You are on page 1of 12

Engineering Applications of Computational Fluid

Mechanics

ISSN: 1994-2060 (Print) 1997-003X (Online) Journal homepage: https://www.tandfonline.com/loi/tcfm20

Evaluation of Sediment Transport in Sewer using


Artificial Neural Network

Isa Ebtehaj & Hossein Bonakdari

To cite this article: Isa Ebtehaj & Hossein Bonakdari (2013) Evaluation of Sediment Transport in
Sewer using Artificial Neural Network, Engineering Applications of Computational Fluid Mechanics,
7:3, 382-392, DOI: 10.1080/19942060.2013.11015479

To link to this article: https://doi.org/10.1080/19942060.2013.11015479

Copyright 2013 Taylor and Francis Group


LLC

Published online: 19 Nov 2014.

Submit your article to this journal

Article views: 731

View related articles

Citing articles: 24 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tcfm20
Engineering Applications of Computational Fluid Mechanics Vol. 7, No.3, pp. 382–392 (2013)

EVALUATION OF SEDIMENT TRANSPORT IN SEWER USING


ARTIFICIAL NEURAL NETWORK
Isa Ebtehaj and Hossein Bonakdari *

Department of Civil Engineering, Razi University, Kermanshah, Iran


*E-Mail: bonakdari@yahoo.com (Corresponding Author)

ABSTRACT: Sedimentation in sewers occurs regularly according to the alternating natural flow. The long term
deposit of material in the sewerage systems increases the risk of changes in the sediments and their consolidation
and cementation. In particular under low flow conditions, permanent settlement similar to that on the sewer bed
alters the nature of velocity and distribution of the boundary shear stress. Consequently, it affects the capacity of
sediment transport and the hydraulic resistance of the sewer. The article reviews the application of Artificial Neural
Network (ANN) in predicting the sediment transport using the concept of self-cleansing of sewer systems. In
comparison with existing methods, the ANN showed acceptable results.
Keywords: self-cleansing, sediment, ANN, sewer, bed-load

1. INTRODUCTION These again, in order to maintain self-cleansing


conditions, depend on minimum slope, minimum
The occurrence of sedimentation in complex velocity of flow or minimum shear stress that the
sanitary sewerage systems, caused by alternation flow should exert on the walls of the pipe (Ashley
in flow, has been a matter of lengthy deliberations et al., 2004).
by many researchers. Low velocities allow debris Effective self-cleansing occurs when the sewer
to settle along the invert of the pipe during has the sediment transport capacity to establish a
periods of minimum discharge. A long term balance between the amounts of deposited and
accumulation of deposit in the sewer increases the eroded materials and which, for an average time
risk of changes occurring in the sediments such as and depth of sediment deposit, minimizes the
consolidation and cementation. In particular, combined costs of construction, operation and
under minimum flow conditions the permanent maintenance (Butler et al., 2003). The most
settlement on the sewer bed alters the nature of important aspect of this definition would be its
velocity and the distribution of the boundary relation to an economic solution, viz if allowing a
shear stress. Accumulation of sediments results in minimum amount of deposit were to accrue less
a loss of carrying capacity that may cause costs, there would be no need for the sewer to be
surcharge or local flooding and the setting up of totally free of sediment deposits. Therefore, by
septic conditions which create odour and considering a minimum deposition depth and by
corrosion problems, (Bonakdari and Larrarte, reducing the gradient required for sediment
2006). The existence of sediment in the sewer transport, it will be possible to increase the
networks can increase the overall hydraulic deposit transport capacity.
roughness of the pipe and reduce the size of the Novak and Nalluri (1975) have expressed the
cross section. This in turn can change the effective factors regarding sediment transport in
hydraulic conditions that have influence on the terms of relative relations, as the parameters of
sediment carrying capacity. These changes transport and flow for circular and rectangular
modify the physical properties of the system and channels. May (1982) developed a theoretical
the turbulence flow structure. In addition, the in- bed-load transport model which is based on forces
pipe deposits can have a substantial impact on the acting on individual sediment particles
flow quality, as at higher flow conditions rate transported at the limit of deposition. May et al.
large quantities of sediment can be released (1989) continued these studies using larger
suddenly into the flow (Ashley et al., 2004). diameter pipes.
The prediction of the probability of sediment Mayerle et al. (1991) did studies on the
deposit in a given sewer network depends on its sedimentation in condition of half-filled flow in a
design, based on a set of conditions called self- smooth tube and in a rectangular channel. Nalluri
cleansing criteria that are related to the flow. and Ab. Ghani (1996) presented new design

