You are on page 1of 23

Water Resources Management

https://doi.org/10.1007/s11269-023-03497-x

A Hybrid Model for Predicting the Energy Dissipation


on the Block Ramp Hydraulic Structures

Mostafa Rahmanshahi1 · Jafar Jafari‑Asl2 · Mahmood Shafai Bejestan3 ·


Seyedali Mirjalili4,5

Received: 4 December 2022 / Accepted: 11 March 2023


© The Author(s), under exclusive licence to Springer Nature B.V. 2023

Abstract
Block ramps are among the environmentally friendly hydraulic structures used for energy dis-
sipation in rivers and waterways. Modeling the energy dissipation on these structures is ever-
challenging in hydraulic engineering. The primary goal of the current study is to propose a
novel metaheuristic-based artificial intelligence (AI) framework for energy dissipation predic-
tion on block ramp structures. An improved African Vultures Optimization Algorithm (AVOA)
is used to optimize the Adaptive Neuro-Fuzzy Inference System (ANFIS) in this investiga-
tion for accurate prediction of the energy dissipation on the block ramps. The performance of
the hybrid ANFIS-IAVOA model is compared with an ANFIS and its hybrid versions using
original AVOA, honey badger algorithm (ANFIS-HBA), grey wolf optimizer (ANFIS-GWO),
monarch butterfly optimization (ANFIS-MBO), and black widow optimization (ANFIS-
BWO) models. A dataset of 210 experiments measured at Shahid Chamran University of
Ahvaz and 241 experiments collected from literature are used to construct the proposed hybrid
models. The results demonstrate the better efficiency of hybrid ANFIS-IAVOA with RMSE
of 0.018–0.020 and R ­ 2 of 0.98–0.98 compared to ANFIS-AVOA (RMSE ~ 0.023–0.25 and
­R ~ 0.97–0.97), ANFIS-HBA (RMSE ~ 0.021–0.025 and ­R2 ~ 0.97–0.97), ANFIS-MBO
2

(RMSE ~ 0.022–0.023 and ­R2 ~ 0.97–0.97), ANFIS-GWO (RMSE ~ 0.022–0.024 and ­R2 ~ 0.97–
0.97), ANFIS-BWO (RMSE ~ 0.027–0.028 and ­R2 ~ 0.96–0.96), and ANFIS (RMSE ~ 0.029–
0.033 and R­ 2 ~ 0.954 − 0.951). The statistical measures show that the proposed ANFIS-IAVOA
performs better than the other metaheuristic-based and standalone ANFIS-developed models.
The impressiveness of the proposed hybrid model demonstrates that it can be used for further
investigations on the probabilistic assessment of the block ramp hydraulic structures.

Keywords Block ramp · Energy dissipation · Metaheuristic · Artificial intelligence ·


AVOA · ANFIS

1 Introduction

Today, one of the crucial priorities of civil construction is to preserve the environment
as much as possible. The construction of concrete hydraulic structures and the high con-
struction cost have many environmental disadvantages. For example, inline concrete

* Jafar Jafari‑Asl
jafar.jafariasl@yahoo.com
Extended author information available on the last page of the article

13
Vol.:(0123456789)
M. Rahmanshahi et al.

structures such as dams, spillways, weirs, and chutes prevent or at least reduce the pos-
sibility of aquatic animals and fish moving upstream. Environmentally friendly struc-
tures made of stone can be a suitable alternative to conventional rough concrete struc-
tures due to their stability and environmental and economic capabilities. Such structures
are more stable and safer than the concrete type due to their flexibility against static and
dynamic loads, lack of force under pressure and piping, and reduced scouring. It is vital
to the water quality viewpoint that organic matter can pass through permeable bodies of
gabion structures. Therefore, these structures minimize sedimentation and eutrophica-
tion and have a positive environmental impact (Rahmanshahi and Shafai Bejestan 2020).
An application of gabion is the construction of energy dissipators such as block ramps.
Due to the rough surface of the rocks and the resulting increase in turbulence and air mix-
ing, the flow passing over the block ramp experiences a reduction in its kinetic energy.
Several researchers have studied block ramps from different aspects, such as energy dis-
sipation (Pagliara and Chiavaccini 2006; Ahmad et al. 2009; Ghare et al. 2010; Ahmad and
Srisvastava 2014), friction factor (Aberle and Smart 2003; Oertel et al. 2011), flow regimes
(Tamagni et al. 2014), block ramp downstream scouring (Pagliara et al. 2009, 2020), drag
coefficient (Kothyari et al. 2009; Oertel et al. 2011), and stability (Weitbrecht et al. 2017).
Pagliara and Chiavaccini (2006) experimentally studied energy dissipation on block
ramps. Their results showed that energy dissipation decreases with increasing slope.
Based on this study, they presented Eq. (1) to calculate the energy dissipation. Accord-
ing to this relationship, the energy dissipation is a function of the block ramp’s material
size, height, slope, and critical depth. The constant coefficients of this equation are dif-
ferent according to the three categories selected for stones granulation.
𝛥E y
𝛥Er = = A + (1 − A)exp(B + C.S) c (1)
E0 H

where ∆Er=(E0-Et)/E0 is relative energy loss (∆E = Energy at the upstream (E0)–Energy at
downstream (Et)), yc is the critical depth, S is the slope of block ramp, H is the ramp height,
and A, B, and C are the constant parameters that depend on the scale of stones’ sizes.
Ahmad and Srisvastava (2014) conducted the energy dissipation on block ramps
in a quasi-real scale laboratory model. Their results showed that the energy dissipa-
tion increases with the slope increase. This result agreed with Rahmanshahi and Shafai
Bejestan (2020) and contradicted Pagliara and Chiavaccini (2006). Rahmanshahi and
Shafai Bejestan (2020) experimental investigate the block ramps’ energy dissipation.
In their research, based on the gene-expression programming model, a relationship for
energy dissipation was presented as follows
[ ( )] yc y ( )
6.14 yc L c yc
𝛥Er = S + 2.25 − 0.3 (2)
L L d50

where L is ramp length, and d50 is the mean diameter of stones.


