Sei sulla pagina 1di 12

Applied Soft Computing 69 (2018) 165–176

Contents lists available at ScienceDirect

Applied Soft Computing


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

Estimating incipient motion velocity of bed sediments using different


data-driven methods
Mohammad Zounemat-Kermani ∗ , Amin Mahdavi Meymand, Mina Ahmadipour
Department of Water Engineering, Shahid Bahonar University of Kerman, Kerman, Iran

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

Article history: In the present research, the data-driven methods (DDMs), are used to estimate the threshold veloc-
Received 23 August 2017 ity of sediment motion. Results of the DDMs used in this research, including artificial neural networks
Received in revised form 14 April 2018 (FFNN & RBNN), adaptive neuro-fuzzy inference system models (ANFIS, ANFIS-GA & ANFIS-IWO), and
Accepted 22 April 2018
wavelet neural network (WaveNet), are compared with those of the mathematical models and exper-
Available online 26 April 2018
imental observations. The obtained results indicate that the WaveNet model with the Nash–Sutcliffe
coefficient of 0.997 has better performance than the other methods. Moreover, in order to specify the rel-
Keywords:
ative importance of the input parameters for the uncertainty of the threshold velocity, sensitivity analysis
Threshold velocity
k-fold cross-validation
is performed, the results of which indicate that the median diameter of the particles and relative density
Invasive weed optimization are the most important parameters affecting the threshold velocity, respectively. In addition, the Monte
Genetic algorithm Carlo simulation is used to quantify the uncertainty of the threshold velocity of motion. The uncertainty
Monte Carlo is expressed using the coefficient of variation (CV). The highest amount of CV is related to the median
diameter of grain size, therefore, this parameter has the maximum effect on variations of the incipient
motion.
© 2018 Published by Elsevier B.V.

1. Introduction the conditions, the flow within the constructs can be designed such
that precipitation of the sedimentary particles is prevented.
1.1. Background
1.2. Literature review
Sediment incipient velocity is known as one of the most impor-
tant parameters for the sediment movement procedure and fluvial Incipient motion can be obtained using theoretical methods
processes in the river engineering. In short, the onset of the motion and dimensional analysis. Since the forces exerted by the fluid
of sediment particles, which can be visible by eyes is known as depend on various parameters, the appropriate equation is com-
the sediment incipient motion. The threshold motion, which leads monly obtained using dimensional analysis along with laboratory
to the movement of sediment grains, takes place when the flow results. Most of the incipient motion criteria have been extracted
velocity exceeds the critical value of incipient velocity [1]. In the from either shear stress or critical velocity.
discussion of sediment hydraulics, it is necessary to predict the So far, numerous studies have been conducted in order to obtain
threshold conditions of particles’ motion. Sedimentation in aquatic the incipient motion of the sediments using dimensional analy-
constructs causes various problems such as reduced capacity of sis and theoretical methods [3,4,5]. Since the forces exerted by
reservoir, channel cross-section and in general, reduced efficiency fluid depend on various parameters, the appropriate equation is
of the hydraulic structure [2]. Besides, sediment precipitation in commonly obtained using dimensional analysis and theoretical
water transfer constructs leads to the reduced flow cross-section methods along with the laboratory results. The majority of the
as well as increased coarseness and, as a result, reduced water- aforementioned studies are based on Shields’ fundamental work
yield of the structure. Incipient motion is of great importance in and providing experimental equations fitted to Shields’ curve [6,7].
designing sustainable watercourses because, in case of having all As mentioned above, the process of incipient motion of sedi-
ments is very complex, thereafter, applying alternative non-linear
methods such as data driven methods (DDMs) can be appropriate
approaches for simulating this process.
∗ corresponding author. DDMs are a good instrument for prediction and estimation;
E-mail address: zounemat@uk.ac.ir (M. Zounemat-Kermani). besides, regarding the limitations of the experimental methods,

https://doi.org/10.1016/j.asoc.2018.04.041
1568-4946/© 2018 Published by Elsevier B.V.
166 M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176

DDMs can be a suitable alternative for estimating the threshold


velocity of motion. Using DDMs in determining complex phenom-
ena in different fields of engineering and especially in hydraulics
engineering, can increase the accuracy and performance of the
simulating process. In the following, examples of the effective
applications of these methods in determining the threshold veloc-
ity of sediment are presented. Sreenivasulu et al. [8] applied radial
base neural networks (RBNN) for estimating the incipient veloc-
ity of sediments and the results were accepTable Sehrawat et al.
[9] used Gaussian process regression (GPR) model for estimating
incipient motion of alluvial channel flow and reported satisfac-
tory results. Safari et al. [2] used three different neural networks
(the feed-forward back propagation (FFNN), generalized regression
(GRNN), and radial basis function (RBNN)) to determine the thresh-
old conditions. The results confirmed the proper performance of
these methods.