Received: 12 Jan. 2013; Revised: 20 May 2013; Accepted: 27 May 2013

382
Engineering Applications of Computational Fluid Mechanics Vol. 7, No. 3 (2013)

criteria for a clean pipe (non-deposition). They the results of some recently applied methods were
compared their equation with the criterion of compared and their application to prediction of
minimum velocity of 0.75 m/s, concluding that sediment transport was evaluated.
the minimum velocity criterion over-designs the
slope for small diameters (D < 500 mm) and 2. SEDIMENT TRANSPORT EQUATIONS
under-designs the slope for larger pipe diameters.
May et al. (1996) considered the experimental Novak and Nalluri (1975) proposed the following
data of others (Ackers et al., 1996) and their own equation by applying the parameter of transport
experiments and modified the general form of the CV VR
( ) and the parameter of flow
equation that was proposed by May et al. (1989).
(Ss  1)gd 50
3
Their regression equations provide an acceptable
estimation of the sediment transport. However, no (Ss  1)d 50
(  ) for both circular and rectangular
exact equation or theory has yet been developed. RS
In the past decade, because of the soft computing channels:
methods capability for analyzing complex
problems, the popularity of using these methods   11.6 2.04 (1)
has enhanced the research in various fields of in which V is the limited velocity, R the hydraulic
science, especially in water resources engineering, radius, d50 the average size of the sediment, SS the
water structures and hydrology (Lin et al., 2006; relative density, g the gravitational acceleration, S
Tayfur and Guldal, 2006; Kisi, 2008; the pipe slope and CV the volumetric
Muzzammil, 2008; Wang et al., 2009; Wu et al., concentration.
2010; Bonakdari et al., 2011; Baghalian et al., The rate of sediment concentration was from 20
2012). The ANN technique is one of the most to 2400 (ppm). By using the resistance formula of
popular soft computing methods. Jain (2001) used Darcy-Weisbach, S=λsV2/8gR, and classifying
ANN to introduce the relation of sediment density parts and changing the Froude number, it is
in Mississippi river and proved that it gives more possible to rewrite Eq. (1) as follows:
reliable results in comparison with other methods.
Cigizoglu (2002) compared the results from ANN V d
 1.77C1V/ 3 ( 50 ) 1 / 3 s2 / 3 (2)
with sediment rating curves to predict the density g(SS  1)d 50 R
of suspended sediments. Kisi (2004) made use of
different ANN methods to predict the same and in which λs is the total friction factor. Mayerle et
declared that MLP offered the best results. Kisi al. (1991) conducted their experiments on circular
(2005) employed ANN to model the suspended and rectangular channels, presenting their formula
sediment load of flow. He also used sediment as follows:
rating curves and a multi-regression model to V d
predict the sediment load and showed that ANN  4.32C0V.23 ( 50 ) 0.68 (3)
produced the best results. Raghuwanshi et al. gd 50 (SS  1) R
(2006) developed ANN models to predict both
Ab. Ghani (1993) used dimensional analysis of
runoff and sediment yields on a daily and weekly
parameters, effective for sediment transport in a
basis for a small agricultural watershed, also
non-deposition state. The expression of the non-
applying regression models for the same. In all
dimensional parameters is found as below:
cases, the ANN models performed better than the
models based on linear regression. To evaluate V R
the relations of total sediment load transport and  f (CV , Dgr , , S ) (4)
g(SS  1)d 50 d 50
to predict the concentration of sediment load in
rivers, Yang et al. (2009) applied ANN and The above equation expresses the effects of the
Memarian and Balasundram (2012) used it to sediment concentration, the size of deposits at
estimate sediment load in a tropical watershed. depth of the flow and the size of the pipe on
Azamathulla et al. (2012) applied the adaptive minimum deposit velocity as a function of non-
neuro-fuzzy inference system (ANFIS), a dimensional numbers. Ab. Ghani (1993)
combination of neural network and fuzzy logic, as presented the following equation for transport of
an alternative approach to estimate the bed load bed?-load at limit of deposition by using the
transport in sewers. The main objective of this regression equation:
study is to predict the sediment transport in pipes
V 0.09 0.21 d 50 0.53 0.21
by using ANN, also an experimental investigation  3.08Dgr CV ( )  s (5)
was conducted to verify the ANN results. Finally, gd 50 R

