• unlimited access with print and download
    $ 37 00
  • read full document, no print or download, expires after 72 hours
    $ 4 99
More info
Unlimited access including download and printing, plus availability for reading and annotating in your in your Udini library.
  • Access to this article in your Udini library for 72 hours from purchase.
  • The article will not be available for download or print.
  • Upgrade to the full version of this document at a reduced price.
  • Your trial access payment is credited when purchasing the full version.
Buy
Continue searching

Monte Carlo based flood risk analysis using a graphics processing unit-enhanced two-dimensional flood model

ProQuest Dissertations and Theses, 2011
Dissertation
Author: Alfred J Kalyanapu
Abstract:
The goal of this dissertation is to improve flood risk management by enhancing the computational capability of two-dimensional models and incorporating data and parameter uncertainty to more accurately represent flood risk. Improvement of computational performance is accomplished by using the Graphics Processing Unit (GPU) approach, programmed in NVIDIA's Compute Unified Development Architecture (CUDA), to create a new two-dimensional hydrodynamic model, Flood2D-GPU. The model, based on the shallow water equations, is designed to execute simulations faster than the same code programmed using a serial approach (i.e., using a Central Processing Unit (CPU)). Testing the code against an identical CPU-based version demonstrated the improved computational efficiency of the GPU-based version (approximate speedup of more than 80 times). Given the substantial computational efficiency of Flood2D-GPU, a new Monte Carlo based flood risk modeling framework was created. The framework developed operates by performing many Flood2D-GPU simulations using randomly sampled model parameters and input variables. The Monte Carlo flood risk modeling framework is demonstrated in this dissertation by simulating the flood risk associated with a 1% annual probability flood event occurring in the Swannanoa River in Buncombe County near Asheville, North Carolina. The Monte Carlo approach is able to represent a wide range of possible scenarios, thus leading to the identification of areas outside a single simulation inundation extent that are susceptible to flood hazards. Further, the single simulation results underestimated the degree of flood hazard for the case study region when compared to the flood hazard map produced by the Monte Carlo approach. The Monte Carlo flood risk modeling framework is also used to determine the relative benefits of flood management alternatives for flood risk reduction. The objective of the analysis is to investigate the possibility of identifying specific annual exceedance probability flood events that will have greater benefits in terms of annualized flood risk reduction compared to an arbitrarily-selected discrete annual probability event. To test the hypothesis, a study was conducted on the Swannanoa River to determine the distribution of annualized risk as a function of average annual probability. Simulations of samples of flow rate from a continuous flow distribution provided the range of annual probability events necessary. The results showed a variation in annualized risk as a function of annual probability. And as hypothesized, a maximum annualized risk reduction could be identified for a specified annual probability. For the Swannanoa case study, the continuous flow distribution suggested targeting flood proofing to control the 12% exceedance probability event to maximize the reduction of annualized risk. This suggests that the arbitrary use of a specified risk of 1% exceedance may not in some cases be the most efficient allocation of resources to reduce annualized risk.

TABLE OF CONTENTS

ABSTRACT……..……………………………………………………………… iii LIST OF TABLES...…………………………………..…………………… …… vii ACKNOWLEDGEMENTS……………………………………………………... viii

Chapter

1.

INTRODUCTION …………………………… .…………………………

1 1.1 Background……………………………………………………… 1 1.1.1 Flood Impacts……………………………………………. 1 1.1.2 Flood Modeling….…………………… …………………. 4 1.1.3 Uncertainty in Flood Modeling………………………….. 9 1.2 Problem Statement………………………………………………. 11 1.3 Research Objectives……………………………………………... 13 1.4 Overview of Dissertation ……………………………………….. 13

2.

GRAPHICS PROCESSING UNIT-BASED TWO-DIMENSIONAL FLOOD MODEL……..…………………………………………………..

15 2.1 Introduction……………………………………………………… 15 2.2 Methodology………………… ………………………………….. 21 2.2.1 Model Description……………………………………….. 21 2.2.2 Model Validation…………………………………………

26 2.2.2.1 Laboratory Scale Dam Break….………………. 26 2.2.2.2 Taum Sauk Dam Break………………………... 28 2.2.3 Model Speedup…………………………………………...

