Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Flow Measurement and Instrumentation 85 (2022) 102161

Contents lists available at ScienceDirect

Flow Measurement and Instrumentation


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

Determination of discharge coefficient of stepped morning glory spillway


using a hybrid data-driven method
Masoud Haghbin a, Ahmad Sharafati b, *, Roozbeh Aghamajidi c,
Seyed Babak Haji Seyed Asadollah b, Mohamadreza Hosseinian Moghadam Noghani b,
María L. Jalón a
a
Department of Structural Mechanics & Hydraulics Engineering, University of Granada, 18001, Spain
b
Department of Civil Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran
c
Department of Civil Engineering, Sepidan Branch, Islamic Azad University, Sepidan, Iran

A R T I C L E I N F O A B S T R A C T

Keywords: The accurate assessment of the discharge coefficient (Cd) of spillways is one of the complex problems in the
Discharge coefficient safety of dams and reservoirs. In this research, the capabilities of Support Vector Regression-Invasive Weed
Support vector regression Optimization (SVR-IWO), Support Vector Regression (SVR), Gaussian Process Regression (GPR), and Kernel
Invasive weed optimization
Ridge Regression (KRR) for estimating discharge coefficients of 21 different layouts of morning glory spillways
Information theory
Bayesian entropy
with the different number of vortex breakers are investigated. In addition, the aleatory and epistemic un­
Morning glory spillways certainties are quantified using classical Mutual information theory and four various Bayesian entropies. The
Froud number (Fr), number of vortex breakers (N), and three dimensionless parameters consisting of the water
height, vortex height, and breaker thickness over the spillway diameter (H/Ds, h/Ds, and t/Ds, respectively),
obtained from 120 experiments, were used as input variables. The findings reveal that SVR-IWO is superior to
other models based on several performance metrics, including R, RMSE, MSE, and MAE. Besides, the SVR-IWO
(R = 0.804, RMSE = 0.131, MSE = 0.017, MAE = 0.074) enhanced the performance indices obtained from
standalone SVR (R = 0.632, RMSE = 0.352, MSE = 0.124, MAE = 0.255) up to 86.13%.

1. Introduction complexities associated with physical and hydraulic conditions that


make the computational assessment of the discharge problem unfeasible
The accurate assessment of the actual discharge capacity of spillways [6].
is a challenging problem. Available evidence reveals that insufficient Experimental studies have been widely employed to solve complex
capacity in spillways and weirs is one of the major reasons for dam problems in water and wastewater engineering [7–12]. Over the last
failures. This risk has been exacerbated in recent years when climate seven decades, extensive experimental studies have been carried out to
change may increase the peak and volume of floods. Therefore, it is determine the discharge coefficient accurately. To reach this aim, many
essential to estimate the actual discharge capacity of spillways accu­ laboratory experiments with different flow conditions and porotype
rately [1]. The morning glory spillway is a suitable t for passing floods in models have been conducted. These investigations focused on obtaining
narrow valleys or steeply slope sites. This spillway includes a circular discharge coefficients using statistical models [7–10,13,14]. However,
inlet, vertical circular shaft, and a pressurized horizontal (or almost these models are restricted when the number of involved variables are
horizontal) tunnel on a mild slope that passes the water from the dam increased. On the other hand, conducting laboratory experiments are
reservoir to the downstream side [2]. However, a major weakness of expensive and time–consuming to build [15].
morning glory spillways is eddy currents, which negatively impact Therefore, there is an immediate need to use an alternative approach
discharge coefficient and spillway performance [3]. The most viable to tackle the limitations mentioned above [16–19]. Several studies
solution to overcome this drawback is installing vortex breakers at the focused on the employment of numerical approaches on hydraulic
inlet of the spillway [4,5]. This introduces several uncertainties and modelings, such as flip bucket cavitation [20] and spillway channel

* Corresponding author.
E-mail addresses: asharafati@gmail.com, asharafati@srbiau.ac.ir (A. Sharafati).

https://doi.org/10.1016/j.flowmeasinst.2022.102161
Received 16 June 2021; Received in revised form 21 January 2022; Accepted 2 March 2022
Available online 7 March 2022
0955-5986/© 2022 Elsevier Ltd. All rights reserved.
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

approach flow pattern [21]. Although numerical methods seem effi­ SVR combined with a popular evolutionary algorithm named Invasive
cient, the data-driven techniques have shown efficiency for solving Weed Optimization (IWO) in dealing with this problem are investigated.
complex problems in water engineering such as scouring, rainfall, water In complex problems, the standalone SVR restricts mapping the relation
quality, sedimentation, wastewater engineering, etc. [17,22,23]. between inputs and output; in this regard, it is essential to enhance its
In the following, we review the previous studies related to applying performance with evolutionary algorithms such as IWO. It was stated in
data-driven models(DDM) for assessing discharge coefficients in the previous paragraphs that accurate determination of discharge co­
different spillways and weirs. Work on this topic was influenced by efficient is a challenging and essential step in designing spillways due to
Bilhan et al. [24], Emiroglu et al. [25] and Salazar et al. [26], who aleatory uncertainties, which are related to hydraulic conditions and
compared different types of artificial neural networks (ANN) with geometry properties of spillways. To tackle this problem, the amount of
multiple linear regression (MLR) and Linear regression (LR) for aleatory uncertainty in the estimation is quantified using Mutual In­
measuring discharge coefficient in rectangular and triangular side weirs formation Theory (MIF). In addition, the epistemic uncertainty, which is
and radial gated spillway. They reported that DDM provided more ac­ related to the physical modeling of the discharge coefficient, is quanti­
curate results for the discharge coefficient as compared to traditional fied using different types of classical and Bayesian entropies.
regression models.
Several studies revealed that ANN-based models such as Multi-layer 2. Material and methods
perceptron (MLP), Radial Basis Function (RBF) and Group Method of
Data Handling (GMDH), Recurrent Neural Networks (RNN) and their In this section, the laboratory experiments and methods for assessing
hybrid versions which combined with evolutionary techniques such as Cd and quantifying their related uncertainties used in this study are
genetic algorithm(GA) or particle swarm optimization(PSO) generated discussed in detail.
high-quality results for estimating discharge coefficient in different hy­
draulic structures. For instance, rectangular and triangular side weir,
circular and rectangular orifices, piano key weirs, labyrinth fuse gate 2.1. The experiment set up and procedure
spillway, and sluices gate spillways [27–34].
However, other researchers have attempted to approach this prob­ This study used a dataset from a Ph.D. dissertation [54] to assess Cd
lem using different types of DDM such as Genetic expression program­ in stepped morning glory spillways. In this way, a 1:50 reduced-scale
ming (GEP), Adaptive Neural Fuzzy Inference System (ANFIS), and physical model set up similar to the morning glory spillway in San
decision tree-based. The reported results confirmed that DDM based Luis forebay dam California, USA [55] was set up for running experi­
models are robust alternatives compared to statistical models [19, ments at the Institute of Hydraulic of Sedimentation, Khuzestan Water
35–41]. and Power Authority, Iran.
Another DDM,that is considered very popular, is the Support Vector The physical model has 3.70 m long, 1.05 m wide, and 3.25 m deep.
Machine (SVM), which proves to be highly efficient in research related In the inside of the physical model, there is a rectangular cubic with
to hydraulic engineering [42]. For example, Parsaie and Haqiabi [43] 1.200 m long, 1.05 m wide, and 0.910 m deep acting as a reservoir of
applied the SVM to this specific topic for the first time. They compared dam, which provides uniform flow using a trapezoidal channel that
SVM and MLP for computing the discharge coefficient of rectangular passes water from dam reservoir to inlet of the spillways. The spillway
side weir. has 1.46 m long, the diameter of the crest is 0.35 m, the thickness of the
Similar investigations were also conducted by Refs. [44–48] using throat is 0.07 m, the bent diameter is identified as 0.1016 m, and the
SVM in different hydraulic structures such as the rectangular, trape­ diameter of the downstream tunnel is 0.0762 m. In addition, a reservoir
zoidal, labyrinth, and oblique side weirs or ogee spillways. In recent with a 2000-L capacity is located downstream for storing water during
years, SVM has been developed by combinations with evolutionary al­ running physical models.
gorithms such as Firefly Algorithm(FA), Genetic Algorithm (GA), and In this physical model, water is passed by a 3-inch pump from the
Particle Swarm Optimization(PSO) for estimating discharge coefficient downstream reservoir to the trapezoidal channel in the dam reservoir,
of side weirs and stepped spillways [49–51]. They used different data­ and it is allowed to pour down through the crest of the spillway after a
sets from the literature containing other physical and hydraulic pa­ short while that the upstream reservoir is filled. After passing through
rameters such as weir heights, weir apex angles, weir lengths, the the spillway, the flow enters the downstream reservoir, the bent, and the
effective height of spillway, Fr, upstream side weir flow depth, Reynolds dam’s tunnel. This process is continuously repeated in the different
Number, etc. A closer look at major investigations in this topic reveals discharge rates generated by the pump. The Images of laboratory setups
that most existing studies focus on different types of side weirs or in different conditions are shown in Fig. 1.
spillways. Very few studies applied DDM to measure the discharge co­ In addition, different vortex breakers (0, 3, 4, and 6) were utilized in
efficient of morning glory spillways. The robustness of DDM for assess­ this survey to understand better the impacts of vortex breakers on
ing the discharge coefficient of morning glory was investigated by discharge coefficient or rate in smoothly stepped morning glory spill­
Ref. [52]. He employed standalone ANN for predicting discharge coef­ ways. Overall, 120 runs with 21 different layouts were considered. The
ficient in smooth spillways with 3,6 and 12 vortex breakers. He reported dataset used in the current study is presented in the supplementary
that ANN has a good capability in this area. In line with this study, material. Besides, the details of layouts and range of flow parameters in
Alfatlawi and Alshaikhli [53] used standalone ANN and Multiple our investigation are summarized in Table 1.
Non-Linear regression (MNLR) in the their investigation. They entered The height of water over spillways varies from (0.012m–0.22 m) in
Fr, number of stairs, head over crest shape and morning glory spillway the smooth morning glory spillways to (0.035m–0.162 m) in stepped
inlet radius as their input variables. They found that ANN provided more spillways. The details about the impacts of different parameters on
accurate results in comparison with MNLR. Except these two recent discharge coefficient are discussed in further sections.
studies, there is no sufficient evidence that research groups focused on
the capabilities and benefits of DDM for measuring discharge coefficient 2.1.1. Dimensional analysis
of morning glory spillways. In this regards, the benefits of DDM for The parameters affecting the discharge coefficient in the morning
morning glory spillway are not considered as fully explored research glory spillway can be presented as a functional expression as follows [4]:
area. f (ρ, μ, σ , H, V, g, N, Ds , h, t, Cd) = 0 (1)
This study attempts to shed light on the capabilities of different DDM
for estimating the discharge coefficient of morning glory spillways based where ρ, μ, σ , H, V, g, N, Ds , h, t, and Cd denote density, surface tension,
on experimental data. On the other hand, the capabilities of a hybrid dynamic viscosity, the height of water over spillways crest, average

