You are on page 1of 18

Ain Shams Engineering Journal 13 (2022) 101732

Contents lists available at ScienceDirect

Ain Shams Engineering Journal


journal homepage: www.sciencedirect.com

Monthly runoff prediction at Baitarani river basin by support vector


machine based on Salp swarm algorithm
Sandeep Samantaray a,b,⇑, Sambit Sawan Das a, Abinash Sahoo b, Deba Prakash Satapathy a
a
Department of Civil Engineering, College of Engineering and Technology, Bhubaneswar, Odisha 751003, India
b
Department of Civil Engineering, National Institute of Technology, Silchar, Assam 788010, India

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

Article history: Accurate monthly runoff prediction is still challenging work regardless of the accessibility of different
Received 6 August 2021 modelling techniques, like the knowledge-driven or data-driven models, and human activities and cli-
Revised 18 January 2022 mate changes. To this context, applicability of hybrid SVM-SSA (Support Vector Machine with Salp
Accepted 1 February 2022
Swarm Algorithm) model and conventional SVM and artificial neural network (ANN) models is investi-
Available online xxxx
gated for runoff prediction in Baitarani river basin, Odisha, India. Potential of proposed techniques is mea-
sured utilising four quantitative indexes, root mean squared error (RMSE), coefficient of determination
Keywords:
(R2), mean absolute error (MAE), and Willmott index (WI). Test results specify that hybrid model gener-
Baitarani River
Runoff
ates better prediction accurateness in comparison to applied conventional methods. The generalization
Support vector machine and robustness of SVM-SSA techniques were very prominent, with R2 values of 0.9847 and 0.9844 for
Salp Swarm Algorithm Anandpur and Champua stations during training phases. Similarly prominent value of WI are 0.9906,
0.9902 and minimum value of RSE and MAE are 20.019, 0.5928 and 0.0769, 0.5934 for Anandpur and
Champua stations respectively. Therefore, SVM-SSA can be recommended for modeling complexity of
interactions for rainfall-runoff process and predicting runoff.
Ó 2022 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Ain Shams Uni-
versity. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/
by-nc-nd/4.0/).

1. Introduction systems [3]. Moreover, precise flow simulation can be a source for
valuation of accessible water resources concerning various circum-
Rainfall-runoff relationship plays a significant part in maintain- stances of impending climate [4]. However, methods used for run-
ing water balance and is one of the fundamental factor of hydrolog- off estimation from rainfall events vary. Generally, two models are
ical cycle [1]. The non-linear process involving prediction of runoff available to simulate rainfall-runoff process: data-driven (DD)
depends on several aspects like duration and intensity of rainfall models and physical-based models [5]. For physical interpretation
and physical features of upstream watershed. It also includes a of processes in a watershed system, physical-based hydrological
huge number of data and other aspects which is needed to be ana- models utilise parametric assumptions and complex physical
lyzed. Modeling rainfall-runoff process in a conventional way has equations [6]. Yet, it is often difficult to understand the relation-
constantly been a very challenging assignment. Yet, their modeling ship amid rainfall and runoff owing to temporal and spatial vari-
is crucial for forecasting future runoff patterns by observation and ability of datasets, meteorological complexities, and topographic
analysis of preceding runoff patterns on basis of rainfall constraints characteristics [7,8]. Moreover, these models necessitate substan-
[2]. Precise rainfall-runoff modelling is essential for appropriate tial efforts during standardization process and may generate insuf-
reservoir management in dry seasons and predictive flood warning ficient results because of structural uncertainty and parameters of
models [9,10]. In recent times, as quality and resolution of avail-
⇑ Corresponding author at: Department of Civil Engineering, National Institute of able data have increased, simulation procedure of physical models
Technology, Silchar, Assam 788010, India. have become computationally intensive and time-consuming.
E-mail address: sandeep1139_rs@civil.nits.ac.in (S. Samantaray). Water pollution is a serious problem that needs to be looked upon
Peer review under responsibility of Ain Shams University. [11-15]. On the other hand, DD models are an efficient alternative
to decrease the precincts of physical models. These models work
on basis of functional relationship between dependent (output)
and independent (input) variables [16,17].
Production and hosting by Elsevier

https://doi.org/10.1016/j.asej.2022.101732
2090-4479/Ó 2022 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Ain Shams University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

In last two decades, DD models made a significant contribution optimisation algorithms for solving various problems related to
towards advancement in hydro-geological studies by giving cost- engineering and science. Population based optimisation algorithms
effective solutions and accurate performance [18]. For hydrological are separated into two main groups: swarm-intelligence and evo-
analysis, conventional DD models consist of artificial neural net- lutionary based algorithms. Laws of natural evolution are basis of
works [19], SVM [20], and random forest [21]. ANNs have been evolutionary algorithms. The well-known evolutionary algorithms
employed adequately to predict nonlinear hydrological processes are differential evolution, genetic programming, genetic algo-
such as water quality, precipitation, rainfall-runoff [22-26], and rithms, etc. Swarm-intelligence algorithms mimic understanding
stream flow modeling [27,28]. These models have always been of flocks, swarms, or herds of natural creatures. Most popular
selected as robust black-box models to simulate runoff and have swarm algorithms [51] are particle swarm optimisation, artificial
been utilised for comparison with other models [29]. Conventional bee colony algorithm, ant colony optimisation. And in recent times,
linear regression models are not compatible for modeling nonlin- some new algorithms are ant lion optimiser [52], whale optimisa-
ear relationship amid input parameters, hence are not usually con- tion algorithm [53], bees algorithm [54,55], grey wolf optimisation
sidered. For establishing reliable and accurate prediction models, [56], grasshopper optimisation algorithm [57], etc. Apart from
more flexible techniques have to be presented. In this study, common control parameters, many algorithms require their speci-
SVM-based prediction model is proposed for solving this problem. fic algorithm-based parameters. Appropriate setting of algorithm-
SVMs are robust and powerful methodology firmly based on prin- based parameters is an essential aspect affecting performance of
ciple of statistical learning theory. It was developed for solving a particular algorithm. Inappropriate arrangement of algorithm-
classification and regression problems. SVM presented preeminent based parameters either generates local optimum solution or
performances in prediction of various hydrological variables, such increases computational work [51].
as flood forecasting [30,31], runoff prediction [32-35], groundwa- Wavelet-based ANN and SVM approaches have been applied for
ter level forecasting [36], sediment analysis [37-39] etc. Lately, modelling rainfall–runoff process at different watersheds [58,59].
there have been very important developments in investigating They found that wavelet-based hybrid techniques modelled
ability of SVM in modeling rainfall-runoff process. rainfall-runoff process more efficiently and reliably. Zhang et al.
Misra et al. [40] used SVM for simulating sediment yield and [60] used SSA to estimate parameter of soil water retention curve.
runoff at different rain gauge stations of Vamsadhara River in India For validating advantages of SSA, obtained results were compared
and compared the obtained results with ANN. The results indicated against PSO, RETC program, and Differential Evolution (DE), which
that SVM performed better and can be an effective alternate to indicated that SSA performed superiorly compared to other applied
ANN. Goyal et al. [41] used SVM models for predicting flow rates techniques. Feng et al. [61] used SVM integrated with quantum-
from two nearby karstic springs in Greece. Their findings showed based PSO for monthly runoff estimation at two reservoirs in
that SVM predicted better flow rates compared to ARIMA, RBFN Yangtze Valley, China. Simulation results indicated that integrated
and GRNN models. Sharma et al. [42] used ANN and SVM tech- model provided better forecasting accurateness than ANN and ELM
niques to model runoff and sediment yield of Kankaimai catch- methods, making it an efficient model for scientifically operating
ment, Nepal. Also, performance of ANN and SVM models was hydropower reservoirs. Nguyen et al. [62] investigated the appli-
investigated for simulating rainfall-runoff process subjective to cation of novel MLNN-SSA (multiple hidden layer neural network)
snow affected Roodak catchment in Tehran region, Iran [43]. Both methods for landslide study in Lai Chau mountainous region, Viet-
studies concluded that SVM provided superior prediction results nam. Their results showed that usage of SSA enhanced search
than ANN. Sharifi et al. [44] applied ANN, ANFIS, local linear effectiveness for landslide study and analysis of other natural haz-
regression, and SVM models for runoff prediction in Amameh ards. Dehghani et al. [63] used SVM-AF (artificial flora) to estimate
catchment in Iran. Results specified that performance of SVM is river flow discharge of Dez River basin located in Iran and com-
superior compared to other applied models. Tasar et al. [45] used pared its outcomes with results of Bayesian-SVM and wavelet-
SVM and ANFIS methods for modelling rainfall-runoff process of SVM. Results revealed that hybrid SVM-AF model performed more
Muskegon basin in USA. Results showed that SVM with low errors efficiently than other models in predicting daily discharge. SVM
modeled rainfall-runoff process more efficiently and consistently technique coupled with GA, PSO and SSA, were developed for pre-
than fuzzy model. Bafitlhile and Li [46] aimed to simulate and dicting strength of cement paste backfill reinforced with fiber (64).
forecast stream flow of three different catchments using ANN Their study revealed that SVM-SSA was most prominent in
and e-SVM. Analysis of results indicated that both models per- strength prediction compared to other hybrid methods. Wu et al.
formed well with SVM performing slightly better than ANN for [65] used integrated SVM-ACO model for runoff prediction at Min-
all catchments. Babaei et al. [47] used ANN and SVM models to jiang River basin. Outcomes showed that runoff predicted by using
predict inflow to Zayandehroud dam. Results showed that SVM SVM-ACO was more accurate than default parameter SVM.
models outperformed those of ANN models. Poonia and Tiwari Even though machine learning models have been employed for
[48] developed FFBP (feed-forward back propagation) and RBFN forecasting runoff, they have certain limitations. Literature reports
models for rainfall-runoff modelling of Hoshangabad catchment application of different optimisation algorithms for tuning SVM
of River Narmada. Results from their study suggested that pro- model parameters, however these algorithms may get trapped in
posed models proved to be important tools in predicting hydrolog- local minima. This research applies data driven models to predict
ical responses of watersheds. Hassan et al. [49] utilised ANN and runoff. A novel optimisation algorithm, SSA is utilised for training
SVM for long-term rainfall-runoff modelling at Muda River, Malay- SVM model. Motivation to select SSA is that it possesses capability
sia. Comparison of results revealed that SVM model showed better of comprehending superior quality results while filling several lim-
performance than ANN. Han and Morrison [50] used ANN, SVM itations of real-world problems [66]. Moreover, SSA displays a good
and long-short term memory approaches for predicting runoff in stability amid exploitation and exploration. Indeed, SSA governed
Russian River, California, USA. Results demonstrated that applied by advanced operatives is utilised as a substitute to conventional
models could efficiently forecast runoff and were useful in regions ANN learning algorithms for determining optimal values of biases
where observation stations were unavailable. and weights. More advantages of using SSA are its ability to avoid
In the meantime, prediction accuracy of SVM models depend on local optima and fast convergence rate [66]. Specifically, the main
correct arrangement of hyper-parameters, and there is no common aim of this research is to develop novel SVM-SSA data-driven
agreement on proper arrangement of these parameters. In this con- model for monthly runoff prediction with different climatological
text, widely adopted approaches are application of nature-inspired characteristics and assess the performance of SVM-SSA model
2
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

