Journal of Process Control 33 (2015) 127139

Contents lists available at ScienceDirect

Journal of Process Control

journal homepage: www.elsevier.com/locate/jprocont

Advanced control with parameter estimation of batch

transesterication reactor
Richard Kern 1 , Yogendra Shastri
Department of Chemical Engineering, Indian Institute of Technology Bombay, Powai, Mumbai 400 076, India

a r t i c l e

i n f o

Article history:
Received 7 August 2014
Received in revised form 25 March 2015
Accepted 4 June 2015
Available online 9 July 2015
Keywords:
Transesterication
Optimal control
Nonlinear model predictive control
Parameter estimation

a b s t r a c t
The objective of this work is to enhance the economic performance of a batch transesterication reactor
producing biodiesel by implementing advanced, model based control strategies. To achieve this goal,
a dynamic model of the batch reactor system is rst developed by considering reaction kinetics, mass
balances and heat balances. The possible plant-model mismatch due to inaccurate or uncertain model
parameter values can adversely affect model based control strategies. Therefore, an evolutionary algorithm to estimate the uncertain parameters is proposed. It is shown that the system is not observable with
the available measurements, and hence a closed loop model predictive control cannot be implemented
on a real system. Therefore, the productivity of the reactor is increased by rst solving an open-loop optimal control problem. The objective function for this purpose optimizes the concentration of biodiesel,
the batch time and the heating and cooling rates to the reactor. Subsequently, a closed-loop nonlinear
model predictive control strategy is presented in order to take disturbances and model uncertainties
into account. The controller, designed with a reduced model, tracks an ofine determined set-point reactor temperature trajectory by manipulating the heating and cooling mass ows to the reactor. Several
operational scenarios are simulated and the results are discussed in view of a real application. With the
proposed optimization and control strategy and no parameter mismatch, a revenue of 2.76 $ min1 can be
achieved from the batch reactor. Even with a minor parameter mismatch, the revenue is still 2.01 $ min1 .
While these values are comparable to those reported in the literature, this work also accounts for the
cost of energy. Moreover, this approach results in a control strategy that can be implemented on a real
system with limited online measurements.
2015 Elsevier Ltd. All rights reserved.

1. Introduction
Growing economies and the increasing need for mobility of people and goods result in a rising demand for energy resources. One
of these energy resources in the transport sector is diesel fuel.
Because petrodiesel is a limited commodity, biodiesel can be used
as a renewable substitute [1]. Biodiesel can be produced from vegetable oils, such as soy bean oil or palm oil, waste cooking oil, animal
fat, as well as algal oil. A drawback is the lack of cost competitiveness against conventional fuels [2]. Therefore, the existing process
steps in the production of biodiesel have to be optimized to make
biodiesel more cost competitive [3]. The transesterication of the
triglyceride from oils using short chain alcohol in the presence of a

Corresponding author. Tel.: +91 22 25767203; fax: +91 22 25726895.

E-mail addresses: richard.kern@tum.de (R. Kern), yshastri@iitb.ac.in (Y. Shastri).
1
Present address: Institute of Automatic Control, Technische Universitt
Mnchen, Faculty of Mechanical Engineering, Building 2/Ground Floor, Boltzmannstr. 15, Munich 85748, Germany.
http://dx.doi.org/10.1016/j.jprocont.2015.06.006
0959-1524/ 2015 Elsevier Ltd. All rights reserved.

catalyst is an essential part of the process. This reaction can be carried out in a continuous or batch reactor [4]. The batch process is
often preferred because it is exible and facilitates the operator to
accommodate variations in feedstock type, composition as well as
quantity, and to satisfy specic product requirements [5]. The initial
capital and infrastructure investments are low [6]. Major drawbacks of the batch reactor are the intensive energy requirements
[6] and the complexity of the system due to the highly nonlinear
and time varying character of the batch process [7]. This behavior
makes an advanced optimization and control strategy a necessity.
Literature shows burgeoning interest in advanced model based
control of a transesterication reactor [5,816]. An approach to
determine an optimal temperature trajectory for a batch reactor was presented in [5]. They formulated and solved an optimal
control problem (OCP) to nd a control law such that a certain
optimality criterion was achieved. The study was extended by
[13], and the productivity of the reactor under feedstock composition uncertainty was maximized. However, constraints were
not directly taken into account. Furthermore, the OCP was solved

128

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

Fig. 1. Overview of the structure of this work.

before the operation of the reactor, and therefore the approach was
an open-loop control strategy, where the output of the controlled
process was not considered. Consequently, deviations between the
desired and actual states occurring during the process cannot be
corrected and disturbances cannot be compensated. A method to
overcome these drawbacks of open-loop control was presented
by [14]. In this work, the economic performance of a semi-batch
reactor for the production of biodiesel was optimized by using a
nonlinear model predictive control (NMPC) strategy. In each time
step, all states were estimated from measurements, and the optimal
methanol feed ow rate and the optimal heat duty were computed.
However, the heat exchange between the reactor and the jacket
was not further investigated in their work. Moreover, the observability of the system was not examined. It is doubtful that there
is enough online information available on real plants to estimate
all states, which is necessary for the NMPC strategy. The literature
review, therefore, shows that although there have been some studies exploring advanced control of the transesterication reaction,
they are limited in terms of their applicability to an actual reactor.
The objective of this paper is to address this research gap.
This work provides a comprehensive and realizable strategy
to optimize and control a batch transesterication reactor. This
is achieved by rst using optimal control theory to develop an
open-loop control strategy using reactor temperature as the control variable. An objective function representing the productivity
of the batch reactor is used for this purpose. Subsequently, this
optimal reactor temperature trajectory is tracked by using nonlinear model predictive control (NMPC). The NMPC problem is solved
using a reduced model of the batch reactor that considers only the
heat balance. The advantage of such an approach is that online
measurements for all the state variables of the reduced model are
available and hence the model is completely observable. This property enables the proposed approach to be implemented on a real
batch reactor. Since model parameter values may change over time
or may not be known accurately, a batch to batch parameter estimation using evolutionary algorithm is proposed, which uses ofine
measurements available at the end of the batch. An overview of the
structure of this work is given in Fig. 1.
This paper is organized as follows: The next section describes
the theoretical foundation of this work. This includes the development of the batch reactor model, discussion on the measurements
and observability, description of the evolutionary algorithm for
parameter estimation, and the formulation of the advanced control problems. Section 3 describes the simulation results of this
work, including those for batch to batch parameter estimation as
well as the optimal control and NMPC application. The approach is

evaluated with respect to its applicability to a real batch reactor.