2
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

Fig. 1. Images of laboratory experiments: a) View of the location of spillways’ reservoir, b) View of Smooth spillway without vortex breakers, c) View of Passing flow
and generating eddy current after 20 s in the smooth spillway, d) View of the smooth spillway with three vortex breakers, e) View of the smooth spillway with four
vortex breakers, f) View of stepped spillway without vortex breakers, g) View of the stepped spillway with six vortex breakers in the condition of passing the
low discharge.

Table 1
The details of layouts and range of flow parameters.
( )
No. layout N of stepped Chamber N of V.B t of V. B (m) h of V.B (m) Ds (m) Length of Spillway (m) lit H of water over spillways Froude
Q
S (m) Number

1 0 0 0 0 0.35 1.10 1.23–3.19 0.069–0.130 1.41–1.51


2 0 3 0.009 0.055 0.35 1.10 1.96–3.15 0.092–0.128 1.3–1.59
3 0 4 0.009 0.055 0.35 1.10 1.84–3.45 0.089–0.133 1.374–1.41
4 0 6 0.009 0.055 0.35 1.10 1.98–3.12 0.110–0.130 1.38–1.50
5 0 4 0.009 0.035 0.35 1.10 1.98–3.01 0.092–.125 1.40–1.51
6 0 6 0.009 0.035 0.35 1.10 1.92–3.18 0.02–0.135 1.31–1.71
7 0 4 0.02 0.035 0.35 1.10 1.93–3.26 0.013–0.147 1.44–1.91
8 0 3 0.02 0.035 0.35 1.10 2.16–3.13 0.027–0.134 1.53–1.91
9 0 6 0.02 0.035 0.35 1.10 2.19–3.56 0.02–0.222 1.34–2.11
10 0 4 0.02 0.050 0.35 1.10 1.92–3.19 0.017–0.065 1.70–3.78
11 0 3 0.02 0.050 0.35 1.10 1.96–3.15 0.014–0.062 1.56–3.81
12 0 6 0.02 0.050 0.35 1.10 1.98–3.08 0.012–0.037 1.35–3.36
13 6 0 0 0 0.35 1.10 1.83–3.24 0.015–0.102 1.34–1.67
14 6 3 0.009 0.050 0.35 1.10 1.73–3.64 0.012–0.162 1.42–3.47
15 6 4 0.009 0.050 0.35 1.10 1.96–2.96 0.05–0.10 1.42–1.84
16 6 6 0.009 0.050 0.35 1.10 1.76–3.37 0.012–0.135 1.37–2.37
17 6 3 0.009 0.035 0.35 1.10 2.15–3.41 0.016–0.146 1.37–3.14
18 6 4 0.009 0.035 0.35 1.10 1.84–2.83 0.0135–0.066 1.56–3.11
19 6 6 0.009 0.035 0.35 1.10 1.686–3.11 0.015–0.104 1.23–2.40
20 6 3 0.02 0.055 0.35 1.10 1.612–3.54 0.014–0.095 1.22–3.11
21 6 4 0.02 0.055 0.35 1.10 2.151–3.12 0.018–0.14 1.34–2.12

velocity, gravitational acceleration, number of vortex breakers, the ( )


H h t
diameter of the spillway, height of vortex breakers, thickness of vortex Cd = f Fr, Re , We , N, , , (2)
breakers and discharge coefficient, respectively. DS DS DS
Employing the Buckingham π theory of the dimensional analysis, the ( )
discharge coefficient Cd can be obtained as follows [4]: where Fr is the Froude number √V̅̅̅̅ , Re is the Reynolds Number
gH

3
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

( ) ( )
ρVD ρV 2 D / )
, We is the Weber Number , N is the number of vortex 1
(− x2
(8)
μ σ
G(x.σ )= √̅̅̅̅̅̅̅ e σ2
H
breakers, DS is the ratio of the height of water over the diameter of the
σ 2π
h
spillway, Ds is the ratio of the height of vortex breaker over the diameter
where x and σ denote input variable and standard deviation, respec­
of spillway and Dst is the ratio of the thickness of the vortex breaker over
tively. The first term σ√1̅̅̅̅ is associated with a normalization constant.
the diameter of the spillway. 2π
The Gaussian Kernel is also named normalized kernel because its inte­
Available evidence [56–58] suggested that Re and We can be elim­
gral is 1 for each value of σ [63].
inated in some ranges specified in Table 2 [4].
In this research, values of Re (varied between 5.33E +05
2.2.2. Gaussian process regression
and 1.69E +06) and We ( > 130) satisfy the suggested ranges in Table 2
A Gaussian Process Regression (GPR) combines different random
for eliminating them. Hence, the final non-dimensional expression for
variables, and a Gaussian distribution can be fitted [64]. A Gaussian
discharge coefficient is obtained as follows:
( ) process (GP) or g(x) is introduced by two principal functions known:
H h t mean h(x) and kernel k(x.x ) functions which are represented in Eqs. (9)

Cd = f Fr, N, , , (3)
DS DS DS and (10), respectively. The later one which is also known as covariance
function is assessed based on occurrence of x and x [65].

