Data driven real-time prediction of urban floods with spatial and temporal distribution

The increase in extreme rainfall events due to climate change, combined with urbanisation, leads to increased risks to urban infrastructure and human life. Physically based urban flood models capable of producing water depth maps with sufficient spatial and temporal resolution are generally too slow for decision makers to react in time during an extreme event. We present a surrogate model with high temporal and spatial resolution for real-time prediction of water levels during a pluvial urban flood. We used machine learning techniques to achieve short computation times. The recursive approach used in this work combines convolutional and fully coupled multilayer architectures. The database for the machine learning was pre-simulated results from a physically based urban flood model. The forcing input of the prediction is precipitation and the output is water level maps with a temporal resolution of 5 min and a spatial resolution of 6 x 6 meters. The prediction performance can be considered promising for testing the model in real operational applications.