Thesis Qian Li

This item was submitted to Loughborough's Research Repository by the author.
Items in Figshare are protected by copyright, with all rights reserved, unless otherwise indicated.
Development of new high-performance simulation approaches for multi-level

urban flood modelling
PLEASE CITE THE PUBLISHED VERSION
PUBLISHER
Loughborough University
LICENCE
CC BY-NC-ND 4.0
REPOSITORY RECORD
Li, Qian. 2022. “Development of New High-performance Simulation Approaches for Multi-level Urban Flood
Modelling”. Loughborough University. https://doi.org/10.26174/thesis.lboro.21617946.v1.
Development of New High-performance
Simulation Approaches for Multi-level Urban
Flood Modelling
by
Qian Li
A doctoral thesis submitted in partial fulfilment of the requirements
for the award of Doctor of Philosophy of Loughborough University
May 2022
© by Qian Li (2022)
Abstract
Urban flooding can threaten people’s lives, damage properties and disrupt infrastructure
networks. With the acceleration of urbanisation and climate change, urban flooding has become
one of the most widespread natural hazards causing huge human and economic losses every
year across the world. It is of great significance to understand the underlying physical processes
of urban flooding and develop effective strategies to manage the flood risk. Numerical
modelling has become an indispensable tool to support this.
Modern cities are highly heterogeneous, and are commonly developed with multi-layer spaces,
e.g. surface layer that contains buildings, streets and grasslands, etc; underground layer
containing underground malls, car parks and metro stations; drainage layer with all of the urban
drainage systems. Modelling the complex dynamic flooding process across the multi-layer
spaces at a city-scale is still beyond the capability of most existing models. Aiming to tackle
this research gap, this thesis develops a new high-performance hydrodynamic modelling
system for large-scale multi-layer urban flood simulations at high-resolution.
Drainage network model is an essential component for complete urban flood prediction and
risk assessment. Different numerical procedures have been used in developing drainage
network models to handle flows in pipes and junctions. In this thesis, a novel numerical scheme
based on the finite volume solution to the two-dimensional (2D) shallow water equations
(SWEs) is proposed to calculate the flow dynamics in junctions, which directly takes into
account both mass and momentum conservation and removes the necessity of implementing
complicated boundary settings for pipe calculations. This new junction simulation method is
then coupled with the widely used two-component pressure approach (TPA) for pipe flow
calculation, leading to a new integrated drainage network model. The developed 1D-2D
coupled drainage network model is validated against an experimental and several idealised test
cases to demonstrate its potential for efficient and stable simulation of flow dynamics in
drainage networks.
The new drainage model is then coupled with a well-established High-Performance Integrated
hydrodynamic Modelling System (HiPIMS) for surface water flood modelling to support dual
drainage urban flood modelling. The dual-drainage model is validated by an experiment test to
i
confirm its predictive capability. The performance of the developed model is statistically
assessed by Nash–Sutcliffe model efficiency coefficient (NSE) and R-squared (R2). Most of
the NSE values are greater than 0.6, and all the R 2 values are higher than 0.8, indicating that
the simulation results have a good agreement with experimental measurements.
To also simulate the flooding process in the underground spaces, the 1D-2D coupled dual
drainage model is extended to deliver an innovative modelling system for simultaneous
simulation of flooding process in all three urban layers, i.e. surface layer, underground space
layer and drainage system. The underground space simulation module is implemented in a
flexible framework to allow the simulation of flood flows in multi-underground spaces of
multiple levels to better reflect the urban design reality.
To overcome the high computational burden of hydrodynamic models for large-scale

applications, high-performance computing techniques are adopted in this thesis. The developed
dual-drainage model is first accelerated by implementing a parallel computing scheme on a
single GPU device. Furthermore, compared with a simulation on a single GPU, multi-GPU
simulations can overcome the limitation of insufficient physical memory on a single GPU and
further improve the model performance for large-scale flood modelling across an entire city or
a large domain. Therefore, the proposed model is further developed to support multi-GPU
simulations. The multi-GPU model is finally applied to reproduce a flood event caused by
Typhoon Morakot in 2009 in Yuhuan county, China. Running on 4 modern GPUs, the flood
simulation is 4.32 times faster than real time, demonstrating the potential of the multi-GPU
model for wider application in large-scale urban flood simulation and risk assessment.
ii
Acknowledgements
I would like to express my sincere and deep gratitude to my supervisor, Prof. Qiuhua Liang,
for his patient guidance and tremendous support throughout my doctoral study. It is my great
honour to be his student and conduct research under his supervision. I am deeply grateful to
him for guiding me into the research area of urban flood modelling and teaching me how to be
an independent researcher. I could never finish my thesis without his inexhaustible help,
especially the help in the last two years of my PhD study during which I was pregnant and
suffered lots of complications. I would also express my warmest gratitude to my supervisor,
Dr. Xilin Xia. I benefit a lot from his profound knowledge and rigorous attitude. I would also
like to thank Dr. Wen Xiao who supervised me in my first year at Newcastle University with
great enthusiasm.
I would also like to thank my colleagues and researchers in Hydro-Environmental Modelling

Laboratory for Natural Hazard Risk & Resilience including Dr. Xiaodong Ming, Dr. Huili
Chen, Dr. Ling Jiang, Dr. Jiaheng Zhao, Dr. Zhenlei Wei, Dr. Xiaoli Su, Dr. Jinghua Jiang, Dr.
Yunsong Cui, Miss. Kaicui Chen, Miss. Xue Tong, Dr. Weihua Zhao, Dr. Yan Xiong, Mr.
Wenyuan Dong, Mr. Haoyang Qin. This is a very active research group, and we have a very
stimulating and enjoyable environment for learning. We can discuss and share our research
freely. Thanks also go to my friends in Newcastle and Loughborough including Mr. Don
Gunton, Mr. David Parker, Miss Jiajun Chen, Dr. Xinxin Liang, Dr. Jingchun Wang, Miss.
Jinyao Yang, Dr. Lanyue Zhang, Dr. Jingyi Zong, Mr. Hao Wang for the help and happiness
they bring to me. I would also like to thank my dear friends from my high school and university
who always encourage me especially during the hardest time of my study. They are Mr. Zhe Ji,
Mr. Yang Xia, Mr. Guangyi Wu, Miss Jiannan Wu, Miss Ting Ouyang, Mr. Haisheng Yu , Dr.
Rong Tang, Dr. Guozhen Wei, Dr. Yongqin Liang.
I am grateful to the staff in School of Architecture, Building and Civil Engineering. Special
thanks go to Mr. Berkeley Young for his kind support in my study at Loughborough University.
My research is funded by Newcastle University, Loughborough University and China

Scholarship Council. Their financial supports are greatly acknowledged. I would also like to
thank Hardship Fund from Loughborough University for financially supporting me in the last
year of my study.
iii
Finally, I would like to express my deep gratitude to my Mother Miss. Shizhen Zhang who
came to UK to help me look after my little girl so that I can focus on writing my thesis, my
father Mr. Benjian Li, my elder brother Mr. Gen Li and my whole family for their great love
and unfailing support in my life. Special thanks go to my baby girl Miss Alissa Yanxi Zhu who
was born last year. She is a treasure to me, and I am so lucky to be her mom. My sincere thanks
also go to my husband Dr. Zuo Zhu for his endless love and accompany since the day he stood
by my side many years ago. Without his encouragement and help, I could never finish my PhD
study.
iv
Table of Contents
Abstract i
Acknowledgements .................................................................................................................. iii
Table of Contents ....................................................................................................................... v
List of Figures ........................................................................................................................ viii
List of Tables ............................................................................................................................ xi
Chapter 1 Introduction .......................................................................................................... 1
1.1 Background ................................................................................................................. 1
1.2 Research aim and objectives ....................................................................................... 7
1.3 Thesis outline .............................................................................................................. 9
Chapter 2 Literature review ................................................................................................ 11
2.1 Surface flow models .................................................................................................. 12
2.2 Drainage models ........................................................................................................ 18
2.3 Dual drainage models ................................................................................................ 20
2.4 Simulation of floods in underground space............................................................... 26
2.5 Acceleration methods in urban flood modelling ....................................................... 29
2.6 Summary ................................................................................................................... 32
Chapter 3 A novel 1D-2D coupled model for hydrodynamic simulation of flows in drainage
networks 34
3.1 Introduction ............................................................................................................... 34
3.2 New 1D-2D coupled drainage network model .......................................................... 34
3.2.1 Pipe model ......................................................................................................... 34
3.2.2 Junction model ................................................................................................... 41
3.3 Results and discussion............................................................................................... 46
3.3.1 Experimental test ............................................................................................... 46
3.3.2 Unsteady flow through different drainage settings ............................................ 50
v
3.3.3 Unsteady flow in V-shape networks .................................................................. 55
3.3.4 A hypothetical drainage network system ........................................................... 57
3.3.5 Discussion .......................................................................................................... 59
3.4 Summary ................................................................................................................... 60
Chapter 4 A single-GPU based 1D-2D dual-drainage model for urban flood simulations 62
4.1 Introduction ............................................................................................................... 62
4.2 Dual drainage model ................................................................................................. 62
4.2.1 HiPIMS .............................................................................................................. 63
4.2.2 Drainage model .................................................................................................. 71
4.2.3 Coupling HiPIMS with the drainage model....................................................... 71
4.3 Implementation on single GPU ................................................................................. 73
4.3.1 Structure of the drainage model on a single-GPU device .................................. 74
4.3.2 Structure of HiPIMs-drainage coupled model on a single-GPU device ............ 75
4.4 Model validation and application .............................................................................. 76
4.4.1 Experiment Test ................................................................................................. 77
4.4.2 A semi-realistic test............................................................................................ 86
4.5 Summary ................................................................................................................... 91
Chapter 5 Multi-layer simulation of flooding into underground spaces ............................ 93
5.1 Introduction ............................................................................................................... 93
5.2 Multi-layer flood modelling framework ................................................................... 93
5.3 Validation of the multi-layer flood model................................................................. 99
5.3.1 Flood in an idealised urban catchment with underground space ....................... 99
5.4 Summary ................................................................................................................. 106
Chapter 6 Multi-GPU implementation and case study ..................................................... 108
6.1 Introduction ............................................................................................................. 108
6.2 Multi-GPU Architecture .......................................................................................... 108
6.3 Comparison against the single-GPU based model .................................................. 112
vi
6.4 Case study (Yuhuan) ............................................................................................... 116
6.4.1 Study site.......................................................................................................... 116
6.4.2 Data and model setup ....................................................................................... 116
6.5 Results and discussion............................................................................................. 121
6.5.1 Model validation .............................................................................................. 121
6.5.2 Comparison against more traditional methods ................................................ 128
6.5.3 Computational efficiency ................................................................................. 130
6.6 Summary ................................................................................................................. 131
Chapter 7 Conclusions and future work ........................................................................... 134
7.1 conclusions .............................................................................................................. 134
7.1.1 Model description of the proposed urban flood modelling system.................. 134
7.1.2 Acceleration techniques of the proposed urban flood modelling system ........ 136
7.2 Future work ............................................................................................................. 137
7.2.1 Couple a water quality module in the urban flood modelling system ............. 137
7.2.2 Consider more interaction processes between different model components in

multi-level framework ................................................................................................... 138
7.2.3 Develop different GPU assigning methods in the multi-GPU based model .... 138
7.2.4 Flood forecasting based on the proposed urban flood modelling system ........ 139
7.2.5 Apply the modelling system to assess the performance of blue-green

infrastructures ................................................................................................................ 139
7.2.6 Apply the modelling system in the multi-objective optimisation of drainage

network rehabilitation measures .................................................................................... 140
List of Publications ................................................................................................................ 141
References ………………………………………………………………………………..142
vii
List of Figures
Figure 1-1 Population change and urbanisation: (a) rural and urban population worldwide; (b)
Percentage of urban population in China. .................................................................................. 3
Figure 2-1 Rainfall-runoff process in urban areas (Bulti and Abebe, 2020) ........................... 11
Figure 3-1 Spatial discretization scheme for a junction........................................................... 42
Figure 3-2 Experimental pressurised flow test: experimental apparatus and set up. ............... 47
Figure 3-3 Flow variables at x = 9.9 m predicted by the new drainage model, TPA-SB model
and TPA-VWR model, in comparison with the VWR experimental measurements: (a) velocity;
(b) pressure............................................................................................................................... 47
Figure 3-4 Flow variables at x = 9.9 m predicted at different grid resolutions: (a) velocity; (b)
pressure. ................................................................................................................................... 48
Figure 3-5 Flow variables at x = 9.9 m predicted using different wave speeds: (a) velocity; (b)
pressure. ................................................................................................................................... 48
Figure 3-6 Flow variables at x = 9.9 m predicted using different Manning coefficients: (a)
velocity; (b) pressure................................................................................................................ 48
Figure 3-7 Flow variables at x = 9.9 m predicted with/without a gate: (a) velocity; (b) pressure.
.................................................................................................................................................. 49
Figure 3-8 Configuration of the idealised drainage system. .................................................... 51
Figure 3-9 The external flow rate imposed at Junction 1. ....................................................... 51
Figure 3-10 Predicted water depths and flow rates at the outfall for Case 1: (a) water depths
under free surface flow condition; (b) flow rates under free surface flow condition; (c) water
depths under pressurised flow condition; (d) flow rates under pressurised flow condition; ‘hP’
denotes the initial water depth in Junction 1 and ‘d’ refers to the pipe diameter. ................... 52
Figure 3-11Predicted water depths and velocities in the two junctions for Case 2: (a) water
depth at Junction 1; (b) flow velocity at Junction 1; (c) water depth at Junction 2; (d) flow
velocity at Junction 2. .............................................................................................................. 53
Figure 3-12 Predicted water depths and velocities in the two junctions for Case 3: (a) water
velocity at Junction 2; ‘PipeL’ denotes pipe length.................................................................. 54
Figure 3-13 V-shape networks of three different connecting angles: 180⁰, 120⁰ and 30⁰. ...... 55
Figure 3-14 Predicted water depths and flow rates at the outfalls of the three V-shape junction-
pipe systems: (a) water depths under free surface flow condition; (b) flow rates under free
viii
surface flow condition; (c) water depths under pressurized flow condition; (d) flow rates under
pressurized flow condition. ...................................................................................................... 56
Figure 3-15 The configuration of a hypothetical drainage network. ....................................... 58
Figure 3-16 External inflow hydrograph imposing at Junction 1. ........................................... 58
Figure 3-17 Predicted water depths and flow rates at Outfall 1 and Outfall 2: (a) water depth;
(b) flow rate.............................................................................................................................. 58
Figure 4-1 Flowchart of the drainage model on a single GPU ................................................ 74
Figure 4-2 Parallel computing process of pipe and junction cells on GPU ............................. 75
Figure 4-3 Flowchart of the HiPIMs-drainage coupled model on GPU .................................. 76
Figure 4-4 Geometry of the experimental setup (unit: m) ....................................................... 78
Figure 4-5 Topography of the street surface (the unit of color bar is in meter) ...................... 79
Figure 4-6 The comparison of discharges in Pipes 1, 2, 3 and 5 in R1 ................................... 81
Figure 4-9 The comparison of discharges in Pipes 2 and 5 in the RR1 ................................... 82
Figure 4-10 The comparison of discharges in Pipes 1, 2 and 5 in the RR2 ............................. 83
Figure 4-11 Time history of water depth in Pipe 4 under free surface condition .................... 85
Figure 4-12 Water depth evolution in Pipe 5 under pressure condition .................................. 86
Figure 4-13 DEM of Yuhuan ................................................................................................... 87
Figure 4-14 Comparison of the inundation maps at different output times predicted by the
coupled model and HiPIMS ..................................................................................................... 88
Figure 4-15 Exchange flow at Junction A, B, C and D ........................................................... 90
Figure 4-16 Surface inundation around zoomed in junctions ................................................. 91
Figure 5-1 Concept of the multi-layer flood model ................................................................. 95
Figure 5-2 The procedure for implementing the multi-layer flood model............................... 96
Figure 5-3 Set-up of the underground building basement test: (a) layout of simulated domain
including surface layer and underground space layer; (b) Layout of the -1 and -2 level under
Building 1............................................................................................................................... 100
Figure 5-4 Rainfall input ........................................................................................................ 101
Figure 5-5 Surface inundation variation ................................................................................ 102
Figure 5-6 Underground space inundation ............................................................................ 103
Figure 5-7 Comparison of average water depth of each room of Building 1 in Scenario (a) and
the hypothetical scenario (S(a) represents Scenario (a) and HS is the abbreviation of
hypothetical scenario) ............................................................................................................ 104
ix
Figure 5-8 Flood inundation of -1 underground level in Scenario (a) and (b), inundation of -2
underground level in Scenario (b).......................................................................................... 105
Figure 6-1 Surface domain and drainage network decompositions for multi-GPU computing
................................................................................................................................................ 110
Figure 6-2 The Multi-GPU architecture ................................................................................ 111
Figure 6-3 Layout of the drainage network and domain division over the drainage system . 113
Figure 6-4 Water depth and flow rate in Pipes 1-5 produced by the single-GPU model and
multi-GPU based model under the free-surface condition .................................................... 114
Figure 6-5 Water depth and flow rate in Pipes 1-5 simulated by the single-GPU model and
multi-GPU based model under pressurised condition ........................................................... 115
Figure 6-6 DEM of the research domain ............................................................................... 117
Figure 6-7 The zoom-in DEM with buildings ....................................................................... 117
Figure 6-8 Land-use type in the simulation domain .............................................................. 118
Figure 6-9 Rainfall intensity and accumulated rainfall of Typhoon Morakot ....................... 119
Figure 6-10 Drainage layout in the research domain ............................................................. 120
Figure 6-11 Locations of the river gauges in the simulation domain .................................... 120
Figure 6-12 Simulated inundation maps: (a) t = 6 hrs; (b) t = 12 hrs; (c) t = 18 hrs; (d) t = 24
hrs; (e) t = 30 hrs; (f) t = 39 hrs ............................................................................................. 122
Figure 6-13 Maximum inundation map ................................................................................. 123
Figure 6-14 Simulated inundation maps in Zone 3: (a) t = 6 hrs; (b) t = 12 hrs; (c) t = 18 hrs;
(d) t = 24 hrs; (e) t = 30 hrs; (f) t = 39 hrs ............................................................................. 125
Figure 6-15 The locations of the outfall gauges .................................................................... 126
Figure 6-16 Water depth and discharge profiles at the outfall gauges .................................. 126
Figure 6-17 Water depth at three river gauges....................................................................... 128
Figure 6-18 Maximum inundation maps (left) and zoomed in maps in Zone 3 and 4 (right): (a)
and (b) proposed model; (c) and (d) model with road sewer coefficient; (e) and (f) model
without considering drainage compacity ............................................................................... 129
Figure 6-19 Domain division in (a) simulation with two GPUs; (b) simulation with three GPUs;
(c) simulation with four GPUs ............................................................................................... 131
x
List of Tables
Table 1-1 The number of cities suffering from flooding in China from 2006 to 2016.............. 2
Table 1-2 Historical flood incidents in underground metro systems ......................................... 5
Table 2-1 Comparison of different types of models in urban flood simulations ..................... 25
Table 3-1 The pipe-junction boundary calculation method ..................................................... 39
Table 3-2 Pipe length (m) ........................................................................................................ 59
Table 3-3 Junction radius (m) .................................................................................................. 59
Table 4-1 Properties of the pipes in the drainage network ...................................................... 79
Table 4-2 Elevations of the junctions in the drainage network ............................................... 79
Table 4-3 Five simulations with different rainfall intensity and runoff inflow ....................... 80
Table 4-4 NSE and R2 calculation under different situations .................................................. 83
Table 6-1 The statistics of the Yuhuan flood ......................................................................... 127
Table 6-2 Maximum water elevation at gauge points ............................................................ 128
Table 6-3 Runtime required by different simulations ............................................................ 131
xi
Chapter 1 Introduction
1.1 Background
Flooding is one of the most widespread natural hazards threatening people’s lives and
assets around the world. According to the Organization for Economic Cooperation and
Development (https://www.nationalgeographic.com/environment/article/floods),
floods cause more than $40 billion of economic loss globally each year. In July 2021
more than 920 people lost their lives due to floods and rain-related hazards around the
world (https://floodlist.com/asia/world-floods-july-2021). Urban areas have been hit
particularly hard by flooding in recent years. For example, on 28th June 2012, a flood
event occurred in Newcastle upon Tyne, UK, affecting 1200 properties (including 500
of them flooded internally) and causing £9.2m worth of damage to the road networks
(Newcastle City Council, 2016). In July 2012, a devastating urban flood hit Beijing,
during which 56,933 people were dispatched, 79 people were killed, and 8200 houses
were destroyed. The direct economic loss was estimated to be more than 10 billion
RMB (https://en.wikipedia.org/wiki/July_2012_Beijing_flood). Table 1-1 shows the
number of cities being flooded at least one time in China from 2006 to 2016,
demonstrating that more than a hundred Chinese cities suffer from different extents of
flooding problems each year (Yan et al., 2020). The more frequent occurrence of urban
floods each year necessitates the development of modelling tools to understand and
predict the whole flood evolution process.
Storm water resulting from rainfall in an urban area is often drained away by drainage
systems and stored in man-made lakes or other waterbodies, or infiltrates into surface
soil. Urban flooding occurs when excessive water flows into a city area more rapidly
than it is discharged. Urban flooding can be classified into four categories: (1) pluvial
flooding, which is referred to the accumulation of the rainfall runoff as a result of the
insufficient drainage capacity to discharge the runoff instantly; (2) river flooding (i.e.,
fluvial flooding), which occurs when the river system cannot promptly discharge the
high river flow caused by rainfall or rapid snowmelt, causing the river banks to breach
or be overtopped and flood water to spread into urban areas; (3) coastal flooding,
1
which is often caused by high tides and storm surges following gale-force winds or
hurricanes; and (4) groundwater flooding, which is most likely to be a problem in the
low-lying cities where the water-bearing rock strata are above the ground surface.
Since pluvial process is one of the most frequent causes of urban floods that disturb
citizens’ safety and normal life, this research will focus on pluvial flooding.
Table 1-1 The number of cities suffering from flooding in China from 2006 to 2016
Year Number of flooded cities
2006 118
2007 109
2008 115
2009 136
2010 258
2011 136
2012 184
2013 243
2014 125
2015 168
2016 192
*The cities in this table include county-level cities, prefecture-level cities and
municipalities.
A number of factors can cause the increasing occurrence of urban floods. Among them,
urbanisation plays an important role. Urbanisation has made the shift in residence of
the human population from rural to urban areas. Figure 1-1(a) illustrates the number
of people living in rural and urban areas worldwide, respectively
(https://ourworldindata.org/urbanization). Since 1980, the number of urban residents
has increased dramatically and overtook the number of people living in the rural areas
in 2007. As released by the United Nations
(https://www.un.org/development/desa/en/news/population/2018-revision-of-world-
urbanization-prospects.html), in 2018, 55% of the world’s population lived in urban
areas and the number is expected to rise to 68% by 2050. As a developing country,
China also experiences rapid urban expansion. The extent of the urbanised area in
China grows from 19.39% in 1980 to 64.72% in 2021, i.e. more than 3 times within
41 years (Figure 1-1 (b), http://www.stats.gov.cn). Fast urbanisation has brought
2
inevitable effects to the natural hydrological water cycle in urban areas and causes the
increased occurrence of urban floods. The construction in cities (e.g. construction of
buildings, streets, outdoor parking lots) has removed most of the vegetation and soil
grounds and replaced them with impervious surfaces. This transformation reduces
water infiltration and then increases the amount and speed of runoff on the ground
surface during rainfall events, potentially leading to flooding problems. In addition,
an increasing amount of reinforced concrete has been used to construct buildings and
infrastructures as the result of rapid urban expansion. Reinforced concrete can absorb
less heat than the vegetarian and grasslands, which leads to more hot air flows over
cities. Hot flows may eventually trigger more precipitation after accumulation and
thickening. Such phenomenon is called urban rain island effect and can further
aggravate urban flood problems in cities.
(a) (b)
Figure 1-1 Population change and urbanisation: (a) rural and urban population worldwide;
(b) Percentage of urban population in China.
Climate change is also a key factor that has increased the frequency of urban flooding.
First, climate change has led to more extreme precipitation events (Abiodun et al.,
2017; Hlavčová et al., 2015; Sun et al., 2021; Tauhid and Zawani, 2018; Zhou et al.,
2019, 2018) since warmer atmosphere can subsequently lead to an increase in water
evaporation and hence more dumping rainwater. Climate change also increases the
frequency of strong storms. In Atlantic Basin, the ruinous hurricanes are expected to
grow by 80% in frequency in the next 80 years (https://www.nrdc.org/stories/flooding-
and-climate-change-everything-you-need-know). These extreme weather events can
inevitably increase surface runoff and potentially trigger the occurrence of flood
hazards in urban areas.
3
Inadequate drainage system capacity is also an important factor that causes urban
flooding. Many cities still rely on combined wastewater and stormwater systems that
were designed and constructed several decades ago. These drainage infrastructures
may be significantly deteriorated and cannot match the contemporary urban
development. The underground drainage systems may lack sufficient maintenance and
upgrading. Collapse of pipes, blockage of the inlets, restrictions in the channel
capacity or failure of the pumps can substantially affect the function of the drainage
network system and potentially lead to urban inundations. Maintaining and upgrading
drainage infrastructure is always expensive. In many cities, especially in the
developing countries, because of a lack of funding the drainage system cannot be
routinely cleaned, imposing high risks during flood events.
Meanwhile, in modern cities, especially in the central districts of large cities, a large
number of underground facilities, such as underground shopping malls, subways and
parking lots have been developed as part of the rapid urbanisation. The vertical space
of cities has shifted from the traditional two layers (surface layer and drainage system
layer) to three layers (i.e. surface, drainage systems and underground spaces). The
development of underground spaces can make full use of the urban spaces and boost
economic development. However, these underground constructions can also bring
damage to the local environment and make cities more vulnerable to flood hazards.
When a city is hit by a flood, flood water may intrude into these underground spaces,
not only causing enormous economic loss but also threatening the lives of people
inside.
Table 1-2 shows some incidents caused by flooding in underground metro systems
worldwide (Lyu et al., 2019). In 2021, a record-breaking rainfall caused severe pluvial
flooding in Zhengzhou in China. The floodwater filled in Line 5 of Zhengzhou
underground metro system instantly and killed 14 passengers
(https://en.wikipedia.org/wiki/2021_Henan_floods). In 2016, an extreme rainfall
event hit Guangzhou in China. The Changpan Station of Line 6 of Guangzhou metro
was inundated, causing 8 deaths (Sun et al., 2022). Hence it is of great significance to
model flood inundation in these underground facilities to more comprehensively
assess urban flood risk and mitigate the impact.
4
Table 1-2 Historical flood incidents in underground metro systems

Date Location Damage Reference
One metro line was
10 May 2016 Guangzhou, China flooded and 8 people Lyu et al., 2016
lost their lives
7 underground metro
tunnels and 3
22-29 Otc. 2012 New York, USA Blake et al., 2013
vehicular tunnels
were inundated
The underground
tunnel system was
attacked by flood
6 Sep. 2003 Virginia, USA Sosa et al., 2014
within 40 minutes
with almost 167 litres
flood water
About 1/3 of Prague
August 2002 Prague underground metro Jakoubek, 2007
system was flooded
The underground Herath and Dutta,
29 June 1999 Fukuoka, Japan
metro was flooded 2004
Floodwater intruded
the metro tunnels and
13 April 1992 Chicago, USA Wagner, 1992
broke down the entire
metro system
The more frequent urban flooding and the following severe consequences necessitate
reliable flood risk assessment and effective management in urban areas. Computer
modelling provides an indispensable tool to understand the underlying physics of the
flood evolution process, which is essential for assessing flood risk and informing the
development of appropriate flood mitigation schemes. Generally, flood models may
be divided into several types, e.g. hydrological model, simplified hydrodynamic model
and full dynamic model. Hydrological models are effective to describe the hydrologic
cycle of water passing through urban areas including precipitation, evaporation,
infiltration and runoff, etc. Simplified hydrodynamic models have been developed and
5
applied in reproducing urban flood events in cities worldwide. However, hydrological

and simplified models cannot accurately predict the highly transient flood waves
interaction with the complex urban geometries, such as buildings, streets, road curbs
and riverbanks which have a direct effect on flood propagation and flow dynamics. It
is important to capture these dynamic interactions and more reliably quantify urban
flood risk, which requires the use of full dynamic models.
During the last two decades, full hydrodynamic models have been developed and
increasingly applied in urban flood simulations, especially in the simulation of floods
induced by intense rainfall events. Using hydrodynamic models, evolving flood
hydrodynamics including spatial and temporal changes of flood depth and velocities
can be accurately estimated to derive flood inundation maps and support the following
risk analysis. Since urban flood simulation is closely linked to and affected by complex
urban geometries and spatial heterogeneity, researchers have found that the quality
and resolution of the topographic data may have a significant influence on the
simulation results (Kim et al., 2021; Re et al., 2019; Xing et al., 2019). Therefore, low-
resolution simulations may be insufficient to represent urban geometries and
topography to provide reliable flood predictions. However, high-resolution
simulations will directly lead to a dramatic increase of computational efforts. To
overcome such a computational challenge, a common approach is to restrict the
research domain to a small area which may be only a part of a city/town. This method
may reduce the computational burden, but it introduces more uncertainties in the
simulation results because during a flood event hydrological and flooding processes
in an urban city/catchment are closely connected with each other due to the
interlinkage of the complex urban environments. For example, runoff or flood routes
may be jointly influenced by the key urban features, e.g. streets, buildings, rivers, and
drainage networks. Simulations on a part of a city require setting the domain
boundaries manually or even arbitrarily, which inevitably cuts off the flow paths
crossing the boundaries and may produce unreliable predictions.
In recent years, high-performance computing technology has experienced rapid

advances, providing a great opportunity to improve the performance of hydrodynamic
models. General-purpose graphics processing units (GPGPUs), one of the most
notable breakthroughs in high-performance computing technology, have been further
developed and applied in scientific computing. Graphic Processing Units (GPUs) are
6
originally designed for picture processing or image rendering which is

computationally intensive but not complex. A GPU-based architecture has been shown
to boost the performance of computation 10 to 20 times compared with the equivalent
Central Processing Units (CPUs). In 2007, NVIDIA released the Compute Unified
Development Architecture (CUDA), allowing direct programming to support
scientific computing on GPU. Since then, the potential of GPU computing power has
been explored by engineers and scientists and has been applied in accelerating
scientific computing in a variety of disciplines, e.g. seismic imaging, computational
biophysics and fluid dynamics (Castro et al., 2011; Nickolls et al., 2008). Due to the
powerful computational capability of GPU, high-resolution large-scale urban flood
simulations with hydrodynamic models can now be conducted on a small server or
even a desktop/laptop with GPU devices rather than a remote supercomputer centre.
This has led to an opportunity for developing and applying hydrodynamic models to
simulate floods across an entire city and systematically investigate the interaction of
flood flows and high-resolution urban topographic features.
In summary, urban flooding has become increasingly frequent, causing more dramatic
economic loss and threatening people’s lives. It is of great importance to better
understand and predict the evolution process of urban flooding to inform the
development of effective mitigation plans to manage flood risk. Cities are highly
heterogeneous and have multi-layer spaces due to the complex urban structural
features, e.g. buildings, roads, drainage networks and underground spaces. Modelling
the complex flood dynamics across the multi-layer spaces at a city-scale is however
beyond the capability of most existing models. There is still an urgent need to develop
hydrodynamic models to support efficient simulation of large-scale urban floods at
sufficiently high resolution to capture the detailed interaction between the flood flow
and the complex multi-layer urban environments. Taking advantage of the rapid
development of high-performance computing technology in recent years, such a high-
resolution flood modelling strategy has become feasible and will be explored in this
work to fill the current research gaps and improve the current urban flood modelling
practice.
7
1.2 Research aim and objectives
The aim of this thesis is to develop a new urban flood modelling system by the virtue
of advanced high-performance computing technology. The developed modelling
system will support large-scale multi-layer high-resolution urban flood simulations to
capture the hydrodynamic process between flood flows and complex urban structural
features. It is expected that the outcomes of this research will provide a new modelling
tool to improve the current practice of urban flood risk modelling and assessment,
contributing to planning and designing flood resilient cities. Specifically, this aim can
be achieved by the following objectives:
➢ Develop a new drainage model for dynamic simulation of drainage flows
Short time of concentration and large highly impervious surface are two key features
in urban flood hydrology. The runoff is often collected and transported through sewer
drainage networks during a storm event. Accurately and robustly predicting drainage
flows is therefore essential when reproducing urban flood events in an urban
modelling system. Numerous numerical schemes have been developed to predict the
flows in pipes. However, the calculation of the flow condition in junctions of the sewer
drainage networks has received much less attention and is traditionally simplified to
only solve the continuity equations. It ignores the momentum exchange in the
junctions and hence cannot provide sufficient boundary conditions for the pipe
calculations. This work proposes a novel numerical scheme to calculate the flow
dynamics in junctions, which directly takes into account both mass and momentum
conservation and removes the necessity of implementing complicated boundary
settings for pipe calculations.
➢ Couple the developed drainage model with an in-house high-

performance hydrodynamic surface flood model
During an urban flood event, the surface flow can be discharged into drainage systems
while surcharge from junctions to the ground surface can also occur as part of the
flooding process. In order to accurately simulate the complex flooding process
featured with surface water and drainage flow interactions, the new drainage model is
coupled with the High-Performance Integrated hydrodynamic Modelling System
8
(HiPIMS) (Xia et al., 2019, 2017) to form a new dual drainage model for high-
resolution urban flood modelling.
➢ Extend the improved HiPIMS for simulating flooding process in

underground spaces
The new high-performance dual drainage modelling system is extended to simulate

different city layers so that the flow on the ground surface, in the drainage systems
and underground spaces can be dynamically linked and the interactions are effectively
captured. This is of great significance for assessing the flood risk in the underground
spaces and helps to provide effective guidelines for mitigating risk during flood events.
➢ Implement the integrated modelling system for GPU high-performance

computing
To match the performance of HiPIMS in high-resolution flood modelling across a

large domain, the new modelling components developed in this work are implemented
on GPUs by adopting the same modelling framework to form the final new high-
performance multi-process urban flood modelling system (HiPIMS-urban).
1.3 Thesis outline
The rest of the thesis is organised as follows:
Chapter 2 provides a comprehensive literature review of the current state of practice

in urban flood simulations, the development of drainage model, the practice of flood
flow simulation in underground spaces and the GPU high-performance computing in
urban flood modelling.
Chapter 3 presents a novel drainage model, which is developed through coupling a

