Intelligent and risk-based early warning system for pluvial floods : Developing a framework to forecast pluvial flooding in real time combining hydrodynamic modelling and deep learning

  • Intelligentes und Risikobasiertes Frühwarnsystem für urbane Sturzfluten

Hofmann, Julian; Schüttrumpf, Holger (Thesis advisor); Jüpner, Robert (Thesis advisor)

Aachen : RWTH Aachen University (2022)
Dissertation / PhD Thesis

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2022


In recent decades, flood risk management has mainly focused on river and coastal flooding, developing reliable models and forecasting systems to mitigate the flood risk due to long-term events. However, the increasing number of heavy rain-induced urban flood events reveals a significant gap in existing early warning systems (EWSs). Also referred to as pluvial flooding, this flood type suddenly emerges in urban areas, causing unforeseen and highly dynamic inundations and posing a severe risk to civilian security and infrastructure. Current warnings of pluvial flooding are limited to forecasts of meteorological parameters or water level monitoring of individual streams. However, recent flood events have demonstrated that this information is insufficient to warn people against the local and terrain-specific risks of pluvial flooding. Although microscale flood models based on hydrodynamic methods provide detailed results, the computational times are too long for operational applications. Existing approaches require supercomputers or are based on simple data-driven models and, thus, do not meet the requirements of an effective and practical EWS. Furthermore, recent studies solely focused on predicting the flood hazard, whereby the vulnerability of urban infrastructure is either entirely overlooked or not sufficiently addressed. To overcome these limitations, this thesis develops a framework for an intelligent and risk-based EWS to predict pluvial floods in real time. For this purpose, new hydrodynamic (HD) modeling approaches and innovative deep learning (DL) techniques are investigated, validated, and combined to develop real-time operating pluvial flood models. Furthermore, taking the flood hazard as well as the damage potential of infrastructures into account, the envisaged EWS identifies the urban areas particularly affected by upcoming heavy rainfall events and issues object-precise flood forecasts. The risk-based EWS comprises a three-component model chain: a rainfall nowcasting system, a high-resolution flood model, and a damage potential model. As a first step, flood simulations could be successfully validated based on in-situ observations of a heavy rainfall event in the city of Aachen on the 29th of May 2018. Next to a good correlation of simulated and observed inundations, it was shown that GPU-supported HD models are not applicable for real-time operations. Consequently, two methodological approaches were conceptualized to overcome the computation time bottleneck. The first method builds on a multi-model concept, subdividing the persistent HD model into several urban sub-catchment (USC) models regarding the local risk. To reduce the computational demand of USC models, a real-time optimization technique was developed considering the user-specific requirements and resources. Finally, a control system (PFA operator) was programmed to activate and run several USC models in parallel and couple them with the rainfall forecasting systems. Performance tests demonstrated high accuracy and promising application for smaller urban areas. Nevertheless, drawbacks arose for large hydrologically complex regions. To overcome this issue, the novel deep learning model floodGAN was developed. The model combines two adversarial Convolutional Neural Networks that are trained on synthetic data generated from rainfall generators and HD models. FloodGAN translates the flood forecasting problem into an image-to-image translation task and therefore, for the first time, enabling translations of rainfall data into high-resolution hazard maps within seconds. Performance tests showed an speedup factor of 10^6 while maintaining high model quality and accuracy as well as good generalization capabilities for highly variable rainfall events. Finally, this thesis presents fundamental insights in innovative flood forecasting and real-time modeling of pluvial flood events. In combination with a first demonstrator, this work provides the proof-of-concept for a new generation of ML-based flood models, building the basement for an intelligent and risk-based EWS. In the future, the floodGAN model will be further developed in terms of accuracy and scalability and transferred for real-time modeling of fluvial flash flood events. In addition, further research is needed in the area of coupled precipitation-flood modeling as well as the communication process to quantify uncertainties within the warning effectively and derive targeted instructions for actions.