Technical Overview
The Heatvelo Forecasting Engine
A thermal demand forecast built from building-specific physics, not generic load profiles. Five stages from raw BMS historian and occupancy data to a BAS-ready staging schedule with per-window setpoint offsets and confidence scores.
From Raw Feeds to Staging Schedule
Data Ingestion
Connect occupancy sensor feeds, weather API, and BMS historian logs. No new hardware required. We support BACnet/IP, Modbus TCP, REST/JSON, and flat-file historian exports. Data is validated for completeness and frequency before entering the model.
Envelope Modeling
Heatvelo characterizes your building's thermal mass, glazing ratio, and orientation. This is the physics layer — it determines how quickly outdoor temperature changes propagate to HVAC demand. Calibrated per building from submitted envelope specs and validated against historical BMS data.
Occupancy Prediction
Historical patterns and calendar data project the 72-hour occupancy curve forward. Building-specific occupancy cycles (weekday/weekend, seasonal, event-driven) are modeled separately. Foot-traffic data from POS or entry counters can substitute for sensor occupancy in retail buildings.
Demand Curve Generation
The engine combines envelope thermal response with predicted occupancy and weather forecast to produce the 72-hour forward HVAC demand curve, with uncertainty bands at the 10th and 90th percentile. Updated daily at 11 PM local time for next-day and +48h/+72h windows.
Staging Schedule Output
Delivered as a JSON staging feed (REST API endpoint) or direct BMS push via BACnet/IP or Modbus TCP. The schedule is structured as discrete staging windows with setpoint offsets and confidence scores — interpretable by any facilities manager, overridable at any point.
What Data Heatvelo Uses
Six input categories. All from systems your building already operates. Zero additional sensors required.
BMS/BAS Sensor Feeds
Zone temperatures, setpoints, AHU runtime, chiller status. Pulled from historian via BACnet/IP or Modbus TCP at 15-minute intervals.
Weather Forecast APIs
Hourly outdoor dry-bulb temperature, dewpoint, solar irradiance, and wind speed from National Weather Service or commercial APIs. 72-hour window.
Historical Occupancy Logs
Access control badge data, WiFi device counts, CO2 sensor readings, or retail POS transaction counts as foot-traffic proxy. Minimum 6 months of history.
Utility Rate Schedules
Time-of-use tariff structure and peak demand window definitions from your utility. Used to weight the staging schedule toward highest-cost intervals.
Building Envelope Specs
Window-to-wall ratio, glazing type, wall construction assembly, roof insulation value, and building orientation. Provided once during pilot setup.
Calendar and Event Data
Building operating schedule, holidays, planned events, and semester calendars (for university buildings). Adjusts the occupancy prediction model for known deviations.
What You Receive
JSON Staging Schedule
A structured schedule delivered to a secure REST endpoint, updated daily. Each staging window specifies start/end time, setpoint offset in Celsius, and a confidence score. Parseable by any BMS controller with network access.
The JSON format is documented and versioned — your operations team can inspect, validate, and optionally override any window before it takes effect.
{
"building_id": "COP-TOWER-02",
"date": "2026-06-19",
"generated_at": "2026-06-18T23:00:00Z",
"staging_windows": [
{
"start_time": "06:00",
"end_time": "07:30",
"setpoint_offset_c": -1.5,
"confidence": 0.91
},
{
"start_time": "14:30",
"end_time": "16:00",
"setpoint_offset_c": -1.0,
"confidence": 0.84
}
]
}Direct BMS Integration
For buildings with network-accessible BMS controllers, Heatvelo can push staging offsets directly via BACnet/IP or Modbus TCP — no intermediate JSON parsing required on your end.
This approach works with any BACnet/IP or Modbus TCP-capable controller. We map the staging offsets to the specific AHU setpoint objects in your BMS during the pilot setup phase.
- BACnet/IP (ASHRAE 135)
- Modbus TCP (Function codes 03/06/16)
- REST/JSON webhook (any BMS with API support)
- Flat-file export (CSV) for manual entry workflows
All integration work is handled by Heatvelo during the pilot setup phase. No IT involvement required on the building side beyond firewall port access.
Forecast Accuracy
Mean Absolute Percentage Error on next-day demand forecasts, validated on held-out building data.
Acceptable planning accuracy for pre-cooling schedule preparation 48–72 hours ahead.
Office, retail, university campus, and mixed-use industrial-flex buildings across cold and mixed climates.
Accuracy is validated using time-series cross-validation on held-out building data — the model is trained on 80% of historical dates and tested on the remaining 20%, preserving temporal order to prevent look-ahead bias. MAPE figures are building-type averages; individual results vary based on occupancy data quality, sensor coverage, and the stability of building operating patterns. Buildings with irregular occupancy (events, construction, tenant turnover) typically show higher MAPE at longer horizons.
Heatvelo does not publish portfolio-average accuracy numbers because accuracy is building-specific. During pilot setup, we run a 30-day backtesting period on your historical data and share the MAPE breakdown before the live forecast begins.
Request pilot data sheet