You are on page 1of 9

Bioresource Technology 328 (2021) 124836

Contents lists available at ScienceDirect

Bioresource Technology
journal homepage: www.elsevier.com/locate/biortech

Advanced nonlinear control strategies for a fermentation bioreactor used


for ethanol production
Emil Petre a, Dan Selişteanu a, *, Monica Roman a
a
Department of Automatic Control and Electronics, University of Craiova, Craiova, A.I. Cuza 13, 200585, Romania

H I G H L I G H T S

• Nonlinear control structures for a bioethanol fermentation process are developed.


• Two adaptive control loops are built for substrate concentration and temperature.
• Design assumptions: unknown kinetics and inputs, noisy data, unmeasurable states.
• Control schemes: linearizing laws, unknown input and state observers, estimators.
• Goal: to increase the ethanol production and to minimize its environmental impact.

A R T I C L E I N F O A B S T R A C T

Keywords: This study addresses the design of advanced control schemes implemented for a continuous fermentation process
Ethanol fermentation processes used to produce ethanol. Due to the inaccuracy of the models that express this complex process, a feasible
Bioreactor temperature control controller is required to maximize the production of ethanol and to minimize its environmental impact, despite
Nonlinear systems
the existence of some significant uncertainties. Therefore, novel estimation and control schemes are designed and
Adaptive control
Observers
tested. These schemes are adaptive control laws including nonlinear estimation algorithms: a sliding mode
observer to estimate the unknown influent concentration, but also state observers and parameter estimators used
to estimate the unknown states and kinetics. Since the temperature is an important factor for an efficient
operation of the process, an algorithm for temperature control in the bioreactor is also developed. To verify the
control algorithms effectiveness, several tests performed via numerical simulations under realistic conditions are
presented.

1. Introduction addressed to ethanol production by using fermentation (Kumar et al.,


2018; Taghizadeh-Alisaraei et al., 2019). Considered as an alternative to
The increase in energy consumption, caused by the growth of the fossil fuels, the bioethanol can be produced from various raw materials
global population and its needs, has also led to excessive consumption as well as by conversion of biomass (Fernandez et al., 2020; Quintero
especially of fossil fuels with adverse consequences on the environment. et al., 2009). Also, the ethanol is viewed as the cleanest liquid fuel being
Since conventional non-renewable energy resources (e.g., fossil fuels) less toxic, easily biodegradable, and producing a smaller amount of air-
are quite limited, in last decades, the development of some renewable borne pollutants than petroleum fuel (Fernandez et al., 2020). Usually,
sources (such as hydropower, solar energy, fusion devices – e.g., ITER, in industry, this kind of processes are carried out in Continuous Stirred
wind, geothermal energy, biomass) for energy production has gained a Tank Reactors (CSTRs). Their mathematical models are obtained by
wide attention (Gaida et al., 2017; Khetkorn et al., 2017; Negrea et al., using mass and energy balances, resulting systems of ordinary differ­
2008; Negrea, 2020; Pandey et al., 2018). ential equations (ODEs). A typical drawback of the continuous fermen­
Along with the improvement of the technologies used to produce tation is the presence of oscillations under certain conditions, the effect
biogas and methane gas (Gaida et al., 2017; Khetkorn et al., 2017; Petre being the decrease of ethanol productivity, the loss of a high quantity of
et al., 2020; Surendra et al., 2015), nowadays a special attention is substrate, and even the reactor instability (Daugulis et al., 1997;

* Corresponding author.
E-mail address: dansel@automation.ucv.ro (D. Selişteanu).

https://doi.org/10.1016/j.biortech.2021.124836
Received 5 January 2021; Received in revised form 7 February 2021; Accepted 8 February 2021
Available online 15 February 2021
0960-8524/© 2021 Elsevier Ltd. All rights reserved.
E. Petre et al. Bioresource Technology 328 (2021) 124836

