GIDPC.Green Infrastructure · Disaster Prevention
Projects
R&D · OptimizationSpatial Optimization · LID

Seocho-gu permeable-pavement AI spatial optimization (XGBoost + NSGA-II + SHAP)

Under a 10-year design storm, full (100%) permeable-pavement coverage reduces runoff by 24.8% — but 73% of that reduction is delivered by the first 30%, marking 30–50% as the cost-effective sweet spot.

Year
2026
Client
Internal R&D
Status
Under review (Climate Risk Management)
Stack
pySWMM · XGBoost · NSGA-II · SHAP · OSM · Copernicus DEM

Background

Urbanization expands impervious surfaces and intensifies urban flood risk. Permeable pavement is a proven LID measure — but where and how much to install, under a fixed budget, is a separate decision problem.

Approach

  • Automated geospatial processing (OSM + Copernicus DEM) builds EPA SWMM for 18 Seocho-gu sub-catchments
  • Scenario sweep across 0–100% coverage under a 10-year design storm
  • XGBoost surrogate (R² = 0.79) trained to replace expensive SWMM runs
  • NSGA-II multi-objective optimization yields the Pareto front (runoff reduction vs installation area)
  • SHAP quantifies per-sub-catchment influence

Key Results

| Metric | Value | |---|---| | Baseline runoff | 614,298 m³ | | Runoff reduction at 100% coverage | 24.8% | | Share delivered by first 30% coverage | 73% | | Cost-effective sweet spot | 30–50% | | Pareto runoff reduction range | 14.7–22.7% | | Pareto installation area range | 3.8–63.5% | | SHAP highest-impact sub-district | Naegok-dong (large catchment) |

Implications

  • Permeable pavement exhibits strong diminishing returns — the planning message is "30–50% is optimal," not "the more, the better."
  • Influence is non-uniform across sub-catchments; SHAP identifies priority investment locations.
  • The workflow is reproducible, scalable, and interpretable — a deployable methodology for LID spatial optimization.

Related solutions

  • Spatial Optimization of Green Infrastructure — core workflow used here
  • Stormwater Risk Analysis (SWMM · Neural ODE) — underlying simulation foundation