Regular article
a r t i c l e i n f o a b s t r a c t
Article history: Dynamic optimization is a very important way to increase the productivity or protability of biochemical
Received 22 December 2015 processes. As an efcient approach for solving these biochemical dynamic optimization problems, control
Received in revised form 28 February 2016 vector parameterization encounters the difculty of selecting an optimal discretization level which bal-
Accepted 12 March 2016
ances the computational cost with the desired solution quality to obtain high accuracy solution. To tackle
Available online 15 March 2016
this issue, a new slope analysis is proposed to analyze the control variables and discretization time grid,
results nd that low slope time grid nodes have less effect on the improvement of performance index
Keywords:
and can be regarded as unnecessary nodes, while high ones are important time points. Based on this, a
Bioprocess optimization
Biochemical processes
novel non-uniform adaptive grid renement control parameterization approach is therefore presented,
Grid renement where the slope analysis is applied to rene or to coarsen the time grid so as to obtain a suitable dis-
Non-uniform grid cretization level with a small number of control intervals. By application in three well-known biochemical
Control parameterization optimization problems, results show that the proposed method is able to achieve similar or even better
Dynamic optimization performance indexes with small numbers of control intervals and lower computational costs.
2016 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.bej.2016.03.006
1369-703X/ 2016 Elsevier B.V. All rights reserved.
64 P. Liu et al. / Biochemical Engineering Journal 111 (2016) 6374
Since the dimension of the NLP problem in CP is much greater Subject to:
than in CVP, this paper focuses on the CVP method. Usually, the
x(t) = f [u(t), x(t), t]
solution quality of CVP method greatly depends on the discretiza-
tion level, accurate solutions require ne discretization. However, E[x0 (t0 ), t0 , x(tf ), tf ] = 0
a very ne discretization may make the discretized NLP problem C[t, x(t|u), u(t)] 0 (2)
become very large scale and/or ill-conditioned [19,20]. Hence, there
t0 t tf
rises a question to select the optimal discretization level which bal-
ances the computational cost due to the number of parameters with u u(t) u
the desired approximation quality.
where x(t) is an (n 1) state vector, u(t) is an (m 1) control vector,
One possible way is the optimization of the discretization grid
J(x(t), u(t)) is the performance index and consists of the perfor-
points. For example, Teo et al. [2426] presented the well-known
mance of the function 0 (x(tf ), tf ) at the terminal time tf and
time-scaling technique to optimize the original time points, where t
the performance of function t f L0 (u(t), x(t), t)dt during t [t0 , tf ].
the length of the time intervals are regarded as additional deci- 0
sion variables. This method is very efcient for bangbang control f [u(t), x(t), t] is the differential equation of dynamic system. The
problems and has been widely used. While, there is a drawback boundary conditions are E[x0 (t0 ), t0 , x(tf ), tf ] and the path con-
that the discretization NLP problem will be more complex and dif- straints are C[t, x(t|u), u(t)]. u and u are the lower and upper
cult to solve [20]. Another possible way is the renement of the bounds on u(t), respectively. The objective of this problem is to nd
discretization mesh. For instance, Schlegel et al. [20] employed the optimal prole u(t) that minimizes J(x(t), u(t)) during t [t0 , tf ].
fast wavelet transformation and the resulting control representa-
tions in wavelet space for grid renement. Garca et al. [27] applied 3. Uniform discretization-based control vector
a renement to all positions in the time grid by halving the step parameterization
size from previous renement iteration until the stopping criteria
were fullled. Hadiyanto et al. [19] proposed a sensitivity-based The idea of control vector parameterization is to discretize and
step size renement method to improve the product quality of bak- approximate the control vector by a basis function with a limited
ing optimization. Bittner et al. [28] presented an automatic density number of parameters [29,30]. Then a low order B-spline function
function-based mesh renement algorithm (DENMRA) for aircraft (for instance piecewise constant function) can be used to represent
trajectory optimization problem. the control vector.
