A control scheme for packed bed reactors having

a changing catalyst activity profile. I: On-line

parameter estimation and feedback control
Larry C. Windes” and W. Harmon Ray+

Department of Chemical Engineering, University of Wisconsin, Madison, Wisconsin 53706,

USA
(Received I 7 June 1991; revised 22 February 1992)

The efficient control of an exothermic packed bed reactor is demonstrated in concert with the distributed
parameter state estimation of temperature and concentration profiles and on-line estimation of the catalyst
activity profile. The position-dependent catalyst activity is estimated using recursive least squares with a
variable forgetting factor, a distributed mass balance model and the current estimated temperature profile.
The temperature level in the reactor is regulated through control of the maximum estimated reactor
temperature using the reactor coolant temperature as the manipulated variable. These methods are demon-
strated both through closed-loop simulation and experiments with a laboratory-scale reactor for the partial
oxidation of methanol. The experiments included both setpoint changes and disturbance rejection. The
estimation and control system effectively overcomes errors in initial conditions, heat transfer parameters
and kinetic parameters.

(Keywords: packed bed reactors; catalyst activity; parameter estimation)

The packed bed chemical reactor is one of the most
difficult challenges for a process control engineer. First,
the reactor is a distributed system: the state variables of
temperature and composition depend upon position as
well as time. Thus, the output variables are limited, the
states to be controlled must be chosen from an infinite
number of possibilities and the control actions must be
based on a limited number of measurements. Often some
form of state estimation must be used in order to control
unmeasured state variables and to overcome measure-
ment errors.
For many such reactors, the maximization of the yield
of the desired product is the key control objective. How-
ever, this can be difficult to measure directly within the
time scale of typical reactor dynamics. In addition, the
manipulated variables often have complicated effects on
reactor yield. These complications are due to the Arrhen-
ius temperature dependence of the reaction rate, non-
linear kinetic expressions, complex reaction networks
and widely varying disturbance propagation rates in the
solid and gas phases of the reactor.
In order to implement model-based control for a dis-
tributed parameter system (DPS) such as a packed bed
tubular reactor, the system must be ‘lumped’ to obtain
numerical results. There are two possible approaches:
early lumping and late lumping of the system model. The
most straightforward method is to discretize the distri-
buted system at the earliest opportunity into an approxi-
mate model usually consisting of a set of ordinary differ-
ential equations. Another approach to discretization is to
pick a limited number of inputs and outputs and identify
an empirical system model relating the outputs to the
inputs. In either case, the control problem can then be
solved by techniques for lumped multivariable systems.
If the control designer desires to apply the full distri-
buted parameter control theory, then ‘late lumping’ dis-
cretizes the equations for the implemented controller
after the control design has been completed’. In almost
all previous work in the literature, early lumping has
been used for control of packed bed reactors.
Several experimental investigations have been con-
ducted for control of different types of packed bed reac-
tors: adiabatic reactor+, autothermal reactors4 and
non-adiabatic wall-cooled reactors@. Mandler et al.‘-*
have used a collocation model based on a laboratory
non-adiabatic reactor for control studies via simulation.
The previous ‘early lumping’ controller design tech-
niques can be divided into two classes:

l Use of the full approximate lumped model

*Current address: Eastman Chemical Company Research Laborator- obtained from the DPS model in the control
ies, Kingsport, TN 37662, USA system design procedure. Based directly on the
fAuthor to whom correspondence should be addressed high order model equations, the system often has a

0959-1524/92/010023-20
0 1992 Butterworth-Heinemann Ltd J. Proc. Cont. 1992, Vol2, No 1 23
Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray
Heater control from computer
//////////“‘////////~~
Speed control from computer
Methanol
feed
Feed pump
Temperatt Ire indica itor
Pressure indicator
Electrical heater
Figure 1 Schematic of the packed bed reactor pilot plant
large and unwieldly number of state variables.
Order-reduction methods can be employed.
l Selection of a limited number of state variables
which are measured directly or determined on-line
via a state estimator. Either the model equations
or identification methods are used to arrive at an
input-output model for this low-order system.
One common approach, first used by Vakil et aLg, is to
discretize the reactor equations by orthogonal colloca-
tion and apply linear-quadratic state estimation and
control. Among those using a small number of input-
output variables, adaptive control has been used for both
SISO systemslo and MIMO application&‘. Internal
model control and structured singular value analysis
have been used to control the reactor outlet tempera-
ture’, and multivariable control design has been applied
in cases where the molar flowrate or reactor conversion
are also consideredg. One example of controlling a
tubular reactor in the presence of catalyst deactivationI
attempted to maintain the reactor outlet conversion con-
stant through use of the pseudo-steady state reactor
equations and did not consider reactor dynamics other
24 J. Proc. Cont. 1992, Vol2, No 1
Jacket and bed
than those resulting from changes in catalyst activity.
Gilles and coworkers (Gilles and Epple13, Gilles and
Ruppel14, Epple15) have developed an approximate distri-
buted model for control problems with constant profile
structure and applied it to packed bed reactors.
In the present paper, we make use of late lumping and
carry out the control system design making use of the full
distributed model. We are able to take advantage of the
distributed features of the reactor behaviour in the
design. In order to demonstrate the applicability and
performance of the design procedure, an experimental
reactor having product yield limitations was controlled
using the proposed design procedure.
The specific reactor studied here is a pilot scale, jacket
cooled, single-tube reactor with an exothermic reaction
taking place within porous catalyst pellets (Figure I). The
reactor is designed to carry out the partial oxidation of
methanol to formaldehyde and suffers from a secondary
reaction of formaldehyde to carbon monoxide and
water. The experimental apparatus, the mathematical
model describing the system and the steady state and
dynamic experiments identifying the model parameters
have been fully described in earlier papers16J7; thus we
only present a brief summary here.
 Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray
The yield of formaldehyde (the intermediate product
in the partial oxidation) is strongly affected by higher
temperatures, which lead to total combustion products;
thus, control of the bed temperature profile is essential
for yield maximization, The material and energy
balances for the proce@ are summarized in the Appen-
dix. Because the fluid species and fluid temperature dyna-
mics are more than 1000 times faster than the bed tem-
perature, they are considered to be in quasi-steady state.
In order to provide a high performance control scheme
for the packed bed reactor, a comprehensive, generally
applicable distributed parameter control strategy has
been developed which is robust to modelling errors and
catalyst deactivation and controls the reactor over a wide
range of operating conditions. The overall control
system package consists of four sections:
l temperature and composition state estimation
0 parameter estimation
0 control
0 optimization.
These run on a process control computer in real time.
This paper emphasizes the parameter estimation and
control studies which utilize the results of the state esti-
mation. The state estimation algorithm has been pre-
sented in an earlier paper’*, and on-line optimization is
described in a companion paperlg.
Design and implementation of the control
scheme
In addition to choosing which strategies and algorithm
to implement, control system structural decisions must
be made for input-output variables, measurements and
control loops. Also, the dynamics of the experimental
control variables and disturbances should be assessed.
Selection of manipulated input variables
Several possibilities for combinations of control inputs,
outputs and disturbances can be observed in the previous
literature on packed bed reactor control. For the reactor
in this study, the choices for manipulated (i.e., control)
variables and disturbances are:
l jacket wall temperature
l inlet feed temperature
l inlet methanol mole fraction
l feed flow rate.
The decision concerning which of these variables to
choose as controls, which to allow to be disturbances,
and which to hold constant is based on the following
‘considerations:
the magnitude of the effect which the control vari-
able has upon the desired output;
creation of ‘interesting’ dynamics (i.e., wrong-way
behaviour) which results in a challenging control
problem;
experimental constraints, especially the dynamic
response of the proposed control variables;
formulating a realistic situation, using controls
commonly available in industrial reactors and
selecting disturbances encountered in industrial
situations;
l reproducibility and successful modelling of the
input response.
In many control problems the goal is to accomplish a
successful setpoint change with a minimum of offset and
interactions. However, for this reactor the setpoint is
chosen to maximize an objective such as reactor yield.
Thus, after reactor start-up, the setpoint does not change
except as reactor parameters change. Here, the more
important task for the control system is to reject distur-
bances and compensate for time-varying parameters.
In order to select the best set of control inputs, a
preliminary set of data is necessary. Because the accuracy
with which the reactor model predicts responses to
changes in input and disturbance variables affects the
performance of the estimator and controller, open-loop
experimentation was conducted for a wide range of
values of all possible manipulated variables. Generally,
any variable that causes an axial shift in the hotspot of
the reactor is more difficult to model because the reactor
response is substantially affected by axial variations in
catalyst activity caused by deactivation. Also, feed flow-
rate causes problems as a control variable because of
parameter variations and additional non-linearities in
the input-output model. Although several types of dis-
turbance cause model parameters to change somewhat, a
constant parameter model has been chosen for model-
based control system design.
Let us review each of the input variables to evaluate
their suitability as a manipulated variable. The reactor
temperature, and thus reactor yield, is highly dependent
on the jacket wall temperature16. This fact favours wall
temperature as a control variable. The wall temperature
is able to control the temperature in general throughout
the reactor, but the entire temperature profile increases
or decreases in response to the corresponding changes at
the wall. The location of the hot spot is affected little by
changes in wall temperature. Thus disturbances which
change the hot spot location (e.g. flow rate, inlet temper-
ature) cannot be completely rejected by manipulating
wall temperature alone. However, the wall temperature
generally provides adequate control of the two most
crucial outputs, the maximum reactor temperature and
the yield at the reactor exit.
The inlet feed temperature is unfavourable as a control
variable in terms of the first criterion because it causes
only small changes in yield. Thus, a very large change in
the inlet temperature would be needed to reject a small
upset in the external wall cooling system or to compen-
sate for deactivating catalyst. Manipulation of the inlet
temperature gives control over location of the hot spot.
However, changes in the inlet temperature can give rise
to ‘wrong-way’ behaviour. The dynamic and steady state
response of the reactor to this variable is highly depen-
dent on the activity profile in the catalyst bed.
The inlet methanol concentration is much like wall
temperature in terms of controllability of the reactor
temperature. However, changes in the inlet concent-
ration have only a small effect on the yield and have a
J. Proc. Cont. 1992, Vol2, No 1 25
Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

