.5465344E+02 ;
SYSTEM.COLUMN.T(4) := 3.5529164E+02 ;
SYSTEM.COLUMN.T(5) := 3.5591793E+02 ;
SYSTEM.COLUMN.T(6) := 3.5658408E+02 ;
SYSTEM.COLUMN.T(7) := 3.5723813E+02 ;

368
SYSTEM.COLUMN.T(8) := 3.5788057E+02 ;
SYSTEM.COLUMN.T(9) := 3.5851183E+02 ;
SYSTEM.COLUMN.T(10) := 3.5913233E+02 ;
SYSTEM.COLUMN.V(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.V(2) := 2.5598822E+01 ;
SYSTEM.COLUMN.V(3) := 2.3500268E+01 ;
SYSTEM.COLUMN.V(4) := 2.1463017E+01 ;
SYSTEM.COLUMN.V(5) := 1.9440153E+01 ;
SYSTEM.COLUMN.V(6) := 2.2879531E+02 ;
SYSTEM.COLUMN.V(7) := 2.2669097E+02 ;
SYSTEM.COLUMN.V(8) := 2.2459850E+02 ;
SYSTEM.COLUMN.V(9) := 2.2251722E+02 ;
SYSTEM.COLUMN.V(10) := 2.2044721E+02 ;
#SYSTEM.COLUMN.FEEDTEMP := 3.0000000E+02 ;
SYSTEM.COLUMN.DPDRY(1) := 0.0000000E+00 ;
SYSTEM.COLUMN.DPDRY(2) := 1.8678982E-09 ;
SYSTEM.COLUMN.DPDRY(3) := 1.5277914E-09 ;
SYSTEM.COLUMN.DPDRY(4) := 1.2295610E-09 ;
SYSTEM.COLUMN.DPDRY(5) := 1.6714719E-07 ;
SYSTEM.COLUMN.DPDRY(6) := 1.6085773E-07 ;
SYSTEM.COLUMN.DPDRY(7) := 1.5486215E-07 ;
SYSTEM.COLUMN.DPDRY(8) := 1.4914144E-07 ;
SYSTEM.COLUMN.DPDRY(9) := 1.4367913E-07 ;
SYSTEM.COLUMN.DPDRY(10) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(1,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.X(1,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(1,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(1,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(2,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(2,1) := 1.0816707E-21 ;
SYSTEM.COLUMN.X(2,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(2,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(2,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(2,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(2,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(2,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(3,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(3,1) := -1.5264713E-22 ;
SYSTEM.COLUMN.X(3,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(3,2) := 2.4869401E-23 ;
SYSTEM.COLUMN.X(3,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(3,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(3,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(3,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(4,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(4,1) := 2.1141297E-21 ;
SYSTEM.COLUMN.X(4,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(4,2) := -2.1385140E-23 ;
SYSTEM.COLUMN.X(4,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(4,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(4,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(4,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(5,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(5,1) := -1.2429451E-21 ;

369
SYSTEM.COLUMN.X(5,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(5,2) := -3.2446314E-24 ;
SYSTEM.COLUMN.X(5,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(5,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(5,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(5,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(6,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(6,1) := 8.0887492E-22 ;
SYSTEM.COLUMN.X(6,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(6,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(6,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(6,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(6,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(6,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(7,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(7,1) := -6.1828866E-22 ;
SYSTEM.COLUMN.X(7,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(7,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(7,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(7,3) := 4.1248869E-24 ;
SYSTEM.COLUMN.X(7,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(7,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(8,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(8,1) := 1.2278487E-21 ;
SYSTEM.COLUMN.X(8,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(8,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(8,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(8,3) := -4.1271025E-24 ;
SYSTEM.COLUMN.X(8,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(8,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(9,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(9,1) := 1.0411440E-21 ;
SYSTEM.COLUMN.X(9,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(9,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(9,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(9,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(9,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(9,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(10,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.$X(10,1) := -8.9506882E-23 ;
SYSTEM.COLUMN.X(10,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(10,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(10,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(10,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.X(10,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.$X(10,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(1,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(1,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(1,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(1,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(2,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(2,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(2,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(2,4) := 0.0000000E+00 ;

370
SYSTEM.COLUMN.Y(3,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(3,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(3,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(3,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(4,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(4,2) := 3.3822172E-24 ;
SYSTEM.COLUMN.Y(4,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(4,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(5,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(5,2) := 5.2878473E-25 ;
SYSTEM.COLUMN.Y(5,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(5,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(6,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(6,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(6,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(6,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(7,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(7,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(7,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(7,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(8,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(8,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(8,3) := 5.8123352E-26 ;
SYSTEM.COLUMN.Y(8,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(9,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(9,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(9,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(9,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(10,1) := 1.0000000E+00 ;
SYSTEM.COLUMN.Y(10,2) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(10,3) := 0.0000000E+00 ;
SYSTEM.COLUMN.Y(10,4) := 0.0000000E+00 ;
SYSTEM.COLUMN.FEEDENTHALPY := -4.5983100E+05 ;
SYSTEM.COLUMN.MWL(1) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(2) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(3) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(4) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(5) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(6) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(7) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(8) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(9) := 7.6000000E+01 ;
SYSTEM.COLUMN.MWL(10) := 7.6000000E+01 ;
SYSTEM.COLUMN.TOTALHOLDUP := 4.6586382E+01 ;
SYSTEM.COLUMN.$TOTALHOLDUP := 0.0000000E+00 ;
SYSTEM.COLUMN.QC := 7.9357010E+05 ;
SYSTEM.MAKEUPENTHALPY.TEMP := 3.0000000E+02 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_VAPOR(1) := 3.4460000E+02 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_VAPOR(2) := 4.0000000E+02 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_VAPOR(3) := 3.2000000E+02 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_VAPOR(4) := 3.1000000E+02 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_LIQUID(1) := -3.0655400E+04 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_LIQUID(2) := -2.5600000E+04 ;
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_LIQUID(3) := -2.7680000E+04 ;