The literature shows that most researchers have developed energy dissipation as a
nonlinear mathematical formula from other dimensionless independent parameters. Due
to the integration of nonlinearity, the presented formulas have special restrictions. In
this regard, due to the complexity of the flow over block ramps, artificial intelligence-
based methods can be used to cover the poor performance of the formulas developed
based on standard regression analysis. Artificial intelligence (AI)-based methods can be
effective surrogate techniques in complex water science problems due to high simula-
tion accuracy for parameters with complex nonlinear relationships, high precision for

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

multi-parameter problems, and problems with low correlation (Jang 1993; Ibrahim et al.
2022; Jutury et al. 2022; Rad et al. 2022). AI-based methods have been used in several
studies on energy dissipation (Khatibi et al. 2014; Kim et al. 2021; Salmasi et al. 2021).
Recently, Mahdavi-Meymand et al. (2022) developed and used a metaheuristic-based
AI model to simulate the energy dissipation on block ramps. Their study results showed
that the accuracy of the metaheuristic-based AI model is better than the standalone AI
models in predicting the energy dissipation on the block ramp.
The main objective of this study is to develop and introduce an accurate AI-based frame-
work for predicting energy dissipation over block ramps based on the metaheuristic learning
models. For this aim, two sets of experimental tests with different hydraulic conditions and
various ramp geometric parameters were considered for the training and testing of all mod-
els. In recent years, many AI-based algorithms have been introduced to predict nonlinearly
complex natural phenomena (Ben Seghier et al. 2021; Mai et al. 2022). Among those mod-
els, the recent applications of the Adaptive Neuro-Fuzzy Inference System (ANFIS) (Jang
1993) model indicated better prediction performance due to integrating the fuzzy system in
an artificial neural network. But Cui et al. (2022) showed that the efficiency of ANFIS model
in modeling nonlinearly complex phenomena could be enhanced significantly by optimizing
its hyperparameters. Varieties of metaheuristic techniques have been introduced to optimize
ANFIS model hyperparameters, improving the prediction performance.
In this work, an effort has been made to use a robust, improved metaheuristic algorithm,
namely the African Vultures Optimization Algorithm (AVOA), to improve the prediction
results of ANFIS model in modeling the energy dissipation on the block ramp energy dis-
sipator structure. The efficiency of ANFIS-IAVOA was compared with five metaheuristic
algorithms (i.e., honey badger algorithm (HBA) (Hashim et al. 2022), grey wolf optimizer
(GWO) (Mirjalili et al. 2014), monarch butterfly optimization (MBO) (Wang et al. 2019),
black widow optimization (BWO) (Hayyolalam and Kazem 2020), and the original version of
AVOA (Abdollahzadeh et al. 2021b)) which has been widely used in recent years for the solv-
ing of many real-world problems. Thus, six hybrid models, ANFIS-IAVOA, ANFIS-AVOA,
ANFIS-GWO, ANFIS-HBA, ANFIS-MBO, ANFIS-BWO, and standalone ANFIS model,
were developed herein.
In the remainder of the paper, Section 2 presents a brief description of datasets and its
dimensional Analysis. The proposed hybrid AI-based model, including ANFIS and IAVOA,
are presented in Section 3. The performance and efficiency of the developed metaheuristic-
based ANFIS models are presented in Section 4. In Section 5, the concluding remarks are
summarized.

2 Dataset Description

2.1 Experimental Setup and Measurements

The data used in this research includes two parts. The first part is the data from the
experiments conducted by the authors, and the second part is the existing experiments
from Ahmad and Srisvastava (2014).
The first experiments were carried out in the Hydraulic Laboratory at Shahid Cham-
ran University of Ahvaz in a flume with a width of 0.3 m and a length of 8 m. The
height of the flume was 0.8 m in the first 2 m and 0.4 m in the last 6 m. Experiments
were carried out for five different slopes (1 V (vertical)∶4 H (Horizontal), 1 V:5 H,

13
M. Rahmanshahi et al.

1 V:6 H, 1 V:8 H, 1 V:12 H) of block ramp with five different uniform material sizes
(1.1, 1.4, 1.7, 2.4, 3 cm). To get rid of any turbulence and instability in the incoming
flow, a mesh wall was placed at the entrance of the flume, and a wave suppressor was
installed at the exit of the feeding pipe. A 0.3 m long broad-crested weir was used at the
upstream boundary of the block ramps. The hydraulic jump at the toe of the ramp block
was regulated by using a sluice gate located at the end of the flume.
The stones were fixed on the ramp using glue. Two hundred and ten tests were per-
formed for flow rates between 8 and 36 L/s. The flow rate was measured by an elec-
tromagnetic flow meter installed on the flume inlet pipe with an accuracy of ± 1%. A
schematic of the block ramp and a laboratory photo is shown in Fig. 1.
The energy dissipation on block ramps can be calculated from the following equation
E0 − E1
𝛥Er = (3)
E0

where E0 = H + v2/2 g, E1 = y1 + v12/2 g, v1 = q/y1 is downstream velocity, and v = q/H is


upstream approach velocity. The continuity equation calculates the average velocity at each
section.
In each test, the block ramp’s upstream depth and the jump’s secondary depth were meas-
ured using a point gauge with an accuracy of ± 0.1 mm. The upstream energy of the ramp
was calculated based on its upstream approach flow depth (h), and the ramp toe energy was
calculated based on the initial jump depth (y1). Due to the fluctuations and disturbances

Fig. 1  a Schematic, and b experimental view of block ramp

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

downstream of the ramp, the primary jump depth was calculated based on the secondary jump
depth (y2) using Belanger equation:
y √
y1 = 2 ( 1 + 8Fr22 − 1) (4)
2
in which Fr is Froud number.
Ahmad and Srisvastava (2014) used pseudo-real block ramps in the second set of experi-
ments. These tests were performed with three slopes, 1 V∶9.87 H, 1 V∶6 H, and 1 V∶3 H,
with nine different material sizes with mean diameters between 64.0 and 133.8. The width of
the flume used was 0.83 m. The height of the block ramps was 0.7 m with lengths of 6.21, 4.2,
and 2.1 m. Details of two sets of experiments are presented in Table 1. Moreover, the histo-
grams of the mentioned parameters are plotted in Fig. 2, which indicates that each parameter
has various values showing the complexity of the energy dissipation prediction problem.

