Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Artificial Neural Network Application inWater Resource Management and Flood
Warning: Case Study North West Malaysia
By
Hassanuddin Mohamed Noor
This thesis is submitted in partial fulfilment of the requirements for thereward of the degree of Doctor of Philosophy University of Portsmouth
i
Abstract
Disasters caused by floods are a major cause of losses of properties and lives. The
unpredictability in weather conditions due to changing weather patterns do not only
lead to flooding but also contribute to water resource management problems. Rapid
development in many tropical countries, like in Malaysia, has resulted in the loss of
natural floodplains leading to an increase in flooding and water shortage. Sufficient
advanced flood warning system that can save lives and properties can be developed
using accurate river model. The work reported in this thesis has made significant
contributions in the prediction of river flow rate based on rainfall rate in the
catchment area using Artificial Neural Network (ANN). The proposed approach
models the non-linear process of the rainfall-runoff in a wide variety of catchment
area conditions. This study demonstrates significant improvement in the accuracy
and reliability of water resource management by using ANN modelling to predict
river flow rate. It also shows ANN as a fast and adaptable approach that is suitable
for river flow rate modelling that does not need detailed geographical information of
the catchment area. Its attractiveness is in its ability to adapt to changing conditions
and therefore does not become outdated like conventional hydrology models. The
research shows that river flow rate is a better parameter to be used for an early flood
warning system as it is more sensitive to rainfall rate compared to the river level
which is used in conventional flood warning systems. The study has also shown that
ANN with a feed-backward network with one hidden layer provides the best results
and it is able to produce river flow rate prediction up to 132 hours with root mean
square error of 0.02 / . This is a significant contribution as the flood warning
system currently used in Malaysia can only predict flooding within 8 to 24 hours.
The work in this thesis can assist the authorities to manage water from dams thereby
effectively managing floods and ensuring sufficient water for domestic and
agricultural use. The findings of this research has already been presented to the
Malaysian government agency responsible for managing waterways and dams.
ii
Acknowledgement
This dissertation would have never been completed without the guidance of my
advisors and committee members, help from friends, and support from my family,
wife and Majlis Amanah Rakyat (MARA) whose sponsoring my study.
I would like to express my deepest gratitude to my study director, Dr David L Ndzi,
for his excellent guidance, encouragement, and providing me with an excellent
atmosphere for doing research. I would like to thank Dr David Sanders and Dr
Shikun Zhou for their services as my annual reviewer panels and helping me to
develop the scope of my research. Also fully thanks to Department of Irrigation and
Drainage (DID) Malaysia for providing the valuable data to be used in this studies.
I am indebted to my friends and the members of the School of Engineering, especially
from Room A2.09 and to Dr Guangguang Yang because this research would not have
been possible without their help. I would like to thank my parents, children, and
friends for their prayers and best wishes. This dissertation is made possible with the
undivided support and encouragement from my wife, Noor Azma. She is always
there and stood by me through the good and bad times. Special thanks also go to my
previous employer, Kolej Kemahiran Tinggi MARA Kemaman (KKTM), Bahagian
Kemahiran dan Teknikal MARA (BKT) and Bahagian Sumber Manusia MARA (BSM)
for supporting my ambition to complete the doctoral degree in University of
Portsmouth.
iii
Declaration: Whilst registered as a candidate for the above Degree, I have not been
registered for any other research award. The results and conclusions embodied in this
thesis are the work of the named candidate and have not been submitted for any other
academic award.
iv
Table of ContentsPage
Abstract …………………………………………………………........................ i
Acknowledgements …………………………………………………………...... iii
Table of Contents………………………………………………………….......... iv
List of Figures ………………………………………………………….............. viii
List of Tables …………………………………………………………................ xi
List of Acronyms………………………………………………………….......... xiii
Publications…………………………………………………………................... xv
CHAPTER 1. INTRODUCTION 1
1.1 Research Background …………………………………………………………... 1
1.2 Introduction of Weather in Malaysia …….……………………………...... 2
1.3 Statement of Aims………………………………………………………………. 3
1.4 Objective ………………………………………………………………………. 4
1.5 Justification………………………………………...…………………………….. 5
1.6 Background and Research Context………………………………………….. 5
1.7 Flood Warning System in Malaysia…………………………………………… 6
1.7.1 Telemetry System…………………………………………………………........ 7
1.7.2 Manual Water River Level Monitoring………………………………………. 7
1.7.3 Flood Warning Board…………………………………………………………… 7
1.7.4 Mass media………………………………………………………………………. 7
1.8 Structure of the Thesis……………………………………………………… 8
1.9. Major Outcome………………………………………………………………….. 9
CHAPTER 2. LITERATURE REVIEW 10
2.1 Existing Flood Forecasting Techniques…………………………………… 10
2.1.1 River and Rain-Gauge Networks………………………………………………. 12
2.1.2 Radar and Geographic Information Systems………………………………… 13
v
2.1.3 Linear Statistical Models……………………………………………………….. 13
2.1.4 Nonlinear Time Series Analysis and Prediction…………………………... 13
2.1.5 Artificial Neural Networks (ANN)………………………………………….. 14
2.2 Weather radar…………………………………………………………………… 14
2.3 Satellite-Based Weather System ……………………………………………… 14
2.4 Introduction to Basic Hydrology Modelling………………………………… 15
2.4.1 Rational method……………………………………………….……………… 17
2.4.2 Time and area diagram (TAD) ……………………………….……………… 18
2.4.3 Unit Hydrograph……………………………….……………………………… 18
2.5 Approaches to Hydrological Modelling………………………………………. 19
2.5.1 Deterministic Models…………………………………………………… 20
2.5.2 Conceptual Models……………………………………………….……………. 20
2.5.3 Empirical Models………………………………………………………………. 21
2.6 Summary…………………………………..……………………………………. 21
2.7 Major Outcome …………………………………..…………………………… 22
CHAPTER 3. ARTIFICIAL NEURAL NETWORKS (ANN) 23
3.1 Introduction.………………………………………………………………… 23
3.2 Nonlinear AutoRegressive Neural Network with eXogenous Input….. 24
3.3 Types of ANN…………………………………………………….………… 24
3.4 Development of an ANN …………………………………………………… 25
3.4.1 ANN Architecture…………………………………………………….………… 26
3.4.2 Training an ANN model……………………………………………………. 28
3.4.2.1 Levenberg-Marquardt (LM) Algorithm……………………………………… 29
3.4.2.2 Bayesian Regularization (BR) Algorithm…………………………………… 30
3.4.3 Network Design Configuration…………………………………………… 30
3.4.4 Variable Selection…………………………………………………………….. 30
3.4.5 Data Pre-Processing. …………………………………………………………… 31
3.4.6 Training, Testing, and Validation. …………………………………..…….. 32
vi
3.5 Neural network paradigms. ………………………………………………. 32
3.6 Evaluation criteria. …………………………………………………………….. 33
3.7 Neural network training. ………………………………………………………. 33
3.8 ANN application in prediction model. ……………………………………… 34
3.9 Implementation………………………………………………………………… 35
3.10 Model Evaluation……………………………………………………………… 37
3.11 Correlation Coefficient……………………………………………………….. 38
3.12 Advantages and Disadvantages of ANN for Hydrological Modelling….. 38
3.13 Summary………………………………………………………………………… 39
3.14 Major Outcome……………………………………………………………… 40
CHAPTER 4 METHODOLOGY
4.1 Study area ……………………………………………………………………… 42
4.2 Mann-Kendall (MK) trend test……………………………………………… 46
4.3 Rain Dynamics…………………………………………………………………… 47
4.4 Catchment Area Characteristic………………………………………………… 51
4.5 Surface-runoff relationship………………………………………………. 52
4.6 ANN-based river flow rate model ……………………………………………. 55
4.7 Comparison between river level and river flow rate measurement ……….. 56
4.8 Summary………………………………………………………………………… 56
4.9 Major Outcome …………………………………………………………………. 57
CHAPTER 5. RESULTS
5.1 Rainfall statistics. …………………………………………………………… 58
5.2 Comparison between level monitoring and flow monitoring for flood
observation……………………………………………………………………… 63
5.2.1 Case 1: River Pelarit Flood event on 30th October to 1st November 2010. 64
5.2.2 Case 2: River Jarum Flood event on 30th October to 1st November 2010…. 66
5.2.3 Case 3: River Pelarit, Flood event on 1 April 2011…………………………. 67
5.2.4 Case 4: River Jarum, Flood event on 1 April 2011……………………………. 67
vii
5.2.5 Summary for comparison between level monitoring and flow
measurement
68
5.3 Rainfall in the catchment area and river flow rate ………..…………………. 69
5.3.1 Dry period before rain against time delay to change flow rate…………… 69
5.3.2 Rain-No change condition……………………………………………………… 70
5.4 ANN-based River Flow Model………………………………………………… 71
5.4.1 Model I: Flow Rate Model for River Pelarit using LM Training Algorithm. 73
5.4.2 Model II River Flow Rate Model for Pelarit using BR Algorithm …………. 76
5.4.3 Model III River Flow Rate for River Pelarit using LM Training Algorithm. 78
5.4.4 Model IV River Flow for River Pelarit using BR Training Algorithm…….. 79
5.4.5 Summary Experiment for River Flow Model River Pelarit………………… 81
5.4.6 Model V River Flow Model for River Jarum -LM Training Algorithm (Oct
2013)
82
5.4.7 Model VI River Flow Model for River Jarum - BR Training Algorithm
(Oct 2013)
83
5.4.8 Model VIII River Flow for River Jarum -LM Training Algorithm (Oct
2010)
88
5.4.9 Summary Experiment for River Flow Model for River Pelarit……………. 89
5.5 Summary …………………………………….…………………………… 89
5.6 Major Outcome…………………………………………………………… 91
Chapter 6: CONCLUSION AND FUTURE WORK 92
6.1 Summary of the Study………………………………………………………… 92
Contributions ……………………………………………………………………. 95
6.3 Recommendations for Future Works in ANN Rainfall-Runoff Modelling... 95
REFERENCES
APPENDICES
Form UPR16- Research Ethics Review Checklist
Certificate of Ethics Review
viii
List of FiguresPage
CHAPTER 1 INTRODUCTION
Fig. 1.1 Monsoon seasons in Peninsular Malaysia. 3
Fig. 1.2 Annual rain distribution for Peninsular Malaysia 5
Fig. 1.3 Flood-prone area in Malaysia 6
Fig. 1.4 Weather warning issued by MMD. 8
Fig. 1.5 Thesis structure 8
CHAPTER 2 LITERATURE REVIEW
Fig. 2.1 The phase of flood event for water reservoir or lake. 10
Fig. 2.2 Peninsular Malaysia weather radar network 15
Fig. 2.3 The time area diagram obtained by splitting the basin into n areas
according to flow time to the outlet.
18
Fig. 2.3 Snyder's synthetic unit hydrograph 19
CHAPTER 3 ARTIFICIAL NEURAL NETWORKS
Fig. 3.1 Layout of feed-forward neural network 23
Fig. 3.2 Two-layer network with abbreviated notation 27
Fig. 3.3 Local and global minima of errors 28
Fig. 3.4 Eight steps in designing neural networks forecasting model 30
Fig. 3.5 Training stopping due to convergence 34
Fig. 3.6 A schematic of the ANN that been applied in the model 36
CHAPTER 4 METHODOLOGY
Fig. 4.1 Development of River Flow Modelling for Timah Tasoh Basin 41
Fig. 4.2 Annual rain distribution for Peninsular Malaysia 42
Fig. 4.3 Location of the hydrological station in Perlis, Malaysia. 43
Fig. 4.4 Land usage in Timah Tasoh Reservoir catchment area 44
Fig. 4.5 The distance of River Pelarit from the gauging station in Kg Batu 45
ix
Limabelas.
Fig. 4.6 River Jarum- gauging station at Kg Masjid to Timah Tasoh Lake 45
Fig. 4.7 Monthly distribution in 2013 for rain stations in Perlis. 48
Fig. 4.8 Location of four rain stations in Timah Tasoh catchment. 48
Fig. 4.9 Different scenarios of rain gauge grid for rain cell detection 50
Fig. 4.10 The direction of rain cell movement within the catchment area 50
Fig. 4.11 Padang Besar daily rainfall in the year of 2010. 52
Fig. 4.12 Relation of rainfall and river flow in the basin of Timah Tasoh 54
Fig. 4.13 Water Level Classification at Flood Warning Centre 56
CHAPTER 5 RESULTS
Fig. 5.1 Monthly average for the accumulated rainfall for 14 meteorology
stations in Perlis from 2010 to 2013
58
Fig. 5.1 Comparison of Rainfall Rate from 2010 to 2013 in Perlis 61
Fig. 5.3 Rain intensity in 2010 to 2013 for 14 weather stations in Perlis 61
Fig. 5.4 Rainfall for Kaki Bukit, level and flow rate of River Pelarit during the
2010 flooding event
64
Fig. 5.5 The plot of Jarum River Level, Flow rate and rainfall at Padang Besar
Weather station.
66
Fig. 5.6 Plot of River Pelarit Level, Flow rate and rainfall at Kaki Bukit weather
station
67
Fig. 5.7 Rainfall during the 2011 flooding event for Padang Besar and River
Jarum at a recording interval of 5 minutes
68
Fig. 5.8 Relation of rainfall and flow in the basin of Timah Tasoh. 69
Fig. 5.9 Simplified diagram of Timah Tasoh Catchment Area 72
Fig. 5.10 Prediction from Model 1: River Pelarit. Deterministic forecast: (a)
observed and predicted time series; (b) residual plot; (c) scatter plot.
75
Fig. 5.11 Prediction from Model 2: River Pelarit for Oct. 2013 dataset.
Deterministic forecast: (a) observed and predicted time series; (b)
77
x
residual plot; (c) scatter plot.
Fig. 5.12 Prediction from Model 3: River Pelarit for Nov 2010 dataset.
Deterministic forecast: (a) observed and predicted time series; (b)
residual plot; (c) scatter plot.
79
Fig. 5.13 Prediction from Model 2: River Pelarit for Oct. 2013 dataset 81
Fig. 5.14 Prediction from Model V: River Jarum for Oct. 2013 dataset 83
Fig. 5.15 Prediction from Model VI: River Jarum for Oct. 2013 dataset 85
Fig. 5.16 Prediction from Model VI: River Jarum for Oct. 2010 dataset 87
Fig. 5.17 Prediction from Model VI: River Jarum for Oct. 2010 dataset 89
CHAPTER 6 CONCLUSIONS AND FUTURE WORK
Fig. 6.1 Simplified analysis for model flood prediction in Timah Tasoh 92
xi
List of TablesPage
Table 3.1 Summary of the settings to be used in this experiment and their
values.
36
Table 4.1 Detail of land use in Timah Tasoh catchment area. 44
Table 4.2 Four significance level and the required number of data n. 47
Table 4.3 Rain gauge stations in Timah Tasoh catchment 48
Table 4.4 Rain records for 4 stations in Timah Tasoh Catchment. 51
Table 4.5 Total of data for each station 55
Table 4.6 Related stations for Pelarit and River Jarum. 55
Table 4.7 Stages level of rivers (meter) in Perlis, Malaysia 56
Table 5.1 The total and duration of rainfall from 2010 to 2013 for 14 stations in
Perlis
59
Table 5.2 Frequency of rainfall rates for 10 to 150 mm/h in Perlis 59
Table 5.3 Descriptive statistic rainfall of 5 minutes for 14 stations in Perlis for
2010 to 2013
60
Table 5.4 The descriptive statistic for daily rainfall of 14 stations in Perlis for
2010 to 2013.
62
Table 5.5 Maximum rainfall rate (mm/h) for 2010 to 2013 for 14 stations in
Perlis
62
Table 5.6 Numbers rain of occasions in 2013 in 14 rain stations in Perlis 62
Table 5.7 Percentage of rain event in 4 different sessions 63
Table 5.8 Stages level of River Pelarit (meter) 64
Table 5.9 River level and flow rate changes during 2010 flood event 65
Table 5.10 Summary of parameters changes during 2010 flood event 66
Table 5.11 Water level as classified by DID for River Jarum 66
Table 5.12 Summary of parameters changes during 2011 flood event 67
xii
Table 5.13 Summary of parameters changes during 2011 flood event 68
Table 5.14 Dry time and Flow delay for River Pelarit and River Jarum. 70
Table 5.15 Rain with no changes in Pelarit and River Jarum 71
Table 5.16 Details of Input and the Output for Pelarit and River Jarum Flow
Prediction Model.
73
Table 5.17 River Pelarit (Training: train LM, Target: Oct 2013). Grey shading
denotes the best performing model.
74
Table 5.18 River Pelarit (Training: train BR, Target: Oct 2013). Grey shading
denotes the best performing model.
76
Table 5.19 Pelarit Prediction Model. (Training: train BR, Target: Nov 2010).
Grey shading denotes the best performing model.
78
Table 5.20 Pelarit (Training: train LM, Target: Oct 2010). Grey shading denotes
the best performing model
79
Table 5.21 Overall of the best prediction model performances based on
different training algorithms for River Pelarit.
81
Table 5.22 Jarum Prediction (Training: train LM, Target: Oct 2010) 82
Table 5.23 Jarum (Training: train BR, Target: Oct 2010) 84
0 Jarum (Training: train BR, Target: Oct 2010) 87
Table 5.25 Model of Prediction for River Jarum (train BR, Target: Oct 2010). 88
Table 5.26 Best River Flow Model performance based on different training
algorithms for River Jarum.
89
xiii
List of AcronymsAMSR-E
ARMA
ANN
ARIMA
BR
R2
CCD
DID
GIS
GMS
GFS
GSMaP
HMM
LM
MMD
MYR
MK
MAPE
NN
NNARX
NLP
QPE
RBFN
RNN
RSME
SOM
SSMI
SEVIRI
TBP-NN
TSDM
TMI
Advanced Microwave Scanning Radiometer for the Earth Observing System
Auto-Regressive Moving Average
Artificial Neural Network
Autoregressive Integrated Moving Average
Bayesian Regularization
Coefficient of determination or Regression
Cold Cloud Durations
Drainage and Irrigation Department
Geographic Information System
Geosynchronous Meteorological Satellite
Global Forecast System
Global Satellite Mapping of Precipitation
Hidden Markov Models
Levenberg-Marquardt
Malaysian Meteorological Department
Malaysian Ringgit
Mann-Kendall
Mean absolute percentage error
Neural Network
Neural Network Autoregressive with Exogenous Input
Nonlinear Prediction
Quantitative Precipitation Estimations
Radial Basis Function Networks
Recurrent Neural Network
Root Mean Square Error
Self-Organizing Map
Special Sensor Microwave Imager
Spinning Enhanced Visible and Infrared Imager
Temporal Backpropagation Network
Time Series Data Mining Methodology
Tropical Measuring Mission Microwave Imager
xiv
Publications
Book Chapter
Mohd-Safar, N. Z., Ndzi, D.L., Sanders, D.A., Noor, H. M. & Kamarudin, L. M.,‘Integration of Fuzzy C-Means and Artificial Neural Network for Short-TermLocalized Rainfall Forecast in Tropical Climate’, in SAI Intelligent SystemsConference (IntelliSys), London, 2016, pp. 930–938.
Conference Papers
Author
Hassanuddin Mohamed Noor, David Ndzi, Guangguang Yang, Noor ZuraidinMohd Safar
‘River Flow Rate Prediction Using NARX for Advanced Flood Warning in NorthernMalaysia.’
13th IEEE Colloquium on Signal Processing and its Applications (CSPA 2017),Penang, Malaysia, 10th to 12th March 2017.
Author
Mohd-Safar, Noor Zuraidin; Ndzi, David Lorater; Sanders, David Adrian; Noor,Hassanuddin Mohamed; Kamarudin, Latifah Munirah / Integration of fuzzy c-means and artificial neural network for short-term localized rainfall forecast intropical climate.
Proceedings of 2016 SAI Intelligent Systems Conference (IntelliSys). IEEE, 2016.
Author
H. Md. Noor, D. L. Ndzi, D. Sanders, N. Z. Mohd-Safar, S. Zarina Ibrahim, L. M.Kamarudin, A. Y. M. Shakaff.
‘Study of Rain Characteristics for Early Flood Warning System using Rain GaugeNetwork’.
4th International Malaysia-Ireland Joint Symposium on Engineering, Science andBusiness (IMiEJS), Penang, Malaysia, 25-26 June 2014.
1
Chapter 1 Introduction
This chapter provides a prologue to the research work presented in this thesis. It
describes the research background and the objectives. It also outlines the structure of
the thesis.
Research Background
Flood prediction and water resource management are major issues, especially in
highly populated areas in tropical countries. Unlike in temperate climate regions,
such as in Europe and North America, high resolution spatial and temporal weather
data is not available in most tropical countries. Rainfall in the tropics are
predominantly convection rain that is characterised by small rain cell size, high
intensity and short durations that easily result in flash floods. Flooding in populated
areas has become a major problem in many developing countries as developments in
flood plains have increased. Accurate and fast adaptive flood prediction systems are
have therefore become indispensable, especially as hydrology models become
outdated as land use changes.