371
SYSTEM.MAKEUPENTHALPY.SPECIFIC_ENTHALPY_LIQUID(4) := -3.3690000E+04 ;
SYSTEM.REACTOR.TOTALMOLS := 5.0000000E+01 ;
SYSTEM.REACTOR.$TOTALMOLS := 0.0000000E+00 ;
SYSTEM.REACTOR.TEMP := 3.0000000E+02 ;
SYSTEM.REACTOR.ENTHALPY := -1.5327700E+06 ;
SYSTEM.REACTOR.$ENTHALPY := -4.3344954E+06 ;
SYSTEM.REACTOR.TOTALFEED := 1.5000000E+01 ;
SYSTEM.REACTOR.QJACKET := 4.8152866E+04 ;
SYSTEM.REACTOR.NO_MOLS(1) := 5.0000000E+01 ;
SYSTEM.REACTOR.$NO_MOLS(1) := -7.5000000E+00 ;
SYSTEM.REACTOR.NO_MOLS(2) := 0.0000000E+00 ;
SYSTEM.REACTOR.$NO_MOLS(2) := 7.5000000E+00 ;
SYSTEM.REACTOR.NO_MOLS(3) := 0.0000000E+00 ;
SYSTEM.REACTOR.$NO_MOLS(3) := 0.0000000E+00 ;
SYSTEM.REACTOR.NO_MOLS(4) := 0.0000000E+00 ;
SYSTEM.REACTOR.$NO_MOLS(4) := 0.0000000E+00 ;
SYSTEM.REACTOR.VOLUME := 4.5454545E+00 ;
SYSTEM.REACTOR.CONCENTRATION(1) := 1.1000000E+01 ;
SYSTEM.REACTOR.CONCENTRATION(2) := 0.0000000E+00 ;
SYSTEM.REACTOR.CONCENTRATION(3) := 0.0000000E+00 ;
SYSTEM.REACTOR.CONCENTRATION(4) := 0.0000000E+00 ;
SYSTEM.REACTOR.FEED_A := 7.5000000E+00 ;
SYSTEM.REACTOR.FEED_B := 7.5000000E+00 ;
SYSTEM.REACTOR.FEED_C := 0.0000000E+00 ;
SYSTEM.REACTOR.X(1) := 1.0000000E+00 ;
SYSTEM.REACTOR.X(2) := 0.0000000E+00 ;
SYSTEM.REACTOR.X(3) := 0.0000000E+00 ;
SYSTEM.REACTOR.X(4) := 0.0000000E+00 ;
SYSTEM.REACTOR.FEED_D := 0.0000000E+00 ;
SYSTEM.REACTOR.FEEDENTHALPY(1) := -5.0553546E+05 ;
SYSTEM.REACTOR.FEEDENTHALPY(2) := -1.9395435E+05 ;
SYSTEM.REACTOR.FEEDENTHALPY(3) := -2.0971314E+05 ;
SYSTEM.REACTOR.FEEDENTHALPY(4) := -2.5524695E+05 ;
SYSTEM.REACTOR.SPECIFIC_ENTHALPY(1) := -3.0655400E+04 ;
SYSTEM.REACTOR.SPECIFIC_ENTHALPY(2) := -2.5600000E+04 ;
SYSTEM.REACTOR.SPECIFIC_ENTHALPY(3) := -2.7680000E+04 ;
SYSTEM.REACTOR.SPECIFIC_ENTHALPY(4) := -3.3690000E+04 ;
SYSTEM.REACTOR.REACTIONRATE(1) := 0.0000000E+00 ;
SYSTEM.REACTOR.REACTIONRATE(2) := 0.0000000E+00 ;
SYSTEM.VLE.VAPORENTHALPY(1) := 9.6490429E+03 ;
SYSTEM.VLE.VAPORENTHALPY(2) := 9.6493013E+03 ;
SYSTEM.VLE.VAPORENTHALPY(3) := 9.7613871E+03 ;
SYSTEM.VLE.VAPORENTHALPY(4) := 9.8713496E+03 ;
SYSTEM.VLE.VAPORENTHALPY(5) := 9.9792593E+03 ;
SYSTEM.VLE.VAPORENTHALPY(6) := 1.0094036E+04 ;
SYSTEM.VLE.VAPORENTHALPY(7) := 1.0206730E+04 ;
SYSTEM.VLE.VAPORENTHALPY(8) := 1.0317422E+04 ;
SYSTEM.VLE.VAPORENTHALPY(9) := 1.0426188E+04 ;
SYSTEM.VLE.VAPORENTHALPY(10) := 1.0533100E+04 ;
SYSTEM.VLE.TEMP(1) := 3.5400141E+02 ;
SYSTEM.VLE.TEMP(2) := 3.5400291E+02 ;
SYSTEM.VLE.TEMP(3) := 3.5465344E+02 ;
SYSTEM.VLE.TEMP(4) := 3.5529164E+02 ;
SYSTEM.VLE.TEMP(5) := 3.5591793E+02 ;

372
SYSTEM.VLE.TEMP(6) := 3.5658408E+02 ;
SYSTEM.VLE.TEMP(7) := 3.5723813E+02 ;
SYSTEM.VLE.TEMP(8) := 3.5788057E+02 ;
SYSTEM.VLE.TEMP(9) := 3.5851183E+02 ;
SYSTEM.VLE.TEMP(10) := 3.5913233E+02 ;
SYSTEM.VLE.LIQUIDENTHALPY(1) := -2.1350957E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(2) := -2.1350699E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(3) := -2.1238613E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(4) := -2.1128650E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(5) := -2.1020741E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(6) := -2.0905964E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(7) := -2.0793270E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(8) := -2.0682578E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(9) := -2.0573812E+04 ;
SYSTEM.VLE.LIQUIDENTHALPY(10) := -2.0466900E+04 ;
SYSTEM.VLE.K(1,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(1,2) := 1.1186637E+00 ;
SYSTEM.VLE.K(1,3) := 3.6783189E-01 ;
SYSTEM.VLE.K(1,4) := 2.0422134E-01 ;
SYSTEM.VLE.K(2,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(2,2) := 1.1186548E+00 ;
SYSTEM.VLE.K(2,3) := 3.6782906E-01 ;
SYSTEM.VLE.K(2,4) := 2.0422232E-01 ;
SYSTEM.VLE.K(3,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(3,2) := 1.1147884E+00 ;
SYSTEM.VLE.K(3,3) := 3.6660608E-01 ;
SYSTEM.VLE.K(3,4) := 2.0464705E-01 ;
SYSTEM.VLE.K(4,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(4,2) := 1.1110220E+00 ;
SYSTEM.VLE.K(4,3) := 3.6541455E-01 ;
SYSTEM.VLE.K(4,4) := 2.0506307E-01 ;
SYSTEM.VLE.K(5,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(5,2) := 1.1073514E+00 ;
SYSTEM.VLE.K(5,3) := 3.6425316E-01 ;
SYSTEM.VLE.K(5,4) := 2.0547070E-01 ;
SYSTEM.VLE.K(6,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(6,2) := 1.1034745E+00 ;
SYSTEM.VLE.K(6,3) := 3.6302637E-01 ;
SYSTEM.VLE.K(6,4) := 2.0590358E-01 ;
SYSTEM.VLE.K(7,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(7,2) := 1.0996952E+00 ;
SYSTEM.VLE.K(7,3) := 3.6183030E-01 ;
SYSTEM.VLE.K(7,4) := 2.0632791E-01 ;
SYSTEM.VLE.K(8,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(8,2) := 1.0960090E+00 ;
SYSTEM.VLE.K(8,3) := 3.6066354E-01 ;
SYSTEM.VLE.K(8,4) := 2.0674404E-01 ;
SYSTEM.VLE.K(9,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(9,2) := 1.0924118E+00 ;
SYSTEM.VLE.K(9,3) := 3.5952480E-01 ;
SYSTEM.VLE.K(9,4) := 2.0715229E-01 ;
SYSTEM.VLE.K(10,1) := 1.0000000E+00 ;
SYSTEM.VLE.K(10,2) := 1.0888996E+00 ;
SYSTEM.VLE.K(10,3) := 3.5841286E-01 ;

