Real time probabilistic inundation forecasts using a LSTM neural network
Fedde J. Hop, Ralf Linneman, Bram Schnitzler, Anouk Bomers, Martijn J. Booij
Abstract
Probabilistic inundation forecasts that consider the uncertainty in rainfall predictions are crucial for flood early warning systems to effectively manage and reduce potential risks posed by pluvial flood events. Timely generation of such forecasts is challenging with physically-based numerical models due to computational demands. In contrast, data-driven models have a relatively low computational cost and can generate results quickly, making them a promising alternative to overcome this issue. This study proposes a long short-term memory (LSTM) neural network that can predict inundation progression over time at a high spatial resolution The network is trained on 1600 hydraulic simulations conducted using a 1D2D hydraulic model. With the trained network, probabilistic inundation forecasts are generated by combining the deterministic inundation predictions of 50 ensemble members of the rainfall forecast. The model is successfully tested for temporally varying rainfall events in a rural study area, and can generate accurate probabilistic inundation forecasts within seconds.