Finally, Section 4 summarizes the main aspects of this contribution
and gives an outlook for further research.
2. Materials and methods
In order to optimize and control the batch transesterication
reactor, a realistic model of the process that describes the relationships among its inputs and outputs is needed. With the help of
the model, the system is analyzed and different optimization and
control strategies are evaluated.
2.1. Transesterication reaction model
Commonly, biodiesel (E) is made by the transesterication reaction between lipids and short chain alcohol such as methanol or
ethanol to produce fatty acid esters. The lipids can be derived from
animal fats or plants, such as soy, palm and rapeseed. Even the oily
part of algae can be extracted and used for biodiesel production. The
reaction is driven by a base or acid catalyst. Typically, the reaction is
carried out with vegetable oil, methanol and an homogenous alkaline catalyst [17]. Thereby, free fatty acid methyl esters (FAME) are
formed from triglyceride (TG) and methanol (A). Three consecutive
and reversible reactions occur and the intermediates diglyceride
(DG) and monoglyceride (MG) as well as the byproduct glycerol
(GL) emerge. These equilibrium reactions can be described by the
following equations:
k1

TG + CH3 OHDG + R1 COOCH3

k2

k3

DG + CH3 OHMG + R2 COOCH3

k4

k5

MG + CH3 OHGL + R3 COOCH3

k6

(1)

(2)

(3)

where R1 , R2 and R3 are long-chain hydrocarbons, also called fatty

acid chains [18]. The reaction rate constant ki is dependent on the
temperature and can be expressed by the Arrhenius equation:

ki (TR ) = k0,i exp

Ea,i

RTR

for i = 1, . . ., 6,

(4)

where k0,i stands for the respective pre-exponential factor, Ea,i

stands for the respective activation energy, R represents the universal gas constant and TR represents the reactor temperature. The
state equations of the concentrations can be obtained by a mass

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

Table 1
Values of k0,i in [m3 kmol1 s1 ] and bi in [K] used in the Arrhenius equation.
k0,i [m3 kmol1 s1 ]
3.92 10
5.77 105
5.88 1012
0.98 1010
5.35 103
2.15 104

k0,1
k0,2
k0,3
k0,4
k0,5
k0,6

Table 2
Values of parameters used in the heat balances.

bi [K]
7

b1
b2
b3
b4
b5
b6

129

6614.83
4997.98
9993.96
7366.64
3231.18
4824.87

V [m3 ]
R [kg m3 ]
MR [kg kmol1 ]
c m,R [kJ kmol1 K1 ]

x 8 =

dTJ
1
=
mJ
dt

Hr [kJ kmol1 ]

AU [kJ min1 K1 ]
mJ [kg]
c W [kJ kg1 K1 ]

1
860
391.40
1277

h Th + m
c Tc (m
h+m
c )x8
m

18 500
450
99.69
4.21

AU
(x8 x7 ) ,
c W
(13)

balance for each chemical species present in the reaction scheme.

Following Noureddini and Zhu [18] and Richard et al. [19], it is
assumed that the reactions are rst-order. Secondary reactions, like
the saponication of esters in the presence of water, are assumed
to be insignicant. Thus, the rate of change of concentrations can be
dened by the following system of ordinary differential equations:
x 1 =

dcTG
= k1 x1 x5 + k2 x2 x4
dt

(5)

x 2 =

dcDG
= k1 x1 x5 k2 x2 x4 k3 x2 x5 + k4 x3 x4
dt

(6)

x 3 =

dcMG
= k3 x2 x5 k4 x3 x4 k5 x3 x5 + k6 x6 x4
dt

(7)

x 4 =

dcE
= k1 x1 x5 k2 x2 x4 + k3 x2 x5 k4 x3 x4 + k5 x3 x5 k6 x6 x4
dt
(8)

x 5 =

dcA
= x 4
dt

x 6 =

dcGL
= k5 x3 x5 k6 x6 x4 ,
dt

(9)
(10)

where x16 represent the respective concentration of each chemical

species. The reaction rate constant is dened as:

 b 
i

ki = k0,i exp

TR

for i = 1, . . ., 6,

(11)

where bi stands for bi = Ea,i R1 . In this work, soybean oil and

methanol are chosen as reactants and sodium hydroxide as catalyst.
The corresponding values can be found in Benavides and Diwekar
[5] and are shown in Table 1.
2.2. Batch reactor model
A scheme of the batch reactor setup considered in this work
can be seen in Fig. 2. With given initial concentrations, the desired
output of a batch reactor for transesterication is mainly inuenced by the reactor temperature and the reactor pressure [5].
Because the batch reactor in Fig. 2 is operated under atmospheric
pressure, this work focuses on the control of the reactor temperature TR , which can be achieved by manipulating the following two
h with a temperature of
inputs. One input is the heating mass ow m
Th = 363.15 K (this value represents the upper temperature level of
c
district heating [20]) and the other input is the cooling mass ow m
with an ambient temperature of Tc = 293.15 K. The mass ows are
adjusted by valves which can be opened and closed continuously
h,c [0, 120] kg min1 . The valves are controlled by
in a range of m
PID controllers with a given set-point from a supervising controller.
To determine the reactor temperature TR and the jacket temperature TJ , a heat balance for each component is needed. This yields
the following state equations:
x 7 =

dTR
MR

=
(V rH
r + AU(x8 x7 ))
VR c m,R
dt

(12)

where x7,8 stand for the respective temperatures, MR for the molar
mass of the reactor content, V for the reactor volume, R for the
density of the reactor content, c m,R for the molar heat capacity of
the reactor content, r for the rate of reaction, Hr for the heat of
reaction, AU for the product of the heat exchange surface A and
the thermal transmittance U, mJ for the mass of water inside the
reactor jacket, and c W for the specic heat capacity of water. The
rate of reaction is modeled as:
r =

dcE
dx4
=
.
dt
dt

(14)

As stated by Fjerbaek et al. [4], only the heat of reaction of the

combined reaction is considered, and the heats of reaction of the
three individual reactions are ignored. The contents in the reactor
are assumed to be homogeneous and uniformly mixed. Moreover,
the molar mass, density and heat capacity of the reactor are each
approximated by an average constant, although the values would
change in reality as the reaction progresses. Furthermore, Wilson et al. [21] showed that the heat loss from the jacket to the
surrounding and the stirrer power are in the same order of magnitude and both of them are small. Therefore, they are not taken
into account. The parameter values, corresponding to Eq. (12), are
reported in Table 2 [4,2224]. It is advisable to model the heat
transfer between the jacket and the reactor for an optimal control
approach, especially when waste heat or district heating is used,
because the delays of the inputs and small temperature differences,
which result in small heat ows, are taken into account. Furthermore, the bounds for the heating and cooling mass ows in a real
plant can be easily incorporated.
2.3. Measurements
It can be seen in Fig. 2 that the available online measurements
are the reactor temperature TR (t) and the jacket temperature TJ (t).
Furthermore, the conversion rate of the reaction can be measured
online [25,26] and expressed through the rate of change of the
alcohol concentration cA (t) inside the reactor. Therefore, the stateto-output matrix C Rny nx is dened as:

0 0

C = 0 0 0
0 0 0

1 0

0 .
1

(15)

The online measured output y(t) Rny can be expressed as:

y(t) = Cx(t).

