RAINFALL-RUNOFF MODELLING USING ARTIFICIAL

NEURAL NETWORK - A CASE STUDY OF BANAS RIVER

CATCHMENT

By
PARMAR GIRISHBHAI SENDHABHAI
Enrollment Number: 190430733014

Guided By
PROF. P. S. RAMANUJ
Associate Professor

A thesis submitted to
Gujarat Technological University in partial fulfilment of the requirement for
the Master of Engineering degree in
Civil Engineering [Water Resources Engineering]

MAY 2022

CIVIL ENGINEERING DEPARTMENT

SHANTILAL SHAH ENGINEERING COLLEGE SIDSAR,
BHAVNAGAR
 CERTIFICATE

This is to certify that the research work embodied in this dissertation titled
“Rainfall-Runoff modelling using Artificial Neural Network- A case
study of Banas River catchment” was carried out by Mr. Girishbhai
Sendhabhai Parmar (Enrollment No. 190430733014) at Shantilal Shah
Engineering College, Bhavnagar (043) for partial fulfilment Master of
Engineering degree in Water Resources Engineering to be awarded by
Gujarat Technological University. This research work has been carried out
under my guidance and supervision and it is up to my satisfaction.

Date:
Place: Bhavnagar

Prof. P. S. Ramanuj Dr. V.M. Patel

Associated Professor, Head of Department
Civil Engg. Department Civil Engg. Department
SSGEC, Bhavnagar SSEGC, Bhavnagar
(Internal Guide)
PRINCIPAL
Dr M.G.Bhatt
Shantilal Shah Engineering College, Bhavnagar

Seal of Institute

ii
 COMPLIANCE CERTIFICATE

This is to certify that research work embodied in this thesis entitled

“Rainfall-runoff modelling using Artificial Neural Network: A case study
of Banas River catchment” has been carried out by Mr. Girishbhai
Sendhabhai Parmar (Enrollment No.190430733014) at Shantilal Shah
Engineering college, Bhavnagar (043) for Partial fulfillment of Master of
Engineering degree in Civil (Water Resource Engineering) to be awarded by
Gujarat Technological University. He has complied with the comments given
by the Dissertation phase – I as well as mid semester thesis Reviewer to my
satisfaction.

Date:
Place: Bhavnagar

Prepared by
Girishbhai Sendhabhi Parmar

Prof. R.S.RAMANUJ Prof. Dr. V. M. PATEL

Associate Professor Head of Department
Civil Engg. Dept. Civil Engg.Dept.
S. S. Engg. College S. S. Engg. College
Bhavnagar Bhavnagar

PRINCIPAL
Dr M.G.Bhatt
Shantilal Shah Engineering College, Bhavnagar

Seal of Institute

iii
 THESIS APPROVAL CERTIFICATE

This is to certify that research work embodied in this thesis entitled

“Rainfall-runoff modelling using Artificial Neural Network: A case study
of Banas River catchment” was carried out by Mr. Girishbhai Sendhabhai
Parmar (Enrollment No.190430733014) at Shantilal Shah Engineering
college, Bhavnagar (043) is approved for the degree of Master of Engineering
with specialization of (Water Resource Engineering) by Gujarat
Technological University.

Date:
Place:

Examiner’s Sign and Name:

_____________ _____________
( ) ( )

iv
 DECLARATION OF ORIGINALITY OF WORK

We hereby certify that we are the sole authors of this thesis and that neither
any part of this thesis nor the whole of the thesis has been submitted for a
degree to any other University or Institution. We certify that, to the best of
our knowledge, the current thesis does not infringe upon anyone’s copyright
nor violate any proprietary rights and that any ideas, techniques, quotations or
any other material from the work of other people included in our thesis,
published or otherwise, are fully acknowledged in accordance with the
standard referencing practices. Furthermore, to the extent that we have
included copyrighted material that surpasses the boundary of fair dealing
within the meaning of the Indian Copyright (Amendment) Act 2012, we
certify that we have obtained a written permission from the copyright
owner(s) to include such material(s) in the current thesis and have included
copies of such copyright clearances to our appendix.
We declare that this is a true copy of thesis, including any final revisions, as
approved by thesis review committee.
We have checked write up of the present thesis using anti-plagiarism database
and it is in allowable limit. Even though later on in case of any complaint
pertaining of plagiarism, we are sole responsible for the same and we
understand that as per UGC norms, University can even revoke Master of
Engineering degree conferred to the student submitting this thesis.

Date:
Place: Bhavnagar

Girishbhai Sendhabhai Parmar Prof. P.S Ramanuj

(Enrollment No: 190430733014) Associate Professor
Civil Engineering Department
Shantilal Shah Engg. College,
Bhavnagar

v
DEDICATED TO
MY PARENTS

vi
 ACKNOWLEDGEMENT

I am glad to offer my sincerest gratitude and respect to my supervisor, Prof. P.

S. Ramanuj, for his guidance from the foundation stage of this research to the
entire dissertation work. I truly appreciate his esteemed guidance and
encouragement throughout the work. I consider myself extremely fortunate to
have had the opportunity of associating myself with him.

I am also thanked Amaratbhai for his support, who is currently working at the
Dantiwada dam.

Last but not the least, I would like to thank my family: my parents and my
brother and sister-in-law for supporting me throughout the dissertation work.

Girishbhai Sendhabhai Parmar

vii
 TABLE OF CONTENTS

Title Page ⅰ
Certificate Page ii
Compliance certificate iii
Thesis approval certificate iv
Declaration of original work v
Dedication page vi
Acknowledgement vii
Table of contents viii
List of figures ix
List of tables xiii
Abstract 1
Chapter-1 Introduction 2
1.1 General 3
1.2 Hydrological cycle 3
1.2.1 Indian Rainfall 4
1.3 Runoff 4
1.3.1 River Runoff 4
1.3.2 Types of Runoffs 5
1.3.3 Factors Affecting Runoff 5
1.4 Rainfall-Runoff Model 7
1.4.1 Empirical model 7
1.4.2 Conceptual Model 8
1.4.3 Physically Based Model 8
1.5 Aim 8
1.6 Scope of the study 9
1.7 Objectives 9

viii
Chapter-2 Literature Review 10
Chapter-3 Study Area & Data Collection 17
3.1 General 18
3.2 Banas River 18
3.3 Dantiwada Dam 22
3.3.1 Characteristic of Dantiwada Dam 23
3.4 Data Collection 26
Chapter-4 Introduction of ANN & MLR 27
4.1 General 28
4.1.1 Neurons 28
4.1.2 ANN 29
4.1.3 Network Architecture 31
4.2 Backpropagation Neural Network 32
4.3 Feedforward Neural Network 34
4.4 Activation Functions 35
4.4.1 Log sigmoid function 35
4.4.2 Threshold function 36
4.4.3 Linear function 37
4.5 Training algorithm 37
4.6 Multiple Linear Regression (MLR) 38
4.6.1 Statistics for MLR 39
Chapter-5 Methodology 40
5.1 General 41
5.2 Flow Chart for Methodology 42
5.3 Model development 46
5.4 Sensitivity Analysis 57
Chapter-6 Result and Analysis 58
6.1 General 59

ix
 6.2 Regression & MSE graph for training, testing and validation 59
6.3 ANN model Analysis 64
6.4 Comparison of predicted data with actual data 65
6.5 Multiple Linear Regression model 66
6.6 Comparison between ANN & MLR model 67
Chapter-7 Conclusion 69
References 71

LIST OF FIGURES

x
Fig 3.1 Location of Entire BANAS River Catchment (Source: CWC) 19

Fig 3.2 Location of Banas River Catchment (Study area Map, Source: ArcGIS) 20

Fig 3.3 BANAS River catchment Map Up to Dantiwada Dam (Source: ArcGIS) 21

Fig 4.1 Structure of Neuron 29

Fig 4.2 Artificial Neural Network Structure 30

Fig 4.3 Activation of a single neuron 31

Fig 4.4 Structure of the Back-propagation Network 32

Fig 4.5 A Feed Forward Neural Network 34

Fig 4.6 log sigmoid function 35

Fig 4.7 binary threshold function 36

Fig 4.8 bipolar threshold function 36

Fig 4.9 linear function 37

Fig 5.1 datasheet containing all the available data 47
Fig 5.2 normalized datasheet 48