373
SYSTEM.VLE.K(10,4) := 2.0755297E-01 ;
SYSTEM.VLE.VAPORPRESSURE(1,1) := 1.0132500E+00 ;
SYSTEM.VLE.VAPORPRESSURE(1,2) := 1.1334860E+00 ;
SYSTEM.VLE.VAPORPRESSURE(1,3) := 3.7270566E-01 ;
SYSTEM.VLE.VAPORPRESSURE(1,4) := 2.0692728E-01 ;
SYSTEM.VLE.VAPORPRESSURE(2,1) := 1.0133002E+00 ;
SYSTEM.VLE.VAPORPRESSURE(2,2) := 1.1335331E+00 ;
SYSTEM.VLE.VAPORPRESSURE(2,3) := 3.7272127E-01 ;
SYSTEM.VLE.VAPORPRESSURE(2,4) := 2.0693853E-01 ;
SYSTEM.VLE.VAPORPRESSURE(3,1) := 1.0352865E+00 ;
SYSTEM.VLE.VAPORPRESSURE(3,2) := 1.1541254E+00 ;
SYSTEM.VLE.VAPORPRESSURE(3,3) := 3.7954232E-01 ;
SYSTEM.VLE.VAPORPRESSURE(3,4) := 2.1186832E-01 ;
SYSTEM.VLE.VAPORPRESSURE(4,1) := 1.0572390E+00 ;
SYSTEM.VLE.VAPORPRESSURE(4,2) := 1.1746158E+00 ;
SYSTEM.VLE.VAPORPRESSURE(4,3) := 3.8633050E-01 ;
SYSTEM.VLE.VAPORPRESSURE(4,4) := 2.1680067E-01 ;
SYSTEM.VLE.VAPORPRESSURE(5,1) := 1.0791551E+00 ;
SYSTEM.VLE.VAPORPRESSURE(5,2) := 1.1950039E+00 ;
SYSTEM.VLE.VAPORPRESSURE(5,3) := 3.9308568E-01 ;
SYSTEM.VLE.VAPORPRESSURE(5,4) := 2.2173476E-01 ;
SYSTEM.VLE.VAPORPRESSURE(6,1) := 1.1028776E+00 ;
SYSTEM.VLE.VAPORPRESSURE(6,2) := 1.2169973E+00 ;
SYSTEM.VLE.VAPORPRESSURE(6,3) := 4.0037366E-01 ;
SYSTEM.VLE.VAPORPRESSURE(6,4) := 2.2708645E-01 ;
SYSTEM.VLE.VAPORPRESSURE(7,1) := 1.1265880E+00 ;
SYSTEM.VLE.VAPORPRESSURE(7,2) := 1.2389034E+00 ;
SYSTEM.VLE.VAPORPRESSURE(7,3) := 4.0763366E-01 ;
SYSTEM.VLE.VAPORPRESSURE(7,4) := 2.3244654E-01 ;
SYSTEM.VLE.VAPORPRESSURE(8,1) := 1.1502862E+00 ;
SYSTEM.VLE.VAPORPRESSURE(8,2) := 1.2607240E+00 ;
SYSTEM.VLE.VAPORPRESSURE(8,3) := 4.1486630E-01 ;
SYSTEM.VLE.VAPORPRESSURE(8,4) := 2.3781482E-01 ;
SYSTEM.VLE.VAPORPRESSURE(9,1) := 1.1739724E+00 ;
SYSTEM.VLE.VAPORPRESSURE(9,2) := 1.2824613E+00 ;
SYSTEM.VLE.VAPORPRESSURE(9,3) := 4.2207220E-01 ;
SYSTEM.VLE.VAPORPRESSURE(9,4) := 2.4319108E-01 ;
SYSTEM.VLE.VAPORPRESSURE(10,1) := 1.1976465E+00 ;
SYSTEM.VLE.VAPORPRESSURE(10,2) := 1.3041169E+00 ;
SYSTEM.VLE.VAPORPRESSURE(10,3) := 4.2925192E-01 ;
SYSTEM.VLE.VAPORPRESSURE(10,4) := 2.4857510E-01 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,1) := 9.6490429E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,2) := 1.1200282E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,3) := 8.9602256E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(1,4) := 8.6802185E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,1) := 9.6493013E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,2) := 1.1200582E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,3) := 8.9604655E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(2,4) := 8.6804509E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,1) := 9.7613871E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,2) := 1.1330687E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,3) := 9.0645499E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(3,4) := 8.7812827E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,1) := 9.8713496E+03 ;

374
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,2) := 1.1458328E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,3) := 9.1666624E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(4,4) := 8.8802042E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,1) := 9.9792593E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,2) := 1.1583586E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,3) := 9.2668688E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(5,4) := 8.9772791E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,1) := 1.0094036E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,2) := 1.1716815E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,3) := 9.3734520E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(6,4) := 9.0805316E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,1) := 1.0206730E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,2) := 1.1847626E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,3) := 9.4781009E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(7,4) := 9.1819102E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,1) := 1.0317422E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,2) := 1.1976113E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,3) := 9.5808907E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(8,4) := 9.2814879E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,1) := 1.0426188E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,2) := 1.2102365E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,3) := 9.6818924E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(9,4) := 9.3793332E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,1) := 1.0533100E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,2) := 1.2226466E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,3) := 9.7811726E+03 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_VAPOR(10,4) := 9.4755110E+03 ;
SYSTEM.VLE.P(1) := 1.0132500E+00 ;
SYSTEM.VLE.P(2) := 1.0133002E+00 ;
SYSTEM.VLE.P(3) := 1.0352865E+00 ;
SYSTEM.VLE.P(4) := 1.0572390E+00 ;
SYSTEM.VLE.P(5) := 1.0791551E+00 ;
SYSTEM.VLE.P(6) := 1.1028776E+00 ;
SYSTEM.VLE.P(7) := 1.1265880E+00 ;
SYSTEM.VLE.P(8) := 1.1502862E+00 ;
SYSTEM.VLE.P(9) := 1.1739724E+00 ;
SYSTEM.VLE.P(10) := 1.1976465E+00 ;
SYSTEM.VLE.X(1,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(1,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(1,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(1,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(2,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(2,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(2,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(2,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(3,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(3,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(3,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(3,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(4,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(4,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(4,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(4,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(5,1) := 1.0000000E+00 ;