large effect on the production rate. Thus inlet concent- benefit was gained by the use of more than three temper-
ration should remain as large as safety constraints allow. ature measurements.
Feed flow rate has a large effect both on the yield and In choosing the most appropriate output variables,
the maximum reactor temperature, and it shifts the hot one must consider the control system objectives. The
spot markedly. The large change in the hot spot location objective function can take a rather general form includ-
might have detrimental consequences for model-based ing constraints and can be a function of a combination of
control if only one or two temperature measurements are several facets of reactor operation depending on the par-
used. Control action by feed flow manipulation would ticular process:utility consumption, exit temperatures,
not recover the original steady state around which a coolant temperature, pressure drop, material properties
controller may have been linearized. In addition, the etc. A special case is an objective which is a function of
flow rate is not necessarily available to change as a mani- weighted values of the products of the reactor. In the
pulated variable, because the amount of flow through the present case, reactor yield is chosen as the objective,
reactor is often specified by upstream process units or formaldehyde is assigned a unit value, and methanol,
design constraints for an entire plant. carbon monoxide, water and air have zero value. Thus
The actual dynamic response of the manipulated vari- the reactor yield of formaldehyde is maximized.
ables is important since it defines the constraints on Although the objective gives the overall goal of the
experimental control studies. Very often step changes are control system, we usually choose other output variables
not possible in manipulated variables due to equipment to be regulated in order to best achieve the control
lags, and realistic changes must be represented by a ramp system objective.
Some possible output variables are:
or other function rather than a step change. For our
l a composite objective based on amount of pro-
experimental reactor the rates of response of the possible
ducts (flow and exit composition), with reactor
manipulated variables are as follows:
yield as an example;
l 5°C mini for wall temperature heating or cooling
l the conversion at an interior axial location (has
at 25&26O”C; slower heating and faster cooling at
greater sensitivity to the manipulated variable
higher temperatures and vice versa;
than at the exit);
l 5°C mini for feed temperature with a time delay
l the axial composition profile;
of about 20 s; shows some overshoot indicating
l the maximum reactor temperature (i.e. hot spot
system of high order;
temperature)
l almost immediate response for changes in the
l the axial position of the hot spot;
methanol feed rate;
l the axial centreline temperature profile;
l a first order response with a time constant of l&
l the entire two-dimensional temperature profile;
30 s for changes in the air flow rate. l the temperature at the thermocouple nearest to the
Among the manipulated variables discussed here, the original steady state hot spot;
wall temperature and the inlet temperature are the most l one temperature from any fixed location (i.e. 5-10
likely to be chosen for control of an industrial reactor. cm upstream from steady state hot spot);
They are easily measurable and normally unaffected by 0 two temperature measurements (one upstream,
upstream or downstream process units. Variations in the one downstream) and the difference between them.
inlet temperature could be caused by use of a constant If the process outputs are directly measured tempera-
energy input preheater with a fluctuating feed flow or by tures (as in the last three examples), then feedback
heating the feed by exchange with a time-varying hot control loops could be implemented in a straightforward
stream. If the primary reactant is metered in as a pure manner. However, most of the outputs listed above can-
stream, then it can serve as a control input. The feed flow not be measured directly for real-time control, and their
rate can normally be measured, but may be subject to values must be computed based on a process model and
frequent transients if the feed is coming directly from an measurements which are readily available. The state esti-
upstream unit with no intermediate storage. In this case mator developed earlieri is used to generate these out-
the feed flow is constrained by the plant production and puts.
is a disturbance rather than a manipulated variable. Thus State estimation enables efficient use of only a few
reactor wall temperature is the primary control input measurements and a model of the process to generate an
chosen. output variable containing more valuable information
than measurements or a process model alone. For the
present reactor, the entire temperature and composition
Selection of output variables profiles of the reactor are estimated. However, it is con-
Previous reactor control experiments reported in the venient to use single numbers in feedback control loops.
literature have commonly used several evenly spaced The maximum reactor temperature, or hotspot tempera-
thermocouple measurements or an estimator with axially ture, is known to correlate well with reactor yield and
weighted concentrations as the output variables. Some- also provide a measure of the safety margin in reactor
times the exit temperature was used as a single measure- operation because it is a good indicator of the severity of
ment for adiabatic reactors; however, this usually is not the conditions under which the reactor is operating.
useful for wall-cooled reactors. Generally, only a small Therefore, the maximum reactor temperature (T,,,,,),

