FlightRisk AI

We built an AI system that predicts flight delays from weather using NOAA METAR and historical US flight delay data. Web app shows live METAR data, model predictions, and simple explanations.

Comment

Inspiration

Weather delays cost the airline industry billions yearly and disrupt millions of travelers. Traditional forecasts are hard to use for quick operational decisions. I built an AI system that turns METAR weather data into a simple 0–100 flight risk score to help travelers, airlines, and operations teams make faster calls.

What it does

FlightRisk AI is a full-stack system that: Fetches NOAA METAR in real time Scores 0–100 per US airport (0 = clear, 100 = severe) Predicts delays using 5 regional XGBoost models Serves results in a React web interface Supports 34+ airports across 5 regions

How we built it

Cleaned AUS METAR data Engineered 17 features (temp spread, gust factor, ceiling/visibility ratio, pressure change, time features) Split training for Northeast, PacificCoast, RockyMountains, CentralPlains, Southeast Combined weather with FAA delay data Integrated NOAA Aviation Weather API Handled VRB wind, haze/fog, missing ceilings Built React + Flask backend Added airport search, color-coded scores, live METAR details, regional model info Resolved import and path issues across projects

Challenges we ran into

1) Low R² Root causes: noise from non-weather delays; weak target scaling; limited January sample; lost hourly granularity Fixes: built weather-first target, filtered data, applied VarianceThreshold, tuned hyperparameters Result: Southeast R² 0.30, others 0.10–0.12 2) METAR edge cases VRB wind direction → defaulted to 180° AUTO reports missing ceiling → added fallback Haze not penalizing visibility → reduced visibility 70% when HZ/FG/BR 3) Cross-project integration

Accomplishments that we're proud of

5 regional models with R² 0.10–0.30 34+ airports supported Live METAR integration with robust parsing React + Flask full-stack deployment

What we learned

Technical XGBoost regression with feature importance and hyperparameter tuning Target design (weather vs general delays) Web APIs (NOAA Aviation Weather) Full-stack: Flask + React Domain METAR parsing and aviation weather Regional weather impact on delays

What's next for FlightRisk AI

Historical analysis dashboard Seasonal patterns and alert threshold tuning Auto-update models with new delay data

Built With

Share this project:

Updates