383
Engineering Applications of Computational Fluid Mechanics Vol. 7, No. 3 (2013)

s  1.13D0gr.01C0V.020c.98 (6) generally very complicated. However, it can be


claimed that the number of neurons in the
in which Dgr (=d50((SS-1)/ν2)1/3) is the size of non- intermediary layer is a function of the input
dimensional particles, ∆(=SS-1) is the relative synapses as well as the maximum number of
mass of sediment in water and  C is the clean zones in the input space, linearly distinguishable.
For this reason, the number of neurons in the
water friction factor in the canal. May et al.
hidden layer is generally obtained experimentally.
(1996) proposed the following equations:
Each neuron is connected to those in the next
D 2 d 50 0.6 layer via its output but not to the neurons in its
CV  3.03 102 ( )( ) own layer. The output of each neuron is defined
A D
(7) by the following equation:
V2 V
( )1.5 (1  t )4 n
g ( S S  1) D V a  f ( p i w j,i  b i ) (11)
i 1
y 0.47
Vt  0.125[ g ( SS  1)d50 ]0.5[ ] (8) in which wj,i is the weight of connection between
d50
neuron Mj of the mentioned layer with the Mi of
By reducing the parameters effective on transport the previous layer, underlining the importance of
of the bed-load, Vongvisessmojai et al. (2010) the connection between two neurons in
conducted their experiments on two PVC pipes of contiguous layers, bj is the Bias weight of neuron
100 and 150 mm in diameter. By applying the Mj, pi the output of neuron Mi and f is the
regression analysis for sediment transport in bed- threshold function of neuron Mj. The transfer
load state, they presented the following equations: functions usually have a sigmoid shape (Rezaeian
Zadeh et al., 2010; Jalalkamali and Jalalkamali,
0.616
V d 2011; Chang and Liao, 2012; Dorofki et al., 2012;
 4.31C0V.226  (9) Moharrampour et al., 2012), defined for any
gd (s  1) R
variable as:
0.542
V d 1
 3.57C0V.21  (10) f (12)
gd (s  1)  y 1  exp( z)
In ANN the models are mostly prepared in two
3. ARTIFICIAL NEURAL NETWORKS stages. The first stage is the training of the
network, which is realized on the basis of the
The concept of using Artificial Neural Networks
number of measured data (80% in this study). In
(ANN) is not new, but its application began from the next stage the data are validated by 20 - 30%.
about 1946 by a person named Hu, who used it to In modeling with the help of ANN and depending
predict the weather. However, he did not have on the input variables and their relation with the
much success as he had no computer and relied intended parameters (which are modeled), around
solely on his own calculations (Beals and
10% of the data are used for testing, evaluating
Jackson, 1998). ANN is composed of an the model and estimating the critical points such
interconnected group of neurons, which can as the reference points, i.e. the maximum or
provide outputs by processing the input data. minimum points of a curve. It is obvious that
Neural networks are generally created in a regular these data should not have been used during
and layered form. The first layer which receives training and validation processes. Fig. 1 shows
the input data is the input layer. The intermediary the architecture of the network used in this paper.
layers are the hidden layers and the last layer that
provides the output for the model is the output
layer. The simplest and most conventional type of
a neural network, applied in many fields of
engineering including the present study, is the
Multi-Layer Perceptron (MLP), which utilizes the
backward propagation error for training.
In this network the number of neurons in the input
layer is equal to the number of synapses in the
output. The accurate and actual analysis to find
the number of neurons in the intermediary layer is Fig. 1 Architecture of network used.

384
Engineering Applications of Computational Fluid Mechanics Vol. 7, No. 3 (2013)

4. ADAPTIVE NEURO-FUZZY 1
A (x)  (18)
INFERENCE SYSTEM (ANFIS) x  ci 2 b i
1 | |
Fuzzy logic is a system based on condition-result
ai
regulations which figures out the effects of the in which ai, bi and ci are the parameters of the
input variables on the output variables by the use function that by varying each , the urceolate
of linguistic variables concept and the fuzzy membership function also changes proportionally
decision making procedure. The base of fuzzy and this in fact brings about different forms of a
regulations, which are a complex of logic membership function for the fuzzy set. The
regulations expressing the relationship between existing parameters in this layer are introduced as
fuzzy variables, is considered as the most premise parameters.
important part of a fuzzy system (Gopakumar and Second layer: Nodes which are placed in this
Mujumdar, 2007). Since the given data always layer are considered as node II, the output of
have definite values, a fuzzy-maker has been used which is calculated as follows:
in order to change them into a fuzzy variable, and wi
a non-fuzzy maker has been used in order to O 2,1  w i  .., i  1,2 (19)
convert the results of a fuzzy system into definite
w1  w 2
values. One of the most important types of fuzzy Third layer: Nodes which are placed in this layer
rules is the Takagi-Sugeno (TS) system, which are nodes with title “N”. The ith node calculates
represents the conclusions by functions (Takagi the ratio of the firepower of ith law to the fire
and Sugeno, 1985; Bardossy and Disse, 1993; Sen power of the sum of laws, as follows:
and Altunkaynak, 2006). A rule of this mode will
wi
be: O3,1  w i  .., i  1,2 (20)
w1  w 2
If x1 is A1 and x2 is A2
(13) For simplification, the output of this layer is
THEN y  f ( x1 , x2 ,..., xn ) presented as “Normalized firepower”.
If the fuzzy logic system has 2 inputs (x, y) and 1 Fourth layer: Each "i" node in this layer is a
output (y), the ordinary regulation set of the fuzzy comparative node with the node functions as
system for first grade TS fuzzy model would be in follows:
the form of two "if-then" laws, as follows: O4,i  w i fi  w i (pi x  qi y  ri ) (21)
Rule1:if x is A1 and y is B1
(14)
O5,1   w i f i 
wf i i
THEN f1  p1 x  q1 y  r1 (22)
i