Fig 5.3 input data in MATLAB software 48

Fig 5.4 output value in MATLAB software 49

Fig 5.5 Neurons vs Accuracy chart 51

Fig 5.6 Neurons vs Accuracy chart 52

Fig 5.7 NN tool image from MATLAB 53

Fig 5.8 regression plot 54

Fig 5.9 MSE value 55

Fig 5.10 neural network results 56

Fig 5.11 Relative importance of various parameters 57

Fig 6.1 Training, validation & testing regression for the LM algorithm 60

xi
Fig 6.2 MSE for LMA training, validation and testing algorithm 61

Fig 6.3 Training, validation & testing regression for a SCG algorithm 62

Fig 6.4 MSE for SCG training, validation and testing algorithm 63

Fig 6.5 scatter plot of actual vs predicted values (Normalized) 65

Fig 6.6 Linear regression model statistics 67

LIST OF TABLES

xii
Table 3.1 Source of the distributaries 22

Table 3.2 General information of Dantiwada Dam 23

Table 3.3 Details of the Dantiwada Dam section 24

Table 3.4 Details of Reservoir of Dantiwada Dam 24

Table 3.5 Details of Spillway 25

Table 3.6 Details of Canal 25

Table 3.7 Villages under command of Dantiwada Dam 25

Table 3.8 Characteristics of CWC rain gauge Station 26

Table 5.1 training variables and their allotted values for the ANN model 43

Table 5.2 Dataset list 46

Table 5.2 Network architecture of ANN model 1 with LM algorithm
[training 70%, testing 15%, validation 15%] 51
Table 5.3 Network architecture of ANN model 2 with SCG algorithm
[training 70%, testing 15%, validation 15%] 52
Table 6.1 Performance evaluation of model on training and testing

period for ANN model 64

Table 6.2 Regression statistics for MLR model 66

Table 6.3 comparison of R and RMSE values of both models 67

xiii
 RAINFALL-RUNOFF MODELLING USING ARTIFICIAL
NEURAL NETWORK- A CASE STUDY OF BANAS RIVER
CATCHMENT
Enrollment No. 190430733014
Girishbhai Sendhabhai Parmar
Shantilal Shah Engineering college (043)
girishparmar625@gmail.com
ABSTRACT
In the last decades, artificial neural networks (ANNs) have become one of the most
promising tools for modelling complex hydrological processes such as the rainfall-runoff
process. Artificial Neural Networks (ANN) can be used in cases where the available data is
limited. The ANN model developed in this study is applied to the Banas River catchment
of Gujarat, India. The present work involves the development of an ANN model using the
Feed-Forward Backpropagation algorithm and a multiple linear regression model for
establishing rainfall-runoff relations. The ANN model and the MLR model were assessed
by correlation coefficient (R), R2, and mean square error (MSE) and later the result of the
ANN model was compared with the MLR model. The results show that the ANN model
gives good results than the MLR model. Hence the present study suggests that ANN
models are an essential tool for predicting the hydrological responses in the Banas River
Catchment.

Keywords: Artificial Neural Network (ANN), Multiple Linear Regression model, Feed-
forward backpropagation neural network, rainfall-runoff modelling, Banas River
catchment.

1
 Chapter 1
Introduction

2
 CHAPTER 1
INTRODUCTION

1.1 GENERAL
Rainfall-Runoff models are widely used in hydrology for a range of applications
and play an important role in the optimal planning and management of water resources in
the river basin. Managing risk in the short term is one of the most important applications of
R-R modelling. A rainfall-runoff model is, by definition, a simplification of a complex,
non-linear, time and space varying hydrological process reality. Such a model contains
parameters that cannot often be measured directly but can only be estimated by calibration
against a historical record of measured data. The system input and output data are often
contaminated by measurement errors. The system input data such as rainfall, temperature,
etc. and output are often contaminated by measured errors. Consequently, predictions
made by such a rainfall-runoff model are far from perfect, in other words, there always
exists a distinction between the model prediction and the corresponding observed data, no
matter how precise the model is and how perfectly the model is calibrated. Thus, the model
errors which are the mismatch between the observed and the simulated system behaviour
are unavoidable in rainfall-runoff modelling.

1.2 HYDROLOGICAL CYCLE

Water occurs on the earth in all its three states, viz. liquid, solid, and gaseous, and in
various degrees of motion. Water moves from one reservoir to another by way of processes
like Evaporation, condensation, precipitation, deposition, runoff, infiltration, sublimation,
transpiration, melting, and groundwater flow. The oceans supply most of the evaporated
water found in the atmosphere. Of this evaporated water, only 91% of it is returned to the
way of ocean basins by way of precipitation. Evaporation of water from water bodies such
as oceans and lakes, formation and movement of clouds, rain and snowfall, streamflow,
and groundwater movement are some examples of the dynamic aspects of water. The
various aspects of water related to the earth can be explained in terms of a cycle known as
the hydrological cycle.

3
1.2.1 Indian Rainfall

Precipitation is any form of solid or liquid water that falls from the atmosphere to the
earth’s surface. Rain, drizzle, hail and snow are examples of precipitation. Generally, the
precipitation in the form of water is called rainfall. In India, rain is the most common form
of rainfall.

India is a tropical country, and its precipitation varies considerably in frequency,

intensity, as well as in duration. The total precipitation occurs either as rainfall or as
snowfall. The average annual rainfall for the entire country is estimated as 117 cm. The
largest amount of precipitation falls as rain during the monsoons; whereas, a moderate
amount mainly as snow, falls during the winter season as winter precipitation.

Besides these, two major precipitation periods occur as a minor quantity of rainfalls,
either as post-monsoon rain or as pre-monsoon rain.

1.3 RUNOFF

Rainfall is one of the significant phases of the hydrologic cycle and is the primary
source of runoff. Runoff can be defined as the portion of the precipitation that makes its
way towards rivers, oceans, etc. For given precipitation, the evapotranspiration, initial loss,
infiltration and detention storage requirements will have to be first satisfied before the
commencement of runoff. Its occurrence and quantity are dependent on the characteristics
of the rainfall event, i.e., intensity, duration, and distribution. Surface runoff can be
generated either by rainfall or snowfall (snowmelt contributes to the streamflow it can be
included with the direct runoff, from rainfall).

1.3.1 River Runoff

Runoff is that portion of the rainfall or irrigation water that leaves a field either as
surface or as subsurface flow. When rainfall intensity reaches the soil, the surface is less
than the infiltration capacity, all the water is absorbed into the soil. As rain continues, the
soil becomes saturated and infiltration capacity is reduced, shallow depression fills with
water, and the overland flow starts as runoff.

4
1.3.2 Types of Runoff

There are three types of runoffs depending on the source: surface flow, interflow, and base
flow are described below;

1. Surface Runoff: The portion of rainfall that enters the stream immediately after the
rainfall. It occurs when all loss is satisfied and rainfall is continued and the rate of
rainfall (intensity) is greater than the infiltration rate.
2. Sub-Surface Runoff (Interflow): That part of the rainfall which first leaches into the
soil and moves laterally without joining the water table, to the stream, river or ocean is
known as sub-surface runoff. It is usually referred to as inter-flow.
3. Baseflow (Ground Water flow): It is delayed flow defined as that part of rainfall,
which after falling on the ground surface, infiltrated into the soil and meets the water
table and flows to the streams, ocean etc. The movement of water in this is very slow.
Therefore, it is also referred to as delayed runoff.

1.3.3 Factors Affecting Runoff

The various factors, which affect the runoff from a drainage basin depend upon the two
main characteristics:

(I) Meteorological characteristics

Type of precipitation (rain, snow, sleet, etc.)

Rainfall intensity
Rainfall amount
Rainfall duration
Distribution of rainfall over the watersheds
Other meteorological and climatic conditions affect evapotranspiration,
such as temperature, wind, relative humidity, and season.

(II) Physical characteristics

Catchment size
Land use
Vegetation
Soil type

5
 Drainage area Basin shape
Topography
 The major factors are described below in brief;

Drainage area Basin shape: for equal size watersheds, runoff decreases as overland flow
length. Longer duration storms, needed to produce runoff from all points in the watershed,
have lower average intensity.

Topography: surface slopes and roughness greatly influence runoff. Steep slopes reduce
the time of concentration and detention time. Roughness increases surface storage and
promotes greater infiltration, both of which decrease runoff.

Soil types: watershed soils influence infiltration and deep storage rates. Infiltration must
be satisfied before runoff begins. In the case of clay soil, low infiltration capacity and
runoff will be more. In the case of sand, a high infiltration rate and runoff will be less.

Vegetation: Vegetation cover reduces the runoff in smaller storms.

Catchment size: The runoff per unit area decreases as watershed size increases for a fixed
return interval. This occurs primarily because the average rainfall amount decreases with
increasing area. Secondary: increased travel time for runoff allows more infiltration and
other losses.

Land use: Land use and land management practices have great effects on the runoff yield.
E.g., an area with forest cover and grasses contributes less runoff because water is
absorbed more into the soil.

Types of Precipitation: It has a great effect on the runoff. E.g., A precipitation that occurs
in the form of rainfall starts immediately as surface runoff depending upon rainfall
intensity while precipitation in the form of snow does not result in surface runoff.

Rainfall Intensity: If the rainfall intensity is greater than the infiltration rate of soil then
runoff starts immediately after rainfall. While in case of low rainfall intensity runoff starts
later. Thus, high intensities of rainfall yield higher runoff.

