AIMS-RIC/PUMCLID

🌡️ HeatAlert Accra — SolarCool CoolTree Early Warning System

Real-time heat stress early warning for Accra, Ghana Imperial-AIMS Global Fellows Programme 2026 | Group 3

What this does

• —Fetches live ERA5 + Open-Meteo weather data for Accra (5.60°N, 0.19°W)
• —Trains a 2-layer LSTM on 90 days of historical data at startup
• —Forecasts Wet Bulb Globe Temperature (WBGT) 48 hours ahead
• —Issues tiered alerts: 🟢 GREEN / 🟡 YELLOW / 🟠 ORANGE / 🔴 RED
• —Sends SMS alerts via Africa's Talking (Ghana mobile networks)
• —Sends email alerts via Resend
• —Manager dashboard for AMA and stakeholder management

How to use

1. 1.Open the app — sidebar shows on the left
2. 2.Click Train / Retrain LSTM — takes ~60 seconds (trains on real ERA5 data)
3. 3.Click Fetch forecast — loads live conditions for Accra
4. 4.View Dashboard tab for current alert and advice
5. 5.View LSTM Forecast tab for 48-hour prediction
6. 6.Add subscribers and broadcast alerts from the other tabs

API Keys (add as Hugging Face Secrets)

Go to your Space → Settings → Variables and secrets → add the above.

Data sources

• —Open-Meteo ERA5 archive — historical hourly data, free, no key needed
• —Open-Meteo forecast API — real-time + 72h forecast, free, no key needed
• —Africa's Talking — SMS delivery to MTN, Vodafone, AirtelTigo (Ghana)
• —Resend — transactional email, 100/day free tier

System architecture

ERA5 Archive (90 days) ──► Feature engineering ──► LSTM training
                                                         │
Open-Meteo Forecast ──────► Current conditions ──────► Predict 48h WBGT
                                                         │
                                              ┌──────────┴──────────┐
                                         Alert level           CoolTree network
                                              │                 pre-charge
                                    ┌─────────┴──────────┐
                                   SMS              Email
                               (Africa's Talking)  (Resend)

LSTM details

• —2-layer stacked LSTM, 64 hidden units per layer
• —8 input features: T2m, RH, wind, solar radiation, hour (sin/cos), day-of-year (sin/cos)
• —48-hour lookback window, 48-hour forecast horizon
• —Trained with Adam optimiser + cosine annealing LR scheduler
• —WBGT target computed via Stull (2011) wet-bulb formula

Group 3 — Imperial-AIMS GFP 2026