Control Engineering Practice 22 (2014) 231–241

Contents lists available at ScienceDirect

Control Engineering Practice

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

Modeling and dynamic optimization of fuel-grade ethanol

fermentation using fed-batch process
Wei Dai a, Daniel P. Word b, Juergen Hahn a,n
a
Department of Biomedical Engineering, Department of Chemical & Biological Engineering, Rensselaer Polytechnic Institute, Troy, New York 12180, USA
b
Department of Chemical Engineering, Texas A&M University, College Station, Texas 77843-3122, USA

a r t i c l e i n f o abstract

Article history: This paper investigates optimization of operational strategies of an industrial ethanol fermentation
Received 10 June 2012 process. One of the challenges associated with this type or process is that most of the measurements
Accepted 19 January 2013 are only taken sporadically, thereby complicating process monitoring and optimization. The one
Available online 22 February 2013
exception to this rule involves temperature measurements, which are readily available. However, an
Keywords: existing model of the plant investigated in this paper does not include an energy balance and,
Ethanol fermentation accordingly, the temperature measurements cannot be used to estimate model parameters. This paper
Kinetic model addresses these deficiencies and proposes modifications to an existing ethanol fermentation model. The
Parameter estimation proposed changes include the derivation of an energy balance, modification of the reaction kinetics to
Dynamic optimization
include additional inhibition terms, and also estimation of model parameters from industrial data. The
new model is validated against plant data and then used for optimization of the process operations. It is
shown that modifications of the input profiles for the cooling rate and the glucoamylase addition can
lead to an approximately 10% increase in ethanol yield. These are promising results, even though these
findings will ultimately need to be validated during real plant operations.
& 2013 Elsevier Ltd. All rights reserved.

1. Introduction which translates to an ethanol concentration of over 160 (g/l)

Fuel-grade ethanol can be produced from plants such as maize,
sugar cane, or sweet sorghum. Producing bioethanol has one
potential advantage in that the amount of greenhouses gases
absorbed during the growing process can reduce the gases
produced during the combustion process. Furthermore, ethanol
can be used as a substitute for fuel additives (Lin & Tanaka, 2006).
As the world’s largest biofuel producer, the U.S. is producing
approximately 57 billion liters of ethanol in 2012, and the pro-
duction is estimated to reach 136 billion liters in 2022 (Walker,
2012). Most of the fuel-grade ethanol in the U.S. is produced from
maize-based plants, and more than 90% of the plants make use of
the dry mill process, where the whole grain is processed in four
steps: milling, liquification, simultaneous saccharification and
fermentation (SSF), and distillation (RFA. Accelerating Industry
Innovation, 2012). Simultaneous saccharification and fermenta-
tion is the most important step in the production process. In this
process, dextrin is broken down into fermentable dextrose by
glucoamylase, and dextrose is converted into ethanol by yeast,
e.g., S. cerevisiae. The theoretical maximum yield of ethanol from
maize starch is 0.364 (kg ethanol/kg dry corn) (Patzek, 2006),
after SFF in the batch fermenter. Currently, the most advanced
commercial plants claim to produce approximately 150 (g/l) of
ethanol after SSF, however, the average yield from these plants
using traditional techniques is less than 140 (g/l). Improving the
operations of the SSF process based on current facilities is
undoubtedly the most economic approach to increase ethanol
yield, but optimization of SSF operations requires a good under-
standing of the process, usually represented by data and models.
Models of the SSF process consist of dynamic balances of
components such as the concentrations of yeast, dextrose, ethanol,
and other substances. A variety of models of different complexity
have been developed, ranging from relatively simple models (Fogler,
2005), to quite detailed ones (De Andres-Toro et al., 1998; Lee, Kim, &
Rhee, 1992; Ochoa, Yoo, Repke, Wozny, & Yang, 2008). The main
differences in the complexity of the models results from
 Addition of intermediate steps in the saccharification process;
 Use of more appropriate kinetics reflecting yeast growth rate,
substrate inhibition, and product inhibition;
 Study of the temperature influence on enzyme and yeast
activity.

n
Corresponding author. Tel.: þ1 518 276 2138; fax: þ 1 518 276 3035. The early work of adding intermediate steps in the sacchar-
E-mail address: hahnj@rpi.edu (J. Hahn). ification process was done by Lee et al. A kinetic model with a

