Ain Shams Engineering Journal (2016) 7, 819826

Ain Shams University

Ain Shams Engineering Journal

www.elsevier.com/locate/asej
www.sciencedirect.com

CIVIL ENGINEERING

Prediction of hydraulic jump length downstream

of multi-vent regulators using Artificial Neural
Networks
Mohamed F. Sauida *

Faculty of Engineering, Beni-Suef University, Egypt

Received 29 June 2015; revised 5 November 2015; accepted 15 December 2015

Available online 7 January 2016

KEYWORDS Abstract Multi-vent regulators are widely used in Egypt. Operation system of gates affects the
Hydraulic jump; characteristics of the flow downstream (DS) of multi-vent regulators. The current study aimed
Multi-vent regulators; toward introducing the artificial intelligence technique as a new modeling tool in the prediction
Modeling; of the characteristics of the flow downstream of multi-vent regulators under the management
Hydraulic structures; and operating systems of multi-gates. Specially Artificial Neural Network (ANN) is utilized in
Artificial Neural Network the current study in conjunction with experimental data to predict the relative length of the sub-
merged hydraulic jump (HJ) occured DS of multi-vent regulators under the different cases of gates
operation. The results show that ANN technique is very successful in simulating the relative sub-
merged hydraulic jump length occurred in stilling basins downstream of multi-vent regulators better
than multiple linear regression (MLR).
2015 Faculty of Engineering, Ain Shams University. Production and hosting by Elsevier B.V. This is an
open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction
The management of the operation system of the multi-gates
plays an important role for precising the characteristics of flow
DS of multi-vent regulators. Fahmy [1], investigated the free
HJ using sills under the gates with different DS slopes and dif-
ferent heights. Different operating systems of gates were
employed. For three-vent regulators, it is preferred to open
* Tel.: +20 55/2312845, mobile: +20 1005196189.
E-mail address: mohamedsauida@eng.bsu.edu.eg
Peer review under responsibility of Ain Shams University.
Production and hosting by Elsevier
http://dx.doi.org/10.1016/j.asej.2015.12.005
2090-4479 2015 Faculty of Engineering, Ain Shams University. Production and hosting by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
the two separate vents instead of the two adjacent ones for
the same flow conditions. Operating the three vents gives max-
imum length of the HJ when compared to opening one vent
and two vents.
Abdel-Aal et al.[2], analyzed the properties of the free HJ
formed just DS of multi-vent regulators in radial stilling basin.
Different operating systems of the multi-vent regulators were
tested under the same flow conditions. Theoretical models
for the symmetrical and asymmetrical operation were derived.
Elfiky [3], studied the effect of end-sill on the characteristics of
the free hydraulic jump created DS of multi-vent regulators in
a radial stilling basin. This end sill was used to control and
increase the jump stability and consequently reduce the con-
struction costs. Abdel-Aal [4], studied the effect of angled solid
deflector in controlling free HJ created DS of multi-vent regu-
lators in a radial SB. Theoretical and statistical models for
820
Nomenclature
B width of channel (L)
bt
P total widths of the opened vents (L)
bd sum of widths of the deflectors from Side view pro-
jection (L)
e expansion ratio, B/bt ()
F1 inflow Froude number ()
Lb basin length (L)
Ld distance between current deflector and the begin-
ning of the basin (L)
LJ the jump length (L)
n no. of the working vents ()
Q discharge (L3 T
1)
S degree of submergence, Y3/Y1 ()
Y1 initial depth of the jump (L)
Y3 back up water depth just downstream the gate (L)
different operating systems were derived. Sauida [5], studied
the effect of sharp radial current deflector in controlling the
submerged hydraulic jump created DS of multi-vent regula-
tors. The different parameters (location, height, angle, width)
of the sharp radial current deflector had been studied experi-
mentally. The management and operating systems of multi-
gates were tested under the same flow conditions. Statistical
model was proposed to compute the length of the submerged
hydraulic jump based on the result of the dimensional analysis
using multiple linear regression analysis.
Sauida [6], studied experimentally the separation of flow for
a sudden expanding stilling basin at the downstream of multi-
vent regulators. He found that the percentage area of reverse
flow is dependent mainly on the expansion ratio (e), and par-
tially on the Froude number under gates (Fg). A statistical
model was developed based on multiple linear regression, to
predict the percentage area of reverse flow at the downstream
of multi-vent regulators.
Sauida [7], investigated the effect of expansion ratio, and
type of gates operation on the discharge of a multi-
submerged flow gates. He found that, increasing the expansion
ratio and the asymmetric operation of gates has a positive
effect on the discharge coefficient. Dimensionless general equa-
tion for computing the coefficient of discharge of the sluice
gate is developed using the multiple regression analysis.
In this research paper, an accurate prediction model using
the Artificial Neural Network was developed to predict the rel-
ative length of the submerged hydraulic jump created DS of
multi-vent regulators under different cases of gates operation.
These results of the ANN model were compared with the sta-
tistical one of Ref. [5].
1.1. Applications of ANN
Artificial Neural Networks were applied in many branches of
engineering sciences. In hydraulic Engineering, Artificial Neu-
ral Networks were applied effectively. Negm et al. [8] investi-
gated an ANN model to predict the characteristics of spatial
hydraulic jump formed at asymmetric sudden expansion and
in symmetric diverging channel. The results of the ANN mod-
els were compared to regression and theoretical models, and
M.F. Sauida
Yt tail water depth at the end of the jump (L)
Zd height of deflector (L)
c specific weight (M L
2 T
2)
g factor related to flow condition ()

