Expert Systems with Applications 36 (2009) 8368–8374

Contents lists available at ScienceDirect

Expert Systems with Applications

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

Application of least square support vector machines in the prediction

of aeration performance of plunging overfall jets from weirs
Ahmet Baylar a,*, Davut Hanbay b, Murat Batan a
a
Firat University, Civil Engineering Department, Elazig 23119, Turkey
b
Firat University, Electronic and Computer Science Department, Elazig 23119, Turkey

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

Keywords: Aeration is a mass transfer process between the atmosphere and water. Aeration is used for water quality
SVM enhancement in sewage treatment plants and in polluted rivers and lakes. This can be enhanced by cre-
Weir ating turbulence in the water. Plunging overfall jets from weirs are a particular instance of producing
Aeration efficiency such turbulence. In this paper, two intelligent models are realized to predict the air entrainment rate
Air entrainment rate
and aeration efficiency of weirs. Least square support vector machine (LS-SVM) is used as intelligent tool.
Threefold cross validation test method is used to evaluate the performance of LS-SVM models. The cor-
relation between predicted and measured values is found 0.99 for air entrainment rate and 0.98 for aer-
ation efficiency. The test results indicate that the LS-SVM can be used successfully in predicting the air
entrainment rate and aeration efficiency of weirs. Moreover, the performances of the LS-SVM models
are compared with multi nonlinear and linear regression models.
Ó 2008 Elsevier Ltd. All rights reserved.

1. Introduction These studies were reviewed by Chanson (1995), Ervine (1998),

Dissolved oxygen (DO) is the amount of oxygen that is dissolved
in water and is the most important water quality parameter in riv-
ers and streams. The dissolved oxygen level is an indication of how
polluted the water is and how well the water can support aquatic
plant and animal life. A higher dissolved oxygen level indicates
better water quality. Many naturally occurring biological and
chemical processes use oxygen, thereby diminishing the dissolved
oxygen concentration in the water. The physical process of oxygen
transfer from the atmosphere acts to replenish the used oxygen.
This process has been termed aeration.
Aeration is used for water quality enhancement in sewage
treatment plants and in polluted rivers and lakes. Hydraulic struc-
tures increase the amount of dissolved oxygen in a river system,
even though the water is in contact with the structure for only a
short time. The same quantity of oxygen transfer that normally
would occur over several kilometers in a river can occur at a single
hydraulic structure. The primary reason for this accelerated oxygen
transfer is that air is entrained into the flow in the form of a large
number of bubbles. These air bubbles greatly increase the surface
area available for mass transfer.
Air entrainment and aeration efficiency of hydraulic structures
have been studied experimentally by a number of investigators.
* Corresponding author. Tel.: +90 424 2370000x5428; fax: +90 424 2415526.
E-mail address: abaylar@firat.edu.tr (A. Baylar).
Gulliver, Wilhelms, and Parkhill (1998) and Wilhelms, Gulliver,
and Parkhill (1992). Recently, Baylar (2002, 2003), Baylar and Bag-
atur (2000, 2001a, 2001b, 2006), Baylar and Emiroglu (2002, 2007),
Baylar, Bagatur, and Tuna (2001a, 2001b), Baylar and Ozkan (2006)
and Emiroglu and Baylar (2003a, 2003b, 2005, 2006) investigated
weirs having different cross-sectional geometry and demonstrated
that the air entrainment rate and aeration efficiency of weirs chan-
ged depending on weir shapes.
In recent years, the developments in intelligent methods make
them possible to use in complex systems modeling (Baylar, Han-
bay, & Ozpolat, 2007, 2008 and Hanbay, Turkoglu, & Demir,
2007). Support vector machines (SVMs) introduced by Vapnik
and his co-workers are a type of intelligent methods. SVMs are a
powerful methodology and have been studied extensively for clas-
sification, regression and pattern recognition (Cortes & Vapnik,
1995). The basic idea of SVMs is mapping the input data points
to a high-dimensional feature space and finds a hyperplane. The
least square (LS) version of the SVM was described in Suykens
and Vandewalle (1999). LS-SVM is widely used in complex system
studies for modeling, regression or parameter prediction. However,
its application to the hydraulic systems is very limited. In this pa-
per, LS-SVM models are realized to predict the air entrainment rate
and aeration efficiency of weirs. The organization of the paper is as
follow; at first theoretical background of the system and LS-SVM
are overviewed. Then realized study is described in details. At last
obtained results are concluded.

