Author(s): Leon Frederik De Vos; Markus Reisenbuchler; Anna Kruspe

Linked Author(s): Markus Reisenbüchler

Keywords: Hydrodynamic Modeling; Artificial Neural Networks; Urban Flood; Extreme Events

Abstract: To tackle the problems of increasing intensity and frequency of extreme rain events and the corresponding damage caused by floods in urban areas the upcoming research project “Inno_MAUS” proposes a holistic approach for the prediction, assessment and management of extreme rain events. Inno_MAUS divides the management in specific sub tasks from (i) formation and prediction of rain events, (ii) estimating the runoff formation and potential of urban retention, (iii) simulating the runoff dynamics to (iv) the final damage to infrastructure and will apply these sub tasks on two different study areas. In this paper, we focus on the project sub task (iii) – simulation of runoff dynamics. The conventional and widely used approach for simulating runoff dynamics in urban areas is a 2D-hydrodynamical model. It is the aim of Inno_MAUS to evaluate whether an Artificial Neuronal Network (ANN) as a faster, simpler and probably even more accurate tool can be installed instead. In the first part of this paper, an overview of the current development or state of the art for the field of hydrodynamic modeling in urban areas is given. This will discuss the advantages and disadvantages of 2D-hydrodynamic models and specify the weaknesses of conventional models that can possibly be overcome by using artificial intelligence. Then, a brief summary of first applications of artificial intelligence in current hydrodynamic modeling is presented. Finally, the methodology of Inno_MAUS is highlighted: As the needed data for the training of such an ANN from past flash flood events are scarce, we propose a methodology based on surrogate data. Therefore, an accurate 2D-hydrodynamical model using Telemac2D for two study areas will be set up as a surrogate of the real world, which will generate a sufficiently large data set for training a neural network to accurately predict key values such as maximum water depth, time of flood peak or flooded area for extreme rain events. We intend to use multi-task Recurrent Neural Networks to be able to predict these dynamic variables over time based on a fusion of various input values, such as e.g. hydrological input data, surface roughness data, DEM, as well as a-priori terrain data. This approach allows to prove the concept of an urban flash flood ANN. In the final step of the project, we analyze the transferability of such a model from one study site to another, and develop approaches to improve this transfer step.

DOI: https://doi.org/10.3850/IAHR-39WC252171192022801

Year: 2022