2.2 Dimensional Analysis

The relative energy dissipation over the block ramp as a function of discharge per unit width, q,
the average diameter of the stones, d50, length of block ramp, L, block ramp slope, S, dynamic vis-
cosity, µ, the density of water, ρ, surface tension, σ, and the gravitational acceleration, g, as follows
𝛥Er = f (q, d50 , L, S, 𝜇, 𝜌, 𝜎, g) (5)
Using dimensional Analysis and also by using the definition of critical depth in rectan-
gular section as (q2/g)1/3, the parameters of Eq. (5) can be made dimensionless as follows:
𝛥E y y
𝛥Er = = f ( c , c , S, Re, We) (6)
E0 d50 L

where, f is a functional symbol, is relative critical flow depth, is a ratio of critical


yc yc
d50 L
depth to total ramp length, Re is Reynolds number, and We is the Weber number.
According to (Novák and Čabelka 1981), if the head over the block ramp is more than
3 cm, the effects of surface tension can be ignored. Therefore, the effect of Weber number
has been ignored in this research.
In all experiments, the Reynolds number was higher than ­104 for a fully turbulent flow (Ferro
1999). Thus, the effect of the fluid viscosity is not essential. Accordingly, the Reynolds number can
be omitted from Eq. (7). Therefore, the relative energy dissipation ∆Er can be expressed as follow:

Table 1  Details of the experimental models of the two datasets used in this research
Dataset No. S d50 Q H L
(-) (mm) (l/s) (m) (m)

Set 1 (This study) 1:4 11, 14, 17, 24, 30 8 to 36 0.46 1.9
1:5 11, 14, 17, 24, 30 0.37
1:6 11, 14, 24, 30 0.31
1:8 11, 14, 24, 30 0.24
1:12 11, 14, 30 0.16
Set 2 (Ahmad and 1:3 64 to 33.8 14 to 106 0.7 6.21
Srisvastava 2014) 1:6 64 to 33.8 4.2
1:9.87 64 to 33.8 2.1

13
M. Rahmanshahi et al.

Fig. 2  Histograms and normal distributions of the input and output variables

𝛥E y y
𝛥Er = = f ( c , c , S) (7)
E0 d50 L

The range and correlation of independent dimensionless parameters of the two datasets
used in this research are presented in Table 2.
The effect of relative critical flow depth, the ratio of crucial depth to total ramp length,
and block ramp slope that influence the energy dissipation on the block ramps are typically
recommended (Rahmanshahi and Shafai Bejestan 2020) through dimensionless groups.
Therefore, it can be concluded that the energy dissipation on the block ramps is a function
of dimensionless groups, including relative critical flow depth d c , the ratio of critical depth
y
50
to total ramp length Lc , and block ramp slope S.
y

3 Prediction Models

This section presents a soft computing-based method and a nature-inspired algorithm,


namely ANFIS and the improved variant of AVOA. By then, an efficient prediction model
named ANFIS-AVOA based on improved AVOA and ANFIS is built to estimate the energy
dissipation on the block ramp.

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

3.1 ANFIS

Introduced by Wang et al. (2019), the ANFIS integrates a fuzzy inference system (FIS)
and ANN to provide a robust learning tool and reduce the difficulty of the ANN model.
Typically, the ANFIS estimates a phenomenon using five main steps (i.e., Fuzzification,
Product, Normalization, De-fuzzification, and Output):

Step 1: Fuzzification: During this stage, the Takagi-Sugeno fuzzification rule produces
fuzzy sets by considering the inputs provided. There are some constant parameters in
the first layer, which adjust during the learning process.
Step 2: Product: The products (w) are generated for input signals in this layer.
Step 3: Normalization: In this layer, the product values are normalized through Eq. (8)
as below:
− w1
w1 =
w1 + w2 (8)

where w1 and w2 are values of generated products for nodes 1 and 2, respectively.
Step 4: Normalization: In this layer, a function is applied to normalized values of prod-
ucts as follows:
− − ( )
wi fi = wi pi x + qi y + ri (9)

where pi, qi and ri are the inference parameters.


Step 5: Output: In this layer, the outputs are estimated using Eq. (10) as below:
∑ −
�− wf
f = wi fi = ∑i i i (10)
i i wi

The ANFIS modeling process involves two transmission procedures: forward transmis-
sion, where a least squares algorithm determines the parameters of the fuzzy rule function,
and backward transmission, where an error backpropagation algorithm adjusts the parameters
of the fuzzy membership function for the inputs (Cui et al. 2022). The efficiency of ANFIS
model depends on the hyperparameter values, including the number of membership functions
and their types. To achieve the proper results in the prediction, optimizing these parameters

Table 2  Minimum, maximum, Dataset No. Parameter Independent dimensionless


and correlation of independent variables
dimensionless parameters of the
two datasets used in this research yc ∕L yc ∕d50 S

Set 1 (This study) Correlation -0.783** -0.757** 0.483**


Min 0.022 1.390 0.083
Max 0.059 10.139 0.25
Set 2 (Ahmad and Correlation -0.657** -0.707** 0.317**
Srisvastava 2014) Min 0.0047 0.24 0.1
Max 0.054 1.58 0.33
**
Significant at the level of 0.01

13
M. Rahmanshahi et al.

is necessary during the learning process of ANFIS. According to the literature review, many
optimization techniques are proposed for training ANFIS (Arya Azar et al. 2021; Cui et al.
2022). Trapping to local minima is the main challenge of optimization techniques in this prob-
lem. Thus, a recently efficient variant of AOVA optimizer is used in the current work to boost
the ANFIS model in the training stage. The efficiency of the proposed IAOVA is compared
with AVOA, HBA, GWO, BWO, and MBO, which have been presented to illustrate good
performance in the optimization of various engineering and natural problem. The IAOVA is
explained below.

