Surrogate modelling of urban pluvial flooding for real time- and planning applications

The Intergovernmental Panel on Climate Change (IPCC) projects future precipitation events to increase in frequency and magnitude. Coupled with growing urbanization, this will increase the risk of urban pluvial flooding. To mitigate these changes, either static adaptation, i.e. changes to the current infrastructure, or dynamic adaptation, where the usage of the existing system’s capacity is optimized, can be implemented. To evaluate performance of different measures, flood models play a key role in the decision-making process. These models are often computed for multiple scenarios, e.g. adaptation measures, climate, city development etc. Current state-of-the-art flood models are coupled hydrodynamic models, which dynamically couple the one-dimensional (1D) drainage network to a two-dimensional (2D) floodplain with bi-directional flows. However, the usage of this model type is limited due to their large computational demands, making them unfit for real time applications and multiple scenario planning. To overcome these issues, cheaper-to-run models may be used as substitutes, or surrogates.

The aim of this thesis is to develop fit-for-purpose surrogate urban pluvial flood models that are numerically more stable and with calculation speeds several orders of magnitude faster than current hydrodynamic models, and with comparable accuracy. This will enable better climate adaptation as the short computation times allow for multiple scenario planning and real time applications. To achieve this goal, three surrogate models were developed and tested in this thesis: a drainage-, surface- and coupled surrogate model.

The drainage surrogate model was developed by lumping the drainage network into compartments, in which the volume of water is modelled by in- and out-going discharges. Several model configurations were examined, and flow, volume and surcharge results were benchmarked against the 1D drainage hydrodynamic model MIKE URBAN. Even the simplest drainage surrogate models, using steady state volume-discharge points from MIKE URBAN, were able to emulate the results of MIKE URBAN. The models were robust to the choice of model structure and parameters. The largest influence was the compartment resolution, which could be increased to achieve better results. The Generalized Likelihood Uncertainty Estimation methodology was applied to identify acceptable models using identified critical indicators and limits of acceptability for three applications: real-time control, warning and planning.

The surface surrogate models dynamically routes input surcharge from manholes from a hydrodynamic drainage model through a pre-defined surface network consisting of watersheds, bluespots and flow paths. Resulting volumes and discharges were used to calculate maximum water levels in bluespots and flow paths, which was benchmarked against the 2D surface hydrodynamic model MIKE 21. The surrogate model was robust to the number of bluespots included in the surface network, although too few lead to an overestimation of flooding downstream due to exclusion of upstream storage capacity. The surrogate models were able to emulate the behaviour of the hydrodynamic model both in terms of spatial distribution and magnitude. Cell-by-cell indicators proved that the model outperforms similar static models for large steep catchments, whereas it struggles with divergent flows in flat areas.

The coupled surrogate model couple the drainage- and surface surrogate models by (i) distributing lumped surcharge to individual manholes using a percentage distribution, (ii) computing lumped inlet discharges from manholes within bluespots using the orifice equation and (iii) introducing transportation time in the surface network. Global transportation time parameters were not suitable for all catchment scales and lead to large deviations in one case. Out of the three surrogate models, results from the coupled model differed the most from the results of the hydrodynamic model, in this case the coupled 1D-2D hydrodynamic model MIKE FLOOD, but it was able to describe overall flooding patterns and damage costs.

The surrogate models all yield speed-ups of five-to six orders of magnitude. Hence, simulation time was reduced from hours and days to milliseconds and seconds depending on the models, catchment size and rain input.
Each of the surrogate models presents novel model features, but the largest contribution is the bi-directional coupling of two conceptual models, which allows for fast and dynamic simulation of interactions between the drainage network and surface floodplain. Initial results show that the developed surrogate models can serve as substitutes for urban flood applications where state-of-the art hydrodynamic model are unfit due to computational requirements. Only the coupled surrogate model is valid for all types of real-time and planning applications. The surrogate models presented are not part of a finished ready-to-use modelling toolbox; improvements and further testing of the models are still required before use. However, in their current state, they should mainly be applied for comparative studies and not as stand-alone models. The surrogate models have great potential for feeding into a climate adaptation framework as they allow for accurate large-scale flood modelling, multiple planning scenarios and real time applications.