against standalone SVM and simple ANN models based on statisti- imum temperature up to 40°  46 °C in April to June (mid). This
cal measures and graphical inspections. See Fig. 1 area is structurally and geologically complex and is characterised
by diverse geomorphic setup leading to broadly deviating hydro-
logic conditions. Rainfall (mm), temperature (oC), stage (m), speci-
2. Study area fic humidity (g/kg), and relative humidity (%) data of selected study
area are collected from various meteorological departments
River Baitarani originates from hill ranges of Keonjhar district, (Fig. 3). Statistical results of collected data are presented in Table 1
Odisha (Fig. 2). It is one of the six main rivers of Odisha with an and 2 for Anandpur and Champua, respectively.
altitude ranging between 32 and 1024 mts. River flows amid
20°350 N to 22°150 N latitude and 85°030 E and 87°030 E longitude 3. Methodology
and through Odisha and Jharkhand states. A significant portion of
the basin is covered by agricultural land, with the basin area of 3.1. ANN
14,351 km2 in Odisha. Monsoon starts from June and lasts until
October. Minimum and maximum annual precipitation are around ANN is a computational tool that has the potential of giving new
800 mm and 2000 mm, respectively, with mean annual precipita- information based on some input information and computations
tion being about 1400 mm. [67-69]. In hydrology, MLP is one of the commonly used ANNs
In present study, Anandapur and Champua stations located in (Fig. 4). It constitutes input, one or more hidden layers, and an out-
Keonjhar District of Odisha are selected for analysis. This district put layer. Hidden layers function as a link amid input and output
contributes flow largely into Baitarani and lies in upper part of layers. Units of input layer have work of distributing input data
the basin. Keonjhar district is situated in northern part of Odisha and responses for output signals is provided by output layer. Num-
and comes under North central plateau zone. Because of humid ber of input and output determine number of neurons in both lay-
tropical climatic conditions, it receives massive to average rainfall ers. Mathematically this can be represented as:
from southwest monsoon between June to September and a little
from Northeast monsoon in December and January. Minimum X
n
y ¼ fð wj xj þ bÞ ð1Þ
temperature in winter falls as low as 10 °C or below, with the max-
i¼1

Where y – output; f - transfer function wj - weight vector, xj -


input vector (j ¼ 1; 2:::; n), b - bias.
In present work, Levenberg–Marquardt (LM) BP learning algo-
rithm is utilised, to train MLP network [70]. Fastest BP algorithm
is the LM algorithm, and hence is very much suggested as a first
choice learning algorithm, even though it does necessitate addi-
tional memory compared to other algorithms [71,72].

3.2. SVM

One of the characteristic machine learning algorithms for


regression and classification of dataset is SVM [20]. Best classifica-
tion is achieved by finding optimum hyper plane comprised of sup-
port vectors for classifying various categories of vectors with a
maximum margin of distance amid data. Using SVM, nonlinear
separable problems can also be solved (Fig. 5). Selecting an appro-
priate kernel function helps to map data points of input feature
space onto high dimension feature space which ultimately aids
in identifying complex input–output relationships in a relatively
simple way. SVM doesn’t face any difficulty of local minima and
has benefit of convex optimization. Several training samples are
not required by SVM for developing a model and can solve prob-
lems with smaller samples. An assumption is made that N number
of samples are involved to train database without affecting gener-
alization. Therefore, SVM technique for regression analysis can be
mathematically defined as:

yi ¼ wi :/ðxi Þ þ bi ; t ¼ 1; 2;    ; N ð2Þ
D
where xi 2 R - input vector and yi 2 R - output vector in ith data
sample. /ðxi Þ– non-linear mapping function for transforming orig-
inal input vector into objective feature space. wi - weight vector
and bi – bias. Both wi and bi can be found using following equation:
8 PN
>
< Min 12 kwk2 þ C: i¼1 ðni þ ni  Þ
sujectedto : w:/ðxi Þ þ bi  yi  e þ ni  ð3Þ
>
:
yi  w:/ðxi Þ  bi  e þ ni
Where ni and ni  - positive slack variable signifying ith sample
training error; e - error tolerance . C - parameter which controls
Fig. 1. Study flow chart. empirical risk of model.
3
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

Fig. 2. a) Location of Odisha over India. b) Map of Odisha showing Baitarani river basin. c) DEM of Baitarani river basin.