3.2 Improved African Vulture Optimization Algorithm

The AVOA was introduced by Abdollahzadeh et al. (2021a) based on African vultures’ behav-
ior in nature. The AVOA simulates the behavior of African vultures using four main steps:

1- Choose the two vultures as optimal solutions


2- Update the optimal solutions in each iteration
3- Simulate the hungry and satisfaction of vultures
4- The exploration and the exploitation phases

In AVOA, the i-th initial random is generated as below:

Xi = Ximin + (Ximax − Ximin ) × rand (11)

where Ximin and Ximax are the lower and upper bounds of the decision variables, respectively.
rand is a function to generate a random number between 0 and 1. i is from 1 to N, in which
N is the number of vultures’ population.
After the all-generated solution, the fitness of the population is evaluated, and an initial
solution matrix is generated as follows:
1 1
⎡ X1,1 ⋯ X1,Dim ⎤
X= ⎢ ⋮ ⋱ ⋮ ⎥ (12)
⎢ 1 1 ⎥
⎣ XN,1 ⋯ XN,Dim ⎦ N×Dim

In this context, “Dim” refers to the dimensionality of the optimization problem being
considered.
According to the fitness of each solution, the population of vultures is categorized into
three groups. The solution with the lowest values of the objective function is recognized
as the best and first group. Then the second-best solution is classified as the second-best
vulture, and other solutions are categorized into the third group. The main task of catego-
rization is to find the role of vultures. According to the nature of vultures’ life, the weakest
vultures are related to the hunger vultures. In contrast, the strongest vultures relate to the
most numerous and the best vulture. Therefore, in the AVOA, all vultures try to be nearby
the strong vultures while avoiding the weakest (Abdollahzadeh et al. 2021b).
Generally, the AVOA simulation can be explained in five steps based on the above
explanation as below steps:

a- The categorizing of population: After determining of the first vulture, the second-best
vulture is determined by Eq. (13):

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

{
BestVulture1 if pi = L1
R(i) =
BestVulture2 if pi = L2 (13)

 Where BestVulture1 and BestVulture2 are the first and second-best vultures, respec-
tively. L1 and L2 represents the two random numbers between 0 and 1, whose sum equals
1. pi is a random number and calculated using the roulette-wheel approach by Eq. (14):
F
pi = ∑n i (14)
i=1 Fi

where Fi represents the fitness of two groups, and n is the number of vultures in the
first and second groups.
b- The starvation rate of vultures: The exploration and exploitation stages can be deter-
mined based on the starving and satisfaction of vultures.
( )
( ) iterationi
Fi = 2 × randi + 1 × z × 1 − +t (15)
maxiteration

in which, Fi is an index to represent the vulture is starved or satisfied. randi is a ran-


dom number between 0 and 1, z is the random number in the range -1 to 1, and t is
estimated by Eq. (16):
( ( ) ( ) )
𝜋 iterationi 𝜋 iterationi
t = h × sinw × + cosw × −1 (16)
2 maxiteration 2 maxiteration

  If |F| < 1, the AVOA is in the exploitation stage, else AVOA searches in the explo-
ration stage. c- Exploration phase: Using two strategies, the vultures can investigate
random locations.
{
R(i) − D(i) × Fi ( ) if randp1 ≤ P1
P(i + 1) =
R(i) − Fi + rand2 × (ub − lb) × rand3 + lb Otherwise (17)

where R(i) represents one of the best solutions in the current iteration based on
Eq. (13), Fi is the satiation rate of vulture based on Eq. (13), lb is the lower and ub
is the upper bounds of decision variables, rand2 and rand3 are two random numbers
between 0 and 1. D(i) is the distance between the vulture and the current optimum one
and calculated as:
D(i) = |X × R(i) − P(i)| (18)
where P(i) and X are the i-th vulture’s position and a random number in the range of 0
and 1.
c- Exploitation phase:
  The exploitation stage in AVOA is provided in to state as follows:
  If | Fi | <1:
{ ( )
D(i) − ( Fi + rand
) 4 − d(t) if randp1 ≤ P2
P(i + 1) =
R(i) − S1 − S2 if randp1 > P2 (19)

where rand4 is a random number in [0,1], and d(t) is the distance between the second
group’s best vulture and vulture. d(t) is calculated as:

13
M. Rahmanshahi et al.

d(i) = R(i) − P(i) (20)


where:
( )
rand5 × P(i)
S1 = R(i) × × cos(P(i)) (21)
2𝜋

( )
rand6 × P(i)
S2 = R(i) × × sin(P(i)) (22)
2𝜋

where rand5 and rand6 are two random numbers between 0 and 1.
  If | Fi | <0.5:
A1 + A2
P(i + 1) = (23)
2
where:
BestVulture1 (i) × P(i)
A1 = BestVulture1 (i) − × Fi (24)
BestVulture1 (i) − (P(i))2

BestVulture2 (i) × P(i)


A2 = BestVulture2 (i) − × Fi (25)
BestVulture2 (i) − (P(i))2

  Also, when the AVOA process is in this stage, the vulture tends to follow the best
vulture to eat the food. Thus, the vultures’ new position is updated as follows:
P(i + 1) = R(i) − |d(t)| × Fi × Levy(d) (26)
  Lévy flight patterns are defined as:
u×𝜎
Levy(d) = 0.001 ×
|v|𝜎 (27)

� � 1
⎛ 𝜋𝛽 ⎞p
⎜ Γ(1 + 𝛽) × sin 2 ⎟
𝜎=⎜ � �⎟ (28)
⎜ Γ(1 + 𝛽) × 𝛽 × 2 × 𝛽−1

⎝ 2 ⎠

where, u and v are two random numbers between 0 and 1, respectively, and 𝛽 is a constant
number equal to 1.5.
The AVOA has an explicit exploration and exploitation mechanism. However, this algo-
rithm still has some shortcomings, such as easily trapping into the local optima and not
having a good balance in transmission from exploration to the exploitation stage.
Fan et al. (2021) improved AVOA (IVOA) by adding a tent chaotic mapping tool in the
population initialization stage and coupling a time-varying mechanism in the location updat-
ing stage. In this improved algorithm, tent chaos is integrated into the solution stage’s initiali-
zation to improve the algorithm’s exploration ability. Then, IAVOA used the vulture’s his-
torical information to upgrade its’ exploration ability. Moreover, two time-varying parameters
are used to create a good balance between the explorations to exploitation phases. One of

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

the proposed time-varying parameters decreases in each iteration, and another increase at the
same iteration. According to the obtained results of IAVOA in solving real-world engineer-
ing design problems by Fan et al. (2021), it has a reasonable convergence rate and a highly
comprehensive global exploration ability. Therefore, as mentioned before, this study tries to
develop an efficient learning tool based on IAVOA for optimizing the hyperparameters of
ANFIS model. It should be noted that for optimizing the hyperparameters of ANFIS, the
objective function was defined as minimizing the root mean square error (RMSE) between
the predicted energy dissipation value on the block ramps through optimized ANFIS and its
measured value.

4 Implementation, Results, and Discussion

The primary goal of this research is to employ metaheuristic-based ANFIS models to pro-
pose a robust framework to estimate the energy dissipation on block ramps. Five statisti-
cal parameters are employed to compare the measured values of the energy dissipation on
the block ramp with the estimated ones using the introduced AI-based framework. These
parameters are used as powerful indicators to demonstrate models’ prediction performance
compared to each other’s. Table 3 presents the statistical indicators, including each param-
eter and the corresponding equation (Ohadi et al. 2023).
In the following, the statistical criteria parameters (i.e., R2, RMSE, MAE, NSE, and
­U95) were assessed in detail to recognize the efficient hybrid AI-based predictive model,
and the results have been presented.
The experimental dataset is then classified into the train (70%) and the test (30%). Then
the training datasets were used to build ANFIS models based on metaheuristic algorithms,
while the testing subset was implemented to validate the performance of models.
First, a primary Sugeno-type fuzzy inference system was constructed in the training
stage, as shown in Fig. 3. The optimal values of ANFIS parameters are also presented in
Table 4, showing the details of the membership function type, the number of fuzzy rules,

Table 3  Statistical parameters description


Metric Definition Equation

RMSE Root mean squared error 1 ∑n � obs �2
n i=1
𝛥Er,i − 𝛥EPre
r,i

MAE Mean absolute error 1 ∑n � obs Pre �


n i=1 ��𝛥E r,i − 𝛥E r,i ��
NSE Nash-Sutcliffe efficiency ∑n � �2
i=1𝛥Eobs Pre
r,i −𝛥E r,i
1− ∑n � �2 − ∞ ≤ NSE ≤ 1
obs ave
i=1 𝛥E r,i −𝛥E r

R2 Correlation coefficient ∑n � �2
i=1 𝛥Eobs Pre
r,i −𝛥E r,i
1− ∑n � �2
i=1 𝛥Eobs Pre
r,i −𝛥E r,i

U95 Bias-Variance 1 ∑n i
n i=1 (ymea − yipre )

𝛥Eobs
r,i
: Observed value of the i-th energy dissipation
𝛥EPre
r,i
: Predicted value of the i-th energy dissipation
r : Average value of the energy dissipation from the measured data
𝛥Eave

13
M. Rahmanshahi et al.

Fig. 3  Schematic representation of a hybrid ANFIS integrated with an optimizer algorithm for energy dis-
sipation forecasting on block ramps

the maximum number of epochs in training, and the output membership function. For
achieving the best results in the modeling using metaheuristic algorithms, using the appro-
priate values of constant parameters of algorithms is recommended. This study adopts
these parameters from their main references (see Table 5).
Figure 3 indicates the proposed framework for forecasting energy dissipation on the
block ramp using the auto-tanned ANFIS with the metaheuristic techniques (i.e., IAVOA,
AVOA, HBA, MBO, GWO, and BWO).
The calculated values of statistical parameters for each model during the training and
test stages are presented in Table 6. The metaheuristic-based ANFIS models performed
very well compared to standard ANFIS models. The relative efficiency of the metaheuris-
tic-based ANFIS models revealed the much better efficiency of IAVOA in training
ANFIS model, with ­R2 = 0.98, RMSE = 0.018, MAE = 0.014, and NSE = 0.98 compared
to other algorithms. According to Table 6, the poorest model is the standalone ANFIS,
with ­R2 = 0.95, RMSE = 0.029, MAE = 0.020, and NSE = 0.95. The convergence curves of
metaheuristic-based ANFIS training tools are plotted in Fig. 4. As can be seen, the ANFIS-
IAOVA outperformed other metaheuristic-based ANFIS models in this case, indicating
the lowest RMSE in the training phase. Considering the complexity of the investigated

Table 4  ANFIS parameters Parameter Value

Fuzzy structure Sugeno-type


Initial FIS for training Genfis3
MF type Gaussian
Output MF Linear
Optimization method Hybrid
Number of fuzzy rules 10
Maximum number of epochs in training 1000

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

Table 5  The parameters of the Algorithm Parameter setting


metaheuristic algorithms
General Setting Pop size = 50
Max iter = 500
GWO Control parameter (a) is [2,0]
HBA C = 6;𝛽 = 1.5
MBO Neighbor (k) = 5; Flap = 30
BWO procreate rate = 0.6; can-
nibalism rate; mutation
rate = 0.4
AVOA 𝛽 = 0.2;𝛼 = 0.8;𝛾 = 2.5

problem and the multiplicity of decision variables in it and, per theory, no-free lunch, most
of the best algorithms effective in solving other problems may not be able to solve the
investigated problem. Based on the results, the chaotic mapping and nonlinear equation
added to AVOA have overcome the main weaknesses, including early convergence and
getting trapped in local optima. In addition to the lowest value of the objective function
(RMSE), the improved version of AVOA algorithm has found the optimal solution faster
than its original version and other algorithms.
The training tool of the ANFIS was trapped in local minima. As a result, the model
couldn’t create a proper relation between input and output variables. On the other hand,
the AVOA, HBA, MBO, GWO, and BWO showed almost similar efficiency in training the
ANFIS model for predicting the energy dissipation on block ramps. Besides, the results
illustrate that the metaheuristic-based ANFIS model has the best performance compared
to the standalone ANFIS for the testing stage. Similar to the training stage, the highest
efficiency was achieved when ANFIS was hybridized with a metaheuristic algorithm. For
the testing datasets (Table 6), as can be seen, the ANFIS-IAVOA yielded the highest pre-
­ 2 = 0.98, RMSE = 0.020, MAE = 0.017, and NSE = 0.98) compared to the
cision (i.e., R
ANFIS-HBA, ANFIS-MOB, ANFIS-GWO, ANFIS-AVOA, and ANFIS models showing
the IAVOA as an investigation method to improve the ANFIS accuracy.

Table 6  The results of the Model Dataset Metrics


developed models for ∆Er
R2 RMSE MAE NSE U95

ANFIS Train 0.954 0.029 0.020 0.953 -2.24E-03


Test 0.951 0.033 0.025 0.948 -6.70E-03
ANFIS-AVOA Train 0.971 0.023 0.017 0.971 1.74E-04
Test 0.971 0.025 0.018 0.970 1.57E-03
ANFIS-IAVOA Train 0.982 0.018 0.014 0.982 3.47E-04
Test 0.981 0.020 0.017 0.981 -3.41E-04
ANFIS-HBA Train 0.977 0.021 0.013 0.977 -1.23E-03
Test 0.971 0.025 0.018 0.970 -4.33E-03
ANFIS-MBO Train 0.975 0.022 0.015 0.974 -6.33E-04
Test 0.974 0.023 0.017 0.974 -2.86E-03
ANFIS-GWO Train 0.973 0.022 0.015 0.972 3.83E-04
Test 0.973 0.024 0.017 0.973 -2.50E-03
ANFIS-BWO Train 0.964 0.027 0.021 0.960 -4.99E-04
Test 0.961 0.028 0.021 0.964 -8.85E-04

13
M. Rahmanshahi et al.

Fig. 4  Convergence curves of hybrid models

The ­U95 index, which indicates the uncertainty bounds of the prediction process, dem-
onstrates that the metaheuristic-ANFIS models provide underestimated forecasting. At the
same time, the ANFIS offers the overestimated values of the energy dissipation on block
ramps. As can be seen, the lowest forecasting uncertainty belonged to the ANFIS-IAVOA
­(U95 = 3.47E-04 for training datasets; ­U95=-3.41E-04 for testing datasets).
Moreover, aggregated statistical indicator (OBJ) was calculated based on the RMSE, R­ 2,
and MAE values through the modeling process (training and testing). It assesses the devel-
oped model’s efficiency (Eq. 29) (Ben Seghier et al. 2023).
( ) ( )
ntrain RMSEtrain + MAEtrain ntest RMSEtest + MAEtest
OBJ = + (29)
n R2train n R2test

The OBJ values of metaheuristic-based ANFIS and ANFIS are plotted in Fig. 5 as a bar
chart. Accordingly, the ANFIS-IAVOA demonstrated the lowest value of OBJ, which equals

Fig. 5  The OBJ values of all


developed models for ∆Er

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

Fig. 6  The training and testing results of different methods in predicting of ∆Er

13
M. Rahmanshahi et al.

0.034, followed by the ANFIS-HBA (OBJ = 0.037), then ANFIS-MBO (OBJ = 0.039), the
ANFIS-GWO (OBJ = 0.04), the ANFIS-AVOA (OBJ = 0.042), the ANFIS-BWO (OBJ = 0.05),
while the highest value of OBJ is 0.054 that was calculated for standalone ANFIS.
Figure 6 shows the scatter plots of measured versus estimated energy dissipation val-
ues on block ramps for all developed ANFIS-based models. According to this figure, the
metaheuristic-based ANFIS model is very capable of modeling 𝛥Er , because many of the
estimated values correlate reasonably with the measured data. As can be seen in Fig. 5,
using the metaheuristic algorithms for training the ANFIS model greatly improves the pre-
diction performance of the ANFIS, with an enhancement rate of 3.77%, 2.41%, 2%, 1.99%,
1.78% and 0.73% using IAVOA, HBA, MBO, GWO, AVOA, and BWO, respectively.
Another graphical plot to show the modeling errors obtained from each developed
ANFIS-based model is plotted in Fig. 7. It can be seen that the relative error for ANFIS-
IAVOA model aggregated around the zero roughly in the range of [0.7, -0.7]. In contrast,
the range of relative error to the ANFIS is almost between 0.1 and -0.1. Consequently, the
ANFIS-IAVOA performs more appropriately compared to the other models.
The Taylor diagram can be a brief criterion for evaluating the prediction performance
of the proposed ANFIS-based models, as shown in Fig. 8. A Taylor diagram combines the
standard deviation, the RMSE, and the correlation coefficient to indicate how the simulated
values are fitted to the measured datasets. As a result, the highest similarity to the meas-
ured dataset corresponded to the best prediction model. From Fig. 8, the ANFIS-IAVOA is
the best prediction model with the closest point to the measured values of the energy dis-
sipation on block ramps in the train and the test stages. After the ANFIS-IAVOA, the rank-
ing of the developed estimate performance is ANFIS-HBA, ANFIS-AVOA, ANFIS-MBO,
ANFIS-GWO, ANFIS- BWO, and ANFIS.

Fig. 7  ∆Er errors for training and testing data sets for different methods

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

Fig. 8  Taylor diagram graphs of the hybrid ANFIS models; a Train, b Test

13
M. Rahmanshahi et al.

Table 7  The results of ANFIS – Model Metrics


IAVOA, Eqs. (1) and (2) for ∆Er
R2 RMSE MAE NSE U95

ANFIS-IAVOA 0.981 0.02 0.017 0.981 -0.00034


Equation (1) 0.85 0.057 0.044 0.823 -0.0238
Equation (2) 0.94 0.037 0.029 0.932 -0.00027

In order to illustrate the efficiency of the developed metaheuristic-based ANFIS models


for accurately modeling the energy dissipation values on block ramps, the statistical crite-
ria parameters of ANFIS-IAVOA model are compared with the existing empirical equa-
tions. A comparative study was carried out in Table 7 based on the ­R2, RMSE, MAE, NSE,
and ­U95 values between the results of the best model in this study and the two empirical
equations, including the Eq. (1), and Eq. (2). Accordingly, it can be concluded that the
developed ANFIS-IAVOA model succeed to accurately simulate the energy dissipation
values on block ramps despite their complexity. By comparing the best existing equation,
Eq. (2) that yields an ­R2, RMSE, MAE, NSE, and ­U95 values equal to 0.939, 0.036, 0.030,
0.936, and -0.0035, respectively, with the best developed metaheuristic-ANFIS techniques,
an enhancement is captured in all statistical parameters of ­R2, RMSE, MAE, NSE, and ­U95
that is clearly defined from Table 7.

Fig. 9  The time series plots using a ANFIS-IAVOA model, b Eq. 1, and c Eq. 2

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

Fig. 10  Comparison between


measured and predicted energy
dissipation using ANFIS-
IAVOA, Eqs. (1) and (2)

The time series plots of ANFIS-IAVOA as the best ANFIS-based model, and the
Eqs. (1) and (2) are shown in Fig. 9. As can be seen from the detailed graphs, the energy
dissipation estimated values proposed are closer to the measured than the two empirical
equations. Figure 10 illustrates the scatterplot comparisons of the mentioned models. This
figure shows that the ANFIS-IAVOA model has a higher ­R2 value with fewer scattered pre-
dictions than Eqs. (1) and (2).
Percentage exceedance for absolute ∆Er error of the ANFIS-IAVOA model and the two
existing equations can be seen in Fig. 11. As can be seen from this diagram, the AI-based
model’s predictions have the lowest absolute error compared to empirical equations. It fur-
ther verifies the efficiency and accuracy of the metaheuristic-based ANFIS model.
As mentioned before, developing and applying an efficient model for forecasting the
energy dissipation values on block ramps is necessary. The developed models (metaheuris-
tic-based ANFIS) performed superior to the standalone ANFIS and the empirical equa-
tions. It can be concluded that the metaheuristic-based ANFIS managed promising effi-
ciency and could be applied to simulate the energy dissipation values on block ramps
structures. The main advantage of the metaheuristic-based learning tools for AI models is
that they avoid from local-minima and, therefore, they can better determine the parameters
of the AI model. Hydraulic engineers can quickly and easily use the ANFIS structure of
Fig. 3 with the proposed model’s optimal parameters to accurately predict the energy dis-
sipation on block ramps structures.

13
M. Rahmanshahi et al.

Fig. 11  Percentage exceedance


for absolute ∆Er error

5 Conclusion

The current study presents the first successful try at using six advanced metaheuris-
tic-based AI models: ANFIS-IAVAO, ANFIS-AVOA, ANFIS-HBA, and ANFIS-MBO
ANFIS-GWO, and ANFIS-BWO for predicting the energy dissipation on the block
ramps. The proposed metaheuristic-based ANFIS models can be investigated the com-
plex relationship between the energy dissipation to the three variables of yc ∕d50 , yc ∕L ,
and S. This work also compared the best proposed AI-based model and existing equa-
tions for predicting the energy dissipation on the block ramp. The results indicated that
all metaheuristic-based AI models could estimate the energy dissipation on the block
ramp robustly and reliably. Metaheuristic-based ANFIS models exhibited outstand-
ing forecasting efficiency with RMSE, and R ­ 2 criteria were 0.02 and 0.98 for ANFIS-
IAVOA, 0.025 and 0.97 for ANFIS-AVOA, 0.025 and 0.97 for ANFIS-HBA, 0.023 and
0.97 for ANFIS-MBO, 0.024 and 0.97 for ANFIS-GWO, 0.028 and 0.96 for ANFIS-
BWO, and 0.025 and 0.97 for standalone ANFIS. This indicated that using the learning
metaheuristic techniques in tuning the ANFIS method in energy dissipation prediction
is necessary. Among the hybrid strategies, ANFIS-IAVOA performed better than the
other metaheuristic-based ANFIS models. Compared with the two existing experimen-
tal equations, the predicted energy dissipation of the developed ANFIS-IAVOA model
is superior. Overall, it can be concluded that the engineers can easily use the proposed
hybrid framework to accurate simulation of the energy dissipation on block ramps
structures.‘
Using clustering techniques based on hybrid metaheuristic models can be recom-
mended for future works to determine the training and validation datasets in modeling
this nonlinear phenomenon. Moreover, an investigation is recommended using the pro-
posed ANFIS-IAVOA model to estimate the energy dissipation of the block ramp under
uncertain conditions.

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

Data Availability The data presented in this study are available on request from the corresponding author.

Declarations
Ethics Approval All work complies with Ethical Standards.

Consent to Participate All authors confirm their co-authorship.

Consent to Publish The authors give their permission to publish.

Conflict of Interest The authors declare that they have no competing interests.

References
Abdollahzadeh B, Gharehchopogh FS, Mirjalili S (2021) African vultures optimization algorithm: a
new nature-inspired metaheuristic algorithm for global optimization problems. Comput Ind Eng
158:107408
Abdollahzadeh B, Soleimanian Gharehchopogh F, Mirjalili S (2021) Artificial gorilla troops optimizer:
a new nature-inspired metaheuristic algorithm for global optimization problems. Int J Intell Syst
36:5887–5958
Aberle J, Smart GM (2003) The influence of roughness structure on flow resistance on steep slopes. J
Hydraul Res 41:259–269
Ahmad Z, Petappa NM, Westrich B (2009) Energy dissipation on block ramps with staggered boulders. J
Hydraul Eng 135:522–526
Ahmad Z, Srisvastava D (2014) Energy dissipation on block ramps with large scale roughness. In: ISHS
2014-hydraulic structures and society-engineering challenges and extremes: Proceedings of the 5th
IAHR International Symposium on Hydraulic Structures. The University of Queensland, pp 1–8
Arya Azar N, Ghordoyee Milan S, Kayhomayoon Z (2021) The prediction of longitudinal dispersion
coefficient in natural streams using LS-SVM and ANFIS optimized by Harris hawk optimization
algorithm. J Contam Hydrol 240:103781. https://​doi.​org/​10.​1016/j.​jconh​yd.​2021.​103781
Ben Seghier MEA, Corriea JAFO, Jafari-Asl J et al (2021) On the modeling of the annual corrosion rate in main
cables of suspension bridges using combined soft computing model and a novel nature-inspired algo-
rithm. Neural Comput Appl 2021 3323 33:15969–15985. https://​doi.​org/​10.​1007/​S00521-​021-​06199-W
Ben Seghier MEA, Golafshani EM, Jafari-Asl J, Arashpour M (2023) Metaheuristic-based machine learning
modeling of the compressive strength of concrete containing waste glass. Struct Concr. https://​doi.​org/​
10.​1002/​suco.​20220​0260
Cui F, Al-Sudani ZA, Hassan GS et al (2022) Boosted artificial intelligence model using improved alpha-
guided grey wolf optimizer for groundwater level prediction: comparative study and insight for feder-
ated learning technology. J Hydrol 606:127384
Fan J, Li Y, Wang T (2021) An improved african vultures optimization algorithm based on tent chaotic map-
ping and time-varying mechanism. PLoS ONE 16:e0260725
Ferro V (1999) Friction factor for gravel-bed channel with high boulder concentration. J Hydraul Eng
125:771–778
Ghare AD, Ingle RN, Porey PD, Gokhale SS (2010) Block ramp design for efficient energy dissipation. J
Energy Eng 136:1–5
Hashim FA, Houssein EH, Hussain K et al (2022) Honey badger algorithm: new metaheuristic algorithm for
solving optimization problems. Math Comput Simul 192:84–110
Hayyolalam V, Kazem AAP (2020) Black widow optimization algorithm: a novel meta-heuristic approach
for solving engineering optimization problems. Eng Appl Artif Intell 87:103249
Ibrahim KSMH, Huang YF, Ahmed AN et al (2022) Forecasting multi-step-ahead reservoir monthly and
daily inflow using machine learning models based on different scenarios. Appl Intell 1–24
Jang J-S (1993) ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans Syst Man Cybern
23:665–685
Jutury D, Kumar N, Sachan A et al (2023) Adaptive neuro-fuzzy enabled multi-mode traffic light control
system for urban transport network. Appl Intell 53:7132–7153
Khatibi R, Salmasi F, Ghorbani MA, Asadi H (2014) Modelling energy dissipation over stepped-gabion weirs
by artificial intelligence. Water Resour Manag 28:1807–1821. https://​doi.​org/​10.​1007/​s11269-​014-​0545-y