Since the solution of CVP method greatly depends on the chosen Firstly, partition the time horizon [t0 , tf ] into N even stages
discretization level of time grid, a new slope analysis is proposed [tk1 , tk ], k = 1, 2, , N, called time grid,
to analyze the relationship between the control parameters and t0 = t0 < t1 < t2 < . . . < tN = tf (3)
time grid, results show that low slope points have less effect on
the improvement of performance index and can be regarded as where tk is the time grid node. Next, the control vector can be
unnecessary nodes, while high ones are important points. Based expressed as
on this, a novel slope analysis-based time grid renement method
N
is therfore proposed to select optimal time grid nodes so as to N (t) =
u(t) u k [t ,t ) (t) (4)
obtain high-quality solution for biochemical dynamic optimization k1 k
k=1
problems with a small number of parameters and low computa-
tional cost. To limit the computational effort, the proposed method where u k u for k = 1, . . ., N and [t ,t ) is dened as
k1 k
starts with a coarse discretization level for the control input, the
coresponding time grid nodes are then automatically rened by 1, ift [tk1 , tk )
subdividing the high slope nodes and/or excluding the low ones. [t ,t ) (t) = , k = 1, 2, ..., N. (5)
k1 k
0, ift
/ [tk1 , tk )
With this renement, optimization is continued for a selected group
of input parameters in successive iterations. The proposed method By using the piecewise-constant basis function, the control vec-
is applied in three well-known biochemical dynamic optimization tor u(t) in time stage [tk1 , tk ) is approximated as follows:
systems to verify the efciency of the proposed method.
This paper is structured as follows. The statement of biochem- N (t) = k , t [tk1 , tk ), k = 1, 2, ..., N.
u(t) u (6)
ical dynamic optimization problem is discussed in Section 2 and N
Although Eq. (6) does not dene the value of u (t) at t = tf , it
the uniform discretization-based control vector parameterization
does not affect the evolution of the state trajectory [31].
is presented in Section 3. Resolution of the time grid is discussed
Through this approximate method, the original Problem (P1) is
in Section 4 and the non-uniform adaptive time grid renement
transformed into a nite-dimensional NLP problem.
procedure is presented in Section 5. Section 6 outlines the imple-
mentation of the proposed approach, and the simulation results are
3.1. Problem (P2)
discussed in Section 7. Finally, the conclusion is drawn in Section
N
8.
tk
minJ = 0 [x(tf ), tf ] + L0 [ k , x(t), t]dt (7)
tk1
2. Statement of dynamic optimization problem k=1
Subject to:
A class of typical biochemical dynamic optimization problems
can be stated as the following Bolza form: N
tk
Subject to:x (t) = L0 k , x (t) t dt
k=1 tk1
E x0 (t0 ) , t0 x tf , tf = 0
2.1. Problem (P1) (8)
N
k k
tf C(t, x(t| ), )(tk1, tk ) (t) 0
minJ(x(t), u(t)) = 0 (x(tf ), tf ) + L0 (u(t), x(t), t)dt (1) k=1
u(t)
t0 u k u,
k = 1, 2, . . ., N
P. Liu et al. / Biochemical Engineering Journal 111 (2016) 6374 65
In general, it is desirable to obtain a solution of Problem (P2) 5.2. Time grid renement strategy
which is close to the true solution of the original Problem (P1).
Based on the theory in [3335], if time grid [tk1 , tk ] 0 with N 5.2.1. Time grid exclusion
, the solution of Problem (P2) will equal to the solution of Problem The idea of time grid exclusion is to eliminate unnecessary grid
(P1). Following this result, a very ne time grid for control vector is nodes in the representation of k , k = 1, 2, , N.
a reasonable choice. However, it is not a feasible option for practical Theorem 1
biochemical optimization due to the following two reasons: T
Suppose 1 = [ 11 , . . ., 1N ] is the solution of single control
(r = 1) Problem (P2) with corresponding time grid nodes {t0 =
(1) Numerical accuracy. It should be noted that a very ne time t0 , t1 , t2 , . . ., tN = tf }, where tf is xed. If
grid means that a large number of decision variables need to
+
be calculated, which may make the discretized NLP problem |s1,k | + |s1,k | = 0, k = 2, , N, (11)
ill-conditioned [20].
