Development of new high-performance simulation approaches for multi-level urban flood modelling

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 ii confirm its predictive capability. The performance of the developed model is statistically assessed by Nash–Sutcliffe model efficiency coefficient (NSE) and R-squared (R 2 ). 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. 

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.