An accurate prediction of river flow rate and level are some ways of reducing the risk
of flooding. This can ensure the effective management of flood barriers, activation of
temporary flood defences and early warning to residents. In principle, the longer the
lead time of advanced forecast, the longer the time for precautions to be taken.
One of the biggest problems in designing a flood forecasting system is that models
that take into account ground cover are easily outdated as a result of development in
the floodplains. Watersheds and floodplains are continuously changing as
development take place.
To predict a river’s characteristics, a detailed understanding of its catchment area is
required. This involves a comprehensive modelling that takes into account external
elements like rainfall, heat transfer and radiation combined with internal factors such
as the state of vegetation, soil moisture, and groundwater levels. The needs to forecast
floods and droughts and manage water resources would therefore, require regular
2
modifications to a hydrological modelling system.
An example of a catchment model is the Lumped Based Basin Model [1], this is a
conventional model based on the assumption that some hydrologic parameters can
represent each sub-basin within a watershed. The parameters are a weighted average
representation of the entire sub-basin. The main hydrologic inputs of this model are
precipitation, depth, and temporal distribution. Various geometric parameters such
as length of the stream, land slope, catchment area, centroid location, soil types, land
use, and absorbency are also required for the development of a traditional lumped
based model. However, these parameters are not always available and do change
which makes the conventional lumped based model inaccurate. Forecasted river flow
rate can be produced with the help of an accurate river model and flooding forecasted
based on the prevailing conditions.
Introduction to weather in Malaysia
Malaysia can generally be divided into two parts, West or Peninsular Malaysia and
East Malaysia (Borneo). Malaysia has a tropical climate with two major monsoon
seasons, North East (NE) monsoon (from November to March) and South West (SW)
monsoons (between May and September). Fig. 1.1 illustrates monsoon seasons rain
direction in Peninsular Malaysia. Both monsoons bring heavy rainfall, and as a result,
Malaysia receives between 2000 to 4000 mm of precipitation annually with 150 mm
to 200 mm of rain per day in some cases [2]. The transition between the NE and SW
monsoons (and vice versa) occurs in April and October [3]. Continuous rainfall for
a number of days is common in West Malaysia.
3
Statement of Aims
This research includes the study of rain dynamics across the catchment area and the
effect on river flow characteristic. Thus, broadly speaking, this research can be
divided into the following:
Study of the relation between rainfall rate and runoff in the study area.
Design and development of a computer-based model to predict river flow rate
based on rainfall in the catchment area.
This study focuses mainly on the application of time series analysis and Artificial
Neural Network (ANN) to build a model to forecast the river flow rate.
Understanding some of the rain characteristics is key to the development of an
efficient early flood prediction water resource management system [4].
NE Monsoon(Nov – Mar)
SW Monsoon(May – Sep)
Inter Monsoon(April and Oct)
4
Objectives
The objectives of this research can be outlined as follow:
To study rain characteristics in a selected area in Malaysia using available rain
gauges data.
To study the relationship between rainfall rate, river level and river flow rate
to identify which river characteristics (level and flow rate) has the strongest
relationship with rainfall;
To identify a possible technique(s) that can be used to predict the identified
river characteristics.
To develop a computer-based model, to predict the river characteristics, test
and analyse its performance using measured data.
The above objectives have been achieved as follows steps:
Perlis, a state in Northwest Malaysia was selected an area of study. River Jarum
and River Pelarit were selected.
It has been shown that the river flow rate has the most significant correlation
with rainfall compared to river level. Therefore, river flow rate has been
selected as the parameter to predict.
Nonlinear Autoregressive Exogenous (NARX) technique, one of the Artificial
Neural Network (ANN) has been selected and used to model the river flow rate
based on rainfall in the catchment.
The implemented method can predict river flow rate based on rainfall rate up
to 132 hours with a correlation coefficient of more than 0.9.
5
Justification
Malaysia has many rivers and receives a high amount of rainfall annually. Fig. 1.2
shows annual rain distribution in the Peninsular Malaysia [5]. However, Malaysia
faces serious water management challenges [6]. Rapid urbanisation has accelerated
changes in catchment hydrology and geomorphology. Development within river
watersheds has resulted in higher runoff and decreasing river capacity that has led to
an increase in flood frequency. Furthermore, modelling rainfall-runoff relationship is
inadequately available in Malaysia. However, there are facilities available to provide
real-time measurement of rainfall and rivers. They are available from the Department
of Irrigation and Drainage (DID).
Background and Research Context
Severe floods occurred in Malaysia in 2006, 2007, 2010, 2011, 2013, 2014 and 2015. In
December 2013, most parts of low-lying areas in states of Johor, Pahang, and
Terengganu were flooded and more than 30,000 people were affected [7]. In
December 2015, the worst flooding for 50 years affected four states (Kelantan,
Terengganu, Pahang and Perak) [8] which led to the evacuation of more than 200,000
6
people and losses estimated at over RM 2.9 Billion Malaysian Ringgit (MYR) or £530
million [9].
Fig. 1.3 shows that there are more than ten areas or river basins in Malaysia that are
prone to flooding. Most of these areas are densely populated, and any flood events
will have a high impact on the community and its economic activities. Therefore, the
success of this study could make a very positive contribution to flood prediction in
Perlis in particular and Malaysia in general.
Flood warning system in Malaysia
Malaysia does not have a coordinated system of advanced flood prediction. It has an
evolved national and regional disaster response system that needs better information
regarding flood prediction models [10]. Flood forecasting and warning services are
concentrated in river basins that are highly populated. There are several types of
flood warning and monitoring systems in Malaysia.
Legend
Flood Prone Area
7
1.7.1. Telemetry System
The telemetry system allows measured data, mainly river levels, to be transferred
wirelessly to an online system managed and controlled by the Department of
Irrigation and Drainage (DID). DID is responsible for collecting and analysing rainfall
and monitoring the river levels.
1.7.2. Manual Water River Level Monitoring
A total of 137 river level real-time monitoring stations have been set up in Malaysia.
When water level exceeds the predetermined critical level, a local observer will
transmit the real-time water level information to the DID state office for further
actions.
1.7.3. Flood Warning Board
Flood Warning Boards have been installed in flood-prone areas in the major river
basins. Warning notices on these boards are correlated to the water level at
observation points upstream. The residents are informed of the impending flood
situation in their area.
1.7.4. Mass media
Fig. 1.4 illustrates the example of warning been published by Malaysia Meteorology
Department (MMD) due to heavy rain at 5 a.m. 2nd December 2016. The blue zone is
where thunderstorms are forecasted, and the red zone is where torrential rains are
forecasted that will last for up to three days. These warnings are published on social
media, mass media and other forms of the electronic broadcasting system to warn the
public.
8
Structure of the thesis
Chapter 1 has introduced the research area. This is followed by Chapter 2 for
literature review discussion. Fig. 1.5 shows the overview of the structure of the thesis.
Chapter 1: Introduction
Chapter 2: Literature Review
Chapter 3: ANN
Chapter 4: Methodology
Chapter 5: Result
Chapter 6: Conclusion
Prologue of thesis
Research literaturereview and theoreticalbackground
Research design andcontributions
Case study, researchresults, analysis andconclusion
THUNDERSTORM WARNING HEAVY RAIN WARNING
Legend
Affected Area
Heavy rain for another 1 to 3 days
Medium rain, sometimes heavy rain will be occurred exceeding 1 day
Heavy rain currently occurred for more than 1 day
9
Major Outcome
This chapter, has discussed the importance of river modelling due to the risk of
flooding especially in highly populated area in Malaysia. Present flood early warning
system in Malaysia has proven to be inadequate to give early warning to the public.
As a result, loss of lives and damages to infrastructure occurs due to series of severe
floods. This shows the importance of developing accurate flood prediction and flood
early warning system for Malaysia to minimize these losses.
10
Chapter 2 Literature Review
This chapter discusses the concept and theories that are related to the objectives of
this study. A review of published studies on rain, hydrological modelling, flood
prediction and flood management systems are presented. Flooding can be defined as
“a general and temporary condition of partial or complete inundation of normally
dry land areas from overflow of inland or tidal waters from the unusual and rapid
accumulation or runoff of surface waters from any source” [11]. Also as “the covering
of normally dry land by water that has escaped or been released from the normal
confines of any lake, or any river, creek or another natural watercourse, whether or
not altered or modified; or any reservoir, canal, or dam” [12].
Fig. 2.1 The phases of rain caused flood event for water reservoir or lake
Fig. 2.1 shows the natural phases that lead to a rain caused flood event. Firstly,
precipitation in the catchment area leads to the water level in the ground increasing
Rain cell bringing precipitation
Water level in the ground increased and becomesaturated
Groundwater flooding occurred when water level inground rise above land surface
Additional rain fall on saturated soil become surfacerun off
Water from surface run off flow to river
Excess water flow into the river and increases flowrate to downstream river
Level of water in dam increased and the dam had torelieve excess water and becoming river flood
11
until the ground become saturated (tipping point). Groundwater flooding occurs
when water levels in the ground rise above the land surface. Normally groundwater
levels rise and fall according to the annual cycle but periods of prolonged rainfall may
cause water levels to rise above the land surface. Additional rainfall on saturated soil
become surface runoff. Excess water flows into rivers and increases their flow rates
and levels downstream. For a river that discharges into a reservoir, dam or lake, it
changes not just the amount of water in the reservoir or dam, but the rate at which
they are filled. Depending on the topology, the high level of water in the reservoir
may result in the flooding upstreaming or in the surrounding low-lying areas.
Flooding predominantly occurs downstream of managed dams where excess water
is not released gradually, especially when inflow into the dam is high and has not
been predicted. This is a widespread problem in South East Asia, in general, and in
Malaysia, in particular, where the dams are used for water supply to homes and are
a key part of the irrigation infrastructure.
2.1 Existing Flood Forecasting Techniques
Flood prediction is a complex process because of the numerous factors that affect
river flow rates and levels such as the topology, rainfall amount and duration, soil
types and size of catchments. The relationship between these factors often needs to
be fully understood. Classical linear Gaussian time series (deterministic) models are
inadequate for the analysis and prediction of complex geophysical phenomena [13].
Linear methods such as Auto-Regressive Moving Average (ARIMA) approach are
unable to identify complex characteristics due to the need to characterise all-time
series observations, the necessity of time series stationarity and, the requirement of
normality and independence of residuals [14]. Nonlinear time series approaches such
as Hidden Markov Models (HMM) [15], Artificial Neural Networks (ANN) [16-17]
and Nonlinear Prediction (NLP) [18] techniques applied to discharge forecasting only
produce accurate predictions for short prediction periods due to the dynamic nature
of rainfall and its complexity.
12
The time dependency of the modelling required has resulted in many studies based
on nonlinear modelling with time delay parameters. These time dependency
techniques also have been applied to varieties of physical applications such as in the
domains of physiology [19][20], financial and economics [21], geophysics [22] and
engineering applications [23]. The Time Series Data Mining Methodology (TSDM) is
based on a variant of time-delayed embedding called the reconstruction of phase
space [30- 31]. TSDM framework combines the methods of phase space reconstruction
and data mining to reveal hidden patterns predictive of future events in nonlinear,
nonstationary time series events. This could be applied to river flow rate or river level
prediction.
In recent years, numerous studies from hydrodynamics, civil engineering, statistics
and data mining have contributed to the area of flood prediction. Some of the
current measurement and techniques used in flood prediction can be categorized as
follow [26]:
River and Rain-Gauge Networks.
Radar and Information Systems.
Linear Statistical Models.
Nonlinear Time Series Analysis and Prediction.
The first two techniques use direct measurements and knowledge of past events to
predict flooding. The linear statistical models use the weighted flood discharge,
precipitation intensity, elevation, stream length, and main channel slope for flood
prediction [27]. The ANN modelling falls in the fourth categories which are Nonlinear
Time Series Analysis and Prediction, which is been selected for this studies.
River and Rain-Gauge Networks
Presently, river flow rate, river levels and precipitation measurement are monitored by
the Department of Irrigation and Drainage (DID) in Malaysia. DID has installed and
operated 525 telemetric stations in 38 river basins. Additionally, 670 manual river
gauges, 1013 stick gauges and 182 flood warning boards have been set up in flood-
13
prone areas to provide crucial information during the flood season. 395 automatic
flood warning sirens are an integral part of the local flood warning system is used
[28]. Data from the monitoring stations are used to predict possible flood events. This
approach often does not provide sufficient warning time as rainfall is intense and
localised which may affect areas downstream very far from it rained.
Radar and Geographic Information Systems
Presently, with the development of the radar system and the Geographic Information
System (GIS) is being used along with the traditional river models for improved
flood forecasting [29]. This complex system consists of simulation programs will
use multiple data source from various sources such as telemetry, gauges to
calculate runoff, infiltration and precipitation volumes using land use and elevation
information. However, the major disadvantages of this system are the complexity is
high, the longer time taken to produce prediction result and the high cost of
implementation and maintenance.
Linear Statistical Models
Linear Statistical Models such as autocorrelation functions, spectral analysis, and
analysis of cross- correlations, linear regression and Autoregressive Integrated
Moving Average (ARIMA) have been studied to access their application in flood
forecasting. However, several studies like Solomatine et al. founds that the use of
ARMA unable provides accurate predictions for long-term prediction [30][31].
Nonlinear Time Series Analysis and Prediction
Nonlinear Prediction (NLP) method [32] has been used in flood prediction by
researchers like Porporato et al. [33]. They have experimented with flood prediction
based on single parameter (river flow) time series [33][34] and also multiple variable
time series as well [35]. Some of the nonlinear time series approaches such as Hidden
Markov Models, and Artificial Neural Networks (ANN) are based on multiple time
series. Laio et al. [36] have performed a comparison of ANN and NLP approaches in
daily discharge forecasting NLP method was found to provide accurate forecasts
14
over a shorter prediction period (1-6 hours), but for the prediction periods exceeding
24 hours, the ANN approach is more accurate [26].
Artificial Neural Networks (ANN)
ANN is widely accepted as a potentially useful way of modelling complex nonlinear
and dynamic systems. ANN i s useful in situations where the underlying input
parameter relationships are not entirely understood or where the nature of the
process being modelled is complex. Although ANN does not remove the need for
knowledge or prior information about the systems of interest, it reduces the model's
reliance on this prior information. This removes the need for an exact specification
of the precise functional form of the relationship that the model seeks to represent.
ANN represent the input-output relation where different parameters such as
rainfall, flow rate, temperature, humidity, wind direction, etc. can be used as input to
the model. Numbers of researchers have used this technique in a variety of geophysical
studies [29]. There is, however, no set algorithm that can be applied to ensure that
the network will always produce an optimal solution and the nonlinear nature of the
ANN model often results in multiple predicted values. Details of ANN will be
discussed in Chapter 3.
2.2 Weather radar
In Malaysia, weather radar or rain radars are used to measure the
rainfall[4][37][38][39]. Twelve Doppler weather radars operate in the location shown
in Fig. 2.1 and northern Borneo (Sabah and Sarawak). These consist of S-band (8 sites)
and C-band radars (4 sites). Radar is used to estimate the distributions of precipitation
amounts over a large area. In addition to the large area, radar coverage allows the
detection of rain events which might not be detected by rain gauges [39]. However,
the application of radar-based rain measurement has limitations. Their ranges for
acceptable accuracy are limited and is reduced as the topology becomes more
complex [40]. The localised nature of rainfall rate in Malaysia means that they are not
a reliable means of measuring rainfall rate due to their limit spatial and temporal
15
resolutions.
Fig. 2.2 Peninsular Malaysia weather radar network [41].
2.3 Satellite-Based Weather System
The weather satellite is a type of satellite that is primarily used to monitor the weather
and climate of the Earth. Satellites can be polar orbiting, covering the entire Earth
asynchronously, or geostationary, hovering over the same spot on the equator.
With advances in satellite technology, some of the present rain forecasting and flood
prediction systems have been designed to incorporate geographical information
system (GIS) and ground sensors such as rain station. Examples include the Japanese
Geosynchronous Meteorological Satellite (GMS) with a spatial resolution of up to 10
km and the United States’ Global Forecast System (GFS) with a spatial resolution of
35 km. These render satellite-based systems inadequate for accurate prediction of
localised floods, especially in the tropical areas. Some of these satellite systems have
a temporal resolution of 6 hours. Studies have shown that the average size of rain cell
in Malaysia is between 1.2 to 1.5 km [42]. Rain cells in the tropical areas also have
high perception intensity which can be up to 120 mm per hour. It is typically localised
rather than widespread due to its convective characteristic and has a lifespan of less
16
than an hour [38][43]. Due to this limitation, flood prediction based on satellite
systems inherently produces poor prediction results [44][45].
Presently, the system with the highest resolution for rain detection system is the
Global Satellite Mapping of Precipitation (GSMaP) by the Japan Aerospace
Exploration Agency (JAXA) [46]. GSMaP has been designed to achieve a temporal
resolution of 1 hour, spatial resolution with grid latitude-longitude of 0.1 degrees,
and makes the data available approximately 4-hour after it is captured (quasi-real-
time basis). This system has been developed to assist developing countries with flood
prediction. However, studies in Chao Phraya River basin in Thailand has concluded
that the satellite-based system is not adequate for near-real-time rainfall monitoring
applications due to the dynamic nature of tropical rain cell which is small and
localized compared to the rain cell size in moderate climate [47]. Overall, the GSMaP
system has been found to be suitable for climate change studies.
Several studies have been conducted to improve the accuracies of satellite rain
prediction systems. One of the systems uses geostationary satellites with thermal
sensors to measure the daily minimum cloud top temperatures or Cold Cloud
Duration (CCD) to rain rates. The limitation of this method is that it assumes that rain
rates are related to cloud duration. This is an assumption that fails in case of high
rain rates that occur over a short period. Another method of calculating the rain rates
is the rate retrieval methods [48]. This uses satellite microwave radiometer
observations based on the fundamental principles of radiative transfer. However, this
is an approximation based on the properties of clouds and is susceptible to
inaccuracies in localised weather patterns and windy conditions [49]. Another
approach used for rain detection is based on soil moisture [50]. It detects the surface
soil moisture using low frequencies of microwave imagery systems such as Tropical
Measuring Mission Microwave Imager (TMI), Advanced Microwave Scanning
Radiometer for the Earth Observing System (AMSR-E), and Special Sensor
Microwave Imager (SSMI). The purpose is to improve the accuracy of GSMaP by
introducing the new method to complement the high-resolution global precipitation
17
prediction.
However, the drawback of most air-based rain detection system such as radar and
satellite is their poor spatial and temporal resolutions [51][52]. Some flood warning
systems prediction is very complex and mainly designed for large area prediction
which makes their application in tropical areas unreliable due to localised flooding
and short duration runoff or flash floods [53]. The nature of convection rain in tropical
areas increases the errors and uncertainties in flood prediction and storm modelling
systems [54][55]. These limitations have been proven in the application of the
topography based hydrological or known as TOPMODEL prediction system for river
level forecasting in Nigeria [53].
2.4 Introduction to Basic Hydrology Modelling
There are several basic hydrology models used to estimate discharge of water base
on simple calculations. Below are the some of the basic hydrology models.
2.4.1. Rational method
The Irish engineer Thomas James Mulvaney (1822-1892) published a simple equation
to determine the maximum discharge for a quantity of rainfall in a catchment. The
equation (2.1) published by Mulvaney in 1851 is the following:
Q = C A R (2.1)
A- Catchment Area.R- Maximum catchment average rainfall intensity.C- An empirical coefficient which is the proportion of rainfall that contributes to runoff.
The equation only predicts the hydrograph peak, not the whole hydrograph. The
maximum discharge, Q is predicted by the equation for a rainfall with the duration is
at least equal to the concentration-time of the catchment.
The method shows the way maximum discharge is presumed to increase the
catchment area and rainfall intensity. The calculation of this maximum discharge is
made in a rational way, hence the name of the method. The rational method is a
popular and easy to use technique for estimating peak flow in any small drainage
18
basin having mixed land use and should not be used in basins larger than 1 square
mile. Although this method does not offer the possibility of obtaining an entire
hydrograph but just its peak (maximum discharge) it is enough for hydrological
engineers to design bridges or dams that can sustain the maximum discharge.
2.4.2. Time and area diagram (TAD)
Using this concept, hydrologist obtains a time-area diagram (TAD) which the area
from which water flows in each time step as shown in Fig. 2.3. The time-area diagram
represents the lag in the time needed for water from each area of the catchment to
reach the outlet.
The time-area concept was used in the USA and in England. The idea is the base of
some newer distributed hydrologic models. Kull and Feldman adapted Clark's
method for use with radar data obtained from the NEXRAD system.
Fig. 2.3 The time area diagram obtained by splitting the basin into n areas according toflow time to the outlet.
2.4.3. Unit Hydrograph
The time-area concept is a method that has been used in hydrologic models. It has a
limitation because the accurate determination of the contributing areas for each time
step is difficult because the overland and groundwater flow speed is not easy to be
determined. Sherman tried to avoid this problem using the lag time needed for water
to reach the catchment outlet as a time distribution without a direct link to catchment
areas known as Unit Hydrograph [56].
