01
Temperature
Indoor and outdoor air, supply and return lines, and surface probes per zone.
Reference data for autonomous buildings
Train building AI on real buildings, not synthetic ones.
BuildingData.ai is a live multi-brand HVAC reference hotel — eight systems, fourteen brands, one hotel envelope, minute-level telemetry with labeled deviations.
8 parallel systems1-min telemetry
samsung · daikin · mitsubishi · vaillant
Hotel
anonymized occupancy use case
0
parallel HVAC systems
10 inverter pumps for water filtration and domestic water supply, air-water and air-air freshwater heat pumps, active ventilation
0
brands compared
>0 weeks
continuous 1-minute telemetry
Sensor coverage
3,000 sensors on 1,500 m² of floor space — indoors and outdoors.
Every monitored area carries up to five probes per category, and consumption is metered for every device and circuit, so each deviation traces back to a specific room, circuit, and sensor — not inferred from aggregate meters.
02
Humidity
Relative humidity indoors and outdoors, measured per climate zone.
03
Consumption
Electrical and thermal energy metered for every device and circuit via CT sensors, per system and per area.
04
Pressure
Differential and static pressure across air and water circuits.
05
Throughput
Air and water flow rates through every measured circuit.
06
Lux levels
Illuminance indoors and outdoors, recorded per area.
The reference building for autonomous HVAC AI.
One envelope, fourteen brands
Same Mediterranean hotel. Same anonymized occupancy. Fourteen different controllers. Isolate brand-specific drift on an identical physical load.
Labeled deviations, not raw noise
Ground-truth deviation labels for every anomaly: +166% office, +106% Samsung 1×3, −23% lounge, +24% Mitsubishi. Fine-tune with signal, not guesswork.
Data statement, not data lake
Every series ships with provenance, sensor calibration, and documented anomalies. GIFT-Eval compatible.
Specification data vs. real-time benchmark
Manufacturer specs meet anonymized hotel occupancy loads. Benchmark rated efficiency and comfort against minute-level behaviour on the actual building.
Platform
Ingest · Model · Serve
01
Ingest
Direct HVAC controller integration, weather and occupancy fusion, minute-level normalization.
02
Model
Deviation labels ready for fine-tuning Chronos-2, Moirai 2.0, Granite TinyTimeMixer, TSPulse.
03
Serve
Parquet + JSON schema, GIFT-Eval submission format, sample downloadable.
Data statement
BuildingData.ai treats data like an engineering artifact, not a byproduct. Every telemetry point is traceable to a calibrated sensor, every anomaly is labeled with its known cause, and every dataset ships with a written statement covering privacy, licensing, and open questions. If a foundation model trained on BuildingData.ai misbehaves, the reason lives in the data statement — not in the void.
What we offer
Ten data layers, one complete record of how the building behaves.
Energy-only feeds cannot explain an HVAC fault. Every BuildingData.ai dataset ships with all ten layers below, so models can reason about cause, not just correlation.
01
Asset and topology
Required fields
Building type, location/climate zone, floor/zone hierarchy, AHU/FCU/plant/equipment IDs, capacities, age, manufacturer/model, network relationships, design drawings or simplified topology.
Why it matters
Lets the model understand what a signal physically represents and reason about causal paths — e.g. which valve, pump or AHU can influence a specific thermal zone.
02
Semantic point model
Required fields
Original BMS point name, normalized tag, unit, entity/equipment association, datatype, sampling interval, min/max expected range, read/write status, quality flags.
Why it matters
Raw BAS names such as AHU3_SAT_SP are not reliable enough by themselves; semantic mapping is essential for cross-building learning.
03
High-resolution telemetry
Required fields
Timestamped values from temperatures, RH, CO₂, VOC/PM where available, flow, pressure, damper/valve position, fan speed, VFD speed, CT-measured power and energy, setpoints, modes, alarms and command states.
Why it matters
Supports forecasting, diagnostics, control analysis and explanation. Energy-only data is useful but insufficient for HVAC root-cause reasoning.
04
Control history
Required fields
Setpoint changes, schedules, overrides, commands issued, supervisory sequence version, controller parameters, manual versus automated action, actor/system identity.
Why it matters
A model must distinguish a fault from intentional operation; this is also the basis for learning safe operational recommendations.
05
Events and faults
Required fields
Alarm history, fault start/end, severity, affected assets/zones, standardized fault taxonomy, suspected and confirmed root cause, confidence, evidence links.
