Background — Why this project
EPA SWMM and most operational urban hydrology models still represent antecedent moisture condition (AMC) as a discrete three-class proxy inherited from 1972 design practice. Compressing a continuous variable into three bins becomes a major error source in East-Asian monsoon climates where rainfall is highly concentrated.
For Seocho-gu (2018–2023), pre-storm volumetric water content extracted from ERA5-Land shows that the canonical AMC II value (θ₀ = 0.170 m³/m³) underestimates observed moisture in every one of 21 events, with a mean bias of 0.224 m³/m³ (RMSE 0.238 m³/m³). Translated through Green-Ampt, that bias amounts to ~21 mm of systematic infiltration overestimation per design storm — about 984,480 m³ of unaccounted runoff across the 47 km² domain, exceeding the prior baseline runoff estimate by 60%.
Approach
The project consists of two modules.
1. GE-Water — distributed urban hydrology engine
- Coupled Green-Ampt infiltration with 2D diffusion-wave surface routing
- Continuous θ(t, y, x) tracked on a 50 m grid
- Generates training data internally — no dependency on proprietary models
2. Physics-informed Neural ODE
- Neural network learns dθ/dt as a function of antecedent state and real-time rainfall
- Trained on 21 observed events × 4 initial moisture conditions = 84 synthetic event simulations
- Training time: 95 seconds
Key Results
| Metric | Value | |---|---| | Validation R² | 0.983 | | MAE | 0.56% (of mean θ) | | Per-event-size MAE | 0.0003 – 0.0060 m³/m³ | | Dry-condition MAE | 0.0004 m³/m³ | | Wet-condition MAE | 0.011 m³/m³ | | Events exceeding AMC III boundary | 100% (21 / 21) |
Structural implication: All 21 analyzed events exceed the AMC III boundary, meaning the three-class system itself is unsuited to Seoul's monsoon climate. This is not a calibration issue — it is a model-structure issue.
What this means for decision-makers
- AMC assumptions alone shift peak flow by −5.6% to +6.2% — equivalent to the impact of 10–15% LID coverage.
- Comparing LID cost-effectiveness under a static AMC assumption conflates intervention effect with model error.
- The framework runs in 95 s of training + 0.01 s of inference — usable for both real-time decision support and ex-ante LID design evaluation.
Related solutions
- Stormwater Risk Analysis (SWMM · Neural ODE) — the core technology stack used here
- Spatial Optimization of Green Infrastructure — Neural ODE outputs feed directly into NSGA-II inputs