26 J. Proc. Cont. 1992, Vol2, No 1

 Control scheme
Parameter
Process
reactor
Input experiments]
or
reactor
model
(simulation]
J-i
T
meas
Figure 2 Schematic for parameter estimation
available through state estimation, has been chosen as
the output variable for the control loop. The axial
location of T,,,,, is a promising second output variable for
a multivariable control strategy.
An alternative to using the maximum temperature esti-
mate for the output variable of the control loop would be
to use the exit composition estimates directly. However,
there are deficiencies in this strategy. The first difficulty is
in specification of an appropriate setpoint for a multi-
component mixture. Because the reaction products are
not independent of each other, only certain response
surfaces are achievable. On the other hand, the reactor
selectivity is not the same in all situations either, so
specification of one component does not ensure the
desired amounts of others. If the overall yield or other
objective to be maximized is the control output variable,
then it has a steady state maximum from which there are
only negative deviations from the setpoint. The resulting
control problem is extremely non-linear, and the value of
the output can be attained by two different values of the
input except at the maximum. The problem of maximiz-
ing objectives such as yield can be dealt with more effecti-
vely through on-line optimization coupled with cascade
control. The output of an outer loop yield optimization
algorithm can be the setpoint for the inner control loop
on T,,,. An on-line optimizing loop built around the
controller described here has been developed and is
demonstrated in companion experiments to the ones
reported here’9.
Another important consideration is control system
performance in the presence of errors in the reaction rate
parameters. The temperature estimates are more reliable
than the composition estimates. If there were substantial
errors in the assumed catalyst activity, direct feedback.of
the composition estimates would cause severe errors in
the control action in attempting to force the estimates,to
for packed bed reactors. I: L. C. Windes and W. H. Ray
estimation
Parameter estimation
I
at t*+tD: 1 at t*:
recursive parametric
I
least 1 sensitivity
squares
lYEST (z*.t*)
I
I 4
ut*+tD)
I T(r.z.t*)
State estimation
X = f(r,z,t)
the setpoint. Because the temperatures are rapidly and
accurately measured, and the temperature profile estima-
tion algorithm operates better for errors in the kinetic
parameters, attaining the estimated maximum tempera-
ture setpoint ensures that the reactor is operating near
the desired conditions even if the composition estimates
are in substantial error.
Parameter estimation
A principal difficulty in achieving good predictions of the
composition profiles in the reactor is the fact that the
catalyst deactivates non-uniformly over time. Therefore,
a necessary task is estimating the catalyst activity profile
on-line. These catalyst activity parameter estimates are
based on: a) infrequent composition measurements with
time delays; and b) the reactor temperature profile taken
directly from the state estimation algorithm. A block
diagram in Figure 2 indicates the structure and data flow
of the parameter estimation algorithm which utilizes
recursive least squares with a variable forgetting factor.
One or more parameters are estimated from each indepe-
nently measured component i’nthe reaction mixture. The
current estimate of the catalyst activity profile is updated
by comparing the model predictions with the compo-
sition measurements. These parameter estimates are used
to modify the reaction rate parameters used by the state
estimation and optimization algorithms.
The composition samples are taken one at a time at a
selected axial location in the reactor, and the compo-
sition analysis is available after a time delay (TV). The
mole fractions of all reacting species are assumed to be at
a quasi-steady state with respect to the instantaneous
temperature profile at the time the sample was taken
from the reactor. The sequence of events in parameter
estimation is given below. At sample time, t*:
J. Pro& Cont. 1992, Vol2, No I 27
Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray
1. obtain a composition sample (at location I*) and
send it to measuring device;
2. at the same time, use the estimated temperature
profile and the quasi-steady state non-linear distri-
buted mass balance equations for each component
to solve for the pseudo-steady composition pro-
file;
3. use interpolation to extract and save the estimated
composition at the measurement locations;
4. compute and save the parametric sensitivity of the
measurement estimates.
When the analysis of the composition
plete:
5. compute the measurement value;
6. compare the measurement with the estimate and
calculate the resulting error;
7. use recursive estimation to update the parameters;
8. pass new parameter values to the state estimation
algorithm where they are used in estimating suc-
ceeding temperature and composition profiles;
9. .pass new parameter values to the optimization
algorithm.
The parameter estimation algorithm is implemented in
the following manner. The predicted value of the kth
measurement is obtained by solving the non-linear alge-
braic equations for the composition at the collocation
points and interpolating to generate values at the mea-
surement location. The solution is based upon all para-
meters at the measurement time (Q, including the esti-
mated parameters which were updated at time tk_ 1 + q,
and which were being used in the model at time tk. The
predicted output of composition at time tk is given by:
where ak are fixed parameters and ok_, are the para-
meters estimated on-line.
Linearizing the equations for the composition measur-
ement output with respect to each estimated parameter:
Yk = flakptkpek-d +
ay
a (ok - ok-l)
the parametric sensitivity vector ay/aO is found by finite
differences through one at a time perturbation of the
parameters about ok_, and solving Equation (1). Taking
the output to be the measured value, and substituting j$,
the linearized equation for the parameters yields:
yk - ik =
ay
s (ok - ok-,)
Defining E as the prediction error, 6 as the change in
the parameter vector for this time step, and + as the
parametric sensitivity, the standard form for the para-
meter estimates becomes:
Ek = d?@k
28 J. Proc. Cont. 1992, Vol2, No I
A recursive least squares algorithm with variable for-
getting factor20 implemented with U-D factorization is
used to estimate the parameters. These and similar esti-
mation algorithms are also discussed by Ydstie and Sar-
gent21. After calculating the prediction and the error as
shown above, the algorithm is:
Gain: pk- I$k
Kk = 1 + 4zPk-,t,,k
(5)
Parameter estimates: 8k = K&k (6)
sample is com- Forgetting factor: hk = 1 - (1 - &?Kk) Ek/x (7)
If h < h,i” then set h = h,,
Covariance: Pk = (1 - Kk+‘)Pk-,/hk (8)
When the algorithm is implemented, Pk is replaced by
(VDU), where U is upper triangular with l’s on the
diagonal, and D is diagonal. The user specifies the
following values:
LA, - the minimum forgetting factor;
1 = oZ,N,, - a measure of the information content,
where 02, is the expected measurement noise vari-
ance, and N,, is the nominal asymptotic memory
length (small values give a sensitive estimator);
Di - the diagonal elements of the initial values for
Pk, i = 1, M where A4is the number of parameters
estimated.
Reactor hot spot temperature control
The block diagram for regulation of the maximum reac-
tor temperature is shown in Figure 3. The jacket wall
temperature setpoint was chosen as the manipulated
input variable in the outer cascade loop to regulate the
maximum estimated reactor temperature. The jacket
(1)
wall temperature is maintained by a heater controlled
with a local PID time-proportional controller in the
inner loop. The state estimation and parameter estima-
tion sections work together to provide good estimates of
the maximum temperature.
The wall temperature was chosen as the manipulated
variable for the following reasons:
(2)
l The monotonic response of the reactor tempera-
ture to wall temperature - an increase in wall
temperature increases the temperature everywhere
in the reactor, and a decrease does the reverse;
l The yield in the reactor is strongly affected by the
wall temperature, and a wide range of control over
the methanol conversion and carbon monoxide
production is possible;
l the wall temperature is often available as the vari-
able used to control the reactor in industrially
relevant situations.
The ‘well-behaved’ input-output relationship between
T, and T,,, is helpful for control design. There is no
inverse response, and the gain is always positive regard-
(4) less of the operating conditions.
 Control scheme for packed bed reacrors. I: L. C. Windes and W. H. Ray

Control of maximum reactor temperature

Process
reactor
(experiments)
L
=U Y
k or meas
fQ= --_L * Parameter
74 - reactor
:s
2 Twall estimation
model
5
(simulation)

1 I
/
T(r.z.t*)
meas
e(t*+tD)

Maximum
-
State
temperature _ Tmax r;

controller .*. estimation

Interpolation T(r.2.t J

L temperature
X = f (r,z,t)

setpoint

Figure 3 Schematic for reactor control of maximum reactor temperature

One feature of the wall temperature input is that it is
not constrained by an absolute value, but by the rate of
change of the input. Since a cascade control system was
used, the rate of change of the wall temperature setpoint
was constrained within the maximum possible rate of
change in the wall temperature itself. A proportional-
integral digital controller algorithm with constraint on
the rate of change of the control action was found to
work satisfactorily for manipulation of the wall tempera-
ture setpoint in controlling maximum reactor tempera-
ture. The controller equations are:
Ek = Yset,k - Yk
Au = Kc (Ek_, - Ek + l/T, L!t(Ek_, + Ek)/2)
If IAuJ > Aumax,then
&k =
w (A4 AumaxlKc
+ Ek-l(l - WCW)
1 + At/(24
The control action, where u is the wall temperature
setpoint constrained within a safe level for maximum
coolant temperature, is:
uk = min(u max,uk-I + 4
The input-output relationship betwen T, and T,,,,,was
observed to be of high order, and therefore the gain must
be kept reasonably low to avoid oscillations. However,
another factor limiting the aggressiveness of the con-
troller gains is the inverse response of T,,,,, caused by
some flow rate disturbances.
For the control experiments in this work, the feed
temperature controller attempted to keep the feed tem-
perature equal to the measured coolant temperature. If
the feed temperature was kept near the wall temperature,
it had little effect on reactor performance due to the
deactivation of the catalyst at the entrance of the bed.
The first portion of the bed functioned partly as a heat
exchanger bringing the gas coming through the bed close
to the wall temperature.
Experimental methods
On-line experiments have been performed to test the
(9) control scheme shown in Figure 3. The state estimation
algorithm was used throughout in order to provide reli-
(10) able estimates of the reactor temperature and compo-
sition profiles. The rationale for the design of the experi-
ments was to provide a basic test of the control system
and then to test the performance under difficult and
complicating circumstances.
(11) An enhanced version of the data acquisition and local
control program which had been used to operate the
reactor during open-loop steady state and dynamic
experiments” was used in closed-loop experimentation.
Important features are the on-line gas chromatograph
(12)
data analysis and extensive communication between the
real-time data acquisition package and the estimation
and control algorithms. The cycle time for the gas chro-
matograph samples was 12.5 min. The amounts of
formaldehyde, carbon monoxide and unreacted metha-
nol were measured for each sample and determined as a
fraction of the methanol fed to the reactor. The time
interval for temperature measurement sets was 6 s. The
experiments were run entirely automatically by the
computer using a protocol specified in a data file.

