Use of Raman spectroscopy for continuous monitoring and
control of lignocellulosic biorefinery processes
Shannon Ewanick, Elliott Schmitt, Rick Gustafson, Renata Bura
School of Environmental and Forest Sciences, University of Washington, Seattle, WA 98195, USA
Abstract
The production of fuels and chemicals from lignocellulosic biomass demands efficient processes to compete with
fossil fuel-derived products. Key biorefinery processes, such as enzymatic hydrolysis of cellulose and microbial
fermentation, can be monitored by advanced sensors in real time, providing information about reactant and product
concentration, contamination, and reaction progress. Spectroscopic techniques such as Raman spectroscopy provide
a means of quickly and accurately assessing many types of reaction mixtures non-destructively, in real-time and
with no costly sample preparation and analysis time. Raman spectroscopy techniques have been developed to
accurately quantify a number of compounds present in lignocellulosic processes, and methods have been developed
to overcome the presence of fluorescent compounds that can increase the spectral background. Online Raman
sensors also can provide the feedback measurements necessary for advanced process controls (APC). Specifically,
model predictive control, a common APC used extensively throughout similar processing industries, is especially
well suited for ensuring optimal production of bio-based chemicals from lignocellulosic material.
INTRODUCTION
The production of fuels and chemicals from biomass is a rapidly growing industry as the economic and
environmental effects of global warming are felt around the world. Biofuel and bio-based chemical production
processes must become more efficient to be competitive with the equivalent processes using fossil fuels. Processes
to produce fuels and chemicals from petroleum have a wealth of online analytical sensors that permit them to
operate at or near capacity with optimal process yields. This hyper efficiency is a necessary condition for
profitability in producing high volume, narrow-profit margin products such as fuels and commodity chemicals. The
need for process efficiency – and hence the need for online sensors – is especially acute in biomass fed biorefineries
due to the complexity and expense of the feedstock. Development of robust sensors for lignocellulosic biorefineries
is as critical as the research for developing the processes themselves, but has received little or no attention. Process
improvements (feedstock pretreatment, microorganisms, enzymes, etc.) will likely reduce costs in the future, but in
both the short and long term, improving the efficiency of existing operations will have the greatest effect on overall
process economics.
A multitude of products can be produced from lignocellulosic biomass such as wood, agricultural residues and
waste products; such products include biofuels, biochemicals and chemical precursors to polymers and other
molecules. The challenge in using lignocellulosic, rather than starch or sucrose based biomass, to produce fuels and
chemicals lies in their structure, which is much more resistant to breakdown by low energy methods. They require
relatively intense pretreatment methods to break down the cell wall and expose and release the sugars contained
within. Pretreatment methods can be chemical (acidic, basic or solvent), thermochemical, or physical in nature [1].
Regardless, the goal of any pretreatment method is to release sugars in hydrolysable and/or fermentable form. The
most commonly used pretreatments are thermochemical and utilize acid, producing a liquid stream containing not
only fermentable, monomeric sugars, but also lignin and a variety of other sugar and lignin degradation products.
The accompanying solid stream contains predominantly lignin and the polymer cellulose, which must be
enzymatically or chemically hydrolysed to monosaccharides before fermentation. Once monosaccharides have been
released, they can be fermented by a variety of microorganisms to generate the desired final product. Most
commonly, Saccharomyces cerevisiae is used to convert hexoses to ethanol, but other organisms have been used to
produce xylitol, acetic acid, butyric acid and many other products [2–5].
The Raman effect was discovered in 1928 and is observed upon irradiating solid, liquid or gaseous samples with
monochromatic light and measuring the inelastic scattering that is exhibited by 10-6 of the incident photons [6].
Unlike similar spectroscopic methods such as near infrared (NIR), water is virtually invisible to Raman which
allows its use in aqueous solutions [7]. In addition, Raman spectra have narrower peak widths and cover a wide
spectral range, improving the potential for analyzing complex matrices of compounds [8]. Benefits of using a
spectroscopic method are its speed, lack of sample preparation, and ability to scale from micro scale up to
production scale. The diversity of compounds present in lignocellulosic processes calls for an analytical method that
can simultaneously measure multiple compounds with high sensitivity at both high and low concentrations. Other
methods such as HPLC, UV-Vis spectroscopy, capillary electrophoresis, gas chromatography/mass spectrometry
and electrochemical methods are unable to quantify multiple compounds simultaneously and/or require destructive
and time-consuming sample preparation and analysis [8].