1D pipe model and a 2D junction model. The drainage model is validated by an
experiment and several idealised tests.
Chapter 4 introduces the GPU-acceleration process of the drainage model and

couples the drainage model with the surface component of HiPIMs. The model is
validated against an experiment test case.
9
Chapter 5 develops an integrated modelling system for simulating surface flow,

drainage flow and flood flow in underground spaces simultaneously in an idealised
example.
Chapter 6 further develops the integrated urban flood modelling system by utilising
multi-GPUs high-performance computing scheme based on the work in Chapters 3
and 4.
Chapter 7 draws the major conclusions from this project and provides some
suggestions for future work.
10
Chapter 2 Literature review
Considerable advances have been made to develop models for urban flood simulations.
These models can be categorised as surface flow model, drainage flow model and dual
drainage model. The selection of a proper flood model to simulate a particular flood
event should be made based on various practical factors. For example, the model
should be able to reproduce the flood evolution process accurately. In addition, the
model should be efficient so that necessary measures can be taken based on the
predicted results before potential floods happen. These factors are important and
should be taken into consideration when choosing or developing a flood model for
urban flood simulation. This chapter will present the literature review related to
developing models for urban flood simulation.
Figure 2-1 Rainfall-runoff process in urban areas (Bulti and Abebe, 2020)
Figure 2-1 shows a schematic diagram of the rainfall-runoff process in urban areas.
Rainfall, which is the main precipitation type, is the main source of runoff on the urban
surface ground. Not all rainfall falling on the urban surface ground will convert to
runoff. It may generate initial loss during this process. Initial loss refers to the rainfall
water that remains on vegetation and building roofs, or the water trapped in the low-
11
lying areas or puddles. Other losses include evaporation and infiltration. The amount
of these two types of loss is related to the property of urban surface ground, for
example, during intense rainfall more impervious surface means less infiltration. The
infiltration is usually high at the beginning of rainfall events and starts to decrease
exponentially when the soil becomes saturated (Horton, 1933). When the amount of
rainfall water exceeds all losses introduced before, runoff will be generated and flow
along the surface topographic properties, e.g. streets, gutters or natural flow paths and
then discharge into the drainage system. When the runoff amount exceeds the capacity
of drainage systems, the flood occurs. It is of great significance for urban flood models
to reproduce these important processes during urban floods. In an urban flood model,
surface flow model and drainage flow model are the two most important components
to simulate surface runoff and flow in the drainage system. They will be reviewed in
Sections 2.1 and 2.2, respectively. Dual drainage models, which aim at simulating the
flows both through the drainage and on the surface, will be reviewed in Section 2.3.
Section 2.4 presents the review of the simulation of floods in underground space while
existing acceleration methods for urban flood modelling are discussed in Section 2.5.
2.1 Surface flow models
Numerous numerical models have been developed to simulate surface runoff flow.
These models can be divided into two broad categories: hydrological model and
hydrodynamic model.
Hydrological model is an effective tool that utilises mathematical analogues, small-

scale physical models and computer simulation skills to describe real-world
hydrological features or processes (Anees et al., 2016). Various hydrological models
have been developed and applied to calculate the surface runoff in urban catchments
for flood simulations. For example, the Storm Water Management Model (SWMM),
which is developed by U.S. Environmental Protection Agency, has been widely used
for calculating and analysing the quantity and quality of runoff during a rainfall event
in urban areas. To explore the capacity of SWMM in flood simulation in small
subtropical urban catchments, Tsihrintzis and Hamid (1998) run the runoff component
of SWMM on four domains with four different land-use types (e.g. low-density
residential area, high-density residential area, commercial area and highways). The
quantity and quality of stormwater runoff produced on these four different land-use
12
domains have been calculated in this study and appropriate input parameters for each
land-use type have been discussed. Guan et al. (2015) utilised SWMM to simulate
flood flows in a developing urban catchment in order to assess how the urban
development affects the change of the urban hydrology. Xu et al. (2019) adopted
SWMM to simulate the runoff on two typical kinds of urban green lands: lawn and
shrub. The interception and infiltration capacities of these two green lands have been
explored and runoff reduction results have been discussed, which promotes the
implementation of low-impact development (LID) in cities. Sun et al. (2014) discussed
the relationship between the surface discretisation methods and the parameterisation
in SWMM and found that the different discretisation schemes of urban catchment had
a great effect on the parameterisation and simulation results. Simulating surface runoff
on urban catchments by using SWMM can also be found in Ahamed and Agarwal
(2019), Barco et al. (2009), Jang et al. (2007), Ji and Qiuwen (2015), Jiang et al. (2015),
Wang and Altunkaynak (2012), Hossain et al. (2019), Vidya (2021). There are also
other hydrological models that have been frequently used to calculate surface runoff,
including HEC-HMS (Chu and Steinman, 2009; De Silva et al., 2014; Duhan and
Kumar, 2017; Katwal et al., 2021; Msaddek et al., 2020; USACE, 2016), MIKE-SHE
(Aredo et al., 2021; Golmohammadi et al., 2014; D. Li et al., 2019; Ma et al., 2016;
Rujner et al., 2018; J. Zhang et al., 2021), the surface runoff component of MOUSE
(Broekhuizen et al., 2019; Thorndahl and Schaarup-Jensen, 2007) and LISFLOOD
(Feyen et al., 2007; Gai et al., 2019; Hirpa et al., 2018; van der Knijff et al., 2010).
These hydrological models are generally simple to implement and can be used to
investigate the water transfer between each physical process within urban hydrological
environment (Parra et al., 2018), and can support flood simulations over large-scale
urban catchments due to their efficient computational capacity (Xia et al., 2019).
However, most of the hydrological models are highly affected by uncertainties of
model parameters partly because a large number of model parameters are involved in
the models’ setting-up (Salvadore et al., 2015). Moreover, without an accurate
representation of detailed flood hydraulics, hydrological models could not reliably
describe the interaction between the dynamic flood flow and the complex urban
structural features.
13
Hydrodynamic model depicts the mechanical behaviour of water in the physical water
system. The hydrodynamic model can be categorised as one-dimensional, two-
dimensional and three-dimensional models.
The shallow water equations (SWEs) are generally used in hydrodynamic models to
depict flow motion. The SWEs are derived by depth-integrating the Navier–Stokes
equations, assuming that a flow has a wave length much larger than the water depth,
which is valid for most of the flows with horizontal extent much larger than vertical
depth. The models can be classified into one- and two-dimensional models according
to the dimensions that represent the spatial domains and flow paths.
Early attempts mainly focused on developing 1D shallow water equation (also known
as 1D St. Venant equation) models for flood simulations in complex river/channel
systems and other hydraulic practices. For example, Gouta and Maurel (2002)
proposed a feasible solver for 1D SWEs and applied this scheme to simulate flood
flows in an actual river. Alias et al. (2011) presented a 1D shallow flow solver which
can calculate dynamic flows in rivers or channels with varying width and bed profiles.
1D SWEs model has also been developed for the urban flood simulations. In the 1D
model, the urban surface floodplain is discretised as a set of nodes which are connected
by different links. The nodes represent the channel junctions or ponds, and the links
are characterised as overland pathways such as urban streets and channels (Bulti and
Abebe, 2020; Henonin et al., 2013; Leandro et al., 2009; Rodríguez et al., 2015). The
urban surface network needs to be extracted before simulation, which can be done
manually or automatically based on urban digital elevation model (DEM) with
different methods. For example, Maksimović et al. (2009) developed a set of GIS
modules to define the overland flow pathways network prior to modelling floods in
urban areas. Other hydrodynamic software packages based on 1D SWEs have also
been developed and applied in the urban flood flow simulations, e.g. HEC-RAS
(Anees et al., 2016; Costabile et al., 2020; Desalegn and Mulu, 2021; Marimin et al.,
2018; Namara et al., 2021; Tamiru and Dinka, 2021; Zeiger and Hubbart, 2021) and
MIKE-11 (Ajiwibowo and Pratama, 2020; DHI, 2017; Doulgeris et al., 2012; Liang et
al., 2015; Liu et al., 2022; Rujner et al., 2018; Vari, 2017). Using 1D SWEs model
enables more efficient urban flood simulations with shorter run-time, however, it
requires more time for data pre-processing in the modelling setup (Henonin et al., 2013;
Rodríguez et al., 2015). Besides, the topography is discretised by cross-sections rather
14
than by surface grid cells in 1D hydrodynamic models. Such discretisation scheme

leads to that the 1D models cannot simulate spreading flows in urban built-up areas
(e.g. commercial and residential areas) and are not capable of visualising flood extents
(Bisht et al., 2016). This includes the inability to compute the lateral diffusion flux of
flood waves. Therefore, the 1D models are not suitable for modelling flow dynamics
over complex urban surface areas.
With the development of computer technology and the availability of high-resolution

topographic data, a large number of 2D hydrodynamic flood models based on the 2D
SWEs have been developed in recent years. 2D SWEs provide an effective
mathematical framework to describe the complex hydrodynamic flow problems such
as overland floods, shallow lake currents, estuarine and coastal flows (Bi et al., 2015).
In 2D SWEs model, the urban catchment is discretised as a large number of structured
or unstructured grid cells. In a structured mesh scheme, the topographic features of
urban surfaces are generally represented by quadrilateral grid cells, whereas in an
unstructured mesh scheme the triangular or mixed triangular and quadrilateral grid
cells are dominant (Kim et al., 2014). Each grid cell holds a point with coordinates (X,
Y, Z) and the related catchment parameters such as Manning coefficient and rainfall
intensity are assumed to be spatially homogeneous inside each cell (Rodríguez et al.,
2015).
The 2D SWEs model can be classified into three types according to the complexity:
kinematic wave, diffusion wave and dynamic wave models. In the kinematic wave
model, the gravity driving force and the resistant force of the flow are assumed to be
balanced with each other therefore the flow can achieve a steady uniform state.
However, the assumptions of the kinematic wave model may be oversimplified in
certain situations (Kazezyilmaz-Alhan, 2012). Simons (2019) indicated that the
kinematic wave model is more suitable in the simulation of the flood waves along
steep channel surface where the slope should be greater than 0.002. The diffusion wave
model is obtained by introducing physical diffusion into the kinematic wave equations
and then the resistant force is assumed to be balanced not only with gravity driving
force but also with pressure force. Diffusion wave models are capable of simulating
flood waves in open channels/rivers or on flood plains with mild slopes. LISFLOOD-
FP (Bates and De Roo, 2000; Neal et al., 2018; Sadeghi et al., 2022; Shustikova et al.,
2020, 2019; Sosa et al., 2020) and JFLOW (Bradbrook, 2006; Bradbrook et al., 2005;
15
Defra and Environment Agency, 2010; Neelz et al., 2010) are two of the most popular
software packages that were developed based on diffusion wave equations, and they
have been widely used in flood simulations. Although in some cases solving the
simplified kinematic wave equations or diffusion wave equations is relatively
straightforward and less computational demanding compared to solving the full SWEs,
these two kinds of models have been limited in the simulation of highly transient flows.
During intense rainfall events, flood flows running over the heterogeneous urban
environment and then shock waves / hydraulic jumps may be produced, leading to
significant numerical errors when using simplified models to mathematically capture
different flow’s hydraulic behaviours.
Another type of 2D hydrodynamic model which solves full 2D SWEs (also known as
2D dynamic wave model) is an alternative way for overland flow simulations.
Compared with simplified kinematic wave model and diffusion wave model, full
SWEs-bassed models can obtain more accurate and robust predictions in simulating
flow dynamics over complex topography. Software packages such as MIKE21
(Fadlillah et al., 2020; X. Li et al., 2020; Safavi et al., 2015; Soltani et al., 2017; L.
Yang et al., 2020; Zolghadr et al., 2011), TELEMAC-2D (Z. Li et al., 2019; Linh et
al., 2018; Liu et al., 2019a, 2019b; Tadesse and Fröhle, 2020), TUFLOW (Defra and
Environment Agency, 2010; Fahad et al., 2020; Jovanovic et al., 2019; Morsy et al.,
2018; Phillips et al., 2005; Quan et al., 2020; WBM, 2001) RiverFLO-2D (Garcia and
Gonzalez, 2013; Lacasta et al., 2014; Martinez et al., 2011) and Infoworks RS-2D
(Falter et al., 2013) have been developed based on full SWEs. Besides the existing
software packages, many researchers have also developed full 2D SWEs model for
handling the complex hydrodynamic flow problems in flood simulations (Cea and
Vázquez-Cendón, 2010; Liang, 2010; Liang and Marche, 2009; Rousseau et al., 2012;
Simons et al., 2014; Tsakiris and Bellos, 2014).
Different numerical methods have been explored and adopted to solve SWEs
numerically, such as finite difference method (FDM) (Miyata, 1986; Peiro and
Spencer, 2005; Rousseau et al., 2012; Tsakiris and Bellos, 2014), finite volume
method (FVM) (Alcrudo and Garcia‐Navarro, 1993; Bradford and Sanders, 2002;
Caleffi et al., 2003; Liang, 2010; Peiro and Spencer, 2005; Rousseau et al., 2012; Yu
and Duan, 2014) and finite element method (FEM) (Giammarco, 1996; Peiro and
Spencer, 2005). Compared with FDM and FEM, the FVM presents distinct advantages
16
since it strictly maintains mass and momentum conservation and can be constructed
for arbitrary geometrics. In the last two to three decades, Godunov-type finite volume
schemes has been widely used in solving SEWs (Alcrudo and Garcia‐Navarro, 1993;
Chippada et al., 1998; Hou et al., 2015, 2017; Liang, 2010; Liang and Marche, 2009;
Nazari and Nair, 2017; Xia et al., 2017; Yu and Duan, 2014). In Godunov-type finite
volume scheme, the computation domain can be discretised by large number of grid
cells and solutions can be obtained through solving local Riemann problems at each
cell interface over these computation grid cells. The finite volume Godunov-type
scheme can provide efficient approaches to simulating various flows including
supercritical flows, subcritical flows, transcritical flows, flow discontinuities over
complex topography and wet-dry fronts. Therefore, it is an idealised option to adopt
finite volume Godunov-type scheme to solve SWEs.
The recently developed High-Performance Integrated hydrodynamic Modelling

System (HiPIMS) (Xia et al., 2019, 2017) is an effective tool for predicting floods and
assessing the risks, which is adopted in this work for urban surface flood simulation.
The surface overland model of HiPIMS utilises a shock-capturing Godunov-type finite
volume scheme to solve the full 2D SWEs. HiPIMS is capable of handling two key
challenges that are always faced in urban flood simulations: (1) it proposes a novel
surface reconstruction method that can deal with flows over highly irregular urban
topography which may involve steep/abrupt slopes; and (2) it develops an implicit
scheme to discretise nonlinear friction terms in SWEs which can achieve stable and
accurate predictions in the limit of disappearing water depth (Xia et al., 2017). The
model is incorporated with the classic Harten, Lax, van Leer with contact wave
restored (HLLC) Riemann solver (Harten et al., 1983; León et al., 2006; Liang and
Marche, 2009; Sanders et al., 2010) and can automatically capture various flood flow
features, e.g. subcritical flow, transcritical flow and supercritical flow. HiPIMS has
been validated using analytical and experimental test cases and applied in reproducing
a series of real rainfall events (Ming et al., 2022, 2020; Xia et al., 2019; Xing et al.,
2022b, 2022a).
Extra attempts have been made to develop three-dimensional models to simulate flood
flows in urban areas. Such model solves the full Navier-Stoke equations and considers
urban flood flow as completely three-dimensional. For example, Rong et al. (2020)
utilised both 2D and 3D hydrodynamic models to simulate floods in a coastal city and
17
then quantify the flood damage to the local environment. The research reveals that the
assumptions and approximations of 2D model weaken when the vertical fluctuations
of flow are enormous, which often can be found within complex urban environment.
Combined with high-resolution urban topographic data, 3D model can better
reproduce floods. Applying 3D model in costal flood simulation can also be found in
Anees et al. (2006), Li et al. (2006), Zhang et al.(2018). It is worth mentioning that
these works focused on small urban areas. Applying such 3D models for large-scale
urban flood simulations may be not feasible due to its low computational efficiency
and high requirement of topographic data.
Overall, the full 2D SWEs model is a feasible option to depict flood flow motion,
especially capturing the dynamic interaction between flood flow and urban small scale
topographic features. However, the computational cost of the full 2D SWEs models
is expensive, making large-scale real-world applications challenging. Therefore,
different acceleration techniques have been proposed to improve the simulation
efficiency, which will be discussed in Section 2.5.
2.2 Drainage models
Drainage network modelling is often an integrated component of an urban flood

simulation system. Normally in drainage model, a virtual storage pond is assumed to
be on the top of manhoel (also called junction) and overflow can be temporarily stored
in this pond. Pipes and junctions are the two essential elements of any sizable urban
drainage network and are commonly calculated by different model components in
drainage network models. To predict the flow dynamics in pipes, the 1D Saint-Venant
equations (or one of the modified/simplified forms) are often used and solved
numerically. In most of the drainage network models, effective approaches have been
developed to handle the transitioning free-surface and pressurised flow conditions in
pipes that repeatedly happen during an urban flood event. One group of these
approaches uses different equations for free-surface and pressurised flows. Examples
include the interface tracking model (Politano et al., 2007; WIGGERT DC, 1972), the
rigid column-based model (Li and McCorquodale, 1999; McCorquodale and Hamam,
1983) and the Illinois transient model (ITM) (Leon et al., 2010a). Another type of
widely used approach solves a single set of equations but is incorporated with
numerical calculation schemes to handle pressurised flows. A typical example is the
18
Pressman slot scheme proposed by Preissmann, which has been widely adopted and
further developed by many researchers (e.g. An et al., 2018; Capart et al., 1999; Chen
et al., 2021; Cunge et al., 1980; Malekpour and Karney, 2014; Maranzoni et al., 2015;
Trajkovic et al., 1999; Wang et al., 2019). This approach is based on a hypothetical
narrow slot on the crown of the pipe and the water depth in the slot varies in terms of
the flow pressure. Choosing an appropriate slot width is difficult when implementing
the Pressman slot scheme and an improper slot width may cause model instability with
spurious numerical oscillations (Trajkovic et al., 1999; Vasconcelos et al., 2006). An
alternative method called the two-component pressure approach (TPA) was also
proposed and reported by Vasconcelos et al. (2006) for simulating transient flows.
TPA models assume that the pipe walls are elastic and subsequently the cross-sectional
area of a pipe may expand when the flow inside is pressurised. TPA models can
effectively simulate various types of unsteady flows including free-surface flow,
mixed flow (partly gravity-partly pressurised flow), pressurised flow, sub-atmospheric
pressure flow as well as flow transitions (Bousso et al., 2013).
The calculation of junction flows, however, has received much less attention despite
the fact that they are an indispensable part of a drainage network model and are
essential to providing the necessary boundary conditions (BCs) for the accurate
calculation of pipe flows. The traditional approach for the junction flow calculation
neglects momentum conservation and considers only the continuity equation to
estimate junction water depth. Such an approach has been widely used in drainage
modelling and implemented in SWMM (Bisht et al., 2016; Burger et al., 2014; Hsu et
al., 2000; Leandro and Martins, 2016; Rossman, 2015) and many other urban drainage
models (e.g. Chang et al., 2015; Leandro and Martins, 2016; Noh et al., 2018, 2016;
Schmitt et al., 2005). Although this traditional approach may be computationally
efficient, it normally requires additional complicated methods to provide sufficient
BCs for transient flow calculations in pipes. For example, a decision tree method was
implemented by Capart et al. (1999) at the interfaces between junctions and pipes to
illustrate possible boundary flow regimes. Sanders and Bradford (2011) extended this
work and developed an improved framework to include different types of BCs for free-
surface and pressurised flows. A similar effort has also been made in modelling the
flow around an island in a river where the flow connections around the island are
represented as junctions/bifurcations to provide inner BCs to connect with the river
flow (Franzini et al., 2018). However, these approaches require the identification of
19
various BCs according to the flow variables (e.g. Froude number, water level, dryness
tolerance, etc.) at each time step, which is difficult to implement and may affect the
computational efficiency and numerical stability of the overall drainage model. León
et al. (2010b) proposed a junction and drop-shaft BC model, which was coupled to an
ITM for the simulation of mixed flows in pipes. Separate Ordinary Differential
Equations (ODEs) derived from the conservation of mass and momentum have been
also used for estimating junction flows (Borsche and Klar, 2014). However, the
resulting approaches are complicated and computationally inefficient as a varying
number of equations must be solved at each junction according to the connecting pipes
and flow conditions.
Computational fluid dynamics (CFD) models have also been adopted to simulate the
complex flow dynamics including turbulence and vorticity inside junctions (Beg et al.,
2018a, 2018b). However, these models are considered to be over-sophisticated for a
city-scale drainage flow simulation where the localised fluid structures will have
limited influence on the broad-scale flow dynamics. Furthermore, the lack of detailed
junction data makes the computationally expensive effort of little practical value.
Preliminary attempts have also been made to use 3D (Hong and Kim, 2011) or 2D
(Bermúdez et al., 2017; Casulli and Stelling, 2013; Herty and Seaïd, 2008) domains
to idealise junction nodes in gas pipe network modelling. However, such an approach
has not been investigated in modelling storm water drainage networks.
2.3 Dual drainage models
The dual-drainage models usually comprise two main components: the minor system
and the major system. The minor system represents the traditional storm drainage
network including manholes, pipes, inlets et. al, which is of great significance in
conveying the flow from surface ground to the downstream waterbodies in urban
areas. The major system is the flow paths along surface ground during flood events
when the rainfall intensity surpasses the conveyance capacity of the minor system
(Djordjević et al., 1999; Jahanbazi and Egger, 2014). The flow exchange between
major and minor systems takes place at inlets, manholes and drainage outfalls. During
rainfall events, inlets are generally supposed to collect surface runoff and discharge it
into underground storm drainage systems. Manholes are always set for maintenance.
If the maximum conveyance capacity of the drainage networks is reached or the inlets
20
are blocked, the surface flow cannot enter into the minor system and will be retained
on the surface major system. Likewise, if the capacity of the drainage network is
surpassed, the water can spill out from the minor system at manholes or inlets.
Exchange also happens at outfalls where the drainage flow will be discharged into the
downstream water bodies. However, when the water level of the water bodies rises
during flood events, backwater can be found at outfalls which further reduces drainage
capacity. Dual drainage model is an effective tool to capture the water movements
between these major and minor systems.
Dual drainage models can be classified into two categories: 1D/1D models and 1D/2D
models. In both two categories, the pipe flow in the underground drainage system is
modelled by 1D governing equations while the surface flow is modelled based on the
surface 1D channel/flow path/pond networks in 1D/1D models, and the 2D surface
grid cells in the 1D/2D models (Leandro et al., 2009). SIPSON, which was developed
by Djordjević et al. (2005), is one of the widely used 1D/1D models (Chen et al., 2016;
Martins et al., 2017). A hydrological model called BEMUS (Radojkovic and
Maksimovic, 1984) was adopted in SIPSON to calculate rainfall-runoff hydrographs
and then the hydrographs were set as the input for the underground drainage systems
at manholes. The 1D hydraulic model was then utilised to capture the flow dynamics
in the 1D drainage system. When the surcharge happened at manholes, the water
flowed outside and then 1D St. Venant equations were solved for modelling flow in
streets or conduits. With the development of the data acquisition methods, the surface
channel/flow path/pond can be better represented, and then the 1D/1D dual drainage
models can predict accurate floods along pathways or inside ponds (Leandro et al.,
2009).
Benefiting from this overland flow network delineation method, Simões et al. (2010)
utilised the 1D/1D model to conduct sensitivity analysis in two urban surface runoff
generation cases for urban flood forecasting. However, the 1D/1D models become less
accurate when the accumulated flood depth is greater than the bank of the flow paths
or the crest of the ponds since the flood flows are no longer confined by the
predetermined pathways and then spread outside. In this situation, the 1D/1D dual
drainage models will not be able to simulate flood propagation properly, which
motivates the development of the 1D/2D models.
21
Hsu et al. (2000) proposed a 1D/2D urban inundation model in which the SWMM was
adopted for the underground 1D drainage flow simulations and a 2D diffusive
overland flow model was adopted for the surface flow simulations. SWMM is
employed to provide surcharged flow hydrographs at manholes when the surface
runoff exceeds the capacity of the drainage systems. The 2D overland model is then
supposed to calculate the inundation extent and water depth on the overland surface
due to the surcharged water from manholes. However, in this approach, the interaction
between surface and drainage systems is unidirectional (from underground drainage
system to the surface ground) and the model failed to handle the situation where the
surface water re-enters the drainage system when the capacity of the pipes recovers.
This may lead to the overestimation of the surface inundation. Hsu et al. (2002)
improved the model by considering the return of the surface flow to the drainage
system but did not take into account the interactions of flow hydrodynamics between
surface and drainage flows. Chen et al. (2007) considered surface flow conditions in
the calculation of the flow exchange between surface and underground drainage
systems, and then developed an integrated 1D/2D model: SIPSON/UIM. As
introduced previously, SIPSON is a 1D model that can simulate flow dynamics inside
1D drainage networks. UIM is a 2D overland flow model which was adopted to
simulate the hydrodynamic flows over the surface ground. Computational challenges
exist when the 2D UIM model is applied to the simulations with a large number of
computational grid cells. To overcome this issue, an adaptive time step was
implemented in the 2D UIM model and carefully synchronised with the 1D component
to improve the computation efficiency. Jahanbazi and Egger (2014) also proposed a
1D/2D integrated HYSTEM-EXTRAN 2D model and applied it in a sub-catchment of
the city Hamburg, Germany to assess urban flooding. Compared with a conventional
method in which the overland inundation was derived from sink volume within GIS,
the HYSTEM-EXTRAN 2D can better predict flood extent as well as depth, and is
more user-friendly. Martins et al. (2017) made a comparison among three 1D/2D dual
drainage models with three different surface flow components. The SIPSON was
adopted in these three models for the 1D sewer flow calculation. For the surface flow
modelling, the full 2D SWEs, the local inertial equations which neglect the convective
acceleration terms in SWEs (known as GWM) and the diffusive wave equations
without inertia which neglect both the convective and local acceleration terms in
SWWs (known as PDWAVE) were adopted, respectively. The study shows that when
22
simulating the flood flows over domains with relatively low slopes, the PDWAVE
predicts higher maximum water depth and presents a slow propagation of water waves.
The GWM has a higher water depth at the downstream of pathways than the SWEs’
since the wavefront of GWM is slower. Overall, the GWM and SWEs have a better
performance than PDWAVE in capturing the flow dynamics. This study analysed the
limitations of the simplified hydrodynamic models in the application of dual drainage
flow simulations, presenting a reference for choosing appropriate 2D overland flow
models with considering both model accuracy and computational efficiency.
Besides the in-house models, commercial software based on coupled 1D/2D models
has also been developed and widely used in urban flood modelling in recent years. For
example, XP-SWMM (Haron et al., 2016; Phillips et al., 2005; Spry and Zhang, 2006;
Van Der Sterren and Rahman, 2011) was developed by coupling XP-SWMM 1D
model and TUFLOW 2D model to simulate flow in the underground drainage system
and surface ground simultaneously. MIKE Urban (Bisht et al., 2016; Hernes et al.,
2020; Salvan et al., 2018) is also one of the 1D/2D commercial software which
integrates MOUSE and MIKE 21 to simulate combined sewer and surface flows. Bisht
et al. (2016) adopted MIKE URBAN to assess urban flood and drainage performance
in Indian Institute of Technology Kharagpur campus, India and summarised that
MIKE URBAN can provide and visualise the direction of overland flow as well as the
inundation depth over the flooded surface. Salvan et al. (2018) utilised MIKK URBAN
to analyse the flood flow in the urbanised areas within the Magnan basin in France to
explore the interaction between the surface inundation and underground drainage
flows. InfoWorks Integrated Catchment Modeling (ICM), which was developed by
coupling 1D hydrodynamic model for river and drainage flow simulation and 2D
hydrodynamic model for surface flooding simulation, is also an integrated commercial
software that has been wildly applied in urban flood management (Cheng et al., 2017;
Sidek et al., 2021; H. Yang et al., 2020; H. Zhang et al., 2021).
When coupling the 1D sewer flow model (SFM) and 2D overland flow model (OFM),
different methods have been tested. The integrated 1D/2D models can be classified
into two categories according to the adopted coupling methods: the SFM/OFM
approach and the OFM/SFM approach (Chang et al., 2015). In the SFM/OFM
approach (e.g. Chen et al., 2007; Seyoum et al., 2012), the surface runoff produced by
23
the rainfall data is first calculated in a hydrological model, which is set as the inflow
discharges of the manholes for the underground pipe flow calculations. The 2D OFM
is triggered by the surcharges from manholes and the inundations are then calculated
in the surface component. The OFM/SFM approach (e.g. Hsu et al., 2002) applies
rainfall on the 2D surface OFM and produces surface runoff directly. The surface
water flows traverse along the terrain pathways and then discharge at inlets or
manholes. Each of these two 1D/2D coupled models has its own advantages. For
example, in some built-up areas, the rainfall collection systems have been installed on
building roofs to drain rainfall from the building roof to the underground drainage
system directly. When rainfall exceeds the roof inlet capacity, the water may be
confined inside roof walls. The SFM/OFM is an ideal tool to consider the effect of
rainfall collection system in reproducing urban flood events. However, it cannot
accurately represent the dynamic rainfall-runoff process over the surface domain,
which can be captured in the OFM/SFM. Chang et al. (2015) analysed the SFM/OFM,
OFM/SFM and mixed models and indicated that the mixed SFM/OFM and OFM/SFM
model can reproduce results closer to the survey data since both the roof effect and the
physical rainfall-runoff process in the urban environment have been considered.
Overall, the application of the 1D/2D coupled dual drainage model over complex
urban topography has been conducted by a large number of researchers (e.g. Bulti and
Abebe, 2020; Chen et al., 2016; Leandro and Martins, 2016; Martins et al., 2017;
Qiang et al., 2021; Schmitt et al., 2005; Tayefi et al., 2007). Table 2-1 compares the
1D/2D coupled dual drainage model with other types of models and demonstrates that
utilising 1D/2D coupled dual drainage model makes it possible to obtain more detailed
flood information (e.g., flood depth, velocity and inundation extent and overflow
location on the surface domain and drainage flow discharges in the underground
drainage system). Such information can help to understand flood evolvement and
contribute to risk management. However, the 1D/2D coupled models always suffer
from a high computational burden. They require a more detailed spatial discretisation
scheme and much shorter time steps than the 1D/1D models. Computational efficiency
is considered to be the main limitation for the 1D/2D models when being applied in
real-time flood simulations (Fan et al., 2017; Leandro et al., 2009; Rodríguez et al.,
2015; Tayefi et al., 2007). Hence these models have often been used in the small urban
24
catchment flood simulations or in the simulations with coarse terrain topographic data
to have an acceptable run time.
Table 2-1 Comparison of different types of models in urban flood simulations

1D/1D dual 1D/2D dual
Models 1D surface 1D drainage 2D surface
drainage drainage
Yes, but Yes, but
represented represented
Yes, Yes,
Overland flow in ponds or in ponds or
No represented represented
representation surface surface
in 2D in 2D
channel channel
networks networks
Yes, but only
with an Yes, but
Yes, but only
approximation only in
Surface in ponds or
within the ponds or
inundation surface Yes Yes
virtual storage surface
extent channel
ponds over the channel
networks
top of networks
junctions
Yes, but only
with an Yes, but
Yes, but only
approximation only in
Surface in ponds or
within the ponds or
inundation surface Yes Yes
virtual storage surface
depth channel
ponds over the channel
networks
top of networks
junctions
Yes, but
Yes, but only Yes, in Yes, in
Surface flow only in
in one- No two- two-
velocities one-
direction direction direction
direction
Yes, but only