0967-0661/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.conengprac.2013.01.005
232 W. Dai et al. / Control Engineering Practice 22 (2014) 231–241
Nomenclature
F Prop Flowrate from the propagation tank (l/h)
F 0Prop Designed flowrate from the propagation tank (l/h)
t start Time point that the enzyme starts to flow into the
fermenter (h)
t drop Time point that the enzyme stops flowing into the
fermenter (h)
F Slurry Flowrate from the liquification tank (l/h)
F 0Slurry Designed flowrate from the liquification tank (l/h)
V Liquid volume in the fermenter (l)
V f ull Volume of the fermenter (l)
V GA Volume of the glucoamylase in the fermenter (l)
F GA Flowrate of glucoamylase (l/h)
yGA Concentration of glucoamylase in the fermenter (g/l)
rGA Density of glucoamylase (g/l)
yDex Concentration of dextrin in the fermenter (g/l)
yIN
Dex Concentration of dextrin from liquification tank (g/l)
ydextrose Concentration of dextrose in the fermenter (g/l)
yactive Concentration of active yeast in the fermenter (g/l)
ylag Concentration of lag yeast in the fermenter (g/l)
y0lag Concentration of lag yeast from propagation tank (g/l)
yEtOH Concentration of ethanol in the fermenter (g/l)
series of Michaelis–Menten equations has been proposed by
introducing a degree of polymerization (Lee et al., 1992). This
approach has subsequently been incorporated into several other
models to provide a more detailed description of the saccharifica-
tion process (Murthy, Johnston, Rausch, Tumbleson, & Singh,
2012). The kinetics of SSF have also been extensively studied:
the growth rate of yeast and the conversion rate between
different substances are usually expressed by a Monod equation,
while Moser and Tessier growth laws have alternatively been
used to provide better agreement with experimental data at the
beginning or end of fermentation. Also, several different types of
equations have been used to account for substrate and product
inhibition in various models (Mulchandani & Luong, 1989).
The operating temperature has been found to have a signifi-
cant impact on the process. Ethanol fermentation is an exother-
mic process and temperature changes will significantly affect
enzyme and yeast activity. This influence of the temperature on
enzyme and yeast activity is usually expressed using Arrhenius-
type equations such as those presented by Carvalheiro et al.
(Banat, Nigam, Singh, Marchant, & McHale, 1998; Phisalaphong,
Srirattana, & Tanthapanichakoon, 2006). Due to this temperature
dependence, glucoamylase has an optimum activity at 140 1F in
the saccharification process, while the optimal temperature for
yeast growth and ethanol fermentation is in the range of 84–
94 1F. These competing optimal process temperatures for the SSF
process highlight the complexity of selecting an optimal tem-
perature profile over the course of a batch. It should be noted that
most commercial plants hold the temperature constant between
86 and 90 1F.
One significant challenge for the use of SSF models in commercial
plants is the lack of data, since lag yeast, active yeast, and dead yeast
cannot be measured independently (dye or counting numbers using a
microscope have proven to be unreliable for this system). In addition,
fed-batch fermenters are operated as a ‘‘black box’’. Dextrose and
ethanol concentrations can be measured only every few hours and
other intermediate substances are difficult to measure. Due to this
lack of data, it is non-trivial to accurately estimate model parameters
from plant data, i.e., a model may fit existing data well, but it is not
possible to determine if the parameters are within a reasonable range
y0EtOH Concentration of ethanol from liquification tank (g/l)
T Temperature inside the fermenter (1F)
T ref Reference temperature used to calculate internal
energy (1F)
Hf eed Internal energy from the slurry flow and yeast flow
(Btu/h)
Hcool Heat taken away by cooling facility (Btu/h)
F cool Flowrate of cooling water (gal/h)
r Density of the liquid in the fermenter (g/l)
CP Heat capacity of the liquid in the fermenter (Btu/1F/l)
r GA Dextrin-dextrose conversion rate (h  1)
ma Dextrose-ethanol conversion rate (h  1)
mmax
a Maximum dextrose-ethanol conversion rate (h  1)
mlag Lag yeast-active yeast conversion rate (h  1)
mmax
lag Maximum lag yeast-active yeast conversion rate
(h  1)
ms Dextrose consumption rate (h  1)
mmax
s Maximum dextrose consumption rate (h  1)
mx Active yeast growth rate (h  1)
mmax
x Maximum active yeast growth rate (h  1)
rd Active yeast death rate (h  1)
fa Ethanol inhibition factor
as the models are overparameterized. The one process variable that is
measured often in commercial plants is the process temperature,
however, an energy balance is required if these temperature data are
to be used for parameter estimation or monitoring of any quantity
other than just the temperature. Unfortunately, none of the existing
models for SSF include an energy-balance equation as SSF is a
complex process and energy balances require a significant number
of parameters and relationships to describe the thermo-physical
properties. This paper addresses the above-mentioned challenges by
 Expanding an existing model to also include an energy-balance
equation with new parameters that are to be estimated from
available data;
 Establishing the relationship between cooling water flowrate,
temperature, and active yeast concentration, so that tempera-
ture data can be used to estimate unknown parameters arising
from balance equations other than the energy balance;
 Computing optimal input profiles for a fermenter used for ethanol
fermentation by using the developed model and a simultaneous
approach for solving the dynamic optimization problem.
The authors believe that this combination of a model, which is
based upon industrial plant data, with the determination of an
optimal input profile for the manipulated variables during a batch
is of interest to practitioners and researchers alike. The paper is
structured as follows: preliminaries about the process, the model, and
dynamic optimization are presented in Section 2. Section 3 discusses
the presented model in detail, including the estimated parameters.
The optimization problem for determining optimal input profiles is
formulated and solved in Sections 4 and 5 presents the conclusions.
2. Preliminaries
2.1. Description of industrial ethanol fermentation
Fuel-grade ethanol is produced in one of two ways, using
either the wet mill or dry mill process. Wet milling involves
 W. Dai et al. / Control Engineering Practice 22 (2014) 231–241 233

separating the grain kernel into different components: germ,

fiber, protein, and starch. The starch then undergoes the processes
of liquification, saccharification, and fermentation separately. This
work focuses on dry-mill based plants, where the entire grain
kernel is ground into flour, and the flour is transported into the
cook tank allowing for complete water penetration. Then, the
starch inside the flour is hydrolyzed into dextrin as it passes
through the liquification tank as shown in Fig. 1. Dextrin, which is Fig. 3. Algorithm for parameter estimation.
a mixture of different long-chain sugars (maltose, D3, D4–D7),
then enters the batch fermenter and undergoes a simultaneous Lopez-Orozco, & Fernandez-Conde (1997) and Iyer and Wunsch
saccharification and fermentation (SSF) process in response to the (2001), and the described phenomena fall into three categories:
addition of enzymes and of yeast from the propagation tank.
A general overview of the SSF process and a timeline describ-
ing batch operation is shown in Fig. 2. Nine pairs of ethanol and  Dynamic balances based upon first-principles: Eqs. (1)–(13), Refer-
dextrose concentration data were collected for the investigated ences De Andres-Toro et al. (1998), Lee et al. (1992), Ochoa et al.
fermentation process, including data at the beginning and the end (2008); Roeva et al. (2004), De Andres-Toro et al. (1997).
of the batch. Temperature data were collected much more  Kinetic expressions: Eqs. (14)–(18), References de Andres-Toro
frequently, usually every few minutes. The cooling system is et al. (1998), Lee et al. (1992), Ochoa et al. (2008),
operating during the entire fermentation, while the enzyme is Mulchandani & Luong (1989), De Andres-Toro et al. (1997),
added only during the filling phase which ranges from the first Iyer, & Wunsch (2001).
10%–20% of the batch time Fig. 3.  Temperature influence: Eqs. (19)–(23), References De Andres-
Toro et al. (1998), Banat et al. (1998), Phisalaphong et al.
(2006), De Andres-Toro et al. (1997).
2.2. Model of the process

