Beruflich Dokumente
Kultur Dokumente
C-VVD
0
time
Fig. 2. Representation of traditional control strategies in
terms of the function.
805
8th IFAC Symposium on Advanced Control of Chemical Processes
Furama Riverfront, Singapore, July 10-13, 2012
1963, Srinivasan et al., 2003). We use the fact that the determinant we find the expression for optimal control on
condition H = 0 implies that its derivatives with respect so-called singular arc, if such exists.
to time are equal to zero as well. We will make use of the
The overall control strategy will not change from the
following equations
previously mentioned one. In this case however, switches
H (c1 , c2 , V, 1 , 2 , 3 ) = 0, (8a) between constrained and singular control trajectories have
H (c1 , c2 , V, 1 , 2 , 3 ) = 0, (8b) to be found by other means. In our study we find them
(c1 , c2 , V, , 1 , 2 , 3 ) = 0. numerically by formulating a simple optimization problem.
H (8c)
to eliminate the adjoint variables . Hence that first
4. RESULTS
two conditions are control variable free. Further it can
be shown that conditions (8) form a system of linear
homogeneous equations in variables 1 , 2 , 3 (Bryson, Jr. The optimality conditions (8a) and (8b) give after some
and Ho, 1975). manipulations condition for singular surface
S = 1 c1 S1 + 2 c2 S2 = 0, (11)
3.1 Optimal control in special cases where Si (for i = 1, 2) is given as
Si = (Ri 1)(q + c1 q1 + c2 q2 ) + q(c1 Ri1 + c2 Ri2 ), (12)
As it will be shown later, the optimal state surface will
be in special cases a function of concentrations only, and
q Ri
S(c1 , c2 ) = 0. Thus, it will be a curve in the concentration qj = Rij = i, j = 1, 2. (13)
space. Once it is found, the corresponding singular control cj cj
can be obtained by considering its derivative with respect Although the singular surface (11) depends on unknown
to time trajectories of adjoint variables, we can eliminate them in
1 , c2 ) = S c1 + S c2 = 0.
S(c (9) some special cases as follows
c1 c2
R1 = 1 (R11 = R12 = 0). This represents a common
Using process differential equations (1) then yields for situation for a macro-solute that does not get through
S S
c1 c1 R1 + c2 c2 R2 the membrane and micro-solute can have arbitrary
(t) = S S
. (10) properties. The optimal curve S(c1 , c2 ) is given as
c1 c1 + c2 c2
(R2 1)(q + c1q1 + c2 q2 ) + q(c1 R21 + c2 R22 ) = 0 (14)
To summarize the results, once the optimal concentration
surface S(c1 , c2 ) is found the optimal operation can be both R1 , R2 are constant (Rij = 0). If both retention
stated as follows: coefficients R1 and R2 are constant and do not depend
on concentrations (for example a perfect membrane
(1) The first step is either pure dilution ( = ) or with R1 = 1, R2 = 0) the optimal curve is given as
pure filtration ( = 0) until state variables arrive at
optimal curve S(c1 , c2 ) = 0. q + c1 q1 + c2 q2 = 0 (15)
(2) The second step is diafiltration with time dependent In both these special cases we can proceed to find expres-
(t) given by (10) maintaining optimal concentration sions for optimal control (10) and use directly the optimal
values. control procedure as stated in Section 3.1.