Overflow from
obtain
drainage No No Yes Yes
overflow
systems
volume
Drainage flow
discharges No Yes No Yes Yes
Computation
Relatively
efficiency Relatively high Low Low Very low
high
25
Accuracy for
Medium-
flood risk Low Low Medium High
high
analysis
2.4 Simulation of floods in underground space
The rapid development of urbanisation and industrialisation globally in recent years

has brought an increase in population, infrastructure, industrial facilities, and
transportation in cities, which requires advanced urban space utilisation. The lack of
surface space due to the expanding urbanisation has promoted the development of
underground space, e.g. underground malls, underground car parks, underground
roads, and subways. These underground facilities have often been constructed in the
centre of the large cities where the populations and properties are concentrated, leading
to a highly complex and heterogeneous urban environment. The occurrence of rainfall
that exceeds the capacity of the drainage system or the overflow from the river during
extreme rainfall events tends to increase in recent years, resulting in flooding over
urban catchments. If the areas with constructed underground space have been affected
by floods, the inundation would occur and then the flow may enter the underground
space, causing serious damage. Under the worst situations, people in the underground
space may be drowned to death. For example, on 18th July 2007, a devastating urban
flood hit Jinan, China. An underground market was flooded within one hour and 25
people inside the market were killed. Therefore, it is of great significance to study
floods in underground space for disaster prevention and mitigation.
Toda et al. (2009) proposed an integrated urban flood model which can simulate both
surface and underground inundation due to heavy rainfall. A 2D unsteady flow model
with unstructured meshes was adopted for the surface flow simulation. A storage pond
model was utilised for analysing the inundation in underground space. The integrated
model has been applied in reproducing the flood occurred in Ikuta River basin in Kobe
city, Japan. The inundation in one underground mall inside this basin has also been
simulated and the effect of setting steps at the mall entrances has been discussed
thoroughly for the flood risk management in underground space. However, the pond
models applied in the underground component cannot capture the hydrodynamic flow
behaviours inside the underground space, which may compromise the accuracy of the
simulations of fast flow spreading. Son et al. (2016) developed an integrated flood
26
model to simulate overland flow as well as flow inside underground space

simultaneously. Two approaches have been applied in linking the overland and
underground space components in the integrated model: boundary-type and pond-type
approaches. In the boundary-type approach, the surface interface that corresponds to
the exit of underground space was treated as an ‘open’ boundary where surface
inundation water can spread into the underground space. The pond-type approach
treated the underground space as an underground pond and then pond terrain was
imposed at the corresponded surface grids. The size of the pond was estimated
appropriately to make sure that water stored in the underground space cannot flow
back to the surface ground. 1D SWMM and a 2D overland flow model were coupled
by SFM/OFM method in this model to test the effect of the underground space in the
urban flood inundation. This integrated model has been applied in reproducing a flood
event on 21 September 2010 in the Hyoja drainage basin, Seoul, Korea. However, this
research only focuses on the effect of the surface inundation but neglects the
computation of the flow spreading inside underground space, which may not be
sufficient to support the flood risk analysis in the urban underground space. Kim et al.
(2018) proposed an adaptive transfer method (ATM) and imposed it on a 2D overland
flow model to analyse the spreading inundation over two floors inside one building in
Korea. The ATM computed the flow flux passing through the interface of the upper
layer and then transferred them to the lower layer. This approach can provide a
numerical analysis of the dynamic flow in the underground space when the building
is inundated during flood events. However, the computational efficiency is still a
challenge in this approach when being applied to multiple buildings with complex
underground spaces. Herath and Dutta (2004) described the floods that affected the
underground facilities in the last three decades in Japan and proposed a modelling
system to simulate urban floods including flow inside underground space. The
modelling system consists of a 2D diffusive model for surface flood and a 1D diffusive
model for river network flow simulation. The inundation in the underground space
was modelled by an experimental regressing equation that determines the amount of
water entering the underground space. This equation has also been applied in Liu et
al. (2018) to link the urban overland flow and the flow in the underground space. Han
et al. (2019) utilised a 2D flood model to analyse the inundation process in an
underground space and assess the flood risk. The inflow for the underground space
was obtained by assuming different flood risk levels. Wu et al. (2013) applied a
27
smoothed particle hydrodynamics (SPH) model to investigate the evolution process of

a dam break flow over the urban surface ground as well as the underground space.
This work is helpful to explore the interaction between surface flow and the flood flow
in underground space.
Recently an increasing number of researchers have focused on calculating the

inundation depth in the underground metro stations to assess the flood risk of the metro
system. For example, Lyu et al. (2019) proposed an equation to qualitatively predict
the inundation water depth in the metro stations based on the water depth over the
surface grids, drainage capacity of the metro stations and the step height of the metro
stations. Forero-Ortiz et al. (2020) proposed a model to simulate flow in the
underground metro which treated the metro as a special type of drainage system. The
metro network was considered a kind of pipe network and simulated using a 1D model.
The metro pipe model was then coupled with a 1D/2D urban flood model so that the
integrated modelling system can simulate surface, drainage and metro flow during
flood events. The spatial and temporal variability of flood water depth and the
velocities in the underground metro system were adopted to assess flood risk of the
metro network. However, the shape of the metro station is non-prismatic and the 1D
model is not able to represent the irregular DEM variety inside the underground metro
network system. Meanwhile, these methods and models can only roughly predict the
flood inundations in the underground metro space. The complex and dynamic flood
spreading process in the underground space cannot be captured, which is of great
significance for flood risk control and making appropriate evacuation schemes inside
the underground infrastructures during urban flood events.
Research that utilises other probabilistic methods to analyse inundation in the

underground space has also been conducted in recent years. For example, Wu et al.,
(2018) proposed an integrated Bayesian Network (BN) framework for fast evaluating
the flood evolution process in the underground space. The Delphi method was used in
this framework to determine the Bayesian conditional probabilities so that the flood
impact, mitigation measures and other related flood information have been considered,
which helps to manage flood risk in the underground space.
Overall, the existing research on the flood inundation in the urban underground space
faces several issues: (1) there is a lack of the experimental data for model validation
28
(most of the developed models was only validated within the surface component); (2)
they neglect the dynamic flow spreading process inside the underground space, which
is important for decision-makers to produce appropriate evacuation schemes when the
underground space is inundated; (3) only single or a small number of underground
areas have been considered in the urban flood simulations, which manually cut off the
underlying links between surface flows and flows in multiple underground areas and
cannot globally assess the flood risk over the whole urban system including urban
underground space; (4) the low computational efficiency is still a key issue when using
hydrodynamic models to depict flood water movements inside underground space,
hence most of the applications is limited to small areas; (5) the underground space
may have multiple layers, which is neglected in the most models and then the space is
treated as a pond or a one-layer domain.
2.5 Acceleration methods in urban flood modelling
In order to overcome computational constraints in urban flood modelling, different

acceleration methods have been explored. For example, an approximation of full
SWEs neglects the inertial terms, which is formulated based on the assumption that
floodplain generally has slow growth. However, it has been shown that such
simplification of the numerical representation usually leads to inaccurate depths and
velocities in complex topography simulation. Therefore, this method compromises the
temporal accuracy of the solution (Néelz et al., 2009; Smith and Liang, 2013) and
increases sensitivity to model parameterisation (Fewtrell et al., 2011; Liang and Smith,
2015). Moreover, compared to the full SWEs models, simplified approaches could be
computationally less effective when utilising high-resolution meshes (Costabile et al.,
2017).
In recent years, the parallel computing scheme has attracted increasing attention in the
modern computing environment to improve the speed of flood inundation calculation.
The parallel computing methods can be divided into two fundamental types according
to their separate structures: shared memory systems and distributed memory systems
(Sanders et al., 2010). With the multi-core processors being wildly used in scientific
computation, shared-memory parallelism has attracted increasing interest and the
Open multiprocessing (OpenMP) is a framework to implement shared memory
multiprocessor applications. Neal et al. (2009) tested the LISFLOOD-FP
29
hydrodynamic model based on the OpenMP Application Programming Interface (API)

with a series of study sites using different domain sizes to assess the parallel efficiency,
and the results suggested that the parallel speedup was greater in larger-area
simulations. As for distributed memory systems with Message Passing Interface
(MPI), model domains are separated into a number of sub-domains, and each sub-
domain runs simulations within its own processor core. Data exchange occurs when
the result in one processor core is required by another one. Pau and Sanders (2006)
implemented an MPI-based Godunov model which meshed domains to unstructured
quadrilateral cells and used up to six 1.33 GHz AMD processors. The performance is
improved and the method is found to be applicable in different parallel computing
environments.
The techniques above utilise multiple central processing units (CPUs) cores to
implement calculation simultaneously to improve computational speed and hence
belong to multiprocessor parallelisation schemes. An alternative method involves
graphic processing units (GPUs), which are capable of solving a large amount of data
with consuming the same or less calculation time, providing a promising tool for
scientific computation, such as seismic imaging, computational biophysics and fluid
dynamics (Castro et al., 2011; Garland et al., 2008). A GPU-based hardware
architecture is predicted to boost performance by 6.7 times compared with that of a
quad-core CPU device with 64-bit floating-point computation (Smith and Liang,
2013). Benefits have been gained from utilising GPU computing in large-scale flood
simulations with high efficiency. An early attempt was made by Harris et al. (2002)
who reproduced a real-time visual dynamic phenomenon using GPU. It is shown that
the simulation is 25 times faster compared to that with CPU. Lamb et al. (2009)
accelerated the implementation of a well-established 2D diffusion wave (zero-inertial)
model JFLOW based on GPU and the computational time for the simulation was
reduced from 18 hours to 9.5 min. Kalyanapu et al. (2011) established a 2D flood
model based on shallow water equations and utilised GPU approach developed in
CUDA for parallelisation. The results showed that a speedup of 80× to 88× was
gained with domains ranging from 65.5k-1.05M cells. Vacondio et al. (2014)
succeeded in addressing the challenge of implementing GPU computation efficiently
with moving wet/dry front and arbitrarily complicated open and closed boundary
conditions. Smith and Liang (2013) demonstrated the potential of CPU and GPU co-
30
processors to expedite two-dimensional flood simulation. Hu and Song (2018)

proposed a full hydrodynamic model for flash flood simulations over mountain
watersheds based on OpenACC GPU computing techniques. The model has been
applied to a real flash flooding-prone area in China and a higher speedup ratio has
been achieved. Liu et al. (2018) developed a fast simulation model for large-scale
floods based on high-performance GPU computing skills with OpenACC
programming scheme. The model was validated by a real-world case study and the
results demonstrated that the computation can be accelerated to one order of
magnitude compared with other CPU-based models. Liang et al. (2016) conducted a
catchment-scale high-resolution flash flood simulation using GPU acceleration power
and the research showed that the proposed model can complete 12-hour flood
simulation which involved 5 million computational grid cells with 5m resolution
within only 2.5 hours, nearly 5 times faster than the physical time. Hou et al. (2020)
developed a GPU-based numerical model which can simulate flood flow, sediment
transport and morphological changes. The model has been tested by flood events and
the result showed that the proposed GPU-based model improved the reconstruction
speed 2.17-12.83 times compared with the traditional CPU-based model. The works
above demonstrate that GPU parallel computing technique is a promising tool for
improving the speed of model simulation, which is of significance when utilising
hydrodynamic models in large-scale high-resolution urban flood simulations.
In recent years increasing multi-GPU based models have been proposed and applied
to shallow water simulations. Compared with the single-GPU based models, the multi-
GPU based model can overcome the limitation of physical memory and boost faster
simulations involving millions of computation cells. Sætra and Brodtkorb (2012)
proposed a multi-GPU framework which can efficiently solve shallow water equations
with finite volume method. The research shows that communications between GPUs
at a single node is efficient, which can confirm a tight boundary cooperation between
two adjacent subdomains. Morales-Hernández et al. (2021) proposed a multi-
architecture 2D hydrodynamic flood model which can be implemented on multiple
CPUs and GPUs. The research indicates that compared with multi-CPU simulation,
multi-GPU model can achieve a higher computational efficiency in reproducing flood
events with larger spatial and temporal scale. Sharif et al. (2020) developed a 2D flood
model solved by two different numerical schemes within a heterogeneous multi-GPU
31
high performance computing environment. The model has been applied in simulating
a 10-day Hurricane Harvery flood event over a large scale with 5 meter resolution.
More multi-GPU implementations in shallow water or its related hazard simulation
can also be found in Delmas and Soulaïmani (2022), Kalyanapu et al. (2014), Viñas
et al. (2013).
A new High-Performance Integrated Hydrodynamic Modelling System (HiPIMS)

proposed by Xia et al. (2019) has taken advantage of the multiple GPUs computing
technology. This model has been applied to reproduce the flood event induced by 2015
Storm Desmond in 2500 km2 Eden catchment in UK with 5 m resolution. The results
demonstrated that the simulation was more than 2.5 times faster than the physical time.
The model has also been successfully applied in real-time flood forecasting by
coupling it with a numerical weather prediction model (Ming et al., 2020). Therefore,
HiPIMS is supposed to be an effective tool for large-scale urban flood modelling based
on high-resolution data.
2.6 Summary
Based on the literature review presented in this chapter, the research gaps in urban
flood modelling can be identified and summarised as follow:
(1) Junction is one of the indispensable components in the underground drainage

system. The flow inside can provide various types of BCs for pipe flow
calculation. However, current numerical methods for junction flow calculation
suffer from various numerical restrictions. The momentum exchange inside the
junctions is often neglected in traditional methods. Further research is needed
to develop alternative approaches to automatically capture dynamic junction
flow features and couple the junction flow model with pipe flow model to
support accurate and computationally efficient drainage flow modelling for
large-scale real-world applications.
(2) 2D SWEs models are capable of describing the overland flood process and
flood wave propagation for the complex urban geometry formed by buildings,
roads, and houses. As one of the 2D models, HiPIMS has been proven to be an
effective and efficient tool for overland flow simulations. However, it lacks
the capability to calculate the flows in drainage systems. Coupling a robust
32
drainage model with HiPIMS could potentially improve the accuracy of the
simulated results and hence provides stakeholder with more accurate
information.
(3) Increasing areas of underground space have been developed in modern cities,
but flood risk in the underground space is often understood in a less accurate
manner due to the lack of detailed flood simulations over the underground
space. A fully hydrodynamic model should be developed to capture the flood
evolution process along the irregular shape of underground space. It is
necessary to model the interactions not only between the surface ground and
the underground space but also between the different layers in the underground
space to fully assess the flood risk and make appropriate evacuation schemes
when the underground space is inundated.
(4) Cities are highly heterogeneous and have multi-layer spaces such as surface
ground layer, drainage layer and underground space layer. Modelling the
complex and dynamic flood process across these three-layer spaces is however
beyond the capability of most existing models and there is an urgent need for
developing the next generation urban flood modelling system. The new
developed modelling system should be capable of simulating floods over these
three-layer space and capture the dynamic exchange process between these
space layers during flood events.
(5) Low computation efficiency is always an obstacle when applying
hydrodynamic models for urban flood simulations. A GPU-based urban flood
modelling system shows great promise in supporting large-scale urban flood
simulations by taking the advantage of high-performance computing
technology. However, most of the existing GPU-based models only focus on
the simulations of surface flow. It is desirable to develop an efficient modelling
system that can support high-resolution flood simulation over large-scale
multiple-layer urban area.
Motivated by the above knowledge gaps, this work will contribute to different parts
of urban flood modelling in the next four chapters, addressing the aims and objectives
shown in the previous chapter.
33
Chapter 3 A novel 1D-2D coupled model for hydrodynamic simulation of flows in drainage networks
Chapter 3 A novel 1D-2D coupled model for

hydrodynamic simulation of flows in drainage
networks
3.1 Introduction
Drainage network modelling is often an essential component in urban flood prediction

and risk assessment. Drainage network models most commonly use different
numerical procedures to handle flows in pipes and junctions. Numerous numerical
schemes and models of different levels of complexity have been developed and
reported to predict flows in pipes. However, calculation of flow conditions in junctions
has received much less attention and has been traditionally achieved by solving only
the continuity equation. This method is easy to implement but it neglects the
momentum exchange in the junctions and cannot provide sufficient boundary
conditions for pipe calculation. In this work, a novel numerical scheme based on the
finite volume solution to the two-dimensional (2D) shallow water equations (SWEs)
is proposed to calculate flow dynamics in junctions, which directly takes into account
both mass and momentum conservation and removes the necessity of implementing
complicated boundary settings for pipe calculations. This new junction simulation
method is then coupled with the widely used two-component pressure approach (TPA)
for pipe flow calculation, leading to a new integrated drainage network model. The
new 1D-2D coupled drainage network model is validated against an experimental and
several idealised test cases to demonstrate its potential for efficient and stable
simulation of flow dynamics in drainage networks.
3.2 New 1D-2D coupled drainage network model
In this section, the proposed 1D-2D coupled model for simulating transient flows in
drainage networks will be introduced in detail.
3.2.1 Pipe model
In this work, TPA model is adopted for pipe flow calculation. In TPA model, the
pressure term calculation under the pressurised condition is considered by assuming
34
that the flow inside is incompressible, the pipe wall is elastic and can be expanded
when pressure occurs. Therefore, the calculation of the pressure term can be divided
into two situations, free surface condition and pressurised condition, respectively. The
pipe shape in this work is assumed to be circular, which is the most widely used in the
urban drainage system. It is worth mentioning that the model has the potential to be
extended to the cases with different pipe shapes.
To implement a TPA model for calculating transitioning flows in pipes, the Saint-
Venant equations are extended to simulate both free-surface and pressurised flows
(Vasconcelos et al., 2006) and can be written in the matrix form of 1D conservation
laws as:
U P FP
+ = S Pb + S Pf (1)
t x
 0   0 
A   QP    
U P =   , FP =  2  , S Pb =  dz  , S Pf =  P QP QP  (2)
 P
Q  P
Q / A + I  − gA − cD
 dx   A2 
where the subscripts P, b and f respectively represent ‘pipe’, ‘bed’ and ‘friction’; t
denotes the time; x is the longitudinal coordinate along the pipe direction; A is the
cross-sectional area; QP is the flow discharge; z is the bottom elevation of the pipe
above an arbitrary datum; cD = gnP2 RP−1/3 is the roughness coefficient with nP being the
Manning coefficient and RP being the hydraulic radius; P is the wetted perimeter;
and I is the pressure term. Under the free-surface flow conditions, I is normally
calculated by I = pA  with p being the fluid pressure at the centroid of cross-
sectional area and  being the fluid density, which may be expanded to become:
1
I ( ) = 3sin ( / 2 ) − sin 3 ( / 2 ) − 3 ( / 2 ) cos ( / 2 )  gd 3 (3)
24
where g is the gravitational acceleration; d is the pipe diameter and  is the wetted
angle related to the water depth h p :
 = 2 arccos (1 − 2 hp d ) (4)
Related to  , the geometrical variables A and top width T are given by
35
1
A = ( − sin  ) d 2 (5)
8
 
T = d sin   (6)
2
Based on which the gravity wave celerity in a pipe is defined as:
gA gd ( − sin  )
c= =
T 8sin ( / 2 ) (7)
The variables A, T and c are used in the calculation of numerical fluxes, which will
be introduced in more detail in the next section.
2)When the flow is under pressurised flow conditions, a different pressure term related
to the surcharge head can be obtained by assuming elastic pipe wall, and I may be
accordingly estimated using

I (H ) = gd 2 ( H + d / 2 ) (8)
4
in which H is the pressurised head calculated by
a 2  A − AP 
H=   (9)
g  AP 
where a is the acoustic wave speed and Ap is the original cross-sectional area of the
pipe under consideration.
The above 1D TPA governing equations (1)-(2) are numerically solved using a first-
order Godunov-type finite volume scheme. The 1D computational domain (i.e., each
of the pipes in a network) is discretised using uniform grids. In an arbitrary cell i , the
following finite volume time-marching formula is used to update the flow variables
from time level n to n + 1:
t n
Fi +1/2 − Fin-1/2  + t ( S nPbi + S nPfi+1 )
n+1
U Pi = U nPi − (10)
x
in which x is the cell length; t is the time step; Fi+1/2 and Fi−1/2 are the numerical
n n
fluxes across the right and left cell interfaces; S Pbi and S nPfi+1 represents the slope and
n
friction source terms, respectively.
36
Flux terms
In order to update the flow variables to a new time level using Eq. (10), the interface
n n
fluxes ( Fi+1/2 and Fi−1/2 ) must be properly evaluated and an HLL approximate Riemann
solver is adopted in this work:
 FL if S L  0

F =  F if S L  0  S R (11)
F if S R  0
 R
in which FL and FR are numerical fluxes defined using the left and right Riemann states
(i.e. the values of the flow variables reconstructed at the left and right cell interfaces,
which are assumed to be the same as the cell-centre values for a first-order scheme),
and F  is calculated using the HLL flux formula:
S R FL − S L FR + S L S R (U R − U L )
F = (12)
SR − SL
where S L and S R are the left and right characteristic wave speeds calculated by:
SL = min(VL − cL ,V  − c ) , SR = max(V  + c ,VR + c R ) (13)
in which V is the averaged flow velocity defined as V = Qp A and V  is calculated
by:
1 1
V = (VL + VR ) + (L − R ) (14)
2 2
where  is a Riemann invariant relating to  and its approximations are given by

(Sanders and Bradford, 2011):
gd  
L , R   sin  L , R ,  = 6.41 (15)
8  4 
and subsequently
1 1
 = (L + R ) + (VL − VR ) (16)
2 2
When     gd 8 , the flow is under free-surface condition and the intermediate

wave speed c is calculated using
37

gd (  − sin   )
c = (17)
8sin (  / 2 )
with
  
  = 4 arcsin   (18)
  gd / 8 
When     gd 8 , the water surface level may reach the crown of the pipe and the
flow becomes pressurised. The intermediate wave speed c is then set to be the
acoustic wave speed a for the pressurised interface flux computation.
When evaluating the fluxes at the interfaces between pipes and junctions, boundary
values of AB , I B , QB and H B (with subtitle B representing ‘boundary’) must be
obtained before the fluxes can be calculated using the approximate Riemann solver.
The boundary variables can be calculated using the water depth hJ , velocities u and
v at the x- and y-directions in the connected junction according to the following two
cases:
1) If hJ < d, the boundary cell is under free-surface flow condition; AB and I B
are calculated using Eq. (5) and Eq. (3), respectively. The term H B does not
exist in this case and QB will be obtained by projecting the flow rate in the
junction along the normal direction of the pipe.
2) If hJ  d , the pipe flow becomes pressurised; HB is an unknown variable at
the boundary interface, and hence IB cannot be calculated using Eq. (8). A new
approach is proposed herein to estimate the necessary boundary variables.
Based on the hydrostatic pressure assumption, the pressure term I B can be
calculated using
 1 
I B = g  hJ − d  AP (19)
 2 
38
Then H B and AB can be deduced from I B and calculated using Eq. (8) and
Eq. (9), respectively; QB can be obtained in the same way as in Case 1.
The detailed implementation of the pipe boundary calculation at a pipe-junction

interface is summarised in Table 3-1, where np denotes the outward unit normal vector
of the pipe interface; nx and ny are respectively the unit vector along the x- and y-
directions in the local junction coordinate system.
Table 3-1 The pipe-junction boundary calculation method
hJ  d hJ  d
Calculate HB first and then AB using Eq.
AB Eq. (4) and Eq. (5)
(9)
QB (u  (n n ) + v  (n
p x p )
n y )  AB (u  (n n ) + v  (n
p x p )
n y )  AB
Calculate IB first and then HB using Eq.

HB 0
(8)
IB Eq. (3) Eq. (19)
Source terms
To update the flow variables using Eq. (10), it is also necessary to properly discretise
the source terms. The bed slope terms are simply estimated using a central difference
scheme and this will not create any numerical issues as the bed slopes of drainage
pipes are commonly gentle and nearly horizontal in practice. For the friction source
terms, an efficient fully implicit scheme originally developed for the 2D SWEs (Xia
and Liang, 2018) is adopted and modified herein for implementation in the current 1D
TPA governing equations. Only the momentum equation in Eq. (10) contains a non-
zero friction term and needs to be considered, which may be rewritten as
 1 n 
QPin +1 = QPin − t   Fi +n1/2 − Fi −n1/2  − S Pbi n +1
 − tS Pfi (20)
 x 
39
4 7 4 7
n +1
( )
where S Pfi = gnP Pi
2 n 3 QPin+1 QPin+1 ( Ain ) 3 . Defining  = gnP2 ( Pi n ) 3 ( Ain ) 3 , Eq. (20)
becomes:
 1 n 
tQPin+1 QPin+1 + QPin+1 − QPin + t   Fi +n1/2 − Fi −n1/2  − S Pbi =0 (21)
 x 
 1 n 
Further defining B = −QPi + t 
n
 Fi +n1/2 − Fi −n1/2  − S Pbi  , the two sets of possible
 x 
roots of the above quadratic equation are
−1 + 1 − 4tB
QPin+(11) = if QPin+1  0 (22)
2t
−1 − 1 − 4tB
QPin +(12) = if QPin +1  0 (23)
2t
and
−1 + 1 + 4tB
QPin +(13) = if QPin +1  0 (24)
−2t
−1 − 1 + 4tB
QPin +(14) = if QPin +1  0 (25)
−2t
Since   0 is always true for any meaningful cases, both Eqs.(22) and (23) are
n +1
negative if B  0 , which is not consistent with the condition of QPi  0 . Also, Eq.(25)
n +1
is positive when B  0 , which is not consistent with the condition of QPi  0 .
Therefore, Eq. (24) is the only admissible root for B  0 . Similarly, Eq.(22) is the only
admissible root for B  0 . The two acceptable roots, Eq. (22) and Eq. (24), can be then
combined to provide a single analytical solution for Eq.(21), given as follows
−1 + 1 + 4t B
QPin +1 = (26)
−2t Sgn( B)
1 if B  0
where Sgn(•) denotes the sign function, i.e., Sgn( B) =  .
−1 if B  0
4 7
Substituting  = gn Pi
2
P ( ) (A )
n 3
i
n 3
into Eq. (26) leads to
40
4 7
−1 + 1 + 4t B gnP2 ( Pi n ) 3 ( Ain )
−
3
QPin +1 = 4 7
(27)
−2t Sgn( B)n 2
P (P ) ( A )
i
n 3
i
n −3
where B can be easily obtained after solving the governing equations without friction
terms using the adopted finite volume scheme.
7
n
If A is excessively small, A
i ( )i
n −3
may create an extremely small value that exceeds
the machine precision limit and hence cause numerical instability. To effectively avoid
7
this, both the numerator and denominator of Eq. (27) are multiplied by A ( )
i
n 3
and the
n +1
final expression for QP is obtained
7 14 4 7
− ( Ain ) 3 + ( Ain ) 3 + 4t B gnP2 ( Pi n ) 3 ( Ain ) 3
QPin +1 = 4
(28)
−2t Sgn( B)n 2
P (P ) i
n 3
3.2.2 Junction model
Free-surface flow conditions commonly apply when calculating junction flows even
when the water depth in the junction submerges all of the connecting pipes and the
pipe flows are pressurised. In this work, each of the junctions in a drainage system is
idealized as a 2D domain and the flow is subsequently calculated using a model that
solves the fully 2D SWEs to 1) automatically take into account mass and momentum
conservation, and 2) avoid setting complicated BCs for calculating pipe flows. For
example, Figure 3-1 illustrates a schematic diagram for a junction connecting three
pipes. The diameter of each pipe is denoted by di (i = 1, 2, 3); P1 and P2 are assumed
to be inflow pipes while P3 is an outflow pipe. Based on the layout of the inflow and
outflow pipes, the junction domain is approximated using an irregular 2D grid cell as
shown in Figure 3-1(b). On such a grid, a cell-centred finite volume scheme is
implemented to solve the 2D SWEs to predict the flow dynamics in the junction. In
this case, the inflows from the two incoming pipes (P1 and P2) are mixed in the
junction and then discharged into the outflow pipe (P3). During a simulation, the cell
edges connecting the pipes are all defined as ‘open’ boundaries, through which the
inflow and outflow discharges (q1, q2 and q3) from the connecting pipes are obtained
from pipe calculations and imposed as the boundary conditions for the 2D junction
41
flow calculation. The inflow and outflow pipes are automatically defined according to
the flow directions predicted by the pipe model. This essentially defines a two-way
dynamic coupling scheme that links seamlessly the junction model with the pipe
model, effectively avoiding the requirement of any complicated BCs for pipe flow
calculation.
The 2D SWEs describing the free-surface flow in a junction may be written in a matrix
form as
U J FJ G J
+ + = R + S Jb + S Jf (29)
t x y
Figure 3-1 Spatial discretization scheme for a junction.

where the vector terms are given by
 
 uhJ   
 hJ    vhJ
 
U J = uhJ  , FJ = u 2 hJ + ghJ2  , G J = 
1
uvhJ 
 2   2 
 vhJ    1
v hJ + ghJ 
2
 uvhJ 
 2 
    (30)
   
0  0 
Qe   
   zb    
R =  0  , S Jb =  − ghJ  and S Jf =  − bx 
 0  
x
   
 − gh zb    by 
 J
y  − 
  
where the subscript J represents junction; u and v are the depth-averaged velocities
along the x- and y-directions, respectively; FJ and GJ are the flux terms; R, SJb and SJf
contain respectively the mass, slope and friction source terms; Qe is the external unit
42
flow rate;  bx and  by are bed friction stresses calculated by  bx = C f u u 2 + v 2 and
 by = C f v u 2 + v 2 , with C f = gnJ / hJ being the bed roughness coefficient and nJ

2 1/3
being the Manning coefficient at the junction.
When implementing the above junction model, a finite volume scheme is employed to
solve the above 2D SWEs and the resulting time-marching formula is written as
t
U nJ +1 = U nJ − P + t ( R n + S nJb + S nJf+1 ) (31)

To couple with the 1D pipe model, the flux terms FJ and GJ in Eq. (30) have been
revised and the new flux term is denoted as P (see Section 3.2.1);  is the cell area
that is set to be the actual junction area, and hence its value is independent of cell
configuration.
Evaluating the flux terms
As illustrated in Figure 3-1, two different fluxes inside a junction cell are considered:
1) flux across the interface between the junction and the connecting pipes, denoted by
Ppk for the k-th pipe; and 2) no-flow flux at the wall interface, denoted by Pw .
Therefore, the flux vector can be written as
N
P =  Ppk + Pw (32)
k =1
where N is the number of the pipes connected to the junction.
1) Fluxes through a pipe-junction interface

To ensure strict mass and momentum conservation between the 1D pipe model and
the 2D junction model, the fluxes obtained from the 1D TPA calculation are converted
into the local junction coordinate system to derive the numerical fluxes through the
corresponding cell interfaces in the 2D junction model:
1 0 
   F1k 
Ppk = 0 n pk n x  F  (33)
0 n pk n y   2k 
 
43
where npk denotes the outward unit normal vector of the k-th pipe interface;
F =  F1k F2 k  contains the mass and momentum fluxes of ‘Pipe k’ predicted by the
T
1D TPA model.
2) Fluxes at a wall interface

At the junction interface without connecting to a pipe, it is effectively a wall boundary
and no flow is allowed to cross the interface. Subsequently, only the pressure terms in
the momentum equations are effective for flux calculation. A novel approach is
proposed and used to evaluate the pressure terms in this work. A junction connecting
three pipes as illustrated in Figure 3-1 is again used as a demonstrative example.
Considering the fluid/water inside this enclosed domain (i.e. the junction), the final
net hydrostatic pressure adding on the entire enclosed fluid boundary must be
physically integrated to zero. Subsequently considering force balance, the total
hydrostatic force acting on all of the interfaces between the pipes and the junction
must be equal to that imposing on the interface between the surrounding wall
(excluding the pipe areas) and the fluid, but at an opposite direction. The net pressure
forces on the pipe-junction interfaces may be then used to deduce the hydrostatic force
adding on the wall interface so that the fluxes can subsequently be derived and given
by
 