19
Snyder's synthetic unit hydrograph is based on the creation of the hydrograph from
characteristics of the rainfall and the catchment. Snyder five characteristics of the
Unit Hydrograph can be obtained: the peak discharge per unit of watershed area, QpR,
the basin lag, TlR, the base time, Tb, and the widths, W (in time units) of the unit
hydrograph at 50 and 75 percent of the peak discharge as Fig. 2.4.
Fig. 2.4 Snyder's synthetic unit hydrograph
Another method to determine a hydrograph is the method proposed by SCS. This
method is used to create a dimensional hydrograph with ordinate values expressed
in a ratio Q/Qp (flow/peak flow) and containing the values of the ratio t/tp (time/time
to peak) on the abscissa. The dimensionless unit hydrograph can be used later to
determine a watershed-specific unit hydrograph knowing some characteristics of the
watershed. The data needed to apply the method are the area of the catchment A, the
time of concentration Tc and the duration of the unit excess rainfall D.
2.5 Approaches to Hydrological Modelling
Hydrological models are used to improve the understanding of the hydrological
processes as well as to make forecasts of the future hydrological events, e.g. flood
events, river levels, etc. taking into account rainfall, snowmelt and evaporation. All
models are simplifications of reality, and there are many different ways to represent
it. Some authors have categorized hydrological modelling approaches into three
main types: process-based; conceptual, and empirical or data-driven (which
includes statistical) [57].
20
This decomposes the physical characteristics of the catchment area into different
constituents and models them [57]. An example of this model is the European
Hydrological System [58]. It has been applied to a range of areas for flood
forecasting and used to examine the effects of land use change and groundwater
modelling in Europe.
Physically-based models represent a catchment as three-dimensional grid-like
global circulation models of the climate using the fundamental laws of
conservation of energy and mass to model water movements on the surface
through the unsaturated and saturated zones to the river. A flood hydrograph is
then dynamically built from the model runoff [59]. These models incorporate as
full an understanding of the catchment processes as possible and when the
conditions change, the models can be used to evaluate the impact of runoff or other
catchment properties [60]. Although these models are the most complex and accurate
of the three model types, they required a significant amount of data which is not
always available for all catchments and long processing time is needed. These model,
therefore, have more value for planning than for use in real-time forecasting.
The conceptual model is also known as the lumped conceptual or geomorphology-
based model [61]. These are viewed as the most successful models for rainfall-runoff
simulation and flood modelling [62]. These models still have a physical basis but have
been structured in such a way as to represent a stream as a network in the
surrounding catchment. These models attempt to represent the main dynamics in the
catchment but require calibration of several parameters [63]. Conceptual models are
therefore less demanding compared to physically-based models but need more
information than empirical data-driven models. Topography-based hydrological
(TOPMODEL) that were used in Nigeria is an example of a simple yet powerful
conceptual model that has been used in numerous applications covering a range of
21
catchment sizes and geographical areas [53].
These models do not use catchment characteristics or other physical parameters [64].
One major advantage of this approach is that these models are simple to use
compared to deterministic or conceptual models. However, the disadvantage is that
these model are did not accept any changes such as changes in land use into account
compared to the deterministic model.
ANN is classified as an empirical model where the data is fed into the model, a
relationship between the inputs and the outputs is then used to produce a forecast.
Therefore, the ANN-based model does not require any understanding of the
physical processes. However, some hydrological knowledge is a prerequisite as it
guides the selection of inputs (i.e. rainfall, previous flows, water levels, and so forth)
and outputs to produce a forecast (e.g. runoff, stream flows, hydraulic
conductivities, etc.). This type of modelling does not always appeal to hydrologists
who prefer the core of the model to be a dynamic, physically-based and a
representation of the processes involved. Research reported has used sensitivity
analysis to determine the most significant inputs to be used in the network [34]. This
can be applied to gain a better understanding of what is happening in the model.
Unit hydrographs are a classic empirical model which capture the relationship
between rainfall and catchment response. One example is the time series models
such as ARMA models, which are frequently used as a type of empirical model to
capture the rainfall-runoff relationship [65].
2.6 Summary
A description of the study area has been given, including the different method of rain
measurement systems and existing flood prediction available in Malaysia. An
introduction to hydrological modelling approaches has been given. Several basic
hydrology models for the estimates discharge of water base also been discussed
where most of the model needs to have some hydrological parameters is a
22
prerequisite as it guides the selection of inputs. The ANN approaches been selected
because it does not require any understanding of the physical processes where data
is fed into the model, a relationship between the inputs (rainfall, previous condition
of dry time and flow rate) and the outputs is then used to produce a forecast river
flow rate.
2.7 Major Outcomes
There are several methods of rain measurement systems and flood prediction based
on the ground and radar-based rain measurement. However, because of the
complexity of measuring rainfall rate in tropical areas due to limited spatial and
temporal resolutions, it will leads to unreliable prediction. Hydrology models
estimate discharge of water base on the calculation but physical data is need before
the model can be used. Therefore, ANN approach been selected because it does not
require any understanding of these physical processes. ANN models are simpler to
use compared to deterministic or conceptual models.
23
Chapter 3 Artificial Neural Networks (ANN)
This chapter describes the application of ANN in hydrology for river flow modelling
as a basis to develop a river flow prediction system. An overview of ANN especially
its structure, model development, and advantages and disadvantages are given. This
chapter concludes with a review of the major themes that have emerged from the
ANN river flow forecasting literature and how the research in this thesis fits within
this context.
IntroductionANN describe a system modelled on a human brain with neurons and
interconnections that enable input to be mapped to the output. The development of
the back-propagation algorithm was an important advancement in neural networks.
Neural Network Autoregressive with Exogenous Input (NNARX) is a version of
ANN that is more capable of for nonlinear systems [66]. Its generalisation capability,
effective learning rate and faster convergence make it preferable compared to other
neural networks.
Fig. 3.1 Layout of feed-forward neural network
There is a need to extract knowledge from data and to understand the connections in
the ANN network architecture. There is also a need to reduce complexity to ensure
faster convergence [61][67]. Fig. 3.1 shows an example of an ANN network
architecture. The network can have many hidden layers, hence become more
complex.
Neuron
Neuron
I1
Neuron
I2
Input Layeri
Hidden Layerj
Output Layerk
W11
W12
W11
W12W13
W21W21
W22
Neuron
O1
W22
W31
W23
W31
O2
NeuronI3 W32
W32
W33
24
Nonlinear AutoRegressive Neural Network with eXogenousInput (NARX)
NARX is a nonlinear model derived from Autoregressive with Exogenous Input
(ARX) model [68][69][70]. It is a recurrent dynamic neural network with feedback
connection enclosing several layers of a neural network. NARX is well suited to
modelling nonlinear systems, especially in time series analysis. One major application
of NARX dynamic neural networks is in control systems [71]. Some important
qualities about NARX networks with gradient-descending learning gradient
algorithm have been reported to be much faster and generalise better than other
networks [70]. The dynamic behaviour of the NARX model is described by Equation
(3.1) where F is a nonlinear function of its arguments such as a polynomial. F can be
a neural network, a wavelet network, a sigmoid network, etc.= ( , , , … … , , , , , … . . ) + Є (3.1)
Here, is the variable of the flow rate of the river and is the externally determined
variable in this study which is rainfall in the related station. In this scheme,
information about will help to predict y, as do previous values of itself. Є is
the error term and the current value of model output is denoted by y (t).
Types of ANNA multilayer perceptron (MLP) network is the most commonly used type of ANN.
Other types of ANN include Radial Basis Function Networks (RBFN) [72], Recurrent
Neural Networks (RNN) [73] and Self-Organizing Map (SOM) [74]. RBFN have
architectures that are similar to MLPs. The RBFN differs in the form of the
activation function in the hidden layer nodes, which are of a radial basis or Gaussian
form. The parameters of the radial basis functions are usually determined first
followed by the weights during the training process. RBNF requires a large training
dataset to achieve high accuracy.
RNN uses a modification of feedforward network to allow for feedbacks between
layers. Different learning algorithms have been developed for this type of network.
It has been used successfully in handwriting recognition applications but less in the
25
hydrological modelling.
Self- Organizing Map (SOM) is a slightly different type of ANN [75]. SOM uses
an unsupervised classification method to cluster data into similar types. Therefore,
only an input data vector is required. It is used for data compression and
visualisation of relationships in the data [76]. In its simplest form, SOM is composed
of a single two-dimensional layer of input neurons of vector w. The training
algorithm randomly selects an input vector and finds the best match based on the
Euclidean distance between the two vectors. This iterates until a stopping condition
is satisfied, e.g. a minimum error is reached. The disadvantage of SOM is that
there is no clear methodology for the selection of the number of nodes (or clusters)
or the values of the learning parameters, e.g. the size of the neighbourhood.
Development of an ANNThere are not many guidelines in the literature on how to develop an optimal ANN
for a given application. Some authors have attempted to provide some guidelines
based on empirical results model development [77]. Decisions to make when
implementing include selecting the network architecture, the number of hidden
layers, the number of hidden nodes, input variables, activation functions, training
algorithm, and which performance measures to use. These decisions are described
in section 3.8
The architecture of an ANN is often selected after many trials as there is no
established methodology to determine the configuration [78][79]. Once the type of
network is chosen, the number of hidden layers and the number of nodes in each
hidden layer must be determined. However, there are no fixed rules, so it is often by
trial and error [80][81]. There are, however, some heuristics in the literature. For
example, Hecht-Nielsen [82] suggested the following upper limit for the number
of hidden layers ensure that the network can approximate any continuous function:≤ 2 + 1 (3.2)
where is t h e number of hidden layers and is the number of inputs. Rogers
and Dowla [83] suggested a second relationship to avoid overfitting as follows:
26
= ≤ (3.3)
where is the number of observations in the training dataset.
Depending on which training algorithm is chosen, the number of parameters needs
to be specified, e.g. the learning rate and momentum.
Another important decision to make is the choice of input variables. One study used
river levels or flows and rainfall [84]. Total rainfall is normally used but effective
rainfall has been used [85]. Effective rainfall is total precipitation minus losses,
which produces the runoff. However, the problem with effective rainfall is that it
is hard to estimate because it depends on the antecedent moisture conditions of the
basin, which changes over time.
Other variables that have been used include temperature, evapotranspiration,
moving average antecedent precipitation, and soil moisture [86]. If too many inputs
that are not independent are used, it increases the size of the network and the
training. There is also a greater likelihood of overfitting the training data because of
the ratio of connection weights to training data increases. Although many different
methods exist, trial and error and correlation analysis are used most frequently to
select inputs [67][87][88].
ANN ArchitectureANN has two structure; feed-forward network and a feedback network. The feed-
forward neural network has no provision for output to be used as an input to a
processing unit in the same layer. A feedback network allows outputs to be used as
an input in the same or preceding hidden layer. An ANN system learns by changing
the weighting of inputs. During training, the network sees the real results and
compares them to the outputs. If the difference is significant, the network then uses
the feedback to adjust the weights of the inputs.
An ANN is comprised of nodes or neurons. Information is passed between nodes
through connections. Weights are then associated with each connection, which
represents the magnitude or strength of that connection. Within the node is a
27
nonlinear transformation function, an activation function, which is applied to the
input signals of the node to produce an output signal. The simplest function that
is usually used is the sigmoid function. The nodes or neurons are arranged into a
series of layers: an input layer; one or more hidden layers; and one output layer as
in Fig. 3.2. A weight matrix W, a bias vector b, and an activation or transfer function
f are associated with each hidden layer.
Fig. 3.2 Two-layer network with abbreviated notation.
The input layer is where external information is received and provided to the
network (e.g. antecedent rainfall or runoff) whilst the output layer produces the
forecasted parameter (e.g. the river flow rate). The representation of nodes in each
layer and the interconnections are more clearly shown in Fig. 3.2. This is an example
of a feedforward network.
The output of each node is obtained by computing the value of the activation
function on the product of the input vector and the weight vector, minus the value of
the bias associated with that node. It is possible to express the forward processing
through the network as a single equation. A network with one hidden layer and
K outputs would have a functional form as in Equation (3.4).( , ) = + ∑ + ∑ (3.4)
f1 f2
Output layerInput layer Hidden layer 1 Hidden layer 2
IW1,1a1
LW2,1
+ + a2
b1 b2
a1 = f1 (IW1,1∙X + b1) a2 = f2 (IW2,1∙a1+b2)
XY
28
Where is the number of inputs, is the number of nodes in the hidden layer, g is
the activation function of the hidden layer nodes, and and are the weights. The
indices and correspond to the output node and hidden layer nodes, respectively.
However, ANN are rarely expressed in this manner as the equation is not
interpretable.
Training an ANN modelAfter the network structure been configured, training is used to find the values of
the weights that minimise the error between the inputs and the outputs in the
training dataset: = ∑ ( − ) (3.5)
where is the sum of the errors squared between the targets, t, and the ANN
estimate, a, for i observations in the input-output dataset.
Fig. 3.3 Local and global minima of errors.
The training iteratively adjusts the weights of each node until a stopping condition
is reached. The initial weights are first randomly selected. In a single training run,
the algorithm may fall into a local minimum on the error surface as illustrated in Fig.
3.3. The most common objective function is to minimise the sum of the errors squared.
This assumes that the model errors are distributed around zero mean with unknown
variance.
29
To achieve an acceptable level of generalisation, the dataset is usually divided into
three subsets which consist of training, validation and test datasets. The test dataset
is reserved and used in the testing phase to measure the performance of the network.
This is not used during the training process.
There are two types of training algorithms, supervised and unsupervised. For a
supervised training algorithm, the output is available, and the network will find the
best set of weights by minimising the error between the output and the target. For the
unsupervised network, only the input data set will be given to the network, which
will then try to find clusters of similar inputs without any previous knowledge.
Although many training algorithms have been investigated in this study, two
algorithms are selected and described. These are the Levenberg-Marquardt (LM) and
Bayesian Regularization (BR) algorithms.
3.4.2.1 Levenberg-Marquardt (LM) Algorithm
The Levenberg-Marquardt (LM) algorithm improves the training speed while
avoiding the Hessian matrix [89]. If the objective function of the ANN is the sum of
squares, then the Hessian matrix can be written as equation (3.6)= + (3.6)= (3.7)
The gradient can be determined as:= (3.8)
where is the Jacobian matrix that contains the first derivatives of the errors of the
ANN with respect to the weights and biases of the network, while e is a error vector.
The LM algorithm approximates the Hessian matrix as follows:= − [ + ] (3.9)
When > 0, equation ( 3.9) becomes a gradient descent with a decreased step
size. Thus, is decreased after each iteration.
30
3.4.2.2 Bayesian Regularization (BR) Algorithm
Bayesian Regularization (BR) is an ANN algorithm that updates the weights and
bias values using the LM algorithm as a basis. It minimises the squared errors and
weights and then determines a suitable combination of the two to produce a network
with the greater capability of generalisation. [90].
Network design configurationThe process of designing the ANN Network is to select the configuration. This
includes the number of layers, nodes, etc. Kaastra and Boyd [21] describe a method
for designing a back-propagation ANN as in Fig. 3.4. All of these steps will be
discussed in the next section.
Fig. 3.4 Eight steps in designing neural networks forecasting model [21]
Variable selectionIn the case of river flow prediction, defining input data is critical since it affects the
generalisation ability of the ANN. For flow rate prediction, rain data from weather
stations used as input data. Previous rainfall and river flow rates are sometimes used,
as inputs and as targets. Moving average and exponential moving average are used
31
to enhance the performance of ANN. A simple moving average (SMA) is expressed
as in equation (3.12) = ⋯ ( ) (3.12)
where is parameter such as the flow rate, is the duration or data resolution, and
is the total time period. The exponential moving average is expressed as in equation
(3.13) = ( ) ( ) ⋯ ( ) ( ) ( )( ) ( ) ⋯ ( ) ( ) (3.13)
Correlation analysis [67] provides a way of reducing the number of inputs, hence a
number of trials by eliminating those input variables that are independent or
unrelated to the output [91] [92]. The problems with this method are that: (i) it
assumes a linear relationship between the variables, whereas most hydrological
problems are non-linear; (ii) it is unclear which correlation value threshold to use and
(iii) it does not take variable independence into account.
Data pre-processingSince neural networks pattern matches, the representation of the data is critical in
designing a successful network. Raw inputs and outputs are rarely fed to the ANN
directly [93][94]. The data was pre-processed to reduce noise, detect patterns, and
normalise data distribution to facilitate ANN network training convergence and
generalisation.
The normalised data is mapped into a predefined range [0, 1]. The min-max method
normalises the values of the attributes X of a data set according to its minimum and
maximum values. Equation (3.14) is used.= (3.14)
This also matches the range of data with the range of transfer functions [93][94]. The
input and output variables are standardised before the network is trained. Unlike
statistical models, it does not need to be normally distributed.
32
Training, testing, and validationThe training data set is the largest dataset compared to testing and validation dataset.
The testing set is used to test the generalisation ability of the trained network and the
validation set is used to check the performance of the network. Kaastra and Boyd
recommended a 10% to 30% of the data as testing dataset [21], and Zhuang et al. [14]
recommended 90% to 70% of the data to be used as training dataset. Therefore, in this
study, 70-15-15 (training-testing-validation) configuration is considered as an
acceptable division and has been used in this research [95][96].
Neural network paradigmsThe number of hidden layers and neurons in the hidden layers plays a vital role in
the performance of Back Propagation Neural Network especially in cases where a
problem related to an arbitrary decision boundary with rational activation functions.
Also, multiple hidden layers can approximate any smooth mapping to any accuracy.
However, the process of deciding the number of hidden layers and number of
neurons in each layer is still not defined [97]. Furthermore, many previous studies
showed that it is better to increase the number of hidden neurons rather than to
increase the number of layers of the network [97][98]. In this studies, following
configuration for the experiment has been used:
The number of input neurons is the same as the number of inputs. Therefore,
one input neuron is assigned to the previous flow rate and another is assigned
for the predict the future flow rate [97] [98].
The number of hidden layers depends on a number of inputs and the nature of
data. It is also acceptable that one layer with a sufficient number of neurons is
enough for good approximations because more layers increase the possibilities
of over-fitting[99] [17].
There is only one output neuron, which is the forecasted river flow rate.
The most commonly used transfer functions are linear and sigmoid functions
33
because of their continuity feature. The hyperbolic tangent sigmoid (Tan-sig)
and pure linear functions are both used in this research. The Tan-sig function
accepts inputs in any range and non-linearly transforms them into a value
between -1 and 1. The function can be expressed as:
( ) = ℎ( , ) = − 1 (3.15)
where u is the input range and α is a variable that controls the shape of the function.
The larger the value of α, the closer the function gets to the step function; and the
smaller the value, the closer it gets to a purely linear function. The variable α is varied
in the experiment to test the effect of different shapes of the Tan-sig function on ANN
performance.
Evaluation criteriaNeural network performance is evaluated using specified performance indicators.
For error measurement, the mean square error (MSE) function was used during
training. Regression technique was also used [100][101] to measure the neural
network performance and prediction performance.
Neural network trainingTraining of the ANN was iterative. The main objective of training is to reach the
global minimum of the error function and generalised the model. In this study,
backpropagation algorithm was used. It was developed to improve training for
different types of problems with the aim to reduce the training time and achieved fast
converging.
The main goal of backpropagation is to optimise the weights which enable the neural
network to learn how to map arbitrary inputs to the outputs correctly. In this studies,
the Bayesian regularisation backpropagation and Levenberg-Marquardt
backpropagation have been selected from a wide range of available methods, after
evaluation [68][69][102][103]. This two algorithms (Bayesian regularisation and
Levenberg-Marquardt) were selected after a study comparing different algorithms
showed that the Levenberg-Marquardt algorithm (trainlm in Matlab) is suitable for
34
small to medium-sized networks [104][105][106].
Another aspect of training is the criteria to decide when to stop training process. Since
the network is trained in iterations, also called epochs, the number of iterations must
be specified to guide the training process. The number of iterations can be determined
before training starts, but this method has two disadvantages. If the network
converges very quickly, overtraining will occur, but if the opposite happens then,
training will stop before the network able to converge. On the other hand, the
convergence method, which was used in this research, solves this problem by
stopping when the performance function stops improving, i.e. the network
converges. Fig. 3.5 illustrates an example of error performance and the criteria to stop
the training process. During training, the maximum epoch is fixed to a high value
that will not likely be reached (e.g. 1000) to allow sufficient iteration to achieve
convergence but at the same time to ensure that the training does not carry on forever.
Since the LM algorithm usually has a faster convergence, training time is typically
shorter compared to Bayesian Regularization. However, recurrent NARX networks
with Bayesian Regularisation tend to produce better results compared to the LM
algorithm.
Fig. 3.5 Training stopping due to convergence.