(2) Computational effort. The computational time required for the the adjacent control vectors are equal, namely,
solving of large scale NLP problem increases signicantly with
the number of parameters [19]. Nevertheless, the solution of 1k = 1k+1 . (12)
a dynamic optimization problem with the sequential method
is strongly correlated to the cost for the sensitivity integration Then, the corresponding time grid nodes tk1 and tk can be com-
[16], which is a very large part of the computational effort. bined, and the time stages are eliminated into N 1 segments with
a new set of nodes {t0 = t0 , t1 , t2 , . . ., tN1 = tf }.
Thus a suitable number of decision variables is desired so that T
the solution of Problem (P2) is close to the solution of Problem Proof. Since 1 = [ 11 , . . ., 1N ] is a solution of Problem (P2), the
(P1). This rises the research question on the optimal selection of the performance index J is,
time grid which balances the computational cost with the desired N
tk
approximation quality. Based on this, an adaptive non-uniform J (
1 ) = 0 [x(tf ), tf ] + L0 [ 1k , x(t), t]dt. (13)
renement strategy, which generates suitable time grid for control tk1
k=1
variables, is presented in the following section.
+
Denote :={1, 2, , N 1}. Assume |s1,j | + |s1,j | = 0, thus,
5. Non-uniform adaptive time grid renement procedure
j+1 j j+1 j
| r r / tj tj1 | + | r r / tj+1 tj | = 0, (14)
5.1. Slope analysis of the control vector
As discussed in [19,20], the solution of Problem (P2) greatly since tj =/ tj1 and tj+1 = j =
/ tj , it is obvious that j+1 (j ). Let
depends on the chosen discretization level of time grid, it is worth I:={j, j + 1}, the performance index J can be rewritten as follows,
analyzing the relationship between the control parameters and
N tk
time grid.
Denition 1: The left slope sr,k
of the r-th control vector r , r = J( 1 ) = 0 [x(tf ), tf ] + L0 [ 1k , x(t), t]dt
tk1
I
k=1,k/
1, 2, , m at time grid node tk , k = 2, , N 1 is,
tk
+
The right slope sr,k of the r-th control vector r , r = 1, 2, , m As j = j+1 ,
at time grid node tk , k = 2, , N 1 is,
tj tj+1
j j+1
+ k+1
r k L0 [ 1 , x(t), t]dt + L0 [ 1 , x(t), t]dt
sr,k = r , (10) tj1 tj
tk+1 tk
tj+1
j
where rk and rk+1 are the control parameters at time intervals = L0 [ 1 , x(t), t]dt. (16)
[tk1 , tk ) and [tk , tk+1 ), respectively. As an example, Fig. 1 illustrates tj1
the slopes of the r-th control vector.
66 P. Liu et al. / Biochemical Engineering Journal 111 (2016) 6374
Substituting (16) into (15) gives desired solution quality, we mainly subdivide the important time
grid nodes. Based on the previous testing, the choice for the time
N tk
grid subdivision tolerance is recommended in the range i = 0.22.
J( 1 ) = 0 [x(tf ), tf ] + L0 [ 1k , x(t), t]dt
I
k=1,k/
tk1 Remark 3. The slope analysis of control variables is separately,
tj+1 which means that each ur (t), r = 1, 2, ..., m has its own rened
+
j
L[ 1 , x(t), t]dt. (17) time grid. For multivariable dynamic optimization problems, these
tj1 rened time grids should be re-integrated into one.
Finally, there are N 1 grid nodes in (17), denote a new set of time 5.3. Time grid renement for multivariable dynamic optimization
nodes as {t0 = t0 , t1 , t2 , . . ., tN1 = tf } and the corresponding con-
T
trol parameters are 1 := [ 11 , . . ., 1N1 ] , the performance index Since there are many multivariable dynamic optimization
is problems in biochemical processes, such as fed-batch bioreactor
N1
problems, the time grid renement is expanded as follows.
tk
J( 1 ) = J ( 1 ) = 0 [x(tf ), tf ] + L0 [ 1k , x(t), t]dt. (18) Algorithm 1. Renement strategy for multivariable dynamic opti-
tk1 mization
k=1
Thus, Theorem 1 is proved. Given: l-th iteration control vector (l) , time grid
T (l) :=[t0 , t1 , ..., tN (l) ], tolerance e , i .
Remark 2. For multivariable dynamic optimization problem,
Denote R to mark the renement.