(16)

Besides the online measurements, the nal concentration of each

chemical species can be determined after each batch, e.g. by the use
of a chromatography system. These settings represent the conditions for real plants [21].
2.4. Observability
The observability of a nonlinear system can be investigated
similar to the procedure of the extended Kalman lter. This lter

130

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

TR,sp

NMPC

uc
M

OCP

uh
FC

FC

M
.
mh

.
mc

Motor

TR

TJ

Reactor
TR

.
Q

cA
Stirrer

Jacket TJ
Fig. 2. Scheme of a batch reactor with controlled heating and cooling system; solid arrows indicate material ow while dashed arrows indicate information ow.

linearizes the system via a rst-order Taylor series approximation

around an operating point at each time step into the following form:

x(t)
= Ax(t) + Bu(t)

(17)

y(t) = Cx(t),

16

k0,16

712

b16

13

R

14

AU

15

c m,R

16

MR

17

Hr

(18)

Rnx nx

Rnx nu

where A
stands for the state matrix and B
for the
input-to-state matrix. It is assumed that the system is in a steady
state at each time step. The resulting system (A, C) is observable if
the following condition, called Hautus criterion, holds true for all
eigenvalues:

rank (Hi ) = rank

Table 3
Estimated parameters.

i I A
C

= nx

i = 1, . . ., nx ,

(19)

where Hi R(nx +ny )nx is the Hautus matrix of a specic eigenvalue.

2.5. Parameter estimation
For a rened optimization and control strategy, it is important
to have an accurate reactor model. However, some model parameters may be difcult to know accurately. Moreover, some model
parameters may change over time due to fouling, changes in load
and the wear of components. Hence, it is necessary to revise the
model periodically to account for such mismatches, especially in
case of open-loop control if no online measurements or estimates
of the state and parameters are available. For the batch reactor we
have modeled, this can be performed after each batch run in a real
plant. We propose that the online and off-line measurements from a
particular batch can be compared with the model predictions after
the batch is completed, and the deviations in the measured and
predicted values be used to perform parameter estimation. This
estimation should preferably be completed before the next batch
starts. The more precise model thus obtained will lead to smaller
deviations between the modeled and predicted results, and this
deviation will go down with each estimation.
For the batch reactor model described here, 17 parameters that
could be uncertain and may change from batch to batch were

identied. These 17 parameters and the corresponding symbols for

their estimates are listed in Table 3. It is assumed that there would
be some error in the initial model parameter values and the goal is
to provide better estimates of these parameters based on the results
for each batch.
The alcohol concentration cA = x5 , the temperature of the reactor
TR = x7 , and the temperature of the jacket TJ = x8 can be measured
online at each time step. After each run, the remaining concentrations can be determined ofine at the nal time tf . Better model
parameter estimates can be provided by minimizing the following
scaled least squares objective function which describes the deviation between the simulated states x,

T obtained via the estimated
= 
17 , and the measured states x,
1 
parameter vector 
obtained via the true parameters:
n

Min JPE

x 5,i
1
=
1
x5,i
n
i=1


+ 1


+

x 8,j

2

1
+
m

2  
+

x8,j
x 6,tf
x6,tf


1

j=1

x 1,tf
x1,tf

x 7,j

2

x7,j

2


+ +

x 4,tf

2

x4,tf

2
,

(20)

where the optimization variables are the estimated parameters

117 , n stands for the number of online alcohol concentration

measurements and m for the number of online temperature measurements. To minimize JPE , an evolutionary strategy, as described
by Rechenberg [27], is applied. This is a heuristic optimization
technique based on the principles of adaption and evolution. The

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

of the decision
operators are mutation and selection. The vector 
variables, model parameters in this case, is called an individual. The rst generation of individuals consists of vectors with
randomly distributed numbers as entries between a dened lower
and upper bound for each parameter. The tness of each individual
is evaluated by computing the value of Eq. (20). Subsequently, the
individuals within each generation are ranked, corresponding to
1 , the sectheir tness value. The ttest individual is denoted as 
2 , and so on. To generate the children, the highest
ond ttest as 
ranked individuals are selected and mutated by adding a standard
normal distributed random value to each vector component. The
added absolute value decreases from generation to generation to
improve the convergence. The size of the population p is dened
1 :
by the number of children k1 of the ttest individual 
p = k12 .

(21)
i

is dened as:
The number of children ki N+ of an individual 
ki = k1 i + 1.

(22)

The total number of children K = k1 + k2 + + kk1 is determined by a triangular number:

K=

k1 (k1 + 1)
.
2

(23)

For example, if the ttest individual has four children, the second
ttest individual has three children, and so on, the total number
of children is K = 10 and the population size is p = 16. Furthermore,
the ttest individual is passed to the next generation without mutation to avoid the possibility of worsening the solution, e.g. in case
that the optimum is found. The remaining ve empty slots of the
population are lled with new random individuals to refresh the
gene pool, which is especially important for non-convex optimization problems. This method can be seen in Fig. 3. The number of
children and the number of generations depend on the available
computational time between two runs and have to be adjusted.
The evolutionary algorithm has several advantages in comparison with derivative-based optimization algorithms. For example,
the evolutionary algorithm is suitable for solving non-convex optimization problems and can provide the global optimum, whereas
the solution of derivative-based algorithms is strongly dependent
on the supplied initial value. Furthermore, due to the mutation
operator, the global optimum can be found even if it is outside the
initially specied bounds.
2.6. Optimal control problem formulation
In order to enhance the cost competitiveness of biodiesel, it
is important to increase the productivity of the batch reactor for
transesterication. In this work, the productivity is dened as:
productivity =
revenue from biodiesel cost of feedstock cost of energy
,
time per cycle
(24)
where the revenue is generated from selling the biodiesel, the cost
of feedstock includes the cost of vegetable oil or animal fat and
alcohol, the cost of energy equals the sum of the cost of heating
and cooling, and the time per cycle is the sum of the batch time
and an additional setup time for each batch. The cost for catalyst is
ignored. The capital cost for building the plant or other expenses for
the operation of the plant except the cost of energy are not included
in Eq. (24). Furthermore, the cost for post-processing is assumed to
be constant and independent of the nal concentration of FAME
due to its small fraction of the total expenses [28]. Moreover, the

131

possibility of methanol recovery, as described by Dhar [29], is not

taken into account. The possible revenue generated from selling
glycerol is also not considered. Therefore, the productivity in Eq.
(24) should be seen as a ratio expressing the economic efciency of
the batch reactor for transesterication and not the complete venture. The neglected costs have to be taken into account to determine
the protability of the plant as a whole. That will make the optimization problem signicantly more complicated and hence has
been ignored in this work. These factors could be indirectly considered by adjusting the coefcient pE in Eq. (31), where increasing the
value of pE will lead to a higher concentration of ester in the product. This adjustment could be based on experiences with the plant
and the cost for post-processing. The productivity, as dened in Eq.
(24), can also be used to evaluate the cost margin of the operating
expenses.
Control problems discussed in literature often do not use the
objective function of productivity as dened above. Some studies
consider only the reaction model and ignore the cost of energy since
the reactor jacket and the associated heating and cooling are not
part of the model [5,13]. Some studies instead ignore the batch time
as an optimization objective. Such objective functions are expected
to be inferior as compared to the objective function dened in Eq.
(24). To illustrate this, we perform simulation studies using two
additional objective functions. The rst objective function does not
consider the batch time and cost of energy and is given as:
productivity = revenue from biodiesel cost of feedstock.