 0 
 
N 
Pw = −   n pk n x I pk  (34)
 k =1 
N 
  n pk n y I pk 
 k =1 
where Ipk denotes the pressure flux at the interface between the k-th pipe and the
junction; when the junction water depth hJ is smaller than the pipe diameter, the pipe
flow is under free-surface conditions and Ipk can be calculated according to Eq. (3);
when hJ rises higher than the crown level of the pipe, pressurised flow occurs, and
Ipk should be computed according to Eq. (19).
Source terms
44
For the source terms in Eq. (31), the mass term R (e.g. rainfall rate) will be calculated
or prescribed. The slope terms are set to be zero since each of the junctions is
approximated as a single cell and the bed elevation is considered to be homogeneous
inside the cell. The fully implicit friction discretisation scheme proposed by Xia and
Liang (2018) is implemented to discretise the friction source terms to ensure stable
simulations when the water depth becomes small.
Stability criteria
Since the finite volume schemes adopted for the 1D pipe model and 2D junction model
are both overall explicit, to ensure the stability of coupled drainage model, the time
step for the final coupled drainage network model is controlled by the CFL condition
and set as the minimum between the time steps of pipe model and junction model:
t = CFL  min(tP , tJ ) (35)
where the CFL number is generally 0  CFL  1 and is set to be 0.5 for all of the
simulations considered in this work; t P and t J are defined as:
  
 min  dx  if Ain  AP

 (
 QPi Ai + a 
 ) 
t P = 
  dx 
min   if Ain  AP


(Q A + ci 
 Pi i ) 
, (36)
 
 i 
t J = min  
(
 ( u )
) ( )
+ ( v ) Ji + ghJ 
2 2
 Ji

Choosing the minimum time step can strictly preserve the stability of the whole
coupled model, but it may suffer computation efficiency when the time step of pipe
model extremely differs from the time step of junction model. Osher and Sanders
(1983) proposed Local Time Stepping (LTS) to solve this problem. LTS involves
different time step for individual grid cell and then synchronized. This method may
reduce the computation amount. LTS has been widely used in SWEs model. For
example, Sanders (2008) developed a robust finite-volume Godunov-type shallow
water model with LTS method to improve computation efficiency. LTS application
45
can also be found in Hu et al. (2019a, 2019b), Li et al. (2020). This method may be an
alternative way to improve model’s efficiency and can be considered in the future.
3.3 Results and discussion
In this section, one experimental and three idealised test cases are simulated to validate
the new drainage model and demonstrate its performance for pipe network simulations.
3.3.1 Experimental test
In order to validate the proposed drainage model for accurate simulation of transitional
flow inside a drainage system, an experimental test case is considered in this section
and the numerical results are compared with the laboratory measurements reported by
Vasconcelos et al. (2006) (VWR experiment), and also the alternative numerical
predictions from the TPA model presented by Vasconcelos et al. (2006) (TPA-VWR)
and another TPA sewer network model proposed by Sanders and Bradford (2011)
(TPA-SB).
Figure 3-2 illustrates the laboratory apparatus, which consists of an acrylic horizontal
pipe connected by two junctions at both ends. The pipe is 14.33 m in length and 9.4
cm in diameter. The upstream junction has a square base of 25 cm side length. The
downstream cylindrical tank is 19 cm in diameter and is supposed to be deep enough
to prevent overflowing. A gate is installed at the downstream end of the pipe to prevent
air from entering the cylindrical junction when the pipe is flooded. A ventilation tower
located just upstream of the gate is also installed to expel air from the pipe when it is
under pressurised condition. During the simulation, the wave speed a is set to 25 m/s
and the Manning coefficient is 0.012 m-1/3s. The pipe is discretised using 20 cells to
give x = 0.7165 m. The simulation begins with the initial water at rest throughout the
whole system, with a depth of 7.3 cm above the pipe invert. A 3.1 L/s flow is imposed
at the upstream junction to create a transient flow into the pipe, which is regulated by
a weir overflow structure integrated into the model as suggested by Sanders and
Bradford (2011).
46
Figure 3-2 Experimental pressurised flow test: experimental apparatus and set up.
(a) (b)
Figure 3-3 Flow variables at x = 9.9 m predicted by the new drainage model, TPA-SB model
and TPA-VWR model, in comparison with the VWR experimental measurements: (a)
velocity; (b) pressure.
Figure 3-3(a) shows the flow velocities at x = 9.9 m (from the left-hand-side edge of
the pipe), in which the numerical predictions from the current drainage model (purple
line), the TPA-VWR model (red line) and the TPA-SB model (blue line) are compared
with the experimental measurements (yellow circle). It is shown that the current model
satisfactorily reproduces the time history of the velocity including peak values and the
results are consistent with the two alternative models (i.e. TPA-SB and TPA-VWR).
Figure 3-3(b) compares the pressure simulated by the three models with measurements
at the same cross-section. All three models produce results that are again consistent
with the experimental measurements although the predictions of pressure surges are
slightly overestimated.
47
(a) (b)
Figure 3-4 Flow variables at x = 9.9 m predicted at different grid resolutions: (a) velocity;
(b) pressure.
(a) (b)
Figure 3-5 Flow variables at x = 9.9 m predicted using different wave speeds: (a) velocity;
(b) pressure.
(a) (b)
Figure 3-6 Flow variables at x = 9.9 m predicted using different Manning coefficients: (a)
velocity; (b) pressure.
48
(a) (b)
Figure 3-7 Flow variables at x = 9.9 m predicted with/without a gate: (a) velocity; (b)
pressure.
Further simulations are carried out to investigate the sensitivity of the simulation
results to relevant model parameters. Figure 3-4 shows the predictions of flow velocity
and pressure at x = 9.9 m using different grid resolutions to discretise the pipe, i.e. N
= 20 and N = 400, respectively, where other model parameters remain the same. The
high-resolution simulation results (N = 400) are in close agreement with the low-
resolution prediction (N = 20) in terms of flow velocity and pressure. Figure 3-5
presents the results obtained using different acoustic wave speeds (i.e., a = 25, 50 and
100 m/s). The predicted velocities are consistent and close to each other for all of the
three selected acoustic wave speeds. However, the pressure produced with a = 100 m/s
presents post-shock oscillations of large magnitude at around t = 8 s. Post-shock
oscillations are commonly observed in the simulations involving mixed flow regimes
using TPA or Pressman slot models due to the existence of discontinuity in the wave
speed. When flow changes from free-surface condition to pressurized condition, the
counterpart acoustic wave increases. As this wave speed goes beyond a certain level,
it makes calculation unstable or even broken. To suppress this numerical oscillation,
several methods have been proposed and can be found in
León, A.S. and Ghidaoui, M.S., 2010. Discussion of “Numerical oscillations in pipe-filling bore
predictions by shock-capturing models” by JG Vasconcelos, SJ Wright, and PL Roe. Journal
of Hydraulic Engineering, 136(6), pp.392-393.
Vasconcelos, J.G., Wright, S.J. and Roe, P.L., 2009. Numerical oscillations in pipe-filling bore
predictions by shock-capturing models. Journal of Hydraulic Engineering, 135(4), pp.296-305.
49
Hydraulic Transient Analysis of Sewer Pipe Systems Using a Non-Oscillatory Two-Component

Pressure Approach
A Mixed Flow Analysis of Sewer Pipes with Different Shapes Using a Non-Oscillatory Two-
Component Pressure Approach (TPA)
For the current case, it is recommended to use a = 25 m/s to reduce the numerical
oscillations in the solution. Figure 3-6 provides further simulation results obtained
using two different Manning coefficients, i.e. 0.012 m-1/3s and 0.02 m-1/3s. The results
show certain level of sensitivity to the Manning coefficient. When increasing the
Manning coefficient from 0.012 m-1/3s to 0.02 m-1/3s, the peak velocity slightly reduces
and there is a small shift change in the temporal profile of the velocity (Figure 3-6 (a)),
which is accordingly reflected in the pressure profile as shown in Figure 3-6(b).
Finally, the effect induced by the gate installed at the downstream end of the pipe is
also investigated, and the results are presented in Figure 3-7. This partially closed gate
may influence the flow hydrodynamics and Sanders and Bradford (2011) suggested to
add a local head loss term to take into account the effect (with the head loss coefficient
set to be 1.25). After incorporating the gate effect, the model produces results that are
compared slightly better with the experimental measurements.
3.3.2 Unsteady flow through different drainage settings
This idealised test is designed to demonstrate the effect of different junction-pipe

settings on the simulation results. Figure 3-8 illustrates a simple drainage system with
two horizontal pipes connecting to two junctions with a radius of 0.5 m and one outfall.
Water inside Junction 1 will flow through Junction 2 and then discharge through the
outfall at the end of Pipe 2. During the simulations, the Manning coefficient in the
whole junction-pipe domain is set to be 0.035 m-1/3s. The pipes are discretised using
uniform grids at 0.5 m resolution. Three cases are considered:
Case 1: Both of the pipes are 6 m long with a diameter of 0.5 m. The upstream
Junction 1 is initialised with different water depths (i.e. 20%, 40%, 60% and
80% of the pipe diameter for free surface flow, and 120%, 140%, 160% and
180% of the pipe diameter for transitional pressurized flow) to generate
different unsteady flows. Water in Junction 1 will be released and move to
connecting pipe.
50
Case 2: The length of the pipes remains to be 6 m but different pipe diameters are
used (i.e. 0.3, 0.4, 0.5 and 0.6 m) to investigate the response of the junction
flow to the change of size ratio between the pipes and junctions. The whole
system is dry initially and an external inflow as given in Figure 3-9 is imposed
at Junction 1. All of the simulations last for 400 s.
Case 3: Different pipe lengths (i.e. 3 m, 6 m and 12 m) are further used to explore the
effect of pipe length on junction flows. The pipe diameter is set to be 0.5 m.
Initial and inflow conditions are set to be the same as Case 2.
Figure 3-8 Configuration of the idealised drainage system.
Figure 3-9 The external flow rate imposed at Junction 1.
Figure 3-10 illustrates the numerical results for Case 1. Figure 3-10(a) and (b) present
the time histories of free-surface water depth and flow rate at the outfall predicted for
different initial depths in Junction 1. With the increase of the initial water depth, the
model predicts higher peaks of both the water depth and flow rate. Figure 3-10(c) and
(d) show the temporal change of water depth and flow rate at the outfall under
51
pressurised flow condition. It is evident that the predicted peaks of both water depth
and flow rate under pressurised condition are much sharper than those produced under
free-surface condition for all of the simulations involving different initial depths. All
of the simulation results are as expected since the higher head at the upstream Junction
1 drives the flow with higher velocity along this simple and straight junction-pipe
system and the pressure in the upstream junction can aggravate this driving force.
(a) (b)
(c) (d)
Figure 3-10 Predicted water depths and flow rates at the outfall for Case 1: (a) water depths
under free surface flow condition; (b) flow rates under free surface flow condition; (c) water
depths under pressurised flow condition; (d) flow rates under pressurised flow condition;
‘hP’ denotes the initial water depth in Junction 1 and ‘d’ refers to the pipe diameter.
Figure 3-11 shows the simulation results in terms of flow depth and velocity in the
two junctions for Case 2, where the pipe diameter varies between 0.3 m and 0.6 m.
Figure 3-11 (a) presents the time histories of water depth in Junction 1. It is observed
that higher water depth in Junction 1 is predicted for smaller pipe diameters, which is
as expected due to the lower discharge capacity for smaller pipes. Figure 3-11 (b) plots
the temporal change of flow velocities in Junction 1. The peak velocity decreases as
pipe diameter increases, which is consistent with the water depth predictions as shown
52
in Figure 3-11 (a). Figure 3-11 (c) and (d) illustrate the predicted water depths and
velocities in Junction 2. In both junctions, the water depth shows a similar shape as
the external flow. Compared with the results in Junction 1, the peak values of the water
depth decrease while the peak velocities increase for the same pipe diameter, which is
again as expected. Intuitively, a smaller pipe diameter will lead to lower drainage
capacity and higher water depth in Junction 1. The higher water depth in turn provides
a larger head difference to drive an unsteady flow with higher momentum in the
downstream connecting pipe (Pipe 1) and junction (Junction 2).
(a) (b)
(c) (d)
Figure 3-11Predicted water depths and velocities in the two junctions for Case 2: (a) water
velocity at Junction 2.
53
(a) (b)
(c) (d)
Figure 3-12 Predicted water depths and velocities in the two junctions for Case 3: (a) water
velocity at Junction 2; ‘PipeL’ denotes pipe length.
Figure 3-12 shows the simulation results for Case 3 where the pipe length changes
between 3 m and 12 m. Figure 3-12(a) and (b) respectively plots the predicted water
depths and flow velocities in Junction 1. The peak water depth in Junction 1 increases
(Figure 3-12(a)) as the pipes become longer but the corresponding peak velocity
decreases (Figure 3-12(b)). The simulation results in Junction 2, as presented in Figure
3-12(c) and (d), are consistent with the results in Junction 1. As the pipe becomes
longer, it takes longer for the flow to reach Junction 2 and the predicted flow velocity
in Junction 2 appears to be more sensitive to the pipe length (Figure 3-12(d)). Since
the drainage system is horizontal, the flow is only affected by the head difference and
friction; driven by the same external flow, longer pipes induce more friction losses
and dissipate more momentum, which subsequently slows down the flow and causes
the water depths inside both junctions to rise and flow velocities to decrease. Overall,
all of the simulation results follow the physical processes of the water flow,
54
demonstrating the capacity of the current model in predicting unsteady flows in a

simple junction-pipe system.
3.3.3 Unsteady flow in V-shape networks
Three V-shape networks with different connecting angles are designed to investigate
the importance of considering momentum exchange in junction flow calculation. As
illustrated in Figure 3-13, two pipes are connected to a common junction at three
different angles, i.e. 180⁰, 120⁰ and 30⁰. Both of the pipes are 10 m long and 2.07 m in
diameter and both of the two junctions have a radius of 4 m. During the simulations,
the Manning coefficient of the whole system is set to 0.035 m-1/3s. The pipes are
discretised using uniform grids of 0.5 m resolution. The initially still water depth in
Junction 1 is set to 1.24 m (60% of the pipe diameter) for free-surface flow simulations
and to 3.31 m (160% of the pipe diameter) for transitional pressurised flow simulations,
respectively. Water in Junction 1 will be released and flow to connecting pipe and
junction and then be discharged through outlet. To quantify the effect of momentum
exchange in junction flow calculation, the predictions with flow velocity inside the
junctions set to zero (i.e. neglecting momentum exchange) are compared with the
simulation results predicted by the current model with momentum conservation fully
captured by the 2D SWE model.
Figure 3-13 V-shape networks of three different connecting angles: 180⁰, 120⁰ and 30⁰.
55
(a) (b)
(c) (d)
Figure 3-14 Predicted water depths and flow rates at the outfalls of the three V-shape
junction-pipe systems: (a) water depths under free surface flow condition; (b) flow rates
under free surface flow condition; (c) water depths under pressurized flow condition; (d)
flow rates under pressurized flow condition.
Figure 3-14 compares the water depths and flow rates predicted by the models with
and without taking into account momentum exchange in the junctions for the V-shape
junction-pipe systems with different connecting angles. For both the free surface flow
and pressurised flow, it is clear that the difference between the peak flow values (i.e.
peak water depth and peak flow rate) predicted by the models with and without
considering momentum transfer becomes more predominant as the connecting angle
increases. This is as expected because an acute connecting angle would cause more
energy loss inside the junction, leading to lower flow velocity/momentum into the
discharging pipe. When the connecting angle reaches 180⁰, the momentum of the flow
from the upstream Pipe 1 will be completely transferred to the middle junction and
then to Pipe 2. The effect of varying connecting angles is evidently captured in the
results produced by the current drainage model. However, the results obtained from
56
the model with zero junction velocity present no difference when the connecting angle
is changed, which is clearly incorrect in practice. This may become particularly
problematic for the simulation of intense rainfall induced flood events in which the
flood hydrodynamics in the drainage networks may be highly transient and can only
be reliably predicted when momentum exchange in junctions is properly taken into
account. Therefore, it is essential to consider momentum conservation in junction flow
calculation to ensure reliable simulation results. This test case effectively confirms
this and demonstrates that the current drainage model can automatically reinforce
momentum conservation in junction calculation.
3.3.4 A hypothetical drainage network system
This final test case is considered to demonstrate the current model’s capability in
simulating flows in a more practical drainage network system. The hypothetical
system consists of 24 pipes, 15 junctions and 2 outfalls that are set up to reflect a
simple but practical urban drainage configuration. As illustrated in Figure 3-15, the
junctions and outfalls have different elevations, creating a slope to allow water to
travel from upstream inflow junction (Junction 1) to the downstream outfalls. The pipe
diameter is 0.5 m, but the lengths of the pipes vary according to the network
configuration, as detailed in Table 3-2. The pipes are discretised using 1 m uniform
grids. The junctions have three different radiuses, i.e. 0.5 m, 0.6 m and 0.75 m, as
detailed in Table 3-3. During the simulation, the Manning coefficient is set to 0.035
m-1/3s over the entire system. An inflow hydrograph as shown in Figure 3-16 is
imposed at Junction 1 to create a flow through the connecting pipes and junctions and
finally discharging through Outfall 1 and Outfall 2.
57
Figure 3-15 The configuration of a hypothetical drainage network.
Figure 3-16 External inflow hydrograph imposing at Junction 1.
(a) (b)
Figure 3-17 Predicted water depths and flow rates at Outfall 1 and Outfall 2: (a) water depth;
(b) flow rate.
58
Table 3-2 Pipe length (m)

Index Length Index Length Index Length Index Length Index Length
1 10 2 10 3 10 4 10 5 10
6 20 7 20 8 10 9 10 10 30
11 30 12 10 13 20 14 20 15 17.32
16 20 17 20 18 20 19 20 20 17.32
21 20 22 20 23 20 24 20
Table 3-3 Junction radius (m)

Index Radius Index Radius Index Radius Index Radius Index Radius
1 0.5 2 0.6 3 0.6 4 0.6 5 0.5
6 0.6 7 0.75 8 0.75 9 0.75 10 0.6
11 0.6 12 0.75 13 0.75 14 0.75 15 0.6
Figure 3-17 presents the simulation results in terms of water depth and flow rate at the
two outfalls. Overall, the time histories of the water depth and flow rate at these two
outfalls are consistent with the inflow hydrograph. Due to the shorter route between
Junction 1 and Outfall 1, the flow arrives earlier in Outfall 1 than in Outfall 2; similarly,
Outfall 1 welcomes the flood peak slightly earlier than Outfall 2. Both of the peak flow
depth and discharge at Outfall 1 are higher than those at Outfall 2. To reach Outfall 2,
the flow must travel longer and more complicated routes that involve more junctions
and pipes, which will potentially lead to more complex flow hydrodynamics involving
more momentum exchange and dissipation, and subsequently lower peaks of water
depth and flow rate. The inflow peaks before t = 1000s and terminates at t = 1800s and
the simulation results clearly reflect the inflow pattern. This indicates accurate
prediction and demonstrates the capability of the current model in predicting flow
hydrodynamics in practical drainage systems involving wet-dry fronts, complex
junction-pipe-outfall connections and dynamic flow transitions.
3.3.5 Discussion
The developed drainage model was validated by one experimental test and three
idealised tests. For the experiment test, the developed model can produce results that
are consistent with the experimental measurements, demonstrating the model’s
capacity in simulation of transitional flow inside a drainage system. In the test of 3.3.2,
drainage flow through different junction-pipe conditions has been simulated. Different
initial water depth in Junction 1 has been simulated to generate different unsteady
59
flows in this junction-pipe system. The results show that with higher initial water depth
(especially when the water depth exceeds pipe diameter and then pressurised condition
occurs), higher water depth and flow rate can be observed in the downstream outfall.
This is due to that higher head at the upstream Junction 1 drives the flow with higher
velocity along this simple and straight junction-pipe system and the pressure in the
upstream junction can aggravate this driving force. When pipe diameter decreases,
water depth and velocities in junctions increases. Smaller pipe diameter will lead to
lower drainage capacity of pipes and more water remaining in the upstream junction.
This provides a larger head difference between junction and pipe and then drive
unsteady flow with higher momentum to the downstream system. When pipe length
increases, hydrographs of water depth in both junctions increase while water velocities
decreases. This is due to that longer pipes lead to more friction loss and then more
momentum dissipation. In 3.3.3, unsteady flow in different V-shape system has been
explored. With shaper connecting angels, lower water depth and velocities can be
observed in downstream outfall. The angle effect is evidently demonstrated in the
simulation results of developed drainage model, which is important in reproducing
intense rainfall induced flood events in which the flood hydrodynamics in the drainage
networks may be highly transient and can only be reliably predicted when momentum
exchange in junctions is properly taken into account. Overall, the developed model
can produce simulation results following the physical processes of the water flow,
demonstrating the capacity of the current model in flood simulation in practical
situations.
3.4 Summary
This chapter has presented a novel 1D-2D coupled model for hydrodynamic
simulation of transient flows in drainage networks. The model adopts a 1D TPA model
to simulate the flow dynamics in pipes, which can automatically capture various
transient flows including free-surface and pressurised flows. For junction calculation,
an innovative approach that treats a junction as a 2D domain is proposed, with the flow
hydrodynamics in the junction calculated using a 2D SWE model. Finally, the two
modelling components are dynamically coupled together to become an integrated
drainage model.
As a summary, the proposed drainage model has the following technical highlights:
60
1. The pipe flow is calculated by a 1D TPA model, which has been improved by
implementing a fully-implicit scheme for discretising the friction source terms
for more accurate and stable simulations.
2. The junction flow is calculated by implementing a novel numerical scheme, in
which junctions are idealised as 2D domains and a 2D SWE model is adopted
to predict the junction flow hydrodynamics. The proposed junction model
automatically takes into account both mass and momentum conservation and
provides complete BCs for pipe flow calculation. When implementing the
junction model, a new approach is also proposed to approximate the
hydrostatic pressure fluxes at the close wall boundaries.
3. The 1D pipe model and 2D junction model are then dynamically coupled
together to develop a new integrated drainage network model.
The new drainage network model is validated against three specifically designed
idealised tests of network configurations and one experimental test. Satisfactory
numerical results are produced. Since the new modelling approach removes the
necessity of any complicated boundary settings, it is thus easy to implement and
potentially more efficient.
61
Chapter 4 A single-GPU based 1D-2D dual-drainage model for urban flood simulations
Chapter 4 A single-GPU based 1D-2D dual-drainage

model for urban flood simulations
4.1 Introduction
With ongoing climate change and rapid urbanisation, pluvial flooding has become one
of the most spreading natural hazards that threaten people’s lives and assets in cities.
An accurate urban flood model is of great significance for understanding the flood
evolution process, which forms the basis of flood risk management and hazard
mitigation. Different types of numerical models have been developed to simulate
urban floods over the past few decades, and one of the widely-used modelling types is
dual-drainage model. One benefit of this kind of models is that it can simulate the
flooding process in the subsurface sewer drainage systems along with the simulation
of the surface flood dynamics so that more reliable predictions can be potentially
achieved. However, the computational burden often overshadows its benefits and may
even prevent large-scale applications. Over the last decade, GPU parallel computing
technique has been applied to flood modelling and it has been proved to be able to
significantly accelerate a flood simulation. In order to take advantage of a high-
performance GPU parallel computing scheme, in this chapter, the drainage model
developed in Chapter 3 is further improved to enable simulations on a single-GPU
device. The accelerated drainage model is then coupled with the GPU-based surface
flood model HiPIMS. An experiment test is considered and presented to verify the
performance of this new dual-drainage model. Finally, the developed dual-drainage
model is applied to simulate a designed rainfall-flooding event in the central part of
Yuhuan County in China, in comparison with an existing flooding model.
4.2 Dual drainage model
In a dual drainage model, there are two main components, namely, surface overland
flow model and drainage model. In the development of a new GPU-accelerated dual
drainage model, the drainage model presented in Chapter 3 is used for simulating
drainage flows, which will be coupled with HiPIMS for predicting surface runoff-
flooding processes. Herein in the rest of this chapter, HiPIMS will be briefly
62
introduced for completion, followed by an introduction of the detailed coupling

process of these two models.
4.2.1 HiPIMS
4.2.1.1 Governing equation
During a flood event, the water depth is commonly much smaller than the horizontal
scale. The full 2D shallow water equations (SWEs), which assumes that a flow has a
wave length much larger than the water depth, can be suitably applied in describing
the hydrodynamics of flood waves. The SWEs are solved in HiPIMS and can be
written in a matrix form as
q f g
+ + = r + sb + s f (37)
t x y
where q, f and g are the vector terms containing conserved flow variables and the
fluxes; r, sb and sf are respectively the rainfall rate, slope and friction terms. These
vector terms are defined as
 uh 
h    vh 
  2 1
q = uh  , f = u h + gh  , g = 
2  uvh 
 2  
 vh  v h + gh 2 
2 2
 uvh 
 
   
    (38)
0  0 
 R − Q e  
z   
r =  0  , sb =  − gh b  and s f =  − bx 

 0  
x
   
 − gh zb    by 
 y  − 
  
where u and v represent the depth-averaged velocities along the x- and y-directions,
respectively; h is the water depth; g is the gravity acceleration; R is the rainfall source
term; Qe is the exchange flow rate between surface and drainage system which will
be calculated in 4.2.3; zb is the bed elevation;  bx and  by are bed friction stresses
calculated by  bx = C f u u 2 + v 2 and  by = C f v u 2 + v 2 where ρ is the water density
and C f = gn2 / h1/3 is the bed roughness coefficient with n being the Manning
coefficient.
63
4.2.1.2 Numerical scheme
The SWEs (37) are solved numerically using a first-order Godunov-type finite volume
scheme, which is considered to be sufficient to capture the complex flow
hydrodynamics on the ground surface (Xia et al., 2019). In this scheme, the time-
marching formula used to update flow variables to the next time step is written as
follow:
qi 1 N
+
t i
 F (q)l
k =1
k k = ri + sbi + s fi (39)
in which the subscript i represent the cell index; k is the index of the cell edge
(Cartesian grids are adopted in HiPIMS and hence N = 4); lk is the length of cell edge
k and i represents the area of cell i; Fk (q) = fk (q)nx + g k (q)ny is the integrated flux
term with nx, ny being defined as the outward normal direction at edge k. With the flux
term F, rainfall term r, slope source term sb calculated explicitly and the friction source
term sf treated implicitly, the final time-marching formula of the adopted finite volume
scheme is written as follow:
t N
q n +1 = q n −
i
 F (q )l
k =1
k
n
k + t (rin + sbin + s nfi+1 ) (40)
where the superscript n denotes the time level and t is the time step.
Calculation of flux and bed slope source terms
The interface fluxes in (40) are evaluated by an HLLC approximate Riemann solver.
The HLLC Riemann solver cannot be directly implemented to preserve the lake at rest
solution, i.e. satisfying the so-called C-property (Bermudez and Vazquez, 1994). A
hydrostatic reconstruction scheme (Audusse et al., 2004) may be adopted to modify
the Riemann states before solving the local Riemann problems. Based on this, Xia et
al. (2017) proposed a surface reconstruction method (SRM) to evaluate Riemann states
across the cell interfaces to: 1) avoid the ‘waterfall effect’ when evaluating Riemann
states over constant slope cells; and 2) effectively handle discontinuous topographic
features (e.g., stairs, road curbs) to maintain the computational stability. The
reconstructed Riemann states then include water elevation  L ,  R ; water depth hL, hR;
and unit-width discharges [hu]L, [hu]R, [hv]L, [hv]R. Herein, the subscripts ‘L’ and ‘R’
represent the left-hand and right-hand sides of the cell edge under consideration. The
64
water elevation is firstly reconstructed to derive other Riemann states. Taking two
adjacent cells ‘i’ and ‘i+1’ as an example, the reconstructed water elevations at the
left-hand and right-hand sides of their common interface are given as follow:
  L = i + max ( 0, min ( bi +1 − bi −  b,i +1 − i ) )

 (41)
 R = i +1 + max ( 0, min ( bi − bi +1 +  b,i − i +1 ) )
in which b is the bed elevation;  b is defined as the difference of bed elevation at the
interface between cells ‘i’ and ‘i+1’ and can be calculated by:
 b = bi +1/2+ − bi+1/2− (42)
where bi +1/2+ and bi +1/2− represent the bed elevation at the right-hand and left-hand
sides of the common cell interface and can be calculated by:
bi +1/2− = bi + 0.5 ( ri ) (bi +1 − bi ) (43)
bi +1/2+ = bi +1 + 0.5 ( ri +1 ) (bi − bi +1 ) (44)
in which  ( r ) is a minmod slope limiter and can be calculated as follows:
b+ − bi
 ( ri ) = max ( 0, min ( ri ,1) ) and ri = (45)
bi − b−
where the subscripts ‘+’ and ‘-’ represent the upstream and downstream cells,
respectively. A detailed explanation of SRM can be found in Xia et al. (2017).
From the reconstructed water elevation  L and  R , Riemann states of water depth can
then be derived as:
 hL = max ( 0, L − b f )

 (46)
hR = max ( 0, R − b f )

in which bf is the single bed elevation at the cell interface and can be defined according
to Audusse et al. (2004) as:
b f = max ( bL , bR )  max ( bi , bi +1 ) (47)
with
65
 bL =  L − hi
 (48)
bR =  R − hi +1
Based on the reconstructed Riemann states of water depth hL and hR, the Riemann
states of unit-width discharges can be further derived as follow:
  hu L = hLui ,  hv L = hLvi
 (49)
 hu R = hRui +1 ,  hv R = hR vi +1
where ui, vi, ui +1 and vi +1 represent the x- and y-direction velocity components at the
centre of cells ‘i’ and ‘i+1’, respectively.
After reconstructing Riemann states at the interface of each cell, HLLC Riemann
solver is implemented to evaluate the fluxes. Based on the character wave speeds
including left wave speed SL, middle wave speed SM and right wave speed SR, four
solution regions can be defined. Taking the x-direction interface fluxes across cells i
and i + 1 ( fi+1/2 ) as an example:
fL 0  SL
f SL  0  S M

fi +1/2 =  L (50)
 f R SM  0  S R

f R SR  0
where fL and fR represent the fluxes at the left and right solution regions. They are
directly calculated according to the left and right Riemann states, e.g. f L = f ( q L ) ,
f R = f ( q R ) ; 𝐟∗𝐿 and 𝐟∗𝑅 are the fluxes in the left and right middle region, which can
be calculated as follow:
 f1   f1 
f R =  f2  f L =  f2  (51)
vR f1  vL f1 
where vL and vR are the reconstructed Riemann states of the tangential velocity
component at either side of the cell interface; f1 and f 2 are the first and second
components of the flux vector f , which can be calculated using the following HLL
formula (Harten et al., 1983):
S Rf L − S Lf R + S L S R ( q R − q L )
f = (52)
SR − SL
66
in which the left characteristic wave speed SL and right wave speed SR can be
calculated as follow:
uR − 2 ghR if hL = 0

SL = 
( )
(53)
 min uL − ghL , u − gh if hL  0

uL − 2 ghL if hR = 0

SR = 
( )
(54)
 min uR − ghR , u − gh if hR  0

in which
1
u = ( uL + uR ) + ghL − ghR (55)
2
2
h =
1 1
g  2
( ghL + ghR + ) 1
4
( uL − uR )

(56)
The middle characteristic wave speed SM can be evaluated by the following formula:
S L hR ( u R − S R ) − S R hL ( u L − S L )
SM = (57)
hR ( u R − S R ) − hL ( uL − S L )
After evaluating the flux terms, the source terms in the SWEs must also be properly
discretised to complete the numerical scheme. Implemented with the SRM, the
following discretisation scheme may be implemented for the slope source terms,
which automatically ensures the C-property of the numerical scheme and handles
discontinuous bed elevations:
 0 
Sbi =  1 
 2 g ( h + h ) (b − b )
1 (58)
 i L ,k i f ,k n k lk 
 i 
in which hL,k indicates the left-hand side Riemann state of water depth at cell edge k,
and b f ,k is the reconstructed face value of the bed elevation and can be calculated
through:
b f = b f − b (59)
 b = max ( 0, b f − i ) if hi +1   h


( )
(60)
b = max 0, min ( b, b f − i ) if hi +1 =  h
67
where  h = 10-10 m is an infinitesimal value to define a dry cell.
Discretisation of the friction source terms

Xia and Liang (2018) proposed a fully implicit scheme for discretising the stiff friction
terms in the SWEs, with two main benefits: 1) it ensures the flow velocity relaxed to
the exact equilibrium state in which the friction and gravity are balanced along the
flow direction in a single time step; and 2) the friction terms can be calculated
explicitly without any iteration process. The momentum components of Eq. (40) can
be expanded to the scalar form as:
7
qxn +1 = qxn + tAx − tgn 2 ( h n ) 3 qxn +1 (q ) + (q )
− n +1 2 n +1 2
x y (61)
7
q yn +1 = q yn + tAy − tgn 2 ( h n ) 3 q yn +1 (q ) + (q )
− n +1 2 n +1 2
x y (62)
where qx = hu and qy = hv are used to represent the unit-width discharges in the x- and
y- direction, respectively; Ax and Ay are the corresponding momentum components of
N
1
−
i
 F (q )l
k =1
k
n
k + sbin . Eqs (61) and (62) can then be rewritten as:
 7

qxn +1 1 + tgn2 ( h n ) 3 (q ) + (q )
− n +1 2 n +1 2
 = qx + tAx
n
x y
(63)
 
 7

q yn +1 1 + tgn2 ( h n ) 3 (q ) + (q )
− n +1 2 n +1 2
 = q y + tAy
n
x y
(64)
 
Dividing (63) by (64) gives
qxn +1 qxn + tAx

= (65)
q yn +1 q yn + tAy
Substituting (65) into (63), we can rearrange the equation to give
 7
 my  n +1 2 
2
q n +1 1 + tgn 2 ( h n )− 3 (q )
n +1 2
+   ( qx ) = mx (66)
x
 x
 mx  
 
with
mx = qxn + tAx (67)
my = q yn + tAy (68)
68
n +1
If qx is positive, Eq. (66) can be written as:
2
7
m 
q n +1
x + (q x )
n +1 2
tgn ( h
2
)
n −3
1 +  y  = mx (69)
 mx 
The roots for Eq. (69) are then
2
7
m 
1 + 1 + 4tgn ( h 2
)
n −3
mx 1 +  y 
n +1  mx 
q x = (70)
2
7
m 
−2tgn 2 ( h )
n −3
1+  y 
 mx 
and
2
7
m 
1 − 1 + 4tgn ( h 2
)
n −3
mx 1 +  y 
n +1  mx 
q x = (71)
2
7
m 
−2tgn 2 ( h )
n −3
1+  y 
 mx 
n +1
However, when solving Eq. (66), qx may be negative, and in this case, the equation
may be rewritten as:
2
7
m 
q n +1
x − (q x )
n +1 2
tgn ( h
2
)
n −3
1 +  y  = mx (72)
 mx 
The roots for Eq. (72) are
2
7
m 
1 + 1 − 4tgn ( h 2
)
n −3
mx 1 +  y 
n +1  mx 
q x = (73)
2
7
m 
2tgn 2 ( h )
n −3
1+  y 
 mx 
and
2
7
m 
1 − 1 − 4tgn ( h 2
)
n −3
mx 1 +  y 
n +1  mx 
q x = (74)
2
7
m 
2tgn 2 ( h )
n −3
1+  y 
 mx 
69
As a result, Eqs. (70), (71), (73) and (74) are the four theoretical roots for (66). If mx
n +1
> 0, qx in (70) is negative, which is not consistent with the prescribed assumption of
n +1 n +1
the positive qx ; qx in Equation (71) is positive, which satisfies the requirements of
n +1
assumption. On the other hand, qx in both (73) and (74) are positive, which
n +1
obviously contradicts the assumption of qx  0 . Therefore, Equation (71) is the only
physically correct root for Equation (66) when mx  0 . Similarly, if mx  0 , Equation

n +1
(74) is the only correct root that is consistent with the assumption of the sign of qx .
The two admissible roots can be combined into one formula as:
7
mx − mx 1 + 4tgn 2 ( h n )
−
n +1
3 mx2 + my2
q x = 7
(75)
−2tgn ( h
2
)
n −3
m +m
2
x
2
y
( )
n −
If hn is excessively small, h 3
in the denominator of (75) may lead to an
exaggerated big value that exceeds the machine precision. To solve this problem, part
7
( )
of h
n −3
is cast into the square root operators and then (75) can be written as:
2
 mx   my 
4 2
mx − mx 1 + 4tgn ( h 2 n −3
)  n  + n 
n +1 h  h 
q x = (76)
2
 mx   my 
4 2
−2tgn 2 ( h n −3
)  hn  +  hn 
   