Quintero et al., 2009). developed algorithms is widely examined by using realistic simulations.
To improve the production of bioethanol via fermentation, the Section 4 concludes the paper.
research in this domain was oriented both to the developing of suitable
mathematical models that describe this complex bioprocess (Daugulis 2. Materials and methods
et al., 1997; Herrera et al., 2016; Quintero et al., 2009), and to design of
appropriate monitoring and control strategies (Ajbar and Ali, 2017; 2.1. Fermentation process modelling issues
Fernandez et al., 2020; Quintero et al., 2009; Petre and Selişteanu, 2014;
Petre et al., 2019) able to assure the stability as well as the productivity The process considered in this study is a continuous alcoholic
increase. Still, the design of effective control methods for these processes fermentation process of Saccharomyces cerevisiae (yeast) bacteria used to
is hampered by the fact that the bioprocess models comprise uncertain produce ethanol (Flores-Hernández et al., 2018; Kumar et al., 2019;
kinetic parameters and yield coefficients (Daugulis et al., 1997; Herrera Pachauri et al., 2017, 2018). The fermentation process takes place in a
et al., 2016; Quintero et al., 2009). Much more, since the fermentation bioreactor equipped with a cooling jacket, schematized in Fig. 1, where
process is an exothermic one, a problem that ought to be solved is the Fin, CSin and Tin are flow rate, the concentration, and the temperature of
regulation of the temperature, which must be maintained at an optimum input substrate (glucose), Fout is the bioreactor outflow rate, Fj and Tinj
value (Kumar et al., 2019; Phisalaphong et al., 2006). Another chal­ are the flow rate and the temperature of the input cooling agent. CS, CX,
lenging problem is to get accurate and reasonably priced sensors for real CE represent the concentrations of substrate, biomass (bacteria) and
time measurement of the bioprocess variables utilized in the control ethanol inside the bioreactor as well as at its output, while Tr denotes the
systems (Daugulis et al., 1997; Herrera et al., 2016; Quintero et al., temperature of the mixed culture medium. Fj is the temperature of the
2009). Despite of all these difficulties, numerous control methods were cooling agent inside the jacket, but also at its output. The model of this
reported: PID techniques, especially for temperature regulation (Kumar process is expressed a set of differential and algebraic equations (Flores-
et al., 2019; Pachauri et al., 2017, 2018), fuzzy control (Flores- Hernández et al., 2018; Kumar et al., 2019; Pachauri et al., 2017, 2018).
Hernández et al., 2018), nonlinear controllers (López-Pérez et al., 2015),
nonlinear predictive techniques (Ajbar and Ali, 2017; Herrera et al., ĊX (t) = μX (⋅)CX CS /(KS + CS )e− KE CE
− Fout /V⋅CX (1)
2016), robust-adaptive strategies (Petre and Selişteanu, 2014; Petre
et al., 2019), etc. Most of these control methods require the use of ĊE (t) = μE CX CS /(KS1 + CS )e− KE1 CE
− Fout /V⋅CE (2)
software sensors (Bastin and Dochain, 1990; Petre et al., 2018; Seliș­
1 C e− KE CE 1 C e− KE1 CE Fin Fout
teanu et al., 2010) (combinations of hardware sensors and software ĊS (t) = − μ (⋅)CX S − μ CX S + CSin − CS (3)
YSX X KS + CS YSE E KS1 + CS V V
estimators) used to estimate the unavailable states needed in control
design and also to estimate the uncertain kinetics.
Fin Fout ΔHr ⋅rO2 KT AT (Tr − Tj )
This study addresses the design of innovative advanced control Ṫ r (t) = (Tin + 273) − (Tr + 273) + −
V V 32⋅ρr ⋅Cheat,r V⋅ρr ⋅Cheat,r
schemes implemented for a continuous fermentation bioprocess used to
(4)
produce ethanol. The fermentation takes place in a CSTR under
isothermal conditions. This bioprocess requires the design of a reliable Fj KT AT (Tr − Tj )
control scheme that could ensure a large production of alcohol and a Ṫ j (t) = (Tinj − Tj ) + (5)
Vj Vj ⋅ρj ⋅Cheat,j
minimal quantity of residual substrate. Moreover, the control system
must avoid the possible reactor instability (caused by the improper
ĊO2 (t) = KL a⋅(COsat2 − CO2 ) − rO2 − Fout /V⋅CO2 (6)
operation), or the unavoidable disturbances appeared due to the load
rate variations (Fernandez et al., 2020; Quintero et al., 2009). To
where CO2 denotes the dissolved oxygen concentration inside the
maintain the temperature inside the bioreactor, a reliable algorithm able
bioreactor and Csat
O2 is the oxygen saturation (equilibrium) concentration.
to regulate the bioreactor temperature at some suitable values is also
The consumption rate of the oxygen during the biomass growth is
developed. The estimation and control algorithms are derived taking
denoted rO2 , and V and Vj are the bioreactor and the jacket volumes.
into account that the bioprocess runs under concrete, realistic
conditions. Remark 1. It must be noted that for a CSTR, Fin = Fout, so that the
In the process of control design, it is considered that the concentra­
tion of the influent substrate is totally unknown (as in real situations),
and the growth kinetics is nonlinear, time varying and unknown. Each
proposed adaptive scheme includes a linearizing control law (LCL), in
which the unknown variables are replaced by their on-line estimates
calculated by using state observers and parameter estimators (which
estimate in real-time the incompletely known kinetic rates). In order to
reconstruct the unknown input concentration from the measurements of
alcohol and internal substrate concentrations, a sliding mode observer
(SMO) is derived. The design of state observers and parameter estima­
tors depends on the information provided on-line by the previous
observer. The efficacy and performance of estimators and of the overall
control schemes are validated via numerical simulations.
For the sake of clarity, it should be specified that the present work is
focused on the design of advanced control schemes for the ethanol
fermentation process, and on the impact of these control schemes. The
main control goal is to obtain a substantial quantity of ethanol and a low
level of the residual (unconsumed) substrate, even in the presence of
some significant uncertainties. Other specific biochemical or techno­
logical issues are not fully targeted in this work.
The paper is organized in the next manner. Section 2 describes the
fermentation process model, and a detailed presentation of the estima­ Fig. 1. Schematic of the bioreactor equipped with a cooling jacket, where the
tion and control strategies is provided. In Section 3, the behaviour of the fermentation process takes place.

2
E. Petre et al. Bioresource Technology 328 (2021) 124836

bioreactor volume V is constant. In the next approaches we will use this Table 1
remark, i.e., Fout = Fin. Process and kinetic parameters – description and values.