(25)

The second objective function considers the batch time but ignores
the cost of energy and is given as:
productivity =

revenue from biodiesel cost of feedstock

,
time per cycle

(26)

To solve the optimal control problems, presented in the next

section, the ACADO Toolkit [30] is used. ACADO uses a multiple shooting method to discretize the optimal control problem.
Subsequently, the dynamics of the system are converted into constraints to match the standard nonlinear programming problem
(NLP) formulation. Finally, the NLP problem is solved with a successive quadratic programming (SQP) technique. The resulting reactor
temperature trajectory TR (t) for t [t0 , tf ] from this ofine openloop optimization is used as a set-point for an NMPC strategy.
2.7. Set-point tracking using NMPC
In general, two different control problems arise for batch reactors. The rst one is end-point property control and the other one
is set-point tracking of an ofine determined state trajectory. If the
system is not observable with the given measurements, the concentrations of the chemical species are not known during the batch
and the optimization of Eq. (24) can be done only ofine. Under
these conditions, end-point property control is not possible. Due
to the measurements of the temperatures, the heat balances of the
batch reactor can be used to track a set-point temperature trajectory via an NMPC strategy. For this approach, described by Nagy
et al. [31], the system of the batch reactor for transesterication is
reduced to the two heat balances
x 7 =

dTR
MR

=
(V rH
r + AU(x8 x7 ))
VR c m,R
dt

x 8 =

dTJ
1
=
mJ
dt

c Tc (m
h+m
c )x8
h Th + m
m

(27)

AU
(x8 x7 ) .
c W
(28)

The rate of reaction r is approximated as a stepwise constant and

computed by the negative nite difference of the measured alcohol

132

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

children from
children from

*1

individual

*2

children from

*3

children from

*4

*1

random individuals

Fig. 3. Individuals of a population of the proposed evolutionary algorithm in a particular generation for k1 = 4.

concentration cA between the time step k 1 and k as well as the

sample time ts :
r =

cA,k cA,k1
ts

(29)

This approach is facilitated by the approximation that the heat

of reaction only accounts for the formation of ester and by the
fact that for each ester molecule that is produced one alcohol
molecule is consumed. If the alcohol concentration cannot be measured, the heat of reaction has to be approximated or cannot be
taken into account. Because the heat of reaction is generally small
[4], the NMPC strategy will still yield a good performance in this
case. Before each run, the OCP which maximizes the productivity,
dened in Eq. (24), is solved. During the run, the optimal reactor temperature trajectory proposed by the solution of the OCP
is tracked by the NMPC controller. The controller uses the measurements of the temperatures and the alcohol concentration to
h,c to be able to track the temperacompute the inputs uh,c = m
ture trajectory. The objective function used in the NMPC strategy
is dened as:

2

Min JNMPC = q1 x7 (kt s + T ) x7,sp (kt s + T )

kt s +T

+
kt s

h (t) m
h (kt s ))
[q2 (m
2

(30)

where q14 denote the weighting factors, T the time horizon and x7,sp the reactor set-point temperature. The rst term

2

x7 (kt s + T ) x7,sp (kt s + T ) prevents deviations between the true

and the desired reactor temperature at the end of the time hori-

3. Results and discussion

To evaluate the performance, robustness and reliability of
the proposed optimization and control strategy, the approach is
examined via simulations in MATLAB and Simulink, linked to
algorithms in C++. The simulations are carried out on a machine
with a 2.4 GHz Intel Core 2 Duo CPU and 8 GB RAM. The settings
affecting the computational speed of the simulation are chosen
according to this performance.
3.1. Observability
It can be shown that the system is not observable if only the
reactor temperature TR , the jacket temperature TJ and the alcohol
concentration cA is measured. This is because the Hautus matrix
Hi R(nx +ny )nx , dened in Eq. (19), is singular for some of the
eigenvalues. For observability, at least two more measurements
are needed, which is not practical for a real reactor. Similar results
regarding the observability of industrial batch reactors have been
reported by [21].
3.2. Parameter estimation

c (kt s )) + q4 (m
c (t)) ]dt
c (t) m
h (t)m
+ q3 (m

minimization of Eq. (30) is done by ACADO with the same methods

as described above.

2

zon. The integration of the penalty term uh,c (t) uh,c (kt s ) avoids
fast change rates of the inputs. This is important due to the fact
that a heavy oscillating or bang-bang behavior of the inputs cannot
be realized and damages the valves. Finally, the integration of the
penalty term (uh (t)uc (t))2 prevents heating and cooling at the same
time. The deviation of the reactor temperature from the set-point
is minimized only at the end of the prediction and control horizon in Eq. (30). This reduces the computational load of the NMPC
optimization problem, which is important for its online implementation. Minimization of deviation over the complete time horizon
will add to the computational load. Moreover, such a formulation provides more exibility for the NMPC solution and results
in smoother dynamics of the reactor in case of model parameter
mismatch, especially if the parameter mismatch is signicant. With
this approach, bounds on the control variable can be taken into
account more easily and the solution is more likely to converge
[31]. A prediction and control horizon of 60 seconds was used
in the simulation studies, and consequently, the error introduced
by this approximation was not observed to be signicant. The

After each batch the evolutionary strategy is applied to minimize

Eq. (20) and estimate the uncertain parameters. For the simulation and evaluation of the evolutionary strategy, the population is
initialized with p = 121 individuals and the number of generations
is set to 20. It is assumed that the real parameters, which have
been presented before, remain constant for all runs. To illustrate
the performance of the algorithm, the assumed mismatch exceeds
the deviations in reality. The initial estimates of the parameters, e.g.
given by the supplier of the feedstock, differ from the real values
by 20% and are given in Table 4.
The convergence of the objective function and the resulting
decrease for ve batch runs can be seen in Fig. 4. The computational
time for one optimization cycle after each run is approximately
i = 11 min, which represents the time needed to empty and rell
the reactor. Therefore, the sum of the computational time in Fig. 4
Table 4
Initial estimates of the uncertain parameters in the model.
k0,1 [m3 kmol1 s1 ]
k0,2 [m3 kmol1 s1 ]
k0,3 [m3 kmol1 s1 ]
k0,4 [m3 kmol1 s1 ]
k0,5 [m3 kmol1 s1 ]
k0,6 [m3 kmol1 s1 ]
R [kg m3 ]
c m,R [kJ kmol1 K1 ]
AU [kJ min1 K1 ]

4.70 107
6.92 105
4.70 1012
1.18 1010
6.42 103
1.72 104
1032
1532.40
360

b1 [K]
b2 [K]
b3 [K]
b4 [K]
b5 [K]
b6 [K]
MR [kg kmol1 ]
Hr [kJ kmol1 ]

5290
6000
12 000
8840
2580
5790
313.12
14 800

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

Table 5
Values of parameters required to calculate the productivity of biodiesel production
as per the optimal control problem objective function.