J. Proc. Cont. 1992, Vol2, No 1 29

Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

The functional form for the axial dependence of the

Catalyst activity profile preexponential factor is:
for methanol oxidation
,lOOO-
2
A,(z) = A,+ - (A,+ - AP)_f& Z”, b,) (13)

CI where
$ 600-

2’ : inflection point
f = atan (a(1 - Z’)) - atan (ct(z - Z”))
Al atan (a(1 - Z’)) - atan ( - crZ”>

cL= tan (0.9x/2)

0 35 70 b, (14)
Axial distance (cm)
Parameters Z” and b, were determined by off-line para-
meter optimization.
Figure 4 Catalyst activity profile for methanol oxidation The axial dependence of the preexponential factor for
CO production A2 has the form:
In addition to the real-time program operating the
packed bed reactor, a unified suite of programs have AZ(z) = A; - (A; - A;)f,,(z) = (1 - z)A; + zA;
been written to interface with it and accomplish the tasks (15)
of state estimation, parameter estimation, control and
optimization, and output of data to the user. where
These programs have the flexibility to operate in three
different modes based on the source of the reactor oper- _~Az(~) = 1 - z
ating ‘data’:
The parameter A,” was set to (0.25 A;) based on off-
0 measurements from a preexisting file of real or line parameter optimization. The parameters estimated
simulated data; on-line in the experimental work were:
l measurements obtained at each time step from a
dedicated reactor simulator running in conjunc- l A,+ - the preexponential factor for methanol oxi-
tion with the main program; dation at the reactor exit (most reactive limit);
l reactor measurements communicated in real time l Ai-’- the preexponential factor for methanol oxi-
from the data acquisition and reactor operations dation at the reactor entrance (least reactive limit);
program. l A;- the preexponential factor for partial oxi-
dation of formaldehyde to produce carbon
The second option is important because it allows
monoxide (at reactor exit).
closed-loop simulation of the model-based control and
optimization slgorithms. A separate but structurally
The parameters A, and Ai’ were estimated from the
identical mathematical model (the ‘simulator’) is solved
methanol conversion measurements and A: from the
as a subroutine to the main program. The closed-loop
amount of carbon monoxide produced. It is also possible
simulation is valuable in testing the algorithms before
to estimate the axial location of the inflection point of the
using them in actual experimentation. The simulator
methanol oxidation activity curve if conversion measure-
allows flexibility because trials can be run which are not
ments are made at various points along the length of the
possible experimentally. Also, the differences between
reactor over a long period of time. This enables tracking
the control model and the ‘simulator’ are exactly as
of a slow-moving deactivation wave moving through the
specified by the user.
reactor.

On-line parameter estimation tests Simulation results

DeJinition of parameters
The parameter estimation algorithm was implemented as
discussed in the previous section. The parameter estima-
tion experiment consisted of a sequence of different oper-
ating conditions during which three parameters were
estimated based on gas chromatograph measurements of
the compositions in the reactor. The estimated para-
meters defined the catalyst activity profile along the axial
dimension of the reactor. The function representing the
catalyst activity profile, illustrated in Figure 4, is deter-
mined by four parameters”: A:, A;, Z” and 6,.
The parameter estimation algorithm was first investi-
gated by simulation in order to show the viability of the
parameter estimation for several classes of problems:
l estimation of constant parameters under steady
conditions;
l estimation and tracking of time-varying para-
meters;
l estimation and tracking of parameters during
time-varying reactor conditions and/or inputs.
Since the estimation of a single well-conditioned para-

30 J. Proc. Cont. 1992, Vol2, No 1

 Control scheme
meter from the carbon monoxide measurements is much
simpler than estimating the activity profile from the
methanol measurements, the former is discussed only
briefly here.
The simplest case of parameter estimation is for sta-
tionary reactor conditions and a single composition mea-
surement location (simulation Pl). A single parameter
indicating overall catalyst activity is easily estimated in
this type of experiment. However, estimation of an
activity profile is dependent on: a) the relative error
estimates of the parameters, and b) the initial guess of the
relative activities at the entrance and exit of the reactor.
The axial catalyst activity profile is not observable with
one set of reactor conditions and one composition mea-
surement location. The estimates of the composition at
the measurement location do converge to the true values,
but the kinetic parameters for methanol may converge to
incorrect values which give the same conversion at the
measurement location as the true parameters. To achieve
good catalyst activity profile estimates, multiple measur-
ement locations and/or multiple operating conditions are
needed. When several composition measurement
locations are utilized and the reactor operates at steady
state, the parameters converge to their accurate values,
and the catalyst activity profile based on two parameters
is uniquely determined.
A more difficult case study was estimation of three
parameters of the methanol oxidation activity profile: A:
(at exit), Ai’ (at entrance) and Z” (inflection point). All
three of these parameters are estimated through utiliza-
tion of composition measurements at several axial
locations in the reactor. A good approximation of the
activity profile was obtained after beginning with an
extremely poor initial guess.
By assuming larger initial uncertainty in the parameter
estimates, the parameter estimation is made more
aggressive. The parameters converge more rapidly, but
they may overshoot their actual values as the estimates
improve, or the relative values of interacting parameters
may be corrupted. In some situations a large initial cor-
rection in the parameters may move the model into a
region of parameter insensitivity where sluggish estima-
tor performance results.
Operating the reactor at several different conditions
provides more information for the parameter estimation
than operating at a single steady state. Conditions which
move the hotspot axially in the reactor provide infor-
mation about the axial catalyst activity profile even if
there is only one composition measurement location
(e.g., at the reactor exit). The parameters tend to con-
verge rather slowly in the case of a gradual transition in
catalyst activity, but the results are better than in the case
of estimation at a single steady state.
In addition to identifying constant kinetic parameters
on-line, the parameter estimation must be able to track
time-varying parameters. Specifically, it must follow the
activity of a decaying catalyst bed. The estimation was
simulated over a period during which the catalyst activity
first declined rapidly and later became constant. Both the
carbon monoxide kinetics and the methanol oxidation
for packed bed reactors. I: L. C. Windes and W. H. Ray
0 20 40 60 80 100 120 140 160
Time (min)
Figure 5 Inputs for parameter estimation experiment: initial reactor
temperature uniform at 260°C; methanol flow begins at time 0
kinetics were accurately estimated as they changed. The
parameter estimates enable excellent tracking of the
changing composition profile within the reactor, and
appropriate changes in operating conditions can be made
to achieve the desirable selectivity and yield.
The parameter estimation experiments were simulated
beforehand using model parameters based on the most
recent reactor operation. The estimation parameter 1
(Equation (7)) was set at a relatively conservative large
value to allow for measurement error and so the estima-
tion was not sensitive to small estimate errors. Exper-
ience has shown that small initial covariance values and/
or a larger minimum forgetting factor give better results
than overshooting the true parameters on the first
update. The parameter estimation correctly estimated
the parameters in simulated experiments beginning with
three different initial guesses: catalyst activity too high,
catalyst activity too low, uniform catalyst activity.
Experimental results
The parameter estimation experiments consisted of a set
of several operating conditions over a time period of 160
min during which kinetic parameters were estimated.
The reactor operating test inputs are shown in Figure 5.
Twelve composition samples.were taken from four ports
in the downstream half of the reactor. The initial values
of the kinetic parameter estimates were chosen conserva-
tively high; therefore, the model/estimator predicts more
severe conditions than might actually exist in the reactor
‘during the initial period when estimates are poor. The
operating conditions were changed every 12.5 min in
order to test the temperature estimation with inaccurate
kinetic parameters and to provide a broader base of data
for the parameter estimates.
As shown in Figure 6~~2, the estimates were close to
the final values after four measurements. The conver-
gence was smooth, with noise causing minimal distur-
ante to occur in the parameter estimates. Figure 6d gives
the initial axial catalyst activity profile and the final
J. Proc. Cont. 1992, Vol2, No 1 31
Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