Theorem 1 is also suitable. Denote k = [ 1k , . . ., m k+1 =
k ] and
for k = 1,2,. . .,N(l) -1 do
[ 1k+1 , . . ., m k and k+1 into (12), the result then
k+1 ], substitute
for r = 1,2,. . .,m do
can be obtained. +
Calculate sr,k = (rk+1 rk )/(tk tk1 ), sr,k = (rk+1
Based on Theorem 1, the time grid nodes, which have no sig- rk )/(tk+1 tk ).
+
nicant effect on the improvement of the approximation accuracy, if |sr,k | + |sr,k | e then
can be excluded in the sense that slopes satisfy Rrk : = 0 (record the exclusion node).
+
else if (|sr,k | i or |sr,k | i )
+
|sr,k | + |sr,k | e , r = 1, 2, ..., m, k = 1, 2, ..., N, (19)
Rrk : = 2 (record the subdivision node).
where e is a given tolerance. This can easily be accomplished by else
combining time intervals [tk1 , tk ] and [tk , tk+1 ] into [tk1 , tk+1 ]. Rrk : = 1 (record the original node).
It should be noted that the choosing of time grid exclusion tol- end if
erance will affect the judgment of the unnecessary time grid nodes. end for
In order to obtain high accuracy optimization result, e could not
Rk max([R1k , R2k , ..., Rm
k
]) .
be a big value. Therefore, the choice for the time grid exclusion
tolerance is recommended in the range e = 00.08. Rene time grid [tk1 , tk ] and [tk , tk+1 ] with new nodes.
end for
5.2.2. Time grid subdivision The rened time grid T (l) [t0 , t1 , t2 , . . ., tN (l) ] (N
(l) is the new
The key point in time grid subdivision grid renement lies in number of time grid nodes).
identifying those regions of the control prole which require a
higher discretization level. To limit the computational effort, the 6. Algorithm outline
renement starts the optimization with a coarse time grid as previ-
ous methods [1921]. However, the coarse time grid is insufcient Based on the specic details of the slope-based non-uniform
to obtain high quality solution. It is needed to rene these regions adaptive time grid renement, Algorithm 2 gives an overview of
to approximate the true control prole of Problem (P1). The imple- the proposed method and the ow chart is shown in Fig. 2.
mentation of time grid subdivision is shown as follows.
Remark 4. Basically, a heuristic stopping criterion is formulated
based on checking the relative improvement of the objective func-
(1) If the left slope sr,k satises
tion if
|sr,k | i , r = 1, 2, ..., m, k = 1, 2, ..., N (20) |J (l) J (l1) | J (22)
where i is a given tolerance (i > e ). Then the time interval where J is a specied tolerance. However, since a very ne time
[tk1 ,tk ] will be subdivided into NE stages. grid may lead to ill-conditioned Problem (P2), another stopping
criterion is proposed if
+
(2) If the right slope sr,k satises min{tj } tmin , j = 1, 2, , N 1 (23)
Table 1 Table 2
Abbreviation description. Test results for case 1.
(a) 2 (b) 5 5
0 0
1.5 5 5
0 0.5 1 0 0.5 1
1
1.5
0.5
1
Initial grid
Refined grid
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Iteration=1 Iteration=2
(c) 5 5 (d) 5 5
0 0
0 0
5 5
5 5
10 10
Left slope Right slope Left slope Right slope
10 10 15 15
0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1
2 2
1.5 1.5
1 1
0.5 0.5
Refined grid Refined grid
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Iteration=3 Iteration=4
0.24
1
0.235
0.23
0.225
Optimal profiles
0.22
0.2 0.3 0.4 0.5 0.6
(a) 2 (b) 2 2
1.8
0 0
2 2
Left slope Right slope
4 4
0 5 10 15 0 5 10 15
1
2
0.8
1.5
0.5
Initial grid Refined grid
0 0
0 5 10 15 0 5 10 15
Iteration=1 Iteration=2
(c) 5 5 (d) 10
10
0 0 0
0
10 10
5 5
20 20
Left slope Right slope Left slope Right slope
10 10 30 30
0 5 10 15 0 5 10 15 0 5 10 15 0 5 10 15
2 2
1.5 1.5
1 1
0.5 0.5
Refined grid Refined grid
0 0
0 5 10 15 0 5 10 15
Iteration=3 Iteration=4
with Table 4
Test results of case 2.
