Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

This article appeared in a journal published by Elsevier.

The attached
copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elsevier’s archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/copyright
Author's personal copy

Bioresource Technology 103 (2012) 300–308

Contents lists available at SciVerse ScienceDirect

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

Estimating biofilm reaction kinetics using hybrid mechanistic-neural


network rate function model
B. Shiva Kumar, Ch. Venkateswarlu ⇑
Chemical Engineering Sciences Division, Indian Institute of Chemical Technology, Hyderabad 500 007, India

a r t i c l e i n f o a b s t r a c t

Article history: This work describes an alternative method for estimation of reaction rate of a biofilm process without
Received 21 June 2011 using a model equation. A first principles model of the biofilm process is integrated with artificial neural
Received in revised form 30 September 2011 networks to derive a hybrid mechanistic-neural network rate function model (HMNNRFM), and this com-
Accepted 2 October 2011
bined model structure is used to estimate the complex kinetics of the biofilm process as a consequence of
Available online 17 October 2011
the validation of its steady state solution. The performance of the proposed methodology is studied with
the aid of the experimental data of an anaerobic fixed bed biofilm reactor. The statistical significance of
Keywords:
the method is also analyzed by means of the coefficient of determination (R2) and model efficiency (ME).
Hybrid model
Biofilm kinetics
The results demonstrate the effectiveness of HMNNRFM for estimating the complex kinetics of the bio-
Artificial neural network film process involved in the treatment of industry wastewater.
Anaerobic biofilm reactor Ó 2011 Elsevier Ltd. All rights reserved.
Rate function model

1. Introduction quence of the validation of the biofilm model with the help of mea-
sured data. This approach is also known as inverse modeling in
Fixed bed biofilm reactors are increasingly used for anaerobic which the model parameters are determined such that the behavior
treatment of industrial wastewater. Since biological reactions of the process model approximates the observed process behavior.
involving the conversion of substrate by microorganisms are highly However, the problem of the uncertainty in the values of kinetic
complex, it is often difficult to analyze the biofilm reactors experi- parameters in Monod/Haldane type equations for single substrate
mentally and mathematical models could be very helpful for their and mixed substrate growth kinetics in biological processes has
analysis. However, mathematical modeling of these reactors, which been recognized in the literature reviewed by Kovarova-kova and
have a microbial film attached to the solid internal packing, is not Thomaselgi (1998).
always straight forward. Process kinetics plays a significant role Successful modeling and optimization of waste water treating
in assessing the rates of wastewater treatment reactors. Effective biofilm reactors require accurate knowledge of the biofilm kinetics.
application of biofilm reactor models to practical problems suffers Despite the fact that several attempts have been made to develop
due to the lack of knowledge of appropriate kinetic models and kinetic models, a detailed understanding of the biofilm reaction
uncertainty in the model parameters. The rates of biological reac- kinetics is still a difficult task. This difficulty is due the complex
tions are usually approximated by Monod kinetics and also some nature of the biofilm, in which a variety of biological, chemical
times by Contois and Haldane models. Several attempts have been and physical processes takes place simultaneously. Thus, estimat-
made in the past to determine the parameters of kinetic models of ing the reaction kinetics and kinetic constants is the most critical
waste water treating biofilm reactors, mostly experimentally issue in the studies related to biofilm process. Most of the pub-
(Livingston and Chase, 1989; Jih and Huang, 1994; Nguyen and lished works on biofilm kinetics employ kinetic equations for
Shieh, 1995; Hu et al., 2002; Tsuneda et al., 2002; Kapdan, 2005; which the parameters are evaluated either experimentally or
Chakraborty et al., 2011; Yuan et al., 2011; Yang et al., 2011). numerically. However, it is difficult to choose a proper model equa-
Numerical evaluation of the kinetic parameters of such biofilm pro- tion from a number of possible kinetic models and accurately fit
cesses is an attractive alternative to the experimental methods as the equation to estimate the reaction rate similar to the experi-
reported by various researchers (Shiraishi et al., 1996; Khorasheh mental reaction rate.
et al., 2002; Sarti et al., 2004; Spigno et al., 2004; Kiranmai et al., The aim of this study is to explore an alternative method to rep-
2005). By this approach, the parameters are determined as a conse- resent the biofilm kinetics. The intension is to estimate the reaction
rates of biofilm processes without using a Monod/Haldane type of
⇑ Corresponding author. Tel.: +91 40 27193121; fax: +91 40 27193626. kinetic equation. A mechanistic model of the biofilm reactor can be
E-mail address: chvenkat@iict.res.in (Ch. Venkateswarlu).
derived from first principles considerations such as mass balances

0960-8524/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.biortech.2011.10.006
Author's personal copy

B. Shiva Kumar, Ch. Venkateswarlu / Bioresource Technology 103 (2012) 300–308 301

Nomenclature

av specific surface area of particle, m1 r radial coordinate, m


c substrate concentration in the bulk fluid, kg/m3 Re Reynolds number (Duq/l)
cs substrate concentration in the biofilm, kg/m3 rs reaction rate, kg/m3/day
COD chemical oxygen demand, kg/m3 Sc Schmidt number (l/qD)
BOD biological oxygen demand, kg/m3 u superficial velocity, m/s
TA total alkalinity, kg/m3 z axial coordinate, m
TVA total volatile acids e porosity of bed
D substrate diffusion coefficient in the bulk fluid, m2/day l viscosity, kg/m/day
Df substrate diffusion coefficient in the biofilm, m2/day n space coordinate in the bio-film, m
dp equivalent particle diameter, m q density of fluid, kg/m3
OLR organic loading rate, kg/m3/day qs density of biomass, kg/m3
HLR hydraulic loading rate, m/day y measured substrate conversion at the exit of the reactor
HRT hydraulic retention time, days ^
y model predicted substrate conversion at the exit of the
Lf thickness of the bio-film, m reactor
kg mass transfer coefficient, m/s

on the process variables. However, the critical factor in determin- ciples models with artificial neural networks (ANNs) is a popular
ing the behavior of the biofilm process is the unknown kinetics modeling paradigm. The widely used ANN paradigm is a multi
of the conversion of substrate to products. The central idea is to layered feed-forward network (MFFN) (Nagata and Chu, 2003;
integrate the mechanistic model of the process with a neural net- Bas et al., 2007; Banerjee et al., 2008). The MFFN provides a map-
work (NN) to form a combined model structure and estimate the ping between an input (x) and an output (y) through a nonlinear
complex kinetics of the biofilm process using this model structure. function f, i.e., y = f(x). The first principles partial model specifies
Thus, a hybrid mechanistic-neural network rate function model process variable interactions from physical considerations. Knowl-
(HMNNRFM) is proposed for biofilm kinetics, in which the neural edge of the process may be used to impose constraints to the neu-
network is designed to provide the reaction rate as a consequence ral network model. The incorporation of the available process
of the validation of the steady state solution of the mathematical knowledge from the partial first principles model into the neural
model of the process. The performance of the proposed strategy network makes the combined model to serve as an estimator of
for estimation of the biofilm kinetics is studied with the aid of unmeasured process parameters that are difficult to be determined
the experimental data of an anaerobic fixed bed biofilm reactor from first principles model alone. The hybrid model has an internal
that involves the treatment of industry waste water. structure which clearly determines the interaction among process
variables and process parameters. This model is easier to analyze
than the standard neural network model and is found to have bet-
2. Theory ter interpolation and extrapolation ability.
Neural network-first principles hybrid models can be developed
2.1. Hybrid mechanistic-neural network rate function model in different combinations. A common approach is having the first
(HMNNRFM) principles model as the basis, while the neural network calculates
the unknown parameters. Another alternative is to learn the net-
Hybrid modeling approach involves the combination of first work using the deviation between the mathematical model output
principles models with black-box techniques such as least squares, and the desired output. Other possible approach is using the deter-
fuzzy logic and artificial neural networks. This approach has signif- ministic model as reinforcement for the function relationship
icant advantages over a black-box modeling methodology. It has between inputs and outputs. Hybrid models are reported for differ-
the flexibility of using both the techniques while making use of ent applications in different domains. Psichogios and Ungar (1992)
their strengths and avoiding the individual limitations. Among have introduced the hybrid neural network-first principles model-
the hybrid modeling formalisms, the combination of the first prin- ing scheme, in which the neural network component is used to

Fig. 1. Schematic of HMNNRFM.


Author's personal copy

302 B. Shiva Kumar, Ch. Venkateswarlu / Bioresource Technology 103 (2012) 300–308

estimate the specific growth rate of a fed batch bioreactor and the kg/m3, chlorides = 0.18 kg/m3, nitrates = 0.7 kg/m3, sulfates = 3.7
first principles model is used to provide the output values of the kg/m3, and BOD/COD = 0.577. A total 13 experiments are carried
process variables. Zahedi et al. (2005) proposed a hybrid modeling out and in each experiment the samples of effluent water exiting
scheme, in which neural network is adopted to estimate the outlet the column are collected and analyzed for COD, BOD, TVA, TA, pH,
concentration of a packed bed catalytic reactor and the first princi- etc. The stable state conditions of each experiment are observed
ple model is used to calculate the state variables characterizing the with respect to the minimal variation in reactor effluent pH, alkalin-
reactor behavior. Ng and Hussain (2004) proposed a hybrid model ity and COD concentrations. Additional details on experimental as-
in which neural network is used to estimate the internal parame- pects and data generation can be referred elsewhere (Rama Rao
ters of a semi batch polymerization reactor and a first principle et al., 2010). The data corresponding to the influent and effluent
model is used to estimate the states of the reactor. COD concentrations, hydraulic retention times (HRT) and organic
Based on the concept of hybrid modeling, in this study, a hybrid loading rates (OLR) of all experiments with respect to both the pack-
mechanistic-neural network rate function model (HMNNRFM) is ing material are given in Table 2. From these experiments, it is ob-
proposed for estimating the complex kinetics of a biofilm process. served that the substrate conversion is considerably influenced by
The motivation for this study has transpired from the fact that bio- the packing material as well as the OLR. The increase in porosity of
film reactions are difficult to model due to the intricate nature of packing is observed to decrease the substrate degradation rate due
reaction mechanisms and the reaction orders need to be postulated to decrease in surface area of biofilm. For either type of packing,
and experimentally verified. The neural network component of the COD removal rate is found to decrease with increase in OLR in
HMNNRFM is designed to provide the reaction rate as a conse- spite of BOD/COD ratio is 0.577.
quence of the validation of the steady state solution of the mathe-
matical model of the process. The process variables resulting from 3.1.2. Mechanistic models of biofilm reactor
the solution of the mathematical model of the fixed bed biofilm Mathematical models of different complexities can be used to
reactor are used as inputs to the network along with the other pro- represent the biofilm reactor.
cess conditions and its output is obtained as the estimate of the
derivative of the current reaction rate. The integrated result of 3.1.2.1. One dimensional model. The rate of mass transfer is consid-
the network output is incorporated into the mathematical model ered to be proportional to the concentration difference between
of the process and the resulting solution from the model is used the interface and the bulk fluid. Under steady state condition, there
as input to the network for the next iteration. An indirect error can be no accumulation of the component at the interface and the
function defined in terms of the model predictions and the mea- mass transported from the bulk liquid to the film interface moves
sured data obtained from an experimental fixed bed anaerobic bio- from interface to the solid surface. The conceptual biofilm model
film reactor is used to train the neural network rate function model showing the substrate concentrate profile in a segment of biofilm
involved in HMNNRFM. The learned hybrid model can be used for reactor can be referred elsewhere (Rama Rao et al., 2010). The
predicting the biofilm reactor behavior and optimizing the process differential equation governing the fluid phase is given by
conditions. The schematic of HMNNRFM is shown in Fig. 1.
dc
u ¼ kg am ðc  css Þ ð1Þ
dz
3. Methods
with the boundary condition:
3.1. Experimental system and mechanistic models c ¼ c0 at z ¼ 0 ð2Þ
In this model, the substrate concentration in the bulk fluid varies
A brief description of experimental and modeling aspects of bio-
only in the axial direction.
film reactor are given as follows.
The differential equation governing the solid phase is given by
 
3.1.1. Biofilm reactor and experimental data Df d dcs
na1  rs ðcs Þ ¼ 0 ð3Þ
A fixed bed biofilm reactor with raschig rings/gravel stones as na1 dn dn
packing material is used to conduct experiments for treating the dis-
with the boundary conditions:
tillery industry wastewater stream with varying organic loading
rates and feed rates with different hydraulic retention times. The dcs
¼ 0 at n ¼ 0
specifications of biofilm reactor and packing material are given in dn
Table 1. The characteristics of distillery wastewater collected from
industry are: COD = 67.46 kg/m3, BOD = 38.9 kg/m3, pH = 6.9, total
Table 2
alkalinity = 1.385 kg/m3 (as CaCO3), total solids = 3.49 kg/m3, total
Experimental data.
dissolved solids = 2.9 kg/m3, total suspended solids = 0.59 kg/m3,
total volatile solids = 0.27 kg/m3 (as CaCO3), phosphates = 0.125 Data OLR HRT Influent Raschig rings Gravel
sets (days) COD
Effluent % COD Effluent % COD
COD reduction COD reduction
Table 1
1 1.24 4 4.95 1.550 68.69 1.380 72.12
Specifications of biofilm reactor and packing material.
2 2.48 2 4.95 1.680 66.06 1.840 62.83
Height of the column 1m 3 1.47 8 11.77 3.680 68.73 3.280 72.13
Diameter of the column 0.1016 m (4 in) 4 2.94 4 11.77 4.280 63.64 3.770 67.97
Total volume of the reactor 8.11 l 5 5.89 2 11.77 4.750 59.64 4.130 64.91
L/D ratio 9.8425 6 1.95 12 23.44 7.620 67.49 7.330 68.73
7 2.93 8 23.44 8.020 65.78 8.250 64.80
Property Gravel stones Raschig rings
8 5.86 4 23.44 9.150 60.96 8.910 61.99
Shape Irregular Hallow cylindrical 9 11.72 2 23.44 11.260 51.96 10.090 56.95
Size 2–3 cm 1–1.5 cm 10 3.39 12 40.70 13.930 65.77 13.490 66.86
Density 1.49 kg/l 0.82 kg/l 11 5.09 8 40.70 17.370 57.32 15.570 61.74
Total packing weight 12.07 kg 6.725 kg 12 10.18 4 40.70 17.890 56.04 16.800 58.72
Porosity 0.605 0.757 13 20.35 2 40.70 21.250 47.79 18.740 53.96
Author's personal copy

B. Shiva Kumar, Ch. Venkateswarlu / Bioresource Technology 103 (2012) 300–308 303

  dcs Lf ¼ a þ b OLR  c HLR ð9Þ


kg css  c ¼ Df at n ¼ Lf ð4Þ
dn
where a, b and c are constants.
In Eq. (3), ‘a’ refers to 1, 2, and 3 for planar, cylindrical and
spherical geometries, respectively. The notation used for the terms 3.2. Design and implementation of HMNNRFM for biofilm kinetics
in the model can be referred in nomenclature.
The inputs to the neural network component of HMNNRFM are
3.1.2.2. Two dimensional model. This model considers the variation the organic loading rate (OLR), hydraulic loading rate (HLR), initial
of the substrate concentration in bulk fluid in both the axial and ra- substrate concentration (c0) and bulk fluid substrate concentration
dial directions. The model accounts for inter-phase resistance to (cs), and its output is the estimate of the derivative of the reaction
mass transfer and intra-biofilm processes. The process taking place rate (drs/dcs). Thus, the data sets of the 13 experiments, each repre-
in biofilm is characterized by physical diffusion and reaction and is sented by xp = [cs, c0, OLR, HLR]T form the inputs to the network. The
described by second order differential equation. Pressure losses input to the bias node is considered to be unity. The network is
along the length of the reactor are neglected. processed using the normalized data of inputs. Normalization of
The differential equation governing the bulk fluid phase is given
by
   Table 4
dc eD @ @c
u ¼ r  kg am ðc  css Þ ð5Þ Optimum inter connection weights of ANN involved in HMNNRFM for the reactor
dz r @r @r with raschig rings packing.

This equation has to be solved subject to the boundary condi- Training with 10 sets Training with 13 sets
tion given in Eq. (2). The solid phase biofilm model is same as in wij (i = 1–5; wjk (i = 1–8; wij (i = 1–5; wjk (i = 1–8;
Eq. (3) with its boundary conditions in Eq. (4). The notation used j = 1–7) j = 1) j = 1–7) j = 1)
for the terms in the model is given in nomenclature. w11 = 0.0189 w11 = 14.612 w11 = 0.0082 w11 = 2. 614
w12 = 0.0218 w21 = 8.069 w12 = 0.0218 w21 = 1.25689
3.1.2.3. Mass transfer coefficient. The transport of substrate mass w13 = 4.2814 w31 = 41.934 w13 = 4.2609 w31 = 21.168
w14 = 0.2445 w41 = 13.926 w14 = 0.204 w41 = 0.8564
between the bulk fluid and biofilm surface depends on the dy- w15 = 0.068 w51 = 6.9248 w15 = 4.0947 w51 = 13.1732
namic properties of the fluid flowing through the reactor and the w16 = 2.2134 w61 = 2.6598 w16 = 0.00214 w61 = 16.472
activity of the biocatalyst particle. The substrate mass transport w17 = 5.9875 w71 = 1.7598 w17 = 1.7173 w71 = 8.158
occurring from the bulk fluid to the biofilm surface across the dif- w21 = 0.2364 w81 = 8.1254 w21 = 2.2136 w81 = 17.054
w22 = 0.00987 w22 = 0.00987
fusion layer is defined by
w23 = 5.2148 w23 = 15.28
kg ðscÞ2=3 w24 = 6.3578 w24 = 11.103
jD ¼ ð6Þ w25 = 4.6587 w25 = 9.159
u w26 = 7.429 w26 = 3.75214
In the above equation, Sc is Schmidt number, the jD-factor is calcu- w27 = 2.128 w27 = 2.013
w31 = 12.236 w31 = 13.148
lated using the following correlation (Dwivedi and Upadhyay, w32 = 0.00548 w32 = 0.07541
1977): w33 = 1.325 w33 = 1.524
w34 = 0.5441 w34 = 0.6441
0:765 0:365
jD e ¼ þ ð7Þ w35 = 7.0610 w35 = 4.0770
Re0:82 Re0:386 w36 = 8.3560 w36 = 1.991
qud w37 = 3.7590 w37 = 2.4578
where Re ¼ l p for 0.01 < Re < 15,000. w41 = 10.6528 w41 = 9.1254
The mass transfer coefficient, kg obtained from Eq. (6) is used in w42 = 0.441 w42 = 0.00987
bulk fluid phase equations, Eqs. (1) and (5). w43 = 1.5369 w43 = 1.0774
w44 = 7.2548 w44 = 6.7546
w45 = 18.809 w45 = 7.201
3.1.2.4. Biofilm thickness. Biofilm thickness (Lf) is an important
w46 = 10.648 w46 = 7.783
parameter, which has significant influence on the substrate con- w47 = 1.1549 w47 = 0.1213
version. It is hypothesized that the biofilm is composed of a static w51 = 2.0015. w51 = 4.2008.
portion and a variable portion in which the later is assumed to in- w52 = 0.02153 w52 = 0.456
w53 = 0.0514 w53 = 2.214
crease with the organic loading rate and decrease with the hydrau-
w54 = 9.2154 w54 = 4.1587
lic loading rate. This reasoning leads to the following expressions w55 = 7.159 w55 = 5.1249
for the biofilm thickness (Kiranmai et al., 2005): w56 = 0.0064 w56 = 0.6498
w57 = 19.258 w57 = 24.2004
Lf ¼ a þ b OLR ð8Þ

Table 3
Neural network parameters and performance measures.

Training parameters Raschig rings Gravel stones


Training with 10 sets Training with 13 sets Training with 10 sets Training with 13 sets
Momentum factor (a) 0.730 0.812 0.790 0.785
Learning rate (b) 0.20 0.19 0.18 0.18
Scaling parameter (/) 0.45 0.45 0.46 0.46
Iterations 1980 2050 1740 1850
Hidden nodes 7 7 7 7
Minimum training error 0.01056 0.01513 0.003551 0.005913
Coefficient of determination (R2) 0.8678 0.8123 0.9534 0.9239
Model efficiency (ME) 0.8593 0.8056 0.9308 0.9136
Author's personal copy

304 B. Shiva Kumar, Ch. Venkateswarlu / Bioresource Technology 103 (2012) 300–308

inputs is performed by selecting their minimum and maximum 1


Opk ¼ ypk ¼  ð13Þ
values within the range of their experimental values. The normal- 1 þ e Spk
ization of cs is considered based on the process knowledge. The
where M refers the number of nodes in hidden layer. The network
network output, drs/dcs is used as such since its value ranges be-
output ypk refers the estimate of the derivative of the reaction rate
tween 0 and 1. The initial interconnection weights, wij are ran-
(drs/dcs). The functional dependence of the derivative of the reaction
domly assigned in the range of 1.0 to 1.0. The number of nodes
rate is a mathematical superposition of the process knowledge and
in the input and output layers of the network is specified by the in-
process conditions as expressed by
put and output variable dimensionalities. However, the number of
hidden nodes is an adjustable structural parameter.Each set of in- dr s
¼ f ðco ; cs ; OLR; HLRÞ ð14Þ
put data xp is multiplied by interconnection weights wij and the dcs
products are summed to obtain the activation state Spj:
The drs/dcs on integration provides the update of the reaction rate as
X
N given by
Spj ¼ wij xpi þ wNþ1;j ð10Þ  
i¼1 drs
rsnþ1 ¼ r sn þ ðcsnþ1  csn Þ ð15Þ
dcs
where N refers the number of nodes in input layer. A sigmoid func-
tion is employed as the node processing function. The output of the where n refers to iteration. The values of drs/dcs and updated rs serve
hidden layer neuron Opj for sigmoid function is calculated as as inputs to the mechanistic model component of HMNNRFM de-
1 fined by one dimensional and two dimensional models of the bio-
Opj ¼ f ðSpj Þ ¼  ð11Þ film reactor given in Section 3.1.2.
1 þ e Spj
The numerical solution of mathematical models is carried out
where f represents the differentiable and non-decreasing function. using the following procedure. The height of the packed column
The output layer of a single hidden layer network performs the (1 m) is divided into 50 equal steps, each step representing
same calculations as above, except that the input vector xp is re- 0.02 m. In one dimensional model, the fluid phase equation, Eq.
placed by the hidden layer output Op and the corresponding (1) along with its boundary condition, Eq. (2) is solved for each
weights wjk: height step in the axial direction by using Runge–Kutta (RK) 4th or-
X
M der method. The solid phase equation for the biofilm given in Eq.
Spk ¼ wjk Opi þ wMþ1;k ð12Þ (3) is solved for each segment using orthogonal collocation on fi-
i¼1 nite elements (OCFE) (Finlayson, 1980). OCFE is implemented by

Fig. 2. Comparison of experimental and HMNNRFM prediction results for raschig rings packing.
Author's personal copy

B. Shiva Kumar, Ch. Venkateswarlu / Bioresource Technology 103 (2012) 300–308 305

considering two finite elements, each with four internal collocation In Eq. (23),
points. In two dimensional model, the fluid phase equation, Eq. (5) ^i ÞOpk ð1  Opk ÞOpj ð1  Opj Þwjk
dpj ¼ ðyi  y ð24Þ
is transformed to ODE by finite difference technique and the result-
ing equation along with its boundary condition, Eq. (2) is solved by The network interconnection weights wjk and wij are updated
using 4th order RK method for each height step of the axial direc- based on the incremental interconnection weights Dwjk and Dwij.
tion. The solution for solid phase equation of this model is same as This finishes the first iteration. Next iteration is performed by using
in one dimensional model. The solution of the mathematical mod- the updated interconnection weights and the input data pattern
els yields the values of the bulk fluid substrate concentration (c) along with the previously evaluated cs and rs. For every iteration,
and biofilm substrate concentration (cs). The substrate conversion, the concentrations evaluated by the model are filtered by using a
y is obtained from the following relation, simple filtering technique of the following form:
 
c csnþ1 ¼ /csnþ1 þ ð1  /Þcsn ð25Þ
y¼ 1 ð16Þ
c0
where / is a tuning parameter which lies between 0 and 1, and n
The network output drs/dcs and its integrated value rs resulting refers iteration. This type of smoothing makes the network to con-
for each input data pattern are incorporated in the mathematical verge faster. The same procedure is repeated in subsequent itera-
model, the solution of which provides the values of c and cs. Thus tions and the network is trained in association with the
the neural network in association with the mechanistic model pro- mathematical model of the process until convergence in solution
vides the model predictions, y ^i for all the input data patterns. An is achieved.
error function based on the actual and estimated substrate conver- The implementation of the procedure based on the two dimen-
sions at the exit of the reactor for all the experiments is defined as sional model of the biofilm reactor is outlined in the following
steps:
1X l
E¼ ^ i Þ2
ðy  y ð17Þ
2 i¼1 i 1. Set column height z = 0, data pattern l = 0 and iteration n = 0.
Assume initial guesses for cs and rs.
where E is the error function, l is the number of data sets, yi refer the
2. Solve the mechanistic model represented by Eqs. (3) and (5)
measured substrate conversions at the exit of the reactor, and y ^i de-
simultaneously while calling the neural network at each collo-
note the corresponding model predicted substrate conversions.
cation point and estimate the values of c and cs at each of these
Thus the observed error between the mechanistic model predic-
points.
tions and the actual measurements is used to update the network
interconnection weights. The commonly used algorithm for training
the network is the delta rule back propagation algorithm, the proce-
Table 5
dure of which is reported elsewhere (Jones and Hoskins, 1987; Optimum inter connection weights of ANN involved in HMNNRFM for the rector with
Girosi and Poggio, 1990). Since target output of the network is not Gravel stones packing.
directly available from the experimental data, the indirect error
Training with 10 sets Training with 13 sets
function defined by Eq. (17) is used to train the network. This error
wij (i = 1–5; wjk (i = 1–8; wij (i = 1–5; wjk (i = 1–8;
is back propagated through the mathematical model of the process
j = 1–7) j = 1) j = 1–7) j = 1)
to update the incremental interconnection weights, Dwjk between
w11 = 4.2814 w11 = 30.778 w11 = 3.9845 w11 = 24.6547
output and hidden layers, and Dwij between hidden and input lay-
w12 = 2.2134 w21 = 1.9854 w12 = 2.011 w21 = 0.9874
ers. The incremental interconnection weights Dwjk are calculated by w13 = 0.0189 w31 = 15.1602 w13 = 0.159 w31 = 13.678
using the equation, w14 = 5.9875 w41 = 0.06598 w14 = 3.753 w41 = 0.012964
w15 = 0.2364 w51 = 7.345 w15 = 2.303 w51 = 6.8887
@E
Dwjk ¼ b þ aDwjk ð18Þ w16 = 0.0218 w61 = 11.645 w16 = 0.54 w61 = 9.8754
@wjk w17 = 0.2445 w71 = 7.569 w17 = 0.02214 w71 = 8.1254
w21 = 0.068 w81 = 4.0418 w21 = 0.097 w81 = 3.648
where a and b are network tuning parameters representing the w22 = 1.123 w22 = 2.321
momentum factor and learning rate, respectively. In the above w23 = 2.6218 w23 = 1.8597
equation, w24 = 7.1257 w24 = 5.5087
w25 = 4.6007 w25 = 6.402
X l   w26 = 17.048 w26 = 13.148
@E @yi
¼ dpk Opj ð19Þ w27 = 13.587 w27 = 13.145
@wjk i¼1
@r s w31 = 10.118 w31 = 9.114
w32 = 0.855 w32 = 0.0505
where w33 = 11.015 w33 = 8.467
w34 = 1.5112 w34 = 15.2014
^i ÞOpk ð1  Opk Þ:
dpk ¼ ðyi  y ð20Þ w35 = 17.111 w35 = 13.236
@yi w36 = 0.0062 w36 = 1.024
in Eq. (19) is evaluated as
@rs w37 = 0.0201 w37 = 0.302

@y  y  yn1 w41 = 0.028 w41 = .2128
¼ n ð21Þ w42 = 4.401 w42 = 1.421
@r s n rsn  r sn1 w43 = 1.1587 w43 = 1.17
w44 = 1.9171 w44 = 0.02171
The incremental interconnection weights Dwij between input to
w45 = 18.809 w45 = 12.951
hidden layers are computed by using the equation, w46 = 10.125 w46 = 8.7564
w47 = 2.109 w47 = 1.912
@E
Dwij ¼ b þ aDwij ð22Þ w51 = 0.296. w51 = 0.958
@wij w52 = 2.743 w52 = 2.0982
w53 = 0.6514 w53 = 1.6124
In above equation, w54 = 11.3198 w54 = 13.156
w55 = 4.159 w55 = 3.458
@E X l
w56 = 0.64 w56 = 1.23
¼ ðxpi dpj Þ ð23Þ
@wij w57 = 17.8 w57 = 15.21
i¼1
Author's personal copy

306 B. Shiva Kumar, Ch. Venkateswarlu / Bioresource Technology 103 (2012) 300–308

3. Set z = z + 1, go to step 2 and repeat the procedure for all the col- estimate the biofilm reaction rate. For anaerobic treatment of dis-
umn heights of z until z = 49. The bulk fluid substrate concen- tillery waste water, the biofilm thickness expression that considers
tration at the exit of the column is c and the biofilm substrate the effect of both OLR and HLR has been found to be effective over
concentration is cs. the expression that considers the effect of OLR only. Thus the film
4. Compute the substrate conversion efficiency by using Eq. (16). thickness expression, Eq. (9) with its reported parameters (Rama
5. Set l = l + 1, go to step 2 and repeat the procedure for all the data Rao et al., 2010) as a = 8.96  105, b = 1.61  104 and c = 4.20 
sets until l = 12. 106 for raschig rings packing, and a = 1.01  104, b = 1.38 
6. Evaluate the error function defined by Eq. (17). 104 and c = 6.34  106 for gravel stones packing is used along
7. Update neural network interconnection weights through error with the 2D model.
back-propagation. The method of HMNNRFM involved in the estimation of biofilm
8. Set n = n + 1, go to step 2 and repeat the procedure. kinetics is first configured for the reactor with raschig rings pack-
9. Stop the iterations when the convergence in objective is ing using the experimental data of 13 sets given Table 2. The num-
obtained. ber of hidden units, the number iterations, the learning rate, b and
the momentum factor, a of the neural network are suitably se-
4. Results and discussion lected as given in Table 3. The iterative convergence of the solution
in association with the mathematical model makes the network to
The proposed HMNNRFM based on its implementation proce- learn and modify the values of interconnection weights between
dure described in the previous section is employed to estimate the nodes of the layers. The optimal interconnection weights of this
the kinetics of the biofilm process involved in the anaerobic treat- network configuration based on 13 sets of data are shown in Table
ment of distillery industry wastewater in fixed bed reactor. The 4. The prediction results of HMNNRFM for the input data of 13 sets
experimental data of the reactor corresponding to two different in Table 2 are shown in Fig. 2(a). Further, HMNNRFM is built by
types of packing material given in Table 2 is used for this study. using the experimental data of 10 sets in Table 2 and its prediction
Previous modeling studies on anaerobic treatment of industry ability is tested with the randomly chosen data of 3 sets which are
waste water in fixed bed bioreactor have shown the better perfor- not involved in training. The network parameters corresponding to
mance of two dimensional (2D) mathematical models over one this configuration is given in Table 3. The optimal interconnection
dimensional (1D) models (Kiranmai et al., 2005). Thus HMNNRFM weights of this network configuration are shown in Table 4. The
involving the 2D model coupled with the neural network is used to predictive performance of HMNNRFM for the input data of 10 sets

Fig. 3. Comparison of experimental and HMNNRFM prediction results for gravel stones packing.
Author's personal copy

B. Shiva Kumar, Ch. Venkateswarlu / Bioresource Technology 103 (2012) 300–308 307

and the randomly chosen data of 3 sets is shown in Fig. 2(b). The estimation of biofilm kinetics involved in the anaerobic treatment
model predicted substrate conversions at the exit of the reactor of industry waste water.
for the input data of 4, 9 and 11 sets in Table 2 are 0.6286,
0.5106 and 0.5809, respectively, which are in good agreement with 5. Conclusions
experimental conversions of 0.6364, 0.5196 and 0.5732, respec-
tively. The values of the coefficient of determination (R2) and mod- The complex nature of biofilm processes makes it difficult to
el efficiency (ME) evaluated based on these model predictions are choose a known form of equation to represent the biofilm kinetics.
0.9756 and 0.9707, respectively. The definitions of R2 and ME are In this work, a hybrid mechanistic-neural network rate function
given in Appendix A. model (HMNNRFM) is presented to estimate the complex kinetics
Further, estimation of biofilm kinetics by HMNNRFM is carried of the biofilm processes without using a model equation. The re-
out for the reactor with gravel stones packing in a similar fashion sults on application to an anaerobic fixed bed biofilm reactor exhi-
as in the case of raschig rings packing. The trained network param- bit the effectiveness of HMNNRFM for estimating the kinetics of
eters, the training iterations and the converged error evaluated biofilm process involved in the treatment of wastewater. This
using 13 sets of experimental data with this packing are given in method is an interesting alternative to the conventional kinetic
Table 3. The optimal interconnection weights of this network con- models for estimation of reaction rates of biofilm processes.
figuration are shown in Table 5. The prediction results of
HMNNRFM for the input data of 13 experiments are shown in Appendix A. Performance measures for model predictions
Fig. 3(a). Also, HMNNRFM is built by using experimental data of
10 sets in Table 2 and its prediction ability is tested with the ran-
domly chosen data of 3 sets. The trained network parameters with
Parameter Description Mathematical
this data are also given in Table 3. The optimal interconnection
definition
weights of this network configuration are given in Table 5. The pre-
PN
2
dictive performance of HMNNRFM for the input data of 10 exper- R2 Coefficient of ðxi  Þ
xÞðy y
determination: It is a R2 ¼ PN i¼1 2 PNi
iments and the data of three experiments not involved in i¼1
ðxi 
xÞ i¼1
 Þ2
ðyi y
training is shown in Fig. 3(b). The model predicted substrate con- statistical measure that
versions at the exit of the reactor for the input data of 4, 9 and will give some
11 sets in Table 2 are 0.6773, 0.5605 and 0.6209, respectively, information about the
which are in close agreement with experimental conversions of goodness of fit of a
0.6797, 0.5695 and 0.6174, respectively. The values of R2 and ME model. It ranges from 0
evaluated based on these model predictions are 0.9870 and to 1. An R2 of 1.0
0.9838, respectively. The results of HMNNRFM for the biofilm reac- indicates that the
tor with raschig rings packing and the gravel stones packing are regression model
further analyzed to find the suitability of the packing to be used perfectly fits the real
with the reactor. The results of the method evaluated with respect data.
PN
to training error, number of iterations, R2 and ME based on 13 sets ME Model efficiency: It is a ðxi yi Þ2
ME ¼ 1  Pi¼1
of data given in Table 3 show the better suitability of the reactor good measure of N
i¼1
2 Þ
ðxi y

with gravel stones packing to that of raschig rings packing. The re- accuracy and variations
sults of HMNNRFM built based on 10 sets of data also show the accounted for by the
better performance of the reactor packed with gravel stones. model. It ranges from 
The HMNNRFM with the estimated biofilm kinetics is further to 1. A higher ME value
compared with the reported methods that involve known form corresponds to closer
of kinetic models. Recently, an inverse model based method using match between
ant colony optimization is reported for estimating the parameters predicted and observed
of kinetic models as a consequence of the validation of the process values
models with the aid of measured data obtained from an experi-
mental fixed bed anaerobic biofilm reactor involving the treat-
ment of distillery waste water (Rama Rao et al., 2010). The References
results of the inverse model based method evaluated for various
mathematical and kinetic modeling configurations as well as dif- Banerjee, A.K., Kiran, K., Murthy, U.S.N., Venkateswarlu, Ch., 2008. Classification and
ferent packing materials have shown the better performance of identification of mosquito species using ANN. Comput. Biol. Chem. 32, 442–447.
Chakraborty, C., Chowdury, R., Bhattacharya, P., 2011. Experimental studies and
the reactor represented by two dimensional mathematical model mathematical modeling of an up-flow biofilm reactor treating mustard oil rich
with Haldane kinetic equation and gravel stone packing. Compar- wastewater. Bioresour. Technol. 102 (10), 5596–5601.
ing HMNNRFM with this recently reported method is more appro- Bas, D., Dudak, F.C., Boyaci, I.H., 2007. Modeling and optimization IV: investigation
of reaction kinetics and kinetic constants using a program in which artificial
priate as the wastewater treatment process, the packing material
neural network (ANN) was integrated. J. Food Eng. 79, 1152–1158.
and the mathematical models used in both these methods are Dwivedi, P.N., Upadhyay, S.N., 1977. Particle-fluid mass transfer in fixed and
same. This comparison also forms a suitable basis to appraise fluidized bed. Ind. Eng. Chem. Process Des. Dev. 16, 157–165.
the performance of the proposed method which does not have a Finlayson, B.A., 1980. Nonlinear Analysis in Chemical Engineering. McGraw-Hill,
New York.
known form of kinetic model of the type Monod/Haldane to a Girosi, F., Poggio, T., 1990. Networks and the best approximation property. Biol.
method that has a known kinetic equation. The performance mea- Cybern. 63, 169–179.
sures of training error, R2 and ME evaluated for the reactor with Hu, W.C., Thayanithy, K., Forster, C.F., 2002. A kinetic study of the anaerobic
digestion of ice-cream wastewater. Process Biochem. 37, 965–971.
gravel stones packing by the inverse model based method are re- Jih, C.G., Huang, J.S., 1994. Effect of bio-film thickness distribution on substrate-
ported as 0.00795, 0.8119, and 0.8103, respectively, where these inhibited kinetics. Water Res. 28, 967–973.
values evaluated for the reactor with the same packing by Jones, W.P., Hoskins, J., 1987. Back Propagation. Byte, 155–162.
Kapdan, I.K., 2005. Kinetic analysis of dyestuff and COD removal from synthetic
HMNNRFM are 0.005913, 0.9239, and 0.9136, respectively. These wastewater in an anaerobic packed column reactor. Process Biochem. 40, 2545–
results thus demonstrate the effectiveness of HMNNRFM for the 2550.
Author's personal copy

308 B. Shiva Kumar, Ch. Venkateswarlu / Bioresource Technology 103 (2012) 300–308

Kavarova-Kova, K., Thomaselgi, 1998. Growth kinetics of suspended microbial cells: Sarti, A., Foresti, E., Zaiat, M., 2004. Evaluation of a mechanistic mathematical model
from single-substrate-controlled growth to mixed-substrate kinetics. Microbiol. of a packed bed anaerobic reactor treating waste water. Lat. Am. Appl. Res. 34,
Mol. Biol. R. 62 (2), 646–666. 127–132.
Khorasheh, F., Azadeh, K., Mireshghi, S.A., 2002. Application of an optimization Shiraishi, F., Hasegawa, T., Kasai, S., Makishita, N., 1996. Characteristics of apparent
algorithm for estimating intrinsic kinetic parameters of immobilized enzymes. kinetic parameters in a packed bed Immobilized enzyme reactor. Chem. Eng.
J. Biosci. Bioeng. 94 (1), 1–7. Sci. 51 (l1), 2847–2852.
Kiranmai, D., Jyothirmai, A., Murthy, C.V.S., 2005. Determination of kinetic Spigno, G., Zilli, M., Nicolella, C., 2004. Mathematical modeling and simulation of
parameters in fixed-film bio-reactors: an inverse problem approach. Biochem. phenol degradation in biofilters. Biochem. Eng. J. 19, 267–275.
Eng. J. 23, 73–83. Tsuneda, S., Auresenia, J., Morise, T., Hirata, A., 2002. Dynamic modeling and
Livingston, A.G., Chase, H.A., 1989. Modeling phenol degradation in a fluidized-bed simulation of a three-phase fluidized bed batch process for wastewater
reactor. AIChE J. 35, 1980–1992. treatment. Process Biochem. 38, 599–604.
Nagata, Y., Chu, K.H., 2003. Optimization of a fermentation medium using neural Yang, Y., Sun, G., Guo, J., Xu, M., 2011. Differential biofilms characteristics of
networks and genetic algorithms. Biotechnol. Lett. 25, 1837–1842. shewanella decolorationis microbial fuel cells under open and closed circuit
Ng, C.W., Hussain, M.A., 2004. Hybrid neural network-prior knowledge model in conditions. Bioresour. Technol. 102 (14), 7093–7098.
temperature control of a semi-batch polymerization process. Chem. Eng. Yuan, Y., Zhao, B., Zhou, S., Zhong, S., Zhuang, L., 2011. Electrocatalytic activity of
Process. 43, 559–570. anodic biofilm responses to pH changes in microbial fuel cells. Bioresour.
Nguyen, V.T., Shieh, W.K., 1995. Evaluation of intrinsic and inhibition kinetics in Technol. 102 (13), 6887–6891.
biological fluidized bed reactors. Water Res. 29, 2520–2524. Zahedi, G.A., Elkamel, A., Lohi, A., Jahanmiri, M.R., Rahimpor, 2005. Hybrid artificial
Psichogios, D.C., Ungar, L.H., 1992. Hybrid neural net work-first principles approach neural network-first principle model formulation for the unsteady state
to process modeling. AIChE J. 38 (10), 1499–1511. simulation and analysis of a packed bed reactor for CO2 hydrogenation to
Rama Rao, K., Srinivasan, T., Venkateswarlu, Ch., 2010. Mathematical and kinetic methanol. Chem. Eng. J. 115, 113–120.
modeling of biofilm reactor based on ant colony optimization. Process Biochem.
45, 961–972.

You might also like