30 2.3 Results and Discussion…………………………………………... 31 2.3.1 Model Validation…………………………………………

31 2.3.1.1 Laboratory Scale Dam Break Simulation………

31 2.3.1.2 Taum Sauk Dam Break Simulation……………. 37 2.3.2 Model Speedup Calculation……………………………... 39 2.3.3 Effect of Domain Minimization…………………………. 41 2.3.4 Effect of Spatial Resolution……………………………... 42 2.3.5 Speedup Comparison to Other Studies…………………...

44 2.3.6 Limitations of GPUs……………………………………...

45

vi vi 2.4 Conclusions……………………………………………………… 47

3.

MONTE CARLO BASED FLOOD MODELING FRAMEWORK FOR ESTIMATING PROBABILITY WEIGHTED FLOOD RISK…………..

49 3.1 Introduction……………………………………………………… 49 3.2 Methodology…………………………………………………….. 53 3.2.1 Monte Carlo Analysis…………………………………….

53 3.2.2 Geospatial Output Analysis………………………………

56 3.2.3 Risk Map Development………………………………….. 57 3.3 Case Study……………………………………………………….. 60 3.4 Results…………………………………………………………… 63 3.5 Summary………………………………………………………… 69

4.

ANNUALIZED RISK ANALYSIS APPROACH TO RECOMMEND APPROPRIATE LEVEL OF FLOOD CONTROL……………… ……...

71 4.1 Introduction………………………………………………………..

71 4.2 Methodology……………………………………………………… 77 4.2.1 General Flood Damage Analysis Approach.…………….. 77 4.2.2 Flood Damage Approach in This Study…………………. 79 4.2.2.1 Input Flow Estimation…………………………. 79 4.2.2.2 Hydraulic Modeling…………………………… 82 4.2.2.3 Damage Modeling……………………………... 84 4.3 Case Study………………………………………………………... 87 4.4 Results and Discussion…………………………………………… 91 4.4.1 No Flood Proofing Alternative………………………....... 91 4.4.2 Flood Proofing Alternatives……………………………... 94 4.4.3 Capital Cost Estimates…………………………………... 97 4.5 Summary………………………………………………………….. 100

5.

CONCLUSIONS……………….....………………....…...………………

102 APPENDICES

A. FIRST-ORDER UPWIND NUMERICAL DISCRETIZATION OF THE ST.VENANT EQUATIONS……………………………………………..

107 B. IMPLEMENTATION STEPS FOR CPU AND GPU FLOOD MODEL FRAMEWORKS…………………………………………………………

115 REFERENCES…………………………………………………………………... 121

vi vi LIST OF TABLES

Table

Page

2.1 Comparison of the configuration of the GPU machines….…………… 26

2.2 Quantified statistics for laboratory scale dam simulation……………... 35

2.3 GPU speedup results for Taum Sauk dam break simulations for three sets of model iterations…………………………………………………

39

2.4 Execution times before and after domain optimization……………….. 42

2.5 Comparative summary of recent model speeds in the literature for storage cell (SC), dynamic wave (DYN), and diffusive wave (DIF) models………………………………………………………………….

45

3.1 Error matrix comparing the flood risk maps using deterministic approach (Map A) and deterministic approach (Map B)……… ..……..

65

4.1 Estimated annual damages calculated for the three cases……………... 97

4.2 Calculation of cost estimates for implementing flood proofing………. 98

ACKNOWLEDGEMENTS

I would like to express my gratitude to my research advisor, Dr. Steven Burian, for all his invaluable guidance, patience and continued support throughout my graduate study. I also sincerely thank him for giving me many opportunities that help me to gr ow professionally. I would also like to thank all my committee members, Dr. Timothy McPherson, Dr. Brian McPherson, Dr. Christine Pomeroy and Dr. Eric Pardyjak, for investi ng their time, effort and support. I would also like to thank all the fellow students and researche rs, past and present, that I have interacted with and befriended here at the Univers ity of Utah. Especially, I would like to thank Siddharth Shankar for his valuable collaborat ion in developing the GPU model, and Dr. Dave Judi for his expert suggestions during the dissertation. I would also like to thank my parents, my two sisters, my brother-in-laws, my brother and my sister-in-law for their support, thoughts and love throughout my gradu ate studies. I extend a special thanks to my cute nephews and two nieces for their love and affection. Finally, I would like to thank Tania, whose influence on me I cannot quantify, who was my inspiration to pursue my doctoral degree, who stood by me during the thick and thin, and believed in me even when I did not believe in myself.