(3) Finally, the third step is again either pure dilution
( = ) or pure filtration ( = 0) until final The expression for singular control in general case can be
concentrations of both components are obtained. derived by calculating the determinant of the homogeneous
system (8). This gives
Any of these three steps can be missing at a particular
(S1 S2 )b3 + S1 b2 S2 b1
problem, depending on process initial and final conditions = , (16)
as well as actual functions Ri (c1 , c2 ), q(c1 , c2 ). S2 a 1 S1 a 2
where expressions ai and bi for i = 1, 2 are given as follows
This result establishes that all traditionally used opera-
tions (C-CVD-C, C-VVD, VVD) are potentially optimal Si Si
ai = c1 q c2 q , (17)
for some special types of problems. However, neither one c 1 c2
of these can be concluded as generally time-optimal. Si
bi = c1 qR1 (qRi1 + Ri q1 )S1 (18)
c1
3.2 Optimal control in general case
Si
+ c2 qR2 (qRi2 + Ri q2 )S2 , (19)
c2
In general it might be not possible to end up with closed
form representation of singular surface without using ad- and
joint variables . However, it is possible to find an expres- b3 = c1 q1 S1 + c2 q2 S2 . (20)
sion for optimal control as a function of concentrations Therefore, optimality conditions provide only control (16)
only. along singular arc but not the state arc itself.
This argument is based on a fact that optimality condi-
tions (8) represent the system of homogeneous equations 5. CASE STUDIES
linear in adjoint variables. This system has a non-trivial
solution only if the determinant of its coefficient matrix In this section we concentrate on selected case studies from
is zero (Srinivasan et al., 2003). Thus by computing the literature on the optimal control of diafiltration processes.
806
8th IFAC Symposium on Advanced Control of Chemical Processes
Furama Riverfront, Singapore, July 10-13, 2012
c2 [g/dL]
= 63.42 12.439 ln c1 7.836 ln c2 (21b) 3
where c1 is concentration of proteins and c2 denotes
concentration of lactose. 2
The optimum concentration curve depends on both con-
centrations and is given by (15) as 1
S(c1 , c2 ) = b0 + b1 + b2 + b1 ln c1 + b2 ln c2 = 0. (22)
Once these optimal concentrations are obtained the con- 0
0 5 10 15 20 25
trol is calculated from (10) c1 [g/dL]
b1
(t) = = 0.61. (23)
b1 + b2
1
We consider to drive concentrations from initial point
[c1,0 , c2,0 ] = [3.3, 5.5] to final point [c1,f , c2.f ] = [9.04, 0.64]. 0.8
To perform this task in minimum time we use a three step
strategy (see state diagram in Fig. 3):
0.6
(1) Start at green circle, horizontal line: concentrate with
The resulting final time in this case is 4.49 hours. This can
min. time
be compared to the operation described in Rajagopalan 0
CCVD
and Cheryan (1991) where two step process (C-CVD) was
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
used. This traditional operation takes for the same initial time [h]
and final conditions 4.74 hours, an increase of 5.3%. As
we can see from the lower diagram in Fig. 3, traditional
CVD step ( = 1) starts earlier but it takes more time Fig. 3. Separation of lactose from proteins: comparison of
to reach the final point as the VVD step ( = 0.61) in minimum time and C-CVD control strategy in con-
minimum time control. There, it is assumed that the last centration diagram (upper plot) and corresponding
step (upward arrow) takes no time. Although this is not control (lower plot).
true in reality, we can simply move the dilution step out
the batch to further processing. restriction implies that optimal pure dilution step will be
actually replaced by CVD step.
5.2 Sucrose sodium chloride separation It is desired to concentrate sucrose and dilute sodium
chloride in solution from their initial concentrations given
This case study is taken from our previous study (Fikar by point [c1,0 , c2,0 ] = [10, 250] to final concentrations
et al., 2010) where we concentrated on utilization of represented by the point [c1,f , c2,f ] = [50, 50].
numerical methods of dynamic optimization to derive the The optimum concentration curve depends on both con-
optimal control of diafiltration process using an economic centrations and is given by (14). This curve was found
cost function. using numerical nonlinear equation solver. Fig. 4 shows for
This case study represents diafiltration system with one comparison of the minimum time and C-CVD-C control
variable retention coefficient (R1 is almost constant and strategy. Results show that minimum time approach takes
equal to one) and the empirical relations for q and R2 as 10.2 hours. This is visualized in lower plot in Fig. 4 as a
functions of feed composition are as follows: dashed line for the purpose of lucidity of actual comparison
of two control approaches.
q = U1 (c2 )eU2 (c2 )c1 , (24a) In contrast to that, a traditional solution with C-CVD-C
V2 (c2 )c1 strategy lasts 14.5 hours, which yields 42% optimality loss.