1.3. Purpose, rationale and contribution

The present study is aimed to estimate the threshold velocity


of the bed sediments’ motion using several DDMs. The applied
DDMs in this study include multi-layer ANN (feed forward back-
propagation, FFNN), radial basis ANN (RBNN), three types of
adaptive neuro-fuzzy inference system [classic ANFIS; integration
of ANFIS and genetic algorithm (ANFIS-GA) & integration of ANFIS
and invasive weed optimization (ANFIS-IWO)], and wavelet neu-
ral network (WaveNet). The efficiency of these models can be
taken into consideration to estimate the appropriate design veloc-
Fig. 1. Sketch of hydraulic conditions around particles on the bed of a channel or
ity in irrigation and transmission channels. Moreover, the threshold river: (a) streamlines around the particle; (b) forces exert to the particle.
velocity of motion is calculated using the experimental methods
presented by Isbach [10] and May [11]. The performance of the
experimental equations and models has been evaluated using the
statistical criteria, NRMSE (normalized root mean square error), movement” are some of the specific terms used by researchers in
MAE (mean absolute error), NSE (Nash-Sutcliffe coefficient of effi- this field [12]. Some researchers consider the onset of the motion
ciency), and determination coefficient (R2 ). of one or more particles, which is visible by eyes, as the particle’s
The sedimentary particle’s motion onset depends on the param- incipient motion [13], while some others consider the particle’s
eters that have uncertainty, leading to the uncertainty of the motion onset as the conditions of a flow, in which the value of
threshold velocity of motion. Such uncertainty might be due to the sedimentary material’s motion toward the channel downstream is
precise quantification of the parameters or measurement errors; insignificant or equal to zero. Increasing the intensity of water flow
thus, in the present study, the uncertainty of the input parame- on the sedimentary bed gradually leads to a status, in which the
ters was also investigated. Accordingly, first, the input parameters hydrodynamic forces exerted by the fluid flow (drag and lift) on
affecting the model’s output were specified using sensitivity anal- the particles exceed the sustainable force caused by the particles’
ysis through simple correlation coefficient ranking method and, weight and, consequently, results in the particles’ motion. Trans-
then, the uncertainty analysis of the parameters of median diame- mission from a state that particles are motionless to the state of
ter and relative density of the particles was performed using Monte primary motion of the sediments is defined as the critical condition
Carlo simulation. In the past researches, WaveNet, ANFIS-GA and or incipient motion.
ANFIS-IWO models have not been applied and compared with When the flow passes through a sediment particle located on
other DDMs and mathematical models for simulating the incipi- the bed of a channel or river, external forces are imposed on the
ent motion of bed sediments. Hence, the contribution of this study particle. Resultant of forces will be the main factor of the starting
is to challenge the applied DDMs in simulating sediment motion point for the movement of a particle. Fig. 1(a) shows a schematic
process. In addition, no similar study has been found to analyze view of streamlines around a sediment particle on the bed of a
the uncertainty of the effective parameters of incipient sediment channel/river. Streamlines are separated from the bed by passing
motion using Monte Carlo method. through the particle and created vortex region at the backside of
the particle. The water pressure at the backside of the particle is
2. Materials and methods higher than pressure at the separation point of the streamlines.
The difference amount of pressure on the opposite sides of the par-
2.1. Incipient motion of sediment particles ticle causes the drag force (FD ) which might be large enough to roll
the particle [14]. Other forces exert to the particle include the lift
Due to the arbitrary nature of sediment motion along the allu- (FL ), the buoyancy (FB ), weight due to gravity (W) and the seepage
vial bed, it is very difficult to exactly determine the conditions of (Se ) forces. A definition sketch of these forces is shown in Fig. 1(b)
a flow, in which a sedimentary particle is moved; therefore, such [15,16,17,18]. In theory, when the resultant actuator forces equal
conditions are more or less dependent on the researcher’s defini- to the resultant inhibitory forces, sediment particles would be at
tion of incipient motion. In other words, definition of the particle’s the threshold of motion. However, as mentioned earlier, determin-
incipient motion or particle’s motion onset is not the same among ing the incipient motion of sediment particles is a very complicated
all the researchers. “Initial motion”, “grain moving”, and “critical and intricate process in hydraulics of sediment transport [16].
M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176 167

2.2. Data-driven methods activation functions of neurons of the hidden layer [31,32]. Details
about the WaveNet structure and operation can be found at [29,33].
2.2.1. Artificial neural network (ANN)
ANN is a data processing system, which imitates the human 2.3. Mathematical models
brain’s behavior by imitating the performance and relationship of
neurons. ANNs indicate nonlinear and complicated functions with The mathematical models that have been proposed so far are
numerous parameters so that the network’s outputs are adjusted mainly based on the theoretical analysis of the laboratory data. The
similar to the measured outputs within the identified datasets. mathematical models used in the present study are mentioned in
Neural network requires a considerable amount of data for train- this section.
ing; accordingly, as soon as achieving the satisfactory training, the
network would be able to obtain the output for the previous non- 2.3.1. Isbach’s equation
observed inputs [19,20]. This method is based on the laboratory data. Isbach presented
All the ANN models proposed over the years are common in an equation for incipient motion as follows [10]:
a structural block known as neuron (processor) and interconnec-   −   0.5
tion of the network. The main differences between various types of s w
Vc = 1.2 2g − 1 D50 (1)
ANNs include network architecture, weight updating methods, and w
their input functions [19]. Several types of ANNs can be considered where Vc , s , w , and D50 indicate threshold velocity of motion, den-
for simulating a hydraulic process. In this study two of the most sity of particles, density of water, and average grain size or median
common and well-known ANNs called multi-layer perceptron neu- diameter, respectively.
ral network (also known as feed forward back propagation neural
network, FFNN) and radial basis neural network (RBNN) have been
2.3.2. May equation
used for estimating the incipient motion velocity of bed sediments.
In 2003, May conducted some investigations on the precipita-
More details about the ANNs can be found at [21,22,23,24].
tion of sediments in an inverted siphon. By constructing a physical
model and performing the experiments, he presented some equa-
2.2.2. Adaptive neuro-fuzzy inference system tions, one of which was the equation of sediment incipient motion,
The ANFIS model takes into account the capabilities of both as follows [11]:
fuzzy theory and neural network which embraces the two neural
VC
 h 0.47
and fuzzy models benefits [25]. By exploiting the training power of = 0.125 (2)
neural networks and lingual advantage of fuzzy systems and also [g (Gs − 1) D50 ]
0.5 D50
using the advantages of the two models for analyzing the compli- In the above relation h stands for the water depth.
cated processes, this system has powerful performance. The fuzzy
inference system is based on the if-then rules so that, by using
2.4. Cross-validation estimates of performance
these rules, the relationship between a number of input and output
variables can be obtained. This system can be also used as a pre-
Cross-validation is a technique in which the entire dataset is
dictor model for the conditions, in which the input or output data
used as training and testing sets.
have high levels of uncertainty. The fuzzy modeling steps include
In the k-fold cross-validation technique, the training process is
determining the membership functions, determining the inference
applied k times considering a fraction of 1/k of the training data
system based on the data, writing the inference rules, combining
left out for the testing phase [34]. In other words, data are divided
them, obtaining the result, and, if necessary, defuzzification. For
into k parts, k-1 parts of which are used as the training set and one
more information about the ANFIS structure, the reader is referred
part is used as the testing set (Fig. 2). Then, the procedure of cross-
to [25,26].
validation is repeated k times. Moreover, modeling is performed
In this study, along with the basic hybrid learning algorithm of
using the training set and performance of the model is assessed
ANFIS, two other meta-heuristic methods, namely invasive weed
using the data of the testing set. Mean of the results of k clusters
optimization and genetic algorithm (ANFIS-IWO & ANFIS-GA), have
is considered as the final estimation. The advantage of this method
been used. Detailed information about these optimization methods
is that all the data have the chance for being trained and assessed
can be found at [27,28].
[35].

2.2.3. WaveNet 2.5. Uncertainty


Simultaneous with the works on modeling neural networks,
considerable progress has been made in developing the math- Uncertainty can be considered as a feature of system that
ematical tools for approximation of function; for example, the describes deficiency of the human’s knowledge on a system and its
wavelet theory and multi-resolution analysis progressed signifi- progress status. Analysis of uncertainty for any phenomenon would
cantly. Furthermore, Wavelet analysis is a powerful instrument not only reveal a semi-real perspective of its uncertain parameters,
for demonstrating the functions; moreover, the combination of but also it would facilitate providing close-to-reality estimation as
Wavelet and ANN creates a new pattern of artificial intelli- well as making decisions on that phenomenon.
gence, called WaveNet, which can be a good alternative for So far, various methods have been introduced and developed
back-propagation neural networks, since it can apply any given for analyzing the uncertainty dominating the aquatic systems, the
nonlinear function. WaveNet was first introduced by Bakhshi and most applicable of which are the first-order analysis and fuzzy and
Stephanopolous [29]. It is defined as a multi-resolution hierarchi- Monte Carlo simulation.
cal neural network with a hidden layer of nodes and local learning, First-order analysis method is based on the linearization of the
the basic functions of which were selected from a group of single functional relations, which describes a dependent random variable
orthogonal wavelets. WaveNet usually has the form of three-layer and a set of independent random variables by expanding Tay-
network. The lower layer represents the input layer, the middle lor series. This method has been widely used in the environment
layer is the hidden layer and the upper layer is the output layer and water engineering problems with uncertainty. In Monte Carlo
[30]. In the WaveNet structure, wavelet functions are used as the simulation, random inputs are produced with their probabilistic
168 M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176

Fig. 2. Schematic figure of data partitioning for k-fold cross-validation.