13
M. Rahmanshahi et al.

Kim S, Salmasi F, Ghorbani MA et al (2021) Energy Dissipation in Rough Chute: Experimental Approach
Versus Artificial Intelligence Modeling. Intell Data Anal Decis Syst Hazard Mitig Theory Pract Haz-
ard Mitig 227–249
Kothyari UC, Hayashi K, Hashimoto H (2009) Drag coefficient of unsubmerged rigid vegetation stems in
open channel flows. J Hydraul Res 47:691–699
Mahdavi-Meymand A, Sulisz W, Zounemat-Kermani M (2022) A comprehensive study on the applica-
tion of firefly algorithm in prediction of energy dissipation on block ramps. Eksploat i Niezawodn
24. https://​doi.​org/​10.​17531/​ein.​2022.2.2
Mai SH, Ben Seghier MEA, Nguyen PL et al (2022) A hybrid model for predicting the axial compression
capacity of square concrete-filled steel tubular columns. Eng Comput 38:1205–1222. https://​doi.​org/​
10.​1007/​s00366-​020-​01104-w
Mirjalili S, Mirjalili SM, Lewis A (2014) Grey wolf optimizer. Adv Eng Softw 69:46–61. https://​doi.​org/​10.​
1016/J.​ADVEN​GSOFT.​2013.​12.​007
Novák P, Čabelka J (1981) Models in Hydraulic Engineering: Physical Principles and Design Applications.
Monographs & surveys in water resources engineering
Oertel M, Peterseim S, Schlenkhoff A (2011) Drag coefficients of boulders on a block ramp due to interac-
tion processes. J Hydraul Res 49:372–377. https://​doi.​org/​10.​1080/​00221​686.​2011.​565695
Ohadi S, Hashemi Monfared SA, Azhdary Moghaddam M, Givehchi M (2023) Feasibility of a novel predic-
tive model based on multilayer perceptron optimized with Harris hawk optimization for estimating of
the longitudinal dispersion coefficient in rivers. Neural Comput Appl 35:7081–7105. https://​doi.​org/​
10.​1007/​s00521-​022-​08074-8
Pagliara S, Chiavaccini P (2006) Energy dissipation on block ramps. J Hydraul Eng 132:41–48. https://​doi.​
org/​10.​1061/​(ASCE)​07339​429(2006)​132:​1(41)
Pagliara S, Palermo M, Lotti I (2009) Sediment transport on block ramp: filling and energy recovery. KSCE
J Civ Eng 13:129–136. https://​doi.​org/​10.​1007/​s12205-​009-​0129-1
Pagliara S, Palermo M, Roy D (2020) Experimental investigation of erosion processes downstream of
block ramps in mild curved channels. Environ Fluid Mech 20:339–356. https://​doi.​org/​10.​1007/​
s10652-​019-​09681-1
Rad MJG, Ohadi S, Jafari-Asl J et al (2022) GNDO-SVR: an efficient surrogate modeling approach for reli-
ability-based design optimization of concrete dams. Structures 35:722–733. https://​doi.​org/​10.​1016/j.​
istruc.​2021.​11.​048
Rahmanshahi M, Shafai Bejestan M (2020) Gene-expression programming approach for development of a
mathematical model of energy dissipation on block ramps. J Irrig Drain Eng 146:4019033. https://​doi.​
org/​10.​1061/​(ASCE)​IR.​1943-​4774.​00014​42
Salmasi F, Sattari MT, Nurcheshmeh M (2021) Genetic programming approach for estimating energy dis-
sipation of flow over cascade spillways. Iran J Sci Technol Trans Civ Eng 45:443–455. https://​doi.​org/​
10.​1007/​s40996-​020-​00541-3
Tamagni S, Weitbrecht V, Boes RM (2014) Experimental study on the flow characteristics of unstructured
block ramps. J Hydraul Res 52:600–613. https://​doi.​org/​10.​1080/​00221​686.​2014.​950610
Wang Z, Ma G, Gong D et al (2019) Application of mind evolutionary algorithm and artificial neural net-
works for prediction of profile and flatness in hot strip rolling process. Neural Process Lett 2019 503
50:2455–2479. https://​doi.​org/​10.​1007/​S11063-​019-​10021-Z
Weitbrecht V, Tamagni S, Boes RM (2017) Stability of unstructured block ramps. J Hydraul Eng
143:4016095. https://​doi.​org/​10.​1061/​(ASCE)​HY.​1943-​7900.​00012​59

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and applicable
law.

13
A Hybrid Model for Predicting the Energy Dissipation on the Block…

Authors and Affiliations

Mostafa Rahmanshahi1 · Jafar Jafari‑Asl2 · Mahmood Shafai Bejestan3 ·


Seyedali Mirjalili4,5
Mostafa Rahmanshahi
mostafa.rahmanshahi@polyu.edu.hk
Mahmood Shafai Bejestan
m_shafai@yahoo.com
Seyedali Mirjalili
ali.mirjalili@gmail.com
1
Department of Civil and Environmental Engineering, The Hong Kong Polytechnic University,
Hong Kong, SAR, China
2
Department of Civil Engineering, Faculty of Engineering, University of Sistan and Baluchestan,
Zahedan, Iran
3
Faculty of Water and Environmental Engineering, Shahid Chamran University of Ahvaz, Ahvaz,
Iran
4
Centre of Artificial Intelligence Research and Optimisation, Torrens University, Brisbane,
Australia
5
Yonsei Frontier Lab, Yonsei University, Seoul, Republic of Korea

13

You might also like