g factor related to operating system of gates ()
W central angle of deflector ()
Abbreviations
ANN Artificial Neural Network
DS downstream
HJ hydraulic jump
MLP multi-layer perceptions
MLR multiple linear regression
PE processing element
also, were compared to experimental data. It was found that
the ANN model is much better than the other developed mod-
els. Esmaeili Varaki and Omid [9] adapted an ANN with mul-
tilayer perception structure, to model conjugate depth and
gradually expanding jump length. The optimal models are cap-
able of predicting conjugate depth and jump length for a wide
range of conditions. Guven et al. [10] developed an ANN mod-
els to simulate the mean pressure fluctuations beneath a
hydraulic jump occurring on sloping stilling basins. Multi-
layers feed a forward neural network with a back-
propagation learning algorithm to model the pressure fluctua-
tions beneath such a type of hydraulic jump (B-jump). The
results of the ANN models are found to be superior to the
regression models and are in good agreement with the experi-
mental results due to relatively small values of error (mean
absolute percentage error).
Abdeen [11] utilize the ANN technique in investigating the
impact of submerged aquatic weeds on the water surface pro-
file in an experimental flume that supplies water to different
distributaries. Several ANN models were developed to predict
the water surface profile pattern when the upstream water
depths in the main vegetated channel are higher than the
smooth channel water depths for the same flow rates. The
results show that the ANN technique is very accurate in simu-
lating the water surface profile with the existence of submerged
aquatic weeds. Salmasi [12] developed ANN models, for calcu-
lating the energy dissipation of flow over stepped spillway
chute. The outcomes are valid for a wide range of stepped
chute geometry and flow conditions. Bonakdari et al. [13] pre-
sented numerical analysis and prediction of flow field in a 90
bend using ANN and GA. Results of ANN that had been
trained using GA and BEP indicate that the velocity field is
predicted with good approximation in both training methods
and it is concluded that the proposed procedures are useful
for velocity prediction in channel bends. Salazar et al. [14] ana-
lyzed the radial-gated spillways, and the Oliana Dam is consid-
ered as a case study and the discharge capacity was predicted
both by the application of a level-set-based free-surface solver
and by the use of traditional empirical formulations. The
results of the analysis were then used for training an ANN
to allow real-time predictions of the discharge in any situation
Prediction of hydraulic jump length
of energy head and gate opening within the operation range of
the reservoir.
Hidayat et al. [15] predicted discharge from water level
information from gauge stations in the river. A hindcast model
was established for a tide-dominated lowland site along the
Mahakam River (East Kalimantan, Indonesia), using an
ANN model. The optimized ANN-based hindcast model pro-
duces a good discharge estimation, as shown by a consistent
performance during both the training and validation periods.
Al Suhili et al. [16] developed ANN models to express the dis-
charge coefficient for a weir with three rectangular bottom
openings as a function of different geometrical and flow vari-
ables. The ANN models produce a good result with range of
correlation coefficient (0.9990.941). Abdeen et al. [17] showed
in a detailed fashion the development process of different
ANN models to accurately predict the hydraulic jump lengths
acquired from different experimental studies. The results
obtained from implementing these models show that ANN
techniques are very successful in simulating the hydraulic jump
characteristics occurred in stilling basins.
This paper presents the results of using the ANN technique
to build a prediction model of the relative length of submerged
HJ DS of multi-vent regulators under different cases of gates
operation in terms of the basic involved parameters in the pre-
viously developed regression model given by Eq. (3).
2. Theoretical background
(Fig. 1) shows a definition sketch for the experimental model
showing the basic variables. The relevant parameters can be
written as follows:
X from wood (Pezaform). The middle part consisted of five rows
fq; g; bt ; B; Zd ; Ld ; V; Lb ; W; bd ; g; g; l; LJ ; Y3 ; Yt ; Y1
P q is the water density, B is the
in which: V is the mean velocity,
total width of the channel, bd is the sum of widths of the
Figure 1 Definition sketch showing flow DS of multi-vent regulators.
821
deflectors from Side view projection, bt is the width of the
opened vents, Y1 is the depth of water at the vena contracta,
HU is the upstream water depth, Y3 is the backup water depth,
Yt is the tail water depth, and l is the dynamic viscosity of
water, g is a factor related to flow condition (symmetric or
asymmetric flow),  g is a factor related to operating system of
gates, Zd is the height of deflector, W is the central angle of
deflector, Lb is the basin length, Ld is the distance between cur-
rent deflector and the beginning of the basin, and LJ is the
jump length.Q
Using -theorem, considering q; V; and Y1 as the
independent variables, and by neglecting the effect of Rn and
Y3
Y1
= (constant value = 4), Eq. (1) takes the following form
Ref. [5]:
 P 
bd Zd Ld Yt
Lj =Y1 f F1 ; e; ; ; ; ; w; g; 
g 2
B Y1 Lb Y1
in which F1 is the inflow Froude number, and e is the expan-
sion ratio (e = B/bt).
3. Experimental collection of training data
The training data were collected from the experiments, which
carried out in the Hydraulic Research Institute (HRI),
National Water Research Center (NWRC), Ministry of Water
Resources and Irrigation (MWRI), Delta Barrages, Egypt. The
model which constructed in HRI simulated new Assuit Bar-
rages in a re-circulating flume laboratory 175 cm wide, 40 cm
deep, and 15 m overall length of the flume . The discharge
was measured by Ultrasonic Flowmeter. The model was made
0 of vents each row has five adjacent vents, each one 25 cm wide,
36 cm deep, and 85 cm length. The width of pier was 10 cm
1
and the width of the wing wall was 5 cm. A control sluice gate
made from the wood (Pezaform) was used to control the
upstream depth and the gate opening. A total of about 1216
822 M.F. Sauida