The original model used in this work has been extensively The differential equations of the model describe the mass and
used for commercial fuel-grade ethanol production. The model component balances (glucoamylase, dextrin, dextrose, lag yeast,
originates from References: De Andres-Toro et al. (1998), Lee et al. active yeast, and ethanol) of the process inside the fermenter.
(1992), Ochoa et al. (2008), Mulchandani & Luong (1989), Banat dF Prop n o1
et al. (1998), Phisalaphong et al. (2006), Roeva, Pencheva, ¼ F Prop þ F 0Prop ð1Hðtt drop ÞÞHðtt Prop Þ ð1Þ
dt t
Hitzmann, & Tzonkov, (2004), De Andres-Toro, Giron-Sierra,
dF Slurry n o1
¼ F Slurry þ F 0Slurry ð1HðVV f ull ÞÞ ð2Þ
dt t
here, t is a time constant that is used to adjust the speed of
response of the flowrate from one set point to another.
dV
¼ F Prop þ F Slurry ð3Þ
dt

dV GA
¼ F GA ð4Þ
dt

dyGA F r
¼ GA GA l3 yGA ð5Þ
dt V

dyDex
¼ l1 yIN
Dex l3 yDex r GA yDex ð6Þ
dt
Fig. 1. Flow chart of the dry mill process.

dydextrose
¼ r GA yDex ms yactive l3 ydextrose ð7Þ
dt

dylag
¼ ðmlag l3 Þylag þ l2 y0lag ð8Þ
dt

dyactive
¼ ðmx r d l3 Þyactive þ mlag ylag ð9Þ
dt

dyEtOH
¼ ma f a yactive l3 yEtOH þ l1 y0EtOH ð10Þ
dt
The description of each variable can be found in the list of
symbols. Additionally, Hð. . .Þ is the Heaviside function and l1 ,
l2 , and l3 are introduced for notational purposes and are defined
as follows:
F Slurry
l1 ¼ ð11Þ
V

F Prop
l2 ¼ ð12Þ
Fig. 2. Overview of SSF process and timeline of batch operation. V
234 W. Dai et al. / Control Engineering Practice 22 (2014) 231–241

F Slurry þ F Prop and (26) represent the dynamic system with known initial
l3 ¼ l1 þ l2 ¼ ð13Þ
V values x0 , the path constraint P and the terminal constraint O.
Eqs. (1)–(10) describe the SSF process from the beginning of the For the model used in this work, the vector xj ðt i Þ includes the
filling time to the end of fermentation. In order to avoid introdu- dextrose concentration ydextrose , the ethanol concentration yEtOH ,
cing more complexity to the model, dextrin is assumed to be and the temperature T. Measurements of these states are used to
broken down into dextrose directly without considering inter- estimate the parameters g. The variable cij is introduced which is
mediate steps. This assumption is a trade-off between model equal to 1 if concentration measurements are available at time t i ,
accuracy and complexity as including additional steps would also otherwise cij is equal to 0. The reason for introducing cij is that
result in additional parameters which would introduce uncer- temperature measurements are collected much more frequently
tainty. The kinetic expressions from the dynamic balances shown than measurements of the other quantities.
above are given by Parameter estimation problems with dynamic systems are
commonly solved by applying an optimization algorithm in an
mmax
GA yGA outer loop, while the evaluation of the objective function and its
r GA ¼ ð14Þ
kGA þ yDex gradients is done by numerical integration of the ODEs in the
inner loop (Tjoa & Biegler, 1991). A commonly-used optimization
mmax
s ydextrose
ms ¼ ð15Þ technique for the outer loop is the Gauss–Newton method. This
ks þydextrose
methodology is a Newton method where the computation of the
mmax ydextrose Hessian matrix has been simplified by assuming that the residuals
a
ma ¼ ð16Þ are small at the solution and thus can be neglected. However, the
ks þ ydextrose
assumption of small residuals is not always valid, and several
mmax
x ydextrose modifications of this method have been proposed. Alternative
mx ¼ ð17Þ
ðkx1 þyEtOH Þðkx2 þ ydextrose Þ methods include the Trust-Region methods, rotational discrimi-
nation, and hybrid methods (Biegler, Damiano, & Blau, 2004;
f a ¼ ð1ayEtOH Þ ð18Þ Chouakri, Fonteix, Marc, & Corriou, 1994).
Eqs. (14)–(17) are modified Michaelis–Menten functions The Trust-Region method is used in this work and is able to
which use Monod kinetics. Eq. (17) takes into account the handle singular Hessian matrices, significant uncertainty in para-
simultaneous substrate and product inhibition on active yeast. meters of the models, and noisy data. In addition, the Trust-
Eq. (18) represents a factor for describing product inhibition on Region method is guaranteed to converge to local optima with
active yeast. much weaker assumptions than line search methods. It is worth
The effect of temperature on several of the model parameters noting that this work uses a fixed-step Runge–Kutta method for
is captured by the following expressions: each integration step in the inner loop, rather than a variable-
step-size Runge–Kutta method, since variable-step-methods will
mmax
GA ¼ d1 g1 ða1 þ b1 Þ ð19Þ introduce additional ‘‘noise’’ to the optimization process that may
greatly reduce the efficiency.
mmax
s ¼ d2 expðg2 ða2 T þb2 Þ þ c2 Þ ð20Þ

mmax ¼ d3 expðg3 ða3 T þb3 Þ þ c3 Þ ð21Þ 2.4. Dynamic optimization of a batch process
a