CHAPTER 1

INTRODUCTION

1.1 Background

1.1.1 Flood Impacts

Knowing that more than 2.8 billion of the world’s population lives within 15 km of rivers (Small and Cohen, 2004), it is no surprise that floods are one of the more frequently occurring and higher impact natural disasters. Every year, on average, 196 million people in more than 90 countries experience to some degree cat astrophic flooding. From 1980 to 2000, floods resulted in more than 170,000 deaths worldwide, an average of nearly 9,000 deaths per year (UNDP, 2004). Although effort s have been directed towards flood management and control, the problem continues to wors en as populations in riparian and coastal areas grow, as more people move to r iparian and coastal areas, and as climate patterns change (Burby, 2001; McCar thy et al., 2001; Montz and Gruntfest, 2002). Flood risk mitigation is a major challeng e facing local, regional and global disaster management agencies (Levy et al., 2005). Mitigating flood risk can be achieved in today’s heavily urbanized wo rld by appropriately using the floodplains through floodplain management practices (Ahmad and Simonovic, 2006; Bedient et al., 2008). In the United States (US), flood management practices have been directed towards flood control, mainly through the

2 construction of levees by the US Army Corps of Engineers, especial ly with the passage of the Flood Control Acts of 1917 and 1936 (Wright, 2000). This has evolved significantly over the last 75 years with the inclusion of various s tructural flood control measures like detention basins, levees, dams and nonstructural measur es like flood proofing, permanent evacuation and relocation, land use management, flood hazard maps, hydrologic/hydraulic models, early warning systems, evacua tion procedures, building codes and National Flood Insurance Program (NFIP) (Levy et al., 2005; Ahmad and Simonovic, 2006). However, achieving complete protection from extreme flood events i s not technically feasible or economically viable, because one cannot pre dict the exact flood magnitude and frequency to be able to design flood control with appropriat e specifications, because flood events are inherently random and vary in s pace and time and also with changing climate. For example, during the great USA flood of 1993 in t he upper Mississippi River basin which was protected by at least 1572 federal and nonfederal levees along the river, 40 levees under federal operation and 1043 nonfederally operated levees failed or overtopped across the states of Minnesota, Illinois, Missouri and Nebraska, resulting in total estimated dama ges up to $20 billion and destroying more than 50,000 homes (Johnson et al., 2004). Thus, in the recent

years, there is shift in the perspective of flood policy from the concept of “flood protection” to “flood risk management” (Schanze, 2006). Flood protection aims at preventing flood hazards up to a certain flood magnitude by providing a certain protection level (e.g., protection against floods with a magnitude equivalent to a 1% exceedance probability). Flood risk managem ent, on the

3 other hand focuses on avoiding flood risk or minimizing the impacts of fl oods where flood damages cannot be avoided, by using a combination of floodplain manag ement practices. In the US, with this goal in mind, the US Army Corps of Engineers (USACE) established the National Flood Risk Management Program (NFRMP). The goal of NFRMP is to create a collaborative platform for all the fede ral, state and local agencies to develop strategies using a combination of floodplain management prac tices in reducing flood risk. One of the examples of flood risk management implementation in the US is the Swannanoa Flood Risk Management Project (SFRMP), i n the Swannanoa watershed in the state of North Carolina. Established in 2005, its obje ctive is to reduce the flood damages and risk of economic losses related to flooding in the Swannanoa valley area. Based on the preliminary engineering studies and c ommunity outreach, 50 potential flood risk management projects were identified, including flood warning stations, dam rehabilitation, emergency spillway reconstruction for hi gher discharge capacity, improvements to river approach and exit of bridges, and flood w all construction. More information about this project can be found at their w ebpage, http://www.swannanoafloods.org. Continued improvement in floodplain management practices and managing flood risk requires effective public education of risk, careful planning based on estimates of flood extent and reliable fast information exchange with the publi c during a flood event emergency (FEMA, 2002). The success of these actions is ba sed on accurate and rapid prediction of flood inundation, which is achieved by flood models (Bates et al., 2004; Bates et al., 2005). By improving the prediction of flood depths, velocities and flood inundation extent, one can apply these results with c onfidence in

4 the design of flood management alternatives like dams, storage ba sins, flood walls, levees, enhance the performance of early flood warning systems, unde rstand the impacts of urbanization and climate change on the increase in floods, and much more.