3.3. SSA lylike bodies. They travel together as chains and exist as colonies.
There exist two major categories of salps, one being leaders and
SSA is a newly developed algorithm, i.e., inspired by nature the other followers. Leaders are those which lead the chain in for-
mimicking swarm behavior of salps (73). In SSA, to solve optimisa- ward course, whereas followers follow their leaders in harmony
tion problems movement behavior is modelled mathematically. and synchronously.
Salps are sea creatures moving around by propelling water through Like every swarm intelligent algorithm, SSA starts by an arbi-
their bodies from one side to another, and have barrel-shaped jel- trarily initialising swarm of N salps. For a particular model, n vari-
4
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

Fig. 3. Range of dataset used for Anandpur and Champua.

(
ables are assumed to be assessed, x signifies location of a salp, and yi þ r1 ððubi  Ibi Þr2 þ Ibi Þ; r 3  0;
y describes food source indicating objective of a swarm in search xi ¼
1
ð4Þ
yi  r1 ððubi  Ibi Þr2 þ Ibi Þ; r 3 < 0;
space. Leader salp apprises its location by using following equation
in search process:

5
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

Table 1
Statistical parameters of applied data Anandpur.

Statistical parameters Training set Testing set Total data set (3 6 0) Training set Testing set Total data set (3 6 0)
(2 5 2) (1 0 8) (2 5 2) (1 0 8)
Rainfall Specific humidity
Min 0 0 0 5.79 7.08 5.79
Max 266.0645161 197 266.06 21.14 21.03 21.14
Mean 40.61487649 43.76234175 41.99 14.66710317 14.69888889 14.56
Kurt 1.942740474 1.226937916 1.317852414 1.504241611 1.527454343 1.490162379
Skew 1.334344651 1.228532494 1.208665359 0.109580082 0.096425755 0.104615164
SD 46.28148404 46.37826658 45.51917017 4.706597973 4.663877159 4.697280464
Temperature Relative humidity
Min 17.60483871 17.24 17.24 33.1 35.52 31.85
Max 34.19516129 34.00766667 34.20 92.7 91.02 92.70
Mean 25.74864935 25.6037963 25.78 68.31035714 68.61083333 67.62
Kurt 0.838966044 0.841251472 0.846955919 1.38569263 1.236216137 1.343812149
Skew 0.199304186 0.341856369 0.217705198 0.152261287 0.252544277 0.145110571
SD 4.237265157 4.275478728 4.172442181 17.65495466 16.71002297 17.50008293
Stage
Min 31.2765 4.469333333 4.47
Max 36.4785 51.63806452 51.64
Mean 33.89240893 31.33590379 33.22
Kurt 0.490082918 9.404315401 37.53616965
Skew 0.899627632 1.624524334 3.900531561
SD 0.858846966 5.156217438 2.789849272

Table 2
Statistical parameters of applied data Champua.

Statistical parameters Training set Testing set Total data set (3 6 0) Training set Testing set Total data set (3 6 0)
(2 5 2) (1 0 8) (2 5 2) (1 0 8)
Rainfall Specific humidity
Min 0 0 0 3.68 5.28 3.60
Max 204.3 202.7096774 204.30 20.58 20.32 20.58
Mean 38.55868075 31.38478059 30.91 12.88134921 13.18888889 12.86
Kurt 1.103708082 3.953169265 2.218758346 1.552930624 1.538918649 1.522653034
Skew 1.303833734 1.854150238 1.629214532 0.071344787 0.078083481 0.070362875
SD 45.51900407 40.99036653 43.11546584 5.358930938 5.13479105 5.312193122
Temperature Relative humidity
Min 15.99806452 15.8116129 15.81 17.53 24.16 15.51
Max 35.49 34.03033333 35.49 91.74 91.29 91.74
Mean 24.96596695 24.79410005 24.95 61.77361111 63.47481481 61.54
Kurt 0.908129006 0.898814195 0.926525805 1.339693617 1.265803156 1.297798055
Skew 0.104949071 0.222806223 0.117467595 0.185029194 0.214273112 0.176476852
SD 4.867673223 4.896714813 4.798352609 22.11406529 20.47619637 21.72732018
Stage
Min 359.1019355 0 0.00
Max 373.375 393.9122581 393.91
Mean 371.268622 347.8595503 362.69
Kurt 56.76130121 27.9753688 64.20380866
Skew 6.031519565 4.231033995 7.131488002
SD 1.131829817 46.71082931 35.30937216

where in ith dimension, xi 1 - location of first salp; yi - food posi- dfinal x  x0


k¼ ; whered ¼ ð7Þ
tion. ubi and Ibi - upper and lower bound, and r1 ; r2 ; r3 - three arbi- d0 t
trary numbers. Assuming that d0 ¼ 0, t - iteration in an optimisation problem,
Amongst three arbitrary numbers, r 1 inhabits leading place above-mentioned expression can be converted into subsequent
since it balances exploitation and exploration during whole search formula:
procedure. It is expressed by
 
xi j ¼ 0:5 xi j þ xi j1 ð8Þ
r 1 ¼ 2eð L Þ
4l 2
ð5Þ
where l signifies present iteration; L – previously defined num- where j  2. Eq. (6) illustrates that follower salps apprise their
ber of most iterations; r2 ; r3 - arbitrary numbers amid ½0; 1. Follow- location on basis of their own position and preceding salps.
ing mathematical expression is utilised for followers for updating If certain salps passage out of the restricted search space, they
their locations in accordance with Newton’s law of motion: will be carried back within the limits using following equation:
8 j j
xi j ¼ 0:5kt2 þ d0 t ð6Þ >
<l ifxi j  l
xi j ¼ uj ifxi j  uj ð9Þ
where  2 , xi j - location of jth salp in ith dimension, t - time, >
: j
xi otherwise
d0 - preliminary speed.
6
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

Fig. 4. ANN structure.

Fig. 5. Architecture of SVM.

All the above updating are iteratively performed till end condi- present solution may discover a better solution. To put in another
tions are satisfied. A point should be noted that sometimes food way, salp chain during optimization possess the capability of mov-
sources can be updated by exploring and exploiting space around ing towards global optimal solution.

7
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

SSA pseudo code training SVM. Hence, each salp location is signified an amalgama-
tion of two parameters (kernel parameters and C) in 2-D search
Input maximum iteration L, population size m, ub, lb, l = 1 space.
Initialize the salp position {u1, u2, u3, . . .. . .. . ., um} Testing: Testing sample is utilised for computing error rate
While (termination condition is not satisfied) (misclassification) of each location. Number of misclassified sam-
Calculate the fitness of each individual (salp) ples (N err ) divided by total number of testing samples (N total ) gives
Sort the salp position according to the fitness value error rate. When value of two afore-mentioned constraints at a
Set F as the best position of current population particular location pi 2 R2 achieved least error rate, optimum solu-
Update C1 by Eqn. (26) tion is positioned at pi 2 R2 ,
For each salp position (ui)
If (i  m=2) update the position of leading salps by Eqn. 25 N err
else update the position of follower salps by Eqn. 27 Minimize : F ¼ ð10Þ
Ntotal
end
end Termination: Entire process is terminated after reaching end
Amend the salp that go beyond the upper and lower bound condition; otherwise, next generation operation continues. In this
end study, SSA is terminated after operation reaching maximum itera-
Output the optimization result tions, or attained best solution remains unchanged within a spec-
ified number of iterations. Here, salp locations are then updated
in accordance with (Eqs. (5) and (8)). Flowchart of SVM-SSA is pre-
sented in Fig. 6.