In (1), the maximum specific growth rate μX (⋅) depends on Tr as (an Parameter Description Value

Arrhenius type law) (Kumar et al., 2019; Pachauri et al., 2018): A1 Exponential factors in Arrhenius law 9.5 × 108
( ) ( ) A2 2.55 × 100.3
Ea1 Ea2 AT Heat transfer area 1 m2
μX = A1 exp − − A2 exp − (7) Cheat,r Heat capacity of culture medium 4.18 J/g/K
R(Tr + 273) R(Tr + 273)
Cheat,j Heat capacity of the cooling agent 4.18 J/g/K
The expression of Csat
O2 from (6) depends on Tr and pH (in fact, the Ea1 Activation energies 55,000 J/mol
Ea2 22,000 J/mol
global effect of ionic strength ΣHk Ik ) as (Kumar et al., 2019; Pachauri
HOH Specific ionic constant of OH 0.941
et al., 2018): HH Specific ionic constant of H − 0.774
HCO3 Specific ionic constant of CO3 0.485
COsat2 = (14.16 − 0.394Tr + 0.00772Tr2 − 0.000064Tr3 )⋅10− ΣHk Ik
(8)
HCl Specific ionic constant of Cl 0.844
HMg Specific ionic constant of Mg − 0.314
where ΣHk Ik is calculated as (Kumar et al., 2019; Pachauri et al., 2018): HCa Specific ionic constant of Ca − 0.303
HNa Specific ionic constant of Na − 0.55
mNaCl MNa mCaCO3 MCa mMgCl2 MMg
ΣHk Ik =0.5HNa +2HCa +2HMg Ii Ionic strengths –
MNaCl V MCaCO3 V MMgCl2 V KLa0 O2 mass-transfer coefficient 38 h− 1
( )
mNaCl mMgCl2 MCl mCaCO3 MCO3 (9) KO2 O2 consumption constant 8.886 mg/l
+0.5HCl +2 +2HCO3
MNaCl MMgCl2 V MCaCO3 V KE Growth inhibition constant by ethanol 0.139 g/l
− pH − (14− pH) KE1 Fermentation inhibition constant by ethanol 0.07 g/l
+0.5HH 10 +0.5HOH 10 KS Substrate constant for growth 1.03 g/l
KS1 Substrate constant for production 1.68 g/l
In (4) and (6) the oxygen consumption rate is expressed as (Kumar KT Heat transfer coefficient 3.6 × 105 J/h/m2/K
et al., 2019; Pachauri et al., 2018): mNaCl Quantity of NaCl 500 g
mCaCO3 Quantity of CaCO3 100 g
rO2 (t) = μO2 /YO2 CX CO2 /(KO2 + CO2 )⋅1000 (10)
mMgCl2 Quantity of MgCl2 100 g
In addition, in (6), KLa is expressed as: MNa Molecular mass of Na 23 g/mol
MCa Molecular mass of Ca 40 g/mol
KL a = KL a0 ⋅1.024(Tr − 20)
(11) MMg Molecular mass of Mg 24 g/mol
MCl Molecular mass of Cl 35.5 g/mol
For the sake of clarity, the definition of process and kinetic param­ MCO3 Molecular mass of CO3 60 g/mol
eters involved in model (1)–(11), together with their measurement units MNaCl Molecular mass of NaCl 58.5 g/mol
are presented in Table 1 (Kumar et al., 2019). It should be noticed that MCaCO3 Molecular mass of NaCO3 100 g/mol
the process model was validated in (Kumar et al., 2019), and several MMgCl2 Molecular mass of NaMgCl2 95.2 g/mol
additional modelling aspects are widely presented in this work. R Universal gas constant 8.31 J/mol/K
The Eqs. (1)–(11) show that the presented process is a very complex YSE Ethanol produced yield 0.435 g/g
YSX Biomass produced yield 0.607 g/g
one including strong interactions between growth process, oxygen V Bioreactor volume 1000 l
consumption and temperature. To accomplish the control goal i.e., to Vj Cooling jacket volume 50 l
obtain a substantial quantity of ethanol and a low level of the residual YO2 O2 consumed/unit biomass produced 0.970 mg/mg
(unconsumed) substrate, even in the presence of some significant un­ ΔHr Fermentation reaction heat 518 kJ/mol O2
certainties, it is essential to maintain the substrate concentration CS in μO2 Maximum specific O2 consumption rate 0.5 h− 1
1
ref Maximum specific fermentation rate 1.79 h−
the bioreactor at a required reference CS ∈ R+ . It must note that an μE
Density of jacket liquid 1000 g/l
efficient operation of the reactor can be achieved only if the temperature
ρj
Density of fermentation medium 1080 g/l
is maintained at an optimum value. Since the temperature in the
ρr

bioreactor may to increase due to the exothermic character of the


growth reaction, it results that it is necessary to maintain the tempera­
( )
ture Tr at a suitable value. As control inputs we will use the substrate V 1 C e− KE CE 1 CS e− KE1 CE
Fin (t) = η1 (CSref − CS ) + μX CX S + μE C X
influent flow rate Fin and the flow rate of input cooling agent Fj. CSin − CS YSX KS + CS YSE KS1 + CS
(12)
2.2. Design of control strategies
guarantees the stability of the closed-loop system described by a first
In this section, for the above presented bioethanol fermentation order linear stable ordinary differential equation:
process, advanced reliable control schemes are developed under some ( )
( )
realistic operation conditions. (13)
ref
ĊS − ĊS + η1 ⋅ CSref − CS = 0,