Although other parameters such as circular inlet radius and angle of
anti-vortex with the flow direction can be considered in calculating Cd h(x) = E(g(x)) (9)
[61], this study has been postponed them for future research.
(10)
′ ′ ′
k(g(x).g(x) .θ) = (E(g(x) − h(x))(g(x ) − h(x )))
2.2. Description of data-driven techniques
where the θ denotes the hyper-parameter that needs to be extracted
2.2.1. Kernel Ridge Regression using the tuning procedure. The g(x) function is subsequently repre­
The Kernel ridge regression (KRR) is originated from a combination sented in Eq. (11) as follows:
of kernel method and ridge regressions [59]. The main advantage of
(11)

KRR is the employment of specific criteria and kernel approaches for g(x) ∼ GPM(h(x).k(x.x ))
extracting a non-linear relationship to overcome the over-fitting concern
in regression-based problems [60]. Eq. (4) represents the mathematical where the GPM indicates the Gaussian Process Model. This suggests that
formation of KRR: g(x) is considered as a stochastic function delivered as a Gaussian pro­
cess with mean and covariance functions of h(x) and k(x.x ) respectively.

1∑ n ⃦ ⃦ In term of regression procedure, any target parameter (y) can be


argmin ‖hi − yi ‖2 + δ⃦h2 ⃦H (4)
2 i=1 assigned to a corresponding regression function g(x) with an associated
Gaussian noise (γ) stated in Eq (12).

n
( )
hi = υj Φ xj .xi (5) y = g(x) + γ (12)
j=1
The γ considered to has a mean value of zero and variance of σ2n
where the ‖⋯‖H denotes the Hilbert normed space, the υ is the depen­ (γ ∼ N(0.σ 2n )).So, the GP noted in Eq. (11) can be reformed as Eq (13)
dent weights of data, and the Φ function is noted the kernel function [66]:
corresponding to x as the predictors. The h and y are respectively the )
(13)

response parameter and regression function. The δ is noted the regula­ g(x) ∼ GPM(h(x).k(x.x ) + σ2n I
rization parameter, which is greater than zero and usually considered a
fixed value. For a kernel matrix (Κ) with the size of (n × n), Eq. (4) can where I note the identity matrix. Due to the diminishing procedure taken
be reformed as: by GP as well as the noise (γ) amplifying nature, the joint distribution is
given in Eq (14) [66]:
(Κ + δnI) = y (6) [ ] ([ ][ ( ) ])
y g(X) k X.X + σ2n k(X.Xte )
In the training phase, KRR evaluates the υ value by attaining a so­ ∼N . (14)
y
̂ g(Xte ) k(Xte .X) k(Xte .Xte )
lution for Eq. (6). The algorithm tries to find optimum values for υ and δ
based on variables set for most optimal performance [61]. The extracted
where the y and ̂y are denoting the training and testing targets. Also, the
υ is subsequently applied to the testing phase as follows:
X and Xte are indictors of corresponding positions in training and testing

n
phases, respectively. Eq. (5) states the predictive distribution:
y=
̂ ̂
υi Φ(xi . X) (7)
j=1 p(̂
y |X.y.Xte ) ∼ N(̂
y mean . var(̂
y )) (15)
In this study, the Gaussian kernel function, as a popular kernel, is
where the ̂ y mean and var(̂ y ) are respectively denoting the predictive
utilized in the KRR to obtain appropriate prediction accuracy [62]. The
mean and variance as follows:
mathematical expression of Gaussian Kernel is determined as follows:
[ ( )]− 1
y mean = g(Xte ) + k(Xte .X) k X.X + σ2n I
̂ (y − g(X)) (16)

[ ( )]− 1
y ) = k(Xte .Xte ) k X.X + σ2n I
var(̂ k(X.Xte ) (17)
Table 2
Suggested range for eliminating the impacts of dynamic viscosity and surface The type of Covariance function (kernel) is a Gaussian kernel in this
tension. work. GPR employs this kernel to determine covariance of the prior
distribution for identifying the likelihood function. Also, GPR can select
Scholar Re We
hyper-parameters based on the gradient ascent method, unlike KRR,
(Anwar, 1966) Re > 2 × 104 We > 100 which uses grid search on means square error. GPR can provide a
(Daggett and Keulegan, 1974) Re > 3 × 104 We > 120
probabilistic model of the target function then determine meaningful
(Jain et al., 1978) Re > 5 × 104 We > 120

4
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

confidence intervals associated with the predicted target while the KRR where σ iter , itermax , σ inital and σ final are the standard deviation for the
only generates predicted values [67]. current iteration, maximum number of iterations, and first and final
measured standard deviations, respectively.
2.2.3. Support vector regression (SVR) The final stage is defined as competitive exclusion. Then, a new
The SVR is a principal branch of the support vector machine, which, population is generated for the next iteration by choosing the superior
unlike the classification branch, has only one set of typifying points. individual around each feasible solution member, evaluated based on
While the SVM aims to maximize the divided space between two or more the fitness function values.
sets of representative points, SVR tends toward the complete opposite
procedure by minimizing the separation between samples and hyper­ 2.2.5. Coupling SVR-IWO
plane [68]. There is a strong demand for a reliable and cost-effective approach
In case of problems with non-linear nature, SVR employs specific for automatically tuning SVR’s user-defined parameters. In this regard,
types of kernel function to elevate the regression problems to spaces we examine the capabilities of IWO to tune constant parameter (C),
with higher dimensions. This approach will lead the problem to a more determined margins of error ε and parameters of the kernel function.
understandable space so that a suitable hyperplane can be extracted, Fig. 2 illustrates a flow chart of the tuning of the SVR parameters uti­
which optimally divides the points of a sample [69]. lizing IWO.
[
1∑ t ∑ t
( )( ) ( )
MAX − si − s*i sj − s*j Φ Di .Dj 2.3. Performance indices
2 i=1 i=j
⎧ ⎫

⎪ ∑ t
( ) ⎪
⎪ The performance of the DDMs employed in this research were
⎪ *
] ⎪
⎪ si − si = 0 ⎪

⎪ measured using four performance indices, Root Mean Square Error
∑(t ∑(
t ⎪
⎨ ⎪

) ) (RMSE) [72,73], Mean Absolute Error (MAE) [73,74], Coefficient of
i=j
− *
si + si ε + *
si − si Yi s.t C (18)
i=1 i=1 ⎪ 0 ≤ si .s*i ≤




⎪ Correlation (R) [75,76], and Mean Square Error(MSE) [77,78]:

⎪ ζ ⎪


⎩ ⎪
⎭ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
i = 1.2. ⋯.ζ ∑n 2
i=1 (XObserved − XEstimated )
RMSE = (23)
In equation (18), the Di and ζ are noting the sample data and its size, n
C denotes the coefficient of penalty, ε is exceedance allowability of error ∑n ⃒⃒ ⃒

related to sample size. Φ(Di .Dj ) considered as the kernel function, while MAE = i=1 XEstimated − XObserved
(24)
the s (si . s*i . ⋯sj .s*j ) represent the optimal solutions. The mathematical n
equation of SVR can be represented as follow [70]: ∑ ∑
n( XObserved XEstimated ) − ( XObserved )( XEstimated )

R = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
[ ∑( 2 ) ∑ ][ ∑( 2 ) ∑ ]̅

t
( ) ( ) n X Observed − ( XObserved )2 n X Estimated − ( XEstimated )2
F(x) = si − s*i Φ Di .Dj + b* (19)
i=j (25)
∑n
as long as the (si − s*i ) part of Equation (18) does not become equal to i=1 (XObserved − XEstimated )2
MSE = (26)
zero, the Di as the sample will be considered a support vector in the non- n
linear problem.
In the above equations, XObserved and XEstimated are the observed and
estimated values, respectively, and n is the number of observed/esti­
2.2.4. Invasive Weed Optimization (IWO)
mated values.
Mehrabian and Luca [71] introduced a new evolutionary optimiza­
tion model that originated from the behavior of weed for making col­
2.4. Uncertainty analysis
onies and their capabilities for finding an appropriate place for growth
and regeneration. IWO includes four main stages; the first stage employs
The classical Mutual Information (MIF) theory, based on Entropy
a randomization function to generate the initial population.
theory, measures the aleatory uncertainties in the discharge coefficient
X = [X1 X2 , …, Xm ] (20) due to the chaos inflow current in spillways. The MIF is a widely-used
toolkit for quantifying the flow of information between involved vari­
The second stage is identified as regeneration or reproduction so that
ables in phenomena originating from one popular branch of information
each member of the previous step creates a population of seeds, and it is
theory named entropy [79]. The MIF between two different variables, X
directly associated with the value of fitness function, which is computed
and Y, is presented as bellows:
as follows:
( ) I(X, Y) = H(X) − H(X|Y) (27)
f − fmin
Weed = floor (smax − smin ) + smin (21)
fmax − fmin In the above relation, I(X, Y), H(X) and H(X|Y) denote the degree of
dependencies between variable X and Y, Shannon entropy of variable X,
In the above relation, floor, f, fmin , fmax , smin and smax denote round and the conditional entropy for X given Y, respectively. The Shannon
down function, the amount of fitness function associated with seeds, the entropy H(X) determines the current information in variable X and its
obtained minimum and maximum amount of fitness, and minimum and inherent uncertainties [80]. Its mathematical expression is presented as
the maximum number of seeds, respectively [71]. follows:
The third stage is spatial dispersal, where the newly produced seeds
layout over solution spaces. The coefficient named the level of difference ∑
N
H(X) = − P(Xi )log P(Xi ) (28)
is determined to measure the distance between the place of the parent
i=1
plant and the point where their seeds are falling on the ground. This
coefficient includes a standard deviation operator. The expression of this In the above relation, P(Xi ) denote the probability values related to
coefficient is presented as follows: the values Xi . Another term H(X|Y), is obtained from the following
relation:
(itermax − iter)n ( )
σ iter = σ inital − σfinal + σ final (22)
(itermax )n

5
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

Fig. 2. The flowchart of SVR-IWO.

N ∑
∑ M
( ) ( ⃒ ) In the above relation α, K, πi , N and nK present the Dirichlet con­
H(X|Y) = − P Xi , Yj logP Xi ⃒Yj (29) centration coefficient, numbers of the determined bin in the fitted dis­
tribution over variables, the measured probability that one of the dataset
i=1 j=1

⃒ samples X is located in the ith bin, number of total samples, and number
where P(Xi ⃒Yj ), N and M are the conditional probability of X fitted on Y,
of samples stores in the ith bin. The α equals zero for ML. The values of
number of variables X and Y, respectively.
this parameter for Bayesian models are considered as 0.5, 1,
The MIF theory determines the most impactful input variable, which √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
Summation of generated results
has the maximum degree of dependency with the discharge coefficient 1
Size of vector of generated results
, and Size of vector of generated results
for Jeffrey,
in this research. After measuring the degree of dependence for each Laplace, SG, and Minimax, respectively.
input variable, the input variables are sorted based on their MIF from
lowest to highest values; then, they are eliminated one by one to
3. Results and discussion
generate several input combinations for assessing the discharge coeffi­
cient of morning glory spillways.
The discharge coefficient of morning glory spillways are estimated
Different approaches are used in this study to measure the epistemic ( )
H h t
uncertainty associated with DDM results. The Maximum likelihood en­ based on the several non-dimensional parameters Fr, N, Ds , Ds, Ds and
tropy, the classic type of entropy, and four Bayesian entropies are
65% of the gathered dataset is selected for the training stage, and re­
employed to find a model with similar information compared to
mains are used for the testing stage. The ranges of input parameters in
observed data. These models are Jeffrey, Laplace, Schurmann-
both stages are presented in Table 3.
Grassberger (SG), and Minimax. The P(Xi ) is limited with biases in
The mutual information theory is used to find the most appropriate
small datasets. Researchers suggest using the Bayesian interface to solve
input combinations. Table 4 shows the degree of dependency or flow of
this problem, applying Dirichlet prior and posterior probabilities to fit ( )
H h t
the data [81,82]. information between the input parameters Fr, N, Ds , Ds, Ds and the
The Dirichlet prior and posterior probabilities are presented as the H
target parameter (Cd). As the table reveals, the shows the highest
following mathematic forms: Ds
t
relevancy with discharge coefficient while the Ds represents the least
Γ(Ka) ∏
K
consistency among the assessed inputs. A schematic of mutual infor­
Dir(α)≜Dir(α1 , α2 , …, aK ) = π αi − 1

mation between input parameters and Cd is illustrated in Fig. 3.


K
Γ(α) i=1
(30) Based on the relevancy analysis in Table 4, several input combina­
∏K
πni i+α− 1
Dir(α)≜Dir(α1 + n1 , …, α + nK ) = Γ(Ka + N) = tions are listed in Table 5. The combinations were structured based on
Γ(a + ni )
i=1 the elimination of the least dependent variable from the C1 to C5
combination, so the first combination (C1) includes all five inputs while

6
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

Table 3 These combinations were subsequently employed as inputs in soft


Ranges of input parameters for train and test stage. computing models. The predicted results considering different algo­
Parameters Training stage Testing stage rithms (SVR-IWO, SVR, GPR, and KRR) for both training and testing
phases are shown in Tables 6 and 7, respectively. Note that these tables
Fr 1.29–3.81 1.22–3.51
N 0.00–6 0.00–6 present the accuracy enhancement of the IWO optimization algorithm
H 0.034–0.634 0.034–0.428 compared to the stand-alone SVR model with nearly 27.21% improve­
DS ment in prediction.
0.00–0.157 0.00–0.157
Regarding the most optimized inputs combination in each algorithm,
h
DS
t 0.00–0.057 0.00–0.057 the C1 ranked as the best inputs structure in SVR-IWO (Rtraining = 0.945.
DS MAEtraining = 0.074.Rtesting = 0.804.MAEtesting = 0.074), SVR (Rtraining =
0.913.MAEtraining = 0.265. Rtesting = 0.632.MAEtesting = 0.255) and KRR
(Rtraining = 0.793, MAEtraining = 0.195, Rtesting = 0.703, MAEtesting =
Table 4 0.233). However, the C3 is considered the most efficient combination of
Calculation of dependency degree between input
variables and discharge coefficient (Cd).
Input variables Cd
Table 5
Input combinations constructed based on relevancy to discharge coefficient.
Fr 0.079
N 0.0521 Input combinations Fr N t/Ds h/Ds H/Ds
t/Ds 0.0304 C1 ✓ ✓ ✓ ✓ ✓
H/Ds 0.165 C2 ✓ ✓ – ✓ ✓
h/Ds 0.066 C3 ✓ – – ✓ ✓
C4 ✓ – – – ✓
C5 – – – – ✓
H
the last one (C5) only includes Ds as the most dependent variable.

Fig. 3. The flow of information between input parameters and Cd.