1.1.2 Flood Modeling

The advancement of computer models to simulate floods has had signi ficant impact on the ability to plan, forecast and respond to flood events (Hunt er et al., 2007). In the US, hydrologic and hydraulic models were first applie d to simulate floods in the 1960s (Crawford and Linsley, 1966), and have since been used to enhance engine ering design, planning, floodplain delineation and emergency response. These model s range from simple mathematical equations (e.g., rational method) to comple x and computationally challenging numerical solutions of partial differe ntial equations, also varying in predictive capacity. The current state of flood modeling involves using computer-based flood models that are freely available like the Hydr ologic Engineering Center’s River Analysis System (HEC-RAS), the National Ce nter for Computational Hydroscience and Engineering’s (NCCHE) CCHE1D, and the US Geol ogical Survey’s (USGS) Multi-Dimensional Surface-Water Modeling System (MD _SMS). There are also commercially available flood models like MIKE11, MIKE-FLOOD by DHI Group, FLO-2D by FLO-2D Software, Inc., etc., that are used in flood m odeling practice. Most of these models have the capability to integrate with Geographic Information Systems (GIS) and storing and managing spatial dat a related to floods, using software packages like ArcGIS ® , MapWindow and MapInfo ® .

5 Flood models typically simulate floods as free surface flows, us ing one- dimensional (1D) dynamic wave simulations, also known as Saint Vena nt equations. The 1D representation assumes that flow is parallel to the ri ver channel or floodplain, i.e., flow lateral to the main channel or perpendicular to the main channel flow dire ction does not occur. It represents the river channel and the floodplain as a series of cross- sections perpendicular to the flow direction. 1D models have been a commonl y used class of flood models as they are relatively simple to build and easy to operate (Fread, 1985). They are also computationally efficient and many engineer s consider their use acceptable to produce reasonably accurate surface water profil es (Buchele et al., 2006). Examples of 1D flood models include the US Environmental Protection Age ncy’s Storm Water Management Model (SWMM), the US Army Corps of Eng ineer’s HEC- RAS, MIKE11 by DHI Group, FLDWAV model by the US National Wea ther Service and, WSPRO developed by the USGS. Solving these models for a flood event on a river reach results in flood depths and velocities at all the cross-sec tions along the river. The flood depths can be interpolated to a water surface elevation, which c an be overlaid on a digital surface of ground elevations (e.g., a digital elevation model in an ESRI ® grid format, a triangular irregular network). Total flood inundation extent can be derived from all locations with water depths above the ground surface, as shown i n Figure 1.1 (Merwade et al., 2008). Even though 1D models are commonly used, the approach of 1D models is not always appropriate, especially when applied to floodplain flows. They f ail to simulate the lateral diffusion of the flood wave. And the discretization of the topography as cross-sections, instead of a continuous surface, is responsible for model uncertaintie s in

6

Figure 1.1: Flood inundation mapping process for 1D flood models

sinuous channels. The location of cross-sections is also subjective and m ay potentially affect the simulated extent of floods (Samuels, 1990). Flood inundation e xtent extraction through 1D models: 1) is not a seamless process and requir es postprocessing and 2) is subjective because water surface elevations are generat ed through interpolations (Bates and De Roo, 2000). Two-dimensional (2D) flood models eliminate these limitations, and va rious 2D numerical schemes have been developed in response (Zhang and Cundy, 1989; Lamb et al., 2009; Judi 2009). Applying 2D models enables higher order topographic representation in the simulations. A postprocessing step is not require d for calculating flood inundation extent because all the locations with flood depths collect ively form the

7 flood inundation extent (Bates et al., 1995). Moreover, with the increasing availability of high resolution and high accuracy Digital Elevation Models (DEM ) for floodplains areas, 2D models can be readily integrated with such data sources (Marks and Bates, 2000). Most researchers agree that, because flows in the floodplain significantly increase in complexity, models based on higher-order equations, such as 2D or 3D equations, should be used (Knight and Shiono, 1996; Bates et al., 1998). The recommendation of NRC (2009) towards Federal Emergency Management A uthority (FEMA) was to promote “greater use of 2D hydraulic models”, where needed by the floodplain topography, including preferential flood pathways and existin g and planned structures. A major limiting factor for applying 2D numerical flood models is their computational intensity (Lamb et al., 2009; Judi, 2009). Despite the advances in computer hardware and technology, it can still take a long time to run 2D models. For example, a Central Processing Unit (CPU) based 2D dynamic wave flood mode l using a first-order upwind finite difference numerical scheme was used to simulate a dam break event for a 62 km 2 area (624 x 1136 grid cell domain at 9.36 m spatial resolution), on a 2.33 GHz Intel™ Core2Duo® desktop with 2 GB RAM and Windows XP Professional