ANN application in prediction modelsIn the early stage, Smith and Eli [107] used a b a s i c three- layer ANN (Input, hidden,
output) to predict runoff based on simulated rainfall patterns in a synthetic
35
catchment. Their results were not very accurate in predicting the river flow rate nor
the time because they could not successfully model the physical characteristic of their
catchment and rainfall because the ANN does not model the individual physical
processes.
Later, Dawson and Wilby [77] also used ANN to forecast flow rates of the Amber
and Mole Rivers in the UK for lead time up to 6 hours. This was found to be
comparable in performance to an existing flood forecasting system in operation.
Different networks have also developed, e.g. a Radial Basis Function Networks
(RBFN) by Mason et al. [108], using a temporal backpropagation network (TBP-NN)
by Sajikumar and Thandaveswara [109]. The issue of lead times was also
considered, e.g. Campolo et al. [110] examined the effect of ANN predictions as
lead time is increased from 1 hour to 5 hours. They noted the decrease in accuracy
with increasing forecasting time.
In 2000, reviewed by the American Society of Civil Engineers Task Force, Maier and
Dandy [67] have targeted the broader area of hydrology. They highlighted the main
challenges in ANN model development such as selection of input variables, an
optimal division of the input data, data pre-processing methods, etc. The study
showed that the application of ANN in hydrology, including rainfall-runoff
modelling, concluded that ANN could perform as well as existing hydrological
models. However, there was a lack of an established methodology for model design
and implementation.
ImplementationTable 3.1 gives a summary of network configuration used in this research. These
network configurations were selected based on detailed research and study of the
problem to be solved, including analysis of the available data.
36
Input variables Rainfall, Flow Rate, Dry time
Output variable River Flow Rate
Training, testing, validation 70-15-15
ANN type RNN (NARX)
Number of input neurons 3
Number of hidden layers 1
Number of neurons in hidden layer 1
Number of output neurons 1
Hidden layer transfers function Hyperbolic tangent sigmoid shape
Output layer transfers function Hyperbolic tangent sigmoid and Purelinear
Training algorithm Bayesian Regularization andLevenberg-Marquardt
backpropagationTraining algorithm learning rate (Mu) 0.001 to 1Training algorithm learning rate updates (Mu stepup and Mu step down)
10 and 0.1
Error (performance) function MSE, RMSE
Prediction (performance) Predicted- Actual Flow Regression
Max epochs 1000
Max training time Undetermined
Table 3.1 Summary of the settings to be used in this experiment and theirvalues.
Fig. 3.6 A schematic of the ANN that been applied in the prediction model.
In this research, an ANN model of the River Pelarit and Jarum were developed to
predict the river flow rate using recordings of 5 minutes interval river flow rate from
related stream gauging stations and rainfall of relevant station. The network
Model Input
Neuron
Neuron
Neuron
Rainfall
Neuron
River flowof n-time
Dry Period
Input Layeri
Hidden Layerj
Output Layerk
W11
W12
W13
W11
W12
W21
W22
W31
W32
W21
W22
W23
W31
W32
W33
PredictedFlow rate
37
consisted of 3 lagged inputs (Rainfall, River flow and Dry Period) from each station,
one hidden layer, i.e. the output of predicted flow rate as in Fig. 3.6.
Model EvaluationThe performance of hydrological models is commonly assessed by measures of
performance or goodness-of-fit statistics. This has also been applied to ANN rainfall-
runoff models. It is suggested that the adopted criteria should not be redundant and
should be sensitive to different types of errors (e.g. errors in peak prediction, errors
in the timing of the hydrograph prediction, etc.) [77][111][112]. However, often only
indices, which are linked to each other such as the correlation root mean squared
error are applied.
Evaluation of prediction performance can be assessed by computing an error
measure, E over a time series:= ∑ ( ̈( ) − ( )) (3.16)E the error between the estimated or predicted element an actual sequence element
[70].
After the testing phase is completed, the results are saved in the workspace, and a
graph is plotted between the actual flow rate and the predicted flow rate over a given
time period. The graph is an effective visual way of comparing the measured and
predicted the result. Performances measurement such as Root Mean Square Error
(RSME), Mean Absolute Percentage Error (MAPE) and Regression (R2) have been
used. RMSE is been shown in Equation (3.17) [113].RSME = ∑ ( ) (3.17)Where is the observed value for the ℎ observation and is the predicted value.
MAE gives the same weight to all errors, while the RMSE penalises variance as it
gives errors with larger absolute values more weight than errors with smaller
absolute values [114]. When both metrics are calculated, the RMSE is by definition
never lower than the MAE.
38
Correlation CoefficientThe correlation coefficient is one of the ways to determine the degree to which two
variables' movements are associated as shown in equation (3.18). The range of values
for the correlation coefficient must be from -1.0 to 1.0. A correlation of -1.0 indicates
a perfect negative correlation, while a correlation of 1.0 indicates a perfect positive
correlation [115].
= (∑ . ) (∑ ).(∑ )∑ (∑ ) . ∑ (∑ ) (3.18)Where is observed data; is predicted data; and is the number of data.
Advantages and Disadvantages of ANN for HydrologicalModelling
One of the key advantages of using ANN in hydrology modelling is that when
developed, the model can be running in near-real-time and continuously use
available data to adapt and improve its prediction. This is very important in an area
where land use changes at a faster pace than normal. Continues and adaptive
techniques such as ANN become more important as they are able to evolve their
prediction with changing conditions. Other advantages include:
Ability to adapt to non-linear problems such as hydrological problems. The
interconnection between the neurons in an ANN generates non-linear data
processing structures that are distributed across the network. This feature
allows intrinsically non-linear processes to be modelled, such as the
transformation of rainfall into runoff.
Ability to model input /output relationships; ANN does not need an explicit
mathematical equation to specify the connection between the inputs and
the outputs.
Adaptability; ANN is adaptable to changes as new data become available. This
feature makes it particularly useful in the treatment of non-stationary
39
processes, where learning strategies can be designed in real-time so that the
model learns continuously [116]. This is important as rainfall patterns changes
due to long-term weather change and land use.
Less sensitive to noise in the data; ANN proven has a greater ability to handle
noise in the input data [26] especially when the amount of data available is
large.
Modularity: ANN can be integrated easily into modular architectures to
solve specific subtasks of the overall problem efficiently [117]. This would
allow the method to be adapted to different sections of the river course.
However, the main disadvantage of ANN is its black box approach to the process,
which makes it less preferable to physically-based and conceptual models,
especially for real-time applications. T rying to open up the black box and find
physical meaning in ANN is an area of research that needs more attention. Some
preliminary studies have begun to address this issue, but it continues to be an
area where further research is needed [57][118]. This particular aspect is its attraction
for catchment modelling as the complexities of the physical process and their changes
do not to be explicitly understood.
SummaryThis chapter has provided an overview of the design of an ANN with particular
emphasis on rainfall-runoff modelling. The advantages of ANN in hydrological
modelling, as well as the disadvantages, have been discussed. The following chapter
will present the methodology that been applied in this research.
40
Major OutcomeThe advantages of using an ANN-based prediction method include the following:
Prior knowledge of the underlying process is not required;
Existing complex relationships among the various aspects of the process under
investigation need not be recognised;
Solution conditions, such as those required by standard optimisation or
statistical models are not pre-set;
Constraints and prior solution structures are neither assumed nor enforced
[119].
41
Chapter 4 Methodology
This chapter describes a detailed study of the catchment area and how rainfall affects
the river characteristics. Statistical analysis of rain in the study area is also presented
to gain a detailed understanding of the local rain dynamics.
Fig. 4.1 shows the development of river flow model for Timah Tasoh which consists
of several stages. Firstly, the rainfall characteristic will be studied using rainfall data
from 14 weather stations in Perlis. Secondly, the relationship between rainfall and
river characteristic (flow rate and level) for River Jarum and River Pelarit, based on
the studies of data from 3 rainfall stations and two river flow stations, will be
determined. Finally, to perform flow modelling, ANN modelling will be introduced
and described how it could be used to predict the river important characteristics.
Fig. 4.1 Development of River Flow Modeling for Timah Tasoh Basin.
42
Study area
Malaysia has two predominant seasons; dry and rainy seasons. The north-west region
is the driest part of the Peninsula during the northeast monsoon period where Perlis
is located [120]. The northeast monsoon has fewer influences in the north-western
region because of the Titiwangsa mountain range, which shelters the region [2]. On
the other hand, the northeast monsoon strongly affects the eastern part of the
Peninsula [120][121]. Fig. 4.1 shows that the annual rains in Peninsular Malaysia
range from 1500 mm to above 4000 mm per year [5].
Fig. 4.2 Annual rain distribution for Peninsular Malaysia [5].
Perlis is the smallest state in the northern part of Peninsular Malaysia which has been
chosen as the case study area for this research. Perlis is located at 6°30′ N 100°15′ E
and has an area of 821 km2 (317 miles2). In 2005, 2010 and 2011, a series of floods
occurred in Perlis resulting in significant damages to properties and some loss of lives
at one time [122]. In 2010 alone, 13,711 people were evacuated from the flooded areas
while the floods claimed four lives [122]. According to the Ministry of Agriculture
and Agro-based Industry of Malaysia, losses to farmers were estimated to be up to
RM 50 million (£10 million) [123] in 2010 alone.
43
Fig. 4.3 Location of the hydrological station in Perlis, Malaysia.
Fig. 4.3 shows the locations of the gauging stations from which data has been
collected and used in this study. They consist two river gauging stations and three of
weather stations.
The main water reservoir in Perlis is known as Timah Tasoh which receives water
supply from two main rivers, the River Tasoh and River Pelarit with a combined basin
area of 191 square km. Because of its small capacity, the water level in the reservoir
often rises quickly after heavy rains. Excess water from the reservoir, when released,
causes flooding downstream of the dam to the main populated area, including the
state capital, Kangar. Any delay in releasing the water also causes a flood to the
surrounding area at the upstream of Timah Tasoh dam [124]. Therefore, monitoring
and predicting the amount of run-off and the water level in the lake or dam are crucial
for the authority to effectively manage the water in the dam. This dam also has dual
functions; to ensure that sufficient amount of water is retained for domestic and
irrigation use whilst making sure that the water level does not rise above a critical
level that will cause floods. During heavy rains or monsoon season, excess water in
1. PADANG BESAR
2. LUBOK SIREH
3. KAKI BUKIT
4. JARUM RIVER
5. PELARIT RIVER
44
the dam has to be released to ensure the safety of the dam and to reduce the risk of
flooding upstream. However, during the dry season, the water level in the dam falls
rapidly due to its small capacity. This can lead to water rationing for domestic water
supply and irrigation.
Fig. 4.4 Land usage in Timah Tasoh Reservoir catchment area.
The land use in the Timah Tasoh catchment area is shown in Fig. 4.4 and Table 4.1
provides details of land usage in the basin area. Most of the land is covered by
vegetation, forest and farmland.
Land use River Jarum River Pelarit Hulu River Tasoh River Pelarit Hilir
Sugar 11.58 - 11.58 -
Settlement 0.74 0.35 2.68 0.35
Mixed crop 2.22 0.21 3.37 0.76
Scrubland 2.79 0.33 3.2 0.38
Rubber 12.94 2.09 25.87 5.13
Rice Field 5.23 0.4 7.62 0.59
Forest 28.9 38.72 46.85 52.11
Grass - 0.13 0.08 0.28
Quarry - 0.5 - 0.5
TOTAL 64.4 42.72 101.25 60.1
Detail of land use ((km2) in Timah Tasoh catchment area [125].
45
The distance between the River Pelarit monitoring station at Batu Limabelas and the
Timah Tasoh reservoir is about 9.87 km as shown in 0. The related rain stations for
River Pelarit are Kaki Bukit and Lubuk Sireh stations.
Fig. 4.5 The distance of River Pelarit from the gauging station in Kg BatuLimabelas.
Fig. 4.6 River Jarum- gauging station at Kg Masjid to Timah Tasoh Lake.
46
The distance of the River Jarum monitoring station to the Timah Tasoh dam is about
7.40 km. The aerial view is shown in Fig. 4.6. The related rain gauge station for River
Jarum is Padang Besar gauging station, located in the northern part of Perlis.
Mann-Kendall (MK) trend test
Mann-Kendall (MK) test is a non-parametric way to detect a trend in a series of values
to study the trend of rainfall in Perlis. This test is suitable for cases where the trend
may be assumed to be monotonic and thus no seasonal or other cycle is present in the
data. The MK test is applicable in cases when the data values xi of a time series can
be assumed to obey the model as equation (4.1) where f(t) is a continuous monotonic
increasing or decreasing function of time and the residuals i can be assumed to
be from the same distribution with zero mean. It is therefore assumed that the
variance of the distribution is constant in time. To test the null hypothesis of no
trend, Ho, i.e. the observations xi are randomly ordered in time, against the
alternative hypothesis, H1, where there is an increasing or decreasing monotonic
trend. xi f (ti) i (4.1)MK test statistic is defined as T = i<j sgn (xj −xj), where sgn (x) =1 if x > 0 0 if x =0 −1 if
x < 0. Under the null hypothesis that there is no trend, T is distributed as a normal
random variable with mean zero and variance assuming no ties between (x1, x2,.,xn).
The MK test statistic S is calculated using the formula (4.2) and (4.3):
= ( − ) (4.2)Where xj and xk are the rainfall values is in specific duration j and k, j > k respectively
and Sgn(xj - xk) = +1 , if xj - xk > 0 (4.3)0, if xj - xk = 0-1, if xj - xk < 0If n is 9 or less, the absolute value of S is compared directly to the theoretical
distribution of S derived by MK. In this case, the two-tailed test is used for four
different significance levels : 0.1, 0.05, 0.01 and 0.001. At certain probability level
47
H0 is rejected in favour of H1 if the absolute value of S equals or exceeds a specified
value S/2, where S/2 is the smallest S which has the probability less than /2 to
appear in case of no trend. If the value of S is positive indicates an upward trend and
if the value of S is negative shows a downward trend. The minimum values of n
with which these four significance levels can be reached are derived from the
probability table for S as follows in Table 4.2.
Significance level, required n
0.1 4
0.05 5
0.01 6
0.001 7
Four significance level and the required number of data n.
The significance level 0.001 means that there is a 0.1% probability that the values
xi are from a random distribution and with that probability, a mistake been made
when rejecting H0 of no trend. Thus the significance level 0.001 means that the
existence of a monotonic trend is very probable. Respectively the significance level
0.1 means that there is a 10% probability that a mistake been made when rejecting
H0. Total of data, n is at least 10 the normal approximation test is used. However,
for the case where a number of data values 10 or more or there are several tied values
(i.e. equal values) in the time series, it may reduce the validity of the normal
approximation when the number of data values is close to 10.
Rain Dynamics
Annual rain data from 4 weather stations located in the catchment area has been
analysed to study the rain characteristic. The data was measured at every 5 minutes
intervals. The list of the stations and the distances between the stations are shown in
Table 4.3 which shows the distance (in meters) between the stations. The closest two
rain gauge stations are Lubuk Sireh and Kaki Bukit which are 1416 meter apart. For
the purpose of this study, these two stations have been used to determine the rain cell
sizes and dynamics. The selection of these stations is also important because Lubuk
48
Sireh and Kaki Bukit are located in the middle of the Timah Tasoh catchment area
and Tasoh is located at the edge of the dam as shown in Fig. 4.8 . Rainfall from Wang
Kelian does not flow into the catchment but its position is important for studying the
dynamics and direction of rain cells within the area.
Rain gauge stations in Timah Tasoh catchment.
Fig. 4.7 Monthly distribution in 2013 for rain stations in Perlis.
Fig. 4.1 compares the monthly rainfall amounts measured by the 4 stations in 2013.
The results show that the wettest month was October with up to 480 mm of rain in
Wang Kelian station. Amongst the 4 stations, Wang Kelian experiences the most
amount of rain followed by Kaki Bukit, Lubok Sireh and then Tasoh. These show that
topology significantly influences the amount of rain that falls in this region. Rain from
the sea is blocked by the hill that separates Wang Kelian from the rest of the rain
gauge stations.
0
100
200
300
400
500
600
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
WANG KELIAN
LUBOK SIREH
KAKI BUKIT
TASOH
WA
NG
KE
LIA
N
LU
BO
KSIR
EH
KA
KI
BU
KIT
WANG KELIAN - 4958LUBOK SIREH 4749 -KAKI BUKIT 4094 1416 -
TASOH 7811 3643 3724
49
Fig. 4.8 Location of four rain stations in Timah Tasoh catchment.
The study of rain dynamics is achieved using the rain gauge grid as described in Fig.
4.9. The box represents the coverage area which is divided into four smaller sections
and each rain gauge is considered to be located in the centre of each segment. Based
on the rain cell size, there are three possible scenarios: (a) the rain cell size is smaller
than or equal to 1.5 km (distance between the 2 closest rain gauge stations) and is
moving across the area; (b) the rain cell is larger than 1.5 km and is moving across the
area and; (c) the rain cell is small and localised. In the first scenario (Fig. 4.9 a), initially
at a time a small sized rain cell travels from outside into the catchment area at time
. In each case, only one rain gauge station is experiencing rain. When the rain cell
enters the catchment area, rain gauge, RG1 will record the rainfall. The second rain
gauge RG2 will record rainfall at time t2. Finally, at time t3, none of the rain gauges in
the catchment area will record the rainfall from the rain cell.
In the second scenario, Fig. 4.9 b), the rain cell size is large covering more than one
weather station at the same time. This is widespread rain which is a characteristic of
rain in temperate climates and is easier to detect and predict. In the third scenario,
Fig. 4.9 c), the rain cell is small and static, where it only rains and stop in the same
location. Fig. 4.9 a) and Fig. 4.9 b) illustrates cases where the rain dynamics is likely
Rainguage
RiverLevelsensor
Wang Kelian
Lubuk Sireh
Kaki BukitTasoh
50
being influenced by the wind. Fig. 4.10 show the direction of rain cell movement in
the catchment area based on the study of rainfall data.
a) b) c)Fig. 4.9 Different scenarios of rain gauge grid for rain cell detection: (a) small
sized rain cell < 1.5 km, (b) large rain cell > 1.5 km, (c) localized cell < 1.5 km.
Fig. 4.10 The direction of rain cell movement within the catchment area.
Fig. 4.10 shows an example the dynamics of the rain cell in the Timah Tasoh
catchment area. It is based on Table 4.4 which indicates that the movement of the rain
Rain station
River Level station
Time RainGauge
RG1RG2RG3RG4
Initial, t0 0 0 0 0
t1 X 0 0 0t2 0 X 0 0t3 0 0 0 0
1. t01 2
3 4
1 2
3 4
1 2
3 4
1 2
3 4
2. t1
3. t2
4. t3
Time Rain GaugeRG1RG2RG3RG4
Initial, t0 0 0 0 0
t1 X 0 X 0t2 0 X 0 Xt3 0 0 0 0
1. t0
1 2
3 4
2. t1
3. t2
4. t3
1 2
3 4
1 2
3 4
1 2
3 4
Time Rain GaugeRG1RG2RG3RG4
Initial, t0 0 0 0 0
t1 X 0 X 0t2 X 0 0 0t3 0 0 0 0
1. t0
1 2
3 4
2. t1
3. t2
4. t3
1 2
3 4
1 2
3 4
1 2
3 4
Wind Direction
Wang Kelian
Lubuk Sireh
Kaki Bukit
Tasoh
51
cell in the North West direction deduced from the analyses of the sequence of the rain
gauges measurements. For this case study, initially, on 30 January 2013 at 11:15 am,
the rain cell travelled from outside the catchment area and at time 11:25 am both Kaki
Bukit and Lubuk Sireh rain gauge stations were experiencing rain. This suggests that
the rain cell that entered the catchment area had a size that was larger than 1.5 km in
diameter based on the distance between the two rain gauge stations. However, the
amount of rain that was recorded is not same. This rain cell has been recorded for 15
minutes at the Lubok Sireh station before it disappeared at 11:35 a.m. Meanwhile, this
rain cell moved in the direction of Wang Kelian station where at 11:35 a.m., the station
started recording the rain. Due to the movement of the rain cell, at time 12:05 p.m. the
rain cell had passed.
SITE WANG KELIAN KAKI BUKIT LUBOK SIREH TASOH
Date Time Amount of rain [mm]
Rain (mm)
Rain (mm)
Rain mm
20130103 11:15 0 0 0 0
20130103 11:20 0 0.1 0.5 0
20130103 11:25 0 0.5 1.5 0
20130103 11:30 0 0.7 1.6 0
20130103 11:35 0.6 1.1 0 0
20130103 11:40 2.6 0.8 0 0
20130103 11:45 0.9 0.2 0 0
20130103 11:50 0.2 0.2 0 0
20130103 11:55 0.2 0 0 0
20130103 12:00 0.1 0 0 0
20130103 12:05 0 0 0 0Rain records for 4 stations in Timah Tasoh Catchment.
Catchment Area Characteristic
In the conventional hydrological method, hydrological processes in the floodplain are
used to estimate the river characteristics e.g. flow rate and level [126][127]. The
catchment's physical characteristics (soils, geology, and the presence different of
types of vegetation) must be known and some assumptions about their spatial
homogeneity made. This is not always possible and therefore an alternative approach
52
has to be sought.
