Our school has its own AI model…
🎯
Forecast accuracy — today so far
❓ How to read this panel?
What we compare Every hour we compare two numbers…
The difference colours The difference (±X µg/m³) shows how far off the model was:
Green: within 1 µg/m³ — nearly perfect.
Amber: the model overestimated…
Red: the model underestimated…
Classes and the ✓ mark PM2.5 levels fall into classes…
The numbers on top Decision accuracy is the share of hours…
Features: hour of day, humidity, temperature…
🧪
Experiment result: which weather forecast helps more?

Our model predicts PM2.5 from the weather forecast…

Days of collection
14
Hours checked against the sensor
334
Pairs of predictions compared
3942
Predicted ahead by Forecast used today WeatherNext 3
1–3 h
4–6 h
7–9 h
10–12 h

Average prediction error in µg/m³ — lower is better.

The forecast we already use came out slightly better…
Why we do not say WeatherNext is worse

Because we measured something narrower than it looks…

Second: over these two weeks the air was unusually clean…

Weather data: WeatherNext 3 (Google DeepMind) and Open-Meteo… CC BY 4.0
📈
PM 2.5 — Last 24 hours
PM2.5 hourly readings.
🔥
Average pollution risk by hour — 2 months of data
Low risk
High risk
LeoAir · PSP Leonarda Piwoni · Szczecin, Polska · 2026 Sensor installation #122437 · data updated hourly