375
SYSTEM.VLE.X(5,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(5,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(5,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(6,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(6,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(6,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(6,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(7,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(7,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(7,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(7,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(8,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(8,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(8,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(8,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(9,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(9,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(9,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(9,4) := 0.0000000E+00 ;
SYSTEM.VLE.X(10,1) := 1.0000000E+00 ;
SYSTEM.VLE.X(10,2) := 0.0000000E+00 ;
SYSTEM.VLE.X(10,3) := 0.0000000E+00 ;
SYSTEM.VLE.X(10,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(1,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(1,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(1,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(1,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(2,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(2,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(2,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(2,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(3,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(3,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(3,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(3,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(4,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(4,2) := 3.3822172E-24 ;
SYSTEM.VLE.Y(4,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(4,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(5,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(5,2) := 5.2878473E-25 ;
SYSTEM.VLE.Y(5,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(5,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(6,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(6,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(6,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(6,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(7,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(7,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(7,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(7,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(8,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(8,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(8,3) := 5.8123352E-26 ;

376
SYSTEM.VLE.Y(8,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(9,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(9,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(9,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(9,4) := 0.0000000E+00 ;
SYSTEM.VLE.Y(10,1) := 1.0000000E+00 ;
SYSTEM.VLE.Y(10,2) := 0.0000000E+00 ;
SYSTEM.VLE.Y(10,3) := 0.0000000E+00 ;
SYSTEM.VLE.Y(10,4) := 0.0000000E+00 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,1) := -2.1350957E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,2) := -1.4799718E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,3) := -1.9039774E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(1,4) := -2.5319781E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,1) := -2.1350699E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,2) := -1.4799418E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,3) := -1.9039535E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(2,4) := -2.5319549E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,1) := -2.1238613E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,2) := -1.4669313E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,3) := -1.8935450E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(3,4) := -2.5218717E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,1) := -2.1128650E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,2) := -1.4541672E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,3) := -1.8833338E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(4,4) := -2.5119796E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,1) := -2.1020741E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,2) := -1.4416414E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,3) := -1.8733131E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(5,4) := -2.5022721E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,1) := -2.0905964E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,2) := -1.4283185E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,3) := -1.8626548E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(6,4) := -2.4919468E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,1) := -2.0793270E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,2) := -1.4152374E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,3) := -1.8521899E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(7,4) := -2.4818090E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,1) := -2.0682578E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,2) := -1.4023887E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,3) := -1.8419109E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(8,4) := -2.4718512E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,1) := -2.0573812E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,2) := -1.3897635E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,3) := -1.8318108E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(9,4) := -2.4620667E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,1) := -2.0466900E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,2) := -1.3773534E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,3) := -1.8218827E+04 ;
SYSTEM.VLE.SPECIFIC_ENTHALPY_LIQUID(10,4) := -2.4524489E+04 ;
SYSTEM.JACKET.TWATER_OUT := 2.9946497E+02 ;
SYSTEM.JACKET.QJACKET := 4.8152866E+04 ;
SYSTEM.JACKET.TEMP_REACTOR := 3.0000000E+02 ;
SYSTEM.LIQENTHALPY.TEMP := 3.0000000E+02 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_VAPOR(1) := 3.4460000E+02 ;

377
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_VAPOR(2) := 4.0000000E+02 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_VAPOR(3) := 3.2000000E+02 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_VAPOR(4) := 3.1000000E+02 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_LIQUID(1) := -3.0655400E+04 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_LIQUID(2) := -2.5600000E+04 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_LIQUID(3) := -2.7680000E+04 ;
SYSTEM.LIQENTHALPY.SPECIFIC_ENTHALPY_LIQUID(4) := -3.3690000E+04 ;

INITIAL
TOTAL_C_OUT = 0;
TOTAL_D_OUT = 0;
WITHIN System.Reactor DO
No_Mols(A) = 50 ;
No_Mols(B) = 0 ;
No_Mols(C) = 0.0 ;
No_Mols(D) = 0.0 ;
Temp=300 ;
END
WITHIN System.Column DO
#FOR I:=2 TO NSTAGE-1 DO
# $M(I)=0;
#END
FOR I:=2 TO NSTAGE DO
X(I,D)= 0.0 ;
X(I,B)= 0.0 ;
X(I,C)= 0.0 ;
END
TotalHoldup=4.6586382E+01 ;
M(2)=2.9237071E+00 ;
M(3)=2.9189519E+00 ;
M(4)=2.9138424E+00 ;
M(5)=3.1682045E+00 ;
M(6)=3.1664972E+00 ;
M(7)=3.1647940E+00 ;
M(8)=3.1630950E+00 ;
M(9)=3.1614000E+00 ;
FOR I:=2 TO NSTAGE DO
SIGMA(X(I,))=1 ;
#SIGMA(K(I,)*X(I,))=1 ;
END
END

# FINAL
#TOTAL_C_OUT>=130 ;
#ChangeMols_A_Out^2 +ChangeMols_B_Out^2+
# ChangeMols_C_Out^2 +ChangeMols_D_Out^2<=0.00001;
#SYSTEM.REACTOR.$ENTHALPY<=0.01 ;
#SYSTEM.REACTOR.$NO_Mols(1)<=0.01 ;
#SYSTEM.REACTOR.$NO_Mols(2)<=0.01 ;
#SYSTEM.REACTOR.$NO_Mols(3)<=0.01 ;
#SYSTEM.REACTOR.$NO_Mols(4)<=0.01 ;
#MOLS_D_OUT<=0.065 ;

SCHEDULE

378
 CONTINUE FOR 100
#CONTINUE FOR FINAL_TIME

END

379
380
References

[1] R. Allgor, Modeling and Computational Issues in the Development of Batch

Processes, PhD thesis, Department of Chemical Engineering, Massachusetts
Institute of Technology, Cambridge, MA, 1997.