On the other hand, in practice, it may happen that the denominator of (76) may be
equal to zero or is excessively small, leading to a singular root for (66). To avoid this
situation, (76) is further modified as:
70
 2
 mx   m y 
4 2
 mx , if tgn ( h
2
)
n −3
 h n  +  h n   10
−10
    

 2
 mx   my 
4 2
qxn +1 =  m − m 1 + 4tgn 2 ( h n )− 3
 x x  hn  +  hn  2
    mx   m y 
2
 4
 , if tgn ( h 2 n −3
)  h n  +  h n   10
−10
 4
m   my 
2 2
   
−2tgn 2 ( h n ) 3  nx
−
  +  hn 
 h   
(77)
The roots for (64) can be derived in a similar way and the final expression for q yn +1
can be obtained as follow:
 2
 mx   m y 
4 2
if tgn 2 ( h n )
−
 my , −10
 h n  +  h n   10
3
    

 2
 mx   m y 
4 2
q yn +1 =  m − m 1 + 4tgn 2 ( h n )− 3
 y y  hn  +  hn  2
    mx   my 
2
 4
 , if tgn ( h 2 n −3
)  h n  +  h n   10
−10
 4
m   my 
2 2
   
−2tgn 2 ( h n ) 3  nx
−
  +  hn 
 h   
(78)
4.2.2 Drainage model
The 1D-2D coupled drainage model introduced in Chapter 3 is adopted in this chapter
to couple with the surface flood model HiPIMS to develop a dual drainage model for
urban flood modelling.
4.2.3 Coupling HiPIMS with the drainage model
The HiPIMS is coupled with drainage model through calculating exchange of

discharge at junctions. In the current study, the inlet can be considered as a type of
junction and assumed to be circular. During simulation, water is assumed to enter
junctions averagely, and then the exchange of discharge between surface and drainage
models can be calculated using weir or orifice equations (Chen et al., 2007; Leandro
et al., 2009; Martins et al., 2017; Rubinato et al., 2017). This discharge is defined to
be positive when water flows from surface to drainage system, and vice versa.
71
4.2.3.1 Surface to drainage
The water flows from the surface to drainage systems when junction water level is
smaller than the surface water elevation. In this process, two conditions can be defined:
1) When junction water level is less than the surface bed elevation
In this case, the free weir equation is generally adopted to calculate the exchange
discharge, given as:
2
Qe = Ci DM ( 2 g ) ( hs )
12 32
(79)
3
where Qe represents exchange discharge; Ci denotes the energy loss coefficient; DM
is the junction diameter and hs is the water depth above the surface elevation.
2) When junction water level exceeds surface bed elevation but is less than surface
water elevation
In this case, a submerged weir equation is used to describe flow exchange:
Qe = Ci DM ( 2 g ) hs ( hs + Z crest − h j )
12 12
(80)
where Z crest is the height difference between junction bottom elevation and the
corresponding surface bed elevation; h j is the water depth in the junction. Eq. (80) is
applied to the situation when hs  Am ( DM ) , where AM is the junction area. When
hs  Am ( DM ) , the junction is considered to be fully submerged and a submerged
orifice equation is adopted to describe flow behaviour as follow:
Qe = Ci AM ( 2 g ) (h + Z crest − h j )
12 12
s (81)
It should be noted that in order to keep surface water depth positive, a restriction is
applied here:
Qe  (hs + Z crest − h j ) d s d s dt (82)
72
where d s represents the length of the surface grid cell and dt denotes the current time
step. If (82) is not satisfied, then Qe is set to be equal to the right-hand side of (82).
4.2.3.2 Drainage to surface
When the water level in the junction exceeds the surface water elevation, drainage-to-
surface exchange occurs. In this case, an orifice equation may be used to estimate the
discharge:
Qe = Ci AM ( 2 g ) h j − ( hs + Z crest )

12 12
(83)
Again, a restriction is imposed here to keep the junction water depth positive:
Qe  (h j − Z crest − hs ) AM dt (84)
If Eq. (84) is not satisfied, then Qe is set to be equal to the right-hand side of Eq. (84).
4.3 Implementation on single GPU
Compute Unified Device Architecture (CUDA) is a programming platform developed

by NVIDIA to enable scientific computing on GPUs. CUDA supports object-oriented
programming with stability and efficiency and it is adopted in this work to develop a
GPU-based dual-drainage model.
73
4.3.1 Structure of the drainage model on a single-GPU device
Figure 4-1 Flowchart of the drainage model on a single GPU

A CUDA programme generally includes a series of host and kernel functions. The host
functions run on CPUs (host) while the kernel functions are performed on GPUs
(device). As illustrated in Figure 4-1, the GPU-accelerated drainage model has a
number of host and kernel functions. The CPU codes are written in C/C++ and
designed to read, process and output data while the GPU codes are developed for
parallelisation computing. The programme starts from the input component, which is
basically used for reading the drainage input data into the host, e.g. the features of
pipes and junctions. In the data pre-processing component, all variables and
parameters are copied to the memory on the device. Then pipe flow and junction flow
are calculated simultaneously through executing the GPU kernel functions. In the GPU
parallel part, the boundary conditions for the pipe flow and junction flow are imposed
first before flux calculation within each time step. In this step, 1D arrays are
respectively used to store the data for the pipe and junction cells.
As illustrated in Figure 4-2, each pipe domain can be discretised into a series of grid
cells with different lengths. All the pipe cells are arranged into a long vector with the
74
boundary positions of each pipe recorded. The flux calculation at each face of the pipe
cells is calculated simultaneously, while extra boundary conditions are imposed at the
upstream and downstream pipe boundaries according to the calculations at the
connecting junctions. As for the junction parallel calculation, each junction is
independent from the others and the junction flux calculation is only affected by the
connecting pipes since all of the junctions are connected through pipes and have no
direct connection. When it is necessary to output the simulation results, the results/data
are copied back to the host memory and written in the output files. All of the
calculation parts of the whole drainage model are executed on the device, which avoids
repeating data exchange between the host and device in order to improve the
computation speed.
Figure 4-2 Parallel computing process of pipe and junction cells on GPU
4.3.2 Structure of HiPIMs-drainage coupled model on a single-GPU device
Figure 4-3 shows the flowchart of the coupled HiPIMs-drainage model on a GPU. It
should be noted that when calculating the exchange flow between HiPIMS and the
drainage model, two basic variables are necessary: surface water depth in the next time
step and junction water depth in the current time step. Therefore, HiPIMs is first run
to obtain surface water depth at the next time step. This water depth is subsequently
adopted to calculate the exchange flow rate between the surface and subsurface models.
75
Then the exchange flow result is transferred to HiPIMS for re-updating the water depth
and the drainage model for drainage flow simulation, respectively. The interface
between surface and drainage model is also developed on GPU, which avoids time-
consuming communication between the CPUs and GPUs.
Figure 4-3 Flowchart of the HiPIMs-drainage coupled model on GPU
4.4 Model validation and application
In this section, the proposed 1D-2D dual drainage model is validated against an
experimental test and applied to reproduce a semi-realistic flood event. Refer to XXX,
to have a quantitive comparison between the numerical results and experimental
measurement records, the Nashe-Sutcliffe efficiency (NSE), correlation coefficient R2
and root-meansquare error (RMSE) are used, which are defined as follows:
 (X − X mn )
N n 2
i =1
NSE = 1 −
o
(85)
 (X )
N 2
i =1
n
o − Xo
76
n
in which N is the number of total time steps, X o is the experimental data at the n time
n
step, X m is the model result at the n time step, X o is the mean of experimental data.
2
 

R2 = 
(
i=1  X mn − X m
N
)( X − Xo 
n
o  ) 
 (86)
  (X )  (X ) 
N 2 N 2
n
− Xm n
− Xo
 i =1 m i =1 o

where X m is the average of simulation results.
4.4.1 Experiment Test
The performance of the coupled model is demonstrated through an experiment test

reported by Fraga et al. (2017). The experiment was conducted on a full-scale street
comprising both surface ground and drainage network. As shown in Figure 4-4, the
surface domain consists of pavement and roadway. A basin was built at the upstream
of the street to generate surface runoff and at the offsite side a channel was installed
to collect the lateral outflow. The topography of the surface domain is shown in Figure
4-5. The drainage network, consisting of 7 pipes of two different diameters and 3
junctions, is linked to the surface component to convey the rainfall discharge from
surface to the downstream outlet basin. The pipe network is illustrated in Figure 4-4,
with the properties of the pipes listed in Table 4-1. Two junctions are located at the
roadway and close to the pavement. The third one is at the end of the lateral outflow
channel. Runoff over the surface ground is collected by these three junctions and
conveyed to the pipes. The elevations of the three junctions are provided in Table 4-2.
Five simulations with different rainfall intensities and runoff inflows as listed in Table
4-3 are performed for both the free and pressurised tests. The rainfall is distributed
across the entire surface domain and listed in Table 4-3 with the average values. The
rainfall and runoff were maintained to be constant during the whole experiment. In the
free surface test, Valve 1 was closed while Valves 2 and 3 were open. Water from the
surface domain and channel entered into the sewer network through the three junctions
and drained out via Valve 3. At the pressurised condition, Valves 2 and 3 were closed
and Valve 1 was open at t = 200s. Five sensors (indicated by the red blocks in Figure
4-4) were installed inside Pipes 1-5 to record the water depths and discharges. Gauges
77
3, 4 and 5 were placed at the middle of the pipes while Gauges 1 and 2 were located 4
meters downstream from Junctions 2 and 1, respectively.
During the simulations, the surface and drainage pipe domains are discretised with
uniform grids of 0.06m and 0.1m resolution, respectively. The roughness coefficient
is set according to Fraga et al. (2017) and the CFL = 0.8 is used for both the surface
and drainage models.
Figure 4-4 Geometry of the experimental setup (unit: m)
78
Figure 4-5 Topography of the street surface (the unit of color bar is in meter)
Table 4-1 Properties of the pipes in the drainage network

Pipe Diameter (mm) Slop (%) Length (m)
1 85 0.40 5.21
2 85 3.13 5.21
3 190 1.03 3.71
4 190 0.91 2.75
5 190 5.07 1.70
6 190 1.26 2.55
7 190 1.57 1.70
Table 4-2 Elevations of the junctions in the drainage network

Junction Top elevation (m) Bottom elevation (m)
1 0.515 0.235
2 0.495 0.215
3 0.358 0.134
79
Table 4-3 Five simulations with different rainfall intensity and runoff inflow
Mean rainfall intensity Runoff inflow
Rainfall event
(mm/h) (L/s)
R1 50 0
R2 75 0
R3 90 0
RR1 50 1.4
RR2 75 2.6
Free surface tests
Figure 4-6 - Figure 4-10 compares the numerical results and experiment measurements
at Gauges 1, 2, 3 and 5 during the Rainfall 1, 2, 3 (R1, R2, R3) and Rainfall-Runoff 1,
2 (RR1, RR2) events, respectively. The discharge in Pipe 4 is similar to that in Pipe 5
and is not presented here since both pipes are at the downstream part of this network
and no extra inflow or outflow is added between them. Overall, the coupled model is
found to correctly reproduce the hydrographs obtained from the experiment data. In
R1 and R2 events, the computed discharge is slightly smaller than the experiment data
in Pipe 1 and higher than that in Pipe 2 (Figure 4-6 and Figure 4-7). These two pipes
directly connect to the surface domain through Junctions 1 and 2, suggesting that the
discharges in Pipes 1 and 2 are sensitive to the water inflow from the surface. A small
deviation of the inflow discharge may produce an obvious difference between
numerical results and experimental data. Such difference diminishes at the
downstream part of the drainage network (e.g. Pipe 5).
In RR1 and RR2, with the increase of the surface runoff, more water inflow was
discharged into the drainage network. A reverse flow was observed in Pipe 3 at the
early stage of the experiment. This reverse flow is correctly captured by the current
coupled model. However, since no experimental measurements were recorded after
the reverse flow in Pipe 3, the comparison of discharge is not made at Gauge 3 in RR1
and RR2. This reverse flow did not occur in Pipe 3 in the R1, R2 and R3 events, and
the comparison shows that flow rates simulated from the coupled model are consistent
with the experimental measurements. Overall, the arrival time of the wet front and the
flow rate under the steady-state conditions are predicted correctly at the most gauge
80
points in all five rainfall events, demonstrating the proposed coupled model is able to
depict flow dynamics in the drainage pipe network system. It should be noticed that,
in the experimental data, the duration of the rising limbs of the hydrographs in the
downstream Pipe 5 nearly doubled from those in Pipes 1 and 2. This flood abatement
has been correctly reproduced in the proposed coupled model.
Figure 4-6 The comparison of discharges in Pipes 1, 2, 3 and 5 in R1
81
Figure 4-9 The comparison of discharges in Pipes 2 and 5 in the RR1
82
Figure 4-10 The comparison of discharges in Pipes 1, 2 and 5 in the RR2

Figure 4-11 indicates the time history of water depth in Pipe 4 during R1, R2, R3, RR1
and RR2 simulations. In RR1 and RR2, extra runoff entered the drainage network from
the street surface, which leads to higher water depth at the steady-state condition and
earlier wet front arrival time compared with those in R1, R2 and R3. Overall, a good
agreement between the simulation results and the experiment data is obtained and
observed in Pipe 4, which indicates the accuracy of the coupled model.
To further analyse the simulation results, NSE and R2 have been calculated and listed
in Table 4-4. Most of the NSE values are greater than 0.6, and all the R 2 values are
higher than 0.8, indicating that the simulation results have a good agreement with
experimental measurements.
Table 4-4 NSE and R2 calculation under different situations
Coorsponding
NSE R2
figure
Rainfall 1 Pipe 1 0.412 0.939 Figure 4-6
Rainfall 1 Pipe 2 -2.110 0.888 Figure 4-6
Rainfall 2 Pipe 2 -4.751 0.799 Figure 4-7
83

Rainfall 1 Runoff 1 Pipe 2 0.860 0.860 Figure 4-9
Rainfall 1 water depth at Pipe 4 0.952 0.954 Figure 4-11
Rainfall 1 Runoff 1 water depth at Pipe 4 0.887 0.966 Figure 4-11
84
Figure 4-11 Time history of water depth in Pipe 4 under free surface condition
Pressurised tests
Valves 2 and 3 are closed at t = 200s to create a pressurised condition in the pipe
network. Figure 4-12 presents the time history of water depth under the pressurised
condition in all five rainfall and runoff events. The duration of rising limbs of the water
depth in RR1 and RR2 is dramatically shorter than that in R1, suggesting the extra
runoff from the street surface can cause surcharge conditions at a fast speed in the
drainage system after closing valves. Overall, the coupled model can reproduce the
variation of water depth as well as the moment when the pressurised condition starts
in all five events.
85
Figure 4-12 Water depth evolution in Pipe 5 under pressure condition
4.4.2 A semi-realistic test
A realistic test is further considered, which is located in Yuhuan, a coastal county in

the east of China. As outlined in Figure 4-13, the available drainage-covered domain
is mainly in the central part of this county where the model will be applied to simulate
flood. In this section, a comparison is made in this area under the same rainfall event
between the proposed coupled dual-drainage model and the original HiPIMS which
does not include a drainage modelling component. The drainage network consists of
493 pipes, 495 junctions and 23 outfalls. Pipe length ranges from 5 to 50 m and
diameter ranges from 0.3 to 1.8 m. In the drainage model, each pipe is discretised by
a varied number of grid cells due to the difference in the pipe length. The boundaries
of all outfalls are set to be ‘free’, which means water flows out freely and will not go
back into the drainage system. The surface domain was discretised by 1181×1334 grid
cells at 3m resolution. Since no land-use date is available, a uniform Manning
coefficient is used and set to be 0.035. A very intense rainfall of 200 mm/h is applied
at the first 1200s and the simulation lasts for 4800s.
86
Figure 4-13 DEM of Yuhuan
(a)
(b)
87
(c)
(d)
Figure 4-14 Comparison of the inundation maps at different output times predicted by the
coupled model and HiPIMS
Figure 4-14 compares the inundation maps predicted by the coupled model and
HiPIMS at t = 1200s, 2400s, 3600s and 4800s, respectively. For better visulisation,
inundation under 0.15m is seto to be inexistence. At t = 1200s, inundation starts to
appear on some of the streets but less difference can be observed between coupled
model and HiPIMS. At t = 2400s and 3600s, the inundation and also water depth
increase, and a more obvious distinction between the results predicted by the two
models can now be observed. On certain streets, although the inundation extents
predicted by the coupled model appear to be similar to those produced by HiPIMS,
the coupled model predicts smaller water depth as the surface flood water is drained
away gradually by the drainage network system. In Figure 4-14(d), the difference
between the inundation predicted by the two models becomes very obvious. In this
case, the drainage system is shown to reduce the flood risk for this case. This also
88
indicates that, for urban flood modelling, a dual drainage model may be necessary to
more reliably predict the flooding process and assess flood risk.
To further analyse the interactions between surface model and drainage model, the
four junctions are chosen to record how much surface runoff run into junctions. The
exchange flow rates of these four junctions are shown in Figure 4-15. It can be
observed that during simulation surface runoff continuously flows into these junctions,
indicating that the drainage system fullfils its function of draining water away. All
these hydrographs of junction exchange flow have a peak around T = 1200 s and then
drop down. This is due to that a fixed intense rainfall is applied from the start to T =
1200s and then stops. The variation of these four exchange flow hydrographs are
consistent with rainfall trend. Figure 4-16 shows the comparison of surface inundation
between the proposed coupled model and HiPIMS around these four junctions.
Distinctions can be found that in the proposed model, surface inundation is not serious
and decreases continuously along with simulation time. At T=3600s and 4800s, most
surface inundation drops to below 0.15 m and is not shown in map. However, in
HiPIMS since there is no drainage system to drain surface inundation away, the
inundation at each output time almost keeps the same. Through comparison at these
junctions, drainage capacity can be clarified.
Junction A
Junction B
Junction C
Junction D
(a) location of junctions
89
(b) Junction A (c) Junction B
(d) Junction C (e) Junction D.

Figure 4-15 Exchange flow at Junction A, B, C and D
(a) Coupled model at T=1200s (b) ) HiPIMS at T=1200s
90
(c) Coupled model at T=2400s (d) HiPIMS at T=2400s
(e) Coupled model at T=3600s (f) HiPIMS at T=3600s
(g) Coupled model at T=4800s (h) HiPIMS at T=4800s

Figure 4-16 Surface inundation around zoomed in junctions
The total run time is 1100s and 918s for the coupled model and HiPIMS, respectively,
which is 4.36 and 5.23 times faster than the real time. The computation efficiency has
been substantially improved through GPU implementation. With the advance of the
improved computational efficiency provided by GPU acceleration, it is realistic to
utilise high-resolution data to simulate urban floods with considering the flow
exchange between surface ground and drainage networks. Therefore, the current GPU-
accelerated dual-drainage model may provide an alternative chance to explore the flow
dynamic evolution in the fine grids and support large-scale urban floods simulations.
4.5 Summary
This chapter introduces a GPU-accelerated dual-drainage model for urban flood

modelling, which can be summarised as follows:
1) A drainage model as proposed in Chapter 3 has been implemented using the

CUDA/C++ platform for high-performance computation on a single GPU. A parallel
91
computation scheme is adopted for calculating the flow dynamics in all pipe and
junction grid cells.
2) The GPU-based drainage model has been coupled with HiPIMS by establishing an
interface between the surface and drainage computational components. Weir and
orifice equations have been used for estimating flow exchange between the two
computational domains.
3) The performance of the developed GPU-accelerated model has been demonstrated

through an experimental test case. Satisfactory agreements have been achieved
between the simulation results and the analytical solutions/experiment measurements.
4) The GPU-based surface-drainage coupled model has been applied to a realistic test
case in Yuhuan, China. The coupled model predicts a smaller inundation extent and
also smaller water depth on some of the street surfaces. Being accelerated by the GPU
simulation, the coupled model completes the simulation up to 4.36 times faster than
real time.
92
Chapter 5 Multi-layer simulation of flooding into underground spaces
Chapter 5 Multi-layer simulation of flooding into

underground spaces
5.1 Introduction
In the downtown of large cities, buildings are crowded on the ground, under which
underground facilities such as underground malls and subways are constructed. If this
densely populated area is flooded, flood water may go into underground space, causing
serious damage and even casualties. This makes it important to assess flood risk in
underground space to prevent potential disasters during heavy rainfall events. To
simulate surface flood flow, underground inundation, and drainage flow
simultaneously, the 1D-2D dual drainage model in Chapter 4 is further developed in
this chapter to capture flood flow in the underground space. The model is accelerated
by GPU to improve the computation efficiency and developed for the purpose of
supporting city-scale simulations.
5.2 Multi-layer flood modelling framework
Figure 5-1 demonstrates the concept of the developed multi-layer flood model, in
which the urban space is divided into several layers including surface layer,
underground space layer and drainage system. Correspondingly, the model includes
three main modules and each of them is connected based on the corresponding
interface settings. In the surface module, the surface flow is predicted by the surface
component of HiPIMS. The drainage flow is simulated in the drainage module and the
interaction of these two layers has been demonstrated in Chapter 4. This chapter
focuses on the underground space layer and how it connects to the surface model.
When surface buildings are inundated during floods, the flood water may intrude into
the underground space below. There are some assumptions in the current multi-layer
framework: 1) the buildings in the surface layer are assumed to have ‘solid boundary’,
which means water cannot intrude inside surface buildings; 2) the turbulence caused
by staircases is not considered in the current work.
As shown in Figure 5-1, the green interfaces in the surface layer represent gates or
doors where flood water can flow into the underground space layer if the buildings are
93
inundated around. The green interfaces in the underground space layer represent the
locations of staircases through which flood water enters the underground space. The
underground space layer in the proposed integrated model is a holistic concept which
can be demonstrated through two aspects: the first one is that the underground space
layer can include a large number of building basements rather than a single one. All
basements, underground tunnels or other facility spaces can be represented on the
DEM of the underground space layer, and then flood water can be modelled based on
this DEM. The second aspect is that the underground space may include more than
one level. Nowadays underground spaces like underground metro stations, malls and
car parks may have multiple levels. The integrated model is developed to support flood
simulations at different levels simultaneously. The number of underground space
levels in the model is flexible (e.g. in Figure 5-1 there are two levels in the
underground space) and can be set by users in the developed model. The interfaces in
the current level represent the staircase entrance and are the key settings to connect
the next adjacent level, and inundation water will transfer through these interfaces.
The flexibility of building different levels in the underground space in the model is
consistent with the complex urban underground environment, which is expected to be
more practical when reproducing or forecasting flood inundation in the underground
space.
94
Figure 5-1 Concept of the multi-layer flood model
95
Figure 5-2 The procedure for implementing the multi-layer flood model
As shown in Figure 5-2, the procedure for implementing the proposed multi-layer
flood model can be divided into two main parts: CPU (host) and GPU (device) portion.
96
Before the simulation, input data of different modules and the interface information
need to be prepared, including underground space data, surface data, the interface data
connecting the underground and the surface, drainage data and the interface data
connecting the surface and the drainage system. The input data of the surface HiPIMS
module include surface DEM, precipitation, land use type, infiltration data, manning
data, etc. The underground space flood simulation is also built and driven by the same
input data as the surface model except for precipitation and infiltration. The
precipitation in underground space does not exist and hence is set to be zero in the
model. The underground space is normally set as an impermeable area; therefore, the
infiltration is also neglected in this module. The interface data connecting the
underground and the surface mainly includes the locations of the gates of the buildings
in the surface layer and underground staircases in the underground space layer. The
drainage data includes pipe layout, junction locations and the attributes of both
junctions and pipes. Before the simulation, all these data need to be pre-processed and
integrated in the CPU device and then will be copied from the CPU to GPU device for
parallel computing (Figure 5-2).
After starting the simulation, the surface, underground space and drainage module are
activated simultaneously. In the surface module, the flux advection, precipitation, and
infiltration are calculated among surface grid cells. As explained before, the
underground space simulation module is the same as the surface part except for the
precipitation and infiltration calculation. The water depth at the surface and
underground space interface grid cells are simulated independently, based on which
the exchange flow rate QE between the surface layer and underground space layer can
be calculated according to the weir equations as follow:
2
QE = Ci Pre ( 2 g ) ( hs )
12 32
(87)
3
in which Ci is the energy loss coefficient; Pre denotes the perimeter of the gate surface
cell; g is the gravity acceleration; hs is the water depth. Normally the underground
space is much larger compared to the junctions, hence no overflow from underground
space to surface ground is considered in this model. The water depths in the surface
ground and underground space domain are then updated via adding a source term in
the governing equations separately based on this exchange flow rate QE. The detailed
equations are demonstrated below:
97
The revised governing equations for surface flood flow and flow in the underground
space computation are shown in Eqs.(88) and (89), respectively:
q f g
+ + = r + sb + s f − e (88)
t x y
q f g
+ + = r + sb + s f + e (89)
t x y
in which
QE / dx 2 
 
e= 0  (90)
 0 
 
where dx is the cell size. The other parts of Eqs. (88) and (89) are the same as
governing equation (Eq. (37)) shown in Chapter 4.
Because flood water flows into the underground space layer from the surface layer,
the source term is negative in the surface equation but positive in the underground one.
After calculation, the water depth at the surface layer as well as the underground space
layer can be updated. The drainage flow calculation is introduced in Chapters 3 and 4.
The surface runoff will also enter the drainage system through the junction interfaces
and discharge to the downstream river or water treatment plants.
The three calculation modules operate separately and are connected by the flow
transfer at their according interfaces. The overall calculation procedure will repeat at
each time step until the end of the simulation. An output time interval is set during the
simulation for writing the simulation files. At each output time point, the flow
variables such as water depths, and velocities in the surface domain, underground
space and drainage system will be copied back from GPU device to CPU memory for
post-processing.
Note that there may be more than one level in the underground space, and water then
may enter the downward level through the interfaces between the adjacent two levels.
To simulate water movement in the different levels inside the underground space,
interfaces that connect the two adjacent levels are first identified. Water depths at the
interfaces in upward level are used to determine the exchange flow according to (87).
Then this flow is set as the input for the downward level and the water depths in both
98
levels are updated based on it. The calculation repeats until completing the flood
calculation of the lowest level layer in the underground space.
5.3 Validation of the multi-layer flood model
5.3.1 Flood in an idealised urban catchment with underground space
This test is designed to simulate the urban flood by considering flood flow in the
underground basement. As shown in Figure 5-3(a), the simulation domain is a 50
m×50 m area, which has a 5% slope from Northwest to Southeast. Two streets cross
at the centre of the domain with the width being 5 m. The street is 0.2 m lower than
the adjacent ground surface. A drainage system involving 6 pipes, 6 junctions and 1
outlet has been built to discharge surface runoff. The pipes are 20 m in length and 1 m
in diameter. The junction diameter is 2 m. There are four buildings in this urban
catchment: Buildings 1 and 2 are located at the South-East corner; Building 3 is in
North-East and Building 4 is in the North-West. All of these four buildings have an
underground basement. In Building 1, the underground basement has two levels: -1
underground level and -2 underground level. And there are 4 and 3 rooms on these
two levels, respectively. The simulation is driven by the rainfall given in Figure 5-4.
The initial water depths in the surface layer, drainage system and underground space
are set to be 0 m. The surface and underground space domain are discretised by 0.1
m-resolution grids and each drainage pipe is 1 m. A uniform Manning coefficient of
0.035 sm-1/3 is used in the domain. The surface boundary is set to be ‘rigid’ except for
the southern boundary of the domain. Therefore, surface water will flow along the
slope from the West to the East and from the North to the South but can only be
discharged at the southern boundary. The outlet of the drainage system is set to be
‘open’. The underground space is ‘rigid’, and no water will flow out from this space.
99
(a)
(b)
Figure 5-3 Set-up of the underground building basement test: (a) layout of simulated domain
including surface layer and underground space layer; (b) Layout of the -1 and -2 level under
Building 1
100
Figure 5-4 Rainfall input

In this test, two scenarios are considered:
1) In the first scenario (Scenario a), Building 1 is assumed to have only one level.
Flood is simulated not only on the surface layer but also in the underground
space layer as well as the drainage system. This scenario is designed to
demonstrate the capability of the developed model to simulate flood flows in
multiple underground spaces.
2) In the second scenario (Scenario b), Building 1 is assumed to extend to -2 level
and floods in both the -1 and -2 levels are considered (Figure 5-3(b)). This
scenario is designed to demonstrate the importance of considering multiple
levels in the underground space.
Figure 5-5 presents the simulated surface inundation. When t = 20 s, there is little
inundation on the streets, around buildings or at the east boundary due to the small
rainfall intensity. The inundation on the surface ground becomes more obvious as time
goes on. When t = 60 s, the inundation area is obviously larger compared to the results
of t = 40s, especially at the southeast corner. During this period, rainfall intensity
reaches its peak and starts to drop. From t = 80 s to t =100 s, the surface inundation
decreases and the water depths at the east side and the northeast corner of Building 1
decline. Overall, the change of inundation is consistent with the change of rainfall.
The water depth in the east is generally larger than that in the west or centre, which is
in line with the slope settings. The simulation results are reasonable.
101
Figure 5-5 Surface inundation variation
102
Figure 5-6 Underground space inundation
Figure 5-6 shows the inundation process in the underground space during the
simulation. During the first 20 s, little water has flowed into the underground space
since on the surface ground little runoff is produced. When t = 40 s, in the rooms that
directly connect to the surface through the interfaces, inundation can be observed and
the water depth is around 0.1 m. In the other rooms, the inundation area is still quite
small. When the simulation reaches 60 s, it can be easily found that the water from the
interface rooms starts to flow to the adjacent rooms. At the end of the simulation, the
water depths are distinct for different underground rooms in different buildings. It is
103
obvious that the inundation of Building 1 is the most significant. For other basements,
the lower the elevation of the building, the more significant the inundation.
In modern cities, an increasing number of underground spaces have been built. The
spatial distribution and connection of these underground areas may influence the
flooding processes not only on the ground but also in the underground layer. To have
a direct understanding of this influence, a comparison has been made between the
current scenario and a hypothetic one. In the hypothetic scenario, the underground
basements of Buildings 2, 3 and 4 are nonexistent and only the basement of Building
1 is considered. The rest settings are the same as Scenario 1. Figure 5-7 presents the
comparison of the average water depth of each room in Building 1 in the current
scenario and the hypothetic scenario. Obvious differences can be observed in all four
rooms. The water depth simulated in the current test is smaller than that in the single
underground space scenario. When rainfall occurs, runoff is produced and flows along
the slope. Water could enter other building basements if the surface around the
building is inundated. Buildings 2, 3, and 4 lie at the upstream of Building 1. Hence if
the underground basements exist, flood water may flow into these basements and then
less flood water enters the downstream Building 1 basement. Therefore, it is of great
significance to consider the spatial distribution and connections of multiple
underground spaces when simulating floods over complex urban environment where
multiple underground spaces are existent.
Figure 5-7 Comparison of average water depth of each room of Building 1 in Scenario (a)
and the hypothetical scenario (S(a) represents Scenario (a) and HS is the abbreviation of
hypothetical scenario)
104
-1 underground level -1 underground level -2 underground level

in Scenario (b) in Scenario (a) in Scenario (b)
t = 20 s
t = 40 s
t = 60 s
t = 80 s
t =100 s
Figure 5-8 Flood inundation of -1 underground level in Scenario (a) and (b), inundation of -
2 underground level in Scenario (b)
In Scenario (b), inundations in different underground levels have been simulated. The
effect of considering multiple underground levels on the simulated results has been
investigated by comparing the inundation process of -1 underground level of Building
1 in Scenario (a) and (b), as shown in Figure 5-8. When t = 20 s, surface runoff has
already entered the underground space through the interface. At this time, a small wave
can be observed at -1 level in both Scenarios (a) and (b). Since the interface that
connects the -1 level and -2 level is located at the southwest corner of Room 3 and this
105
room still keeps dry at this time, no inundation is found in -2 level. At t = 40 s, water
in Room 1 rises and starts to flow to the adjacent Room 2 and Room 3. The inundation
area in Room 3 is small and may not reach the interface of -2 level, leaving almost no
water flowing into the -2 level at this time. When simulation runs to 60 s, Room 3 is
totally inundated in -1 level, which leads to water transfer between -1 and -2 levels. A
small inundation area can be observed in Room 1 in -2 level. When t = 80 s, the water
depth in Room 3 in -1 level keeps increasing. Compared with single level simulation
in Scenario (a), the inundation water depth is smaller in Scenario (b). This is because
part of the inundation water has been transferred into the -2 level through the interface.
In -2 level, Room 1 is totally inundated and water starts to flow from Room 1 into
Room 2. At the end of the simulation, the inundation water depth in all four rooms in
-1 level is obviously smaller in Scenario (b) than that in Scenario (a), indicating that a
certain amount of inundation has been transferred into -2 level. In -2 level, water depth
in Room 1 and Room 2 rises to around 0.2 and 0.1 m, respectively. And a wave
propagation can be found in Room 3.
This scenario has shown that flood water in the underground space may transfer
through the connecting interface to different levels, leading to changes in inundation
water depth at different levels. This is a common situation in modern cities since
underground spaces such as car parks, malls and metro may have multiple levels.
Considering this water transfer between multiple levels is more practical and is of
significance in order to make appropriate evacuation schemes when the underground
space is inundated.
5.4 Summary
To simulate the flooding process in the underground spaces, the 1D-2D coupled dual
drainage model developed in Chapter 4 is further extended to deliver an innovative
modelling system for simultaneous simulation of flooding process in urban surface
ground, underground spaces and drainage system. The underground inundation is
calculated by the same algorithm as the surface module but without considering
rainfall or infiltration. Interfaces are established to link surface module, drainage
module and underground space module together. The exchange flow among these
interfaces is calculated by the weir equation.
106
The proposed model is validated by several idealised tests. The results show that there
is a distinct difference when simulating floods with single and multiple underground
basements. Different building layouts and their spatial connection may have an effect
on not only surface flow route but also the underground space inundation. When
considering the upstream building basement, the flooding in the underground space of
the downstream building changes and the flood depth decreases. Also, buildings like
underground malls and car parks may have more than one layer. When considering
multiple layers in these underground buildings, flood processes are different. The
flood water depth decreases since water flows from the current level to the adjacent
downward level. Taking the flooding of multiple levels into consideration is important
to assess the flood risk in these key underground buildings or infrastructure during
flood events.
The proposed model is developed by C++ and CUDA program, which benefits from
GPU acceleration and can significantly improve computational speed. The proposed
model is supposed to support large-scale high-resolution urban flood simulation with
the capability of considering underground space flooding and drainage flow.
107
Chapter 6 Multi-GPU implementation and case study
6.1 Introduction
Flooding is one of the major hazards that take place in many cities around the world
every year. As stated before, flooding can cause severe economic loss and affect
millions of residents in cities. Being a natural hazard, urban flooding cannot be
avoided. However, the losses can be relieved by proper flood mitigation measures. It
is essential to use well-suited flood models to better understand urban flooding process
so that proper emergency planning and risk management strategies can be developed
to mitigate flood losses. 2D hydrodynamic models represent the state-of-the-art in
simulating complex flood hydrodynamics and inundation processes and have been
widely used in predicting different types of flood events. However, the computational
intensity and complexity limit the model applications in large-scale urban flood
simulations. In recent years, high-performance computing technologies have
undergone great progress in improving computing speed and have been applied in
flood simulations. In Chapter 4, a single GPU-enhanced dual drainage model has been
developed for urban flood simulation and applied to simulate the flooding process in
the central part of Yuhuan County, China. However, the single GPU-based model may
still not be able to support extremely large-scale high-resolution urban flood
simulation due to the limitation of physical memory and the still high computational
cost. The high-performance computing application based on multiple GPUs is a
potential choice to overcome the memory issue and can enhance the computation
speed to a greater level. In this chapter, the dual-drainage model introduced in Chapter
4 is further developed to fully take advantage of the acceleration power of multiple
GPUs. The developed multi-GPU based model is firstly validated by comparison with
the single GPU model developed and then applied in a real case study to demonstrate
its capability in reproducing large-scale urban flood events.
6.2 Multi-GPU Architecture
GPUs were originally designed for rending complex 3D scenes in video games and
animations, which involves a large amount of repeating simple parallel computations.
In recent years, the computational power of GPUs has been explored in general-
108
purpose computing including scientific computations. As introduced before, CUDA