runs were performed, during the course of the experiments, 5. Modeling of length of submerged HJ using ANN
and all different cases of multi-gates management and operat-
ing systems were tested under the same flow conditions for 5.1. Basic definitions
symmetrical and asymmetrical flow. The operating system
includes opening Five, Four, three, two, and one vent and
Activation function is an element of a neuron in a Multi-
the expansion ratio e = (B/bt). The range of the experimental
layer Perceptron. A smooth nonlinear function was used
data was as follows: Q = (1040 L/s), Froude numbers
to modify the summed inputs to the neuron.
(F1 = 2.05.5), Submergence ratio (S = Y3/Y1 = 4). The dis-
Back propagation is an aspect of the learning algorithm
charge was adjusted to the desired value and the gate was
used in a Multi-layer Perceptron. It refers to use the error
opened to the desired opening for the required flow conditions.
generated at the output layer of neurons to alter weights
When the flow became at the steady state conditions, the
between neurons in previous layers. Multi-layer Perceptron
upstream water depth, the sequent water depth, the length of
referred to as back propagation network. Neural connec-
jump and the water depth DS the pier were measured.
tions utilize both conjugate gradient and steepest descent.
The deflector models had been changed and examined as
Conjugate gradient is a learning algorithm that is used by
follows: Ld/Lb = 0.10, 0.20, 0.40, 0.60, and 0.80. Moreover,
Multi-layer Perceptron. It is a more sophisticated algorithm
Six deflection angles of the deflector were W = 5, 20, 45,
than steepest descent but requires more computation.
70, 90 and the deflector relative heights Zd/Y1  were  1.43,
P Global solution is the best model that can be built for a par-
bd
1.90, 2.38. Finally, Five relative deflector widths of ticular problem, defined as the solution with the lowest
B
error.
0.20, 0.40, 0.50, 0.80, 1.00 were tested. Hidden layer is a layer of nodes in a neural network that has
no direct connections outside the neural network.
4. Modeling of relative length of jump using (MLR) Learning algorithm is a rule that indicates how learning
should take place. The selection of a learning algorithm
The multiple linear regression tool that is available in the neu- influences how well a neural network becomes able to solve
ral connection software was used to develop a prediction a problem.
model to predict the length of the submerged hydraulic jump Multi-layer Perceptron (MLP) is a supervised neural net-
created DS of multi-vent regulators. Several models were tried. work. This is the most common neural network technique
According to Ref. [5], the following model has the best perfor- and is built from several layers of connected nodes.
mance compared to all tested models of different forms: Neuron is a small processing element from which a neural
 P network is built. Each neuron performs a simple calcula-
LJ bd Yt Zd
exp  0:0257F1
0:117 0:0497
0:009149 tion. It is by arranging neurons in connected layers that
Y1 B Y1 Y1
 neural networks become able to solve complex problems.
Ld Normalization is a technique for limiting the variance of a
0:135
0:025e
0:05986g 0:000913g 3:016 :
Lb field.
3 Overtraining means if training is allowed to progress for too
long, overtraining is said to has taken place. In these cases,
The values of (g,g) are presented in Table 1. Eq. (3) has a a neural network loses the ability to generalize.
square correlation coefficient (R2) of 73.10%.