[2] R. Allgor and P. Barton, Accurate solution of batch distillation models:

Implications of ill-conditioned corrector iterations, AIChE 1995 Annual Meet-
ing, Miami, (1995).

[3] , Mixed-integer dynamic optimization I: Problem formulation, Computers

chem. Engng., (submitted, 1997).

[4] U. Ascher and L. Petzold, Collocation software for boundary-value

differential-algebraic equations, SIAM J. Sci. Comput., 15 (1994), pp. 938–952.

[5] R. Bachman, L. Brüll, T. Mrziglod, and U. Pallaske, On methods for

reducing the index of differential algebraic equations, Computers chem. Engng.,
14 (1990), pp. 1271–1273.

[6] A. Back, J. Guckenheimer, and M. Myers, A dynamical simulation fa-

cility for hybrid systems, Hybrid Systems, Lecture Notes in Computer Science,
736 (1993), pp. 255–267.

[7] J. Banga and J. Casares, Integrated controlled random search: Application

to a wastewater treatment plant model, Inst. Chem. Eng. Symp. Ser., 100 (1987),
pp. 183–192.

[8] J. Banga and W. Seider, Global optimization of chemical processes using

stochastic algorithms, in State of the Art in Global Optimization: Computa-
tional Methods and Applications, Princeton, NJ, April 1995, Princeton Univer-
sity.

[9] J. Barlow and U. Vemulapati, Rank detection methods for sparse matrices,
SIAM J. Matrix Anal. Appl., 13 (1992), pp. 1279–1297.

[10] M. Barrera and L. Evans, Optimal design and operation of batch processes,
Chem. Eng. Comm., 82 (1989), pp. 45–66.

381
[11] A. Barrlund and B. Kägström, Analytical and numerical solutions to
higher index linear variable coefficient DAE systems, J. Comp. Appl. Math, 31
(1990), pp. 305–330.

[12] M. Bartholomew-Biggs, A penalty method for point and path state con-
straints in trajectory optimization, Opt. Cont. Appl. and Meth., 16 (1995),
pp. 291–297.

[13] P. Barton, The Modelling and Simulation of Combined Discrete/Continuous

Processes, PhD thesis, University of London, London, U.K., 1992.

[14] P. Barton, R. Allgor, W. Feehery, and S. Galán, Dynamic optimiza-

tion in a discontinuous world, Ind. Eng. Chem. Res., 37 (1998).

[15] P. Barton and C. Pantelides, Modeling of combined discrete/continuous

processes, AIChE Journal, 40 (1994), pp. 966–979.

[16] M. Bazaraa, H. Sherali, and C. Shetty, Nonlinear Programming: The-

ory and Applications, Wiley, New York, 1993.

[17] J. Beaumgarte, Stabilization of constraints and numerical solutions to higher

index linear variable coefficient DAE systems, J. Comp. Appl. Math., 31 (1990),
pp. 305–330.

[18] R. Bellman, Dynamic Programming, Princeton University Press, Princeton,

New Jersey, 1957.

[19] T. Bhatia and L. Biegler, Dynamic optimization in the planning and

scheduling of multi-product batch plants, Ind. Eng. Chem. Res., 35 (1996),
pp. 2234–2246.

[20] J. Boston, H. Britt, S. Jiraponghan, and V. Shah, An advanced system

for simulation of batch distillation columns, in Proceedings of the Conference on
Computer-Aided Process Design, R. Mah and W. Seader, eds., 1981, pp. 203–
207.

[21] K. Brenan, S. Campbell, and L. Petzold, Numerical Solution of Initial

Value Problems in Differential-Algebraic Equations, SIAM, Philadelphia, PA,
1996.

[22] L. Brüll and R. Pallaske, On consistent initialization of differential-

algebraic equations with discontinuities, in Proceedings of the Fourth European
Conference on Mathematics in Industry, Teubner, Stuttgart, 1992, pp. 213–217.

[23] R. Brusch and R. Schapelle, Solution of highly constrained optimal control

problems using nonlinear programming, AIAA J., 11 (1973), pp. 135–136.

[24] A. Bryson and Y. Ho, Applied Optimal Control, Hemisphere, New York,
1975.

382
[25] P. Bujakiewicz, Maximum weighted matching for high-index differential al-
gebraic equations, PhD thesis, Technical University of Delft, The Netherlands,
1994.
[26] S. Campbell, Consistent initial conditions for singular nonlinear systems,
Circuits, Systems and Signal Processing, 2 (1983), pp. 45–55.
[27] , Least squares completions for nonlinear differential algebraic equations,
Numer. Math., 65 (1993), pp. 77–94.
[28] , Numerical methods for unstructured higher-index DAEs, Annals of Nu-
merical Mathematics, 1 (1994), pp. 265–278.
[29] S. Campbell and E. Griepentrog, Solvability of general differential-
algebraic equations, SIAM J. Sci. Comput., 16 (1995), pp. 257–270.
[30] M. Caracotsios and W. Stewart, Sensitivity analysis of initial value prob-
lems with mixed ODEs and algebraic constraints, Computers chem. Engng., 9
(1985), pp. 359–365.
[31] M. Carver, Efficient integration over discontinuities in ordinary differential
equation simulations, Math. Comput. Sim., XX (1978), pp. 190–196.
[32] J. Casares and J. Rodriguez, Analysis and evolution of a wastewater treat-
ment plant by stochastic optimization, Appl. Math. Model, 13 (1989), pp. 420–
424.
[33] M. Charalambides, Optimal Design of Integrated Batch Processes, PhD the-
sis, University of London, London, U.K., 1996.
[34] J. Cuthrell and L. Biegler, On the optimization of differential-algebraic
process systems, AIChE J., 33 (1987), pp. 1257–1270.
[35] W. De Backer, Jump conditions for sensitivity coefficients, in Sensitivity
methods in control theory, L. Radanović, ed., Dubrovnik, Aug. 31 - Sep. 5
1964, Pergamon Press, pp. 168–175.
[36] M. de Pinho, R. Sargent, and R. Vintner, Optimal control of nonlinear
DAE systems, in Proceedings of 32nd IEEE Conference on Decision and Control,
1993.
[37] W. Denham and A. Bryson, Optimal programming problems with inequality
constraints. II: Solution by steepest ascent, AIAA J., 2 (1964), pp. 25–34.
[38] J. Dennis and R. Schnabel, Numerical Methods for Unconstrained Opti-
mization and Nonlinear Equations, Prentice Hall, Englewood Cliffs, NJ, 1983.
[39] P. Deuflhard, Recent advances in multiple shooting techniques, in Computa-
tional Techniques for Ordinary Differential Equations, Academic Press, 1980,
pp. 217–272.

