Catching stuffy air six hours before people feel it
Research2025Random Forest · Time series · Python
Problem
Indoor environmental quality is managed after the fact. CO₂ climbs through a full lab, someone notices, ventilation responds, and the hour where it mattered has already passed.
Approach
Treat it as a short-horizon forecasting problem. Six hours ahead is long enough for a facilities team to act and short enough that the signal in recent sensor history still carries.
What I built
- A six-hour-ahead forecasting system built on random forests over campus IEQ telemetry.
- An evaluation harness covering 11 labs, scored against a persistence baseline for every target-lab pair rather than a single aggregate number.
Results
- Beat the persistence baseline on 73% of target-lab pairs.
- CO₂ forecast improvement of up to 43%.