One of the main setbacks of utilizing advanced process control (APC) in bioprocesses has been the lack of
suitable in vitro measurement technology. With the development of chemometric methods and Raman technology,
we now are at a point where this paradigm could change. Implementation of APC technologies can increase profit
margins through increases in throughput, process stability improvement, energy consumption reduction, process
safety and regulation, and increased yield of valuable products. This has already been seen for decades in similar
industries such as in refining, petrochemical, and pulp and paper. Using online Raman measurements of key process
parameters will allow for the development of more robust system models, which in turn can be used for prediction
and optimization. Control methods, such as model predictive control, use these prediction models to ensure
regulation at an optimal setpoint.
APPLICATION OF RAMAN TO CONTINUOUS, ONLINE MEASUREMENTS
The primary advantages of using Raman, or other spectroscopic methods, over traditional analysis methods are
increased speed of analysis and a reduction in user interaction. With online Raman spectroscopy, a full spectrum can
be collected in less than a minute to provide immediate data regarding the concentration of any compounds of
interest, once appropriate models have been developed. This real-time, continuous output of information enables
application of advanced process controls. While the use of Raman spectroscopy as an analytical technique is widely
used in other disciplines, its application for continuous monitoring of fermentation of lignocellulosics is still being
developed. Small scale ethanol fermentations have been measured by Raman spectroscopy through periodic removal
and measurement of samples [9], and continuous monitoring of nano-scale reactions with in situ measurement [10]
has been conducted.
The method used for sample collection is an important consideration both for building a robust process sensor
and for improving the quality of data. Following excitation with a laser beam, the Raman scattered light is filtered
and collected on a detector. A Raman microscope or immersion probe can be used to measure individual samples, or
to continuously monitor a reaction by pumping the reaction mixture through a sampling loop. A probe can also be
immersed in the reaction solution to enable continuous remote sampling. Such probes are typically composed of a
fiber optic portion connected to a stainless steel probe fitted with a window at the tip. Compared to a standard probe
with a flat window, the use of a sapphire-tipped hemispherical ball probe tip reduces the focal length and increases
the number of Raman events that can be measured [11]. In reactions subject to high levels of fluorescent
interference, the reduced focal length results in less noise and higher quality data.
CHALLENGES OF USING RAMAN TO MEASURE LIGNOCELLULOSIC BIOPROCESSES
While Raman spectroscopy has the distinct advantage of being practical in aqueous solutions, it is subject to
interference by other means. Such interference falls into two categories – biological and chemical. Both cause an
elevated spectral background signal that can mask the desired Raman features of the spectra representing the
compounds of interest. Biological interference is present in any system containing living cells, and results from the
heterogeneous nature of biological systems. The second means of interference is the influence of compounds present
in the media, even in the absence of any cells. In lignocellulosic fermentations, these are often represented by ligninderived compounds. Lignin is made up of highly conjugated phenolic groups that fluoresce when excited by the
laser [12]. In addition to process interference, measurement noise caused by instrumentation, background light
sources, and cosmic rays must also be filtered before deducing the true spectra. Figure 1 shows how typical Raman
spectra and its constituent components are broken down.
Fig. 1: Components of simulated Raman spectra. The Spectra = Signal + Background + Noise. The goal is to reduce the spectra
to its true signal.
The impact of this background interference can be minimized in a number of ways. Cells can be physically
removed from the media prior to measurement by means of a filter [9] and chemical interference can be minimized
by using low fluorescence medium [10]. Alternatively, many different chemometric methods have been developed
to mathematically remove the fluorescent background and noise from the collected spectra and reveal the spectral
components of the compounds of interest [13,14]. In addition, the choice of excitation wavelength is important. In
solutions without the presence of fluorophores, lower wavelength visible light such as 514 nm provides high
sensitivity but greatly increased fluorescence [10,15]. Higher wavelength, near-IR excitation, may be used to
minimize fluorescence, but at the expense of peak intensity. For lignocellulosics, 785 nm wavelength is most often
utilized [8,16,17], but 993 nm and 1064 nm have also been used to further reduce the fluorescent background
[18,19].