3.4. SVM-SSA
3.5. Processing and preparation of data
Here, developed SVM-SSA model is described for computing
optimal parameters of SVM using following steps The objective of present study is to predict runoff considering
Initialisation: Subsequent parameters of SSA were initialised different climatic parameters. For runoff prediction, monthly rain-
(number of salps, maximum iterations, velocity, and position of fall and stage were collected from CWC, Bhubaneswar, and
salps). monthly temperature, specific humidity, and relative humidity
Training: In SVM-SSA model, SVM is provided with two values from IMD, Pune for a period of 1991–2020. 70% of total dataset
of input by SSA, RBF kernel parameter and penalty parameter C, for (1991–2011) are selected and utilised to train, and remaining

Fig. 6. Flowchart of SSA algorithm.

8
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

30% of dataset (2012–2020) are utilised to test the three data- PN  obs 2
i¼1 Ai  Bpre
i
driven models. WIðWillmott1981Þ ¼ 1  ½  
     2
The subsequent step of model development is establishment of PN  pre   obs
obs  

obs 
the database. Scaling of collected data are done using following i¼1 ðBi  Ai  þ Ai  Ai Þ
formula:  ð0  1Þ ð16Þ
y  ymin
yscaled ¼ ð11Þ 
ymax  ymin Where, Aobs
i and Bpre are measured and predicted i th
runoff, Aobs
i
i

where yscaled = scaled data; y = actual data prior to scaling; and and Bpre
i are mean of measured and predicted runoff
ymax and ymin = maximum and minimum data respectively prior
to scaling.
Assortment of appropriate inputs for creating model structure is 4. Results and discussions
significant (74). Correlation benchmark is employed as a statistical
index for choosing input parameters. Correlation amid runoff at Recalling that 70% of data were utilised to train and 30% were
time t and rainfall, temperature, stage, specific humidity, and rela- utilised to test the models for runoff prediction. Different inputs
tive humidity with one month lag time was analyzed, and param- include Pt, Tt, St, AHt, RHt, Pt-1, Tt-1, St-1, AHt-1, and RHt-1 variables
eters associated to lag time having eloquent correlation were and the target (output) variable is runoff (Qt). Results of runoff pre-
identified. Hence predicted runoff is a function of rainfall, temper- dictions are listed in Table 4 in training and testing phases for ANN,
ature, stage, specific humidity, and relative humidity at times t and SVM, and SVM-SSA. Results reveal that the prominent values of R2,
t  1 Table 3 WI are 0.9847, 0.9906, and 0.9844, 0.9902 during training phases
Q t ¼ P t ; Tt ; St ; AHt ; RHt ; Pt1 ; Tt1 ; St1 ; AHt1 ; RHt1 ð12Þ for Anandpur and Champua gauge station respectively. Whereas
the minimum value of RMSE is 0.019 and 0.0769 for Anandpur
where Q t = predicted runoff at time t; P t = rainfall at time t; and Champua gauge stations, respectively. It is found that the para-
Tt = temperature at time t; St = stage at time t; AHt = specific mount value of R2, WI in context to minimum value of RMSE and
humidity at time t;RHt = relative humidity at time t; Pt1 = one MAE found while scenario 10 is considered as input combination.
month lag precipitation; Tt1 = one month lag temperature; The performance value of all four statistical indicators for all sce-
St1 = one month lag stage; AHt1 = one month lag specific humid- narios conditions during both training and testing phases at
ity; RHt1 = one month lag relative humidity; f = function producing Anandpur and Champua gauge station is given in Table 4. From
output from inputs. Table 4, all the statistical indices for performance assessment spec-
ify that SVM-SSA prediction model has better fitness and provides
3.6. Model performance evaluation more reliable and accurate runoff projections.
In contrast, SVM simulates observed runoff better than ANN. It
For getting optimum models established to calculate runoff at is also found that scenario 10 found more prominent than other
stations of Baitarani River basin, various statistical measures are nine scenario input combinations. For a given lag time, ANN model
discussed. The measures utilised are R2, RMSE, MAE, and WI which performance in testing period is inferior to in training period,
are expressed in Eq. (13), 14, 15, and 16, respectively: which is suggested by higher RMSE and lower WI values. This spec-
PN   2 ifies over fitting of data by ANN models in training stage. SVM and
obs
2 i¼1 ðAi  Aobs i ÞðBi
pre
 Bpre i Þ SVM-SSA models also have related problems, however to consider-
R ¼ ðsffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiÞ ð0  1Þ ð13Þ
PN  2 N
P obs  2 ably smaller magnitude, which clarifies an opinion that SVM and
ðAobsi  A obs
i Þ ðBi  Bobs i Þ SVM-SSA models outperform ANN models in both training and
i¼1 i¼1
testing stages. On basis of different statistical measures, findings
reveal that applied hybrid method has the potential of producing
N  
1X  obs pre  improved prediction results compared to conventional methods
MAE ¼ A  Bi  ð14Þ
N i¼1 i in different scenarios. Therefore, an efficient method is recom-
mended for solving complex hydrological forecasting problems
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N  2 under changing environmental conditions.
u1 X
RMSE ¼ t Bpre
i  Aobs i ð0  1Þ ð15Þ It can be observed from Fig. 7 (a) that best R2 values for runoff
N i¼1 estimations in testing period are equal to 0.905, 0.9255, and 0.9771

Table 3
Models and their corresponding inputs.

Output variable Input combinations Model description


ANN SVM SVM-SSA
Qt Pt ANN:1 SVM:1 SVM1:SSA
Qt Pt , Tt ANN:2 SVM:2 SVM2:SSA
Qt Pt , Tt, St ANN:3 SVM:3 SVM3:SSA
Qt Pt , Tt, St, AHt ANN:4 SVM:4 SVM4:SSA
Qt Pt , Tt, St, AHt, RHt ANN:5 SVM:5 SVM5:SSA
Qt Pt , Tt, St, AHt, RHt, Pt-1 ANN:6 SVM:6 SVM6:SSA
Qt Pt , Tt, St, AHt, RHt, Pt-1, Tt-1 ANN:7 SVM:7 SVM7:SSA
Qt Pt , Tt, St, AHt, RHt, Pt-1, Tt-1, St-1 ANN:8 SVM:8 SVM8:SSA
Qt Pt , Tt, St, AHt, RHt, Pt-1, Tt-1, St-1, AHt-1 ANN:9 SVM:9 SVM9:SSA
Qt Pt , Tt, St, AHt, RHt, Pt-1, Tt-1, St-1, AHt-1, RHt-1 ANN:10 SVM:10 SVM10:SSA

9
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

Table 4a
Comparative performance of ANN, SVM and SVM-SSA methods with various structures for Anandpur station.