7
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

Table 6
Performance evaluators for all predictive models in training phase.
Predictive model R RMSE MSE MAE

SVR-IWO-C1 0.945 0.131 0.0172 0.074


SVR-IWO-C2 0.921 0.156 0.0244 0.094
SVR-IWO-C3 0.915 0.161 0.0261 0.100
SVR-IWO-C4 0.896 0.177 0.0315 0.113
SVR-IWO-C5 0.744 0.267 0.0716 0.1912

SVR-C1 0.913 0.3282 0.1077 0.2651


SVR-C2 0.9128 0.3222 0.1038 0.2602
SVR-C3 0.903 0.3105 0.0964 0.2479
SVR-C4 0.862 0.2812 0.0790 0.2222
SVR-C5 0.779 0.2628 0.0691 0.1799

GPR-C1 0.646 0.307 0.0943 0.260


GPR-C2 0.734 0.277 0.076 0.225
GPR-C3 0.932 0.168 0.028 0.132
GPR-C4 0.902 0.173 0.301 0.120
GPR-C5 0.741 0.270 0.730 0.196

KRR-C1 0.793 0.246 0.060 0.195


KRR-C2 0.748 0.267 0.0713 0.212
KRR-C3 0.730 0.275 0.0756 0.226
KRR-C4 0.705 0.284 0.0810 0.232
KRR-C5 0.521 0.342 0.117 0.293

Table 7
Performance evaluators for all predictive models in testing phase.
Predictive model R RMSE MSE MAE

SVR-IWO-C1 0.804 0.131 0.0172 0.074


SVR-IWO-C2 0.743 0.293 0.086 0.206
SVR-IWO-C3 0.798 0.2636 0.0695 0.203
SVR-IWO-C4 0.802 0.2613 0.0683 0.1921
SVR-IWO-C5 0.612 0.3463 0.120 0.246

SVR-C1 0.632 0.3524 0.1242 0.2547


SVR-C2 0.617 0.3613 0.1305 0.2637
SVR-C3 0.6147 0.3643 0.1327 0.2647
SVR-C4 0.607 0.3645 0.1329 0.2661
SVR-C5 0.458 0.4012 0.1609 0.3017

GPR-C1 0.618 0.347 0.121 0.270


GPR-C2 0.658 0.331 0.110 0.252
GPR-C3 0.737 0.318 0.101 0.263
GPR-C4 0.799 0.262 0.698 0.250
GPR-C5 0.630 0.346 0.120 0.257

KRR-C1 0.703 0.275 0.0972 0.233


KRR-C2 0.668 0.325 0.105 0.244
KRR-C3 0.667 0.326 0.106 0.235
KRR-C4 0.684 0.321 0.103 0.230
KRR-C5 0.457 0.395 0.156 0.300

the GPR predictive algorithm (Rtraining = 0.902.MAEtraining = 0.120.


Rtesting = 0.799.MAEtesting = 0.250).
While these tables almost clarified the best combination for each
predictive model, a more comprehensive evaluation between the algo­
rithms must be carried out to clarify the best predictive model. To
achieve this goal, the Heat-map plot is represented in Fig. 4. The Heat-
map is considered a useful approach for comparing the predictive
combinations, which operates based on simultaneous normalization of
several performance indices. These normalized values, then addressed Fig. 4. Comprehensive comparison between predictive model employed in the
by a color range representing the models’ suitability, illuminate the current study using Heat-map for a) Training and b) Testing.
worst to best.
Fig. 4 shows the training phase performs better than the testing phase than SVR-IWO, the GPR proved to be a highly efficient predictive model
in all investigated AI algorithms. It has been evident that the SVR-IWO- over both studied phases.
C1 obtained the highest normalized value in both phases and proved While heat maps have proven to be an excellent visual comparison
more accurate than other models. Also, with slightly lower performance approach, several other graphical accuracy measurement tools, such as

8
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

the scatter plot and the Taylor diagram, have also been highly efficient
in previous literature.
The scatter plot is considered a visual evaluation method utilized in a
wide range of scientific literature. This plot shows the scattering setup of
each predictive model output compared to observed values shown as a
straight line known as Fit-line. The model that has a closer scattering
pattern to this Fit-line considered a model with better accuracy and also
a higher coefficient of determination (R2 ). As shown in Fig. 5, the SVR-
IWO with the C1 combination has more alignment with observed data in
both training and testing phases.
The Taylor diagram is highly applicable in engineering studies as
another graphical presentation. It combines the R and RMSE as error
indices with normalized standard deviation and proposes a more
comprehensive comparison between predictive models. Fig. 6 shows the
Taylor diagram extracted from the current study for training and testing

Fig. 6. Predictive model assessment using Taylor diagram over a) Training and
b) Testing stage.

datasets. In this case, the model which takes the nearest position to the
experimental (dark blue dot) is considered the best model. As Fig. 6
shows, the green dot noted the SVR-IWO-C1 is the closest to the
experimental values in both studied phases.

3.1. Validation of SVR-IWO

In this section, the obtained results from SVR-IWO are compared


with the other two almost very similar studies as well as two popular
other hybrid SVR models such as Support Vector Regression-Genetic
Fig. 5. Graphical presentation of data scattering compare to fit-line for a) Algorithm(SVR-GA) and Support Vector Regression- Particle Swarm
Training stage and b) Testing stage. Optimization(SVR-PSO) to validate the performance of proposed