QUANTIFICATION OF RAMAN DATA
Robust models must be constructed to use the large amounts of spectral data generated in a given process and to
convert Raman signal intensities into units of concentration. Unlike chromatographic methods, any compound
measured by Raman can present multiple peaks, many of them overlapping. In addition, some compounds are strong
scatterers, presenting high intensity peaks, while others scatter weakly. Ethanol, for example, presents a strong, wellisolated C-C bond stretching peak at 883 cm-1. This single peak allows for univariate analysis of ethanol by
measuring the peak area of a number of samples and comparing the areas to the concentration of the same samples
measured by a validation method such as HPLC [19]. This calibration model is then validated by predicting the
concentration of a second set of spectra, and comparing the predicted values to the actual values. For weaker
scatterers or complex mixtures of samples with overlapping peaks, multivariate methods must be employed.
Similarly, in the presence of a high background, even strong scattering compounds at low concentrations may
require multivariate analysis [8]. Typically these models are constructed using partial least squares (PLS) methods.
The goal of PLS is to predict a set of dependent variables, Y, from a set of independent variables, X. In the case of
developing a model for Raman spectra for lignocellulosics, the independent variables are the Raman spectra and the
dependent variables are concentration of substrates and products usually measured on an HPLC. A PLS algorithm is
then implemented to simultaneously decompose these variables into scores, T and U, loadings, P and Q, and a
matrix of error, E, with the goal of explaining as much as possible of the covariance between X and Y. A linear
relationship is established according to:
(1)
(2)
(3)
where
is the regression coefficient vector of the linear model.
USE OF RAMAN SPECTROSCOPY TO MEASURE ENZYMATIC HYDROLYSIS AND
FERMENTATION OF SUGARS TO ETHANOL
After pretreatment of lignocellulosics to break down the biomass structure, the resulting solids must be
enzymatically hydrolyzed to saccharify cellulose to glucose. Tightly controlling the progress of enzymatic
hydrolysis is important to minimize the use of costly cellulytic enzymes. The presence of significant quantities of
lignin can compromise use of Raman to monitor hydrolysis. For simple enzymatic hydrolysis of pretreated solids,
chemical interference can come from lignin present on the pretreated solids; the amount of lignin present depends on
the type of pretreatment used. Acid pretreatments leave a considerable amount of lignin on the pretreated solids,
while up to 85% of lignin can be removed using an alkaline treatment such as aqueous ammonia [8]. Any residual
lignin can leach into the enzymatic hydrolysate, increasing the Raman spectroscopic limit of detection from 4 g L -1
glucose in aqueous ammonia-pretreated corn stover to 20 g L-1 in the same material pretreated with dilute acid [8].
The conversion of glucose to ethanol is one of the oldest forms of fermentation and has been a widely studied
application of Raman spectroscopy for continuously measuring fermentation. During fermentation with
Saccharomyces cerevisiae, measurable components include ethanol, glycerol and the cell concentration, all of which
can be quantified to measure productivity, stress, and process efficiency. Shaw et al [9] used a 780 nm laser to
measure the progress of a fermentation of pure glucose by continuously pumping the fermentation solution through
a quartz cell in a Raman microscope. Samples were filtered through a 0.2 µm filter prior to analysis every 0.5 to 1
hour, removing much of the background fluorescence from the cells. A multivariate PLS model developed from this
fermentation was then used to predict glucose and ethanol concentrations in a different fermentation with a root
mean square error of prediction (RMSEP) of 8.3% and 1.8% for glucose and ethanol, respectively [9]. Ávila et al
[20] developed a multivariate PLS model based on measurements of glucose, ethanol, glycerol and cell
concentrations by 785 nm Raman taken at 15 minute intervals through a flow cell pumped continuously from the
fermentation reactor without any filtration. Their model found RMSEP values of 0.5%, 0.3%, 0.02% and 1 g L -1 for
glucose, ethanol, glycerol and cell concentration respectively. The limit of detection (LOD) can be calculated by
doubling the RMSEP.