i w i i

Rule2 :if x is A2 and y is B2


(15) 5. DATA COLLECTION
THEN f 2  p2 x  q2 y  r2
Ab. Ghani (1993) conducted tests on three pipes,
Fig. 2a shows the reasoning mechanism for TS
sized 154, 305 and 450 mm in diameter. He used
model and Fig. 2b shows the ANFIS
all the three sizes for smooth rigid beds and the
corresponding to this fuzzy logic.
305 mm pipe for the rough rigid beds. For weak
First layer: Each node in this layer is an
boundary conditions, he used the 450 mm pipe.
Adaptive node accompanied by Node Function.
The maximum input discharge was equivalent to
O1,i  Ai (x),.....for.....i  1,2,...or (16) 20.5 l/s and the maximum slope of the flume 0.06.
The limits of the data in these tests are shown in
O1,i  Bi2 ( y),.....for.....i  3,4 (17) Table 1. The data used in this study relate to 137
tests for transport of the load on the bed at
x, y: Input to node i deposition level, 80% of which (109 samples)
Ai, Bi-1: Linguistic label which belongs to Node i. were used for training, 10% (14 samples) for
O1,i: Membership grade. evaluation and the remaining 10% for testing.
For membership function “A”, a triangular, Vongvisessomjai et al. (2010) conducted their
trapezoidal, urceolate or Gauss type membership tests on two pipes of 100 and 150 mm in diameter
function could be selected. For example the and 16 m in length. Tests of Vongvisessomjai’s et
urceolate function is defined as follows: al. (2010) were conducted in deposited state.
Table 2 shows the range of the data used in their
tests.

385
Engineering Applications of Computational Fluid Mechanics Vol. 7, No. 3 (2013)

Fig. 2 (a) Reasoning mechanism for Takagi-Sugeno model; (b) Scenario of ANFIS model. .

6. RESULTS AND DISCUSSION x i  x min


x̂  (23)
Most researchers considered the effective factors x max  x min
in sediment transport as: limited velocity, VS, the The results of validating the data presented by
size of non-dimensional particles, Dgr, the ANN are shown in Table 3. As shown, the
hydraulic radius, R, the average size of sediment, equation in row 1 (R2 = 0.98 and RMSE =
d50, the relative mass of sediment in water, SS, the 0.0011) give the best results. Fig. 3 shows the
total friction factor, λS, and the friction factor of effect of different parameters in calculating
clean water in canal, λc. To assess the effects of V/√(∆gd50) and predicts the results of the given
these parameters on the results of the ANN phases using ANN.
presented, we have considered 5 different According to the results from Table 3, showing a
situations (Table 3) in order to select the best very small difference among the 5 cases, case 1
model. Also in Table 3, the ANN results are offers the best results. Based on this, one may
compared with the ANFIS model obtained by conclude that using the size of dimensionless
Azamathulla et al. (2012) to show the high particles (Dgr) causes no significant effect on
precision of the ANN model presented. calculating V/√(∆gd50). As known, the only
The noteworthy point in using data in ANN is that variable parameter in estimating (Dgr) is the
the data must be normalized before network average particle size (d50) which effect is
training. This will bring about the uniformity of considered on the parameter d50/R. Considering
the data values for the network and increase the Fig. 3 and the results from Table 3, one may
speed and accuracy of the network. Furthermore, understand that the bed friction coefficient (λs)
for the training data the smallest and the largest produces the main effect on the results and
data are used for obtaining results from the consequently offers the least exactness. It is
network, which are more accurate than the ones in recognized from Fig. 4 that, in general, the given
the situation where the training data do not results seem accurate in every condition, as the
include these two quantities. The following smallest correlation coefficient is 0.92, and it
equation is used to normalize the data: shows the results of predicting sediment transfer
by using ANN in the best given condition (case
1).