In this study, the alternative method chosen utilises rainfall and river data. The data
covers a 4-year period from 1 Jan 2010 to 31 Dec 2013. The aim of the analyses is:
To study the relationship between rainfall and river characteristics such as
river flow rate.
To identify the dependency on the flow rate of the river and rainfall in terms
of rainfall duration and intensity.
To analyse how the river flow responds after rainfall in the catchment area
after a dry period and the effects of percolation into the soil.
Surface-runoff relationship
Rain-runoff is a process that occurs in both space and time domains. Spatial domain
is defined by catchment boundaries, rainfall distribution, topography, soil layers,
river network, and all that have any tangible effect on the resulting estimation. The
time domain is defined by the variation of the estimate for each discrete rainfall point
a long time.
Fig. 4.11 Padang Besar daily rainfall in the year of 2010.
For small catchment areas likes Timah Tasoh, lumped sum model approach is the
most suitable model to be used [128]. In this method, the complexity of parameters
53
such as topology, type of soil and type of vegetation in the catchment area can be
avoided [129][130].
The water level in the Timah Tasoh reservoir is monitored and controlled manually.
Therefore, the level would change artificially when the water is released irrespective
of rain condition. If too much water is released, it could result in water shortage if
inflow from the two rivers is not sufficient. A prediction of the river flow rate would
enable a more efficient water release management to be put in place, to ensure the
safety the dam and balance the mitigation flooding against retaining sufficient water
for domestic and agricultural use. The flow rates of the rivers are affected, naturally,
by rainfall in the catchment area. To understand the catchment area, further analysis
to find the relationship between rainfall and water discharge into the dam were
carried out using the available data in the catchment area which are:
Padang Besar rainfall station
Lubok Sireh rainfall station
Kaki Bukit rainfall station
River Pelarit flow rate
River Jarum flow rate
54
Tdry,
Dry Period-Duration before rain
Tdelay
Flow Delay-Time between the stat of rain and a change in the river flow rate.
Train
Rain Period -Duration of rain.
TF Change Time duration for the flow rate to return to base flow rate.
Fig. 4.12 Relation of rainfall and river flow in the basin of Timah Tasoh.
Fig. 4.12, illustrates the relations of rainfall and river flow. The rainfall affects the river
flow rate at a time duration that depends on the topology, distance of weather station
from the river, hydrology of the catchment, etc. As the rain starts to get heavy, the
soil becoming saturated. Thus there is an increase in surface runoff and water starts
flowing into the river channel, causing the rate of discharge to increased. A very steep
rise would indicate a fast increasing discharge rate meaning that water is flowing into
the river channel at a rapid rate, and thus imply heavy rainfall in the river catchment.
When the river flow reaches the maximum level (peak discharge), and the rain stops
the river discharge starts to decrease, the curve will start to slope gradually (gradient
is usually less steep compared to the lag time). However, the flow rate can peak when
it still raining, in which case the river flow rate starts to increase.
This analysis is necessary to determine the saturation time for the basin. This could
also describe the soil moisture variation based rainfall cumulated over a fixed period
Time
TdelayTdry TF Change
T rain
River Flow rate
55
and the time evolution of these variables. In other words, this analysis is to determine:
Duration of time for when flow changes based on the amount of rainfall in the
basin after a dry period.
Duration of time that the flow rate affected by rain.
The minimum time to predict the river flow after the rain occurred.
Table 4.5 shows the number of available data from stations in Timah Tasoh basin
which consist of rainfall and river flow rate data from 2010 to 2013 with 5 minutes
intervals.
Station Measurement parameter Amount ofdata
1 Padang Besar Rainfall (mm/h) 420,7682 Lubok Sireh Rainfall (mm/h) 420,7683 Kaki Bukit Rainfall (mm/h) 420,7684 River Pelarit Flow rate (m3 / sec) 420,7685 River Jarum flow rate (m3 / sec) 420,768
Total of data for each station.
The next step is to group the related weather and river flow stations according to the
river catchment as shown in Table 4.1.
Station Rivercatchment
1 Padang Besar, River Jarum Jarum2 Lubok Sireh, Kaki Bukit, River.
PelaritPelarit
Related stations for Pelarit and River Jarum.
ANN-based river flow rate model
Satisfactory flood prediction requires a model that represent the parameters of the
river. As discussed in Chapter 3, ANN falls under data-driven hydrological models.
ANN is sensitive to the data that is used and there are several conditions that must
be fulfilled to ensure that the process to produce river flow rate model can be
generalised. This includes selecting appropriate parameters as input based on their
relevance and to reduce the possibility of overfitting during the training phase.
56
Comparison between river level and river flow rate measurement.
Current practice in flood monitoring, river level will be constantly monitored as well
as rainfall rate and duration in order to warn the public the possibility of flooding. In
Malaysia, reliance on water level monitoring often results in very short lead time
warning. The water level markers are used as shown in Fig. 4.13. This study aims to
detect events that lead to an increase in river or dam water to improve flood forecast.
Table 4.7 provides the details of the current practices. It should be noted that these
level are located specifically.
Fig. 4.13 Water Level Classification at Flood Warning Centre [131].
Station Name(Cross-section)
DistrictRiverBasin
(Trend)
Normallevel
AlertLevel
WarningLevel
DangerLevel
River PelaritKakiBukit River Perlis 35.60 38.60 38.72 39.00
River Jarum Beseri River Perlis 30.00 33.30 33.39 33.60
Stages level of rivers (meter) in Perlis, Malaysia from DID website forflood monitoring system [132].
Summary
This chapter has broadly discussed the methodology. It has discussed the rain
dynamics within the catchment area. The discussion of rainfall in the catchment area
has broadly reviews regarding the how to determine has been provided. Most
important, ANN, which has been selected for use in this research has been described.
A brief introduction to the current flood event monitoring used in Malaysia has been
57
provided. The following chapter describes in detail the distribution of rain in the
studied area, the advantages of river flow monitoring compared to river level
monitoring and the implementation of ANN modelling to predict river flow rate.
Major Outcome
To develop the river model in this studies, several stages have been involved. Firstly,
the land coverage had to be studied. In this case, it is found that most of the studied
area consists of farmland and forest. The next phase, the rainfall characteristic have
studied to determine the rain characteristic in the tropical area. Next, the relationship
between rainfall and river characteristic (flow rate and level) for River Jarum and
River Pelarit, based on the studies of data from 3 rainfall stations and two river flow
stations, will be determined. Finally, to perform flow modelling, ANN modelling will
be introduced and described how it could be used to predict the river flow rate.
58
Chapter 5 Result
The results of the study are presented in this chapter. The aim of this study was to
use rainfall data in the catchment area to predict river flow rate which can be used to
estimate the amount of water as well as predict and prevent flooding. To achieve this
rainfall in the catchment area has been studied. The initial studies were to gain an
understanding of the rain pattern in the catchment area. The existing method of
measuring the water level in the rivers to predict flooding is considered inadequate
as it does not provide sufficient advanced warning.
To analysis the trends of precipitation based on 5 minutes data for all the stations,
Mann-Kendall (MK) test been used [133][134][135]. The result shows that there is no
fixed seasonal pattern of rainfall in Perlis based on data from 2010 to 2013. This is
evident in Fig. 5.1. The North East Monsoon (NEM) period should be the wettest
months in Perlis and the driest during the South West Monsoon (SWM) period.
However, Fig. 5.1 shows that the rainy season in 2010 shifted several months earlier
and in 2011, the double-peaked rainy season occurred in March and September which
is abnormal compared to other years within the four years.
Fig. 5.1 Average Monthly rainfall for 14 weather stations in Perlis from 2010 to 2013.
Rain analysis has been carried out on the 4-year data obtained from Perlis. The results,
based on measurement in the catchment area using data from 4 stations are
Average Monthly Rainfall for 2010 to 2013
59
summarised in Table 5.1 and Table 5.2.
Year Number of datasamples (5 min)
Number of data sample withrainfall (14 stations)
Average rain period (hours)
2010 1,471,680 52079 7442011 1,471,680 39049 5582012 1,475,712 40274 5752013 1,471,680 36559 522
Table 5.1 The total and duration of rainfall from 2010 to 2013 for 14 stations inPerlis.
From Table 5.1, in 2010 the highest average duration of rain for each station with 744
hours and 2013 had the lowest. An assumption is made in this analysis that when rain
is recorded, it will last for the duration of 5 minutes (quantisation).
Rain R
ate (mm
/h)
WA
NG
KE
LIA
N
PA
DA
NG
BE
SAR
LU
BO
KSIR
EH
KA
KI B
UK
IT
TA
SOH
BK
T.
TE
MIA
NG
LD
G. P
ER
.SE
L
AB
I KG
.B
AH
RU
NG
OL
AN
G
GU
AR
NA
NG
KA
UL
U P
AU
H
PA
DA
NG
KA
TO
NG
AR
AU
KG
. BE
HO
RL
AT
EH
Frequency
10 11922
9194 9056 10883
8875 15625
8718 13206
8591 9537 8987 8903 8780 914920 1042 820 900 942 887 675 808 760 803 939 837 811 815 86030 508 383 445 458 444 330 383 364 420 430 424 413 396 43440 245 215 210 193 215 156 204 165 178 220 213 184 196 21650 153 143 117 152 145 107 128 129 106 145 116 136 131 14360 102 91 90 84 84 65 96 73 91 98 93 92 85 10570 50 57 46 40 54 49 55 51 44 51 62 73 68 7680 39 47 38 29 35 27 35 31 45 40 30 32 27 5190 42 35 12 18 23 30 25 30 21 25 29 32 28 42
100 14 13 18 17 14 8 18 6 16 18 19 12 14 24110 11 11 10 4 10 12 13 7 10 9 7 13 12 16120 5 18 6 6 11 3 8 8 10 6 11 7 5 8130 3 5 1 7 7 4 3 3 2 5 3 5 5 4140 5 1 2 3 2 1 2 3 3 0 4 3 2 3150 8 7 1 5 4 6 3 4 2 0 0 3 6 4
TOTALL
14149
11040
10952
12843
10810
17098
10499
14840
10342
11523
10835
10719
10570
11135Table 5.2 The frequency of rainfall rates for 10 to 150 mm/h in Perlis.
60
WANG KELIAN
PADANG BESAR
LUBOK SIREH
KAKI BUKIT
TASOH
BKT. TEMIANG
LDG. PER. SEL
ABI KG. BAHRU
NGOLANG
GUAR NANGKA
ULU PAUH
PADANG KATONG
ARAU
KG. BEHOR LATEH
Number ofrain
occurrences of5 min interval
14149 11040 10952 12843 10810 17098 10499 14840 10342 11523 10835 10719 10570 11135
Mean (mm/h) 6.7 7.4 7.0 6.4 7.4 4.4 7.2 5.2 7.2 7.1 7.2 7.3 7.2 7.8
Median(mm/h)
2.4 2.4 2.4 2.4 2.4 1.2 2.4 1.2 2.4 2.4 2.4 2.4 2.4 2.4
Standard dev(mm/h)
14.3 14.3 12.6 12.3 13.6 10.2 13.6 11.1 13.3 12.8 13.3 13.8 13.6 14.6
Min (mm/h) 1.2 1.2 1.2 1.2 1.2 1.2 1.2 1.2 1.2 1.2 1.2 1.2 1.2 1.2
Max (mm/h) 189.6 216.0 148.8 166.8 230.4 192.0 177.6 170.4 164.4 129.6 136.8 246.0 165.6 166.8
Table 5.3 Descriptive statistic rainfall of 5 minutes for 14 stations in Perlis for 2010to 2013.
Descriptive statistics of rainfall rate for 14 stations have been summarised in Table
5.3. The highest rain duration per year is measured at Bukit Temiang station with
rain duration of 356 hours per year. However, the maximum rainfall rate was
recorded at Padang Katong station in 2013 with rain rate of more than 150mm/hour.
One of the most important indications for potential flood event is high daily rainfall.
It has been found that, during the 2010 flood event, the daily rainfall rate was 246 mm
within 24 hours [136]. Fig 5.2 compares the distribution of rain rate from 2010 to 2013.
61
Fig. 5.2 Comparison of Rainfall Rate from 2010 to 2013 in Perlis
Rain intensity has been categorised into four groups [136] [137][138] as follows:
light precipitation; R ≤ 10mm/h moderate: 10 < R ≤ 25 mm/h heavy: 25 < R ≤ 50 mm/ h extreme: R > 50 mm/ h.
Fig. 5.3 shows the number of rain occurrences for each category. 85% of the rainfall is
categorised as light, 9% as moderate, 4% as heavy and 2% for extreme rainfall. Heavy
and extremely heavy rainfall account for 6% which at these intensities, will severely
affect services like a high frequency of communication (> 10 G Hz) and also lead to
flash floods.
Fig. 5.3 Rain intensity in 2010 to 2013 at 14 weather stations in Perlis.
020000400006000080000
100000120000140000160000
10 mm 25 mm 50 mm Exceed 50 mm
Freq
uenc
y
Rain rate in 4 years in Perlis85%
62
Rainfall amount varies from 1 mm per day to 225 mm per day as shown in Table 5.4.
The highest daily rainfall of 228.9 mm was recorded in Padang Besar (which is in the
catchment area) rain station on 01/11/2010. On 02/11/2010, the large flood event
occurred downstream of the catchment area indicating that the daily cumulative
rainfall can be used to predict potential flood [124].
WA
NG
KE
LIA
N
PAD
AN
GB
ESA
R
LU
BO
KSIR
EH
KA
KI B
UK
IT
TA
SOH
BK
T.
TE
MIA
NG
LD
G. PE
R.
SEL
AB
I KG
.B
AH
RU
NG
OL
AN
G
GU
AR
NA
NG
KA
UL
U PA
UH
PAD
AN
GK
AT
ON
G
AR
AU
KG
.BE
HO
RL
AT
EH
Mean (mm/day) 14.0 13.1 11.7 12.4 12.4 12.7 13.1 12.6 11.9 12.2 11.7 12.4 11.8 13.6
Median (mm/day) 7.0 5.9 6.3 6.6 6.4 7.4 7.0 7.4 6.1 6.5 6.4 6.0 6.5 7.0
Standard dev (mm/day) 19.5 20.1 17.0 18.4 18.4 17.5 17.6 16.5 16.4 16.7 15.5 17.6 16.1 18.2
Min (mm/day) 0.1 0.3 0.4 0.1 0.2 0.1 0.1 0.1 0.2 0.1 0.1 0.2 0.3 0.1
Max (mm/day) 162.7 228.9 168.4 174.1 198.7 130.3 191.5 137.7 155.4 180.4 149.1 138.3 152.4 139.5
Table 5.4 The descriptive statistic for daily rainfall of 14 stations in Perlis for 2010to 2013.
YEAR
WANG
KELIAN
PADANGBESAR
LUBOKSIREH
KAKI BUKIT
TASOH
BKT.TEM
IANG
LDG. PER.SEL
ABI KG.BAHRU
NGOLANG
GUARNANGKA
ULU PAUH
PADANGKATONG
ARAU
KG. BEHORLATEH
2010 160 216 120 160 146 112 118 125 132 130 137 149 166 126
2011 139 128 149 148 143 109 126 122 134 119 136 128 156 136
2012 158 158 133 167 143 192 157 170 119 128 120 134 127 167
2013 162 246 125 162 164 121 145 138 178 190 144 230 109 145
Table 5.5 Maximum rainfall rate (mm/h) for 2010 to 2013 for 14 stations in Perlis.
Session Time Frequency Percentage (%)
Early Morning (EM) 1 a.m. to 6 a.m. 943 16.24
Morning (M) 7 a.m. to 12 p.m. 1173 20.20
Afternoon (A) 1 p.m. to 6 p.m. 2168 37.33
Night (N) 7 p.m. to 12 a.m. 1524 26.24
Table 5.6 Numbers rain of occasions in 2013 in 14 rain stations in Perlis
63
KG. BEHOR LATEH
PADANG KATONG
GUAR NANGKA
ARAU
NGOLANG
ULU PAUH
ABI KG. BAHRU
BKT. TEMIANG
LDG. PER. SEL.
WANG KELIAN
KAKI BUKIT
TASOH
LUBOK SIREH
PADANG BESAR
Average (%)
EM 16 18 16 16 15 17 15 16 16 21 16 14 15 18 16
M 18 22 21 19 24 19 22 18 23 24 20 21 20 23 20
A 39 36 38 39 37 39 33 35 31 30 32 38 38 30 37
N 27 25 24 26 25 26 30 30 31 26 31 27 28 29 26
Table 5.7 Percentage of a rain event in 4 different sessions: Early Morning (EM),Morning (M); Afternoon (A), Night (N) for 14 rain stations in Perlis.
To analyse rain occurrences on a daily cycle, the day was divided as shown Table 5.7.
This shows that most rainfall occur are in the afternoon with 37% of the rain events
dominated by convective rain [139]. Understanding these statistics or daily variation
is important as it shows that localised flash flood is most likely in the late afternoon
and early evening, whilst widespread flooding is most likely going to start at night
due to the time it takes for runoff to result in widespread flood.
Four cases of flood events are studied to establish the effectiveness of standard river
level based flood monitoring approach compared with the proposed river flow rate
monitoring approach for River Pelarit and River Jarum. These four flood events are
used to justify the use of flow rate as a river parameter to be predicted as an indicator
for a potential flood.
64
Case 1: River Pelarit: 30th October to 1st November 2010
Fig. 5.4 Rainfall for Kaki Bukit, level and flow rate of River Pelarit during the2010 flooding event.
Fig. 5.4 shows the observed discharge hydrographs between 31 October and 04
November 2010. The river flow rate, the river level and the Kaki Bukit rain gauge
station reading are shown. The flood started on 2nd November downstream of Timah
Tasoh reservoir. There was a period of persistent rainfall in the Timah Tasoh
catchment area between 31 October and 02 November. It is found that the cumulative
rainfall per day was the highest in 2010 before the flood event started. It could be
concluded that the main cause of the flooding was from diffuse precipitation covering
large spatial and temporal scales within the catchment. The amount of rain meant the
Timah Tasoh reservoir exceeded its capacity within a short space of time leading to
flooding both upstream and downstream.
Station Name Alert Level Warning Level Danger Level
River Pelarit at Kaki Bukit 38.60 38.72 39.00
Table 5.8 Stages level of River Pelarit (meter)
Table 5.8 shows the water level and the warning level of River Pelarit that are
assigned by the authorities (DID). Table 5.9 gives a summary of changes in the level
and the flow rate of the River Pelarit during the period shown in Fig. 5.9 . From the
0.3 m3/sec
47,23 m3/sec
65
Fig. 5.9 illustrates the rising limb of Pelarit river flow rate which started 19:00 on 31
Oct. 2010. The flow rate changed from 0.30 cubic m3/sec and rose to 47.23 cubic m3/sec
which is more than 156 times the normal flow rate. By comparison, the river level
only increased from 35 m to 39.90 m which is a 14 % change as given in Table 5.9.
Note that the river level is referenced to the mean height above sea level rather .than
the depth of the river. It should also be noted that river level is a function of the terrain
and change in height alone is a poor indicator
Normal Peak Percentage changesRiver Level 35 m 39.90 m 14Flow rate 0.30 m3/s 47.23 m3/s 15643.33
Table 5.9 River level and flow rate changes during 2010 flood event.
The result shows that the river flow rate is more sensitive to rainfall rate than river
level. The river level also depends on the location of a river course.
In this case, the level of River Pelarit only reaches and exceeds danger level (shown
in Table 5.8) for 12 hours on 1st November. From Fig. 5.3, it can be seen that the
conventional method of monitoring flood by measuring the river level is not sensitive
to changes in the river leading to short warning lead-time.
66
Case 2: River Jarum, Flood event on 30th October to 1st
November 2010
Fig. 5.5 illustrates the observed discharge hydrographs, (the river flow rate and the
water level) of the River Jarum, and the rain gauge reading at Padang Besar station
between 31 Oct and 04 November 2010. Table 5.10 gives a summary of the changes in
the level and the flow rate of the river during the period shown in Fig. 5.5.
Fig. 5.5 Jarum River Level, Flow rate and rainfall at Padang Besar weather station.
Normal Peak % changesRiver Level 30 m 34 m 13Flow rate 0.30 m3/sec 59.90 m3/sec 198666
Table 5.10 Summary of parameters changes during 2010 flood event
Station Name Alert Level Warning Level Danger Level
River Jarum 33.30 33.39 33.60
Table 5.11 Water level as classified by DID for River Jarum.
From Fig. 5.5, the river flow rate started to rise at around 19:00 on 31st October. The
total measured rainfall was 401.4 mm from 31st October to 02 November. The flow
rate changed from 0.29 m3/s and peaked at 59.90 m3/s meanwhile the river level
increased from 30 to 34 meter as in Table 5.10.
67
Case 3: River Pelarit, Flood event on 1 April 2011
Fig. 5.6 The plot of River Pelarit Level, Flow rate and rainfall at Kaki Bukitweather station.
