PHD Zagonjolli PDF

Dam Break Modelling, Risk
Assessment and Uncertainty

Analysis for Flood Mitigation
Dam Break Modelling, Risk
Assessment and Uncertainty
Analysis for Flood Mitigation
DISSERTATION
Submitted in fulfillment of the requirements of

the Board for Doctorates of Delft University of Technology
and of the Academic Board of the UNESCO-IHE Institute for Water Education
for the Degree of DOCTOR
to be defended in public
on 24 September 2007 at 10:00 hours
in Delft, the Netherlands
by
Migena ZAGONJOLLI
born in Tirana, Albania

Bachelor of Science in Hydraulic Engineering, Polytechnic University of Tirana
Master of Science in Hydroinformatics, UNESCO-IHE
This dissertation has been approved by the promotor: Prof.dr.ir. A.E. Mynett
Members of the Awarding Committee:
Chairman Rector Magnificus TU Delft, the Netherlands

Vice Chairman Rector UNESCO–IHE, the Netherlands
Prof.dr.ir. A.E. Mynett TU Delft / UNESCO-IHE, the Netherlands, Promotor
Prof.dr.ir. H.J. de Vriend TU Delft / WL | Delft Hydraulics, the Netherlands
Prof.drs.ir. J.K. Vrijling TU Delft, the Netherlands
Prof.dr. N. Wright UNESCO–IHE / TU Delft, the Netherlands
Prof.dr. Y. Zech Université Catholique de Louvain, Belgium
Prof.dr. K. Takeuchi ICHARM Centre, PWRI, Japan
Prof.dr.ir. G. S. Stelling TU Delft, the Netherlands, Reserve Member
Copyright © 2007 Taylor & Francis Group plc, London, UK
All rights reserved. No part of this publication or the information contained herein
may be reproduced, stored in a retrieval system, or transmitted in any form or by
any means, electronic, mechanical, by photocopying, recording or otherwise,
without written prior permission from the publisher.
Although all care is taken to ensure the integrity and quality of this publication
and the information herein, no responsibility is assumed by the publishers nor the
authors for any damage to property or persons as a result of operation or use of
this publication and/or the information contained herein.
Published by A.A. Balkema Publishers, a member of Taylor & Francis Group plc.
www.balkema.nl and www.tandf.co.uk
ISBN 9 780415 455947 (Taylor & Francis Group)

Keywords: dam break modelling, flood risk reduction, flood mitigation, data
mining techniques, graph theory, constraint based approach
This thesis is dedicated to my parents, for their endless love,
encouragement and support, and for always being near me
even when I was far away.
Summary
More than 800,000 dams and thousands of kilometers of dikes have been constructed
around the world. However, the history of the construction of dams and dikes (in
this thesis referred to as ‘structures’) coexists with the history of their collapse. Hun-
dreds of dam failure events were reported the past centuries, but still today dikes
breach every year due to high water levels, often with catastrophic consequences. In
the Netherlands the storm surge of 1953 with 1850 casualties led to the construction
of the famous Delta Works. One more recent example is the breaching of the New
Orleans levee systems during hurricane Katrina in August 2005, which caused pro-
longed flooding with 1,300 casualties as well as tens of billions of dollars of economic
and social damage.
Effects of climate change are likely to cause more severe flow conditions within the
life span of existing structures, leading to increased safety concerns. If a structure
fails, the release of (large quantities of) water may threaten the lives of people as
well as property in the downstream areas. Likely loss of life depends on actual water
depth and flow velocity, the geographical distribution of the population, warning
time necessary to reach them and their awareness at the time of disaster. Warning
messages released in advance can be an important factor for saving lives. Hence,
developing and improving risk assessment and flood mitigation models is becoming
increasingly important and can be considered a necessity to reduce human casualties
and economic damage.
The aim of this research is to develop a framework and explore techniques for mod-
elling dam and dike failure events, as well as to develop novel approaches for risk
assessment. Numerical, statistical and constraint based methods are applied to
breach modelling and flood water mitigation. A new breach model (BREADA) is
developed for simulating the gradual failure of a structure due to overtopping which
is validated against historical dam failure events. In order to explore the accuracy
of breach models in a different way, and to try to extract possible additional in-
formation from available data of recorded dam failure events, we use data mining
techniques that have been successfully applied in the field of hydroinformatics.
While physically based methods require proper understanding of all processes occur-
ring during dam breaching, data mining methods rely only on recorded data of dam
failure events. In this research we apply data mining techniques including Artificial
Neural Networks and Instance Based Learning for predicting dam breach character-
istics and peak outflows. Despite the shortage of documented data, this research
demonstrates there is a possibility to improve the presently available empirical rela-
tions and prediction capabilities of physically based models by complementing with
data mining techniques.
viii Summary
The analysis of a potential failure event for existing structures (especially for large
dams) is essential for planning and organizing emergency procedures that anticipate
and mitigate downstream damages in case of disaster. In this thesis we analyse a
hypothetical failure event for the Bovilla Dam near Tirana, Albania, and explore
potential mitigation measures in case of a worst case scenario. A comparison is
made between the results of the developed BREADA model, other available breach
modelling formulations, and empirical techniques in order to get a range of peak
outflows. Flood routing is carried out using WL | Delft Hydraulics’ hydrodynamic
modelling package Sobek 1D2D. Sensitivity analysis is carried out to identify uncer-
tainties associated with dam failure analysis. The model is used to identify areas
prone to flooding, to assess the risk involved, as well as to take measures to reduce
flood damage and develop emergency plans.
Clearly, structural failure events pose a significant threat not only to human life but
also to the environment and in general also to economic development. With such
catastrophic consequences in mind, it is essential to investigate not only mechanisms
for predicting these failure events but also to reduce their risk of occurrence. Tra-
ditional approaches focus on establishing very small probabilities of occurrence of
extreme events. However, if such event were to occur - despite its low probability
value - the consequences may be very severe, like in the case of New Orleans.
In this research, instead of focusing on methodologies to minimise the failure prob-

ability of a structure, we consider an alternative approach that aims at decreasing
the consequences of a flood event. A numerical–constraint based model is developed
for evaluating risk and mitigating consequences in a system of polders or low–lying
areas. The model is capable of simultaneously evaluating different flood mitigation
scenarios in a very short time by utilizing algorithms based on ‘graph theory’. The
results of a case study which takes into account different objective functions such
as storage capacities and economical values of a multiply connected polder system,
look quite promising for flood risk mitigation.
The approach developed in this thesis can be used to complement existing practices
of flood modelling, which are traditionally carried out by simulating the consequences
of a forecasted or assumed flood event and elaborated into a few typical ‘what - if’
scenarios. The ‘lightweight’ numerical–constraint based technique proposed in this
thesis is capable of evaluating many scenarios in a very short period of time by first
determining the most feasible scenarios, which can then be modelled in more detail
using a conventional hydrodynamic simulation approach. In this way computation
time is considerably reduced while focusing on the most feasible options is ensured.
Clearly a combination of these two methods can either be achieved by enhancing a
hydrodynamic modelling package with an optional numerical–constraint based ap-
proach, or vice versa.
Samenvatting
Wereldwijd zijn meer dan 800,000 dammen en vele duizenden kilometers dijk aan-
geled. De geschiedenis van dammen en dijken gaat echter hand in hand met de
geschiedenis van hun falen. Honderden gevallen van het bezwijken van dijken hebben
zich de afgelopen eeuwen voorgedaan en ook vandaag de dag bezwijken er dammen
en dijken onder de druk van het water - met alle gevolgen vandien. In Nederland
leidde de stormvloed van 1953 waarbij 1850 mensen omkwamen, tot de aanleg van de
fameuze Deltawerken. Een recent voorbeeld van het bezwijken van waterkeringen be-
treft de orkaan Katrina in augustus 2005 die aanzienlijke overstromingen tot gevolg
had met 1300 doden en tientallen miljarden aan economische en sociale schade.
De verwachting is dat de gevolgen van klimaatverandering de komende decennia

zullen leiden tot meer bedreigende omstandigheden gedurende de levensduur van
waterkeringen wat aanleiding is tot toegenomen aandacht voor veiligheid. Immers,
als een waterkering het begeeft, dan kan dat grote waterstromen veroorzaken die een
gevaar vormen voor mens en haard benedenstrooms. Eventueel verlies aan mensen-
levens hangt af van waterdiepte en stroomsnelheid, waarschuwingstijd en de aan-
wezigheid van bewoners ten tijde van een eventuele ramp. Tijdige waarschuwing
speelt een belangrijke rol bij het voorkomen van verlies aan mensenlevens. Het
ontwikkelen en verbeteren van modellen voor risicobenadering en het afwenden van
gevolgen van overstromingen, wordt steeds belangrijker bij het toetsen van de sterkte
van reeds bestaande waterkeringen.
Het doel van dit onderzoek is om een raamwerk op te zetten en technieken te on-
twikkelen voor het modelleren van het bezwijken van waterkeringen. Ook worden
mogelijke nieuwe benaderingen op het gebied van risico analyse onderzocht. Naast
numerieke en statistische benaderingen zijn ‘constraint based’ technieken ontwikkeld
voor het modelleren van dijkdoorbraken en het afwenden van overstromingsgevaar.
Er is een nieuw model voor dijkdoorbraak (BREADA) ontwikkeld voor het geleidelijk
bezwijken van waterkeringen ten gevolge van het ’over’stromen. Dit is gevalideerd
op basis van beschibare veldmetingen. Om de nauwkeurigheid van eerdere modellen
na te gaan en om eventuele nieuwe informatie aan de beschikbare meetgegevens te
onttrekken, zijn ‘data mining’ technieken gebruikt die al geruime tijd in de hydroin-
formatica met succes zijn beproefd op toepassingen in waterbouw en waterbeheer.
Op fysica gebaseerde modellen vragen om een goed begrip van onderliggende pro-
cessen om een wiskundige beschrijving te kunnen opstellen; deze wordt vervolgens
getoetst aan de hand van beschikbare meetgegevens. Data mining technieken gaan
uit van dezelfde meetgegevens, maar proberen het fysisch proces hieraan te ont-
trekken zonder vooraf een bepaald model op te leggen. In dit onderzoek worden
data mining technieken gebruikt om de maximale waarde van uitstroom bij dijkdoor-
x Samenvatting
braak te voorspellen. Ondanks de beperkte hoeveelheid gedocumenteerde gegevens

laat dit onderzoek zien dat het mogelijk is om de bestaande empirische modellen te
verbeteren.
Een adequate analyse van een mogelijke catastrofale gebeurtenis als een grote damdoor-
braak is essentieel om noodmaatregelen voor te bereiden die de benedenstroomse
gevolgen van een ramp kunnen beperken. In dit proefschrift wordt een analyse uit-
gevoerd naar de een eventuele doorbraak van Bovilla Dam bij Tirana in Albanie, en
worden mogelijke maatregelen onderzocht om een ramp af te wenden. Resultaten
van BREADA en andere modellen worden gebruikt om de variatie in uitstroomcon-
dities na te gaan. Met WL | Delft Hydraulics’ Sobek 1D2D numerieke modelsysteem
worden overstromingen gesimuleerd en gevoeligheidsanalyses uitgevoerd om onzek-
erheden vast te stellen. Voor mogelijk bedreigde gebieden worden noodmaatregelen
onderzocht.
Het is duidelijk dat het falen van waterkeringen levensbedreigend kan zijn met grote
economische gevolgschade. Daarom is het noodzakelijk om niet alleen een faalkans
vast te stellen, maar ook om mogelijke gevolgen te kunnen beperken. Immers, mocht
een faalkans worden overschreden, dan kunnen de gevolgen verstrekkend zijn, zoals
bijvoorbeeld bij New Orleans. In dit proefschrift wordt daarom een alternatieve route
bewandeld met de nadruk op het verkleinen van de gevolgen van overstromingen.
Daartoe is een numeriek model gecombineerd met een ‘constraint’ aanpak gebaseerd
op ‘graph theory’ wat het mogelijk maakt om meerdere scenario’s in zeer korte tijd
te evalueren. De resultaten van een toepassing in een gebied met meerdere polders
met verschillende doelfuncties zoals bergingscapaciteit en economische waarde (ook
wel systeemwerking genoemd) zien er veelbelovend uit.
De benadering die in dit proefschrift is gekozen kan gebruikt worden als aanvulling op
de huidige aanpak die veelal is gebaseerd op het simuleren van een groot aantal mo-
gelijkheden waarvan pas later blijkt of deze relevant zijn of niet. De hier voorgestelde
‘lichte benadering’ van gecombineerde ‘numerical–constraint based’ technieken is
flexibel genoeg om zeer veel mogelijkheden in korte tijd te evalueren en om de meest
waarschijnlijke scenario’s te selecteren die vervolgend met behulp van een conven-
tionele hydrodynamische aapak in meer detail kunnen worden onderzocht. Op deze
manier wordt de rekentijd sterk gereduceerd en alleen gebruikt om de meest waarschi-
jnlijke opties te onderzoeken. Een combinatie van beide methoden is mogelijk door
bestaande hydrodynamische pakketten uit te breiden met opties voor een ‘numeri-
cal–constraint based’ benadering, of vice versa.
Acknowledgment
This dissertation would not have been possible without the continuous encourage-
ment of my promotor Prof. Arthur Mynett. I am very grateful to him not only for
directing my research, but also for the invaluable moral support I received through-
out this project. His guidance and supervision enabled me to complete my work
successfully.
I would like to acknowledge WL | Delft Hydraulics for their financial support and
for providing the professional environment to carry out this research. Many thanks
go to all my friends and colleagues at UNESCO-IHE, Delft University of Technology
and WL | Delft Hydraulics: your friendship and professional collaboration meant
a great deal to me. I am grateful to Dr. Hans Goossens and Dr. Henk van den
Boogaard of the Strategic Research Department of WL | Delft Hydraulics for con-
tributing their time and expertise to this project. Special thanks go to Prof. Roland
Price and Prof. Dimitri Solomatine of UNESCO-IHE for the informal and fruitful
discussions.
I would like to express my gratitude to the thesis committee members for their in-
terest and valuable comments on my work and to Prof. Nigel Wright for helping
improve the thesis with his useful observations and suggestions.
This list of acknowledgments would not be complete without all the people to whom
I am indebted at a personal level. My friends and relatives have provided invaluable
moral support during the four years of this research. I am very grateful to each and
every one of you. I wish to express my special appreciation to my dear friend Merita
Hatibi Serani who introduced me to the field of hydraulic engineering. Although she
untimely passed away, the memories and creative energy she left behind are always
with me.
I am as ever, especially grateful to my family. To my brothers for their love and

encouragement. To my mother for her continuous support in my objective of enrich-
ing my knowledge, despite the pain of being away from her. To the happy memory
of my father, who always believed in me but unfortunately did not live to see this
thesis being completed, but who provides a persistent inspiration for my journey in
life.
Migena Zagonjolli,
24 September 2007
xiii
Contents
Summary vii
Samenvatting ix
Acknowledgment xi
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Management of the flood risk caused by structural failure . . . . . . 4
1.3 Scope of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Water Retention and Flood Defence Structures 9

2.1 Design criteria and failure modes for dams . . . . . . . . . . . . . . . 9
2.2 Design criteria and failure modes for dikes . . . . . . . . . . . . . . . 13
2.3 Breach modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.1 Breach shape . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3.2 Breach development . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.3 Breach formation mechanisms . . . . . . . . . . . . . . . . . . 21
2.3.4 Hydraulics of flow over the dam . . . . . . . . . . . . . . . . . 23
3 Current Approaches for Dealing with Flooding 25

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 The Netherlands’ long history of battle against floods . . . . . . . . 26
3.3 Japan’s experience in dealing with floods . . . . . . . . . . . . . . . . 34
3.4 Modelling of propagation of flood caused by structural failure . . . . 36
3.5 Uncertainty associated with flood modelling . . . . . . . . . . . . . . 41
4 Current Approaches for Breach Modelling 43

4.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2 Experimental work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3 Empirical equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
xiv Contents
4.4 Physically based models . . . . . . . . . . . . . . . . . . . . . . . . . 50
5 Data Mining Techniques in Dam Breach Modelling 57

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.2 Main principles of data mining techniques . . . . . . . . . . . . . . . 58
5.2.1 Linear regression . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.2.2 Artificial Neural Networks (ANNs) . . . . . . . . . . . . . . . 60
5.2.3 Instance Based Learning (IBL) . . . . . . . . . . . . . . . . . 64
5.3 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.1 Data preparation and model evaluation . . . . . . . . . . . . 67
5.3.2 Predicting breach width based on water depth behind the dam 69
5.3.3 Predicting peak outflow . . . . . . . . . . . . . . . . . . . . . 71
5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
6 BREADA model: Breach Model for Earthfill Dams 75

6.1 Development of a dam breach model . . . . . . . . . . . . . . . . . . 75
6.1.1 First formulation of dam breaching development . . . . . . . 79
6.1.2 Second formulation of dam breaching development . . . . . . 80
6.2 Validation of the BREADA model . . . . . . . . . . . . . . . . . . . 81
6.3 Failure analysis of an earthfill dam . . . . . . . . . . . . . . . . . . . 85
6.3.1 Modelling of breach development . . . . . . . . . . . . . . . . 89
6.3.2 Modelling of flood propagation . . . . . . . . . . . . . . . . . 95
6.4 Uncertainty in dam break analysis . . . . . . . . . . . . . . . . . . . 98
6.5 Discussion and conclusions . . . . . . . . . . . . . . . . . . . . . . . . 101
7 A Numerical-Constraint based Model 103

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.2 Description of the numerical–constraint based approach . . . . . . . 104
7.3 Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
7.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
8 Conclusions and Recommendations 119

8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.1.1 Traditional and novel approaches for predicting dam breach
characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
8.1.2 Identification of dam failure flood extent, uncertainty, and risk
reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
8.1.3 A numerical–constraint based approach for flood mitigation . 122
8.2 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Bibliography 125
List of symbols 139

Contents xv
List of abbreviations 141
Curriculum vitae 143

Chapter 1
Introduction
If any one be too lazy to keep his dam in proper condition, and
does not so keep it; if then the dam break and all the fields be
flooded, then shall he in whose dam the break occurred be sold
for money, and the money shall replace the corn which he has
caused to be ruined.
The Code of Hammurabi (18th century BC)
1.1 Background
The history of water defence and water retention structures coexists with the his-
tory of their failures. Around the world thousands of dams have been constructed
over many centuries. But also, hundreds of dams have failed and every year many
dikes breach due to high flows in the rivers, sea storm surges, etc. often leading to
catastrophic consequences. By far the world’s worst dam disaster occurred in Henan
province in China, in August 1975, when the Banqiao Dam and the Shimantan Dam
failed catastrophically due to the overtopping caused by torrential rains. Approx-
imately 85,000 people died from flooding and many more died during subsequent
epidemics and starvation; millions of residents lost their homes (Qing, 1997). This
catastrophic event is comparable to what Chernobyl and Bhopal represent for the
nuclear and chemical industries (McCully, 1996). In the Netherlands, in February
1953, a high–tide storm caused the highest water levels observed up to date and
breached the dikes in more than 450 places, causing the death of nearly 1,900 people
as well as enormous economic damage (Gerritsen, 2005).
During recent years, the observation is made that the global impact of climate change
will be devastating (see e.g. Intergovernmental Panel on Climate Change (IPCC) as-
sessment reports at www.ipcc.ch; Millennium Ecosystem Assessment (2005); Archer
(2006); Flannery (2006); Wentz et al. (2007)). According to the IPCC, “the ob-
served warming trend is unlikely to be entirely natural in origin”. We witness the
Climate Change 2001: The Scientific Basis Page 1 of 2
Climate Change 2001:

Working Group I: The Scientific Basis
2 Table of contents Go Section 1.1. Background
Other reports in this collection
Figure Figure 11.12: Global

1.1: Global averagesea
average sea level
levelriserise
1990for
to 2100
theforperiod
the SRES1990
scenarios. Thermal
to 2100 expansion
for different
and land ice changes were calculated using a simple climate model calibrated separately for each of
scenarios (IPCC,
seven AOGCMs, 2001).
and contributions from changes in permafrost, the effect of sediment deposition and the
long-term adjustment of the ice sheets to past climate change were added. Each of the six lines
appearing in the key is the average of AOGCMs for one of the six illustrative scenarios. The region in
dark shading shows the range of the average of AOGCMs for all 35 SRES scenarios. The region in light
shading shows the range of all AOGCMs for all 35 scenarios. The region delimited by the outermost
rising of temperatures,
lines shows the range ofwhich in itself
all AOGCMs does not
and scenarios always
including meanin land-ice
uncertainty nicer sunny
changes, weather
permafrost
but more likelychanges and sediment
enhances deposition.increase
the potential Note that this
of range does not allow
the intensity offor uncertainty
flash floods, heat
relating to ice-dynamical changes in the West Antarctic ice sheet. See 11.5.4.3 for a full discussion. The
waves, bars
mudslides and droughts, leading to catastrophic social,
show the range in 2100 of all AOGCMs for the six illustrative scenarios environmental and
economic damages. The declining ice extent in Arctic, the increase in melting rate
of Greenland ice sheet, and rising globalBack to text
mean sea level (Figure 1.1) are the most
evident consequences of global warming. The number of high category hurricanes
Table of contents
has increased,
Other reports in while the first ever hurricane was recorded in 2004 in the South At-
this collection
lantic (South Coast of Brazil).
The failure of the New Orleans’ levee system during hurricane Katrina of
http://www.grida.no/climate/ipcc_tar/wg1/fig11-12.htm 2005
6/2/2007
contributed to prolonged flooding causing 1,300 casualties and billions of dollars
of economic damage. In August 2002, floods caused by torrential rains in Europe
claimed the lives of at least 109 people. The Elbe exceeded its 200 year flood return
period in Dresden, flooding much of the city, and a flood return period of 500 years
was estimated in Prague.
Nowadays there are more than 800,000 dams and thousands of kilometers of dikes,
constructed around the world for different purposes: flood control (the most com-
mon purpose), irrigation, electricity generation, water supply, recreation, etc. Dams
Chapter 1. Introduction 3
and dikes are often designed based on the statistical distribution of recorded river
flows or sea storm surge measurements. However, climate change has introduced
uncertainty related to increasing maximum flows within the life span of dams and
dikes, leading to safety concerns. Many dams and dikes previously considered ade-
quate (safe) now exhibit a potential to experience overflowing (overtopping) during
high (flash) flood events. If a dam or dike fails, loss of life and economic damage are
direct consequences of such an event, depending on the magnitude of water depth
and velocity, warning time, and presence of population at the time of the event.
Early warning is crucial for saving lives in flood prone areas.
Costa (1985) compared the loss of life during two dam failure events: the Teton
Dam in Idaho (93m high) that failed at midday on 5 June 1976 and Laurel Run
Dam in Pennsylvania (12.8m high) that failed on 20 July 1977 at 4:00 a.m. The
latter event claimed the lives of 1 out of every 4 people exposed to floodwaters, while
only one out of 3,000 people exposed to floodwaters suffered from casualties in the
case of the Teton Dam failure. The database compiled by the Centre for Research
on Epidemiology Disasters (www.em-dat.net) identifies floods as the second most
endangering factor after drought in terms of loss of human lives around the world
for the period 1900-2007.
When the failure of a tailing dam happens, the socio–environmental implications

might affect the flooded region for a long period of time. The dam failure at the
Los Frailes mine in Spain, in April 1998, released between 5 to 7 million tonnes
of toxic mud carrying heavy metals and highly acid compounds in the Guadiamar
River spreading in large areas of the floodplains. Large scale pollution of the Doñana
World Heritage nature park, one of Europe’s primary wildlife sites, was narrowly
avoided thanks to an emergency dike constructed by the Spanish authorities. How-
ever, the environmental disaster was immense and its long–term consequences are
feared to affect the region for many years to come (Olı́as et al., 2005).
The construction of dams and dikes (hereafter referred to as structures) leads people
to believe that the floods are fully controlled, and therefore an increased urban and
industrial development in the floodplains usually takes place. Hence, if the struc-
ture fails, the damage caused by flooding might be much greater than it would have
been without the structure’s presence. Having the historical failures of structures
in mind, one might pose the question what can be done in order to reduce the risk
posed from a dam or dike failure event.
4 Section 1.2. Management of the flood risk caused by structural failure
1.2 Management of the flood risk caused by struc-

tural failure
The traditional approach of preventing the impact of flood through flood protec-
tion, is more recently being replaced by the flood management approach (Mynett
and de Vriend, 2005; de Vriend, 2005; Samuels et al., 2005; Simonovic and Ahmad,
2005) as a result of the recognition that absolute flood prevention is unachievable
and unsustainable, due to high costs and inherent uncertainties. The purpose of
flood risk management is to protect the people facing risk up to a certain accept-
able (affordable) level, and to reduce the consequences of an extreme event exceeding
the acceptable level, so that a disaster can be avoided. Thus, risk management is
defined to be the process of assessing and reducing the risk. Assessment of the
risk involves recognizing the plausible failure modes for a structure or the plausi-
ble flooding events, quantifying probabilities and consequences (socio-economic and
environmental) for all (or only plausible) failure modes or flooding events, and evalu-
ating the risk by comparing the posed risk to the predefined risk criteria (acceptable
or non acceptable risk level). Usually a threshold criterion is applied that estab-
lishes a level over which risk is considered unacceptable. International Commission
on Large Dams (2005) defines the tolerable risk as follows:
A risk within a range that society can live with so as to secure certain
net benefits. It is a range of risk that we do not regard as negligible or
as something we might ignore, but rather as something we need to keep
under review and reduce it still further if and as we can.
The key approach in achieving tolerable risk is reducing risk as low as reasonably
practicable. It is usually defined for each structure or system of structures (dike
rings) rather than as a general criteria. It is based on a case specific evaluation
of all possible risk measures, and different criteria might be applied for life safety
and other consequences in different countries. A tolerable level of life risk is often
evaluated in reference to risk of loss life due natural hazards or disease (Shortreed
et al., 1995).
Different methods can be applied for reduction of the risk posed by a dam failure
according to Bowles (2001):
1. Avoid the risk before or after the dam is built. If the safety of the dam is
questionable, then decommissioning of the dam can be proposed as a solution
to the problem.
2. Reduce failure probability occurrence through structural and non–structural
measures.
3. Reduce (mitigate) consequences through transferring the risk, effective emer-
gency evacuation planning or relocation of population at risk. The ‘Peace
Dam’ is constructed about 125 kilometers northeast of Seoul, South Korea for
the purpose to mitigate the flood water in case of collapse of North Korea’s Im-
nam or Mount Geumgang Dam or sudden release of water through the outlet
works.
4. Retain (accept) the risk, but protection measures should be taken in the down-
stream area against the flood water as well as other measures should be applied
to mitigate the flood water from the most populated or economically valuable
areas.
The risk associated with flooding is generally expressed as the product between the
probability of the event occurrence and the monetary value of its consequences.
Expressing loss of life in monetary terms usually is not morally acceptable, thus,
the population at risk is commonly taken into account as a decisive element in
risk analysis. Risk expression might lead to the equivalence of an event with low
probability and high consequences to an event that has high probability of occurrence
but very low consequences. Generally, risk reduction measures try to reduce the
probability of flooding, though minimizing the probability of a flood might come at
the price of increasing its destructive power. The Indian Ocean tsunami of December
26, 2004 demonstrated that the consequences should be a triggering element of any
risk analysis, instead of focusing on the probability of the event only. The impact of
the Asian tsunami could have been lower in terms of human live losses if an effective
warning system would have been implemented and operated.
The total failure probability of a structure involves the combination of individual

probabilities estimated for different factors and loading conditions that contribute to
its failure (Hartford and Baecher, 2004). Correctly defining the overall probability,
while taking into account all factors that could lead to the failure of the structure
is questionable, especially when extreme probabilities are deduced from small data
samples. A new extreme flow record or new development at the areas potentially
affected by the structure failure will make the calculated probability value outdated.
Furthermore, in most cases the failure modes are not independent and therefore the
failure probabilities are not simply additive.
Nowadays, as a result of economic development and population growth in flood

prone areas, the potential flood damage increases as well. Furthermore, climate
change caused by global warming might lead to more devastating flooding than
ever. Higher river dikes are now seen more as a contributor to major flooding than a
protection against it, in addition to having negative effects on nature, landscape and
cultural heritage of the surrounding areas. In the Netherlands, the standard policy
of raising dike crest levels in order to maintain the required level of flood protection
is being abandoned in favour of the ‘Room for the River’ policy: widening river cross
sections by relocating dikes further away from the river and/or by lowering the river
floodplains, constructing floating houses, etc. (van Schijndel, 2005).
6 Section 1.3. Scope of the thesis
Unfortunately, we might fail to stop an event from happening due to our incapability,
negligence or unpredictability of extreme natural events, but we can develop models
and tools for fast response to any event in order to reduce its consequences. The
aim is to find measures that reduce the probability of flooding and minimize the
potential consequences.
1.3 Scope of the thesis

Developing and improving flood propagation, risk assessments and flood mitigation
models for already constructed dams and dikes is becoming a necessity for a variety
of reasons such as decreasing human casualties and economic damage. In this thesis,
instead of focusing on methodologies to estimate and lower the failure probability of
hydraulic structures, we propose approaches that cope with hazards caused by struc-
tural failure events by decreasing their consequences. We consider events, though
not likely to happen in any given year, if occurring are extremely catastrophic and
have enormous socio–economic impact.
We address the problem of dam and dike breach analysis as well as simulation and
mitigation of the flood water caused by failure of these structures. Formation of
the breach in a structure is a complex process that depends on various hydraulic,
hydrologic, geotechnical factors. In this thesis, we develop a framework and tech-
niques for modelling dam and dike failure events as well as propose several novel
approaches for dam breach modelling. Furthermore, we introduce and apply several
numerical, statistical and constraint based methods in particular related to dam and
dike breach modelling and flood water mitigation.
In this thesis, a ‘lightweight’ numerical–constraint based technique is proposed. This

technique offers advantages of simultaneous evaluation of different flood mitigation
scenarios. Through constraints we optimize the strategy for choosing the most fea-
sible flood propagation scenario that minimizes economic consequences.
The objectives of this thesis can be summarized as follows:

1. Review and comparison of different existing methods (mainly physically and
statistically based) for dam and dike failure modelling.
2. Development of a dam breach model for breach formation in earthfill dams
and its validation against a real dam failure event.
3. Development of a methodology for predicting dam breach characteristics and
peak outflow during the failure event by using statistical and data mining
techniques.
4. Development of an approach for reduction of flood consequences caused by a
dam failure event, identification of areas prone to flooding, identification of
uncertainties (input, model, and completeness uncertainty), and proposal of

risk reduction measures.
5. Development of a numerical constraint based model for flood mitigation in

low–lying areas subject to flooding.
1.4 Outline of the thesis

The thesis is composed of eight chapters. Chapter 2 presents a literature review on
earthfill dams and dikes, focusing on their failure analysis and the processes involved
during the breaching of a structure.
Chapter 3 describes in detail current state–of–the–art approaches used for dealing

with floods in particular in the Netherlands and Japan. Furthermore, it provides
in depth overview of the applied methods, namely their characteristics, advantages,
and disadvantages.
Chapter 4 compares the theoretical and practical aspects of available approaches

and/or models used for breach modelling of water defence and water retention struc-
tures.
Chapter 5 introduces a new approach for the estimation of dam breach character-
istics. It starts with an overview of the theory behind the data mining techniques
used in this research and demonstrates their application to the dataset of dam failure
events.
Chapter 6 presents the mathematical model of a dam breach tool developed during
this research as well as its application to the modelling of a hypothetical failure of
an earthfill dam. It provides an evaluation of breach characteristics using different
approaches, a sensitivity analysis of important breach parameters and an evaluation
of uncertainties.
Chapter 7 describes the development of a numerical–constraint based model which

is capable to simulate different flood mitigation scenarios taking into account the
social and economic value of areas that could be prone to inundation.
Chapter 8 presents the conclusions of the research and future recommendations.