g1 (t) = 4.75g3 (t)/(0.12 + g3 (t))
Iter. SNR-CVP UR-CVP
g2 (t) = x4 (t)e5x4 (t) /(0.1 + x4 (t)) (28)
P J(max) tcpu (s) P J(max) tcpu (s)
g3 (t) = 21.87x4 (t)/((x4 (t) + 0.4)(x4 (t) + 62.5)) 1 20 32.4533 8.41 20 32.4533 8.83
2 23 32.6309 12.83 40 32.6326 26.38
where x1 (t) and x2 (t) are the concentration of the secreted protein 3 33 32.6866 20.27 60 32.6887 64.84
and the total protein (g/L), x3 (t) is the culture cell concentration 4 49 32.7146 47.50 80 32.6902 123.30
(g/L), x4 (t) is the substrate concentration (g/L), x5 (t) is the holdup Total tcpu (s) 89.01 223.35
volume (L), u(t) is the glucose feed rate (L/h). The objective is to
maximize the mass of protein produced with t = 15 h. The dynamic
optimization problem has the form:
maxJ = x1 (tf )x5 (tf ) achieves a better maximal protein product of 32.7146 g with
u(t) fewer parameters and lower computation cost compared with
Subject to : Dynamic equations(27) the UR-CVP method. Luus et al. [44] used an iterative dynamic
x(0) = [0 0 1 5 1]T (29) programming algorithm to obtain a value of 32.46. Tholuder et al.
[9] achieved a value of 32.47 by employing a ltered iterative
tf = 15h dynamic programming (FIDP) procedure with 3651 s CPU time.
0 u(t) 2.5 Eva et al. [11] presented a mesh rening technique with CVP
method to obtain a value of 32.480 with 316 s computation time.
During the test, N (0) = 20 is set and T (0) = Angira et al. [45] applied a trigonometric differential evolution
[0, 3/4, 6/4, 9/4, ..., 15], the initial control parameter u(0) = 0.5 (TDE) approach for solving this problem with a value of 32.684494
is set according to DOTcvp [39], and lmax is 4. Results obtained by and 308.1328 s CPU time. Chen et al. [1] proposed a hybrid CVP
UR-CVP and SNR-CVP with tolerance J = 105 , e = 0.5, i = 0.06 to solve this problem with a value of 32.68712. It can be found
are shown in Table 4. It can be seen that SNR-CVP approach that the result obtained by the SNR-CVP method is better than the
P. Liu et al. / Biochemical Engineering Journal 111 (2016) 6374 71
Table 5 Table 6
Comparison of test results for case 2. Test results of case 3.
Park-Ramirez Ind. Eng. Chem. IDP 32.46 P J(max) tcpu (s) P J(max) tcpu (s)
bioreactor Res. [44]
Int. J. Contr. [9] FIDP 32.47 3651 1 20 6.1477 20.48 20 6.1477 25.16
Comput. Chem. second-order 32.480 316 2 24 6.1507 18.94 40 6.1508 41.47
Eng. [11] CVP 3 34 6.1512 37.51 60 6.1511 112.85
Comput. Chem. TDE 32.684494 308.1328 4 74 6.1514 27.58 80 6.1512 182.65
Eng. [45] Total tcpu (s) 104.51 362.13
#
IEEE Trans. ndCVP-HGPSO 32.68712
Automat. Sci.
Eng. [1] Table 7
Current work SNR-CVP 32.7146 89.01 Comparison of test results for case 3.
15 dx6 /dt = k1 x6
x1 [g/L]
x2 [g/L] dx7 /dt = k2 (1 x7 )
x3 [g/L]
x4 [g/L] T
x(0) = [ 1 0.1 40 0 0 1 0]
x5 [L]
10 t0 = 0
tf = 10(h)
0 u1 (t) 1
5
0 u2 (t) 1
(30)
where xi (t), (i = 1, 2, ..., 7) are the state variables and the detailed
f f
0 parameters (x3 , x5 , x6 , x7 ), Cn , Y 1 , Rfp , Ci , k1 , k2 can be found in
0 5 10 15 [46,47] and are shown in Appendix. The ratio of the cost of inducer
TIme [h] to the value of the protein production Q = 0 is considered (Table 5).
