Occupancy Forecasting
Occupancy Forecasting 101: Training AI for Indoor Air Quality Optimization
A practical overview of how data, labels, and evaluation methods turn past room usage patterns into smarter real-time heat recovery ventilation decisions.
What “occupancy forecasting” means in buildings
In a commercial office, HVAC schedules often assume occupancy stays near a static pattern. Forecasting replaces that assumption by estimating, for each time window, how likely spaces are to be occupied and how densely people are present.
When paired with heat recovery ventilation, these predictions help the system choose ventilation rates that balance indoor air quality (IAQ), thermal comfort, and energy efficiency. Instead of reacting after air gets stale, the control logic can anticipate demand and adjust supply air and exhaust recovery earlier.
1) Start with the right training signals
- Occupancy proxies: badge swipes, desk booking logs, meeting room calendar events, motion sensor summaries, or CO₂ trend data when available.
- Temporal patterns: day-of-week, holidays, tenant-specific routines, and recurring meeting blocks.
- Weather and outdoor air: temperature, humidity, and forecast windows that influence ventilation demand and economizer behavior.
- Building context: floor layout, zone boundaries, and equipment constraints for each heat recovery ventilation unit.
2) Choose labels that match control decisions
Forecasting can predict a simple count, a probability of occupancy, or a categorical level (low, medium, high). For control, you want targets that map cleanly to ventilation setpoints and scheduling policies.
For example, if your IAQ strategy uses ventilation thresholds, it is often more practical to train the model to predict occupancy bands for each zone than to predict exact headcounts that may be noisy.
Training tip: align your label window with the control update cadence (e.g., 15-minute or 30-minute horizons) so your model outputs can be used directly.
3) Modeling approach: keep it interpretable first
You do not need the most complex architecture to see value. A strong baseline can outperform static schedules while remaining easier to validate across tenants and seasons.
- Seasonal baselines: learn day-of-week and holiday patterns and adjust with recent signals.
- Probabilistic models: estimate likelihood of occupancy bands to support risk-aware control.
- Sequence models: use past occupancy proxies and weather forecasts to capture time dependencies.
Whichever you choose, make sure the model can handle missing sensor data and can be audited at the zone level.
4) Evaluate like an operator, not a dashboard
Forecast accuracy matters, but the more important question is how forecasts affect ventilation decisions and IAQ outcomes.
Model metrics
- For probabilities: calibration, Brier score, or reliability curves.
- For bands: per-class precision/recall and confusion over horizons.
Control-aware metrics
- IAQ proxy outcomes (e.g., CO₂ excursions) relative to targets.
- Energy impact and ventilation variability across seasons.
Treat forecasts as inputs to a control system. Validate end-to-end behavior, especially for sudden changes like winter arrivals, tenant onboarding, or scheduling policy updates.
5) Deployment loop: retrain, monitor drift, and close the gap
Occupancy patterns shift. A model trained for last quarter may degrade when tenant behaviors change or when building operations update.
- Monitor drift: compare predicted occupancy bands versus observed proxies over time.
- Retrain on schedule: refresh model weights using the most recent reliable history.
- Use a safe fallback: blend forecasts with conservative defaults when confidence is low.
This is where AI scheduling becomes dependable. Over time, the scheduling policy learns the building’s real rhythms and supports predictive maintenance triggers when equipment performance deviates.
Practical next steps for building teams
- Define the occupancy forecasting horizon that matches your heat recovery ventilation control cycle.
- Pick occupancy labels that translate directly into ventilation strategies (bands or probabilities).
- Build a baseline model and validate it end-to-end with IAQ and energy metrics.
- Set up drift monitoring and a retraining cadence before scaling to more zones.
If you are building a subscription offering for commercial office buildings, these steps also guide onboarding. Tenants need fast results, consistent reporting, and clear visibility into how forecasts influence real-time system adjustments and monthly performance reports.