Fig. 5.6 illustrates the observed discharge hydrographs (flow rate and the water level)
for River Pelarit together with Kaki Bukit rain gauge station reading between 27
March and 02 April 2011.
Normal Peak % changesRiver Level 35 m 38.76 m 10.74Flow rate 0.30 m3/sec 47.01 m3/sec 15570
Table 5.12 Summary of parameters changes during 2011 flood event
Table 5.12 shows that the river flow rate rose from 28 March 2011. The flow rate
changed from 0.30 m3/s and peaked at 47.01 m3/s. The river level increased from 35 m
to 38.76 m.
Case 4: River Jarum, Flood event on 1 April 2011
Fig. 5.7 illustrates the measured river level and flow rate, and rainfall at the Padang
Besar station from 27 March to 02 April 2011. The flow rate changes from (0.30 m3/s
to the first peak of 48.35 m3/s) and the river level increased from 30 to 33.61.
68
Fig. 5.7 Rainfall during the 2011 flooding event for Padang Besar andRiver Jarum at a recording interval of 5 minutes.
Normal Peak % changesRiver Level 30 m 33.61 m 12.03Flow rate 0.30 m3/sec 48.35 m3/sec 16016
Table 5.13 Summary of parameters changes during 2011 flood event
Summary for comparison between level monitoring and
flow rate measurement.
Flood forecasters rely heavily on real-time data of rainfall and river water levels as
well as rainfall forecasts. They use hydrology models to estimate runoff in the
catchment area and hence potential for flooding.
The case studies of the two major flood events (October 2010 and April 2011), for both
River Pelarit and River Jarum, show that changes to the flow rates were more
pronounced compared to river levels. Flow rate can also be used to calculate the
volume of water which is important for water resource management. Therefore, this
study focuses on predicting river flow rate to address both flooding as well as water
resource management.
0.3 m3/s
48.35 m3/s
69
Fig. 5.8 Relationship of rainfall and river flow rate.
The purpose of this analysis is to determine the duration of effectiveness for river
flow rate prediction. Fig. 5.8 illustrates the rainfall and river flow rate relationship.
The objective is to ascertain the length of time it take to change the river flow rate
after rainfall has been recorded at specific rain gauge stations. This also depends on
the soil moisture or dry period before the rain.
Dry (no rain) period and time delay for flow rate to change
Table 5.14 shows the dry or no rain period and the duration of time delay between
the start of rain and the start of a change in river flow rate. The maximum dry period
was in 2011 when there was no rain for 912 hours or 38 days. The average time delay,
between start of rain and the change in the river flow rate is 18.64 hours for River
Jarum and 32.56 for River Pelarit.
Tdelay
Time
Tdry
T rain
TF Change
Tdry, Dry Period Duration without raining before the Flow goes higher than the average.Tdelay Flow Delay Duration between the Rain start and the change in the Flow rate.T rain Rain Period Duration of rain start and end.
TF Change Duration when the flow rate increase higher than the base flow to flow decrease to lower than base flow
70
Year
Dry period (Hours) Flow delay (Hours)
max average max average
2010 356.27 30.62 118.84 18.52
2011 912.42 30.80 108.45 17.69
2012 201.04 29.91 130.52 19.34
2013 427.14 28.22 125.76 19.04
i. River Jarum and Padang Besar Rain Station.
Year
Dry period (Hours) Flow delay (Hours)
max average max average
2010 327.79 33.18 118.09 12.01
2011 263.83 25.82 114.84 13.04
2012 501.85 34.08 118.42 15.31
2013 599.12 37.18 115.84 14.23
ii. River Pelarit and Kaki Bukit Rain Station
Table 5.14 Dry period and river flow change delay for Pelarit and River Jarum.
The time delay depends on the location of the rain gauge stations. It is worth noting
that whilst the longest dry period for Padang Besar rain gauge station was in 2011
(912 hours), for the Kaki Bukit rain gauge station it was in 2013 (599 hours). These
differences illustrate the small rain cell size in the area. From the table, there is no
strong statistical relationship between the length of the dry period before it rains and
the time delay before the river flow rate is affected by runoff. However, it is thought
to be because of the soil moisture, hence runoff cannot be successfully predicted
based on an annual average. Consideration of cumulative rainfall could be better.
This has not been explored any further in this thesis.
Rain-No change condition
Year
Dry time (Hours) Flow delay (Hours)
min max Average min max Average
2010 0.75 522.92 67.69 0.50 37.20 4.67
2011 0.92 362.10 54.48 0.50 37.20 4.67
2012 0.08 783.91 56.28 0.50 37.20 4.67
2013 0.42 355.69 52.42 0.50 37.20 4.67a) River Jarum (Padang Besar Rain Station) Period without rain and flow delay.
71
Year
Dry time (Hours) Flow delay (Hours)
min max Average min max Average
2010 0.83 592.37 84.89 0.40 24.90 4.79
2011 1.00 653.17 71.68 0.50 20.30 3.44
2012 0.75 372.51 55.87 0.50 41.30 6.31
2013 1.00 260.92 68.36 0.40 75.50 7.68b) River Pelarit (Kaki Bukit Rain Station) Period without rain and flow delay
Table 5.15 Rain with no changes in Pelarit and River Jarum
Table 5.15 shows the dry period before rain and the time between the start of the rain
event and when changes in the flow rates for Pelarit and River Jarum occurred. For
River Pelarit, the maximum dry period was 783.91 hours or 32 days which occurred
in 2012. For River Jarum, the maximum dry period was 653.17 hours or 27 days which
occurred in 2011. For the flow delay, the maximum time delay was 37.20 hours, from
when the rain started for the River Pelarit flow rate to start to increase. Normally, this
duration occurs during the dry season where the ground will absorb the rainwater
until it is saturated and the underground water starts to flow into the river. However,
during the wet season when the soil cannot absorb rainfall, the duration will be
reduced and it takes only 0.5 hours for Jarum and 0.45 hour for River Pelarit flow to
be affected when the rainfall is recorded in the catchment area.
ANN has been used to study the degree of significance of the chosen factors on the
river flow rate prediction.
72
Fig. 5.9 Simplified diagram of Timah Tasoh Catchment Area.
In this study, eight river flow rate models have been developed. The first four models,
(Model I, Model II, Model III and Model IV) have been designed for River Pelarit and
the rest (Model V, Model VI, Model VII and Model VIII) have for River Jarum. Two
different datasets have been chosen, (Nov 2010 and Oct 2013) which are periods prior
to the flood events. Table 5.16 shows the details of the stations where data were used
as input and output for each river that in this study.
Station Description
PB (input 1) Padang Besar Rainfall
SJ (input 2) Previous Flow rate of River Jarum of t
SJ (target) Predicted Flow rate of River Jarum
a) Model for River Pelarit Prediction-(Model I to IV)
TIMAHTASOH
RESERVOIR
RIVERPELARIT
RIVERJARUM
TIMAHTASOH DAM
RAINFALLKAKI BUKIT
RAIN STATION
RAINFALLPADANG BESARRAIN STATION
9.87 km 7.40 km
73
Station Description
LS (input 1) Lubok Sireh Rainfall
KB (input 2) Kaki Bukit Rainfall
SP (input 3) Previous Flow rate of River Pelarit of t
SP (output) Predicted Flow rate of River Pelarit
b) Model for River Jarum Prediction. (Model V to VIII)
Table 5.16 Details of Input and the Output for Pelarit and River Jarum FlowPrediction Model.
For October 2010 prediction, 87552 data samples have been used and for November
2010 prediction, 88552 data samples have been used. Thus, the 70-30 (training and
test datasets) configuration has been adopted. Regarding data training, the algorithm
randomly selects 70% samples from the whole dataset for each attempt. The value of
the root means square error (RMSE) depends on the selected data. 30 % of the
remaining data is used for validation and testing.
In this study, two training algorithms were used. Levenberg-Marquardt (LM)
algorithm for the models I, III, V and VII, Bayesian Regularisation (BR)
backpropagation algorithms in models II, IV, VI and VIII to determine the prediction
period. The selection of two algorithms has been discussed in Section 3.7. The
prediction periods were increased in step of 12 hours until the model was not able to
produce the predicted flow rate using number of performance indicators (MAPE,
RMSE and Regression). In this study, the best prediction was selected to be the one
with the lowest value of MAPE, the lowest value of RSME and the correlation
coefficient or that has a regression close to value 1.
74
In this configuration, the LM training algorithm was used for eleven different
prediction periods as outlined in Table 5.17. Table 5.17 shows that the best result for
the prediction model is only up to 12 hours. When the prediction period is increased,
this model fails to produce good prediction results.
Prediction Period Performance MeasurementTrails Hours MAPE RMSE Regression
1 12 0.001 0.006 0.995
2 24 163.509 328.210 -0.0033 36 130.865 306.647 04 48 165.430 57.6355 -0.3095 60 40.173 125.363 0.0606 72 57.529 184.802 -0.0367 84 65.531 177.280 0.0028 96 83.783 195.204 -0.0409 108 92.585 195.423 -0.034
10 120 108.840 205.217 011 132 105.033 189.562 -0.031
Table 5.17 River Pelarit (Training: train LM, Target: Oct 2013). Grey shading denotesthe best performing model.
a) Plot of the Predicted and measured Flow rate of River Pelarit for 12
Hours lead time using LM training algorithm, (Oct 2013)
Flow
rat
e(m
3 /s)
75
b) Error Plot of the predicted and actual flow rate from NARX model
c) Regression plot of actual and forecasted flow of River Pelarit.
Fig. 5.10 Prediction using Model 1: River Pelarit. Deterministic forecast: (a)observed and predicted time series; (b) residual plot; (c) scatter plot.
Fig. 5.10 a) shows a comparison of 12 hours prediction results and measured flow
rate. The graph shows that the predicted and measured result are very similar,
illustrating that within that period, the model is accurate. Fig. 5.10 b) shows the error
between predicted river flow rate and the actual flow rate. The values of root mean
square error (RMSE), MAPE and correlation coefficient or regression is 0.006, 0. 002
and 0. 99, respectively.
Pred
icte
dFl
ow
(m3 /s
)
(m3 /s
)
76
In this prediction model, Bayesian Regularisation (BR) training function was used.
Eleven different prediction period is outlined in Table 5.18 were tested.
Prediction Period Performance MeasurementExperiment Hours MAPE RMSE Regression
1 12 0.003 0.009 0.9902 24 0.004 0.012 0.9993 36 0.028 0.005 0.9994 48 0.011 0.252 0.9995 60 0.016 0.497 0.9996 72 0.014 0.401 0.9997 84 0.449 0.019 0.9998 96 0. 018 0.437 0.9999 108 0.016 0.379 0.999
10 120 99.697 191.484 -0.03311 132 106.740 192.409 -0.027
Table 5.18 River Pelarit (Training: train BR, Target: Oct 2013). Grey shading denotesthe best performing model.
a) The plot of the Predicted and Measured Flow rate of River Pelarit for 108 Hours Lead
Time using BR training algorithm. (Oct 2013)
Flow
rat
e(m
3 /s)
77
b) Error Plot of the predicted and actual flow rate from ANN model
c) Regression plot of actual and forecasted flow of River Pelarit.
Fig. 5.11 Simulation for Model 2: River Pelarit for Oct. 2013 dataset. Deterministicforecast: (a) observed and predicted time series; (b) residual plot; (c) scatter
plot.
Fig. 5.11 0 shows the result of 108 hours prediction for River Pelarit and the actual
flow rate in October 2013. Comparing with the results in Fig 5.11 (River Pelarit with
LM Training algorithm), the BR algorithm is able to increase the prediction period up
to 108 hours or 4.5-day period. By comparing both predicted and the actual flow rate,
the error is in the acceptable range with the maximum value of 1.2 cubic m3/s at the
highest peak of flow rate (39 m3/s). The results also demonstrate the excellent
performance of the proposed algorithm with the values of RMSE, MAPE and
regression of 0.378, 0. 017 and 0. 99, respectively.
Pred
icte
dFl
ow
Flow
rat
e(m
3 /s)
78
Prediction Period Performance MeasurementExperiment Hours MAPE RMSE Regression
1 12 0.018 0.0567 0.9992 24 54.589 202.198 -0.4873 36 0.0915 0.0163 0.9994 48 30.941 185.314 -0.1745 60 21.072 146.514 -0.2646 72 12.832 111.467 0.0067 84 8.635 98.479 0.0138 96 14.630 62.436 0.0099 108 57.026 95.349 -0.001
10 120 84.479 95.201 0.00111 132 0.744 22.117 0.892
Table 5.19 Pelarit Prediction Model. (Training: train BR, Target: Nov 2010). Greyshading denotes the best performing model.
a) Plot of Predicted and Measured Flow rate of River Pelarit for 12 Hours lead time using
LM training Algorithm, (Nov 2010)
b) Error Plot of the predicted and actual flow rate from NARX model
79
c) Regression of actual and forecasted flow of River Pelarit.
Fig. 5.12 Prediction of Model 3: River Pelarit for Nov 2010 dataset. Deterministicforecast: (a) observed and predicted time series; (b) residual plot; (c) scatter plot.
Fig. 5.12 0 shows the result of 12 hours prediction period using the November 2010
dataset. The values of RMS error, MAPE and correlation coefficient are 0.06, 0.027 and
0. 999, respectively.
Prediction Period Performance MeasurementExperiment Hours MAPE RMSE Regression
1 12 0.0191 0.0598 0.999762 24 0.0162 0.0606 0.999913 36 0.0133 0.0741 0.999954 48 0.0089 0.1087 0.999985 60 0.0095 0.2037 0.999996 72 0.0075 0.2219 0.999917 84 0.0279 0.7508 0.999738 96 0.0465 2.1526 0.998539 108 0.0397 1.9065 0.99884
10 120 0.0581 3.2330 0.9969411 132 0.7440 22.1169 0.89261
Table 5.20 Pelarit (Training: train LM, Target: Oct 2010). Grey shading denotes thebest performing model.
Fore
cast
ed
80
a) Plot of the Predicted and Measured Flow rate for River Pelarit 108 Hours Prediction using
BR training algorithm, Prediction: Oct 2010
b) Error Plot of the predicted and actual flow rate from the output of NN model.
81
c) Regression of actual and simulated flow of River Pelarit.
Fig. 5.13 Prediction from Model 2: River Pelarit for Oct. 2013 dataset
Fig. 5.13 a) shows the result of 108 hours prediction using the October 2010 dataset.
The values of RMSE, MAPE and regression, are 0. 0397, 1.065 and 0.9988, respectively.
Experiment Prediction Period (October 2013) Performance MeasurementModel algorithm Hours Days MAPE RMSE Regression
I LM 12 0.5 0.001 0.006 0.995II BR 108 4.5 0.016 0.378 0.999
Prediction Period (November 2010) Performance Measurementalgorithm Hours Days MAPE RMSE Regression
III LM 12 0.5 0.0181 0.056 0.999IV BR 108 4.5 0.040 1.906 0.998
Table 5.21 Overall of the best prediction model performances based on differenttraining algorithms for River Pelarit.
The results of river flow rate prediction for River Pelarit are summarised in Table
5.21. When LM training algorithm is used acceptable prediction of the flow rate up to
12 hours is achieved. Using the BR algorithm, acceptable results up to 132 hours (4 ½
days) can be obtained. Based on the processing speed, on average LM is able to
produce quicker results which are up to 3 to 4 times faster compared to the model
that is using LM training function. This is because the LM algorithm has a faster
convergence and the training time is typically shorter compared to Bayesian
Regularization algorithm. However, taking into account, the run-time of the system,
Fore
cast
edFl
ow
82
the LM training algorithm based model, on average, is 2 to 4 times faster than BR
training algorithm based model. Given that the runtime is in minutes for higher
accuracy, the model using BR training algorithm is considered to be superior.
Prediction Period Performance MeasurementExperiment Hours MAPE RMSE Regression
1 12 0.026 0.080 0.9972 24 0.250 0.997 03 36 111.248 200.715 -0.9084 48 362.444 148.123 05 60 133.354 368.767 -0.0026 72 116.792 378.476 07 84 0.009 0.102 0.9998 96 0.617 10.945 09 108 173.526 379.840 -0.06110 120 232.229 399.939 0.04711 132 239.562 369.233 0
Table 5.22 Jarum Prediction (Training: train LM, Target: Oct 2010)
a) Plotting of the Predicted and Measured Flow rate of River Jarum with 12 Hours Ahead
using LM training algorithm (Oct 2013)
Flow
rate
incu
bic
met
erpe
rsec
83
b) Error Plot of the predicted and actual flow rate from NARX model
c) Regression of actual and forecasted flow of River Pelarit.
Fig. 5.14 Prediction from Model V: River Jarum for Oct. 2013 dataset
Fig. 5.14 a) shows the result of 12 hours river flow rate prediction for River Pelarit.
The values of RMSE, MAPE and correlation coefficient, are 0.080, 0.026 and 0.997,
respectively. The 12 hours prediction is the best result obtained from the models that
use the Levenberg-Marquardt (LM) training algorithm.
Actual Flow1.5 2 2.5 3 3.5
Pred
icte
dFl
ow
1.5
2
2.5
3
3.5
Regression Sungai Jarum-Actual and Forecasting: R=0.99769
DataFit
Erro
rofF
low
rate
(cub
icm
eter
pers
ec)
84
Table 5.23 Jarum (Training: train BR, Target: Oct 2010)
a) Plot of Predicted and Measured Flow rate of River Jarum with 132 Hours prediction using
BR training algorithm (Oct 2013)
b) Error Plot of the predicted and measured flow rate
Erro
rofF
low
rate
(cub
icm
eter
pers
ec)
Prediction Period Performance MeasurementExperiment Hours MAPE RMSE Regression
1 12 236.818 394.962 -0.1952 24 0.013 0.045 0.9993 36 0.019 0.132 0.9994 48 0.018 0.213 0.9995 60 0.021 0.271 0.9996 72 0.021 0.277 0.9997 84 0.023 0.255 0.9998 96 0.025 0.267 0.9999 108 180.513 395.459 -0.06410 120 0.267 0.025 0.99911 132 0.207 0.020 0.999
Flow
rate
incu
bic
met
erpe
rsec
85
c) Regression of actual and forecasted flow of River Pelarit.
Fig. 5.15 Prediction from Model VI: River Jarum for Oct. 2013 dataset.
Fig. 5.15 a) shows a result of 132 hours prediction and measured flow rate using the
October 2010 data set. The values of RMSE, MAPE and correlation coefficient are
0.019, 0.206 and 0.999, respectively. When the prediction period increases in 12 hours
step, the model is able to produce good prediction results up to 132 hours where the
RMSE, MAPE and regression coefficient are in acceptable range.
Actual Flow5 10 15 20 25 30
Pred
icte
dFl
ow
5
10
15
20
25
30Regression Sungai Jarum-Actual and Forecasting: R=0.99966
DataFit
86
Prediction Period Performance MeasurementExperiment Hours MAPE RMSE Regression
1 12 0.00471 0.0084 0.9992 24 222.405 403.116 -0.3683 36 166.214 405.339 -0.3054 48 58.1244 185.44 0.0255 60 41.8196 149.812 -0.1746 72 26.7131 124.528 0.0457 84 61.0448 331.454 0.0008 96 32.2485 185.795 -0.0719 108 104.695 159.367 0.00010 120 146.216 132.434 0.0008569211 132 217.637 164.111 -0.071199
Table 5.24 Jarum (Training: train BR, Target: Oct 2010)
a) The plot of the Predicted and Measured Flow Rate of River Jarum with 12
Hours Ahead using LM training algorithm (Oct 2010).
Hour0 1 2 3 4 5 6 7 8 9 10 11 12 13
1.3
1.35
1.4
1.45
1.5
1.55
1.6
1.65
1.7
1.75Sungai Jarum-Comparing of Actual and Forecasting Nov 2010
PredictionsActual
87
b) Error Plot of the predicted and actual flow rate.
c) Regression of actual and forecasted flow of River Jarum.
Fig. 5.16 Prediction from Model VI: River Jarum for Oct. 2010 dataset
Fig. 5.16 0 shows the result of 12 hours prediction using the October 2010 dataset.
Using this configuration, the best prediction is only up to 12 hours with RMSE, MAPE
and regression value of 0. 0083, 0. 0047 and 0.999 respectively. When the prediction
period is increased beyond 12 hours, the model fails to produce good prediction
results.
Erro
rofF
low
rate
(cub
icm
eter
pers
ec)
Pred
icte
dflo
w
88
Prediction Period Performance MeasurementExperiment Hours MAPE RMSE Regression
1 12 0.004 0.007 0.9992 24 0.006 0.021 0.9993 36 0.016 0.235 0.999934 48 0.017 0.331 0.999965 60 0.018 0.613 0.99996 72 0.015 0.565 0.99987 84 0.032 1.859 0.999948 96 0.052 2.459 0.998869 108 0.026 1.359 0.99972
10 120 0.027 1.567 0.9996411 132 0.023 1.283 0.99974
Table 5.25 Model of Prediction for River Jarum (train BR, Target: Oct 2010).
a) Mapping of the Predicted and Actual Flow rate River Jarum 132 Hours Prediction using
BR train, Prediction: Oct 2010.
b) Error Plot of the predicted and actual flow rate from NARX model.