Model Name RMSE R2 MAE WI RMSE R2 MAE WI


Training Testing
ANN:1 29.802 0.9045 3.7778 0.9105 44.86 0.8889 12.807 0.8945
ANN:2 27.305 0.9064 3.7109 0.911 42.37 0.8907 12.6683 0.8961
ANN:3 26.227 0.9078 3.6332 0.9132 41.298 0.8922 12.0072 0.8983
ANN:4 25.103 0.9096 3.507 0.915 40.421 0.8931 11.836 0.8996
ANN:5 24.232 0.9104 3.3365 0.9163 39.216 0.8952 11.559 0.9014
ANN:6 22.499 0.9117 3.2781 0.9174 38.106 0.8974 11.376 0.903
ANN:7 20.774 0.9133 3.2174 0.9186 36.978 0.8996 11.2074 0.9052
ANN:8 18.367 0.9152 3.0116 0.9198 35.581 0.9015 11.0021 0.9088
ANN:9 17.501 0.9169 2.989 0.9207 32.73 0.9032 10.908 0.9109
ANN:10 17.18 0.918 2.9273 0.9242 29.001 0.905 10.8655 0.9117
SVM:1 15.338 0.9322 2.542 0.9372 23.118 0.9111 9.512 0.9173
SVM:2 14.889 0.9348 2.4881 0.9395 20.317 0.9137 9.4877 0.9195
SVM:3 14.36 0.9363 2.4125 0.941 19.142 0.9143 9.331 0.9209
SVM:4 13.743 0.9382 2.3019 0.9438 17.88 0.9166 9.1504 0.922
SVM:5 13.406 0.9399 2.2856 0.9453 17.336 0.9174 9.0012 0.9238
SVM:6 12.901 0.9405 2.2108 0.9471 16.901 0.9199 8.9003 0.9256
SVM:7 12.375 0.9429 2.1734 0.949 16.634 0.9206 8.7196 0.9279
SVM:8 12.229 0.9437 2.006 0.9506 16.228 0.9219 8.518 0.9292
SVM:9 11.408 0.9461 1.9997 0.9519 16.012 0.9231 8.4839 0.9304
SVM:10 11.103 0.9474 1.9402 0.9537 15.773 0.9255 8.4263 0.9317
SVM1:SSA 5.72 0.9753 1.638 0.9805 10.374 0.9667 4.1072 0.9722
SVM2:SSA 3.903 0.9761 1.3973 0.9818 9.906 0.9686 3.991 0.9748
SVM3:SSA 2.089 0.9778 1.229 0.9823 9.5593 0.9695 3.749 0.9754
SVM4:SSA 1.004 0.9786 1.1037 0.9839 9.154 0.9703 3.2805 0.9769
SVM5:SSA 0.891 0.979 1.0076 0.9857 8.882 0.9718 3.2297 0.9776
SVM6:SSA 0.612 0.9802 0.925 0.986 7.995 0.9732 3.1883 0.9781
SVM7:SSA 0.31 0.9815 0.7994 0.9872 6.907 0.974 3.1648 0.9795
SVM8:SSA 0.1885 0.9829 0.7091 0.9881 4.012 0.9759 3.0071 0.9807
SVM9:SSA 0.102 0.9832 0.682 0.9894 3.015 0.9764 2.904 0.9821
SVM10:SSA 0.019 0.9847 0.5928 0.9906 2.721 0.9771 2.8156 0.983

Table 4b
Comparative performance of ANN, SVM and SVM-SSA methods with various structures for Champua station.

Model Name RMSE R2 MAE WI RMSE R2 MAE WI


Training Testing
ANN:1 33.209 0.9025 4.286 0.9093 45.583 0.886 4.773 0.8922
ANN:2 31.497 0.9036 4.1077 0.9108 45.219 0.8884 4.3281 0.8945
ANN:3 28.229 0.9054 3.952 0.9124 44.002 0.8892 4.1055 0.8963
ANN:4 26.784 0.9077 3.4004 0.9156 43.905 0.8905 3.993 0.8997
ANN:5 25.58 0.9092 3.2841 0.9164 39.997 0.8939 3.6008 0.9011
ANN:6 23.496 0.911 3.118 0.9179 38.773 0.8963 3.317 0.903
ANN:7 22.001 0.9129 3.0176 0.919 37.445 0.8977 3.1884 0.9049
ANN:8 21.208 0.9131 3.0025 0.9195 36.307 0.9004 3.0052 0.9061
ANN:9 18.95 0.9148 2.999 0.9217 34.882 0.9021 2.996 0.9086
ANN:10 17.992 0.9163 2.9273 0.9223 31.497 0.9036 2.8270 0.9094
SVM:1 15.566 0.9312 3.319 0.9486 24.638 0.9103 3.3328 0.9168
SVM:2 15.04 0.9346 3.1198 0.9404 22.97 0.9114 3.0196 0.9179
SVM:3 14.591 0.9354 2.905 0.9428 21.85 0.913 2.9925 0.9196
SVM:4 14.007 0.9367 2.7794 0.9433 19.609 0.9141 2.778 0.9203
SVM:5 13.902 0.9381 2.493 0.9449 18.003 0.9155 2.695 0.9214
SVM:6 13.55 0.9399 2.1137 0.945 17.294 0.9176 2.5003 0.9237
SVM:7 13.117 0.9405 2.089 0.9477 17.005 0.9198 2.4487 0.9251
SVM:8 12.54 0.9428 2.0041 0.9482 16.887 0.9202 2.381 0.9265
SVM:9 12.002 0.9443 1.996 0.9497 16.376 0.9217 2.2904 0.928
SVM:10 11.807 0.945 1.9402 0.9501 15.99 0.9249 2.2619 0.9302
SVM1:SSA 8.41 0.972 0.967 0.9777 10.901 0.9644 1.0071 0.9685
SVM2:SSA 7.09 0.9736 0.9386 0.9795 10.68 0.9665 0.991 0.9713
SVM3:SSA 6.446 0.9743 0.9073 0.9806 10.117 0.967 0.9674 0.972
SVM4:SSA 4.669 0.9757 0.8867 0.9819 10.008 0.9682 0.9106 0.9741
SVM5:SSA 3.668 0.9762 0.7275 0.9823 9.019 0.9704 0.8999 0.9768
SVM6:SSA 1.73 0.9781 0.6991 0.9848 8.119 0.9721 0.8563 0.9774
SVM7:SSA 0.705 0.9798 0.672 0.9854 8.013 0.9728 0.817 0.9786
SVM8:SSA 0.407 0.9806 0.6388 0.9871 7.414 0.9733 0.8005 0.979
SVM9:SSA 0.208 0.9819 0.609 0.988 6.018 0.9749 0.7962 0.9802
SVM10:SSA 0.0769 0.9844 0.5934 0.9902 5.117 0.9756 0.7698 0.9817

for ANN, SVM, SVM-SSA models, respectively, at station Anandpur. tive order. As previously stated, models with R2 value closer to 1
Correspondingly during training phase, the value of R2 is 0.918, are more accurate models for prediction. Likewise, R2 values for
0.9474, and 0.9847 for ANN, SVM, and SVM-SSA models, in respec- runoff projections at Champua gauge station during training–test-

10
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

ing phases are given in Fig. 7 (b). Final prediction outcomes are prediction modeling can provide precise monthly runoff prediction
attained by SVM-SSA model, provide a synergetic impact and results for researchers that can assist as commendable practical
improve performance of hybrid prediction model. On the whole, references in hydrological predictions for global research commu-
combination of swarm intelligence and data decomposition into nity. In a nutshell, our findings prove that SVM and ANN models

Fig. 7. Scatter plots between observed and predicted runoff using ANN, SVM and SVM-SSA models at (a) Anandpur (b) Champua.

11
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

have limitations in correctly simulating runoff when many peak station. For Champua station, runoff values forecasted by ANN,
values (extreme events) are accessed. However hybrid SVM-SSA SVM, and SVM-SSA are, respectively, 287.5368, 296.7959, and
machine learning model was remarkably capable of accurately 310.0233 m3/sec, whereas the observed peak runoff is
simulating extreme events. 314.94 m3/sec. From Fig. 8, it is clear that ANN gives poor runoff
Outputs obtained from soft computing models in training and projections for the whole period of study, whereas SVM model pro-
testing phases were studied against actual range of values. Results vided superior results than ANN. On the other hand, SVM-SSA
reveal that the value of predicted runoff for ANN, SVM, and SVM- showed best performance compared to SVM and ANN in all
SSA are, respectively, 988.2221, 1019.545, and 1059.909 m3/sec, months. It is evident from graphical representations in Fig. 8 that
whereas the observed peak runoff is 1076.38 m3/sec at Anandpur SVM-SSA is the best performing model in runoff estimation

Fig. 8. Actual and predicted runoff time series of (a) Anandpur and (b) Champua stations obtained by ANN, SVM and SVM-SSA.