Duration of Rainfall: It is directly related to the volume of runoff because soil infiltration
rate decreases with the duration of rainfall. Therefore, medium intensity rainfall even
results in a considerable amount of runoff if the duration is longer.

6
Rainfall Distribution: Runoff from a watershed depends very much on the distribution of
rainfall. It is also expressed as a “distribution coefficient”. Near the outlet of the
watershed, a runoff will be more.

Other Climate Factors: Other factors such as temperature wind velocity, relative
humidity, annual rainfall, etc. affect the water losses from the watershed area.

1.4 RAINFALL-RUNOFF MODELLING

Rainfall-runoff models describe a portion of the water cycle. Modelling runoff helps
gain a better understanding of hydrologic phenomena and how changes affect the
hydrological cycle. Runoff models visualize what occurs in water systems due to changes
in previous surfaces, vegetation, and meteorological events. Rainfall-runoff modelling can
be carried out within a purely analytical framework based on observations of a catchment
area’s inputs and outputs. Before undertaking rainfall-runoff modelling for a particular
storm, assessing the quantity and quality of available data is advisable. Quite often, the
available data dictate the type of model to be used more than the problem itself.

There has been extensive research on the rainfall-runoff relationship with different
models that can be classified into three main groups, namely fully distributed physically-
based models, lumped conceptual models, and black-box models.

1.4.1 Empirical model

These contain no physically based transfer function to relate input to output: in
other words, no consideration of the physical processes is involved. Such models usually
depend upon establishing a relationship between input and output, calibrated from existing
hydro-meteorological records. Within the range of calibration data, such models may be
highly successful, often because the formal mathematical structure carries with it an
implicit understanding of the underlying physical system. Black box models were
developed and extensively applied before advances in computer technology made it
possible to use more physically correct (and thus more complex) models. Today, black box
principles are more often used to form components of a larger model, e.g., the unit
hydrograph is often used for streamflow routing in conceptual rainfall-runoff models.

7
1.4.2 Conceptual models
These occupy an intermediate position between the fully physically-based
approach and empirical black-box analysis. Such models are formulated based on a
relatively small number of components, each of which is a simplified representation of one
process element in the system being modelled. The purpose of this breakdown is primarily
to enable the runoff from ungauged catchments to be estimated using standard parameters
together with data on the actual areal distribution between the four sub-catchments and
meteorological time series.

1.4.3 Physically-based models

These are based on our understanding of the physics of the hydrological processes
which control catchment response and use physically-based equations to describe these
processes. From their physical basis, such models can simulate the complete runoff regime,
providing multiple outputs (e.g., river discharge, phreatic surface level, and evaporation
loss) while black-box models can offer only one output. Also, almost by definition,
physically-based models are spatially distributed since the equations from which they are
formed generally involve one or more space coordinates. They can therefore simulate the
spatial variation in hydrological conditions within a catchment as well as simple outflows
and bulk storage volumes. On the other hand, such models make huge demands in terms of
computational time and data requirements and are costly to develop and operate.

1.5 AIM

The ultimate aim of the research work is to study the rainfall-runoff interaction for
a Banas River catchment, in Gujarat, India. And develop the rainfall-runoff model

1.7 OBJECTIVES

1. To develop an ANN model for rainfall-runoff relationship for Banas River

catchment.
2. To compare the ANN model with the Multiple Linear Regression model (MLR).

8
1.6 SCOPE OF THE STUDY

This study is forced to use either the Black box or lumped conceptual model as they
are simple to use and require less data as compared to the physical hydrological model.
This study will be performed with the help of the ANN model to better simulation of the
runoff for the Banas catchment. Using this model, the availability and variation of water
resource potential within the study area can be analyzed. This analysis involves some
degree of uncertainty due to the inadequacy of the required data.

9
 Chapter 2
Literature Review

10
 CHAPTER 2
LITERATURE REVIEW

Rainfall-runoff model using artificial neural network approach (2003)

Author: S. Raid, J. Mania, L. Bouchaou, Y. Najjar

(Journal of Science Direct, Elsevier)

Raid et.al, established an ANN model for the rainfall-runoff relationship for a catchment
located in a semiarid climate in Morocco. The purpose of this study was to predict river
runoff by using a multilayer perceptron (MLP) neural network. This study aimed to
develop the rainfall-runoff model in the Ourika catchment located in a semiarid climate in
Morocco, using a black box type model based on ANN methodology. In the study present,
the flow and rainfall series observed in the Ourika basin at Aghbalou station in Morocco
were analyzed using the ANN model. For this study data were taken for a period of seven
years (1990 to 1996). The entire database was represented by 2250 daily values of rainfall
& runoff pairs. In this paper, last year’s data (1996) was used for model testing, while the
other remaining data (1990 to 1995) was used for model training/ calibration. In this
research, the ANN was optimal at 600 iterations with 12 hidden nodes to evaluate the
performance of ANN the multilayer regression techniques were applied with the same data
sets used in the ANN model. The results and comparative study indicated that the artificial
neural network method is more suitable to predict river runoff than the classical regression
method.

Comparative analysis of event-based rainfall-runoff modelling

techniques -deterministic, statical, and Artificial Neural Networks (2003)

Authors: Ashu Jain, KV Prasad Indurthy

(Journal of hydrologic engineering, ASCE)

In this present study, the unit hydrograph (UH) theory, statistical regression, and ANNs
were used to model an event-based rainfall-runoff process. Four different storm events
from the Salado creek at Bitters Road, San Antonio were employed in this study for model
calibration and validation. Two storms were used for developing the unit hydrographs. The

11
data interval was 5 min. The whole data set consisting of 245 data points was divided into
two parts-a training or calibration set consisting of 208 data points corresponding to the
first three storms, and a validation or testing set consisting of 37 data points corresponding
to the last storm. Two types of regression model structures were investigated namely, a
linear multiple regression model structure (LMR) and a nonlinear multiple regression
model structure (NLMR). A back-propagation ANN with the generalized data rule as the
training algorithm was employed in this study. Two different types of ANN models were
developed- the single hidden layer ANN model and the multiple hidden layer ANN model.
The models were assessed by correlation coefficient (R) and average absolute relative error
(AARE). In terms of AARE, the LMR model performed the worst, whereas the ANN
model performed the best. In this study, as a result, an event-based rainfall-runoff process
determined peak discharge and time to the peak discharge very accurately.

Rainfall-runoff modelling using artificial neural networks:

Comparison of network types (2005)
Authors: A. R. Senthil Kumar, K. P. Sudheer, S. K. Jain, P. K. Agarwal
(Wiley InterScience)
Kumar et.al, developed rainfall-runoff modelling of two Indian River basins and evaluated
the performance of MLP- and RBF-type neural network models. The Malaprabha
catchment has been taken for this study. The Malaprabha river is a right-bank tributary of
the Krishna River, India. In this paper, the daily rainfall (areal average) and runoff values
were taken during the monsoon season (June to October) from 1987 to 1991. In this paper,
Determining the input variables involves finding the lags of rainfall-runoff that have a
significant influence on the predicted flow. The current study analyzed different
combinations of antecedent rainfall values and runoff values, and the appropriate input
vector has been selected based on the analysis of residuals. This experiment was done for
developing MLPs; then, based on the performance of various trained MLPs with different
input combinations, the three best networks were selected. The corresponding input vectors
in the selected MLPs were used to develop the RBF. In this study, the rainfall and
discharge data during the years 1988 and 1989 are used for training the ANN. The
rationale behind the selection of this data set for training is that the selected training period
contained the peak flow values for the entire period being studied; As ANNs are reported
to be not good in extrapolation. The data for other years (1987, 1990 and 1991) is used for
validating the model. The analysis utilized the statistical properties of the model output,

12
such as root-mean-square error (RMSE) and the coefficient of correlation between the
computed and actual outputs. Two additional storm-specific evaluation measures were also
considered: percentage error in predicted peak flow and percentage error in total runoff
volume. A total of 18 combinations of input variables are investigated, Three RBF models
are also trained using the input vectors that have been identified from the above analysis.
The other input combinations were not considered, as they are unable to improve the
performance of MLPs; an RBF performs better than an MLP in rainfall-runoff modelling.
Although the results presented in this paper indicate that an MLP is more accurate than an
RBF, it requires a longer time for training. However, a clear judgment on which is superior
may require a few more applications and analyses. A comparison of the rainfall-runoff
modelling skill of two ANN configurations, i.e., an MLP and an RBF, is presented. The
results suggest that the choice of the type of network certainly has an impact on the model
prediction accuracy. However, a judgment on which was superior is not possible from this
study. Both networks have merits and limitations. the results of the study indicate that the
generalization properties of RBF networks are poor compared with those of MLPs in
rainfall-runoff modelling.
Rainfall-runoff modelling using artificial neural network: modelling and
understanding (2008)

Author: Aman Mohmmad Kalteh

(ResearchGate, Caspian Journal of environment science)

In the present study, the author develops rainfall-runoff modelling using an ANN approach
for a watershed located in northern Iran. The author also describes different approaches
including a Neural Interpretation diagram, Garson’s algorithm, and randomization
approaches to understanding the relationship learned by an ANN model. In this paper, a
watershed from northern Iran was used as a study area. For a feed-forward multi-layer
perceptron model the data series of 1982-83 to 1996-97 was taken for input and output
variables (9 input variables and 1 output variable was selected). In this study relative
importance of input, variables were also evaluated with the help of Garson’s algorithm.
The result of the ANN model was assessed by correlation coefficient (r) and RMSE (root
mean square error). At the end of this research, the author found out that the ANN model is
to be reasonably accurate.

