Clustering networks: reducing the complexity of urban hydrology models with graph partitioning for fast and flexible simulations

Graph partitioning algorithms separate nodes of a graph into clusters, resulting in a smaller graph that maintains the connectivity of the original. In this study we use graph partitioning to produce reduced complexity sewer networks that can be simulated by a novel urban hydrology model. We compare a variety of algorithms, including spatial clustering, spectral clustering, heuristic methods and we propose two novel methods. We show that the reduced network that is produced can provide accurate simulations in a fraction of the time (100-1000x speed up) of typical urban hydrology models. We address some likely use cases for this approach. The first is enabling a user to pre-specify the desired size of the resultant network, and thus the fidelity and speed of simulation. The second is enabling a user to preserve desired locations that must remain in their own cluster, for example, locations with complex hydraulic structures or where monitoring data exists. The third is a case where detailed sewer network data is not available and instead the network must be simulated hundreds of times in a random sampling of network parameters, something that is only possible with the speed gains that our method allows. We envisage that this reduced complexity approach to urban hydrology will transform how we operate and manage sewer systems, enabling a far wider range of model applications than are currently possible, including optimisation and scenario analysis.