Table 1 The values of (g, 

g) of Eq. (3).
Cases of gates operation g (Factor of flow g (Factor of
condition) operation)
G1 (First gate from left bank is opened) 1 3
G2 (Second gate is opened) 1 4
G3 (Third gate is opened) 2 5
G12 (First and Second gates are opened) 1 6
G13 (First and Third gates are opened) 1 7
G14 (First and Second gates are opened) 1 8
G15 (First and Fifth gates are opened) 2 9
G23 (Second and Third gates are opened) 1 10
G24 (Second and Fourth gates are opened) 2 11
G123 (First, Second, and Third gates are opened) 1 12
G124 (First, Second, and Fourth gates are opened) 1 13
G125 (First, Second, and Fifth gates are opened) 1 14
G124 (First, Second, and Fourth gates are opened) 1 15
G135 (First, Third, and Fifth gates are opened) 2 16
G234 (Second, Third, and Fourth gates are opened) 2 17
G1234 (First, Second, Third, and Fourth gates are opened) 1 18
G1235 (First, Second, Third, and Fifth gates are opened) 1 19
G1245 (First, Second, Fourth, and Fifth gates are opened) 2 20
G12345 (First, Second, Third, Fourth, and Fifth gates are opened) 2 21
Prediction of hydraulic jump length 823