Operating System (basic system characteristics expected for m ost modelers). The model took 9.1 h to simulate approximately 15 min of flood wave. This limits the number of scenarios, spatial extent and/or level of detail that is expecte d for a particular flood problem, especially in the development of Monte Carlo based risk asses sment methods for flood modeling (Sayers et al., 2000; Buijs et al., 2003; Lamb et a l., 2009; NRC, 2009). They require multiple scenario simulations that tremendously incr ease the

8 computational intensity in generating flood depths, velocities, extents etc. For discussion’s sake, let us consider a flood emergency and evacuation stud y for the above-mentioned dam break simulation. The lead time for flood warning is very small, maybe around 2 – 3 hours, but it takes 10 hours to run the numerical model. So, m odel application would not be sufficient for emergency and evacuation study. Computational intensity is further affected by using higher resolut ion digital topographic data, and by applying models at regional scales. Bates and de Roo (2000) found that performance of models with a 100 m grid deteriorated signi ficantly when compared to a 25 m grid, and this was much less for a 50 m grid. So, t here is a difficult choice for the modeler, whether to use a lower spatial represe ntation and not take advantage of enhanced topographic data or to use a higher resolution but s uffer with computational intensity. In addition to computational intensity, flood estim ations from models that can be run quickly, while maintaining accuracy, can be use d in a more “near or better” than real-time fashion to include dynamic conditions and e nhance the emergency management and decision-making capability. 2D flood models have been implemented in a high-performance parallel

computing architecture including Flo2DH by Hluchy et al., (2002), RMA by Rao (2005), CalTWiMS by Pau and Sanders (2006), TRENT by Villanueva and Wrig ht (2006), FloodMap-Parallel by Yu (2010) and LISFLOOD-FP by Neal et al . (2009). However, converting existing codes to execute on a parallel programming comput er cluster is a complex process requiring significant programming effort (Tran and Hluchy, 2004), and high-performance computing infrastructure is typically limited to academic and

9 government institutions, limiting their availability to private consul ting and engineering firms. There is a need for a simpler and a cost effective way to build 2D flood mode ls.

1.1.3 Uncertainty in Flood Modeling

Numerical models are but conceptualizations of reality and hence reduce physical complexity by simplifications through the systems of e quations (Wagener and Gupta, 2005). The model parameters, input data and model structure are cal ibrated with observed data, and when needed, the parameters are altered or for ced (within their acceptable ranges) to fit the model simulations to the observed d ata, and the validated model is used for prediction, as illustrated in Figure 1.2. However, prior estimation of feasible ranges of parameters does not guarantee the model prediction within a close range of observations, especially when it is extrapolated to other problem locations, like in the case of the two red dots outside the observation data in Figure 1.2. The lack of correlation between model par ameters and physical floodplain characteristics results in significant unce rtainty in prediction, especially if the model is extrapolated to predict the system behavior at a different location and/or flood event. Flood model uncertainty is thus of critical concern when modeling res ults are used to set policy, decision making and emergency planning. Failure to acknowledge uncertainty could result in wasteful overdesign of flood protection/mi tigation systems, or could lead to inadequate preparation for potential situations and even fa ilure of these systems.

10

Figure 1.2. Typical approach for model calibration, validation and prediction ( adapted from Smith, 2007 )

Thus, a careful and detailed calibration and uncertainty analysis is required for a successful application of flood inundation models (Duan et al., 1992; Beven and Binley, 1992; Yang et al., 2008). For the past two decades, several studies on unce rtainty analysis of models have been conducted, including Aronica et al. (2002), Bates et al. (2004), Werner et al. (2005), Pappenberger et al. (2005), studying the effe ct of surface roughness coefficient, Aronica et al. (1998), the grid cell size, and Purvis et al. (2008) for flow characteristics etc. However, the common representation of flood model simulation results re mains a deterministic flood inundation map based on a single simulation, or at best, a few scenarios (Apel et al., 2006). These limitations in the analysis have been due to the lack of data, and lack of higher dimensional modeling capabilities. Unfortunat ely, these deterministic approaches rely on the use of a single or limite d parameter sets and does not account for the uncertainties in the modeling process (Bates et al., 2004) and may