386
Engineering Applications of Computational Fluid Mechanics Vol. 7, No. 3 (2013)

Table 1 Range of Ab. Ghani (1993) data. Table 2 Range of Vongvisessomjai et al. (2010) data.

d50 V R CV d50 V y R Cv
No λs No
(mm) (m/s) (m) (m) (ppm)
(mm) (m/s) (m) (ppm)
1 0.97 0.821 0.276 0.0429 320 1 0.2 0.237 0.03 0.017 4
2 0.97 0.517 0.0256 0.0749 14 2 0.2 0.336 0.03 0.017 22
3 0.97 0.613 0.0243 0.0641 49 3 0.2 0.433 0.03 0.025 24
4 0.97 0.697 0.0253 0.0585 125 4 0.2 0.326 0.03 0.012 42
5 0.97 0.844 0.0271 0.0478 262 5 0.3 0.237 0.03 0.017 6
6 0.97 0.973 0.0227 0.0525 379 6 0.3 0.336 0.06 0.017 29
7 2 0.675 0.0284 0.0508 161 7 0.3 0.464 0.06 0.028 30
8 2 0.516 0.0251 0.075 13 8 0.3 0.568 0.03 0.028 71
9 0.43 0.237 0.03 0.017 6
9 2 0.614 0.0238 0.064 61
10 0.43 0.336 0.06 0.017 34
10 2 0.693 0.0257 0.0588 129
11 0.43 0.464 0.03 0.028 35
11 2 0.84 0.0277 0.0479 318 12 0.43 0.411 0.06 0.017 79
12 2 0.699 0.0294 0.033 318 13 0.43 0.568 0.03 0.028 83
13 2 0.997 0.0218 0.0611 235 14 0.2 0.247 0.03 0.018 4
14 4.2 0.671 0.0294 0.051 252 15 0.2 0.349 0.05 0.018 21
15 4.2 0.821 0.0277 0.0429 437 16 0.2 0.466 0.03 0.028 25
16 4.2 0.694 0.0308 0.0364 562 17 0.2 0.427 0.03 0.018 46
17 4.2 0.802 0.0281 0.0472 419 18 0.3 0.247 0.06 0.018 5
18 4.2 0.57 0.0256 0.0725 37 19 0.3 0.362 0.06 0.032 7
19 4.2 0.519 0.0258 0.0748 15 20 0.3 0.511 0.03 0.032 31
20 4.2 0.69 0.0268 0.0589 207 21 0.3 0.427 0.06 0.018 57
21 4.2 0.831 0.0283 0.0482 542 22 0.3 0.626 0.06 0.032 74
22 4.2 0.69 0.0301 0.0333 586 23 0.43 0.247 0.03 0.018 8
23 4.2 1 0.0216 0.0609 313 24 0.43 0.362 0.06 0.032 9
24 5.7 0.669 0.0291 0.0511 254 25 0.43 0.511 0.06 0.032 40
25 5.7 0.806 0.0291 0.0434 662 26 0.43 0.427 0.03 0.018 69
26 5.7 0.849 0.0267 0.0564 366 27 0.43 0.626 0.06 0.032 90
27 5.7 0.705 0.0296 0.036 617 Min 0.2 0.237 0.03 0.012 4
28 5.7 0.798 0.0286 0.0473 537 Max 0.43 0.626 0.06 0.032 90
29 5.7 0.573 0.0253 0.0724 31
30 5.7 0.691 0.0303 0.0333 745 The results of the empirical relations obtained by
31 5.7 0.971 0.0231 0.0525 443 linear regression, and also those obtained by ANN
32 8.3 0.844 0.0271 0.0566 516 are shown in Table 4. Based on this Table, it can
33 8.3 0.688 0.0316 0.0366 867 be shown that Eq. (12) with the average relative
34 8.3 0.788 0.0294 0.0477 705 error of 14.73% has the least error, and also that
35 8.3 0.571 0.0252 0.0725 30 the error of the empirical relations reaches up to
36 8.3 0.824 0.0289 0.0485 765 60% in some cases. Eq. (2) has more changes, so
37 8.3 0.678 0.0316 0.0337 923 the minimum of its errors is 1.7% and the
38 8.3 0.95 0.025 0.0532 837 maximum 56.94%. It can be seen that the results
39 8.3 0.974 0.0235 0.0619 583 obtained by ANN are in appropriate accordance
40 8.3 0.486 0.0158 0.076 0.759 with the experimental data and the average of
Min 0.97 0.486 0.0158 0.033 0.759 relative errors by equalizing is 7.31%, offering
Max 8.3 1 0.276 0.076 923 less error than by the other methods.
Table 3 Study of effects of different parameters on reducing errors of presented model.

Dependent variables Independent variables R2 ANN RMSE ANN R2 ANFIS RMSE ANFIS
1 V/√(∆gd50) d50/R, λs, Dgr, Cv 0.98 0.0011 0.98 0.0024
2 V/√(∆gd50) d50/R, λs, Dgr 0.94 0.037 0.85 2.3765
3 V/√(∆gd50) d50/R, λs, Cv 0.95 0.067 0.89 1.458
4 V/√(∆gd50) d50/R, Dgr, Cv 0.92 0.07 0.93 0.1489
5 V/√(∆gd50) λs, Dgr, Cv 0.97 0.052 0.96 0.0567

387
Engineering Applications of Computational Fluid Mechanics Vol. 7, No. 3 (2013)

(a) Case 1 of Table 3 (b) Case 2 of Table 3

(c) Case 3 of Table 3 (d) Case 4 of Table 3

(e) Case 5 of Table 3


Fig.3 Studying different parameters to predict dimensionless V/√(∆gd50).