mmax
x ¼ d4 g4 expða4 T þ b4 Þ ð22Þ
r d ¼ d5 expða5 T þ b5 Þ ð23Þ
Arrhenius-type equations are used to describe the temperature
effect on the enzyme and yeast activity, where a1,y,5, b1,y,5, c2,3,
d1,y,5, are known parameters, and g1,y,4, are parameters that
need to be estimated. The parameters to be estimated were
determined by local sensitivity analysis. Data for estimation of
parameters of this model were collected at higher frequency than
normal and the fermenter was operated at different, but constant,
temperatures. It is important to note that this model describes the
effect of temperature on the system, however, the effect of the
exothermic reactions and cooling of the process is not captured in
the model as the model contains no energy balance.
2.3. Parameter estimation of dynamic systems
The parameter estimation problem for a dynamic system of
ODE’s can be formulated mathematically as follows:
XX
min cij ðxj ðt i Þxnij Þ2 ð24Þ
g
i j
s:t: x_ ¼ Fðx, g,uÞ, xð0Þ ¼ x0 ð25Þ
Pðx,uÞ r0, Oðxðt f ÞÞ r0 ð26Þ
here, xj ðt i Þ is the simulated value of component j at sampling
time t i ; xnij is the experimental data of component j at sampling
time t i ; g are the unknown parameters; u are the inputs; Eqs. (25)
In the dynamic process described by Eqs. (1)–(23), the state
variables undergo significant changes during SSF. There is no steady
state and thus no constant setpoints over the course of a batch.
Therefore, the major objective is to have the controlled variables
follow a reference trajectory which maximizes a performance index,
e.g., the ethanol yield at the end of the run (Srinivasan, Palanki, &
Bonvin, 2003). This problem belongs to the category of dynamic
optimization problems that can be formulated as follows:
max y ¼ xEtOH ðt f Þ ð27Þ
u
s:t: x_ ¼ Fðx, g,uÞ, xð0Þ ¼ x0 ð28Þ
Pðx,uÞ r0, Oðxðt f ÞÞ r0 ð29Þ
here, y is the performance index; Eqs. (28) and (29) represent the
dynamic system with known initial values x0 , known parameters g,
path constraint P, and terminal constraint O. For this work, the
performance index is the final ethanol concentration and the most
important path constraint is to maintain the dextrose concentration
within a given range. Other path constraints include the flowrate
and total volume of glucoamylase, and the temperature. There is no
terminal constraint in the model. The ultimate goal of this work is to
determine a profile of the input u, i.e., the glucoamylase flowrate
and cooling water flowrate profiles, that maximizes the final ethanol
concentration at the end of the batch.
Two different approaches have been reported in the literature
for direct optimization methods, depending on whether the
dynamic system is integrated explicitly or implicitly: the Control
Vector Parameterization (CVP) approach and the simultaneous
 W. Dai et al. / Control Engineering Practice 22 (2014) 231–241
approach. The CVP approach parameterizes the inputs using a
finite number of decision variables, integrates the system states
to compute the performance index, and then uses an optimization
algorithm, e.g., Quasi-Newton methods, to update the values of
the decision variables. Although the CVP approach is straightfor-
ward to implement, it tends to be slow, especially when dealing
with inequality path constraints (Vassiliadis, Sargent, &
Pantelides, 1994a, b). The simultaneous approach discretizes both
the inputs and the ODEs, so that the differential equations are
satisfied only at a finite number of time points (typically via
orthogonal collocation). This step transforms the dynamic opti-
mization problem into a nonlinear programing problem (NLP)
involving a large number of algebraic equations (Kameswaran &
Biegler, 2006). The transformed problem can be of significant size
for stiff systems due to the large number of discretization points
required, and will require a powerful NLP solver.
It should be pointed out that parameter estimation of dynamic
systems can also be performed via the simultaneous approach.
However, the sequential approach described in Section 2.3 is
easier to implement while it was found to be sufficiently fast for
the estimation of model parameters performed in this work. On
the other hand, computing a time-varying profile of an input that
maximizes the ethanol yield was found to be sufficiently more
difficult than the parameter estimation and therefore made use of
the simultaneous approach presented in this subsection.
3. Model development
The model presented in Section 2.2 provides a starting point for
the modeling effort presented in this work. Specifically, there are
several areas where changes are made to the existing model:
reformulate the flowrate equations, which reduces the stiffness of
the system and simplifies numerical solution of the model; use more
appropriate kinetic functions to reflect the synergistic substrate and
product inhibition; incorporate an energy balance equation to better
reflect the connection between the temperature and the reactions
occurring inside the tank; and most importantly, estimate para-
meters using the large amount of temperature data. These tasks will
be described in more detail in the following subsections.
3.1. Reformulate the flowrate equations
Eqs. (1) and (2) describe the slurry flowrate from the liquifica-
tion tank and the enzyme flowrate from the propagation tank. The
shape of the flowrate profile is similar to a step function, and the
slope of the profile is controlled by adjusting the parameter t.
Here, t reflects the speed of response of the transient from one set
point to another. However, considering the fact that the flowrate
is maintained constant for most of the time, and the transient
behavior of these two flow rates at the start up and shut down is
almost negligible compared to the filling time, it is reasonable to
replace Eqs. (1) and (2) with the following expressions:
F Prop ¼ F 0Prop ð1Hðtt drop ÞHðtt start ÞÞ
F Slurry ¼ F 0Slurry ð1HðVV f ull ÞÞ
Even though Eqs. (30) and (31) can no longer reflect the
transient behavior of the flow rates, this substitution has sig-
nificantly reduced the stiffness of the system as all states now
have time constants that are of the same order of magnitude. This
change can also have a positive effect on the speed of the
numerical integration and other routines, such as parameter
estimation, which make use of the model.
235
3.2. Simultaneous substrate and product inhibition
Simultaneous elevated concentrations of dextrose and ethanol
during the SSF process exhibit synergistic stress on the yeast and
can cause incomplete fermentation. The fermentation process
needs to be carefully managed to minimize this type of stress.
As dextrose is produced and consumed during the process, its
concentration reaches its highest level at some point during the
batch, whereas the ethanol concentration is steadily increasing
during the process as ethanol represents the final product of this
process. Eq. (17) uses a Monod kinetics expression to reflect the
substrate inhibition; however, this expression has been replaced
in this work by a Haldane kinetics term to better represent
substrate and product inhibition
mmax ydextrose
x
mx ¼ ð32Þ
ðkx1 þ yEtOH Þðkx2 þydextrose þ y2dextrose =kx3 Þ
Furthermore, maintaining the dextrose concentration within a
certain range forms the most important path constraint for
dynamic optimization of this process. The use of Haldane kinetics
can indirectly simplify the optimization problem as the inhibition