Why it matters
A labeled “known cause” is substantially more valuable than an unexplained anomaly.
06
Intervention outcomes
Required fields
Work orders, technician notes, inspection findings, repairs, parts replaced, recommissioning steps, time/cost, before/after KPI values, recurrence.
Why it matters
This turns anomaly detection into actionable diagnostic learning: what was done, did it work, and how do we know?
07
Occupancy and use
Required fields
Occupancy estimate, schedules, room/zone use, booking/events, tenant activity where lawful, special operating modes.
Why it matters
Prevents models from treating a genuine load change as a mechanical fault.
08
External context
Required fields
Local weather, solar irradiance, outdoor air quality, tariff/carbon intensity, holidays, building calendar.
Why it matters
Necessary for normalizing performance and comparing sites fairly.
09
Documents and knowledge
Required fields
O&M manuals, sequences of operation, commissioning reports, as-builts, BMS graphics exports, maintenance manuals, SOPs, audit reports, energy certificates.
Why it matters
Provides the language-and-engineering context needed for retrieval-augmented generation, troubleshooting assistants and agentic workflows.
10
Human feedback
Required fields
Operator query, model recommendation, accepted/rejected action, correction, rationale, safety escalation.
Why it matters
Allows preference learning and makes model performance improve against real operator judgement.
Use cases
Real-building data for energy management, BMS software and digital twins.
Building energy management systems
Validate BEMS forecasting, fault detection and optimisation against measured consumption from eight HVAC systems under anonymized hotel occupancy, with every anomaly labeled.
Building management system software
Test BMS software and controls logic across fourteen brands on one identical load. See how each controller behaves before you integrate it in the field.
Digital twin of a building
Calibrate and validate a building digital twin with minute-level telemetry, weather and anonymized occupancy from a real 640 m² Mediterranean hotel envelope.
Questions
What is BuildingData.ai exactly — dataset, platform, or product?+
Both. A research-grade reference building producing labeled multi-brand HVAC telemetry, and the software platform that ingests, labels and serves it.
Which brands and models are instrumented?+
Fourteen brands and models — including Airwoods, Samsung WindFree (1×1, 1×2, 1×3 wall units), Daikin, Mitsubishi, and Vaillant DHW, plus ESPA, Grundfos, and Generic inverter pumps — operate in parallel inside a single 640 m² Mediterranean hotel envelope: ten inverter pumps for water filtration and domestic water supply, air-water and air-air freshwater heat pumps, and active ventilation.
How dense is the sensor instrumentation?+
3,000 sensors cover 1,500 m² of floor space, indoors and outdoors. Every monitored area carries up to five probes per category — temperature, humidity, consumption, pressure, throughput, and lux levels — and consumption is metered for every device and circuit, so each deviation can be traced to a specific sensor, device, and location.
How is consumption measured, and how does it relate to operating state?+
Every device and circuit is metered with current transformers at minute resolution. Because control history — setpoints, schedules, modes and overrides — is captured alongside, consumption can be correlated directly with the infrastructure's operational mode: you can attribute a load change to an operating state change, not to noise.
Under what license is the data available?+
Research access under a permissive, attribution-required license. Commercial or product-integration licensing is handled bilaterally.
How does BuildingData.ai handle privacy for the hotel envelope?+
Telemetry describes equipment behaviour. Occupancy information is anonymized — aggregated and stripped of anything that could link a series to identifiable persons.
Can we get a sample before signing anything?+
Yes. A sample window with schema and data statement is shared on request, before any agreement.
How does BuildingData.ai differ from ASHRAE Great Energy Predictor, Building Data Genome, LBNL, and Trane/BrainBox customer telemetry?+
Those sources are either single-vendor, coarse-grained, or aggregated across heterogeneous buildings. BuildingData.ai holds the building constant and varies the controller — the exact setup needed to measure cross-brand generalization.
Can I use BuildingData.ai to test a building energy management system (BEMS)?+
Yes. Replay minute-level consumption, comfort and anonymized occupancy data from a real hotel through your BEMS logic and check whether it catches the documented, labeled deviations.
Is BuildingData.ai a digital twin of the building?+
It provides the ground truth a digital twin needs: one real envelope, eight HVAC systems, weather and anonymized occupancy. Use it to calibrate and validate your twin against measured behaviour.
Building an autonomous HVAC model? Talk to us about a dataset partnership.
Request access