Proceedings
of
the
American Control Conference
Albuquerque,
New
Mexico
June
1997
0-7803-3a32-41971$j
0.00
o
I
997
AACC
On-line Optimization
Of
the Tennessee Eastman Challenge Process
Ming Yan
Department
of
Chemical Engineering
University
of
Washington
Seattle, WA
98195
Ming.Yan@mwk.com
Abstract
A
real-time optimization (RTO) strategy that elimi-
nates the steady-state-wait requirement of conventional
RTO is described. To ensure its success, four critical
components-dynamic modeling, on-line parameter and
state estimation,
a
plant-wide control system, and
an
op-
timization algorithm-must be well coordinated. It is es-
pecially effective when conventional RTO cannot be used,
and sustained disturbances occur on
a
time scale that is
close to the plant response time. This strategy was ap-
plied to the Tennessee Eastman challenge process, where
RTO saved up to
6%
of the operating cost relative to the
cases without RTO.
1
Introduction
Real-time optimization (RTO) has generated
a
great deal
of interest in both acedemia and industry. Currently,
most plantwide RTO systems are turned on only when
the plant is within
a
prescribed neighborhood of steady-
state. This incurs
a
substantial loss for plants where
steady-state is rarely achieved. The fundamental diffi-
culty comes from on-line estimation of parameters to be
used in
RTO.
Most such parameter estimation methods
are only valid for steady-state. Another limiting factor
for
process optimization comes from the process control
system.
Poor
control causes high variability of controlled
variables, which moves the plant away
from
its optimal
operating point and reduces the optimization benefit.
All
the above issues motivate us to consider a better integra-
tion and coordination among modeling, parameter esti-
mation, control, and optimization.
Large scale challenge problems have been published re-
cently by several industrial research groups. Among them,
the Tennessee Eastman (TE) problem
[l]
has the broadest
range
of
interesting and practical features. Most previous
work on the TE problem has been concerned with control
strategies. The present paper focuses on the real-time
optimization
of
the
TE
process.
A major difficulty faced in this work is that the com-
N.
Lawrence Ricker
Department
of
Chemical Engineering
University
Of
Washington
Seattle,
WA
98195
ri cker
Q
cheme. was hingt on. edu
plexity of the TE process makes
it
impossible
to
do ex-
haustive simulations for all the scenarios one might wish
to consider. Thus, some decisions derive from qualitative
analysis instead of computation. In many situations, we
can only speculate that one option is better than another.
Moreover, several promising emerging approaches were
not tested, such
as
receding horizon state estimation
[3,
21.
Despite these limitations, the results suggest that an inte-
grated
modeling/control/optimization
approach can suc-
ceed in
a
challenging, realistic environment.
2
TE
Process
Implementation
2.1
Model
See
[I]
for details of the process, typical operating condi-
tions, scenarios, etc. A mechanistic model of the process
had been developed earlier to support work on nonlinear
MPC
[6].
This model
was
adequate for
MPC,
but exhib-
ited shortcomings when used
for
optimization. In partic-
ular, over-simplified modeling of the stripper caused the
separator temperature to be driven to unrealistically low
values. Thus, improvements were made in the stripper
model, and
a
simple compressor model
was
added. The
modified version used here comprises
30
nonlinear, ordi-
nary differential equations in the generic state space form,
Eq.
1
and 2.
X
=
f(x,
U,
d)
Y
=
dx,
U,
4
(1)
(2)
where
x
is the state vector,
U
is the input vector,
d
is the
parameter vector, and
y
is the output vector. There are
30
states,
11
inputs, 21 parameters, and 31 outputs.
2.2
Parameter and
state
estimation
We use an Extended Kalman filter (EKF) formulation for
parameter and state estimation. In the TE process, purge
and feed compositions measurements are available every
6
minutes with
a
6-minute delay, product compositions are
available every 15 minutes with
a
15-minute delay, and
2960