released by NVIDIA is one of the widely used programming languages to harness
GPUs for scientific computing and is adopted in this work to develop the proposed
multi-GPU based flood model.
A domain decomposition method is adopted in the HiPIMS surface flood model for
parallel computing across multiple GPUs (Xia et al., 2019). As illustrated in Figure
6-1, the whole surface computation domain is first divided into several subdomains
based on the principle that each surface sub-domain should contain a similar number
of computational cells. Accordingly, the drainage network is then spatially divided
according to the surface domain decomposition. Simulations on these sub-domains are
run separately on different GPUs. For the surface module, an internal boundary layer
is defined at the boundary connecting two neighbouring sub-domains. A ‘ghost
boundary layer’ is set outside the inner boundary layer and denoted as an external
boundary layer. The internal and external boundary layers compose the shared layer
for data exchange between different GPU devices. During the simulation, flow
variables (e.g. flow depth and unit-width discharge) in the computational cells of the
internal layer will be copied to the neighbouring sub-domain external layer so that the
external layer cells can provide sufficient boundary conditions for the calculation. In
this way, all cells in each of the sub-domains except for those in the external layers
can proceed with the calculation to the next time step. It should be noticed that
calculations on the external layer cells are unnecessary since the flow variables in the
external layer cells are only updated by the internal layer cells in the neighbouring
sub-domain to provide boundary conditions.
To develop the multi-GPU based drainage model, a similar domain decomposition

method is implemented. As illustrated in Figure 6-1, ghost computational cells are set
in each sub-domain at the end of the pipes which are cut artificially by surface domain
division and these ghost cells in drainage module are denoted as domain link nodes
(DLNs). Flow variables including water depth, discharge, cross-sectional area in the
pipe boundary cells will be transferred to the DLNs in the neighbouring sub-domain
so that DLNs can provide boundary conditions for pipe calculations. Similar to the
case in the surface external layer cells, these DLNs do not need to perform calculations
and only provide boundary conditions for the connecting pipe cells. In the multi-GPU
based drainage model, the upstream of a pipe can be either a junction or a DLN; and
109
the downstream can be either junction, DLN or outfall. This differs from the single-
GPU based drainage model and therefore both the data structure and logic connection
need to be revised accordingly.
Figure 6-1 Surface domain and drainage network decompositions for multi-GPU
computing
To implement the domain decomposition method in the multi-GPU based drainage

model, the C++ 11 multi-threading library is adopted. The C++ 11 library is part of
the standard template library (STL) and it can be well compatible with CUDA to
launch multi-GPU computing. Figure 6-2 demonstrates the architecture of the multi-
GPU based model. When the simulation starts, the developed multi-GPU based model
will launch multiple threads. The number of threads keeps the same as the number of
available GPU devices. Each thread can control the corresponding GPU device to
conduct the assigned computation task independently. Memories on CPU and GPU
are allocated in advance for each sub-domain and then the necessary input data can be
read and stored separately. At the beginning of the simulation, similar to the single-
GPU based model, all of the flow variables will be copied from CPU to GPU for
paralleling computing. In each thread, the flow variables both in the surface model and
drainage model are calculated and updated toward the next time step and this process
is similar to the single-GPU based model. One of the most important parts of multi-
GPU based model is the data exchange between different GPU devices. To fulfil this
function, unified virtual addressing (UVA) is adopted in this research. UVA provides
a single virtual memory address space for all memories in the system. It can make
pointers to be accessed from the GPU code no matter where they reside in the system.
110
By using UVA, the variables in the external layer cells of each sub-domain are updated
by the flow variables in the neighbouring internal layer cells in the surface model.
Meanwhile, in the drainage model, the DLNs in each sub-domain are updated by the
connecting pipe cells in the neighbouring domains. This process will repeat at each
time step and all of the threads will suspend until the variables in all external layer
cells and DLN have been updated.
Figure 6-2 The Multi-GPU architecture
Note that the multi-GPU based model performs computation separately on each thread.
The computational time step typically differs from one sub-domain to another. To
maintain consistency between different threads during the computation, a criterion
should be made to update the time step to the next round of calculation. To do so, the
smallest time step among the threads is adopted and broadcast to all the sub-domains
as the uniform global time step for the next calculation. By doing so, the time steps
between all sub-domains are synchronised within each step.
111
All of the threads will conduct a similar calculation procedure during the simulation
until the maximum simulation time is reached. Output time intervals are prescribed by
the user and at these time points flow variables like water depth and discharges will
be copied back from the GPUs to CPU for writing the output files for post-processing.
Manuscripts have been prepared to combine the output results of each sub-domain into
a single file.
6.3 Comparison against the single-GPU based model
An idealised test is first used to validate the developed multi-GPU based drainage
model. In this test, the drainage system consists of 5 pipes, 5 junctions and 1 outfall.
The layout is presented in Figure 6-3. All of the pipes have the same properties, that
is, 0.5 m in diameter and 3 m in length. Each pipe is divided into three cells with the
length of each cell being 1m. The radius of junctions is 0.5 m. The whole system is
horizontal, and the manning coefficient of all computational cells is set to be 0.035.
The outfall is set to be ‘rigid’ hence there is no water loss from this drainage system.
Two cases are considered in this test to validate the flow condition under the free
surface and pressurised circumstances, respectively. In the first case, the initial water
depth in Junction 1 is 0.4 m and the rest of the pipes and junctions are set to be ‘dry’
at the beginning. The difference of water elevation between Junction 1 and Pipe 1 will
cause water to flow from Junction 1 to the outfall until the water elevation inside the
whole system reaches an equilibrium. No pressurised condition will be produced in
this case since the initial water depth in Junction 1 is less than the diameter of all pipes.
In the second test, the water depth in Junction 1 and other computational cells are set
to be 1.4 m and 0.5 m, respectively. Such a large elevation difference between Junction
1 and Pipe 1 will introduce pressurised flow in the whole system and this pressure will
propagate back and forth until an equilibrium is reached. To confirm whether the
developed multi-GPU based drainage model can obtain similar simulation results with
the single-GPU based drainage model, a comparison has been made between the
results produced by these two models. In the multi-GPU based model, to be consistent
with the pipe discretisation, the simulation domain is divided into three parts: Domain
0, Domain 1, and Domain 2. Water depth and flow rates are recorded during
simulations with both single-GPU and multi-GPU based models.
112
Figure 6-3 Layout of the drainage network and domain division over the drainage system
(a) (b)
(c) (d)
113
(e) (f)
Figure 6-4 Water depth and flow rate in Pipes 1-5 produced by the single-GPU model and
multi-GPU based model under the free-surface condition
Figure 6-4 shows the comparison results of water depth and flow rate in all 5 pipes for
Test 1. The solid and dot lines represent the simulation results produced by the multi-
GPU based model and single GPU model, respectively. The lines corresponding to the
same pipe overlap, indicating that the results simulated by the two models are identical,
as expected. This verifies that the developed multi-GPU based drainage model is
capable of simulating drainage flow under the free surface condition without
introducing any extra errors. Therefore, the domain decomposition method and the
method adopted to transfer data between different GPUs are verified and confirmed to
accurately fulfill their functions.
(a)
114
(b)
(c)
Figure 6-5 Water depth and flow rate in Pipes 1-5 simulated by the single-GPU model and
multi-GPU based model under pressurised condition
Since the pipes are full under the pressurised condition during the whole simulation in
the second case, only the flow rate is compared in Figure 6-5. In all the domains, the
simulation results obtained from multi-GPU based model coincide precisely with the
results from single-GPU based model.
A multi-GPU model will further improve the computational efficiency for large-scale
simulations, which may be essential for high-resolution flood modelling or forecasting
across a large domain, e.g. an entire city, and therefore widen the model’s application.
The computational performance of the multi-GPU model will be further investigated
and confirmed in a real-world case study in the next section.
115
6.4 Case study (Yuhuan)
6.4.1 Study site
Yuhuan, a coastal county in the East of China, as shown in Figure 6-6, is chosen to be
the study site in this chapter. There are more than 20 rivers in this county and the total
length of these rivers reaches 26.17 km. Yuhuan is vulnerable to flooding from
multiple sources, e.g., coastal flooding, fluvial flooding and pluvial flooding. The
heavy rainfall brought by typhoons is the main cause that triggers flooding. During
1956-2002, 170 typhoons have been recorded to bring 50mm accumulated rainfall and
19 of them brought in more than 100 mm rainfall. July, August and September are the
three months when typhoons most frequently happen in this area. On 6th August 2009,
powerful Typhoon Morakot landed in Yuhuan and induced extremely serious rainfall.
In this event, no breaching of coastal defences was reported and coastal flooding was
therefore not a driver of flooding. A large number of houses, roads, schools and
commercial buildings were flooded during this typhoon event. This flooding also
affected 343.17 km2 crops, 320 thousand people, and caused a direct economic loss of
51 million Yuan RMB. Yuhuan is selected as the case study in this work to reproduce
the Morakot flood. The east side of the computational domain faces the sea and the
other three sides are surrounded by mountains, essentially defining an enclosed
catchment with natural boundaries.
6.4.2 Data and model setup
The necessary data required to support flood simulation includes digital elevation
model (DEM), land-use type, gridded rainfall data, drainage data, etc. Besides,
observed data such as measured flood depth and flow rate are also available for model
validation.
• DEM
The original bare earth terrain data is available at 3 m resolution and the resulting
DEM is shown in Figure 6-6. Buildings are the key characteristics of urban
environment and have a great influence on the flooding dynamics, which are added
onto the bare earth terrain data to improve the DEM as presented in Figure 6-7, with
buildings, roads and rivers clearly shown.
116
Figure 6-6 DEM of the research domain
Figure 6-7 The zoom-in DEM with buildings
• Land-use type
117
The land-use in the study site can be classified into four types: buildings, roads,
grasslands (crops) and water bodies (see Figure 6-8), which are used to assign friction
coefficients for flood modelling. The Manning coefficient adopted in this study is
based on Xing et al. (2019), that is, (1) 0.05 for buildings; (2) 0.02 for roads; (3) 0.03
for grassland; (4) 0.035 for bare grounds and water bodies.
Figure 6-8 Land-use type in the simulation domain
• Precipitation
On 6th August 2009, Typhoon Morakot directly brought heavy rainfall of 453 mm
within 40 h and led to severe urban flooding in Yuhuan. Figure 6-9 shows the rainfall
intensity and accumulated rainfall data of this event. It is clearly seen that rainfall
intensity reached a peak of 44.8 mm/h at around 16 h.
118
Figure 6-9 Rainfall intensity and accumulated rainfall of Typhoon Morakot
• Drainage system
As shown in Figure 6-10, the available drainage system data covers mainly the centre
and north parts of the research domain. From the zoom-in map, it is clear that the pipes
are connected by either junction or outfall at upstream/downstream sides. There are
644 pipes, 640 junctions and 47 outfalls. The drainage information, e.g., pipe diameter,
pipe length, pipe slope, junction coordinates, junction size, junction elevation, junction
maximum depth, outfall coordinates, outfall type as well as topological connections
of pipes, junctions and outfalls, are pre-processed to set up drainage modelling
component.
119
Figure 6-10 Drainage layout in the research domain
• Measured data
River gauge measurements and post-event investigations are useful in evaluating

model performance. In the current study, the maximum water depths measured by
three river gauges are collected from the local authority. As shown in Figure 6-11,
these gauges are located at Qingnantang River, Rongliu River and Gushun River,
respectively.
Figure 6-11 Locations of the river gauges in the simulation domain
120
6.5 Results and discussion
The proposed multi-GPU based model is used to reproduce the flood event induced
by Typhoon Morakot in the research domain with an area of 43.798 km2. The flood
event is first simulated using two GPU devices and the model is validated by
comparing with the available measured data. Then we make another comparison of
the inundation maps between the current model and a more traditional method (i.e.,
the drainage capacity is set as a drainage loss among the computational cells along
roads). Finally, further simulations are run on different numbers of GPUs (e.g. one to
four) to compare the runtime with the single GPU simulation.
6.5.1 Model validation
The simulation to reproduce the Typhoon Morakot flood is carried out on 3m DEM,
involving 5.97 million effective computational grid cells. Zero infiltration is assumed
considering that the soil is often saturated during the rainy season and most of the
surface ground in the city is not permeable. Without considering infiltration, Manning
coefficient is the only parameter which is set according to Section 6.4.2. Since the
entire research domain is surrounded by mountains in the north, east and south and by
the sea in the West, ‘open’ boundary conditions are imposed during all of the
simulations. The initial water depth and flow discharge are both set to be zero in the
surface ground grid cells except for water bodies. The initial water depth in rivers is
set to be 1.5 m according to the local investigation report and the initial flow discharge
in rivers is 0 m/s. The simulation is run for 39 h from10:00 AM on 6th August to 01:00
AM on 8th August 2009.
121
(a) (b)
(c) (d)
(e) (f)
Figure 6-12 Simulated inundation maps: (a) t = 6 hrs; (b) t = 12 hrs; (c) t = 18 hrs; (d) t =
24 hrs; (e) t = 30 hrs; (f) t = 39 hrs
Figure 6-12 shows the inundation maps at t = 6, 12, 18, 24, 30, 39 hrs. The maps only
plot the areas with a water depth deeper than 0.15 m for better visualisation. At t =
6hrs (Figure 6-12(a)), inundation occurs mainly along the roads or at some low-lying
locations. As time increases, the rainfall becomes more intense. At t = 12 hrs (Figure
6-12(b)), more areas have become inundated in the north and south parts of the domain.
Figure 6-12(c) indicates the inundation map at t = 18 hrs when the rainfall intensity
just passed its peak. Inundation extent continues to increase, and the flood depth has
also obviously risen compared with the results at the previous output time. Note that
at this point the centre of the research domain has already been severely affected by
the flood with the water depth generally over 0.3m and even reaching 0.8m in some
places. After the rainfall peak, the flooded area starts to shrink as shown by a smaller
122
extent at t = 24 hrs and 30 hrs, respectively (Figure 6-12(d) and Figure 6-12(e)). Whilst
the flood extent decreases as the rainfall passes its peak, the drainage system also
drains the surface water away from inundated locations gradually. At the end of the
simulation (t = 39 hrs; Figure 6-12(f)), the flood extent has decreased significantly
compared with the previous results although the inundation can still be observed at
some places.
Zone 3
Zone 1
Zone 4
Zone 2
Zone 5
Figure 6-13 Maximum inundation map
Figure 6-13 shows the predicted maximum inundation map over the whole research
domain, with the five most severely affected places marked. Zone 1 is a port with an
estuary. Zone 2 is an estuary with crop fields in the north. The flood is serious across
these two places possibly because they are located at the downstream of the rivers with
low elevations. Since tidal data are not available for this event, coastal flooding due to
high tide or surge is not considered in this case. Zone 3 is the old town where most of
the roads and streets are inundated. According to the information from the local
authority, this place frequently suffered from flooding in the past when heavy rainfall
123
or Typhoon occurred. Zones 4 and 5 are grasslands or crop fields along rivers. They
are inundated seriously, which may be due to the flat and low surface elevation.
Figure 6-14 shows the inundation maps in Zone 3 at different output times. Only a
small inundation area can be observed at t = 6hrs. Inundation extent and water depth
increase when reaching t = 12hrs, especially in the north area of the main road. Water
depth at some locations in the north is even above 1 m at this time. At t = 18 hrs, most
of the roads and streets are flooded. The inundation becomes extremely serious not
only in the north but also on the main street at the centre. According to the report from
the local authority, the flooding disrupted traffic and created chaos, substantially
disturbing people’s life. The water depth tends to be deeper in the area farther away
from the main streets. This may be because the drainage system only exists along the
main streets. When t = 24 hrs and 30 hrs, inundation recedes obviously in the West
town centre in this zone. At the end of the simulation, the surface inundation near the
main street in the West and the overland flow in the East and South around rivers
almost disappears. However, some roads and streets in the north still suffer from the
inundation with a water depth of more than 0.5 m, which can still disturb local traffic.
(a) (b)
124
(c) (d)
(e) (f)
Figure 6-14 Simulated inundation maps in Zone 3: (a) t = 6 hrs; (b) t = 12 hrs; (c) t = 18
hrs; (d) t = 24 hrs; (e) t = 30 hrs; (f) t = 39 hrs
Four outfalls are selected to further analyse the flooding processes in the drainage
system in this zone and their locations are marked in Figure 6-15. Figure 6-16 shows
the water depths and discharges at these outfalls. At each outfall, water depth and
discharge show a similar variation trend. Compared to other outfalls, Outfall gauge 1
is worst affected where the maximum water depth reaches over 1m at around 18 hrs.
Among these outfalls, the area near Outfall gauge 3 is less inundated with water depth
below 0.5m during the typhoon event. Note that all of the presented depth and
discharge hydrographs have two peaks. The first one is around 18 hrs and the second
one is around 27 hrs, consistent with the rainfall profile. As shown in Table 6-1, some
statistic numbers have also been reported. During the whole simulation, 97 junctions
have been pressurised and among them 33 junctions discharged flow from drainage
system to surface ground. At the end of simulation, there is 7.385  106 m3 runoff
125
remained at the surface ground as well as rivers in the research domain; 2.642  106
m3 runoff has entered into drainage system through junctions during simulation and
2.029  106 runoff has been discharged through outfalls.
Figure 6-15 The locations of the outfall gauges
Figure 6-16 Water depth and discharge profiles at the outfall gauges
126
Table 6-1 The statistics of the Yuhuan flood

Junctions that has been pressurised 97
Junction that has discharge from drainage system to surface 33
Runoff remained at the end of simulation 7.385  106 m3
Runoff that has entered into drainage system through junctions 2.642  106 m3
during simulation
Runoff that discharged through outfalls during simulation 2.029  106 m3
To further test the accuracy of the proposed model, another comparison has been made
between the measured data and the simulation results. The river gauges were installed
at Qingnantang River, Rongliu River and Gushun River, respectively, where
maximum water elevations are available during the Typhoon Morakot flood, as listed
in Table 6-2. The predicted maximum water elevations are 3.863 m, 3.125 m and 3.721
m at these three gauges, respectively. The predicted water elevations at both
Qingnantang River gauge and Gushun River gauge are reasonably close to the
investigation records. However, the simulated water elevations at all three gauges are
lower than the records. This may be because water being discharged through outfalls
is set to be disappeared from the research domain in the proposed model. Since the
data of outfalls are not adequate, e.g. from the collected data we cannot judge whether
outfalls connect to a downstream water body or other municipal infrastructures or
pipes. Such outfall setting is an alternative to simulate the capacity of drainage outfalls
when data is insufficient, but it may lead to less surface water flowing into the
river/water bodies. Figure 6-17 presents the variation of water elevation simulated by
the model at the three river gauges. During the first 10 hours, all of the three time
series have a similar variation trend, i.e., dropping dramatically first and then
increasing slightly. This is because in the first 5 hours the precipitation is not intense
and water in the rivers flows to the outside of the research domain and then disappears
in the model. With the increase of the precipitation, water accumulates in the rivers
and the elevation rises. Note that all of the three curves show a similar variation trend
and have two peaks at around 18 hrs and 27 hrs, respectively. This is consistent with
the variation of precipitation.
127
Overall, the simulated results are consistent with the observation data, demonstrating
that the proposed model is capable of simulating urban floods over a city-scale
considering both surface and drainage flows.
Table 6-2 Maximum water elevation at gauge points

Maximum water elevation (m)
Gauge points
Investigation record Model result
Qingnantang River Gauge 3.9 3.853
Rongliu River Gauge 3.3 3.125
Gushun River Gauge 3.8 3.721
Figure 6-17 Water depth at three river gauges
6.5.2 Comparison against more traditional methods
When drainage data is not available in urban flood simulation, a widely used
alternative method is to represent the drainage capacity through a constant infiltration
on the road grid cells to account for urban drainage loss. Simulation is also conducted
based on this method and the drainage loss rate is set to be 13.2 mm/h according to
Xing et al. (2019). Another treatment for a drainage system is to entirely ignore it
during the simulation. This is in consideration that the drainage system is often
paralysed shortly after extremely intense rainfall occurs. Comparisons are made
between these two methods and the proposed model. Figure 6-18 shows the maximum
inundation maps produced by different models.
128
Zone 1 Zone 3
Zone 4
Zone 2
Zone 5
(a) (b)
(c) (d)
(e) (f)
Figure 6-18 Maximum inundation maps (left) and zoomed in maps in Zone 3 and 4
(right): (a) and (b) proposed model; (c) and (d) model with road sewer coefficient; (e) and
(f) model without considering drainage compacity
From Figure 6-18, it is clear that different simulation methods produce different
simulation results in terms of both inundation extent and water depth. Among the
129
zones considered before, Zones 1, 2 and 5 are relatively far away from the drainage
system and hence are less affected by the drainage model settings. In view of this, the
inundation is mainly compared in Zones 3 and 4 and the results are zoomed-in in
Figure 6-18(b), (d) and (f). The flood inundation predicted by the proposed model is
only slightly relieved in Zone 3 compared with that by the other two models. This is
because the drainage system only exists along several main streets in Zone 3, and in
most of this area there is no available drainage data. Hence, drainage flow is only
calculated in a small area in Zone 3. In Zone 4, a clear distinction can be observed
between the results simulated by the proposed and the other two models. The
inundation results from Figure 6-18 (d) and Figure 6-18 (f) are similar, indicating that
the drainage coefficient adopted may not be able to represent the practical drainage
capability in Yuhuan.
6.5.3 Computational efficiency
All of the simulations are run on a GPU server with different numbers of Nvidia Tesla
P100 GPUs and the runtime is listed in Table 6-3. All of the simulations are much
faster than the actual event time. For the single GPU simulation, the required runtime
is 26.93 hrs and is 12.07 hrs faster than real time. The domain is then divided into two
parts to test the required runtime. As shown in Figure 6-19(a), the division is along the
North-South direction. The required runtime in two GPUs-simulation reduces to 18.55
hrs and the computational efficiency is significantly improved compared with the
single GPU simulation. When the domain is divided into three parts (Figure 6-19 (b))
and simulated with three GPU devices, the required runtime further declines to 10.83
hrs. For the four-GPUs simulation (Figure 6-19(c)), the consumed runtime is 9.02 hrs.
In this case, the runtime does not reduce significantly compared with the three-GPU
simulation. This is due to the fact that when the surface domain is divided into four
parts, the drainage system is extremely asymmetrical in each of the sub-domains. Note
that no drainage system is assigned in Domain 0, which does not harmonize with the
other three domains. Therefore, when adopting multi-GPU model for improving the
computation efficiency, it is essential to consider the computational amount in each
sub-domain and assign it as evenly as possible. Overall, the results effectively
demonstrate the potential of the proposed multi-GPU HiPIMS-drainage model for
large-scale high-resolution urban flood simulations.
130
Table 6-3 Runtime required by different simulations

Device Event Runtime
duration
Test 1 1 Nvidia Tesla P100 39 hrs 26.93 hrs
(a) (b)
(c)
Figure 6-19 Domain division in (a) simulation with two GPUs; (b) simulation with three
GPUs; (c) simulation with four GPUs
6.6 Summary
In this chapter, the single-GPU based drainage model is further developed into a model
implemented on multiple GPUs for large-scale urban flood simulations. A domain
decomposition method is used in the drainage model to divide the drainage network
131
into several parts and each part belongs to the according surface sub-domain. In the
modified domain decomposition method, DLN is introduced as a new type of pipe
boundary cell in the drainage model and is set at the connection points between
different sub-domains. Therefore, the upstream node of a pipe can be either a junction
or DLN and the downstream node of a pipe can be a junction, outfall or DLN. The
data structure of pipe cells has been revised accordingly compared to the single-GPU
based model. Flow variables, e.g. water depth, cross-section area and water flow rate
in DLNs will only be used for providing boundary conditions for the connecting pipe
cells and will not participate in computation. DLN cells are updated by the connecting
pipe cells in the neighbouring sub-domain and this process is completed through UVA,
which is an advanced method to promote data transfer between different GPUs. The
developed multi-GPU based drainage model is then coupled with HiPIMS to become
a modelling system which can simulate both urban surface inundation as well as
drainage flow over a large-scale domain.
The modelling system is first validated by an idealised test in which the multi-GPU
based model is compared with the single-GPU based model. The results demonstrate
that the multi-GPU based model can produce exactly the same results as the single-
GPU based model under both free surface and pressurised conditions.
A case study has also been considered to explore the capability of the multi-GPU based
model in large-scale urban flood simulation. The developed modelling system is
applied to predict an extreme flood event caused by the 2009 Typhoon Morakot in
Yuhuan county in China. The developed modelling system successfully reproduces
both the inundation process on the urban surface ground and the drainage flows in the
underground drainage network system. The simulation results are compared with the
local investigation records and reasonable agreements have been achieved. For
simulation of a 39 hrs rainfall event over the 43.798 km2 research domain, the
consumed time is 26.93 hrs, 18.55 hrs, 10.83 hrs and 9.02 hrs for the single GPU, 2-
GPU and 3-GPU and 4-GPU models, respectively. Note that when using 4 GPUs, the
runtime is not dramatically reduced compared with the simulation based on 3-GPU
server, and this is because the drainage system is extremely asymmetrical in the
divided sub-domains and so the computation amount is heterogeneous across different
sub-domains. Overall, this case effectively demonstrates the potential of the multi-
132
GPU modelling system for large-scale high-resolution urban flood simulation and
real-time flood forecasting.
133
Chapter 7 Conclusions and future work
This thesis presented a high-performance urban flood modelling system for city-scale
simulation of urban flooding across multi-layer spaces. A novel hydrodynamic
drainage network model was developed and coupled with HiPIMS to form a dual
drainage model for simulating the interactive flooding process on the urban ground
surface and in drainage networks. Then, a multi-layer model was developed by
extending HiPIMS to include an additional layer for simulating flood flow in
underground spaces. The whole modelling system was accelerated by taking
advantage of the computational power of modern GPUs. This chapter summarises all
of the modelling technologies developed through this thesis, draws key conclusions
and suggests future development and further application of the current modelling
system.
7.1 conclusions
7.1.1 Model description of the proposed urban flood modelling system
7.1.1.1 Drainage model
A novel 1D-2D coupled model has been developed for hydrodynamic simulation of
transient flows in drainage networks. The model adopts a 1D TPA model to simulate
the flow dynamics in pipes, which can effectively capture free-surface and pressurised
transient flows. For junction calculation, an innovative approach that treats a junction
as a 2D domain has been proposed, with the flow hydrodynamics in the junction
calculated using a 2D SWE model to automatically take into account both mass and
momentum conservation. The new coupled drainage model has been validated against
one experimental and several idealised test cases. In one of the test cases, the
numerical predictions were also compared with the results obtained by neglecting
momentum exchange inside junctions to demonstrate the advantage of the new
junction simulation approach. The proposed drainage model provides a potential tool
for accurate simulation of transient flow hydrodynamics in urban drainage systems
and has the following technical highlights:
134
1)The use of 2D SWE model for junction calculation automatically ensures mass and
momentum conservation and removes the necessity of implementing complicated
boundary settings for pipe calculation.
2)With a finite volume shock-capturing scheme implemented in the TPA model for
pipes and an innovative 2D SWE model for junctions, the new drainage model is able
to reliably predict highly transient flow hydrodynamics involving wet-dry fronts,
complex junction-pipe-outfall connections and dynamic transition between free
surface and pressurised flows.
3)Without the necessity of implementing complicated boundary settings for pipe
calculation, the resulting coupled model is potentially easier to implement and
computationally more efficient, providing a robust tool for application in large-scale
drainage modelling.
7.1.1.2 Coupled HiPIMS-drainage model
The developed drainage model was further coupled with HiPIMS to form a dual
drainage urban flood model. The exchange of mass between these two models was
realised at manholes, through which the surface flood flows calculated by the HiPIMS
enter the drainage system via manholes and vice versa. Weir and orifice equations
were adopted to estimate the flow rate exchanged between the two models at the
manholes. Two situations were considered: (a) surface to drainage exchange meaning
that the drainage system still has the available capacity; (b) drainage surcharge at
manholes occurs when the drainage system has reached its maximum capacity and the
surface water can no longer enter the drainage system.
To ensure numerical stability of the integrated model, the time steps of HiPIMS and
the drainage model are separately calculated according to CFL condition and the
smaller one between the two is then adopted as the final time step for the integrated
model. The coupled HiPIMS-drainage model was validated by an experiment test. To
further analyse the simulation results, NSE and R2 have been calculated and listed in
Table 4-4. Most of the NSE values are greater than 0.6, and all the R2 values are higher
than 0.8, indicating that the simulation results have a good agreement with
experimental measurements, demonstrating the potential of the model for more
complex real-world applications.
135
7.1.1.3 Integrated modelling system for simulation of flooding in underground

spaces
HiPIMS has been further extended to become a multi-layer model and coupled with
the new drainage model to compose an integrated urban flood modelling system. This
new multi-layer modelling system can support flood simulations simultaneously over
urban ground surface and in drainage systems as well as underground spaces.
For the underground space modelling, the water flow is calculated by the same 2D
SWE solver as in HiPIMS except for neglecting rainfall and infiltration. The new
multi-layer model has two key features:
1) Compared to most models that only consider flood inundation in a single

underground space (e.g. a single building), the proposed model can simulate flood
flows over multiple underground spaces with different locations and different sizes to
better reflect the realistic urban environment.
2) Interfaces were established not only between the surface layer and the
underground space layer, but also between different levels inside the underground
spaces. Surface flood flow can enter an underground space through the connecting
interfaces and water in the underground space can further flow into downward level
via their staircases. The levels of the underground space can be set according to the
real-world situation.
3) The integrated model can capture the dynamic interactions of the flood flow
between the surface and underground space layers, which is of great significance for
managing flood risk and informing appropriate evacuation schemes to save lives.
7.1.2 Acceleration techniques of the proposed urban flood modelling system
To overcome the computational burden induced by the adopted hydrodynamic

modelling approaches, GPU parallel computing scheme has been adopted and
implemented.
7.1.2.1 Single-GPU based integrated modelling system
A single-GPU based integrated modelling system has been first developed. The
specific contribution of this thesis lies in the use of GPU power to facilitate the
136
computation of the drainage model and the calculations at the interfaces. HiPIMS is
employed for the surface model, where high-performance GPU parallel computing has
been already implemented in previous work.
To be consistent with HiPIMS, the CUDA programming platform is adopted to

develop the code for the drainage model. A parallel computation scheme was
implemented for calculating the flow dynamics in all pipe and junction cells so that
the drainage model can independently perform simulations. The calculations of
exchanging masses between surface and drainage system and between surface and
underground spaces were also realised in the CUDA programme.
7.1.2.2 Multi-GPU based HiPIMS-drainage coupled model
The single-GPU based HiPIMS-drainage coupled model has been further developed
for implementation on multi-GPUs to support larger-scale high-resolution urban flood
simulations. A modified domain decomposition method has been adopted to divide
the drainage system into several parts and each part corresponds to the corresponding
surface sub-domains created in HiPIMS. DLN is introduced in the multi-GPU based
drainage model to store the flow variables (e.g. water depth, cross-section area, water
flow rate) to provide boundary conditions for pipe calculation at the interfaces between
different sub-domains.
The model was applied to reproduce a flood event caused by the 2009 Typhoon
Morakot in the 43.798 km2 Yuhuan county in China. The runtime was 26.93 hrs, 18.55
hrs, 10.83 hrs and 9.02 hrs, respectively, for the 39-hour rainfall event when using a
single GPU, 2 GPUs, 3 GPUs and 4 GPUs, all faster than real time. The simulation
results effectively demonstrate the potential of the multi-GPU coupled HiPIMS-
drainage model for large-scale high-resolution urban flood simulations.
7.2 Future work
7.2.1 Couple a water quality module in the urban flood modelling system
During a flood event, the pollution of the combined sewer overflows and stormwater
discharges may have a severe effect on urban environments and bring health risks to
residents. Developing numerical models to simulate pollutant transportation and
assess stormwater quality in urban areas has received wide attention and become
137
another key component in the urban flood simulation. The proposed urban flood
modelling system has demonstrated its potential for large-scale high-resolution urban
flood simulations. Coupled with a water quality module, the modelling system can be
further extended to simulate stormwater quality and track pollutant transportation
paths along the surface ground and in the drainage system at the city-scale.
7.2.2 Consider more interaction processes between different model