Steepest descent is a learning algorithm that is used by neurons within the layer. A typical network of this type of size
Multi-layer Perceptrons. It is a simpler algorithm than con- 9-7-1 is shown in Fig. 2.
jugate but takes more steps to reach the solution. The network consists of three layers. Input layer of 9 neu-
Sub-optimal solution is a stable solution that does not have rons represents the variables F1, Rbd/B, Yt/Y1, Zd/Y1, Ld/Lb, e,
an error as the global solution. W, g and  g. An output layer represents the single output vari-
Supervised learning is a training process in which a teaching able LJ/Y1. The hidden layer between the input and the output
mechanism is used to make neural network associate targets layer receives the input, performs the computations and sends
with inputs. Training data must contain targets for super- the outputs to the output layer. The hidden layer uses a trans-
vised learning to take place. fer or activation function to modify the input to the neuron.
Target data are the data field that a neural network is trying The transfer function may be linear, sigmoid or tanh. Sigmoid
to learn = output data. and tanh are smooth nonlinear functions and one of them is
Test data are a set of complete data that are kept aside to normally chosen because the learning algorithm requires a
confirm the performance of an application. response function with a continuous, single-valued first deriva-
Training data are the data records that are used to train a tive. Building the MLP network is known as training.
neural network. Training proceeds in the following way. First the weights
Validation data are the data that are used to prevent a neu- and biases in the network are initialized, usually to small ran-
ral network from overtraining. It is important that the val- dom values. A training pattern is then applied to the input
idation data and training data both cover the full range of layer and the activation of neurons in the first hidden layer
the problem. is calculated. The output produced by these neurons via the
Weights are the values associated with the connections transfer function is then fed to the neurons of the following
nodes in a neural network. Adjusting the weights causes layer. This forward pass process is repeated at each layer until
the neural network to learn. an output signal from the neurons of the output layer is
obtained. The difference between the actual (target) and the
desired output values is measured, and the network model con-
5.2. Basics of ANN nection strength is modified such that the desired output from
the network becomes more close to the actual values. This is
Artificial Neural Networks or simply neural networks are a achieved by a backward pass process during which connection
new way of analyzing data. They have a unique ability to learn modifies are propagated back through the network, starting
complex patterns and trends in the data emulating the human with the connections to the output layer and ending with those
ability to solve complicated problems. Therefore, neural com- at the input layer (from which comes the name
puting is based on the way that biological system, such as
human brain works. The human brain is made up of many
neurons. Each of the brain cells is connected to many others
in a network that adopts and changes as the brain learns. In
neural computing, processing elements (PE) replace the neu-
rons and these PEs are linked together to form neural net-
works. Each PE performs a simple task. It is the connections
between the PEs that give neural networks the ability to learn
patterns and interrelationship in data through the certain
weights associated with the connections of PEs. The most
important tool of ANN applied in this study is the multi-
layer perceptrons (MLP) which is a neural network modeling
tool that is optimized for prediction and forecasting applica-
tions, Neural Connection [18]. The MLP can be used to clas-
sify patterns or to predict values from data. The MLP learns
from examples by providing the data to the neural network
through an input tool or input layer which consists of a num-
ber of neurons equal to the number of input variables or fields
such that each PE represents one variable or one filed. These
data are processed in a second layer called hidden layer
because of all the processing and computations through this
core layer are hidden from the users. More than one hidden
layer could be used depending on the complexity of the appli-
cation (if no hidden layer is used, it is a simple perceptron).
The processed data are sent to the output layer that consists
of a number of PEs equal to the number of targets (or number
of desired outputs) in the application. All the PEs or neurons
of each layer are connected to the PEs of the preceding and fol-
lowing layers by links or connections without feedback (there- Figure 2 The developed network of size 9-7-1 to predict the
fore they are MLP foreword networks) but not to other length of submerged jump DS of multi-vent regulators.
824 M.F. Sauida

