GIDPC.Green Infrastructure · Disaster Prevention
Projects
Education · Rural AIRural Climate AI

Rural micro-catchment LSTM rainfall–runoff 6-hour forecast demo

Built for a rural-leaders' lecture: a 5 ha virtual farm catchment where an LSTM ingests 24 hours of rainfall and soil moisture and predicts 6-hour runoff in 0.01 s — and verifies LID intervention effects.

Year
2026
Client
Rural leaders' lecture (2026-05-22)
Status
Demonstration / education (SNU GE Lab)
Stack
PyTorch · LSTM · Windows Scheduler · Kakao alerts · SWMM·SCS-CN

Background

Rural micro-catchments are highly vulnerable to flash flooding — peak flows rise quickly, threatening farmland and irrigation infrastructure. Physics-based models like SWMM and SCS-CN are accurate but parameter-heavy and computationally expensive, making them impractical for farmers to operate directly.

LSTM is a deep-learning model that learns temporal patterns — once trained, real-time prediction is essentially free. Think of it as a "village AI forecaster."

Demonstration Catchment

| Property | Value | |---|---| | Location | Virtual rural village (GE Lab demo) | | Area | 5 ha (~7 soccer fields) | | Land cover | Farmland 60% / Village 15% / Forest 25% | | Observation period | 3 yr synthetic (2025–2027), 1-hour intervals, 26,280 steps | | Inputs | Hourly rainfall (mm/hr), soil moisture (%) | | Output | 6-hour ahead runoff (m³/s) |

Key Results

| Metric | Baseline | LID Applied | |---|---|---| | NSE (1.0 = perfect) | 0.641 (adequate) | 0.880 (excellent) | | RMSE (m³/s) | 0.020 | 0.008 | | MAE (m³/s) | 0.004 | 0.002 | | PBIAS (%) | −9.3 | −1.8 | | Peak-flow error (%) | +17.1 | −3.6 |

LID Effects (observation-based)

  • Peak flow: 0.60 → 0.33 m³/s (−45%)
  • Annual runoff: 436,000 → 390,000 m³ (−10.4%)
  • Peak delay: +5 hours — buys decision time for pumps, drainage gates, evacuation

Field Deployment Scenarios

  1. Real-time 6-hour forecast — rain observation → LSTM inference (under 1 s) → Kakao-talk alert
  2. Ex-ante LID effect analysis — train and compare scenarios with vs without vegetated swales / retention ponds
  3. Decision support — pumping-station activation, drainage-gate operation, resident-evacuation timing
  4. Public-data integration — extendable with KMA AWS + Ministry of Environment hydrological monitoring

Limits and Extension Paths

  • Current demo uses synthetic data. Real village CSV can plug in with the same column schema for county-scale deployment.
  • Single-seed results → quantify uncertainty via ensemble / Monte-Carlo dropout
  • Input extensions: Antecedent Precipitation Index (API), groundwater level, satellite SAR soil moisture
  • Model extensions: Seq2Seq multi-horizon prediction, attention LSTM, GNN coupling
  • Operational integration: SWMM–LSTM hybrid (physics + data)

Related solutions

  • Rural Climate-Crisis AI · IoT Response — direct extension of this demo. Pilot-site partners welcomed.
Projects