13
Modelling of rainfall-runoff correlations using artificial neural network
(2017)

Author: Ajay B Patel, Geeta S Joshi

(Civil Engineering Journal)

Patel et.al, studies involve the development of an ANN model using a feed-forward
backpropagation algorithm to establish monthly and annual rainfall-runoff correlation. The
ANN model developed in this study was applied to the Dharoi reservoir watershed of the
Sabarmati River basin of India. For this study twenty-nine years (1986-2014) hydrologic
data was applied. The whole computation was performed by using MATLAB capability of
developing ANN by using nntoolbox. In this paper, the trial-and-error method was applied
for different input data sets to train, test, and validation. Model performance was tested in
form of RMSE and correlation of coefficient (R) for the monthly rainfall-runoff ANN
model. Also, for the annual rainfall-runoff ANN model. It revealed that ANN could
capture the nonlinearity of rainfall-runoff modelling very well with good predictive power
for simulation in hydrological models.

Rainfall-runoff modelling using artificial neural network technique

(2017)

Author: Lateef Ahmed Dar

(International Research Journal of Engineering and Technology)

In this study, ANN techniques were used to develop the rainfall-runoff models, to predict
the runoff discharge at Padshahibagh station. Various types of neural networks viz., BPN
and RBF were developed. The present study was carried out in the upper Jhelum
catchment. The discharge data at Padshahibagh gauging station data from 1991-to 2013
and precipitation data were used in this study. The data sets for the years 2001-2009 were
used for training the model. Various statistical indices like R2 and root mean square error
(RMSE) were computed for comparing these models. From the values of R2 and RMSE, it
was observed that radial basis networks (RBF) were able to predict the runoff more

14
accurately than backpropagation networks (BPN). ANNs have proven to be especially
good in modelling complex and nonlinear systems.

Rainfall-runoff modelling and simulating using remote sensing and

Hydrological model (2020)

Author: Anant Patel

(springer)

The main objective of this study was to determine the water level and submerged area by
using a GIS-based hydrological model in the Banas River basin. The Banas River was
taken as a study area for rainfall-runoff simulation. In this study, two models were
explored, a rainfall-runoff model that converts rainfall excess to surface flow & river
runoff, and a second model as a hydraulic model that covers unsteady-state flow through
the river channel network. In this study, rainfall data was collected for flood events that
occurred during the Gujarat-2017 flood in July month. Rainfall data were also collected for
the particular event starting from 21 July to 31 July 2017. In this study, the rain gauge
station data was collected and average rainfall was carried out based on the arithmetic
average rainfall method for a particular date. In this research, HEC-RAS 5.0.3 has been
used for R-R modelling. The result of this study will benefit flood management and also be
used in future modelling for disaster forecasts.

Rainfall-runoff modelling for the Hoshangabad Basin of Narmada River using an

artificial neural network (2020)

Author: Vikas Poonia & Hari Lal Tiwari

Arabian Journal of Geosciences

Poonia et.al, established an ANN-based rainfall-runoff relationship model for a

Hoshangabad catchment of the Narmada River in Madhya Pradesh. This study aimed to
estimate daily streamflow by using different NN models i.e., Radial basis function (RBF)
and Feedforward backpropagation (FFBP) neural network model. This study aimed to
develop the rainfall-runoff model in the Hoshangabad catchment located in Madhya
Pradesh, based on ANN methodology. In this study, the discharge and rainfall series were
utilized for developing ANN models by using various combinations of input parameters.
For this study data were taken for a period of ten years (2004 to 2013). For better results,

15
only monsoon season data (Rainfall and runoff data) has been considered in the present
study. The three evaluation criteria used in this study are R 2 (Coefficient of Determination)
value, average absolute residual error (AARE) and Mean Square Error (MSE). Both FFBP
and RBF networks were compared based on different performance evaluation criteria and
demonstrated the best model variant out of both networks for the Hoshangabad catchment.
Results demonstrate that RBF network results are better than Backpropagation networks.

16
 Chapter 3

Study Area & Data Collection

17
 CHAPTER 3
STUDY AREA & DATA COLLECTION

3.1 GENERAL

The Banas River basin is taken as a study area for rainfall-runoff modelling by using
Artificial Neural Networks. Banas River is the longest river in the north Gujarat region.
The Dantiwada dam and Sipu dam are two main irrigation structures existing on the main
channel of the Banas River.

3.2 BANAS RIVER

The Banas River rises near Pindwara village in the Sirohi district of Rajasthan. The
total length of Banas River is 266 km of which 78 km is in Rajasthan and the remaining
188 km is in Gujarat state. The total length of the river from origin to its outfall into the
little Rann of Kachchh. It drains an area of 8,674 sq km out of which nearly 38 % lies in
Rajasthan State and the remaining 62 % falls in Gujarat state. The basin of the river lies
between 24° 19’ 55.98” to 24° 54’ 51.02” North latitude and 72° 35’ 22.07’’ to 73° 10’
03.50’’ East longitudes. The river flows in a south-westerly direction and empties into little
Rann of Kachchh. The number of principal tributaries, which contribute significantly, is
seven. Sipu is the only major tributary on the right bank. The other six tributaries namely
Batria, Sukli, Sewaran, Suket, Balaram and Khari drain into the main channel from the left
bank. Hence draining system on the left bank of the Banas River is more extensive as
compared to the right bank area. The average rainfall in the Banas basin is 921 mm. Owing
to topographical characteristics, the climate is variable.

Right Bank Stream: Sipu

Sipu is the principal tributary of the Banas rising from Sirohi and Mount Abu hills in the
Sirohi district of Rajasthan state. Abu hills fall between the Banas and Sipu sub-basin.
About 30% of Mount Abu hill’s direct runoff drains into the Sipu River while about 70%
of mount Abu hill’s direct runoff flows into the Banas River. The confluence of the Sipu
River and Banas River is 12 km downstream of the Dantiwada dam.

18
Left Bank Stream: Khari

Khari River rises from Palanpur (Banaskantha district) and drains into the Banas River
through Patan district at 80 km downstream of the Dantiwada dam

Left Bank Stream: Sukli

The Sukli tributary rises from Aravalli hills near Pindwara of Sirohi district (Rajasthan)
and drains into the Banas River downstream of Swaroopganj dam and 9 km upstream of
Abu Road of Rajasthan state.

Left Bank Stream: Batria

The river rises near the Ambaji hills of the Aravalli range and drains into Banas, 3 km
upstream of Abu Road.

Figure.3.1 Location of Entire BANAS River Catchment (Source: CWC)

19
Figure.3.2 Location of Banas River Catchment (Study area Map, Source: ArcGIS)

20
 Figure.3.3 BANAS River catchment Map Up to Dantiwada Dam (Source: ArcGIS)

The catchment area and length of the distributaries are indicated in the following table.

21
 Table 3.1 Source of the distributaries

Name of river Bank Length (Km) Catchment area (Km2)

Banas Main 266 8674

Sipu Right 75 1420

Sukli Left 38 438

Batria Left 24 218

Balaram Left 40 345

Khari Left 88 1391

Suket Left 15 79

3.3 DANTIWADA DAM

The Dantiwada Dam is an earthen (mud) and masonry dam on the Banas River near
Dantiwada, Banaskantha district of northern Gujarat in India. The dam was constructed in
1965 mainly for irrigation and flood control. The maximum height of the dam above the
lowest point of the foundation is 61 meters and the length of the top of the dam is 4832
meters. The area of the catchment of this scheme is 2862 km2.