12
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

Fig. 8 (continued)

amongst proposed machine learning models. Results suggested still yield better model performances when uncertainty in input is
that uncertainty of model parameters is underestimated to some considered. Also, based on Fig. 8, from the starting month to the
extent when impacts of input uncertainty are ignored. An increase end, it is clear that produced runoff results by SVM-SSA model
in uncertainty of model parameters occurs since deviations in are very precise, and there exist lesser error amid observed and
model parameters can be somewhat accommodated by differences predicted runoff. It is also seen that in low-flow conditions, accu-
in inputs of model. Hence, a broader range of parameter values can rate runoff predictions occur. It appears these outcomes specify

13
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

Fig. 9. Box plot for observed and estimated data at (a) Anandpur (b) Champua.

that hybrid approaches have high acceptable performances in pre- For the Anandpur and Champua station, the results (Fig. 11)
diction of low flows. showed that SVM-SSA model presented better results than the
Hence, for testing robustness of proposed models, distribution classical ANN and SVM models. The results indicated that the per-
of prediction results is presented in Fig. 9, which compares robust formance of SVM-SSA in terms of WI and R2 value is superior com-
SVM-SSA model against the standalone SVM and the traditional pared to the SVM and ANN models, respectively. The proposed
ANN models. To assess skewness of distribution and degree of dis- robust SVM-SSA enhances model performance with better accu-
persion, a box plot representation is displayed for runoff predic- racy and more reliability than other hybridization of SVM
tion. It can be observed that even though conventional ANN and approaches [75-77]. The implementation of the models at testing
SVM models attained distributions more skewed in context to stage showed that the accuracy of the models for proposed gauge
actual runoff, hybrid SVM-SSA model showed less skewness than stations decreases as compared to training phase [61,78]. Results
its counterpart models. Hence, results indicate that robust SVM- demonstrated that developed hybrid model provided better pre-
SSA model performs superiorly to simple SVM and ANN models. diction outcomes than many existing techniques with five statisti-
Histograms of predicted runoff and observed runoff for each cal indices. Hence, proposed technique based on parameter
station are shown in Fig. 10 for ANN, SVM, and SVM-SSA models, optimization and data modelling strategies is a favorable alterna-
respectively. It can be understood that prediction ranges by SVM- tive to improve forecasting results of short or long-term runoff pre-
SSA include observed runoff for all conditions, whereas, for ANN, dictions. Also, for water resources planning and management, the
peak runoff constantly fall outside the interim prediction bounds. efficiency of the SVM-SSA model was fairly acceptable to estimate
In field of hydrological modelling, decision makers always aspire the runoff or flow discharge.
for precise predictions with fine prediction intervals when further Application of intelligent optimisation algorithms (SVM-SSA) is
management and analysis is necessary on basis of modelling/fore- a significant parameter optimisation (Liu et al., 2009). On the other
casting outcomes. Based on assessment of results between SVM- hand, certain case studies have shown that accurateness of data
SSA, SVM, and ANN, it can be established that SVM-SSA generally driven models cannot always be improved by utilising optimisa-
produces more reliable and accurate predictions, particularly in tion algorithms [79-81]. This is due to use of optimised combined
long lag-time forecasting. models for comparing with actual models that depend upon expe-

14
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

Fig. 10. Histogram plot of observed vs predicted runoff for ANN, SVM and SVM-SSA at (a) Anandpur (b) Champua.

rience of users regarding manual parameters. In present research, a Therefore, findings from this research signify that performance of
comparison of optimised SVM model (SVM-SSA) is performed since SVM-SSA model is superior compared to that of conventional
parameter arrangement of SVM and ANN models is very restricted. models.
15
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

Fig. 11. Comparison of various model.

5. Conclusion and future recommendation (iii) By comparing SVM-SSA, SVM, and ANN models, results dis-
play that ANN gives worst accuracy; SSA significantly and
This research explored usability of three data-driven tech- expressively increases accuracy of monthly runoff predic-
niques, ANN, SVM, and hybrid SVM-SSA, for runoff prediction. tion. Results of this study shed new light on the role of
Firstly, proposed techniques were applied, taking into considera- SSA in runoff prediction by data-driven models.
tion several input scenarios, and then using cross validation, more (iv) The study showed that proposed data-driven techniques
robust predictions were obtained. Assessment of results provided (e.g., SVM-SSA, SVM, ANN) which utilise a limited number
following inferences: of parameters could be convenient for ungauged watersheds
where information regarding runoff is not available for cali-
(i) It was evident that models with only preceding rainfall val- brating advanced physical-based models.
ues as input were the least accurate models without other
input variables. However, there still exist certain limitations that need to be
(ii) Inclusion of additional variables considerably enriched the addressed in future study. For future researches, more data from
models accuracies by remarkably decreasing RMSE and various basins can be utilised to evaluate efficiency of proposed
MSE and increasing NSE. Among all proposed data-driven methods. Then again, rainfall-runoff process is mostly affected by
models, SVM-SSA showed better accuracy than SVM and different aspects of human society and natural environments, pre-
ANN. senting complex, vibrant evolutionary characters [82-85]. Hence, a