Table 4 Comparison between ANN and sediment transport equations with experimental data of Ab. Ghani (1993).

No Exp ANN Eq. 2 Eq. 3 Eq. 5 Eq. 7 Eq. 9 Eq. 10


1 1.53 1.69 2.03 2.22 2.05 1.93 1.87 2.04
2 9.38 8.99 9.55 12.18 11.24 10.72 10.8 10.85
3 7.45 7.01 6.11 9.63 8.78 7.95 7.9 8.13
4 4.73 4.92 3.79 7.33 6.59 5.59 5.62 5.65
5 3.44 3.66 4.36 3.93 3.80 3.71 3.64 3.7
6 1.91 2.04 2.53 2.34 2.40 1.78 1.76 1.94
7 4.77 5.36 2.37 7.69 6.51 6.06 6.12 6.07
8 5.61 6.25 2.41 8.07 6.75 6.41 6.44 6.5
9 3.44 3.02 2.67 4.89 4.23 3.93 3.85 3.96
10 2.59 2.74 2.70 3.51 3.07 3.05 2.98 2.95
11 2.69 2.93 3.05 3.36 3.01 3.15 3.18 3.22
12 2.22 2.25 2.46 2.86 2.55 2.51 2.43 2.55
13 2.65 2.70 3.11 3.22 2.91 2.93 2.98 3.04
14 2.62 2.32 3.62 2.99 2.79 2.95 2.89 2.92

388
Engineering Applications of Computational Fluid Mechanics Vol. 7, No. 3 (2013)

of the Froude number was obtained by the given


relations.
According to Fig. 5, it can be shown that for
Froude numbers less than ~ 3, the relations (7),
(5) and (3) presented almost similar results,
having reasonable agreement with the
experimental data. For higher Froude numbers,
the result of Eq. (7) is more suitable in
comparison with the other equations. Eq. (2) has
larger errors in most cases and the errors are more
than 50% in some cases. It can be seen that the
Fig. 4 Predicted non-dimensional amount of obtained results from Eq. (7) and ANN for Froude
V/√(∆gd50) against experimental data numbers less than 4 are almost equal to and
(Training). consistent with the Ab. Ghani (1993) data. As the
Froude number increases, the accuracy of the
results obtained by ANN increases when
compared with Eq. (7).
Fig. 6 compares the ANN results and the
sediment transport equations with the
Vongvisessmojai et al. (2010) data. It can be
observed that the ANN results and the equations
proposed by Vongvisessmojai et al. (2010) (Eqs.
(9) and (10)) return the best results. At some
points, Eq. (10) has large errors. Although Eq. (9)
returns good results, the ANN model gives greater
accuracy as confirmed by Table 5. It should be
noted that Eq. (9) returns acceptable results only
for the Vongvisessmojai et al. (2010) data, while
Fig. 5 Comparison between ANN and relation of for Froude numbers of over 3 its errors increase
sediment transport in limit of deposition for for the Ab. Ghani (1993) data, , whereas the
bed load with data of Ab. Ghani (1993). ANN returns the best results for both sets of data.

7. CONCLUSIONS

Sediment transport in a pipe is a complex


phenomenon. In this study, a significantly
improved model for predicting sediment transport
is presented, by using the Artificial Neural
Network. The ANN model presented here can
predict it as close to actual measurements as
possible. To overcome the complexities of
calculation of the sediment transport and to
consider also the effective parameters, this study
demonstrates that ANN predicts the sediment
transport in pipes with good accuracy. The
adequate adhesion factors demonstrate the good
Fig.6 Comparison between ANN and relations of conformity that ANN has with the measured data.
sediment transport in limit of deposition for The comparison between the experimental results
bed load with data of Vongvisessmojai et al.
and empirical relations show the same
(2010).
comparable results for Froude numbers of less
Fig. 4 compares the relations of sediment than 3. Also the results show that the error
transport at the limit of deposition for the bed occured by ANN is in the bounds of the minimum
load with the Ab. Ghani (1993) data. The relative error. Among empirical relations, Eq.
horizontal axis of Froude number in Fig. 5 was (12) agrees comparably better than the others with
drawn from the experimental results obtained the experimental results. Therefore, with the use
from Ab. Ghani (1993) data and the vertical axis of ANN, the sediment transport in a sewage

389
Engineering Applications of Computational Fluid Mechanics Vol. 7, No. 3 (2013)

Table 5 Comparison between ANN and sediment transport equations with experimental data of Vongvisessmojai et al.
(2010).