Chapter 2
Water Retention and Flood

Defence Structures
Engineering is the professional art of applying science to the
optimum conversion of natural resources to the benefit of man.
Ralph J. Smith
2.1 Design criteria and failure modes for dams

A dam is a barrier made of earth, rock, or concrete or a combination thereof that
is constructed across a river for impounding or diverting the flow of water. The
history of dam construction dates back to 2900 B.C. with the oldest dam in the
world believed to be constructed in Wadi el–Garawi, 30 km south of Cairo, Egypt
(Singh, 1996). The dam was built for irrigation purposes and had a crest length of
106m and a maximum height of 11.3m. It collapsed the first winter it was in use, but
its remains are still present today. The Alicante Dam in Spain was 46m high when
completed in 1594 and remained the world’s highest dam for 300 years. Currently,
hundreds of high dams are in operation worldwide. The Nurek Dam in Tajikistan
(300m high) is the highest and the Three Gorges Dam in China with a reservoir
storage of 39.3 billion m3 is the largest dam in the world.
The construction of dams is often seen as a solution for providing water supply, flood
control or for ‘green’, renewable electricity. The dams must be designed, built, and
operated so that they make a positive contribution to socio–economic development,
while having minimal impact on the environment. There are, however, different
perspectives on this issue and discussions are taking place about the controversial
impacts of dam construction. Resettlement of people that have to withdraw from
their social and cultural identity is often seen as a major impact of dam construction,
especially for large dams. Irreversible degradation of natural habitat, degradation of
10 Section 2.1. Design criteria and failure modes for dams
water quality, sedimentation of reservoirs and the downstream effects are other neg-
ative impacts of dam construction. The reader is referred to the report by the World
Commission on Dams (2000) for a detailed description of the complex impacts of
dams. To alleviate the impacts, measures are usually taken for adequate mitigation
of the natural habitat e.g. by creating nearby protected areas for wildlife, or new
fisheries within reservoirs, etc. The dam proponents propose more hydropower dam
construction to play “a major role in reducing greenhouse gas emissions in terms of
avoided generation by fossil fuels” (Lafitte, 2001). On the opposite side, environ-
mentalists insist that any large dam, including hydropower dams, emits greenhouse
gases (GHGs) due to the rotting of the flooded organic matter. However, the science
of quantifying GHGs reservoir emissions is uncertain and controversial conclusions
are drawn related to hydropower emissions in comparison to those from fossil fuels,
viz. the difference between the pre–dam emissions from the undamed catchment and
the post–dam emissions. In this thesis we focus on a dam as a structure, while its
construction impacts - though very important - are out of the scope of this thesis.
The interested reader is referred to e.g. Galy-Lacaux et al. (1999); Rosa and dos
Santos (2000); Soumis et al. (2005); Tremblay et al. (2004).
Constructed dams can be categorized in two large groups: gravity and arch dams
(Figure 2.1). Gravity dams rely on their weight to resist the forces imposed upon
them. Arch dams, with the arch pointing back into the water, use abutment re-
action forces to resist the water pressure force. They can be made of concrete or
masonry. Gravity dams can consist of concrete or earth, rock, a mixture of these
materials, or masonry. Dams are designed to have a low probability of failure during
their construction and operation life span. Dam design criteria require the dams
to withstand different loads, namely construction and reservoir water load, with or
without seismic load. The greater the chance of loss of life or damage to valuable
property in case of the failure event, the safer the design should be. Despite this,
dams do frequently fail.
The failure of a structure can be partial or complete. The failure of the structure
to fulfill its purpose is another type of failure. In this thesis, the term ‘dam failure’
indicates the partial or complete collapse of the dam or its foundation, leading to
uncontrolled release of water in the downstream areas. Landslides in the reservoir
might cause the release of water in the downstream areas, despite no failure of the
dam structure occurs. A wave estimated to be 100m high overtopped the Vaiont
arch dam in Italy in 1963 when a massive rockslide of 240 million m3 fell into the
reservoir at a velocity of approximately 30m/s (Pugh and Harris, 1982). Only minor
damage to the dam crest was observed though the wave that reached a height of
70m at 1.6km downstream the dam caused the loss of 2,600 human lives. Failure
of a dam can be sudden or gradual. A sudden failure is associated with concrete
Chapter 2. Water Retention and Flood Defence Structures 11
Cross section
Cross section Plan view
Plan view Cross
Cross section
section Planview
Plan view
Upstream
Upstream slopeslope
Upstream
Upstream
Downstream slope Downstream

Downstream slope Downstream
abutment
abutment
Toe
Toe
abutment
abutment
(a) (b)
Figure 2.1: Schematic picture of (a) gravity and (b) arch dams.
dams, gravity or arch dams∗ . If breaching is initiated, the further development is

faster than for earthfill dams under the same conditions. This observation allows
modelling of the concrete dam failure events simply as a sudden (gate opening) pro-
cess. Rockfill and earthfill dams, termed embankment dams, constitute the largest
percentage of constructed dams around the world and not surprisingly the largest
number of dam failure accidents occur with these dams, particularly with earthfill
dams. Their failure, depending on the triggering factors, is mostly a gradual process
rather than a sudden one. In this thesis we investigate the failure of embankment
dams, in particular of earthfill dams.
Failure of an earthfill dam can be triggered by different factors, e.g. overtopping,

foundation defect, seepage and piping. Overtopping is one of the most common
failure modes for earthfill dams. It can be triggered by inflows higher than the de-
sign inflow, malfunctioning or a mistake in the operation of the spillway or outlet
structure, inadequate carrying capacity of spillways, settlement of the dam or as a
result of landslides into the reservoir. According to National Performance of Dams
Program in USA (NPDP, 2007), 245 of 256 dam failure events recorded in the USA
during the year 1994 happened due to high inflow discharges. Any embankment
dam will fail if the spillway capacity is too small and flood waters rise high enough
to flow over the top of the dam for a considerable amount of time. In August 1979,
a flood two to three times larger than the design flood triggered the failure of the
Machhu II dam in India, causing more than 2,000 casualties (Hagen, 1982). Once
∗ The Malpasset Dam in southern France, an arch dam of 66.5m in height and maximum designed
reservoir capacity of 55 million m3 , is described to have failed explosively on 2 December 1959

(Hervouet, 2000).
12 Section 2.1. Design criteria and failure modes for dams
Reservoir water surface
Pervious foundation
Flow line
Impervious layer
Figure 2.2: Seepage on the foundation of a dam.
an initial breach channel is created, and the high reservoir water levels persist, the
breaching will continue to develop and any effort made to stop it will be unsuccess-
ful. Overtopping may not result in structural failure, but still presents a major flood
hazard as happened during the overtopping of the Vaiont Dam in Italy (Pugh and
Harris, 1982). Similarly, rapid release of reservoir water in order to lower the water
level within safe limits, can be a big concern in downstream areas.
Water penetrating through the dam’s interior body or its foundation might progres-
sively erode soil from the embankment or its foundation leading to the failure of the
dam. Here, we define piping failure as a failure mode caused by water penetrating
through the dam’s body, carrying with it small particles of dam material, continu-
ously widening the gap. If the initial piping can be detected before it reaches the
critical condition, remedy might be possible. Penetration of water in the dam body
can cause slope failure. To prevent this type of failure, appropriate instrumentation
is needed to estimate the rate of infiltration within an embankment.
Seepage failure (Figure 2.2) or foundation failure occurs due to the saturation of
the foundation material leading to either washout of the material or a weakening of
the rock towards a sliding failure. The flow of water through a pervious foundation
produces seepage forces as a result of the friction between the percolating water and
the walls of the pores of the soil through which it flows. Figure 2.2 shows how water
flows through the pervious foundation of a dam.
Earthquakes or sabotage are yet two other causes of dam failure. Earthquakes that
have stimulated immediate failure of a dam appear to be very rare. The upstream
slope of the Lower San Fernando Dam in California (USA) failed due to liquefaction
during the earthquake in 1971. The dam was constructed by fill soil mixed with a
large amount of water, transported to the dam site by pipeline, deposited on the em-
bankment in stages, allowing the excess water to drain away. The fill that remained
was loose, and was subject to liquefaction as a result of the earthquake. Fortunately,
the reservoir level was low at the time of the earthquake and no flooding occurred.
Failure due to an earthquake might result in a higher threat to the population down-
stream rather than the overtopping failure. In the first case a sudden breach of the
dam would cause a flood wave moving downstream while the population might have
no clue about the structural failure of the dam and the approaching flood water. On
the other hand, prior to an overtopping failure, rising flood waters often give reason
for concern to the residents in the floodplain area and lead to issuing flood warning.
The historical database of dam failure events shows that the number of failures
caused by sabotage is small. One example is the British bombardment of the Ger-
man dams on the Ruhr River during World War II. The Dnieprostroy Dam (43m
high) on the Dnieper River was also destroyed for the purpose of preventing the
movement of German troops.
2.2 Design criteria and failure modes for dikes

A dike is a barrier built along the shore of a sea or lake or along a river with the
objective of holding back water and preventing flooding (Figure 2.3). Dikes are often
constructed in the floodplain for the purpose of protecting from flooding. They are
usually built from sand, clay or a combination of them, or from peat (like in many
places in the Netherlands, where local soils were used for construction).
There are many aspects that are important in making decisions related to dike con-
struction. Economic, environmental and other social interests have to be considered,
involving different parties in the process of decision making. Different methods exist
for the design of flood defence systems or water impounding structures: probability
and risk–based design methods.
In a probabilistic approach, dikes are designed based on a water level with a partic-
ular frequency of being exceeded. The design flood levels as well a safety level or a
margin ensure the dikes’ integrity. Depending on the probabilistic method used in
the analysis and on the available data records, the design levels might be different.
Thus, including uncertainty in the estimation of the design levels is necessary, and
the margin can deal to some degree with the small uncertainties. During the years
the design levels might change as a results of changing flow conditions or as a result
of development in the area protected by dike. Therefore different measures are im-
plemented to ensure the structural integrity that take into account new developing
conditions.
Dikes might fail due to different reasons and triggering factors. The common failure
modes that are similar to the dam failure modes are piping, seepage, overtopping
caused by sabotage is small. One example is the British bombardment of the
German dams on the Ruhr River during World War II. Other factors involve
dynamiting of dams through military action. Russian engineers realised in 1941 the
destruction of the Dnieprostroy Dam (43m high) near Kichkas on the Dnieper River.
Dikes and their failure modes
***The words dike and polder came from Dutch***Dikes are usually constructed
along the river banks and sea shores for the purpose of protecting an area from
flooding. Dikes are classified based on their function and construction material.
14 They are usually built from Section
sand, clay,
2.2.and combination
Design criteriaof them
and or from
failure peat.for
modes Sea
dikes
dikes have a seaward slope in order to reduce the wave run-up and the erodible
effect of the waves.
(a) (b)
(c) (d)
Figure 2.3: Dike along a (a) coastline, (b) river, (c) channel, and (d) on the flood-
Figure 2.3. Dike along a channel and on the floodplain.
plain.
In the Netherlands, there are 53 areas enclosed by continuous system of dikes, high
ground areas and other hydraulic structures (gates, sluices, locks, pumping stations
etc.). These areas are called dike ring areas and for each of them the flooding
12
Figure 2.4: Possible failure modes for dikes (TAW, 1998).

Figure 2.5: Failure of a canal dike in Wilnis, the Netherlands (2003).
(overflowing), uplift, and slope failure. For a typical dike, various potential failure
mechanisms are distinguished as shown in Figure 2.4. Water is not the only one
triggering factor of dike failure. In August 2003, after a very dry and warm summer,
one section of a peat dike (Figure 2.5) constructed along a canal in the Netherlands
suffered a horizontal sliding, causing flooding of about 600 houses and evacuation
of 2,000 people. In specific cases, the integrity of a dike might be affected by the
activity of burrowing animals (field mouse, gopher, mole and fox). In case of their
presence, actions are usually taken to control their activity through wire screening
or traps placing along the structure.
The dike failure probability is calculated on the basis of a so–called reliability function
Z. If the resistance R is determined and also the load S is known, the determination
whether the structure will fail is simply Z = R − S. Here load and strength are
both stochastic variables. If we consider the stochastic nature of the strength and
load, the probability that the construction will fail is the probability of P {Z < 0}.
If Z = 0, the limit state is reached, which constitutes the failure boundary.
Calculating the failure probability is as complex as modelling the dike failure pro-
cesses. The factors affecting accurate estimation and modelling are:
Variation of the properties of structural material (strength, deformation, per-

meability, time–dependent consolidation, compaction of the material, different
quality of construction works, length of the dike).
Lack of a reliable mathematical description of the failure processes needed for

16 Section 2.3. Breach modelling
accurate determination of the limit states. As for overtopping and overflowing,

these processes are understood with some degree of accuracy, but much less is
known for other failure modes, such as piping.
Uncertainty in estimating the overall failure probability of the dike. The total
failure probability of a structure involves the combination of individual proba-
bilities estimated for different factors and loading conditions that contribute to
its failure. Correctly defining the total probability, while taking into account
all the factors that could lead to the failure of the structure seems question-
able. In most cases the failure modes are not independent and therefore the
failure probabilities are not simply additive.
For the probabilistic safety analysis to be meaningful the accurate computational

models and sufficient (statistical) data are needed. Different techniques are avail-
able for determining the failure probability given a reliability function and statistical
characteristics of the basic variables. The uncertainties associated with them “are
often even greater than the uncertainties due to the intrinsically stochastic charac-
ter of load and strength” (CUR/TAW, 1990). As a result, the calculated value of
failure probability is often used as a relative indicator for implementing priorities in
maintenance of the dike ring system. For detailed analysis the reader is referred to
CUR/TAW (1990) and Thoft–Christensen and Baker (1982).
Due to the shortcomings of the probabilistic approach (CUR/TAW, 1990), the risk
analysis approach is drawing more attention. The risk based design approach con-
siders the probability as well as the consequences of inundation in case of a failure
of a flood defence system. Breaching of the dikes and consequent flooding have
claimed many lives and caused enormous economical damages worldwide. Based on
the risk estimation method the magnitude of the damage or loss is considered during
the design of the dike. For detailed description see CUR/TAW (1990) and Vrijling
(2001).
In this thesis we only consider the overflowing and overtopping failure modes that are
the most common failure modes for embankment dams and dikes. In the following
section we give a general description of the breach development characteristics and
emphasize the differences between modelling of the breaching processes in earthfill
dams and dikes.
2.3 Breach modelling

Breaching of a structure is a time–dependent and non–linear phenomenon. Wa-
ter–soil interaction together with often non–homogeneous and specific material prop-
erties for each structure lead to the difficulty of accurately modelling the processes
involved in breach development. Hydrodynamics, sediment transport mechanics,
and geotechnical aspects are all present in the breach formation and their accurate
modelling is very important for the accurate prediction of breach outflow. The devel-
opment of effective emergency action plans and the design of early warning systems
heavily rely on these prediction results.
There are similarities and differences in the processes involved during the breaching
of embankment dams and dikes, as summarized below:
Dikes in comparison to embankment dams are typically longer than high.

While the breaching of a dam might develop to the limit of its geometry, for
dikes only one section is usually breached. Also, the settlement along the dike
length might vary in different sections, resulting in different crest elevations
along the dike length.
The breaching of embankment dams depends mostly on reservoir volume rather

than river inflow. During the breaching of river dikes and sea dikes, the river
flow and sea surge determine the breach development respectively. A relatively
finite volume of water is involved in dam breaching and river dike breaching,
opposite to sea dike breaching where an infinite and periodic volume of water
is caused by tides.
The flow in the river is parallel to the river dike axis, while for embankment
dams and sea dikes the flow direction is perpendicular to the structure axis.
Hydraulic load behind a dam is usually larger than the one for a dike.
Numereous methods have been developed for the purpose of modelling breach de-
velopment as further discussed in Chapter 4. Here we elaborate on specific breach
characteristics and the processes involved during breaching.
2.3.1 Breach shape

The mathematical description of the interaction between dam material and water
flow is not yet fully accurate (Morris, 2005). The opening formed in the structure
during the failure process - from here on defined as breach shape - depends on that
interaction. For accurate modelling, the soil mechanic parameters should be known.
Nevertheless, they can be determined only with limited degree of accuracy. Assump-
tions are made concerning the breach shape in order to avoid the non–linearity in
the equations. Models usually predefine the shape of the breach. Constant breach
shape and uniform erosion of the breach section throughout the whole breaching
development time, is usually assumed. The breach cross section is often considered
to be triangular, rectangular, trapezoidal or parabolic (Figure 2.6).
Johnson and Illes (1976), after analysing the data from approximately 100 case
studies concluded that the breach develops initially in ‘V ’ shape, three to four times
Figure 2.6: Erosive patterns of various breach shapes.
Top width
Height
Figure 2.7: Parameters of the breach opening shape.
wider than deep, later developing in the lateral direction, once the apex reaches the
hardest material of the dam core or its foundation. The lateral erosion continues un-
til either the dam is completely washed out, or the reservoir is emptied. MacDonald
and Langridge-Monopolis (1984) observed from the collected historical dam failure
events that in most of the cases the ultimate breach shape is of trapezoidal shape.
They concluded that for embankment dams, the breach shape can be assumed to be
triangular up to the time that the base of the embankment is reached. Once the apex
of the triangle reaches the foundation level, the breach develops forming a trape-
Reservoir
zoidal section water surface
extending due to lateral erosion. The conclusions drawn from several
field and laboratory tests performed within the IMPACT project is that breach sides
preserve the vertical angle during the breach development (Morris, 2005). However,
a factor influencing the conclusion can be that the rectangular initial breach shape
is predefined in all experiments. Data related to the progressing breach shape de-
Pervious foundation
velopment (in time) during real dam failure events are still missing.
Flow line
The parameters that specify the shape of a breach channel are: the breach depth
hb or the vertical extent of the breach measured from the dam crest down to the
breach bottom, width at the top Bt and bottom Bbot of the breach channel, and the
Impervious
breach side slope factorlayer
z (see Figure 2.7).
2.3.2 Breach development
Breach development in time

Breach initiation time is defined as the time of duration starting with the first ob-
servable flow over or through the structure that might initiate warning, evacuation,
or awareness, and ending with the start of the breach formation phase. During the
breach initiation phase, the outflow is relatively small, and if it can be stopped the
structure might not fail. Typical breach initiation times may range from minutes to
days. Especially piping failure might be preceded by a prolonged initiation phase.
The breach formation phase is considered to begin at the point where the structure
failure is imminent and ends when the breach has reached its maximum size. For
small reservoirs, the peak outflow from a dam break may occur before the breach
fully develops due to significant drop in reservoir levels during the formation of the
breach, whereas in larger reservoirs the peak outflow may occur when the breach has
reached its maximum size. During the breach formation phase, outflow and erosion
are rapidly increasing; while for a dike it might be possible to stop the breaching,
it is unlikely that the outflow and failure can be stopped in case of an embankment
dam. Several small springs were noticed near the right abutment of the Teton Dam,
one day before its failure. All efforts made to close the sinkholes while the leak
was rapidly growing, failed. In contrast, dike breaching can be stopped by human
intervention; the famous Hans Brinker story† is a typical example.
The rate of breach formation depends on soil material properties (cohesive, non
–cohesive, compaction, etc.) and embankment condition. Breach formation in em-
bankment dams is highly dependent on the reservoir capacity and continues till
either the reservoir is depleted or the dam can withstand further. According to the
historical data, the breach formation phase for embankment dams ranges from 0.1
to 4 hours. The breach formation in the dike structure depends on the river flood
or sea storm conditions as well as on the dike material. Cohesive dikes are likely to
breach slower than non–cohesive dikes.
Breach development in space

All breach models assume an initial channel to have been created on the struc-
ture body (Figure 2.8) either parallel to the structure crest (DEICH A (Broich,
1998)), with rotation around the downstream toe, or parallel to the downstream
face (BREACH (Fread, 1988), BRES (Visser, 1998)). The initial channel forms the
starting condition for the breaching process. If no initial channel exists then the
subsequent stages of the breaching process will not occur. The initial characteristics
of the breach channel define further breaching development.
† The legend tells the story of the brave Dutch boy thought to be named Hans Brinker, who
supposedly prevented the flooding of Harlem city by pressing his finger in the dike opening (Dodge,
1997).
(c)
Figure 2.8: Modelling modes of breach growth: (a) parallel to the structure crest,
(b) with rotation around downstream dam toe, and (c) parallel to downstream face
(Broich, 1998).
It is rather difficult to predict where the initial breach channel will be formed as
this depends on many factors e.g. structure coverage, flow characteristics, bad com-
paction at any point in the structure. Initial erosion may begin anywhere in the
structure. For embankment dams, the developed models commonly assume a breach
located at the center of the dam. However, some historical failure events showed
that breaching might occur near an abutment as well. Examples include the fail-
ures of Teton Dam (Figure 2.9), Baldwin Hills Dam, etc. The breach development
and outflow from a centrally located breach will most likely be different from a ‘side’
Figure 2.9: Teton Dam failure (Rogers, 2007).
breach in terms of time to peak discharge, peak value, and hydrograph shape. When
lateral growth is restricted in one direction, erosion rates in the other direction do
not compensate (Morris, 2005). Therefore, to predict the initial breach location
there is a need to undertake local surveys to identify weaknesses in the structure or
sub surface geology by visual means, sensors or remote sensing techniques. Within
the scope of this thesis, no attention is given to the processes that might define the
development of the initial channel location.
The structure might be homogeneous or heterogeneous with an impervious core at

the center of the structure (embankment dams). Different types of material may
lead to different channel slopes. When breaching is initiated at the downstream face
of the dam, a steep slope can be observed at the first stage of dam breaching due
to overtopping, but as the breaching continuous, the slope might remain about the
same or even decrease.
2.3.3 Breach formation mechanisms

Two breach formation mechanisms are identified: erosion and headcut erosion. The
latter is the process of removal of structural material by the combined effect of
the erosive force of water flow and by mass wasting (see Figure 2.10). Laboratory
experiments and observations of real earthfill structure failures show that erosion
is predominant for non–cohesive structures without a cohesive core. Headcut ero-
sion is observed to be predominant during the breaching of structures with cohesive
filling material, or with non–cohesive filling material but with a cohesive core (see
formed by erosion headcutting, and breaches formed by erosion of the dam material as a
result of flowing water. Dam breaching mechanisms differ depending on the dam’s
material (cohesive or non-cohesive). For non-cohesive embankments without a cohesive
core, breaching occurs mostly because of erosion, which depends on the flow velocity
and the embankment material. This erosion process progresses in time until water level
in the reservoir is low enough to cause further breach erosion. For cohesive
embankments or rockfill embankments with cohesive core, breach can be formed by
headcutting and erosion. This process starts when the protective layer on the
embankment dam fails. Headcutting typically starts at the toe of the dam or in the
knickpoint at the downstream edge of the crest (as shown in Figure 2-4).
FigureFigure
2.10: Headcut erosion process
2-4 Headcutting in a cohesive
in a cohesive soil embankment.
soil embankment
Powledge et al.(1987);
e.g. Ralston (1989)Dodge
describe flowPowledge
(1988); over an embankment, with lowetoral.no(1999);
et al. (1989); Hanson tail water,
through the (2005).
Morris following Bythree zones:the modeling results versus field and laboratory ex-
validating
periments carried out during the IMPACT project, the breach models that predict
breach growth considering the headcut erosion processes rather than only erosion,
were argued to perform better (Morris, 2005) than the models that consider only
erosion.
Modelling of the headcut erosion is not 9trivial, and while many experiments are
carried out to gain insight into this process, the mathematical modelling of this pro-
cess is just at the initial stages (see e.g. Temple and Hanson (1994); Temple and
Moore (1997); Wu et al. (1999); Robinson and Hanson (1994); Hanson et al. (2001);
Alonso et al. (2002)). Most breach models either do not consider headcut erosion,
or consider this process using very simplified assumptions, usually modelling it as
an energy dissipation process.
Erosion is modelled using the sediment transport equations that are conventionally
derived for steady subcritical flow conditions, specific types and certain diameter
ranges of sediment (Yalin, 1972; van Rijn, 1993; Bogárdi, 1974). During structure
breaching, the flow might develop into unsteady, supercritical flow and if these condi-
tions apply, the use of unsteady non–uniform sediment transport equations is more
appropriate. However, due to their absence the Meyer-Peter and Müller (1948);
Exner (1925), Einstein–Brown (Brown, 1950), and the modified Meyer–Peter and
Müller formula adapted by Smart (1984) are commonly used.
The rate of erodibility is given by:
Er = k(τ − τc )a
where, Er presents the erosion rate, k and a are two correlation coefficients, τ
presents the flowing water tractive stress and τc is the critical tractive stress for the
erodible material.
The rate of erosion is commonly assumed to be uniform throughout the channel

Subcritical flow Supercritical flow
Different erosion rates
Figure 2.11: Description of the flow over an embankment as from Powledge et al.
(1989).
section that is the submerged part of the breach channel sides is supposed to erode
at the same rate as the unsubmerged part and the breach channel bottom. Most
of the mathematical models deal with the homogeneous dams or dikes. The failure
modelling of the heterogeneous structures is commonly done through averaging the
characteristics
a) of soil properties.
b) c) d)
2.3.4 Hydraulics of flow over the dam

The breach outflow hydrograph is of crucial importance for the assessment of the
flooding characteristics in the downstream areas. The available breach models sim-
ulate flow through the breach channel using either the orifice equation (at the
initial phase of piping failure) and the weir equation, or the one dimensional de
Saint–Venant equations.
The weir equation that estimates the unit discharge for the free flow (low tailwater)
condition is generally expressed as:
q = CH 1.5 (2.1)
where q is the discharge per unit width, C is the discharge coefficient that depends
on the weir (breach) shape, and H is the total (energy) head above the crest. The co-
efficient C ranges between 1.60 and 2.15 in SI metric units. If the crest is submerged
by tailwater, then Eq. 2.1 becomes
q = Cs H 1.5 (2.2)
where, Cs is a coefficient that considers the submergence effect. Powledge et al.

(1989) describe the three flow zones (Figure 2.11) observed during overflowing of
the dam with no tailwater effects. The first zone is the movement of the flow from
the calm reservoir (static energy head) to subcritical velocity state (static and dy-
namic head) over the upstream portion of the dam crest. In this zone the hydraulic
forces and flow velocities are low. The small energy slope of the subcritical flow range
imposes small tractive stresses too. In the second zone, the flow travels through crit-
ical velocity on the crest to supercritical flow across the remainder of the dam crest,
to the downstream slope. In this zone the tractive stresses might become significant.
In the third zone a rapidly accelerating turbulent supercritical flow is observed on
the steep downstream slope. Here the energy levels increase significantly as the flow
proceeds along the downstream slope of the dam. Due to the steep energy slope, the
vertical velocities might increase significantly and the tractive stress will be large.
The downstream slope of the dam is a steep slope in hydraulic terms and a correc-
tion must be made in the calculation of gravitational forces.
The weir formula is often used to calculate the flow along the crest. The steady
non–uniform flow equations have also been used to compute the water depths, ve-
locities, and energy slope on the downstream slope despite the short reach of the
breach channel and its steep slope because of their relatively simpler computations
compared to full de Saint–Venant equations.
Chapter 3
Current Approaches for

Dealing with Flooding
Happy Holland, had we not dug and even diked,

We were now living above the rivers,
That must cut the land, but now run over it in man–made
channels,
Whose bottom is ever higher raised by falling silt,
And so rise ever more above the land and in force and violence
their dikes can overwhelm.
Willem Bilderdijk (XIX Century)
3.1 Introduction
Nishat (2006) defines flood as “the process of inundating normally dry areas and
causing damages”. Inundation does not always pose a risk and not necessarily has
only negative impacts. It can have positive effects on soil fertility, ecosystems, etc.
(Nishat, 2006). Floods∗ , on the other hand, can be very severe and have enormous
consequences in terms of economic, ecological and social values. Floods’ destructive
forces are a threat that has been faced by humans for generations already. Noah’s
legend is one of the thousands of legends related to floods and their devastating
consequences.
Therefore, novel strategies and methods are continuously developed worldwide for
dealing with floods (Knight et al., 2006). Several common types of measures are
distinguished, such as:
1. Measures for preventing flooding: dam or dike construction, maintenance and
improvement, river dredging, etc.
∗ Reader is referred to FLOODsite (2005) for definition of flood and risk related terms.
26 Section 3.2. The Netherlands’ long history of battle against floods
2. Measures for reducing flood impact: retention and detention basins, flood-
ways, flood forecasting, spatial planning, awareness raising (games, role–plays,
brochures, etc.).
3. Measures for dealing with an approaching flood and during a flooding period:
Decision Support Systems (DSS), warning and emergency plans, evacuation
and local emergency protection, etc.
4. Measures taken after the flooding occurred: aftercare, compensation, insurance

and restoring of the flooded area. The loss or damage of property, and in some
cases the inability to return home for a period of time causes great stress
and disruption to people. The metro of Prague, was unable to run for several
months after the 2002 torrential flood that hit large parts of Europe. Measures
should be taken to deal not only with direct but also with indirect (long term)
consequences.
These measures taken against flooding are categorized in two groups of strategies:
1. Flood controlling strategies referred to as resistance strategies, and
2. Flood damage reduction strategies referred as resilience strategies (Vis et al.,

2003).
The first strategy aims at fully preventing floods. However, as flood prevention is
not always possible and feasible, nowadays more emphases is given to strategies that
improve the coping capacity, resilience, and adaptability, and offer more flexibility
for future interventions. Disaster mitigation is suggested as a priority at numerous
international events such as the Fourth World Water Forum in Mexico† in 2004,
the Conference of International Center for Water Hazard and Risk Management‡
in 2006, etc. In the following sections we describe flood dealing strategies in two
countries that are similarly vulnerable to water related disasters.
3.2 The Netherlands’ long history of battle against

floods
“...Dutch created the Netherlands”, is the Dutch expression that emphasizes the con-
tinuous battle against water to expand or protect the Netherlands territory through
which three major European rivers flow into the sea: the Rhine with its branches
the Waal and the Neder Rijn flowing from Germany, and the Maas (a branch of the
Meuse) and the Schelde, flowing from Belgium. Being one of the most densely pop-
ulated deltas in the world with 25% of its land area below sea level and 65% prone
† Reader is referred to www.worldwaterforum4.org.mx
‡ Reader is referred to www.icharm.pwri.go.jp/html/meetings/iwfrm2006
alternative resilience strate- distinguished:
d assessed for their hydrau- – ‘room for rivers’, a strategy focusing on measures to lower
ty criteriaʼ. the water levels despite a rise in the design discharge, e.g. by
enlarging the discharge capacity of the channel itself and/or
through lowering the floodplains (5);
– a resilience strategy for the protected areas based on mini-
mizing flood risks by controlled flooding and by limiting the
ds experienced periods of un- flood-damage through dedicated physical planning in tune
euse Rivers reached very high with flood frequencies (6).
n 1926 had a higher discharge Although both strategies have to address similar issues such as
er 250 000 people were evacu- the distribution of discharge from the Rhine River over the vari-
Chapter 3. Current Approaches for Dealing with Flooding 27
when the stability of the dikes ous river branches and the relationship between river manage-
scussions on The Netherlands’

sified and extended to include
elevant factor for the long term.
n the discharge regime of the
crease in the so-called design
a probability of 1/1250 years)
owards 16 800 (minimum sce-
m scenario) by 2100. Addition-
, sea level rise may hamper the
sea level rise is currently esti-
bove present. Finally, in deltas
ge and oxidation of extensiveFigure 3.1: The land and sea level throughout the years in the Netherlands (Klijn
of large areas, a process which
et al., 2004). The1.dashed
Figure line
The land presents
level sinks the forecasted
whereas estimates.
the sea level rises, ham-
and maintained by a history of pering the discharge of rivers in deltaic areas and increasing the
1). vulnerability of the protected areas.
al changes, social to flooding, the Netherlands has no other choice of existence but to fight against sea
demographic and and river high water levels. The land has been continuously subsiding (Figure 3.1)
etherlands is such as a result of drainage during the Middle Ages’ reclamation and the action of both
bly affected by a shrinkage and oxidation of peat layers. Since the eleventh century, when the first
00 the population dikes are known to have been built (van de Ven, 2004), nowadays, 53 rings of dikes
5 million inhabit- protect the delta area from flooding.
was over 16 mil-
concentrated in In 1927, after the severe flood of 1916, the implementation of the Zuiderzee Works
hereas nowadays started (see Figure 3.2). A system of artificial dams, land reclamation and water
vailable room for drainage works was built. The Afsluitdijk or the ‘closure dike’ (in English) has been
re, infrastructure, the longest man–made dam§ for 74 years. Some 32km long, 90m wide and 19m high
er vast areas, and (some 7.25m above sea level), the dam is essential for protecting the land from the
and-surface area. Figure 2. Floodplain
North
and (b) 2000.
area and
Sea. After the land-use intensity
completion of thearound
works, Arnhem around
the coastline (a) 1830
length was decreased
by almost 300km and a freshwater lake was created that provides drinking water as
well as reclaiming
of Sciences land
2004 for agriculture and urban development.
© Royal Swedish Academy
http://www.ambio.kva.se
141
During the February 1953 storm (Gerritsen, 2005), the greatest storm surge on
record for the North Sea, the Afsluitdijk prevented flooding on the Zuiderzee coast.
However, on the southwest of the country, the storm lead to the worst ever disaster
in the Netherlands when 800km of dikes protecting the area were breached causing
1,835 casualties and flooding 2,000 km2 of land. Afterwards, the construction of the
Delta Works (Figure 3.3), the most spectacular water related mega structure in the
world, took place.
Numerous dams and dikes, sluices, locks, and storm surge barriers were built in the
Netherlands to protect the land from sea and river flows. However, in 1993 and 1995,
§ The Saemangeum Seawall, located on the southwest coast of the South Korea, is the world’s
longest man–made dike from the time of its inauguration in 2006, being 0.5km longer than the
Afsluitdijk dike.
(a) (b)
Figure 3.2: Afsluitdijk - the ‘closure dike’ - in the Netherlands.
(a) (b)
Figure 3.3: Eastern Schelde storm surge barrier (Rijkswaterstaat).