During the test, N (0) = 20 and T (0) = [0, 1/2, 1, 3/2, ..., 10] are
Fig. 9. Optimal state proles for case 2.
set, the initial control parameters u1 (0) = 0.5 and u2 (0) = 0.5
are set according to [39], the maximum number of renement
best result reported in [1], which shows a good solution quality iterations lmax is 4. Table 6 shows the results of SNR-CVP and UR-
(Table 5). CVP methods, where tolerance J = 10-5 , e = 0.2, i = 0.01 are
Fig. 7 shows the time grid renement process of SNR-CVP adopted. It can be seen that the proposed renement method is ef-
approach, which reveals the adjustment of time grid during the cient to solve the multivariable dynamic optimization problem. By
optimization. Fig. 8 illustrates the control prole of SNR-CVP and using the non-uniform renement strategy, fewer parameters are
the corresponding state proles are shown in Fig. 9. increased in successive iteration when compared with the uniform
renement CVP method. Thus, the computation time is efciently
7.3. Case 3: multivariable optimal control of Lee-Ramirez decreased. Eva et al. [11] proposed a novel renement method using
bioreactor restricted second-order information to solve this problem with a
result of 6.15. Sarkar et al. [48] presented a non-dominated sorting
Case 3 considers the multivariable dynamic optimization of a genetic algorithm to solve this problem and obtained a performance
fed-batch reactor for induced foreign protein production by recom- index of 6.1504. Logist et al. [47] obtained a same result as Sarkars
binant bacteria, as presented by Lee and Ramirez [46]. The objective by employing the ACADO Toolkit. The results of literatures and SNR-
is to maximize the protability of the process using the nutrient and CVP method are reported in Table 7, it can be seen that the proposed
the inducer feeding rates (u1 (t) and u2 (t)) as the control vector. The method is able to obtain similar or even better performance index.
72 P. Liu et al. / Biochemical Engineering Journal 111 (2016) 6374
Slopes of u1
Slopes of u1
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 5 10 0 5 10
0.8 0.8
left slope Right slope
1
Slopes of u2
Slopes of u2
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 5 10 0 5 10
2
0
Initial grid Refined grid
0 2
0 2 4 6 8 10 0 2 4 6 8 10
Iteration=1 Iteration=2
Slopes of u1
Slopes of u1
Slopes of u1
Slopes of u1
1 2
2 2
0.5 1
0 0
0 0
0.5 1 2 2
0 5 10 0 5 10 0 5 10 0 5 10
2 4 6 6
Slopes
left slope of u1 Slopes of u1
Right slope Slopes
left slope of u1 Slopes of u1
Right slope
Slopes of u2
Slopes of u2
Slopes of u2
Slopes of u2
4 4
1 2
2 2
0 0
0 0
1 2 2 2
0 5 10 0 5 10 0 5 10 0 5 10
2 2
Slopes of u1 Slopes of u1 Slopes of u1 Slopes of u1
0 0
Refined grid Refined grid
2 2
0 2 4 6 8 10 0 2 4 6 8 10
Iteration=3 Itertaion=4
(a) SNRCVP
[7] L. Xinggao, C. Long, H. Yunqing, Solution of chemical dynamic optimization
1 using the simultaneous strategies, Chin. J. Chem. Eng. 21 (2013) 5563.
URCVP
[8] R. Luus, On the application of iterative dynamic programming to singular
optimal control problems, IEEE Trans. Autom. Control 37 (1992)
0.8 18021806.
[9] A. Tholudur, W.F. Ramirez, Obtaining smoother singular arc policies using a
Nutrient feed rate
126.
[16] F. Logist, P. Van Erdeghem, J. Van Impe, Efcient deterministic multiple
objective optimal control of (bio) chemical processes, Chem. Eng. Sci. 64
(2009) 25272538.
0.4 [17] F. Logist, F. Assassa, J. Van Impe, W. Marquardt, Exploiting grid adaptation and
structure detection in multi-objective dynamic optimization problems,
Proceedings of the 22nd ESCAPE-European Symposium on Computer Aided
Process Engineering (2012) 782786.
