Deep Learning for Modelling of Urban Drainage Networks: A Physics-informed Surrogate Model Using Measured and Simulated Data

City-wide climate adaptation for pluvial flood mitigation requires fast and reliable simulation tools. Considering the limitations of hydrodynamic models at city-scale simulations, data driven models have high potential in the development of surrogate tools. This study explores the Google DeepMind WaveNet™ model architecture to map hydrological response of catchments onto hydraulic parameters of the pipe network in a physically informed approach to deep learning. The WaveNet-based surrogate model successfully predicted hydraulic head and pipe flow in the network at average Normalized Nash-Sutcliffe Model Efficiency Indices of above 0.8, while boosting simulation speed by a factor of 1000. The developed AI model can be used for different assessment and optimization studies on the drainage network, thanks to its physics-informed structure.