2.2.1. Exact linearizing control law where η1 > 0 is a design parameter.


First, it is considered that the whole process that takes place inside The LCL (12) assures a linear dynamic for the tracking error
the bioreactor is completely known. Also, the temperature and pH values ref
CS = CS − CS provided by C
̃ ̃˙ S = − η CS , which for η > 0 has an expo­
are known and are chosen so that to assure a proper operation of the 1 1

bioreactor. This means that the dynamical model (1)–(11) is entirely nential stable point at CS = 0.
̃
known, and furthermore all the state variables and inflow rates are Regarding the temperature control in this type of reactors, most
measurable (on-line). In fact, this (unrealistic) case will be considered as proposed techniques used different variants of PID algorithms (Kumar
a benchmark (ideal) case and will be used as backbone to design et al., 2019; Pachauri et al., 2017, 2018), but also fuzzy control (Flores-
advanced adaptive control schemes under some realistic conditions. Hernández et al., 2018). In this study we will use a strategy based on the
As the Eq. (3) of the model (1)–(11) has the relative order equal to linearization method. The problem is not very simple, because the model
one, then, the following exact linearizing control law (LCL) (1)–(11) does not contain any equation that assures a direct relation
between the input Fj and the output Tr. To avoid the use of an input-

3
E. Petre et al. Bioresource Technology 328 (2021) 124836

output model of second order (difficult to be implemented), we will


̂˙ E (t) = μE C CS Fin
exploit some practical aspects of the bioreactor. More exactly, tacking in C ̂X e− KE1 CE
− ̂ E)
CE + ω1 (CE − C (18)
KS1 + CS V
account that the cooling jacket volume Vj is much small than the
bioreactor volume V, it can be considered that the dynamics of the
where C ̂ X denotes the estimation of CX. Then, this estimation will be
cooling process is very fast and then the singular perturbation technique
calculated by using the next relation:
(Bastin and Dochain, 1990) can be used. Then, from the steady state of
the cooling process, i.e., Ṫ j (t) = 0, we obtain: ̂˙ X (t) = γ 1 ⋅μE CS /(KS1 + CS )e−
C KE1 CE ̂ E)
(CE − C (19)

Tr − Tj =
ρj ⋅Cheat,j
(Tinj − Tj )⋅Fj (14) Thus, the Eqs. (18) and (19) constitute a novel state observer for the
KT AT estimation of unknown CX, in which ω1 and γ1 are tuning parameters.
The substitution of Tr − Tj from (14) into (4) leads to: The selection of these parameters can be achieved by a trial-and-error
method (for example, like in the case of an observer-based estimator).
Ṫ r (t) =
Fin
(Tin − Tr ) +
ΔHr rO2

ρj Cheat,j
(Tinj − Tj )⋅Fj , (15) Estimation of CSin. To estimate the unknown influent concentration
V 32 ρr Cheat,r V ρr Cheat,r CSin, a suitable sliding mode observer (SMO) (Petre et al., 2019, 2020;
Sbarciog et al., 2018) will be derived, by using the generalized super-
which represents a first order input-output model. twisting observer designed by Moreno (2013). To design the particular­
Then, the following exact LCL ized SMO, the nonlinearities from the first three model equations will be
ρ Cheat,r V
(
ΔHr rO2 Fin
) eliminated. To do this, the variable Θ is defined as a linear combination
Fj (t) = r η2 (Trref − Tr ) − − (Tin − Tr ) of CX, CE and CS, as:
ρj Cheat,j Tinj − Tj 32 ρr Cheat,r V
(16) Θ = (1/YSX )CX + (1/YSE )CE + CS . (20)

assures for the controlled cooling process a dynamic behaviour given by By using the model Eqs. (1)–(3), the dynamic of variable Θ is given
the next first order linear stable ODE: by the next linear ODE:

Θ̇(t) = − (Fin /V) Θ + (Fin /V) CSin . (21)

̂˙ = − (Fin /V) Θ + (Fin /V) C


Θ ̃
̂ Sin − λΘ κ1 φΘ1 (Θ),
˙̂ ̃
C Sin = − λ2Θ κ2 (Fin /V) φΘ2 (Θ), λΘ , κ1 , κ2 > 0, (22)

Based on (21), the sliding mode observer, able to provide the on-line
ref
(Ṫ r − Ṫ r ) + η2 ⋅(Trref − Tr ) = 0, (17) estimation C
̂ Sin is given by:

where η2 > 0 is a design parameter. where Θ ̂ − Θ, and:


̃ =Θ

̃ 1/2 sign (Θ)