Time=8 and 20 min Time = 33 and 45 min Time = 5s and 70 mir

;m$g

0 35 70 0 35 70

0.081 I
A2 at exit

0.04
0.02 c 1
0 20 40 60 80 100 120 140 160

Time (min)

Axial distance (cm)

Time = 83 and 95 min Time = 108 and 12Omin Time=133 and 145mi1
,$ 1500;4 .O

.8

.6

.4

.2

0 35 70
Final estimated profile

I I I 1 1 I I I
0 10 20 30 40 50 60 70

Axial distance [cm)

Figure 6 Parameter estimation experiment: estimated parameter

values
Axial distance (cm)

estimated profile. Estimation of the preexponential
factor for carbon monoxide production based on the
fractional carbon monoxide conversion was quite reli-
able.
Figure 7 shows the estimated composition profiles in
comparison to the data at the times indicated. The
numbers in the legends refer to the operating conditions
given in Table 1. Composition profiles l-4 show metha-
nol conversions and by-product carbon monoxide pro-
duction considerably greater than the data during the
period of high estimated activity. However, beginning
with the fifth profile, good agreement is attained between
data and estimates. Thus, the parameter estimates
quickly converge to values giving accurate estimates of
the entire composition profile in the reactor. It is import-
ant to notice that not only is the exit composition of the
reactor predicted properly, but also the compositions at
other sampling locations.
Figure 8 shows the corresponding estimated tempera-
ture profiles in comparison to the data at the times indi-
cated on the graphs. The profiles are at steady state at
each of the different operating conditions of the experi-
ment. The kinetic parameter estimates are erroneous
enough to cause ‘runaway’ in the model alone if the
temperature measurements had not been used to esti-
mate the temperature profile. Thus, the state estimation
algorithm generates acceptable temperature estimates
even in the face of significant parameter bias. This
robustness is important because it gives confidence that
Figure 7 Parameter estimation experiment: estimated composition
profiles compared with data points at the time of parameter updates
the composition estimates are based upon a reasonable
temperature profile. Note that the control scheme
provides a separation between the parameter and state
estimation with resulting improved stability of the esti-
mates.
In examining the results in Figure 8 at the time of 25
min, we see that the first parameter update at 20 min
allows significant improvement in the estimated tempera-
ture profile versus the profile at a time of 12 min. How-
ever, more severe conditions at lower flow rates (33 min)
emphasize that the kinetics are still in error. After 50 min
of operation, the temperature profiles are quite well
represented under a variety of steady state conditions.
Note that even the more severe conditions of a wall
temperature of 270°C and a low flow rate are accurately
estimated at time = 137 min.
It is concluded that reliable estimates of the reactor
behaviour and a reasonable approximation of the cata-
lyst activity profile in the reactor can be obtained on-line
after about five composition measurements. This perfor-
mance is good even in the case of a poor guess for the
preexponential factors of the kinetic rate constants.
Operating conditions can be changed as needed, and
there is no requirement that the system be at steady state.
The estimated maximum reactor temperature is needed
for control and optimization, and the maximum temper-

32 J. Proc. Cont. 1992, Vol2, No 1

 Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

Table 1 Parameter estimation experiments

Number Time (min) T, = T;(‘C) u, CH,OH Flow (g SK’) Sample time (min) Sample location (cm) Comments

1 0.0-12.5 260 0.05 1.0 8.0 65 Reactor at steady state

2 12.5-25.0 260 0.05 1.4 20.5 35 Flow change
3 25.k37.5 260 0.05 0.5 33.0 55 Flow change
4 37.5-50.0 250 0.05 0.5 45.5 45 TWaII change
5 50.0-62.5 250 0.05 1.4 58.0 65 Flow change
6 62.5-75.0 250 0.03 0.5 70.5 35 Feed% and flow change
I 75.0-87.5 260 0.05 0.7 83.0 55 T,,,,, feed%, flow
8 87.5-100.0 270 0.05 0.7 95.5 45 T,,,, change
9 100&l 12.5 270 0.03 1.0 108.0 65 Feed% and flow change
10 112.5-125.0 270 0.03 0.5 120.5 35 Flow change
11 125.0-137.5 270 0.05 0.5 133.0 55 Feed% change
12 137.5-150.0 270 0.05 1.4 145.5 45 Flow change
13 150.&160.0 260 0.05 0.5 158.0 65 T,,!,, flow; repeat # 3

Initial conditions for reactor and estimation: T, = T, = 260, Y, = 0.05, flow = 1.O; temperature uniform at 26o’C; methanol flow begins at time = 0

350 350 reactor based on three temperature sensors and compo-

Time = 12 min 3501 Time = 25 min Time = 33 min
sition measurements taken every 12.5 min. All compo-
sition samples were taken close to the reactor exit at 65
cm from the inlet.
For all of these experiments, the control scheme was as
indicated in Figure 3. The controller settings were chosen
from model simulations and were not changed during the
350rl 3501
TI ime = 75 min
experiments. The values used were as follows:

l inner loop control of the recirculating jacket oil

temperature
a. proportional band of 3°C.
b. resets/min = l/r, = 1.6 min-I
0 35 70
c. derivative rate: zd = 0.1 min
Axial distance (cm)
0 outer loop control of reactor maximum tempera-
ture by jacket wall temperature (setpoint)
a. Kc = 0.7
320
6 b. integral action: l/q = 0.5 min-I
4 300 c. maximum rate of change of setpoint: 5°C min-1
5
f;; 280
L A description of the experiments is shown in Table 2.
i 260 0 35 70 35 70
These experiments demonstrate four types of control
problems:

Cl - setpoint change and disturbance rejection;

C2 - control with unmeasured wall temperature;
C3 - control with errors in heat transfer para-
meters;
C4 - combined parameter estimation and control
with errors in kinetic parameters.
Axial distance (cm)

Figure 8 Parameter estimation experiment: comparison between mea-

The purpose of the first experiment, designated Cl, is
sured and estimated temperature profiles straightforward testing of the control loop. The experi-
ment is divided into eight parts (a-h) of 15 min duration
ature estimates improve as the parameter estimates are each with inputs as shown in Figure 9. Note that the
appropriately updated. performance of both the inner loop (Tw,,J and the outer
loop (T,,,,,) is shown. The control performance in experi-
ment Cl is shown in Figure IOa-h. Figure 10a demon-
Maximum reactor temperature controller tests strates reactor start-up from 240°C to a maximum reac-
tor temperature of 325°C with the estimation and control
Having demonstrated the performance of the state esti- module given a poor initial guess of reactor conditions.
mation and parameter estimation algorithms, we can In order to start up the reactor efficiently, the wall tem-
now test the complete control system. The goal of the perature is increased at its maximum rate for 45 min,
control experiments was to demonstrate satisfactory and then decreases before reaching steady state after 10
closed-loop control of the maximum temperature in the min. The T,,,,, estimate rapidly converges and tracks the

J. Proc. Cont. 1992, Vol2, No 1 33

Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

Table 2 Control experiments

Experiment Time (mm) r, CH,OH Flow (g s- ‘) TmnisetW) Sample time Sample location Comments

CIA CL15 0.05 0.70 325 6.0 65 Reactor start-up

ClB 15-30 0.05 0.70 340 18.5 65 Setpoint change
ClC 3&45 0.05 0.70 325 31.0,43.5 65 Setpoint change
ClD 4560 0.05 0.40 325 56.0 65 Flow disturbance
ClE 60-75 0.05 1.40 325 68.5 65 Flow disturbance
ClF 75-90 0.05 0.70 320 81.0 65 Flow + setpoint
ClG 90-98 0.045 0.70 320 93.5 65 Unmeas. Y,disturb.
ClG’ 98-106 0.040 0.70 320 106.0 65 Unmeas. Y,disturb.
ClH 106-120 0.050 0.70 320 118.5 65 Unmeas. Y,disturb.
c2 o-15 0.05 0.70 325 No wall T meas. feedback
c3 o-15 0.05 0.70 325 - Poor heat transfer param.
C4A O-12.5 0.05 0.7 325 8.0 65 Setpoint change
C4B 12.5-25.0 0.05 1.4 325 20.5 65 Flow disturbance
c4c 25.0-37.5 0.05 0.5 325 33.0 65 Flow disturbance
C4D 37.5-50.0 0.05 1.0 325 45.5 65 Flow disturbance
C4E 50.062.5 0.04 1.0 310 58.0 65 Y, + setpoint
C4F 62.5-75.0 0.05 0.7 325 70.5 65 Y, + setpoint