16
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

solitary prediction model may fail to provide satisfactory outcomes [26] Samantaray S, Sahoo A. Prediction of runoff using BPNN, FFBPNN, CFBPNN
algorithm in arid watershed: A case study. Int J Knowledge-Based Intell Eng
in certain cases. For enhancing performance of proposed technique,
Syst 2020;24. doi: https://doi.org/10.3233/KES-200046.
a number of research directions justify careful consideration in [27] Sahoo A, Samantaray S, Ghose DK. Stream flow forecasting in mahanadi river
future. Firstly combination of effective post-processing or data- basin using artificial neural networks. Procedia Comput. Sci., vol. 157, 2019.
processing methods into prediction process of developed method. https://doi.org/10.1016/j.procs.2019.08.154.
[28] Sridharam S, Sahoo A, Samantaray S, Ghose DK. Assessment of Flow Discharge
Secondly, development of more efficient search strategies for in a River Basin Through CFBPNN, LRNN and CANFIS. vol. 134. 2021.
improving local exploration and global exploitation abilities of https://doi.org/10.1007/978-981-15-5397-4_78.
SSA technique for network optimization. [29] Kisi O, Nia AM, Gosheh MG, Tajabadi MRJ, Ahmadi A. Intermittent streamflow
forecasting by using several data driven techniques. Water Resour Manag
2012;26:457–74.
[30] Sahoo A, Singh UK, Kumar MH, Samantaray S. Estimation of Flood in a River
Declaration of Competing Interest
Basin Through Neural Networks: A Case Study. vol. 134. 2021. https://doi.org/
10.1007/978-981-15-5397-4_77.
The authors declare that they have no known competing finan- [31] Wu J, Liu H, Wei G, Song T, Zhang C, Zhou H. Flash flood forecasting using
cial interests or personal relationships that could have appeared support vector regression model in a small mountainous catchment. Water
2019;11:1327.
to influence the work reported in this paper. [32] Samantaray S, Sahoo A, Mohanta NR, Biswal P, Das UK. Runoff Prediction Using
Hybrid Neural Networks in Semi-Arid Watershed, India: A Case Study. vol.
134. 2021. https://doi.org/10.1007/978-981-15-5397-4_74.
References [33] Zaini N, Malek MA, Yusoff M, Mardi NH, Norhisham S. Daily river flow
forecasting with hybrid support vector machine–particle swarm optimization.
[1] Li H, Zhang Y, Zhou X. Predicting surface runoff from catchment to large region. IOP Conf. Ser. Earth Environ. Sci., vol. 140, IOP Publishing Ltd.; 2018, p. 1315–
Adv Meteorol 2015;2015. 755.
[2] Seyam M, Othman F, El-Shafie A. Prediction of stream flow in humid tropical [34] Bell B, Wallace B, Zhang D. Forecasting river runoff through support vector
rivers by support vector machines. MATEC Web Conf., vol. 111, EDP Sciences; machines. 2012 IEEE 11th Int. Conf. Cogn. Informatics Cogn. Comput., IEEE;
2017, p. 1007. 2012, p. 58–64.
[3] Hu C, Wu Q, Li H, Jian S, Li N, Lou Z. Deep learning with a long short-term [35] Okkan U, Serbes ZA. Rainfall–runoff modeling using least squares support
memory networks approach for rainfall-runoff simulation. Water vector machines. Environmetrics 2012;23:549–64.
2018;10:1543. [36] Ebrahimi H, Rajaee T. Simulation of groundwater level variations using
[4] Keteklahijani VK, Alimohammadi S, Fattahi E. Predicting changes in monthly wavelet combined with neural network, linear regression and support vector
streamflow to Karaj dam reservoir, Iran, in climate change condition and machine. Glob Planet Change 2017;148:181–91.
assessing its uncertainty. Ain Shams Eng J 2019;10(4):669–79. [37] Samantaray S, Sahoo A, Ghose DK. Assessment of Sediment Load Concentration
[5] Wang W. Stochasticity, nonlinearity and forecasting of streamflow processes. Using SVM, SVM-FFA and PSR-SVM-FFA in Arid Watershed, India: A Case
Ios Press 2006. Study. KSCE J Civ Eng n.d.:1–14.
[6] Seo Y, Kim S, Singh VP. Machine learning models coupled with variational [38] Sahoo A, Barik A, Samantaray S, Ghose DK. Prediction of Sedimentation in a
mode decomposition: a new approach for modeling daily rainfall-runoff. Watershed Using RNN and SVM. vol. 134. 2021. https://doi.org/10.1007/978-
Atmosphere (Basel) 2018;9:251. 981-15-5397-4_71.
[7] Doycheva K, Horn G, Koch C, Schumann A, König M. Assessment and weighting [39] Samantaray S, Sahoo A. Assessment of Sediment Concentration Through RBNN
of meteorological ensemble forecast members based on supervised machine and SVM-FFA in Arid Watershed, India. vol. 159. 2020. https://doi.org/10.1007/
learning with application to runoff simulations and flood warning. Adv Eng 978-981-13-9282-5_67.
Informatics 2017;33:427–39. [40] Misra D, Oommen T, Agarwal A, Mishra SK, Thompson AM. Application and
[8] Lin G-F, Jhong B-C, Chang C-C. Development of an effective data-driven model analysis of support vector machine based simulation for runoff and sediment
for hourly typhoon rainfall forecasting. J Hydrol 2013;495:52–63. yield. Biosyst Eng 2009;103:527–35.
[9] Choi C, Kim J, Han H, Han D, Kim HS. Development of water level prediction [41] Goyal MK, Sharma A, Katsifarakis KL. Prediction of flow rate of karstic springs
models using machine learning in wetlands: A case study of Upo wetland in using support vector machines. Hydrol Sci J 2017;62:2175–86.
South Korea. Water 2020;12:93. [42] Sharma N, Zakaullah M, Tiwari H, Kumar D. Runoff and sediment yield
[10] Fan H, Jiang M, Xu L, Zhu H, Cheng J, Jiang J. Comparison of long short term modeling using ANN and support vector machines: a case study from Nepal
memory networks and the hydrological model in runoff simulation. Water watershed. Model Earth Syst Environ 2015;1:1–8.
2020;12:175. [43] Sedighi F, Vafakhah M, Javadi MR. Rainfall–runoff modeling using support
[11] Basheer AA. Chemical chiral pollution: impact on the society and science and vector machine in snow-affected watershed. Arab J Sci Eng 2016;41:4065–76.
need of the regulations in the 21st century. Chirality 2018;30(4):402–6. [44] Sharifi A, Dinpashoh Y, Mirabbasi R. Daily runoff prediction using the linear
[12] Basheer AA. New generation nano-adsorbents for the removal of emerging and non-linear models. Water Sci Technol 2017;76:793–805.
contaminants in water. J Mol Liq 2018;261:583–93. [45] Tașar B, Unes F, Varcin H. Prediction of the Rainfall-Runoff Relationship Using
[13] Basheer AA, Ali I. Stereoselective uptake and degradation of (±)-o, p-DDD NeuroFuzzy and Support Vector Machines. 2019” Air Water-Components
pesticide stereomers in water-sediment system. Chirality 2018;30 Environ Conf Proc 2019:237–46.
(9):1088–95. [46] Bafitlhile TM, Li Z. Applicability of e-support vector machine and artificial
[14] Ali I, Jain CK. Groundwater contamination and health hazards by some of the neural network for flood forecasting in humid, Semi-Humid and Semi-Arid
most commonly used pesticides. Curr Sci 1998;75(10):1011–4. Basins in China. Water 2019;11:85.
[15] Ali I, Khan TA, Hussain I. Treatment and remediation methods for arsenic [47] Babaei M, Moeini R, Ehsanzadeh E. Artificial neural network and support
removal from the ground water. Int J Environmental Eng 2011;3(1):48–71. vector machine models for inflow prediction of dam reservoir (case study:
[16] Kim J, Han H, Johnson LE, Lim S, Cifelli R. Hybrid machine learning framework Zayandehroud Dam Reservoir). Water Resour Manag 2019;33:2203–18.
for hydrological assessment. J Hydrol 2019;577:123913. [48] Poonia V, Tiwari HL. Rainfall-runoff modeling for the Hoshangabad Basin of
[17] Liang J, Li W, Bradford SA, Šimůnek J. Physics-informed data-driven models to Narmada River using artificial neural network. Arab J Geosci 2020;13:1–10.
predict surface runoff water quantity and quality in agricultural fields. Water [49] Hassan Z, Rosdi SZ, Kamarudzaman AN, Rahim MA, Ghazaly ZM. Comparison
2019;11:200. of Artificial Neural Network and Support Vector Machine for Long-Term
[18] Mosavi A, Ozturk P, Chau K. Flood prediction using machine learning models: Runoff Simulation. IOP Conf. Ser. Earth Environ. Sci., vol. 476, IOP Publishing;
Literature review. Water 2018;10:1536. 2020, p. 12119.
[19] McCulloch WS, Pitts W. A logical calculus of the ideas immanent in nervous [50] Han H, Morrison RR. Data-driven approaches for runoff prediction using
activity. Bull Math Biophys 1943;5:115–33. distributed data. Stoch Environ Res Risk Assess 2021:1–19.
[20] Vapnik VN. The nature of statistical learning. Theory 1995. [51] Rao R. Jaya: A simple and new optimization algorithm for solving constrained
[21] Breiman L. Random forests. Mach Learn 2001;45:5–32. and unconstrained optimization problems. Int J Ind Eng Comput
[22] Kassem AA, Raheem AM, Khidir KM, Alkattan M. Predicting of daily Khazir 2016;7:19–34.
basin flow using SWAT and hybrid SWAT-ANN models. Ain Shams Eng J [52] Mirjalili S. The ant lion optimizer. Adv Eng Softw 2015;83:80–98.
2020;11(2):435–43. [53] Mirjalili S, Lewis A. The whale optimization algorithm. Adv Eng Software
[23] Khadr M, Elshemy M. Data-driven modeling for water quality prediction case 2016;95:51–67.
study: The drains system associated with Manzala Lake Egypt. Ain Shams Eng J [54] Maroufi H, Mehdinejadiani B. A comparative study on using metaheuristic
2017;8(4):549–57. algorithms for simultaneously estimating parameters of space fractional
[24] Latif SD, Birima AH, Ahmed AN, Hatem DM, Al-Ansari N, Fai CM, et al. advection-dispersion equation. J Hydrol 2021;602:126757.
Development of prediction model for phosphate in reservoir water system [55] Mehdinejadiani B. Estimating the solute transport parameters of the spatial
based machine learning algorithms. Ain Shams Eng J 2021. fractional advection-dispersion equation using Bees Algorithm. J Contam
[25] Jimmy SR, Sahoo A, Samantaray S, Ghose DK. Prophecy of Runoff in a River Hydrol 2017;203:51–61.
Basin Using Various Neural Networks. vol. 134. 2021. https://doi.org/10.1007/ [56] Mirjalili S, Mirjalili SM, Lewis A. Grey wolf optimizer. Adv Eng Softw
978-981-15-5397-4_72. 2014;69:46–61.