term will more accurately reflect the lower reaction rates and
thereby more easily keep the dextrose concentration within its
bounds. However, it should be noted that this modification of the
model was made to more accurately reflect the process and not
for ease of computation purposes.
3.3. Development of an energy balance for the fermentation process
While several papers describing models of the SSF process can
be found in the open literature, none of the models include an
energy balance which reflects the relationship between tempera-
ture, the heat released by the ongoing reactions, and the cooling
mechanism. The main reason for this is that deriving an energy
balance for this process is non-trivial due to the complex nature of
the process. However, there are significant benefits of using an
energy balance which lead to the development of the balance
equation in this work: (1) temperature can be frequently and easily
measured during SSF, and these data can be used for parameter
estimation not only of parameters of the energy balance but also
other parameters included in the model; (2) the energy balance
incorporates the effect that the cooling water flowrate has on the
process temperature and thereby reflects the time constants at
which the fermenter temperature can change.
Several assumptions are made for deriving the energy balance:
 The density r and the heat capacity C P are constant,
 Reaction heat is constant, and only the heat generated by the
fermentation (Eq. (10)) is included,
 Heat loss of the system to the surrounding area is neglected;
this assumption obviously does not ignore the heat transferred
via the cooling system.
Based upon these assumptions, the energy balance of the
ð30Þ
fermenter can be written as
dfrVC P ðTT ref Þg
ð31Þ ¼ ðHf eed Hcool Þ þHreaction ð33Þ
dt
Hreaction ¼ ðDHÞma f a yactive V ð34Þ
here, T ref denotes the reference temperature that is used to
calculate the internal energy of each substance; the initial
temperature T0 of the mash slurry from the liquification tank is
usually used as T ref ; Hf eed denotes the total internal energy of the
mash slurry from the liquification tank and that of the yeast flow
236 W. Dai et al. / Control Engineering Practice 22 (2014) 231–241
Table 1
The estimated value of unknown parameters.
g1 g2 g3 g4
1.2778 0.4319 0.2174 0.9395
Table 2
The initial conditions of the SSF model.
V 0 (gal) V 0GA (gal) y0GA (g/l) y0lag (g/l)
5000 0 0 0 0
from the propagation tank; Hcool represents the heat removed by
the cooling water and has a linear relationship with F cool ; Hreaction
represents the energy produced during the fermentation process.
Rearranging Eq. (33) and making substitutions results in
dT Hf eed Hcool ðDHÞma f a yactive V
¼ þ l3 ðTT ref Þ
dt rC P V rC P V
where l3 represents the volume fraction expressed in Eq. (13).
While the energy balance shown in Eq. (35) is straightforward,
it includes several parameters whose values are not easy to
determine as they are not constant. For example, as the reaction
proceeds, solid particles inside the slurry dissolve into the solu-
tion and gas products escape from the solution, changing the
concentration of each substance. Therefore, density and heat
capacity vary with changes of the components in the fermenter,
even though one of the assumptions made above was that these
properties are constant. In addition to generating heat, energy
from the consumption of dextrose is also used for the growth of
active yeast and the production of other byproducts. This effect is
also influenced by the concentrations of the components and
other conditions in the fermenter. In order to capture the
uncertainties and correlate them with concentrations of compo-
nents in the fermenter, the factors c1 and c2, shown in Eq. (36),
have been introduced.
dT Hf eed Hcool mf y
¼ C1 þ C2 a a active l3 ðTT ref Þ
dt rC P V rC P
here,
C1 ¼ G1 þ G2 yEtOH
C2 ¼ G3 þ G4 yEtOH þ G5 y2EtOH
While there is no specific reason why these factors are
expressed in this form, it has been found that this type of
expression provided significantly better results for the tempera-
ture than other expressions, such as higher or lower polynomials,
polynomials which also depended upon other concentrations, etc.,
when applied to the available plant data sets. The approach that
was used for selecting the structure for the polynomials involved
that the coefficients of the polynomials were estimated using the
procedure described in Section 2.3. The model predictions for the
temperature and the concentrations, involving each of the deter-
mined correlations, were compared to plant data not used for
estimating the correlation parameters. A comparison of the MSE
between the predictions and the data was made and the best
correlations, as represented by Eqs. (37) and (38), was chosen.
3.4. Parameter estimation and model validation
The model is composed of Eqs. (3)–(16), (18)–(23), (30)–(32)
and (36)–(38). The model includes nine unknown parameters
which need to be estimated: g1,y,g4 deal with uncertainty in the
G1 G2 G3 G4 G5
1.0385  0.0051 647.6748  7.5731 0.0285
y0active (g/l) y0Dex (g/l) y0dextrose (g/l) y0EtOH (g/l) T0 (1F)
244 10 4 92
ð35Þ
Fig. 4. Normalized cooling water flowrate over the course of a batch. Note that
batch time is shown in percentage of total batch time.
kinetic expressions, while G1,y,G5 capture the uncertainty in the
energy balance. The data set used for training includes dextrose
concentration, ethanol concentration, and temperature from
Batch 001. Since dextrose concentration and ethanol concentra-
tion are measured sporadically during the entire SSF process
while temperature is recorded every few minutes, there are
significantly more data points for the temperature than for the
concentrations. Therefore, a coefficient a is introduced to adjust
the numerical weight placed on the data during parameter
estimation. The sum of the squared error of the temperature is
ð36Þ multiplied by a, and the sum of the squared error of component
concentrations is multiplied by (1 a). a is then selected so that
each data set is weighted equally in the estimation.
ð37Þ Using the Trust-Region approach with the fixed-step 5th order
Runge–Kutta method mentioned in Section 2.2, the parameter
ð38Þ values have been estimated and are shown in Table 1. Using these
parameters, the initial conditions presented in Table 2, and the
normalized cooling profile shown in Fig. 4, the system can be
simulated. Note that the cooling profile has been normalized as
this information is derived from proprietary plant data. The
simulation results of the several key factors of the system are
shown in Fig. 5. It can be seen that the simulated dextrose
concentration, ethanol concentration, and temperature match
the experimental data with a good degree of accuracy for this
type of process. Furthermore, simulations have been carried out
for four other batches for which plant data were available (Batch
002, Batch 003, Batch 004, and Batch 005). For all of these batches,
the simulations fit the data well as can be seen in Fig. 6 where the
ethanol and dextrose concentrations are shown.
4. Determination of optimal input profiles
4.1. Optimization problem formulation
As the derived model has been validated in Section 3, it is now
possible to use this model for determining improved operating
 W. Dai et al. / Control Engineering Practice 22 (2014) 231–241 237