11 lead to an inaccurate hazard assessment (Di Baldassarre et al., 2010). To remedy the shortcomings of a single deterministic simulation of a floodplain, pr obabilistic modeling approaches are emerging. Probabilistic flood mapping, using Monte Carlo framework, is designed to incorporate uncertainty from input data and model parameters, represent spatial and temporal risk, and present flood m aps in terms of probabilities and percentages (Aronica et al., 2002; Romanowicz and Beven, 2003; Bates et al., 2004; Hall et al., 2005; Pappenberger et al., 2006; Di Bald assarre et al., 2010). However, these approaches are mainly based on simple planar hydr aulic models and model implementation using the full shallow water wave equations ha s not been found in the literature. Incorporating a physically-based 2D hydra ulic model would improve the model evaluations because of the improved spatial represent ation and accuracy of flood depths and velocities compared to 1D and simplified 2D planar models (Bates et al., 2004). This would also aid in better understandin g the sources of flood risk that is essential for flood risk management.

1.2 Problem Statement Advancement of computer models to simulate floods has significant im pacts on the ability to plan, forecast and respond to flood events. 1D flood models ar e commonly used in flood modeling applications; some of them, like HEC-RAS, WSP RO etc., are even available for free. However, they are prone to limitations, including simplified topographic representation through cross-sections, interpolation in wate r surface elevation generation and problems simulating complex lateral floodplain flows, etc. 2D flood models alleviate these limitations with their ability to s imulate complex flows and

12 with higher order topographic representation. However, a major limiting factor for applying 2D numerical flood models is their computational intensity. H igh-performance computing facilities are available to only a few academic re search institutes and a few consulting firms because of tremendous financial costs related t o installing and maintaining these computing facilities. There is a need to sea rch for simpler and more cost-effective alternatives. Another critical challenge facing flood modeling is the uncertai nty involved in modeling floods. It is a significant challenge, especially wh en models are used in determining policy, decision making and emergency planning. Failure to i ncorporate uncertainty could potentially result in wasteful overdesign of flood prot ection/mitigation measures, or even worse, lead to inadequate preparation for flood protec tion/mitigation. Incorporating uncertainty is imperative for successful use of hydr ologic and hydraulic models. However, flood models are being operated in a determinist ic (or single simulation) fashion relying on the use of a single or limited param eter sets and they in general do not account for the uncertainties in the modeling proces s. Probabilistic modeling approaches are emerging to reduce this limitation but thes e approaches are mainly based on simple planar hydraulic models (Di Baldassarre e t al., 2010). Flood models using the full shallow water wave equations have not been found i n the literature. Thus, there is a need to use physically-based 2D flood models in a pr obabilistic framework and demonstrate the usefulness in flood risk management.

13 1.3 Research Objectives The goal of the proposed research is to improve flood risk management by enhancing flood model computational capability and incorporating flood model ing uncertainty and demonstrating the importance of acknowledging uncertai nty in flood risk applications. This is accomplished by (1) developing a computationally efficient Graphics Processing Unit (GPU) based 2D flood model by using an e fficient and robust upwind numerical scheme to solve the complete 2D Saint Venant equations , (2) developing a Monte Carlo based probabilistic flood modeling framework to i ncorporate data and parameter uncertainties and generate probability weight ed flood risk, and (3) applying the Monte Carlo based framework to study the benefits of implementing flood risk management alternatives.

1.4 Overview of Dissertation The proposed goals of the research and the details of the methods and r esults are presented in the following chapters. The research included the dev elopment of a computationally efficient GPU-based 2D flood model. It is explained i n the Chapter 2 and information is provided on the GPU framework, the development of a f lood model and the optimization of the model for computational performance. The flood model is validated by comparing results from a laboratory exercise to show its ability to accurately estimate flood depths and velocities. The model is then further val idated with a case study using high water mark data collected for the Taum Sauk dam break event. The computational improvement of the GPU-based flood model is observed, and the performance of GPU as a function of domain size and spatial resolution is explored.