0.25

pE [$ l1 ]
pTG [$ l1 ]
pA [$ l1 ]
ph [$ MJ1 ]
T0 [K]
TG [kg m3 ]
E [kg m3 ]

0.2

0.15

0.89
0.57
0.19
0.00330
293.15
926
880

ME [kg kmol1 ]
MTG [kg kmol1 ]
MA [kg kmol1 ]
pc [$ l1 ]
Tin [K]
A [kg m3 ]

292.2
920
32.04
8.9554 1004
363.15
791.8

0.1

0.05

Batch Run
Fig. 4. Value of the objective function JPE after each batch.

is = 55 min. The possibility to run the parameter estimation algorithm during the run in parallel with the NMPC strategy is not
considered but advisable, especially when a machine with better
hardware is used. Moreover, the computational time is expected to
reduce with every batch since the starting guess would be closer
to the actual parameter values, and hence the convergence would
be faster. The objective function in Fig. 4 decreases by almost an
order of magnitude in only ve runs, which indicates an excellent
performance by the proposed approach.
It is expected that larger errors in initial parameter estimates
would require more batch runs for the objective function to reduce.
In such cases, either a larger population size or more generations
could improve convergence, as long as the optimization problem
can be solved within the limited time available between each batch.
The proposed approach can also be useful in handling changes in
feedstock. A change in feedstock will primarily affect the composition, such as the composition of different TGs (different chain
lengths and saturation levels), with subsequent impacts on kinetic
parameter values. The compositional details of the new feedstock
can be obtained from the supplier, literature, or by performing
experimental analysis. The new data can be used to update the
initial guesses and dene the search region of the evolutionary algorithm. This will ensure that the evolutionary algorithm converges
faster. If such data are not available, parameter values from the previous run must be used, which will delay the convergence. The rst
batch in such cases is expected to be inferior.
In summary, the advantages of an evolutionary approach and
its good performance in this case make this a highly useful strategy
for parameter estimation.

The optimal inputs uh,c (t) for the OCP and the resulting FAME
concentration and temperature proles can be seen in Figs. 5 and 6,
respectively. The nal concentration of FAME equals cE (tf ) =
0.8317 mol l1 . The maximum value of the objective function
equals JI = 138.55 $. If this value is divided by the sum of the batch
time tf = 100 min and a setup time between the cycles tset = 10 min
[5] the following value, which expresses the economic performance
of the batch reactor, is obtained:
JI

tf + tset

138.55 $
$
.
= 1.26
110 min
min

(32)

The batch time is an important parameter of the economic performance of a batch reactor [2]. However, it is neglected in JI . Moreover,
since the cost of heating and cooling is neglected, the respective
proles show uctuations as well as simultaneous use of heating
and cooling. This is highly undesirable and must be avoided.

120

90
uh(t) [kg/min]

Objective Function JPE

133

60

30

25

50

75

100

75

100

Time [min]
(a) Heating mass flow

3.3. Economic performance

120

3.3.1. Case A: Revenue maximization

The objective function which maximizes the revenue from
biodiesel is dened as:
Max

ME
MTG
MA
JI = pE x4 (tf )
V pTG x1 (t0 )
V pA x5 (t0 )
V,
E
TG
A

(31)

where pE,TG,A stand for the respective price. These prices and other
required parameter values are reported in Tab. 5 [32,33,5,34,1,22].

90
uc(t) [kg/min]

The optimization of the productivity is conducted in three steps.

Objective function JI maximizes the revenue from biodiesel only,
the nal time is xed and the cost of energy neglected. Objective
function JII maximizes the revenue from biodiesel and minimizes
the batch time but does not consider the cost of energy. Finally,
objective function JIII maximizes the productivity as described in
Eq. (24). There is no parameter mismatch assumed. Taxes and the
cost for distribution are excluded from the price of biodiesel [28].

60

30

25

50
Time [min]

(b) Cooling mass flow

Fig. 5. Optimal mass ows of heating and cooling liquids recommended by the OCP
solution using JI objective.

134

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

120

90
uh(t) [kg/min]

cE(t) [mol/L]

0.8

0.6

0.4

60

30

0.2

25

50

75

100

10
15
Time [min]
(a) Heating mass flow

20

10
15
Time [min]
(b) Cooling mass flow

20

Time [min]
(a) Concentration of FAME
120
350

uc(t) [kg/min]

T(t) [K]

90

320

60

30
TJ(t)
TR(t)
290

25

50

75

100

Time [min]
(b) Temperatures
Fig. 6. Optimal FAME concentration and reactor and jacket temperature proles for
the OCP problem solution using JI objective.

3.3.2. Case B: Revenue maximization and batch time

minimization
To overcome the issue of batch time, the following objective
function optimizes the nal yield of FAME as well as the batch time:
M

Max JII =

TG

tf + tset

3.3.3. Case C: Productivity maximization considering batch time

and cost of energy
Finally, an objective function is introduced which represents the
productivity, dened in Eq. (24):

pE x4 (tf )  E V pTG x1 (t0 )  TG V pA x5 (t0 )  A V

E

Fig. 7. Optimal mass ows of heating and cooling liquids recommended by the OCP
solution using JII objective.

(33)

Please note that the cost of heating and cooling is again ignored.
Because of the non-convexity of the problem, the initial values
for the optimization are selected by an automated trial and error
method. The optimal inputs uh,c (t) as well as the resulting trajectories can be seen in Fig. 7 and 8, respectively. The nal concentration
of FAME equals cE (tf ) = 0.7162 mol l1 and the optimal batch time
is tf = 23.30 min. The maximum value of the objective function

equals JII = 3.14 $ min1 . Comparison with results for case A show
that the nal concentration of FAME is about 14% lower for case
B. However, the nal productivity in $ min1 is still almost 150%
more for case B. This is due to the signicant reduction of batch
time by 76.5%. From an application perspective, this result recommends operating more batches of shorter time even though the
nal concentration of FAME for each batch is lower. The time saved
by operating shorter batches will compensate for the lost revenue
due to lower nal FAME concentration. Please note that this ignores
the cost of FAME separation, which will be more if the nal concentration is lower.
Fig. 7 shows that the issue of simultaneous heating and cooling
is still encountered for this case. Furthermore, the mass ows and
therefore the valves are exposed to oscillations. This is due to the
fact that the cost of heating and cooling is neglected. For a real
application, these cost have to be included.

