This project develops GNN-SWS, an end-to-end spatiotemporal graph neural network surrogate for real-time hydraulic prediction in urban stormwater systems. The work addresses a key limitation of earlier stormwater surrogates by learning directly from rainfall inputs rather than relying on precomputed runoff from a separate hydraulic model.

Key contributions

• Built an end-to-end GNN that jointly models rainfall-runoff generation and hydraulic flow routing
• Predicted both junction-level and conduit-level hydraulic states using a heterogeneous message-passing architecture
• Added physics-guided constraints and the pushforward trick to improve physical consistency and long-horizon forecasting stability

Impact

The resulting framework supports real-time stormwater prediction, flood-risk assessment, and resilient urban water-system operations while avoiding the cost of repeated SWMM-style simulations.