Modeling Studies For The Determination of Completely Mixed Activated Sludge Reactor Volume: Steady-State, Empirical and Ann Applications

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/286156517
Modeling studies for the determination of completely mixed activated sludge

reactor volume: Steady-state, empirical and ann applications
Article in Neural Network World · January 2010
CITATIONS READS
16 311
1 author:
Kaan Yetilmezsoy
Yildiz Technical University
119 PUBLICATIONS 2,262 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Butterfly optimization algorithm View project
Oman Cement Company View project
All content following this page was uploaded by Kaan Yetilmezsoy on 02 May 2018.
The user has requested enhancement of the downloaded file.

MODELING STUDIES
FOR THE DETERMINATION OF
COMPLETELY MIXED ACTIVATED SLUDGE
REACTOR VOLUME: STEADY-STATE,
EMPIRICAL AND ANN APPLICATIONS
Kaan Yetilmezsoy∗
Abstract: This paper presents an empirical model and a three-layer (7:11:1) ar-
tificial neural network (ANN) approach for the determination of completely mixed
activated sludge reactor volume (CMASRV). CMASRV values were estimated by
a new mathematical formulation and a three-layer ANN model for 1,000 differ-
ent artificial scenarios given in a wide range of seven biological variables. The
predicted results obtained from each stochastic approach were compared with the
well-known steady state volume model based on mass balance equations. The com-
putational analysis showed that the proposed empirical model and ANN outputs
were obviously in agreement with the steady-state volume model and all the pre-
dictions proved to be satisfactory with a correlation coefficient of about 0.9989 and
1, respectively. The maximum volume deviations from the steady-state volume
equation were recorded as only 7.17% and 6.89% for the proposed model and ANN
outputs respectively. In addition to volume comparison, waste sludge mass flow
rates (PX ), food to mass ratios (F/M ), hydraulic retention times (HRTs), volu-
metric organic loads (LV ) and oxygen requirements (ORs) were also compared for
each model, and significant points of proposed approaches were evaluated.
Key words: Activated sludge, completely mixed reactor, steady-state model,

empirical model, artificial neural network
Received: September 26, 2009

Revised and accepted: May 13, 2010
∗ Kaan Yetilmezsoy
Department of Environmental Engineering, Yildiz Technical University, Faculty of Civil En-