the Netherlands was again under threat of significant flooding, when the Rhine and
the Maas Rivers reached very high levels and the stability of the dikes seemed no
longer guaranteed. Some 200,000 people were evacuated during the January 1995
storm.
Nowadays, there are over 16,000km of dikes and 300 structures that protect the
Netherlands from flooding. About 2,400km of constructed dikes are designated as
primary dikes - dikes in direct contact with water bodies (sea, lakes and rivers) -
and 14,000km are secondary dikes - dikes sorrounding the canals. The design of
sea and river dikes is traditionally based on a design water level with a particu-
lar frequency of exceedance. Following the disaster of 1953, the Delta Committee
(1960) defined the design water levels for the sea dikes, which are built to withhold
the 10,000 year return period storm surge levels. The Commission on River Dikes
(1977) recommended that river dikes should be designed and improved to retain
water levels associated with a governing rate of discharge of the Rhine River at
Lobith (entry location of the Rhine River in the Netherlands), which is exceeded
with a frequency of about 1,250 year (CUR/TAW, 1990). The design levels keep
increasing, as the current 1,250 year return period discharge at Lobith corresponds
to 16,500m3 /s while further increase in the design discharge is forecasted to be from
16,800 (minimum scenario) to 18,000m3 /s (maximum scenario) by the year 2100
(Kwadijk and Rotmans, 1995). Still, those values are uncertain as the 1,250 year
flood event is forecasted based on only 100 years of records.
Innovation in flood protection, prevention, and management is continuously devel-

oping as the perception that the Netherlands can be subject to flooding anytime
has not perished. Climate change has increased the uncertainty about river flows
while sea level rise might impede the rivers’ discharge. Moreover, sedimentation that
occurs on the floodplains along the rivers and the subsidence of the land behind the
dikes, lead to larger differences between floodplain and hinterland level (Asselman
and Middelkoop, 1995). In other words: when the river capacity is decreased, the
flood hazard is increased. Construction of dikes leads people to believe that they are
safe, and as such the urbanization and economic development of the areas continue
to grow, increasing in turn the flood risk. This development along the rivers makes it
more difficult to implement flood measures, such as compartmenting areas, creating
water retention basins or floodways, viz. developing so–called green rivers.
In the Netherlands, it is typical to have continuous system of dikes that together

with high ground areas and other hydraulic structures (gates, sluices, locks, pump-
ing stations, etc.) create an enclosed area. These areas are called dike ring areas
and currently there are 53 of them identified in the Netherlands (Figure 3.4). The
largest dike ring surface area of 600 km2 is in the Betuwe, in the so–called central
river land. The Flood Defences Act of 1996 defined the safety standard for each dike
ring area, “expressed as the average exceedance probability–per year–of the highest
ments and hence sustainability. We go into the advantages but hazard has substantially increased.
also the far-reaching implications. However, we first briefly The above-mentioned dike-rings consist of
introduce the present situation along The Netherlands’ Rhine ly closed rings of dikes or of dikes which con
River branches. grounds. Their surface area ranges from less t
about 600 km2 for the largest, the Betuwe. This
large differences in potential flood damage throug
BRIEF HISTORY AND PRESENT SITUATION tude. Also the potential flooding depths differ sub
Before 1000 AD inhabitants of the delta and alluvial plains lived much less than 2 m for dike-rings adjacent to h
on natural levees, other natural heights, or they built mounds. more than 6 m in the lower parts of the enclosed
Since then, the Dutch started to construct dikes, resulting in a differences become even more accentuated whe
closed pattern of dikes from about 1400 AD onwards. Today, status and development rate are taken into acco
this pattern of dikes is largely unchanged. It consists of 53 so- largest dike-ring with the largest potential depth
called dike-rings for the whole country, 10 of which adjoin the vulnerable land-use types are now being develop
Rhine River branches (Fig. 3). industry, and (inter)national transport.
Although on the one hand flood protection ha
sound, a further heightening of the dikes meets v
This partly follows from the notion: “the highe
bigger the disaster”, but also because of the neg
protective structures on natural, cultural, and sc
values. Even more importantly, the improved
tains the negative spiral of further economic inv
protected areas. The enhanced ‘notion of safet
economic investments in the protected areas, wh
a social call for dike reinforcements, etc. The op
reinforcement therefore calls for a change of stra
escape this negative spiral.
A summary of the current situation would re
structure from 1400 AD, equal safety standards
the river, and a steadily increasing flood-damag
cause of a still largely unbridled 21st century lan
nomic development rate. The urgency of changin
of future developments is acknowledged and incr
trated by the many reports and the advice to the
research institutions (3, 10), dedicated committe
water management authorities.
LIMITING THE RISKS OF FLOODING

A resilience strategy for flood-risk management r
ber of principles and related measures. Firstly,
phasize that flood-risk management is not the sam
trol. Flooding risks result from a flood hazard in
Figure 3. Dike-rings
Figure 3.4: Dike (numbered, referring
ring areas to Tables
of the 1 and 2) and
Netherlands. is vulnerable. Without a vulnerable society or eco
safety levels along the lower Rhine River in The Netherlands.
may be regarded as a natural phenomenon which
ified in a normative way as posing ‘a risk’. This a
The height of the dikes along the Rhine River has, however, the concept of flood risk embraces the concepts
steadily
water level, which theincreased
primaryover waterthe course
defence of history.
must be Initially,
capablethis of
meant
withstanding from often, it is expressed as a funct
and vulnerability;
adapting to the last-experienced flood level.
the outside, while taking into account other factors which determine theFrom the 1950s on- product,
waterto be
de-precise). In The Netherlands, how
wards, a more scientific approach has been followed, based on mon practice (12) to define flood risk as flood pro
fensive capability” (Floris,
hydraulic 2005).forFor
modelling a dikedischarge’
a ‘design ring, itswithweakest link defines its safety
a recurrence damage. Obviously, the approach is the same, o
intervalRepresenting
level or its strength. of 1/1250 years.the Thissafety
design level
discharge
of aisdike
determined
ring by by itsquantification
weakest link differs. The difference lies primari
a statistical analysis of all floods recorded in the last
is economically undesirable, as the failure probability of the entire dikeflooding 100 years. ring isdepth
high,is included:
The 1993 and 1995 floods have caused the design discharge to
when only partrise of from
it has15high
000 toprobability
16 000 m3 s-1of failure in
, resulting andthe needs
appliedparticular
linear Rf = attention.
Hf x V = Pf x Df
regression becoming slightly tilted. This emphasizes the suscep- Rf = flood risk; Hf = flood hazard; V = vulnerab
Water retentiontibility
and ofdetention areas
this statistical are seen
method as two
to climatic potentialwhether
fluctuations, solutions, probability;
both beingDf = flood damage; with both Hf an
normal or resulting
applicable to particular situations. from global
Floodchange
water(7). It also emphasizes
detention on flooding
areas are useful for depth.
at-
tenuating floodthe uncertainty
peaks, of forecasts
provided of rare events
the storage volume (1/1250 years),
is large when
compared toThe
thevulnerability
flood of an area depends on the n
based on only a 100 years of records. The 1/1250 years discharge ple, the economic value (invested capital and
peak and the detention occurs in time of peak discharge. The area is temporarily
can only—with 90% certainty—be estimated to lie between 13 and/or ecological values. In the downstream par
flooded and can 500be andrestored
18 000 mor s (5, 7).
3 -1be available for land use or other functions again
River basin, an adequate Flood Early Warning S
The fixation of
later on. For flood reduction purposes, the river between
they thearedikes since about
considered more1400efficient
is operational.
than re-Presently, the water levels can be
AD in combination with the steady heightening of the dikes has
tention areas that serve to store water, not allowing it to flow downstream. Water
retention basins
142along the Rhine River are not seen as©aRoyal
feasible solution
Swedish Academy as they
of Sciences 2004 Ambio Vo
http://www.ambio.kva.se
are considered to contribute only to the attenuation of low to medium peak flows,
and not in the case of extreme or prolonged events. Detention areas far upstream in
the Rhine basin are “not considered very effective in lowering extreme floods that
endanger the downstream areas” (Hooijer et al., 2004). Clearly, some flood control
solutions cannot be transferred to different places and are very much depending on
the particular location and situation.
Dividing a dike ring area into compartments might significantly reduce the risk.
During the Flood Risks and Safety (Floris) project, the re–estimation of flooding
probabilities and consequences of flooding for 16 dike ring areas (out of 53 in total),
together with in-situ examination of the dike rings was accomplished. The objective
was to check the reliability of the dikes and the hydraulic structures, as well as to
identify and reduce uncertainty. The risk was calculated for each dike ring area
separately ignoring hydraulic interactions effects, that are not taken into account
in the current design and safety assessment practice in the Netherlands despite the
awareness of its importance. The damage in the province of South Holland varied
approximately from e280 million to e37 billion (depending of the location of the
breach) when the presence of embankments (highway, railroads, etc.) and natural
terrain was considered, instead of e290 billion calculated assuming the flooding of
the whole area without partitioning. For more details the reader is referred to the
report by the Dutch Ministry of Transport, Public Works and Water Management
(VenW, 2005).
Dutch are currently implementing the plan to give back some of the land to water.
Creating room for the river means more storage for flood water by relocating the
dikes further inland, lowering the floodplains, creating bypasses or so–called green
rivers, creating detention areas along the river, etc. (see Figure 3.5). The green river
definition is used since this solution is believed to give more priority to ecological
benefits rather than only economic. It is a general belief that the ‘Room for the
River’ policy is effective. However, its implementation at a large scale might be dif-
ficult especially when periodical flooding of inhabited areas is involved or when the
water retention basins highly affect the economic development of the community.
The idea of amphibious floating houses and roads is currently on its way to imple-
mentation. The first floating village, expected to accommodate 20,000 people, is
under construction at a semi aquatic city on IJburg near Amsterdam. This tech-
nology allows the houses to float on the water if the water levels rise during a flood
event.
Vis et al. (2003) and Klijn et al. (2004) compared the controlling strategies of suc-
cessively heightening and strengthening of dikes for coping with the Rhine River
floods with two resilience strategies consisting of
1. Using detention areas of low economic value along the river and partitioning
of a dike ring area into compartments with different flooding probabilities; or
2. Increasing river discharge capacity either by enlarging the floodplain area or

by adding retention compartments of low economic value that have a high
probability of flooding.
Towards sustainable flood risk management in the Rhine and Meuse river basins
Illustration
Figure of typical location
3.5: Illustration of flood risk
of typical management
location measures
of flood risk within the river basin.
management measures within
(From Project 5).
the river basin (Hooijer et al., 2004).
The following criteria were assessed (Klijn et al., 2004):

1. Costs (investment and maintenance),
2. Flexibility (to adapt to changes in boundary conditions, normative views, etc.),
3. Resulting (expected) flood damage,
4. Economic impact,
5. Ecological impact,
6. Landscape qualities (scenery and cultural heritage).

The first (traditional) strategy was found economically favorable to implement being
cheaper than the other two resilience strategies. While the first strategy involves
the heightening of the dikes, the other two incorporate the construction of new
dikes, adaptation of infrastructure, and financial compensation to the owners and
inhabitants of areas where economic development will be affected as a result of in-
creased allowed flooding frequency. The incremental costs of heightening existing
dikes, is expected to be lower since the original costs of constructing dikes are not
included in the
Some ‘Room foroverall estimation
Rivers’ measures or comparison.
practised Klijn
or considered along et al.Rhine
the lower (2004)
river:concluded that
in long–term, thetheresilience
(From ‘Room for Rhine’) strategies have fewer disadvantages than the flood con-
trolling strategy of dike heightening. However, the implementation of the resilience
IRMA-SPONGE Summary Report - 10

Towards sustainable flood risk management in the Rhine and Meuse river basins
A map of the Rhine basin with flood trav

the Dutch-German Dutch border. Also sh
contributions to floods from different pa
basin
(Result of Room for the Rhine).
An improved flood early warning system
a forecasting time of 4 days to The Neth
(Result of Project 12).
Figure 3.6: A map of the River Rhine basin with flood travel times (in days) to the
Dutch–German border (Hooijer et al., 2004).
strategies requires large investments in the short–term whereas their revenues will
become clear only after relatively long period of time.
Different types of flood risk management measures were considered during the IRMA
–SPONGE project (Hooijer et al., 2004), focusing on flood prevention measures, de-
velopment of methodologies to assess the impact of flood risk reduction measures
and climate change scenarios. The objective was “to support the spatial planning
process in establishing alternative strategies for an optimal realization of the hy-
draulic, economic and ecological functions of the Rhine and Meuse River Basins”
(Wolters et al., 2001). Alternative resilience strategies have been elaborated and
assessed for their hydraulic functioning and ‘sustainability’ in risk management cri-
teria. The IRMA–SPONGE project (Hooijer et al., 2004) concluded that for the
Rhine River:
34 Section 3.3. Japan’s experience in dealing with floods
1. The timing of flood peaks from tributaries to the main stream is

highly complex.
2. The storage volume available in detention areas is utilized most
effectively for peak shaving provided accurate timing of detention
can be achieved during flood peaks and not during the earlier stages
of floods.
3. The further upstream retention and detention areas are created, the
less effective they are for reducing extreme floods downstream.
Downstream the Rhine River basin (Figure 3.6), the Flood Early Warning System
(FEWS) is in operation. The water levels can currently be predicted with adequate
accuracy 2 days ahead of a flood, and efforts are currently being made to lengthen
this period up to 3 to 4 days.
As the flood risk increases, public awareness becomes a key factor for saving lives
and it is important to have better participation of people in the decision making
related to measures needed to decrease the risk. The population living in risk zones
should be aware of the flood hazard maps. This is something that is not gener-
ally implemented in the Netherlands, but has been widely adapted in some other
countries (e.g. disseminated by the Internet in UK).
3.3 Japan’s experience in dealing with floods

Japan is a country where the occurrence of disasters is quite common. Earthquakes,
floods, tidal waves, volcano eruptions and typhoons are part of every day life in
Japan. Enclosed in all directions by the sea, it is vulnerable to storm surges, high
waves, and tsunami on the coasts. Preserving the coastline from these hazards is
crucially important while at the same time preserving various ecosystems. More
than 50% of the population and more than 70% of the nation’s assets are concen-
trated in the floodplains.
Storm surges by typhoons have caused catastrophic damages in Japanese coastal

areas throughout the years. The typhoon Vera in 1959 caused Japan’s greatest
storm disaster accounting for nearly 5,000 casualties, and leaving some 1.5 million
people homeless due to the enormous damage from wind, floods and landslides. If it
were not for the construction of coastal dikes and improvement of weather forecast
systems, the number of casualties from storm surge floods can continue to be high
nowadays. However, the storm surge flood of 1999, the worst flood disaster since
1959, indicated that there is still a need to improve the measures against storm
surges.
Rivers with their short length, steep slopes and narrow catchment areas are another
source of flooding disaster. In response, engineers have taken disaster prevention and
Breach width
Breach depth
Breach side slope
[Before the construction of a super levee]

Back slope of the riverbank
River zone
[After the construction of a super levee]
Existing back slope of riverbank is used
Super levee special zone
Figure 3.7: Example of super levee (www.rfc.or.jp).
conventional dike conventional dike conventional dike

Reservoir water surface
Pervious foundation
Figure 3.8: Behaviour of conventional dike (above), and super levee (below) subject
Flow line
to overflow, seepage and earthquake shaking.
Impervious
management measures. layer control and management is implemented through a
Flood
number of structural and non–structural measures. In addition to these mentioned
in Section 3.1, artificial underground channels, stimulation of flood–proof building
construction, installation of flood sensors that can assist in deciding on evacuation,
and other options described in detail below are taken in Japan.
The underground floodways and regulating reservoirs are two solutions applied in
Japan to deal effectively with urban flooding. The construction of super levees
(Figure 3.7) is seen as a good solution to prevent flooding in urban areas where
a potential dike break may have catastrophic consequences. In the battle against
floods and earthquakes (Figure 3.8), they seem to offer stronger foundation and
better protection. A super levee is considered resistant to overflow and seepage.
Having a width of about thirty times their height, super levees have the strength to
withstand severe flooding. Super levees provide usable land and space for dwellings
and restores access to the riverfront.
The design tide level for coastal dikes is estimated based on the recorded values or
the expected maximum water level creating uncertainty about any storm stronger
than the designated storm. To deal with these exceptional events, early warning sys-
tems are in operation and evacuation plans are utilized. Urbanization near coastal
Breach depth
Breach side slope
[Before the construction of a super levee]

Back slope of the riverbank
River zone
[After the construction of a super levee]
Existing back slope of riverbank is used
36 Section 3.4. Modelling of propagation

Super levee special zone of flood caused by structural failure
Number Flood Frequency

of people 500-1000 years
1 hour 6 Flood Elevation (m) 100 years
50 years
People aware People not 30 years
of hazard aware of 5 10 years
maps hazard maps 5 years
3 years
4
Time
1
Evacuation Evacuation
announcement order 0
(a) Number (b)

of people
People aware 1 hour People not

of hazard aware of
Figure 3.9: (a) Effectiveness of flood maps
hazard map during
hazard mapstorrential flood in
Fukushima and (b) Expression of the safety degree against flood water levels (Mer-
abtene et al., 2004).
Evacuation Time
areas and extensive expansion of mega cities in underground
announcement Evacuation
order
levels make people more
vulnerable to potential flooding. For warning the population of the risk posed due
to flood hazard, vulnerability maps are made available and are distributed to public.
Merabtene et al. (2004) confirms that warning the population using up to date in-
formation technology and preparatory work of flood hazard map distribution among
them proves to be effective to mitigating flood damage and reduce the numbers of
casualties. During the torrential rain of 1998 in Fukushima, people of Sukagawa city
that were aware of flood hazard maps evacuated 1 hour earlier than those who were
not (Merabtene et al., 2004). Figure 3.9a shows the difference in evacuation time
among the people with and without knowledge about flood hazard maps.
A comprehensible Flood Risk Indicator has been developed in Japan to indicate the
degree of safety/risk against flood damage through the means of a chart depicting
the frequency of floods and inundation level versus the height of people and houses
(see Figure 3.9b).
3.4 Modelling of propagation of flood caused by

structural failure
To correctly estimate the consequences derived from a structural failure the mod-
elling of flood propagation should be of high accuracy. Identification of the inundated
areas, inundation depth, speed and duration, as well as the impact that flood water
characteristics (salt, freshwater, contaminated water, etc.) can have on the inun-
dated areas, are very important for decision making, emergency evacuation and early
warning.
Most of the research on modelling dam break flood wave propagation has been fo-
cused on the movement of clear water in the downstream valley, excluding debris
flows and sediment transport. Clear water dam break floods have been studied ex-
tensively experimentally, analytically and numerically.
The De Saint–Venant (1871) equations or shallow water equations are used for mod-
elling dam break flood propagation. These equations consist of the mass and mo-
mentum conservation equations, and assume that the vertical velocities are much
smaller than the horizontal velocities, which leads to hydrostatic pressure distribu-
tion in a channel cross section (Derivation of these equations can be found in Stoker
(1957); Abbott (1979); Cunge et al. (1980); Chow (1959); Chaudhry (1993)).
At first, dam break flood analysis was assuming instantaneous and complete fail-
ure process rather than gradual failure of a dam. The sudden failure results into a
highly unsteady flow, with a forward (positive) wave advancing in the channel down-
stream the dam and a back (negative) disturbance propagating into the still water
upstream the dam (see Figure 3.10). Owing to their mathematical complexity, the
analytical integration of de Saint–Venant equations in an unsteady flow situation
can be obtained only for very few idealized situations. The first explicit solution to
a dam break problem was given by Ritter (1892) who solved the de Saint–Venant
equations analytically for the case of dam break flow in a horizontal and infinitely
large rectangular channel neglecting the hydraulic resistance caused by stream bed
friction and turbulence.
When the wave is considered to advance into still water of appreciable depth, the
flow is approximated by a shock wave with jump conditions of mass, momentum,
and energy conservation. The solution is deduced by evaluating the relation between
hydrostatic forces and rate of change of momentum at the wave front. Many authors
have carried out experimental investigations for the purpose of checking the validity
of the analytical solution Schoklitsch (1926); Dressler (1954); U.S. Army Corps of
Engineers (1960, 1961); Montuori (1965), etc.
The experiments of Schoklitsch (1917) indicated that the velocities for the forward
wave may be as low as 40 percent of the result obtained by Ritter’s solution, while
having good agreement for the tip of the negative wave. Later, Dressler (1952)
considered the wave propagation on a dry horizontal channel bed taking into con-
sideration the effect of hydraulic resistance in the forward part of the flow. The
hydraulic resistance caused by bed friction and turbulence in the shallow front re-
gion of the flow were considered dominant over the effect of the small slopes that
were investigated later by Dressler (1958).
38 Section 3.4. Modelling of propagation of flood caused by structural failure
Removed dam wall

y
Negative wave
Advancing positive surge
Figure 3.10: Dam break wave in a horizontal wet channel.
The numerical solution of the de Saint–Venant equations in natural streams was first
achieved using the methods of characteristics. Later, Ré (1946) analyzed the dam
break flood wave for a non–horizontal initially wet channel with a Chézy resistance
coefficient using a finite–difference method applied to the characteristic equations.
Different numerical schemes have been developed to solve the de Saint–Venant equa-
tions in 1 and 2 dimensional forms. Finite–difference (Abbott, 1979), finite–element
methods (see e.g. Hervouet (2007)), and lately finite–volume (Toro, 1999) methods
are generally preferred for their ability to reproduce discontinuous solutions.
Depending on the addressed problem and on the numerical scheme used to solve the
shallow–water equations, conservative or non–conservative forms of the equations
can be used. For different cross sections - arbitrary shape (natural shape), prismatic
or rectangular - different forms of the shallow–water equations can be written. In
Table 3.1 the conservative and non–conservative form of the shallow equations in
arbitrary shape cross section are presented. For further information on modelling of
dam break induced flows in complex topographies the reader is referred to Soares
Frazão (2002); Chanson (2005).
During dam breaching, reservoir trapped sediment and dam material is moved away
together with the water and is deposited in the downstream valley. According to
Costa and Schuster (1988), historical floods from dam failure events have induced
Table 3.1: Shallow water equations in one dimension.

Represented forms Continuity & momentum equation
Conservative form on
arbitrary cross section, where:
A = cross sectional area
∂A ∂Q
Q = discharge ∂t + ∂x = ql
I1 = first moment of the cross
section
I2 = spatial variation of the cross
∂Q Q2
section width ∂t + g ∂I
∂x +
1 ∂
∂x A =gA (S0 − Sf ) + gI2
S0 = bottom slope
Sf = friction slope
ql = lateral inflow per unit length
∂h
Non-conservative form on ∂t + U ∂h ∂U
∂x + h ∂x = 0
arbitrary cross section,
where: U = velocity
∂U
∂t + U ∂U ∂h
∂x + g ∂x =g (S0 − Sf )
severe soil movements in various forms: debris flows, mud flows, floating debris and
sediment–laden currents. Intense erosion and deposition is expected to highly affect
the morphology of the valley that in turn affects the flow. Still, nowadays, the prop-
agation of flood wave induced by a structural failure is commonly modelled ignoring
any presence of debris and sediment. It is the complexity of the processes involved,
that limits the modelling of floods induced by structural failure to clear water wave
propagation. Despite the awareness of the importance that the proper modelling of
flood water and sediment/debris interaction have on the hazard evaluation, mathe-
matical or numerical modelling of this process is not yet satisfactory. For information
on ongoing work on modelling of debris and sediment transport during structural
failure see e.g. Leal et al. (2002); Nsom (2002); Spinewine and Zech (2007).
There are a number of commercial and non–commercial hydrodynamic packages

available worldwide. Sobek 1D2D developed at Delft Hydraulics (www.sobek.nl),
MIKE FLOOD at DHI (www.dhisoftware.com/mikeflood), Telemac at EDF-DRD
(www.telemacsystem.com) are few of the most widely used commercial hydrody-
namic computation products. There are also freely available software packages,
mostly used for one dimensional modelling of the steady or unsteady flow.
In this thesis we use Sobek 1D2D, an integrated one– and two–dimensional numerical
simulation package developed by WL | Delft Hydraulics for modelling the propaga-
40 Section 3.4. Modelling of propagation of flood caused by structural failure
tion of flood wave caused by a structural failure. Sobek 1D2D solves the shallow
water equations using the so–called ‘Delft Scheme’ (Stelling et al., 1998; Stelling
and Duinmeijer, 2003). The 1D schematization for flow through a river channel is
combined with a 2D schematization for overland flow, bringing the model’s behavior
closer to the real physical behavior (see Figure 3.11). A control volume approach is
used for the 1D calculation points of the channel in combination with the 2D grid
cells. The flow in the 1D channel below the 2D grid bed level is treated as 1D flow,
while the flow above the 2D grid level is treated as 2D flow within the area of the
2D grid cell.
Sobek 1D2D is capable of simulating the dynamics of overland flow over an initially
dry land, as well as flooding and drying processes on every kind of geometry (Dhon-
dia and Stelling, 2002). It can correctly simulate the transition between sub and
supercritical flows and vice versa (Verwey, 2001). Stelling and Duinmeijer (2003)
give a description of the grid scheme used in the two dimensional overland flow
model of Sobek 1D2D, and present a comparison between numerical and experi-
mental results of the dam break flood propagation experiments under dry and wet
conditions. A GIS based tool is associated with Sobek 1D2D for data input and
output processing.
Figure 3.11: Schematization of the hydraulic model: Combined 1D/2D staggered

grid.
3.5 Uncertainty associated with flood modelling

The processes involved in structural failure analysis are neither completely under-
stood nor accurately described in mathematical terms. Uncertainty analysis is neces-
sary for identifying sources of uncertainty and reducing it. Identifying the reducible
and the irreducible uncertainties in a model is an important step in the overall
process of uncertainty reduction. There are several types of uncertainties classified
differently by various researchers (Funtowicz and Ravetz, 1990; Hoffman and Ham-
monds, 1994; Rowe, 1994; Krzysztofowicz, 2001; van der Sluijs, 1997). Vesely and
Rasmuson (1984) classified uncertainty as:
1. Input or data uncertainty.
2. Modelling uncertainty, that is a result of incomplete understanding of the

modelled phenomena, or numeral approximations used in mathematical repre-
sentation.
3. Completeness uncertainty. The uncertainty associated with all omissions that

occur due to lack of knowledge.
Tools for dealing with uncertainty are sensitivity analysis, error propagation equa-
tions, Monte Carlo analysis, expert elicitation, scenario analysis, Bayesian belief
networks, etc. Saltelli et al. (2000) and Saltelli (2004) define the Sensitivity Analy-
sis (SA) as the study of how uncertainty in the output of a model can be apportioned
to different sources of uncertainty in the model input, and how a given model de-
pends on the information fed into it. There are three types of sensitivity analysis:
(i) Screening, (ii) Local Sensitivity Analysis, and (iii) Global Sensitivity Analysis.
Local Sensitivity Analysis consist in varying one parameter at a time and observing
its influence on the output, while Global Sensitivity Analysis consist in varying all
parameters and evaluating their contribution to the variance of the output.
Janssen et al. (1994) defines uncertainty analysis as the study of the uncertain as-
pects of a model and their influence on the model output. Stochastic modelling
is traditionally used to analyze uncertainty and uncertainty propagation associated
with input data and model parameters that are represented by probability distribu-
tion functions rather than by a single value.
Monte Carlo Simulation and Latin Hyper Cube sampling are some of the methods
for stochastic model calculations. Due to their time and calculation complexity, less
time consuming sampling methods (best case, mean case, and worst case sampling
methods) have been designed, that consequently provide less accurate information
on the uncertainty in the outcome. The lack of information about the shape of the
probability distribution functions of the values of the input data and model parame-
ters leads to the idea of uncertainty of the uncertainty methods such as Monte Carlo
or Latin Hyper Cube. Alcamo and Bartnicki (1987) argued that the uncertainty
42 Section 3.5. Uncertainty associated with flood modelling
about the shape of probability distribution functions might be negligible.
In flood risk modelling, usually a sensitivity analysis is undertaken to detect the

influence of the inputs on the output of the model. In this thesis we apply local
sensitivity analysis to the models developed.
Chapter 4
Current Approaches to
Dam/Dike Breach Modelling
There are lies, damned lies, and statistics.

Benjamin Disraeli
4.1 Background
One important step in dam break modelling is the accurate prediction of the breach
outflow hydrograph. Significant effort has been made in developing models that
accurately predict breach characteristics; still, many uncertainties related to breach
modelling are still existent (Franca and Almeida, 2004; Morris, 2005; Zagonjolli
et al., 2005; Zagonjolli and Mynett, 2005a). Due to the incomplete understanding
of breach formation processes and hence the limited capabilities of mathematical
description of dam breaching mechanisms, the presently available models rely on
several assumptions as mentioned in Section 2.3.
In fact, no model is exclusively represented from first principles into mathematical

equations and there always exist some degree of approximation and empiricism. In
this thesis a distinction is made between empirical models and models created using
data mining techniques, where a model is derived from the data itself, and physical
models, where data are used to fit the equations based on physical principles.
During the IMPACT project, a European project investigating extreme flood pro-
cesses and uncertainties∗ , several breach models were assessed using field and lab-
oratory scale experimental data. Though these models offer improved capabilities
and reliability in comparison with earlier models, their prediction accuracy is still far
∗ The reader is referred to the extra issue of Journal of Hydraulic Research, VOL 45, 2007.
44 Section 4.2. Experimental work
from satisfactory. A 30% uncertainty band was identified for predicting peak out-
flows in the field and laboratory tests. When applied to the historical failure event
of the Tous Dam in Spain, a 50% uncertainty band was observed. Poor prediction
of the outflow hydrograph and unreliable prediction of the breach development in
time was observed in the experimental results (Morris, 2005) suggesting that further
research is needed to increase the capabilities of the existing physically based mod-
els. In this chapter we describe the present tools and methodologies used for breach
modelling.
4.2 Experimental work

For better understanding the processes involved and to correctly describe the com-
plex phenomena of embankment breaching, various field and laboratory experiments
were carried out during the previous century. At first, the experimental work was
focused on identifying the scale of flooding resulting from a sudden failure of the
dam structure. This approach, though appropriate for concrete and arch dams that
usually exhibit a failure within a relatively short duration of time - similar to sudden
failure - did not provide insight into the breach modelling processes except for the
flood wave movement in the downstream channel (Dressler, 1954). A thin plate rep-
resenting the dam was located in a rectangular ‘channel’, and its removal simulated
instantaneous failure of the dam.
During recent years, several field tests and laboratory experiments were carried out
modelling gradual failure of dams. In the Netherlands, research regarding breach
growth in dikes started with some field experiments on non cohesive sand dikes
in 1989. Two fields tests were carried out at ‘Het Zwin’ located in Zeeuwsch-
Vlaanderen, at the boundary between the Netherlands and Belgium. The exper-
iment was focused on modelling sea sand dike breaching processes neglecting the
effect of waves on the test results. The results of two other tests performed on 6th
and 7th October 1994 are reported by Visser et al. (1996). Later, in 1996, large-scale
field experiments were carried out for investigating the three dimensional aspects of
breach growth in sand dikes (Visser, 1998).
Different failure modes and triggering conditions have been simulated in small scale
laboratory tests and large scale field tests with the aim to better understand the
breaching mechanisms and obtain data for calibration and validation of mathemat-
ical models. For further information on some of the experiments undertaken the
reader is referred to Tinney and Hsu (1961); Tingsanchali and Hoai (1993); Cole-
man et al. (2002); Rozov (2003); Zhu et al. (2006). Here we focus on the latest
experimental work carried out during the IMPACT project (IMPACT, 2004).
Several large scale field tests have been performed on non–cohesive and cohesive em-
Chapter 4. Current Approaches for Breach Modelling 45
Table 4.1: Information about five field tests undertaken in Norway.