Max JIII =

pE x4 (tf )  E V pTG x1 (t0 )  TG V pA x5 (t0 )  A V

E

ph

 tf
t0

TG

tf + tset

uh (t)cW (Tin T0 )dt + pc

tf + tset

 tf
t0

uc (t)dt

(34)

where c W = 4.205 kJ kg1 K1 denotes the specic heat capacity of

water. The objective function considers the revenue from biodiesel,
the cost of feedstock, the batch time, and the cost of energy. The
values of other parameters required to determine the cost of energy
are reported in Table 5 [35,36].
Because of the non-convexity of the problem, the initial values
for the optimization are selected by an automated trial and error
method. The solution for the optimal inputs uh,c (t) and the resulting trajectories can be seen in Figs. 9 and 10, respectively. As it can
be seen in Fig. 9, it is more economical to heat the reactor as long
as possible and cool it in the end of the batch run to prevent overheating. This is due to the low price of cooling water compared to
the price of FAME and the shorter batch time. The cooling input is
slightly oscillating because of numerical inaccuracies.
The nal concentration is cE (tf ) = 0.7277 mol l1 , the optimal
batch time is tf = 24.64 min and the maximum value of the objec = 2.76 $ min1 . It is important to note that the
tive function is JIII
absolute value of objective function JII is bigger than the absolute
. This is due to the fact that the cost for
value of objective function JIII
heating and cooling is not considered in objective function JII . If it

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

120

90
uh(t) [kg/min]

cE(t) [mol/L]

0.8

0.6

0.4

60

30

0.2

135

10

15

20

10

15

Time [min]

Time [min]

(a) Concentration of FAME

(a) Heating mass flow

20

120
350

uc(t) [kg/min]

T(t) [K]

90

320

60

30
TJ(t)
TR(t)
290

10

15

20

Time [min]

had been included, the actual value of objective function JII would
equal JII = 2.01 mol l1 . This is due to the excessive costs for heating
and cooling, as shown in Fig. 7.
Objective function JIII is superior to the objective functions discussed before because all important variables for the optimization
of a batch reactor for transesterication are regarded. Furthermore,
it is applicable to a real reactor. Therefore, objective function JIII is
used in the following NMPC approach.

3.4. Set-point tracking NMPC

Three different operational scenarios are investigated and the
respective results are presented. The results from the simulation
of the different operational scenarios are analyzed and discussed
to validate and verify the optimization and control strategy with
respect of its applicability to a real reactor.

3.4.1. Case I scenario: No parameter mismatch

In the case I scenario, it is assumed that there is no parameter
mismatch between the model and the real reactor. The sample time
is set to ts = 6 s, the length of the discrete time horizon N = T/ts to
N = 10 and the predicted as well as the controlled horizon account
for T = 60 s. These settings are chosen as a trade-off between the
computational demand and the performance of the set-point tracking. The resulting temperature proles of the set-point reactor
temperature TR,sp (t), computed ofine by the OCP, and the actual
h,c (t)
reactor temperature TR (t) as well as the inputs uh,c (t) = m
can be seen in Fig. 11. The deviation between TR,sp (t) = x7,sp (t) and

10

15

20

Time [min]

(b) Temperatures
Fig. 8. Optimal FAME concentration and reactor and jacket temperature proles for
the OCP problem solution using JII objective.

(b) Cooling mass flow

Fig. 9. Optimal mass ows of heating and cooling liquids recommended by the OCP

solution using JIII

objective.

TR (t) = x7 (t) is expressed by the root mean square (RMS) of this

difference:



N
 1 

2
RMS = 
x7 (k) x7,sp (k) .
N+1

(35)

k=0

Fig. 11(a) shows that the controller tracks the desired set-point
temperature with high precision. The deviation accounts for
RMS = 0.1551 K. It should be noted that the closed-loop response
generally slightly differs from the open-loop solution due to
the different time horizons and because in this case different
objective functions for the open-loop and closed-loop NLP problem are used. Therefore, a value of RMS = 0 is most unlikely
[37]. The nal concentration of FAME predicted by the OCP is
cE (tf ) = 0.7320 mol l1 , while that achieved by simulation of the
NMPC strategy is cE (tf ) = 0.7331 mol l1 . Fig. 12 shows that the
computational time (k) that is needed to solve the optimization
problem, described in Eq. (30), is one order of magnitude smaller
than the sample time. This is crucial for the real-time feasibility of
an NMPC strategy. The productivity accounts for JOCP = 2.76$ min1 .
It was seen that the deviation error between the desired temperature trajectory and the actual reactor temperature within the time
horizon was negligible due to the fact that the prediction and control time horizon was short.
3.4.2. Case II scenario: Minor parameter mismatch
The case II scenario simulates a slight parameter mismatch to
reproduce realistic operating conditions. This is due to the fact
that it is impossible that the evolutionary algorithm for parameter
estimation fully converges in nite time. Furthermore, there will

136

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

cE(t) [mol/L]

0.8

0.6

0.4

0.2

10

15

20

Time [min]

(a) Concentration of FAME

T(t) [K]

350

320

TJ(t)
TR(t)
290

10

15

20

Time [min]

(b) Temperatures

Fig. 10. Optimal FAME concentration and reactor and jacket temperature proles for the OCP problem solution using JIII
objective.

always be a mismatch due to the alteration from run to run and the
highly uncertain operating conditions. All other settings are similar to the case I scenario. The estimated parameters, as described in
Table 3, are assumed to differ from the real parameters as follows:
117 = 117 0.1117 ,


(36)

where the sign is arbitrary. The parameter mismatch does

not only affect the controller but also the computation of the

open-loop set-point temperature trajectory. Fig. 13 shows the

resulting temperature prole and the inputs.
It can be seen that due to the parameter mismatch the batch
time tf = 41.73 min is longer than in the case without parameter
mismatch. One reason for this is that the estimated parameter AU,
which stands for the product of the heat exchange surface A and the
thermal transmittance U, is expected to be smaller than the actual
parameter. The true value is AUtrue = 14 = 450 kJ min1 K1 and the

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

340

TR
TR,sp

330

Reactor Temperature [K]

Reactor Temperature [K]

340

320

310

300

290

10

TR
TR,sp

330

320

310

300

290

20

137

10

Time [min]

(a) Reactor temperature

uc

80
60
40
20

uc

80
60
40
20

10

20

10

Time [min]

20

30

40

Time [min]

(b) Heating and cooling mass flow

Fig. 11. Results of NMPC application assuming no parameter mismatch: Temperature proles of the set-point reactor temperature and the actual reactor
temperature, and the heating and cooling inputs.