0957-4174/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2008.10.061
 A. Baylar et al. / Expert Systems with Applications 36 (2009) 8368–8374
2. Weir air entrainment
Tsang (1987) described four basic air entrainment mechanisms
with reference to a free falling jet from a weir. Tsang’s four air
entrainment mechanisms are smooth, rough, oscillating, and disin-
tegrated. For small drop heights, a water jet with a relatively
smooth surface issues from the weir. The major source of air sup-
ply is visualized as a thin layer surrounding the jet and carried into
the water upon impact, and therefore the air entrainment capacity
is limited. The water surface in the receiving pool is relatively
undisturbed. As the drop height increases, the surface of the jet be-
comes roughened. The air supply can be considered as coming lar-
gely from small air pockets entrapped between the jet surface
roughness and the receiving water. At impact, the jet produces rip-
ples on the pool surface. Compared with the smooth jet under sim-
ilar conditions, this mechanism results in shallower bubble
penetration but increased entrainment rate, because the bubbles
are more densely packed in the biphasic zone. As the drop height
is increased further, the jet begins to oscillate during the fall. The
primary air source originates from large air pockets entrapped be-
tween the undulating jet and the pool surface. The pool surface is
considerably agitated, and air may also be entrained by surface
roller action and splashing. Large air pockets are transported from
the surface into the water depth and broken down due to turbu-
lence. With an even larger drop height, the jet breaks up into dis-
crete droplets. The pool surface is intensely agitated, and air is
entrained by the action of surface rollers and by the engulfing of
air pockets as jet fragments hit the pool surface. The bubbles are
generally only transported to relatively shallow depths. Disinte-
grated jets have the advantage over solid jets of greater surface
area; however, air entrainment rate QA and bubble penetration
are significantly reduced because of energy loss to the surrounding
atmosphere during the fall (Wormleaton & Tsang, 2000).
3. Air–Water mass transfer
The mass balance equation relating the instantaneous rate of
change in DO concentration (dC/dt) to the rate of oxygen mass
transfer between air and water is
dC A maximum water flow rate of 4 L/s. The water was pumped from
¼ K L ðC s  CÞ; ð1Þ
dt V a pump to the stilling tank. The water jet from the test weir
where C is the DO concentration, KL the liquid film coefficient for
oxygen, A the surface area associated with the volume V, over which
transfer occurs, Cs the saturation concentration and t is the time. Eq.
(1) does not consider sources and sinks of oxygen in the water body
because their rates are relatively slow compared to the oxygen
transfer that occurs at most hydraulic structures due to the increase
in free-surface turbulence and the large quantity of air that is nor-
mally entrained into the flow.
The predictive relations assume that Cs is constant and deter-
mined by the water–atmosphere partitioning. If that assumption
Deoxgenation Grid
Mixer chemical
input
Flow Direction Upstream
channel
Water
flow
meter
Stilling tank
Storage tank
P P
Water feed line
Water Flow
pump control
valves
Fig. 1. Experimental equipment.
8369
is made, Cs is constant with respect to time, and the aeration effi-
ciency (oxygen transfer efficiency), E may be defined as (Gulliver,
Thene, & Rindels, 1990):
Cd  Cu
E¼ ; ð2Þ
Cs  Cu
where u and d are subscripts indicating upstream and downstream
locations respectively, Cu the dissolved oxygen concentration at up-
stream of hydraulic structure and Cd is the dissolved oxygen con-
centration at downstream of hydraulic structure.
A value of E > 1 means the downstream water has become
supersaturated (i.e., Cd > Cs). A transfer efficiency value of 1.0
means that the full transfer up to the saturation value has occurred
at the structure. No transfer would correspond to E = 0.0. The sat-
uration concentration in distilled, deionized water may be ob-
tained from charts or equations. This is an approximation
because the saturation DO concentration for natural waters is often
different from that of distilled, deionized water due to the impact
of trace contaminants and salinity.
Oxygen transfer efficiency is sensitive to water temperature,
and investigators have typically employed a temperature correc-
tion factor. For hydraulic structures, Gulliver et al. (1990) devel-
oped a relationship for temperature correction factor:
E20 ¼ 1  ð1  EÞ1=f ; ð3Þ
where E is the aeration efficiency at the water temperature of mea-
surement, E20 the aeration efficiency at the 20 °C and f is the expo-
nent described by
f ¼ 1:0 þ 0:02103ðT  20Þ þ 8:261  105 ðT  20Þ2 ; ð4Þ
where T is the water temperature.
4. Experiments
4.1. Experimental methodology
The data used in this study were taken from study conducted by
Baylar and Bagatur (2006) on a large model. Schematic representa-
tion of the experimental setup is shown on Fig. 1. The experimental
channel was 3.40 m long, 0.60 m wide, and 0.50 m deep with a
plunged into a downstream water pool, whose height could be ad-
justed using a pulley arrangement. The water depth in the down-
stream water pool was controlled by an adjustable weir. The
plan-view dimensions of the downstream water pool were
1.20  1.20 m. An air flow meter installed on the air-hood was
used to measure air entrainment rate QA. The plan-view dimen-
sions of the air-hood were 0.75  0.60 m. The test weir featured
three exchangeable weir elements: 45° triangular sharp-crested
weir, 90° triangular sharp-crested weir and 135° triangular
sharp-crested weir.
Weir
DO
Free overfall jet
Air flow
QA meter
Drop
DO height
DO measurement
Q. A
Tailwater
Receiving depth
pool Release
Bubble valve
trap Adjustable
weir
Water
pump
8370 A. Baylar et al. / Expert Systems with Applications 36 (2009) 8368–8374

