Author(s): Roberto Bentivoglio; Alexander Garzon; Zoran Kapelan
Linked Author(s):
Keywords: Urban drainage networks; Deep learning; Surrogate model; Metamodels
Abstract: Urban drainage models are computationally demanding, especially for uncertainty analyses, optimization and real-time evaluations. Deep learning surrogates are increasingly explored as an alternative due to their speed and accuracy. However, there are no guidelines on which models are most suited as metamodels. This study evaluates four architectures, namely Multilayer Perceptrons (MLP), Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), and Graph Neural Networks (GNN), trained to predict node hydraulic heads and conduit flow rates over time for a real urban drainage system, under multiple rainfall scenarios. Results indicate that LSTMs provide the most accurate results with mean absolute errors for hydraulic heads and flow rates of 0.32cm and 0.04 L/s and execution time 10 times faster on CPU than the physically-based numerical model, highlighting their suitability as surrogates.
Year: 2026