The use of Raman spectroscopy to follow the progress of reactions using pretreated lignocellulosic material is
just beginning. Initial work by Shih and Smith [17] measured the progress of ethanol production from the enzymatic
hydrolysate of both dilute acid and ammonia pretreated corn stover using S. cerevisiae by analyzing offline aliquots
of the fermentation broth. In the pretreated corn stover hydrolysate, LOD’s for glucose and ethanol were determined
to be 8 and 6 g L-1 respectively based on quantification of peaks at 1128 cm-1 and 883 cm-1 for glucose and ethanol
respectively.
Ewanick et al. at the University of Washington continuously measured the fermentation of both pure glucose in
water and a switchgrass-derived lignocellulosic hydrolysate [16]. Measurement of both reactions was accomplished
by means of an immersion ball probe inserted in a fast loop circulating media continuously from the fermentation
reactor, with measurements taken every minute. Both reactions required mathematical removal of the fluorescent
background through use of a polynomial fitting algorithm. In the pure glucose fermentation most of the background
could be attributed to the cell biomass. The background was four times greater in the lignocellulosic fermentation
due to lignin-derived components of the switchgrass hydrolysate (Fig. 2). With proper signal processing, however,
this additional background in the lignocellulosic fermentation only increased the glucose limit of detection (LOD)
from 1 to 2 g L-1 and the ethanol LOD from 0.8 to 1.2 g L-1. Both ethanol and glucose were modeled based on
principle component analysis (PCA) of the spectrum and PLS comparison of the main spectral components to HPLC
data. These results showed for the first time that Raman could successfully follow the progress of fermentation of a
highly fluorescent lignocellulosic liquid stream [16]. By collecting data continuously and comparing it to HPLC,
data, a model was developed with a low LOD compared to previous research.
Raman spectroscopy will continue to be investigated as a means to measure the progress of both enzymatic
hydrolysis and fermentation. In particular, monitoring of simultaneous saccharification and fermentation (SSF) will
prove highly valuable to improve the efficiency of a widely used industrial process [21]. Raman monitoring of SSF
of corn mash (starch-based) has already been successfully proven and shown to be able to measure not only
monomeric glucose, but oligomeric and polymeric glucose as well as ethanol [19]. Application of these techniques
to lignocellulosic hydrolysis and SSF will prove more challenging, but holds the potential to reduce reaction time
and cost by increasing process efficiency. Future work at the University of Washington related to Raman
spectroscopic measurement of lignocellulosic biorefinery processes will focus on simultaneous quantification of
multiple sugars, particularly xylose, and on alternative microbial products such as acetic acid. The use of
pretreatment hydrolysates from diverse feedstocks and different pretreatment processes will introduce new
combinations of fluorophores that may affect the Raman measurements and require new models and techniques to
reduce the fluorescent background.
A
Glucose Ethanol
Glucose
B
Ethanol
Fig. 2. Raman spectra of pure glucose fermentation (A) and switchgrass hydrolysate (B) with the untreated spectra in the top
right inset and the spectra after polynomial fitting in the main image.
Characteristic Raman Bands for Fermentation of Lignocellulosic Components
A review of the literature regarding Raman spectroscopic methods for fermentation of bio-based substances
shows common Raman bands (cm-1) can be found. Table 1 shows several common components including: ethanol,
glucose, fructose, cellulose, and glycerol; for an exhaustive list of Raman bands for biological components the
reader is referred to [22]. Characteristic modes were assigned for several bands depending on the component. Most
researchers find a strong visible peak for ethanol at approximately 880 cm-1 representing C-C stretching. Often this
peak is used to monitor the progress of fermentation, as it is a strong peak that can be easily deciphered from the
data. While other peaks for ethanol, such as those at the region near 2900 cm-1 are often disregarded since most
organic compounds present similar bands [23]. Similarly, glucose has a strong distinguishable peak around 1120
cm-1, which in turn can be followed. However, when dealing with multiple sugar systems there exists overlapping
Raman bands from various sugars due to spectral interference between peaks making isolation difficult. For
fermentation of sugars the bands between 300 – 1500 cm-1 are rich with structural information. Raman bands in the
920 – 960 cm-1 region are strong from α-1,4 glycosidic linkages in glucose oligomers and polysaccharides [19,23].