Table 1
Data sources used in the present research.

Data set Reference Bed materials Number of data Reynolds number range Flume width (m) Froude number range

1 Dey & Raju [40] gravel and coal 48 [10773,27745] 0.51 [0.776, 2.992]
2 Dey & Raikar [41] gravel 54 [65750,224250] 0.9 [0.83, 1.09]
3 Afzalimehr et al. [42] gravel 20 [74194,173843] 0.6 [0.23, 0.68]
4 Dey et al. [43] gravel 4 [68600,177100] 0.6 [0.42, 0.61]
5 Brownlie [44] sand 25 [2577,15079] 0.1575 [0.37, 073]
gravel 26 [21370,140070] 2.44 & 0.61 [0.14, 0.31]
Gravel & sand 51 [20415,195730] 0.9 [0.018, 1.52]
sand 12 [11021,107113] 1.372 [0.38, 0.201]
sand 16 [3930,24088] 0.25, 0.5 & 0.75 [0.4, 0.97]

distributions and, then, given to the theory or mathematical mod- ingly, the bigger the absolute value of simple correlation coefficient
els in order to obtain the random model’s output. Afterwards, the of the random input, the higher the importance of that random
produced outputs are statistically analyzed to quantify the out- input for the output uncertainty would be [39].
put uncertainty. There are numerous cases of uncertainty analysis
using Mont Carlo simulation in the field of environment and water 2.7. Data presentation
engineering [36,37,38].
In estimating the threshold velocity of motion, there are several Data used in the present study were collected and clustered from
factors with uncertainty. The factors with different uncertainties the investigated papers on this field, which included the laboratory
that affect the threshold velocity of motion are called random data on the incipient motion. Table 1 summarizes the data sources
inputs. In this study, the uncertainty of the threshold velocity of used in this research.
motion has been quantified using Monte Carlo simulation [39]. Experiments were performed in rectangular flumes with dif-
ferent dimensions and sedimentary bed using uniform materials;
2.6. Sensitivity analysis accordingly, first, the sedimentary bed was saturated mildly with
water and, then, the flow rate or the channel slope was adjusted to
Sensitivity analysis provides valuable information about the establish the initial motion (incipient motion) conditions.
level of model’s sensitivity to the input parameters so that, identi- In the present study, the input parameters of the DDMs for
fying the effect rate of the input parameters on the precision of the estimating the threshold velocity of motion included the mean
output prediction would lead to the identification of the effective diameter of sediment particles d50 , water depth h, bed width B,
parameters. When a large number of input parameters are used slope of watercourse bed S0 , and density of sediment particles Sg .
to predict the output, sensitivity analysis is performed to specify In general, using wide range of different sources of data can chal-
the degree of effectiveness of each random input for the output lenge the ability of the intended models. Hence, in this paper, 9
uncertainty; in other words, sensitivity analysis selects the ran- series of data from 5 experimental studies were gathered. It is worth
dom inputs with maximum importance for the output uncertainty. mentioning that all the gathered data were obtained in rectangu-
Therefore, the number of random inputs is reduced and, thus, a lar sections of the experimental flumes with cohesionless particles.
more perfect sensitivity analysis can be performed based on the Also, 256 data used in the present research included a wide range
Monte Carlo simulation [39]. of almost uniform natural and synthetic sedimentary particles with
Effect of the random inputs on the output can be measured by the mean diameter of 0.1 < d50 <35 mm and particle specific gravity
ranking the simple correlation coefficient between the jth input vec- of 1.3 < Sg < 2.7. Furthermore, these data were related to the uniform
(j) (j) water flows and smooth bed conditions.
tor (x1 , ..., xi ) and the corresponding output vector (y1 , ..., yi ). The
Parameters such as mean diameters of particles, water depth,
simple correlation coefficient between the jth random input and
and width and slope of the bed floor are called random parame-
output is calculated as follows:
ters. The statistical features of the data used in this research are

n
(j) (j)
presented in Table 2.
(xi − 
ˆ x )(yi − 
ˆ y)

SCC(x(j) , y) =
i=1 3. Models application and results
 0.5 (3)

n
(j) (j) 2

n
2
(xi − 
ˆx ) (yi − 
ˆ y) 3.1. Implementation of the methods
i=1 i=1
Threshold velocity of motion was modeled using the mathe-
j
where  ˆ x, 
ˆ y , and n indicate the mean random input sample of matical models, ANN, ANFIS, and WaveNet. Initially, to perform
xj , mean output sample of y, and sample size, respectively. Accord- modeling, the available data were sorted completely randomly;
M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176 169

Table 2 Table 3
Statistical features of the used parameters. Statistical evaluation of estimating threshold velocity of motion using mathematical
models.
Input parameters Minimum Maximum Average Standard
deviation Mathematical model Error measurement criteria

Median particle diameter (mm) 0.19 31.75 7.68 6.76 NRMSE MAE NSE R2
Water depth (m) 0.0119 0.61 0.117 0.089
Bed width (m) 0.1575 2.438 0.836 0.532 Isbach 0.556 0.370 −0.813 0.721
Channel slop 0 0.027 0.0055 0.005 May 0.81 0.52 −2.67 0.301
Particle density 1.36 2.67 2.6 0.43
Threshold flow velocity (m/s) 0.049 1.217 0.583 0.291
using the aforementioned statistical criteria, the results of which
are presented in Table 3.
moreover, modeling was performed using the training data and Accordingly, the closer to 0 the values of NRMSE and MAE and
also the model’s performance was assessed using the testing data. the closer to 1 the values of determination coefficient and NSE, the
Moreover, k-fold cross-validation was carried out by separating more desirable the performance of the estimated method would be.
the data into five clusters so that four clusters (80%) were used In the phase of testing the mathematical models, the method pro-
for training and the remaining cluster was used for testing. The posed by Isbach with NRMSE of 0.556, MAE of 0.370, NSE of −0.813,
training procedure was repeated five times; every time, one of the and determination coefficient of 0.721 had a better performance
clusters was used as the testing set. Finally, the results of the five compared to May’s method. In order to reveal the performance
models were averaged. The advantage of the k-fold cross-validation of the mathematical models in the testing phase, scatter plot was
technique is that all the data contribute to training and testing. plotted for the methods.
Performance of the used models was assessed using multiple Regarding the scatter plots plotted for testing data in Fig. 3, it can
statistical tests (criteria), which describe the model-related errors. be seen that the points of scatter plot in Isbach’s method (Fig. 3(b))
In the present study, four statistical criteria, including NRMSE, MAE, are closer to the 1:1 agreement line; therefore, it has better per-
NSE, and R2 , were used, the equations of which are as follows. formance than May’s mathematical model. On the other hand, the
Among these criteria, NRMSE and MAE demonstrated better per- points of scatter plot of May’s mathematical model are more dis-
formance of the results (the closer to zero, the better). The value of tant from the 1:1 agreement line, indicating its poor performance
NSE ranges between negative infinity and unity so that the closer to in estimating the threshold velocity of motion. According to the
unity, the more precise the model would be. Furthermore, the closer results of Fig. 3, both mathematical models have a tendency to
to unity the value of determination coefficient, the more precise the under-estimate the simulated results.
estimations of the model would be. The formulation of these criteria
is written in the following: 3.2.2. The FFNN model