components in multi-level framework
In Chapter 5, a multi-layer model framework was proposed which can support flood
simulations simultaneously over urban ground surface, urban underground spaces and
drainage system. While in the current developed model, surface runoff can only flow
to underground space layer without returning back. Furthermore, drainage system in
the underground space layer is not simulated in the current model. Simulating the
interactions between underground spaces and drainage system is of significance when
the underground spaces are inundated. For example, in a hilly area, a building
basement may connect to a lower ground entrance. During intense floods, the
basement may be flooded by the surcharge from connecting drainage system at the
lower ground entrance as well as the drainage inside. This situation can be commonly
found in simulating floods in urban underground metro system. The interactions
between underground space flow and drainage flow and how water transfers between
each model component should be explored in the future.
7.2.3 Develop different GPU assigning methods in the multi-GPU based model
In the current multi-GPU based model, the GPUs are assigned according to the
division of the surface domain and the division of the drainage domain follows the
same rule. In this way, each GPU is assigned to calculate not only the surface but also
the drainage flows. However, drainage data is always hard to obtain and may only
exist in some of the sub-domains. This current GPU assignment method may lead to
the situation that the total computing amount of surface and drainage is not averaged
across different sub-domains and hence compromise the computation speed of the
whole model. More efficient GPU assigning methods can be developed to overcome
this issue. For example, when the drainage system data is not sufficient and may only
concentrate on a certain piece of the urban area, one GPU can be assigned to conduct
138
the drainage flow calculation and the rest of GPUs can be used for the surface flow
simulation.
7.2.4 Flood forecasting based on the proposed urban flood modelling system
Flood forecasting provides crucial information for decision-makers and at-risk

population to prepare for the coming flood hazard and reduce the potential damage.
The traditional flood forecasting systems are mostly based on hydrological models,
which cannot capture the flow dynamics over the heterogeneous urban environment.
The traditional hydrodynamic models are restricted by low computation efficiency.
The proposed GPU-based hydrodynamic modelling systems can effectively overcome
this issue and therefore provide a potential tool for urban flood forecasting. When
coupled with the weather prediction products, the multi-GPU modelling system may
be able to support real-time urban flood forecasting at a high resolution.
7.2.5 Apply the modelling system to assess the performance of blue-green

infrastructures
In recent years, more sustainable flood management approaches such as blue/green

infrastructure (BGI) have been widely promoted to manage flood risk. BGI measures
may include natural and man-made wetlands, restored floodplains, green roofs,
bioswales, etc. They can not only contribute to reducing flood risk but also providing
additional benefits such as controlling water pollutants inside the urban water cycle
and improving the capacity of recovering from flood hazards. However, evidence
demonstrates that the BGI is mainly successfully applied in laboratories or small-scale
urban communities (Thorne et al., 2018). Similarly, modelling and assessing the
performance of BGI measures have been mostly done at small scales. Therefore, the
applicability of the BGI for city-scale implementation is still an open question. To
holistically evaluate the performance of BGI, well-suited flood models must be
applied to understand flooding process and assess the flood risk after the installation
of BGI in cities. The models must be able to simulate BGI measures at an adequate
resolution to capture their interaction with flood hydrodynamics to quantify the flood
risk. The models should be able to capture the different types of water flows (e.g.
subcritical, transcritical and supercritical flow) induced by intense rainfall or over-
topping/breaching of riverbanks. The proposed modelling system can satisfy all these
139
requirements and can potentially support the performance assessment of BGI

measures.
7.2.6 Apply the modelling system in the multi-objective optimisation of

drainage network rehabilitation measures
Drainage network system plays an important role in discharging surface water during
a flood event. The increase in rain intensity may lead to insufficient capability of
existing drainage network systems that were designed and built several decades ago.
This problem may be mitigated by rehabilitating the drainage network through
different reconstruction strategies. These strategies include using pipes with greater
diameter to replace failed ones, installing pumps and sewer tanks, etc. However, all of
these strategies are expensive, and the economic cost should be considered when
choosing reconstruction measures. Another factor that should not be ignored is the
performance of these measures. This introduces a new problem about how to choose
the appropriate rehabilitation measures to improve the drainage capacity to reduce
flood risk in urban areas. When coupled with a multi-objective optimisation model,
the proposed urban flood modelling system can be applied in analysing different
drainage rehabilitation measures and help to find the optimised combination of these
strategies.
140
List of Publications
Journal papers
Q. Li, Q. Liang, X. Xia (2020). A novel 1D-2D coupled model for hydrodynamic
simulation of flows in drainage networks, Advances in Water Resources, 137, 103519
Conference papers
Q. Li, X. Xia, Q. Liang, W. Xiao (2018). Drainage network modelling with a novel
algorithm for junction calculation. 13th International Conference on Hydroinformatics,
Palermo, Italy.
141
References
Abiodun, B.J., Adegoke, J., Abatan, A.A., Ibe, C.A., Egbebiyi, T.S., Engelbrecht, F.,
Pinto, I., 2017. Potential impacts of climate change on extreme precipitation over
four African coastal cities. Clim. Change 143. https://doi.org/10.1007/s10584-
017-2001-5
Ahamed, S.M.F., Agarwal, S., 2019. Urban flood modeling and management using
SWMM for new R.R. pet region, Vijayawada, India. Int. J. Recent Technol. Eng.
7.
Ajiwibowo, H., Pratama, M.B., 2020. Hydrodynamic changes impacted by the

waterway capital dredging in cikarang bekasi laut channel, west java, indonesia.
Water Pract. Technol. 15. https://doi.org/10.2166/wpt.2020.032
Alcrudo, F., Garcia‐Navarro, P., 1993. A high‐resolution Godunov‐type scheme in

finite volumes for the 2D shallow‐water equations. Int. J. Numer. Methods Fluids
16. https://doi.org/10.1002/fld.1650160604
Alias, N.A., Liang, Q., Kesserwani, G., 2011. A Godunov-type scheme for modelling
1D channel flow with varying width and topography. Comput. Fluids 46.
https://doi.org/10.1016/j.compfluid.2010.12.009
An, H., Lee, S., Noh, S.J., Kim, Y., Noh, J., 2018. Hybrid numerical scheme of
Preissmann slot model for transient mixed flows. Water (Switzerland) 10.
https://doi.org/10.3390/w10070899
Anees, M.T., Abdullah, K., Nawawi, M.N.M., Ab Rahman, N.N.N., Piah, A.R.M.,
Zakaria, N.A., Syakir, M.I., Mohd. Omar, A.K., 2016. Numerical modeling
techniques for flood analysis. J. African Earth Sci.
https://doi.org/10.1016/j.jafrearsci.2016.10.001
Aredo, M.R., Hatiye, S.D., Pingale, S.M., 2021. Impact of land use/land cover change
on stream flow in the Shaya catchment of Ethiopia using the MIKE SHE model.
Arab. J. Geosci. 14. https://doi.org/10.1007/s12517-021-06447-2
Audusse, E., Bouchut, F., Bristeau, M.O., Klein, R., Perthame, B., 2004. A fast and
stable well-balanced scheme with hydrostatic reconstruction for shallow water
142
flows. SIAM J. Sci. Comput. 25. https://doi.org/10.1137/S1064827503431090
Barco, J., Wong, K.M., Stenstrom, M.K., 2009. Closure to “Automatic Calibration of
the US EPA SWMM Model for a Large Urban Catchment” by Janet Barco,
Kenneth M. Wong, and Michael K. Stenstrom. J. Hydraul. Eng. 135.
https://doi.org/10.1061/(asce)hy.1943-7900.0000121
Bates, P.D., De Roo, A.P.J., 2000. A simple raster-based model for flood inundation
simulation. J. Hydrol. 236. https://doi.org/10.1016/S0022-1694(00)00278-X
Beg, M.N.A., Carvalho, R.F., Leandro, J., 2018a. Effect of surcharge on gully-
manhole flow. J. Hydro-Environment Res. 19.
https://doi.org/10.1016/j.jher.2017.08.003
Beg, M.N.A., Carvalho, R.F., Tait, S., Brevis, W., Rubinato, M., Schellart, A.,
Leandro, J., 2018b. A comparative study of manhole hydraulics using
stereoscopic PIV and different RANS models. Water Sci. Technol. 2017.
https://doi.org/10.2166/wst.2018.089
Bermúdez, A., López, X., Vázquez-Cendón, M.E., 2017. Treating network junctions
in finite volume solution of transient gas flow models. J. Comput. Phys. 344.
https://doi.org/10.1016/j.jcp.2017.04.066
Bermudez, A., Vazquez, M.E., 1994. Upwind methods for hyperbolic conservation
laws with source terms. Comput. Fluids 23. https://doi.org/10.1016/0045-
7930(94)90004-3
Bi, S., Song, L., Zhou, J., Xing, L., Huang, G., Huang, M., Kursan, S., 2015. Two-
dimensional shallow water flow modeling based on an improved unstructured
finite volume algorithm. Adv. Mech. Eng. 7.
https://doi.org/10.1177/1687814015598181
Bisht, D.S., Chatterjee, C., Kalakoti, S., Upadhyay, P., Sahoo, M., Panda, A., 2016.
Modeling urban floods and drainage using SWMM and MIKE URBAN: a case
study. Nat. Hazards 84. https://doi.org/10.1007/s11069-016-2455-1
Blake, E.S., Kimberlain, T.B., Berg, R.J., Cangialosi, J.P., Beven, J.L., 2013. Tropical
Cyclone Report Hurricane Sandy. J. Chem. Inf. Model. 53.
Borsche, R., Klar, A., 2014. Flooding in urban drainage systems: Coupling hyperbolic
143
conservation laws for sewer systems and surface flow. Int. J. Numer. Methods
Fluids 76. https://doi.org/10.1002/fld.3957
Bousso, S., Daynou, M., Fuamba, M., 2013. Numerical Modeling of Mixed Flows in
Storm Water Systems: Critical Review of Literature. J. Hydraul. Eng. 139.
Bradbrook, K., 2006. JFLOW: A multiscale two-dimensional dynamic flood model.

Water Environ. J. 20. https://doi.org/10.1111/j.1747-6593.2005.00011.x
Bradbrook, K., Waller, S., Morris, D., 2005. National floodplain mapping: Datasets
and methods - 160,000 km in 12 months. Nat. Hazards 36.
https://doi.org/10.1007/s11069-004-4544-9
Bradford, S.F., Sanders, B.F., 2002. Finite-Volume Model for Shallow-Water

Flooding of Arbitrary Topography. J. Hydraul. Eng. 128.
https://doi.org/10.1061/(asce)0733-9429(2002)128:3(289)
Broekhuizen, I., Muthanna, T.M., Leonhardt, G., Viklander, M., 2019. Urban drainage
models for green areas: Structural differences and their effects on simulated
runoff. J. Hydrol. X 5. https://doi.org/10.1016/j.hydroa.2019.100044
Bulti, D.T., Abebe, B.G., 2020. A review of flood modeling methods for urban pluvial
flood application. Model. Earth Syst. Environ. https://doi.org/10.1007/s40808-
020-00803-z
Burger, G., Sitzenfrei, R., Kleidorfer, M., Rauch, W., 2014. Parallel flow routing in
SWMM 5. Environ. Model. Softw. 53.
https://doi.org/10.1016/j.envsoft.2013.11.002
Caleffi, V., Valiani, A., Zanni, A., 2003. Finite volume method for simulating extreme
flood events in natural channels. J. Hydraul. Res. 41.
https://doi.org/10.1080/00221680309499959
Capart, H., Bogaerts, C., Kevers-Leclercq, J., Zech, Y., 1999. Robust numerical
treatment of flow transitions at drainage pipe boundaries. Water Sci. Technol. 39.
https://doi.org/10.2166/wst.1999.0455
Castro, M.J., Ortega, S., de la Asunción, M., Mantas, J.M., Gallardo, J.M., 2011. GPU
computing for shallow water flow simulation based on finite volume schemes.
144
Comptes Rendus - Mec. https://doi.org/10.1016/j.crme.2010.12.004
Casulli, V., Stelling, G.S., 2013. A semi-implicit numerical model for urban drainage
systems. Int. J. Numer. Methods Fluids 73. https://doi.org/10.1002/fld.3817
Cea, L., Vázquez-Cendón, M.E., 2010. Unstructured finite volume discretization of

two-dimensional depth-averaged shallow water equations with porosity. Int. J.
Numer. Methods Fluids 63. https://doi.org/10.1002/fld.2107
Chang, T.J., Wang, C.H., Chen, A.S., 2015. A novel approach to model dynamic flow
interactions between storm sewer system and overland surface for different land
covers in urban areas. J. Hydrol. 524.
https://doi.org/10.1016/j.jhydrol.2015.03.014
Chen, A.S., Djordjević, S., Leandro, J., Savić, D.A., 2007. The urban inundation
model with bidirectional flow interaction between 2D overland surface and 1D
sewer networks, in: Novatech 2007. https://doi.org/2042/25250
Chen, A.S., Leandro, J., Djordjević, S., 2016. Modelling sewer discharge via
displacement of manhole covers during flood events using 1D/2D SIPSON/P-
DWave dual drainage simulations. Urban Water J. 13.
https://doi.org/10.1080/1573062X.2015.1041991
Chen, S., Zheng, F., Liu, X., 2021. Augmented HLL Riemann solver including slope
source term for 1D mixed pipe flows. J. Hydroinformatics 23.
https://doi.org/10.2166/hydro.2021.155
Cheng, T., Xu, Z., Hong, S., Song, S., 2017. Flood Risk Zoning by Using 2D
Hydrodynamic Modeling: A Case Study in Jinan City. Math. Probl. Eng. 2017.
https://doi.org/10.1155/2017/5659197
Chippada, S., Dawson, C.N., Martinez, M.L., Wheeler, M.F., 1998. A Godunov-type
finite volume method for the system of shallow water equations. Comput.
Methods Appl. Mech. Eng. 151. https://doi.org/10.1016/S0045-7825(97)00108-
4
Chu, X., Steinman, A., 2009. Event and Continuous Hydrologic Modeling with HEC-
HMS. J. Irrig. Drain. Eng. 135. https://doi.org/10.1061/(asce)0733-
9437(2009)135:1(119)
145
Costabile, P., Costanzo, C., Ferraro, D., Macchione, F., Petaccia, G., 2020.
Performances of the new HEC-RAS version 5 for 2-D hydrodynamic-based
rainfall-runoff simulations at basin scale: Comparison with a state-of-the art
model. Water (Switzerland) 12. https://doi.org/10.3390/W12092326
Costabile, P., Costanzo, C., Macchione, F., 2017. Performances and limitations of the
diffusive approximation of the 2-d shallow water equations for flood simulation
in urban and rural areas. Appl. Numer. Math. 116.
https://doi.org/10.1016/j.apnum.2016.07.003
Cunge, J.A., Holly, F.M., Verwey, A., 1980. Practical aspects of computational river
hydraulics.
De Silva, M.M.G.T., Weerakoon, S.B., Herath, S., 2014. Modeling of Event and
Continuous Flow Hydrographs with HEC–HMS: Case Study in the Kelani River
Basin, Sri Lanka. J. Hydrol. Eng. 19. https://doi.org/10.1061/(asce)he.1943-
5584.0000846
Defra, Environment Agency, 2010. Benchmarking of 2D Hydraulic Modelling

Packages, SC080035/SR2.
Delmas, V., Soulaïmani, A., 2022. Multi-GPU implementation of a time-explicit finite

volume solver using CUDA and a CUDA-Aware version of OpenMPI with
application to shallow water flows. Comput. Phys. Commun. 271.
https://doi.org/10.1016/j.cpc.2021.108190
Desalegn, H., Mulu, A., 2021. Mapping flood inundation areas using GIS and HEC-
RAS model at Fetam River, Upper Abbay Basin, Ethiopia. Sci. African 12.
https://doi.org/10.1016/j.sciaf.2021.e00834
DHI, 2017. MIKE 11 River and Channel Modelling: Short Introduction - Tutorial.
Water Environ.
Djordjević, S., Prodanović, D., Maksimović, Č., 1999. An approach to simulation of

dual drainage, in: Water Science and Technology. https://doi.org/10.1016/S0273-
1223(99)00221-8
Djordjević, S., Prodanović, D., Maksimović, Č., Ivetić, M., Savić, D., 2005. SIPSON
- Simulation of interaction between pipe flow and surface overland flow in
146
networks. Water Sci. Technol. 52. https://doi.org/10.2166/wst.2005.0143
Doulgeris, C., Georgiou, P., Papadimos, D., Papamichail, D., 2012. Ecosystem
approach to water resources management using the MIKE 11 modeling system
in the Strymonas River and Lake Kerkini. J. Environ. Manage. 94.
https://doi.org/10.1016/j.jenvman.2011.06.023
Duhan, S., Kumar, M., 2017. Event and Continuous Hydrological Modeling with
HEC- HMS: A Review Study. Int. J. Eng. Technol. Sci. Res. IJETSR 4.
Fadlillah, L.N., Widyastuti, M., Sunarto, Marfai, M.A., 2020. Comparison of tidal
model using mike21 and delft3d-flow in part of Java Sea, Indonesia, in: IOP
Conference Series: Earth and Environmental Science.
https://doi.org/10.1088/1755-1315/451/1/012067
Fahad, M.G.R., Nazari, R., Motamedi, M.H., Karimi, M.E., 2020. Coupled
Hydrodynamic and Geospatial Model for Assessing Resiliency of Coastal
Structures under Extreme Storm Scenarios. Water Resour. Manag. 34.
https://doi.org/10.1007/s11269-020-02490-y
Falter, D., Vorogushyn, S., Lhomme, J., Apel, H., Gouldby, B., Merz, B., 2013.
Hydraulic model evaluation for large-scale flood risk assessments. Hydrol.
Process. 27. https://doi.org/10.1002/hyp.9553
Fan, Y., Ao, T., Yu, H., Huang, G., Li, X., 2017. A coupled 1D-2D hydrodynamic
model for urban flood inundation. Adv. Meteorol. 2017.
https://doi.org/10.1155/2017/2819308
Fewtrell, T.J., Duncan, A., Sampson, C.C., Neal, J.C., Bates, P.D., 2011.
Benchmarking urban flood models of varying complexity and scale using high
resolution terrestrial LiDAR data. Phys. Chem. Earth 36.
https://doi.org/10.1016/j.pce.2010.12.011
Feyen, L., Vrugt, J.A., Nualláin, B.Ó., van der Knijff, J., De Roo, A., 2007. Parameter
optimisation and uncertainty assessment for large-scale streamflow simulation
with the LISFLOOD model. J. Hydrol. 332.
Forero-Ortiz, E., Martínez-Gomariz, E., Porcuna, M.C., Locatelli, L., Russo, B., 2020.
147
Flood risk assessment in an underground railway system under the impact of
climate change-A case study of the Barcelona Metro. Sustain. 12.
https://doi.org/10.3390/su12135291
Fraga, I., Cea, L., Puertas, J., 2017. Validation of a 1D-2D dual drainage model under
unsteady part-full and surcharged sewer conditions. Urban Water J. 14.
https://doi.org/10.1080/1573062X.2015.1057180
Franzini, F., Hoedenaeken, D., Soares-Frazão, S., 2018. Modeling the Flow Around
Islands in Rivers Using a One-Dimensional Approach, in: Springer Water.
https://doi.org/10.1007/978-981-10-7218-5_9
Gai, L., Nunes, J.P., Baartman, J.E.M., Zhang, H., Wang, F., de Roo, A., Ritsema, C.J.,
Geissen, V., 2019. Assessing the impact of human interventions on floods and
low flows in the Wei River Basin in China using the LISFLOOD model. Sci.
Total Environ. 653. https://doi.org/10.1016/j.scitotenv.2018.10.379
Garcia, R., Gonzalez, N., 2013. The importance of using two-dimensional flexible
meshes to simulate dam-break flooding over complex terrain, in: Association of
State Dam Safety Officials Annual Conference 2013, Dam Safety 2013.
Garland, M., Le Grand, S., Nickolls, J., Anderson, J., Hardwick, J., Morton, S., Phillips,
E., Zhang, Y., Volkov, V., 2008. Parallel computing experiences with CUDA.
IEEE Micro 28. https://doi.org/10.1109/MM.2008.57
Giammarco, P., 1996. A conservative finite elements approach to overland flow: the
control volume finite element formulation. J. Hydrol. 175.
https://doi.org/10.1016/0022-1694(95)02855-2
Golmohammadi, G., Prasher, S., Madani, A., Rudra, R., 2014. Evaluating three
hydrological distributed watershed models: MIKE-SHE, APEX, SWAT.
Hydrology 1. https://doi.org/10.3390/hydrology1010020
Gouta, N., Maurel, F., 2002. A finite volume solver for 1D shallow-water equations
applied to an actual river. Int. J. Numer. Methods Fluids 38.
https://doi.org/10.1002/fld.201
Guan, M., Sillanpää, N., Koivusalo, H., 2015. Modelling and assessment of
hydrological changes in a developing urban catchment. Hydrol. Process. 29.
148
https://doi.org/10.1002/hyp.10410
Han, Y.S., Shin, E.T., Eum, T.S., Song, C.G., 2019. Inundation risk assessment of
underground space using consequence-probability matrix. Appl. Sci. 9.
https://doi.org/10.3390/app9061196
Haron, S, Khalid, K., Ali, M., Abd Rahman, N.F., Mispan, R., Haron, Siti, Rasid, Z.,
M.H, N., Kamaruddin, H., 2016. Application of the SWAT hydrologic model in
Malaysia: recent research, in: The Challenges of Agro-Environmental Research
in Monsoon Asia.
Harris, M.J., Coombe, G., Scheuermann, T., Lastra, A., 2002. Physically-based visual
simulation on graphics hardware, in: Proceedings of the
SIGGRAPH/Eurographics Workshop on Graphics Hardware.
https://doi.org/10.1145/1198555.1198791
Harten, A., Lax, P.D., Leer, B. van, 1983. On Upstream Differencing and Godunov-
Type Schemes for Hyperbolic Conservation Laws. SIAM Rev. 25.
https://doi.org/10.1137/1025002
Henonin, J., Russo, B., Mark, O., Gourbesville, P., 2013. Real-time urban flood
forecasting and modelling - A state of the art. J. Hydroinformatics 15.
Herath, S., Dutta, D., 2004. Modeling of urban flooding including underground space.
2nd APHW Conf. 5-8 July 2004 Singapore /Asia Pacific Assoc. Hydrol. Water
Resour.
Hernes, R.R., Gragne, A.S., Abdalla, E.M.H., Braskerud, B.C., Alfredsen, K.,
Muthanna, T.M., 2020. Assessing the effects of four SUDS scenarios on
combined sewer overflows in Oslo, Norway: Evaluating the low-impact
development module of the Mike Urban model. Hydrol. Res. 51.
https://doi.org/10.2166/nh.2020.070
Herty, M., Seaïd, M., 2008. Simulation of transient gas flow at pipe-to-pipe
intersections. Int. J. Numer. Methods Fluids 56. https://doi.org/10.1002/fld.1531
Hirpa, F.A., Salamon, P., Beck, H.E., Lorini, V., Alfieri, L., Zsoter, E., Dadson, S.J.,
2018. Calibration of the Global Flood Awareness System (GloFAS) using daily
149
streamflow data. J. Hydrol. 566. https://doi.org/10.1016/j.jhydrol.2018.09.052
Hlavčová, K., Lapin, M., Valent, P., Szolgay, J., Kohnová, S., Rončák, P., 2015.
Estimation of the impact of climate change-induced extreme precipitation events
on floods. Contrib. to Geophys. Geod. 45. https://doi.org/10.1515/congeo-2015-
0019
Hong, S.W., Kim, C., 2011. A new finite volume method on junction coupling and
boundary treatment for flow network system analyses. Int. J. Numer. Methods
Fluids 65. https://doi.org/10.1002/fld.2212
Horton, R.E., 1933. The Rôle of infiltration in the hydrologic cycle. Eos, Trans. Am.
Geophys. Union 14. https://doi.org/10.1029/TR014i001p00446
Hossain, S., Hewa, G.A., Wella-Hewage, S., 2019. A comparison of continuous and
event-based rainfall-runoff (RR) modelling using EPA-SWMM. Water
(Switzerland) 11. https://doi.org/10.3390/w11030611
Hou, J., Kang, Y., Hu, C., Tong, Y., Pan, B., Xia, J., 2020. A GPU-based numerical
model coupling hydrodynamical and morphological processes. Int. J. Sediment
Res. 35. https://doi.org/10.1016/j.ijsrc.2020.02.005
Hou, J., Liang, Q., Zhang, H., Hinkelmann, R., 2015. An efficient unstructured
MUSCL scheme for solving the 2D shallow water equations. Environ. Model.
Softw. 66. https://doi.org/10.1016/j.envsoft.2014.12.007
Hou, J. ming, Wang, R., Jing, H. xiao, Zhang, X., Liang, Q. hua, Di, Y. yan, 2017. An
efficient dynamic uniform Cartesian grid system for inundation modeling. Water
Sci. Eng. 10. https://doi.org/10.1016/j.wse.2017.12.004
Hsu, M.H., Chen, S.H., Chang, T.J., 2002. Dynamic inundation simulation of storm
water interaction between sewer system and overland flows. J. Chinese Inst. Eng.
Trans. Chinese Inst. Eng. Ser. A/Chung-kuo K. Ch’eng Hsuch K’an 25.
https://doi.org/10.1080/02533839.2002.9670691
Hsu, M.H., Chen, S.H., Chang, T.J., 2000. Inundation simulation for urban drainage
basin with storm sewer system. J. Hydrol. 234. https://doi.org/10.1016/S0022-
1694(00)00237-7
Hu, P., Lei, Y., Han, J., Cao, Z., Liu, H., He, Z., 2019a. Computationally efficient
150
modeling of hydro-sediment-morphodynamic processes using a hybrid local time
step/global maximum time step. Adv. Water Resour. 127.
https://doi.org/10.1016/j.advwatres.2019.03.006
Hu, P., Lei, Y., Han, J., Cao, Z., Liu, H., He, Z., Yue, Z., 2019b. Improved Local Time
Step for 2D Shallow-Water Modeling Based on Unstructured Grids. J. Hydraul.
Eng. 145. https://doi.org/10.1061/(asce)hy.1943-7900.0001642
Hu, X., Song, L., 2018. Hydrodynamic modeling of flash flood in mountain
watersheds based on high-performance GPU computing. Nat. Hazards 91.
https://doi.org/10.1007/s11069-017-3141-7
Jahanbazi, M., Egger, U., 2014. Application and comparison of two different dual
drainage models to assess urban flooding. Urban Water J. 11.
https://doi.org/10.1080/1573062X.2013.871041
Jakoubek, M., 2007. Flood Protection of Prague Metro after the 2002 flood, in:
“Proceedings of the 33rd ITA-AITES World Tunnel Congress - Underground
Space - The 4th Dimension of Metropolises.”
Jang, S., Cho, M., Yoon, J., Yoon, Y., Kim, S., Kim, G., Kim, L., Aksoy, H., 2007.
Using SWMM as a tool for hydrologic impact assessment. Desalination 212.
https://doi.org/10.1016/j.desal.2007.05.005
Ji, S., Qiuwen, Z., 2015. A GIS-based Subcatchments Division Approach for SWMM.
Open Civ. Eng. J. 9. https://doi.org/10.2174/1874149501509010515
Jiang, L., Chen, Y., Wang, H., 2015. Urban flood simulation based on the SWMM
model, in: IAHS-AISH Proceedings and Reports. https://doi.org/10.5194/piahs-
368-186-2015
Jovanovic, D., Gelsinari, S., Bruce, L., Hipsey, M., Teakle, I., Barnes, M., Coleman,
R., Deletic, A., Mccarthy, D.T., 2019. Modelling shallow and narrow urban salt-
wedge estuaries: Evaluation of model performance and sensitivity to optimise
input data collection. Estuar. Coast. Shelf Sci. 217.
https://doi.org/10.1016/j.ecss.2018.10.022
Kalyanapu, A.J., Ghafoor, S.K., Marshall, R.J., Dullo, T.T., Judi, D.R., Shankar, S.,
2014. Benchmark Exercise for Comparing the Computational Performance of
151
Two-Dimensional Flood Models in CPU, Multi-CPU, and GPU Frameworks, in:
World Environmental and Water Resources Congress 2014: Water Without
Borders - Proceedings of the 2014 World Environmental and Water Resources
Congress. https://doi.org/10.1061/9780784413548.133
Kalyanapu, A.J., Shankar, S., Pardyjak, E.R., Judi, D.R., Burian, S.J., 2011.
Assessment of GPU computational enhancement to a 2D flood model. Environ.
Model. Softw. 26. https://doi.org/10.1016/j.envsoft.2011.02.014
Katwal, R., Li, J., Zhang, T., Hu, C., Rafique, M.A., Zheng, Y., 2021. Event-based
and continous flood modeling in Zijinguan watershed, Northern China. Nat.
Hazards 108. https://doi.org/10.1007/s11069-021-04703-y
Kazezyilmaz-Alhan, C.M., 2012. An improved solution for diffusion waves to

overland flow. Appl. Math. Model. 36.
https://doi.org/10.1016/j.apm.2011.11.045
Kim, B., Sanders, B.F., Schubert, J.E., Famiglietti, J.S., 2014. Mesh type tradeoffs in
2D hydrodynamic modeling of flooding with a Godunov-based flow solver. Adv.
Water Resour. 68. https://doi.org/10.1016/j.advwatres.2014.02.013
Kim, H.J., Rhee, D.S., Song, C.G., 2018. Numerical computation of underground
inundation in multiple layers using the adaptive transfer method. Water
Kim, S., Shen, H., Noh, S., Seo, D.J., Welles, E., Pelgrim, E., Weerts, A., Lyons, E.,
Philips, B., 2021. High-resolution modeling and prediction of urban floods using
WRF-Hydro and data assimilation. J. Hydrol. 598.
https://doi.org/10.1016/j.jhydrol.2021.126236
Lacasta, A., Morales-Hernández, M., Murillo, J., García-Navarro, P., Garcia, R., 2014.
High Performance GPU Speed-Up Strategies For The Computation Of 2D
Inundation Models, in: 11th International Conference on Hydroinformatics.
Lamb, R., Crossley, M., Waller, S., 2009. A fast two-dimensional floodplain
inundation model. Proc. Inst. Civ. Eng. Water Manag. 162.
https://doi.org/10.1680/wama.2009.162.6.363
Leandro, J., Chen, A.S., Djordjević, S., Savić, D.A., 2009. Comparison of 1D/1D and
152
1D/2D Coupled (Sewer/Surface) Hydraulic Models for Urban Flood Simulation.
J. Hydraul. Eng. 135. https://doi.org/10.1061/(asce)hy.1943-7900.0000037
Leandro, J., Martins, R., 2016. A methodology for linking 2D overland flow models
with the sewer network model SWMM 5.1 based on dynamic link libraries. Water
Sci. Technol. 73. https://doi.org/10.2166/wst.2016.171
Leon, A.S., Ghidaoui, M.S., Schmidt, A.R., Garcia, M.H., 2010. A robust two-
equation model for transient-mixed flows. J. Hydraul. Res. 48.
https://doi.org/10.1080/00221680903565911
León, A.S., Ghidaoui, M.S., Schmidt, A.R., García, M.H., 2006. Godunov-Type
Solutions for Transient Flows in Sewers. J. Hydraul. Eng. 132.
León, A.S., Liu, X., Ghidaoui, M.S., Schmidt, A.R., García, M.H., 2010. Junction and
Drop-Shaft Boundary Conditions for Modeling Free-Surface, Pressurized, and
Mixed Free-Surface Pressurized Transient Flows. J. Hydraul. Eng. 136.
Li, B., Phillips, M., Fleming, C.A., 2006. Application of 3D hydrodynamic model to
flood risk assessment. Proc. Inst. Civ. Eng. Water Manag. 159.
https://doi.org/10.1680/wama.2006.159.1.63
Li, D., Liang, Z., Li, B., Lei, X., Zhou, Y., 2019. Multi-objective calibration of MIKE
SHE with SMAP soil moisture datasets. Hydrol. Res. 50.
https://doi.org/10.2166/nh.2018.110
Li, J., McCorquodale, A., 1999. Modeling Mixed Flow in Storm Sewers. J. Hydraul.
Eng. 125. https://doi.org/10.1061/(asce)0733-9429(1999)125:11(1170)
Li, W., Zou, J.Y., Hu, P., 2020. Urban flood simulation based on porosity and local
time step. Zhejiang Daxue Xuebao (Gongxue Ban)/Journal Zhejiang Univ.
(Engineering Sci. 54. https://doi.org/10.3785/j.issn.1008-973X.2020.03.023
Li, X., Huang, M., Wang, R., 2020. Numerical simulation of Donghu Lake
hydrodynamics and water quality based on remote sensing and Mike 21. ISPRS
Int. J. Geo-Information 9. https://doi.org/10.3390/ijgi9020094
Li, Z., Liu, J., Mei, C., Shao, W., Wang, H., Yan, D., 2019. Comparative analysis of
153
building representations in TELEMAC-2D for flood inundation in idealized
urban districts. Water (Switzerland) 11. https://doi.org/10.3390/w11091840
Liang, J., Yang, Q., Sun, T., Martin, J.D., Sun, H., Li, L., 2015. MIKE 11 model-based
water quality model as a tool for the evaluation of water quality management
plans. J. Water Supply Res. Technol. - AQUA 64.
https://doi.org/10.2166/aqua.2015.048
Liang, Q., 2010. Flood Simulation Using a Well-Balanced Shallow Flow Model. J.
Hydraul. Eng. 136. https://doi.org/10.1061/(asce)hy.1943-7900.0000219
Liang, Q., Marche, F., 2009. Numerical resolution of well-balanced shallow water
equations with complex source terms. Adv. Water Resour. 32.
Liang, Q., Smith, L.S., 2015. A High-Performance Integrated hydrodynamic

