Event-based spatio-temporal graph convolutional networks for rainfall–Runoff prediction in HUC12 watersheds

Abstract Graph neural networks are increasingly used in rainfall–runoff modeling, with sub-watersheds as nodes and edges describing how they are linked. Edges are usually based on geographic distance, but in canal-regulated basins water follows canals and control structures rather than proximity. Whether a drainage-based graph outperforms a distance-based graph has rarely been tested with all other modeling choices held fixed. This study develops the Event-Based Spatio-Temporal Graph Convolutional Network (E-STGCN), which couples graph convolution with a recurrent neural network trained only on storm events, and uses it as a controlled experiment for this comparison. The framework was applied to five adjacent 12-digit Hydrologic Unit Code (HUC12) sub-watersheds of the Kissimmee River–Canal C-41A system in south-central Florida, using 111–252 storm events per extraction setting at 30-minute resolution from a decade of South Florida Water Management District DBHYDRO records. A full factorial design of 72 configurations varies the graph (drainage topology or k-nearest-neighbor proximity), the recurrent network (Long Short-Term Memory, Gated Recurrent Unit, or SimpleRNN), the loss function, and the event-extraction rule, each with five random seeds. The drainage-based graph outperformed the distance-based graph in 23 of 24 scenarios (combinations of event set, recurrent network, and loss function), with an average advantage about twice the variation between repeated runs. No configuration outperformed persistence, a forecast that holds discharge at its last observed value, so the results compare design choices rather than demonstrate forecasting skill. The drainage-based advantage was largest in the three basins with the lowest discharge variability, although five basins are too few to confirm this pattern. Models trained with a logarithmic loss failed, and the three recurrent networks performed similarly. Overall, in this canal-regulated network, which basins are linked in the graph affected accuracy more than the choice of recurrent network.