As shown previously in Figure 2, substrate and metabolite bands can usually be identified over time. As the
fermentation progresses, the substrate is consumed and metabolites are produced showing major peaks disappearing
and being formed, but generally chemometric methods are still required to deconvolute these systems into
corresponding components of interest.
Table 1: Common Raman bands for major constituents of lignocellulosic fermentation
Component
Band (cm-1)
Characteristic Mode
Source
874, 879, 881, 883
C-C stretch
[9], [24], [20], [19], [10], [16]
1030, 1040, 1054
C-O stretch
[24], [19], [10]
1079, 1080, 1096
CH3 rocking
[24], [19], [9], [10]
2878
CH2
[10], [17]
2925, 2929
CH3 stretch
[9], [10]
2972
CH3 stretch
[10], [17]
Ethanol
Glucose
437, 448, 451
C-CO bend
[9], [8], [24]
1056, 1060, 1063
C-O bend
[9], [25], [24]
1118, 1123, 1124, 1126, 1128
C-C stretch
1336, 1347
CH2 twist
[25], [9]
626, 629
C-CO bend
[22], [24]
818, 821
C-C stretch
[22], [24]
870, 872
C-C stretch
[24], [22]
1060, 1065, 1068
C-OH bend
[22], [24], [25]
896, 899
ring stretch
[22], [26]
1098
COC stretch
[26], [22]
1120
COC stretch
[26], [22]
1049, 1055
COH stretch
[20], [22]
1109, 1110
COH stretch
[20], [22]
1254, 1257
CH2 twist
[20], [22]
[9], [8], [25], [20], [10], [16], [17], [24]
Fructose
Cellulose
Glycerol
ADVANCED PROCESS CONTROL
The goal of process control is to keep conditions of a process at a desired or optimal value in the presence of
process variations. Advanced process control (APC) extends this by handling multivariable and nonlinear systems,
with constraints if necessary, while achieving real-time optimization. APC systems are common practice throughout
various industries such as, petrochemical, refining, and pulp and paper. Benefits include increasing profit margins
through increases in throughput, process stability improvement, energy consumption reduction and increased yield
of valuable products [27]. However, the literature concerning control systems for bioreactors is relatively scarce. A
review of controls for bioreactors from Rani and Rao [28], discussed various control techniques including dynamic
programming, online adaptive control, online optimization, nonlinear control, and optimal control. One of the major
reasons for the minimal application of APC for bioreactors is lack of reliable technology for online measurements of
major chemical and biochemical concentrations [29]. Raman spectroscopy has the ability to overcome this
shortcoming, by allowing for real-time measurements of major state variables such as reactant, product, and cell
biomass concentration.
The type of control strategy implemented depends on the bioreactor operating mode (batch, fed-batch,
continuous), availability of online measurements, and the accuracy of the bioreactor model [29]. Development of
control systems that use feedback measurements from bioreactors is critical for lignocellulosic biorefineries. Over
the last decade model predictive control (MPC) or receding horizon control (RHC) has become commonplace in
many chemical and manufacturing industries. In 2003, it was reported that there were over 5,000 applications of
MPC, with majority of the applications being in refining, petrochemicals, chemicals, and pulp and paper industries
[30]. More recently, an international survey of 66 APC experts, 38 APC users, and 28 APC suppliers showed that
over two thirds of the respondents used MPC [27]. In the literature, the majority of the applications of MPC to
fermentation systems have been based on simulation. These include both fed-batch [31,32] and continuous [33–35]
systems, all of which used some form of nonlinear MPC or neural network. Though academic research continues to
develop the theory and algorithms for MPC, there exists a wealth of experience that can be applied to biorefining
from related industries. By providing a mechanism for feedback, such as with Raman technology, robust output
feedback MPC strategies for fermentation can be employed.
MODEL PREDICTIVE CONTROL
The following provides a brief overview of linear MPC, for further discussion of linear MPC see [36,37], and for
nonlinear MPC [38,39]. MPC is an online model-based control for real-time optimization of a process. A model of
the plant is used to make predictions of the future behavior of the system, and then use those predictions to
determine the optimal input to achieve a setpoint or regulation. It can include disturbance rejection, as well as linear
or nonlinear constraints within an optimal cost function. It can handle linear or nonlinear equality or inequality
constraints, multiple inputs and outputs (MIMO), and linear or nonlinear models.