gineering, 34220, Davutpasa Campus, Esenler, Istanbul, Turkey, Tel.: +90 212 383 53 76,
Fax: +90 212 383 53 58, E-mail: yetilmez@yildiz.edu.tr, kyetilmezsoy@gmail.com, Web:
http://www.yildiz.edu.tr/∼yetilmez
c
°ICS AS CR 2010 559
Neural Network World 5/10, 559-589
1. Introduction
The activated sludge process is the most generally applied biological wastewater
treatment method. In the activated sludge process, a bacterial biomass suspension
is responsible for the removal of pollutants. Depending on the design and the
specific application, an activated sludge wastewater treatment plant can achieve
removal of biological nitrogen and biological phosphorus, besides removal of organic
carbon substances [1]. The principle in activated sludge plants is the following: a
mass of activated sludge consisting largely of bacteria is kept moving in water
by stirring or aeration [2]. In the process, large quantities of air are bubbled
through wastewaters that contain dissolved organic substances in open aeration
tanks. Oxygen is required by bacteria and other types of microorganisms present
in the system to live, grow, and multiply in order to consume the dissolved organic
“food”, or pollutants in the waste.
The activated sludge process has been employed extensively throughout the
world in its conventional form and modified forms, all of which are capable of
meeting secondary treatment effluent limits. Step aeration, contact stabilization,
completely mixed activated sludge, extended aeration and closed-loop reactor (also
known as oxidation ditch) can be included in modified forms of the conventional
activated sludge process. Aeration time ranges between 3 and 6 hours. Recircu-
lation ratios in a completely-mixed system will range from 50 to 150% [3]. The
term “completely mixed reactor” usually refers to a continuously stirred tank re-
actor (CSTR) concept, a simplified representation that is often used to represent
activated sludge tanks. A CSTR has one input, and indeed the concentrations
are assumed to be homogeneous throughout the tank due to vigorous mixing, but
certainly not because wastewater and sludge are introduced uniformly over the
tank.
In order to achieve an efficient biological treatment, the most appropriate pro-
cess design should be performed. Biological variables used in the design should
be representative for the activated sludge process as far as possible. Activated
sludge processes are based on many projection criteria, such as wastewater flow
rate (Q), mixed liquor volatile suspended solids (MLVSS) concentration, biochemi-
cal oxygen demand (BOD5 ) load, hydraulic retention time (HRT), sludge age (θC ),
oxygen requirement (OR) and other some different biological variables related to
the process.
Modeling is a valuable tool in both design and operation of biological treatment
plants. In addition, modeling of wastewater treatment plants can also be used for
process optimization and testing of control strategies in order to meet effluent
quality requirements at a reasonable cost. Hence, modeling helps to develop a bet-
ter understanding of the treatment processes and provides a significant potential
for solving operational problems as well as reducing operational cost in biological
wastewater treatment process [4]. Moreover, model results can be evaluated for
different operating data before transferring the concepts to a full-scale plant. The
volume of an activated sludge reactor can be calculated by using different mathe-
matical models in terms of biological variables. It should be noted that some of the
kinetic coefficients presented in literature differ extremely from each other. Hence,
an appropriate data set depending on ranges of input biological variables should be
560
Yetilmezsoy K.: Modeling studies for the determination of completely. . .
considered in order to obtain efficient and reliable outputs for the proposed volume
model.
The artificial neural networks (ANNs) are intended to interact with the ob-
jects of the real world in the same way as does the biological nervous system [5].
The advantage of ANNs is to perceive complex interactions between the input and
output in multivariate systems. ANN techniques have been succesfully used in
many application fields, such as modeling, simulation, learning, recognition and
forecasting [6]. Stanley [7] has reported that ANNs are tools that can also quan-
tify pattern recognition. Because they do not require any prior knowledge of a
solution, ANN modeling has been applied to various environmental problems in
the past decade. Although ANN-based artificial intelligence techniques have been
considered as black-box analyses, however, in order to construct a good modeling
algorithm, it will be better to know some reasonable relationships between input
and target variables.
Rogers and Dowla [8] developed a new approach to nonlinear groundwater man-
agement methodology to optimize the aquifer remediation with the aid of ANNs.
The authors suggested that the ANN approach has many advantages, such as
greater influence of hydrogeologic insight and ability to recycle flow and transport
simulations. Hsu et al. [9] proposed a nonlinear ANN model approach to pro-
vide a better representation of a rainfall-runoff process. They presented a new
procedure entitled linear least squares simplex (also known as LLSSIM) for iden-
tifying the structure and parameters of three-layer feed-forward ANN models and
demonstrated the potential of such models for simulating the nonlinear hydrologic
behavior of watersheds. They emphasized that ANN approach provides a viable
and effective alternative to the linear time series approach for developing input-
output simulation and forecasting models in situations that do not require modeling
of the internal structure of the watershed. The authors reported that the nonlinear
ANN model approach provided a better representation of the rainfall-runoff rela-
tionship of the medium-size Leaf River basin near Collins, Mississippi, than the
linear autoregressive moving average with exogenous inputs (also known as AR-
MAX) time series approach or the conceptual Sacramento soil moisture accounting
model (SAC-SMA). Scardi [10] compared two conventional empirical models with a
new approach, based on ANNs for phytoplankton production modeling. Although
very simple ANNs were used in the study, it was concluded that ANNs provided
a much better fit to observed data than the conventional models. Yabunaka et al.
[11] carried out studies on a novel application of a back-propagation ANN model
for the prediction of algal bloom in Lake Kasumigaura, Japan. They concluded
that the ANN model achieved a reasonable effectiveness with respect to learning
the relationship between the set of selected water quality parameters and algal
bloom.
Another study was undertaken by Djebbar and Katoda [12] to estimate sanitary
flows under dry weather conditions, using ANNs. The validation showed that the
ANN model resulted in an average error less than 16%, when compared with the
measured data. Hall and Minns [13] applied a regional flood frequency analysis
using ANNs for the data from the southwest of England and Wales. The authors
concluded that the root mean square error (RMSE) of the ANN outputs was 39%
lower for the mean annual flood, and 30% lower for the 50-year flood than the
561
estimated values given by the UK Flood Studies Report (FSR) of 1975. Rodriguez
and Sérodes [14] presented two empirical models (a linear autoregressive model with
external inputs [also known as ARX] and a non-linear ANN model) for simulating
and forecasting residual chlorine concentrations within urban water systems. They
concluded that the ANN model gave better predictions than the ARX model for
specific water treatment conditions (very high and very low chlorine doses) due to
its unique ability to detect non-linear complex relationships between data.
Campolo et al. [15] proposed an ANN model forecasting river flow rate during
low-flow periods. The study concluded that model predictions were found to be ac-
curate with RMSE on the predicted river flow rate less than 8% over the entire time
horizon of prediction. Karul et al. [16] used a three-layer Levenberg–Marquardt
feed-forward learning algorithm to model the eutrophication process in three water
bodies of Turkey (Keban Dam Reservoir, Mogan and Eymir Lakes). Despite the
very complex and extraordinary nature of Keban Dam Reservoir, they observed
a relatively good correlation coefficient (between 0.60 and 0.75) between the mea-
sured and predicted values. For Mogan and Eymir Lakes, predictions between the
measured and ANN outputs proved to be satisfactory with a maximum correlation
coefficient of about 0.95. They emphasized that ANNs were able to model non-
linear behavior in eutrophication process reasonably well and could successfully
estimate some extreme values from validation and test data sets which were not
used in training the ANN.
Sengorur and Oz [17] investigated water pollution in fishponds by using an ANN
model. They trained the network, using 24 inputs, 9 outputs and 38 hidden neu-
rons with a learning coefficient of 0.2. The results of the model were found to be
compatible with the real data obtained by experiments, with an accuracy of 94.8%.
Anctil et al. [18] focused on the impact of the length of observed records on the
performance of MLPs type ANNs and of conceptual parsimonious rainfall-runoff
forecasting models. They discussed ANN modeling to help modelers to create and
train efficient rainfall-runoff models in a systematic way. The authors suggested
that longer training sets were more beneficial to the MLPs than to the conceptual
model. Hamed et al. [19] developed ANN models to predict the performance of a
wastewater treatment plant (WWTP) in the Greater Cairo district, Egypt, with
an average flow rate of 1 million m3 /day. They obtained daily records of Biochem-
ical Oxygen Demand (BOD) and Suspended Solids (SS) concentrations through
various stages of the treatment process over 10 months from the plant laboratory.
In the study, two ANN-based models for prediction of BOD and SS concentra-
tions in plant effluent are presented. The authors concluded that the ANN-based
models were found to provide an efficient and robust tool in predicting WWTP
performance. Nunnari et al. [20] modeled SO2 concentration at a point by inter-
comparing several statistical techniques, such as ANNs, fuzzy logic, generalized
additive techniques. Because the ANN models worked better in predicting critical
episodes, they recommended the ANN technique for the implementation of a warn-
ing system for air quality control. Onkal-Engin et al. [21] used an ANN trained
with a back-propagation algorithm to determine the relationship between sewage
sample odors and their related BOD values. They reported that ANNs could be
used to classify the sewage samples collected from different locations of a WWTP.
562
Karaca and Ozkaya [22] developed an ANN-based model (NN-LEAP) for con-
trolling daily leachate flow rate in a municipal solid waste (MSW) landfill site.
They predicted daily leachate flow rates in MSW landfill area, with a correlation
coefficient of about 0.95 and a mean squared error (MSE) of 0.00168. The au-
thors concluded that the proposed ANN model provided an effective prediction of
daily leachate discharges and showed a reliable and fast outputs in controlling and
management of the flow rate levels in a MSW landfill area. Yetilmezsoy [23] deter-
mined optimum body diameter (OBD) of air cyclones using a two layer ANN and
a new empirical model for 505 different artificial scenarios given in a wide range
of five operating variables selected in the cyclone design. He compared the pre-
dicted results obtained from each proposed approach with the well-known Kalen
and Zenz’s model. The study concluded that maximum diameter deviations from
Kalen and Zenz’s model were recorded as only 1.3 cm and 0.0022 cm for the em-
pirical model and ANN outputs, respectively. Although both approaches produced
promising results, the ANN model provided a speedy and practical manner and
exhibited more robust outputs in determining OBD values. Finally, a two-layer
ANN approach and a linear-nonlinear study were performed for the determination
of single droplet collection efficiency of countercurrent spray towers in the work of
Yetilmezsoy and Saral [24]. They concluded that predicted results obtained from
linear–nonlinear regression analysis and the ANN model were in agreement with
the theoretical data, and all predictions proved to be satisfactory with a correlation
coefficient of about 0.921 and 0.99, respectively.
In the literature, however, a limited number of studies are available on acti-
vated sludge process using the ANN approach. One of the recent applications,
based on the modeling of activated sludge process and using ANNs, can be found
in the study of Moral [25]. He modeled three activated sludge-type treatment
plants by using ANN models. Initially, a hypothetical wastewater treatment plant
constructed in the Simulation of Single-Sludge Processes for Carbon Oxidation,
Nitrification & Denitrification (SSSP) program, which is an implementation of the
Activated Sludge Model No 1 (ASM1), was used as the source of input and output
data. The other systems were actual treatment plants, Ankara Central Wastewater
Treatment Plant, ACWTP and Iskenderun Wastewater Treatment Plant. The de-
fault dynamic data set of SSSP having two types of data processing (steady-state
solution and dynamic solution) was used in the modeling process. When develop-
ing ANN model from the SSSP Simulator default dataset, three system variables,
such as flow rate, particulate products and autotrophic biomass, were taken into
consideration. The author concluded that ANN could accurately model activated
sludge process, as indicated by simultaneous studies using the SSSP program.
Another study on a biological treatment process using ANNs was devoted to
the relationships between the effluent biological activity of a contact stabilization
process of Hawalli WWTP in Kuwait, with a flow rate of 360 m3 /day and micro
fauna communities distribution by Al-Mutari et al. [26]. They utilized ANNs to
investigate such relationships using both back-propagation and general regression
algorithms. In the study, the micro fauna distribution data of a contact stabilization
process were used in an ANN system to model and predict the biological activity
of the effluent. The authors optimized the architecture of the back-propagation
neural network (BPNN) model in four steps. They examined the test set size,
563
learning and momentum rates, number of neurons in the hidden layers and the
initial weight, respectively. In the optimization study, six different neural network
architectures were trained to find the optimum architecture for the BPNN model.
In addition, they also performed a sensitivity study to test the general capability
of the optimum BPNN model, by estimating the significance of each input variable
individually. The study concluded that the genetic adaptive general regression
neural network (GRNN) model could be used as a powerful and simple tool for the
modeling process.
Environmental preservation efforts and technological developments have re-
sulted in stringent discharge standards. To achieve an optimum control and man-
agement in wastewater treatment plants, new concepts involving efficient operation
and design should be developed and understood. A model is a simplified represen-
tation of a system at some particular point in time or space intended to promote
understanding of the real system. One of the real benefits of modeling is its ability
to accomplish a time and space compression between the inter-relationships within
a system. Hence, a high quality representative model can provide a favorable so-
lution in the process control of wastewater treatment plants, helps to explain the
real process performance and develop a continuous control strategy for this type of
technologies.
The ANN models can show remarkably good success in the modeling of highly
nonlinear systems into which the activated sludge process can also be included.
However, special attention should be paid when choosing the input variables and
their demonstrated relationships to the output parameters of the model. Moral [25]
emphasized that the ANNs need intensive measurement of system variables only for
the model development phase. This measurement period is needed for introducing
system into an ANN model being developed. After the development of an ANN
model of the system is complete, measurements will be done less frequently than
before, mainly for cross checking the response of the ANN model with the actual
system response. He also stated that the current activated sludge models, such
as ASM1, ASM2, ASM3 and ASM2d, are deterministic in character and are con-
structed by proceeding from fundamental biokinetics. However, calibrating these
models is extremely time-consuming and laborious.
Although deterministic models (also called white-box models) give a good in-
sight into the mechanism, they require a lot of hard work before being applied to
a specific wastewater treatment plant. In addition, kinetic parameters and influ-
ent wastewater characteristics may display some fluctuations in different periods
of time. Because the required reactor volume is highly dependent on both kinetic
parameters and influent wastewater characteristics, an ANN model will provide
a speedy and practical manner for the control engineer to make a modification
and/or a supplement in the volume of the reactor, and also to develop an effective
continuous monitoring tool with regard to the discharge standards. Therefore, ar-
tificial intelligence techniques using neural networks have become a popular and a
very attractive issue for modeling and controlling environmental processes. Fur-
thermore, calibration of ANN models is easier than in white-box models, as there
are fewer parameters used in the model development process. When the measured
variables start showing difference with the response of ANN, the model can be
re-trained, using the newer data employed for cross checking. This process can be
564
automated by embedding the ANN model in an expert system that controls the
complete system.
Consequently, potential advantages and benefits of the ANN technique have
been highlighted by many substantial research activities above. For this reason, an
ANN model, a preferable and popular modeling technique, has also been attempted
in the determination of CMASRV values in addition to the empirical study. The
major contributions of this study are (1) the formulation of a model for completely
mixed activated sludge reactor volume (CMASRV); (2) implementation of vari-
ous ANN schemes for obtaining values of the parameters in the model; and (3)
comparison of obtained results.
2. Definition of the Models

2.1 The set of steady-state activated sludge tank design
equations
The set of steady-state activated sludge tank design equations in simple terms
describe the relationship between important measurable quantities, such as the
wastewater flow rate, the sludge concentration in the activated sludge tanks and
the volume of the aerobic activated sludge tanks to be used for organic matter
removal. In consulting companies involved in designing activated sludge plants,
such sets of steady-state design rules (or rules of thumb, because that is what
these simplified mass balances represent) would, indeed, be used to obtain a plant
design given a set of wastewater specifications. These rules are then typically
included in a spreadsheet program, allowing the designer to modify the value of
the constants used. Such modifications are necessary, for example to describe the
effect of temperature on the decay rate of the biomass. For the designer, the whole
point of applying these empirical design equations is to have at least some kind of
justification of a plant design proposed to a client: Indeed, the influence of selected
parameters (i.e. decay rate, biomass yield, etc.) on the plant design (e.g. activated
sludge tank volume) can be justified in a design spreadsheet.
A schematic diagram of a completely mixed activated sludge process is illus-
trated in Fig. 1. The mass balance includes the flow rates for the mass of the con-
stituent entering and/or leaving the system, and appropriate reaction rate terms
for the depletion or production of the constituent within the system [27]. The mass
balance equation for the net cell production can be written as follows:
Waste sludge mass flow rate = Cell input rate − Cell output rate → PX =
= QW Xr = Y Q(Si − Se ) − kd XV, (1)
where PX is the waste (excess) sludge mass flow rate (kg MLVSS/day), QW
(m3 /day) is the waste sludge volumetric flow rate, Xr is the concentration of the
cells in recirculation line (kg MLVSS/m3 ), Y is the growth yield coefficient (mg
MLVSS/mg BOD5 ), Q is the wastewater flow rate (m3 /day), Si is the influent sub-
strate (BOD5 ) concentration (mg BOD5 /L), Se is the effluent substrate (BOD5 )
concentration (mg BOD5 /L), kd is the microorganism endogenous decay coefficient
(day−1 ), X is the concentration of the cells in the reactor and V is the volume of
the completely mixed activated sludge reactor.
565
Fig. 1 The flow diagram of a completely mixed activated sludge process.
In Eq. (2), the mean cell residence time (MCRT) or sludge age (θC ) is written
as a ratio between the mass of cells in the reactor and mass of cells wasted per day
by neglecting the concentration of the cells in effluent (Xe).
M ass of cells in the reactor XV XV XV