Initial conditions for Cl:

Reactor: T, = T, = 240, Y, = 0.0, F = 0.7, reactor uniformly at 240°C
Controller: T, = T, = 240, Y, = 0.0, F = 0.7, reactor uniformly at 300°C
Initial conditions for C2:
Reactor: T, = T, = 260, Y, = 0.05, F = 0.7, steady state
Controller: T, = T, = 270, incorrect initial temperature profile (based on model at T, = 270, Y = 0.05, F = 1.4)
Initial conditions for C3:
Reactor: T, = T, = 240, Y, = 0.0, F = 0.7, reactor uniformly at 240°C
Controller: T, = Ti = 240, Y = 0.0, F = 0.7, reactor uniformly at 300°C
Parameters: Pe = 40 (32) Le = 8 (6.5), Hfs = 0.015 (0.025)
Initial conditions for C4:
Reactor: T, = T, = 250, Y, = 0.05, F = 0.7, steady state
Controller: correct inputs, incorrect initial temperature profile (based on model at T,, = 260, Y, = 0.05, F = 0.5)

Good control of T,,,,, was observed for a disturbance

350 consisting of a feed flow rate decrease in the experiment
34o-a (Figure 1Od). As observed for open-loop step changes in
330- 1_,
320-
flow rate, there was little steady state change in T,
needed in order to maintain T,,, constant. Too tightly
tuned local control of the oil jacket caused some oscilla-
tion in T,, but this had negligible effect on T,,,. A
disturbance consisting of a feed flow rate increase (Figure
IOe) required greater control action, and the response of
T,,,,, was as good as could be expected in view of the
wrong-way behaviour of T,,, in response to the flow rate
change. The overshoot as T,,,,, achieved its final value
was minimal. Good control was also achieved during
LL
simultaneous flow and T,, setpoint changes (Figure
I1 I I I II 11 1 ’
0 IO 20 30 40 50 60 70 80 90 100 110 120 ZOj). The setpoint is reached within 5 min and with less
Time (min) than 1” error in the estimated T,,,,,.
The compensating control action to maintain constant
Figure 9 Inputs for control experiment Cl. Initial conditions: T, = T, reactor temperature during unmeasured disturbances in
= 24o’C, Y, = 0.0, F = 0.7 g s-l; reactor temperature = 24o’C;
reactor temperature estimate = 300°C inlet mole fraction methanol (Figure 10g,h) is entirely
dependent on sensing declining reactor temperatures.
Substantial changes were made in the wall temperature,
reactor temperature data. Good control performance in but some steady state offset existed due to the pro-
this case was indicated by the control variable operating portional nature of the gains in the temperature estima-
at its constraint and then decreasing smoothly to its final tion module. In the case of increasing feed mole fraction,
value. The steady state deviation between the estimate the maximum temperature overshot the setpoint because
and data for T,,,,, was small (less than 2°C). the dynamics of the disturbance was more rapid than the
Good control performance was also observed for step compensating wall temperature manipulation. However,
changes to increase and decrease the maximum reactor the ultimate control action was quite good.
temperature from 325°C to 340°C and back to 325°C (cf Figure II demonstrates that during the experiment Cl,
Figure ZOb,c). The control input was at the constraint for both the hotspot temperatures and the entire tempera-
its rate of change and then lined out, and the estimated ture profile are accurately known and controlled.
T,,,,, did not overshoot the setpoint. Another important feature of the estimation and control

34 J. Proc. Cont. 1992, Vol2, No 1

 Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

a
,/--
/
.___
..__. e--;__w--_
265

:,:
320 260

260

240
I/
I- ----
-------

5
Tmaximum
Estimate
Setpoint

10
255

250

0
-
-------

5
Twall
Setpoint

10 60
----
-------

65
Estimate
Setpoint

70
255-
60
I
65
------- Setpoint
I
70 75
Time (min) Time (min) Time (min) Time [min)
345 340, 262
b If
340 - ..___.... ~~..--~~~~
- Twall
------- Setpoint

- Twall
.------- Setpoint

I I
320 L
15 20 25 30 ‘“:- 80 85 90 75 80 85

Time (min) Time (min) Time (min) Time (min)

275 322 266

C C
_ Tmaximum - Twall g
---- Estimate ------- Setpoint 320 _____.______.
________
._._____
340 7 ------- Setpoint
‘$1
318

316
--- Estimat e
It
314

325 --__-_
312 - Twall
------- Setpoint

320 1 I
30 35 40 45

Time (min) Time (min) Time [min) Time (min)

345 266
d Id - Twall
- Twall
-------- Setpoint ;::p .------_ Setpoin
340 ____--_ Setpoint

335 A

330

325

320 I I I
45 50 55 60

Time (min) Time (min) Time (min) Time (mini

Figure 10 Control experiment Cl: response of maximum temperature control loop

scheme is that it accurately represent the yield perfor- monoxide are well represented as shown by the compari-
mance of the reactor. Both methanol conversion to son between the composition estimates and experimental
formaldehyde and the formaldehyde oxidation to carbon data in Figure 12.

J. Proc. Cont. 1992, Vol2, No 1 35

Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

320

Time = 27 min
Tset = 340
Yi = 0.05
F = 0.7

260
0 35

Time = 57 min
Tset = 325
320 320

280

260 260
0 35 0 35 70

Axial distance (cm) Axial distance (cm)

Figure 11 Control experiment Cl: comparison between measured and estimated temperature profiles

Carbon monoxide value within 12 min. The true wall temperature was
o”“---y-Ei changed rapidly until it coincided with the setpoint out-
put of the controller, and then it quickly was adjusted to
the original steady state value. For the sampling time
interval of the experiment, no complicating factors were

!+I-+
0
i 100
Figure 12 Control experiment Cl: comparison between
data and estimates at exit
The second control experiment, C2, demonstrates per-
formance when the boundary conditions (wall tempera-
ture) are not accurately known by the estimation and
control program. There is no direct measurement of the
wall temperature transmitted to the process control
computer. In this case, the wall temperature setpoint was
sent to the local controller, and the response of the local
loop is represented within the estimation and control
program as a simple first-order system. In addition, the
experiment verifies the accuracy of the closed-loop
control simulation which uses a first-order dynamic
model to represent the experimental response of the local
wall temperature control loop. The results of experiment
C2 are shown in Figure 13, where an initial error of 10°C
in the wall temperature estimate caused a 4°C upset in the
maximum temperature, which returned to its original
observed due to using the modelled rather than the mea-
sured wall temperature response. The reactor tempera-
ture estimates quickly recovered from the original erro-
neous boundary conditions, and the T,,,,, estimates were
within 2°C of the data after rapid initial convergence.
Experiment C3 tests the estimation and control system
in the presence of errors in the parameters for heat
transfer. The radial Peclet number was chosen 25%
larger than indicated by off-line analysis of previous
data, and the interphase heat transfer coefficient was
chosen 40% lower than the base case model. These inac-
composition curate parameters cause higher estimated reactor tem-
peratures as well as a more sensitive model, and the state
estimator must compensate accordingly. It is important
for the state estimator to be able to compensate adequa-
tely for a somewhat mismatched model in order that the
controller set the desired reactor temperatures, especially
the T,,,. The procedure of experiment C3 was the same
as C I a: the start-up of the reactor of a maximum temper-
ature of 325°C.
The results of this experiment were similar to Cla
except with more overshoot in the estimated T,,,,,and the
wall temperature (cf Figure 14). This result could be
expected, since less heat transfer gives a more sensitive
model having a higher steady state gain but a similar
time constant. Due to the model errors causing incorrect
temperature estimates, the reactor was operated with 4-
5°C lower wall temperature than the previous experi-
ments without deliberate model mismatch. This perfor-
mance is a great improvement over the 1520°C errors

