Professional Documents
Culture Documents
C
)
Figure 2: Simulated temperature profile vs process data (diamonds)
10
4. Linearisation
The steady state model was used to generate the relationship between the refluxes and product
qualities. The results for two such analyses are shown in Figure 3. Both top and bottom
qualities are shown to be highly non-linear in the composition region of interest. Some
improvement in linearity is observed when a logarithmic transform is applied.
Logarithmically transformed quality variables are therefore used for the remainder of this
study.
0
0,005
0,01
0,015
0,02
100250 100750 101250 101750 102250
Major Reflux (STDm3/hr )
P
r
o
p
a
n
e
m
o
l
%
i
n
t
o
p
-3,5
-3
-2,5
-2
-1,5
T
r
a
n
s
f
o
r
m
e
d
A
x
i
s
Transform
0
0,01
0,02
0,03
0,04
0,05
18,5 18,6 18,7 18,8 18,9 19
Minor Ref lux (m3/hr )
P
r
o
p
y
l
e
n
e
m
o
l
%
i
n
b
o
t
t
o
m
-4
-3
-2
-1
0
T
r
a
n
s
f
o
r
m
e
d
A
x
i
s
Transform
Figure 3: Nonlinear steady state relationships between reflux and composition
5. Steady State Gain Analysis
Gain relationships between minor and major reflux and the transformed top and bottom
qualities are generated from linear regressions of the data in Figure 3. After scaling, the
slopes are assembled into the 22 steady state gain matrix as follows:
=
07 . 1 56 . 0
64 . 2 50 . 1
G
This matrix has an RGA number of 11.7, which could be considered acceptable. However the
gain ratios (-0.57 versus 0.52) are not largely dissimilar when the uncertainty surrounding
the regression step is considered. There is some risk that the actual gain relationship is
colinear. A better conditioned alternative is obtained when a reflux/distillate ratio is
employed as a manipulated variable in preference to a reflux alone. For example, the gain
11
matrix for R
min
/(R
min
+D) and major reflux (R
maj
) as independent variables yields the following
matrix with RGA number of 1.8:
=
77 . 0 35 . 0
37 . 3 99 . 0
G
This matrix has an RGA number of 1.8 and gain ratios of 0.29 and 0.46. Therefore a
control design that employs reflux ratio and major reflux as slaves can be considered. In this
case, a new basic control loop that controls reflux ratio in cascade with minor reflux is
required.
6. Dynamic Modelling
Starting from the steady state model, a dynamic simulation was constructed in HYSYS by
specifying additional engineering details including pressure/flow relationships and equipment
dimensions. In addition, all basic controllers were included in the model and configured
exactly as they are in the plant. The dynamic model was checked for consistency and
calibrated against process data. The engineering time for the tasks associated with dynamic
simulation totalled about 2 weeks of a Control Engineer with no experience in dynamic
simulation.
A dynamic simulation of this size and complexity is numerically intensive. At best the
simulation could run 75 times faster than real time on a PC with 2 Gigahertz CPU and 512
Megabytes of RAM.
7. Step Test
The step test was performed within the HYSYS environment by employing the event
scheduler to step each manipulated and feedforward variable in turn. One up and one down
step was performed on each of the 3 manipulated and 3 feedforward variables. Between each
step the process was allowed 40 hours to achieve steady state. As each HYSYS run took
several hours, and during a run CPU resources are consumed entirely by HYSYS, it was
found most practical to perform the test runs at night while the PC was otherwise not in use.
12
Figure 4: Dynamic Matrix for Propylene/Propane Splitter
8. Identification
Data recorded during the test run was exported to Aspentechs DMCplus Model identification
package using the .clc format, which is conveniently available within the HYSYS
environment. Logarithmic transforms were applied to composition vectors. FIR and
subspace models were then generated from the step data. Given that the step data is noise
free, the identified models were observed to be very clean. Figure 4 shows a typical dynamic
matrix generated by this analysis in which scales have been removed for clarity.
