GIDPC.Green Infrastructure · Disaster Prevention
Solutions

Cloudburst · Urban Flooding

Stormwater Risk Analysis (SWMM · Neural ODE)

EPA-SWMM coupled with Green-Ampt and a physics-informed Neural ODE captures continuous soil-moisture dynamics — overcoming the structural limits of static AMC assumptions.

EPA-SWMMpySWMMGreen-AmptNeural ODEERA5-LandPyTorch

What this solves

The largest structural error source in urban stormwater modeling is the three-class AMC framework — a categorical proxy inherited from 1972 design practice. In East-Asian monsoon climates, soil moisture varies sharply year-round; compressing that into three bins makes it impossible to distinguish LID intervention effects from model error.

How we analyze

| Step | Tools | Output | |---|---|---| | 1. Urban sewer network + sub-catchments | GIS, OSM, EPA-SWMM | Real-conduit model | | 2. Rainfall data | KMA AWS / ERA5-Land | Hourly rainfall + antecedent moisture | | 3. Infiltration modeling | Green-Ampt + dynamic IMD coupling | Continuous infiltration flux | | 4. Soil-moisture dynamics | Physics-informed Neural ODE | Learned dθ/dt, R² = 0.983 | | 5. LID scenario evaluation | pySWMM + XGBoost-SHAP | Peak, volume, lag-time metrics |

Demonstrated case

For 47 km² of Seoul Seocho/Gangnam:

  • All 21 storm events exceed the standard AMC III boundary — empirical proof of structural misfit
  • Dynamic AMC coupling quantifies 39.6% annual / 33.8% peak runoff reduction at Mixed LID 40%
  • 95-second training + 0.01-second inference — usable in real-time decision pipelines

Typical deliverables

  • SWMM model of your site's urban sewer network
  • Calibrated antecedent-moisture dataset (ERA5-Land based)
  • Quantitative LID scenario comparison (millimeter-scale reduction)
  • Neural ODE training results + reproducible pipeline