[18] B.A. Conway, A survey of methods available for the numerical optimization of
0 continuous dynamic systems, J. Optim. Theory Appl. 152 (2012) 271306.
[19] H. Hadiyanto, D. Esveld, R. Boom, G. Van Straten, A. Van Boxtel, Control vector
0 2 4 6 8 10 parameterization with sensitivity based renement applied to baking
Time [h] optimization, Food Bioprod. Process. 86 (2008) 130141.
[20] M. Schlegel, K. Stockmann, T. Binder, W. Marquardt, Dynamic optimization
Fig. 11. Optimal control proles for Lee-Ramirez bioreactor. using adaptive control vector parameterization, Comput. Chem. Eng. 29
(2005) 17311751.
[21] F. Assassa, W. Marquardt, Dynamic optimization using adaptive direct
multiple shooting, Comput. Chem. Eng. 60 (2014) 242259.
Table A1 System parameters of case 3. [22] T. Hirmajer, E. Balsa-Canto, J. Banga, Dotcvp: Dynamic optimization toolbox
with control vector parameterization approach for handling continuous and
Parameter Value Parameter Value mixed-integer dynamic optimization problems, Tech. rep., Instituto de
Investigaciones Marinas (2008).
max 1.0 h 1
fmax 0.233 h1 [23] X. Liu, Y. Hu, J. Feng, K. Liu, A novel penalty approach for nonlinear dynamic
CN 14.35 g/L fIO 0.0005 g/L optimization problems with inequality path constraints, IEEE Trans. Autom.
s 111.5 g2 /L2 I 0.022 g/L Control 59 (2014) 28632867.
cr 0.22 g/L Ci
f
4.0 g/L [24] K.L. Teo, L.S. Jennings, H.W.J. Lee, V. Rehbock, The control parameterization
f enhancing transform for constrained optimal control problems, J. Aust. Math.
Cn 100 g/L ll 0.09 h1
Soc. Ser. B 40 (1999) 314335.
Y 0.51 klx 0.034 g/L
[25] R. Li, K. Teo, K. Wong, G. Duan, Control parameterization enhancing transform
for optimal control of switched systems, Math. Comput. Model. 43 (2006)
13931403.
Appendix B. Supplementary data [26] R.C. Loxton, K.L. Teo, V. Rehbock, Optimal control problems with multiple
characteristic time points in the objective and constraints, Automatica 44
Supplementary data associated with this article can be found, in (2008) 29232929.
[27] M.S.G. Garca, E. Balsa Canto, A.A. Alonso, J.R. Banga, Computing optimal
the online version, at http://dx.doi.org/10.1016/j.bej.2016.03.006. operating policies for the food industry, J. Food Eng. 74 (2006) 1323.
[28] P. Bittner, M. Bruhs, F. Richter, An automatic mesh renement method for
aircraft trajectory optimization problems, in: AIAA Guid. Nav. Control. Confer.,
References Boston, MA, 2013, pp. 110.
[29] V.S. Vassiliadis, R.W.H. Sargent, C.C. Pantelides, Solution of a class of
[1] X. Chen, W. Du, H. Tianeld, R. Qi, W. He, F. Qian, Dynamic optimization of multistage dynamic optimization problems. 1. Problems without path
industrial processes with nonuniform discretization-based control vector constraints, Ind. Eng. Chem. Res. 33 (1994) 21112122.
parameterization, IEEE Trans. Autom. Sci. Eng. 11 (2014) 12891299. [30] Y. Zhou, X. Liu, Control Parameterizationbased adaptive particle swarm
[2] X. Chen, W. Du, F. Qian, Multi-objective differential evolution with approach for solving chemical dynamic optimization problems, Chem. Eng.
ranking-based mutation operator and its application in chemical process Technol. 37 (2014) 692702.
optimization, Chemom. Intell. Lab. Syst. 136 (2014) 8596. [31] Q. Lin, R. Loxton, K.L. Teo, The control parameterization method for nonlinear
[3] R. Baronas, J. Kulys, A. Z ilinskas, A. Lancinskas, D. Baronas, Optimization of the optimal control: a survey, J. Ind. Manage. Optim. 10 (2014) 275309.
multianalyte determination with biased biosensor response, Chemom. Intell. [32] R. Loxton, K.L. Teo, V. Rehbock, Robust suboptimal control of nonlinear
Lab. Syst. 126 (2013) 108116. systems, Appl. Math. Comput. 217 (2011) 65666576.
