Weather Predictor

XGBoost + LightGBM

V2 runs two parallel model families. An XGBoost regression model handles temperature forecasting at +6h, +12h, and +24h horizons.

A LightGBM multiclass classifier predicts weather codes (Clear / Rain / Snow) at +3h, +6h, and +12h.

Both models are trained independently and run in parallel during each hourly worker cycle.

Training Dataset

Over 120,000 city-aggregated hourly observations spanning November 2023 to March 2026 across 6 Finnish cities — Oulu, Helsinki, Tampere, Turku, Rovaniemi, and Vaasa.

Data sourced from FMI Open Data via dual WFS endpoints — hourly for temperature, wind and pressure, and 10-min for cloud cover and dew point.

Train/val/test split uses a seasonal gap strategy — validation and test sets are both full Finnish winters from different years to ensure reliable evaluation of snow and rain detection.

Feature Set

• Pressure Tendency: 1h / 3h / 6h change — signals approaching frontal systems.
• Cloud Cover + Dew Point: From 10-min endpoint — critical for fog and snow detection.
• Wind Decomposition: U/V components derived from speed and direction.
• Coastal Distance: Static feature encoding proximity to coastline — differentiates maritime vs inland cities.
• Seasonal Encoding: Day-of-year sine/cosine captures annual temperature cycles.

Model Performance

Evaluated on a held-out winter test set (Nov 2025 → Mar 2026). Hover the buttons to see per-horizon metrics.

Hourly Worker

GitHub Actions triggers every hour

Each run fetches the last 8 hours of observations (for lag features), aggregates 10 stations into 6 cities, engineers features, and runs inference on the latest row per city.

FMI Harmonie NWP forecasts are fetched alongside and stored in MongoDB for side-by-side comparison on the dashboard.

Stack

MongoDB Atlas: Stores one document per city per hour. Each document contains current observations, our model predictions, and FMI Harmonie forecasts.

Azure App Service: Flask app served from a Docker container, reads latest city documents from MongoDB and renders the dashboard.

fmiopendata: Open-source Python library for WFS queries — github.com/pnuu/fmiopendata

Source Code: Full project available at github.com/Doug-Vo/Weather-Predictor