̃ = λ1 |Θ| ̃ ,
̃ + λ2 Θ
2.2.2. Novel adaptive control strategies φΘ1 (Θ)
Tacking in account that the whole knowledge concerning the bio­ 2 (23)
process assumed before is not realistic, it is necessary to consider real­ ̃ = λ2 sign (Θ)
φΘ2 (Θ) ̃ + 3λ1 λ2 |Θ|
̃ 1/2 sign (̃
ψ ) + γ22 ψ̃ ,
2 2
istic conditions related to process inputs, state variables and reaction
kinetics, as follows: with λ1 , λ2 > 0. The SMO (22) and (23) is an algorithm independent of
the unknown reaction kinetics and requires five tuning positive pa­
• the influent substrate concentration CSin is unknown and time- rameters λΘ , κ1 , κ2 , λ1 , λ2 , whose values can be chosen, usually, by a
varying. trial-and-error method.
• the process state variable CX is not measurable.
Remark 2.. It should be noted that, practically, to calculate Θ, the un­
• the biomass growth rate is incompletely unknown and time-varying
because of uncertainties related to the maximum specific growth rate known CX from (20) will be substituted by its estimation C
̂ X given by the
μX and of unavailability of CX. observer (18) and (19).
• the variables CS, CE and CO2 are on-line available. Estimation of specific reaction rate μX CX . As it can be seen, the
• the temperatures Tin and Tinj are time-varying. exact LCL (15) involves both the unmeasurable state CX and the uncer­
• all other process parameters are known. tain maximum specific growth rate ψ = μX CX . Since the unavailable
state CX can be replaced by its on-line estimations, it results that a
In these practical conditions, a novel adaptive multivariable control suitable estimator is needed. This estimator should be able to reconstruct
scheme will be built. In fact, the exact LCLs (12) and (16) will be used; the unknown information concerning μX . Considering that CX is not
these LCLs become two adaptive control laws if all the uncertain and available, then the specific reaction rate μX CX = ψ will be treated as an
unavailable variables are replaced with their on-line estimates provided unknown parameter. Thus, the linearizing control law (12) becomes:
by observers and parameter estimators. ( )
Estimation of CX. By using the fact that CS and CE are on-line F in (t) =
V 1 C e− KE CE
η1 (CSref − CS ) + ψ (⋅) S
1
+ μE C X
CS e− KE1 CE
measurable, to estimate the unmeasurable state CX a novel state CSin − CS YSX KS + CS YSE KS1 + CS
observer is proposed. This is based on the dynamics of CE from (2). If C
̂E (24)
is the influence of unknown CX upon the variable CE, its dynamics can be The unknown parameter ψ (t) is substituted by the on-line estimate
rewritten as ψ̂ (t) computed with a parameter estimator implemented for CS. By using
(3), where μX (⋅)CX = ψ (t), the proposed on-line estimation algorithm for

4
E. Petre et al. Bioresource Technology 328 (2021) 124836

̂˙ S (t) = − 1 C e− KE CE 1 ̂ X CS e
− KE1 CE
C ψ̂ (t) S − μ C ̂ Sin − CS ) + ω2 (CS − C
+ (Fin /V)⋅( C ̂ S ), (25)
YSX KS + CS YSE E KS1 + CS

ψ (t) is a modified observer-based estimator (Bastin and Dochain, 1990)


given by:
The schematic of the resulting adaptive controlled closed-loop bio­
process is shown in Fig. 2. The control system contains two control
loops, one for the bioreactor temperature and one for the substrate
1 C e− KE CE
ψ̂˙ (t) = − ψ̂ (t) S ̂ S)
⋅γ ⋅(CS − C (26) concentration. Both control loops are based on adaptive approaches,
YSX KS + CS 2
with exact linearizing control laws as backbone of the control structure
In (25) and (26), the estimation C ̂ Sin is provided by the SMO (22), and containing state observers and parameter estimators.
C X is provided the observer (18) and (19) and the estimator’s design
̂
3. Results and discussion
parameters ω2 and γ 2 are selected to ensure its stability and tracking
properties. In fact, the stability is guaranteed if the design parameters
In the present section, the behaviour of the adaptive control laws
are selected as ω2 < 0 and γ2 > 0 (see (Petre et al., 2019)).
(27) and (28), by comparison to the exact LCLs (12) and (15) (consid­
The bioprocess adaptive control structure. After the design of
ered as benchmark) is tested and analysed by using realistic simulation
these algorithms, the overall adaptive control structure is in fact a
scenarios.
combination of the state observer (18) and (19), the SMO (22) and (23),
All simulations were achieved by using the dynamic process model
the reaction rate estimator (25) and (26) and the control law (24)
given by the Eqs. (1)–(11). The values chosen for the bioprocess yield
written as:
and kinetic coefficients are presented in Table 1 and are based on the
( )
V 1 CS e− KE CE 1 ̂ X CS e
− KE1 CE
model developed by Kumar et al. (2019).
Fin (t) = η1 (CSref − CS ) + ψ̂ + μE C .
̂ Sin − CS
C YSX KS + CS YSE KS1 + CS The operating conditions of the bioreactor are specified by Kumar
(27) et al. (2019) and Pachauri et al. (2017) and are summarized in Table 1.
For example, the volume of the bioreactor V (in fact the total volume of
The temperature adaptive control structure. Regarding the
reaction medium) is 1000 l, and the volume of the jacket is Vj = 50 l.
bioreactor temperature control, one can see that the exact LCL (16) in­
Other steady-state operating parameters are the inlet flow temperature
cludes the oxygen consumption rate rO2 , in whose structure CX is un­
T0in = 25 ◦ C, the nominal temperature of the input cooling agent T0inj =
measurable. Then, this control law becomes an adaptive one if the 15 ◦ C, the pH = 6, and the oxygen mass-transfer coefficient KLa0 = 38
unknown CX is substituted by the on-line estimate C ̂ X given by the
h− 1. Also, other nominal operating parameters for the bioreactor in
observer (18) and (19). The adaptive control algorithm takes the form: steady state (without control action), are the input and output flows

( )
ρr Cheat,r V ΔHr μO ̂ C O2 Fin
Fj (t) = η (T ref − Tr ) − × 1000 2 C X − (Tin − Tr ) (28)
ρj Cheat,j Tinj − Tj 2 r 32 ρr Cheat,r YO2 KO2 + CO2 V

Fin = 51 l/h, Fout = 51 l/h, and the flow rate of the cooling agent Fj =
18 l/h (Kumar et al., 2019).
Since this process can easily destabilized by an inadequate operation
(Daugulis et al., 1997; Herrera et al., 2016; Quintero et al., 2009), the
designed control schemes must avoid the process instability, assure a
substantial production of alcohol, a minimal quantity of residual sub­
strate and, at the same time, must regulate the temperature inside
bioreactor at suitable values.
The performance of the adaptive controlled system, by comparison
to the benchmark (ideal case – the bioprocess is fully known) is analysed
in the next conditions:

• the variable CX is not measurable.