14 = 405 kJ min1 K1 . Therefore, the

estimated value is AU est = 
open-loop optimization predicts a lower heat transfer than it is the
case in reality. Despite the parameter mismatch, the controller is
still able to track the set-point temperature trajectory satisfactorily
and the deviation accounts for RMS = 0.4019 K. The nal concentration of FAME, predicted by the OCP, is cE (tf ) = 0.7755 mol l1 .
The achieved value in the simulation of the NMPC strategy
is cE (tf ) = 0.8011 mol l1 . The computational time (k), shown in
Fig. 14, is only slightly increased compared to the case I scenario and
fullls the real-time requirements. The value of the productivity is
JOCP = 2.01$ min1 .

(b) Heating and cooling mass flow

Fig. 13. Results of NMPC application assuming minor parameter mismatch: Temperature proles of the set-point reactor temperature and the actual reactor
temperature, and the heating and cooling inputs.

The case II scenario proves that the control strategy still yields
excellent results in case of a slight parameter mismatch. However,
it also shows that the solution of the open-loop OCP, which determines the set-point temperature prole, is strongly dependent on
the estimated parameters. Therefore, a parameter mismatch in the
system can substantially inuence the operation of the batch reactor and the product quality.
3.4.3. Case III scenario: Signicant parameter mismatch
In the case III scenario, a signicant parameter mismatch is
simulated to evaluate its effect on the optimization and control

0.8

1.4

0.7

1.2
Computational Time [s]

Computational Time [s]

40

uh

100
Mass Flow [kg/min]

Mass Flow [kg/min]

120

uh

100

0.6
0.5
0.4
0.3

1
0.8
0.6
0.4
0.2

0.2
0.1

30

(a) Reactor temperature

120

20
Time [min]

0
0

100

200

Time Step
Fig. 12. Computational time (k) for the NMPC problem solution assuming no
parameter mismatch.

100

200

300

400

Time Step
Fig. 14. Computational time (k) for the NMPC problem solution assuming minor
parameter mismatch.

138

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

TR
TR,sp

330

5
Computational Time [s]

Reactor Temperature [K]

340

320

310

300

290

4
3
2
1

20

40

60

80

100

120

Time [min]

200

400

600

800

1000

1200

Time Step

(a) Reactor temperature

Fig. 16. Computational time (k) for the NMPC problem solution assuming signicant parameter mismatch.

120

uh
uc

Mass Flow [kg/min]

100

mismatch can lead to an increased computational time. The condition ts is nearly violated at some time steps and the real-time
feasibility is no longer given. Therefore, the case III scenario shows
the limits of the presented optimization and control strategy. The
reaction kinetics k0,16 and b16 should be chosen as exact as possible to successfully compute a set-point temperature trajectory.
The remaining estimated parameters may differ more from the real
parameters. However, this leads to an oscillating behavior of the
system which is undesirable, causes damage to the heating and
cooling components, and will worsen the productivity of the reactor substantially. Therefore, it is advisable to increase the weights
q2,3 on the change rate of inputs in Eq. (30) and/or use a lter in
order to smoothen the output signals of the NMPC strategy.

80
60
40
20
0

20

40

60

80

100

120

Time [min]

(b) Heating and cooling mass flow

Fig. 15. Results of NMPC application assuming signicant parameter mismatch:
temperature proles of the set-point reactor temperature and the actual reactor
temperature, and the heating and cooling inputs.

strategy. The other settings regarding the simulation are in agreement with the case I scenario. Unfortunately, the solution of the
open-loop OCP cannot converge if the parameter mismatch of
the reaction kinetics k0,16 and b16 exceeds a certain limit of
112 > 112 0.1112 . Therefore, the parameter mismatch of

the reaction kinetics is set to:
112 = 112 0.1112 .


(37)

The mismatch of the other parameters is assumed to be:

1317 = 1317 0.51317 .


(38)

For example, the estimated value of the parameter AU is

AUest = 14 = 225 kJ min1 K1 , which is 50% less than the true value
of AU. The resulting temperature prole and the inputs are shown
in Fig. 15.
The batch time is tf = 134.05 min and differs signicantly from
the case I scenario. It can be seen in Fig. 15 that the system
starts to oscillate, especially after t = 60 min, and the controller is
no longer able to track the set-point temperature smoothly. This
phenomenon is caused by the fact that the predicted states differ signicantly from the real states due to model uncertainties.
Therefore, the input signals, computed by the controller using an
inadequate model, are chosen incorrectly. The deviation between
the actual reactor temperature and the set-point reactor temperature accounts for RMS = 1.1284 K. The nal concentration of FAME,
predicted by the OCP, is cE (tf ) = 0.8585 mol l1 . The achieved value
in the simulation of the NMPC strategy is cE (tf ) = 0.8243 mol l1 . The
productivity is reduced to JOCP = 0.89$ min1 . Fig. 16 shows the computational time (k). It can be seen that the signicant parameter

4. Conclusions
This work addresses the optimization and control of a batch
reactor for transesterication to enhance the cost competitiveness
of biodiesel. The objective of this work is to develop an advanced
control strategy that can be implemented on a real batch reactor
performing transesterication. To that effect, a nonlinear dynamic
model considering the reaction kinetics, mass balances and heat
balances of the jacketed reactor is developed. It is recognized that
the system is not observable since no online information on the concentrations of the chemical species, except for alcohol, is available.
Consequently, closed loop advanced control of a full state model
is not feasible. The strategy proposed in this work is to rst solve
an optimal control problem ofine based on a full state model to
obtain a set-point reactor temperature trajectory. This trajectory is
tracked online by a nonlinear model predictive controller (NMPC),
based on a reduced model, during the operation of the batch reactor. The values of the inputs of the heating and cooling mass ows,
computed by the NMPC controller, can be easily used as reference
variables for conventional PID controllers. Since an accurate model
is desired, a reliable parameter estimation strategy is introduced to
enhance the model of the system.
The simulation results show that the maximum value of the productivity in case of no parameter mismatch between the model and
the plant with a batch time of tf = 24.64 min is JOCP = 2.76$ min1 . If
the batch time is increased and the reactor temperature kept at its
highest possible value, the maximum achievable yield of FAME is
cE (tf ) = 0.8318 mol l1 . However, this value is not advisable because
these settings reduced the productivity. The simulation of the different scenarios showed that the NMPC strategy is robust in case
of parameter mismatch but the solution of the open-loop OCP is
strongly dependent on the estimated parameters. It is important
to note that the resulting set-point trajectory may be optimal for

