Mitra Aelami

University of Montpellier, in progress

Supervisors: V. Guinot, G. Toulemonde; co-supervision: C. Delenne and R. Hostache.

ORCID; theses.fr/s403144.

“Neural network study of urban flood risk and the impact of extreme spatio-temporal rainfall events”

Abstract. Decision-making and flood risk assessment primarily rely on in-situ water level observations and weather forecasts. Hydrodynamic models are rarely used in practice due to their complexity and computational demands. While one-dimensional models are relatively fast, they are limited to modeling channeled flows and cannot accurately represent overflow events. Accurately considering the geometry of an area, especially in urban zones, necessitates a two-dimensional spatial model with mesh sizes of about one square meter, restricting simulations to the neighborhood scale. Despite the substantial computing power available today, real-time simulations for entire urban areas remain unattainable. Deep learning models, such as LSTMs (long short-term memory networks), have proven effective in hydrological contexts, for example, for rainfall-runoff modeling. However, very few studies have focused on flood prediction in terms of extent and water depth. One of the main challenges is the need for extensive training data, in a field where validation data is almost non-existent. The aim of this thesis is to develop an artificial intelligence model that primarily relies on the results of hydraulic models for training. Special attention will be given to adhering to the physical laws governing shallow water flows, namely the conservation of mass and momentum, placing this work in the realm of hybrid AI. Applying this model to the city of Montpellier will enable impact studies based on various scenarios.