3.3.1 Characteristics of Dantiwada Dam

Table 3.2 General information of Dantiwada Dam

22
 Village: Dantiwada
Location Taluka: Dantiwada
District: Banaskantha

Purpose Irrigation & Flood Control

River Banas

Catchment Area 2862 km2

Mean Annual Rainfall 940 mm

Completion of Year 1965

Table 3.3 Details of the Dantiwada Dam section

23
 Type of Dam Earthen & Masonry

Bedrock Granite

Maximum height above the

61 m
lowest point of the foundation

Length of the top of the dam 4832 m

Masonry volume content 0.394 mm3

Concrete volume content 0.393 mm3

Earthwork content 0.621 mm3

Table 3.4 Details of Reservoir of Dantiwada Dam

The area at full reservoir level 40.47 km2

Gross storage capacity 464 mm3

Effective storage capacity 444 mm3

The area under submergence:

a) Forest b) Wasteland c) culturable

a) 215 ha b) 810 ha c) 2025 ha

No. of the village under submergence 12 Partial

Table 3.5 Details of Spillway

24
 Type Ogee

Length 165 m

Energy dissipater Roller & Ski-jump buckets

Maximum Discharge 7504 m3/s

Type, Nos. and size of gate Radial, 11, (12.5 m × 8.23 m)

Table 3.6 Details of Canal

Length of canal 48 km

Capacity 31 m3/s

Gross command area 80939 ha

Culturable command area 59895 ha

Table 3.7 Villages under command of Dantiwada Dam

a) District b) Taluka c) No. of villages

Banaskantha Palanpur 18

Dantiwada 2

Deesa 31

Kankrej 8

Patan Patan 49

Sidhpur 3

Total 111

3.4 DATA COLLECTION

25
 The study area consists of seven rain gauge stations, to monitor rainfall. These rain
gauge stations; six rain gauge stations are located upstream side of the Dantiwada Dam .
Out of seven three stations measure discharge also. Moreover, other climatic variables such
as temperature, relative humidity, and wind velocity, are also measured at some of these
stations in particular time intervals. All the stations are operated by Central Water
Commission, whose characteristics are given in Table

The precipitation data from the year 2004 to 2018 (2009 to 2011 was missing due
to some reasons) is available for this study. For the present study, the monsoon storm data
(precipitation, weather data, and inflow are available) are adequate for the development of
the model and other seasons (winter and summer) data were inadequate to fit the model.
The monsoon data was considered from June to October months (from 2012 to 2018). The
missing data were omitted from the datasheet.

Table 3.8 Characteristics of CWC rain gauge Station

Sr No Location Name Local River Latitude Longitude

1 Swaroopganj Banas 24 41 28 72 55 52

2 Abu road Banas 24 29 38 72 47 30

3 Sarotry Banas 24 22 04 72 32 48

4 Chitrasani Balaram 24 17 20 72 29 54

5 Dantiwada dam Banas 24 20 14 72 29 54

6 Bhakudar Sipu 24 23 56 72 18 02

26
 Chapter 4

Introduction of Artificial Neural Network & MLR

Model

CHAPTER 4

27
INTRODUCTION OF ARTIFICIAL NEURAL NETWORK &
MLR MODEL

4.1 GENERAL

The rainfall-runoff process has been described quantitatively since the nineteenth
century. However, it is only in the last two decades or so that ANNs have been applied to
the problem. Nonetheless, ANNs have been in existence since McCulloch and Pitts (1943)
introduced the concept of the artificial neuron. Since that time neural network research has
evolved in three distinct phases. The first era involved preliminary work on the
development of the artificial neuron until Minsky and Papert (1969) identified several
limiting factors. The second era began with the rediscovery and popularization of the
backpropagation training algorithm. The third era is characterized by more rigorous
assessments of network limitations and generalizations, fusion with other technologies
(such as genetic algorithms and fuzzy logic) and the implementation of ANNs using
dedicated hardware.

Artificial neural networks are the most powerful learning models. They have the
versatility to approximate a wide range of complex functions representing multi-
dimensional input-output maps. A trained neural network can be thought of as an "expert"
in the category of information it has been given to analyze.

4.1.1 Neurons

A neuron is the basic processing unit in a neural network sitting in our brain. It consists of

1. Nucleus

2. Axon- Output node

3. Dendrites-Input node

4. Synaptic junction

The dynamics of this synaptic junction are complex. We can see the signal inputs from the
action of a neuron and through synaptic junction an output is actuated which is carried over
through dendrites to another neuron. Here, these are the neurotransmitters. We learned
from our experience that these synaptic junctions are either reinforced or sense that they

28
behave in such a way that the output of the synaptic junction may excite a neuron or inhibit
the neuron. This reinforcement of the synaptic weight is a concept that has been taken to an
artificial neural model.

Fig. 4.1 Structure of Neuron

The objective is to create an artificial machine and this artificial neural network is
motivated by certain features that are observed in the human brain.

4.1.2 Artificial Neural Networks

The terminology of artificial neural networks has developed from a biological

model of the brain. The biological neuronal network comprises many interconnections
between the neurons.

The influence of neurons arises from their combined performance in a network

where all neurons are organized. A neural network consists of a set of connected cells and
neurons. The neurons receive impulses from either input cells or other neurons and
perform some kind of transformation of the input and transmit the outcome to other
neurons or output cells. The neural networks are built from layers of neurons connected so
that one layer receives input from the preceding layer of neurons and passes the output on
to the subsequent layer.

29
 Fig. 4.2 Artificial Neural Network Structure

Input layer: This layer resembles the number of neurons to the number of neural network
inputs. This layer also contains a passive node that does not takes part in the actual signal
adjustment but only conveys the signal to the corresponding layer.

Hidden layer: This layer has a random number of neurons for the random number of
layers. The signal adjustment takes part in this layer and therefore they are active.

Output layer: The neuron number in the output layer connects to the number of neural
network output values. The active nodes are present in this layer. The basic Artificial
Neural Network structure is shown in fig.4.2 Statistics flow in an artificial neural network
is unidirectional.

30
4.1.3 Network Architecture
ANN may be described as a network of interconnected neurons (sometimes called
nodes). Figure 4.3 presents the structure of an individual neuron. Each neuron consists of
several input arcs (stemming from other neurons or from outside the network; u 1, , , ,un)
and several output arcs (which in turn lead to other neurons or the ‘outside world’). A
neuron computes an output, based on the weighted sum of all its inputs (S j), according to
an activation function (f(Sj)). These activation functions may be logistic sigmoid (Figure
4.3), linear, threshold, Gaussian or hyperbolic tangent functions, depending on the type of
network and training algorithm employed.

Fig 4.3 Activation of a single neuron

In most studies, the logistic sigmoid function or hyperbolic tangent functions are used.
The logistic sigmoid activation function (below Formula) – in which x represents the
weighted sum of inputs to the neuron and f(x) the neuron’s output – is often used because
it is continuous and relatively easy to compute (as is its derivative). It maps the outputs
away from extremes, and it introduces nonlinear behaviour to the network:
1
f ( x )= −x
1+e
Advantages of ANN
 ANN handle noisy or missing data.
 ANN can work with a large number of variables or parameters.
 ANN provide a general solution with good predictive accuracy.
 ANN deal with the non-linearity in the world in which we live.

31
  ANN is an empirical model which depends on empirical observations rather than
on mathematically describable relationships of the system modelled
 ANN generates estimated functions that are unknown
 ANN method is simple to runoff, flood, etc.

Disadvantages of ANN
 The neural network needs the training to operate
 It has a black box image. So not shown about the conversion of layer
 The architecture of a neural network is different from the architecture of
microprocessors and therefore needs to be emulated.
 Requires high processing time for large neural networks.

4.2 BACKPROPAGATION NEURAL NETWORK

Backpropagation is perhaps the most popular algorithm for training ANNs. The
network architecture used in a backpropagation (BP) algorithm is made up of
interconnected nodes arranged in at least three layers (Fig. 4.4). The input layer receives
the input data patterns. The output layer produces the result. The hidden layers sequentially
transform the input into the output. A maximum of three hidden layers are adequate for
most problems in water resources applications. The number of neurons in the hidden layers
is usually fixed by trial and error.

32
 Fig 4.4 Structure of the Back-propagation Network
The transfer function used in BP networks is usually a sigmoid function (Fig. 4.3). During
training, back-propagation networks process the patterns in a two-step procedure. In the
first or forward phase of back-propagation learning, an input pattern is applied to the
network, with initially assumed weights to provide the output at the output layer. The error
is estimated from the corresponding output value provided in the training set. In the second
or backward phase, the error from the output is propagated back through the network to
adjust the interconnection weights between layers. This process is repeated until the
network’s output is acceptable. Back-propagation learning is this process of adapting the
connection weights. When an unknown input pattern is presented, the trained network
produces a corresponding output pattern. Details of the back-propagation method are
described by Freeman and Skapura.

33
4.3 FEED FORWARD NEURAL NETWORK
In feed-forward networks, the connections between neurons flow in one direction:
from an input layer, through one or more hidden layers, to an output layer (Figure 4.5).
While some studies direct predicted output back to the input side of a network to make
further predictions, this is still feed-forward as only one forward pass is made through the
network for each prediction. Two feed-forward network types have been widely used to
model the rainfall-runoff process: the multilayer perceptron (MLP) and the radial basis
function (RBF). These networks typically consist of three or four connected layers of
neurons (as shown in Figure 4.5). The number of neurons in the input and output layer is
specified by the problem to which the network is applied (i.e., the number of predictors
and predictands, respectively).