R. Kern, Y. Shastri / Journal of Process Control 33 (2015) 127139

the model but not for the real plant. Uncertainties in the model
will alter the solution of the open-loop OCP substantially and may
worsen the protability of the production of biodiesel.
Acknowledgements
The authors would like to thank the German Academic Exchange
Service (Grant name and number: A new passage to India,
56049289) who funded the rst authors stay at IIT Bombay, India.
The authors would also like to acknowledge the valuable inputs
provided by Prof. Sharad Bhartiya, IIT Bombay, India, and Prof. Boris
Lohmann, Technische Universitt Mnchen, Germany during the
course of this research.
References
[1] Y. Zhang, M.A. Dube, D.D. McLean, M. Kates, Biodiesel production from waste
cooking oil: 2. Economic assessment and sensitivity analysis, Bioresour. Technol. 90 (2003) 229240.
[2] D. Klein-Marcuschamer, H.W. Blanch, Survival of the ttest: an economic perspective on the production of novel biofuels, AIChE J. 59 (2013) 44544460.
[3] N. Nasir, W. Daud, S. Kamarudin, Z. Yaakob, Process system engineering in
biodiesel production: a review, Renew. Sustain. Energy Rev. 22 (2013) 631639.
[4] L. Fjerbaek, K.V. Christensen, B. Norddahl, A review of the current state of
biodiesel production using enzymatic transesterication, Biotechnol. Bioeng.
102 (2009) 12981315.
[5] P.T. Benavides, U. Diwekar, Optimal control of biodiesel production in a batch
reactor. Part I: Deterministic control, Fuel 94 (2012) 211217.
[6] B. He, J.V. Gerpen, Reactors for biodiesel production, National biodiesel education program, 2013 http://www.extension.org/pages/26630/reactors-forbiodiesel-production
[7] J. Valappil, C. Georgakis, Nonlinear model predictive control of end-use properties in batch reactors, AIChE J. 48 (2002) 20062021.
[8] F.S. Mjalli, M.A. Hussain, Approximate predictive versus self-tuning adaptive control strategies of biodiesel reactors, Ind. Eng. Chem. Res. 48 (2009)
1103411047.
[9] F.S. Mjalli, L. Kim San, K. Chai Yin, M. Azlan Hussain, Dynamics and control of a
biodiesel transesterication reactor, Chem. Eng. Technol. 32 (2009) 1326.
[10] H.Y. Kuen, F.S. Mjalli, Y.H. Koon, Recursive least squares-based adaptive control of a biodiesel transesterication reactor, Ind. Eng. Chem. Res. 49 (2010)
1143411442.
[11] C.S. Bildea, A.A. Kiss, Dynamics and control of a biodiesel process by reactive
absorption, Chem. Eng. Res. Des. 89 (2011) 187196.
[12] W. Wali, A. Al-Shammaa, K.H. Hassan, J. Cullen, Online genetic-ANFIS temperature control for advanced microwave biodiesel reactor, J. Process Control 22
(2012) 12561272.
[13] P.T. Benavides, U. Diwekar, Studying various optimal control problems in
biodiesel production in a batch reactor under uncertainty, Fuel 103 (2013)
585592.
[14] A.S.R. Brsio, A. Romanenko, J. Leal, L.O. Santos, N.C.P. Fernandes, Nonlinear
model predictive control of biodiesel production via transesterication of used
vegetable oils, J. Process Control 23 (2013) 14711479.

139

[15] R.M. Ignat, A.A. Kiss, Optimal design, dynamics and control of a reactive DWC
for biodiesel production, Chem. Eng. Res. Des. 91 (2013) 17601767.
[16] W. Wali, K. Hassan, J. Cullen, A. Shaw, A. Al-Shammaa, Real time monitoring and intelligent control for novel advanced microwave biodiesel reactor,
Measurement 46 (2013) 823839.
[17] D.Y.C. Leung, X. Wu, M.K.H. Leung, A review on biodiesel production using
catalyzed transesterication, Appl. Energy 87 (2010) 10831095.
[18] H. Noureddini, D. Zhu, Kinetics of transesterication of soybean oil, J. Am. Oil
Chem. Soc. 74 (1997) 14571463.
[19] R. Richard, S. Thiebaud-Roux, L. Prat, Modeling the kinetics of transesterication reaction of sunower oil with ethanol in microreactors, Chem. Eng. Sci.
(2012).
[20] P. Lauenburg, J. Wollerstrand, Adaptive control of radiator systems for a lowest
possible district heating return temperature, Energy Build. 72 (2014) 132140.
[21] D.I. Wilson, M. Agarwal, D.W.T. Rippin, Experiences implementing the extended
Kalman lter on an industrial batch reactor, Comput. Chem. Eng. 22 (1998)
16531672.
[22] T.W. Patzek, A rst law thermodynamic analysis of biodiesel production from
soybean, Bull. Sci. Technol. Soc. 29 (2009) 194204.
[23] R. Luus, O.N. Okongwu, Towards practical optimal control of batch reactors,
Chem. Eng. J. 75 (1999) 19.
[24] E. Sorguven, M. zilgen, Thermodynamic assessment of algal biodiesel utilization, Renew. Energy 35 (2010) 19561966.
[25] W.M. Clark, N.J. Medeiros, D.J. Boyd, J.R. Snell, Biodiesel transesterication
kinetics monitored by pH measurement, Bioresour. Technol. 136 (2013)
771774.
[26] S. Zabala, G. Arzamendi, I. Reyero, L. Ganda, Monitoring of the methanolysis
reaction for biodiesel production by off-line and on-line refractive index and
speed of sound measurements, Fuel 121 (2014) 157164.
[27] I. Rechenberg, Evolutionsstrategie: Optimierung Technischer Systeme Nach
Prinzipien Der Biologischen Evolution, Frommann-Holzboog, 1973.
[28] Piedmont Biofuels, The price of biodiesel, 2013 http://www.biofuels.coop/
price-of-biodiesel
[29] B.R. Dhar, K. Kirtania, Excess methanol recovery in biodiesel production process
using a distillation column: a simulation study, Chem. Eng. Res. Bull. 13 (2010)
5560.
[30] B. Houska, H.J. Ferreau, M. Diehl, ACADO toolkit an open source framework
for automatic control and dynamic optimization, Optim. Control Appl. Meth.
32 (2011) 298312.
[31] Z.K. Nagy, B. Mahn, R. Franke, F. Allgwer, Evaluation study of an efcient output
feedback nonlinear model predictive control for temperature tracking in an
industrial batch reactor, Control Eng. Pract. 15 (2007) 839850.
[32] U.S. Department of Energy, Clean cities alternative fuel price report, 2013
http://www.afdc.energy.gov/uploads/publication/afpr jul 13.pdf
[33] National Biodiesel Board, Chemical weight and formula, 2013 http://
www.biodiesel.org/docs/ffs-performace usage/chemical-weight-formula.
pdf?sfvrsn=4
[34] SImetric, Density, mass, specic gravity of liquids, 2013 http://www.simetric.
co.uk/si liquids.htm
[35] Strategic Energy Technologies Information System, Background report on EU27 district heating and cooling potentials, 2012 https://setis.ec.europa.eu/
system/les/1.DHCpotentials.pdf
[36] California Energy Commission, Cost for cooling water, 2013 http://www.
energy.ca.gov/2006publications/CEC-500-2006-034/CEC-500-2006-034.PDF
[37] F. Allgwer, R. Findeisen, Z.K. Nagy, Nonlinear model predictive control: from
theory to application, J. Chin. Inst. Chem. Eng. 35 (2004) 299316.