 n To estimate the threshold velocity of motion of the bed sedi-

ments, the feed forward back-propagation multi-layer ANN (FFNN)

(yio − yip )2

was used; for this purpose, several different structures were applied

i=1
NRMSE =
(4) for the multi-layer ANN. In each of these structures, different

 n
(y )2
stimulant functions, various weight updating methods, as well as
io
different numbers of neurons were used. Fig. 4 shows the dia-
i=1
gram of NRMSE variations of the testing data for different numbers

1
n of neurons for the first cluster of the data neural network with
MAE = |yio − yip | (5) the stimulant function of logsig and conjugate gradient weights
n
i=1 updating. Moreover, for different iterations, up to maximum 500
simulations were repeated. Then, in each structure, the best itera-

n
tion was considered as the result of that structure with regard to
(yio − yip )2
the NRMSE criterion in order to specify the best performance of
i=1 each structure. Modeling the FFNN was performed with five dif-
NSE = 1 − (6)

n
ferent clusters (folds) of training data; consequently, the modeling
2
(yio − ȳip ) results for the optimum number of neurons and different types of
i=1 training algorithms were determined for all the five clusters of the
n 2 testing data. The statistical criteria were calculated for the cluster

(yio − ȳo )(yip − ȳp ) of training data as well as modeling test.
In the present study, the two hyperbolic tangent-sigmoid
i=1
R2 = (7) (tansig) and log-sigmoid (logsig) stimulant functions, the linear

n

n
stimulant function (purelin), and the two Levenberg-Marquardt
(yio − ȳo )2 (yip − ȳp )2
and conjugate gradient learning algorithms were used for the
i=1 i=1
hidden layer, output layer, and weights updating, respectively.
where yio , yip , n, ȳo , and y indicate the observed (measured) val- Table 4 shows the performance of each structure for any of the
ues, predicted values, number of data, mean observational data, five data clusters with logsig stimulant function and conjugate
and mean predicted data, respectively. Results and analysis of each gradient learning algorithm (traincgb); furthermore, the average
method will be presented in the following sections. performance of each cluster of cross-validation of each structure is
presented in Table 5. According to the statistical evaluation pre-
3.2. Analysis and results sented in Table 5, it can be seen that the Levenberg-Marquardt
learning algorithm (trainlm) had a slightly better performance in
3.2.1. The mathematical models comparison to the conjugate gradient algorithm.
In this section, the threshold velocity of motion would be esti- In order to specify the performance of each network, scatter
mated using the mathematical equations by Isbach [10] and May plots of each structure for the cluster of the first 20% of the test-
[11]. Performance of each of these two methods was evaluated ing data are plotted in Fig. 5. According to Fig. 5, the majority of
170 M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176

Fig. 3. Scatter plots for testing data of mathematical models, (a) May’ method and (b) Isbach’s method.

Table 4
Evaluation results of the multi-layer perceptron neural network with stimulant function logsig and conjugate gradient weights training algorithm.

Number of fold (cluster) Activation function (Middle layer) Learning algorithm Train set Test set

NRMSE MAE NSE R2 NRMSE MAE NSE R2

First logsig traincgb 0.020 0.010 0.998 0.998 0.078 0.043 0.97 0.975
second logsig traincgb 0.040 0.020 0.991 0.992 0.139 0.082 0.90 0.978
Third logsig traincgb 0.042 0.022 0.991 0.991 0.072 0.039 0.97 0.982
Fourth logsig traincgb 0.043 0.022 0.990 0.990 0.202 0.109 0.82 0.972
Fifth logsig traincgb 0.022 0.011 0.997 0.997 0.09 0.043 0.95 0.964
Average logsig traincgb 0.033 0.017 0.993 0.993 0.116 0.062 0.92 0.973

Table 5
Statistical performance results of the multi-layer perceptron ANN (FFNN) structures.

ANN Features Training phase Testing phase


a 2
Activation function (Middle layer) Learning algorithm NRMSE MAE NSE R NRMSE MAE NSE R2

logsig traincgb 0.033 0.017 0.991 0.993 0.116 0.062 0.921 0.973
logsig trainlm 0.033 0.015 0.990 0.992 0.119 0.066 0.920 0.973
tansig traincgb 0.040 0.019 0.988 0.989 0.127 0.072 0.911 0.975
tansig trainlm 0.016 0.0075 0.995 0.996 0.118 0.066 0.926 0.983
a
Note: traincgb: conjugate gradient learning algorithm; trainlm: Levenberg-Marquardt training algorithm.

the points of the scatter plot are above the 1:1 agreement line,
indicating the over-estimation of the multi-layer neural network
method.

3.2.3. The RBNN model


Similar to the multi-layer ANN (FFNN), the five different clus-
ters of the training and testing data were created and the threshold
velocity of each cluster was estimated for different radii of the radial
basis ANN. As a result, the best value of radius was determined for
each data cluster and the arithmetic mean of the statistical crite-
ria of five categories was considered as the final performance of
the network. Table 6 represents the results of the radial basis ANN Fig. 4. NRMSE of test data for different numbers of neurons (set of the first 20% of
(RBNN) obtained by averaging the results of the best radius for the the tesingt data).
five testing and training clusters.
The NRMSE of the testing data for different radii for the cluster
of the first 20% is plotted in Fig. 6. As seen in Fig. 6, the minimum 3.2.4. The ANFIS model
value of the NRMSE per radius is 0.1. The scatter plots of the testing In the present research, the Sugeno fuzzy inference system was
and training data for the cluster of the first 20% for the best radius produced using fuzzy c-means clustering. The sub-cluster algo-
are plotted in Fig. 7. rithm with the radius of 0.5 as well as maximum and minimum
Fig. 7(a) shows that the points of the scatter plot of the train- input and output specified the number of clusters; furthermore,
ing data are close to the 1:1 agreement line, while the points of the center of each cluster was specified using the decreasing clus-
the scatter plot of the testing data are farther from the 1:1 agree- tering method. The number of clusters determined the number of
ment line. However, the scatter plot of the testing data is acceptable. rules as well as the membership function. The simple GAUSSMF
According to Fig. 7, the radial basis ANN shows an over-estimation (Gaussian membership functions) and linear membership function
characteristic. were considered for the input and output parameters, respectively.
M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176 171

Fig. 5. Scatter plots of testing data for the cluster of the first 20%, FFNN neural network: (a) stimulant function logsig and traincgb weights updating metho; (b) stimulant
function logsig and trainlm weights updating method; (c) stimulant function tansig and traincgb weights updating method; (d) stimulant function tansig and trainlm weights
updating method.