60 60
o
θ = 45
o θ = 135

Q = 1 L/s Q = 1 L/s
50 50
Q = 2 L/s Q = 2 L/s

Q = 3 L/s Q = 3 L/s
40 40
(m 3 /s )

Q = 4 L/s Q = 4 L/s

Q Ax 10 - 4 (m 3 /s )
-4

30 30
Q Ax 10

20 20

10 10

0 0
0.00 0.15 0.30 0.45 0.60 0.75 0.90 1.05 0.00 0. 15 0.30 0.45 0.60 0.75 0.90 1.05
h (m) h(m)
Fig. 2. Variation in air entrainment rate with drop height for h = 45°. Fig. 4. Variation in air entrainment rate with drop height for h = 135°.

Free overfall jet from the triangular sharp-crested weir plunged

through the atmosphere, impinged on the downstream water pool
below and entrained air bubbles which were in turn trapped in the
air-hood. Then, QA was measured by means of the air flow meter
installed on the air-hood. During the experiments, DO measure-
ments upstream and downstream of the weirs were taken using
calibrated portable HANNA Model HI 9142 oxygen meters at the
locations identified in Fig. 1. In all of the experiments for all weirs,
flow paths of entrained air bubbles in the receiving pool were
uninterrupted, i.e., the receiving pool depth (tailwater depth) was
greater than the bubble penetration depth. Each triangular sharp-
crested weir was tested under flow rates Q varying from 1.0 to
4.0 L/s in 1.0 L/s steps. The drop height h, defined as the difference
between the water levels upstream and downstream of the trian-
gular sharp-crested weir, was varied between 0.15 and 0.90 m in
0.15 m steps.
4.2. Experimental results
Baylar and Bagatur (2006) experimental results for the air
entrainment rate QA and aeration efficiency E20 in weirs are given
in Figs. 2–7. Regression analyses were performed by Baylar and
Bagatur (2006) using the nonlinear regression (NLR) module.
Empirical correlations predicting the air entrainment rate QA and
aeration efficiency E20 for the triangular sharp-crested weirs are gi-
ven in