Fig. 5. Simulation result with estimated parameters (Batch 001).

Fig. 6. Comparing model predictions to plant data for different batches where the parameter values were determined based upon data from Batch 001.

conditions of the process. The process improvement is based upon discretized in the simultaneous approach and are treated as
the solution of a dynamic optimization problem using the optimization variables. The discretized model is then included
simultaneous approach. All variables, i.e., states and inputs, are in the form of algebraic constraints in the optimization problem.
238 W. Dai et al. / Control Engineering Practice 22 (2014) 231–241
Both the control profile and the system’s states are discretized
using a three-point Radau collocation on finite elements (Zavala,
2011). Using this discretization technique the differential equation
dx
¼ f ðx,tÞ ð39Þ
dt
results in a set of algebraic equations
X an upper bound. In this way, the capacity and performance of the
xi,j ¼ xi1,3 þ h ai,j f ðxi,k ,t i,k Þ ð40Þ
k A CP
here, x denotes the state vector consisting of entries of the states at
different points in time t. In the algebraic equations, i is the finite
element number, j and k are collocation point numbers, and CP is
the set of collocation points. The finite element length is h, and a is
a matrix of collocation equation coefficients.
For the model developed in Section 3, t is the time horizon of
the entire SSF process. It is discretized on finite elements with the
length h, ranging from 0.027% to 0.833% of the time horizon t,
which is determined by the accuracy of the Radau collocation, the
stiffness of the system, and the size of the problem after
discretization. After gradually relaxing the length limit when
implemented in the simultaneous approach, 0.556% was deter-
mined as the element length h. Then, the total number of finite
elements nf e is 180. Within each finite element, the number of
 
collocation points is 3. Therefore, the set f e ¼ 1, . . .,nf e repre-
sents the entire SSF time horizon; the filling phase and other time
fraction can be defined as a subset of f e; f1, 2, 3g is the collocation
point set CP. The matrix of collocation coefficients a used in the
discretized model is found from solving for the roots of the
Lagrange interpolating polynomials:
0 1
0:19681547722366 0:39442431473909 0:37640306270047
B 0:29207341166523 0:51248582618842 C
a ¼ @ 0:06553542585020 A the glucoamylase near the end of the filling time rather than
0:02377097434822 0:04154875212600 0:11111111111111
ð41Þ
By discretizing the continuous differential equation model into
an algebraic model, the model can be formulated as a nonlinear
programming problem as is shown in Table 3. This optimization
problem is implemented in AMPL (Robert, David, & Brian, 2002),
which is an algebraic modeling language that provides first and
second order derivatives through automatic differentiation, and
Table 3
Formulation of the dynamic optimization problem when the model is discretized
(expressed with pseudo code).
Objective function:
maximize ethanol : y_ethanol½lastðfeÞ, lastðCPÞ
Manipulated variables:
var F_ga {fe, CP}
var F_cool {fe, CP}
Discretized model (for example, ydextrose is shown as below)
dy_dextrose_dt[i, j] {i in fe, j in CP}¼
r_ga[i,j]*y_dex[i, j]-mu_s[i,j]*y_active[i,j]-
F_slurry[i,j] þF_prop[i,j]*y_dextrose[i,j]/V[i,j];
y_dextrose[i,j] {i in 1.0.1, j in CP} ¼
y_dextrose_init þ h*sum{k in cp} a[k,j]*y_dextrose_dt[i,k]; (Radau Collocation
at initial point)
y_dextrose[i,j] {i in fe diff{1}, j in CP}¼
y_dextrose[i-1,last(cp)] þ h*sum{k in cp} a[k,j]*y_dextrose_dt[i,k]; (Radau
Collocation at other points)
State variable constraint (for example, T is shown as below)
var T{fe, CP}¼90.0, 4 ¼ 88.0, o ¼100;
param T_initial¼92;
Manipulated variable constraint (for example H_cool is shown as below)
diff_F_cool[i] {i in fe}o ¼ upper bound
F_cool[i] {i in fe}o ¼ upper bound*
the optimization problem is solved using the interior-point non-
linear solver IPOPT (Wächter & Biegler, 2006).
The constraint of the manipulated variable can be applied at any
discretized point in the set f e and each collocation point in the set CP
separately, either on the variable itself or on its first-order derivative.
For example, in Table 3, the cooling water flowrate can be given an
upper bound, and the flowrate change per unit time can also be given
cooling facility can be reflected in the constraints. Meanwhile, the
constraint of the state variables can also be applied, not only on
the discretized points, but also on the collocation points. For example,
the temperature profile can be controlled within a reasonable range
through the entire SSF process.
4.2. Dynamic optimization result and analysis
Solution of the optimization problem formulated in Section 4.1
returns the profiles shown in Fig. 7. It can be seen that the optimal
temperature profile maintains a high temperature during the
filling time, but then rapidly drops down to the lower operating
condition towards the end of the filling phase and remains there
for most of the batch time. Towards the end of the batch, the
temperature slightly rises as no cooling is taking place and the
fermentation still produces a small amount of heat. This optimal
profile behavior can be explained by the fact that a high
temperature during the filling phase favors saccharification to
generate a sufficient amount of dextrose for fermentation, and a
lower temperature favors yeast growth and fermentation for the
remaining time of the batch. The temperature increase at the end
of fermentation is to stimulate saccharification to prevent a halt
in fermentation, since the dextrose concentration is very low near
the end of fermentation. The optimal glucoamylase profile injects
throughout the entire filling phase, so that the dextrose concen-
tration will be controlled within a certain range.
The fact that the temperature slightly increases towards the
end of the batch time leads to investigation of what benefits a
more significant increase in temperature towards the end of the
batch can return. In order to investigate this, the optimization
problem was reformulated such that cooling as well as heating of
the fermenter was allowed. It was found that the glucoamylase
profile returned by the optimization is exactly the same as the
one shown in Fig. 7, demonstrating that heating the fermenter
will not change the glucoamylase injection strategy. While the
temperature profile with both heating and cooling shows a
similar pattern as the profile with just cooling, there are some
differences that can be seen. For example, the temperature
reaches the upper bound of the temperature constraint during
the filling time. This result reinforces the interpretation of the
optimal temperature profile given above – that a high tempera-
ture during the filling time favors saccharification, while lower
temperature favors fermentation, and the final increase in tem-
perature is for stimulation of saccharification. While allowing
heating and cooling of the fermenter increases the temperature
change required of the facility, there is only a marginal benefit in
the final yield of ethanol (shown in Table 4). Based upon this, the
benefits provided by heating and cooling the fermenter are likely
not worth the additional costs associated with this change and
only cooling the fermenter will be sufficient in practice.
It can be seen from the optimal temperature profile that the
temperature remains at the lower bound for a majority of the
batch time. Due to this, it is important to investigate what effect a
change in the lowest allowable temperature has on the ethanol
yield. Fig. 8 shows the optimal temperature profile for a lower
bound of 86 1F. It can be seen that the temperature profiles retain
a similar shape as before but that the profile still remains at the
 W. Dai et al. / Control Engineering Practice 22 (2014) 231–241 239