383
[40] P. Deuflhard, E. Hairer, and J. Zugck, One-step and extrapolation
methods for differential-algebraic systems, Numer. Math., 51 (1987), pp. 501–
516.

[41] L. Dixon and M. Bartholomew-Biggs, Adjoint-control transformations

for solving practical optimal control problems, Opt. Cont. Appl. and Meth., 2
(1981), pp. 365–381.

[42] S. Dreyfus, Variational problems with state variable inequality constraints, J.

Math. Anal. Appl., 4 (1962), pp. 297–308.

[43] I. Duff, On algorithms for obtaining a maximum transversal, ACM Trans.

Math. Software, 7 (1981), pp. 315–330.

[44] I. Duff, A. Erisman, and J. Reid, Direct Methods for Sparse Matrices,
Oxford University Press, 1986.

[45] I. Duff and J. Reid, MA48, a FORTRAN code for direct solution of sparse
unsymmetric linear systems of equations, Tech. Rep. RAL-93-072, Rutherford
Appleton Laboratory, Oxon, UK, 1993.

[46] W. Feehery, J. Banga, and P. Barton, A novel approach to dynamic

optimization of ODE and DAE systems as high-index problems, AIChE Annual
Meeting, Miami, (1995).

[47] W. Feehery and P. Barton, A differentiation-based approach to dynamic

simulation and optimization with high-index differential-algebraic equations, in
Computational Differentiation: Techniques, Applications, and Tools, M. Berz,
C. Bischof, G. Corliss, and A. Griewank, eds., SIAM, Philadelphia, Penn., 1996,
pp. 239–252.

[48] , Dynamic simulation and optimization with inequality path constraints,

Computers chem. Engng., 20 (1996), pp. S707–S712.

[49] W. Feehery, J. Tolsma, and P. Barton, Efficient sensitivity analysis of

large-scale differential-algebraic equations, Applied Numerical Mathematics, 25
(1997), pp. 41–54.

[50] R. Felder and R. Rousseau, Elementary Principles of Chemical Processes,

Wiley, New York, 1986.

[51] A. Feng, C. Holland, and S. Gallun, Development and comparison of

a generalized implicit Runge-Kutta method with Gear’s method for systems of
coupled differential and algebraic equations, Computers chem. Engng., 8 (1984),
pp. 51–59.

[52] C. Floudas, Nonlinear and Mixed-Integer Optimization, Oxford University

Press, New York, 1995.

384
[53] R. Franks, Modeling and Simulation in Chemical Engineering, Wiley-
Interscience, New York, 1972.

[54] C. Führer and B. Leimkuhler, Numerical solution of differential-algebraic

equations for constrained mechanical motion, Numer. Math., 59 (1991), pp. 55–
69.

[55] S. Galán, W. Feehery, and P. Barton, Parametric sensitivity functions

for hybrid discrete/continuous systems, submitted to Applied Numerical Math-
ematics, (1997).

[56] R. Gani and T. Cameron, Modelling for dynamic simulation of chemical

processes- the index problem, Chem. Engng. Sci., 47 (1992), pp. 1311–1315.

[57] C. Gear, Maintaining solution invariants in the numerical solution of ODEs,

SIAM J. Sci. Statist. Comput., 7 (1986), pp. 734–743.

[58] , Differential-algebraic equation index transformation, SIAM J. Sci. Statist.

Comput., 9 (1988), pp. 39–47.

[59] C. Gear, G. Gupta, and B. Leimkuhler, Automatic integration of the

Euler-Lagrange equations with constraints, J. Comp. Appl. Math., 12 (1985),
pp. 77–90.

[60] C. Gear and L. Petzold, ODE methods for the solution of differential-
algebraic equations, SIAM J. Numer. Anal., 21 (1984), pp. 716–728.

[61] P. Gill, W. Murray, and M. Saunders, Large-scale SQP methods and

their application in trajectory optimization, in Computational Optimal Control,
R. Bulirsch and D. Kraft, eds., Birkhäuser Verlag, Basel, 1994, ch. 1, pp. 29–42.

[62] P. Gill, W. Murray, and M. Wright, Practical Optimization, Academic

Press, New York, 1981.

[63] C. Goh and L. Teo, Control parametrization: A unified approach to optimal

control problems with general constraints, Automatica, 24 (1988), pp. 3–18.

[64] V. Gopal and L. Biegler, A successive linear programming approach for

initialization and reinitialization after discontinuities of differential algebraic
equations, SIAM Journal on Scientific Computing, (to appear).

[65] A. Griewank, On automatic differentiation, in Mathematical Programming:

Recent Developments and Applications, T. Coleman and Y. Li, eds., Kluwer
Academic Publishers, 1989, pp. 83–108.

[66] D. Gritsis, The Dynamic Simulation and Optimal Control of Systems De-
scribed by Index Two Differential-Algebraic Equations, PhD thesis, University
of London, 1990.

385
[67] D. Gritsis, C. Pantelides, and R. Sargent, Optimal control of systems
described by index-2 differential-algebriac equations, SIAM J. Sci. Comput., 16
(1995), pp. 1349–1366.
[68] T. Gronwall, Note on the derivatives with respect to a parameter of the so-
lutions of a system of differential equations, Annals of Mathematics, 20 (1919),
pp. 292–296.
[69] J. Guckenheimer and S. Johnson, Planar hybrid systems, Hybrid Systems
II, Lecture Notes in Computer Science, 999 (1995), pp. 202–225.
[70] E. Hairer, C. Lubich, and M. Roche, The numerical solution of
differential-algebriac systems by Runge-Kutta methods, in Lecture Notes in
Mathematics, Springer-Verlag, 1989.
[71] E. Hairer, S. Nørsett, and G. Wanner, Solving Ordinary Differential
Equations, vol. I Nonstiff Problems, Springer-Verlag, Berlin, 1987. pp. 94-98.
[72] E. Haug and P. Ehle, Second-order design sensitivity analysis of mechanical
system dynamics, Internat. J. Numer. Methods Engrg., 18 (1982), pp. 1699–
1717.
[73] E. Haug, R. Wehage, and N. Barman, Design sensitivity analysis of pla-
nar mechanism and machine dynamics, Trans. ASME, 103 (1981).
[74] D. Jacobson and M. Lele, A transformation technique for optimal control
problems with a state variable inequality constraint, IEEE Trans. Automatic
Control, 5 (1969), pp. 457–464.
[75] D. Jacobson, M. Lele, and J. Speyer, New necessary conditions of opti-
mality for control problems with state variable inequality constraints, J. Math.
Anal. Appl., 35 (1971), pp. 255–284.
[76] R. Jarvis and C. Pantelides, A differentiation-free algorithm for solving
high-index DAE systems, AIChE Annual Meeting, Miami, (1992).
[77] B. Keeping and C. Pantelides, On the implementation of optimisation al-
gorithms for large-scale transient systems on a MIMD computer, AIChE Annual
Meeting, Miami, (1995).
[78] J. Keiper and C. Gear, The analysis of generalized backward-difference for-
mula methods applied to Hessenberg form differential-algebraic equations, SIAM
J. Numer. Anal., 28 (1991), pp. 833–858.
[79] H. Keller, Numerical Methods for Two-Point Boundary Value Problems,
Blaisdell, Waltham, MA, 1968.
[80] H. Kelley, A trajectory optimization technique based upon the theory of the
second variation, Progress in Astronautics and Aeronautics, 14 (1964), pp. 559–
582.

