BijliGuard- Vehari Load shedding Optimizer

Inspiration

Vehari, Punjab (pop. 1M+) experiences 8-12 hours of unscheduled load shedding daily. The official MEPCO schedule has <40% accuracy. As a developer and student, I lost 3-4 hours of productivity daily - online classes drop, commits fail, UPS dies without warning.

Talking to 20+ families in Vehari, the pattern was identical: farmers can't run tube-wells, students can't study, shopkeepers close early. The problem isn't electricity generation - it's information asymmetry. No real-time ground truth exists.

We realized: If Waze can crowdsource traffic, we can crowdsource electricity.

What it does

BijliGuard is a real-time, predictive load-shedding intelligence platform for Vehari.

1-Click Crowdsource:Users report "Light Gayi / Light Aayi" with location. Firebase Realtime DB updates live map in <100ms.
Live Vehari Grid Map: Leaflet-based heatmap showing Green (Power ON) / Red (Power OFF) zones across Vehari tehsils - Luddan, Mailsi, Burewala, Tibba Sultan Pur.
AI Prediction Engine: Time-series analysis on last 14 days of crowd data + MEPCO history. Outputs: "85% probability outage tomorrow 7-9 PM. Charge devices now."
Student/Work Mode:Optimization algorithm suggests best 2-hour window for high-power tasks based on predicted uptime and UPS battery curve.
MEPCO vs Reality Dashboard: Data integrity layer exposing delta between official schedule and actual ground reports.
PWA + WhatsApp Alerts: Works on 2G, offline-first, pushes alert 15 mins before predicted outage.

How we built it

Architecture (Production-Grade): Frontend: Next.js 14 (App Router) + TypeScript + Tailwind CSS + Framer Motion. Deployed on Vercel Edge with ISR for map tiles. Backend: Firebase Realtime Database for low-latency state, Firestore for analytics, Firebase Auth (anonymous for frictionless reporting). Maps: Leaflet + OpenStreetMap, custom tile layer for Vehari district boundaries. Prediction: Python microservice (FastAPI) hosted on Render - uses Exponential Moving Average + hour-of-day / day-of-week seasonality. Falls back to heuristic if cold start. Notifications: Twilio WhatsApp API + Web Push. PWA: Workbox for offline queue - reports sync when internet returns.

Data Pipeline: Crowd event -> RTDB trigger -> Cloud Function validation (spam filter, geo-fencing to Vehari bounding box 30.04N,72.35E) -> Aggregation job -> Prediction model retrain every 6 hours via Vercel Cron.

Challenges we ran into

Cold Start Problem:No historical data on day 1. Solved with synthetic MEPCO schedule + 72-hour bootstrapping with 15 test users in Vehari.
Spam & False Reporting:One user spamming could break map. Built trust scoring - user credibility decays if reports deviate from cluster consensus. Geo-fencing restricts to Vehari district.
Realtime on Low Bandwidth:Firebase RTDB