R2 = V1 (c2 )e , (24b)
Although that two-step approach (C-CVD) would result in
where U1 , U2 , V1 , V2 are second order polynomials which faster process it yields unacceptably high concentrations
coefficients were determined from laboratory experiments of salt (during the process run) out of the range studied
with the process solution in Kovacs et al. (2009). For in Kovacs et al. (2009). This can be observed from upper
operational reasons, it is assumed that [0, 1]. This plot in Fig. 4. For C-CVD approach, it would be neces-
807
8th IFAC Symposium on Advanced Control of Chemical Processes
Furama Riverfront, Singapore, July 10-13, 2012
550
500
0.5
450
400
0.4
350
c2 [mol dm3 ]
300
c2 [g/dL]
0.3 min.time
250 CCVDC
200 0.2
150
S(c ,c )=0 0.1
100 1 2
50 min. time
CCVDC 0
0
0 10 20 30 40 50 60
10 15 20 25 30 35 40 45
c1 [mol dm3 ] c1 [g/dL]
1
1 min.time
0.9 CCVDC
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
min. time min tf 0.1
0 CCVDC min tf
0
0 2 4 6 8 10 12 14
0 0.1 0.2 0.3 0.4 0.5
time [h] time [h]
Fig. 4. Comparison of minimum time and C-CVD-C con- Fig. 5. Comparison of minimum time and C-CVD-C con-
trol strategy for sucrosesodium chloride separation. trol strategy for radiopaque ethylene glycol separa-
Upper plot concentrations diagram, lower plot (t) tion. Upper plot concentrations diagram, lower plot
trajectory. (t) trajectory.
sary to follow the first part of the red trajectory until
c1 = 50 mol dm3 . This would result in inadmissibly high This example represents a situation when we are not able
concentration c2 . VVD approach is clearly sub-optimal to obtain expression for optimal concentration surface ana-
since it takes 22.8 hours (124% optimality loss). lytically. We proceed as suggested previously and derive an
expression for singular optimal control from (16). Then we
use numerical optimization to find corresponding lengths
5.3 Radiopaque ethylene glycol separation of intervals for boundary values of control as well as for
singular one. Results indicate that the optimal control
In this case study we treat modified case study taken trajectory consists of three parts: pre-concentration, singu-
from Lutz (1997) where filtration using reverse osmo- lar arc, and post-concentration step. Numerical procedure
sis membrane was studied to treat a solution containing determines lengths of all these parts. Once the structure
12 g/dL of radiopaque component (c1 ) and 0.5 g/dL of and lengths of respective intervals are fixed, we can operate
ethylene glycol (c2 ) to end up with the product with the process optimally with singular control (16) in the
concentrations: 40 g/dL of radiopaque and 0.01 d/dL of middle part.
ethylene glycol. Experimentally obtained membrane char- Fig. 5 shows optimal evolution of concentrations under
acteristics are as follows minimum time control , sketched as well. When com-
q = 29.19 ln c1 + 118.1 (25) pared with traditional control strategies, minimum time
R1 = 1 (0.01c1 + 0.25c2 + 0.1) (26) strategy saves 4.5% of process time in comparison with C-
CVD-C and 18% of process time when compared to VVD
R2 = 1 (0.0073c1 + 0.813) (27) control strategy. C-CVD-C is graphically compared to the
For the purpose of this example constants which character- minimum time one in Fig. 5. Dashed line is used, for lower
ize rejection of radiopaque were slightly changed to reflect plot, to clearly distinguish the end-point of minimum time
the situation where rejection R1 is not close to one. strategy.
808
8th IFAC Symposium on Advanced Control of Chemical Processes
Furama Riverfront, Singapore, July 10-13, 2012
809