17
S. Samantaray, S. Sawan Das, A. Sahoo et al. Ain Shams Engineering Journal 13 (2022) 101732

[57] Saremi S, Mirjalili S, Lewis A. Grasshopper optimisation algorithm: theory and [85] Liu Y, Freer J, Beven K, Matgen P. Towards a limits of acceptability approach to
application. Adv Eng Softw 2017;105:30–47. the calibration of hydrological models: Extending observation error. J Hydrol
[58] Komasi M, Sharghi S. Hybrid wavelet-support vector machine approach for 2009;367:93–103.
modelling rainfall–runoff process. Water Sci Technol 2016;73:1937–53.
[59] Tayyab M, Zhou J, Dong X, Ahmad I, Sun N. Rainfall-runoff modeling at Jinsha
River basin by integrated neural network with discrete wavelet transform.
Meteorol Atmos Phys 2019;131:115–25. Sandeep Samantaray received B.Tech in civil engi-
[60] Zhang J, Wang Z, Luo X. Parameter estimation for soil water retention curve neering from Biju Patnaik University of Technology,
using the salp swarm algorithm. Water 2018;10:815. Odisha in 2013, and M.Tech in water resources engi-
[61] Feng Z, Niu W, Tang Z, Jiang Z, Xu Y, Liu Y, et al. Monthly runoff time series neering from Veer Surendra Sai University of Technol-
prediction by variational mode decomposition and support vector machine ogy, Odisha in 2016. He is currently working as a Ph.D.
based on quantum-behaved particle swarm optimization. J Hydrol student in the Department of Civil engineering at the
2020;583:124627. National Institute of Technology, Silchar, Assam, India.
[62] Nguyen H-D, Pham V-D, Nguyen Q-H, Pham V-M, Pham MH, Vu VM, et al. An His research interests focus on watershed management,
optimal search for neural network parameters using the Salp swarm hydrological forecasting, hydrologic modeling, and
optimization algorithm: a landslide application. Remote Sens Lett developing sustainable means of managing the envi-
2020;11:353–62. ronment. He has published 27 international journals
[63] Dehghani R, Poudeh HT, Younesi H, Shahinejad B. Daily streamflow prediction
and six book chapters and attended around 35 inter-
using support vector machine-artificial flora (SVM-AF) hybrid model. Acta
national conferences (SCOPUS Indexing). He also published 2nos Indian patents. He
Geophys 2020;68:1763–78.
[65] Wu Y, Wang Q, Li G, Li J. Data-driven runoff forecasting for Minjiang River: a also authored two book published in Taylor & Francis.
case study. Water Supply 2020;20:2284–95.
[66] Faris H, Mafarja MM, Heidari AA, Aljarah I, Ala’M A-Z, Mirjalili S, et al. An
efficient binary salp swarm algorithm with crossover scheme for feature Sambit Sawan Das received B.Tech in civil engineering
selection problems. Knowledge-Based Syst 2018;154:43–67. from Parala Majharaja Engineering College, Berhampur,
[67] Lee S, Ryu J-H, Lee M-J, Won J-S. The application of artificial neural networks to Odisha in 2018, and M.Tech in water resources engi-
landslide susceptibility mapping at Janghung. Korea. Math Geol neering from College of Engineering and Technology,
2006;38:199–220. Bhubaneswar, Odisha in 2021. His research interests,
[68] Samantaray S, Sahoo A, Ghose DK. Assessment of runoff via precipitation using including hydrological forecasting, hydrologic model-
neural networks: Watershed modelling for developing environment in arid ing, watershed management
region. Pertanika. J Sci Technol 2019;27.
[69] Sahoo A, Singh UK, Kumar MH, Samantaray S. Estimation of Flood in a River
Basin Through Neural Networks: A Case Study. Singapore: Springer; 2021. p.
755–63.
[70] De Vos NJ, Rientjes THM. Constraints of artificial neural networks for rainfall-
runoff modelling: trade-offs in hydrological state representation and model
evaluation. Hydrol Earth Syst Sci 2005;9:111–26.
[71] Adeloye AJ, De Munari A. Artificial neural network based generalized storage–
yield–reliability models using the Levenberg–Marquardt algorithm. J Hydrol
2006;326:215–30. Abinash Sahoo received B.Tech in civil engineering
[72] Campisi-Pinto S, Adamowski J, Oron G. Forecasting urban water demand via from Siksha ’O’ Anusandhan University, Odisha in 2014,
wavelet-denoising and neural network models. Case study: city of Syracuse, and M.Tech in water resources engineering from Veer
Italy. Water Resour Manag 2012;26:3539–58. Surendra Sai University of Technology, Odisha in 2016.
[75] Kwon M, Kwon H-H, Han D. A Hybrid Approach Combining Conceptual He is currently working as a Ph.D. scholar in the
Hydrological Models, Support Vector Machines and Remote Sensing Data for Department of Civil engineering, National Institute of
Rainfall-Runoff Modeling. Remote Sens 2020;12:1801. Technology Silchar, Assam, India. He has authored or co-
[76] Liang Z, Li Y, Hu Y, Li B, Wang J. A data-driven SVR model for long-term runoff
authored more than 30 refereed journal articles and
prediction and uncertainty analysis based on the Bayesian framework. Theor
conference papers in his research interests, including
Appl Climatol 2018;133:137–49.
hydrological forecasting, hydrologic modeling, water-
[77] Kan G, Li J, Zhang X, Ding L, He X, Liang K, et al. A new hybrid data-driven
model for event-based rainfall–runoff simulation. Neural Comput Appl shed management, applications of AI, and optimization
2017;28:2519–34. techniques. He also co-authored a book published in
[78] Roushangar K, Alizadeh F, Nourani V. Improving capability of conceptual Taylor & Francis and one no. of Indian patent.
modeling of watershed rainfall–runoff using hybrid wavelet-extreme learning
machine approach. J Hydroinformatics 2018;20:69–87.
[79] Goyal MK, Bharti B, Quilty J, Adamowski J, Pandey A. Modeling of daily pan Deba Prakash Satapathy is currently working as an
evaporation in sub tropical climates using ANN, LS-SVR, Fuzzy Logic, and Associate professor at College of Engineering and
ANFIS. Expert Syst Appl 2014;41:5267–76. Technology, Bhubaneswar, obtained his Bachelor’s
[80] Piri J, Mohammadi K, Shamshirband S, Akib S. Assessing the suitability of degree in Civil Engineering from NIT, Rourkela, Odisha.
hybridizing the Cuckoo optimization algorithm with ANN and ANFIS
He has published 3 research publications, 5 conference
techniques to predict daily evaporation. Environ Earth Sci 2016;75:246.
papers in international conferences. He also is involved
[81] Shamshirband S, Amirmojahedi M, Gocić M, Akib S, Petković D, Piri J, et al.
in many consultancy areas like bridge design, soil
Estimation of reference evapotranspiration using neural networks and cuckoo
search algorithm. J Irrig Drain Eng 2016;142:4015044. investigation, water harvesting, and ground water
[82] Pang R, Xu B, Kong X, Zou D, Zhou Y. Seismic reliability assessment of earth- detection. He presently guiding 3 Ph.D students and
rockfill dam slopes considering strain-softening of rockfill based on successfully guided 31 master’s students and on-going 6
generalized probability density evolution method. Soil Dyn Earthq Eng
2018;107:96–107.
[83] Zhu F, Zhong P, Sun Y, Yeh WW. Real-time optimal flood control decision
making and risk propagation under multiple uncertainties. Water Resour Res
2017;53:10635–54.
[84] Liu D, Guo S, Shao Q, Liu P, Xiong L, Wang L, et al. Assessing the effects of
adaptation measures on optimal water resources allocation under varied
water availability conditions. J Hydrol 2018;556:759–74.

18

You might also like