No. Description of the test
1 6m high clay and silt embankment (D50 = 0.01 mm) with 25% clay
and less than 15% of sand to fail due to overtopping.
2 5m high non-cohesive embankment with D50 about 5mm (less than 5%
fines) to fail due to overtopping.
3 6m rock fill embankment with moraine core to fail due to overtopping.
4 6m rock fill embankment with moraine core to fail due to piping.
5 4.5m moraine fill embankment to fail due to piping.
Table 4.2: Information about five laboratory tests undertaken at HR Wallingford

(Morris, 2005).
Tests Description of the test
9 Overtopping failure of homogeneous, non–cohesive embankments for
various grain size, breach location and embankment geometry.
8 Overtopping failure of homogeneous, cohesive embankments with
different material type and grading, compaction, water content
and geometry.
5 Piping failure through homogeneous, cohesive embankments.
bankments failing from overtopping and piping failure modes. The five large scale
field tests (see Figure 4.1) conducted in Norway are summarized in Table 4.1. The
information about the twenty two laboratory tests conducted at a 1:10 scale at the
HR Wallingford laboratories are summarized in Table 4.2.
These field and laboratory results lead to extensive validation of the numerical mod-
els being used and under development. In contrast to other experimental findings,
during the field tests in Norway it was concluded that the breach shape is rectan-
gular (vertical walls), not trapezoidal as often assumed previously. Morris (2005)
suggest that the trapezoidal shapes might develop after the breaching is finalized,
during the drying process of the embankment material. However, it is important
to mention that in all laboratory experiments and field tests an initial rectangular
breach channel was imposed as initial condition, which might have influenced the
further development of the breach shape.
46 Section 4.2. Experimental work
(a)
(b)
Figure 4.1: Development of the breach in the field tests in Norway for (a) piping
failure and (b) overtopping failure (Vaskinn, 2003).
4.3 Empirical equations

An alternative, conservative way of estimating dam breach characteristics is the use
of empirical predictor equations obtained from regression analysis of historical dam
failure events. They are often used for providing ultimate breach extent information
to physically based models (see e.g. Table 4.3), or as a comparative tool for the
output obtained from physically based methods. However, their application to a
particular dam should be used cautiously making sure that dam and reservoir char-
acteristics are within the range of the dataset upon which the empirical equation is
based.
Numerous empirical formulations have been developed for predicting dam breach
characteristics and peak outflows based on hydraulic and geometrical characteristics
of dams and reservoirs e.g. reservoir storage, depth and volume of water in the
reservoir at the failure time, dam height, etc. In the following, some of the available
empirical equations are described.
The U.S. Bureau of Reclamation (1988) suggested that for earthfill dams the ulti-
mate width of a rectangular dam breach shape equals three times the initial water
depth in the reservoir measured to the breach bottom elevation assumed to be at
the stream bed elevation. This relationship was used as a guideline in the National
Weather Service Simplified Dam Break Model (SMDBRK).
Bavg = 3hw (4.1)
where, Bavg is the ultimate average breach width and hw is the water depth in the
reservoir initiating the failure.
Hagen (1982) analysed 18 historical events of dam failure from overtopping. The
product of volume Vw and depth hw of water in the reservoir triggering the failure,
hereafter called dam factor DF , was plotted versus peak discharge Qp . As a result,
Hagen provided an envelope equation:
Qp = 1.205(Vw hw )0.48 (4.2)
The peak outflow discharge varied from 730 to some 83,000m3 /s while the dam factor
Vw hw ranged from 2.9×106 to 4.8×1010 m4 .
MacDonald and Langridge-Monopolis (1984) analysed 42 dams: 30 earthfill and 12

non–earthfill dams (rockfill and other dams with protective concrete surface layers
or core walls). The height of the dams varied from 6 to 93m. The product of water
volume outflow Vout from the reservoir with the depth hw−b was plotted versus
the peak outflow for 18 earthfill dams and the following envelope and regression
equations were obtained:
Qp = 3.85(Vout hw )0.41 (4.3)
48 Section 4.3. Empirical equations
Table 4.3: Estimation of breach width and time of failure as a function of dam height
Hd , crest length CL and different dam types.
Type of dam Breach width √ Formation Time
Earthfill & rockfill 3Hd √Hd
Concrete gravity CL/3 Hd
Arch CL Hd /50
Qp = 1.15(Vout hw )0.41 (4.4)

The volumeVout includes the reservoir storage at the failure time Vw plus the inflow
associated with the rising limb of the inflow hydrograph.
Singh and Snorrason (1982, 1984) analysed some historical earthfill dam failure
events due to overtopping. One finding of this analysis was the identification of a
strong correlation between breach width and dam height. For most of the cases the
breach width fell between the two bounding lines: B = 2hd and B = 5hd . The max-
imum water depth above the dam crest triggering failure varied in a range from 0.15
to 0.61m, and the failure time from inception to completion of the breach, fell in the
range between 0.25 to 1 hour. Singh and Snorrason (1982, 1984) analysed the hypo-
thetical failure of eight earthfill dams in Illinois (USA) with a height hd ranging from
4.4m to 28m. Six out of eight dams were simulated for failing from overtopping and
two of them were simulated for failing from piping assuming a Probable Maximum
Flood inflow in the reservoir. US Army Corps of Engineers HEC-1 and National
Weather Service BREACH (Fread, 1988) models were used for predicting the peak
outflow. In both models, the maximum failure time, maximum breach depth and
width, and initial water depth above the dam crest (for overtopping failure) were
predefined by the user equal to 0.5 hours, hd , 4hd and 0.15m, respectively. The
obtained peak outflows were plotted versus the reservoir storage S and dam height
hd and the following regression equations were obtained:
Qp = 1.78S 0.47 (4.5)
Qp = 13.4h1.89
d (4.6)
Costa (1985) analysed 31 historical dam failure events. The height of the dams
considered in the analysis varied from 1.8m to 83.8m, while the volume of the reser-
voir at the failure time ranged from 3.8×103 m3 to 7.0×108 m3 . No distinction was
made between different failure modes and dam types. The following envelope curves
encompassing 29 data points were derived:
Qp = 1.12Vw 0.57 (4.7)
Qp = 2.63(Vw hd )0.44 (4.8)

100000
Peak discharge [m /s]
10000
3
1000
100
10
1.0E+03 1.0E+05 1.0E+07 1.0E+09 1.0E+11
4
Dam factor [m ]
Costa MacDonald & Langridge-Monopolis Froehlich
Figure 4.2: Dam factor vs. peak discharge for dams analysed by MacDonald and
Langridge-Monopolis (1984), Costa (1985), and Froehlich (1987).
while, the regression relations are proposed as follows:
Qp = 10.5hd 1.87 (4.9)
Qp = 1.27Vw 0.48 (4.10)
Qp = 0.981(Vw hd )0.42 (4.11)

There is no significant difference between the regression equations obtained by Mac-
Donald and Langridge-Monopolis (1984) and Costa (1985), though the first authors
include in the analysis only earthfill dams, while the second author includes earthfill
and non–earthfill dams in his analysis.
Froehlich (1987) analysed 22 embankment dam failures with hw ranging from 3.4m
to 77.4m and Vw ranging from 0.1 to 310 million m3 . The regression equation ob-
tained is expressed as:
Qp = 0.607(Vw 0.295 hw )1.24 . (4.12)
In all analysis, more than 90% of the analysed dam failure events belong to dams in
the USA. In Figure 4.2, peak outflow regression relations suggested by MacDonald
and Langridge-Monopolis (1984); Costa (1985); Froehlich (1987) are presented along
with measured peak outflows.
Broich (1998) analysed 39 embankment dam failures and obtained two regression
50 Section 4.4. Physically based models
equations which relate peak outflow to the depth and volume of water in the reservoir
at the initial failure time in the following way:
Qp = 255.859(Vw hw )0.449 (4.13)
Qp = 72.611(Vw hw 4 )0.256 , (4.14)
where Vw is in 106 m3 .
Wahl (2004) carried out an uncertainty analysis of the empirical equations using
a compiled database of 108 dam failure events. In the analysis no distinction was
made between different failure modes (the same stands for Costa (1985) analysis). It
is clear that for peak outflow prediction based on dam height and reservoir storage,
the distinction between different failure modes is important. The water depth in the
reservoir can be lower than the dam height in case of a piping failure mode, and
higher than the dam height during an overtopping failure mode.
4.4 Physically based models

Cristofano (1965) as reported by Singh (1996) developed a dam breach model with
the following characteristics: Breach Shape is considered trapezoidal during entire
breach development; Bottom Width of the breach channel is assumed constant; Top
Width is enlarged with the maximum width defined by the maximum side angle
equal to the angle of repose of the compacted material of the dam. The initial
channel is supposed to be parallel to the dam bottom and the (bottom) slope of the
breach channel is assumed equal to the internal friction angle.
The flow through the channel is calculated using the broad–crested weir flow equation
with a discharge coefficient ranging from 2.9 at the beginning of failure to 2.2 at peak
flow
Q = CLH 1.5 , (4.15)
where C is the discharge coefficient, H is the water depth above the breach bottom,
L is the length of the breach channel, and Q is the discharge.
Sediment erosion is deduced by relating the force of the flowing water through the
breach channel to the resistive shear strength of the soil particles acting on the
bottom surface of the breach channel. The rate of erosion is expressed as a function
of the rate of change of water flowing through the breach.
Harris and Wagner (1967) presented a simple model for the breaching of embankment
dams due to piping and overtopping. The breach shape is assumed parabolic with
top width of the parabolic shape Btop assumed to be 3.75 times the depth D (see
Figure 4.3). The flow through the breach channel, assuming critical flow over a
Cross section Plan view
Upstream slope
Downstream slope
abutment
Btop
Figure 4.3: Parabolic breach shape.
broad–crested weir and accounting for the parabolic breach shape, is expressed in
SI units as
Q = 5.54H 2.5 . (4.16)
Three modes of sediment erosion were adopted in the model: bed load, suspended
load and saltation load. The Schoklitsch bed load formula (Schoklitsch, 1917) was
modified to estimate the bed load transport along the breach channel. The rate of
sediment transport was computed as
G = 86.7S 1.5 Qd−0.5 ,
where G is the sediment transport rate, S is the hydraulic slope, and d is the soil grain
diameter (expressed in English units). The suspended load had to be estimated based
on the turbulence intensity and fall velocity. A constant sediment concentration is
assumed and no slope stability or failure mechanism is considered.
Brown and Rogers (1977, 1981) developed the BRDAM model that simulates the
breach erosion of earthfill dams due to piping or overtopping. The initial channel is
created at the dam crest, when simulating overtopping failure. The parabolic breach
shape of the notch is assumed constant during the whole process of dam breaching.
Some 45◦ side slopes and a channel slope of 5◦ up to 20◦ is assumed, depending on
the fill material of the dam. The outflow during overtopping failure is calculated by
the broad–crested weir formula:
√
Q = gBtop yc 1.5 . (4.17)
The critical depth of the flow over the crest is taken to be 75% of the water depth
above the breach bottom level. Like Harris and Wagner (1967), the Schoklitsch
(1934) bed load formula is used to estimate the solid load which is transported
along the breach channel bed assuming constant sediment concentration. The slope
stability and lateral erosion mechanisms are not taken into account in the model.
Ponce and Tsivoglou (1981) presented a model for gradual failure of earthfill dams
due to overtopping. The initial location of the breach is assumed at a weak point
on the crest and downstream face. The top width of the breach opening is related
to the flow rate.
Lou (1981) developed a mathematical model that is similar to the Ponce and Tsivoglou
(1981) model. The breach grows in space and time until an equilibrium section is
attained. The most efficient stable section derived for describing the breach shape
is expressed as:
π
y = hmax cos( x),
Btop
where y is the depth at a distance x, hmax is the maximum depth at the center,
and x is the distance from the center. De Saint–Venant equations are applied to
model the unsteady flow of water during breaching of an earth dam. The dam is
divided into n-reaches, with the discharge flowing from the reservoir routed down-
stream from the dam. During dam overtopping, the sediment discharge is computed
using three approaches: 1)Du Boy’s equation (Du Boys, 1879) for bed load sediment
transport combined with Einstein’s equation for suspended load sediment transport
to calculate the total sediment transport; 2) a sediment transport formula derived as
a function of failure duration, erodibility index, and water velocity, 3) Cristofano’s
equation for computing the erosion rate. Du Boy’s equation is written as:
qbl = Kτ0 (τ0 − τc ), (4.18)
where qbl is the rate of soil transported as bed load per unit width, (τ0 − τc ) is the
excess shear stress, and K is a parameter dependent on the thickness of the moving
bed layer, velocity of the layer as well as critical shear stress τc .
Einstein’s (1942) suspended load rate per unit width can be expressed as

30.2h
qsl = Ks 2.303log I1 + I2 , (4.19)
d65
where h is the flow depth, I1 and I2 are integral quantities depending upon the flow
depth, d65 is a representative grain size diameter, Ks = 11.6Ca us abl , abl is thickness
of the bed layer where the suspended load starts, us = (gRS)0.5 and Ca is the
sediment concentration at a known distance from the bed.
Lou (1981) as described in Singh (1996) derived an expression for solid transport in
embankments by relating the kinetic energy to the erosion process, with the following
expression:
Ce
Er = td u 4 ,
12
where Ce is a constant, td is the duration of failure, u is the water velocity and Ms is
the mass of soil lost during erosion. Slope stability mechanisms are not considered,
while the lateral erosion stops once the peak flow is reached.
The National Weather Service model BREACH developed by Fread in 1984 and
later revised in 1988 (Fread, 1988) simulates overtopping and piping failure modes
for cohesive and non–cohesive embankment dams. The dam may consist of two
materials with different properties in the inner core and outer zone of the dam or
otherwise be homogeneous. Erosion initially occurs along the downstream face of
the dam where a small rectangular–shaped rivulet is assumed to exist along the dam
face. The sides of the breach channel are assumed to collapse when the depth of the
breach reaches a critical value and then the breach is transformed into a trapezoidal
shape channel. Maximum values of breach dimensions are defined by the user (dam
height, maximum top and bottom breach width limited from the valley cross sec-
tion). The flow through the breach channel is calculated by the broad–crested weir
equation and the orifice flow relationship (for piping). The sediment transport rate
is calculated by the Meyer-Peter and Müller (1948) sediment transport relation as
modified by Smart (1984) for steep channels.
DEICH (Dam Erosion with Initial breach Characteristic) models are developed at
the University of the Federal Armed Forces Munich (Broich, 1998). Depending on
their calculation approach, they are named: DEICH A (analytical model), DEICH P
(parametrical model), and DEICH N1/N2 (1D/2D numerical model), respectively.
These models calculate the overtopping failure of cohesive and non–cohesive embank-
ment dams with or without covering core. Initially a channel or pipe is created at the
dam crest. The breach develops parallel to the dam crest. A constant trapezoidal
breach shape with side slopes of 45◦ and a constant relationship between lateral and
bottom erosion is assumed. Breach discharge is calculated using the broad crested
weir formula enhanced to take into account backwater effects. DEICH A and DE-
ICH P calculate the total sediment load, while DEICH N1/N2 distinguish between
bed load and suspended load. DEICH A uses Ponce–Tsivoglou’s transport formula
(Ponce and Tsivoglou, 1981), a combination of the Exner with the Meyer-Peter and
Müller (1948) equation. Slope stability mechanisms are not considered in all models.
Several commercial hydrodynamic modelling packages have incorporated breach

models. However, most of the time the final breach characteristics and rate of
breach growth have to be predefined by the user who often bases his decision on
expert judgment or empirical equations that relate breach parameters to dam and
reservoir characteristics. Mike11, the Danish Hydraulic Institute’s modelling pack-
age, includes a simple dam breach model. In this package the failure of the dam
can initially take place from overtopping or as a piping failure. The breach will
continue to develop until it has reached the breach geometry limit, which is defined
by final bottom depth, width and the breach slope on each side of the breach. The
development of the breach can take place in two different ways:
1. Time Dependent: The development of the dam breach is specified by the user
in terms of breach level, width and slope as functions of time.
Initial width
5 4 1 4 5
Initial crest level 2 Maximum breach height
Reference level
Figure 4.4: Breach development in Sobek 1D2D, first in vertical direction (step 1,
2, and 3) and then in horizontal direction (step 4 and 5).
2. Erosion Based: MIKE 11 calculates the breach development by using a sedi-

ment transport formula.
In Sobek 1D2D, WL | Delft Hydraulics’ modelling package, a dam break can be

modelled either in 1D or 2D, using the 1D Flow Dam Break Reach (mainly for flood
defense structures along the water body) or the 2D-Breaking Dam Node (mainly for
water retaining structures) respectively. The 1D Flow Dam Break Reach model (Ver-
heij, 2002) can simulate breach development in clay or sandy dikes. The breach shape
is considered rectangular during the entire breach development. The breach will first
deepen in vertical direction only, until the maximum breach depth is reached, after
which the breach grows in lateral direction only. Figure 4.4 presents the different
breach development stages during the breach growth. The breach development can
take place as either:
1. Irrespective of actual occurring flow conditions in the breach, or
2. Depending on the occurring hydraulic conditions and the parameters defined.
In the first case, the breaching development is user defined. The user must provide
the initial breach width, and initial crest level of the dam, maximum breach depth,
the starting time of the breaching process, the duration of the breaching process,
the maximum breach–width or the maximum duration for reaching the maximum
breach width. The discharge through the breach channel is calculated using stan-
dard weir formulas.
In the second case, the user must supply information about the lowest level that the
breach will reach, the maximum flow velocity for which erosion does not occur, the
starting time of the breaching process and the duration of the breaching process. In
this case the increase in crest width as a function of time is not known beforehand,
but depends on the occurring hydraulic conditions and the parameters defined.
For information about other breach models the reader is referred to Table 4.4 and
Singh (1996); Singh and Quiroga (1987); Loukola and Huokuna (1998). To fur-
ther understand the mechanisms associated with embankment failure, we examine a
database of the historical events. In the following sections, we present a methodol-
ogy for estimating breach shape and breach outflow characteristics based on a data
mining approach.
Table 4.4: Breach model characteristics. 56
Model Geometry Hydraulics Sediment
Cristofano (1965) Trapezoidal with side angle Broad-crested weir Developed empirical
equal to angle of repose of formula
the compacted material
Harris and Wagner (1967) Parabolic with top width Broad-crested weir Modified bed load
3.75 times the depth Schoklitsch formula
and suspended load formula
based on turbulence and
fall velocity
Brown and Rogers (1981) Parabolic with side slope of 45 Broad-crested Modified bed load,
BRDAM model Schoklitsch formula
Ponce and Tsivoglow (1981) Top width flow rate relation de Saint-Venant Exner with
Meyer-Peter-Muller
Lou (1981) Most effective stable section de Saint-Venant 1. Du Boy & Einstein,
(Cosine curve shape) 2. Lou,
3. Christofano.
Nogueira Effective shear stress section de Saint-Venant Exner equation with
(Cosine curve shape) Meyer-Peter-Muller
Fread (1988) Rectangular and trapezoidal Broad crested weir Meyer-Peter-Muller
NWS BREACH orifice flow modified by Smart
Singh & Quiroga (1987) Trapezoidal Broad crested weir Einstein-Brown
BEED
Loukola & Houkuna (1998) Trapezoidal Broad crested weir Meyer-Peter-Muller
EDBREACH
Broich (1998) Diffusion approach de Saint-Venant Several transport eqs.
DEICH N1/N2
Section 4.4. Physically based models
Chapter 5
Data Mining Techniques in

Dam Breach Modelling
First get your facts, then you can distort them at your leisure.
Mark Twain
5.1 Introduction
While capabilities of computational modelling systems are continuously advancing,
data from sensor measurements, satellites and computational operations are rapidly
expanding. Hence, data mining techniques are becoming increasingly more popular
nowadays and are being applied to different fields including water science and en-
gineering (Minns and Hall, 1996; Price, 2000; Babovic et al., 2001; Mynett, 2002,
2004a,b, 2005; Solomatine, 2002).
Data mining techniques can be used to extract knowledge from large amounts of
data. Despite the scarcity of reliable data related to historical dam failure events, it
was hypothesized that present day data mining techniques might prove to be useful
to some extent for predicting breach characteristics and peak outflow resulting from
dam failure. Ultimate breach width and dam peak outflow might be predicted based
on dam and reservoir characteristics at the failure time. In this chapter, we present
a methodology for predicting dam failure characteristics and peak outflow using a
range of data mining techniques, including linear and non–linear regression methods.
The algorithms and underlying theory are introduced in the following section.
58 Section 5.2. Main principles of data mining techniques
5.2 Main principles of data mining techniques
The engineering of intelligent machines has been a focus of scientific research for a
long time. Already in 1950 Alan Turing (1950) proposed a test that would measure
machine’s capability to perform human–like conversation. It was John McCarthy
(1956) who first introduced the term ‘Artificial Intelligence (AI)’ as the topic of the
Dartmouth Conference in 1956, the first conference devoted to this subject, defin-
ing AI as “the science and engineering of making intelligent machines, especially
intelligent computer programs”. The ultimate goal of the artificial intelligence is to
imitate human intelligence.
The field of machine learning is concerned with the question of how to construct com-
puter programs that automatically improve with experience. It is a field based on
computer science, psychology, neuroscience and engineering. In recent years many
successful machine learning applications have been developed, ranging from data
mining programs, to information filtering systems, all aiming to program comput-
ers in the way that they could “learn to improve automatically with experience”
(Mitchell, 1997). There are different manners of machine learning, namely: super-
vised, unsupervised and reinforcement learning.
Supervised learning is a machine learning technique that creates a function based on

training data of input–output pairs. The output of the function can be a continuous
value (called regression), or a class label of the input object (called classification).
The objective is to correctly predict the output given a new input after having seen
a number of training examples.
In the unsupervised learning the model is fit to input data with no a priori output.
The learner’s task is to represent the inputs in a more efficient way, as clusters, cate-
gories or reduce the dimensions. Unsupervised learning is used for data compression,
outlier detection, classification, etc.
Reinforcement learning (RL) is another learning technique much closer to supervised

than unsupervised learning. It concerns learning from the consequences of action.
The learner receives information that what it did is appropriate or not. Contrary
to supervised learning where the learner is ‘advised’ on exactly what it should have
done, reinforcement learning only says that the behavior was inappropriate and (usu-
ally) how inappropriate it was.
In this chapter we use the supervised learning techniques and algorithms, namely lin-
ear and non–linear techniques, for extracting information from the available dataset
of historical dam failure events.
Chapter 5. Data Mining Techniques in Dam Breach Modelling 59
5.2.1 Linear regression

A linear model is written as
Y = Xw + , (5.1)
where w is the k–dimensional parameter vector, i is the error vector usually as-
sumed to be normally distributed with zero mean. The objective is to compute the
parameter vector w = (w1 , w2 , ..., wk ) that best fits the data by the linear model.
The linear model should be robust and not exhibit an unstable behavior that can
greatly amplify even small changes in the dataset.
At the end of the 18th century, Gauss (1809) and Legendre (see Stigler, 1981) pro-
posed the method of Least Squares (LS) regression that consists of minimizing the
sum of the squared residuals with respect to the coefficient vector w as
k
X
minimize ri2 , (5.2)
w
i=1
where the residuals ri are the differences between what is actually observed and
what is estimated. However, this method is known to be highly influenced from
the presence of outliers in the dataset. Thus other versions of this estimator were
proposed such as replacing the squared residuals by the median of squared resid-
uals (Rousseeuw, 1984; Rousseeuw and Leroy, 2003). The least absolute deviation
method introduced in 1757 by Boscovich (Birkes and Dodge, 1993) (also known as
least absolute value method) and least median of squares linear regression methods
are considered better methods for regression and often recommended as robust or
outlier resistant alternatives of LS, though they can exhibit instabilities for small
changes in the data.
The least absolute value method, presented by Edgeworth (1887) is expressed as

k
X
minimize |ri |. (5.3)
w
i=1
In contrast to the LS method, it minimizes the sum of the absolute values of the
residuals. Here, we use the Least Median of Squares (LMS) method (Rousseeuw,
1984) that is considered a less sensitive and more robust fitting technique compared
with the simple linear regression method. It is also considered to be the most widely
used robust estimator. It minimizes the median of the squared residuals with respect
to the coefficient vector w:
minimize med ri2 . (5.4)
w i
Thus the largest residuals in the sample (i.e. those whose absolute values are larger
than the median) are ignored, making this technique more robust in the presence of
outliers or noisy data. At least 50% of the data would need to be corrupt in order
to skew the result, while for ordinary least squares a single corrupt data sample can
already give the resulting regression line an excessively large slope.
5.2.2 Artificial Neural Networks (ANNs)

McCulloch and Pitts (1943) were the first to introduce the concept of how the brain
could produce complex patterns by using basic cells called neurons that are con-
nected to each other. McCulloch and Pitts (1943) presented an artificial neuron
model with binary input and output and an activation threshold.
Artificial Neural Networks (ANNs) are envisioned to have similarities to the human
brain functioning. As the latter are built of very complex webs of interconnected
neurons, ANNs are built to form complex interconnected sets of units, each of which
takes a number of real–valued inputs and produces a single real–valued output (Fig-
ure 5.1). They can be used for approximating discrete–valued, real–valued and
vector–valued target functions.
The perceptron is a type of artificial neural network invented in 1957 by Rosenblatt

(1958). The perceptron (Figure 5.2) takes a vector of real–valued inputs, calculates
a linear combination of these inputs, and then outputs results based on some ac-
tivation function. Using the McCulloch and Pitts (1943) threshold function, the
output is 1 if the result is greater than the threshold value and 0 (or −1) otherwise.
Therefore, given inputs x1 through xn , the output o (x1 , ..., xn ) computed by the
perceptron is

1 if w0 + w1 x1 + w2 x2 + ... + wn xn > 0
o (X1 , ..., Xn ) = , (5.5)
0 otherwise
where each wi is a real–valued constant referred to hereafter as weight, determining

the contribution of input xi to the perceptron output. Teaching a perceptron network
involves establishing values for the weights w0 , ..., wn . Depending on the problem,
one of the activation functions showed in Figure 5.3 can be used.
Several algorithms are known to solve this learning problem e.g. perceptron rule,
delta rule, etc. (Mitchell, 1997). The perceptron rule finds a successful weight vector
when the training examples are linearly separable. It revises the weight wi associated
with input xi according to the rule:
wi ← wi + ∆wi , (5.6)
where
∆wi = η (t − o) Xi . (5.7)
Here t is the target output for the current training example, o is the output generated
by the perceptron, and η is a positive constant called the learning rate. The role of
hw
Vw Qp
hw
Hd
Vw Qp
Hd
Input layer Hidden layer Output layer
Figure 5.1: Architecture of a three layer neural network with three inputs and one
output.
Input layer Hidden layer Output layer
X1
X1
w1
w1
Inputs Inputs
X2 w2 X2
∑
w2
∑ Output Outpu
Treshold Sigmoid activation function

wn Treshold Sigmoid activation function
wn
Xn
Xn
Figure 5.2: A perceptron.
Subcritical flow Supercritical flow
Different erosion rates
(a) (b) (c) (d)
Figure 5.3: Commonly used activation functions: (a) Binary (threshold step func-
tion), (b) Sigmoid, (c) Piecewise Linear, and (d) Gaussian activation functions.
the learning rate is to control (lessen) the degree to which weights are changed every
time step. The rule converges within a finite number of applications of the perceptron
training rule to a weight vector that correctly classifies the training data, provided
they are linearly separable and provided a sufficiently small η is used (Mitchell,
1997). Contrary to the perceptron rule that can fail to converge when the data are
not linearly separable, the delta rule (or sometimes the LMS rule, Adaline rule, or
Widrow-Hoff rule - after its inventors) converges towards a best–fit approximation
to the target concept in that case. The delta rule uses the gradient descent to
search the hypothesis space of possible weight vectors to find the weights that best
fit the training data. This is done by specifying a measure for the training error of
a hypothesis (weight vector) relative to the training examples such as:
1 X
E(w) ≡ (to − od ), (5.8)
2
m∈M
where m is the set of training examples, to is the target output for training example
m, and o is the output of the linear unit for training example m. To understand
the gradient descent algorithm, we visualize the entire hypothesis space of possible
weight vectors and their associated E values, as illustrated in Figure 5.4.
During the gradient descent search, a weight vector is determined that minimizes
E by starting with an arbitrary initial weight vector, then repeatedly modifying
the weights in small steps, during the learning phase. During each step, the weight
vector is altered in the direction that produces the steepest descent along the error
surface shown in Figure 5.4. The direction can be found by computing the derivative
of E with respect to each component of the vector w. This vector derivative is called
the gradient of E with respect to w, written ∆Ew.