60 0.6
o o
θ = 90 θ = 45

Q = 1 L/s Q = 1 L/s
50 0.5
Q = 2 L/s Q = 2 L/s

Q = 3 L/s Q = 3 L/s
40 0.4
Q = 4 L/s Q = 4 L/s
Q Ax 10 - 4 (m 3 /s )

E 20

30 0.3

20 0.2

10 0.1

0 0.0
0.00 0.15 0.30 0.45 0.60 0.75 0.90 1.05 0.00 0.15 0.30 0.45 0.60 0.75 0.90 1.05
h (m) h (m)

Fig. 3. Variation in air entrainment rate with drop height for h = 90°. Fig. 5. Variation in aeration efficiency with drop height for h = 45°.
 A. Baylar et al. / Expert Systems with Applications 36 (2009) 8368–8374 8371

0.6 1995. It is applicable to both classification and regression. The

o SVMs are based on the principle of Structural Risk Minimization
θ = 90
(Cortes & Vapnik, 1995).
0.5
Q = 1 L/s Consider a given training set fxk ; yk gNk¼1 with input data xk 2 Rn
Q = 2 L/s and output data yk 2 R with class labels yk 2 f1; þ1g and linear
classifier
Q = 3 L/s
0.4
Q = 4 L/s
yðxÞ ¼ sign½wT x þ b: ð7Þ
When the data of the two classes are separable one can say
 
20

wT xk þ b P þ1; if yk ¼ þ1
E

0.3
: ð8Þ
wT xk þ b 6 1; if yk ¼ 1
These two sets of inequalities can be combined into one single
0.2
set as follows:
yk ½wT xk þ b P 1 k ¼ 1; . . . ; N: ð9Þ
0.1 SVM formulations are done with in a context of convex optimi-
zation theory. The general methodology is to start formulating the
problem as a constrained optimization problem, next formulate
0.0 the Lagrangian and then take the conditions for optimality, finally
0.00 0.15 0.30 0.45 0.60 0.75 0.90 1.05 solve the problem in the dual space of Lagrange multipliers. With
resulting classifier
h (m)
" #
XN
Fig. 6. Variation in aeration efficiency with drop height for h = 90°. yðxÞ ¼ sign ak yk xtk x þ b : ð10Þ
k¼1

"  0:172 #4:598 This linear SVM classifier was extended to non-separable case
0:293 h
Q A ¼ 0:171h Q 0:158 cos ; ð5Þ by Cortes and Vapnik (1995). It is done by taking additional slack
2 variable in the problem formulation. One modifies the set of
"  0:206 #1 inequalities into
1:341 0:280 h
E20 ¼ 1  1 þ 0:149h Q sin ; ð6Þ  
2 yk wT xk þ b P 1  nk ; k ¼ 1; . . . ; N: ð11Þ

where h is the drop height (m), Q the discharge (m3/s), and h is the
angle in triangular sharp-crested weir (deg.).
5. Least square support vector machine (LS-SVM)
The SVM is a supervised learning technique which is developed
by Vladimir Vapnik and co-workers at AT&T Bell Laboratories in
The SVM has been used for linear and nonlinear function esti-
mation too. The details of them can be found in literature. The least
square version of the SVM classifier was described by Suykens and
Vandewalle (1999). The LS-SVM considers equality type con-
straints instead of inequalities as in the classic SVM approach. This
reformulation greatly simplifies a problem such that the LS-SVM
solution follows directly from solving a set of linear equations
rather than from a convex quadratic program. For a LS-SVM classi-
fier, in the primal space it takes the form,