386
[81] E. Kikkinides and R. Yang, Simultaneous SO2 /NO2 removal and SO2 re-
covery from flue gas by pressure swing adsorption, Ind. Eng. Chem. Res., 30
(1991), pp. 1981–1989.
[82] D. Kirk, Optimal Control Theory: An Introduction, Prentice-Hall, Englewood
Cliffs, New Jersey, 1970.
[83] P. Kokotović and R. Rutman, Sensitivity of automatic control systems
(survey), Automation and Remote Control, 4 (1965), pp. 727–749.
[84] D. Kraft, On converting optimal control problems into nonlinear programming
problems, Comp. Math. Prog., 15 (1985), pp. 261–280.
[85] , Algorithm 733: TOMP- Fortran modules for optimal control calculations,
ACM Trans. Math. Software, 20 (1994), pp. 263–281.
[86] M. Kramer and J. Leis, The simultaneous solution and sensitivity analy-
sis of systems described by ordinary differential equations, ACM Trans. Math
Software, 14 (1988), pp. 45–60.
[87] A. Lefkopoulos and M. Stadtherr, Index analysis of unsteady-state
chemical processes- I. an algorithm for problem formulation, Computers chem.
Engng., 17 (1993), pp. 399–413.
[88] J. Leis and M. Kramer, Sensitivity analysis of systems of differential and
algebraic equations, Computers chem. Engng., 9 (1985), pp. 93–96.
[89] M. Lentini and V. Pereyra, An adaptive finite-difference solver for non-
linear two-point boundary value problems with mild boundary layers, SIAM J.
Numer. Anal., 14 (1977), pp. 91–111.
[90] J. Logsdon and L. Biegler, Accurate solution of differential-algebraic op-
timization problems, Ind. Eng. Chem. Res., 28 (1989), pp. 1628–1639.
[91] R. Luus, Piecewise linear continuous optimal control by iterative dynamic pro-
gramming, Ind. Eng. Chem. Res., 32 (1993), pp. 859–865.
[92] L. Lynn, E. Parkin, and R. Zahradnik, Near-optimal control by trajectory
approximation, Ind. Eng. Chem. Fundam., 18 (1979).
[93] C. Majer, W. Marquardt, and E. Gilles, Reinitialization of DAEs after
discontinuities, Computers chem. Engng., 19 (1995), pp. S507–S512.
[94] T. Maly and L. Petzold, Numerical methods and software for sensitivity
analysis of differential-algebraic systems, Applied Numerical Mathematics, 20
(1996), pp. 57–79.
[95] W. Marquardt, Numerical methods for the simulation of differential-
algebraic process models, tech. rep., Rheinisch-Westfälische Technische
Hochschule Aachen, 1994.

387
 [96] R. März, Numerical methods for differential-algebraic equations, Acta Numer-
ica, (1991), pp. 141–198.

[97] S. Mattsson and G. Söderlind, Index reduction in differential-algebraic

equations using dummy derivatives, SIAM J. Sci. Stat. Comput., 14 (1993),
pp. 677–92.

[98] R. Mehra and R. Davis, A generalized gradient method for optimal control
problems with inequality constraints and singular arcs, IEEE Trans. Autom.
Control, (1972).

[99] A. Mengel, Optimum trajectories, in Proceedings of Project Cyclone Sympo-

sium 1 on REAC Techniques, New York, March 1951, Reeves Instrument Corp.
and U.S. Navy Special Devices Center, pp. 7–13.

[100] A. Miele, B. Mohanty, P. Venkataraman, and Y. Kuo, Numerical

solution of minimax problems of optimal control, part 2, JOTA, 38 (1982),
pp. 111–135.

[101] M. Mohideen, J. Perkins, and E. Pistikopoulos, Optimal synthesis

and design of dynamic systems under uncertainty, AIChE Journal, 42 (1994),
pp. 2251–2272.

[102] , Optimal synthesis and design of dynamic systems under uncertainty, Com-
puters chem. Engng., 20 (1996), pp. S895–S900.

[103] , Towards an efficient numerical procedure for optimal control, Computers

chem. Engng., 21 (1997), pp. S457–S462.

[104] K. Morison and R. Sargent, Optimization of multistage processes described

by differential-algebraic equations, Lect. Notes Math., 1230 (1986), pp. 86–102.

[105] K. R. Morison, Optimal Control of Processes Described by Systems of

Differential-Algebraic Equations, PhD thesis, University of London, 1984.

[106] C. Neuman and A. Sen, A suboptimal control algorithm for constrained prob-
lems using cubic splines, Automatica, 9 (1973), pp. 601–603.

[107] C. Pantelides, The consistent initialization of differential-algebraic systems,

SIAM J. Sci. Stat. Comput., 9 (1988), pp. 213–231.

[108] C. Pantelides, D. Gritsis, K. Morison, and R. Sargent, The math-

ematical modelling of transient systems using differential-algebraic equations,
Computers chem. Engng., 12 (1988), pp. 449–454.

[109] C. Pantelides, V. Vassiliadis, and R. Sargent, Optimal control of mul-

tistage systems described by high-index differential-algebraic equations, in Com-
putational Optimal Control, R. Bulirsch and D. Kraft, eds., Birkhaüser, Basel,
Germany, 1994, pp. 177–191.

