HyPhyNet: Hydrology GNN with Structure Augmentation and Physical Knowledge Integration for Flood Forecasting

Flood forecasting in small catchments remains challenging compared to larger rivers due to sparse sensor networks and faster response dynamics. Purely physics-based models are computationally expensive for real-time forecasting, while purely data-driven models often struggle to generalize under limited observations and may produce physically inconsistent predictions. We propose HyPhyNet, a hydrology aware graph neural network (GNN) designed for flood forecasting in data-scarce environments. HyPhyNet augments the hydrological graph with auxiliary upstream nodes that represent ungauged contributing areas, enabling rainfall-only observations to inuence downstream predictions without requiring imputation of missing ow measurements. We further incorporate domain knowledge by deriving edge attributes from hydrological and topographic features to parameterize message passing to reflect physically meaningful water routing behaviour. Temporal dynamics are captured using a Transformer-based encoder that models nonlinear rainfall runo dependencies from historical observations and forecast rainfall forcing. We evaluate HyPhyNet on multiple flash-flood prone catchments in the UK with minimal sensor coverage. Results show that HyPhyNet consistently outperforms baseline GNN models, achieving improved predictive accuracy on both full test periods and flood event subsets, with particularly strong gains during extreme events. Ablation studies demonstrate that integrating graph structure augmentation, physically informed spatial propagation, and temporal attention mechanisms together provides an effective framework for flood forecasting in small, data-scarce catchments.