9
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

methodology. In addition, the impact of the different datasets on the Table 8


obtained results is evaluated in the current section. Comparison between results obtained from the current study and previous
Christodoulou et al. [13] conducted similar experimental research on parametric equations.
the Cd for a morning glory spillway inlet and proposed the following Scholar Phase R RMSE MSE MAE
regression equation: Current study Training 0.945 0.131 0.0172 0.074
[ ( )] Alfatlawi and Alshaikhli 0.3808 1.3885 1.928 1.0632
H
Cq = 0.515 1 − 0.2 × (31) [53]
R Christodoulou et al. [13] 0.4543 0.396 0.156 0.337
Current study Testing 0.804 0.131 0.0172 0.074
Also, Alfatlawi and Alshaikhli [53] predicted the discharge coeffi­
Alfatlawi and Alshaikhli 0.500 1.371 1.879 1.092
cient of circular steeped morning glory spillway with models consisting [53]
of 4, 5, and 6 steps. The prediction phase was proceeded using ANN and Christodoulou et al. [13] 0.459 0.434 0.188 0.3324
Multiple Nonlinear Regression (MNLR) techniques, and the parametric
regression equation has been proposed as:
( )0.118 Table 9
H
Cqc = 0.784 × N 0.132 × × Fr1.068 (32) Comparison between results obtained from SVR-IWO and other hybrid SVR
R models.
From Table 8, it can be concluded that the proposed SVR-IWO-C1 Scholar Phase R RMSE MSE MAE
shows significant superiority compared to regression equations (31) SVR-IWO-C1 Training 0.945 0.131 0.0172 0.074
and (32) outputs over both training and testing phases. SVR-GA-C1 0.916 0.161 0.0259 0.128
The Genetic Algorithm (GA) and Particle Swarm Optimization(PSO) SVR-PSO-C1 0.923 0.154 0.0238 0.116
optimization approaches were compared with the IWO used in the SVR-IWO-C1 Testing 0.804 0.131 0.0172 0.074
SVR-GA-C1 0.705 0.310 0.0963 0.222
current study for further investigation. The GA and PSO are widely used
SVR-PSO-C1 0.738 0.295 0.0871 0.220
to optimize the hyper-parameters of standalone ANN, ANFIS SVR, etc.
[51,83]. To this end, the best input combination (C1) was entered into
GA and PSO models. The results revealed that SVR-IWO outperformed uncertainties and chaos associated with the flow pattern and geometry
SVR-GA and generated slightly better results than SVR-PSO. The details of hydraulic structures. In this study, four different types of DDM are
of this comparison are shown in Table 9. named Support Vector Regression-Invasive Weed Optimization (SVR-
The available datasets about Cd of stepped morning glory spillway IWO), Support Vector Regression (SVR), Gaussian Process Regression
are limited in the literature. However, the different samples are pro­ (GPR), and Kernel Ridge Regression(KRR) are employed for assessing
vided from the original dataset to assess the reliability of the superior discharge coefficients of different types of morning glory spillways. To
model (SVR-IWO) in prediction modeling. In this way, the original this end, data set from a 1:50 reduced-scale physical model similar to the
dataset is randomized again. Then different portions of data, i.e., 10%, morning glory spillway of San Luis forebay dam (USA) is used, and 120
15%, and 20%, are eliminated from original data to obtain three new experiments are simulated in different flow conditions. Using Bucking­
(
samples named New Sample 1 (NS1), NS2, NS3, respectively. Afterward, H
ham Π theorem, four non-dimensional parameters Fr, N, Ds ,
those new samples (i.e., NS1, NS2, and NS3) and SVR-IWO are used for
)
the prediction modeling. Fig. 7 shows the reduction in the correlation h t
coefficients obtained from new samples compared with this metric ob­ Ds and Ds are selected as input variables. Whereas these input vari­
tained from the original dataset. From Fig. 7, it is obvious that the ables directly impact aleatory uncertainty of the discharge coefficient,
maximum reduction in correlation coefficient in other datasets over the the Mutual information theory is employed to measure the degree of
testing stage is 10%. Hence, the SVR-IWO is a reliable model for esti­ dependency between these input variables and the discharge coefficient.
mating the discharge coefficient of stepped morning glory spillway with Then, different input combinations are generated based on mutual in­
vortex breakers. formation theory. The results showed that the ratio of the height of
water over the crest to the radius of the spillway with 0.103 bits had the
3.2. Uncertainty analysis highest degree of dependency among other input variables. In addition,
the DDM has been applied, which results showed that SVR-IWO
The epistemic uncertainty is measured for the four different esti­ (Rtraining = 0.945, MAEtrainint = 0.074, Rtesting = 0.804, MAEtesting = 0.195)
mators (standalone SVR, SVR-IWO, GPR, and KRR) in the input com­ had powerful capabilities among other models. Finally, the epistemic
binations that provided the best performance. The Entropy package in R uncertainties were measured using Maximum likelihood, Jeffery, SG,
software is employed to measure different entropies [84]. Minimax, and Laplace entropies. The results of measured undertrained
The obtained results are shown in Fig. 8. The results demonstrate confirmed that SVR-IWO had the closest information content to other
that the entropy values of the observed discharge coefficient have models. Therefore, it can be selected as the most reliable model for
overall closest agreements with generated results by SVR-IWO. (percent predicting the discharge coefficient of morning glory spillways.
difference of ML = 0.983, Jeffery = 0.256, SG = 4.084, Minimax
= 0.723 and Laplace = 0.1156) for stage one(Train) and Authors’ contributions
(ML = 0.584, Jeffery = 0.2736, Laplace = 0.135, SG ­
= 0.608 and Minimax = 0.595) for stage two (Test)). Therefore, it can Masoud Haghbin: Conceptualization; Methodology; Writing –
be concluded that information in observed data sets is similar to SVR- original draft. Ahmad Sharafati: Supervision; Methodology; Validation;
IWO results, which makes it a superior model to others. Writing – review & editing. Roozbeh Aghamajidi: Data curation,
Validation, Writing – original draft. Seyed Babak Haji Seyed Asa­
4. Conclusion dollah: Software; Formal analysis. Mohamadreza Hosseinian Mog­
hadam Noghani: Investigation; Writing – original draft. Maria Lourdes
New approaches such as data-driven models (DDM) are available to Jalon Ramirez: Visualization; Writing – review & editing.
solve different hydraulic engineering problems. The accurate estimation
of discharge coefficient (Cd) is a real challenge due to significant

10
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

Fig. 7. Reduction in correlation coefficient based on the different new samples.

Fig. 8. The epistemic uncertainties associated with DDM a) Training and b) Testing stage.

11
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

Declaration of competing interest [27] S. Bagheri, A.R. Kabiri-Samani, M. Heidarpour, Discharge coefficient of rectangular
sharp-crested side weirs, Part I: traditional weir equation, Flow Meas. Instrum. 35
(2014) 109–115.
The authors declare that they have no known competing financial [28] H. Bashiri, B. Dewals, M. Pirotton, P. Archambeau, S. Erpicum, Towards a New
interests or personal relationships that could have appeared to influence Design Equation for Piano Key Weirs Discharge Capacity, 2016.
the work reported in this paper. [29] I. Ebtehaj, H. Bonakdari, A.H. Zaji, H. Azimi, A. Sharifi, Gene expression
programming to predict the discharge coefficient in rectangular side weirs, Appl.
Soft Comput. 35 (2015) 618–628.
References [30] A. Eghbalzadeh, M. Javan, M. Hayati, A. Amini, Discharge prediction of circular
and rectangular side orifices using artificial neural networks, KSCE J. Civ. Eng. 20
(2016) 990–996.
[1] N. Aein, M. Najarchi, S.M.M. Hezaveh, M.M. NajafiZadeh, E. Zeighami, Application
[31] M.A. Ghorbani, F. Salmasi, M.K. Saggi, A.S. Bhatia, E. Kahya, R. Norouzi, Deep
of 3D numerical model and intelligent systems in discharge coefficient estimation
learning under H2O framework: a novel approach for quantitative analysis of
of combined weir-gate, Water Resour. 47 (2020) 537–549.
discharge coefficient in sluice gates, J. Hydroinf. 22 (2020) 1603–1619.
[2] M.R. Enjilzadeh, E. Nohani, Numerical modeling of flow field in morning glory
[32] M. Nouri, M. Hemmati, Discharge coefficient in the combined weir-gate structure,
spillways and determining rating curve at different flow rates, Civ. Eng. J. 2 (2016)
Flow Meas. Instrum. 75 (2020) 101780.
448–457.
[33] M. Yasi, Z. Gholami, Performance evaluation of discharge coefficient in physical
[3] A. Parsaie, A. Haghiabi, The effect of predicting discharge coefficient by neural
models of labyrinth fusegate spillways with intellectual and statistical models, Iran,
network on increasing the numerical modeling accuracy of flow over side weir,
J. Irrig. Drain. 11 (2017) 798–809.
Water Resour. Manag. 29 (2015) 973–985.
[34] A.H. Zaji, H. Bonakdari, S. Shamshirband, S.N. Qasem, Potential of particle swarm
[4] F. Sayadzadeh, S.H. Musavi-Jahromi, H. Sedghi, A. Khosrojerdi, Pyramidal vortex
optimization based radial basis function network to predict the discharge
breakers influences on the flow discharge of morning glory spillway, Ain Shams
coefficient of a modified triangular side weir, Flow Meas. Instrum. 45 (2015)
Eng. J. 11 (2020) 455–463.
404–407.
[5] A. Parsaie, A. Haghiabi, Hydraulic investigation of finite crested stepped spillways,
[35] O. Kisi, O. Bilhan, M.E. Emiroglu, Anfis to estimate discharge capacity of
Water Supply 21 (5) (2021) 2437–2443. https://doi.org/10.2166/ws.2021.078.
rectangular side weir, in: Proc. Inst. Civ. Eng. Manag., Thomas Telford Ltd, 2013,
[6] M. Jamei, I. Ahmadianfar, X. Chu, Z.M. Yaseen, Estimation of triangular side orifice
pp. 479–487.
discharge coefficient under a free flow condition using data-driven models, Flow
[36] Y. Mehri, N. Abbasi, Estimating the discharge coefficient of the type B piano-key
Meas. Instrum. (2020) 101878.
side weir at a 120 curve using RBF and ANFIS models in comparison with artificial
[7] W.E. Wagner, Morning-glory shaft spillways: a symposium: determination of
neural networks, Irrig. Sci. Eng. 43 (2020) 89–100.
pressure-controlled profiles, Trans. Am. Soc. Civ. Eng. 121 (1956) 345–368.
[37] A.H. Azimi, A. Rajabi, S. Shabanlu, Optimized ANFIS-genetic algorithm-particle
[8] C.A. Fattor, J.D. Bacchiega, Design conditions for morning-glory spillways:
swarm optimization model for estimation of side orifices discharge coefficient,
application to potrerillos dam spillway, in: Adv. Water Resour. Hydraul. Eng.,
J. Numer. Methods Civ. Eng. 2 (2018) 27–38.
Springer, 2009, pp. 2123–2128.
[38] M. Zounemat-Kermani, A. Mahdavi-Meymand, Hybrid meta-heuristics artificial
[9] X.-L. Dong, K.-L. Yang, X.-L. Guo, Y.-X. Guo, Hydraulic mechanism and application
intelligence models in simulating discharge passing the piano key weirs, J. Hydrol.
of swirling device in morning glory shaft spillway, Shuili Xuebao(J. Hydraul. Eng.
569 (2019) 12–21.
42 (2011) 14–18.
[39] H. Bonakdari, A.H. Zaji, New type side weir discharge coefficient simulation using
[10] M.H. Mirabi, H. Akbari, M. Alembagheri, Detailed vibrational analysis of
three novel hybrid adaptive neuro-fuzzy inference systems, Appl. Water Sci. 8
unbalanced morning glory spillways using coupled finite volume-finite element
(2018) 10.
method, SN Appl. Sci. 3 (2021) 1–16.
[40] F. Salmasi, J. Abraham, Discharge coefficients for ogee weirs including the effects
[11] M. Amouamouha, G.B. Gholikandi, Assessment of anaerobic nanocomposite
of a sloping upstream face, Water Supply 20 (4) (2020) 1493–1508. https://doi.
membrane bioreactor efficiency intensified by biogas backwash, Chem. Eng.
org/10.2166/ws.2020.064.
Process. Intensif. 131 (2018) 51–58.
[41] F. Salazar, B.M. Crookston, A performance comparison of machine learning
[12] G.B. Gholikandi, B.I. Beklar, M. Amouamouha, The Technical and Economical
algorithms for arced labyrinth spillways, Water 11 (2019) 544.
Assessment of the Different Electrode Materials for pH Recovery in the Anaerobic
[42] A. Parsaie, A.H. Haghiabi, A. Moradinejad, Prediction of scour depth below river
Baffled Reactor on a Lab-Scale, Desalin. Water Treat., Under Press, 2018.
pipeline using support vector machine, KSCE J. Civ. Eng. 23 (2019) 2503–2513,
[13] A. Christodoulou, A. Mavrommatis, T. Papathanassiadis, Experimental study on the
https://doi.org/10.1007/s12205-019-1327-0.
effect of piers and boundary proximity on the discharge capacity of a morning
[43] A. Parsaie, A.H. Haqiabi, Development and evaluating of two-neural network
glory spillway, in: In: 1st IAHR Eur. Congr. Edinburgh, Scotl., 2010.
model (MLP1 and SVM2) to estimate the Side weir discharge coefficient, Intl. J.
[14] Z.-P. Liu, X.-L. Guo, Q.-F. Xia, H. Fu, T. Wang, X.-L. Dong, Experimental and
Agric. Crop Sci. 5 (2013) 2804.
numerical investigation of flow in a newly developed vortex drop shaft spillway,
[44] H.M. Azamathulla, A.H. Haghiabi, A. Parsaie, Prediction of side weir discharge
J. Hydraul. Eng. 144 (2018) 4018014.
coefficient by support vector machine technique, Water Sci. Technol. Water Supply
[15] A. Sharafati, M. Haghbin, D. Motta, Z.M. Yaseen, The application of soft computing
16 (2016) 1002–1016.
models and empirical formulations for hydraulic structure scouring depth
[45] A.H. Zaji, H. Bonakdari, Optimum support vector regression for discharge
simulation: a comprehensive review, assessment and possible future research
coefficient of modified side weirs prediction, Ina. Lett. 2 (2017) 25–33.
direction, Arch. Comput. Methods Eng. (2019) 1–25.
[46] K. Roushangar, A. Foroudi Khowr, M. Saneie, Prediction of discharge coefficient for
[16] M. Haghbin, A. Sharafati, D. Motta, N. Al-Ansari, M.H.M. Noghani, Applications of
ogee spillway with curve Axis using support vector machine by comparison with
soft computing models for predicting sea surface temperature: a comprehensive
adaptive neuro fuzzy inference System, Iran, J. Irrig. Drain. 11 (2017) 647–657.
review and assessment, Prog. Earth Planet. Sci. 8 (2021) 1–19.
[47] M. Majedi Asl, M. Fuladipanah, Application of the evolutionary methods in
[17] A. Sharafati, M. Haghbin, S.B. Haji Seyed Asadollah, N.K. Tiwari, N. Al-Ansari, Z.
determining the discharge coefficient of triangular labyrinth weirs, JWSS-Isfahan
M. Yaseen, Scouring depth assessment downstream of weirs using hybrid
Univ. Technol. 22 (2019) 279–290.
intelligence models, Appl. Sci. 10 (2020) 3714.
[48] R. Norouzi, H. Arvanaghi, F. Salmasi, D. Farsadizadeh, M.A. Ghorbani, A new
[18] I. Ebtehaj, H. Bonakdari, A.H. Zaji, H. Azimi, F. Khoshbin, GMDH-type neural
approach for oblique weir discharge coefficient prediction based on hybrid
network approach for modeling the discharge coefficient of rectangular sharp-
inclusive multiple model, Flow Meas. Instrum. 76 (2020) 101810.
crested side weirs, Eng. Sci. Technol. an Int. J. 18 (2015) 746–757.
[49] K. Roushangar, S. Akhgar, A. Erfan, J. Shiri, Modeling scour depth downstream of
[19] F. Salmasi, J. Abraham, Expert System for determining discharge coefficients for
grade-control structures using data driven and empirical approaches, J. Hydroinf.
inclined slide gates using genetic programming, J. Irrigat. Drain. Eng. 146 (2020)
18 (6) (2016) 946–960, https://doi.org/10.2166/hydro.2016.242.
6020013.
[50] H. Karami, S. Karimi, M. Rahmanimanesh, S. Farzin, Predicting discharge
[20] A. Parsaie, S. Dehdar-Behbahani, A.H. Haghiabi, Numerical modeling of cavitation
coefficient of triangular labyrinth weir using support vector regression, support
on spillway’s flip bucket, Front. Struct. Civ. Eng. 10 (2016) 438–444.
vector regression-firefly, response surface methodology and principal component
[21] A. Parsaie, A. Moradinejad, A.H. Haghiabi, Numerical modeling of flow pattern in
analysis, Flow Meas. Instrum. 55 (2017) 75–81.
spillway approach channel, Jordan J. Civ. Eng. 12 (2018).
[51] K. Roushangar, S. Akhgar, Particle swarm optimization-based LS-SVM for
[22] Z.M. Yaseen, M.I. Ghareb, I. Ebtehaj, H. Bonakdari, R. Siddique, S. Heddam, A.
hydraulic performance of stepped spillway, ISH J. Hydraul. Eng. 26 (2020)
A. Yusif, R. Deo, Rainfall pattern forecasting using novel hybrid intelligent model
273–282.
based ANFIS-FFA, Water Resour. Manag. 32 (2018) 105–122.
[52] A.A. Kamanbedast, The investigation of discharge coefficient for the morning glory
[23] M. Haghbin, A. Sharafati, B. Dixon, V. Kumar, Application of soft computing
spillway using artificial neural network, World Appl. Sci. J. 17 (2012) 913–918.
models for simulating nitrate contamination in groundwater: comprehensive
[53] T.J. Alfatlawi, H.I. Alshaikhli, Prediction the coefficient of discharge for stepped
review, assessment and future opportunities, Arch. Comput. Methods Eng. (2020)
morning glory spillway using ANN and MNLR approaches, Int. J. Civ. Environ. Eng.
1–23.
37 (2015), 1701–8285.
[24] O. Bilhan, M.E. Emiroglu, O. Kisi, Application of two different neural network
[54] R. Aghamajidi, Simulations Effect of Vortex Breaker and Stepped Chamber through
techniques to lateral outflow over rectangular side weirs located on a straight
Hydraulic Characteristics of Morning Glory Spill Way, ISLAMIC AZAD
channel, Adv. Eng. Software 41 (2010) 831–837.
UNIVERSITY Science and Research Branch, 2012.
[25] M.E. Emiroglu, O. Bilhan, O. Kisi, Neural networks for estimation of discharge
[55] USBR, Hydraulic Model Studies of San Luis Forebay Dam Spillway-San Luis Unit-
capacity of triangular labyrinth side-weir located on a straight channel, Expert
West San Joaquin Division—Central Valley Project, 1961. Report no. Hyd-517.
Syst. Appl. 38 (2011) 867–874.
California.
[26] F. Salazar, R. Morán, R. Rossi, E. Oñate, Analysis of the Discharge Capacity of
[56] H.O. Anwar, Formation of a weak vortex, J. Hydraul. Res. 4 (1966) 1–16.
Radial-Gated Spillways Using Numerical Modeling Application to Oliana Dam,
2011.