Fig. 7. Optimal input profiles and simulation result after dynamic optimization.

Table 4 process of saccharification proceeds. The simulation results sup-

Comparison of different scenarios regarding ethanol yield.
porting this idea are shown in Fig. 9.
Temperature Cooling/ Maize Optimization status Ethanol As a consequence of applying high-amount-solid mash, higher
bound (1F) heating solid (g/L) yield (g/L) concentrations of dextrin and dextrose would be reached, which
expose the active yeast to osmotic stress and even substrate
88–100 Cooling only 327.5 Before optimization 133.0 inhibition. This impact could be mitigated by optimizing the
88–100 Cooling only 337.5 Before optimization 137.2
Cooling only After optimization 142.2
glucoamylase injection strategy and temperature profile with
88–100 327.5
Cooling and After optimization the same approach used in previous discussion. However, the
88–100 327.5 142.5
heating subsequent problem is that the increased amount of solids in the
86–100 Cooling only 327.5 After optimization 147.9 mash will also increase the burden on other processes. For
example, more maize particles in the solution requires a longer
cooking time in the cooking tank to allow for complete penetra-
lower bound for extended periods of time. Similar results are tion of water, and the separation of more solid residues requires
found when the lower bound is further relaxed to a temperature larger distillation columns. The changes of the production process
of 84 1F. In both of these cases the ethanol yield has improved and the facility will definitely increase the expense of fuel-grade
significantly over the original one, which suggests that keeping ethanol production. In order to get the optimal solids amount,
the temperature at a reasonably low level shortly after the filling modeling of both upstream and downstream processes is
phase is something that could be considered. That being said, it is required and no final conclusions can be drawn from this work
important to note that the model parameters were estimated regarding this point.
from data which came from processes that operated at a tem-
perature at or above 88 1F, and that optimization results show
that a process should be operated at a lower temperature should 5. Conclusions
be taken with a grain of salt. Nevertheless, the results are
suggestive and point towards some modifications that can poten- In this study, a novel model of a simultaneous saccharification
tially improve ethanol yield. and fermentation process is presented that incorporates an
It can also be seen from Fig. 7 that the optimal dextrose energy-balance equation into a model currently used in several
concentration is lower than the one where no optimization was commercial ethanol plants. This new model allows the use of a
performed. This allows to consider increasing the source of large amount of historical temperature data for parameter esti-
dextrose so that more ethanol would be produced while main- mation. The estimation algorithm presented here is based on a
taining sufficiently low dextrose levels throughout the batch. One Trust-Region approach with a fixed-step Runge–Kutta method.
approach is to increase the total volume of glucoamylase. How- The simulation results fit the original experimental data well for
ever, simulations have shown that there is no difference in the both the training and the testing set.
final ethanol yield when the total volume of glucoamylase is This model was subsequently used for optimizing the process.
increased. The reason for this is that all of the dextrin has been The optimal control profile (cooling water flowrate and glucoa-
broken down into dextrose before the end of fermentation, and mylase flowrate) of the SSF model is found by solving a dynamic
the extra amount of glucoamylase does not affect the overall optimization problem using a simultaneous approach. The opti-
outcome. Another approach is to increase the amount of solids mal temperature profile has a high temperature during the filling
(maize particles) in the slurry mash at the very beginning before time and a low temperature for the rest of the process. The
the liquification tank. This would provide more dextrose as the optimal glucoamylase profile shows that it should be injected into
240 W. Dai et al. / Control Engineering Practice 22 (2014) 231–241
Fig. 8. Comparison of the results of the optimized profiles for a process where only cooling is allowed, the process where cooling and heating can be performed, and the
original profiles.
Fig. 9. Comparison between the simulation results with designed amount of maize solids and that of increased amount of maize solids.
the fermenter near the end of the filling time rather than at a
constant rate during the entire filling time. The dextrose concen-
tration after optimization is well controlled within a reasonable
range, and the ethanol concentration is increased after optimiza-
tion by as much as 7%. Furthermore, relaxing the lower bound of
the temperature constraint proves to be the most effective way to
further increase the final yield (increasing production by as much
as 11% when the lower bound is set to 86 1F). Heating and cooling
the fermenter will not result in a significantly higher ethanol yield
but will require more energy and additional facilities. Increasing
the solids in the slurry flow will increase the final ethanol yield,