[4] Y. Li, Y. Liu, D. Song, Y. Zhou, L. Wang, S. Xu, Y. Cui, Particle swarm [33] C.J. Teo, A Unied Computational Approach to Optimal Control Problems,
optimization-based protocol for partial least-squares discriminant analysis: Longman Scientic & Technical, Essex, 1991.
application to 1H nuclear magnetic resonance analysis of lung cancer [34] R. Martin, K.L. Teo, M. DIncalci, Optimal Control of Drug Administration in
metabonomics, Chemom. Intell. Lab. Syst. 135 (2014) 192200. Cancer Chemotherapy, World Scientic, Singapore, 1994.
[5] P. Zhang, H. Chen, X. Liu, Z. Zhang, An iterative multi-objective particle swarm [35] G. Li, X. Liu, Optimal control problems with incomplete and different integral
optimization-based control vector parameterization for state constrained time domains in the objective and constraints, Nonlinear Dyn. 78 (2014)
chemical and biochemical engineering problems, Biochem. Eng. J. 103 (2015) 28832895.
138151. [36] D. Gunn, W. Thomas, Mass transport and chemical reaction in multifunctional
[6] Y. Zhang, L. Tang, H. Zou, Q. Yang, X. Yu, J. Jiang, H. Wu, R. Yu, Identifying catalyst systems, Comput. Chem. Eng. 20 (1965) 89100.
protein arginine methylation sites using global features of protein sequence [37] P. Liu, G. Li, X. Liu, Fast engineering optimization: a novel highly effective
coupled with support vector machine optimized by particle swarm control parameterization approach for industrial dynamic processes, ISA
optimization algorithm, Chemom. Intell. Lab. Syst. 146 (2015) 102107. Trans. 58 (2015) 248254.
74 P. Liu et al. / Biochemical Engineering Journal 111 (2016) 6374
[38] F. Logist, M. Vallerio, B. Houska, M. Diehl, J. Van Impe, Multi-objective optimal [43] S. Park, W. Fred Ramirez, Optimal production of secreted protein in
control of chemical processes using ACADO toolkit, Comput. Chem. Eng. 37 fedbatch reactors, AIChE J. 34 (1988) 15501558.
(2012) 191199. [44] R. Luus, O. Rosen, Application of dynamic programming to nal state
[39] T. Hirmajer, E. Balsa-Canto, J.R. Banga, DOTcvp: Dynamic Optimization constrained optimal control problems, Ind. Eng. Chem. Res. 30 (1991)
Toolbox with Control Vector Parameterization Approach for Handling 15251530.
Continuous and Mixed-integer Dynamic Optimization Problems, Instituto De [45] R. Angira, A. Santosh, Optimization of dynamic systems: a trigonometric
Investigaciones Marinas, IIM-CSIC, Vigo, Spain, 2010. differential evolution approach, Comput. Chem. Eng. 31 (2007) 10551063.
[40] G. Li, X. Liu, Comments on optimal use of mixed catalysts for two successive [46] J. Lee, W.F. Ramirez, Optimal fedbatch control of induced foreign protein
chemical reactions, J. Optim. Theory Appl. 165 (2015) 678692. production by recombinant bacteria, AIChE J. 40 (1994) 899907.
[41] J. Banga, R. Irizarry-Rivera, W. Seider, Stochastic optimization for optimal and [47] F. Logist, D. Telen, B. Houska, M. Diehl, J. Van Impe, Multi-objective optimal
model-predictive control, Comput. Chem. Eng. 22 (1998) 603612. control of dynamic bioprocesses using ACADO toolkit, Bioprocess Biosyst.
[42] R. Irizarry, A generalized framework for solving dynamic optimization Eng. 36 (2013) 151164.
problems using the articial chemical process paradigm: applications to [48] D. Sarkar, J.M. Modak, Pareto-optimal solutions for multi-objective
particulate processes and discrete dynamic systems, Comput. Chem. Eng. 60 optimization of fed-batch bioreactors using nondominated sorting genetic
(2005) 56635681. algorithm, Comput. Chem. Eng. 60 (2005) 481492.