• the influent substrate concentration CSin is time-varying as is shown
in Fig. 3 and is presumed totally unknown;
• the biomass growth rate is incompletely unknown and time-varying.
• the states CS, CE and CO2 are on-line accessible.
• all the other bioprocess parameters are known constants.

Much more, Tin and Tinj are supposed time-varying as:

Tin (t) = Tin0 (1 + 0.25sin(πt/12)), (29)


Fig. 2. Structure of the double control loops for the adaptive closed-loop
controlled system.

5
E. Petre et al. Bioresource Technology 328 (2021) 124836

Fig. 3. Evolution of the main bioprocess variables for the adaptive control system – the first simulation scenario: (a) Time profile of output CS; (b) Time evolution of
output Tr; (c) Time profile of control input Fin; (d) Time profile of control input Fj; (e) Estimation of the unknown variable CSin; (f) Evolution of the ethanol con­
centration CE.

0
Tinj (t) = Tinj (1 + 0.5sin(πt/20)), (30) the dynamics considered in the present simulation.
First case scenario. The behaviour of the closed-loop system with
and the maximum specific consumption rate μO2 is also time varying, as the adaptive laws (27) and (28) is shown in Fig. 3. The graphics in
follows: μO2 (t) = μ0O2 (1 − 0.05cos(πt/20)). Fig. 3a and b represent the behaviour of controlled variables CS and Tr,
In (29) and (30) Tin0 = 25 ◦ C and Tinj0 = 15 ◦ C. while the graphics plotted in Fig. 3c and d are related to the control
Some discussions about the time factor in the alcoholic fermentation inputs Fin and Fj. In this scenario, the setpoints of the two controlled
ref
process can be done. Depending on the type of alcoholic fermentation, variables were fixed at two constant values, respectively CS = 35 mg/l
several time ranges were reported in industrial practice and research and Tref
r = 31.5 C.

works. For example, the continuous alcoholic fermentation process of By analysing Fig. 3a and b it results that the values of both temper­
Saccharomyces cerevisiae (yeast) bacteria used to produce ethanol, ana­ ature Tr and substrate concentration CS inside the bioreactor are main­
lysed in several works (Kumar et al., 2019; Nagy, 2007; Pachauri et al., tained at their references, despite the large variations of CSin, Tin and Tinj.
2017), evolves on a time span of about 400 h. Other processes, like the The graphics from Fig. 3c and d show that both the input flow rate Fin
continuous alcoholic fermentation from Zymomonas mobilis described in and input flow rate Fj are finite and remain in some physical limits.
works such as (Geng et al., 2020; Quintero et al., 2009) have a faster The shape of the estimates of unknown CSin produced by the SMO
evolution of about 50–150 h. Of course, the relevant changes in the (22) and (23) is given in Fig. 3e, from which it is clear that the SMO
dynamics of the process concentrations depend on the particular oper­ works well even if CSin variations are large and the dynamic is rapid. The
ating conditions and could be slower or faster than the time interval and evolution of ethanol concentration is presented in Fig. 3f. The obtained

6
E. Petre et al. Bioresource Technology 328 (2021) 124836

Fig. 4. Time profiles of the outputs and control inputs for the adaptive control system – the second simulation scenario: (a) Evolution of the substrate concentration
CS; (b) Time profile of the reactor temperature Tr; (c) Time profile of control input Fin; (d) Time profile of control input Fj.

values for this concentration are similar with the values reported in The behaviour of the controlled closed-loop system using the adap­
other works (see, for example, Phisalaphong et al., 2006). It should tive laws (27) and (28) under this scenario is shown in Figs. 4 and 5. The
remind here that the control laws use the estimated values of the un­ graphics in Fig. 4a and b show the behaviour of controlled variables CS
measurable state CX calculated by the state observer (18) and (19) and and Tr, while the graphics plotted in Fig. 4c and d represent the control
the estimated values of the unknown parameter ψ (t) given by the OBE inputs Fin and respectively Fj. Fig. 4a and b shows that the concentration
(25) and (26). of the substrate CS tracks its reference, and the temperature Tr inside the
The gains η1 and η2 of control laws (27) and (12) and respectively of bioreactor is maintained at its setpoint, despite the large variations of
control laws (28) and (16) have been set to: η1 = 2.5, η2 = 2.5. The CSin, Tin and Tinj. The graphics plotted in Fig. 4c and 4d show that both
tuning parameters of the state observer (18) and (19) and of the OBE the input flow rate Fin and input flow rate Fj are finite and remain in
(25) and (26) have been set to: ω1 = − 0.5, γ 1 = 0.75, ω2 = − 5, γ2 = some physical limits.
25. For the SMO (22) and (23), the tuning parameters were chosen as: The estimated values of the unmeasurable state CX calculated by the
λΘ = 2.5, κ1 = 0.075, κ2 = 2.5, λ1 = 5, λ2 = 25. state observer (18) and (19) are depicted in Fig. 5a, indicating a proper
From all the profiles depicted in Fig. 3 it can be assessed that the behaviour of this observer. The shape of the estimates of unknown CSin
behaviour of adaptive controlled bioprocess is suitable, very close to the produced by the SMO (22) and (23) is given in Fig. 5b. The estimated
behaviour from the benchmark case i.e., when the LCLs (12) and (16) values of the unknown parameter ψ(t) given by the OBE (25) and (26)
are used (known process model). are presented in Fig. 5c. There is a time variance between actual
Second case scenario. To increase the quantity of ethanol we pro­ parameter and estimated parameter (the specific reaction rate) depicted
ceeded to the increase of the substrate consumption. That is why in this in Fig. 5c. This behaviour is due to the quite slow convergence of the
scenario the simulations were carried out considering that the reference modified OBE used for the estimation of the specific reaction rate, and to
ref
CS is planned to have a decreasing evolution, in constant steps. Thus, the different initial conditions of the actual and estimated variables. The
the setpoint values are as follows: estimation algorithm operates in harsh conditions since it uses the
⎧ influent substrate concentration estimates provided by the SMO and