∂E ∂E ∂E
∆E(w) ≡ , , ..., . (5.9)
∂w0 ∂w1 ∂wn
The gradient specifies the direction that produces the steepest increase in E. The
negative of the vector therefore gives the direction of steepest decrease. Single per-
ceptrons can only express linear decision surfaces. In contrast, multilayer perceptron
(MLP) networks trained with a Backpropagation algorithm are capable of capturing
Figure 5.4: Hypothesis space of possible weight vectors and their associated error
values E (Mitchell, 1997).
a variety of non–linear decision surfaces. The Backpropagation algorithm is the most

commonly used ANN learning technique. Gradient descent provides the basis of the
algorithm. It has been found to be appropriate for many problems, as summarized
from Mitchell (1997), as follows:
Instances are represented by many input attributes being correlated or inde-

pendent of each other.
The target function output might be discrete–valued, real–valued or a vector

of several real–or discrete–valued attributes.
Training examples might contain errors or noisy data.
Training time is relatively long compare to other methods, depending on the

number of weights in the network, the number of training examples considered,
and the settings of various learning algorithm parameters.
Despite the biological motivation of the neural networks, the learned neural
networks are less easily communicated to humans and should be used when
the readability of the results is not important.
Backpropagation algorithm uses a gradient approach descent to tune network pa-

rameters to best fit the training set of input–output pairs.
5.2.3 Instance Based Learning (IBL)
In contrast to above described learning methods which produce a generalization as

soon as the data are introduced, there are the so–called instance based learning
methods or ‘lazy’ learning methods that postpone the description of the target func-
tion until a new instance (query) is requested.
Each time a new query instance is encountered, its relationship to the previously
stored examples is examined in order to assign a target function value for the new
instance. These methods fit the training data only in a region around the location
of the query instance instead of estimating the target function once for the entire in-
stance space. Therefore, even a very complex target function can still be described,
by constructing it from a collection of less complex local approximations using the
instance based learning methods (Mitchell, 1997). These methods can construct a
different approximation to the target function for each specific query instance re-
quiring classification.
However, there are disadvantages associated with these kinds of methods such as:
the cost of classifying new instances can be high since the computation takes place
at classification time rather than when the training examples are first encountered.
And since all attributes of the instances are considered when attempting to retrieve
similar training examples from memory, the instances that are truly most ‘similar’
may well be a large distance apart.
The instance based learning methods include Nearest neighbor, locally weighted
regression methods, and case-based reasoning methods. The Nearest neighbor
method (IBk) chooses the closest points near the query point and uses its output.
The weighted average model averages the outputs of neighbouring points, inversely
weighted by their distance to the query point. The locally weighted learning re-
gression model (LWL) weighs the outputs of neighbouring points according to their
distance to the query point and then performs a local linear regression based on the
weighted data.
The k-NEAREST NEIGHBOR method

The most basic instance-based method is the k–Nearest neighbor algorithm that
computes the classification of each new query instance as needed. It assumes all
instances as points in the n–dimensional space <n and the nearest neighbours of an
instance are defined in terms of the standard Euclidean distance.
Let an arbitrary instance x be described by the feature vector
ha1 (x), a2 (x), ..., an (x)i, (5.10)

where ar (x) denotes the value of the r–th attribute of instance x. Then the distance
between two instances xi and xj is defined to be d(xi , xj ), where
v
u n
uX 2
d(xi , xj ) ≡ t (ar (xi ) − ar (xj )) . (5.11)
r=1
In Nearest neighbor learning the target function may be either real–valued or

discrete–valued. When approximating real–valued (continuous valued) target func-
tions, the k–Nearest neighbor algorithm calculates the mean value of the k near-
est training examples, where:
Pk
f (xi )
F (xq ) ← r=1 (5.12)
k
is a function that approximates the function f (x) and that returns the mean value
of f among the k training examples nearest to xq .
For approximating a discrete–valued target function, the k–Nearest neighbor

algorithm denotes the k instances x1 , ..., xk from the training examples that are
nearest to the query instance xq and returns the most common value rather than
the mean value of the k nearest training examples, where
v
u k
uX
F (xq ) ← argmaxt δ (v, f (xi )) (5.13)
v∈V r=1
is a function that approximates the function f (x) and that returns the most common
value of f among the k training examples nearest to xq . If k is chosen equal to 1
then the 1–Nearest neighbor algorithm assigns to F (xq ) the value f (xi ) where
xi is the training instance nearest to xq . For larger values of k, the algorithm assigns
the most common value among the k nearest training examples.
The algorithm can better be explained through the graphical presentation of the
Voronoi diagram (Figure 5.5). The instances are points in a two–dimensional space
with the target function being boolean valued with positive and negative training
examples shown by ‘+’ and ‘-’ respectively. The 1–Nearest neighbor algorithm
classifies the query instance xq , as positive, whereas the 5–Nearest neighbor
algorithm classifies it as a negative example. The diagram shows the shape of the
decision surface induced by 1–Nearest neighbor over entire instance space.
The distance between instances is calculated based on all attributes of the instance.
As such it is important to select the most important defining attributes. Large
number of non important (irrelevant) attributes might influence the distance between
neighbours and as a result lead to misclassification. The k–Nearest neighbor
algorithm is robust to noisy training data and quite effective when it is provided a
hc hc H HH
d
Hd
---
-
+ + - + -
+ +
xq + xq
+ - -
+
+ +
- -
(a) (b)
Figure 5.5: (a) Example of Nearest neighbor and (b) Voronoi diagram.
sufficiently large set of training data.
Distance-weighted NEAREST NEIGHBOR

This algorithm weighs the contribution of each of the k neighbors according to their
distance to the query point xq giving greater weight to closer neighbors. Thus the
algorithm formulation for approximating a real–valued target function changes to
Pk
r=1 wi f (xi )
F (xq ) ← Pk . (5.14)
r=1 wi
In a similar way, the algorithm formulation for approximating a discrete–valued

target function changes to
k
X
F (xq )←argmax wi δ(v, f (xi )), (5.15)
v∈V r=1
where
1
wi ≡ 2. (5.16)
d(xq , xi )
Considering all training instances while approximating the function without the cost
of misleading is possible with distance weighting as the very distant examples will
have very little effect on F (xq ). The only disadvantage of considering all examples
is the ‘longer’ processing time. Furthermore, by taking the weighted average, the
impact of noisy training examples will be smoothed.
Locally weighted regression

The locally weighted regression method constructs an explicit approximation to f
over a local region or only on data near the query point xq . The contribution of each
training example is weighted by its distance from the query point. The distance-
weighted training examples are used to form a local approximation to f . The target
function in the neighborhood surrounding the instance xq can be approximated
using a linear function, a quadratic function, a multilayer neural network, or some
other functional form. Here, let us consider the case when the target function f is
approximated near xq using a linear function of the form
F (x) = w0 + w1 a1 (x) + ... + wn an (x). (5.17)
As before, ai (x) denotes the value of the ith attribute of the instance x. Fitting the
local training examples is done through an error criterion E (xq ), expressed in three
different manners.
1. Minimize the squared error over just the k nearest neighbors:
1 X 2
E1 (xq ) ≡ (f (x) − F (x)) . (5.18)
2
x∈knrs
2. Minimize the squared error over the entire set of training instances M , while
weighing the error of each training example by some decreasing function N of
its distance from xq :
1X 2
E2 (xq ) ≡ (f (x) − F (x)) N (d (xq , x)). (5.19)
2
x∈M
3. Combine 1 ands 2:
1 X 2
E3 (xq ) ≡ (f (x) − F (x)) N (d (xq , x)). (5.20)
2
x∈knrs
Second criterion allows every training example to influence the classification of xq ,

but at the same time increases the computation time. The third criterion minimizes
the squared error only over the nearest neighbors and thus the computation cost
depends only on the number of k neighbors rather than the number of all training
examples.
5.3 Application
5.3.1 Data preparation and model evaluation
During this research, we use the Waikato Environment for Knowledge Analysis
(WEKA) modelling package (Witten and Frank, 2000). It is a comprehensive suite
68 Section 5.3. Application
of Java class libraries that implement many state–of–the–art machine learning and
data mining algorithms. Tools are provided for preprocessing data, creating mod-
els using a variety of learning schemes, and analyzing the resulting classifiers and
their performance. The techniques described in the previous section are applied to
the relatively small database of 108 historical dam failure events compiled by Wahl
(2004). For each dam failure event the information consists of:
1. Dam characteristics before failure occurred: dam type, height and length,
width at crest and bottom, upstream and downstream slope;
2. Reservoir characteristics: area and storage capacity;
3. Failure characteristics: failure mode, volume and depth of water at initial

failure time, final breach characteristics (breach shape, top, bottom or average
breach width and breach height, breach side slope factor), material removed
from the dam, breach formation time and peak outflow.
Data are often incomplete, lacking attribute values, noisy, containing errors and
outliers, or containing discrepancies in their information. Dataset does not provide
enough information for prediction of time needed to initiate a breach and the rate of
breach formation. Still, it might be useful for the prediction of breach characteristics
and peak outflow.
We apply data preprocessing for transforming the raw data into a format that will
be more easily and effectively used by data mining techniques. The first step in this
analysis consists of visual inspection of the data, dealing with missing and inaccurate
values and calculating statistical properties. Domain knowledge is applied for filling
in missing values and resolving inconsistencies as shown in Table 5.1. In the first row
for each of the dams some information is missing or is considered inaccurate. The
given water depth and volume in the reservoir at the failure time for Swift Montana
dam that failed from overtopping are substituted with values higher than the dam
height and reservoir storage respectively. For Kendall dam the missing values of
the water volume and depth in the reservoir at the time of the failure are assumed
to be larger than maximum dam specification, for the reason that it failed due to
overtopping.
The correlation between available attributes is calculated by dividing their covari-

ance by the product of their standard deviations. We measure the accuracy with
which one variable can be predicted from knowledge of another one by correlation
coefficient. Dam height, reservoir storage and the water depth behind the dam at
the initial failure time are used for the prediction of the breach characteristics and
peak outflow. However, due to the absence of adequate information for all the dams,
we use only a subset of the original database for experiments.
The problem of dealing with small datasets is that there is not enough data for train-
Table 5.1: Examples of missing values and inconsistencies.

Dam name (location) hd hw S Vw
Failure mode
Swift Montana (USA) 57.61 47.85 3.7E+07 3.7E+07
Overtopping ≥ 57.61 >3.7E+07
Kendall Lake (USA) 5.49 ? 7.28E+05 ?
Overtopping ≥ 5.49 7.28E+05 ≥ 7.28E + 05
ing and testing the model. Thus, we consider a 10-fold cross validation procedure
in this study. In the cross validation method, the result does not depend on how
the dataset is divided: every instance is used only once in testing. In this case, the
dataset is randomly reordered and then split into ten folds of (approximately) equal
size. During each iteration one fold is used for testing and the rest for training the
model. The test results are collected and averaged over all folds. The Root Mean
Square Error (RMSE) is used to measure the model performance:
v
u n
u1X 2
RM SE = t (yi − ŷi ) , (5.21)
n i=1
where yi and ŷi represent the observed and computed values for the calculated
attribute, whether it is average breach width or peak outflow. Lower RMSE values
guarantee better performance of the model.
5.3.2 Predicting breach width based on water depth behind

the dam
The first problem addressed in this section is the prediction of average breach width
based on the water depth behind the dam at the initial failure time. The database
(Wahl, 1998) provides a ‘training’ sample of previously observed cases in which both
water depth in the reservoir at initial failure time and the average breach width have
been recorded. Here, we construct a data–driven model for predicting likely average
breach width values for future dam failure events, when only the water depth in the
reservoir at the initial failure time is known.
For this analysis, only the overtopping failure mode is considered. In the dataset,
most of the 23 dams belong to earthfill, mixed earthfill and rockfill types with
or without revetment on the dam slopes. For 20% of the dams no information
is available on their material type. Table 5.2 shows the performance of different
algorithms applied to the available dataset. We apply the LMS (Section 5.2.1)
algorithm to the available data using the cross validation procedure and use the
equation obtained from LMS on the dataset (23 dams). The RMSE obtained is about
Table 5.2: Prediction of average breach width for 23 dams.

Method RMSE
LMS1: Bavg = 4.6hw − 5.7 15.0
USBR: Bavg = 3hw 22.5
180
160 LMS1
140 USBR
120
Predicted
100
80
60
40
20
0
0 20 40 60 80 100 120 140 160 180
Observed
Figure 5.6: Average breach width prediction based on the water depth in the reser-
voir.
30% lower than the error when using the USBR equation on the same dataset, thus
making former more applicable for predicting the average breach width in case of
dam failure from overtopping. Figure 5.6 shows the observed and predicted average
breach width computed using the LMS algorithm and the USBR equation. As can
be seen from Figure 5.6, the USBR prediction tends to underestimate the average
breach width, while LMS prediction results in a correct or overestimated average
breach width.
Since in the dataset, 84% of the dams failing from overtopping belong to the low
and medium dam categories, another experiment is performed for predicting breach
width excluding the high dams from the dataset. It can be seen from Table 5.3
and Figure 5.7 that the LMS model performance increases by 30%, while the USBR
equation performs 40% better than in the previous experiment. Both LWL and LMS
algorithms perform 20% and 10% better, compared to the USBR equation.
From both experiments different equations (LMS1 and LMS2) are obtained, the
second one demonstrating better performance on the dataset used for this analysis.
It can be seen that for the prediction of breach width for non–high earthfill dams
Table 5.3: Prediction of average breach width for 18 low and medium dams.
Method RMSE
LWL using inverse–distance weighting kernels 9.6
LMS2: Bavg = 3.8hw − 1 10.9
USBR: Bavg = 3hw 12
50
LWL
40 LMS
USBR
Predicted
30
20
10
0
0 10 20 30 40 50
Observed
Figure 5.7: Average breach width prediction based on the water depth behind the
dam at initial failure time for low and medium dams.
the LMS2 and USBR equations are performing almost the same with very small
difference in RMSE.
5.3.3 Predicting peak outflow

In this section we address the issue of estimating the peak outflow based on dam
height and reservoir storage. A total of 27 dams are included in the analysis for es-
timating peak outflow as a function of dam height. No distinction is made between
different dam materials and failure modes. We consider this assumption as valid due
to small dataset size, though the breach outflow hydrograph and consequently the
peak outflow is likely to be affected by geotechnical characteristics of the embank-
ment.
As can be seen from Table 5.4, models created with IBk and MLP perform compa-
rably better than the Singh & Snorrason equation in case of peak outflow prediction
based on dam height only. Two neighbours are used for classification in case of IBk
application, while no changes are made in the default settings for the MLP network
in WEKA. The decision to use default parameters is motivated by the scarcity of
Table 5.4: Prediction of peak outflow Qp = f (hd ).

Method RMSE
IBk–2 neighbours 1000
MLP 1367
Singh and Snorrason (1982)-Eq. 4.6 2342
12000
IBk
10000
MLP
8000
Predicted
6000
4000
2000
0
0 2000 4000 6000 8000 10000 12000
Observed
Figure 5.8: Peak outflow prediction based on dam height.
data and therefore the difficulty of having a separate validation set for parameter
estimation. Though IBk RMSE estimate is high, it is 57% smaller compared to the
one obtained from the Singh & Snorrason’s empirical equation. The performance of
IBk and MLP model and the Singh & Snorrason equation is presented in Figure 5.8.
Due to missing data, a smaller dataset of 22 dams is used for predicting peak out-
flow based on reservoir storage and dam height. The performance of the developed
data mining models is compared with the available empirical formulations of Singh
& Snorrason, Hagen and Costa (Table 5.5).
Using the same data mining algorithms the results obtained by including the reser-
voir storage in the analysis are compared to the empirical equations of Costa (Figure
5.9). Here, better model performance is observed using MLP rather than with IBk
algorithm. Higher errors are observed when Hagen and Costa equations are applied
to the whole dataset. However, better results can be expected if the distinction
between failure types is taken into consideration, which in turn implies the necessity
of more data.
Table 5.5: Prediction of peak outflow Qp = f (hd , S).

Method RMSE
IBk–2 neighbours 1200
MLP 1135
Hagen, Eq. 4.2 8400
Costa, Eq. 4.11 1300
10000
MLP
8000 IBk
Costa
Predicted
6000
4000
2000
0
0 2000 4000 6000 8000 10000
Observed
Figure 5.9: Peak outflow prediction based on dam height and reservoir storage.
5.4 Conclusions
Despite the shortage of documented data from historic dam failure events, in this
chapter we demonstrate that it is possible to improve currently available empirical
relations and prediction capabilities for breach characteristics by applying data min-
ing techniques. Encouraging results are obtained for predicting the average breach
width based on the water depth behind the dam at the initial failure time. The
applied data mining algorithms, LMS and MLP, perform better than the available
empirical equation of USBR (1988).
A 30% improved performance of the LMS and MLP models is observed when the
complete dataset is used. Since 84% of the dams in the dataset belong to low and
medium category, another experiment is carried out excluding the instances belong-
ing to high dams. An improvement of the LMS and MLP model performance (9.6
and 10.9 RMSE respectively) and USBR equation (12.0 RMSE) is observed, leading
to the conclusion that the data mining models seem to perform slightly better in
experiments for predicting average breach width for low and medium earthfill dams
74 Section 5.4. Conclusions
failing from overtopping.
Furthermore, the prediction of peak outflow based on dam height and reservoir stor-
age provides better results compared to the prediction of peak outflow based on
dam height only. The results obtained by using IBk and MLP algorithms are better
than the ones obtained by applying the empirical equations developed by Singh &
Snorrason, that predict peak outflow based on dam height only, and of Hagen and
Costa, whose equations use dam height, depth and volume of water in the reservoir
for peak outflow prediction. Clearly, the relatively small number of recorded dam
failure events contained in the database, limits the capabilities of data mining tech-
niques. However, with increasing amount of information about historic (or future)
dam failure events, better results might be expected.
Chapter 6
BREADA model: Breach

Model for Earthfill Dams
One certainly cannot predict future events exactly if one
cannot even measure the present state of the universe precisely.
Stephen Hawking
6.1 Development of a dam breach model

In this chapter we propose a method for modelling the overtopping failure of earthfill
dams. In dike failure literature a distinction is made between overtopping and
overflowing failure. The first is identified as the failure initialized from the waves
(e.g. wind waves) flowing over the structure’s crest for considerable amount of
time, though the water depth in the water body is (still) lower than or equal to
the structure’s height. The overflowing failure is caused due to continuously rising
water level (in the water body) above the structure’s crest in absence of waves. The
overflowing definition is almost never used in (embankment) dam failure literature.
Instead, the term overtopping is generally used to describe the flow over the dam’s
crest irrespective of the waves’ presence. Here, we also use the overtopping term
though we do not consider waves in our model, but only the free flow over the dam,
initiating a failure. The overtopping failure mode is known to be the most common
failure mode (Costa, 1985; Singh, 1996) for embankment dams caused as a result of
extreme inflows, malfunction of any hydraulic structure designated to release part
of the hydraulic load, etc.
Controversial conclusions are drawn by different researchers (Johnson and Illes, 1976;
MacDonald and Langridge-Monopolis, 1984; Morris, 2005) related to the initial,
progressive, and ultimate breach shape, since proper monitoring of a real dam failure
event is difficult to achieve so far. During laboratory and field experiments the
76 Section 6.1. Development of a dam breach model
initial breach shape is commonly predefined and thus influences in some way its
further development (see Section 2.3.1). In the model BReaching of the EArthfill
DAm (BREADA) it is possible to choose between two breaching developments. One
formulation assumes a trapezoidal breach shape to be initially formed at the crest
of the dam, which progresses to the dam bottom elevation (usually assumed at
streambed elevation at the dam toe). Once the dam bottom is reached, the breaching
develops in a lateral direction only. In the second formulation, a triangular breach
channel is initially formed and progresses to the dam bottom (see Figure 6.1). Once
the dam bottom is reached, the breaching develops in lateral direction having a
trapezoidal shape. The latter is considered by several authors (Johnson and Illes,
1976; MacDonald and Langridge-Monopolis, 1984; Macchione and Rino, 1989) to be
a common type of breach evolution during historical dam failure events.
The initial breach location, though an important component in breach modelling,

is usually difficult (if not impossible) to predict beforehand, since many factors can
affect any potential initial location, such as:
Bad compaction quality of the dam material during construction, or inappro-

priate material used in dam construction.
Presence of internal structures and weak points inside the dam’s body.
Damage to the dam structure that can occur during operation life, possibly at
the slopes or at the crest.
The breaching of a dam usually initiates at the weakest point on the dam crest or
downstream slope when overtopping occurs. In our model, the initial breaching is
assumed to occur at the dam crest in the middle of the dam length (unless specified
otherwise) and further breaching develops parallel to the dam crest (the implemented
breach development mode is described in Section 2.3.2).
For routing of the flow through the breach, the reservoir routing principle is applied:
dV
= Qin − Qb − Qc − Qoutlet , (6.1)
dt
where Qin , Qb , Qc and Qoutlet are the inflow into the reservoir, the flow through
the breach, the flow over the crest, and the flow through the outlet, respectively and
dV /dt represents the change of reservoir volume in time.
Significant uncertainty is associated with experimental results obtained from labo-

ratory and field experiments that determine the dominant sediment transport mech-
anisms during dam breaching (Section 2.3.3). For non–cohesive embankment dams,
the erosion in a breach channel is considered to be similar to a morphological bed
evolution (Broich, 2003). While in rare cases the erosion of dam material is assumed
to be similar to suspended load transport, most of the time it is simulated as a bed
Chapter 6. BREADA model: Breach Model for Earthfill Dams 77
δβ δβ hc hc
δ δβ β
hc hc
H hd
H hd H Hhd hd
y y y y
hc hc H H
hd hd
hc hc H Hhd hd
(a) (b)
Figure 6.1: Breach development in both formulations: (a) triangular shape to the
bottom of the dam and afterwards trapezoidal shape and (b) trapezoidal shape with
vertical and lateral erosion to the bottom of the dam and afterwards lateral erosion
only.
load transport (see Section 4.4 and Table 4.4). In the BREADA model, the erosion
of the embankment material is calculated using an empirical formula presented by
Meyer-Peter and Müller (1948).
Meyer-Peter and Müller (1948) performed the experiments on uniform bed mate-
rial as well as with mixtures of particles with diameter ranging from 0.4 to 29mm,
channel slope from 0.0004 to 0.02, water depth of 0.1 to 1.2m and specific weight
of particles in range from 0.25 to 3.2N/m3 . The bed load transport rate [m2 /s]
empirical equation obtained was written as
1/2
qbl = φb [(s − 1) g] d3/2
m , (6.2)
where s is the relative density ρs /ρ, ρs is soil density, dm is the mean particle
diameter [m] considered 1.1 to 1.3 times the d50 for nearly uniform material and
equal to the d50 for mixture of materials (van Rijn, 1993), and φb is the bed load
transport rate expressed in dimensionless form as
3/2
φb = 8 (µθ − θcr ) , (6.3)
where θcr is generally interpreted as the critical mobility parameter that for mixtures
of sediments on stream beds it is usually taken equal to 0.047. During a dam breach
erosion, θcr might be considered negligible compared to θ that is the dimensionless
particle mobility parameter estimated as
RI
θ= . (6.4)
(s − 1) dm
A bedform factor or efficiency factor µ is calculated with

3/2
k
µ= , (6.5)
k0
where k 0 is the grain–related Strickler coefficient and k is the overall Strickler coeffi-
cient. Yalin (1972) emphasized that if the plane is flat and the flow is two dimensional
then the ratio k/k 0 becomes 1.
Substituting Eq. 6.3, 6.4, 6.5, into Eq. 6.2 we obtain

3/2
qbl = ker (γRS) , (6.6)
where ker is an erodibility coefficient equal to

8
ker = ,
(γs − γ) g 1/2
R is the hydraulic radius, γ is the water specific weight, γs is the soil specific weight,
and S is the energy slope that can be calculated with Strickler’s equation
u2
S= ,
k 2 R4/3
where u is the mean flow velocity. Assuming that the solid discharge entering our
domain from the reservoir is null and considering a space step ∆x = 1m, we can
write the following mass conservation:
dAb dqbl c
= (6.7)
dt dx
as
dAb
= qbl c, (6.8)
dt
where c is the wetted perimeter.
We assume that the initial conditions - water depth in the reservoir, initial breach
depth and the inflow hydrograph - are known. For each time step, the flow over the
crest, through the spillway and the outlet is calculated based on reservoir elevation.
The difference between inflow and outflow is then used to calculate the change
in reservoir storage and the new reservoir elevation. The reservoir stage–volume
relationship is provided to the model through a mathematical expression. Based on
this relationship the reservoir water level is extracted for every estimated reservoir
storage at each time step. Other flow characteristics specific for different breach
shapes are presented in the following.
6.1.1 First formulation of dam breaching development

The breach shape is assumed to be trapezoidal. At first the breaching of the dam
progresses in vertical and lateral direction until the bottom of the dam is reached.
The breach opening area is calculated as
2
(Hd − y)
Ab = B (Hd − y) + (tan β + tan δ) , (6.9)
2
where B is the bottom breach width. The development of breach cross section in
time can be written:
dAb dAb dy dy
= = − [B + (Hd − y) (tanβ + tanδ)] . (6.10)
dt dy dt dt
Flow in the breach channel is assumed a critical flow and the outflow discharge is
calculated as
1/2 3/2
g (2B + hc (tan β + tan δ)) hc
Qb = , (6.11)
B + hc (tan β + tan δ) 2
where hc is expressed as:
−3B + 2 (H − y) (tan β + tan δ) + R1/2

hc = , (6.12)
5 (tan β + tan δ)
and
2
R = 9B 2 + 8B (H − y) (tan β + tan δ) + 4 (H − y) (tan β + tan δ).
Combining the Eqs. 6.6, 6.8 and 6.10, the breach bottom elevation can be estimated
every time step with following equation
 3/2
1 1
dy ker (γg)
3/2
(2B + hc (tan β + tan δ))hc  B + hc cos β + cos δ
=− 3  .
dt k 2 B + (Hd − y) (tan β + tan δ) B + hc (tan β + tan δ)
(6.13)
For the development of the breach only in lateral direction (once the dam bottom
is reached) the model equations for the breach characteristics are given as follows.
The breach area is estimated as

Hd
Ab = B + − y (tan β + tan δ) Hd , (6.14)
2
and the development of the breach only in lateral direction is expressed as

2
dy ker 3/2 hc [2B − (2y − hc ) (tan β + tan δ)] cos β + cos δ
= − 3 (γg)
dt k 2Hd sin (β + δ)
h i1/2 (6.15)
B − y (tan β + tan δ) + hc cos1 β + cos1 δ
× 3/2
.
[B + (hc − y) (tan β + tan δ)]
The discharge through the breach opening is estimated to be
r 3/2
g [hc (2B − (2y − hc ) (tan β + tan δ))]
Q= . (6.16)
2 2 [B + (hc − y) (tan β + tan δ)]1/2
Assuming β = δ we write a dependency between the lateral and vertical erosion as

∆B
= Const
∆y
where ∆B and ∆y represent the change in width (bottom or top) and in depth of
the breach opening.
6.1.2 Second formulation of dam breaching development

In this formulation, a triangular breach channel is initially formed and progresses
to the dam bottom (see Figure 6.1). The area of the triangular cross section of the
breach channel (Figure 6.1), for β 6= δ is:
2
(Hd − y) (tanβ + tanδ)
Ab = . (6.17)
2
We differentiate Ab with respect to time t as follows
dAb dAb dy dy
= = − (tanβ + tanδ) (Hd − y) . (6.18)
dt dy dt dt
The discharge through the breach channel is calculated as
g 1/2 tan β + tan δ
Qb = h5/2
c , (6.19)
2 2
where, the critical depth is
4
hc = (H − y) . (6.20)
5
Combining the Eqs. 6.6, 6.8 and 6.18, the breach bottom elevation can be estimated
every time step with
3/2
h2c

dy 1 ker 3/2 cos β + cos δ
=− 3
(γg) . (6.21)
dt 2 k Hd − y sin (β + δ)
Once the dam bottom is reached, the breaching develops in lateral direction into a
trapezoidal shape. The breach area cross section for trapezoidal breach shape with
y < 0 is equal to
Hd
Ab = (Hd − 2y) (tan β + tan δ) . (6.22)
2
The wetted area is
hc
A = (hc − 2y) (tan β + tan δ) , (6.23)
2
where hc is the critical depth derived as
p
3y + 2H + 9y 2 − 8Hy + 4H 2
hc = . (6.24)
5
Using the above expressions the discharge through the breach channel may be written
as r
g tan β + tan δ 3/2 −1/2
Qb = [hc (hc − 2y)] (hc − y) . (6.25)
2 2
The breach bottom elevation can be estimated at every time step with
2
1/2
dy ker 3/2 hc (hc − 2y) cos β + cos δ cos β + cos δ
= − 3 (γg) h c − y . (6.26)
dt k 2Hd (hc − y)3/2 sin (β + δ) sin (β + δ)
Based on the formulations described above, a software package (see Figure 6.2) is
developed using Borland Delphi. The program has user–friendly graphical interface,
which allows the user to load the necessary information and to run the model.
6.2 Validation of the BREADA model

We validate the BREADA model against a historical dam failure event: the breach-
ing of the Schaeffer Dam in the USA that failed due to overtopping in 1921. To
extend even further the comparison with the real measurements (or known informa-
tion), we also compare our results to the output of the commercially available model
BREACH (Fread, 1988). We analyze the breach outflow time series generated by
both models and compare the peak outflow values with the results obtained from
empirical equations (introduced in Section 4.3) and data mining techniques (intro-
duced in Section 5.3).
The Schaeffer Dam was a 30.5m high non–cohesive earthfill dam constructed across
Beaver Creek in USA. The dam failed due to heavy rains described in Follansbee and
Jones (1922) as ‘cloud-bursts’, which led to water levels in the reservoir higher than
the dams top elevation. Some of the dam and reservoir characteristics are presented
in Table 6.1.
According to Follansbee and Jones (1922), the dam was washed away in 30 minutes.
82 Section 6.2. Validation of the BREADA model
Figure 6.2: Main window of the BREADA program.
Table 6.1: Schaeffer Dam characteristics.

Dam Characteristics
Dam height [m] 30.5
Embankment width at crest [m] 4.6
Embankment length at crest [m] 335
Upstream slope [H:V] 1:3
Downstream slope [H:V] 1:2
Grain size D50 [mm] 2.5
Uniform factor d90/d30 12.4
Porosity 0.5
Unit weight [KN/m3 ] 18.5
Friction angle 25 - 40
Hydraulic Characteristics
Reservoir [m3 ] 3.92 million
Initial water depth in the reservoir [m] 30.5
Inflow [m3 /s] 270
Initial breach depth [m] 0.3
5000
4000
Discharge [m /s]
3
3000
2000
1000
0
00:00 00:15 00:30 00:45 01:00 01:15 01:30
Time [hr:mm]
Figure 6.3: Sensitivity analysis for breach side angle with the BREADA model.
At the end of failure the breach opening was observed to be of a trapezoidal shape
with a top width of about 210m (some 70% of dam crest length) and side slopes of
approximately 1:2.25 [H:V]. A peak outflow of 4,500m3 /s was recorded.
The available record of parameters for the Schaeffer Dam is incomplete and, there-
fore, a sensitivity analysis is undertaken to identify which of the (unknown) input
data influences the outflow discharge. The inflow into the reservoir was estimated by
Follansbee and Jones (1922) to be 270m3 /s and belongs to the maximum discharge
estimated for Beaver Creek, due to the storm that occurred more than 24 hours
before the failure of the dam. It includes the rainfall and tributary inflow. In our
experiment a reservoir stage–area curve is created based on the available information
for some representative reservoir levels. Model parameters, namely the initial breach
depth, breach side slope and erodibility coefficient are part of the sensitivity analy-
sis for the BREADA model and the internal friction angle is part of the sensitivity
analysis for the BREACH model. In this section, we briefly discuss the influence of
side slope and internal friction angle on breach outflow hydrograph.
Figure 6.3 presents the resulting outflow hydrograph simulated for sensitivity anal-
ysis of the breach channel side slope (β = δ) by the BREADA model. The angle
that the breach sides create with the vertical is varied from 25◦ to 40◦ . It can be
observed that the variation in the side slope influences the peak outflow magnitude
and its time of occurrence. This is expected since the breach channel area depends
on the breach channel side slope. Smaller is the angle that the breach side creates
with the vertical, smaller is the initial breach channel cross section area. This leads
to smaller initial breach outflow meaning an increase of water depth in the reservoir.
84 Section 6.2. Validation of the BREADA model
12000
10000
Discharge [m /s]
8000
3
6000
4000
2000
0
00:08 00:20 00:32 00:44
Time [hr:mm]
Figure 6.4: Sensitivity analysis for internal friction angle with the BREACH model.
This process continues for few minutes. Then the water depth in the reservoir that
is increased as a result of lower outflow values, initiates higher erosion rates, thus
higher discharge rates flowing through the breach channel. In Figure 6.3 we can
observe that the angle of 25◦ is producing a more realistic breach outflow hydro-
graph, with peak values similar to the observed one and a flood duration of about
30 minutes. This result is consistent with the observed final side slope of the breach
channel of 1:2.25.
Figure 6.4 presents the sensitivity analysis of the BREACH model for variable in-
ternal friction angle. The angle of 40◦ generates a peak outflow about 40% higher
than the recorded value, but lower than the peak outflow obtained for angles of
25◦ and 30◦ . For the internal friction angle equal to 60◦ , the obtained side slope
angle is 26◦ coinciding with the observed one and the corresponding peak outflow is
4,000m3 /s. This happens due to the same reason as for the BREADA model. The
internal friction angle value affects the enlargement of the breach width. During all
the simulations, the duration of the hydrograph rising limb in the BREACH model
is very short, fast reaching the maximum peak outflow.
In order to evaluate the applicability of data mining techniques to dam breach mod-
elling, the models created with IBk and MLP algorithms in Section 5.3.3 are used to
predict the peak outflow for the Schaeffer Dam and the resulting peak outflow val-
ues are analyzed. The results of peak outflow estimation for the Schaeffer Dam are
presented in Figure 6.5. The peak outflow estimated by the IBk model is 4,000m3 /s
viz. about 10% smaller than the observed value, while a 34% lower peak outflow is
predicted by the MLP model.
12000
Measured
10000 MLP
IBk
Peak flow [m /s]
8000
3
6000
4000
2000
0
0 5 10 15 20 25 30 35 40
Dam height [m]
Figure 6.5: Peak outflow estimation based on dam height.
Table 6.2: Peak outflow estimation using empirical equations for Schaeffer Dam.
Reference Peak outflow
(m3 /s)
Regression Equations
Singh and Snorrason (1982) - Eq. 4.5 2230
MacDonald and Langridge-Monopolis (1984) - Eq. 4.4 2584
Costa (1985) - Eq. 4.11 2422
Froehlich (1987) - Eq. 4.12 3843
Envelope Equations
Hagen (1982) - Eq. 4.2 9700
MacDonald and Langridge-Monopolis (1984) - Eq. 4.3 8460
Costa (1985) - Eq. 4.7 6991
- Eq. 4.8 9435
Using the empirical relations (see Table 6.2), the peak outflow varies from 2,230m3 /s
to 3,843m3 /s excluding the upper limits from the envelope equations. Froehlich’s
equation estimating peak outflow based on depth and volume of water in the reser-
voir at failure time, performs better than the other equations, where the estimated
peak outflow is 15% lower than the recorded peak outflow.
6.3 Failure analysis of an earthfill dam

Failure of large dams is most likely to result in the release of enormous quantities
of water, endangering people’s lives and property in downstream areas. The conse-
86 Section 6.3. Failure analysis of an earthfill dam
Figure 6.6: View of the Bovilla Dam facing upstream.
quences are particularly high in terms of human casualties when no warning system
is available. A dam break analysis is required for every constructed large dam for
the purpose of identifying the impact of failure in the downstream areas and the
necessary measures for mitigating the consequences.
In this section, we undertake a dam break analysis for the hypothetical failure of
Bovilla Dam (Figure 6.6) that was constructed on the Terkuze river, 15km north-
east of Tirana City, Albania. It is an earthfill homogeneous dam, 81m high, and
135m long at the crest. The dam’s filling material is composed of gravel and sand,
taken from alluvial deposits of the Terkuze River upstream of the dam. The nor-
mal capacity of the reservoir (Figure 6.6) is 80 million m3 at the pool elevation of
318m from which 50 million m3 are used for water supply and the rest for irrigation.
Bovilla Reservoir can accommodate the average annual volume that the Terkuze
River brings, that is evaluated to be nearly 105 million m3 water. The purpose
for constructing the Bovilla Dam is to provide enough water to the continuously
expanding population of Tirana city and to satisfy the demands for irrigation pur-
poses.
Though dams are meant to function for a long period of time, the probability of their
failure is never zero. The Bovilla Dam is classified into the high hazard category
due to populated areas located just 10 km from the dam. The dam construction
lead to permanent displacement of more than 400 families from their homes and
traditional livelihood in 1993. However, a year later, people started to settle near
Bovilla Reservoir
Bovilla Reservoir Disolv1.shp
Disolv1.shp Populated areas 2001.shp
Populated Areas 1994.shp Schools, hospitals, government offices
Commercial Low density informal housing area
Construction area Mix of apartments and single-family housing
Existing rural construction Residential multi-story construction
Hospitals Singe-family detached housing
Industrial area Residential and commercial
New construction Very low density rural settlement
Terkuza River Medium density informal housing
Retail, wholesale,office or hotel
Airport
Terkuza River
(a) (b)
Figure 6.7: Land use data for the area downstream the dam showing the populated
areas for the years (a) 1994 and (b) 2001.
the river banks downstream the dam. Figure 6.7 shows the land use data for two
different years: 1994, a year after the work for construction of the dam had been
started and 2001, five years after the work was completed. A potential failure of the
Bovilla Dam may lead to highly devastating consequences in the areas downstream.
Therefore, a dam break analysis needs to be conducted to identify the potential
flood propagation scenarios and consequences in case of a failure event. Modelling
of the flood wave propagation induced by a potential dam break event can help the
implementation of emergency plans, risk assessment, and future development of the
areas.
The pertinent properties of the dam and the reservoir are presented in Table 6.3.
The impenetrability of the upstream slope of the dam is achieved by mean of a
geo-textile material. There is a bottom outlet, used for water release for flood con-
trol and irrigation purposes. The total discharge capacity of the bottom outlet is
150m3 /s and is designed to handle 1,000 year return period flood event with peak
discharge in the river of 700m3 /s. This flood brings an increase of the water level
in the reservoir to maximum of 320m. Two gates are installed at the outlet tower:
the principal gate and the emergency gate. Next to the outlet tower an intake tower
releases water on a regular basis for water supply.
The hydrologic data consist of the flood discharge time series for return period events
of 10, 100 and 1,000 years (Albinfrastrukture, 1996). The relationship between reser-
voir surface area and water elevation (Figure 6.9) is obtained from digital elevation
data (Figure 6.8) using Arc–GIS. An area 32km downstream of the dam is consid-
ered in this analysis. The terrain elevation varies from 240m above sea level at the
Table 6.3: Bovilla Dam Pertinent Data.