No Exp ANN Eq. 2 Eq. 3 Eq. 5 Eq. 7 Eq. 9 Eq. 10


1 4.16 4.19 6.26 5.08 4.8 4.24 4.00 3.97
2 5.73 5.65 11.33 6.89 6.48 7.30 5.50 6.50
3 4.82 4.90 10.62 6.08 5.61 6.15 4.89 7.03
4 2.84 2.85 5.55 3.31 3.24 2.87 2.74 2.85
5 4.34 4.25 6.49 5.28 4.95 4.84 4.15 3.98
6 6.13 6.20 11.27 7.73 6.96 8.23 6.04 7.41
7 7.50 7.40 14.62 9.26 8.18 8.89 7.21 6.63
8 3.54 3.50 6.11 4.22 4.03 4.86 3.40 4.86
9 5.19 5.17 9.39 6.74 5.85 5.17 5.23 5.22
10 6.13 6.18 13.72 7.39 6.65 8.59 5.89 8.10
11 2.96 2.94 6.39 3.68 3.55 3.56 3.03 3.03
12 4.34 4.27 9.06 5.59 4.93 4.41 4.44 4.52
13 6.12 6.18 14.86 7.88 6.70 6.20 6.21 6.19
14 5.12 5.16 12.97 6.04 5.53 6.03 4.93 4.76

system can be predicted with significantly higher REFERENCES


accuracy than by the use of other comparable
empirical relations reported in the literature. 1. Ab. Ghani A (1993). Sediment Transport in
Sewers. Ph.D. Thesis, University of
ACKNOWLEDGEMENTS Newcastle Upon Tyne, UK.
2. Ackers JC, Butler D, May RWP (1996).
The authors would like to express their Design of Sewers to Control Sediment
appreciation to anonymous reviewers for their Problems. Report No. CIRIA 141,
helpful comments and to Ellen Vuosalo Tavakoli Construction Industry Research and
for final editing of the English text. Information Association, London.
3. Ashley RM, Bertrand-Krajewski JL, Hvitved-
NOMENCLATURE Jacobsen T, Verbanck M (2004). Solid in
sewers-20 years of investigation. Water
bj bias weight of neuron Mj Science and Technology 52(3):73-84.
D pipe diameter 4. Azamathulla HMd, Ab. Ghani A, Seow YF
Dgr size of non-dimensional particles (2012). ANFIS-based approach for predicting
CV volumetric concentration sediment transport in clean sewer. Applied
d50 average size of the sediment Soft Computing 12(3):1227-1230.
f threshold function of neuron Mj 5. Baghalian S, Bonakdari H, Nazari F, Fazli M
Frm densimetric Froude number (2012). Closed-form solution for flow field in
G gravity acceleration curved channels in comparison with
pi the output of neuron Mi experimental and numerical analyses and
R hydraulic radius artificial neural network. Engineering
S pipe slope Applications of Computational Fluid
SS relative density Mechanics 6(4):514-526.
V limited velocity 6. Bardossy A, Disse M (1993). Fuzzy rule-
wj,I the weight of connection between two based models for infiltration. Water
layers Resources Research 29(2):373-382.
We effective width 7. Beals R, Jackson T (1998). Neural
∆ relative mass of sediment in water Computing: An Introduction. Department of
λs total friction factor Computer Science. University of York, IOP
λs clean water friction factor in the canal Publishing Ltd.
ψ flow parameter 8. Bonakdari H, Larrarte F (2006). Experimental
φ transport parameter and numerical investigation on self cleansing
and shear in sewers. 2nd International IWA

