⚡ Grid Outage Forecaster

T2.3 · AIMS KTT Hackathon 2026 · Kigali, Rwanda
0.176
Brier Score
61.2
MAE (min)
2.79h
Avg Lead Time
Business:
💇 Beauty Salon
🧊 Cold Room
🧵 Tailor Shop
⚠️ OFFLINE MODE — Forecast last updated . Plan valid for 6 hours from generation. After 13:00 without refresh, treat HIGH-risk hours as confirmed. Call 0788-GRID for live status.
📈 Forecast
🔌 Appliance Plan
📱 SMS Digest
ℹ️ About
P(outage)
Uncertainty band
LOW risk <12%
MEDIUM 12–25%
HIGH >25%

Hourly Risk — click a cell to drill into plan

Hour 0 · 00:00

Hour 0
Shedding Logic: Luxury → Comfort → Critical (never shed during peak unless P > 0.50). Within category: lowest revenue shed first. Critical always ON during business peak hours.

📱 Morning Digest — Feature Phone SMS

Sent at 06:30 CAT. Max 3 messages × 160 chars. Works on any GSM phone. No internet required. Language: Kinyarwanda/English mix for maximum reach.

🔕 Offline Fallback Protocol
If no internet refresh by 13:00: Device shows last cached plan with a red ⚠️ staleness banner. Risk budget: plan valid for 6 hours from generation time. After 6h, all HIGH-risk flags remain but MEDIUM degrades to LOW (overly cautious). Maximum acceptable staleness before stopping to trust the plan: 8 hours. Owner sees: "PLAN STALE — use generator, call 0788-GRID."
🔊 Illiteracy Adaptation — Voice + LED Relay
Design choice: Colored LED relay board (3 LEDs per appliance slot).
🟢 GREEN = ON safe · 🟡 YELLOW = shed if load high · 🔴 RED = OFF now.
Board connects via GPIO to a ≈USD 8 ESP32 running cached plan. No reading required. Physical override switch lets owner override any LED. Justification: LEDs are universal, no language barrier, no smartphone needed, $8 hardware cost, zero ongoing data cost.

Technical Notes

Model
LightGBM classifier for P(outage) + regressor for E[duration | outage]. Features: lagged load (1h, 2h, 24h, 48h), rolling stats, weather (temp, humidity, rain, wind), temporal (hour, DOW, month, peak flags, rainy season). Training: 150-day window. Evaluation: rolling 30-day held-out.
Performance
Brier score: 0.1756 (naïve base rate = ~0.212)
Duration MAE: 61.2 min
Avg lead time on true outages: 2.79h
Inference latency: <300ms CPU
Retraining time: <10 min
Constraints Met
✅ CPU-only · ✅ <10 min retrain · ✅ <300ms serve
✅ 50KB static UI · ✅ Feature phone SMS digest
✅ Offline fallback protocol · ✅ Illiteracy adaptation
✅ 3 business archetypes · ✅ Critical-before-luxury rule
Hardest Trade-off
Chose LightGBM over Prophet: faster retrain, handles irregular time steps, natively supports tabular weather features. Trade-off: less interpretable seasonality decomposition. Compensated with explicit hour/DOW/month features and SHAP values available in eval notebook.