Days0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5
0
10
20
30
40
50
60
70
Step ahead=1584Number of neuron=1Delay=1RMSE=1.28356MAPE=0.0228199
Sungai Jarum-Comparing of Actual and Forecasting Nov 2010Network PredictionsExpected Outputs
Erro
rofF
low
rate
(cub
icm
eter
pers
ec)
89
c) Regression of actual and forecasted flow of River Jarum.
Fig. 5.17 Prediction from Model VI: River Jarum for Oct. 2010 dataset
Fig. 5.17 a) shows the result of 132 hours prediction using the October 2010 dataset.
This is the highest predicting period that can be achieved for this model. The values
of RMSE, MAPE and regression are 1.283, 0.023 and 0. 999, respectively.
Prediction Period (October 2013) Performance MeasurementTrainingalgorithm
Hours MAPE RMSE Regression
I LM 12 0.026 0.08 0.99II BR 132 0.20 0.02 0.99
Prediction Period (November 2010) Performance Measurementalgorithm Hours MAPE RMSE Regression
III LM 12 0.01 0.01 0.99IV BR 132 0.02 1.28 0.99
Table 5.26 Best River Flow Rate Model performance based on different trainingalgorithms for River Jarum.
Table 5.26 shows the best prediction for River Jarum based on various training
algorithms. Using LM training algorithm, the best result is only up to 12-hour
prediction. However, using BR algorithm, the flow rate can be predicted up to 132
hours or 5.5 days.
Studies of rainfall rate in the tropical area have shown that rainfall intensity can reach
and exceed 150 mm/h. Rainfall in Perlis is characterised by highly diverse low and
extremely high rainfall rate where the daily accumulated rainfall can reach 200 mm.
The studies also reveal that there was no consistency in annual rainfall pattern in the
Actual Flow10 20 30 40 50
Prer
dict
edFl
ow
5
10
15
20
25
30
35
40
45
50
55
Regression Sungai Jarum-Actual and Forecasting: R=0.99974
DataFit
90
four years from 2010 to 2013. Therefore, predicting flood events based on previous
flood event will not produce reliable results.
Analyses of rain occurrences on a daily cycle found that the highest rainfall occurs in
the afternoon (12 pm to 6 pm) with 37% of the total yearly rain. This is due to the fact
that rainfall type is dominated by convective rain in tropical areas and the risk of
flood occurrence is higher in the evening and night.
Considering the current flood detection system which uses river level, it has been
shown that the proposed method of using river flow rate is better as an early warning
system because it is more sensitive to rainfall compared to river level. This is a major
finding of this study.
The study of the catchment area characteristics using the rainfall and river flow data,
has shown that the time delay between the start of the rain and when the flow rate
starts to changes in the dry season (or long no rain period) is 18.64 hours for River
Jarum and 32.56 hours for River Pelarit on average. Therefore, the prediction time is
in addition to the delay time, for the locations of the rain gauge stations used.
However, during the wet or rainy season when the soil is saturated, the time delay is
shorter and is down to 0.5 hours for Jarum and 0.45 hour for River Pelarit from the
start of rainfall. It should, however, be noted that this duration can be modified or
will change proportionally with rain intensity and duration.
One of the most significant contributions of this thesis is the development of the river
flow prediction model. Using the ANN-based techniques, the model can produce
short-term (<12 hrs.), medium-term (< 3 days), long-term (4-10 days) flow rate
forecasting of the rivers [140][141]. In general, for very fast approximation, the
Levenberg-Marquardt (LM) training algorithm provides the fastest convergence.
However, as the number of weights in the network increases, the advantage of this
algorithm decreases. Bayesian Regularization (BR) training algorithm although
slower, provides more accurate results, both for short and long-term.
Finally, in this study, the need to know the surface topology has been eliminated as
it is taken into account by the implemented ANN model.
91
This study have revealed that there was no consistency in annual rainfall pattern in
the period from 2010 to 2013. Therefore, predicting flood events based on previous
rain events will not produce reliable results. This study has shown that river flow rate
prediction provides faster response time to rainfall compared to river level.
92
Chapter 6 Discussion and Future Work
This chapter summarises the contributions of this study and makes recommendations
for future works.
Summary of the Study
One of the biggest problems in designing a flood and water runoff forecasting system
is that models that take into account ground cover easily become outdated as a result
of development in the flood plains. In hydrology modelling, a detailed understanding
of the catchment area is required. A complex system that models all the processes
which consider all external elements like rainfall combined with internal factors such
as the state of vegetation, soil moisture, and groundwater levels is often used. The
parameters are difficult and time-consuming to obtain but easily become outdated in
areas where development is taking place.
Simplified analysis of catchment model in Timah Tasoh Catchment Area
In this research, the black box approach has been used as shown in Fig. 6.1. The main
hydrologic parameters of this model are precipitation, river flow rate and temporal
rainfall distribution. Catchment area modelling based on rainfall-runoff data does not
consider the physiographical factors or the catchment area characteristics explicitly
and therefore does not require them to be measured. Because of the black box features,
this system can be operated in a near real-time prediction system because it can
“update” itself based on the real-time input to produce real-time forecasting.
These concepts are suitable for small and medium-sized catchment areas but can be
extended to include larger catchment areas with the use of high spatial resolution rain
gauge network. This avoids the complexities of knowing the values of parameters
such as topology, type of soil and type of vegetation in the catchment area and their
“Black Box Concept”Timah Tasoh Basin
Input (Rainfall incatchment area)
Output (RiverFlow rate)
93
variations. The time-serial correlation between rainfall and the river flow rate has
been extensively used in this study.
River flow rate predictions can be made with the help of an accurate river model. The
relationship between river flow rate and the volume of water, therefore, can be used
to forecast flooding by taking into account the prevailing conditions. From this study,
it has been found that the main influence of rainfall in the River Pelarit and River
Jarum catchment areas is on their flow rates and not on the river levels as used in a
conventional method to predict flooding. The study has shown that the use of river
flow rate will provide a more advanced warning in term of lead time for the
application of flood control and to manage water resources compared to the use of
river level as an indicator.
This research’s aim was to develop an artificial intelligence based computer model
that can be used to improve flood prediction and water resource management in NW
Malaysia. By considering the application of artificial neural networks, a nonlinear
autoregressive machine learning technique was adopted. The developed model has
been trained using preceding conditions such as rainfall values, dry period and
previous flow rate to predict the future river flow rates of two rivers that discharge
into the reservoir. The chosen methodology was to evaluate the performance of
various architectures with a view to determining the most appropriate predictive
model. Study of the results of two rivers prediction models shows that the recurrent
neural network which uses the Bayesian Regularization training algorithm is the
most efficient neural network architecture of the two for river flow rate modelling.
The best prediction period for both River Pelarit and River Jarum obtained is 108
hours and 132 hours, respectably. The network was configured to have one hidden
layer.
Overall, the results show that rainfall data can be used to predict impending floods
or water shortage by using river flow rate as the parameter of measurement. Whilst
the accuracy of the results reported in this thesis is high for up to four day forecast, it
starts to decrease for longer term forecasts. However, the model, if implemented, can
94
regularly update the forecast as new data comes in and adapt to changes in the
catchment area. This type of real-time or near real-time model that uses new data to
adapt its prediction is envisaged to be more appropriate in the phase of climate
change or changing weather patterns.
Overall, the objectives of this research were outlined as follows:
To study rain characteristics in a selected area in Malaysia using available rain
gauges data.
To study the relationship between rainfall rate, river level and river flow rate
to identify which river characteristics (level and flow rate) has the strongest
relationship with rainfall;
To identify a possible technique(s) that can be used to predict the identified
river characteristics.
To develop a computer-based model, to predict the river characteristics, test
and analyse its performance using measured data.
The objectives have been achieved by:
Analysis the rainfall characteristic and pattern in two rivers’ (River Jarum and
River Pelarit) catchment areas in Perlis, Northwest Malaysia.
Establish the relationship between rainfall and river flow rate with the river
flow rate being the best parameter that can be used as an indicator for advanced
flood prediction and water resource availability management.
Identified Artificial Neural Network (ANN) technique, developed and
implementing a model that has been used to predict river flow rate based on
rainfall. This is the first river model that been developed and implemented
based on ANN technique in this catchment area.
Built a model that can predict river flow rates based on rainfall. The result
shows that it can predict river flow rates up to 132 hours in advance with
95
correlation coefficients of more than 0.9. Currently, the is no river model that
has been develop in for the studied area. The dam is manually controlled based
on the experience of the operator. Furthermore, several studies on river
modelling from different areas in Malaysia shows that the best-predicted
model is from 8 to 24 hours in advance [142][143][144].
Contributions
Although several studies have been carried out in Malaysia on flood prediction, the
systems that have been put in place, have proven to be ineffective in providing
sufficient lead time during floods. The actions that are often taken in the management
of water reservoirs or dams are also uncoordinated and decisions are made based on
experience rather than system based. This thesis has made significant and novel
contributions in two major areas; in the river modelling based on ANN technique that
can be used to predict floods and also in the prediction of water inflow into a reservoir
that can be used to manage water resources.
The developed ANN-based model, although other authors have used them in related
areas, is considered to be one of the first to be applied in this context. The long
prediction time and its high accuracy mean that it can be used to manage water dams,
safeguard water supply and to control flooding both upstream and downstream.
Some of the results of this study have already been presented to some members of
the Malaysian Department of Irrigation and Drainage (DID) and aspects of the study
e.g. the use of river flow rate rather than river level as an indicator for potential floods
is being considered for adoption. The current practice of measuring the river level has
does not allow sufficient lead time to issue warning and for evacuation.
Recommendations for Future Works in River Modelling
Although the model developed and used produces good results, the author feels that
several areas of the research carried out can be further strengthened. These include:
The development of a model that relates the position of the rain gauges in the
catchment area to the time delay of the impact of the rainfall on the river flow
96
rate;
The development of a model that will take readings from multiple rain gauges,
providing better spatial resolution, for river flow rate prediction.
The development and deployment of a self-adapting real-time version of the
model and integration with any flooding or water resource management
system.
97
References[1] D. Ruelland, V. Larrat, and V. Guinot, “A comparison of two conceptual
models for the simulation of hydro-climatic variability over 50 years in a largeSudano- Sahelian catchment,” Glob. Chang. facing risks Threat. to water Resour.,no. October, pp. 25–29, 2010.
[2] W. Z. W. Z. & A. A. J. Jamaludin Suhaila, Sayang Mohd Deni, “Trends inPeninsular Malaysia Rainfall Data During the Southwest Monsoon andNortheast Monsoon Seasons : 1975 – 2004,” vol. 39, no. 4, pp. 533–542, 2010.
[3] N. S. Muhammad, “Probability Structure and Return Period Calculations forMulti-Day Monsoon Rainfall Events at Subang, Malaysia,” Colorado StateUniversity, 2013.
[4] K. Sene, Flash Floods: Forecasting and Warning, XIII. Dordrecht: SpringerNetherlands, 2013.
[5] Malaysian Meteorological Department, “Urban Stormwater ManagementManual,” Percetakan Nasional Malaysia Berhad, 2013. [Online]. Available:http://www.water.gov.my/urban-stormwater-mainmenu-564/188.
[6] Urban Stormwater Management Manual for Malaysia, 2012th ed. Department ofIrrigation and Drainage, 2012.
[7] T.N.Alagesh, H. Mohd, and H. Reduan, “Flood: Worsening situation inPahang, Terengganu,” News Straits Times, 2013. [Online]. Available:http://www.nst.com.my/latest/font-color-red-flood-font-worsening-situation-in-pahang-terengganu-1.424948.
[8] Z. A. Akasah and S. V Doraisamy, “2014 Malaysia flood : impacts & factorscontributing towards the restoration of damages,” J. Sci. Res. Dev. 2, vol. 2, no.14, pp. 53–59, 2015.
[9] “Floodlist.” [Online]. Available: http://floodlist.com/asia/malaysia-floods-kelantan-worst-recorded-costs.
[10] M. S. Bin Khalid and S. B. Shafiai, “Flood Disaster Management in Malaysia:An Evaluation of the Effectiveness Flood Delivery System,” Int. J. Soc. Sci.Humanit., vol. 5, no. 4, pp. 398–402, 2015.
[11] “United Nations Platform for Space-based Information for DisasterManagement and Emergency Response - UN-SPIDER.” [Online]. Available:http://www.un-spider.org/about/what-is-un-spider. [Accessed: 31-Jan-2017].
[12] G. Australia, “What is a Flood?,” 2011. [Online]. Available:http://www.ga.gov.au/scientific-topics/hazards/flood/basics/what.
98
[13] M. Dong, “A tutorial on nonlinear time-series data mining in engineering assethealth and reliability prediction: Concepts, models, and algorithms,” Math.Probl. Eng., vol. 2010, 2010.
[14] R. J. Povinelli, “Time Series Data Mining: Identifying Temporal Patterns forCharacterization and Prediction of Time Series Events,” p. 193, 1999.
[15] B. R. Kumar, “Forecasting and control of future values on Spatial-Temporalprediction,” pp. 31–36, 2001.
[16] S. H. Elsafi, “Artificial Neural Networks (ANNs) for flood forecasting atDongola Station in the River Nile, Sudan,” Alexandria Eng. J., vol. 53, no. 3, pp.655–662, 2014.
[17] R. Chuentawat, S. Bunrit, C. Ruangudomsakul, and N. Kerdprasop, “ArtificialNeural Networks and Time Series Models for Electrical Load Analysis,” vol. I,2016.
[18] D. Labat, R. Ababou, and A. Mangin, “Linear and nonlinear input/outputmodels for karstic springflow and flood prediction at different time scales,”Stoch. Environ. Res. Risk Assess., vol. 13, no. 5, pp. 337–364, 1999.
[19] C. M. Zealand, D. H. Burn, and S. P. Simonovic, “Short term streamflowforecasting using artificial neural networks,” J. Hydrol., vol. 214, no. 1, pp. 32–48, 1999.
[20] F. Guenther, “Neural networks: Biological models and applications.” pp. 1–7,2003.
[21] I. Kaastra and M. Boyd, “Designing a neural network for forecasting financialand economic time series.pdf.” 1995.
[22] H. Tabari, P. Hosseinzadeh Talaee, and P. Willems, “Short-term forecasting ofsoil temperature using artificial neural network,” Meteorol. Appl., vol. 22, no. 3,pp. 576–585, 2015.
[23] O. Awodele and O. Jegede, “Neural Networks and its Application inEngineering,” Proc. Informing Sci. IT Educ. Conf. 2009, 2009.
[24] T. Htet, H. San, and M. M. Khin, “River Flood Prediction using Time SeriesModel,” vol. 1, no. 2, pp. 265–269, 2015.
[25] P. R.J., “Identifying Temporal Patterns for Characterization and Prediction ofTime Series Events,” Marquette University, Milwaukee, 1999.
[26] C. Damle, “Flood forecasting using time series data mining,” 2005.
[27] P. Supriya, M. Krishnaveni, and M. Subbulakshmi, “Regression Analysis ofAnnual Maximum Daily Rainfall and Stream Flow for Flood Forecasting inVellar River Basin,” Aquat. Procedia, vol. 4, pp. 957–963, 2015.
99
[28] ESCAP / WMO Typhoon Committee ( MALAYSIA ), “Members Report,” 2011.
[29] E. Baltas, “Advances in Geosciences The combined use of weather radar andgeographic information system techniques for flood forecasting,” Inf. Syst., pp.117–123, 2007.
[30] D. Solomatine, C. Rojas, S. Velickov, and H. Wust, “Chaos theory in predictingsurge water levels in the North Sea,” Proc. 4th Int. Conf. …, no. July, pp. 1–8,2000.
[31] M. Kim and J. Kim, “Predicting IGS RTS Corrections Using ARMA NeuralNetworks,” vol. 2015, 2015.
[32] J. D. Farmer and J. J. Sidorowich, “Predicting chaotic time series,” Phys. Rev.Lett., vol. 59, no. 8, pp. 845–848, Aug. 1987.
[33] a. Porporato and L. Ridolfi, “Nonlinear analysis of river flow time sequences,”Water Resour. Res., vol. 33, no. 6, pp. 1353–1367, 1997.
[34] B. Sivakumar, “Chaos theory in hydrology: important issues andinterpretations,” J. Hydrol., vol. 227, no. 1, pp. 1–20, 2000.
[35] L. Cao, A. I. Mees, and K. Judd, “Dynamics from multivariate time series,”Phys. D Nonlinear Phenom., vol. 121, pp. 75–88, 1998.
[36] F. Laio, a. Porporato, R. Revelli, and L. Ridolfi, “A comparison of nonlinearflood forecasting methods,” Water Resour. Res., vol. 39, no. 5, p. n/a-n/a, 2003.
[37] R. F. Adler and A. J. Negri, “A satellite infrared technique to estimate tropicalconvective and stratiform rainfall,” J. Appl. Meteorol., vol. 27, no. 1, pp. 30–51,1988.
[38] K. S. Takahashi, Tsutomu, “Tropical Rain Characteristics and Microphysics ina Three-Dimensional Cloud Model,” Am. Metrol. Soc., vol. Volume 61, no. 1984,pp. 2817–2845, 2004.
[39] W. Arm, A. Msm, D. Ara, A. Aw, I. Ar, and H. Saliza, “Improved WatershedRunoff Estimation Using Radar-Derived Rainfall in Peninsular Malaysia,” Int.Conf. Recent Emerg. Technol., 2009.
[40] S. Burcea, S. Cheval, A. Dumitrescu, B. Antonescu, A. Bell, and T. Breza,“Comparison Between Radar Estimated And Rain Gauge MeasuredPrecipitation In The Moldavian Plateau,” Environ. Eng. Manag. J., vol. 11, no. 4,pp. 723–731, 2012.
[41] R. Adriyanto, “CURRENT STATUS OF WEATHER RADAR DATAEXCHANGE,” World Meteorol. Organ., no. Workshop On Radar DataExchange, 2013.
[42] J. D. A. R. Nor Hisham Haji Khamis, “Determination of Rain Cell Size
100
Distribution for Microwave Link Design in Malaysia,” RF Microw. Conf., pp.38–40, 2004.
[43] Shuyi S. Chen and R. A. H. Jr, “Diurnal variation and life-cycle of deepconvective systems over the tropical Pacific warm pool,” Q. J. R. Meteorol. Soc.,vol. 123, no. 538, pp. 357–388, 1997.
[44] S. Ramli and W. Tahir, “Radar Hydrology: New Z/R Relationships forQuantitative Precipitation Estimation in Klang River Basin, Malaysia,” Int. J.Environ. Sci. Dev., vol. 2, no. 3, pp. 223–227, 2011.
[45] A. M. Jafri, Z. Hashim, M.L.Kavvas, Z.Q.Chen, and N.Ohara, “Development ofAtmospheric Based Flood Forecasting and Warning System for Selected RiverBasins in Malaysia.,” J. Hydrol., vol. 356, no. 3–4, pp. 1–18, 2010.
[46] T. Ushio et al., “A Kalman Filter Approach to the Global Satellite Mapping ofPrecipitation (GSMaP) from Combined Passive Microwave and InfraredRadiometric Data,” J. Meteorol. Soc. Japan, vol. 87A, no. November 2008, pp.137–151, 2009.
[47] W. Veerakachen, M. Raksapatcharawong, and S. Seto, “Performanceevaluation of Global Satellite Mapping of Precipitation (GSMaP) products overthe Chaophraya River basin, Thailand,” Hydrol. Res. Lett., vol. 8, no. 1, pp. 39–44, 2014.
[48] R. a. Roebeling and I. Holleman, “SEVIRI rainfall retrieval and validationusing weather radar observations,” J. Geophys. Res., vol. 114, no. D21, p.D21202, Nov. 2009.
[49] R. A. Roebeling and I. Holleman, “Validation Of Rain Rate Retrievals FromSEVIRI Using Weather Radar Observations,” EUMETSAT Meteorol. Satell.Conf., 2008.
[50] S. Seto, T. Tsunekawa, and T. Oki, “A new rain detection method tocomplement high-resolution global precipitation products,” Japan Soc. Hydrol.Water Resour., vol. 86, pp. 82–86, 2012.
[51] W. K. Al-Assadi, S. Gandla, S. Sedigh, and I. P. Dugganapally, “Design of aflood prediction system,” 2009 12th Int. IEEE Conf. Intell. Transp. Syst., pp. 1–6,Oct. 2009.
[52] J. N. Ghazali and A. Kamsin, “A Real Time Simulation and Modeling of FloodHazard,” 12th WSEAS Int. Conf. Syst. Heraklion, Greece, pp. 438–443, 2008.
[53] E. Coppola, B. Tomassetti, M. Verdecchia, F. S. Marzano, and G. Visconti,“Small-catchment flood forecasting and drainage network extraction usingcomputational intelligence,” 2006 IEEE Int. Jt. Conf. Neural Netw. Proc., pp. 851–858, 2006.