390
Engineering Applications of Computational Fluid Mechanics Vol. 7, No. 3 (2013)

Conference, Sewer Operation and 20. Kisi O (2008). River flow forecasting and
Maintenance, SOM 06, Vienna, Austria, 19- estimation using different artificial neural
26. network techniques. Hydrology Research
9. Bonakdari H, Baghalian S, Nazari F, Fazli M 39(1):27-40.
(2011). Numerical analysis and prediction of 21. Lin JY, Cheng CT, Chau KW (2006). Using
the velocity field in curved open channel support vector machines for long-term
using artificial neural network and genetic discharge prediction. Hydrological Sciences
algorithm. Engineering Application of Journal 51(4):599-612.
Computational Fluid Mechanics 5(3):384- 22. May RWP (1982). Sediment Transport in
396. Sewers. Hydraulic Research Station,
10. Butler B, May R, Ackers J (2003) Self- Wallingford, England, Report IT 222.
cleansing sewer design based on sediment 23. May RWP, Ackers JC, Butler D, John S
transport principles. Journal of Hydraulic (1996). Development of design methodology
Engineering ASCE 129(4):276-282. for self-cleansing sewers. Water Science and
11. Chang CL, Liao CS (2012). Parameter Technology 33(9):195-205.
sensitivity analysis of artificial neural 24. May RWP, Brown PM, Hare GR, Jones KD
network for predicting water turbidity. World (1989). Self-Cleansing Condition for Sewers
Academy of Science, Engineering and Carrying Sediment. Hydraulic Research Ltd
Technology 70:657-660. (Wallingford), Report SR 221.
12. Cigizoglu HK (2002). Suspended sediment 25. Mayerle R, Nalluri C, Novak P (1991).
estimation for rivers using artificial neural Sediment transport in rigid bed conveyances.
networks and sediment rating curves. Turkish Journal of Hydraulic Research 29(4):475-
Journal of Engineering Environmental 496.
Sciences 26(1):27-36. 26. Memarian H, Balasundram HS (2012).
13. Cigizoglu HK (2002). Suspended sediment Comparison between multi-layer perceptron
estimation and forecasting using artificial and radial basis function networks for
neural networks. Turkish Journal Engineering sediment load estimation in a tropical
Environmental Sciences 26(1):15–25. watershed. Journal of Water Resource and
14. Dorofki M, Elshafie AH, Gaafar O, Karim Protection 4:870-876.
OA, Mastura S (2012). Comparison of 27. Moharrampour M, Kherad M, Abachi N,
artificial neural network transfer functions Zoghi M, Asadi Asad Abad MR (2012).
abilities to simulate extreme runoff data. Comparison of artificial neural networks
International Conference on Environment, ANN and statistics in daily flow forecasting.
Energy and Biotechnology, Singapore. Advances in Enviromental Biology 6(2):863-
15. Gopakumar R, Mujumdar PP (2007). A fuzzy 868.
dynamic wave routing model. Hydrological 28. Muzzammil M (2008). Application of Neural
Processes 21:458-467. Networks to scour depth prediction at the
16. Jain SK (2001). Development of integrated bridge abutments. Engineering Application of
sediment rating curves using ANNs. Journal Computational Fluid Mechanics 2(1):30-40.
of Hydraulic Engineering ASCE 127(1):30- 29. Nalluri C, Ab. Ghani A, EI. Zaemey AKS
37. (1994). Sediment transport over deposited
17. Jalalkamali A, Jalalkamali N (2011). beds in sewers. Water Science Technology
Application of hybrid neural modeling and 29(2):125-133.
radial basis function neural network to 30. Novak P, Nalluri C (1975). Sediment
estimate leakage rate in water distribution transport in smooth fixed bed channels.
network. Journal of World Applied Science Journal of Hydraulic Division, ASCE
15(3):407-414. 101(9):1139-1154.
18. Kisi O (2004). Multi-layer perceptrons with 31. Raghuwanshi NS, Singh R, Reddy LS (2006).
Levenberg-Marquardt optimization algorithm Runoff and sediment yield modeling using
for suspended sediment concentration artificial neural networks: Upper Siwane
prediction and estimation. Hydrological River India. Journal of Hydraulic
Science Journal 49(6):1025–1040. Engineering ASCE 11(1):71-79.
19. Kisi O (2005). Suspended sediment 32. Rezaeian Zadeh M, Amin S, Khalili D, Singh
estimation using neuro-fuzzy and neural VP (2010). Daily outflow prediction by multi
network approaches? Hydrological Science layer perception with logistic sigmoid and
Journal 50(4):683–696. tangent sigmoid activation functions. Journal

391
Engineering Applications of Computational Fluid Mechanics Vol. 7, No. 3 (2013)

of Water Resources Mangement 24(11):2673-


2688.
33. Sen Z, Altunkaynak A (2006). A comparative
fuzzy logic approach to runoff coefficient and
runoff estimation. Hydrological Processes
20:1993-2009.
34. Takagi T, Sugeno M (1985). Fuzzy
identification of systems and its applications
to modeling and control. Proceedings of the
IEEE Transactions on Systems, Man and
Cybernetics 15:116-132.
35. Tayfur G, Guldal V (2006). Artificial neural
networks for estimating daily total suspended
sediment in natural streams. Nordic
Hydrology 37(1):69-79.
36. Vongvisessomjai N, Tingsanchali T, Babel
MS (2010). Non-deposition design criteria for
sewers with part-full flow. Urban Water
Journal 7(1):61-77.
37. Wang WC, Chau KW, Cheng CT, Qiu L
(2009). A comparison of performance of
several artificial intelligence methods for
forecasting monthly discharge time series.
Journal of Hydrology 374(3-4):294-306.
38. Wu CL, Chau KW, Fan C (2010). Prediction
of rainfall time series using modular artificial
neural networks coupled with data-
preprocessing techniques. Journal of
Hydrology 389(1-2):146-167.
39. Yang CT, Reza M, Aalami MT (2009).
Evaluation of total load sediment transport
using AAN. International Journal of
Sediment Research 24(3):274-286.

392

You might also like