θC = = = =
M ass of cells wasted per day QW Xr PX Y Q(Si − Se ) − kd XV
(2)
The mean cell residence time (sludge age) equation is rearranged and the volume
of the reactor (V ) is obtained in Eq. (5).
1 Y Q(Si − Se ) − kd XV
= (3)
θC XV
1 Y Q(Si − Se )
+ kd = (4)
θC XV
Y Q(Si − Se )θC
V = . (5)
(1 + kd θC ) X
The hydraulic retention time (hour), food to mass ratio (kg BOD5 /kg MLVSS.day)
and volumetric organic load (kg BOD5 /m3 .day) are calculated by the following
equations:
V
Hydraulic Retention Time (HRT) = θ = (6)
Q
F Q(Si − Se )
F ood to Mass ratio = = (7)
M XV
Q(Si − Se )
V olumetric organic load = LV = . (8)
V
The overall goal of the activated-sludge process is to remove substances that have
a demand for oxygen from the system. This is accomplished by the metabolic
reactions (synthesis-respiration and nitrification) of the microorganisms. Bacteria
constitute a majority of microorganisms present in activated sludge. Some bacteria
are strict aerobes, they can only live in the presence of oxygen. Hence, an aeration
source is required to ensure that adequate oxygen is fed into the tank(s) and that
566
the appropriate mixing takes place. This source may be provided by pure oxygen,
compressed air or mechanical aeration. Just as there are modifications in the shape
and number of aeration tanks that can be used in the activated-sludge process,
different equipment systems exist to deliver air or oxygen into aeration tanks [28].
Briefly, the process design and the mode of plant operation are highly dependent
on oxygen requirements. Theoretical oxygen requirement can be calculated by
knowing BOD of wastewater and the amount of organisms wasted from the system:
µ ¶ · ¸ · ¸
kg ultimate BOD utilized BOD of organisms
O2 = − . (9)
day per day wasted per day
The BOD of a mole of cells is assumed as follows:

bacteria
C5 H7 N O2 + 5O2 −→ 5CO2 + 2H2 O + N H3 (10)
µ ¶
O2 5 × 32 g 160 g O2
= = = 1.42 . (11)
cells 113 g 113 g − cell
Consequently, the BOD of cells is equal to
BOD = 1.42(cells) (12)
µ ¶ · ¸ · ¸
kg ultimate BOD utilized organisms
O2 = − 1.42 (13)
day per day wasted per day
µ ¶ µ ¶ µ ¶ µ ¶ µ ¶
kg dS dX Q(Si − Se ) XV
O2 = − 1.42 = − 1.42 , (14)
day dt U dt f θC
where f is the ratio between BOD5 and ultimate BODU for a known substrate
decay rate (k = 0.1 day −1 ) and the value of 1.42 is the ultimate biochemical oxygen
demand of per kg of bacteria cells (molecular formula: C5 H7 NO2 ).
Using Eq. (14), the actual air requirement of the system can be calculated for
a certain oxygen diffusion efficiency. Eqs. (1), (6), (7), (8) and (14) were used
in head-to-head performance comparisons between the steady-state design set, the
ANN model and the proposed empirical model outputs for various waste sludge
mass flow rates (PX ), food to mass ratios (F/M ), hydraulic retention times (HRTs),
volumetric loads (LV ) and oxygen requirements (ORs), respectively. These com-
parisons are presented in Results and discussion. Projection parameters and some
kinetic coefficients considered in the design of conventional, completely mixed ac-
tivated sludge reactors are given in Tab. I. Based on the kinetic coefficients of the
conventional, completely mixed activated sludge process, the type of wastewater
considered in this study is the municipal wastewater (k = 0.1 day−1 , Y = 0.6 mg
MLVSS/mg BOD5 , kd = 0.060 day−1 ).
2.2 The proposed model

The proposed empirical model was developed on the basis of seven main biolog-
ical variables: Wastewater flow rate (Q), growth yield coefficient (Y ), sludge age
567
Projection Parameters and Kinetic Unit Operating

Coefficients Ranges
Hydraulic retention time (HRT) [hour] 4–8
Sludge age (θC ) [day] 5 – 15
Volumetric organic load (LV ) [kg BOD5 /m3 .day] 0.8 – 2.0
Food to mass ratio (F/M) [kg BOD5 /kg MLVSS.day] 0.2 – 0.6
Concentration of mixed liquor sus- [mg MLSS/L] 3000 – 6000
pended solids (MLSS)
Recirculation ratio (r) [%] 25 – 100
−1
Microorganism endogenous decay [day ] 0.040 – 0.060
rate (kd )
Growth yield coefficient (Y ) [mg MLVSS/mg BOD5 ] 0.4 – 0.8
Tab. I Projection parameters and some kinetic coefficients considered in the design
of CMASRV [29].
(θC ), microorganism decay coefficient (kd ), biomass concentration (X), influent

substrate (BOD5 ) concentration (Si ) and effluent substrate (BOD5 ) concentration
(Se ). The computational analysis was carried out for 35 different wastewater flow
rates (Q), 41 different conversion coefficients (Y ), 41 different sludge ages (θC ),
36 different microorganism endogenous decay coefficients (kd ), 61 different biomass
concentrations (X), 41 different influent substrate (BOD5 ) concentrations (Si ), and
71 different effluent substrate (BOD5 ) concentrations (Se ).
In the first step of the computational analysis, wastewater flow rates were se-
lected between 4000 – 50000 m3 /day. Growth yield coefficients were run between
0.40 – 0.80 mg MLVSS/mg BOD5 . Sludge ages were considered between 5 – 15
days. Microorganism endogenous decay coefficients were carried out between 0.040
– 0.075 day−1 . Biomass concentrations were selected between 2400 – 5400 mg
MLVSS/L. Influent and effluent substrate concentrations were run between 200 –
400 mg BOD5 /L and 6 – 20 mg BOD5 /L, respectively. Some extreme values for
each variable were also considered with their respective ranges to enhance the pre-
diction flexibility of the models. When calculating the volume of reaction zone, no
specific safety factor has been considered in model construction.
In the second step, seven parameter equations presented in Tab. II were de-
veloped, based on the power-law correlation for results obtained from the compu-
tational analysis. The parameter equations were derived, based on the selected
particular values (Q = 50000 m3 /day, Y = 0.60, θC = 10 days, kd = 0.060 day−1 ,
X = 4000 mg MLVSS/L, Si = 300 mg BOD5 /L, Se = 20 mg BOD5 /L) of main
biological variables (Tab. III).
Parameter equations given in Tab. II were combined (Eq. (15)) and new results
were calculated computationally for different biological variables selected in the
first step of computational analysis. Since the combined equations were obtained
as the geometric mean of seven parameter equations, the common exponent was
determined at about 0.143. The multiplication of individual coefficients (0.2625 ×
21875 × 2920.3 × . . . × 14923) yields a common coefficient of about 1.696 × 1024 .
Finally, the combined equation developed from parameter equations is written as
568
Constant Parameters Parameter Equations
Variable Parameters R2
Q Y θC kd X Si Se
Q (m3 /day) — 0.60 10 0.060 4000 300 20 V = 0.2625 Q 1.0000
Y (mg MLVSS/mg BOD5 ) 50000 — 10 0.060 4000 300 20 V = 21875 Y 1.0000
0.6481
θC (day) 50000 0.60 — 0.060 4000 300 20 V = 2920.3 θC 0.9974
kd (day −1 ) 50000 0.60 10 — 4000 300 20 V = 4789.9 kd −0.3571 0.9971
X (mg MLVSS/L) 50000 0.60 10 0.060 — 300 20 V = 50000000 X −1 1.0000
Si (mg BOD5 /L) 50000 0.60 10 0.060 4000 — 20 V = 28.291 Si 1.0762 1.0000
Se (mg BOD5 /L) 50000 0.60 10 0.060 4000 300 — V = 14923 Se −0.0413 0.9745
Tab. II Relationship between reactor volume and main biological variables used in the computational analysis.
Main Biological Variables Value