101
[54] C. R. Jackson and J. R. Apel, Synthetic Aperture Radar Marine User’s Manual, Sep2004. National Oceanic and Atmospheric Administration (NOAA), 2004.
[55] P. Campling, A. Gobin, K. Beven, and J. Feyen, “Rainfall-runoff modelling of ahumid tropical catchment: the TOPMODEL approach,” Hydrol. Process., vol.16, no. 2, pp. 231–253, Feb. 2002.
[56] Z. Yang, “Derivation of unit hydrograph using a transfer function approach,”vol. 42, pp. 1–9, 2006.
[57] C. W. Dawson and R. L. Wilby, “Hydrological modelling using artificial neuralnetworks,” Prog. Phys. Geogr., vol. 25, no. 1, pp. 80–108, 2001.
[58] M. B. Abbott, J. C. Bathurst, J. A. Cunge, P. E. O’Connell, and J. Rasmussen,“An introduction to the European Hydrological System — SystemeHydrologique Europeen, ‘SHE’, 2: Structure of a physically-based, distributedmodelling system,” J. Hydrol., vol. 87, no. 1–2, pp. 61–77, Oct. 1986.
[59] F. Fenicia, D. Kavetski, and H. H. G. Savenije, “Elements of a flexible approachfor conceptual hydrological modeling: 1. Motivation and theoreticaldevelopment,” Water Resour. Res., vol. 47, no. 11, p. n/a-n/a, Nov. 2011.
[60] R. Allen Freeze, “Modelling changes in forest,” Hydrol. Forecast., vol. 12, no. 2,pp. 214–215, 1985.
[61] P. Taylor, C. W. Dawson, and R. Wilby, “An artificial neural network approachto rainfall-runoff modelling An artificial neural network approach to rainfall-runoff modelling,” Hydrol. Sci. J., vol. 43, no. August 2012, pp. 37–41, 1998.
[62] V. Koren, S. Reed, and M. Smith, “Combining physically-based and conceptualapproaches in the development and parameterization of a distributed system,”no. 282, pp. 101–108, 2003.
[63] T. Wagener, M. J. Lees, and H. S. Wheater, “A toolkit for the development andapplication of parsimonious hydrological models,” Math. Model. smallwatershed Hydrol., vol. 2, 2001.
[64] I. G. Pechlivanidis, B. M. Jackson, N. R. Mcintyre, and H. S. Wheater,“Catchment Scale Hydrological Modelling: A Review Of Model Types,Calibration Approaches And Uncertainty Analysis Methods In The Context OfRecent Developments In Technology And Applications,” Glob. NEST J., vol. 13,no. 3, pp. 193–214, 2011.
[65] C. Garc, “Time Series Analysis : Autoregressive , MA and ARMA processes,”2012.
[66] C. L. COCIANU and H. GRIGORYAN, “An Artificial Neural Network forData Forecasting Purposes,” Inform. Econ., vol. 20, no. 2/2015, pp. 34–45, 2015.
102
[67] R. May, G. Dandy, and H. Maier, “Review of Input Variable Selection Methodsfor Artificial Neural Networks,” Artif. Neural Networks - Methodol. Adv. Biomed.Appl., no. August 2016, p. 362, 2011.
[68] F. Ahmad Ruslan, A. M. Samad, Z. M. Md Zain, and R. Adnan, “Flood waterlevel modeling and prediction using NARX neural network: Case study atKelang river,” Signal Process. its Appl. (CSPA), 2014 IEEE 10th Int. Colloq., no. 1,pp. 204–207, 2014.
[69] Haviluddin and R. Alfred, “Performance of Modeling Time Series UsingNonlinear Autoregressive with eXogenous input ( NARX ) in the NetworkTraffic Forecasting,” vol. 2013, no. June 2013, pp. 164–168, 2015.
[70] E. Diaconescu, “The use of NARX neural networks to predict chaotic timeseries,” WSEAS Trans. Comput. Res., vol. 3, no. 3, pp. 182–191, 2008.
[71] Y. Liu, G. Ma, and X. Jiang, “A design method for adaptive inverse controlusing NARX neural networks,” Fifth World Congress on Intelligent Control andAutomation (IEEE Cat. No.04EX788), vol. 1. p. 459–463 Vol.1, 2004.
[72] K. C. Gupta, “Neural Network Structures,” Neural Networks RF Microw. Des.,pp. 61–103, 2000.
[73] C. Bergmeir and J. M. Benitez, “Neural Networks in R Using the StuttgartNeural Network Simulator: RSNNS,” J. Stat. Softw., vol. 46, no. 7, pp. 1–26,2012.
[74] T. Attia, “A Comparative Study of Three Different Topologies of NeuralNetwork – based Multiuser Detectors of WCDMA signals,” no. 6, pp. 61–70,2006.
[75] S. M. Guthikonda, “Kohonen Self-Organizing Maps,” no. December, 2005.
[76] a. K. Mitra, S. Nath, and a. K. Sharma, “Fog forecasting using rule-basedfuzzy inference system,” J. Indian Soc. Remote Sens., vol. 36, no. 3, pp. 243–253,2008.
[77] C. W. Dawson and R. Wilby, “An artificial neural network approach torainfall-runoff modelling,” Hydrol. Sci. J., vol. 43, no. 1, pp. 47–66, Feb. 1998.
[78] J. M. Ortiz-rodríguez, J. M. Ortiz-rodríguez, M. R. Martínez-blanco, and M.Manuel, “Robust Design of Artificial Neural Networks Robust Design ofArtificial Neural Networks Methodology in Neutron SpectrometryMethodology in Neutron Spectrometry.”
[79] P. Hosseinzadeh Talaee, “Multilayer perceptron with different trainingalgorithms for streamflow forecasting,” Neural Comput. Appl., vol. 24, no. 3–4,pp. 695–703, 2014.
103
[80] P. P. Balestrassi, E. Popova, a. P. Paiva, and J. W. Marangon Lima, “Design ofexperiments on neural network’s training for nonlinear time seriesforecasting,” Neurocomputing, vol. 72, no. 4–6, pp. 1160–1178, 2009.
[81] B. Makki and M. N. Hosseini, “Some refinements of the standardautoassociative neural network,” Neural Comput. Appl., vol. 22, no. 7–8, pp.1461–1475, 2013.
[82] N. Sajikumar and B. S. Thandaveswara, “A non-linear rainfall–runoff modelusing an artificial neural network,” J. Hydrol., vol. 216, no. 1–2, pp. 32–55, Mar.1999.
[83] C. Engineering, “Artificial intelligence techniques in flood forecasting m ~BRISTOL,” no. January, 2005.
[84] C. M. Zealand, D. H. Burn, and S. P. Simonovic, “Short term streamflowforecasting using artificial neural networks,” vol. 214, no. May 1998, pp. 32–48,1999.
[85] A. Jain and S. Srinivasulu, “Integrated approach to model decomposed flowhydrograph using artificial neural network and conceptual techniques,” J.Hydrol., vol. 317, no. 3–4, pp. 291–306, 2006.
[86] I. Durre, J. M. Wallace, and D. P. Lettenmaier, “Dependence of extreme dailymaximum temperatures on antecedent soil moisture in the contiguous UnitedStates during summer,” J. Clim., vol. 13, no. 14, pp. 2641–2651, 2000.
[87] T. Chaipimonplin and T. Vangpaisal, “Comparison of the Efficiency of InputDetermination Techniques with LM and BR Algorithms in ANN for FloodForecasting, Mun Basin, Thailand,” Int. J. Comput. Electr. Eng., vol. 6, no. 2, pp.90–94, 2014.
[88] T. Chaipimonplin, L. See, and P. Kneale, “Improving neural network for floodforecasting using radar data on the Upper Ping River,” no. December, pp. 12–16, 2011.
[89] S. Roweis, “Levenberg-Marquardt Optimization,” Notes, Univ. Toronto, 1996.
[90] F. C. Kruse, Predictions Nonlinearities and Portfolio. BoD – Books on Demand,2012, 2012.
[91] W. Wang, P. H. a. J. M. Van Gelder, J. K. Vrijling, and J. Ma, “Forecasting dailystreamflow using hybrid ANN models,” J. Hydrol., vol. 324, no. 1–4, pp. 383–399, Jun. 2006.
[92] J. Tao and A. P. Barros, “Prospects for flash flood forecasting in mountainousregions – An investigation of Tropical Storm Fay in the SouthernAppalachians,” J. Hydrol., Mar. 2013.
104
[93] L. S. Maciel and R. Ballini, “Design a Neural Network for Time Series FinancialForecasting : Accuracy and Robustness Analisys,” Inst. Enonomia. Univ.Estadual Campinas, no. 1986, 2008.
[94] S. C. Nayak, B. B. Misra, and H. S. Behera, “Impact of Data Normalization onStock Index Forecasting,” Int. J. Comput. Inf. Syst. Ind. Manag. Appl., vol. 6, no.2014, pp. 257–269, 2014.
[95] A. I. Erdi TOSUN, Kadir AYDIN, Simona Silvia MEROLA, “Estimation OfOperational Parameters For A Direct Injection Turbocharged Spark IgnitionEngine By Using Regression Analysis And Artificial Neural Network,” p. 3.
[96] U. Edward J. Brouch, Captain, “Artificial Neural Network Prediction OfChemical-Disease Relationships Using Readily Available ChemicalProperties,” Air Force Institute Of Technology, 2014.
[97] S. Karsoliya, “Approximating Number of Hidden layer neurons in MultipleHidden Layer BPNN Architecture,” Int. J. Eng. Trends Technol., vol. 3, no. 6, pp.714–717, 2012.
[98] F.-J. Chang, Y.-M. Chiang, and L.-C. Chang, “Multi-step-ahead neuralnetworks for flood forecasting,” Hydrol. Sci. J., vol. 52, no. 1, pp. 114–130, 2007.
[99] G. Panchal, A. Ganatra, Y. Kosta, and D. Panchal, “Behaviour analysis ofmultilayer perceptrons with multiple hidden neurons and hidden layers,” Int.J. Comput. Theory Eng., vol. 3, no. 2, pp. 332–337, 2011.
[100] D. L. Bailey and D. Thompson, “Developing Neural-network Applications,” AIExpert, vol. 5, no. 9, pp. 34–41, Jun. 1990.
[101] K. G. Sheela and S. N. Deepa, “Review on methods to fix number of hiddenneurons in neural networks,” Math. Probl. Eng., vol. 2013, no. Article ID425740, 2013.
[102] M. A. Mohammed, Lubna. B. Hamdan and E. A. Abdelhafez, “Hourly SolarRadiation Prediction Based on Nonlinear,” Jordan J. Mech. Ind. Eng., vol. 7, no.1, pp. 11–18, 2013.
[103] T. Chaipimonplin, “Investigation internal parameters of neural network modelfor Flood Forecasting at Upper river Ping, Chiang Mai,” KSCE J. Civ. Eng., vol.20, pp. 478–484, 2015.
[104] H. Demuth, “Neural Network Toolbox,” Networks, vol. 24, no. 1, pp. 1–8, 2002.
[105] Ö. Kisi and E. Uncuoglu, “Comparison of three back-propagation trainingalgorithm for two case study,” Indian J. Eng. Mater. Sci., vol. 12, no. October,pp. 434–442, 2005.
[106] J. Eriksson, Optimization and Regularization of Nonlinear Least Squares Problems.
105
1996.
[107] J. Smith and R. N. Eli, “Neural-Network Models of Rainfall-Runoff Process,” J.Water Resour. Plan. Manag., vol. 121, no. 6, pp. 499–508, 1995.
[108] C. W. Dawson and R. L. Wilby, “A comparison of artifical neural networksused for river flow forecasting,” Hydrology and Earth System Sciences, vol. 3, no.4. pp. 529–540, 1999.
[109] N. Sajikumar and B. S. Thandaveswara, “A non-linear rainfall–runoff modelusing an artificial neural network,” J. Hydrol., vol. 216, no. 1–2, pp. 32–55, Mar.1999.
[110] L. Xiong, K. M. O’Connor, and S. Guo, “Comparison of three updatingschemes using artificial neural network in flow forecasting,” Hydrol. Earth Syst.Sci., vol. 8, no. 2, pp. 247–255, 2004.
[111] R. J. DAWSON, “PERFORMANCE-BASED MANAGEMENT OF FLOODDEFENCE SYSTEMS,” no. May, 2003.
[112] Y. A. Pachepsky, G. Martinez, F. Pan, T. Wagener, and T. Nicholson,“Evaluating Hydrological Model Performance using Information Theory-based Metrics,” Hydrol. Earth Syst. Sci. Discuss., no. February, pp. 1–24, 2016.
[113] T. Chai and R. R. Draxler, “Root mean square error (RMSE) or mean absoluteerror (MAE)? -Arguments against avoiding RMSE in the literature,” Geosci.Model Dev., vol. 7, no. 3, pp. 1247–1250, 2014.
[114] C. Res, C. J. Willmott, and K. Matsuura, “Advantages of the mean absoluteerror ( MAE ) over the root mean square error ( RMSE ) in assessing averagemodel performance,” vol. 30, pp. 79–82, 2005.
[115] N. Fenton and M. Neil, “Correlation coefficient and p-values: what they areand why you need to be very wary of them,” Risk Assess. Decis. Anal. withBayesian Networks, 2012.
[116] P. Coulibaly and C. K. Baldwin, “Dynamic neural networks for nonstationaryhydrological time series modeling,” in Practical Hydroinformatics, SpringerBerlin Heidelberg, 2009, pp. 71–85.
[117] A. B. Bhattacharya, D. K. Tripathi, T. Das, and A. Nag, “StatisticalCharacteristics Of Tropical Rain Rate And Rain Intensity From Radar AndRain Gauge Measurements,” vol. 4, no. 1, pp. 53–64, 2011.
[118] A. Tayebiyan, T. A. Mohammad, and A. H. Ghazali, “SCIENCE &TECHNOLOGY Artificial Neural Network for Modelling Rainfall-Runoff,”vol. 24, no. 2, pp. 319–330, 2016.
[119] S. French, “Cynefin, statistics and decision analysis,” J. Oper. Res. Soc., vol. 64,
106
no. 4, pp. 547–561, 2013.
[120] A. R. and K. A. R. Negin Vaghefi, Mad Nasir Shamsudin, “Impact of ClimateChange on Rice Yield in the Main Rice Growing Areas of PeninsularMalaysia,” Res. J. Environ. Sci., vol. 7, no. 2, pp. 59–67, 2013.
[121] Z. A. Mohtar, A. S. Yahaya, F. Ahmad, S. Suri, and M. H. Halim, “Trends forDaily Rainfall in Northern and Southern Region of Peninsular Malaysia,” J.Civ. Eng. Res., vol. 4, no. 3A, pp. 222–227, 2014.
[122] D. U. Lawal, A. Matori, A. M. Hashim, K. W. Yusof, and A. Chandio, “NaturalFlood Influencing Factors : A Case Study of Perlis, Malaysia,” Int. Conf. Civil,Offshore Environ. Eng. (ICCOEE 2012), pp. 1–6, 2012.
[123] M. I. Phang Kun Liong, “A Case Study On The Sever Rain Occured InNorthwestern Peninsular Malaysia From 31 October Till 2 November 2010,”JMM Res. Publ., vol. No. 11/201, 2012.
[124] C. P. Diman and W. Tahir, “Dam Flooding Caused A Prolonged Flooding,”Int. J. Civ. Environ. Eng. IJCEE-IJENS Vol12 No06, no. 6, 2012.
[125] W. R. Ismail, “Timah Tasoh Reservoir, Perlis, Malaysia.”
[126] D. for E. F. & R. Affairs, How to model and map catchment processes when flood riskmanagement planning. Enviroment Agency, 2016.
[127] J. M. Wright, “Chapter 4: Flood Risk Assessment,” Floodplain Manag. Princ.Curr. Pract., pp. 1–25, 2008.
[128] K. J. Beven and M. J. Kirkby, “A physically based, variable contributing areamodel of basin hydrology,” Hydrol. Sci. Bull., vol. 24, no. 1, pp. 43–69, 1979.
[129] C. Reszler, G. Bloschl, and J. Komma, “Identifying runoff routing parametersfor operational flood forecasting in small to medium sized catchments,”Hydrol. Sci. J., vol. 53, no. 1, pp. 112–129, 2008.
[130] C. Zhang, Y. Peng, J. Chu, C. A. Shoemaker, and A. Zhang, “Integratedhydrological modelling of small-and medium-sized water storages withapplication to the upper fengman reservoir basin of China,” Hydrol. Earth Syst.Sci., vol. 16, no. 11, pp. 4033–4047, 2012.
[131] Paridah Anun bt Tahir, “Malaysia Water Resorce Management (MyWRMS)Forum 2014,” 2014.
[132] Department of Irrigation and Drainage (JPS), “Infobanjir.” [Online]. Available:http://infobanjir.water.gov.my/waterlevel_page.cfm?state=PLS.
[133] H. B. Mann, “Nonparametric Tests Against Trend,” Econometrica, vol. 13, no. 3,pp. 245–259, 1945.
107
[134] M. G. Kendall, Rank Correlation Methods, 4th Editio. 1975.
[135] R. M. Hirsch, J. R. Slack, and R. A. Smith, “Techniques of trend analysis formonthly water quality data,” Water Resour. Res., vol. 18, no. 1, pp. 107–121,Feb. 1982.
[136] A. S. D. and G. T. V. K. Andy Y. Kwarteng, “Analysis of a 27-year rainfall data(1977–2003) in the Sultanate of Oman,” Int. J. Climatol., vol. 4, no. December2007, p. 1549–1555., 2008.
[137] I. Conference, E. Science, E. Ipcbee, and I. Press, “Radar Hydrology : New Z / RRelationships for Klang River Basin , Malaysia,” vol. 8, no. 3, pp. 248–251,2011.
[138] Y. C. Gao and M. F. Liu, “Evaluation of high-resolution satellite precipitationproducts using rain gauge observations over the Tibetan Plateau,” Hydrol.Earth Syst. Sci., vol. 17, no. 2, pp. 837–849, 2013.
[139] C. M. Taylor, R. A. M. de Jeu, F. Guichard, P. P. Harris, and W. A. Dorigo,“Afternoon rain more likely over drier soils,” Nature, vol. 489, no. 7416, pp.423–426, Sep. 2012.
[140] C. E. M. Tucci and W. Collischonn, “Flood forecasting,” WMO Bull., vol. 55,no. July, pp. 179–184, 2006.
[141] Z. Liu, “Flood Forecasting and Warning in China.”
[142] R. Ruslan, Fazlina Ahmat; Samad, Abd Manan; Zain, Zainazlan Md; Adnan,“Flood water level modeling and prediction using NARX neural network :Case study at Kelang river.,” Proc. - 2014 IEEE 10th Int. Colloq. Signal Process. ItsAppl. CSPA 2014. IEEE Comput. Soc. 2014., 2014.
[143] Y. Abou Rjeily, O. Abbas, M. Sadek, I. Shahrour, and F. Hage Chehade, “Floodforecasting within urban drainage systems using NARX neural network,”Water Sci. Technol., Jul. 2017.
[144] K. C. Keong, M. Mustafa, A. J. Mohammad, M. H. Sulaiman, and N. R. Hasma,“Artificial Neural Network Flood Prediction for Sungai Isap Residence,” p. 11.
Certificate Code: DE21-1DCB-316A-0A2A-5BFA-18FE-9E7A-F600 Page 1
Certificate of Ethics ReviewProject Title: Disaster Prediction, Management and Analysis System
User ID: 673723
Name: Hassanuddin Bin Mohamed Noor
Application Date: 27/01/2015 09:40:01
You must download your referral certificate, print a copy and keep it as a record of this review.
The FEC representative for the School of Engineering is Giles Tewkesbury
It is your responsibility to follow the University Code of Practice on Ethical Standards and anyDepartment/School or professional guidelines in the conduct of your study including relevant guidelinesregarding health and safety of researchers including the following:
• University Policy• Safety on Geological Fieldwork
It is also your responsibility to follow University guidance on Data Protection Policy:
• General guidance for all data protection issues• University Data Protection Policy
SchoolOrDepartment: ENGPrimary Role: PostgraduateStudent SupervisorName: Dr David LNdziHumanParticipants: NoPhysicalEcologicalDamage: NoHistoricalOrCulturalDamage: No HarmToAnimal: NoHarmfulToThirdParties: NoOutputsPotentiallyAdaptedAndMisused: NoConfirmation-ConsideredDataUse: ConfirmedConfirmation-ConsideredImpactAndMitigationOfPontentialMisuse: ConfirmedConfirmation-ActingEthicallyAndHonestly: Confirmed
Certificate Code: DE21-1DCB-316A-0A2A-5BFA-18FE-9E7A-F600 Page 2
Supervisor ReviewAs supervisor, I will ensure that this work will be conducted in an ethical manner in line with the University Ethics
Policy.
Supervisor signature:
Date: