AI Olympiad: VoltPlan
Electric-vehicle charging demand prediction and infrastructure planning, built with team notokens for the Czech AI Olympiad. Second place in the Prague regional round, and a place in the national final.
// evidence
- result
- 2nd place
- demand MAE
- 24.03 kWh
- vs baseline
- −38.8%
- classifier
- 84.7% vs 38.7%
- rows
- 2.29M
- size
- 263 MB
// The task
Assignment AIO_PHA-02-PHA, Prague regional round. Three people, roughly four hours.
// Demand prediction
A LightGBM model predicting daily charging demand for 2030 at MAE 24.03 kWh — 38.8 percent below a population baseline learned from the training split alone.
// The number that got corrected
The competition submission reported 23.57 kWh and 40.0 percent. That version early-stopped on the same 517 zones it then measured MAE on, which means the tree count was selected on the measurement set.
Re-run on a fixed tree budget, the honest figure is 24.03 kWh and 38.8 percent. That is the number quoted here, and the competition number is not.
// Charger type classification
Classifying charger type reached 84.7 percent against a 38.7 percent baseline.
// Simulation and interface
An auditable historical simulation of managed charging across 2025, and a Streamlit dashboard reading real model outputs rather than mock data.
// Data
2.29 million rows of hourly grid and charging history, 263 MB, read lazily through Polars.