Fig
4.5 A Feed Forward Neural Network
Inputs to the network (predictors) are passed from the input layer of neurons, through
the hidden layer(s) of neurons, to the output layer (Figure 4.5) where they become
predictands. Neurons in the input layer do no more than disperse all predictors to each
neuron in the hidden layer. The network operates by applying weights to values as they
pass from one layer to the next and calculating outputs for each of the neurons in all other
layers

34
4.4 ACTIVATION FUNCTION
The activation function is used to convert the input to output, it is also called the transfer
function.
I/P – horizontal axis shows the sum of inputs
O/P – vertical axis shows the value output (value of function ‘f’ produce)
The function f is designed to produce the values between 0 to 1
The most commonly used functions are described below:

4.4.1 Log Sigmoid function

The log sigmoid function is plotted as an s-shaped graph. Log sigmoid is a transfer
function, Transfer functions calculate a layer’s output from its net input.
The sigmoid function is achieved using an exponential equation,

Fig 4.6 log sigmoid function

logsig(n) = 1 / (1 + exp(-n))

The nonlinear notice that x value lies between -1to 1, y values is very steep. This means
small changes in x could also bring about large changes in the value of y.

35
4.4.2 Threshold function
In the threshold function output is set at one of two levels, depending on whether the total
input is greater or less than some threshold value.
Binary threshold:
The output of a binary threshold function produces;
Y= f(i)=1, if the weighted sum of inputs is +ve
f(i)=0, if the weighted sum of output is -ve

threshold

Fig 4.7 binary threshold function

Bipolar threshold:
The output of a bipolar threshold function produces;
Y= f(i)=1, if the weighted sum of inputs is +ve
f(i)=-1, if the weighted sum of output is -ve

36
 Fig 4.8 bipolar threshold function
4.4.3 Linear function
In linear regression, a linear activation function transforms the weighted sum inputs of the
neuron to an output using a linear function.
One of the major problems in linear function is, that it is not possible to backpropagation
as the derivative of the function is a constant and has no relation to the input.

Fig 4.9 linear function

4.5 TRAINING ALGORITHM

Levenberg Marquardt algorithm
This algorithm typically requires more memory but less time. Training automatically stops
when generalization stops improving, as indicated by an increase in the mean square error
of the validation samples.
Bayesian Regularization algorithm
This algorithm typically requires more time but can result in good generalization for
difficult, small or noisy datasets. Training stops according to adaptive weight minimization
(regularization).
A scaled conjugate gradient algorithm
This algorithm requires less memory. Training automatically stops when generalization
stops improving, as indicated by an increase in the mean square error of the validation
samples.

37
4.6 MULTIPLE LINEAR REGRESSION MODEL (MLR)
Multiple linear regression (MLR) is used to model the linear relationship between a
dependent variable and one or more independent variables. It is also used to (understand
which independent variables are related to the dependent variable, and to) determine the
form of these relationships. The Multilinear regression method is generally based on least
squares: the model is fit such that the sum-of-squares of differences of actual and
forecasted value is minimized. In this model, the value of independent variable x is
associated with the value of the dependent variable y. the model is fit to a period – the
calibration period- for which climatic and tree ring data overlap. In the process of fitting,
or estimating, the model, statistics are computed that summarize the accuracy of the
regression model for the calibration period. The performance of the model on data not used
suitable the model is usually checked in some way by a process called validation.
Multiple linear regression is not strictly a “time series” method. The most important
point in application to time series is that observations are typically not independent of one
another. As an outcome, special attention must be paid to a regression assumption about
the independence of the residuals.
The dependent variable is sometimes also called the predicted, and the independent
variables are the predictors. The predictors in my regression problem might be inter-
related. This so-called multi-collinearity does not exclude the use of regression but can
make it impossible or difficult to assess the relative importance of individual predictors
from the estimated coefficients of the regression equation.
In MLR, the function is a linear equation, i.e., straight-line, in the form:

𝒚 = 𝛽0 + 𝛽1 𝒙1 + 𝛽2 𝒙2 +………+ 𝛽n 𝒙n+ ei

where 𝒚 will be represented the output variable, 𝛽0; ...; 𝛽n will be representing the
regression coefficient to be determined, and 𝒙1; ...; 𝒙n will be representing the input
variables.
ei = error term

38
4.6.1 Statistics for MLR
The sum of the square term: several regression statistics are computed as functions of the
sums-of-square terms:
n
Sum of squares, error SSE =∑ ei
2

i=1

n
Sum of squares, total SST = ∑ ( y i− ^y )
2

i=1

n
Sum of squares, Regression SSR = ∑ (~
y i− ^y )2
i=1

n = sample size (no. of observation in calibration period)

 Partition of variation
The regression equation is determined such that the total sum of squares can be partitioned
into components due to regression and residuals:
SST = SSR + SSE
Coefficients of determination: the explanatory power of the regression is summarized by
its ‘R squared’ value, computed from the sums of squares terms as

SSR SSE
R2 = =1-
SST SST

Also, the coefficient of determination is often described as the proportion of variance R 2

described by regression. The relative sizes of the sum of squares terms indicate how R 2
“good” the regression is in terms of fitting the calibration data. If the regression is
“perfect”, all residuals are zero, SSE will become zero and R 2 equals 1. If the regression is
total failure R2 = 0.

39
 Chapter 5
Methodology

40
 CHAPTER 5
METHODOLOGY

5.1 GENERAL
The methodology is one of the significant parts of research work. In this study, two
different models namely, Artificial Neural Network and the Multiple linear regression
model are being used. The focus of the dissertation is to evaluate the ANN and MLR
model to predict the runoff of the Banas River catchment up to the Dantiwada dam,
Gujarat. Hence all the Available Rainfall data, discharge data and other meteorological
data were collected from the Central Water Commission and State Water Data Centre,
Gandhinagar. Total 1836 no. of data pairs was available, but due to inadequacy of data row
here only 357 data pairs were considered. The training data sets are used to adjust
connection weight according to its error. Validation is used to measure network
generalization, and to halt training when generalization stops improving. testing does not
affect training and so provides an independent measure of network performance during and
after training. Precipitation, wind velocity, humidity, temperature was being used to predict
the runoff of the catchment area. The target and input data forms are divided into three
parts; training (70%), validation (15%), and testing (15%).

The artificial neural network with one input layer, one hidden layer, and one output
layer is taken into consideration. For the selection of the best network architecture with an
optimum number of hidden layer neurons, many trials have to be carried out. Different
training algorithm such as Levenberg Marquardt (LM), and scaled conjugate gradient has
been used to train the model. In the network architecture, many trials have been done with
changes number of neurons in the hidden layer. The performance of the model can be
evaluated in the terms of accuracy refers to the ability of the model to reduce calibration
error consistency is used for representing the characteristics of the model whereby the level
of accuracy and estimation of the parameter’s values persist through a different sample of
data. A versatile model is defined as a model which is accurate and consistent when used
for different applications.

41
5.2 FLOW CHART FOR METHODOLOGY
Problem Identification

Objective of the study

Data Collection

Import the data (normalized data) in

MATLAB workspace

Use the neural network tool box (nntool box)

Import the input and output dataset in data manager and

create network

Select network types and number of neurons for hidden layer

Run the model

Training Training Training

NO
Performance measure check
whether MSE  0 and R  1

yes

De-normalization & Compare ANN data

with actual
yes data

Comparison between ANN and MLR Model

42
 Sensitivity Analysis

Table 5.1 training variables and their allotted values for the ANN model

Training variables Allotted values

Neural network types Feedforward backpropagation

Number of hidden neurons 2 to 16

Levenberg Marquardt / scaled conjugate

Training algorithm
gradient

Hidden layer transfer function Sigmoid

Output layer transfer function Purelin

Epoch Random

Step:1 Normalization

All the parameters should be normalized to ensure they receive equal attention during the
training process. Also, the normalization of data is important for the efficiency of the
training algorithm. For this research work range of data, normalization is taken from 0 to 1.

R−Rmin
The formula for normalization; Rnorm =
¿¿

Where,

Rnorm = normalized value

R = actual value

Rmin = minimum value in the record

Rmax = Maximum value in the record

43
Step:2 Network architecture

The network architecture of the ANN model is comprised of the input layer, hidden
layer(s), and output layer. The hidden layer is to be specified with a various number of
neurons as required for the model, also transfer function is specified before running the
model. The error goal of the model in terms of Mean Square Error (MSE) and a maximum
number of epochs are set down.

Step:3 Network training

The normalized data (input and output sets) are presented to the network in the MATLAB
environment. Training multiple times will generate different results due to different initial
conditions and sampling. At the end of the training, the network output and error
calculation can be measured.

Step:4 Network testing

The trained network is presented testing the input dataset for simulating the observed
output set.

Step:5 Statistic indices

Various indices can measure the performance of the trained network and are calculated to
check model capability to simulate desired output.

Mean Square Error (MSE)

MSE is a network performance function. It measures the network’s performance according

to the mean of squared errors. A lower MSE value indicates greater similarity between
input and output.