but will also increase the burden on other processes such as the
upstream liquification process and the downstream distillation
process. While the derived model used in this investigation is
based upon plant data, it is nevertheless required to validate the
model predictions by actually running a plant in this fashion
before any definite conclusions can be drawn. That being said, this
work provided several suggestions that indicate that they can
result in increased ethanol yields, and the effect of these sugges-
tions on other parts of the plants has also been discussed.
Acknowledgments
The authors gratefully acknowledge partial financial support
from Rockwell Automation, the Process Science Technology
Center at UT Austin, and from the National Science Foundation
(Grant CBET#0941313). We are also grateful to Carl Laird and
Bijan Sayyar-Rodsari for discussions and input.
References
Banat, I., Nigam, P., Singh, D., Marchant, R., & McHale, A. (1998). Review: ethanol
production at elevated temperatures and alcohol concentrations: Part I–yeasts
in general. World Journal of Microbiology and Biotechnology, 14, 809–821.
Biegler, L., Damiano, J., & Blau, G. (2004). Nonlinear parameter estimation: a case
study comparison. AIChE Journal, 32, 29–45.
Chouakri, N., Fonteix, C., Marc, I., & Corriou, J. (1994). Parameter estimation of a
Monod-type model. Part I: theoretical identifiability and sensitivity analysis.
Biotechnology Techniques, 8, 683–688.
De Andres-Toro, B., Giron-Sierra, J., Lopez-Orozco, J., Fernandez-Conde, C.,
Peinado, J. M., & Garcia-Ochoa, F. (1998). A kinetic model for beer production
under industrial operational conditions. Mathematics and Computers in Simula-
tion, 48, 65–74.
De Andres-Toro, B., Giron-Sierra, J., Lopez-Orozco, J., & Fernandez-Conde, C. (1997).
Evolutionary optimization of an industrial batch fermentation process. Europan
Control Conference, ECC’97.
Fogler, H. S. (2005). Elements of chemical reaction engineering (4th ed.). Prentice
Hall, Inc., NJ, USA.
Iyer, M. S., & Wunsch, D. C. (2001). Dynamic re-optimization of a fed-batch
fermentor using adaptive critic designs. IEEE Transactions on Neural Networks,
12, 1433–1444.
Kameswaran, S., & Biegler, L. T. (2006). Simultaneous dynamic optimization
strategies: recent advances and challenges. Computers & Chemical Engineering,
30, 1560–1575.
Lee, C. G., Kim, C., & Rhee, S. (1992). A kinetic model and simulation of starch
saccharification and simultaneous ethanol fermentation by amyloglucosidase
and Zymomonas mobilis. Bioprocess and Biosystems Engineering, 7, 335–341.
Lin, Y., & Tanaka, S. (2006). Ethanol fermentation from biomass resources: current
state and prospects. Applied Microbiology and Biotechnology, 69, 627–642.
 W. Dai et al. / Control Engineering Practice 22 (2014) 231–241
Mulchandani, A., & Luong, J. (1989). Microbial inhibition kinetics revisited. Enzyme
and Microbial Technology, 11, 66–73.
Murthy, G. S., Johnston, D. B., Rausch, K. D., Tumbleson, M., & Singh, V. (2012).
A simultaneous saccharification and fermentation model for dynamic growth
environments. Bioprocess and Biosystems Engineering, 1–16.
Ochoa, S., Yoo, A., Repke, J. U., Wozny, G., & Yang, D. R. (2008). Modeling and
parameter identification of the simultaneous saccharificationfermentation
process for ethanol production. Biotechnology Progress, 23, 1454–1462.
Patzek, T. W. (2006). A statistical analysis of the theoretical yield of ethanol from
corn starch. Natural Resources Research, 15, 205–212.
Phisalaphong, M., Srirattana, N., & Tanthapanichakoon, W. (2006). Mathematical
modeling to investigate temperature effect on kinetic parameters of ethanol
fermentation. Biochemical Engineering Journal, 28, 36–43.
RFA. Accelerating Industry Innovation (2012). Ethanol industry outlook. Washington
DC: Renewable Fuel Association.
Robert, Fourer, David, M.Gay, & Brian, W.Kernighan (2002). AMPL: A modeling
language for mathematical programming (2nd ed.). Duxbury Press.
Roeva, O., Pencheva, T., Hitzmann, B., & Tzonkov, S. (2004). A genetic algorithms
based approach for identification of Escherichia coli fed-batch fermentation.
Bioautomation, 1, 30–41.
241
Srinivasan, B., Palanki, S., & Bonvin, D. (2003). Dynamic optimization of batch
processes: I. Characterization of the nominal solution. Computers & Chemical
Engineering, 27, 1–26.
Tjoa, I. B., & Biegler, L. T. (1991). Simultaneous solution and optimization strategies
for parameter estimation of differential-algebraic equation systems. Industrial
& Engineering Chemistry Research, 30, 376–385.
Vassiliadis, V., Sargent, R., & Pantelides, C. (1994a). Solution of a class of multistage
dynamic optimization problems. 1. Problems without path constraints. Indus-
trial & Engineering Chemistry Research, 33, 2111–2122.
Vassiliadis, V., Sargent, R., & Pantelides, C. (1994b). Solution of a class of multistage
dynamic optimization problems. 2. Problems with path constraints. Industrial
& Engineering Chemistry Research, 33, 2123–2133.
Wächter, A., & Biegler, L. T. (2006). On the implementation of an interior-point
filter line-search algorithm for large-scale nonlinear programming. Mathema-
tical Programming, 106, 25–57.
Walker, G. M. (2012). 125th anniversary review: fuel alcohol: current production
and future challenges. Journal of the Institute of Brewing, 117, 3–22.
Zavala, V. M. (2011). Computational strategies for the optimal operation of large-scale
chemical processes. ProQuest UMI Dissertation Publishing, MI, USA.