0.6 yðxÞ ¼ sign½wT x þ b; ð12Þ

o
θ = 135 where b is a real constant. For nonlinear classification, the LS-SVM
Q = 1 L/s
classifier in the dual space takes the form
0.5 " #
Q = 2 L/s XN
yðxÞ ¼ sign ak yk Kðx; xk Þ þ b ; ð13Þ
Q = 3 L/s k¼1
0.4
Q = 4 L/s where ai is the positive real constants and b a real constant, in gen-
eral, Kðxi ; xÞ ¼ h/ðxi Þ; /ðxÞi, h; i the inner product, and /ðxÞ the
nonlinear map from original space to high-dimensional space. For
0.3
E 20

function estimation, the LS-SVM model takes the form

X
N
yðxÞ ¼ ak Kðx:xk Þ þ b: ð14Þ
0.2 k¼1

When radial basic function (RBF) kernels are used, two tuning
parameters (c, r) are added. Where c is the regularization constant
0.1
and r the width of RBF kernel.

0.0 6. LS-SVM modeling

0.00 0.15 0.30 0.45 0.60 0.75 0.90 1.05
All program codes were written in MATLAB by using LS-SVM
h (m)
toolbox (http://www.esat.kuleuven.ac.be/sista/lssvmlab). The
Fig. 7. Variation in aeration efficiency with drop height for h = 135°. parameters considered in the study are discharge Q, drop height
8372 A. Baylar et al. / Expert Systems with Applications 36 (2009) 8368–8374

60 0.6
Measured
Predicted
50 0.5

40 0.4
-4
Q A x 10

E 20
30 0.3

20 0.2

10 0.1
Measured
Predicted
0 0.0
0 5 10 15 20 25 0 5 10 15 20 25
Sample Sample

Fig. 8. QA performance of LS-SVM model. Fig. 9. E20 performance of LS-SVM model.

h, angle in triangular sharp-crested weir h, air entrainment rate QA
and aeration efficiency E20. The parameters, Q, h, and h were used
as inputs to the LS-SVM models to predict QA and E20, respectively.
72 experimental data sets were used for the SVM model. Before
training and testing of LS-SVM model for QA and E20, the data set
was normalized by using
xi  xmin
xi ¼ þ b;
xmax  xmin
where xmin and xmax are the minimum and maximum of the input
data and b is the bias value (b = 0.1).
In the first part of this study, the LS-SVM model with RBF kernel
is trained for air entrainment rates QA. Grid search algorithm is
used for determining the optimal values of RBF kernel parameters.
For QA, the determined optimal c value is 511.35 and optimal r va-
lue is 1.9914. The test performance of LS-SVM model is graphically
presented in Fig. 8. As seen in Fig. 8, LS-SVM model predicts the
measured values at high accuracy rate. Threefold cross validation
method is applied to the 72 experimental data sets for computing
the validation of LS-SVM models. The correlation between mea-
ð15Þ sured QA and predicted by LS-SVM model is 0.99.
Moreover, the measured air entrainment rates QA are compared
with the LS-SVM model, the nonlinear regression (NLR) technique
and the linear regression (LR) technique results. It can be obviously
seen from Table 1 that the LS-SVM model approximates the mea-
sured QA values with a quite high accuracy. The LS-SVM model

Table 1
Absolute relative error (ARE) statistics for computed QA using LS-SVM, NLR and LR models.

