Conference Agenda

For decades numerical modelling has been a cornerstone for achieving accurate flood simulations, however, such numerical methods present a severe trade-off between accuracy and computational expense. In recent years, simulation approaches based on neural networks have made large strides in all branches of science enabling fast simulations while also generalizing to permutations of the forcing terms. Several applications for the simulation of hydraulics exist. These include dam break scenarios, river flooding, or flows through water distribution and sewer networks. In this paper, we develop and validate a fully convolutional neural network architecture based on the U-net architecture for the spatio-temporal simulation of pluvial flooding in scenarios where the rainfall has spatial and temporal variation, using various different terrains. This case distinguishes itself by the accumulation of very small surface flows across large areas, resulting in an extremely skewed distribution of water depths. We demonstrate neural network surrogates that achieve CSI scores over 0.8, considering terrains and rain events that were not included in the training data. A conservative estimate for computation speed-up compared to the numerical model is factor 70. Further work will focus on integrating physical constraints and small-scale terrain variations into the architecture.