1 2
MSE = (x − y )
N i i

Where, Xi = N no of input data

yi = N no of desired data

Root mean square error (RMSE)

44
The root mean square error measures the forecasted accuracy of the established model. It
equates difference between forecasted and actual parameters values. It is a positive value
ranging from zero (0) to infinite (∞). The perfect fit model has zero RMSE and the
increased value of RMSE show a higher deviation between forecasted and actual values.
The root mean square error is calculated using the following equation,

RMSE=
√ 1
N
( x i− y i )
2

xi= N no of Actual output

yi= N no of Forecasted output

Coefficient of correlation (R)

The coefficient of correlation is denoted by R, “R” measures the strength of the linear
relationship developed by a particular model. The correlation coefficient shows the amount
of closeness between actual and forecasted values. The value of r close to 1.0 represents
good model performance & 0 represents poor model performance. It is calculated using the
following equation

∑ ( x −xmean )( y − y mean )
i=1
R=

√∑
N N
( x−x mean) 2
∑ ( y− y mean)2
i=1 I=1

Where, X= actual output

Xmean =mean of actual output

Y= forecasted output

ymean= mean of forecasted output

Step:6 Simulation

It helps to predict the behaviour of a system. The performance indices for observed and
simulated output are worked out. These measures indicate the ability of a network to
reproduce output for inputs known to the network.

Step:7 De-normalization of data

45
The training and testing datasets are denormalized & checked for the performance indices.
Where the results are acceptable their next step is executed otherwise network architecture
is changed and again the procedure is carried out (followed by stage:2).

Step:8 Freezing the Architecture

When all the performance measures are acceptable, the final network architecture in terms
of link weights, neuron bias and trains are frozen.

5.3 MODEL DEVELOPMENTS

Step:1 preparation of the datasheet is the primary stage of ANN. In the datasheet inputs
and output data are prepared. The inputs and output parameters are listed below.

Table 5.2 Dataset list

Model Input layer data Output layer data Data availability

ANN MODEL Precipitation of Runoff 2012 to 2018

variation station monsoon data
Wind velocity

Temperature

Humidity

 Present the model as a chart:

MODEL

ANN

RAINFALL
I
N
P W.V / TEMP. OUTPUT AS A
U RUNOFF
T
HUMIDITY

46
Step:2 The data series consists of seven-year data from (2012 to 2018) such as daily
rainfall, temperature, humidity, wind velocity (from June to October, rainy storms). The
input and output parameters in MS Excel include all the available data, which are below.

Fig 5.1 datasheet containing all the available data

47
Step:3 the normalized data (only including monsoon seasonal data) in MS Excel is shown
below

Fig 5.2 normalized datasheet

Step:4 import the input data into MATLAB software shown below

48
 Fig 5.3 input data in MATLAB software

Step:5 import the output values in the MATLAB environment

Fig 5.4 output value in MATLAB software

Step:6 after completion of importing inputs and output data sets, go to the command
window and type nntool

Step:7 in the neural network dialogue box import the input and target data and then create
a new network

Step:8 Select the network type and number of layers, after that select the number of
neurons for the hidden layer. After selecting transfer functions networks will be created by
clicking on creating option.

49
Step:9 in this step datasets are divided for training, validation, and testing purpose. The
data is divided in form of percentage, 70%, 15%, 15% for training, testing, and validation
respectively.

Step:10 artificial neural networks are a kind of black-box type; it means we do not know
the structure but just regard it as behaviour in practice. During the training phase, a
different number of neurons is taken in the hidden layer. Training, testing, and validation
are performed in MATLAB and obtained statistical indices such as R and MSE. The
regression graph is also obtained and get the value of “R” in the MATLAB environment.

The different values of the different models are obtained by the trial-and-error
method. One of the tables containing values of statistical indices and network architects is
presented below. The architect pattern 18-8-1 indicates, 18 numbers of the input value, 9
number of neurons, and 1 is the output value, remaining architects follow the same as
previously explained.

50
Table 5.2 Network architecture of ANN model 1 with LM algorithm [training 70%, testing
15%, validation 15%]

Model
Training Testing Validation All MSE value Epoch
Architecture
18-2-1 0.995 0.9846 0.7561 0.9315 0.00003076 22

18-4-1 0.9971 0.3624 0.9637 0.9357 0.000038038 14

18-6-1 0.9953 0.3682 0.9057 0.9317 0.0008694 13

18-8-1 0.9773 0.7909 0.9971 0.9763 0.00031839 9

18-10-1 0.9357 0.9456 0.1456 0.9318 0.00055397 9

18-12-1 0.9897 0.5667 0.847 0.9778 0.000001197 10

18-14-1 0.9693 0.5515 0.425 0.8825 0.00025413 8

18-16-1 0.9419 0.975 0.6922 0.9057 0.00154284 9

0.8

0.6
R value

0.4

0.2

0
1 2 3 4 5 6 7 8
no. of neurons

training R validation R Testing R Aggregate R

Fig 5.5 Neurons vs Accuracy chart

51
 Table 5.3 Network architecture of ANN model 2 with SCG algorithm [training 70%,
testing 15%, validation 15%]
Model
architectur Training Testing Validation All MSE value Epoch
e
18-2-1 0.9121 0.9841 0.6282 0.779 0.0005499 22
18-4-1 0.8685 0.0368 0.9568 0.8503 0.00012737 16
18-6-1 0.9377 0.2917 0.9002 0.8965 0.000004007 20
18-8-1 0.8786 0.813 0.992 0.8699 0.0001645 32
18-10-1 0.8794 0.9937 0.2578 0.974 0.0003374 66
18-12-1 0.987 0.5478 0.7606 0.9628 0.017304 40
18-14-1 0.973 0.4654 0.2563 0.9295 0.0024571 41
18-16-1 0.7306 0.9218 0.2166 0.7596 0.0032145 15

1
0.9
0.8
0.7
0.6
R value

0.5
0.4
0.3
0.2
0.1
0
1 2 3 4 5 6 7 8
no of nurons

Training R validation R testing R Aggregate R

Fig 5.6 Neurons vs Accuracy chart

All the above tables and graphs show that 9 neurons give the best percentage of training,
testing, and validation values for the ANN model. For the ANN model accuracy testing
value of R is considered for the best model. Hence the model architect 18-9-1 is consider
the best fitting for accuracy in this study.

52
Step:11 during the training phase two types of training algorithm is used, namely
Levenberg-Marquardt (LM) and scaled conjugate gradient (SCG). The training of the
model stops automatically when the network is fitted for the best.

Step:12 the model is retrained until the R-value comes to nearer unity and the Mean
Square Error near zero. In the training the value of R near 1 means a close relationship
between target output and simulated output, same the MSE value nearer to zero indicates
errors are negligible.

Step:13 the neural network training tool (nntraintool) shows the number of iterations,
performance chart, training algorithm, gradient, epoch, etc.

53
 Fig 5.7 NN tool image from MATLAB
Step:14 one of the regression plots shows the relations between targets and output as
follows.

Fig 5.8 regression plot

Step:15 the validation performance graph gives the Mean Square Error value, form this
best validation is to be optimized and given below.

54
Fig 5.9 MSE value

55
Step:16 at the end of the model development program saves the result and later the
different files like net, info, error and output files need to be added to the workspace.

Fig 5.10 neural network results

Step:17 finally the simple or advanced script need to be generated. The output values
obtained (described in step 16) will be the predicted value.
Step:18 compares the actual value with the predicted value.

56
5.4 SENSITIVITY ANALYSIS
The artificial neural network performs well in forecasting or predicting the observed
choice. It has lack methods for interpreting the significance of input variables because
ANN learns the parameters by encoding them in numeric connection weights. ANN gives
the black box image so it can be removed by finding the relative importance of the
variables.

The relative importance of input parameters is obtained by the portioning of weights

(Garson algorithm,1991). For the sensitivity analysis, keep all the variables at their mean
value except one variable, and find the relative change in the output concerning change in
that variable from mean up to half the value of standard deviation. Repeat the procedure
for all parameters. This analysis is only done for the ANN model, obtained results are
shown below.

Sensitivity Analysis
25.00%

20.00%

15.00%

10.00%

5.00%

0.00%
i( p
) p) p) p) p) p) j(p) ity) ity) ity) .s) .s) .s) mp) mp) mp)
j bu( da( ni ( ry ( ad( n i d i d i d ( w ( w ( w
ba sa rot ga ni ad b t
e te te
t.a a ro m m m ry b b
am m ntiw itra sa abu roop i (hu (hu d(hu asa arot u ro bul bul bul
da ch a n try a hitr s a b w et w et w et
sw rasa aro u ro c
ti s b i( y( d(
ch a
s an rotr roa
a sa u
itr ab
ch

Fig 5.11 Relative importance of various parameters

The above figure shows that the precipitation of Chitrasani, Sarotry, Abu Road have a
much importance on runoff generation, but also Ambaji, mt. Abu, Dantiwada, and
Swaroopganj has a good impact on runoff. Meanwhile other weather data such as wind

57
velocity, temperature, and humidity does not show the importance on generation of runoff
for this study area. From the above analysis, Chitrasani and Sarotry contribute 21.32% and
18.19% of the precipitation value to generate runoff in the Banas River catchment, Gujarat,
India.

