The standard inputs for a commercial building energy model are well established: building geometry, envelope thermal properties, weather data, lighting power density, plug load assumptions, and HVAC system parameters. Most energy simulation tools — from simplified spreadsheet models to detailed whole-building simulation engines — handle these inputs with reasonable precision.
What most of them get wrong is occupancy. Not because occupancy is technically hard to model — occupants generate sensible and latent heat loads, consume outside air for ventilation, and drive the demand timing that determines when HVAC stages up. The problem is that occupancy in real buildings doesn't behave the way the model assumes. And the gap between assumed occupancy and actual occupancy is a primary driver of HVAC demand forecast error.
The Static Schedule Problem
Building energy models and BMS control sequences typically represent occupancy as a fixed weekly schedule: occupied hours are Monday through Friday 7 AM–7 PM, unoccupied otherwise. This schedule is set at commissioning and rarely revisited. It reflects the design intent, not the operational reality.
In a multi-tenant commercial office building, actual occupancy on any given weekday might range from 35% of design capacity on a Friday in late summer to 110% of design capacity on a Monday conference day. That 3:1 ratio of actual-to-minimum occupancy has enormous implications for internal heat gain. At design occupancy (approximately 250 Btu/hr per person of combined sensible and latent heat), 500 occupants in a 100,000 sq ft building generate 125,000 Btu/hr of internal load. At 35% occupancy, that's 43,750 Btu/hr — a 65% reduction in a load category that drives cooling demand and ventilation requirements simultaneously.
An energy model with a fixed occupancy schedule treats every weekday as a design-occupancy day. The HVAC system staggers up for full occupancy even when the building has 180 people instead of 500. Demand charges get set on days when the model had no idea occupancy would be low — and could have staged down accordingly.
What Occupancy Data Sources Are Actually Available
The good news is that most commercial buildings already collect multiple streams of occupancy-correlated data. The challenge is integrating them into the energy model in real time, rather than treating them as separate operational systems.
Access control logs. Badge entry and exit data from electronic access control systems is the highest-fidelity occupancy dataset available in most office buildings. It records individual entry and exit times, allowing reconstruction of moment-by-moment headcount for any zone or floor that has controlled access points. The limitation is coverage — access control only captures programmatic occupancy at access-controlled entry points. It misses visitors, contractor staff, and occupants who tailgate through controlled doors.
Wi-Fi probe counts. Network access infrastructure logs the number of unique device probes per access point, which correlates with zone-level occupancy. It's coarser than access control — one person might appear as 2–3 devices, and the signal-to-occupant ratio varies by device type — but it provides real-time zone-level data without requiring physical sensor installation. Several BMS integration platforms provide Wi-Fi occupancy feeds from network infrastructure that can be consumed directly.
CO2 concentration sensors. Installed CO2 sensors in return air streams or zone-level locations provide an indirect occupancy signal through the relationship between occupant count and CO2 generation (approximately 0.3 L/min per person at moderate activity). CO2 concentration in a zone is a lagged indicator — it takes 10–20 minutes to respond to occupancy changes — but it's highly reliable and directly relevant to HVAC ventilation control because ASHRAE 62.1 minimum ventilation rates are tied to occupant count.
Calendar integration. Building management teams with access to space booking systems or corporate calendar tools can forecast occupancy ahead — not just observe it in real time. A conference center booking system that shows 400 registrants for a Thursday event is the highest-value occupancy forecast input available, and it's available days in advance. Calendar-based occupancy forecasts are less precise than sensor-based real-time counts but provide the advance signal needed for pre-conditioning decisions.
Occupancy as a Forecast Input vs. a Real-Time Input
There's an important distinction between using occupancy data to improve real-time HVAC control and using occupancy data to improve next-day demand forecasts. Both matter, but they use different data streams and operate on different timescales.
Real-time occupancy control — reducing ventilation rates in unoccupied zones, adjusting zone setpoints based on current headcount — is a well-established HVAC optimization technique covered in ASHRAE Guideline 36. It reduces energy consumption and improves zone-level comfort. But it doesn't address demand charges, because by the time the control system observes current occupancy, it's reacting — not pre-positioning HVAC for the demand peak that's coming in 2 hours.
Demand charge management requires occupancy forecasting: knowing tomorrow's expected headcount before the building opens, so the pre-conditioning window can be sized appropriately. Charging up full thermal mass pre-cooling for a 35%-occupancy day is a waste of off-peak energy. Running standard pre-conditioning for a 150%-capacity conference day is insufficient. The pre-conditioning decision depends on tomorrow's occupancy estimate, which requires a predictive model — not a real-time sensor feed.
Real-time occupancy sensing improves comfort and energy efficiency. Occupancy forecasting is what reduces demand charges. They're different tools that solve different problems, and the building energy field often conflates them.
Building the Occupancy Forecast Model
A practical occupancy forecast model for commercial buildings doesn't need to be highly complex to be useful. The most important predictors are day-of-week (weekday vs. weekend, with Friday systematically lower in hybrid-work environments), week-of-year (summer Fridays and holiday-adjacent days systematically lower), and explicit event signals from calendar systems.
For a typical multi-tenant office building with 6 months of access control history, a regression model using these features can achieve occupancy forecast errors in the 15–25% range for next-day predictions. That's adequate precision for sizing the pre-conditioning window correctly — the difference between an 80%-occupancy day and a 50%-occupancy day represents about 75 kW of internal heat gain in a mid-size office building, which directly affects how deep and how long the pre-cooling window needs to be.
Improving occupancy forecast accuracy below 15% error requires additional signal sources — calendar integration, recurring event pattern detection, visitor management system data — and the incremental benefit in demand charge reduction needs to be weighed against integration complexity. In our pilot work, buildings with calendar integration typically achieve 8–12% better occupancy forecast accuracy than access-control-only models, and that accuracy gain translates to measurable reductions in both over-conditioning (wasted pre-cooling energy on low-occupancy days) and under-conditioning (missed demand charge reduction on high-occupancy days).
Privacy and Data Minimization
Occupancy data — particularly access control badge logs and Wi-Fi device tracking — raises legitimate privacy questions that building operators and tenants should take seriously. Individual-level location tracking within a building creates GDPR and CCPA obligations in many jurisdictions and can create employee relations issues if the data collection scope isn't clearly communicated.
For energy forecasting purposes, the relevant signal is aggregate headcount by floor or zone — not individual location tracking. Anonymized, aggregate-level occupancy data is sufficient for thermal demand forecasting and carries a substantially different privacy profile than individual movement tracking. BMS integrations that consume Wi-Fi probe counts or CO2 sensor readings in aggregate never touch personally identifiable information and can be scoped explicitly to exclude individual-level data.
The buildings that have been most successful in our pilot program are those where the facilities team clearly scoped their data collection to aggregate occupancy counts — floor-level or building-level — and communicated that scope to tenants in advance. When occupancy data collection is positioned as "we're measuring how many people are in the building so we can condition it correctly" rather than "we're tracking individual movements," adoption is straightforward.