37.5 g/l, for t ∈ [0, 100]h moreover the measurements of some variables are vitiated by mea­
ref
CS =
35 g/l, for t ∈ [100, 200]h
(31) surement noises. However, even in these conditions, the estimation al­

⎪ 32.5 g/l for t ∈ [200, 300]h gorithm works and provides useful estimates of the reaction rate, used in

30 g/l for t ∈ [300, 400]h
the adaptive control law.
All the other operating conditions of the bioreactor remained the The evolution of ethanol concentration for the second simulation
same as in the first scenario. Note also that values of all gains and tuning scenario is presented in Fig. 5d. It can be remarked that the ethanol
parameters are the same as in the first scenario. Furthermore, the concentration increases by comparison with its value obtained in the
measurements of variables CS and CE are vitiated by measurement noises first scenario. Here it is worthy to mention the correlation between the
(an additive white noise was considered, with zero average, 5% of its residual substrate evolution and the time profile of ethanol concentra­
nominal values). tion, which can be seen for the first case in Fig. 3a and f, and respectively

7
E. Petre et al. Bioresource Technology 328 (2021) 124836

Fig. 5. Evolution of some estimated bioprocess variables and of the ethanol concentration, the second simulation scenario: (a) Estimation of the unknown variable
CX; (b) Estimation of the unknown variable CSin; (c) Time profile of the estimates of ψ ; (d) Evolution of the ethanol concentration CE.

for the second case in Figs. 4a and 5d: if the residual substrate level is CRediT authorship contribution statement
low, the amount of alcohol is greater.
Overall, this is a multivariable control scheme, containing a control Emil Petre: Conceptualization, Methodology, Investigation, Soft­
loop for the substrate concentration and a control loop for the reactor ware, Writing - original draft. Dan Selişteanu: Supervision, Formal
temperature, an important factor for the efficient operation of the bio­ analysis, Validation, Writing - original draft. Monica Roman: Visuali­
process. One can remark the proper regulation and the ability of adap­ zation, Software, Funding acquisition, Writing - review & editing.
tive control laws to maintain the bioprocess output CS close to its
reference, despite the high time variation of CSin and of the process
parameters. The behaviour of the controlled system remains proper even Declaration of Competing Interest
if the measurements of variables CS and CE are vitiated by measurement
noises. The authors declare that they have no known competing financial
The simulation experiments showed a good behaviour of the pro­ interests or personal relationships that could have appeared to influence
posed adaptive control system; however, the overall efficacy of the the work reported in this paper.
control scheme needs additional test with experimental laboratory
setups. Acknowledgement

4. Conclusions This work was partially supported by European Regional Develop­


ment Fund Competitiveness Operational Program, project TISIPRO, ID:
An innovative adaptive control scheme was developed and applied P_40_416/105736 (2016–2021).
for an ethanol production bioprocess. Both control loops for substrate
concentration and temperature used linearizing control laws, coupled
References
with estimators ensuring the whole functionality of control system in
harsh, realistic operating conditions (unknown influent substrate con­ Ajbar, A., Ali, E., 2017. Study of advanced control of ethanol production through
centration, time-varying and uncertain kinetics, noisy measurements). continuous fermentation. J. King Saud Univ. Eng. Sci. 29 (1), 1–11.
Despite these uncertainties and disturbances, the adaptive control Bastin, G., Dochain, D., 1990. On-line Estimation and Adaptive Control of Bioreactors.
Elsevier, NY.
structure behaved well, and the control goal was accomplished: a sub­
Daugulis, A.J., McLellan, P.J., Li, J., 1997. Experimental investigation and modeling of
stantial quantity of ethanol and a low level of residual substrate. The oscillatory behaviour in the continuous culture of Zymomonas mobilis. Biotechnol.
reported simulation results are encouraging, and the proposed control Bioeng. 56 (1), 99–105.
Fernandez, M.C., Pantano, M.N., Serrano, E., Scaglia, G., 2020. Multivariable tracking
structure can be adapted to other types of fermentation processes.
control of a bioethanol process under uncertainties. Math. Prob. Eng. 2020,
8263690.
Flores-Hernández, A., Reyes-Reyes, J., Astorga-Zaragoza, C., Osorio-Gordillo, G., García-
Beltrán, C., 2018. Temperature control of an alcoholic fermentation process through
the Takagi-Sugeno modelling. Chem. Eng. Res. Des. 140, 320–330.