Q: Wastewater flow rate (m3 /day) 50000
Y : Growth yield cefficient (mg MLVSS/mg BOD5 ) 0.60
θC : Sludge age (day) 10
−1
kd : Microorganism endogenous decay coefficient (day ) 0.060
X: Biomass concentration (mg MLVSS/L) 4000
Si : Influent substrate concentration (mg BOD5 /L) 300
Se : Effluent substrate concentration (mg BOD5 /L) 20
Tab. III Particular values of main biological variables used in derivation of both parameter and correction equations.
569
follows:
· µ ¶µ ¶ µ ¶¸0.143
¡ ¢ 0.6481 1 1 1.0762 1
Vce = 1.696x1024 (Q) (Y ) (θC ) (Si ) ,
kd0.3571 X Se0.0413
(15)
where Vce is the completely mixed activated sludge reactor volume (CMASRV)
(m3 ), Q is the wastewater flow rate (m3 /day), Y is the growth yield coefficient (mg
MLVSS/mg BOD5 ), θC is the sludge age (day), kd is the microorganism endogenous
decay coefficient (day −1 ), X is the biomass concentration (mg MLVSS/L), Si is
the influent substrate concentration (mg BOD5 /L), Se is the effluent substrate
concentration (mg BOD5 /L).
In the last step, new results obtained from Eq. (15) were corrected for seven
correction equations of each biological variable (Q, Y, θC , kd , X, Si and Se ). First,
the particular values of each biological variable (Tab. III) were divided by the
minimum values of their respective variables (i.e. 50000/4000 = 12.5 for Q = 4000
m3 /day, 0.6/0.4 = 1.5 for Y = 0.40, 10/5 = 2 for θC = 5 days, 0.06/0.04 = 1.5
for kd = 0.040 day−1 , 4000/2400 = 1.6667 for X = 2400 mg MLVSS/L, 300/200
= 1.5 for Si = 200 mg BOD5 /L, and 20/6 = 3.3333 for Se = 6 mg BOD5 /L).
Then, CMASRV values for the minimum values of each variable were determined
from both Eqs. (5) and (15). For instance, the computation of CMASRV values
from Eqs. (5) and (15) yields 1050 and 9163.66 m3 for the minimum flow rate
value of 4000 m3 /day. As given in Tab. II, other variables were kept at their
particular levels: Y = 0.60, θC = 10 days, kd = 0.060 day−1 , X = 4000 mg
MLVSS/L, Si = 300 mg BOD5 /L, Se = 20 mg BOD5 /L. Following this step,
volume ratios obtained for the minimum values of each biological variable (others
were kept at their particular levels) were computed (i.e. 9163.66/1050 = 8.727 for
Q = 4000 m3 /day, 12409.26/8750 = 1.418 for Y = 0.40, 12329.09/8076.923 = 1.526
for θC = 5 days, 13424.79/15000 = 0.895 for kd = 0.040 day−1 , 14146.57/21875
= 0.647 for X = 2400 mg MLVSS/L, 12354.04/8437.5 = 1.464 for Si = 200 mg
BOD5 /L, and 13245.37/13781.25 = 0.961 for Se = 6 mg BOD5 /L). Finally, the
exponents were calculated for each biological variable by using the predetermined
values (i.e. 12.5a = 8.727, a = 0.8578, Q = 4000 m3 /day). Computation of other
exponents yields 0.8617 for Y, 0.6102 for θC , -0.2736 for kd , -0.8533 for X, 0.9404
for Si , and -0.033 for Se . After determining exponents, seven correction equations
were derived. Transferring correction equations to Eq. (15) yields the corrected
combined equation as follows:
h¡ ¢ ³ ´¡ ¢ ³ ´i0.143
0.6481 1 1 1.0762 1
1.696x1024 (Q) (Y ) (θC ) 0.3571
kd X (S i ) 0.0413
Se
Vcce = ³ ´0.8578¡ ¢ ³ ´0.6102³ ´−0.2736¡ ¢ ³ ´0.9404³ ´−0.033
50000 0.60 0.8617 10 0.060 4000 −0.8533 300 20
Q Y θC kd X Si Se
(16)
The exponents of each variable in Eq. (16) were rounded to simplify the model
structure. For instance, the exponents of Q and Y were calculated as 1.0008 and
1.0047, respectively. Similarly, the exponents of θC , kd and X were determined
to be 0.7029, 0.3247 and 0.9963, respectively. Furthermore, the computation of
exponents yields 1.0943 for Si , and 0.03891 for Se . After rounding the exponents
in the corrected combined equation (Eq. (16)), the proposed model is rewritten
as a function of the main biological variables and the final form of the proposed
570
model is obtained as follows:

0.70 1.1
(0.294) (Q) (Y ) (θC ) (Si )
Vm = 0.33 0.04 . (17)
(X) (kd ) (Se )
Fouad and Bhargava [30] developed a steady-state mathematical model for describ-
ing the completely mixed biofilm-activated sludge reactor (also known as the hybrid
reactor) by considering Monod kinetics expressions and Fickian’s diffusion theory
for substrate in biofilm. They focused on deriving an easy mathematical model
that can be used to design and control such a system based on fundamental prin-
ciples and kinetics of both attached and suspended cultures simultaneously. In the
study, the system was assumed to run under a steady-state condition for biomass
and substrate, with a rate limiting substrate concentration through the reactor.
They developed the model based on mass balance equations and also presented a
simple set of curves to obtain key parameters (φx and φf ). The study showed that
representing a model in more than one form may provide an explicit solution of the
system and also help to design and operate the reactor under different conditions
for any given set of input parameters. Since modeling can be considered a good
tool to explain the performance and the characteristics of the system, a number of
attempts at developing a representative model to be a different function of process
parameters can promote the existing calculation method under different operating
conditions.
Another study by Fouad and Bhargava [31] was devoted to sludge production
and settleability in biofilm-activated sludge process. They ran a steady-state hybrid
reactor at different stages of suspended biomass concentration (X) under constant
values of influent substrate concentration (Si ) and hydraulic retention time (HRT).
The authors supported their experimental study by a mathematical expression for
describing the sludge yield in the system under different conditions. They derived
their mathematical equations and made substitutions based on the well-known
mass balance equations cited in the literature.
In both studies of Fouad and Bhargava [30,31], the substrate balance for a
hybrid system (completely mixed biofilm-activated sludge reactor) under steady-
state conditions is computed in following equations as follows:
KSe
VX = Q(Si − Se ) − aV J (18)
Ks + S e
KSe
VX = QR(Si − Se ) (19)
Ks + S e
aV J = Q(1 − R)(Si − Se ), (20)
where V is the reactor volume (m3 ), X is the suspended biomass concentration (kg
MLVSS/m3 ), K is the maximum specific rate of substrate utilization (day−1 ), Se is
the effluent substrate concentration (kg BOD5 /m3 ), Ks is the Monod half-velocity
coefficient (kg/m3 ), Q is the wastewater flow rate (m3 /day), R is the fraction of
substrate utilized by suspended biomass, Si is the influent substrate concentration
(kg BOD5 /m3 ), a is the specific surface area of biofilm reactor (m−1 ), and J is the
substrate flux into biofilm per unit area (kg/m2 .day).
571
The ratio of the suspended culture to the total mass of wasted sludge and the
ratio of active mass of biofilm to the total mass of wasted sludge were assumed
as φx and φf , respectively. Although the parameters φx and φf have no physical
meaning (dummy parameters), they have been considered key parameters that
capture the system [30]. In the study, the key parameters and the suspended
biomass concentration are defined as follows:
VX
φx = (21)
QY (Si − Se ) − V Xkd − Y aV Jkd /bt
Y aV J/bt
φf = (22)
QY (Si − Se ) − V Xkd − Y aV Jkd /bt
φx QY (Si − Se )
X= , (23)
V (1 + kd φx + kd φf )
where φx is the first dummy variable assumed as the ratio of the suspended culture
to the total mass of wasted sludge (day), φf is the second dummy variable assumed
as the ratio of active mass of biofilm to the total mass of wasted sludge (day), kd
is specific decay rate (day−1 ), bt is the sum of specific decay and shear loss rates
(day−1 ), Y is the growth yield coefficient (mg MLVSS/mg BOD5 ) and V, X, Q, Si ,
Se , a, J are the other parameters defined in Eqs. (18), (19) and (20).
When the hybrid reactor runs at a high concentration of suspended biomass (X)
and low level of substrate (Si ), the substrate flux into biofilm (J) is negligible and
the biofilm is inactive for COD removal irrespective of its amount in the reactor [31].
This definition was also emphasized in the other study of Fouad and Bhargava [30].
They stated that the hybrid reactor would act as a pure attached or pure suspended
growth reactor if there is no attached biofilm. It was reported that each of φx and
φf must have a real value for a hybrid system. However, φf will be equal zero if the
substrate flux into biofilm (J) is negligible. In addition, the key parameter φx will
be equal to the sludge age of the pure suspended growth (θC ) by neglecting the J
value. This expression is presented in Eq. (24), where bs is the specific loss rates
(day−1 ). If the J value equals zero, φf will be equal to zero in Eq. (22). Hence,
θC will be equal to φx in Eq. (24) at last.
³ ´
VX
QY (Si −Se )−V Xkd −Y aV Jkd /bt φx
θC = h ³ ´ i= . (24)
1− Y aV J/bt
b (1 − φf bs )
QY (Si −Se )−V Xkd −Y aV Jkd /bt s
Therefore, the hybrid reactor will run as pure suspended-growth reactor under
these conditions (J = 0, R = 1 and φf = 0). After these definitions, an illustrative
comparison was made between the results of the proposed empirical model given
in Eq. (17) and the steady-state mathematical study by Fouad and Bhargava [31]
for a pure suspended-growth condition. This comparison is presented in Results
and discussion.
2.3 Artificial Neural Network (ANN) approach

