Deep learning models have historically been difficult to interpret or explain (aka black-box models) but have the potential to be improved by constraining them using process-based or first principles understanding of physical system processes. This approach takes advantage of the automatic differentiation capabilities of deep learning frameworks to simultaneously optimize the physics formulations describing dynamics in a system and their neural network operator approximations through shared parameters. Applications of this approach to water systems modeling have largely focused on rainfall runoff, groundwater, and large river and reservoir systems modeling. However, little work has been done on applying this approach to hydraulic routing in stormwater and wastewater collection systems that are characterized by networks comprised of built infrastructure including pipes, weirs, pumps, storage basins, etc. Here, we present experiments that employ this paradigm for solving the mass, momentum, and energy conservation routing equations over stormwater and wastewater collection systems using Graph Neural Networks (GNN). GNNs are a logical framework to apply to stormwater systems as collection system networks can be considered directed graphs. Results from a GNN hydraulic routing model are compared to analytical solutions and results from the Environmental Protection Agency’s Stormwater Management Model (SWMM) to investigate its ability to capture complex flow dynamics such as flow reversals and hydraulic jumps. This presentation will explore advantages of GNNs and highlight areas that need further development or investigation including handling of arbitrary shaped hydraulic networks and the heterogenous infrastructure that comprise stormwater and wastewater collection systems.