36 J. Proc. Cont. 1992, Vol2, No 1

 Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

332 7
a
- Twall
270
- -- Estimate
----- Setpoint

268

266

264

260

- Tmax
258
--- Estimate
------- Setpoint

3221 I I I I I 2561 I I I I
0 2.5 5.0 7.5 10.0 12.5 0 2.5 5.0 7.5 10.0 1;

Time (mini Time (min)

Figure 13 Control experiment C2: response of maximum temperature control loop

- Tmax - Twall
--- Estimate ----- Setpoint
_---_-- Setpoint

5 10 15

Time (min) Time (min)

Figure 14 Control experiment C3: response of maximum temperature control loop

which occur in the model in the absence of state estima-
tion.
The fourth section of control experiments (C4 a-f)
demonstrates control of the maximum reactor tempera-
ture in the presence of significant errors in the kinetic
parameters. These experiments also demonstrate the esti-
mation of the kinetic parameters in a closed-loop control
setting, and the improvements in the composition state
estimates resulting from adjustments in the kinetic para-
meters. The maximum temperature setpoints and inputs
for this experiment are shown in Figure 1.5. Experiment
C4a demonstrates start-up of the reactor to a maximum
temperature of 325°C. The controller then attempts to
hold the maximum temperature at 325°C during step
changes in the flow rate (C4b,c,d). Then, in part C4e, an
operating change is made in the feed mole fraction meth-
anol from 0.05 to 0.04, and the setpoint for maximum
temperature is reduced from 325 to 310°C. This type of
change is typical of what occurs if there is known change
in the feed, and experience or off-line modelling indicates
a corresponding change to be made in the setpoint. The
final part of the experiment (C4f), is a change back to the
setpoint of Cla to demonstrate the difference in perfor-
mance as the estimates of the kinetic parameters
improve.
First, this sequence was tested through closed-loop

J. Proc. Cont. 1992, Vol2, No 1 37

Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray
gi=ypyg
0 10 20 30 40 50 60 70
Time (min)
Figure 15 Inputs for control experiment C4. Initial conditions:
T, = 250, Y, = 0.05, F = 0.7 g s-1; temperature profile at steady state
simulation, which allows comparison of the estimates of
temperature and composition with the ‘true’ reactor as
represented by the simulation. The estimated tempera-
ture profiles are compared with the actual values of the
reactor simulator in Figure 16. The catalyst activity pro-
file parameter estimates are shown in Figure I7a-c and
the resulting estimates of the exit compositions are com-
pared with the ‘true’ values from the simulator in Figure
Z7d,e. The quality of the estimates shows dramatic
improvement during the course of the simulation due to
the modifications of the inaccurate kinetic parameters.
The preexponential factor for carbon monoxide produc-
tion has reached its true value after five measurements,
and the improvement in the carbon monoxide estimate is
noticeable at the time of each parameter update.
Although methanol conversion also converges to its
actual values, the kinetic parameters do not converge to
their true values during the elapsed time of this experi-
ment. This is because a range of catalyst activity profiles
will produce the observed methanol conversion at the
single measurement location (the exit). Due to the chang-
ing operating conditions being used, there is observabi-
lity of the catalyst activity due to the varying parametric
sensitivity under different operating conditions. How-
ever, these effects are not sufficient for rapid convergence
of the estimator.
Next the sequence is repeated with the experimental
reactor. The experimental response of the maximum
reactor temperature for experiment C4 is shown in
Figure ISa-J When the kinetic parameter estimates are
in error, the temperature estimates are adversely affec-
ted. Even though the estimates of the maximum tempera-
ture are much better than if predicted by the model
alone, the estimated T,,, are initially as much as 8-10” in
error (Figure Z&z). The control response of the estimates
is good, but the actual reactor is controlled to an incor-
rect steady state at the beginning of the experiment. For
step changes in flow rate, the responses are similar to
experiment C 1. A major improvement in control perfor-
38 J. Proc. Cont. 1992, Vol2, No 1
mance is realized as the kinetic parameters are updated
to more accurate values and agreement between the data
and the temperature estimates is improved. The control
action taken as kinetic parameters change is most clearly
seen in Figure 186 at the time of 20 min. The estimated
Tmax quickly changes toward the actual T,,, and
increases the deviation from the setpoint. Then the con-
troller takes corrective action, decreases the wall temper-
ature, brings the estimate to the setpoint, and moves the
actual T,,,,, closer to the setpoint. This process is repeated
each time the parameter estimates are changed, and the
estimated T,,,,, moves to within 2” of the true value in
Figure ISfcompared with an 8” error for the same oper-
ating conditions in Figure 18~. The comparison of the
estimated and actual temperature profiles (Figure 29)
80
shows good estimates after only one composition sample
and gradual improvement thereafter. Noteworthy is the
substantial improvement between 12 min and 75 min
T, = operating at the same feed conditions at a setpoint of
T max= 325°C. The kinetic parameter estimates obtained
in experiment C4 (Figure 20) are consistent with those
obtained in the earlier parameter estimation experiments
and in fact somewhat better than the simulation results
of Figure 17. The parameters are close to their final
values after only three composition measurements. The
comparison between the composition estimates and the
data (Figure 21) indicates the relatively poor estimates of
the compositions early in the experiment and the relati-
vely good estimates later in the experiment. Substantial
changes in the composition estimates occur just after the
parameter updates (about 1 min before a new compo-
sition measurement is taken, at 20, 32 and 45 min).
Summary and conclusions
Successful experiments have been performed for on-line
parameter estimation and maximum temperature
control of a packed bed reactor. These applications used
a distributed parameter state estimation algorithm in
conjunction with a two-dimensional heterogeneous reac-
tor model. In parameter estimation, the axial catalyst
activity profile for the principal reaction and the rate of
by-product production were estimated. Good estimates
of initially erroneous parameters were obtained after five
to seven composition measurements. Long-term tracking
of slowly changing parameters was also demonstrated.
Using only a few temperature measurements, it is
possible to control efficiently the maximum reactor tem-
perature. The true temperature maximum during experi-
ments was determined by additional auxiliary tempera-
ture measurements from the reactor. For a step change in
the maximum temperature setpoint, the controller (in
this case, a recirculating oil loop heater) operated at the
constraint until the steady state value was reached, and
there was little overshoot in the maximum temperature
in the reactor. Because of the wrong-way behaviour of
the dynamic response to flow rate changes, the tempera-
ture control for flow rate disturbances could be improved
by implementation of model predictive control.
 Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

340
rime = 12 min -‘-I Time = 37 min

..... .. . . . .. .
320 ‘. .
I’ ’ \\
/ \\
300 / \\
z I’
e 280
r

..
?
3
?\
m
!g 260 70 0 35
0 35
E
‘, 340 340
Time = 50 min Time = 62 min Time = 75 min
.z
z r~ . . .
I
I

A
: 320 ‘\ 320
t
s

300
/
280 ., /
/

+ 70

Axial distance (cm)

Figure 16 Control simulation C4: comparison between measured and estimated temperature profiles

\ zfyyy-g
100 ___--_-----_-------- ------- - _--_

“Eb-1

0 10 20 30 40 50 60 70 80
Time (min]
Methanol Carbon monoxide’

o*750w 80 o~02%i%hi-
0 0

Time (min) Time (min)

Figure 17 Control simulation C4: estimates of catalyst activity and comparison between composition data and estimates at exit

The methodology described in this paper is useful for buted systems, for preventing reactor runaway, and for
controlling packed bed reactors and other similar distri- Qptimizing reactor performance measures such as pro-

J. Proc. Cont. 1992, Vol2, No 1 39

Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

- Tmaximum
---- Estimate
------- Setpoint
G
0
; 315-

- Twall - Twall
245[ , ---I_--- :etpo:nt ------- Setpoint
_---- Setpoint
300 I I , I
0 2.5 5.0 7.5 10.0 12. .5 0 2.5 5.0 7.5 10.01 50 52.5 55 57.5 60 62 i 50 52.5 55 57.5 60 62.5
Time (min) Time (min) Time (min) Time (min)