In this study, Neural Network Toolbox V4.0 of MATLAB° R was used to predict
CMASRV values. Main biological parameters, which are effective in determining
572
completely mixed activated sludge reactor volume (CMASRV), such as wastewater

flow rate, conversion coefficient, sludge retention time, microorganism decay coef-
ficient, biomass concentration, influent and effluent substrate concentrations, were
taken as model input. Volume values calculated from the steady-state model were
used as a target vector for the ANN approach. The ANN model used an inputs
vector (7 × 1000) of seven main biological variables and target vector (1 × 1000) to
predict volume values obtained from the theoretical model. Seven input variables
(Q, Y, θC , kd , X, Si and Se ) and their operating ranges considered in the scenarios
are defined in Tab. IV. The data were divided into an input matrix [p] and a target
matrix [t].
Variables Units Min. Max. Range Average SDa Median

Wastewater flow [m3 /day] 3000 100000 97000 34545 25600 24600
rate (Q)
Growth yield coef- [mg MLVSS/mg 0.35 0.80 0.45 0.58 0.13 0.59
ficient (Y ) BOD5 ]
Sludge age (θC ) [day] 5 15 10 8.68 2.55 9
Microorganism [day −1 ] 0.040 0.080 0.040 0.057 0.011 0.056
endogenous decay
coefficient (kd )
Biomass concen- [mg MLVSS/L] 2000 5500 3500 4115 850 3980
tration (X)
Influent substrate [mg BOD5 /L] 150 400 250 276 81 279
concentration (Si )
Effluent substrate [mg BOD5 /L] 4 20 16 11 4.61 10.1
concentration (Se )
a
Standard deviation.
Tab. IV Input variables and their operating ranges considered in the ANN
modelling.
In the first step, the data were loaded into the workspace from Microsoft° R
Excel, used as ODBC (Open Database Connectivity) data. Original network inputs
and targets given in the matrices p and t were normalized, using an algorithm code,
prestd. The normalized inputs and targets, pn and tn were gained as zero means
and unity standard deviation. The mean and standard deviations of the original
inputs and targets were defined before the network had been trained. Moreover,
the matrices containing the transformed input vectors and the principal component
transformation matrix were defined, respectively. All these vectors were used in
transformation for future inputs.
The next step, principal component analysis (PCA), was performed as an ef-
fective procedure for the determination of input parameters. In some situations,
the dimension of the input vector is large, but the components of the vectors are
highly correlated (redundant). It is useful in this situation to reduce the dimension
of input vectors. This technique eliminates those components that contribute the
least to the variation in the data set. Hence, an algorithm code, prepca, which
performs PCA was used to eliminate those components in the data set. Principal
components that contribute less than 0.1% to the total variation in the data set
573
were eliminated or it can be said that principal components which accounted for
99.9% of the variation were used. It was observed that there was no redundancy
in the data set and size of the transformed data after computation. This step was
followed by dividing the original data into training, validation and test subsets.
Most of ANN models are characterized by the summation function, activation
function, network topology and learning rule they employ [32]. In this study,
a three-layer neural network with tangent sigmoid transfer function (tan-sig) at
hidden layer and a linear transfer function (purelin) at output layer were used.
The data were randomly divided into training, validation and testing subsets. One
fourth of the data was randomly taken for the validation set, one fourth for the
testing set, and one half for the training set. Hence, 500 input sets were used for
the training and 250 input sets were used for the validation and test, respectively.
Application of some performance optimization techniques for the training data
set may help to further minimize the estimation error and to achieve better gener-
alizing ability, but the choice of the method mainly depends on the desired criterion
and main purpose for estimating generalization error, as well as on the dimension
of the data set. Cross-validation can be used simply to estimate the generalization
error of a given model, or it can be used for model selection by choosing one of
several models that has the smallest estimated generalization error. This method
may also be used to choose the number of hidden units or to choose a subset of
the inputs (subset selection). In k-fold cross-validation, the data are divided into
k subsets, and then the network is trained k times, each time leaving out one of
the subsets from training. On the other hand, in the single random split method
(or split-sample validation), only a single subset (the validation set) is used to
estimate generalization error, instead of k different subsets [33]. Briefly, the cru-
cial distinction between cross-validation and single random split methods is that
with cross-validation, after guessing the best subset, the linear regression model
is trained for that subset, using all training cases. But, with single random split
validation, only the training set is ever used for training. However, it is noted that
choosing the best subset is not the same thing as getting the best generalization
[33]. The distinction between cross-validation and single random split validation
is also extremely important because cross-validation is markedly superior for small
data sets, as demonstrated by Goutte [34]. Since the main purpose of the present
work was not really to choose the best subset as done in cross-validation, split-
sample validation method was used to estimate generalization error on the present
validation data. Moreover, considering the high dimension of the present data set,
the single random split method was chosen instead of cross-validation in this study,
as was the case in several previous studies [22-24, 35-37].
In this study, Multi Layer Perceptron (MLP) architecture of ANN models was
used with various training algorithms. The MLP-based ANN model consists of
main operational units called neurons. The neurons are arranged in parallel as
layers, which are an input layer having NI number of neurons, a hidden layer with
NH neurons, and an output layer with NO neurons. There may be more than
one hidden layer in the general network architecture. But using more than one
hidden layer generally causes the network to memorize the training data set and
lose the ability to generalize the estimation capacity on the test data set [38]. One
of the most common types of MLP neural networks is the feed-forward, where
574
the information is transmitted in a forward direction only. The weighted sum of

the inputs is transferred to the hidden neurons, where it is transformed, using an
activation function. The outputs of the hidden neurons, in turn, act as inputs to
the output neuron where it undergoes another transformation [19]. Hamed et al.
[19] have reported that the linear activation function for the output neuron can
be selected since it is appropriate for continuous valued targets. Moreover, the
tangent sigmoid activation functions for the input and hidden neurons are needed
to introduce non-linearity into the network. Without non-linearity, hidden layers
would not make nets more powerful than just plain perceptrons. Based on the
above-mentioned facts, this study employed a three-layer feed-forward MLP-based
ANN model (7: NH : 1), with tangent sigmoid transfer function (tansig) at hidden
layer with NH neurons, and a linear transfer function (purelin) at output layer.
Their mathematical definitions of tansig and purelin are given in Eqs. (25) and
(26):
2
f (n) = −1 (25)
(1 + e−2n )
f (n) = n. (26)
The training set makes it possible to train the network on a representative set
of input/target pairs and obtain good forecasting results. This property of ANN
provides for getting a new output similar to correct output. The function train
carries out such a loop of calculation. In each pass, the function train proceeds
through a specified sequence of inputs, calculating the output, error and network
adjustment for each input vector in the sequence as the inputs are presented. It
is very difficult to know which training algorithm will be the fastest for a given
problem. It will depend on many factors, including the complexity of the problem
and the number of data points in the training set. Hence, a number of benchmark
comparisons of the various training algorithms were performed in this study.
Learning rate parameter may also play an important role in the convergence of
the network, depending on the application and network architecture. The learning
rate can be used to increase the chances of preventing the training process from
being trapped in a local minimum instead of the global minimum [19]. The larger
the learning rate, the bigger the step. If the learning rate is made too large, the
algorithm becomes unstable. If the learning rate is set too small, the algorithm
takes a long time to converge. Furthermore, momentum allows a network to re-
spond not only to the local gradient, but also to recent trends in the error surface.
Without momentum, a network may get stuck in a shallow local minimum [22]. In
this study, the learning rate and the momentum constant were used as 0.1 and 0.9,
respectively.
Almasri and Kaluarachchi [39] described the optimization of a neural network
as an important task of neural network-based studies. This operation plays an
important role in the performance of the network. Hence, an optimization was
carried out between the neuron number and mean squared error (MSE). Then, the
three-layer neural network was evaluated by the best training algorithm, selected
as the best of eight algorithms for training (Tab. VI).
In the last step, some analysis of the network response was carried out. The
entire data set was put through the network (training, validation and test) and
575
a linear regression between the network outputs and the corresponding targets
was performed. Firstly, the trained network was simulated. In the next step, the
network outputs were un-normalized and converted back into the original units
with an algorithm code, poststd. Briefly, the un-normalized network outputs (a)
were obtained in the same units as the original targets (t).
3. Results and Discussion

The proposed empirical model was solved for 1,000 different artificial scenarios
given in a wide range of biological variables. Scenarios were run for wastewater be-
tween 3,000 and 10,0000 m3 /day, for growth yield coefficient between 0.35 and 0.80
mg MLVSS/mg BOD5 , for sludge age between 5 and 15 days, for microorganism
endogenus decay coefficient between 0.040 and 0.080 day−1 , for biomass concentra-
tion between 2,000 and 5,500 mg MLVSS/L, for influent substrate concentration
between 150 and 400 mg BOD5 /L and effluent substrate concentration between 4
and 20 mg BOD5 /L. Results obtained from the predicted model agreed remarkably
well with the theoretical volume model based on mass balance equations.
The proposed model predicted very well volume values for completely mixed
activated sludge reactors and can be used in biological treatment process design
for various operating and biological variables. In computational calculations, very
small volume deviations were recorded, with a maximum deviation of about 7.17%
from the steady-state volume equation. The relationship between the proposed
model and the steady-state volume equation shows an excellent correlation (R2 =
0.9989), with the best linear equation of y = 1.0125x − 65.227 for CMASRV values
calculated computationally.
Consider a specific wastewater having the following characteristics and kinetic
coefficients [29]: Si = 250 mg BOD5 /L, Se = 6 mg BOD5 /L, kd = 0.060 day−1 ,
Y = 0.50 mg MLVSS/mg BOD5 , X = 3500 mg MLVSS/L, Q = 21600 m3 /day and
θC = 10 days. Assuming the reactor running in pure suspended growth conditions
(J = 0, R = 1 and φf = 0), the volume of the reactor yields
VX VX
φx = → J = 0 → φx =
QY (Si − Se ) − V Xkd − Y aV Jkd /bt QY (Si − Se ) − V Xkd
(27)
φx
θC = → φf = 0 → θC = φx (28)
(1 − φf bs )
VX Y Q(Si − Se )θC
θC = →V = . (29)
QY (Si − Se ) − V Xkd (1 + kd θC ) X
Fouad and Bhargava [30] reported the mathematical expression in Eq. (29) as the
well-known Lawrence and McCarty’s model for pure suspended growth condition.
Applying all known parameters in Eqs. (29) and (17), while considering the units
defined for both model structures, we obtain the following formulas:
Y Q(Si − Se )θC (0.50)(21600)(0.25 − 0.006)(10) ∼

Vsteady−state = = = 4706 m3
(1 + kd θC ) X [1 + (0.060)(10)] (3.50)
(30)
576
0.70 1.1
(0.294) (Q) (Y ) (θC ) (Si )
Vempirical = 0.33 0.04 =
(X) (kd ) (Se )
0.70 1.1
(0.294) (10000) (0.50) (10) (300) ∼
= 0.33 0.04 = 4651 m3 . (31)
(3500) (0.060) (6)
This illustrative comparison showed that the steady-state mathematical model was
in agreement with the proposed empirical model, with an approximation error
of about 1.2% for this specific wastewater characteristics and kinetic coefficient
values given as an example. In addition to this illustrative comparison, an example
set of illustrative comparison of the models for various random data derived from
operating ranges given by Muslu [29] is summarized in Tab. V.
In the ANN study, a three-layer network was formed and trained after the
definition of subsets. The training algorithms are a subset of the ones available in
the neural nets toolbox. (The names indicate the type: gradient with momentum,
named after the two authors, and Bayesian reduction.) ANNs are inspired by
biology, and attempt to emulate the learning processes in the brain. They contain
no preconceptions of what the model shape will be, so they are ideal for cases
with low system knowledge. They are useful for functional prediction and system
modeling where the physical processes are not understood or are highly complex.
However, the disadvantage of neural nets is that they require a lot of data to ensure
confidence in the results, so they are not suitable for small data sets. Also, with
higher numbers of inputs, the number of connections and hence the complexity
tend to increase rapidly. Moreover, as the complexity of the problem grows, more
neurons are needed in the recurrent layer for the network to do a good job.
It is very difficult to know which training algorithm will be the fastest for a
given problem. It will depend on many factors, including the complexity of the
problem, the number of data points in the training set, the number of weights and
biases in the network, the error goal, and whether the network is being used for
pattern recognition (discriminant analysis) or function approximation (regression).
Networks are used with a variety of different architectures and complexities, and
trained to a variety of different accuracy levels. The performance of the various
algorithms can be affected by the accuracy required of the approximation depend-
ing on the mean square error versus for several representative algorithms. When
the problem formulation has a combinatorial nature, definition of each process pa-
rameter results in a complex interaction of variables used in calculations. Hence,
a number of benchmark comparisons of the various training algorithms is needed
to choose the best suited algorithm to achieve good performance when address-
ing the laborious problems having an interative and non-linear nature. Generally
speaking, the LMA does not perform as well on pattern recognition problems as it
does on a realistic function approximation (or nonlinear regression) problems, and
is designed for least squares problems that are approximately linear. For this type
of problems, the LMA is best suited and the approximation must be very accurate.
However, the conjugate gradient algorithm and resilient algorithm (RP or Rprop ) all
provide fast convergence on the pattern recognition problems under consideration.
Moreover, the RP algorithm works well with all the pattern recognition problems.
Consequently, there are several algorithm characteristics that can be deduced from
the experiments. In general, in case of function approximation problems (or non-
577
Q Y (mg θC kd X Si Se Vsteady−state Vempirical Appr.
(m3 /day) MLVSS/mg (day) (day−1 ) (mg (mg (mg (m3 ) (m3 ) Error
BOD5 MLVSS/L BOD5 /L) BOD5 /L) (± %)
10000 0.76 14.2 0.048 5400 380 6.8 4435 4603 3.78
17200 0.44 10.0 0.062 3800 290 7.4 3474 3467 0.21
24800 0.50 11.0 0.055 3800 210 8.0 4518 4415 2.26
86100 0.35 9.0 0.048 3950 180 9.5 8175 7865 3.80
74000 0.40 8.5 0.052 3690 200 10.2 8975 8664 3.46
68000 0.64 7.0 0.066 4950 190 11.0 7535 7220 4.18
23300 0.77 6.3 0.061 5210 230 12.3 3412 3312 2.94
65400 0.62 9.9 0.044 3980 320 19.9 21084 21123 0.18
39700 0.51 11.1 0.070 5420 380 18.2 8442 8728 3.38
41000 0.58 13.0 0.071 4780 250 15.0 7903 8216 3.96
50400 0.69 11.0 0.079 4860 150 14.3 5715 5797 1.44
23500 0.70 13.0 0.062 4920 344 10.8 8019 8312 3.65
38000 0.78 13.3 0.040 5010 246 6.7 12291 12173 0.96
17600 0.80 12.6 0.041 4000 238 7.8 6732 6630 1.52
4000 0.36 7.0 0.053 5000 368 9.9 527 528 0.33
5200 0.46 8.9 0.068 3980 248 10.0 793 778 1.90
14700 0.55 7.9 0.079 3870 302 14.0 2927 2901 0.87
12300 0.58 6.6 4060 398 16.0 3235 3248 0.42
15800 0.71 7.7 0.050 4200 400 13.2 5744 5785 0.73
29800 0.53 10.1 0.048 4580 154 13.3 3300 3203 2.96
32600 0.40 10.5 0.074 5030 160 15.5 2213 2223 0.43
43400 0.61 11.0 0.061 4010 178 16.6 7014 6990 0.34
6000 0.63 6.0 0.072 2700 192 14.0 1044 1005 3.77
Tab. V An example set of illustrative comparison of the models for various random data derived from operating ranges given
by Muslu [29].
578
linear regression) that have combinatorial natures, the LMA will have the fastest
convergence. In many cases, trainlm is able to obtain lower mean square errors
than any of the other algorithms tested. However, trainlm performance is rela-
tively poor on pattern recognition problems. The trainbfg performance is similar
to that of trainlm. It does not require as much storage as trainlm, but the compu-
tation required does increase geometrically with the size of the network, since the
equivalent of a matrix inverse must be computed at each iteration.
Eight training algorithms were compared to select the best-fit training algo-
rithm. For all training algorithms, a three-layer network with a tangent sigmoid
transfer function (tan-sig) at hidden layer and a linear transfer function (purelin)
at output layer, were used. Ten neurons were used at the hidden layer for all train-
ing algorithms. The LMA, with a minimum mean squared error (MSE), was found
as the best of eight training algorithms (Tab. VI). Tab. VI also lists the training
algorithms that are tested and the acronyms used to identify them. Comparison
of the training algorithms is shown in Fig. 2.
The optimization scheme showed a loss on the optimality of the estimates/results
produced by some training algorithms. The benchmark comparison study resulted
in that the LMA was able to provide smaller MSE compared to other training algo-
rithms, such as the resilient (Rprop ) algorithm and conjugate gradient algorithms.
As shown in Tab. VI, the smallest MSE was obtained of about 0.000340851 for the
trainlm function. This was followed by the trainbfg with a MSE of 0.00165388. The
trainrp function is the fastest algorithm on pattern recognition problems. How-
ever, both trainrp and conjugate gradient algorithms, such as traincgf, traincgp
and traincgb, do not perform well on function approximation problems. These
algorithms yield more errors than the LMA. The loss of the optimality of the
estimates/results produced by some training algorithms was caused by the com-
binatorial nature and non-linear structure of the problem under scrutiny. Hence,
the complexity analysis of the problem was validated by the results of the training
algorithms used in the benchmark comparison.
The neural network used for training and prediction was implemented on the
Neural Network Toolbox V4.0 of MATLAB°. R The predictive performance was
measured by computing the mean squared error (MSE, a measure of the prediction
Mean Iteration
Training algorithms Algorithm Acronyms R2 Squared Error Number
(MSE)
Resilient (Rprop ) trainrp RP 0.993 0.00943153 100
Fletcher-Reeves conjugate gradient traincgf CGF 0.993 0.00955303 42
Polak- Ribiére conjugate gradient traincgp CGP 0.992 0.0116533 33
Powell-Beale conjugate gradient traincgb CGB 0.996 0.00400851 66
Levenberg-Marquardt trainlm LM 0.999 0.000340851 58
Scaled Conjugate Gradient trainscg SCG 0.997 0.00249949 98
BFGS Quasi-Newton trainbfg BFG 0.997 0.00165388 91
One Step Secant trainoss OSS 0.996 0.00626484 71
Tab. VI Comparison of eight training algorithms.
579
Fig. 2 Comparison of training algorithms for ten neurons used in the hidden layer.
precision). Basically, the learning in the network is achieved through an iterative