back-propagation neural network). The connection strength is

Table 2 Weights versus R2 for all data case.
modified according to the produced output of the network. If
the network output is correct, no change in the weights takes Weights Training data (R2) Validation Test data (R2)
place. If the network output is larger than the desired output data (R2)
at any neuron, then the connection weights between this neu- (+/
)0.1 0.6992 0.7261 0.7203
ron and all the input neurons decrease and vice versa. It should (+/
)0.4 0.7014 0.7253 0.7202
be noted that the training algorithm should be run from many (+/
)0.8 0.6978 0.7256 0.7175
different starting values to find a good optimum solution (+/
)1.0 0.6991 0.7258 0.7209
(lower minimum mean square error). This is because the gradi- (+/
)1.5 0.6984 0.7241 0.7190
(+/
)2.0 0.7056 0.7199 0.7147
ent descent used by the algorithm finds only the nearest local
(+/
)4.0 0.5890 0.5490 0.5918
minimum in the mean square error from any given set of initial
connection values.

5.3. Results of the developed ANN model

Table 3 Activation function versus R2 for all data case.
The following parameter (LJ/Y1) was modeled using all the Activation Training data Validation data Test data
data from the experimental work. The following steps were fol- function (R2) (R2) (R2)
lowed to develop the network.
tan h 0.6991 0.7258 0.7209
Linear 0.6991 0.7265 0.7181
(1) The data set for the specified case was prepared in a file Sigmoid 0.6961 0.7249 0.7169
with the format suitable for the Neural Connection
software.
(2) The Neural connection was operated and the needed
tools from the toolbar were connected together to form
general topology of the network. Table 4 MLP structure (topology) versus R2 for all data case.
(3) The input tool was used to specify and allocate the Topology Training data Validation data Test data
input data. In this stage the order of the data was ran- (R2) (R2) (R2)
domized and divided into three subsets: training data 9-2-1 0.7407 0.7475 0.7748
(70%), validation data (15%) and test data (15%). 9-4-1 0.7750 0.7626 0.7871
All of them were selected randomly. 9-6-1 0.8298 0.8074 0.8032
(4) The MLP tool was used to specify the following 9-7-1 0.8790 0.8855 0.8460
parameters: 9-8-1 0.8701 0.8879 0.8440
a. The normalization method of the input fields (zero 9-10-1 0.8739 0.8798 0.8317
mean unit standard deviation was used).
b. The activation function (obtained by trail and
error).
c. The normalization method of the target fields (zero Table 5 No. of iterations versus R2 for all data case.
mean unit standard deviation was used).
d. The range of the initial weights of the links No. of Training data Validation data Test data
iteration (R2) (R2) (R2)
between neurons (obtained by trail and error).
e. The learning algorithm (conjugate gradient is 10 0.7301 0.7641 0.7584
used). 40 0.8648 0.8625 0.8405
f. The maximum updates or maximum number of 50 0.8677 0.8671 0.8529
60 0.8677 0.8595 0.8259
iterations (obtained by trail and error).
70 0.8721 0.8791 0.8276
200 0.8571 0.8511 0.8231
(5) The network was allowed to train and the validation
system error is observed. Once the minimum validation
system error was reached, the training is stopped. The
output tool was used to save the output results for The best range to initialize the weight matrix is (1.0),
training, validation and test in three separate files. as shown in Table 2.
(6) The output files were transferred to another software as The best activation function is tanh, as shown in
Excel or SPSS to enable other plotting of results, com- Table 3.
parisons and computations of the square correlation The best number of neurons in the hidden layer is 7
coefficient (R2) between the measured (target) and the neurons, as shown in Table 4.
predicted (network output). The best number of iterations is 50 iterations (maxi-
(7) So many runs were tried to obtain the best range of ini- mum updates), as shown in Table 5.
tial weight, the best activation function, the best num-
After determining the neural network structure, the stabil-
ber of neurons of the hidden layer and the best number
ity of such model was investigated. Several computer experi-
of iteration. Typical results for predicting LJ/Y1 for the
ments were conducted using the pre-determined of the
all data case are as follows:
network but in each case the data sets were chosen randomly
Prediction of hydraulic jump length 825