388
[110] T. Park and P. Barton, State event location in differential-algebraic models,
ACM Transactions on Modeling and Computer Simulation, 6 (1996), pp. 137–
165.
[111] H. Pesch, Real-time computation of feedback controls for constrained optimal
control problems. Part 2: A correction method based on multiple shooting, Opt.
Cont. Appl. and Meth., 10 (1989), pp. 147–171.
[112] L. Petzold, Differential/algebraic equations are not ODEs, SIAM J. Sci. Stat.
Comput., 3 (1982), pp. 367–384.
[113] , A description of DASSL: A differential/algebraic equation solver, in Sci-
entific Computing, R. Stepelman, ed., North-Holland, 1983, pp. 65–68.
[114] L. Petzold, Y. Ren, and T. Maly, Regularization of higher-index
differential-algebraic equations with rank-deficient constraints, SIAM J. Sci.
Comput., 18 (1997), pp. 753–774.
[115] L. Petzold, J. Rosen, P. Gill, L. Jay, and K. Park, Numerical optimal
control of parabolic PDEs using DASOPT, Proc. IMA Workshop on Large-Scale
Optimization, (1996).
[116] P. Piela, ASCEND: An Object-Oriented Computer Environment for Modeling
and Analysis, PhD thesis, Carnegie–Mellon University, Pittsburgh, 1989.
[117] J. Ponton and P. Gawthrop, Systematic construction of dynamic models
for phase equilibrium processes, Computers chem. Engng., 15 (1991), pp. 803–
808.
[118] L. Pontryagin, V. Boltyanskii, R. Gamkrelidze, and E. Mishenko,
The Mathematical Theory of Optimal Processes, Interscience Publishers Inc.,
New York, 1962.
[119] H. Rabitz, M. Kramer, and D. Dacol, Sensitivity analysis in chemical
kinetics, Ann. Rev. Phys. Chem., 34 (1983), pp. 419–61.
[120] W. Ray, Advanced Process Control, McGraw-Hill, New York, 1981.
[121] K. Reinschke, Multivariable control: A graph theoretic approach, in Lecture
Notes in Control and Information Sciences, M. Thoma and A. Wyner, eds.,
vol. 108, Springer-Verlag, 1988.
[122] S. Roberts and J. Shipman, Two Point Boundary-Value Problems: Shooting
Methods, American Elsevier, New York, 1972.
[123] O. Rosen and R. Luus, Evaluation of gradients for piecewise optimal control,
Computers chem. Engng., 15 (1991), pp. 273–281.
[124] E. Rozenvasser, General sensitivity equations of discontinuous systems, Au-
tomation and Remote Control, (1967), pp. 400–404.

389
[125] I. Russ, Sensitivity function generation from nonlinear systems with jump dis-
continuities, Buletinul Institutului Politehnic “Gheorghe Gheorghiu-Dej” Bu-
curesti. Seria Electrotehnica, 39 (1977), pp. 91–100.

[126] C. Sagan, Cosmos, Random House, New York, 1980.

[127] K. San and G. Stephanopoulos, Optimization of fed-batch penicillin fer-

mentation: A case of singular optimal control with state constraints, Biotech-
nology and Bioengineering, 34 (1989), pp. 72–78.

[128] R. Sargent and G. Sullivan, The development of an efficient optimal con-

trol package, in Proc. 8th IFIP Conf. Optimization Tech. Pt. 2, 1977.

[129] R. Schopf and P. Deuflhard, OCCAL a mixed symbolic-numerical optimal

control calculator, in Computational Optimal Control, R. Bulirsch and D. Kraft,
eds., Birkhaüser, Basel, Germany, 1994.

[130] R. Sincovec, A. Erisman, E. Yip, and M. Epton, Analysis of descriptior

systems using numerical algorithms, IEEE Trans. Automatic Control, AC-26
(1981), pp. 139–147.

[131] G. Stewart, Rank degeneracy, SIAM J. Sci. Stat. Comput., 5 (1984), pp. 403–
413.

[132] P. Tanartkit and L. Biegler, Stable decomposition for dynamic optimiza-

tion, Ind. Eng. Chem. Res., 34 (1995), pp. 1253–1266.

[133] K. Teo, C. Goh, and K. Wong, A Unified Computational Approach to Op-

timal Control Problems, Pitman Monographs and Surveys in Pure and Applied
Mathematics, Wiley, New York, 1991.

[134] J. Tolsma and P. Barton, Efficient calculation of sparse Jacobians, sub-

mitted to SIAM Journal on Scientific Computing, (1997).

[135] , On computational derivatives, Computers chem. Engng., (in press, 1997).

[136] R. Tomović and M. Vukobratović, General Sensitivity Theory, American

Elsevier, New York, 1972.

[137] T. Tsang, D. Himmelblau, and T. Edgar, Optimal control via collocation

and non-linear programming, Int. J. Control, 21 (1975), pp. 763–768.

[138] Y. Z. Tsypkin and R. S. Rutman, Sensitivity equations for discontinuous

systems, in Sensitivity methods in control theory, L. Radanović, ed., Dubrovnik,
Aug. 31 - Sep. 5 1964, Pergamon Press, pp. 195–196.

[139] J. Unger, A. Kröner, and W. Marquardt, Structural analysis of

differential-algebraic equation systems- theory and application, Computers
chem. Engng., 19 (1995), pp. 867–882.

390
[140] F. Valentine, The problem of Lagrange with differential inequalities as added
side conditions, in Contributions to the Calculus of Variations, Chicago Uni-
versity Press, 1937, pp. 407–448.

[141] S. Vasantharajan and L. Biegler, Simultaneous strategies for optimiza-

tion of differential-algebraic systems with enforcement of error criteria, Com-
puters chem. Engng., 14 (1990), pp. 1083–1100.

[142] V. Vassiliadis, Computational Solution of Dynamic Optimization Problems

with General Differential-Algebraic Constraints, PhD thesis, University of Lon-
don, London, U.K., 1993.

[143] V. Vassiliadis, R. Sargent, and C. Pantelides, Solution of a class of

multistage dynamic optimization problems. 2. Problems with path constraints,
Ind. Eng. Chem. Res., 33 (1994), pp. 2123–2133.

[144] O. Von Stryk, Numerical solution of optimal control problems by direct col-
location, in Optimal Control and Variational Calculus, R. Bulirsch, A. Miele,
J. Stoer, and K. Well, eds., Birkhaüser, Basel, Germany, 1993.

[145] H. Wu, Generalized maximum principle for optimal control of generalized state-
space systems, Int. J. Control, 47 (1988), pp. 373–380.

[146] A. Xing and C. Wang, Applications of the exterior penalty method in con-
strained optimal control problems, Opt. Cont. Appl. and Meth., 10 (1989),
pp. 333–345.

[147] V. Yen and M. Nagurka, Optimal control of linearly constrained linear

systems via state parameterization, Opt. Cont. Appl. and Meth., 13 (1992),
pp. 155–167.

391