330. 266

330
- Twall
G .------- Setpoint
o 325 - Twall
f ________
Setpoir E :
1
zb 320 z :’
I 320.. :
x :
2 I
a 315
I- E
- Tmaximum k- - Tmaximum
315
3 1IJ _--_ Estimate
258 -
_____-_ Setpoint

3051 ’ ’ ’ ’ 1 2561 310 1 k I ! 2561

12.5 15.0 17.5 20.0 22.5 25.0 12.5 15.0 17.5 20.0 22.5 25.0 62.5 65 67.5 70 72.5 62.5 65 67.5 70 72.5

Time (min) Time (min) Time (min) Time (min)

---- Estimate
------- Setpoint

. - Twall
------- Setpoint
320 I I 2541
25 27.5 30 32.5 35 37.5 25 27.5 30 32.5 35 3 7.5

Time (min) Time (min)

330

G
0 325
klf
,
z 320
%
E
aI 315
c-
_ Tmaximum
310 254 - - Twall
------- Setpoint
305 252 I I I , .
37.5 40 42.5 45 47.5 50 37.5 40 42.5 45 47.5 50

Time (min) Time (min)

Figure 18 Control experiment C4: response of maximum temperature control loop

duct yield. The estimation and control system effectively oxidation of methanol for formaldehyde production, is
overcame incorrect initial conditions, erroneous heat itself an important industrial packed bed reactor process
transfer parameters, erroneous kinetic parameters and operated around the world and is representative of wall-
unmeasured disturbances. cooled packed bed processes. Thus the approach demon-
The example system chosen for the study, the partial strated in this work should be applicable to other packed

40 J. Proc. Cont. 1992, Vol2, No 1

 Control scheme for packed bed reactors. I: L. C. Windes and W. H.

340
Time = 25 min 1

E
3
L”
$ 0 35 70

q
E
m
; 340 340
.E Time = 50 min Time = 75 min

t; 320 320 ,
u” f .

300 300

280 260 ’
.
.’
/*
7 2600%
2600 35 70 35

Axial distance (cm1

Figure 19 Control experiment C4: comparison between measured and estimated temperature profiles

Methanol Carbon monoxide

1 .o 0.12
Al at exit 0 Data
- Estimates
”
E
3
u 800
[a I I I I I I I da
m
c 0.9
.o
?
:
c
8
1 0.8
A2 at exit .o
”
2 0
II 0 Data
- Estimate!
0.7 I I I I I I
0 20 40 60 0 20 40 60 80
Time (mini
Time (min)

Figure 20 Control experiment C4: estimates of catalyst activity

Figure 21 Control experiment C4: comparison between composition
data and estimates at exit
bed processes of practical interest. The detailed pro-
cedures outlined here and elsewhereI should aid these 9 Vakil, H. B., Michelsen, M. L. and Foss, A. S. Znd, Eng. Chem.
applications. Fund. 1973, 12, 328
10 Harris, T. J., MacGregor, J. F. and Wright, J. D. Technometrics
1980,22, 153
11 McDermott, P. E., Mellichamp, D. A. and Rinker, R. G. Proc.
References American Control Conference, 1984, 1614-1620
12 Hong, J. C. and Lee, H. H. AZChE J 1985,31,302
Soliman, M. A. 1979 JACC - Plant Control & Optimization, 1979, 13 Gilles, E. D. and Epple, U. in ‘Modeling of Chemical Systems’ (Eds
140-145 K. Ebert, P. Deuflhard, and W. Jlger) Springer Verlag, Berlin, 1981
Silva, J. M., Wallman, P. M. and Foss, A. S. Ind. Eng. Chem. Fund. 14 Gilles, E. D. and Ruppel, W. in ‘Dynamics and Modeling of Reac-
1979, 18,383 tive Systems’ (Eds C. Conley, W. H. Ray, W. E. Stewart) Academic
Sorenson, J. P., Chem. Eng. Sci. 1977,32, 763 Press, New York 1980, 37-58
Bonvin, D., Rinker, R. G. and Mellichamp, D. A. Chem. Eng. Sci. 15 Epple, U. Proc. IFAC Symposium Dynamics and Control of Che-
1980,35,603 mical Reactors and Distillation Columns, Bournemouth, UK, 1986
Jutan, A., Wright, J. D. and MacGregor, J. F. AIChE J 1977, 23, 279-284
751 16 Windes, L. C., Schwedock, M. J. and Ray, W. H. Chem. Eng.
Lee, K. S. and Lee, W. K. AZChE J 1985,31,667 Comm. 1989,78, 1
Mandler, J. A., Morari, M. and Seinfeld, J. H. Chem. Eng. Sci. 17 Schwedock, M. J., Windes, L. C. and Ray, W. H. Chem. Eng.
1986,41, 1577 Comm. 1989,78,45
Mandler, J. A., Morari, M. and Seinfeld, J. H. Znd. Eng. Chem. 18 Windes, L. C., Cinar, A. and Ray, W. H. Chem. Eng. Sci. 1989,44,
Fund. 1986,25,645 2087

J. Proc. Cont. 1992, Vol2, No 1 41

Control scheme for packed bed reactors. I: L. C. Windes and W. H. Ray

19 Windes, L. C. and Ray, W. H. J. Proc. Cont. 1992,2,43 Boundary conditions:

20 Fortescue, T. R., Kershenbaum, L. S. and Ydstie, B. E. Automatica,
1981, 17, 831
21 Ydstie, B. E. and Sargent, R. W. H. IFAC Symposium on Identifi-
r= 1: ?&=O
cation & System Parameter Estimation, Washington, DC, 1981
ar
22 Schwedock, M. J. ‘Modelling and identification of a Catalytic
Packed Bed Reactor’, Ph.D. Thesis, University of Wisconsin aT, _
Madison, 1983
x - B&XT, - Tf)
23 Windes, L. C. ‘Modelling and Control of a Packed Bed Reactor’,
PhD. Thesis, University of Wisconsin Madison, 1986 ar, -
BUT, - TJ
ar

z = 0: aaZ = Pe,,Cyi - yjf)

Nomenclature
dTf_ 1
array of all fixed model parameters -
a,7
- 2Pefz(l + CJ)(T, r,) = \Iro(Tr K)
kinetic preexponential factor
az -
total flow rate through the reactor (g s-1)
controller gain aZ 1~0’s - T)
inverse information matrix for parameter estimation
time z = 1: !ZYi=()
temperature aZ
maximum temperature in the reactor
wall temperature (equivalent to coolant temperature)
control input
ar, = ;Pefz( 1 - o)(T, 7-w)= W‘-r Tw)
measured variable output az
dimensionless mole fraction based on feed mole fraction
az -
z axial location in reactor
az c(dTs - TeJ
z” axial position of catalyst activity inflection point

Greek letters Here (J and p0 are parameters accounting for radial

E error
transport at the inlet and exit
0 estimated parameters
k parameter estimation gain
h forgetting factor
z information content 24Bi,p x
(3= 1+
rD time delay for composition measurement 1 Pef,Pe&%f + 3) I

1 I4=0
r1 integral time for controller
4 parametric sensitivity 6Bi,p %
cl0 = Pe,,
1Pe,,Pe,,(Bi,, + 3)
Subscripts
co carbon monoxide Dimensionless parameters:
M methanol
set setpoint

Tr = -tuo cj = 5 y, z A T, = g
L ‘MO y MO 0
Superscripts
estimated quantity I
+ at reactor exit Tr = 7 z=Z ,=r a=L
0 at reactor entrance 0 L rt rt

Appendix

Packed bed reactor model

Mass balance on speciesj: l-methanol, 2X0: per, = “p Bi,, = h!$!? pe,, = Gy
fr fr sz

o _ 1 a2yi I a 1a ,ayi_ &

Pe,, az* Pem, r ar i ar 1 aZ

Energy balance on the fluid:

B, = ( - AH~)rloKd?‘o, TOIL

GC,Jo

Energy balance on the solid: Da = VORMCPO,TO)L = TIoMPO~TO)L~~

couo Go

St = SvW & = t1 - Eb)(PC&O = c1 - Eb)(PC&

h
CC,, wQ>f Pmcpf

42 J. Proc. Cont. 1992, Vol2, No I