Inputs QA (m3/s  104) ARE (%)

3
Q (m /s) H (m) h (deg.) Measured LS-SVM NLR LR LS-SVM NLR LR
0.001 0.45 45 6.66 6.23 7.12 9.72 6.5 7.0 45.9
0.001 0.90 45 15.57 16.58 18.13 23.76 6.5 16.4 52.6
0.002 0.45 45 12.30 11.92 11.79 13.81 3.1 4.2 12.3
0.002 0.90 45 32.44 30.53 29.99 27.85 5.9 7.6 14.2
0.003 0.45 45 12.07 14.62 15.83 17.90 21.1 31.1 48.3
0.003 0.90 45 39.83 39.98 40.26 31.94 0.4 1.1 19.8
0.004 0.45 45 16.02 18.68 19.50 21.99 16.6 21.7 37.2
0.004 0.90 45 51.50 47.67 49.62 36.03 7.4 3.6 30.0
0.001 0.45 90 5.98 7.21 5.77 5.27 20.5 3.6 11.9
0.001 0.90 90 12.47 14.30 14.67 19.31 14.7 17.6 54.8
0.002 0.45 90 10.72 11.39 9.54 9.36 6.2 11.0 12.7
0.002 0.90 90 26.34 25.72 24.27 23.40 2.4 7.8 11.2
0.003 0.45 90 12.64 12.70 12.64 13.44 0.5 0.0 6.4
0.003 0.90 90 31.42 33.33 31.42 27.48 6.1 0.0 12.5
0.004 0.45 90 16.92 14.88 15.79 17.53 12.0 6.7 3.6
0.004 0.90 90 42.03 39.43 40.16 31.57 6.2 4.4 24.9
0.001 0.45 135 3.50 4.75 3.55 0.82 35.7 1.4 76.7
0.001 0.90 135 10.04 7.85 9.03 14.86 21.8 10.1 48.0
0.002 0.45 135 7.11 6.76 5.87 4.90 4.9 17.4 31.0
0.002 0.90 135 11.39 14.37 14.94 18.94 26.2 31.1 66.3
0.003 0.45 135 7.33 7.32 7.88 8.99 0.2 7.5 22.7
0.003 0.90 135 19.80 19.17 20.05 23.03 3.2 1.3 16.3
0.004 0.45 135 7.84 8.61 9.71 13.08 9.8 23.9 66.9
0.004 0.90 135 26.63 23.53 24.71 27.12 11.7 7.2 1.8
 A. Baylar et al. / Expert Systems with Applications 36 (2009) 8368–8374 8373

Table 2
Absolute relative error (ARE) statistics for computed E20 using LS-SVM, NLR and LR models.

Inputs E20 ARE (%)

3
Q (m /s) H (m) h (deg.) Measured LS-SVM NLR LR LS-SVM NLR LR
0.001 0.45 45 0.29 0.27 0.30 0.29 7.0 3.9 1.2
0.001 0.90 45 0.50 0.52 0.52 0.51 3.7 4.4 1.9
0.002 0.45 45 0.26 0.25 0.26 0.26 3.5 0.8 0.9
0.002 0.90 45 0.46 0.49 0.47 0.49 7.3 2.9 5.5
0.003 0.45 45 0.22 0.24 0.24 0.24 8.7 9.4 8.2
0.003 0.90 45 0.42 0.46 0.45 0.46 10.5 6.0 9.8
0.004 0.45 45 0.20 0.24 0.23 0.21 19.2 13.1 6.8
0.004 0.90 45 0.45 0.44 0.43 0.44 2.9 5.5 3.0
0.001 0.45 90 0.30 0.26 0.28 0.27 14.8 8.3 11.1
0.001 0.90 90 0.48 0.51 0.49 0.49 7.2 2.2 2.0
0.002 0.45 90 0.27 0.23 0.24 0.24 15.5 11.8 10.2
0.002 0.90 90 0.50 0.49 0.44 0.47 2.7 11.6 6.9
0.003 0.45 90 0.23 0.20 0.22 0.22 11.2 5.1 5.1
0.003 0.90 90 0.44 0.45 0.41 0.44 2.1 5.8 0.3
0.004 0.45 90 0.21 0.19 0.20 0.19 9.3 2.5 7.7
0.004 0.90 90 0.42 0.41 0.39 0.42 2.0 6.0 0.8
0.001 0.45 135 0.30 0.26 0.26 0.25 14.9 11.9 17.7
0.001 0.90 135 0.46 0.49 0.48 0.47 5.8 3.6 2.2
0.002 0.45 135 0.25 0.22 0.23 0.22 12.5 8.6 10.9
0.002 0.90 135 0.42 0.46 0.43 0.45 8.5 2.0 6.1
0.003 0.45 135 0.16 0.18 0.21 0.20 13.8 30.6 24.0
0.003 0.90 135 0.44 0.41 0.40 0.42 6.2 8.9 4.2
0.004 0.45 135 0.16 0.15 0.20 0.17 3.8 22.5 8.8
0.004 0.90 135 0.36 0.37 0.38 0.40 2.2 6.1 10.3