Table 6 Seeds versus R2 for different data sets for predicting Lj/y1.
Data set Exp. 1 seed 1 Exp. 2 seed 2 Exp. 3 seed 3 Exp. 4 seed 5 Exp. 5 seed 7
Training 0.8236 0.8106 0.8408 0.8677 0.8458
data (R2)
Validation data (R2) 0.7816 0.8175 0.8216 0.8671 0.8005
Test data (R2) 0.7800 0.8465 0.8224 0.8529 0.7806

35
36

30
Predicted (Lj/Y1)

32

Predicted (Lj/Y1)
25 Test Data
Regression
28 ANN
20

15 24
15 20 25 30 35
Measured (Lj/Y1)

20
Figure 3 Comparison between prediction of the ANN model
20 24 28 32 36
and the experimental data for the test data set. Measured (Lj/Y1)

Figure 5 Comparison between predictions of ANN model,

regression model Eq. (3) and the measured experimental data.
2
1.5
1 mental data were compared to the prediction of different
0.5 ANN and MLR models for LJ/Y1 as shown in Fig. 5. Clear
Residuals

Test Data
0 inspection of the figure shows that the prediction of the
-0.5 15 20 25 30 35 ANN model is very accurate and closer to the experimental
-1 data compared to the MLR model.
-1.5
-2 7. Conclusions
Predicted (Lj/Y1)