Dam Characteristics
Dam top elevation, [m a.s.l] 321
Height above streambed [m] 81
Length at the crest [m] 135
Width of top crest [m] 8
Upstream slope 1.6
Downstream slope with 4 berms 1.6
Reservoir Characteristics
Normal pool elevation [m a.s.l] 318
Maximal pool elevation [m a.s.l] 319
Normal pool area [km2 ] 4.1
Normal pool capacity [m3 ] 75 million
Maximum pool capacity [m3 ] 84.3 million
Catchment area [km2 ] 98
Bottom Outlet
Type armoured concrete surface
Bottom elevation at the entrance (m a.s.l) 275
Principal gate dimension [m] 3.5×3.7
Emergency gate dimension [m] 3.5×4.0
Length of outlet gallery [m] 156
Maximum capacity [m3 /s] 150
Freeboard
Normal pool [m] 3
Bovilla Reservoir
Terkuza River
Grid
300 - 1050
250 - 300
180 - 250
140 - 180
120 - 140
80 - 120
70 - 80
60 - 70
50 - 60
40 - 50
30 - 40
20 - 30
10 - 20
No Data
Figure 6.8: Digital elevation grid for the area downstream the dam.
330
320
310
Elevation (m)
300
290
280
270
260
250
240
0 1 2 3 4 5
Area (km2)
(a) (b)
Figure 6.9: (a) Surface area-water elevation relationship for Bovilla reservoir and
(b) Surface boundary for different elevations starting from inner boundary at 240m,
250m, 270m, 280m, 300m, 321m elevation, respectively
dam site to 15m above sea level at the downstream boundary. The floodplain at
the first kilometers from the dam is quite narrow and steep, but expands rapidly
afterwards.
6.3.1 Modelling of breach development

The first step in this dam break analysis is the prediction of the breach outflow hy-
drograph that will constitute the upstream boundary condition of the flood routing
model. The main goal is the modelling of the flood wave propagation in case of
the dam’s failure, regardless of the expected probability of occurrence of such event.
Here, we describe in detail the overtopping failure of the dam that might represent
the worst flooding scenario (high magnitude of flood conditions) and discuss the
piping failure.
Different scenarios can lead to dam overtopping, namely the inflow into the reservoir
is higher than the spillway’s capacity, malfunctioning of the spillway, etc. Currently,
there is no methodology that is able to determine the extent of overtopping (flow
depth and duration of flow over the crest) that the earth dam can withstand. Dam
failure depends on the resistance of the embankment soil to water flow, type of dam
covering material, slope of the embankment, which in turn influences the flow ve-
locities, etc. In this study, we conduct a sensitivity analysis for the inflows into the
reservoir assuming the 1,000 year and the 100 year return period flood events to
inflow into the reservoir when the dam erosion is initiated. We ignore the effect of
other small stream and rainfall inflows into the reservoir and the hydrologic charac-
teristics of the basin.
Other assumptions used in modelling are the following:
Final breach bottom level corresponds to the stream bed elevation. Failure
of a high dam, might not always develop down to the bottom as the width
of the dam is usually much wider at that location. However, we make this
assumption for the sake of simplicity.
Water depth in the reservoir at the initial time is assumed at the dam top
elevation or 0.2m above.
We start with a sensitivity analysis of the two different breach development formu-
lations: (a) trapezoidal breach shape throughout simulation (see Section 6.1.2), and
(b) triangular breach shape till dam foundation is reached and trapezoidal breach
shape afterwards (see Section 6.1.1). In Figure 6.10 we can observe that the hydro-
graphs do not differ significantly from each other in terms of peak outflow magni-
tude, but differ in timing. The trapezoidal breach development formulation exhibits
slower breaching of the dam. While this might seem counterintuitive, the result can
be expected due to the fact that the ratio between the vertical and lateral erosion is
limited to a constant factor (in this case ∆B/∆y = 2), signifying slow enlargement
of the breach width. For the following simulations we use the second formulation of
breach development unless specified otherwise.
Figure 6.11 presents the sensitivity analysis for two different inflows into the reser-
voir. It can be observed that the magnitude of inflow into the reservoir does not
influence notably the flood hydrograph. The peak of the flood hydrograph is reached
about 6 minutes faster and is only 3% higher for the 1,000 year return period flood
event compared to the 100 year return period event. That is expected as the inflow
peak discharge for the 1,000 year return period flood is just 700m3 /s compared to
430m3 /s for the flood with return period of 100 years, with both values being sig-
nificantly lower than the breach outflow. For further analysis we assume the inflow
in the reservoir at the failure time to be 1.05 times higher than the peak discharge
of 1,000 year return period flood used for the design of the spillway, thus consider-
ing dam (and the spillway) to be able to accommodate lower inflow floods into the
reservoir.
Two different stage–volume relationships are created for the reservoir: one based on
the GIS data and the other based on the data provided by the designers of the dam
(Albinfrastrukture, 1996). The latter does not include the reservoir dead storage
information, but only the active storage capacity. The peak outflow obtained by the
model for different reservoir stage–volume relationships differs as much as 15% (see
Figure 6.12). This emphasizes the uncertainty related with the actual volume of
70000
1st Formulation
60000
2nd Formulation
50000
Discharge [m /s]
3
40000
30000
20000
10000
0
00:00 00:30 01:00 01:30 02:00 02:30 03:00 03:30
Time [hr:min]
Figure 6.10: Breach outflow hydrograph for first and second breach development
formulations.
60000
T=100
50000 T=1000
Discharge [m /s]
40000
3
30000
20000
10000
0
0:00 0:30 1:00 1:30 2:00 2:30 3:00
Time [hr:min]
Figure 6.11: Breach outflow hydrograph for 1,000 and 100 year return period flood
event.
70000
GIS Data
60000
Active Storage Capacity
Discharge [m /s]
50000
3
40000
30000
20000
10000
0
0:00 0:30 1:00 1:30 2:00 2:30 3:00
Time [hr:min]
Figure 6.12: Breach outflow hydrograph for two reservoir stage–volume relationship.
water in the reservoir at the failure time, which is affected as well from the sediment
mass accumulated in the basin throughout the years, though not considered in this
analysis.
Two user–defined parameters of the BREADA model, initial breach channel depth
D and the angle that the breach channel sides create with the vertical β = δ, are
subject to a sensitivity analysis. Different values for initial breach channel depth
(D=1m, 0.5m, and 0.3m) are considered in order to understand the influence of
the initial conditions on the breach development. As shown in Figure 6.13, D does
not affect the breach outflow magnitude, but has an impact on the peak timing.
Higher initial depth in the breach opening results in faster breach development. In
contrast, β has an influence on the breach outflow hydrograph (Figure 6.14). For
different values of β the magnitude of the peak discharge changes. An increase of
β leads to a decrease of the peak discharge and vice versa similarly as observed in
Section 6.2. Compared to the angle of 35◦ , the magnitude of the breach outflow for
angles of 40◦ and 45◦ , decreases 11.8% and 15.8% respectively.
Furthermore, we compare the dam breach modelling results obtained from the
BREADA and BREACH model. In Figure 6.15 it can be observed that though the
peak outflows differ in range of only 10%, the timing, which is a very important fac-
tor for proper warning of the population downstream from the dam, is significantly
different. In the BREACH model the peak outflow occurs after about 39 minutes
from the beginning of initial breach, while the peak outflow in the BREADA model
60000
D = 1.0m
50000
D = 0.5m
Discharge [m /s]
40000 D = 0.3m
3
30000
20000
10000
0
0:00 0:30 1:00 1:30 2:00 2:30 3:00 3:30
Time [hr:min]
Figure 6.13: Sensitivity analysis for different initial breach depth D (β=40◦ ; Q1% ).
60000
β = 35°
50000 β = 40°
β = 45°
Discharge [m /s]
40000
3
30000
20000
10000
0
0:00 0:30 1:00 1:30 2:00 2:30 3:00 3:30
Time [hr:min]
Figure 6.14: Sensitivity analysis for different breach angle β (D=0.5m; Q1% ).
80000
BREADA
70000 BREACH
60000
Discharge [m /s]
3
50000
40000
30000
20000
10000
0
0:00 0:30 1:00 1:30 2:00
Time [hr:mm]
Figure 6.15: Breach outflow hydrograph for the BREACH and BREADA model.
occurs about 70 minutes later. A sharper rising limb is produced by the BREACH
model compared to a more gradual development of the breach in the BREADA
model.
The breach cross section could be a possible reason for this discrepancy - the
BREACH model develops a rectangular breach shape from the beginning till the
moment when the sides of the breach channel collapse, creating a trapezoidal shape.
This moment in time depends on the dam material internal friction angle that in
turn controls the side slope. Instead, the BREADA model develops a triangular
shape till the bottom of the dam is reached and then develops into a trapezoidal
breach shape. This leads to a slower development of the breach in the latter model.
However, the duration of the rising limb produced by the BREACH model is only
6 minutes, and that might not be realistic. The results obtained from both breach
models represent two different scenarios, one being less catastrophic to the down-
stream area due to slower development despite 10% higher peak outflow, and the
other predicting a disastrous situation with faster development of breaching. In
both models we do not take into account the cover layer at the upstream slope of
the dam, which might influence the results. Peak outflows obtained by the BREACH
and BREADA models are 63,570m3 /s and 68,740m3 /s respectively.
To conclude, we predict the peak outflow using the empirical equations. From the
results in Table 6.4, it can clearly be seen that only Hagen’s formula produces almost
the same peak outflow as predicted using the dam breach models. The upper curve
Table 6.4: Peak outflow estimation based on dam height and reservoir storage.
Reference Peak outflow
(m3 /s)
Hagen (1982) - Eq. 4.2 67,230
MacDonald and Langridge-Monopolis (1984) - Eq. 4.3 43,630
- Eq. 4.4 13,030
Costa (1985) - Eq. 4.9 38,910
- Eq. 4.11 13,950
- Eq. 4.7 39,560
- Eq. 4.8 58,950
Froehlich (1987) - Eq. 4.12 31,980
Broich (1998) - Eq. 4.13 14,260
- Eq. 4.14 21,040
of the envelope equation proposed by Costa is 17% and 9% lower than the peak
outflow produced by the BREADA and BREACH models respectively. These re-
sults were expected since Hagen’s formula is the only formula derived from historical
data of earthfill dams failing due to overtopping, with a range of characteristics that
covers the characteristics of the Bovilla Dam. The dam considered in this study, is
higher than all of the 22 dams that Froehlich used for deducing his equation. More-
over, other empirical formulas were extracted from data sets where no distinction
was made between dam material or failure type.
6.3.2 Modelling of flood propagation

The breach outflow hydrograph generated from the BREADA and BREACH models
is routed in the areas downstream the dam using Sobek 1D2D, an integrated one–
and two–dimensional numerical simulation package developed by WL | Delft Hy-
draulics (see Section 3.4). A 1D model is created for the 32km reach downstream of
the dam with an average bed slope of 0.01. The 2D modelling area is approximately
19km×17km with grid cell size of 30m×30m. Since the purpose of this analysis is
the identification of flooded areas in case of a hypothetical failure of the Bovilla
Dam, rather than the identification of the flood duration period, the simulation is
finalized when the flood wave reaches the downstream boundary.
We ignore the reservoir sediment or the dam material, which are flushed away to-
gether with the water stored in the reservoir, but instead consider propagation of
clear water in the valley. The flooded areas are treated as rigid, meaning neither
erosion nor deposition processes are modelled despite the awareness that these pro-
cesses will occur during a real dam failure flood wave propagation. The presence of
the debris material in the flood water, which will affect the flow pattern significantly,
Figure 6.16: Roughness grid.
is taken into account through higher values of roughness coefficients.
The roughness of the floodplain area is subject to sensitivity analysis. The extent
of the dam break flood event is unprecedented in the natural history of the valley,
thus the calibration of the roughness coefficients is not possible. Sobek 1D2D has
an option to apply constant or variable roughness coefficients for the channel reach
and the overland area. A constant Manning’s roughness coefficient n equal to 0.07
is assumed for the river channel. Constant and variable roughness coefficients are
used for the overland area. The roughness coefficients for different land use cate-
gories (Figure 6.16) are deduced from the Tables of Manning’s n presented in Chow
(1959); van der Sande et al. (2003); Huang (2005).
The flood wave propagation scenario for the breach outflow hydrograph generated
by the BREADA model is presented in Figure 6.17. The extent of flooding for two
time intervals is given for the purpose of comparing with the flood patterns when the
breach outflow hydrograph of the BREACH model is used as upstream boundary
condition in the flood modelling analysis. Simulation results show that the flood
wave reaches the location 26 km downstream of the dam (downstream of the river
reach presented in Figures 6.17 and Figure 6.18) in approximately 5, 4, and 3 hours
time span for constant Manning n = 0.15, n = 0.1, and for roughness based on land
use data. It reaches the populated areas located 9 km downstream the dam in 1hr
and 40 minutes, enough time to warn population for immediate evacuation. At 13
km away from the dam, flood peaks are dampened out as the wave reaches the wide
flat areas; still, the water depth at the villages, including the international airport
(a) (b)
Figure 6.17: Routing of the BREADA model outflow hydrograph using Sobek 1D2D
with Manning coefficient in the floodplain area equal to 0.15 (a) 2hr 38min after the
breach has been initiated and (b) 4hr 38min after the breach has been initiated.
exceeds 2m. The flood speed reaches 7m/s at the populated areas when taking into
account the land use based roughness.
The flood wave propagation scenario for the breach outflow hydrograph generated
by the BREACH model is presented in Figure 6.18. Owing to the fast breach de-
velopment in the BREACH model, the flood wave reaches 26 km downstream of
the dam in 4, 3, and 2 hours time span for constant Manning n = 0.15, n = 0.1,
and for land use based roughness. In this scenario the time interval from the initial
breaching to the time the flood wave approaches the first residential areas is only 40
minutes. Higher water depths are observed in some locations in comparison to the
BREDA model scenario (see Figure 6.19).
The different roughness maps provide different flood hazard scenarios. The results of
a sensitivity analysis for constant and land use based Manning roughness coefficients
on the floodplain are presented in Figure 6.19. The water depths are shown for var-
ious locations in the valley downstream the dam (see Figure 6.17). As expected, the
increase in floodplain roughness coefficient decreases the flood propagation speed,
thereby delaying the flooding of a particular area. An increase in water depth is
observed at the high elevation locations (e.g. location A which correspond to a res-
idential area) while there is no notable change in water depth at the low elevation
areas (location C that corresponds to the airport area). The water depth in A ob-
tained using the BREACH model hydrograph as upstream boundary condition is
lower than the one obtained by using the BREADA model hydrograph, associated
98 Section 6.4. Uncertainty in dam break analysis
(a) (b)
Figure 6.18: Routing of the BREACH model outflow hydrograph using Sobek 1D2D
when Manning coefficient in the floodplain area equals to 0.15: (a) 2hr 18min after
the breach has been initiated and (b) 4hr 18min after the breach has been initiated.
this with the higher peak outflow values for the latter. These differences are damp-
ened at the location C further away from the dam where the floodplain area is wide.
According to the flood maps the water depths in residential areas are in the order
of 5m. The failure of Bovilla Dam would jeopardize the lives of thousands of peo-
ple living downstream and extensively affect the environment and the economy of
the area. Based on preliminary population data for the regions inundated by flood
water, the population at risk consists of at least 10,000 people. In case of a dam
failure event, the population faces significant risk. Immediate construction of a flood
warning system is necessary to avoid human casualties, in particular for residents
living as close as 9km away from the dam.
The inundation maps provide an estimate regarding the direct consequences, but
the indirect (long term) consequences as a result of dam failure will be enormous as
well. The development of the areas should be planned by taking into consideration
the above mentioned flood hazard maps ensuring better protection of the people
facing the risk in a potential dam failure event, as well as decreasing economic and
environmental consequences as a result of the failure.
6.4 Uncertainty in dam break analysis

When physically based methods are used for forecasting the flood wave propaga-
tion of a dam failure event, no data are usually available for its validation. Instead
the focus is on the identification of sources of uncertainty and the range to which
9
A: n = L A: n = 0.10 A: n = 0.15
8 C: n = L C: n = 0.10 C: n = 0.15
7
Water Depth [m]
6
5
4
3
2
1
0
1:00 2:00 3:00 4:00 5:00
Time [hr:mm]
(a)
9
A: n = 0.10 B: n = 0.10 C: n = 0.10
8 A: n = 0.15 B: n = 0.15 C: n = 0.15
7
Water Depth [m]
6
5
4
3
2
1
0
01:00 02:00 03:00 04:00
Time [hr:mm]
(b)
Figure 6.19: Sensitivity analysis of the roughness coefficient at locations 11km, 15km,
and 19km downstream the dam (namely A, B, and C respectively) for (a) BREADA
outflow hydrograph (L stands for land use data) and (b) BREACH outflow hydro-
graph.
100 Section 6.4. Uncertainty in dam break analysis
they affect the model results. Identifying the uncertainty associated with the pre-
diction of floods imposed by dam failure is necessary for effective flood management
and emergency planning. Three sources of uncertainty are identified in Section 3.4
namely, input uncertainty, model uncertainty, and completeness uncertainty. The
latter, often referred as ignorance uncertainty, is associated with all omissions that
occur due to lack of knowledge.
Input uncertainty
Neither the inflow into the reservoir nor the water depth in the reservoir at the fail-
ure time are known for the analysis of a hypothetical dam failure. Both values are
assumed depending on the event that will be simulated e.g. inflow is higher than
the spillway design capacity or the dam design flood for the overtopping failure, and
the water depth in reservoir corresponds to the maximum depth. These characteris-
tics are a source of uncertainty. However, the sensitivity analysis shows that inflow
hydrograph influence is not significant for the large Bovilla Dam. An increase in the
inflow with 60% (from Q1% to Q0.1% ) for Bovilla reservoir resulted in an increase of
the peak outflow with only 1%.
Although the reservoir area and storage volume are considered to be deterministic
quantities, the calculated reservoir area and storage volume at a given elevation may
vary due to different measurement and computational techniques used for their es-
timation. Moreover the reservoir area and storage volume might be different from
what is assumed, depending on the sedimentation in the reservoir during the years
before the failure occurs. The reservoir characteristics, namely the approximated
shape, volume, and area, all affect the final results.
Dam material characteristics at the failure time are part of the sensitivity analysis
and are very important for the correct estimation of the breach development and
characteristics. Roughness coefficient and Digital Elevation Model (DEM) data for
the floodplain area influence the scale of flooding in terms of magnitude and time
and are part of sensitivity analysis.
Model uncertainty
Model uncertainty is associated with different formulations implemented for mod-
elling the processes involved during the breaching of the structures, mathematical
and numerical description of the processes, etc. Different models result in differ-
ent output for the common problem due to various approximations made in the
description of the processes. The model related parameters influence significantly
the scale of flooding. We observe that two formulations of the BREADA model, us-
ing different breach shapes for breach development, lead to different breach outflow
hydrographs with differences in timing, which is a very important factor in flood
analysis. The same is observed when comparing the BREACH and the BREADA
models.
Completeness uncertainty
The completeness uncertainty represents unknown contributions and hence is a qual-
itative analysis rather than quantitative. Processes that are currently considered
irrelevant for the dam failure flood analysis, might prove to be important if inves-
tigated in greater detail. There are many processes that we omit in modelling due
to current lack of knowledge or awareness (e.g. wind can have impact on the failure
of the dam but its influence is usually not considered in any analysis for high dams
except for dikes). For the overtopping of an earthfill dam, the wind direction might
influence the failure process. The wind blowing toward the dam during high flood
might contribute to higher depth of water overtopping the dam.
6.5 Discussion and conclusions

This chapter starts by describing the model developed for predicting the breach
characteristics of an earthfill dam failing due to overtopping. A sensitivity analysis
is undertaken for assessing the influence of the model parameters on the computed
breach outflow hydrograph. It is observed that the initial breach channel depth,
that is a user defined parameter, has an influence on peak timing. The breach side
slope affects both the peak timing and its magnitude.
Two different methods are used for predicting breach outflow hydrograph in case of
hypothetical failure of an earthfill dam. We compare the resulting breach outflow
hydrographs obtained from the BREADA (Zagonjolli et al., 2005) and the BREACH
model (Fread, 1988). Despite only 10% difference in the peak outflow values obtained
from the two breach models, significant differences are observed in timing and shape
of the hydrograph.
Furthermore, we compare the resulting peak outflows of the BREACH and BREADA
models with the range of peak outflows obtained using the empirical formulations de-
veloped during the last decades and data mining techniques (see e.g. Zagonjolli and
Mynett (2005b,a)). Though empirical formulations can be used for a ‘rough/fast’
prediction of the peak outflow values, they are not applicable for dam break flood
forecasting where knowledge about breach development in time is important. Ha-
gen’s empirical formula produces almost the same peak outflow as predicted using
physically based numerical models. Other empirical formulas do not show good
agreement, that can be explained from the fact that the Bovilla Dam characteristics
are not within the range of dam characteristics in the datasets used for developing
those formula.
The resulting breach outflow hydrographs, constituting the upstream boundary con-
dition for the 1D2D hydrodynamic model, led to two different flood progression sce-
narios in the areas downstream from the dam. For numerical simulation of flood
102 Section 6.5. Discussion and conclusions
propagation downstream of the dam, the breach models are coupled with the Sobek
1D2D modelling package developed by WL | Delft Hydraulics (the Netherlands).
The choice of erodibility factor for the BREADA model turns out to have a high
influence on the breach outflow hydrograph timing and shape, but here we assume
a scenario where the breach development is slower, simulating flood propagation for
the fastest (BREACH) and slowest flood wave (BREADA). Roughness in the areas
subject to flooding influences flood speed and characteristics. A constant and vari-
able roughness based on land use data is used.
Based on preliminary population data for several regions inundated by flood water,
the Population at Risk consists of at least 10,000 people. Immediate construction
of a flood warning system is proposed, as a key source for avoiding loss of human
lives, for residents living as close as 9km away from the dam; since in case of a dam
failure event, this population faces significant risk.
Chapter 7
A Numerical-Constraint
based Model
The more constraints one imposes, the more one frees one’s
self. And the arbitrariness of the constraint serves only to
obtain precision of execution.
Igor Stravinsky
7.1 Introduction
During past decades different methodologies were proposed to prevent and manage
floods. The traditional approach to prevent the impact of flooding through flood
protection is more recently being replaced by a flood management approach (Plate,
2000; Brinkhuis–Jak et al., 2003; de Vriend, 2005; Samuels et al., 2005; Simonovic
and Ahmad, 2005) recognizing that absolute flood prevention is unachievable and
unsustainable, due to high costs and inherent uncertainties. In the Netherlands, the
standard policy of raising dikes crest levels in order to maintain the required level
of flood protection is being abandoned in favor of a new policy of creating ‘Room
for the River’ that involves widening river cross sections by relocating dikes further
away from the river, lowering floodplains, etc. (van Schijndel, 2005). Partitioning of
the area at risk into compartments, utilizing the highway and railroad embankments
as well as the natural terrain, could possibly lead to an increase in warning time and
a reduction of the flooded area.
The risk associated with flooding is usually defined as the product of probability
and consequences. This expression leads to a situation where the total risk of an
event with low probability and high consequences is equal to the total risk of an
event that has high probability of occurrence but low consequences. Generally, risk
reduction measures aim to reduce the probability of flooding and might often be
104 Section 7.2. Description of the numerical–constraint based approach
easier to implement, although minimizing the probability of a flood can come at the
expense of increasing its destructive power (when it happens), thus increasing the
consequences.
Simulation of the consequences of a predicted or assumed flood is typically elabo-

rated in what–if scenarios. Every what–if scenario requires specification of the initial
state and the configuration of the system. Considering different scenarios makes the
number of required simulations grow increasingly large due to the possible combina-
tions of choices for the initial state and the configuration of the system. Therefore,
in practice, only a small fraction of all possible strategies can be explored.
In this chapter we present a numerical–constraint based approach for flood risk re-
duction and decision support as well as a case study illustrating its application to a
system of polders at risk (Zagonjolli et al., 2006). The model developed is able to
simulate different flood mitigation scenarios taking into account the social and eco-
nomic value of the areas that could be prone to inundation. The impact of flooding
is minimized by selecting the most feasible mitigation scenarios.
The hydrodynamic modelling packages require extensive and lengthy calculations,

while the proposed ‘lightweight’ numerical–constraint based technique offers advan-
tages of simultaneous evaluation of different flood mitigation scenarios, taking into
account different economic and social aspects that a traditional numerical system is
not able to use. In this chapter we demonstrate the possibility of integrating both
approaches, viz. the usage of the developed method for providing the most feasible
mitigation scenario and detailed hydrodynamic modelling that can be carried out at
a later stage. This in turn allows for inclusion of the proposed approach, into any
existing hydrodynamic modelling package.
7.2 Description of the numerical–constraint based

approach
The key elements of the developed numerical–constraint based technique are the
graph algorithms that are applied to simulate different possible flood propagation
scenarios occurring in areas prone to inundation. Thus, a flood mitigation problem
is converted into a problem of finding the most feasible path for water propagation,
which can be addressed within the existing framework of graph algorithms (Corman
et al., 2001).
Figure 7.1 shows the transformation of a flood prone area into a graph. The area
surrounded by outer dikes (either wet or dry boundary dikes) consists of 12 polders
separated by inner dikes (dikes that belong to more than one polder). Each of the
polders, represented as a vertex v of the graph, has a maximum capacity to store
Chapter 7. A Numerical-Constraint based Model 105
L K
L
I K
J I
H
J
H
F
G F
D G
E
C D
E C
B A
B A
Figure 7.1: Transformation of a system of polders into the graph.
flood water. For each inner dike there is an edge connecting two vertices (polders).
Therefore, all possible ways of flood propagation in the area prone to flooding are
represented as all possible paths from the initially flooded polder to the polder that
needs to be protected. For example, we can assume an initial breaching to occur
at the river dike in polder C as a result of hydraulic conditions or as a result of
the decision to deliberately initiate breaching at that particular location. We also
assume that polder J is estimated (based on the given evaluation) to be the most
important polder that requires maximum protection. Then the algorithm finds the
optimal route for the flood water propagation avoiding the most valuable polder(s)
or a particular polder that needs to be protected. Therefore, one of the main ques-
tions arising while addressing a flood mitigation problem, that is find all possible
paths for flood water before it arrives at the protected area, is efficiently addressed
by our numerical–constraint based model, utilizing graph algorithms with different
selection criteria as discussed later in this chapter.
A notable advantage of this approach over traditional numerical models is the calcu-
lation of a single (or small number) of paths found by the graph algorithm to be most
feasible flood propagation scenarios, resulting in minimal damage. Using a ‘forward
approach’ involving traditional numerical simulation would require a ‘brute’ force
approach of evaluating all (many) possible ways of the flood propagation, thus con-
ducting extensive simulations and only later providing the user with the possibility
to manually select the most appealing scenario. This might lead to unnecessary
complications during the simulations, resulting in high computation cost and com-
plexity of the model.
We distinguish two ways of modelling flood propagation by:

1. Ignoring economic, social and environmental values of the area, thus reflecting
a ‘naturally’ occurring flood propagation scenario; or
2. Considering economic, social and environmental values of the area prone to
flooding and proposing a preferred route for the flood water to follow.
Another important feature of the proposed method is that different constraints can
be taken into account for social, economic and environmental values of the areas
prone to flooding. The model is able to determine the most valuable polder(s) to be
protected, based on user supplied information. While the value of economic assets
can be determined with some degree of certainty, it is much more difficult to ac-
curately estimate the value of non–economic assets such as environmental, cultural
and social considerations. Objects having national significance can be taken as an
important constraint for inundation of a particular area, and - together with the
population at risk - will contribute to the overall score of the polder’s value.
Let us denote the directed graph in Figure 7.2 as G = (V, E) where V is the vertex
set corresponding to the available polders and E is the edge set and any edge corre-
sponding to the to link between adjacent polders. In case of directed graphs every
edge (u, v) ∈ E is represented through a set of ordered pairs of vertices. In our case,
the direction of each of the edges is not decided beforehand, but according to the
hydraulic conditions in each polder at the every time step, as explained later in the
chapter. We distinguish two types of vertices in our graph: a source s and a sink t.
If polder C is the initial flooded polder then this polder will be the source vertex in
the graph network and polder J, the most valuable polder in the area which we aim
to protect from flooding, will be the sink vertex. Other vertices in the graph that
are neither sources nor sinks are called intermediate vertices. Water flows through
the intermediate vertices and might be stored in them. We assume that every vertex
lies on some path from the source to the sink vertex. The rate at which water enters
a vertex must equal the rate at which it leaves the vertex and the amount remaining
in the vertex, so mass conservation is preserved.
A non–negative real–valued function p is defined on the edges set and represents the
capacity function of our network G. Its value on a particular edge is the capacity of
the edge e that represents the maximum rate of flow from one polder in the other.
The capacity function of the network is related to the initial flow conditions at polder
C.
The flow in the network should satisfy the following constraints