Table 6
Evaluation results of the radial basis neural network.

Number of fold (cluster) Train Test

NRMSE MAE NSE R2 NRMSE MAE NSE R2

First 0.022 0.0477 0.996 0.997 0.129 0.158 0.906 0.980


second 0.022 0.0467 0.996 0.997 0.178 0.206 0.840 0.994
Third 0.022 0.0485 0.996 0.997 0.107 0.138 0.934 0.983
Fourth 0.018 0.0435 0.994 0.998 0.154 0.157 0.907 0.992
Fifth 0.021 0.0520 0.996 0.997 0.117 0.180 0.931 0.966
Average 0.021 0.0478 0.996 0.997 0.136 0.168 0.903 0.983

Table 7
Inference methods of the applied ANFIS model used in the present research.

Inference Method Operator

And Prod
Or Prober
Implication prod
Aggregation sum
Defuzzification watver

tion of the least squares and back propagation decreasing gradient


methods was used to train the parameters of the fuzzy inference
system functions in order to model a set of input and output data.
The inference method used in the adaptive neuro-fuzzy model is
Fig. 6. NRMSE of testing data for different radii (set of the first 20% of testing data).
represented in Table 7.
The threshold velocity of motion was evaluated for all the
Moreover, the adaptive neuro-fuzzy network applied the com- five clusters of 5-fold cross-validation of data using the ANFIS,
bined learning algorithm to determine the parameters of the ANFIS-IWO and ANFIS-GA. Table 8 represents the values of the
single-output membership functions; in other words, the combina- statistical criteria for the ANFIS models, indicating that these mod-
172 M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176

Fig. 7. Scatter plot of radial basis network for the set of the first 20% of testing data: (a) training data; (b) testing data.

Table 8
Evaluation results of the ANFIS, ANFIS-IWO & ANFIS-GA models.

method Number of fold (cluster) Train Test


2
NRMSE MAE NSE R NRMSE MAE NSE R2

ANFIS First 0.02 0.0066 0.997 0.998 0.041 0.014 0.99 0.991
second 0.057 0.018 0.983 0.986 0.159 0.0388 0.867 0.922
Third 0.043 0.0123 0.990 0.991 0.046 0.0169 0.988 0.988
Fourth 0.023 0.0088 0.997 0.997 0.036 0.0141 0.994 0.995
Fifth 0.019 0.0077 0.998 0.998 0.026 0.0104 0.996 0.996
Average 0.033 0.010 0.993 0.994 0.062 0.019 0.967 0.979
ANFIS-IWO First 0.071 0.035 0.975 0.975 0.073 0.041 0.971 0.972
second 0.071 0.036 0.975 0.975 0.071 0.038 0.974 0.974
Third 0.074 0.038 0.973 0.973 0.07 0.038 0.974 0.977
Fourth 0.072 0.038 0.973 0.973 0.102 0.043 0.959 0.957
Fifth 0.07 0.037 0.975 0.975 0.076 0.037 0.974 0.974
Average 0.072 0.037 0.974 0.974 0.079 0.039 0.971 0.971
ANFIS-GA First 0.027 0.012 0.996 0.996 0.038 0.021 0.992 0.992
second 0.023 0.011 0.997 0.997 0.035 0.017 0.994 0.994
Third 0.026 0.013 0.997 0.997 0.032 0.016 0.995 0.995
Fourth 0.026 0.012 0.997 0.997 0.052 0.018 0.99 0.989
Fifth 0.023 0.011 0.997 0.997 0.03 0.015 0.996 0.996
Average 0.025 0.012 0.997 0.997 0.037 0.017 0.993 0.993

Fig. 8. Scatter plot of ANFIS-GA model for the cluster of the first 20%: (a) training data; (b) testing data.

els have good performance in both training and testing phases. 3.2.5. The WaveNet model
As can be seen, the best performance is achieved by the ANFIS- In this section, the estimated results obtained using the
GA model. The scatter plots of the ANFIS-GA is shown in Fig. 8. WaveNet are presented. Similar to the previous sections, the
According to the plot, the ANFIS-GA is indifferent for over- threshold velocity of motion was estimated for five different clus-
estimation and under-estimation because the point of the scatter ters of the training and testing data. In this network, the Gaussian
plot is dispersed quite equally at both sides of the 1:1 agreement wavelet was used and the simulation was performed for the
line. precisions of m = 0, −1, −2. Furthermore, the best precision was
M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176 173

Table 9
Evaluation results of the WaveNet model.

Number of fold (cluster) Train Test

NRMSE MAE NSE R2 NRMSE MAE NSE R2

First 0.0026 0.0010 0.999 0.999 0.022 0.007 0.997 0.997


second 0.0025 0.0011 0.999 0.999 0.027 0.0096 0.996 0.996
Third 0.0022 0.0009 0.999 0.999 0.019 0.0068 0.998 0.998
Fourth 0.0028 0.0012 0.999 0.999 0.034 0.0106 0.994 0.994
Fifth 0.0024 0.0011 0.999 0.999 0.018 0.0063 0.998 0.998
Average 0.0025 0.001 0.999 0.999 0.024 0.008 0.997 0.997

Fig. 9. Scatter plot of WaveNet for the cluster of the first 20%: (a) training data; (b) testing data.

Table 10
Results of the sensitivity analysis of input parameters.