8
E. Petre et al. Bioresource Technology 328 (2021) 124836

Gaida, D., Wolf, C., Bongards, M., 2017. Feed control of anaerobic digestion processes for Pachauri, N., Singh, V., Rani, A., 2018. Two degrees-of-freedom fractional-order
renewable energy production: a review. Renew. Sustain. Energy Rev. 68 (2), Proportional–Integral – derivative-based temperature control of fermentation
869–875. process. J. Dyn. Syst. Meas. Control 140 (7), 071006.
Geng, B.Y., Cao, L.Y., Li, F., Song, H., Liu, C.G., Zhao, X.Q., Bai, F.W., 2020. Potential of Pandey, A., Chang, J.-S., Soccol, C.R., Lee, D.J., Chisti, Y. (Eds.), 2018. Biomass, Biofuels,
Zymomonas mobilis as an electricity producer in ethanol production. Biotechnol. Biochemicals: Biofuels from Algae, second ed. Elsevier (S&T) NY.
Biofuels 13 (36), 1–11. Petre, E., Selişteanu, D., 2014. A robust-adaptive control strategy for a continuous
Herrera, W., Rivera, E., Alvarez, L.A., Plazas Tovar, L., Tamayo Rojas, S., Yamakawa, C., alcoholic fermentation process. In: Proceedings of the 18th Int. Conf. System Theory,
Bonomi, A., Maciel Filho, R., 2016. Modelling and control of a continuous ethanol Control and Comp. (ICSTCC), 2014, Sinaia, Romania, pp. 424–429.
fermentation using a mixture of enzymatic hydrolysate and molasses from Petre, E., Selișteanu, D., Roman, M., 2018. Nonlinear robust adaptive control strategies
sugarcane. Chem. Eng. Trans. 50, 169–174. for a lactic fermentation process. J. Chem. Technol. Biotechnol. 93 (2), 518–526.
Khetkorn, W., Rastogi, R.P., Incharoensakdi, A., Lindblad, P., Madamwar, D., Pandey, A., Petre, E., Selişteanu, D., Roman, M., 2020. Control schemes for a complex biorefinery
Larroche, C., 2017. Microalgal hydrogen production – a review. Bioresour. Technol. plant for bioenergy and biobased products. Bioresour. Technol. 295 (122245), 1–12.
243, 1194–1206. Petre, E., Șendrescu, D., Selişteanu, D., Roman, M., 2019. Estimation based control
Kumar, M., Prasad, D., Giri, B.S., Singh, R.S., 2019. Temperature control of fermentation strategies for an alcoholic fermentation process with unknown inputs. In:
bioreactor for ethanol production using IMC-PID controller. Biotechnol. Rep. 22 Proceedings of the 23rd Int. Conf. System Theory, Control and Comp. (ICSTCC),
(e00319), 1–10. 2019, Sinaia, Romania, pp. 224–229.
Kumar, V., Nanda, M., Joshi, H.C., Singh, A., Sharma, S., Verma, M., 2018. Production of Phisalaphong, M., Srirattana, N., Tanthapanichakoon, W., 2006. Mathematical modeling
biodiesel and bioethanol using algal biomass harvested from freshwater river. to investigate temperature effect on kinetic parameters of ethanol fermentation.
Renew. Energy 116, 606–612. Biochem. Eng. J. 28, 36–43.
López-Pérez, P.A., Cuevas-Ortiz, F.A., Gómez-Acata, R.V., Aguilar-López, R., 2015. Quintero, O., Amicarelli, A.A., Scaglia, G., Sciascio, F., 2009. Control based on numerical
Improving bioethanol production via nonlinear controller with noisy measurements. methods and recursive bayesian estimation in a continuous alcoholic fermentation
Chem. Eng. Commun. 202 (11), 1438–1445. bioprocess. Bioresour. Technol. 4 (4), 1372–1395.
Moreno, J.A., 2013. On discontinuous observers for second order systems: properties, Sbarciog, M., Giovannini, G., Chamy, R., Vande Wouwer, A., 2018. Control and
analysis and design. In: Bandyopadhyay, B. (Ed.), Advances in Sliding Mode Control. estimation of anaerobic digestion processes using hydrogen and volatile fatty acids
Lect. Notes Contr. Inf. 440. Springer, Berlin, pp. 243–265. measurements. Water Sci. Technol. 78 (10), 2027–2035.
Nagy, Z.K., 2007. Model-based control of a yeast fermentation bioreactor using optimally Selișteanu, D., Roman, M., Șendrescu, D., 2010. Pseudo bond graph modelling and on-
designed artificial neural networks. Chem. Eng. J. 127 (1–3), 95–109. line estimation of unknown kinetics for a wastewater biodegradation process. Simul.
Negrea, M., Petrișor, I., Weyssow, B., 2008. On revisited models of L-H transition for Model. Pract. Theory 18 (9), 1297–1313.
tokamak plasmas. J. Optoelectron. Adv. Mater. 10 (8), 1946–1949. Surendra, K.C., Sawatdeenarunat, C., Shrestha, S., Sung, S., Khanal, S.K., 2015.
Negrea, M., 2020. Diffusion of ions in an electrostatic stochastic field and a space Anaerobic digestion-based biorefinery for bioenergy and biobased products. Ind.
dependent unperturbed magnetic field. Plasma Sci. Technol. 22 (1), 015101. Biotechnol. 11 (2), 103–112.
Pachauri, N., Rani, A., Singh, V., 2017. Bioreactor temperature control using modified Taghizadeh-Alisaraei, A., Motevali, A., Ghobadian, B., 2019. Ethanol production from
fractional order IMC-PID for ethanol production. Chem. Eng. Res. Des. 122, 97–112. date wastes: adapted technologies, challenges, and global potential. Renew. Energy
143, 1094–1110.

You might also like