Modelling System for urban flood simulations. J. Hydroinformatics 17.
Liang, Q., Xia, X., Hou, J., 2016. Catchment-scale High-resolution Flash Flood
Simulation Using the GPU-based Technology, in: Procedia Engineering.
https://doi.org/10.1016/j.proeng.2016.07.585
Linh, N.T.M., Tri, D.Q., Thai, T.H., Cao Don, N., 2018. Application of a two-
dimensional model for flooding and floodplain simulation: Case study in Tra
Khuc-Song Ve river in Viet Nam. Lowl. Technol. Int. 20.
Liu, J., Li, Z., Mei, C., Wang, K., Zhou, G., 2019a. Urban flood analysis for different
design storm hyetographs in Xiamen Island based on TELEMAC-2D. Kexue
Tongbao/Chinese Sci. Bull. 64. https://doi.org/10.1360/N972018-01180
Liu, J., Li, Z., Mei, C., Wang, K., Zhou, G., 2019b. TELEMAC-2D. Kexue
Tongbao/Chinese Sci. Bull. 64.
Liu, Q., Qin, Y., Li, G., 2018. Fast simulation of large-scale floods based on GPU
parallel computing. Water (Switzerland) 10. https://doi.org/10.3390/w10050589
Liu, R., Li, Z., Xin, X., Liu, D., Zhang, J., Yang, Z., 2022. Water balance computation
and water quality improvement evaluation for Yanghe Basin in a semiarid area
of North China using coupled MIKE SHE/MIKE 11 modeling. Water Supply 22.
154
https://doi.org/10.2166/ws.2021.214
Liu, Y., Niu, S., Zhang, W., Xu, J., 2018. Integrated modeling and calculation of urban
surface and underground flooding. Ekoloji 27.
Lyu, H.-M., Shen, S.-L., Yang, J., Yin, Z.-Y., 2019. Scenario-based inundation
analysis of metro systems: a case study in Shanghai. Hydrol. Earth Syst. Sci.
Discuss.
Lyu, H.M., Wang, G.F., Shen, J.S., Lu, L.H., Wang, G.Q., 2016. Analysis and GIS
mapping of flooding hazards on 10 May 2016, Guangzhou, China. Water
Lyu, H.M., Shen, S.L., Zhou, A. and Yang, J., 2019. Perspectives for flood risk
assessment and management for mega-city metro system. Tunnelling and
Underground Space Technology, 84, pp.31-44.
Ma, L., He, C., Bian, H., Sheng, L., 2016. MIKE SHE modeling of ecohydrological
processes: Merits, applications, and challenges. Ecol. Eng. 96.
https://doi.org/10.1016/j.ecoleng.2016.01.008
Maksimovi, Č., Prodanovi, D., Djordjevi, S., 2007. Urban Pluvial Flooding :
Development of GIS Based Pathway Model for Surface Flooding and Interface
with Surcharged Sewer Model. Novatech.
Maksimović, Č., Prodanović, D., Boonya-Aroonnet, S., Leitão, J.P., Djordjević, S.,
Allitt, R., 2009. Overland flow and pathway analysis for modelling of urban
pluvial flooding. J. Hydraul. Res. 47.
https://doi.org/10.1080/00221686.2009.9522027
Malekpour, A., Karney, B., 2014. A non-oscillatory preissmann slot method based
numerical model, in: Procedia Engineering.
https://doi.org/10.1016/j.proeng.2014.11.461
Maranzoni, A., Dazzi, S., Aureli, F., Mignosa, P., 2015. Extension and application of
the Preissmann slot model to 2D transient mixed flows. Adv. Water Resour. 82.
Marimin, N.A., Razi, M.A.M., Ahmad, M.A., Adnan, M.S., Rahmat, S.N., 2018.
HEC-RAS hydraulic model for floodplain area in Sembrong River. Int. J. Integr.
155
Eng. 10. https://doi.org/10.30880/ijie.2018.10.02.029
Martinez, C., Garcia-Martinez, R., Miralles-Wilhelm, F., 2011. A two-phase debris

flow model with boulder transport. Int. J. Saf. Secur. Eng. 1.
https://doi.org/10.2495/SAFE-V1-N4-389-402
Martins, R., Leandro, J., Chen, A.S., Djordjević, S., 2017. A comparison of three dual
drainage models: Shallow water vs local inertial vs diffusive wave. J.
Hydroinformatics 19. https://doi.org/10.2166/hydro.2017.075
McCorquodale, J.A., Hamam, M.A., 1983. Modeling surcharged flow in sewers,.
Ming, X., Liang, Q., Dawson, R., Xia, X., Hou, J., 2022. A quantitative multi-hazard
risk assessment framework for compound flooding considering hazard inter-
dependencies and interactions. J. Hydrol. 607.
Ming, X., Liang, Q., Xia, X., Li, D., Fowler, H.J., 2020. Real-Time Flood Forecasting
Based on a High-Performance 2-D Hydrodynamic Model and Numerical
Weather Predictions. Water Resour. Res. 56.
https://doi.org/10.1029/2019WR025583
Miyata, H., 1986. Finite-difference simulation of breaking waves. J. Comput. Phys.

65. https://doi.org/10.1016/0021-9991(86)90011-2
Morales-Hernández, M., Sharif, M.B., Kalyanapu, A., Ghafoor, S.K., Dullo, T.T.,
Gangrade, S., Kao, S.C., Norman, M.R., Evans, K.J., 2021. TRITON: A Multi-
GPU open source 2D hydrodynamic flood model. Environ. Model. Softw. 141.
https://doi.org/10.1016/j.envsoft.2021.105034
Morsy, M.M., Goodall, J.L., O’Neil, G.L., Sadler, J.M., Voce, D., Hassan, G., Huxley,
C., 2018. A cloud-based flood warning system for forecasting impacts to
transportation infrastructure systems. Environ. Model. Softw. 107.
Msaddek, M., Kimbowa, G., Garouani, A. El, 2020. Hydrological Modeling of Upper
OumErRabia Basin (Morocco), Comparative Study of the Event-Based and
Continuous-Process HEC-HMS Model Methods. Comput. Water, Energy,
Environ. Eng. 09. https://doi.org/10.4236/cweee.2020.94011
156
Namara, W.G., Damisse, T.A., Tufa, F.G., 2021. Application of HEC-RAS and HEC-
GeoRAS model for Flood Inundation Mapping, the case of Awash Bello Flood
Plain, Upper Awash River Basin, Oromiya Regional State, Ethiopia. Model.
Earth Syst. Environ. https://doi.org/10.1007/s40808-021-01166-9
Nazari, F., Nair, R.D., 2017. A Godunov-type finite-volume solver for nonhydrostatic
Euler equations with a time-splitting approach. J. Adv. Model. Earth Syst. 9.
https://doi.org/10.1002/2016MS000888
Neal, J., Dunne, T., Sampson, C., Smith, A., Bates, P., 2018. Optimisation of the two-
dimensional hydraulic model LISFOOD-FP for CPU architecture. Environ.
Neal, J., Fewtrell, T., Trigg, M., 2009. Parallelisation of storage cell flood models
using OpenMP. Environ. Model. Softw. 24.
Néelz, S., Pender, G., Britain, G., 2009. Desktop review of 2D hydraulic modelling
packages, Environment Agency. https://doi.org/978-1-84911-079-2
Neelz, S., Pender, G., Wright, N.G., 2010. Benchmarking of 2D Hydraulic Modelling
Packages protecting and improving the environment in England and,
Benchmarking.
Newcastle City Council, 2016. Local Flood Risk Management Plan [WWW
Document]. URL https://www.newcastle.gov.uk/sites/default/files/2019-05/126
Newcastle Local Flood Risk Management Plan.pdf
Nickolls, J., Buck, I., Garland, M., Skadron, K., 2008. Scalable parallel programming
with CUDA. Queue 6. https://doi.org/10.1145/1365490.1365500
Noh, S.J., Lee, J.H., Lee, S., Kawaike, K., Seo, D.J., 2018. Hyper-resolution 1D-2D
urban flood modelling using LiDAR data and hybrid parallelization. Environ.
Noh, S.J., Lee, S., An, H., Kawaike, K., Nakagawa, H., 2016. Ensemble urban flood
simulation in comparison with laboratory-scale experiments: Impact of
interaction models for manhole, sewer pipe, and surface flow. Adv. Water Resour.
97. https://doi.org/10.1016/j.advwatres.2016.08.015
157
Osher, S., Sanders, R., 1983. Numerical approximations to nonlinear conservation
laws with locally varying time and space grids. Math. Comput. 41.
https://doi.org/10.1090/s0025-5718-1983-0717689-8
Parra, V., Fuentes-Aguilera, P., Muñoz, E., 2018. Identifying advantages and
drawbacks of two hydrological models based on a sensitivity analysis: a study in
two Chilean watersheds. Hydrol. Sci. J. 63.
https://doi.org/10.1080/02626667.2018.1538593
Pau, J.C., Sanders, B.F., 2006. Performance of Parallel Implementations of an Explicit

Finite-Volume Shallow-Water Model. J. Comput. Civ. Eng. 20.
Peiro, J., Spencer, S., 2005. Finite Difference, finite elements and finite volume
methods for partial differential equations, in: Handbook of Materials Modeling.
Phillips, B.C., Yu, S., Thompson, G.R., Silva, N. De, 2005. 1D and 2D Modelling of
Urban Drainage Systems using XP-SWMM and TUFLOW. 10th Int. Conf.
Urban Drain.
Politano, M., Odgaard, A.J., Klecan, W., 2007. Case Study: Numerical Evaluation of
Hydraulic Transients in a Combined Sewer Overflow Tunnel System. J. Hydraul.
Qiang, Y., He, J., Xiao, T., Lu, W., Li, J., Zhang, L., 2021. Coastal town flooding upon
compound rainfall-wave overtopping-storm surge during extreme tropical
cyclones in Hong Kong. J. Hydrol. Reg. Stud. 37.
https://doi.org/10.1016/j.ejrh.2021.100890
Quan, N.H., Hieu, N.D., Van Thu, T.T., Buchanan, M., Canh, N.D., da Cunha Oliveira
Santos, M., Luan, P.D.M.H., Hoang, T.T., Phung, H.L.T., Canh, K.M., Smith,
M., 2020. Green infrastructure modelling for assessment of urban flood reduction
in Ho Chi Minh city, in: Lecture Notes in Civil Engineering.
https://doi.org/10.1007/978-981-15-0802-8_177
Radojkovic, M., Maksimovic, C., 1984. DEVELOPMENT, TESTING AND

APPLICATION OF BELGRADE URBAN DRAINAGE MODEL.
Re, M., Kazimierski, L.D., Badano, N.D., 2019. High-resolution urban flood model
158
for risk mitigation validated with records collected by the affected community. J.
Flood Risk Manag. 12. https://doi.org/10.1111/jfr3.12524
Rodríguez, S.O., Onof1, C., Maksimović1, and Č., Li-Pen Wang1, 2, Willems2, P.,
Assel3, J. Van, Auguste, Gires4, , Abdella Ichiba4, 5, Bruni6, G., Veldhuis, and
M.-C. ten, 2015. Urban pluvial flood modelling: current theory and practice.
Review document related to Work Package 3 – Action 13. J. Hydrol. 13.
Rong, Y., Zhang, T., Zheng, Y., Hu, C., Peng, L., Feng, P., 2020. Three-dimensional
urban flood inundation simulation based on digital aerial photogrammetry. J.
Hydrol. 584. https://doi.org/10.1016/j.jhydrol.2019.124308
Rossman, L.A., 2015. STORM WATER MANAGEMENT MODEL USER’S

MANUAL Version 5.1. EPA/600/R-14/413b, Natl. Risk Manag. Lab. Off. Res.
Dev. United States Environ. Prot. Agency, Cincinnati, Ohio.
Rousseau, M., Cerdan, O., Delestre, O., Dupros, F., 2012. Overland flow modelling
with the Shallow Water Equation using a well balanced numerical scheme:
Adding efficiency or just more complexity? J. Hydrol. Eng. 12.
Rubinato, M., Martins, R., Kesserwani, G., Leandro, J., Djordjević, S., Shucksmith, J.,
2017. Experimental calibration and validation of sewer/surface flow exchange
equations in steady and unsteady flow conditions. J. Hydrol. 552.
Rujner, H., Leonhardt, G., Marsalek, J., Viklander, M., 2018. High-resolution
modelling of the grass swale response to runoff inflows with Mike SHE. J. Hydrol.
562. https://doi.org/10.1016/j.jhydrol.2018.05.024
Sadeghi, F., Rubinato, M., Goerke, M., Hart, J., 2022. Assessing the Performance of
LISFLOOD-FP and SWMM for a Small Watershed with Scarce Data
Availability. Water (Switzerland) 14. https://doi.org/10.3390/w14050748
Sætra, M.L., Brodtkorb, A.R., 2012. Shallow water simulations on multiple GPUs, in:
Lecture Notes in Computer Science (Including Subseries Lecture Notes in
Artificial Intelligence and Lecture Notes in Bioinformatics).
https://doi.org/10.1007/978-3-642-28145-7_6
Safavi, S., Shamsai, A., Saghafian, B., Bateni, S., 2015. Modeling Spatial Pattern of
159
Salinity using MIKE21 and Principal Component Analysis Technique in Urmia
Lake. Curr. World Environ. 10. https://doi.org/10.12944/cwe.10.2.28
Salvadore, E., Bronders, J., Batelaan, O., 2015. Hydrological modelling of urbanized
catchments: A review and future directions. J. Hydrol.
Salvan, L., Abily, M., Gourbesville, P., 2018. Hydrodynamic Coupling Method for
Stormwater Studies in Suburban Catchments—Study Case of the Magnan Basin,
Nice, in: Springer Water. https://doi.org/10.1007/978-981-10-7218-5_49
Sanders, B.F., 2008. Integration of a shallow water model with a local time step. J.
Hydraul. Res. 46. https://doi.org/10.3826/jhr.2008.3243
Sanders, B.F., Bradford, S.F., 2011. Network Implementation of the Two-Component

Pressure Approach for Transient Flow in Storm Sewers. J. Hydraul. Eng. 137.
Sanders, B.F., Schubert, J.E., Detwiler, R.L., 2010. ParBreZo: A parallel, unstructured
grid, Godunov-type, shallow-water code for high-resolution flood inundation
modeling at the regional scale. Adv. Water Resour. 33.
Schmitt, T.G., Thomas, M., Ettrich, N., 2005. Assessment of urban flooding by dual
drainage simulation model RisUrSim. Water Sci. Technol. 52.
https://doi.org/10.2166/wst.2005.0141
Seyoum, S.D., Vojinovic, Z., Price, R.K., Weesakul, S., 2012. Coupled 1D and
Noninertia 2D Flood Inundation Model for Simulation of Urban Flooding. J.
Hydraul. Eng. 138. https://doi.org/10.1061/(asce)hy.1943-7900.0000485
Sharif, M.B., Ghafoor, S.K., Hines, T.M., Morales-Hernändez, M., Evans, K.J., Kao,
S.C., Kalyanapu, A.J., Dullo, T.T., Gangrade, S., 2020. Performance Evaluation
of a Two-Dimensional Flood Model on Heterogeneous High-Performance
Computing Architectures, in: Proceedings of the Platform for Advanced
Scientific Computing Conference, PASC 2020.
https://doi.org/10.1145/3394277.3401852
Shustikova, I., Domeneghetti, A., Neal, J.C., Bates, P., Castellarin, A., 2019.
160
Comparing 2D capabilities of HEC-RAS and LISFLOOD-FP on complex
topography. Hydrol. Sci. J. 64. https://doi.org/10.1080/02626667.2019.1671982
Shustikova, I., Neal, J.C., Domeneghetti, A., Bates, P.D., Vorogushyn, S., Castellarin,
A., 2020. Levee breaching: A new extension to the LISFLOOD-FP model. Water
(Switzerland) 12. https://doi.org/10.3390/W12040942
Sidek, L.M., Jaafar, A.S., Majid, W.H.A.W.A., Basri, H., Marufuzzaman, M., Fared,
M.M., Moon, W.C., 2021. High‐resolution hydrological‐hydraulic modeling of
urban floods using infoworks icm. Sustain. 13.
https://doi.org/10.3390/su131810259
Simões, N.E., Leitão, J.P., Maksimović, Č., Sá Marques, A., Pina, R., 2010. Sensitivity
analysis of surface runoff generation in urban flood forecasting. Water Sci.
Technol. 61. https://doi.org/10.2166/wst.2010.178
Simons, D.B., 2019. Open channel flow, in: Introduction to Physical Hydrology.
https://doi.org/10.4324/9780429273339-11
Simons, F., Busse, T., Hou, J., Özgen, I., Hinkelmann, R., 2014. A model for overland
flow and associated processes within the Hydroinformatics Modelling System. J.
Hydroinformatics 16. https://doi.org/10.2166/hydro.2013.173
Smith, L.S., Liang, Q., 2013. Towards a generalised GPU/CPU shallow-flow

modelling tool. Comput. Fluids 88.
Soltani, P., Askar, M.B., Bahrami, H., Pour, S.H., 2017. Evaluation of Sediment
Transport in the Naiband Gulf Area Using Mike21. Open J. Geol. 07.
https://doi.org/10.4236/ojg.2017.72012
Son, A.L., Kim, B., Han, K.Y., 2016. A simple and robust method for simultaneous
consideration of overland and underground space in urban flood modeling. Water
Sosa, E., Thompson, G., Barbero, E., 2014. Testing of full-scale inflatable plug for
flood mitigation in tunnels. Transp. Res. Rec. https://doi.org/10.3141/2407-06
Sosa, J., Sampson, C., Smith, A., Neal, J., Bates, P., 2020. A toolbox to quickly prepare
flood inundation models for LISFLOOD-FP simulations. Environ. Model. Softw.
161
123. https://doi.org/10.1016/j.envsoft.2019.104561
Spry, R.B., Zhang, S., 2006. Modelling of Drainage Systems and Overland Flowpaths
at Catchment Scales. 7th Int. Conf. Urban Drain. Model. 4th Int. Conf. Water
Sensitive Urban Des. B. Proc.
Sun, N., Hall, M., Hong, B., Zhang, L., 2014. Impact of SWMM Catchment
Discretization: Case Study in Syracuse, New York. J. Hydrol. Eng. 19.
https://doi.org/10.1061/(asce)he.1943-5584.0000777
Sun, X., Li, R., Shan, X., Xu, H., Wang, J., 2021. Assessment of climate change
impacts and urban flood management schemes in central Shanghai. Int. J.
Disaster Risk Reduct. 65. https://doi.org/10.1016/j.ijdrr.2021.102563
Sun, D., Wang, H., Lall, U., Huang, J. and Liu, G., 2022. Subway travel risk evaluation
during flood events based on smart card data. Geomatics, Natural Hazards and
Risk, 13(1), pp.2796-2818.
Tadesse, Y.B., Fröhle, P., 2020. Modelling of flood inundation due to levee breaches:
Sensitivity of flood inundation against breach process parameters. Water
Tamiru, H., Dinka, M.O., 2021. Application of ANN and HEC-RAS model for flood
inundation mapping in lower Baro Akobo River Basin, Ethiopia. J. Hydrol. Reg.
Stud. 36. https://doi.org/10.1016/j.ejrh.2021.100855
Tauhid, F.A., Zawani, H., 2018. Mitigating climate change related floods in urban
poor areas: Green infrastructure approach. J. Reg. City Plan. 29.
https://doi.org/10.5614/jrcp.2018.29.2.2
Tayefi, V., Lane, S.N., Hardy, R.J., Yu, D., 2007. A comparison of one- and two-
dimensional approaches to modelling flood inundation over complex upland
floodplains. Hydrol. Process. 21. https://doi.org/10.1002/hyp.6523
Thorndahl, S., Schaarup-Jensen, K., 2007. Comparative analysis of uncertainties in

urban surface runoff modelling. 6th Int. Conf. Sustain. Tech. Strateg. Urban
Water Manag. , June 25-28 2007.
Thorne, C.R., Lawson, E.C., Ozawa, C., Hamlin, S.L., Smith, L.A., 2018. Overcoming
uncertainty and barriers to adoption of Blue-Green Infrastructure for urban flood
162
risk management. J. Flood Risk Manag. 11. https://doi.org/10.1111/jfr3.12218
Toda, K., Kawaike, K., Yoneyama, N., Fukakusa, S., Yamamoto, D., 2009.
Underground inundation analysis by integrated urban flood model, in: Advances
in Water Resources and Hydraulic Engineering - Proceedings of 16th IAHR-APD
Congress and 3rd Symposium of IAHR-ISHS. https://doi.org/10.1007/978-3-
540-89465-0_31
Trajkovic, B., Ivetic, M., Calomino, F., D’Ippolito, A., 1999. Investigation of
transition from free surface to pressurized flow in a circular pipe, in: Water
Science and Technology. https://doi.org/10.1016/S0273-1223(99)00222-X
Tsakiris, G., Bellos, V., 2014. A Numerical Model for Two-Dimensional Flood
Routing in Complex Terrains. Water Resour. Manag. 28.
https://doi.org/10.1007/s11269-014-0540-3
Tsihrintzis, V.A., Hamid, R., 1998. Runoff quality prediction from small urban
catchments using SWMM. Hydrol. Process. 12.
https://doi.org/10.1002/(SICI)1099-1085(199802)12:2<311::AID-
HYP579>3.0.CO;2-R
USACE, 2016. Hydrologic Modeling System HEC-HMS User’s Manual. Hydrol. Eng.
Center, Davis, CA 1.
Vacondio, R., Dal Palù, A., Mignosa, P., 2014. GPU-enhanced finite volume shallow
water solver for fast flood simulations. Environ. Model. Softw. 57.
van der Knijff, J.M., Younis, J., de Roo, A.P.J., 2010. LISFLOOD: A GIS-based
distributed model for river basin scale water balance and flood simulation. Int. J.
Geogr. Inf. Sci. 24. https://doi.org/10.1080/13658810802549154
Van Der Sterren, M., Rahman, A., 2011. The impact of rainwater tanks on stormwater
runoff from a single lot development in Western Sydney based on XP-SWMM
modelling, in: 34th IAHR Congress 2011 - Balance and Uncertainty: Water in a
Changing World, Incorporating the 33rd Hydrology and Water Resources
Symposium and the 10th Conference on Hydraulics in Water Engineering.
Vari, 2017. MIKE 11 Reference Manual. Environ. Capacit. Guohe River Water Transf.
163
Proj. from Yangtze River to Huaihe River Based a MIKE 11 Model. 32.
Vasconcelos, J.G., Wright, S.J., Roe, P.L., 2006. Improved Simulation of Flow
Regime Transition in Sewers: Two-Component Pressure Approach. J. Hydraul.
Vidya, K.N., 2021. Runoff assessment by storm water management model (Swmm)-
a new approach. J. Appl. Nat. Sci. 13. https://doi.org/10.31018/jans.v13iSI.2813
Viñas, M., Lobeiras, J., Fraguela, B.B., Arenaz, M., Amor, M., García, J.A., Castro,
M.J., Doallo, R., 2013. A multi-GPU shallow-water simulation with transport of
contaminants, in: Concurrency and Computation: Practice and Experience.
https://doi.org/10.1002/cpe.2917
Wagner, J., 1992. The great Chicago flood of 1992. ANNA J. / Am. Nephrol. Nurses"
Assoc. 19.
Wang, K.-H., Altunkaynak, A., 2012. Comparative Case Study of Rainfall-Runoff

Modeling between SWMM and Fuzzy Logic Approach. J. Hydrol. Eng. 17.
https://doi.org/10.1061/(asce)he.1943-5584.0000419
Wang, Y., Zhang, C., Li, Z., Sun, B., Zhou, H., 2019. Applicability of Preissmann box
scheme for calculation of transcritical flow in pipes. Water Sci. Technol. Water
Supply 19. https://doi.org/10.2166/ws.2019.010
WBM, B., 2001. TUFLOW User Manual. 6th Conf. Hydraul. Civ. Eng. State Hydraul.
35.
WIGGERT DC, 1972. Transient flow in free- surface, pressurized systems. ASCE J
Hydraul Div 98. https://doi.org/10.1061/jyceaj.0003189
Wu, J., Fang, W., Hu, Z., Hong, B., 2018. Application of Bayesian approach to
dynamic assessment of flood in urban underground spaces. Water (Switzerland)
10. https://doi.org/10.3390/w10091112
Wu, J.S., Zhang, H., Yang, R., Dalrymple, R.A., Hérault, A., 2013. Numerical
modeling of dam-break flood through intricate city layouts including
underground spaces using GPU-based SPH method. J. Hydrodyn. 25.
https://doi.org/10.1016/S1001-6058(13)60429-1
164
Xia, X., Liang, Q., 2018. A new efficient implicit scheme for discretising the stiff
friction terms in the shallow water equations. Adv. Water Resour. 117.
Xia, X., Liang, Q., Ming, X., 2019. A full-scale fluvial flood modelling framework
based on a high-performance integrated hydrodynamic modelling system
(HiPIMS). Adv. Water Resour. 132.
https://doi.org/10.1016/j.advwatres.2019.103392
Xia, X., Liang, Q., Ming, X., Hou, J., 2017. An efficient and stable hydrodynamic
model with novel source term discretization schemes for overland flow and flood
simulations. Water Resour. Res. 53. https://doi.org/10.1002/2016WR020055
Xing, Y., Chen, H., Liang, Q., Ma, X., 2022a. Improving the performance of city-scale
hydrodynamic flood modelling through a GIS-based DEM correction method.
Nat. Hazards. https://doi.org/10.1007/s11069-022-05267-1
Xing, Y., Liang, Q., Wang, G., Ming, X., Xia, X., 2019. City-scale hydrodynamic
modelling of urban flash floods: the issues of scale and resolution. Nat. Hazards
96. https://doi.org/10.1007/s11069-018-3553-z
Xing, Y., Shao, D., Liang, Q., Chen, H., Ma, X., Ullah, I., 2022b. Investigation of the
drainage loss effects with a street view based drainage calculation method in
hydrodynamic modelling of pluvial floods in urbanized area. J. Hydrol. 605.
Xu, Z., Xiong, L., Li, H., Xu, J., Cai, X., Chen, K., Wu, J., 2019. Runoff simulation
of two typical urban green land types with the Stormwater Management Model
(SWMM): sensitivity analysis and calibration of runoff parameters. Environ.
Monit. Assess. 191. https://doi.org/10.1007/s10661-019-7445-9
Yan, Di., Liu, J., Shao, W., Mei, C., 2020. Evolution of Urban Flooding in China, in:
Proceedings of the International Association of Hydrological Sciences.
https://doi.org/10.5194/piahs-383-193-2020
Yang, H., Chen, Y., Zhou, W., Sun, Z., 2020. Optimal Scheduling of Conbined
Drainage Tanks Based on InfoWorks ICM Model. Shenyang Jianzhu Daxue
Xuebao (Ziran Kexue Ban)/Journal Shenyang Jianzhu Univ. (Natural Sci. 36.
165
https://doi.org/10.11717/j.issn:2095-1922.2020.03.21
Yang, L., Li, J., Kang, A., Li, S., Feng, P., 2020. The Effect of Nonstationarity in
Rainfall on Urban Flooding Based on Coupling SWMM and MIKE21. Water
Resour. Manag. 34. https://doi.org/10.1007/s11269-020-02522-7
Yu, C., Duan, J., 2014. Two-Dimensional Hydrodynamic Model for Surface-Flow
Routing. J. Hydraul. Eng. 140. https://doi.org/10.1061/(asce)hy.1943-
7900.0000913
Zeiger, S.J., Hubbart, J.A., 2021. Measuring and modeling event-based environmental
flows: An assessment of HEC-RAS 2D rain-on-grid simulations. J. Environ.
Manage. 285. https://doi.org/10.1016/j.jenvman.2021.112125
Zhang, H., Yang, Z., Cai, Y., Qiu, J., Huang, B., 2021. Impacts of climate change on
urban drainage systems by future short-duration design rainstorms. Water
Zhang, J., Zhang, M., Song, Y., Lai, Y., 2021. Hydrological simulation of the jialing
river basin using the mike she model in changing climate. J. Water Clim. Chang.
12. https://doi.org/10.2166/wcc.2021.253
Zhang, T., Peng, L., Feng, P., 2018. Evaluation of a 3D unstructured-mesh finite
element model for dam-break floods. Comput. Fluids 160.
Zhou, Q., Leng, G., Huang, M., 2018. Impacts of future climate change on urban flood
volumes in Hohhot in northern China: benefits of climate change mitigation and
adaptations. Hydrol. Earth Syst. Sci. 22. https://doi.org/10.5194/hess-22-305-
2018
Zhou, Q., Leng, G., Su, J., Ren, Y., 2019. Comparison of urbanization and climate
change impacts on urban flood volumes: Importance of urban planning and
drainage adaptation. Sci. Total Environ. 658.
https://doi.org/10.1016/j.scitotenv.2018.12.184
Zolghadr, M., Hashemi, M.R., Zomorodian, S.M.A., 2011. Assessment of MIKE21

model in dam and dike-break simulation. Iran. J. Sci. Technol. Trans. B Eng. 35.
166

Thesis Qian Li

Uploaded by

Document Information

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

Thesis Qian Li

Uploaded by

Copyright:

Available Formats

This item was submitted to Loughborough's Research Repository by the author.

Development of new high-performance simulation approaches for multi-level

A doctoral thesis submitted in partial fulfilment of the requirements

for the award of Doctor of Philosophy of Loughborough University

To overcome the high computational burden of hydrodynamic models for large-scale

I would also like to thank my colleagues and researchers in Hydro-Environmental Modelling

My research is funded by Newcastle University, Loughborough University and China

Acknowledgements .................................................................................................................. iii

Table of Contents ....................................................................................................................... v

List of Figures ........................................................................................................................ viii

List of Tables ............................................................................................................................ xi

Chapter 1 Introduction .......................................................................................................... 1

1.1 Background ................................................................................................................. 1

1.2 Research aim and objectives ....................................................................................... 7

1.3 Thesis outline .............................................................................................................. 9

Chapter 2 Literature review ................................................................................................ 11

2.1 Surface flow models .................................................................................................. 12

2.2 Drainage models ........................................................................................................ 18

2.3 Dual drainage models ................................................................................................ 20

2.4 Simulation of floods in underground space............................................................... 26

2.5 Acceleration methods in urban flood modelling ....................................................... 29

2.6 Summary ................................................................................................................... 32

3.1 Introduction ............................................................................................................... 34

3.2 New 1D-2D coupled drainage network model .......................................................... 34

3.2.1 Pipe model ......................................................................................................... 34

3.2.2 Junction model ................................................................................................... 41

3.3 Results and discussion............................................................................................... 46

3.3.1 Experimental test ............................................................................................... 46

3.3.2 Unsteady flow through different drainage settings ............................................ 50

3.3.4 A hypothetical drainage network system ........................................................... 57

3.3.5 Discussion .......................................................................................................... 59

3.4 Summary ................................................................................................................... 60

4.1 Introduction ............................................................................................................... 62

4.2 Dual drainage model ................................................................................................. 62

4.2.1 HiPIMS .............................................................................................................. 63

4.2.2 Drainage model .................................................................................................. 71

4.2.3 Coupling HiPIMS with the drainage model....................................................... 71

4.3 Implementation on single GPU ................................................................................. 73

4.3.1 Structure of the drainage model on a single-GPU device .................................. 74

4.3.2 Structure of HiPIMs-drainage coupled model on a single-GPU device ............ 75

4.4 Model validation and application .............................................................................. 76

4.4.1 Experiment Test ................................................................................................. 77

4.4.2 A semi-realistic test............................................................................................ 86

4.5 Summary ................................................................................................................... 91

Chapter 5 Multi-layer simulation of flooding into underground spaces ............................ 93

5.1 Introduction ............................................................................................................... 93

5.2 Multi-layer flood modelling framework ................................................................... 93

5.3 Validation of the multi-layer flood model................................................................. 99

5.3.1 Flood in an idealised urban catchment with underground space ....................... 99

5.4 Summary ................................................................................................................. 106

Chapter 6 Multi-GPU implementation and case study ..................................................... 108

6.1 Introduction ............................................................................................................. 108

6.2 Multi-GPU Architecture .......................................................................................... 108

6.3 Comparison against the single-GPU based model .................................................. 112

6.4.1 Study site.......................................................................................................... 116

6.4.2 Data and model setup ....................................................................................... 116

6.5 Results and discussion............................................................................................. 121

6.5.1 Model validation .............................................................................................. 121

6.5.2 Comparison against more traditional methods ................................................ 128

6.5.3 Computational efficiency ................................................................................. 130

6.6 Summary ................................................................................................................. 131

Chapter 7 Conclusions and future work ........................................................................... 134

7.1 conclusions .............................................................................................................. 134

7.2 Future work ............................................................................................................. 137