Parameter Median grain size Water depth Bed width Bed slope Relative density of particles

Simple correlation coefficient 0.46860 0.00001 −0.0048 0.00026 0.1555

determined per data cluster and the arithmetic mean of the sta- ity of motion was specified using sensitivity analysis. In this section,
tistical criteria of the five clusters was considered as the final the uncertainty analysis of the threshold velocity of motion of
performance of the network (Table 9). the bed sediments is performed with regard to the model’s input
According to the scatter plots of the WaveNet, it is seen that parameters using Monte Carlo simulation.
all the points of the diagrams were placed on the bisector (the 1:1 First, the random input involved in uncertainty was produced
agreement line), indicating high power of the network for estimat- based on the Monte Carlo simulation and other random parame-
ing the threshold velocity of motion (Fig. 9). ters were considered constant. For instance, when uncertainty was
caused by the median diameter of sediment particles, this param-
3.3. Sensitivity analysis eter would be produced using Monte Carlo simulation; thus, the
produced random input would be used for estimating the output.
In this section, the sensitivity analysis was performed using sim- Besides, the predicted outputs would be used for statistical analysis.
ple correlation coefficient. Table 10 represents the values of simple In this section, regarding the obtained results of the sensitivity
correlation coefficient values for the input parameters. analysis (see Table 10), the uncertainty analysis of the thresh-
According to Table 10, the mean diameter of particles, relative old velocity of motion was investigated considering the two most
density, bed width, bed slope, and water depth had maximum abso- effective input parameters of the median diameter of particles
lute values of simple correlation coefficient, respectively. Results (correlation coefficient = 0.47) and relative density of particles (cor-
of sensitivity analysis indicated that the median size of sediment relation coefficient = 0.16).
particles, relative density, bed width, bed slope, and water depth First, the median diameter of particles was produced as the
were the most important determinant parameters of the threshold random input involved in the uncertainty of threshold velocity
velocity of motion, respectively. In the present study, the effect of of motion based on the Monte Carlo simulation. By the use of
uncertainty of the median size of sediment particles and density of Monte Carlo simulation, 1000 random data with normal distribu-
particles on the threshold velocity of motion of the sediments was tion, mean of 7.68, and standard deviation of 0.17 were produced
investigated. for the median diameter of particles. Further, other inputs were
considered constant. Then, the median diameter of the particles
produced based on normal distribution was used to estimate the
3.4. Uncertainty
threshold velocity of motion through WaveNet method.
Afterwards, considering the relative density of particles as the
Inability to quantify as well as lack of sufficient precision in mea-
random input involved in the uncertainty, all the above-mentioned
suring the parameters would result in uncertainty in the model’s
steps were similarly performed on this parameter as well; sub-
input parameters for estimating the threshold velocity of motion.
sequently, 1000 random data with normal distribution, mean of
Relative importance of the inputs in estimating the threshold veloc-
174 M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176

Table 11
Results of the uncertainty analysis of measuring the input parameters in the present study.

Uncertainty parameters Estimated average of Estimated standard deviation Estimated variation coefficient
threshold velocity of threshold velocity of threshold velocity

Simulation without considering uncertainty 0.585 0.22290 0.49928


Median grain size 0.787 0.00222 0.00282
Relative density of particles 0.634 0.00014 0.00023

Table 12 of 0.301 was the weakest method for estimating the threshold
Comparing the performance of different data-driven models used in the present
velocity of motion according to the applied data in this study.
study.
In addition to the mathematical comparison of the applied mod-
Applied methoda Error measurement criteria els (based on the statistic criteria), the Mann-Whitney U test is also
NRMSE MAE NSE R2 used to compare differences between the outcomes of each model
and the observed data. This test can be utilized for two independent
May (empirical) 0.810 0.520 −2.67 0.301
Isbach (empirical) 0.556 0.370 −0.813 0.721 groups when the dependent variable are not normally distributed
RBNN 0.136 0.168 0.903 0.983 (in this study, preliminary calculations using the Anderson-Darling
FFNN-tansig- traincgb 0.127 0.072 0.911 0.975 test confirmed that the data were not normal). Results implied
FFNN-logsig- trainlm 0.119 0.066 0.920 0.973 that unlike the mathematical models, the estimating results of
FFNN-tansig- trainlm 0.118 0.066 0.926 0.983
FFNN-logsig- traincgb 0.116 0.062 0.921 0.973
all the DDMs are not statistically different at 0.01 and 0.05 levels
ANFIS-IWO 0.079 0.039 0.971 0.971 (Table 13).
ANFIS 0.062 0.019 0.967 0.979 Tables 12 and 13 reveal the fact that DDMs are much more effi-
ANFIS-GA 0.037 0.017 0.993 0.993 cient than mathematical models in estimating the incipient motion
WaveNet 0.024 0.008 0.997 0.997
velocity. Needless to say that the physics of the threshold velocity
a
Note: FFNN: feed forward back-propagation multi-layer perceptron neural net- of motion is dependent on many effective parameters and is also
work; RBNN: radial basis function neural network; ANFIS: adaptive neuro-fuzzy complicated. The main reason for this matter can be sought in the
inference system; ANFIS-GA: Hybrid ANFIS and Genetic Algorithm; ANFIS-IWO:
Hybrid ANFIS and Invasive Weed Optimization: WaveNet: wavelet neural network;
capability of DDMs in simulating complex processes.
the applied methods are ordered from worst to best based on the NRMSE criterion. Having a better insight of the outcomes of this study, Table 14
compares the simulation procedure and the major outcomes of
the similar reported studies for simulating the incipient motion of
2.446, and standard deviation of 0.2 were produced. Results of the bed sediments using data-driven models versus the present study.
uncertainty analysis of threshold velocity of motion are presented Based on the obtained results of these studies, it can be concluded
in Table 11. that using data-driven models seems to be a promising approach
A critical analysis from the perspective of uncertainty test, for modeling controversial topics in hydraulic engineering such as
Table 11 shows that the median diameter of particles is the most simulating incipient motion of the bed sediments.
important factor affecting the mean and standard deviation of the
threshold velocity of motion; furthermore, the biggest coefficient
of variations is related to the median diameter of particles, which is 5. Conclusion
nearly 0.28%. As a result, the parameter of density of sediment par-
ticles has minimum effect on variations of the threshold velocity of In the present research, the threshold velocity of motion of
motion. sediment particles was estimated using the data-driven and mathe-
matical models. The data-driven methods employed in the present
study included multi-layer perceptron ANN, radial basis ANN, adap-
4. Discussion tive neuro-fuzzy inference system (ANFIS, ANFIS-IWO & ANFIS-GA),
and WaveNet. The parameters affecting the incipient motion were
In order to compare the performance of the methods used in considered as the models’ inputs, which included median diam-
the present study to estimate the threshold velocity of motion, the eter of particles, water depth, bed width, bed slope, and relative
statistical criteria of each method for the testing data are presented density of particles. Moreover, the mathematical model of Isbach
in Table 12. and May was used in the present study to estimate the threshold
Regarding the NRMSE of 0.024 and NSE of 0.997, the WaveNet velocity of motion. In Isbach’s method, the relative density of parti-
method was the best model compared to other models; further- cles should be higher than 2, which is a limitation for this method.
more, May’s mathematical model with the NRMSE of 0.81 and NSE The obtained results indicated that May’s method had the poor-

Table 13
Mann-Whitney U test to compare differences between the final results of the applied models and the observations in the test phase.