Given a linear, time invariant, discrete MIMO system defined by:
(4)
(5)
where
are the state, control input, and output vectors, respectively, and
and
are appropriately-dimensioned matrices relative to
and
. A finite horizon implementation is typically based on
the solution of the following quadratic open-loop optimization problem:
(6)
Subject to process constraints
(7)
where,
(
denotes the length of the prediction horizon, and
denotes the length of the control horizon
). The weighting matrices Q and R are used to penalize setpoint tracking and manipulated input
movement, respectively. The constant matrix Q is positive semi-definite (
), and the matrix R is positive
definite (
), while the constants
, and
are tuning parameters for the MPC. The first term in
equation 6 controls the predicted system state, and the second term controls the input sequence. In equation 7, the
inequality constraint is placed on both the state and input, though in general, MPC is capable of handling both linear
and nonlinear constraints. Figure 3 shows schematically how the MPC is implemented, and is described by the
following algorithm:
(1)
(2)
(3)
(4)
Obtain measurement of current state,
Solve the open-loop finite horizon optimization problem subject to constraints,
Apply the first control action of the optimal input sequence,
Go back to step 1.
Fig. 3: Basic concept of model predictive control (MPC), demonstrating how at time step k, the predicted output (yk) is calculated
over the prediction horizon (Np), to calculate the optimal input sequence (uk). This process is repeated at every time step.
MODEL AND CONTROL FOR A BIOREACTOR SYSTEM
In addition to sensor development, deriving appropriate bioreactor system models is critical for MPC. In general,
models for most bioreactor systems can be described by the following reaction:
Given this scheme, a mass balance for a batch bioreactor system can be described by a system of nonlinear ordinary
differential equations. A model of Moorella thermoacetica, an anaerobic bacterium that demonstrates a high
conversion of carbohydrates to acetate, was developed for batch fermentation as follows:
(7)
(8)
(9)
where X, S, and P, are the biomass, substrate and product concentration, respectively; μ is the specific growth rate of
the biomass; Kd is rate of cell death (constant); q is the rate of substrate consumption; and
is the rate of product
formation associated with growth. In this scheme product formation takes place simultaneously with cell growth.
The models presented in equations 7 – 9 are called unstructured-unsegregated models, and are the simplest
representation and most useful for control systems. The cell growth kinetics are often given by some type of Monod
relationship or derivation thereof, of which there exists numerous variations throughout the literature. For this work,
growth kinetics include both substrate and product inhibition characteristics,
(10)
(11)
(12)
where
is the maximum growth rate;
is the substrate inhibition constant;
is the maximum product
concentration before total inhibition; is a constant;
and
are the cell yield and product yield constants,
respectively; and
is the cell maintenance coefficient. The model was fitted to data from [40] and the nine
parameters were estimated using a nonlinear least squares fit in MATLAB (R2011). The results from the fitted
model are shown in Fig. 4 and the estimated parameter values are shown in Table 1, and also include the mean
squared error (MSE).
Fig. 4: Batch kinetic model of M. thermoacetica, including biomass (X), Glucose (S), and Acetate (P) concentrations, fitted to
literature data from [40].
Table 1: Parameter estimates for batch kinetic model (shown in Fig. 4) of M. thermoacetica after being fitted to literature data
from [40].
Substrate
MSE
Parameter
μmax
Ks
Kd
Ms
Yxs
Ypx
Pmax
n
Initial Guess
Glucose
8.25
Value
0.14
1.6
0.01
0.02
0.2
3.3
48.0
1.0
0.12
2.9
0.01
0.015
0.18
0.25
56.0
1.1
The batch model was then used to make comparison between a system using the same organism for continuous
fermentation in a cell-recycle system [41]. The model Equations 7 – 9 were modified to yield the following model
for a membrane cell-recycle bioreactor (MCRB):
(13)
(14)
(15)
where D is the dilution rate; Sf is the substrate feed concentration; and B is the bleed ratio. The model was used to
calculate productivity, steady state glucose and steady state acetate concentrations at various dilution rates. Figure 5
shows the results from the model compared to literature data from [41]. It can be seen that even using limited data
from literature the model is fairly accurate in calculating steady state behavior of the system.
Fig. 5: Comparison of model of an MCRB for fermentation of glucose to acetate by M. thermoacetica to literature data from
[41], (Left Side) cell concentration of X0 = 5 g L-1, feed concentration Sf = 20 g L-1; (Right Side) cell concentration of X0 = 5 g
L-1, feed concentration Sf = 30 g L-1
In general, bioreactor models are highly nonlinear, resulting in parametric and unstructured uncertainty. The full
nonlinear MPC problem involving both nonlinear equations, coupled with system constraints, forms a nonlinear
optimization problem that can often be computationally difficult to solve reliably in a reasonable amount of time.
Even then, a solution might not exist. For this initial controller development we use the method of successive
linearization for MPC detailed in [42]. This method results in linear approximations of the nonlinear reactor system
at the current system state. The result is a quadratic program (QP), for which there are efficient numerical
optimization methods for finding a solution. Given the nonlinear system
represented in equations 13 – 15,
the following nonlinear MPC problem is defined:
(16)
Such that,
(17)
where ∆ = x – xsetpoint the difference between the current state (x) and some reference trajectory or setpoint (x setpoint)
and the constraints in equation 17 are implemented to ensure input and output values achieve realistic values. The
manipulated variables for the bioreactor were the dilution rate, substrate feed concentration, and bleed ratio. The
model was initiated by running in open-loop until a steady state was reached, which were then used for initial
conditions for the closed-loop system. The setpoints were set at 25 g L-1 for acetate concentration (P) and glucose
concentration (S) was regulated to 0.1 g L-1. Furthermore, it was assumed that a 10% model mismatch between the
plant and controller model existed and full state feedback is provided by online sensors. The results of a set point
change are shown in Fig. 6 below.
Fig. 6: Simulation of MCRB using MPC to regulate glucose concentration (S) to 0.1 g L-1and acetate concentration (P) to 25 g
L-1, by controlling the inputs: dilution rate (D), bleed ratio (B), and substrate feed concentration (Sf). Optimization was solved
through successive linearization in MATLAB (2011).
The controller was able to achieve steady state performance at the desired setpoint, while reducing the glucose
concentration in the system. Further work will include disturbance rejection and application to various fermentation
reactor systems. Having a closed loop feedback system from online sensors of the system will greatly reduce model
uncertainty by providing accurate measurements of current system states.
CONCLUSION
Online measurements of bioprocesses are critical for the success of the biorefining industry as the production of
fuels and chemicals from lignocellulosic biomass increases. As measurement technologies, such as Raman
spectroscopy, are further developed to continuously and accurately quantify multiple reactant and product
concentrations in enzymatic hydrolysis and fermentation reactions, the processing of lignocellulosic biomass can
take advantage of APC systems that increases process efficiency. This coupling of advanced sensors and process
control has the potential to increase efficiency and improve profit margins.
REFERENCES
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
Galbe, M., Zacchi, G. Biomass and Bioenergy. 46, 70–78 (2012).
Bura, R., Ewanick, S.M., Gustafson, R.R. Tappi J. 11, 59–66 (2012).
Vajzovic, A., Bura, R., Kohlmeier, K., Doty, S. J. Ind. Microbiol. Biotechnol. 39, 1453–63 (2012).
Balasubramanian, N., Kim, J.S., Lee, Y.Y. Appl. Biochem. Biotechnol. 91-93, 367–376 (2001).
He, G., Kong, Q., Chen, Q., Ruan, H. J. Zhejiang Univ. Sci. B. 6, 1076–1080 (2005).
Lewis, I.R., Edwards, H.G.M. Handbook of Raman spectroscopy: from the research laboratory to
the process line. In Lewis, IR and Edwards, HGM, eds. 28, CRC Press, New York, NY (2001).
Brooker, M.H., Hancock, G., Rice, B.C., Shapter, J. J. Raman Spectrosc. 20, 683–694 (1989).
Shih, C.-J., Lupoi, J.S., Smith, E.A. Bioresour. Technol. 102, 5169–5176 (2011).
Shaw, A.D., Kaderbhai, N., Jones, A., Woodward, A.M., Goodacre, R., Rowland, J.J., Kell, D.B.
Appl. Spectrosc. 53, 1419–1428 (1999).
Picard, A., Daniel, I., Montagnac, G., Oger, P. Extremophiles. 11, 445–452 (2007).
Battaglia, T.M., Dunn, E.E., Lilley, M.D., Holloway, J., Dable, B.K., Marquardt, B.J., Booksh,
K.S. Analyst. 129, 602–606 (2004).
Agarwal, U.P. In Argyropoulos, DS, ed., Advances in lignocellulosics characterization. , TAPPI
Press, Atlanta GA, 201–225 (1999).
Afseth, N.K., Segtnan, V.H., Wold, J.P. Appl. Spectrosc. 60, 1358–1367 (2006).
Lieber, C.A., Ahadevan-Jansen, A.M. Appl. Spectrosc. 57, 1363–1367 (2003).
Shope, T.B., Vickers, T.J., Mann, C.K. Appl. Spectrosc. 41, 908–912 (1987).
Ewanick, S.M., Thompson, W., Marquardt, B., Bura, R. Biotechnol. Biofuels. 6, 1 - 8 (2013).
Shih, C.-J., Smith, E.A. Anal. Chim. Acta. 653, 200–206 (2009).
Mendes, L.S., Oliveira, F.C.C., Suarez, P.A.Z., Rubim, J.C. Anal. Chim. Acta. 493, 219–231
(2003).
Gray, S.R., Peretti, S.W., Lamb, H.H. Biotechnol. Bioeng. 110, 1654–1662 (2013).
Avila, T.C., Poppi, R.J., Lunardi, I., Tizei, P.A.G., Pereira, G.A.G. Biotechnol. Prog. 28, 1598–
1604 (2012).
Olofsson, K., Bertilsson, M., Lidén, G. Biotechnol. Biofuels. 1, 7 (2008).
Gelder, J. De, Gussem, K. De, Vandenabeele, P., Moens, L. J. Raman Spectrosc. 38, 1133–1147
(2007).
Almeida, M., Alves, R. Anal. Bioanal. Chem. 397, 2693–2701 (2010).
Uysal, R.S., Soykut, E.A., Boyaci, I.H., Topcu, A. Food Chem. 141, 4333 - 4343 (2013).
Kacuráková, M., Mathlouthi, M. Carbohydr. Res. 284, 145–157 (1996).
Cao, Y., Lu, Y., Huang, Y. J. Mol. Struct. 693, 87–93 (2004).
Bauer, M., Craig, I.K. J. Process Control. 18, 2–18 (2008).
Yamuna Rani, K., Ramachandra Rao, V. Bioprocess Biosyst. Eng. 21, 77–88 (1999).
Henson, M. IEEE Control Syst. Mag. , 54–62 (2006).
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
Qin, S.J., Badgwell, T.A. Control Eng. Pract. 11, 733–764 (2003).
Ashoori, A., Moshiri, B., Khaki-Sedigh, A., Bakhtiari, M.R. J. Process Control. 19, 1162–1173
(2009).
Kiran, A.U.M., Jana, A.K. Bioprocess Biosyst. Eng. 32, 801–808 (2009).
Ramaswamy, S., Cutright, T.J., Qammar, H.K. Process Biochem. 40, 2763–2770 (2005).
Nagy, Z.K. Chem. Eng. J. 127, 95–109 (2007).
Mhaskar, P., Aumi, S. Can. J. Chem. Eng. 85, 416–423 (2007).
Morari, M., Lee, J.H. 23, 667–682 (1999).
Mayne, D.Q., Rawlings, J.B., Rao, C. V, Scokaert, P.O.M. Automatica. 36, 789–814 (2000).
Findeisen, R., Imsland, L., Allgower, F., Foss, B. Eur. J. Control. 9, 179–195 (2003).
Cannon, M. Annu. Rev. Control. 28, 229–237 (2004).
Sugaya, K., Tusé, D., Jones, J. Biotechnol. Bioeng. 28, 678 – 683 (1986).
Parekh, S.R., Cheryan, M. Enzyme Microb. Technol. 16, 104–109 (1994).
Lee, J., Ricker, N. Ind. Eng. Chem. Res. 33, 1530–1541 (1994).