9. Final Design
Analysis of plant historical data showed that the top quality specification was never exceeded
whereas the bottom quality was somewhat erratic. It was therefore decided to stagger the
implementation of advanced controls into 2 stages. The first stage was to implement a
simplified MISO strategy in the DCS in which bottom quality only is controlled by minor
reflux, with feed rate and atmospheric temperature providing feedforward compensation. The
second stage was the full DMCplus implementation with 3 manipulated, 3 feedforward and 3
controlled variables.
13
The block diagram for this MISO scheme is shown in Figure 5. Subspace models from the
previous step were used to tune PID regulator as fast as possible, but still within acceptable
stability limits. Dynamic compensators for the feedforward variables were derived from the
quotient of the subspace model pertaining to feedforward and manipulated variable
respectively. In this case, static compensation was sufficient.
10. Closed Loop Simulation
The simplified MISO scheme was implemented in HYSYS using the PID and lead/lag
function blocks. The regulatory responses of the controller to changes in measured (feed rate,
atmospheric temperature) and unmeasured (feed composition) disturbances were simulated.
Fine tuning and testing of the control scheme was performed.
Testing and tuning of the DMCplus design is scheduled for implementation as part of the
FCCU APC revamp project in October this year. For now, a basic simulation of the closed
loop performance was done using Aspentechs DMCplus Simulate program.
Analyser
Validation
PID
Regulator
Static
Compensator
Static
Compensator
Atmospheric
Temperature
Feed Rate
Propylene
mole %
+
Minor
Reflux SP
SP
PV
OP
Logarithmic
Transform
Figure 5: MISO control structure showing feedforward components
14
11. Commissioning
Commissioning time for the MISO controller was relatively quick. To date, no on line
retuning has been performed. Figure 6 shows the variation in bottom quality before and after
commissioning. Note that spikes to the left of the plot are due to erratic behaviour of the
analyser and not only poor control. Clearly a vast reduction in bottoms quality variation has
been observed.
The insert in Figure 6 shows that the average bottom quality meets the 1.5% specification
despite a periodic oscillation of about 1 %. On closer examination, this is observed to occur
at a frequency of exactly 1 oscillation per day. Clearly the diurnal temperature variation is the
source of this disturbance. Considering that changes in reflux take far longer than 12 hours to
take effect, it is physically impossible to suppress the effect of atmospheric and other
disturbances that occur at these high frequencies. Some further tuning of the atmospheric
temperature feedforward compensator may however be worthwhile.
0
20
40
60
80
100
120
4/1/03 3/2/03 5/3/03 4/4/03 4/5/03 3/6/03 3/7/03
Ti me
P
r
o
p
y
l
e
n
e
m
o
l
e
%
COMMISSIONING
Figure 6: Before/after trend of propylene mole % in bottom product
15
Conclusions
A design procedure for advanced process controllers utilising first principles steady state and
dynamic models as an alternative to empirical models identified from plant tests has been
presented.
The procedure was illustrated on the challenging control problem posed by the
propylene/propane splitter for which it was argued that a classical plant step test was not
feasible.
The procedure is based on the premise that a realistic dynamic simulation of the process and
every variable that participates in the control scheme can be developed, like most of the
distillation units with long settling times. Unfortunately this is not always the case, like in
reactor units where the modelling efforts are more time consuming than plant step testing.
References
1. Svrcek, W.Y., Mahoney, D.P., Young, B.R., page-7 A Real-Time Approach to
Process Control, 2000, J ohn Wiley & Sons, ISBN 0-471-80452-5.
2. Beaver, W.H., Davies, P., Minimising Plant Testing For Multivariable Predictive
Control Of An Acrylonitrile Plant, ISA 2001, 10-13 September, Houston TX.
3. Mather, U., Conroy, R.J ., Successful Multivariable control without plant tests,
Hydrocarbon Processing, J une, 2003.
4. McMillan, G.K., Cameron, R.A., Chapter 1, Models Unleashed, Virtual Plant and
Model Predictive Control Applications, 2004, ISA, ISBN 1-55617-857-3.
5. Laurier, G.C., Predictive Control Robustness Testing Through Dynamic Simulation,
ISA 2001, 10-13 September, Houston TX.
6. Kister, H.Z., Distillation Operation, 1990, McGraw-Hill, ISBN 0-07-034910-X.
16