0 ≤ g (e) ≤ p (e) , ∀ e ∈ E (7.1)
s1
s2 t1
S’ s3 t2 t’
s4 t3
s5
Figure 7.2: A network with multi source and multi sink vertices.
and
g − (v) = g + (v) , ∀ v ∈ V. (7.2)
The value g (e) expresses the rate at which the water flow travels along e. The
upper bound in condition (7.1) is called the capacity constraint; it imposes the
natural restriction that the flow along an edge cannot exceed the capacity of the
edge or cannot be higher than the available water in the polder from where water
is flowing out. Condition (7.2), called the conservation condition, requires that
for any intermediate vertex (polder), the volume of water flowing into v is equal
to the volume flowing out of v. The conservation condition implies that the flow
entering Polder C is equal to the sum of the volumes of water distributed over the
intermediary and sink vertex. Our aim is, given the network G with source s and
sink t, to find the best way of accommodating a flood in the paths consisting of
polder vertices that have a low economic value. As polder J is the polder to be
protected from being flooded, we assume that the graph has many sinks that are
the polders adjacent to the polder to be protected, and name polder J the supersink
t.
Inundation of areas is based on mass conservation (continuity equation):
dVp
= Qin − Qout , (7.3)
dt
where V is the polder volume, t is time and Qin and Qout are the flow rates in each
direction into and out of the cell (see Figure 7.3), corresponding to flow through the
breaching dike.
Main Properties of proposed Numerical Model
Qoutflow
dV
= Q up + Qdown + Qleft + Q right
dt
Figure 7.3: Mass conservation principle in an area compartmentalised in many pold-
ers. Each cell in the domain represents a polder.
There are different factors that might initiate the failure of a dike structure as de-
scribed in Section 2.2. PC–Ring (Vrouwenvelder, 1999; Vrouwenvelder et al., 2001)
is a tool developed and used in the Netherlands to calculate the total failure proba-
bility of a flood defence system consisting of dikes. A dike ring system is as weak as
its weakest link (dike section). Thus, the failure probabilities are calculated for the
weakest links in a system of dikes and not for every dike section. The total failure
probability includes the failure probabilities for overtopping and overflowing, insta-
bility of inner or outer slope, uplifting/piping. The failure probabilities are usually
calculated for dikes along a river, sea or other water body; it is common that the
failure probability of dikes that are not exposed to water, i.e. tertiary dikes is not
known.
Currently, in our model, we consider the overflowing failure mode as the most likely
failure mode for dikes during a flood event. During overflowing, the water level
exceeds (even in the absence of waves) the crest level of the dike and the water
flowing over the dike and along the inner (landward) slope induces structural fail-
ure. The probability of water levels exceeding the dike crest level is expressed as
P (Z < 0), where Z is a reliability or state function that represents the difference
Strength − Load. For Z < 0, the load is higher than the strength of the dike and as
such it fails, and vice versa (for Z > 0 the dike does not fail). The state when Z = 0
is called the limit state. For overflowing of the dikes, Z is expressed as follows:
1/3
q 2c

Z = hd + − hw , (7.4)
0.36g
where hd is the dike height [m], qc is the critical discharge [m3 /s] or the maximum
discharge flowing over the dike for which failure of the dike is not initiated, and hw
is the occurring water level [m]. The second term in Eq. 7.4 expresses the critical
depth on a broad crested weir.
The flow exchange between two polders in case of a dike failure from overflowing is
estimated using the BREADA model (Zagonjolli et al., 2005; Zagonjolli and Mynett,
2006a,b) presented in Chapter 6 that is adapted for dike failure by increasing the
erodibility coefficient and taking into account the submergence of the breach flow.
Since the BREADA model is applicable for non–cohesive dams, that means that all
the dikes in the domain are assumed to be non–cohesive dikes.
The mutual dependence between the safety of different compartments (polders or

dike ring areas) comes from hydraulic and economic interactions as summarized in
the following:
Due to high discharge in the river, deliberate failure of one of the dike sec-
tions can be induced, which need not be the weakest dike section along the
river. This deliberate failure leads to flooding of a less valuable area while it
might lead to the attenuation of the flood wave in the more valuable areas
downstream, preventing other potential failures.
When two polders or dike ring areas share a common dike, its failure leads to
flooding of both adjacent areas.
Failure of one dike section might prevent failure of another dike section due to
a decrease of the water level in the polder.
From an economic point of view: inundation of one area can impose damage to
another non–inundated area due to economic links (e.g. infrastructure, etc.).
The hydraulic interaction between compartments (dike ring or polder areas) influ-
ences the protection level of the area that might depend on another one. Further-
more, risk reduction in one particular area might be followed by an increased risk
of flooding in the other one. Thus, a detailed assessment has to be made, including
the interaction between different compartments. Within the framework of the pro-
posed approach the interactions between compartments of a dike ring are taken into
account. To minimize flooding impact on the complex hydraulic system of dike ring
areas, different objectives could be taken into consideration during the simulation
process:
1. Natural flooding: flooding of the compartments naturally occurring under spe-
cific hydraulic conditions
2. Longest Path: intentional flooding of compartments for accommodating the
water, that is finding the longest path for the flood water propagation before
it reaches the protected polder.
3. Maximum Storing Volume Capacity: selection of polders to be flooded based

on their maximum volume capacity.
4. Minimal Total Damage: selection of polders to accommodate the flood water

based on their socio–economic, cultural and environmental value. Less valuable
polders to be flooded first.
5. Combination of the above objectives (2-4).
Below we describe in more detail how these objectives can be satisfied within the
proposed methodology.
Natural flooding
We define the flooding of the area that occurs without deliberate intervention as oc-
curring naturally. In a system of polders it is usually preferable to take into account
the probability of several dike failure modes as well as the associated uncertainties.
Knowing the failure probability of all dike sections is beneficial when identifying the
potential breach location. Moreover, when deliberate breaching of a dike is con-
sidered, it provides the possibility to choose the dike section with highest failure
probability. Our modelling tool is built to take into consideration this information,
if provided by the user. However, in our example application we consider only dike
elevation and water depth in the polder as the primary source for breach initiation.
Once the water level in the polder raises higher than the top elevation of any of the
surrounding dikes, overflow breaching of that dike is assumed to be initiated leading
to the flooding of the polder. Another important element for dike breach initiation
is the duration of the overflowing water, which should be large enough to initiate
erosion. We do not specify the overflow duration time in our model, but ensure that
the duration is long enough as the filling up of the first polder coincides with the
rising limb of the inflow hydrograph, which implies that further increase of water
level in the polder is expected.
Longest path flooding

Probably one straightforward way to prevent flood water reaching the most valuable
polder, is the requirement to accommodate as much water as possible along the flood
path. That is, to select the longest path from the source polder to the protected one.
Our numerical–constraint based model employs for that purpose a graph algorithm
that generates all possible paths between them and selects the longest possible one.
The algorithm returns a list of vertices (polders), including the start and the end
vertices (polder C and J) comprising the path. The same vertex does not occur
more than once in the returned path (no cycles are allowed). In some cases the
algorithm can return more than one unique longest path, if several are available. In
these situations, our model randomly selects a single path or provides the user the
possibility to select a preferred one.
Then the numerical model for natural flooding uses the suggested path to simulate
controlled flood propagation. This means that previously not breached dikes will be
breached for the purpose of accommodating the excess water and controlling flood
propagation to protect the most valuable polder. In case the flood could be retained
within the polders located on its path before reaching most valuable polder, then J
will be dry, otherwise it will be the last polder to be flooded, thus having more time
for evacuation.
Maximum storing volume capacity flooding

Another possibility to defend the most valuable polder as well as to cause minimal
damage to the rest, is to accommodate the flood water in the polders with maximum
storage capacity. Each of the polders in the domain has a maximum storage capacity
which is assumed to be equal to the polder area multiplied by the minimum height
of its surrounding dikes. The algorithm tries to find the best solution for accom-
modating the flood water in the polders that have higher storage capacities. For
this purpose we again utilize a graph algorithm that finds the path, such that its
storage volume capacity is maximal. Then the numerical–constraint based model
uses this path to accommodate flood water and to protect the most valuable polder
from flooding, in the same way as discussed previously.
It can be observed that in some cases both the longest path and the maximum stor-
age volume objectives could lead to the same solution. Obviously, in case of a unique
longest path, it possesses the largest storage capacity than any other one. However,
the situation is completely different in case a restriction on the number of polders
that should be used to accommodate flooding is imposed.
Minimal total damage flooding

Finally, the most economically feasible flood propagation scenario could be imposed
by enforcing a minimal total damage constraint. The key idea behind this approach
is to identify and deliberately flood areas that have the least economic value. Thus,
controlled flooding is aimed to produce the least total damage although deliberately
flooded areas might have low flooding probability. The graph algorithm finds a path
along the least valuable areas in the domain based on the user defined input data
in terms of socio–economic, cultural and environmental value. Below we present an
application of our method using synthetic data to demonstrate its main features and
usability.
7.3 Application
The domain area of the case study considered in this section is schematized in Figure
7.4 together with the information about dike height and top elevation as well as the
polder’s maximum storage capacities and economic values. The value is expressed in
L 2.3; 8.8 K
2; 8.5 (0.18; 0.7)L K
5
2; 8. 2; 8.5 (0.16; 0.7)
2.4; 8.9 I 2; 8.5 (0.25; 0.8)I

J H
(0.15; 1)J
2.6; 9.1
2.3; 8.8 H
F (0.07; 0.1)
2; 8.5
G 2.2; 8.7 F(0.13; 0.6)
2.2; 8.7 D
(0.11; 0.7)G
E 2; 8.5 2; 8.5
2.9; 9.4 2.8; 9.3 C D(0.16; 0.4)
3.1;
9.6 3; 9.5
.9 (0.1; 0.5)E C
;8
2.4 (0.08; 0.1)
B A
2.4
; 10
;
3.5
8.9
(0.384; 0.3)B A(0.184; 0.3)

5.5; 12
Figure 7.4: Transformation of the case study area into the graph. Dike height and
the dike top elevation [m] as well as polder area [km2 ] and its relevant value are
presented.
relative terms varying in a scale from 0 to 1. However, the overall score is calculated
based on a particular economic value of each polder and its social and environmental
aspects are considered as contributing elements, although not expressed in mone-
tary terms. These values are not depending on the water level or velocities in the
polder, but are used as indicators for the polders that are more valuable and require
most protection. Including a depth–damage function as one of the constraints in our
model is foreseen as future work and can be incorporated into existing framework.
TAW (2002) provides guidelines for the critical discharge on overtopped sea dikes of
clay or sand material with either good or bad grass cover. However, to our knowl-
edge, no information exists concerning the critical discharges of overflowed inner
dikes. In the case study, qc = 0.16m3 /s is assumed , which leads to a water depth of
0.2m above the dike crest to initiate dike failure. Due to the fact that the application
is used for demonstration purposes of the developed method, we may argue that our
assumption can be considered reasonable for the purpose of dike breach initiation.
The initial dike to breach is the river dike of Polder C. The breach outflow for that
dike is assumed rather than estimated using a breach model. The hydrograph has
maximum peak outflow of 400 m3 /s. As can be observed in Figure 7.4 the border
dikes have higher heights to satisfy the property of rigid boundaries that never suffer
failure. The same condition is applied for the other river dikes as well.
2.5
Water depth [m]
2.0
1.5
1.0
0.5
Time [s]
0.0
0 2000 4000 6000 8000 10000
C D E F G J L K I H
Figure 7.5: The natural flooding of the polders - water depth in polders as a function
of time.
In Figure 7.5 the rate of change of water depth in every polder for a natural prop-
agation of the flood in the domain is shown. It can be observed that Polder J is
flooded. The breaching of the dikes is initiated when the water depth is higher than
hd + 02m. For simplicity, our domain is considered flat with bottom elevation of the
dikes to be the same.
Using the Longest Path objective, for the area presented in Figure 7.4, the longest
path suggested by the algorithm is [C, A, B, D, E, G, F, I, H, K, L, J] (see Figure
7.6 and Figure 7.7). Note, that the first and the last polders are considered as most
crucial for the model, however the intermediate polders might not necessarily be
flooded based on their ranking order in the path. We observe that in this case the
polder we aim to protect is dry at the end of the simulations. The flood water is
directed towards the other polders and the result is the expected flood propagation
scenario at the end of the simulation.
The path obtained when selecting the polders with the largest maximum volume
capacity is [C, A, B, D, E, G, F, I, H, K, L, J] which corresponds to the one ob-
tained for the longest path. This could be expected since the longest path and the
maximum volume objective appeared to be the same objectives in terms of storage
capacity.
A very different flood path propagation (compared to the longest and maximum
storage paths) is obtained for the minimum total damage scenario, thus selecting
2.0 Water depth [m]
1.5
1.0
0.5
Time [s]
0.0
0 2000 4000 6000 8000 10000
C D E F G I L H K
Figure 7.6: The longest path and the maximum polder volume flooding scenarios.
Natural Flooding Longest Path

L K L K
I I
J J
H H
F F
G G
D D
E C E C
B A B A
C-D-E, D-F, E-G, F-G-J-L-K-(IH) C-D-E, D-F, F-G, E-G, F-I-(HK), K-H, K-L
Figure 7.7: The natural and longest path flooding of the polders in graph represen-
tation.
2.5
Water depth [m]
2.0
1.5
1.0
0.5
Time [s]
0.0
0 2000 4000 6000 8000 10000
C D E G F B A
Figure 7.8: The minimum damage flooding scenario.
Largest Storing Volume Smallest Value

L (0.18) K (0.16) L (0.7) K (0.7)
I (0.25) I (0.25)
J (0.15) J (1)
H (0.07) H (0.1)
F (0.13) F (0.6)
G (0.11) G (0.7)
D (0.16) D (0.4)
E (0.1) C (0.081) E (0.5) C (0.1)
B (0.384) A (0.184) B (0.3) A (0.3)
C-D-E, D-F, F-G, E-G, F-I-(HK), K-H, K-L C-D-E-G-F, E-B-A
Figure 7.9: The flooding of the polders based on the largest storage volume and the
minimum value objective.
the less valuable polders in our domain to be flooded. The polders are found to be
[C, A, B, E, D, F, G, J] and the flooding pattern is shown in Figure 7.8 and Figure
7.9. The Polder J is dry at the end of this simulation.
7.4 Conclusions
The main idea behind the proposed approach of using graph theory to develop con-
straints for numerical simulations is to manage the flooding process within a system
of polders or dike ring areas divided into a number of compartments. For this pur-
pose we use a numerical–constraint based model that takes into account not only
physical processes involved in dike breaching and flow propagation but also social,
economic and environmental values of the areas prone to inundation. We simulate
different scenarios of flooding based on the hydraulic processes and the constraints
introduced to minimize the flood damage in each protected area.
By temporarily storing the water in compartments, the total risk might be mini-
mized. The flooding probability is increased for some compartments, while the total
flood risk is decreased. The filling process of the compartments depends on the fail-
ure process of the dike and on the condition whether a deliberate breaching of the
dike is involved or ‘natural’ failure of hydraulic structure occurs. In both cases the
same breaching model is used, but in the former case dikes are deliberately lowered.
There is a possibility to include faster breach development of the dikes that might
not have a slow breaching process depending on the methods used to remove (open)
a dike section. Furthermore, the model can be enhanced by including the presence
of additional hydraulics structures.
The key difference of our method in comparison to existing ones, is that instead
of assigning a particular probability to a scenario, we propose an alternative ap-
proach of evaluating and mitigating consequences based on a numerical–constraint
based approach. In particular, we evaluate different ways for flood mitigation dur-
ing a flood event. In the simulation we take into account compartments created by
the available dikes, roads and natural high grounds in the areas prone to flooding.
Through constraints, we optimize the strategy for choosing the most feasible flood
propagation scenario that minimizes socio–economic consequences. This is done by
utilizing algorithms based on graph theory that efficiently transform dike ring areas
into a graph and find different optimal paths applicable for each objective function.
The main strategy consists of navigating floodwater ‘away’ from the most populated
or economically valuable regions.
Based on the considerations presented above, we may suggest that flood risk man-
agement could draw significant benefit from a constraint based methodology by using
a set of measures that ‘keeps the system dry’. Including targets and goals inside
the support system, allows better prediction of different scenarios since in this case
the model will provide optimal solutions. However, the case might be that some
extreme floods in the low-lying areas, caused by prolonged heavy rainfall might not
be prevented as a result of measures taken or the method used.
Chapter 8
Conclusions and
Recommendations
. . . the river . . .
Keeping his seasons and rages, destroyer, reminder
Of what men choose to forget. Unhonoured, unpropitiated
By worshippers of the machine, but waiting, watching and
waiting.
T.S. Eliot
8.1 Conclusions
In this thesis, various approaches are proposed and applied for simulating dam and
dike failure events, flood water routing in downstream areas, and flood risk reduction,
providing a unified framework for addressing flood related events. Numerical, data
mining, graph theory and constraint based methods are applied to the problem of
breach modelling and flood water mitigation. The findings related to each of the
proposed approaches are listed below.
8.1.1 Traditional and novel approaches for predicting dam

breach characteristics
For every dam or dike there is a limit to which the structure is built to withstand
the forces applied. Unexpected or unforeseen events might trigger failure, when the
strength of the structure is weaker than the acting load. In this thesis a range of
modelling techniques is explored to deal effectively with the failure event of hydraulic
structures such as earthfill dams and dikes.
We focus on dams and dikes composed of earth material since they constitute the
largest percentage of total number of dams and dikes built around the world and the
largest number of failures occurs in these structures as well. For structures with co-
hesive material, two processes are most likely to develop: erosion and headcutting.
The latter was found to be predominant during field and laboratory experiments
(Morris, 2005). The cohesive bond between the particles of the cohesive soil presents
higher resistance to the flow, while for the non–cohesive soil the physical properties
of the particles (size, shape, density, porosity, and fall velocity) contribute to the
resistance to the flow. Because the particles are not bound together, the erosion
process is believed to be the predominant breaching mechanism. However, any in-
stability in part of the structure might also result in mass failure. In this thesis, a
dam breach model is developed for simulating overtopping failure of non–cohesive
dams.
BREADA model
The BREADA model developed in this thesis assumes two different breach shape
evolutions in time and space with two important time intervals. In the first formu-
lation, the initial trapezoidal breach shape develops in vertical and lateral direction,
until the ground is reached and then progresses in lateral direction only. In the
second formulation, a triangular breach shape is initially assumed and continues to
expand in vertical direction until the bottom level of the structure is reached. During
the second phase, the trapezoidal breach progresses in lateral direction until either
the water flow force can not cause any further erosion or the breach has reached its
maximum allowed dimensions that are limited by the dam geometry. Flow through
the breach channel is calculated as a weir flow and the erosion of the dam material is
estimated with a bed load transport empirical equation. The model developed is val-
idated against the historical failure event of Schaeffer Dam in USA. Good agreement
is observed in terms of peak outflow prediction for a range of model parameters.
Data mining models

We apply data mining techniques for deriving dam breach characteristics and peak
outflow from available data. These methods generally rely on large amount of data,
but in the applications presented in this thesis, the information about historical
dam failure events is sparse. Still, performance of data mining techniques is found
to be slightly better than the current empirical equations, suggesting that enriching
the databases with more data about real dam failure events is one possible way
to improve the models that can further be used to evaluate the capabilities of the
physically based methods by providing the range of peak outflow and breach char-
acteristics or to provide some of the input information needed for the simulation of
the semi–physically based models.
Uncertainty in dam breach modelling

The uncertainty associated with each of the processes involved in forecasting flood
risk occurring due to breaching of structures is identified and quantified in relative
Chapter 8. Conclusions and Recommendations 121
terms. We distinguish between model uncertainty and input uncertainty.
Four different methods to calculate dam breach peak outflow are evaluated against
the documented failure event of the Schaeffer Dam. We compare the breach outflow
hydrograph and the peak outflow obtained from our model with the BREACH model
as well as with the range of peak outflows obtained from empirical equations and
data mining techniques. The results show a wide range of predicted peak outflows.
When using empirical equations, the peak outflow discharge varies in a range of
±50%. Though the best performing empirical equation for the Schaefer dam can be
identified, this does not imply its direct applicability to other failure events. When
the dam breach models BREACH and BREADA are used the predicted peak outflow
varies in a narrower range depending on the assumptions and selected parameters.
Good results are obtained from the IBk data mining model that predicts only a 10%
lower peak outflow than the recorded one.
Many dam material characteristics are needed for modelling, but their (precise) val-
ues are not always available. The largest uncertainty is observed in the hydrograph
shape, particularly in the timing of the breach development. While the BREACH
model exhibits a very sharp rising limb meaning fast development of the breaching, a
smooth rising limb is observed for the BREADA model, the former leading to worst
downstream flood conditions while the slower wave produed by the latter will leave
more warning time to the population downstream.
8.1.2 Identification of dam failure flood extent, uncertainty,

and risk reduction
We simulate the hypothetical failure of the Bovilla Dam, an earthfill dam of 81m
high and observe the impact of two outflow hydrographs obtained from the BREADA
and BREACH models, in the downstream area. In terms of flood duration, the fast
rising breach outflow hydrograph leads not only to faster flood propagation, but
also to higher water depths in the flat areas further downstream the dam. The slow
developing flood wave leaves room for longer warning period for the population and
the lower water depth values in the flat areas, but higher water depths are observed
in the first kilometers from the dam. The models, one leading to a less catastrophic
scenario than the other, are examples of the uncertainty that current breach mod-
elling techniques possess with respect to the most important element - timing. Peak
outflows of the models do not differ much, and there is a good agreement with the
empirical equation predictions.
Other sources of uncertainty also affect the modelling of dam break flood propaga-
tion, especially when failure of large dam is involved. The roughness of the areas
subject to flooding is not known beforehand. The resulting flood propagation sce-
narios depend on the roughness coefficient used (either constant or land use based
one). Moreover, we ignore the reservoir sediment or the dam material, which are
flushed away together with the water stored in the reservoir, but instead consider
propagation of clear water in the valley. The presence of the material attached to the
water, which will affect the flow pattern significantly is taken into account through
higher values of roughness coefficients.
Once the potential risk is identified, the task is to find the ways of reducing this risk
and take measures for warning the population downstream in case of a structural
failure event. Our modelling results can be used as flood hazard maps and can assist
communities in planning future developments in areas that are prone to flooding.
8.1.3 A numerical–constraint based approach for flood miti-

gation
The flood risk associated with the failure of a high dam with a large volume of water
stored in the reservoir and the flood risk associated with the failure of a river dike
during a high discharge event is highly dependent on the vicinity of the structure
to the populated areas and the measures taken for dealing with potential flood-
ing. Here, we introduce a numerical–constraint based model for flood mitigation
in low–lying areas that are subject to flooding during a flood event. Instead of as-
signing a particular probability to a scenario, we propose an alternative approach
of evaluating and mitigating consequences based on a numerical–constraint based
model. During the simulation we take into account compartments created by the
presence of dikes, roads and natural high grounds in the areas prone to flooding.
Through constraints, we optimize the strategy for choosing the most feasible flood
propagation scenario that minimizes socio–economic consequences. This is done by
utilizing graph based algorithms that efficiently transform compartmentalized areas
into a graph and find different optimal paths applicable for each objective function.
The strategy consists of navigating floodwater ‘away’ from most populated or eco-
nomically valuable regions.
We believe that flood risk management could benefit from a constraint based ap-
proach by using a set of measures that keep the system dry. Including objectives
in a hydrodynamic modelling package, allows better prediction of different scenarios
since in this case the model will provide optimal solutions. However, it might still be
the case that some extreme floods or floods caused due to prolonged heavy rainfall
over the domain area might not be prevented as a result of measures taken or the
method used.
Chapter 8. Conclusions and Recommendations 123
8.2 Recommendations
The following recommendations for further research are suggested.
During the past three decades different dam breach modelling methods have
been proposed for predicting the outflow hydrograph and peak outflow value.
However, none of the methods can guarantee to obtain accurate breach charac-
teristics. The complexity of breach development usually motivates researchers
to make many approximations to obtain simpler mathematical description of
the processes involved, thus increasing the uncertainty related to the structural
failure development. This is the case when modelling erosion and headcut ero-
sion processes. Many laboratory and field experiments have been performed
to provide in depth understanding of these processes. However, the complete
physical description is far from perfect. A possible direction for future inves-
tigation could be the analysis of erosion processes in a river stretch during a
high flow and evaluating its application to a dam breaching event.
For high dams the monitoring and recording of failure events has not been
satisfactory so far, making the calibration of breaching models difficult and
leading to uncertainty in the prediction results. The drawback of the exper-
imental tests is that the weakest spot on the dam structure is predefined,
influencing further development of the breaching. The optimal solution is real
time dam monitoring and aerial pictures of breach development, but these
proposed measures might be difficult to implement, since breaching caused by
overtopping usually occurs during bad weather conditions, therefore limiting
the monitoring possibilities. However, due to technological advances both in
hardware and software sensor technology, one might expect (real time) moni-
toring of dikes and dams during breach development to become possible and
feasible.
Scale models are currently the main source of validation data for dam/dike
breach modelling. Although hydraulic conditions, such as water velocity, etc.
are effectively replicated in laboratories, the exact scaling factors for the soil
properties and sediment transport are usually not available. Therefore, an
interesting direction for future research is to investigate scaling of soil related
properties. Furthermore, simulation and visualization of breaching processes
in 3D might lead to better descriptions of the phenomena.
Improving the accuracy of existing physically based models by using the in-
formation extracted from a database of recorded dam failure events is another
direction of this research. Currently, the available database consists of about
100 dam failure events and suffers from inconsistencies between the different
sources. Increased efforts in documenting and sharing the information related
124 Section 8.2. Recommendations
to historical dam dike failure events is necessary. This will lead not only to bet-
ter validation of the physically based models, but also to more accurate data
mining models for discovering patterns and relationships among dam failure
events.
During structural breaching, the dam material is flushed away together with
the water, changing the morphology in the areas downstream. The modelling
of flood propagation is usually accomplished by assuming clear water and
rigid terrain throughout the flood water propagation domain. However, in-
cluding the interaction of debris flow and non–rigid terrain boundaries may
significantly affect the flooding pattern. This problem has recently received
significant attention and further research in this area is necessary.
The numerical–constraint based approach proposed in this thesis can be ex-

tended by introducing different objective functions that are applicable for spe-
cific scenarios. Furthermore, its capabilities can easily be boosted by using
an advanced numerical engine of any well known hydrodynamic package. On
the other hand, it is also possible to include the constraint–based part of the
method into existing hydrodynamic packages, which are often part of the deci-
sion making process for proposing measures needed to protect important areas
in terms of their socio–economic and cultural value.
125
Bibliography
Abbott, M. B.: 1979, Computational hydraulics: Elements of the theory of free

surface flows, Pitman Publishing Ltd., London.
Albinfrastrukture: 1996, Diga e Bovilles (in albanian and italian), Technical report,
Ministry of Construction, Albania.
Alcamo, J. and Bartnicki, J.: 1987, A framework for error analysis of a long-range
transport model with emphasis on parameter uncertainty, Atmospheric Environ-
ment 21(10), 2121–2131.
Alonso, C. V., Bennett, S. J. and Stein, O. R.: 2002, Predicting head cut erosion and
migration in concentrated flows typical of upland areas, Water Resources Research
38(12), 39.1–39.15.
Archer, D.: 2006, Global warming: Understanding the forecast, Blackwell Publishing
Ltd.
Asselman, N. E. M. and Middelkoop, H.: 1995, Floodplain sedimentation: Quanti-

ties, patterns and processes, Earth Surface Processes and Landforms 20(6), 481–
499.
Babovic, V., Keijzer, M., Aguilera, D. R. and Harrington, J.: 2001, An evolutionary
approach to knowledge induction: Genetic programming in hydraulic engineer-
ing, ASCE, Proc. of the World Water and Environmental Resources Congress,
Orlando, Florida, USA, pp. 64–64.
Birkes, D. and Dodge, Y.: 1993, Alternative methods of regression, John Wiley, New
York.
Bogárdi, J.: 1974, Sediment transport in alluvial streams, Akadémiai Kiadó, Bu-
dapest, Hungary.
Bowles, D. S.: 2001, Evaluation and use of risk estimates in dam safety decision
making, Proc. of the United Engineering Foundation Conference on Risk-Based
Decision-Making, Santa Barbara, California, USA, pp. 17–32.
126 Bibliography
Brinkhuis–Jak, M., Holterman, S. R., Kok, M. and Jonkman, S. N.: 2003, Cost
benefit analysis and flood damage mitigation in the Netherlands, Paper for the
ESREL Conference, Maastricht, the Netherlands, pp. 16–18.
Broich, K.: 1998, Mathematical modelling of dam break erosion caused by overtop-
ping, Proc. of the 2nd CADAM Meeting, Universität der Bunderswehr, München,
Germany. Available at www.hrwallingford.co.uk/projects/CADAM/.
Broich, K.: 2003, Sediment transport in breach formation process, Proc. of the
3rd IMPACT Workshop, Université Catholique de Louvain, Louvain–la–Neuve,
Belgium. Available at www.impact-project.net.
Brown, C.: 1950, Sediment transportation, Engineering Hydraulics pp. 769–857.
Brown, R. J. and Rogers, D. C.: 1977, A simulation of the hydraulic events during
and following the Teton Dam failure, Proc. of Dam–Break Flood Routing Model
Workshop, U.S. Water Resource Council, Bethesda, Maryland, pp. 131–163.
Brown, R. J. and Rogers, D. C.: 1981, Users manual for program BRDAM, U.S.
Bureau of Reclamation, Denver, Colorado.
Chanson, H.: 2005, Applications of de Saint–Venant equations and method of char-
acteristics to the dam break wave problem, University of Queensland, Australia.
Chaudhry, M. H.: 1993, Open-channel flow, Prentice Hall, New Jersey.
Chow, V. T.: 1959, Open channel hydraulic, McGraw-Hill, New York.
Coleman, S., Andrews, D. and Webby, M.: 2002, Overtopping breaching of noncohe-
sive homogeneous embankments, Journal of Hydraulic Engineering 128(9), 829–
838.
Commissie Rivierdijken: 1977, Rapport Commissie Rivierdijken (in Dutch). The
Hague, the Netherlands.
Corman, T. H., Leiserson, C. E. and Rivest, R. L.: 2001, Introduction to algorithms,
MIT Press, Cambridge, USA.
Costa, J. E.: 1985, Floods from dam failures, Open-file report, U.S. Department of
the Interior Geological Survey, Denver, USA.
Costa, J. E. and Schuster, R. L.: 1988, The formation and failure of natural dams,
Geological Society of American Bulletin 100(7), 1054–1068.
Cristofano, E. A.: 1965, Method of computing erosion rate for failure of earthfill
dams, U.S. Department of the Interior, Bureau of Reclamation, Denver, Colorado.
Cunge, J. A., Holly, F. M. and Verwey, A.: 1980, Practical aspects of computational
river hydraulics, Pitman Publishing Ltd., London.
Bibliography 127
CUR/TAW: 1990, Probabilistic design of flood defences, Report 141, Technical Ad-
visory Committee on Water Defences, Center for Civil Engineering Research and
Codes, Gouda, the Netherlands.
de Saint-Venant, A. J. C.: 1871, Théorie du mouvement non-permanent des eaux,

avec application aux crues des rivières et á lintroduction des marées dans leur lit,
Comptes Rendus des Séances de l’Académie des Sciences 73, 147–154.
de Vriend, H.: 2005, State of the art in flood management research, Proc. of the 3rd
International Symposium on Flood Defence, Nijmegen, the Netherlands.
Deltacommissie: 1960, Rapport Deltacommissie (in Dutch), Staatsuitgeverij.
Dhondia, J. F. and Stelling, G. S.: 2002, Application of one dimensional–two dimen-

sional integrated hydraulic model for flood simulation and damage assessment,
Proceedings of the 5th International Conference in Hydroinformatics pp. 265–276.
Dodge, M. M.: 1997, Hans Brinker or the silver skates, The World Wide SchoolTM,
Seattle, Washington, USA.
Dodge, R. A.: 1988, Overtopping flow on low embankment dams–Summary report

of model tests, REC–ERC–88–3, U.S. Bureau of Reclamation, Denver, USA.
Dressler, R. F.: 1952, Hydraulic resistance effect upon dam-break functions, Journal
of Research, National Bureau of Standards, USA 49(3), 217–225.
Dressler, R. F.: 1954, Comparison of theories and experiments for the hydraulic dam-
break wave, Vol. 3, Proc. of the International Association of Scientific Hydrology,
Rome, Italy, pp. 319–328.
Dressler, R. F.: 1958, Unsteady Non-Linear Waves in Sloping Channels, Proceedings

of the Royal Society of London. Series A, Mathematical and Physical Sciences
247(1249), 186–198.
Du Boys, M. P.: 1879, Études du régime du Rhône et de laction exercée par les
eaux sur un lit à fond de graviers indéfiniment affouillable, Annales des Ponts et
Chaussées 5(18), 141–195.
Edgeworth, F. Y.: 1887, On observations relating to several quantities, Hermathena