Models p-value Significantly Significantly


different (95%) different (99%)

Measurements vs. Isbach 0.0000 YES YES


Measurements vs. May 0.0000 YES YES
Measurements vs. FFNN-logsig- traincgb 0.593 NO NO
Measurements vs. FFNN-logsig- trainlm 0.442 NO NO
Measurements vs. FFNN-tansig- traincgb 0.243 NO NO
Measurements vs. FFNN-tansig- trainlm 0.414 NO NO
Measurements vs. RBNN 0.198 NO NO
Measurements vs. ANFIS 0.937 NO NO
Measurements vs. ANFIS-IWO 0.953 NO NO
Measurements vs. ANFIS-GA 0.989 NO NO
Measurements vs. Wavenet 0.997 NO NO
M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176 175

Table 14
Comparison of the simulation procedure and the major outcomes of the similar reported studies for simulating the incipient motion of bed sediments using data-driven
models versus the present study.

Authors Method Remarks Data Statistical indices Major outcomes


for model
verificationa

Sreenivasulu et al. [8] Radial basis neural Predicting the incipient a) from their own R2 RBF approach gives an
network (RBNN) motion in an alluvial experiments & b) other approximation route of
channel sources designing the system and at
the same time Manning’s
equation can be avoided while
designing the incipient motion
in alluvial channels
Valyrakis et al. [45] Adaptive neuro-fuzzy Predicting the incipient Data sets obtained RMSE, CC MAE Acceptable results (CC = 0.79) &
inference system entrainment of a coarse from experimental demonstrates data-driven
(ANFIS) particle at low mobility tests using LDV models’ potential use in a
uniform flow conditions broader array of phenomena in
geomorphology
Kumar et al. [46] Multi-gene symbolic Sediment Several data sources NSE, R2 , Id It was observed that models
genetic programming transport-vegetated flow, have been used developed therein showed
incipient motion and total satisfactory efficiency
bed load prediction according to all three criteria
and also showed better
efficiency compared to
empirical relations
Safari et al. [2] Three types of ANN Modelling incipient a) from their own RMSE, MAPE, CC Acceptable results: CC = 0.97
techniques including deposition of sediment in experiments & b) other for FFNN; CC = 0.90 for GRNN &
FFNN, GRNN & RBNN rigid boundary channels sources CC = 0.86 for RBNN. As a
conclusion, it is found that
appropriately constructed ANN
and regression models can be
successfully applied for the
estimation of flow velocity at
the incipient deposition
condition
Ebtehaj et al. [47] Group method of data Prediction of the Three different sources MAPE, RMSE It can be stated that applied
handling based on densimetric Froude of experimental data GMDH-GA model presents
genetic algorithm number (Fr) for incipient densimetric Froude, essential
(GMDH-GA) motion in rigid rectangular for incipient motion, with
storm water channel relatively higher accuracy
when compared with the
existing mathematical
equations
The present study Two types of ANN Simulating incipient Several data sources NRMSE, MAE, NSE, R2 Promising results of
techniques including motion of sediment in have been used data-driven methods (In
FFNN and RBNN as well alluvial channels general R2 > 0.97 & NSE > 0.92).
as ANFIS & WaveNet According to the sensitivity
analysis, the mean diameter of
particles, relative density, bed
width, bed slope, and water
depth had maximum absolute
values of simple correlation
coefficient, respectively.
a
Note: CC: correlation coefficient; MAE: the mean average error; RMSE: the root mean square error; MAPE: he mean absolute percentage error; R2 : coefficient of multiple
determination; Id : index of agreement; NSE: Nash–Sutcliffe Efficiency; NRMSE: normalized root mean square error.

est performance in estimating the threshold velocity of motion; two input parameters of the median diameter of particles and rel-
furthermore, comparison of the results showed that the WaveNet ative density on the threshold velocity of motion was investigated,
method with the highest NSE of 0.997 was the best method for esti- indicating that the parameter of the median diameter of sediment
mating the threshold velocity of motion. Following by the WaveNet particles had the highest effect on the variations of the threshold
method, the ANFIS-GA with the NSE of 0.993 was better than the velocity of motion, compared to the parameter of relative density.
other methods. Moreover, the multi-layer perceptron ANN had bet-
ter performance than the radial basis function ANN; besides, the
radial basis function ANN with the least NSE value of 0.903 had the References
poorest performance in estimating the threshold velocity of motion
among the DDMs. [1] V.A. Vanoni, Sedimentation Engineering, ASCE, New York, 1975.
[2] M.J.S. Safari, H. Aksoy, M. Mohammadi, Artificial neural network and
With regard to the sensitivity analysis using the simple correla-
regression models for flow velocity at sediment incipient deposition, J.
tion coefficient ranking method, the effect of each input parameter Hydrol. 541 (2016) 1420–1429.
on the threshold velocity of motion was investigated, the results of [3] Z. Cao, G. Pender, J. Meng, Explicit formulation of the Shields diagram for
incipient motion of sediment, J. Hydraul. Eng. 132 (2006) 1097–1099.
which indicated that, among these parameters, the median grain
[4] H. Zhu, P. Cheng, B. Zhong, B.D. Wang, The mechanisms of contaminants
size had the highest effect on the threshold velocity of motion. How- release due to incipient motion at sediment-water interface, Sci. China Phys.
ever, the least effect was related to the water depth. Afterwards, the Mech. Astron. 57 (2014) 1563–1568.
uncertainty analysis of the threshold velocity of motion was per- [5] R. Bravo, P. Ortiz, J. Luis Pérez–Aparicio, Analytical and discrete solutions for
the incipient motion of ellipsoidal sediment particles, J. Hydraul. Res. (2017)
formed with regard to the sensitivity analysis. Uncertainty of the 1–15.
176 M. Zounemat-Kermani et al. / Applied Soft Computing 69 (2018) 165–176

[6] A. Sheilds, Anwendung der Ahnlichkeitsmechanik und Turbulenzforschung [28] M. Shahlaei, A. Madadkar-Sobhani, L. Saghaie, A. Fassihi, Application of an
auf Geschiebebewegung, Mitteilungen der Preuss. Versuchsanst.f. Wasserbau expert system based on Genetic Algorithm–Adaptive Neuro-Fuzzy Inference
u. Schiffbau, Heft 26, Berlin, 1936. System (GA–ANFIS) in QSAR of cathepsin K inhibitors, Expert Syst. Appl. 39
[7] A.A. Beheshti, B. Ataie, Analysis of threshold and incipient conditions for (2012) 6182–6191.
sediment movement, J. Costal Eng. 55 (2008) 423–430. [29] B.R. Bakshi, G. Stephanopoulos, Wave-net: a multiresolution hierarchical
[8] G. Sreenivasulu, B. Kumar, A.R. Rao, RBF modeling of incipient motion of plane neural network with localized learning, AIChE J. 39 (1993) 57–81.
sand bed channels, international journal of computer electrical, automation, [30] B. Shokooh Saljooghi, A. Hezarkhani, Comparison of WAVENET and ANN for
Control Inf. Eng. 2 (2008) 2688–2693. predicting the porosity obtained from well log data, J. Petrol. Sci. Eng. 123
[9] J. Sehrawat, M. Patel, B. Kumar, Gaussian process regression to predict (2014) 172–182.
incipient motion of alluvial channel, Proceedings of Fourth International [31] M. Thuillard, A Review of Wavelet Networks, Wavenet, Fuzzy Wavenets and
Conference on Soft Computing for Problem Solving 336 (2015) 435–441. Their Applications, ESIT, Aachen, Germany, 2000.
[10] S.V. Isbach, Construction of Dams by Depositing Rock in Running Water, [32] Q. Zhang, Regressor selection and wavelet network construction, Decision
Transactions of the Second Congress on Large Dams, Communication No. 3, and Control, in: Proceedings of the 32nd IEEE Conference, San Antonio, TX,
U.S. Government Printing Office, Washington, D.C, 1936. USA, USA, 1993.
[11] R.W.P. May, preventing sediment deposition in inverted sewer siphons, J. [33] Y. Chen, B. Yang, J. Dong, Time-series prediction using a local linear wavelet
Hydraul. Eng. 129 (2003) 283–290. neural network, Neurocomputing 69 (2006) 449–465.
[12] C.T. Yang, Sediment Transport: Theory and Practice, McGraw-Hill Companies, [34] Y. Bengio, Y. Grandvalet, No unbiased estimator of the variance of k-fold
New York, 1996. cross-validation, J. Mach. Learn. Res. (2004) 1089–1105.
[13] A.M. Salem, The effects of the sediment bed thickness on the incipient motion [35] T.T. Wong, Performance evaluation of classification algorithms by k-fold and
of particles in a rigid rectangular channel, in: Seventeenth International leave-one-out cross validation, Pattern Recogn. 48 (9) (2015) 2839–2846.
Water Technology Conference, IWTC17, Istanbul, 2013. [36] J.A. Sanchez-Cabeza, A.C. Ruiz-Fernandez, J.F. Ontiveros-Cuadras, L.H. Perez
[14] A.J. Sutherland, Proposed mechanism for sediment entrainment by turbulent Bernal, C. Olid, Monte Carlo uncertainty calculation of 210Pb chronologies
flows, J. Geophys. Res. 72 (1967) 6183–6194. and accumulation rates of sediments and peat bogs, Quat. Geochronol. 23
[15] S.K. Bose, S. Dey, Sediment entrainment probability and threshold of (2014) 80–93.
sediment suspension: exponential-based approach, J. Hydraul. Eng. 139 [37] H.M. Baaloush, Using Monte Carlo simulation to estimate natural
(2013) 1099–1106. groundwater recharge in Qatar, Model. Earth Syst. Environ. 87 (2016) 2–7.
[16] J.H. Jang, H.Y. Ho, C.L. Yen, Effects of lifting force on bed topography and [38] B. Rajasekhar, I.M. Nambi, S.K. Govindarajan, Human health risk assessment
bed-surface sediment size in channel bend, J. Hydraul. Eng. 137 (2011) of ground water contaminated with petroleum PAHs using Monte Carlo
911–920. simulations: a case study of an Indian metropolitan city, J. Environ. Manage.
[17] S. Dey, Sediment threshold, J. Appl. Math. Modell. 23 (1999) 399–417. 205 (2018) 183–191.
[18] A. Armanini, Incipient sediment motion at high slopes in uniform flow [39] J.D. Salas, H.S. Shin, Uncertainty analysis of reservoir sedimentation, J.
condition, J. Water Resour. Res. 41 (2005) W12431. Hydraul. Eng. 125 (1999) 339–350.
[19] S. Palani, S.Y. Liong, P. Tkalich, An ANN application for water quality [40] S. Dey, U. Raju, Incipient motion of gravel and coal beds, Sadhana 27 (2002)
forecasting, Mar. Pollut. Bull. 56 (2008) 1586–1597. 559–568.
[20] M. Zounemat-Kermani, O. Kisi, T. Rajaee, Performance of radial basis and [41] S. Dey, R.V. Raikar, Characteristics of loose rough boundary streams at near
LM-feed forward artificial neural networks for predicting daily watershed threshold, J. Hydraul. Eng. 133 (2007) 288–304.
runoff, Appl. Soft Comput. 13 (2013) 4633–4644. [42] H. Afzalimehr, S. Dey, P. Rasoulianfar, Influence of decelerating flow on
[21] ASCE Task Committee, Artificial neural networks in hydrology. I: preliminary incipient motion of a gravel-bed stream, Sadhana 32 (2007) 545–559.
concepts, J. Hydrol. Eng. 5 (2000) 115–123. [43] S. Dey, S. Sarkar, L. Solari, Near-bed turbulence characteristics at the
[22] M. Zounemat-Kermani, Hydrometeorological parameters in prediction of soil entrainment threshold of sediment beds, J. Hydraul. Eng. 137 (2011) 945–958.
temperature by means of artificial neural network: case study in Wyoming, J. [44] W.R. Brownlie, Compilation of Alluvial Channel Data: Laboratory and Field,
Hydrol. Eng. 18 (2012) 707–718. 1981, Report. W. M. Keck Laboratory of Hydraulics and Water Resources.
[23] K. Gurney, An Introduction to Neural Networks, UCL Press, 1997. [45] M. Valyrakis, P. Diplas, C.L. Dancey, Prediction of coarse particle movement
[24] J. Liu, Radial Basis Function (RBF) Neural Network Control for Mechanical with adaptive neuro-fuzzy inference systems, Hydrol. Processes 25 (2011)
Systems, Springer-Verlag Berlin Heidelberg, 2013. 3513–3524.
[25] J.S. Jang, ANFIS: adaptive-network-based fuzzy inference system, IEEE Trans. [46] B. Kumar, A. Jha, V. Deshpande, G. Sreenivasulu, Regression model for
Syst. Man Cybern. 23 (1993) 665–685. sediment transport problems using multi-gene symbolic genetic
[26] M. Zounemat-Kermani, M. Teshnehlab, Using adaptive neuro-fuzzy inference programming, Comput. Electron. Agric. 103 (2014) 82–90.
system for hydrological time series prediction, Appl. Soft Comput. 8 (2008) [47] I. Ebtehaj, H. Bonakdari, F. Khoshbin, C.H.J. Bong, A. Ab Ghani, Development of
928–936. Group Method of Data Handling based on Genetic Algorithm to predict
[27] A.R. Mehrabian, C. Lucas, A novel numerical optimization algorithm inspired incipient motion in rigid rectangular storm water channel, Sci. Iranica Trans.
from weed colonization, Ecol. Inf. 1 (2006) 355–366. A Civil Eng. 24 (2017) 1000–1009.

Potrebbero piacerti anche