seems to have similar performance to multi nonlinear regression
model. The accuracy of the linear regression model is lower than
the LS-SVM and multi nonlinear regression models.
In the second part of this study, the LS-SVM model with RBF
kernel is trained for aeration efficiencies E20. Grid search algorithm
is used for determining the optimal values of RBF kernel parame-
ters. For E20, the determined optimal c value is 76.517 and optimal
r value is 3.0647. The test performance of LS-SVM model for E20 is
graphically presented in Fig. 9. As it can be seen from Fig. 9, LS-
SVM model predicts the measured values at high accuracy. Three-
fold cross validation method is applied to the 72 experimental data
sets for computing the validation of LS-SVM models. The correla-
tion between measured E20 and predicted by LS-SVM model is 0.98.
The measured aeration efficiencies E20 are compared with the
LS-SVM model, the nonlinear regression (NLR) technique and the
linear regression (LR) technique results. It can be obviously seen
from Table 2 that the LS-SVM model approximates the measured
E20 values with a quite high accuracy. The LS-SVM model seems
to have similar performance to multi nonlinear and linear regres-
sion models.
7. Conclusions
The dominant aeration mechanism in free overfall jet from a weir
is that of air bubbles being carried into the biphasic zone when the jet
plunges into the receiving pool. The degree of aeration will depend
upon the amount of air entrained. In this study, LS-SVM models were
developed to determine the air entrainment rate and aeration effi-
ciency of weirs. It was found from the test results that LS-SVM mod-
els can be used successfully in predicting the air entrainment rate
and aeration efficiency of weirs. The most important aspect of the
intelligent modeling is the ability of self-organization without
requirements of programming and the immediate response of a
trained net during real-time applications. These features make the
intelligent models suitable for complex systems modeling.
References
Baylar, A. (2002). Study on the effect of type selection of weir aerators on oxygen
transfer, PhD thesis. Firat University, Elazig, Turkey (in Turkish).
Baylar, A. (2003). An investigation on the use of venturi weirs as an aerator. Water
Quality Research Journal of Canada, 38(4), 753–767.
Baylar, A., & Bagatur, T. (2000). Aeration performance of weirs. Water SA, 26(4),
521–526.
Baylar, A., & Bagatur, T. (2001a). Aeration performance of weirs. Water Engineering &
Management Part 1, 148(3), 33–36.
Baylar, A., & Bagatur, T. (2001b). Aeration performance of weirs. Water Engineering
& Management, Part 2, 148(4), 14–16.
Baylar, A., & Bagatur, T. (2006). Experimental studies on air entrainment and oxygen
content downstream of sharp-crested weirs. Water and Environment Journal,
20(4), 210–216.
Baylar, A., & Emiroglu, M. E. (2002). The effect of sharp-crested weir shape on air
entrainment. Canadian Journal of Civil Engineering, 29(3), 375–383.
Baylar, A., & Ozkan, F. (2006). Applications of venturi principle to water aeration
systems. Environmental Fluid Mechanics, 6(4), 341–357.
Baylar, A., Bagatur, T., & Tuna, A. (2001a). Aeration performance of triangular notch