14 Chapter 3 focuses on development of a new Monte Carlo based probabilistic flood modeling framework. The framework is used to incorporate data and pa rameter uncertainty and estimate probability weighted flood risk. The chapte r includes a description of the framework and an application to determine the flood risk of the Swannanoa River in North Carolina. The ability of the framework in pr oviding detailed flood risk compared to a deterministic flood modeling approach is demonst rated. Chapter 4 focuses on the application of the probabilistic flood modeling framework to the an alysis of flood risk management alternatives. The financial impact in ter ms of reduction in flood damages by implementing flood proofing for different design flood eve nts is studied. The annualized risk concept is used in the formulation of design alternati ves. Chapter 5 summarizes the results of the dissertation and presents the conclusions.

CHAPTER 2

GRAPHICS PROCESSING UNIT-BASED TWO-DIMENSIONAL FLOOD MODEL

2.1 Introduction

With more than 21% of the world’s population living within 30 km of the coas t (Gommes et al., 1997), catastrophic flood events continue to cause an increasing amount of casualties, economic impact and infrastructure damage. On avera ge, 196 million people in more than 90 countries are exposed to flooding each year (UND P, 2004), while in the United States (US), by 2005, flood damages increased to $ 6 bi llion per year (FEMA, 2002; Levy et al., 2005). While already significant, Pielke et al. (2002) have analyzed US flood damage statistics between 1934 and 2000, found them to be increasing and projected them to continue to increase because of population growth and migration patterns and changing storm event patterns. Floodplain management actions are implemented to prevent and mitigate flood impacts on humans, ecology and the economy (Ahmad and Simonovic, 2006). Flood management has evolved over the last 50 years with key areas of adva nce including development of flood hazard maps, hydrologic/hydraulic models, flood warni ng systems, evacuation procedures, and flood insurance programs (Levy et al., 2005). Continued improvement requires effective public education of flood risk, careful pl anning based on

16 estimates of flood inundation extent and reliable and fast informati on exchange with the public during an emergency (FEMA, 2002). The success of these actions is in part based on accurate and rapid flood modeling. Flood modeling can be used in various flood

Full document contains 140 pages
Abstract: The goal of this dissertation is to improve flood risk management by enhancing the computational capability of two-dimensional models and incorporating data and parameter uncertainty to more accurately represent flood risk. Improvement of computational performance is accomplished by using the Graphics Processing Unit (GPU) approach, programmed in NVIDIA's Compute Unified Development Architecture (CUDA), to create a new two-dimensional hydrodynamic model, Flood2D-GPU. The model, based on the shallow water equations, is designed to execute simulations faster than the same code programmed using a serial approach (i.e., using a Central Processing Unit (CPU)). Testing the code against an identical CPU-based version demonstrated the improved computational efficiency of the GPU-based version (approximate speedup of more than 80 times). Given the substantial computational efficiency of Flood2D-GPU, a new Monte Carlo based flood risk modeling framework was created. The framework developed operates by performing many Flood2D-GPU simulations using randomly sampled model parameters and input variables. The Monte Carlo flood risk modeling framework is demonstrated in this dissertation by simulating the flood risk associated with a 1% annual probability flood event occurring in the Swannanoa River in Buncombe County near Asheville, North Carolina. The Monte Carlo approach is able to represent a wide range of possible scenarios, thus leading to the identification of areas outside a single simulation inundation extent that are susceptible to flood hazards. Further, the single simulation results underestimated the degree of flood hazard for the case study region when compared to the flood hazard map produced by the Monte Carlo approach. The Monte Carlo flood risk modeling framework is also used to determine the relative benefits of flood management alternatives for flood risk reduction. The objective of the analysis is to investigate the possibility of identifying specific annual exceedance probability flood events that will have greater benefits in terms of annualized flood risk reduction compared to an arbitrarily-selected discrete annual probability event. To test the hypothesis, a study was conducted on the Swannanoa River to determine the distribution of annualized risk as a function of average annual probability. Simulations of samples of flow rate from a continuous flow distribution provided the range of annual probability events necessary. The results showed a variation in annualized risk as a function of annual probability. And as hypothesized, a maximum annualized risk reduction could be identified for a specified annual probability. For the Swannanoa case study, the continuous flow distribution suggested targeting flood proofing to control the 12% exceedance probability event to maximize the reduction of annualized risk. This suggests that the arbitrary use of a specified risk of 1% exceedance may not in some cases be the most efficient allocation of resources to reduce annualized risk.