6(13), 279–285.
Einstein, H. A.: 1942, Formulas for the transportation of bed load, ASCE Transac-
tions 107, 561–573.
Exner, F. M.: 1925, Ber die wechselwirkung zwischen wasser und geschiebe inflssen,
Sitzungber. Acad. Wissenscaften Wien Math. Naturwiss, Abt. 2A, 134 pp. 165–
180.
128 Bibliography
Flannery, T.: 2006, The Weather Makers: how man is changing the climate and
what it means for life on earth, Atlantic Monthly Press.
FLOODsite: 2005, Language of Risk-Project Definitions, Consortium Report T32-
04-01. Available at www.floodsite.net.
Follansbee, R. and Jones, E. E.: 1922, The Arkansas River flood of June 3-5, 1921,
Water-Supply Paper 487 .
Franca, M. J. and Almeida, A. B.: 2004, A computational model of rockfill
dam breaching caused by overtopping (RoDaB), Journal of Hydraulic Research
42(2), 197–206.
Fread, D. L.: 1988, BREACH: an erosion model for earthen dam failures, Hydrologic
Research Laboratory, National Weather Service, NOAA.
Froehlich, D. C.: 1987, Embankment-dam breach parameters, Proc. of the ASCE
National Conference on Hydraulic Engineering, Williamsburg, Virginia, USA,
pp. 570–575.
Funtowicz, S. O. and Ravetz, J. R.: 1990, Uncertainty and quality in science for
policy, Springer.
Galy-Lacaux, C., Delmas, R., Kouadio, G., Richard, S. and Gosse, P.: 1999, Long-
term greenhouse gas emissions from hydroelectric reservoirs in tropical forest re-
gions, Global Biogeochemical Cycles 13(2), 503–517.
Gauss, C. F.: 1809, Theoria motus corporum coelestium in sectionibus conicus solem
ambientium, Vol. Translation reprinted as Theory of motions of the heavenly bodies
moving about the sun in conic sections. Dover, New York, 1963, Perthes et Besser.
Gerritsen, H.: 2005, What happened in 1953? The Big Flood in the Netherlands in
retrospect, Philosophical Transactions: Mathematical, Physical and Engineering
Sciences 363(1831), 1271–1291.
Hagen, V. K.: 1982, Re-evaluation of design floods and dam safety, Vol. 1, Proc.
of 14th Congress of International Commission on Large Dams, Rio de Janeiro,
Brasil, pp. 475–491.
Hanson, G. J., Robinson, K. M. and Cook, K. R.: 2001, Prediction of headcut
migration using a deterministic approach., Transactions of the ASAE 44(3), 525–
531.
Hanson, G. J., Temple, D. M. and Cook, K. R.: 1999, Dam overtopping resistance
and breach processes research, Proc. of the Annual Conference Association of
State Dam Safety Officials, St. Louis, Missouri, USA.
Harris, G. W. and Wagner, D. A.: 1967, Outflow from breached earth dams, Uni-
versity of Utah, Salt Lake City, Utah, USA.
Bibliography 129
Hartford, D. N. D. and Baecher, G. B.: 2004, Risk and uncertainty in dam safety,
Thomas Telford.
Hervouet, J. M.: 2000, A high resolution 2-D dam-break model using parallelization,
Hydrological Processes 14(13), 2211–2230.
Hervouet, J. M.: 2007, Hydrodynamics of free surface flows: Modelling with the finite
element method, John Wiley & Sons.
Hoffman, F. O. and Hammonds, J. S.: 1994, Propagation of uncertainty in risk
assessments: The need to distinguish between uncertainty due to lack of knowledge
and uncertainty due to variability, Risk Analysis 14(5), 707–712.
Hooijer, A., Klijn, F., Pedroli, G. B. M. and van Os, A. G.: 2004, Towards sustain-
able flood risk management in the Rhine and Meuse river basins: synopsis of the
findings of IRMA-SPONGE, River Research and Applications 20(3), 343–357.
Huang, Y.: 2005, Appropriate modeling for integrated flood risk assessment, PhD
thesis, University of Twente, the Netherlands.
IMPACT: 2004, Investigation of Extreme Flood Processes and Uncertainties, EC
Contract EVG1-CT-2001-00037, www.impact-project.net.
International Commission on Large Dams: 2005, Risk assessment in dam safety
management: a reconnaissance of benefits, methods and current applications,
CIGB/ICOLD. Bulletin 130.
IPCC: 2001, Climate Change 2001: The scientific basis, Third Assessment Report
of the IPCC, Cambridge University Press.
Janssen, P. H. M., Heuberger, P. S. C. and Sanders, R.: 1994, UNCSAM: A tool for
automating sensitivity and uncertainty analysis, Environmental Software 9(1), 1–
11.
Johnson, F. A. and Illes, P.: 1976, A classification of dam failures, International
Water Power and Dam Construction 28(12), 43–45.
Klijn, F., van Buuren, M. and van Rooij, S. A. M.: 2004, Flood-risk management
strategies for an uncertain future: Living with Rhine River floods in the Nether-
lands, AMBIO: A Journal of the Human Environment 33(3), 141–147.
Knight, D., Cao, S., Liao, H., Samuels, P., Wright, N., Liu, X. and Tominaga, A.:
2006, Floods are we prepared?, Journal of Disaster Research 1(2), 325–333.
Krzysztofowicz, R.: 2001, The case for probabilistic forecasting in hydrology, Journal
of Hydrology 249(1), 2–9.
Kwadijk, J. and Rotmans, J.: 1995, The impact of climate change on the river
Rhine: A scenario study, Climatic Change 30(4), 397–425.
130 Bibliography
Lafitte, R.: 2001, Hydropower, The 20th edition of the Survey of Energy Resources,
World Energy Council.
Leal, J., Ferreira, R. M. L. and Cardoso, A. H.: 2002, Dam-break waves on movable
bed, Vol. 2, River Flow 2002, Proc. of the International Conference on Fluvial Hy-
draulics, Louvain-la-Neuve, Belgium, Swets & Zeitlinger, Lisse, the Netherlands,
pp. 981–990.
Lou, W. C.: 1981, Mathematical modeling of earth dam breaches, PhD thesis, Col-
orado State University, USA.
Loukola, E. and Huokuna, M.: 1998, A numerical erosion model for embankment
dams failure and its use for risk assessment, Proc. of the 2nd CADAM Meeting,
Universität der Bunderswehr, München, Germany.
Macchione, F. and Rino, A.: 1989, Dimensionless analytical solutions for dam-breach
erosion, Journal of Hydraulic Research 27(3), 447–452.
MacDonald, T. C. and Langridge-Monopolis, J.: 1984, Breaching charateristics of

dam failures, Journal of Hydraulic Engineering 110(5), 567–586.
McCarthy, J., Minsky, M. L., Rochester, N. and Shannon, C. E.: 1956, A proposal
for the Dartmouth summer research project on Artificial Intelligence, Retrieved
from www-formal.stanford.edu/jmc/history/dartmouth.
McCulloch, W. S. and Pitts, W.: 1943, A logical calculus of the ideas imminent in
nervous activity, Bulletin of Mathematical Biophysics 5, 115–133.
McCully, P.: 1996, Silenced rivers: The ecology and politics of large dams, Zed
Books, London & New Jersey.
Merabtene, T., Yoshitani, J. and Kuribayashi, D.: 2004, Managing flood and water-
related risks: A challenge for the future, Proc. of the AOGS Joint 1st Annual
Meeting & 2nd APHW Conference, Singapore.
Meyer-Peter, E. and Müller, R.: 1948, Formulas for bed-load transport, Proc. of the
2nd Meeting of the International Association for Hydraulic Structures Research,
pp. 39–64.
Millennium Ecosystem Assessment: 2005, Ecosystems and human well-being: Syn-

thesis, Island Press.
Minns, A. W. and Hall, M. J.: 1996, Artificial neural networks as rainfall–

runoff models, Hydrological Sciences Journal/Journal des Sciences Hydrologiques
41(3), 399–417.
Mitchell, T. M.: 1997, Machine learning, McGraw-Hill Higher Education.

Bibliography 131
Montuori, C.: 1965, Introduction d’un débit constant dans un canal vide, Proc. of
the 11th IAHR Congress, Leningrad, Russia, pp. 1–7.
Morris, M. W.: 2005, IMPACT: Final technical report, Proc. of the 1st IMPACT
Workshop, HR Wallingford, UK.
Mynett, A. E.: 2002, Hydroinformatics in aquatic resources management, Proc.
of the NATO Advanced Research Workshop on “New Paradigms in River and
Estuary Management”, Moscow, Idaho, USA.
Mynett, A. E.: 2004a, Artificial Intelligence techniques in environmental hydroin-
formatics, Proc. of the IAHR-Asian and Pacific Division Conference, Hong Kong.
Mynett, A. E.: 2004b, Hydroinformatics tools for ecohydraulics modelling, Proc. of
the 6th International Conference on Hydroinformatics, Singapore.
Mynett, A. E.: 2005, Hydroinformatics applications in hydroscience and engineering,
Chapter 2, Encyclopedia of Hydrological Sciences, Willey Publishers.
Mynett, A. E. and de Vriend, H. J.: 2005, Next generation information and commu-
nication technology (ICT) for integrated water management of the Yellow River
Basin, Proc. of the 2nd Yellow River Forum, Zhengzou, China.
Nishat, A.: 2006, Experiences from Bangladesh on disaster management with focus
on floods, Meeting Report on Regional Journalist Workshop on Water Issues in
Asia, Bangkok, Thailand.
NPDP: 2007, Dam incidents database, National Performance of Dams Program,
Stanford University, California, USA. Available at npdp.stanford.edu.
Nsom, B.: 2002, Horizontal viscous dam–break flow: Experiments and theory, Jour-
nal of Hydraulic Engineering 128(5), 543–546.
Olı́as, M., Cerón, J., Fernández, I., Moral, F. and Rodrı́guez-Ramı́rez, A.: 2005,
State of contamination of the waters in the Guadiamar Valley five years after the
Aznalcóllar spill, Water, Air, & Soil Pollution 166(1), 103–119.
Plate, E. J.: 2000, Flood management as part of sustainable development, Vol. 1,
Proc. of the International Symposium on River Flood Defence, Kassel, Germany.
Ponce, V. M. and Tsivoglou, A. J.: 1981, Modeling gradual dam breaches, Journal
of Hydraulic Division 107(7), 829–838.
Powledge, G. R., Ralston, D. C., Miller, P., Chen, Y. H., Clopper, P. E. and Temple,
D. M.: 1989, Mechanics of overflow erosion on embankments. Part II. Hydraulic
and design considerations, Journal of Hydraulic Engineering 115(8), 1056–1075.
Price, R. K.: 2000, Hydroinfomatics for river flood management, Vol. 71, NATO
Science Series 2 Environmental Security, Kluwer Academic Publishers, pp. 237–
250.
132 Bibliography
Pugh, C. A. and Harris, D. W.: 1982, Prediction of landslide generated water waves,
Vol. Q. 54, Proc. of the 14th International Congress on Large Dams (ICOLD),
Rio de Janerio, Brazil, pp. 1056–1075.
Qing, D.: 1997, The River Dragon has come!: Three Gorges dam and the fate of
China’s Yangtze River and its people, ME Sharpe.
Ralston, D.: 1987, Mechanics of embankment erosion during overflow, Proc. of the
ASCE National Conference in Hydraulic Engineering, pp. 733–738.
Ré, R.: 1946, Etude du lacher instantané d’une retenue d’eau dans un canal par
la méthode graphique (Study of the Sudden Water Release from a Reservoir in a
Channel by a Graphical Method.), Jl La Houille Blanche (in French) 1(3), 181–
187.
Ritter, A.: 1892, Die Fortpflanzung der Wasserwellen (Propagation of waves),
Zeitschrift des Vereines Deutscher Ingenieure 36(33), 947–954.
Robinson, K. M. and Hanson, G. J.: 1994, A deterministic headcut advance model,
Transactions of the ASAE 37(5), 1437–1443.
Rogers, J. D.: 2007, Background on failure of Teton Dam near Rexburg,
Idaho on June 5, 1976, University of Missouri-Rolla, USA. Retrieved from
web.umr.edu/˜rogersda/teton dam.
Rosa, L. P. and dos Santos, M. A.: 2000, Certainty & uncertainty in the science of
greenhouse gas emissions from hydroelectric reservoirs, Report prepared for the
World Commission on Dams.
Rosenblatt, F.: 1958, The perceptron: A probabilistic model for information storage
and organization in the brain., Psychological Review 65(6), 386–408.
Rousseeuw, P. J.: 1984, Least median of squares regression, Journal of the American
Statistical Association 79(388), 871–880.
Rousseeuw, P. J. and Leroy, A. M.: 2003, Robust regression and outlier detection,
Wiley-Interscience.
Rowe, W. D.: 1994, Understanding uncertainty, Risk Analysis 14(5), 743–750.
Rozov, A. L.: 2003, Modeling of washout of dams, Journal of Hydraulic Research
41(6), 565–577.
Saltelli, A.: 2004, Global Sensitivity Analysis: An Introduction, pp. 27–43.
Saltelli, A., Chan, K. and Scott, E. M.: 2000, Sensitivity analysis, Wiley.
Samuels, P., Klijn, F. and Dijkman, J.: 2005, River flood risk management policies
in different parts of the world: Synthesis of special session by NCR, Proc. of the
3rd International Symposium on Flood Defence, Nijmegen, the Netherlands.
Bibliography 133
Schoklitsch, A.: 1917, Über Dambruchwellen, Sitzungberichten der Königliche

Akademie der Wissenschaften, Vienna, Austria 126, 1489–1514.
Schoklitsch, A.: 1926, Geschiebebewegung in Flüssen und an Stauwerken, J. Springer.
Schoklitsch, A.: 1934, Der Geschiebetrieb und die Geschiebefracht, Wasserkraft und
Wasserwirtschaft, 29, Jahrgang.
Shortreed, J., Dinnie, K. and Belgue, D.: 1995, Risk criteria for public policy,
Proc. of the 1st Biennial Conference on “Process Safety and Loss Management in
Canada”, Institute for Risk Research, University of Waterloo, Waterloo, Ontario,
Canada, pp. 131–158.
Simonovic, S. P. and Ahmad, S.: 2005, Computer-based model for flood evacuation
emergency planning, Natural Hazards 34(1), 25–51.
Singh, K. P. and Snorrason, A.: 1982, Sensitivity of outflow peaks and flood stages to
the selection of dam breach parameters and simulation models, Technical Report
289, State Water Survey Division at the University of Illinois, USA.
Singh, K. P. and Snorrason, A.: 1984, Sensitivity of outflow peaks and flood stages
to the selection of dam breach parameters and simulation models, Journal of
Hydrology 68, 295–310.
Singh, V. P.: 1996, Dam breach modeling technology, Kluwer Academic Pub.
Singh, V. P. and Quiroga, C. A.: 1987, A dam-breach erosion model: I. Formulation,
Water Resources Management 1(3), 177–197.
Smart, G. M.: 1984, Sediment transport formula for steep channels, Journal of
Hydraulic Engineering 110(3), 267–276.
Soares Frazão, S.: 2002, Dam–break induced flows in complex topographies: Theoret-
ical, numerical and experimental approaches, PhD thesis, Université Catholique
de Louvain, Belgium.
Solomatine, D. P.: 2002, Applications of data-driven modelling and machine learning
in control of water resources, Computational Intelligence in Control, Idea Group
Publishing, pp. 197–217.
Soumis, N., Lucotte, M., Duchemin, a., Canuel, R., Weissenberger, S., Houel, S.
and Larose, C.: 2005, Hydroelectric reservoirs as an anthropogenic sources of
greenhouse gases, John Wiley & Sons, Ohio.
Spinewine, B. and Zech, Y.: 2007, Small-scale laboratory dam-break waves on mov-
able beds, Journal of Hydraulic Research 45(Extra Issue), 73–86.
Stelling, G. S. and Duinmeijer, S. P. A.: 2003, A staggered conservative scheme for
every Froude number in rapidly varied shallow water flows, International Journal
for Numerical Methods in Fluids 43(12), 1329–1354.
134 Bibliography
Stelling, G. S., Kernkamp, H. W. J. and Laguzzi, M. M.: 1998, Delft flooding system:
a powerful tool for inundation assessment based upon a positive flow simulation,
Proc. of 5th International Conference on Hydroinformatics, pp. 449–456.
Stigler, S. M.: 1981, Gauss and the invention of least squares, The Annals of Statis-
tics 9(3), 465–474.
Stoker, J. J.: 1957, Water waves: The mathematical theory with applications, Pure
and Applied Mathematics 4.
TAW: 1998, Fundamentals on water defences, Technical report, Technical Advisory

Committee on Water Defences, the Netherlands.
TAW: 2002, Wave run-up and wave overtopping at dikes, Technical report, Technical
Advisory Committee on Water Defences, the Netherlands.
Temple, D. M. and Hanson, G. J.: 1994, Headcut development in vegetated earth

spillways, Applied Engineering in Agriculture 10(5), 677–682.
Temple, D. and Moore, J.: 1997, Headcut advance prediction for earth spillways,
Transactions of the ASAE 40(3), 557–562.
Thoft–Christensen, P. and Baker, M. J.: 1982, Structural reliability theory and its
applications, Springer.
Tingsanchali, T. and Hoai, H. C.: 1993, Numerical modelling of dam surface ero-
sion due to flow overtopping, Vol. 1, Advances in HydroScience and Engineering,
Washington, USA, pp. 883–890.
Tinney, R. E. and Hsu, H. Y.: 1961, Mechanics of washout of an erodible fuse plug,
Journal of the Hydraulics Division 87(3), 1–30.
Toro, E. F.: 1999, Rieman solvers and numerical methods for fluid dynamics: A
practical introduction, Springer-Verlag, Heidelberg, Berlin.
Tremblay, A., Lambert, M. and Gagnon, L.: 2004, Do hydroelectric reservoirs emit
greenhouse gases?, Environmental Management 33, 509–517.
Turing, A. M.: 1950, Computing machinery and intelligence, Mind 59(236), 433–460.
U.S. Army Corps of Engineers: 1960, Floods resulting from suddenly breached dams:
Conditions of minimum resistance, Report 1(374), U.S. Army Engineers Water-
ways Experiment Station, Vicksburg, Mississippi.
U.S. Army Corps of Engineers: 1961, Floods resulting from suddenly breached dams:
Conditions of high resistance, Report 2(374), U.S. Army Engineers Waterways
Experiment Station, Vicksburg, Mississippi.
Bibliography 135
U.S. Bureau of Reclamation: 1988, Downstream hazard classification guidelines,

ACER Technical Memorandum No. 11, Assistant Commissioner-Engineering and
Research, U.S. Department of the Interior, Denver, Colorado.
van de Ven, G. P.: 2004, Man-made lowlands: History of water management and
land reclamation in the Netherlands, Matrijs, Utrecht, the Netherlands.
van der Sande, C. J., de Jong, S. M. and de Roo, A. P. J.: 2003, A segmentation
and classification approach of IKONOS-2 imagery for land cover mapping to assist
flood risk and flood damage assessment, International Journal of Applied Earth
Observation and Geoinformation 4(3), 217–229.
van der Sluijs, J. P.: 1997, Anchoring amid uncertainty: On the management of
uncertainties in risk assessment of anthropogenic climate change, PhD thesis,
Utrecht University, the Netherlands.
van Rijn, L. C.: 1993, Principles of sediment transport in rivers, estuaries and
coastal seas, Aqua Publications.
van Schijndel, S.: 2005, The Planning Kit, a decision making tool for the Rhine
branches, Proc. of the 3rd International Symposium on Flood Defence, Nijmegen,
the Netherlands.
Vaskinn, K. A.: 2003, Stability and failure mechanisms of dams, Proc. of the 3rd
IMPACT Workshop, Université Catholique de Louvain, Louvain-la-Neuve, Bel-
gium.
VenW: 2005, Floris study-Full report, Flood Risks and Safety in the Netherlands
(Floris) Project, Dutch Ministry of Transport, Public Works and Water Manage-
ment.
Verheij, H. J.: 2002, Modification breach growth model in HIS–OM, Q3299, Tech-
nical report, WL | Delft Hydraulics (in Dutch).
Verwey, A.: 2001, Latest developments in floodplain modelling–1D/2D integration,

Proc. of 6th Conference on Civil Engineering Hydraulics, Hobart, Australia.
Vesely, W. E. and Rasmuson, D. M.: 1984, Uncertainties in nuclear probabilistic

risk analyses., Risk Analysis 4(4), 313–322.
Vis, M., Klijn, F., de Bruijn, K. M. and van Buuren, M.: 2003, Resilience strategies
for flood risk management in the Netherlands, International Journal of River
Basin Management 1, 33–40.
Visser, P. J.: 1998, Breach growth in sand–dikes, PhD thesis, Delft University of
Technology.
136 Bibliography
Visser, P. J., Smit, M. J. and Snip, D. W.: 1996, Zwin ’94 experiment: Meetop-
stelling en overzicht meetresultaten, Report 4–96, Delft University of Technology,
the Netherlands.
Vrijling, J. K.: 2001, Probabilistic design of water defense systems in the Nether-
lands, Reliability Engineering and System Safety 74(3), 337–344.
Vrouwenvelder, A. C. W. M.: 1999, Theoretical manual of PC-Ring, Part C: Calcu-

lation methods (in Dutch), TNO-report 98-CON-R1204, Delft, the Netherlands.
Vrouwenvelder, A. C. W. M., Steenbergen, H. M. G. M. and Slijkhuis, K. A. H.:

2001, Theoretical manual of PC-Ring, Part B: Statistical models (in Dutch), TNO-
report 98-CON-R1431, Delft, the Netherlands.
Wahl, T. L.: 1998, Prediction of embankment dam breach parameters: A litera-

ture review and needs assessment, U.S. Department of the Interior, Bureau of
Reclamation, Dam Safety Office.
Wahl, T. L.: 2004, Uncertainty of predictions of embankment dam breach parame-

ters, Journal of Hydraulic Engineering 130.
Wentz, F. J., Ricciardulli, L., Hilburn, K. and Mears, C.: 2007, How much more
rain will global warming bring?, Science 317(5835), 233–235.
Witten, I. H. and Frank, E.: 2000, Data mining: Practical machine learning tools
and techniques with Java implementations, Morgan Kaufmann.
Wolters, H. A., Platteeuw, M. and Schoor, M. M.: 2001, Guidelines for rehabilitation
and management of floodplains, NCR publication.
World Commission on Dams: 2000, Dams and development: A new framework for
decision-making, Earthscan.
Wu, W., Wang, S. S. Y., Jia, Y. F. and Robinson, K. M.: 1999, Numerical simulation
of two-dimensional headcut migration, ASCE, Proc. of the International Water
Resources Engineering Conference, Seattle, WA, USA.
Yalin, M. S.: 1972, Mechanics of sediment transport, Pergamon Press.
Zagonjolli, M., Goossens, H. and Mynett, A. E.: 2006, A numerical-constraint based

approach for flood mitigation in dike ring areas, Proc. of 7th International Con-
ference on Hydroinformatics, Nice, FRANCE.
Zagonjolli, M. and Mynett, A.: 2006a, Dam breaching uncertainty and its effect
in downstream areas, Proc. of the 7th International Conference on Hydroscience
and Engineering, Philadelphia, USA, Michael Piasecki and College of Engineering,
Drexel University.
Bibliography 137
Zagonjolli, M. and Mynett, A. E.: 2005a, Dam breach analysis: A comparison

between physical, empirical and data mining models, Proc. of the 29th IAHR
Congress Seoul, South Korea, p. 753.
Zagonjolli, M. and Mynett, A. E.: 2005b, Data mining techniques in dam breach
modelling, Submitted to the IAHR Journal of Hydraulic Research .
Zagonjolli, M. and Mynett, A. E.: 2006b, Failure analysis of a high earthfill dam:
Scenario development using numerical simulation and GIS, Proc. of 7th Interna-
tional Conference on Hydroinformatics, Nice, FRANCE.
Zagonjolli, M., Mynett, A. E. and Verwey, A.: 2005, Dam break modelling analysis
of Bovilla Dam near Tirana, Albania, Proc. of the 3rd International Symposium
on Flood Defence, Nijmegen, the Netherlands.
Zhu, Y., Visser, P. J. and Vrijling, J. K.: 2006, Laboratory observations of em-
bankment breaching, Proc. of the 7th International Conference on Hydroscience
and Engineering, Philadelphia, USA, Michael Piasecki and College of Engineering,
Drexel University.
List of Symbols
Roman Letters
Symbol Description SI Units
abl Thickness of the bed layer [m]

Bavg Ultimate average width of the breach channel [m]
Bbot Bottom width of the breach channel [m]
Bt Top width of the breach channel [m]
C Weir discharge coefficient [-]
Cs Submergence coefficient on the weir [-]
d Soil grain diameter [mm]
D Height of the breach channel [m]
DF Dam factor [m4 ]
d50 , d65 Representative grain size diameter [mm]
Fr Froude number [-]
g Gravitational acceleration [m2 /s]
h Flow depth [m]
H Water depth above the breach bottom [m]
hb Vertical extent of the breach channel [m]
hd Dam height or dike height [m]
hw Depth of water in the reservoir triggering the failure [m]
k and a Correlation coefficients [-]
I1 and I2 Integral quantities depending upon the flow depth [-]
L Length of the breach channel [m]
Ms Mass of soil eroded during breaching [kg/s]
q Discharge per unit width [m2 /s]
Qp Peak discharge [m3 /s]
td Duration of failure [hr]
u Water velocity [m/s]
Vout Volume of water outflow from the reservoir [m3 ]
Vw Volume of water in the reservoir triggering the failure [m3 ]
Y Elevation of breach bottom [m]
z Breach side slope factor [-]
Greek letters
Symbol Description SI Units
β, δ Angle that the breach side creates with vertical [◦ ]

γ Unit weight of water [N/m3 ]
140 List of Symbols
τ Tractive stress of the flowing water [N/m2 ]

τc Critical tractive stress for erodible material [N/m2 ]
(τ0 − τc ) Excess shear stress [N/m2 ]
Mathematical symbols and operators
Symbol Description
≡ Equivalent to (or defined to be)

argmaxf (x) Value of x that leads to the maximum value of f(x)
x
List of Abbreviations
AI Artificial Intelligence
ANNs Artificial Neural Networks
BREADA BReaching of the EArthfill DAm
Floris FLOod RIsks and Safety
GHGs GreenHouse Gases
GIS Geographical Information System
IBL Instance Based Learning
IPCC Intergovernmental Panel on Climate Change
LMS Least Median of Squares
LS Least Squares
LWL Locally Weighted Learning
MLP Multi Layer Perceptron
NPDP National Performance of Dams Program in USA
PAR People At Risk
RL Reinforcement Learning
RMSE Root Mean Square Error
SMDBRK National Weather Service Simplified Dam Break Model
TAW Technical Advisory Committee on Water Defences in the Netherlands
USBR U.S. Bureau of Reclamation
WCD World Commission on Dams
WEKA Waikato Environment for Knowledge Analysis
Curriculum Vitae
Migena Zagonjolli was born in 1975 in Tirana, Albania. She studied hydraulic en-
gineering at the Polytechnic University of Tirana, Albania, where she graduated in
1999. From then until 2001 she worked at the Department of Hydraulic Engineering
as an assistant lecturer in Hydraulics and Water Resources Management. During
this period she visited the University of Limerick, Republic of Ireland, as a guest
researcher.
From 2001 to 2003 she studied at UNESCO–IHE in Delft, the Netherlands, where
she obtained her MSc degree in Hydroinformatics on a research topic related to dam
break modelling. In 2004 she started her PhD research within the collaboration
framework between WL | Delft Hydraulics, UNESCO–IHE and Delft University of
Technology. She addressed a range of research topics, including data mining and
analysis, computational modelling of dam and dike breaching, sediment transport,
and flood risk assessment. One of the novel aspects of her research is the design of
a numerical–constraint based model for flood propagation and risk mitigation.
Her experience includes software development related to dam/dike breaching, flood

wave propagation, constraint based modelling, uncertainty analysis and flood mit-
igation. She has published several papers in international journals and conference
proceedings. Starting 1st October 2007 she will be employed as a researcher/advisor
at the Department of Inland Water System at WL | Delft Hydraulics.

PHD Zagonjolli PDF

Uploaded by

Copyright:

Available Formats

PHD Zagonjolli PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PHD Zagonjolli PDF

Uploaded by

Copyright:

Available Formats

Dam Break Modelling, Risk

Assessment and Uncertainty

Submitted in fulfillment of the requirements of

born in Tirana, Albania

Members of the Awarding Committee:

Chairman Rector Magnificus TU Delft, the Netherlands

Copyright © 2007 Taylor & Francis Group plc, London, UK

ISBN 9 780415 455947 (Taylor & Francis Group)

In this research, instead of focusing on methodologies to minimise the failure prob-

De verwachting is dat de gevolgen van klimaatverandering de komende decennia

braak te voorspellen. Ondanks de beperkte hoeveelheid gedocumenteerde gegevens

I am as ever, especially grateful to my family. To my brothers for their love and

2 Water Retention and Flood Defence Structures 9

3 Current Approaches for Dealing with Flooding 25

4 Current Approaches for Breach Modelling 43

4.4 Physically based models . . . . . . . . . . . . . . . . . . . . . . . . . 50

5 Data Mining Techniques in Dam Breach Modelling 57

6 BREADA model: Breach Model for Earthfill Dams 75

7 A Numerical-Constraint based Model 103

8 Conclusions and Recommendations 119

List of symbols 139

List of abbreviations 141

Curriculum vitae 143

Climate Change 2001:

Figure Figure 11.12: Global

lantic (South Coast of Brazil).

When the failure of a tailing dam happens, the socio–environmental implications

1.2 Management of the flood risk caused by struc-

The total failure probability of a structure involves the combination of individual

Nowadays, as a result of economic development and population growth in flood

1.3 Scope of the thesis

In this thesis, a ‘lightweight’ numerical–constraint based technique is proposed. This

The objectives of this thesis can be summarized as follows:

uncertainties (input, model, and completeness uncertainty), and proposal of

5. Development of a numerical constraint based model for flood mitigation in

1.4 Outline of the thesis

Chapter 3 describes in detail current state–of–the–art approaches used for dealing

Chapter 4 compares the theoretical and practical aspects of available approaches

Chapter 7 describes the development of a numerical–constraint based model which

Chapter 8 presents the conclusions of the research and future recommendations.

Water Retention and Flood

2.1 Design criteria and failure modes for dams

Downstream slope Downstream

dams, gravity or arch dams∗ . If breaching is initiated, the further development is

Failure of an earthfill dam can be triggered by different factors, e.g. overtopping,

reservoir capacity of 55 million m3 , is described to have failed explosively on 2 December 1959

Reservoir water surface

Figure 2.2: Seepage on the foundation of a dam.

2.2 Design criteria and failure modes for dikes

Dikes and their failure modes

Figure 2.4: Possible failure modes for dikes (TAW, 1998).

Figure 2.5: Failure of a canal dike in Wilnis, the Netherlands (2003).

 Variation of the properties of structural material (strength, deformation, per-

 Lack of a reliable mathematical description of the failure processes needed for

accurate determination of the limit states. As for overtopping and overflowing,

For the probabilistic safety analysis to be meaningful the accurate computational

2.3 Breach modelling

 Dikes in comparison to embankment dams are typically longer than high.

 The breaching of embankment dams depends mostly on reservoir volume rather

Variation of the properties of structural material (strength, deformation, per-

Lack of a reliable mathematical description of the failure processes needed for

Dikes in comparison to embankment dams are typically longer than high.

The breaching of embankment dams depends mostly on reservoir volume rather