algorithm that minimizes the MSE between the desired and actual outputs. As
a result of this, the optimal architecture of the ANN model and its parameter
variation were chosen, based on the minimum value of the MSE of the training and
prediction set. All the processing in the neural network was done in the hidden and
output layers. The error of the network was computed by finding the MSE. This
is calculated as a difference between the target output and the network output as
follows:
Q Q
1 X 2 1 X 2
MSE = e (k) = [t (k) − a (k)] , (32)
Q Q
k=1 k=1
where MSE is the mean squared error, Q is the size of the input vector, t is
the corresponding target output, a is the network output, and e is the difference
between the target output and the network output.
The train function applies the inputs to the new network, calculates the out-
puts, compares them to the associated targets, and calculates a mean square error.
If the error goal is met, or if the maximum number of epochs is reached, the train-
ing is stopped, and train returns the new network and a training record. Otherwise
train goes through another epoch. During the adaptation phase, the training algo-
rithm received part of the data (inputs and outputs) and automatically developed
the ANN model. After development, the model could generate the appropriate
responses for simulations with varying levels of data input. When the learning was
complete, the neural network was used for prediction.
In optimization of the network, six neurons were used in the hidden layer as
an initial guess. With increasing neuron number, the network gave several local
580
minimum values, and different MSE values were obtained for the training set. Fig.
3 illustrates the effect of the number of hidden neurons on the MSE. It can be
easily seen that the MSE of the network was much higher for the six (MSE =
0.00807041) and seven (MSE = 0.00657556) numbers of neurons than those with
8 (MSE = 0.00388318) and 9 (MSE = 0.00370059). At 10 hidden neurons, the
error decreased significantly, from 0.00370059 to 0.000340851. At 11 hidden neu-
rons, the MSE decreased much more, from 0.000340851 to 0.000197303. When the
number of neurons exceeded 11, the MSE increased significantly, from 0.000197303
to 0.00319102. Hence, the neural network containing 11 hidden neurons (MSE =
0.000197303) was chosen as the best case (Fig. 3). In this optimization study, MSE
increased significantly when the number of neurons exceeded 11. This increment
would probably relate to the MSE performance index, and the characteristics of
the input vector [p]. Finally, it can be stated that networks are sensitive to the
number of neurons in their hidden layers. Too few neurons can lead to underfitting.
Too many neurons can contribute to overfitting, in which case all training points
are well fitted, but the fitting curve takes wild oscillations between these points.
Therefore, a number of different optimization strategies should be implemented to
obtain the most appropriate architecture of the network for its future performance.
Finally, the optimal neural network structure for prediction of CMASRV values
is shown in Fig. 4: a three-layer network, with tangent sigmoid transfer function
(tan-sig) at hidden layer with 11 neurons and a linear transfer function (purelin)
at output layer.
Fig. 3 Effect of number of hidden neurons on the mean squared error MSE for the
LMA.
The training was stopped after 93 iterations (TRAINLM, Epoch 93/100) for the
LMA because the differences between training error and validation error started
to increase. Detailed descriptive statistics of the proposed ANN model subsets are
given in Tab. VII. The linear regression between ANN testing outputs and corre-
581
Fig. 4 Optimal neural network structure for prediction of CMASRV values.
sponding targets showed that ANN outputs (predicted values) were obviously in
agreement with the steady-state volume model (Fig. 5). Fig. 6 shows the agree-
ment between ANN testing outputs (predicted values) and the steady-state volume
model. ANN outputs showed a very small volume deviations with a maximum de-
viation of about 6.89% from the theoretical data. The R2 value was almost 1.
It is believed that ANN-based techniques, which have recently been applied to
various problems, may provide a good alternative to statistical and theoretical tech-
niques and also iterative problems because of their speed and capability of learn-
ing, robustness, predictive capabilities, non-linear characteristics, non-parametric
regression capabilities, generalization properties and easiness of working with high-
dimensional data. The last decade has shown that this technology can provide
useful tools for solving practical engineering problems [40,41].
In addition to volume comparison, head-to-head performance comparisons be-
tween the steady-state design set, the ANN model and the proposed empirical
model outputs for various waste sludge mass flow rates (PX ), food to mass ra-
tios (F/M), hydraulic retention times (HRTs), volumetric loads (LV ) and oxygen
requirements (ORs) were also performed in this study. Maximum approximation
errors between proposed volume models (empirical and ANN) and the steady-state
volume equation are given in Tab. VIII. Agreements between proposed approaches
and the steady-state design set for these projection parameters are depicted in
Figs. 7 and 8.
582
Descriptive statistics ANN subsets
Training Validation Testing
Data Statistics
Size 500 250 250
Regression variable results (y = ax + b)
Coefficient of multiple determination (R2 ) 0.9998 0.9988 0.9994
Adjusted coefficient of multiple determination (Ra2 ) 0.9997 0.9988 0.9994
Regression Coefficient, RC (a) 0.9998 0.9928 1.0006
Standard error for RC 6.335 × 10−5 2.15 × 10−3 1.557 × 10−3
t-ratio for RC 1578.35 461.65 642.81
p value for RC 0.0 0.0 0.0
Constant Term, CT (b) 0.4866 49.21 1.1045
Standard error for CT 6.536 21.937 15.985
t-ratio for CT 0.0745 2.2432 0.0691
p value for CT 0.941 0.026 0.945
Residual statistics
Residual tolerance 1 × 10−10 1 × 10−10 1 × 10−10
Sum of residuals 1.801 × 10−4 6.389 × 10−6 -5.622 × 10−4
Average residual 3.603 × 10−7 2.56 × 10−8 -2.25 × 10−6
Minimum residual -420.28 -1978.54 -620.11
Maximum residual 487.62 898.37 1566.86
Residual sum of squares (absolute) 473717.9 13831381.7 6966643.4
Standard error of the estimate 97.702 236.16 167.61
Durbin-Watson statistics 1.9896 1.8571 1.9335
Tab. VII Detailed descriptive statistics of the proposed ANN model subsets.
583
Fig. 5 The correlation between ANN testing outputs and the steady-state volume
model.
Fig. 6 Agreement between ANN testing outputs and the steady-state volume model.
4. Conclusions
A three-layer ANN approach and a new empirical model that could predict the com-
pletely mixed activated sludge reactor volume (CMASRV) were described. 1,000
different artificial scenarios given in a wide range of seven main biological vari-
584
Fig. 7 Head-to-head performance comparisons of the steady-state design set, the

ANN model and the proposed empirical model outputs for various waste sludge mass
flow rates (PX ) and food to mass ratios (F/M).
ables were solved, based on both proposed empirical equation and the ANN model.
Results obtained from these approaches were compared with the volume model
based on a set of steady-state design equations. Moreover, significant points of the
proposed approaches were evaluated.
Both development of a new mathematical model and creation of an ANN model
on CMASRV values have promoted the existing calculation method and given fa-
vorable results for desired values. The steady-state volume equation was rewritten
as a new function of wastewater flow rate, conversion coefficient, sludge retention
time, microorganism decay coefficient, biomass concentration, influent and effluent
substrate concentrations. Both the proposed empirical model and ANN outputs
have shown a significant prediction on CMASRV values under 1,000 different ar-
tificial scenarios. Choosing an appropriate data set depending on ranges of input
variables has shown a favorable performance for stochastic modeling approaches
presented in this study.
585
Fig. 8 Head-to-head performance comparisons of the steady-state design set, the

ANN model and the proposed empirical model outputs for various hydraulic reten-
tion times (HRTs), volumetric loads (LV ) and oxygen requirements (ORs).
586
Max. approximation
errors (%)
Projection parameters Units Evaluation ranges
Empirical ANN
model approach
Waste sludge mass flow [kg MLVSS/day] 265.87 – 18280 6.57 6.10
rate (PX )
Food to mass ratio [kg BOD5 /kg 0.20 – 0.60 6.17 5.57
(F/M) MLVSS.day]
Hydraulic retention [hour] 3–5 6.20 5.70
time (HRT)
Volumetric organic load [kg BOD5 /m3 .day] 0.81 – 1.97 6.06 5.57
(LV )
Oxygen requirement [kg O2 /day] 663.51 – 37118.18 6.44 5.73
(OR)
Tab. VIII Approximation errors between proposed volume models (empirical and
ANN) between the steady-state design set.
A number of attempts at developing a representative prediction model may obvi-