12
M. Haghbin et al. Flow Measurement and Instrumentation 85 (2022) 102161

[57] L.L. Daggett, G.H. Keulegan, Similitude in free-surface vortex formations, [71] A.R. Mehrabian, C. Lucas, A novel numerical optimization algorithm inspired from
J. Hydraul. Div. 100 (1974) 1565–1581. weed colonization, Ecol. Inf. 1 (2006) 355–366.
[58] A.K. Jain, R.J. Garde, K.G. Ranga Raju, Vortex formation at vertical pipe intakes, [72] M. Amouamouha, G. Badalians Gholikandi, Characterization and antibiofouling
J. Hydraul. Div. 104 (1978) 1429–1445. performance investigation of hydrophobic silver nanocomposite membranes: a
[59] Y. Zhang, J. Duchi, M. Wainwright, Divide and conquer kernel ridge regression, in: comparative study, Membranes (Basel) 7 (2017) 64.
Conf. Learn. Theory, PMLR, 2013, pp. 592–617. [73] R.A. Adewoyin, P. Dueben, P. Watson, Y. He, R. Dutta, TRU-NET: a deep learning
[60] Y. You, J. Demmel, C.-J. Hsieh, R. Vuduc, Accurate, fast and scalable kernel ridge approach to high resolution prediction of rainfall, Mach. Learn. 110 (2021)
regression on parallel and distributed systems, in: Proc. 2018 Int. Conf. 2035–2062.
Supercomput, 2018, pp. 307–317. [74] A.M. Barragán-Montero, M. Thomas, G. Defraene, S. Michiels, K. Haustermans, J.
[61] M. Ali, R. Prasad, Y. Xiang, Z.M. Yaseen, Complete ensemble empirical mode A. Lee, E. Sterpin, Deep learning dose prediction for IMRT of esophageal cancer:
decomposition hybridized with random forest and kernel ridge regression model the effect of data quality and quantity on model performance, Phys. Med. 83
for monthly rainfall forecasts, J. Hydrol. 584 (2020) 124647, https://doi.org/ (2021) 52–63.
10.1016/j.jhydrol.2020.124647. [75] G.B. Gholikandi, B.I. Beklar, M. Amouamouha, Performance prediction and
[62] A. Shahsavar, M. Jamei, M. Karbasi, Experimental evaluation and development of upgrading of electroanaerobic baffled reactor using neural-fuzzy method,
predictive models for rheological behavior of aqueous Fe3O4 ferrofluid in the J. Environ. Chem. Eng. 9 (2021) 106029.
presence of an external magnetic field by introducing a novel grid optimization [76] W. Mao, W. Wang, L. Jiao, S. Zhao, A. Liu, Modeling air quality prediction using a
based-Kernel ridge regression supported by sensitivity, Powder Technol. 393 deep learning approach: method optimization and evaluation, Sustain. Cities Soc.
(2021) 1–11. 65 (2021) 102567.
[63] M.K. Chung, Gaussian Kernel Smoothing, Statistical and Computational Methods in [77] S.A. Shetty, T. Padmashree, B.M. Sagar, N.K. Cauvery, Performance analysis on
Brain Image Analysis (2013) 416. https://doi.org/10.1201/b15056. machine learning algorithms with deep learning model for crop yield prediction,
[64] C.E. Rasmussen, Gaussian processes in machine learning, in: Summer Sch. Mach. in: Data Intell. Cogn. Informatics, Springer, 2021, pp. 739–750.
Learn., Springer, 2003, pp. 63–71. [78] S.Q. Salih, I. Alakili, U. Beyaztas, S. Shahid, Z.M. Yaseen, Prediction of dissolved
[65] E. Momeni, M.B. Dowlatshahi, F. Omidinasab, H. Maizir, D.J. Armaghani, Gaussian oxygen, biochemical oxygen demand, and chemical oxygen demand using
process regression technique to estimate the pile bearing capacity, Arabian J. Sci. hydrometeorological variables: case study of Selangor River, Malaysia, Environ.
Eng. 45 (2020) 8255–8267. Dev. Sustain. 23 (2021) 8027–8046.
[66] J.J. Li, A. Jutzeler, B. Faltings, S. Winter, C. Rizos, Estimating urban ultrafine [79] V. Nourani, G. Andalib, D. Dąbrowska, Conjunction of wavelet transform and SOM-
particle distributions with Gaussian process models, Res. Locate14. (2014) mutual information data pre-processing approach for AI-based Multi-Station
145–153. nitrate modeling of watersheds, J. Hydrol. 548 (2017) 170–183.
[67] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, [80] C. Malings, M. Pozzi, Conditional entropy and value of information metrics for
M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, Scikit-learn: machine learning in optimal sensing in infrastructure systems, Struct. Saf. 60 (2016) 77–90.
Python, J. Mach. Learn. Res. 12 (2011) 2825–2830. [81] E. Archer, I.M. Park, J.W. Pillow, Bayesian and quasi-Bayesian estimators for
[68] Q. Quan, Z. Hao, H. Xifeng, L. Jingchun, Research on water temperature prediction mutual information from discrete data, Entropy 15 (2013) 1738–1755.
based on improved support vector regression, Neural Comput. Appl. (2020) 1–10. [82] M. Hutter, M. Zaffalon, Distribution of Mutual Information for Robust Feature
[69] M. Awad, R. Khanna, Support vector regression, in: Effic. Learn. Mach., Springer, Selection, 2002.
2015, pp. 67–80. [83] M.J. Asadi, S. Shabanlou, M. Najarchi, M.M. Najafizadeh, A hybrid intelligent
[70] H. Drucker, C.J.C. Burges, L. Kaufman, A.J. Smola, V. Vapnik, Support vector model and computational fluid dynamics to simulate discharge coefficient of
regression machines, in: Adv. Neural Inf. Process. Syst., 1997, pp. 155–161. circular side orifices, Iran, J. Sci. Technol. Trans. Civ. Eng. (2020) 1–26.
[84] J. Hausser, K. Strimmer, M.K. Strimmer, Package ‘entropy, 2015.

13

You might also like