weirs at recirculating system. Water Quality Research Journal of Canada, 36(1),
121–132.
Baylar, A., Bagatur, T., & Tuna, A. (2001b). Aeration performance of triangular-notch
weirs. Journal of the Chartered Institution of Water and Environmental
Management, 15(3), 203–206.
Baylar, A., Hanbay, D., & Ozpolat, E. (2007). Modeling aeration efficiency of stepped
cascades by using ANFIS. CLEAN – Soil, Air, Water, 35(2), 186–192.
Baylar, A., Hanbay, D., & Ozpolat, E. (2008). An expert system for predicting aeration
performance of weirs by Using ANFIS. Expert Systems with Applications, 35(3),
1214–1222.
Baylar, A., & Emiroglu, M. E. (2007). The role of weir types in entrainment of air
bubbles. International Journal of Science and Technology, 2(2), 143–154.
Chanson, H. (1995). Predicting oxygen content downstream of weirs, spillways and
waterways. Proceedings of the Institution of Civil Engineers – Water, Maritime and
Energy, 112(1), 20–30.
Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 20(3),
273–297.
Emiroglu, M. E., & Baylar, A. (2003a). Experimental study of the influence of
different weir types on the rate of air entrainment. Water Quality Research
Journal of Canada, 38(4), 769–783.
Emiroglu, M. E., & Baylar, A. (2003b). The effect of broad-crested weir shape on air
entrainment. Journal of Hydraulic Research, 41(6), 649–655.
Emiroglu, M. E., & Baylar, A. (2005). The influence of included angle and sill slope on
air entrainment of triangular planform labyrinth weirs. Journal of Hydraulic
Engineering, ASCE, 131(3), 184–189.
Emiroglu, M. E., & Baylar, A. (2006). Closure of ‘the influence of included angle and
sill slope on air entrainment of triangular planform labyrinth weirs’. Journal of
Hydraulic Engineering, ASCE, 132(7), 748.
Ervine, D. A. (1998). Air entrainment in hydraulic structures: A review. Proceedings
of the Institution of Civil Engineers – Water, Maritime and Energy, 130(3),
142–153.
Gulliver, J. S., Thene, J. R., & Rindels, A. J. (1990). Indexing gas transfer in self-aerated
flows. Journal of Environmental Engineering, ASCE, 116(3), 503–523.
Gulliver, J. S., Wilhelms, S. C., & Parkhill, K. L. (1998). Predictive capabilities in
oxygen transfer at hydraulics structures. Journal of Hydraulic Engineering, ASCE,
124(7), 664–671.
8374 A. Baylar et al. / Expert Systems with Applications 36 (2009) 8368–8374

Hanbay, D., Turkoglu, I., & Demir, Y. (2007). Predicting chemical oxygen demand
(COD) based on wavelet decomposition and neural networks. CLEAN – Soil, Air,
Water, 35(3), 250–254.
Suykens, J. A. K., & Vandewalle, J. (1999). Least squares support vector machine
classifiers. Neural Processing Letters, 9(3), 293–300.
Tsang, C. C. (1987). Hydraulic and aeration performance of labyrinth weirs, PhD Thesis.
University of London, London.
Wilhelms, S. C., Gulliver, J. S., & Parkhill, K. (1992). Reaeration at Low-Head
Hydraulic Structures. Tech. Rep. Hl-91, US Army Engineer Waterways
Experiment Station, Vicksburg, Miss.
Wormleaton, P. R., & Tsang, C. C. (2000). Aeration performance of rectangular
planform labyrinth weirs. Journal of Environmental Engineering, ASCE, 126(5),
456–465. <http://www.esat.kuleuven.ac.be/sista/lssvmlab>.