ously represent the real condition and help explain the real treatment performance
and develop a continuous control strategy for treatment technologies. Although
the model results give promising outputs, its advantages should be utilized by de-
veloping new training algorithms or including new variables to evaluate the effects
of each model component on the corresponding target values that will be consid-
ered in future studies. For this purpose, a sensitivity analysis will be conducted
to determine the degree of effectiveness of a specific variable using the ANN-based
model. Moreover, performance of various combinations of input variables will also
be tested by the optimal ANN structure, using the best-fit training algorithm in a
future study.
It can be pointed out that a modern activated sludge plant design involves much
more than aerobic activated sludge tank for organic matter removal, which is the
process that is examined (and designed) in this paper. In modern activated sludge
tanks, aerobic zones (aerated) need to be combined with anoxic and anaerobic
tanks to achieve biological removal of nitrogen and phosphorus. While this study
highlights the feasibility of modeling in CMASRV, it could be useful to investigate
the effects of other different parameters on the modern process design of further
works. Therefore, investigation of new model configurations could yield further
insights into the process design, and the mode of plant operation.
To conclude, the stochastic approaches given in this study can provide realistic
results for volume of completely mixed activated sludge reactors. The results of
this study are expected to provide a useful source of information for other relevant
theoretical and experimental future studies and prove to be helpful for wastewater
engineers, designers, academicians and technical audiences who are concerned with
activated sludge processes and modeling in environmental engineering.
587
References
[1] Gernaey K. V., Van Lossdrecht M. C. M., Henze M., Lind M., Jorgensen S. B.: Activated
sludge wastewater treatment plant modelling and simulation: State of the art. Environ.
Model. Soft., 19, 2004, pp. 763-783.
[2] Henze M., Harremoës P., Jansen J., Arvin E.: Wastewater treatment - Biological and chem-
ical processes. Springer-Verlag, Berlin, Heidelberg, New York, 1995.
[3] U. S. Army Corps of Engineers, Activated Sludge Plants, Chapter 13, TM 5-814-3/AFM
88-11, Volume III. Available online: www.hnd.usace.army.mil/techinfo/ UFC/UFC3-240-
09FA/TM58143/chap13.pdf
[4] Jungblut J., Möller D. P. F.: Modelling and simulation of new biological wastewater treat-
ment plants. UKSim99: Final Programme, Session-I: Simulation in Control & Dynamical
Systems, St Catharine’s College, Cambridge, England, 1999.
[5] Kohonen T.: State of the Art in Neural Computing. In: IEEE First International Conference
On Neural Networks, 1, San Diego, 1993, pp. 77-91.
[6] Goktepe A. B., Agar E., Lav A. H.: Backcalculation of mechanical properties of flexible
pavements using neural networks. ITU Dergisi, 4, 2005, pp. 31-42.
[7] Stanley, J.: Introduction of neural networks. California Scientific Software, 1988, p. 255.
[8] Rogers L. L., Dowla F. U.: Optimization of groundwater remediation using artificial neural
networks with parallel solute transport modelling. Wat. Resour. Res. 30, 1994, pp. 457-481.
[9] Hsu K., Gupta H. V., Sorooshian S.: Artificial neural network modelling of the rainfall runoff
process. Wat. Resour. Res., 31, 1995, pp. 2517-2530.
[10] Scardi M.: Artificial neural networks as empirical models for estimating phytoplankton pro-
duction. Mar. Ecol. Ser., 139, 1996, pp. 289-299.
[11] Yabunaka K., Hosomi M., Murakami A.: Novel application of a backpropagation artificial
neural network model formulated to predict algal bloom. Water Sci. Tech., 36 1997, pp.
89-97.
[12] Djebbar Y., Katoda P. T.: Estimation of Sanitary Flows using neural networks. Wat. Sci.
Technol., 38, 1998, pp. 215-222.
[13] Hall M. J., Minns A. W.: Regional flood frequency analysis using artificial neural networks.
In: Proc. 3rd Int. Conference on Hydroinformatics (Copenhagen, Denmark), V. Babovic and
L. C. Larsen (Eds.), Balkema, Rotterdam, 2, 1998, pp. 759-763.
[14] Rodriguez M. J., Sérodes J. B.: Assessing empirical linear and non-linear modelling of resid-
ual chlorine in urban drinking water systems. Environ. Model. Soft., 14, 1998, pp. 93-102.
[15] Campolo M., Soldati A., Andreussi P.: Forecasting river flow rate during low flow periods
using neural networks. Wat. Resour. Res., 35, 1999, pp. 3547-3552.
[16] Karul C., Soyupak S., Cilesiz A. F., Akbay N., Germen E.: Case studies on the use of neural
networks in eutrophication modeling. Ecolog. Model., 134, 2000, pp. 145-152.
[17] Sengorur B., Oz C.: Determination of the effects of water pollution of aquacultures using
neural networks. Turkish J. Eng. Environ. Sci. 26, 2002, pp. 95-105.
[18] Anctil F., Perin C., Andreassian V.: Impact of the length of observed records on the perfor-
mance of ANN and of conceptual parsimonious rainfall-runoff forecasting models. Environ.
Model. Soft., 19, 2004, pp. 357-368.
[19] Hamed M. M., Khalafallah M. G., Hassanien E. A.: Prediction of wastewater treatment plant
performance using artificial neural networks. Environ. Model. Soft., 19, 2004, pp. 919-928.
[20] Nunnari G., Dorling S., Schlink U., Cawley G., Foxall R., Chatterton T.: Modelling SO2
concentration at a point with statistical approaches. Environ. Model. Soft., 19, 2004, pp.
887-905.
[21] Onkal-Engin G., Demir I., Engin S. N.: Determination of the relationship between sewage
odour and BOD by neural networks. Environ. Model. Soft., 20, 2005, pp. 843-850.
588
View publication stats
[22] Karaca F., Ozkaya B.: NN-LEAP: A neural network-based model for controlling leachate
flow-rate in a municipal solid waste landfill site. Environ. Model. Soft., 21, 2006, pp. 1190-
1197.
[23] Yetilmezsoy K.: Determination of optimum body diameter of air cyclones using a new em-
pirical model and a neural network approach. Environ. Eng. Sci., 23, 2006, pp. 680-690.
[24] Yetilmezsoy K., Saral A.: Stochastic modeling approaches based on neural network and
linear–nonlinear regression techniques for the determination of single droplet collection effi-
ciency of countercurrent spray towers. Environ Model. Assess. 12, 2007, pp. 13-26.
[25] Moral H.: Modeling of activated sludge process by using artificial neural networks. MSc in
Environmental Engineering, Graduate School of Natural and Applied Sciences, Middle East
Technical University, 2004.
[26] Al-Mutairi N., Kartam N., Koushki P., Al-Mutairi M.: Modeling and predicting biological
performance of contact stabilization process using artificial neural networks. J. Comput. in
Civ. Eng., 18, 2004, pp. 341-349.
[27] Metcalf and Eddy, Wastewater Engineering Treatment and Reuse, Fourth Editions, McGraw
Hill Higher Education, New York, 2003.
[28] UCLA College of Letters and Science: Microorganisms and their role in the activated-sludge
process, Student project, USA.
http://www.college.ucla.edu/webproject/micro7/studentprojects7/Rader/asludge2.htm
[29] Muslu Y.: Wastewater Treatment, Vol. 1, ITU Publishing Corp., Istanbul, 1996 (in Turkish).
[30] Fouad M., Bhargava R.: Mathematical model for the biofilm–activated sludge reactor. J.
Envir. Engrg., 131, 2005, pp. 557-562.
[31] Fouad M., Bhargava R.: Sludge production and settleability in biofilm-activated sludge
process, J. Envir. Engrg., 131, 2005, pp. 417-424.
[32] Massong E., Wang J.: Introduction to computation and learning in artificial neural networks.
Europ. J. Oper. Res., 47, 1990, pp. 1-28.
[33] Sarle W. S.: What are cross-validation and bootstrapping, FAQ Archieves, AI FAQ/neural
Nets, Part 3, Section-12, February 2010.
[34] Goutte C.: Note on free lunches and cross-validation. Neural Comp., 9, 1997, pp. 1211-1215.
[35] Yetilmezsoy K., Demirel S.: Artificial neural network (ANN) approach for modeling of Pb(II)
adsorption from aqueous solution by Antep pistachio (Pistacia Vera L.). J. Hazard. Mater.,
153, 2008, pp. 1288-1300.
[36] Yetilmezsoy K., Sapci-Zengin Z.: Stochastic modeling applications for the prediction of
COD removal efficiency of UASB reactors treating diluted real cotton textile wastewater.
Stoc. Environ. Res. Risk Assess., 23, 2009, pp. 13-26.
[37] Akkoyunlu A., Yetilmezsoy K., Erturk F., Oztemel E.: A neural network-based approach
for the prediction of urban SO2 concentrations in the Istanbul Metropolitan Area, Int. J.
Environ. Pol., 40, 2010, pp. 301-321.
[38] Kanat G., Saral A.: Estimation of biogas production rate in a thermophilic UASB reactor
using artificial neural networks, Environ. Model. Assess., 14, 2009, pp. 607-614.
[39] Almasri M. N., Kaluarachchi J. J.: Modular neural networks to predict the nitrate distri-
bution in ground water using the onground nitrogen loading and recharge data. Environ.
Model. Soft., 20, 2005, pp. 851-871.
[40] Roth M. W.: Neural network technology and its applications. John Hopkins Appl. Tech.
Dig., 1988, pp. 242-253.
[41] Stevenson W. J.: The use of artificial neural nets in mechanical engineering. Council for
Scientific and Industrial Research Tech. Rep. AERO 91/198, 1991.
589

Modeling Studies For The Determination of Completely Mixed Activated Sludge Reactor Volume: Steady-State, Empirical and Ann Applications

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Modeling Studies For The Determination of Completely Mixed Activated Sludge Reactor Volume: Steady-State, Empirical and Ann Applications

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Modeling studies for the determination of completely mixed activated sludge

Article in Neural Network World · January 2010

Butterfly optimization algorithm View project

Oman Cement Company View project

The user has requested enhancement of the downloaded file.

Key words: Activated sludge, completely mixed reactor, steady-state model,

Received: September 26, 2009

Department of Environmental Engineering, Yildiz Technical University, Faculty of Civil En-

2. Definition of the Models

Fig. 1 The flow diagram of a completely mixed activated sludge process.

M ass of cells in the reactor XV XV XV

The BOD of a mole of cells is assumed as follows:

BOD = 1.42(cells) (12)

2.2 The proposed model

Projection Parameters and Kinetic Unit Operating

(θC ), microorganism decay coefficient (kd ), biomass concentration (X), influent

Main Biological Variables Value

model is obtained as follows:

2.3 Artificial Neural Network (ANN) approach

completely mixed activated sludge reactor volume (CMASRV), such as wastewater

Variables Units Min. Max. Range Average SDa Median

the information is transmitted in a forward direction only. The weighted sum of

3. Results and Discussion

Y Q(Si − Se )θC (0.50)(21600)(0.25 − 0.006)(10) ∼

Tab. VI Comparison of eight training algorithms.

precision). Basically, the learning in the network is achieved through an iterative

Fig. 4 Optimal neural network structure for prediction of CMASRV values.

Fig. 7 Head-to-head performance comparisons of the steady-state design set, the

Fig. 8 Head-to-head performance comparisons of the steady-state design set, the

A number of attempts at developing a representative prediction model may obvi-

Yetilmezsoy K.: Modeling studies for the determination of completely. . .

You might also like