Figure 4 Variation of the residuals with the predicted (LJ/Y1) of
the ANN model for the test data set.
in different ways. The results of five experiments for all data
model (LJ/Y1) using different seeds of 1, 2, 3, 5, and 7 are
shown in Table 6. Out of 5 computer experiments, the R2 of
the data of experiments deviated by a value does not exceed
4% from the average R2 of all the 5 experiments indicating rel-
atively stable network, as shown in Table 6.
The output of the developed network (LJ/Y1) after stability
was related to the measured and residuals data as shown in
Figs. 3 and 4. From all of these figures, it is clear that the out-
put of the developed network is well compared to the mea-
sured data. The residuals are being symmetrical around the
line of zero error and are uncorrelated to the network output
as the correlation coefficient between the residuals and the net-
work output (estimated and predicted values) is very small
(mostly zero or near to zero).
6. Comparisons
The predicted equation developed by regression analysis was
used to compare the results of the ANN model. The experi-
The ANN was used to develop a prediction model to predict
the relative length of submerged hydraulic jump DS of
multi-vent regulators under different cases of gates operation.
The ANN model was trained, verified and tested using exper-
imentally collected data. The results of the ANN model were
compared to those of the previously developed model using
multiple linear regression analysis (MLR). The ANN model
is very successful in predicting the relative submerged hydrau-
lic jump downstream of multi-vent regulators better than
MLR model, and an improvement by about 12% in R2 is
obtained compared to MLR model.
References
[1] Fahmy MR. Design of stilling basins downstream of multi-vents
regulators. Unpublished Ph. D. Egypt: Faculty of Engineering,
Zagazig University; 2001.
[2] Abdel-Aal GM, Elfiky MM, Negm AM. Analysis of free
hydraulic jump characteristics DS multi-vent regulators. Cairo,
Egypt: Scientific Bulletin, Faculty of Engineering, Ain Shams
University, vol. 39(3); 2004. p. 47594 [September 30].
[3] Elfiky MM. Effect of end sill on free hydraulic jump character-
istics DS multi-vent regulators. Cairo, Egypt: Scientific Bulletin,
Faculty of Engineering, Ain Shams University, vol. 39(4); 2004. p.
33143 [December].
826
[4] Abdel-Aal GM. Control of hydraulic jump DS multi-vent
regulators in a radial stilling basin. Cairo, Egypt: Scientific
Bulletin, Faculty of Engineering, Ain Shams University, vol. 40
(1); 2005. p. 390416 [March 31].
[5] Sauida MF. Characteristics of flow in diverging stilling basin
downstream of regulators under vertical deformation conditions.
Unpublished Ph.D. Egypt: Faculty of Engineering, Zagazig
University; 2009.
[6] Sauida MF. Reverse flow downstream multi-vent regulators.
Cairo, Egypt: Scientific Bulletin, Faculty of Engineering, Ain
Shams University, vol. 4(2); 2013. p. 20714 [June].
[7] Sauida MF. Calibration of submerged multi-sluice gates. Alexan-
dria, Egypt: Scientific Bulletin, Faculty of Engineering, Alexan-
dria University, vol. 53(3); 2014. p. 6638 [September].
[8] Negm AM, Ibrahim AA, Abdel-Aal GM, El-Saiad AA. Modeling
of characteristics of hydraulic jumps at asymmetric sudden
expansions and symmetric divergence using neural networks,
vol. 37(4), Cairo, Egypt: Scientific Bulletin, Faculty of Engineer-
ing, Ain Shams University, Faculty of Engineering; 2002. p. 195
219.
[9] Esmaeili Varaki M, Omid MH. A neural network model to
predict characteristics of gradually expanding hydraulic jumps. J
Agric Sci Nat Resources 2005;12(3):2132.
[10] Guven A, Gunal M, Cevik A. Prediction of pressure fluctuations
on sloping stilling basins. Can J Civ Eng 2006;33(11):137988.
[11] Abdeen M. Predicting the impact of vegetations in open channels
with different distributaries operations on water surface profile
using Artificial Neural Networks. J Mech Sci Technol
2008;22:183042.
M.F. Sauida
[12] Salmasi Farzen. An Artificial Neural Network (ANN) for
hydraulics of flows on stepped chutes. Eur J Sci Res 2010;45
(3):4507.
[13] Bonakdari H, Baghalian S, Nazari F, Fazli M. Numerical analysis
and prediction of the velocity field in curved open channel using
Artificial Neural Network and genetic algorithm. J Eng Appl
Comput Fluid Mech 2011;5(3):38496.
[14] Salazar Fernando, Moran Rafael, Rossi Riccardo, Onate Euge-
nio. Analysis of the discharge capacity of radial-gated spillways
using CFD and ANN Oliana Dam case study. J Hydraul Res
2013;51(3):24452.
[15] Hidayat H, Hoitink A, Sassi M, Torfs P. Prediction of discharge
in a tidal river using Artificial Neural Networks. J Hydrol Eng
2014;19(8):04014006.
[16] Al Suhili RH, Al Bidhani JH, Al Mansori NJ. Hydraulic
characteristics of flow over rectangular weir with three rectangular
openings using ANN. J Babylon Univ 2014;4(22).
[17] Abdeen M, Abdin A, Abaas W. Simulation and prediction of
energy dissipaters and stilling basins design using
artificial intelligence technique. J Cogent Eng 2015;2:
1018705.
[18] Neural Connections. Artificial Neural Networks software
and user manuals. SPSS Inc./Recognition Systems Inc.; 1998.
Mohamed F. Sauida is a currently lecturer in Faculty of Engineering,
Beni-Suef University. He received his B.Sc. (Civil) degree, M.Sc. (Civil)
degree, and Ph.D. degree from Zagazig University in 1995, 2003, and
2009 respectively.