Chapter 6
Results And Analysis

58
 CHAPTER 6
RESULT AND ANALYSIS

6.1 GENERAL

The result is one of the significant parts of the research work. With the help of the results,
the accuracy of the model can be measured. In this study ANN gives the best results as
compared to the MLR model, furthermore, an ANN gives results as a regression graph and
mean square error graphs. Various ANN model gives different results, here the architecture
M1- 18-9-1 containing training (70%), testing (15%), and validation (15%) possessed the
best result in this study.

6.2 REGRESSION AND MSE GRAPH FOR TRAINING, TESTING,

AND VALIDATION

The regression graph and Mean Square Error graph of ANN model architecture 18-9-1
(LM Algorithm) and 18-9-1 (SCG Algorithm) are chosen as better models.

As earlier decided the best model with training, testing, and validation data distributed
70%, 15%, and 15% respectively for both different algorithm LM and SCG using
feedforward Backpropagation Neural network obtained from the MATLAB software are
shown below;

59
Fig 6.1 Training, validation and testing regression for the Levenberg-Marquardt algorithm

60
Fig 6.2 MSE for LMA training, validation and testing algorithm

61
Fig 6.3 Training, validation and testing regression for a scaled conjugate gradient
algorithm

62
 Fig 6.4 MSE for SCG training, validation and testing algorithm

6.3 ANN MODEL ANALYSIS

63
To Forecast catchment runoff at the Dantiwada dam outlet for a Banas River catchment
using an Artificial neural network (ANN), here two different models namely an Artificial
neural network model using a feed-forward backpropagation algorithm and a Multiple
linear regression model were developed. Here for the ANN models ‘Levenberg-Marquardt’
Algorithm was used as a learning algorithm with the Transfer function sigmoid in the
hidden layer and the Transfer function purelin in the output layer, only the hidden neuron
was changed and the best ANN architecture was selected based on high regression value
and low mean square error.

Table 6.1 Performance evaluation of model on training and testing period for ANN model

Model
algorithm Training Testing validation All MSE value
architecture

0.9773 0.7909 0.9971 0.9763 0.000318

LM 18-8-1
0.8786 0.813 0.992 0.8699
SCG 18-8-1 0.0001645

From the Statistical value of table 6.1, it is clear that Model 1 having precipitation as input
to predict the catchment runoff gives comparatively more reliable and quite accurate
results than Model 2 which has a different algorithm for training model that gives
satisfactory results, hence model 1 consider as a good model using ANN.

64
6.4 COMPARISION OF PREDICTED DATA WITH ACTUAL DATA
The below shown Scattered graph has been plotted for comparison of ANN model
predicted catchment runoff with actual runoff.

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig 6.5 scatter plot of actual vs predicted values (Normalized)

The results showing that above 75% of data are being matched (around 268 pairs value are
match well with the actual data).

6.5 MULTIPLE LINEAR REGRESSION (MLR) MODEL

65
Regression models describe the relationship between a dependent variable and one or more
independent variables. In the present study, Linear regression with multiple predictor
variables is being used during the development model. For the regression learner, Train
regression models to predict data using supervised machine learning were used.

The below table shows the regression statistics;

Table 6.2 Regression statistics for MLR model

Regression statistics

R (square root of R2) 0.8831

R square (R2) 0.78

MSE 0.02272

Standard error 1.6317

Observation 117

Adjusted R square 0.6541

Where, R =correlation between dependent and independent variables,

R2 = coefficient of determination
Adjusted R = used if more than one variable
Standard error = standard error of regression
Observation = number of data used for regression

66
 Fig 6.6 Linear regression model statistics

6.6 COMPARISON BETWEEN ANN AND MLR MODELS

Table 6.3 comparison of R and RMSE values of both models

ANN model MLR model

R=0.9763 R=0.8831

RMSE=0.000318 RMSE = 0.15073

The Statistical value of Table 6.2 shows the comparison of the ANN and MLR model.
From the table, it is distinguished that both the ANN model and MLR model give quite
Precise and reliable results. The ANN model of catchment runoff prediction has input as
precipitation from six up steam gauging stations of Dantiwada dam and other
meteorological data, hence it has 18 inputs so to predict the catchment runoff uses the
feedforward backpropagation algorithm. During the building of the model different
training algorithms, such as Levenberg Marquardt, and scaled conjugate gradient were
used, and the number of neurons was arranged as required during model development. For
choosing the best model for this study obtained various statistical indices like R
(considering the testing R value as a good model), MSE, and RMSE.

The best model is sought with a 0.9763 value of R (coefficient of correlation) and a
corresponding MSE value of 0.000318, whereas the MLR model gives R and MSE values
of 0.8831 and 0.02272 respectively. The number of hidden neurons in ANN was taken

67
differently to obtain a good result of the model in which increases in neurons number lead
to more training time and a greater number of epochs, later the decreased number of
neurons and selected the best model architect to fit for this study. From the above result, it
is clear that both the ANN and MLR model were able to predict the catchment runoff with
adequate accuracy. The ANN and MLR models result demonstrate that the ANN model is
the best fit for the prediction of catchment runoff for the present study. Moreover, the
Model1, which has 18 inputs parameters ANN model provides a slightly better result with
Levenberg Marquardt than the ANN model with a Scaled conjugate gradient algorithm.

68
Chapter 7

Conclusion

CHAPTER 7

69
 CONCLUSION

The ANN hydrological model with Feed-Forward Back Propagation network is

developed in the present study for Banas River catchment, Gujarat, India. The performance
of the developed models was evaluated by statistical evaluation measurements, such as
correlation coefficient (R), Mean Square Error, and Root Mean Square Error (RMSE). For
the ANN models, the evaluation shows correlation coefficient (R) has been obtained as
0.9763. Also, for an MLR model, the correlation coefficient (R) was obtained as 0.8831.
The RMSE values for the ANN and MLR models were 0.000318 and 0.15073
respectively. The architect 18-9-1 ANN model was the best model among all of the models
developed in this study.

The important parameter, when predicting a catchment runoff are the precipitation
of Chitrasani and Sarotry rain gauge station have a significant role. The results indicate
that the ANN model had a good ability to capture the relationship between input/output
i.e., Rainfall/Runoff better than an MLR model. The result concluded that ANN can
capture the nonlinearity of rainfall-runoff modelling very well with good predictive power
in hydrological models.

REFERENCES
BIBLIOGRAPHY

70
[1] S. Raid, J. Mania, L. Bouchaou, Y. Najjar, Rainfall-runoff model using artificial neural
network approach (2003), Journal of Science Direct, Elsevier.
[2] Ashu Jain, KV Prasad Indurthy, Comparative analysis of event-based rainfall-runoff
modelling techniques - deterministic, statical, and Artificial Neural Networks (2003),
Journal of hydrologic engineering, ASCE.
[3] A. R. Senthil Kumar, K. P. Sudheer, S. K. Jain, P. K. Agarwal, Rainfall-runoff
modelling using artificial neural networks: comparison of network types (2005), Wiley
InterScience
[4] Ajay B Patel, Geeta S Joshi, modelling of rainfall-runoff correlations using artificial
neural network (2017), Civil Engineering Journal.
[5] Lateef Ahmed Dar, Rainfall-runoff modelling using artificial neural network technique
(2017), International Research Journal of Engineering and Technology.
[6] Anant Patel, Rainfall-runoff modelling and simulating using remote sensing and
Hydrological model (2020), springer.
[7] Vikas Poonia, Hari Lal Tiwari, Rainfall-runoff modeling for the Hoshangabad Basin of
Narmada River using artificial neural network (2020), Saudi Society for Geosciences,
Springer.
[8] A. M. Kalteh, Rainfall-runoff modelling using artificial neural networks (ANNs):
modelling and understanding (2008), Caspian Journal of Environmental Sciences,
Research Gate.

WEBOGRAPHY
1. MathWorks “Neural Network Introduction”
(https://in.mathworks.com/discovery/neural-network.html)
2. Banas River, Narmada, Water Resources, Water Supply and Kalpsar Department
(https://guj-nwrws.gujarat.gov.in/showpage.aspx?contentid=1586&lang=english)

BOOKS
[1] Ven Te Chow, David R Maidment, Larry W Mays. Applied Hydrology. McGraw Hill
book company, international edition 1988.
[2] K. Subramanya. Engineering Hydrology. McGraw Hill book 2008.

OTHER REFERENCES
1. Keith Beven, Rainfall-runoff modelling the primer, A John Wiley & sons, Publication

71
2. Laurene Fausett. Fundamental of Neural Networks.

3. Phil Kim. MATLAB deep learning with machine learning, neural network and artificial
Intelligence. 2017

72